perf/core: Split context's event group list into pinned and non-pinned lists
[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 static struct list_head *
293 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
294 {
295         if (event->attr.pinned)
296                 return &ctx->pinned_groups;
297         else
298                 return &ctx->flexible_groups;
299 }
300
301 /*
302  * Add a event from the lists for its context.
303  * Must be called with ctx->mutex and ctx->lock held.
304  */
305 static void
306 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
307 {
308         struct perf_event *group_leader = event->group_leader;
309
310         /*
311          * Depending on whether it is a standalone or sibling event,
312          * add it straight to the context's event list, or to the group
313          * leader's sibling list:
314          */
315         if (group_leader == event) {
316                 struct list_head *list;
317
318                 list = ctx_group_list(event, ctx);
319                 list_add_tail(&event->group_entry, list);
320         } else {
321                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
322                 group_leader->nr_siblings++;
323         }
324
325         list_add_rcu(&event->event_entry, &ctx->event_list);
326         ctx->nr_events++;
327         if (event->attr.inherit_stat)
328                 ctx->nr_stat++;
329 }
330
331 /*
332  * Remove a event from the lists for its context.
333  * Must be called with ctx->mutex and ctx->lock held.
334  */
335 static void
336 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
337 {
338         struct perf_event *sibling, *tmp;
339
340         if (list_empty(&event->group_entry))
341                 return;
342         ctx->nr_events--;
343         if (event->attr.inherit_stat)
344                 ctx->nr_stat--;
345
346         list_del_init(&event->group_entry);
347         list_del_rcu(&event->event_entry);
348
349         if (event->group_leader != event)
350                 event->group_leader->nr_siblings--;
351
352         update_event_times(event);
353
354         /*
355          * If event was in error state, then keep it
356          * that way, otherwise bogus counts will be
357          * returned on read(). The only way to get out
358          * of error state is by explicit re-enabling
359          * of the event
360          */
361         if (event->state > PERF_EVENT_STATE_OFF)
362                 event->state = PERF_EVENT_STATE_OFF;
363
364         /*
365          * If this was a group event with sibling events then
366          * upgrade the siblings to singleton events by adding them
367          * to the context list directly:
368          */
369         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
370                 struct list_head *list;
371
372                 list = ctx_group_list(event, ctx);
373                 list_move_tail(&sibling->group_entry, list);
374                 sibling->group_leader = sibling;
375         }
376 }
377
378 static void
379 event_sched_out(struct perf_event *event,
380                   struct perf_cpu_context *cpuctx,
381                   struct perf_event_context *ctx)
382 {
383         if (event->state != PERF_EVENT_STATE_ACTIVE)
384                 return;
385
386         event->state = PERF_EVENT_STATE_INACTIVE;
387         if (event->pending_disable) {
388                 event->pending_disable = 0;
389                 event->state = PERF_EVENT_STATE_OFF;
390         }
391         event->tstamp_stopped = ctx->time;
392         event->pmu->disable(event);
393         event->oncpu = -1;
394
395         if (!is_software_event(event))
396                 cpuctx->active_oncpu--;
397         ctx->nr_active--;
398         if (event->attr.exclusive || !cpuctx->active_oncpu)
399                 cpuctx->exclusive = 0;
400 }
401
402 static void
403 group_sched_out(struct perf_event *group_event,
404                 struct perf_cpu_context *cpuctx,
405                 struct perf_event_context *ctx)
406 {
407         struct perf_event *event;
408
409         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
410                 return;
411
412         event_sched_out(group_event, cpuctx, ctx);
413
414         /*
415          * Schedule out siblings (if any):
416          */
417         list_for_each_entry(event, &group_event->sibling_list, group_entry)
418                 event_sched_out(event, cpuctx, ctx);
419
420         if (group_event->attr.exclusive)
421                 cpuctx->exclusive = 0;
422 }
423
424 /*
425  * Cross CPU call to remove a performance event
426  *
427  * We disable the event on the hardware level first. After that we
428  * remove it from the context list.
429  */
430 static void __perf_event_remove_from_context(void *info)
431 {
432         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
433         struct perf_event *event = info;
434         struct perf_event_context *ctx = event->ctx;
435
436         /*
437          * If this is a task context, we need to check whether it is
438          * the current task context of this cpu. If not it has been
439          * scheduled out before the smp call arrived.
440          */
441         if (ctx->task && cpuctx->task_ctx != ctx)
442                 return;
443
444         raw_spin_lock(&ctx->lock);
445         /*
446          * Protect the list operation against NMI by disabling the
447          * events on a global level.
448          */
449         perf_disable();
450
451         event_sched_out(event, cpuctx, ctx);
452
453         list_del_event(event, ctx);
454
455         if (!ctx->task) {
456                 /*
457                  * Allow more per task events with respect to the
458                  * reservation:
459                  */
460                 cpuctx->max_pertask =
461                         min(perf_max_events - ctx->nr_events,
462                             perf_max_events - perf_reserved_percpu);
463         }
464
465         perf_enable();
466         raw_spin_unlock(&ctx->lock);
467 }
468
469
470 /*
471  * Remove the event from a task's (or a CPU's) list of events.
472  *
473  * Must be called with ctx->mutex held.
474  *
475  * CPU events are removed with a smp call. For task events we only
476  * call when the task is on a CPU.
477  *
478  * If event->ctx is a cloned context, callers must make sure that
479  * every task struct that event->ctx->task could possibly point to
480  * remains valid.  This is OK when called from perf_release since
481  * that only calls us on the top-level context, which can't be a clone.
482  * When called from perf_event_exit_task, it's OK because the
483  * context has been detached from its task.
484  */
485 static void perf_event_remove_from_context(struct perf_event *event)
486 {
487         struct perf_event_context *ctx = event->ctx;
488         struct task_struct *task = ctx->task;
489
490         if (!task) {
491                 /*
492                  * Per cpu events are removed via an smp call and
493                  * the removal is always successful.
494                  */
495                 smp_call_function_single(event->cpu,
496                                          __perf_event_remove_from_context,
497                                          event, 1);
498                 return;
499         }
500
501 retry:
502         task_oncpu_function_call(task, __perf_event_remove_from_context,
503                                  event);
504
505         raw_spin_lock_irq(&ctx->lock);
506         /*
507          * If the context is active we need to retry the smp call.
508          */
509         if (ctx->nr_active && !list_empty(&event->group_entry)) {
510                 raw_spin_unlock_irq(&ctx->lock);
511                 goto retry;
512         }
513
514         /*
515          * The lock prevents that this context is scheduled in so we
516          * can remove the event safely, if the call above did not
517          * succeed.
518          */
519         if (!list_empty(&event->group_entry))
520                 list_del_event(event, ctx);
521         raw_spin_unlock_irq(&ctx->lock);
522 }
523
524 /*
525  * Update total_time_enabled and total_time_running for all events in a group.
526  */
527 static void update_group_times(struct perf_event *leader)
528 {
529         struct perf_event *event;
530
531         update_event_times(leader);
532         list_for_each_entry(event, &leader->sibling_list, group_entry)
533                 update_event_times(event);
534 }
535
536 /*
537  * Cross CPU call to disable a performance event
538  */
539 static void __perf_event_disable(void *info)
540 {
541         struct perf_event *event = info;
542         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
543         struct perf_event_context *ctx = event->ctx;
544
545         /*
546          * If this is a per-task event, need to check whether this
547          * event's task is the current task on this cpu.
548          */
549         if (ctx->task && cpuctx->task_ctx != ctx)
550                 return;
551
552         raw_spin_lock(&ctx->lock);
553
554         /*
555          * If the event is on, turn it off.
556          * If it is in error state, leave it in error state.
557          */
558         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
559                 update_context_time(ctx);
560                 update_group_times(event);
561                 if (event == event->group_leader)
562                         group_sched_out(event, cpuctx, ctx);
563                 else
564                         event_sched_out(event, cpuctx, ctx);
565                 event->state = PERF_EVENT_STATE_OFF;
566         }
567
568         raw_spin_unlock(&ctx->lock);
569 }
570
571 /*
572  * Disable a event.
573  *
574  * If event->ctx is a cloned context, callers must make sure that
575  * every task struct that event->ctx->task could possibly point to
576  * remains valid.  This condition is satisifed when called through
577  * perf_event_for_each_child or perf_event_for_each because they
578  * hold the top-level event's child_mutex, so any descendant that
579  * goes to exit will block in sync_child_event.
580  * When called from perf_pending_event it's OK because event->ctx
581  * is the current context on this CPU and preemption is disabled,
582  * hence we can't get into perf_event_task_sched_out for this context.
583  */
584 void perf_event_disable(struct perf_event *event)
585 {
586         struct perf_event_context *ctx = event->ctx;
587         struct task_struct *task = ctx->task;
588
589         if (!task) {
590                 /*
591                  * Disable the event on the cpu that it's on
592                  */
593                 smp_call_function_single(event->cpu, __perf_event_disable,
594                                          event, 1);
595                 return;
596         }
597
598  retry:
599         task_oncpu_function_call(task, __perf_event_disable, event);
600
601         raw_spin_lock_irq(&ctx->lock);
602         /*
603          * If the event is still active, we need to retry the cross-call.
604          */
605         if (event->state == PERF_EVENT_STATE_ACTIVE) {
606                 raw_spin_unlock_irq(&ctx->lock);
607                 goto retry;
608         }
609
610         /*
611          * Since we have the lock this context can't be scheduled
612          * in, so we can change the state safely.
613          */
614         if (event->state == PERF_EVENT_STATE_INACTIVE) {
615                 update_group_times(event);
616                 event->state = PERF_EVENT_STATE_OFF;
617         }
618
619         raw_spin_unlock_irq(&ctx->lock);
620 }
621
622 static int
623 event_sched_in(struct perf_event *event,
624                  struct perf_cpu_context *cpuctx,
625                  struct perf_event_context *ctx,
626                  int cpu)
627 {
628         if (event->state <= PERF_EVENT_STATE_OFF)
629                 return 0;
630
631         event->state = PERF_EVENT_STATE_ACTIVE;
632         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
633         /*
634          * The new state must be visible before we turn it on in the hardware:
635          */
636         smp_wmb();
637
638         if (event->pmu->enable(event)) {
639                 event->state = PERF_EVENT_STATE_INACTIVE;
640                 event->oncpu = -1;
641                 return -EAGAIN;
642         }
643
644         event->tstamp_running += ctx->time - event->tstamp_stopped;
645
646         if (!is_software_event(event))
647                 cpuctx->active_oncpu++;
648         ctx->nr_active++;
649
650         if (event->attr.exclusive)
651                 cpuctx->exclusive = 1;
652
653         return 0;
654 }
655
656 static int
657 group_sched_in(struct perf_event *group_event,
658                struct perf_cpu_context *cpuctx,
659                struct perf_event_context *ctx,
660                int cpu)
661 {
662         struct perf_event *event, *partial_group;
663         int ret;
664
665         if (group_event->state == PERF_EVENT_STATE_OFF)
666                 return 0;
667
668         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
669         if (ret)
670                 return ret < 0 ? ret : 0;
671
672         if (event_sched_in(group_event, cpuctx, ctx, cpu))
673                 return -EAGAIN;
674
675         /*
676          * Schedule in siblings as one group (if any):
677          */
678         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
679                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
680                         partial_group = event;
681                         goto group_error;
682                 }
683         }
684
685         return 0;
686
687 group_error:
688         /*
689          * Groups can be scheduled in as one unit only, so undo any
690          * partial group before returning:
691          */
692         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
693                 if (event == partial_group)
694                         break;
695                 event_sched_out(event, cpuctx, ctx);
696         }
697         event_sched_out(group_event, cpuctx, ctx);
698
699         return -EAGAIN;
700 }
701
702 /*
703  * Return 1 for a group consisting entirely of software events,
704  * 0 if the group contains any hardware events.
705  */
706 static int is_software_only_group(struct perf_event *leader)
707 {
708         struct perf_event *event;
709
710         if (!is_software_event(leader))
711                 return 0;
712
713         list_for_each_entry(event, &leader->sibling_list, group_entry)
714                 if (!is_software_event(event))
715                         return 0;
716
717         return 1;
718 }
719
720 /*
721  * Work out whether we can put this event group on the CPU now.
722  */
723 static int group_can_go_on(struct perf_event *event,
724                            struct perf_cpu_context *cpuctx,
725                            int can_add_hw)
726 {
727         /*
728          * Groups consisting entirely of software events can always go on.
729          */
730         if (is_software_only_group(event))
731                 return 1;
732         /*
733          * If an exclusive group is already on, no other hardware
734          * events can go on.
735          */
736         if (cpuctx->exclusive)
737                 return 0;
738         /*
739          * If this group is exclusive and there are already
740          * events on the CPU, it can't go on.
741          */
742         if (event->attr.exclusive && cpuctx->active_oncpu)
743                 return 0;
744         /*
745          * Otherwise, try to add it if all previous groups were able
746          * to go on.
747          */
748         return can_add_hw;
749 }
750
751 static void add_event_to_ctx(struct perf_event *event,
752                                struct perf_event_context *ctx)
753 {
754         list_add_event(event, ctx);
755         event->tstamp_enabled = ctx->time;
756         event->tstamp_running = ctx->time;
757         event->tstamp_stopped = ctx->time;
758 }
759
760 /*
761  * Cross CPU call to install and enable a performance event
762  *
763  * Must be called with ctx->mutex held
764  */
765 static void __perf_install_in_context(void *info)
766 {
767         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
768         struct perf_event *event = info;
769         struct perf_event_context *ctx = event->ctx;
770         struct perf_event *leader = event->group_leader;
771         int cpu = smp_processor_id();
772         int err;
773
774         /*
775          * If this is a task context, we need to check whether it is
776          * the current task context of this cpu. If not it has been
777          * scheduled out before the smp call arrived.
778          * Or possibly this is the right context but it isn't
779          * on this cpu because it had no events.
780          */
781         if (ctx->task && cpuctx->task_ctx != ctx) {
782                 if (cpuctx->task_ctx || ctx->task != current)
783                         return;
784                 cpuctx->task_ctx = ctx;
785         }
786
787         raw_spin_lock(&ctx->lock);
788         ctx->is_active = 1;
789         update_context_time(ctx);
790
791         /*
792          * Protect the list operation against NMI by disabling the
793          * events on a global level. NOP for non NMI based events.
794          */
795         perf_disable();
796
797         add_event_to_ctx(event, ctx);
798
799         if (event->cpu != -1 && event->cpu != smp_processor_id())
800                 goto unlock;
801
802         /*
803          * Don't put the event on if it is disabled or if
804          * it is in a group and the group isn't on.
805          */
806         if (event->state != PERF_EVENT_STATE_INACTIVE ||
807             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
808                 goto unlock;
809
810         /*
811          * An exclusive event can't go on if there are already active
812          * hardware events, and no hardware event can go on if there
813          * is already an exclusive event on.
814          */
815         if (!group_can_go_on(event, cpuctx, 1))
816                 err = -EEXIST;
817         else
818                 err = event_sched_in(event, cpuctx, ctx, cpu);
819
820         if (err) {
821                 /*
822                  * This event couldn't go on.  If it is in a group
823                  * then we have to pull the whole group off.
824                  * If the event group is pinned then put it in error state.
825                  */
826                 if (leader != event)
827                         group_sched_out(leader, cpuctx, ctx);
828                 if (leader->attr.pinned) {
829                         update_group_times(leader);
830                         leader->state = PERF_EVENT_STATE_ERROR;
831                 }
832         }
833
834         if (!err && !ctx->task && cpuctx->max_pertask)
835                 cpuctx->max_pertask--;
836
837  unlock:
838         perf_enable();
839
840         raw_spin_unlock(&ctx->lock);
841 }
842
843 /*
844  * Attach a performance event to a context
845  *
846  * First we add the event to the list with the hardware enable bit
847  * in event->hw_config cleared.
848  *
849  * If the event is attached to a task which is on a CPU we use a smp
850  * call to enable it in the task context. The task might have been
851  * scheduled away, but we check this in the smp call again.
852  *
853  * Must be called with ctx->mutex held.
854  */
855 static void
856 perf_install_in_context(struct perf_event_context *ctx,
857                         struct perf_event *event,
858                         int cpu)
859 {
860         struct task_struct *task = ctx->task;
861
862         if (!task) {
863                 /*
864                  * Per cpu events are installed via an smp call and
865                  * the install is always successful.
866                  */
867                 smp_call_function_single(cpu, __perf_install_in_context,
868                                          event, 1);
869                 return;
870         }
871
872 retry:
873         task_oncpu_function_call(task, __perf_install_in_context,
874                                  event);
875
876         raw_spin_lock_irq(&ctx->lock);
877         /*
878          * we need to retry the smp call.
879          */
880         if (ctx->is_active && list_empty(&event->group_entry)) {
881                 raw_spin_unlock_irq(&ctx->lock);
882                 goto retry;
883         }
884
885         /*
886          * The lock prevents that this context is scheduled in so we
887          * can add the event safely, if it the call above did not
888          * succeed.
889          */
890         if (list_empty(&event->group_entry))
891                 add_event_to_ctx(event, ctx);
892         raw_spin_unlock_irq(&ctx->lock);
893 }
894
895 /*
896  * Put a event into inactive state and update time fields.
897  * Enabling the leader of a group effectively enables all
898  * the group members that aren't explicitly disabled, so we
899  * have to update their ->tstamp_enabled also.
900  * Note: this works for group members as well as group leaders
901  * since the non-leader members' sibling_lists will be empty.
902  */
903 static void __perf_event_mark_enabled(struct perf_event *event,
904                                         struct perf_event_context *ctx)
905 {
906         struct perf_event *sub;
907
908         event->state = PERF_EVENT_STATE_INACTIVE;
909         event->tstamp_enabled = ctx->time - event->total_time_enabled;
910         list_for_each_entry(sub, &event->sibling_list, group_entry)
911                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
912                         sub->tstamp_enabled =
913                                 ctx->time - sub->total_time_enabled;
914 }
915
916 /*
917  * Cross CPU call to enable a performance event
918  */
919 static void __perf_event_enable(void *info)
920 {
921         struct perf_event *event = info;
922         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
923         struct perf_event_context *ctx = event->ctx;
924         struct perf_event *leader = event->group_leader;
925         int err;
926
927         /*
928          * If this is a per-task event, need to check whether this
929          * event's task is the current task on this cpu.
930          */
931         if (ctx->task && cpuctx->task_ctx != ctx) {
932                 if (cpuctx->task_ctx || ctx->task != current)
933                         return;
934                 cpuctx->task_ctx = ctx;
935         }
936
937         raw_spin_lock(&ctx->lock);
938         ctx->is_active = 1;
939         update_context_time(ctx);
940
941         if (event->state >= PERF_EVENT_STATE_INACTIVE)
942                 goto unlock;
943         __perf_event_mark_enabled(event, ctx);
944
945         if (event->cpu != -1 && event->cpu != smp_processor_id())
946                 goto unlock;
947
948         /*
949          * If the event is in a group and isn't the group leader,
950          * then don't put it on unless the group is on.
951          */
952         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
953                 goto unlock;
954
955         if (!group_can_go_on(event, cpuctx, 1)) {
956                 err = -EEXIST;
957         } else {
958                 perf_disable();
959                 if (event == leader)
960                         err = group_sched_in(event, cpuctx, ctx,
961                                              smp_processor_id());
962                 else
963                         err = event_sched_in(event, cpuctx, ctx,
964                                                smp_processor_id());
965                 perf_enable();
966         }
967
968         if (err) {
969                 /*
970                  * If this event can't go on and it's part of a
971                  * group, then the whole group has to come off.
972                  */
973                 if (leader != event)
974                         group_sched_out(leader, cpuctx, ctx);
975                 if (leader->attr.pinned) {
976                         update_group_times(leader);
977                         leader->state = PERF_EVENT_STATE_ERROR;
978                 }
979         }
980
981  unlock:
982         raw_spin_unlock(&ctx->lock);
983 }
984
985 /*
986  * Enable a event.
987  *
988  * If event->ctx is a cloned context, callers must make sure that
989  * every task struct that event->ctx->task could possibly point to
990  * remains valid.  This condition is satisfied when called through
991  * perf_event_for_each_child or perf_event_for_each as described
992  * for perf_event_disable.
993  */
994 void perf_event_enable(struct perf_event *event)
995 {
996         struct perf_event_context *ctx = event->ctx;
997         struct task_struct *task = ctx->task;
998
999         if (!task) {
1000                 /*
1001                  * Enable the event on the cpu that it's on
1002                  */
1003                 smp_call_function_single(event->cpu, __perf_event_enable,
1004                                          event, 1);
1005                 return;
1006         }
1007
1008         raw_spin_lock_irq(&ctx->lock);
1009         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1010                 goto out;
1011
1012         /*
1013          * If the event is in error state, clear that first.
1014          * That way, if we see the event in error state below, we
1015          * know that it has gone back into error state, as distinct
1016          * from the task having been scheduled away before the
1017          * cross-call arrived.
1018          */
1019         if (event->state == PERF_EVENT_STATE_ERROR)
1020                 event->state = PERF_EVENT_STATE_OFF;
1021
1022  retry:
1023         raw_spin_unlock_irq(&ctx->lock);
1024         task_oncpu_function_call(task, __perf_event_enable, event);
1025
1026         raw_spin_lock_irq(&ctx->lock);
1027
1028         /*
1029          * If the context is active and the event is still off,
1030          * we need to retry the cross-call.
1031          */
1032         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
1033                 goto retry;
1034
1035         /*
1036          * Since we have the lock this context can't be scheduled
1037          * in, so we can change the state safely.
1038          */
1039         if (event->state == PERF_EVENT_STATE_OFF)
1040                 __perf_event_mark_enabled(event, ctx);
1041
1042  out:
1043         raw_spin_unlock_irq(&ctx->lock);
1044 }
1045
1046 static int perf_event_refresh(struct perf_event *event, int refresh)
1047 {
1048         /*
1049          * not supported on inherited events
1050          */
1051         if (event->attr.inherit)
1052                 return -EINVAL;
1053
1054         atomic_add(refresh, &event->event_limit);
1055         perf_event_enable(event);
1056
1057         return 0;
1058 }
1059
1060 void __perf_event_sched_out(struct perf_event_context *ctx,
1061                               struct perf_cpu_context *cpuctx)
1062 {
1063         struct perf_event *event;
1064
1065         raw_spin_lock(&ctx->lock);
1066         ctx->is_active = 0;
1067         if (likely(!ctx->nr_events))
1068                 goto out;
1069         update_context_time(ctx);
1070
1071         perf_disable();
1072         if (ctx->nr_active) {
1073                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1074                         group_sched_out(event, cpuctx, ctx);
1075
1076                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1077                         group_sched_out(event, cpuctx, ctx);
1078         }
1079         perf_enable();
1080  out:
1081         raw_spin_unlock(&ctx->lock);
1082 }
1083
1084 /*
1085  * Test whether two contexts are equivalent, i.e. whether they
1086  * have both been cloned from the same version of the same context
1087  * and they both have the same number of enabled events.
1088  * If the number of enabled events is the same, then the set
1089  * of enabled events should be the same, because these are both
1090  * inherited contexts, therefore we can't access individual events
1091  * in them directly with an fd; we can only enable/disable all
1092  * events via prctl, or enable/disable all events in a family
1093  * via ioctl, which will have the same effect on both contexts.
1094  */
1095 static int context_equiv(struct perf_event_context *ctx1,
1096                          struct perf_event_context *ctx2)
1097 {
1098         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1099                 && ctx1->parent_gen == ctx2->parent_gen
1100                 && !ctx1->pin_count && !ctx2->pin_count;
1101 }
1102
1103 static void __perf_event_sync_stat(struct perf_event *event,
1104                                      struct perf_event *next_event)
1105 {
1106         u64 value;
1107
1108         if (!event->attr.inherit_stat)
1109                 return;
1110
1111         /*
1112          * Update the event value, we cannot use perf_event_read()
1113          * because we're in the middle of a context switch and have IRQs
1114          * disabled, which upsets smp_call_function_single(), however
1115          * we know the event must be on the current CPU, therefore we
1116          * don't need to use it.
1117          */
1118         switch (event->state) {
1119         case PERF_EVENT_STATE_ACTIVE:
1120                 event->pmu->read(event);
1121                 /* fall-through */
1122
1123         case PERF_EVENT_STATE_INACTIVE:
1124                 update_event_times(event);
1125                 break;
1126
1127         default:
1128                 break;
1129         }
1130
1131         /*
1132          * In order to keep per-task stats reliable we need to flip the event
1133          * values when we flip the contexts.
1134          */
1135         value = atomic64_read(&next_event->count);
1136         value = atomic64_xchg(&event->count, value);
1137         atomic64_set(&next_event->count, value);
1138
1139         swap(event->total_time_enabled, next_event->total_time_enabled);
1140         swap(event->total_time_running, next_event->total_time_running);
1141
1142         /*
1143          * Since we swizzled the values, update the user visible data too.
1144          */
1145         perf_event_update_userpage(event);
1146         perf_event_update_userpage(next_event);
1147 }
1148
1149 #define list_next_entry(pos, member) \
1150         list_entry(pos->member.next, typeof(*pos), member)
1151
1152 static void perf_event_sync_stat(struct perf_event_context *ctx,
1153                                    struct perf_event_context *next_ctx)
1154 {
1155         struct perf_event *event, *next_event;
1156
1157         if (!ctx->nr_stat)
1158                 return;
1159
1160         update_context_time(ctx);
1161
1162         event = list_first_entry(&ctx->event_list,
1163                                    struct perf_event, event_entry);
1164
1165         next_event = list_first_entry(&next_ctx->event_list,
1166                                         struct perf_event, event_entry);
1167
1168         while (&event->event_entry != &ctx->event_list &&
1169                &next_event->event_entry != &next_ctx->event_list) {
1170
1171                 __perf_event_sync_stat(event, next_event);
1172
1173                 event = list_next_entry(event, event_entry);
1174                 next_event = list_next_entry(next_event, event_entry);
1175         }
1176 }
1177
1178 /*
1179  * Called from scheduler to remove the events of the current task,
1180  * with interrupts disabled.
1181  *
1182  * We stop each event and update the event value in event->count.
1183  *
1184  * This does not protect us against NMI, but disable()
1185  * sets the disabled bit in the control field of event _before_
1186  * accessing the event control register. If a NMI hits, then it will
1187  * not restart the event.
1188  */
1189 void perf_event_task_sched_out(struct task_struct *task,
1190                                  struct task_struct *next)
1191 {
1192         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1193         struct perf_event_context *ctx = task->perf_event_ctxp;
1194         struct perf_event_context *next_ctx;
1195         struct perf_event_context *parent;
1196         struct pt_regs *regs;
1197         int do_switch = 1;
1198
1199         regs = task_pt_regs(task);
1200         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1201
1202         if (likely(!ctx || !cpuctx->task_ctx))
1203                 return;
1204
1205         rcu_read_lock();
1206         parent = rcu_dereference(ctx->parent_ctx);
1207         next_ctx = next->perf_event_ctxp;
1208         if (parent && next_ctx &&
1209             rcu_dereference(next_ctx->parent_ctx) == parent) {
1210                 /*
1211                  * Looks like the two contexts are clones, so we might be
1212                  * able to optimize the context switch.  We lock both
1213                  * contexts and check that they are clones under the
1214                  * lock (including re-checking that neither has been
1215                  * uncloned in the meantime).  It doesn't matter which
1216                  * order we take the locks because no other cpu could
1217                  * be trying to lock both of these tasks.
1218                  */
1219                 raw_spin_lock(&ctx->lock);
1220                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1221                 if (context_equiv(ctx, next_ctx)) {
1222                         /*
1223                          * XXX do we need a memory barrier of sorts
1224                          * wrt to rcu_dereference() of perf_event_ctxp
1225                          */
1226                         task->perf_event_ctxp = next_ctx;
1227                         next->perf_event_ctxp = ctx;
1228                         ctx->task = next;
1229                         next_ctx->task = task;
1230                         do_switch = 0;
1231
1232                         perf_event_sync_stat(ctx, next_ctx);
1233                 }
1234                 raw_spin_unlock(&next_ctx->lock);
1235                 raw_spin_unlock(&ctx->lock);
1236         }
1237         rcu_read_unlock();
1238
1239         if (do_switch) {
1240                 __perf_event_sched_out(ctx, cpuctx);
1241                 cpuctx->task_ctx = NULL;
1242         }
1243 }
1244
1245 /*
1246  * Called with IRQs disabled
1247  */
1248 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1249 {
1250         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1251
1252         if (!cpuctx->task_ctx)
1253                 return;
1254
1255         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1256                 return;
1257
1258         __perf_event_sched_out(ctx, cpuctx);
1259         cpuctx->task_ctx = NULL;
1260 }
1261
1262 /*
1263  * Called with IRQs disabled
1264  */
1265 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1266 {
1267         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1268 }
1269
1270 static void
1271 __perf_event_sched_in(struct perf_event_context *ctx,
1272                         struct perf_cpu_context *cpuctx)
1273 {
1274         int cpu = smp_processor_id();
1275         struct perf_event *event;
1276         int can_add_hw = 1;
1277
1278         raw_spin_lock(&ctx->lock);
1279         ctx->is_active = 1;
1280         if (likely(!ctx->nr_events))
1281                 goto out;
1282
1283         ctx->timestamp = perf_clock();
1284
1285         perf_disable();
1286
1287         /*
1288          * First go through the list and put on any pinned groups
1289          * in order to give them the best chance of going on.
1290          */
1291         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1292                 if (event->state <= PERF_EVENT_STATE_OFF)
1293                         continue;
1294                 if (event->cpu != -1 && event->cpu != cpu)
1295                         continue;
1296
1297                 if (group_can_go_on(event, cpuctx, 1))
1298                         group_sched_in(event, cpuctx, ctx, cpu);
1299
1300                 /*
1301                  * If this pinned group hasn't been scheduled,
1302                  * put it in error state.
1303                  */
1304                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1305                         update_group_times(event);
1306                         event->state = PERF_EVENT_STATE_ERROR;
1307                 }
1308         }
1309
1310         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1311                 /* Ignore events in OFF or ERROR state */
1312                 if (event->state <= PERF_EVENT_STATE_OFF)
1313                         continue;
1314                 /*
1315                  * Listen to the 'cpu' scheduling filter constraint
1316                  * of events:
1317                  */
1318                 if (event->cpu != -1 && event->cpu != cpu)
1319                         continue;
1320
1321                 if (group_can_go_on(event, cpuctx, can_add_hw))
1322                         if (group_sched_in(event, cpuctx, ctx, cpu))
1323                                 can_add_hw = 0;
1324         }
1325         perf_enable();
1326  out:
1327         raw_spin_unlock(&ctx->lock);
1328 }
1329
1330 /*
1331  * Called from scheduler to add the events of the current task
1332  * with interrupts disabled.
1333  *
1334  * We restore the event value and then enable it.
1335  *
1336  * This does not protect us against NMI, but enable()
1337  * sets the enabled bit in the control field of event _before_
1338  * accessing the event control register. If a NMI hits, then it will
1339  * keep the event running.
1340  */
1341 void perf_event_task_sched_in(struct task_struct *task)
1342 {
1343         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1344         struct perf_event_context *ctx = task->perf_event_ctxp;
1345
1346         if (likely(!ctx))
1347                 return;
1348         if (cpuctx->task_ctx == ctx)
1349                 return;
1350         __perf_event_sched_in(ctx, cpuctx);
1351         cpuctx->task_ctx = ctx;
1352 }
1353
1354 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx)
1355 {
1356         struct perf_event_context *ctx = &cpuctx->ctx;
1357
1358         __perf_event_sched_in(ctx, cpuctx);
1359 }
1360
1361 #define MAX_INTERRUPTS (~0ULL)
1362
1363 static void perf_log_throttle(struct perf_event *event, int enable);
1364
1365 static void perf_adjust_period(struct perf_event *event, u64 events)
1366 {
1367         struct hw_perf_event *hwc = &event->hw;
1368         u64 period, sample_period;
1369         s64 delta;
1370
1371         events *= hwc->sample_period;
1372         period = div64_u64(events, event->attr.sample_freq);
1373
1374         delta = (s64)(period - hwc->sample_period);
1375         delta = (delta + 7) / 8; /* low pass filter */
1376
1377         sample_period = hwc->sample_period + delta;
1378
1379         if (!sample_period)
1380                 sample_period = 1;
1381
1382         hwc->sample_period = sample_period;
1383 }
1384
1385 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1386 {
1387         struct perf_event *event;
1388         struct hw_perf_event *hwc;
1389         u64 interrupts, freq;
1390
1391         raw_spin_lock(&ctx->lock);
1392         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1393                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1394                         continue;
1395
1396                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1397                         continue;
1398
1399                 hwc = &event->hw;
1400
1401                 interrupts = hwc->interrupts;
1402                 hwc->interrupts = 0;
1403
1404                 /*
1405                  * unthrottle events on the tick
1406                  */
1407                 if (interrupts == MAX_INTERRUPTS) {
1408                         perf_log_throttle(event, 1);
1409                         event->pmu->unthrottle(event);
1410                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1411                 }
1412
1413                 if (!event->attr.freq || !event->attr.sample_freq)
1414                         continue;
1415
1416                 /*
1417                  * if the specified freq < HZ then we need to skip ticks
1418                  */
1419                 if (event->attr.sample_freq < HZ) {
1420                         freq = event->attr.sample_freq;
1421
1422                         hwc->freq_count += freq;
1423                         hwc->freq_interrupts += interrupts;
1424
1425                         if (hwc->freq_count < HZ)
1426                                 continue;
1427
1428                         interrupts = hwc->freq_interrupts;
1429                         hwc->freq_interrupts = 0;
1430                         hwc->freq_count -= HZ;
1431                 } else
1432                         freq = HZ;
1433
1434                 perf_adjust_period(event, freq * interrupts);
1435
1436                 /*
1437                  * In order to avoid being stalled by an (accidental) huge
1438                  * sample period, force reset the sample period if we didn't
1439                  * get any events in this freq period.
1440                  */
1441                 if (!interrupts) {
1442                         perf_disable();
1443                         event->pmu->disable(event);
1444                         atomic64_set(&hwc->period_left, 0);
1445                         event->pmu->enable(event);
1446                         perf_enable();
1447                 }
1448         }
1449         raw_spin_unlock(&ctx->lock);
1450 }
1451
1452 /*
1453  * Round-robin a context's events:
1454  */
1455 static void rotate_ctx(struct perf_event_context *ctx)
1456 {
1457         struct perf_event *event;
1458
1459         if (!ctx->nr_events)
1460                 return;
1461
1462         raw_spin_lock(&ctx->lock);
1463         /*
1464          * Rotate the first entry last (works just fine for group events too):
1465          */
1466         perf_disable();
1467         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1468                 list_move_tail(&event->group_entry, &ctx->pinned_groups);
1469                 break;
1470         }
1471
1472         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1473                 list_move_tail(&event->group_entry, &ctx->flexible_groups);
1474                 break;
1475         }
1476         perf_enable();
1477
1478         raw_spin_unlock(&ctx->lock);
1479 }
1480
1481 void perf_event_task_tick(struct task_struct *curr)
1482 {
1483         struct perf_cpu_context *cpuctx;
1484         struct perf_event_context *ctx;
1485
1486         if (!atomic_read(&nr_events))
1487                 return;
1488
1489         cpuctx = &__get_cpu_var(perf_cpu_context);
1490         ctx = curr->perf_event_ctxp;
1491
1492         perf_ctx_adjust_freq(&cpuctx->ctx);
1493         if (ctx)
1494                 perf_ctx_adjust_freq(ctx);
1495
1496         perf_event_cpu_sched_out(cpuctx);
1497         if (ctx)
1498                 __perf_event_task_sched_out(ctx);
1499
1500         rotate_ctx(&cpuctx->ctx);
1501         if (ctx)
1502                 rotate_ctx(ctx);
1503
1504         perf_event_cpu_sched_in(cpuctx);
1505         if (ctx)
1506                 perf_event_task_sched_in(curr);
1507 }
1508
1509 static int event_enable_on_exec(struct perf_event *event,
1510                                 struct perf_event_context *ctx)
1511 {
1512         if (!event->attr.enable_on_exec)
1513                 return 0;
1514
1515         event->attr.enable_on_exec = 0;
1516         if (event->state >= PERF_EVENT_STATE_INACTIVE)
1517                 return 0;
1518
1519         __perf_event_mark_enabled(event, ctx);
1520
1521         return 1;
1522 }
1523
1524 /*
1525  * Enable all of a task's events that have been marked enable-on-exec.
1526  * This expects task == current.
1527  */
1528 static void perf_event_enable_on_exec(struct task_struct *task)
1529 {
1530         struct perf_event_context *ctx;
1531         struct perf_event *event;
1532         unsigned long flags;
1533         int enabled = 0;
1534         int ret;
1535
1536         local_irq_save(flags);
1537         ctx = task->perf_event_ctxp;
1538         if (!ctx || !ctx->nr_events)
1539                 goto out;
1540
1541         __perf_event_task_sched_out(ctx);
1542
1543         raw_spin_lock(&ctx->lock);
1544
1545         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1546                 ret = event_enable_on_exec(event, ctx);
1547                 if (ret)
1548                         enabled = 1;
1549         }
1550
1551         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1552                 ret = event_enable_on_exec(event, ctx);
1553                 if (ret)
1554                         enabled = 1;
1555         }
1556
1557         /*
1558          * Unclone this context if we enabled any event.
1559          */
1560         if (enabled)
1561                 unclone_ctx(ctx);
1562
1563         raw_spin_unlock(&ctx->lock);
1564
1565         perf_event_task_sched_in(task);
1566  out:
1567         local_irq_restore(flags);
1568 }
1569
1570 /*
1571  * Cross CPU call to read the hardware event
1572  */
1573 static void __perf_event_read(void *info)
1574 {
1575         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1576         struct perf_event *event = info;
1577         struct perf_event_context *ctx = event->ctx;
1578
1579         /*
1580          * If this is a task context, we need to check whether it is
1581          * the current task context of this cpu.  If not it has been
1582          * scheduled out before the smp call arrived.  In that case
1583          * event->count would have been updated to a recent sample
1584          * when the event was scheduled out.
1585          */
1586         if (ctx->task && cpuctx->task_ctx != ctx)
1587                 return;
1588
1589         raw_spin_lock(&ctx->lock);
1590         update_context_time(ctx);
1591         update_event_times(event);
1592         raw_spin_unlock(&ctx->lock);
1593
1594         event->pmu->read(event);
1595 }
1596
1597 static u64 perf_event_read(struct perf_event *event)
1598 {
1599         /*
1600          * If event is enabled and currently active on a CPU, update the
1601          * value in the event structure:
1602          */
1603         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1604                 smp_call_function_single(event->oncpu,
1605                                          __perf_event_read, event, 1);
1606         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1607                 struct perf_event_context *ctx = event->ctx;
1608                 unsigned long flags;
1609
1610                 raw_spin_lock_irqsave(&ctx->lock, flags);
1611                 update_context_time(ctx);
1612                 update_event_times(event);
1613                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1614         }
1615
1616         return atomic64_read(&event->count);
1617 }
1618
1619 /*
1620  * Initialize the perf_event context in a task_struct:
1621  */
1622 static void
1623 __perf_event_init_context(struct perf_event_context *ctx,
1624                             struct task_struct *task)
1625 {
1626         raw_spin_lock_init(&ctx->lock);
1627         mutex_init(&ctx->mutex);
1628         INIT_LIST_HEAD(&ctx->pinned_groups);
1629         INIT_LIST_HEAD(&ctx->flexible_groups);
1630         INIT_LIST_HEAD(&ctx->event_list);
1631         atomic_set(&ctx->refcount, 1);
1632         ctx->task = task;
1633 }
1634
1635 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1636 {
1637         struct perf_event_context *ctx;
1638         struct perf_cpu_context *cpuctx;
1639         struct task_struct *task;
1640         unsigned long flags;
1641         int err;
1642
1643         if (pid == -1 && cpu != -1) {
1644                 /* Must be root to operate on a CPU event: */
1645                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1646                         return ERR_PTR(-EACCES);
1647
1648                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1649                         return ERR_PTR(-EINVAL);
1650
1651                 /*
1652                  * We could be clever and allow to attach a event to an
1653                  * offline CPU and activate it when the CPU comes up, but
1654                  * that's for later.
1655                  */
1656                 if (!cpu_online(cpu))
1657                         return ERR_PTR(-ENODEV);
1658
1659                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1660                 ctx = &cpuctx->ctx;
1661                 get_ctx(ctx);
1662
1663                 return ctx;
1664         }
1665
1666         rcu_read_lock();
1667         if (!pid)
1668                 task = current;
1669         else
1670                 task = find_task_by_vpid(pid);
1671         if (task)
1672                 get_task_struct(task);
1673         rcu_read_unlock();
1674
1675         if (!task)
1676                 return ERR_PTR(-ESRCH);
1677
1678         /*
1679          * Can't attach events to a dying task.
1680          */
1681         err = -ESRCH;
1682         if (task->flags & PF_EXITING)
1683                 goto errout;
1684
1685         /* Reuse ptrace permission checks for now. */
1686         err = -EACCES;
1687         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1688                 goto errout;
1689
1690  retry:
1691         ctx = perf_lock_task_context(task, &flags);
1692         if (ctx) {
1693                 unclone_ctx(ctx);
1694                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1695         }
1696
1697         if (!ctx) {
1698                 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1699                 err = -ENOMEM;
1700                 if (!ctx)
1701                         goto errout;
1702                 __perf_event_init_context(ctx, task);
1703                 get_ctx(ctx);
1704                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1705                         /*
1706                          * We raced with some other task; use
1707                          * the context they set.
1708                          */
1709                         kfree(ctx);
1710                         goto retry;
1711                 }
1712                 get_task_struct(task);
1713         }
1714
1715         put_task_struct(task);
1716         return ctx;
1717
1718  errout:
1719         put_task_struct(task);
1720         return ERR_PTR(err);
1721 }
1722
1723 static void perf_event_free_filter(struct perf_event *event);
1724
1725 static void free_event_rcu(struct rcu_head *head)
1726 {
1727         struct perf_event *event;
1728
1729         event = container_of(head, struct perf_event, rcu_head);
1730         if (event->ns)
1731                 put_pid_ns(event->ns);
1732         perf_event_free_filter(event);
1733         kfree(event);
1734 }
1735
1736 static void perf_pending_sync(struct perf_event *event);
1737
1738 static void free_event(struct perf_event *event)
1739 {
1740         perf_pending_sync(event);
1741
1742         if (!event->parent) {
1743                 atomic_dec(&nr_events);
1744                 if (event->attr.mmap)
1745                         atomic_dec(&nr_mmap_events);
1746                 if (event->attr.comm)
1747                         atomic_dec(&nr_comm_events);
1748                 if (event->attr.task)
1749                         atomic_dec(&nr_task_events);
1750         }
1751
1752         if (event->output) {
1753                 fput(event->output->filp);
1754                 event->output = NULL;
1755         }
1756
1757         if (event->destroy)
1758                 event->destroy(event);
1759
1760         put_ctx(event->ctx);
1761         call_rcu(&event->rcu_head, free_event_rcu);
1762 }
1763
1764 int perf_event_release_kernel(struct perf_event *event)
1765 {
1766         struct perf_event_context *ctx = event->ctx;
1767
1768         WARN_ON_ONCE(ctx->parent_ctx);
1769         mutex_lock(&ctx->mutex);
1770         perf_event_remove_from_context(event);
1771         mutex_unlock(&ctx->mutex);
1772
1773         mutex_lock(&event->owner->perf_event_mutex);
1774         list_del_init(&event->owner_entry);
1775         mutex_unlock(&event->owner->perf_event_mutex);
1776         put_task_struct(event->owner);
1777
1778         free_event(event);
1779
1780         return 0;
1781 }
1782 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1783
1784 /*
1785  * Called when the last reference to the file is gone.
1786  */
1787 static int perf_release(struct inode *inode, struct file *file)
1788 {
1789         struct perf_event *event = file->private_data;
1790
1791         file->private_data = NULL;
1792
1793         return perf_event_release_kernel(event);
1794 }
1795
1796 static int perf_event_read_size(struct perf_event *event)
1797 {
1798         int entry = sizeof(u64); /* value */
1799         int size = 0;
1800         int nr = 1;
1801
1802         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1803                 size += sizeof(u64);
1804
1805         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1806                 size += sizeof(u64);
1807
1808         if (event->attr.read_format & PERF_FORMAT_ID)
1809                 entry += sizeof(u64);
1810
1811         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1812                 nr += event->group_leader->nr_siblings;
1813                 size += sizeof(u64);
1814         }
1815
1816         size += entry * nr;
1817
1818         return size;
1819 }
1820
1821 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
1822 {
1823         struct perf_event *child;
1824         u64 total = 0;
1825
1826         *enabled = 0;
1827         *running = 0;
1828
1829         mutex_lock(&event->child_mutex);
1830         total += perf_event_read(event);
1831         *enabled += event->total_time_enabled +
1832                         atomic64_read(&event->child_total_time_enabled);
1833         *running += event->total_time_running +
1834                         atomic64_read(&event->child_total_time_running);
1835
1836         list_for_each_entry(child, &event->child_list, child_list) {
1837                 total += perf_event_read(child);
1838                 *enabled += child->total_time_enabled;
1839                 *running += child->total_time_running;
1840         }
1841         mutex_unlock(&event->child_mutex);
1842
1843         return total;
1844 }
1845 EXPORT_SYMBOL_GPL(perf_event_read_value);
1846
1847 static int perf_event_read_group(struct perf_event *event,
1848                                    u64 read_format, char __user *buf)
1849 {
1850         struct perf_event *leader = event->group_leader, *sub;
1851         int n = 0, size = 0, ret = -EFAULT;
1852         struct perf_event_context *ctx = leader->ctx;
1853         u64 values[5];
1854         u64 count, enabled, running;
1855
1856         mutex_lock(&ctx->mutex);
1857         count = perf_event_read_value(leader, &enabled, &running);
1858
1859         values[n++] = 1 + leader->nr_siblings;
1860         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1861                 values[n++] = enabled;
1862         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1863                 values[n++] = running;
1864         values[n++] = count;
1865         if (read_format & PERF_FORMAT_ID)
1866                 values[n++] = primary_event_id(leader);
1867
1868         size = n * sizeof(u64);
1869
1870         if (copy_to_user(buf, values, size))
1871                 goto unlock;
1872
1873         ret = size;
1874
1875         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1876                 n = 0;
1877
1878                 values[n++] = perf_event_read_value(sub, &enabled, &running);
1879                 if (read_format & PERF_FORMAT_ID)
1880                         values[n++] = primary_event_id(sub);
1881
1882                 size = n * sizeof(u64);
1883
1884                 if (copy_to_user(buf + ret, values, size)) {
1885                         ret = -EFAULT;
1886                         goto unlock;
1887                 }
1888
1889                 ret += size;
1890         }
1891 unlock:
1892         mutex_unlock(&ctx->mutex);
1893
1894         return ret;
1895 }
1896
1897 static int perf_event_read_one(struct perf_event *event,
1898                                  u64 read_format, char __user *buf)
1899 {
1900         u64 enabled, running;
1901         u64 values[4];
1902         int n = 0;
1903
1904         values[n++] = perf_event_read_value(event, &enabled, &running);
1905         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1906                 values[n++] = enabled;
1907         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1908                 values[n++] = running;
1909         if (read_format & PERF_FORMAT_ID)
1910                 values[n++] = primary_event_id(event);
1911
1912         if (copy_to_user(buf, values, n * sizeof(u64)))
1913                 return -EFAULT;
1914
1915         return n * sizeof(u64);
1916 }
1917
1918 /*
1919  * Read the performance event - simple non blocking version for now
1920  */
1921 static ssize_t
1922 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1923 {
1924         u64 read_format = event->attr.read_format;
1925         int ret;
1926
1927         /*
1928          * Return end-of-file for a read on a event that is in
1929          * error state (i.e. because it was pinned but it couldn't be
1930          * scheduled on to the CPU at some point).
1931          */
1932         if (event->state == PERF_EVENT_STATE_ERROR)
1933                 return 0;
1934
1935         if (count < perf_event_read_size(event))
1936                 return -ENOSPC;
1937
1938         WARN_ON_ONCE(event->ctx->parent_ctx);
1939         if (read_format & PERF_FORMAT_GROUP)
1940                 ret = perf_event_read_group(event, read_format, buf);
1941         else
1942                 ret = perf_event_read_one(event, read_format, buf);
1943
1944         return ret;
1945 }
1946
1947 static ssize_t
1948 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1949 {
1950         struct perf_event *event = file->private_data;
1951
1952         return perf_read_hw(event, buf, count);
1953 }
1954
1955 static unsigned int perf_poll(struct file *file, poll_table *wait)
1956 {
1957         struct perf_event *event = file->private_data;
1958         struct perf_mmap_data *data;
1959         unsigned int events = POLL_HUP;
1960
1961         rcu_read_lock();
1962         data = rcu_dereference(event->data);
1963         if (data)
1964                 events = atomic_xchg(&data->poll, 0);
1965         rcu_read_unlock();
1966
1967         poll_wait(file, &event->waitq, wait);
1968
1969         return events;
1970 }
1971
1972 static void perf_event_reset(struct perf_event *event)
1973 {
1974         (void)perf_event_read(event);
1975         atomic64_set(&event->count, 0);
1976         perf_event_update_userpage(event);
1977 }
1978
1979 /*
1980  * Holding the top-level event's child_mutex means that any
1981  * descendant process that has inherited this event will block
1982  * in sync_child_event if it goes to exit, thus satisfying the
1983  * task existence requirements of perf_event_enable/disable.
1984  */
1985 static void perf_event_for_each_child(struct perf_event *event,
1986                                         void (*func)(struct perf_event *))
1987 {
1988         struct perf_event *child;
1989
1990         WARN_ON_ONCE(event->ctx->parent_ctx);
1991         mutex_lock(&event->child_mutex);
1992         func(event);
1993         list_for_each_entry(child, &event->child_list, child_list)
1994                 func(child);
1995         mutex_unlock(&event->child_mutex);
1996 }
1997
1998 static void perf_event_for_each(struct perf_event *event,
1999                                   void (*func)(struct perf_event *))
2000 {
2001         struct perf_event_context *ctx = event->ctx;
2002         struct perf_event *sibling;
2003
2004         WARN_ON_ONCE(ctx->parent_ctx);
2005         mutex_lock(&ctx->mutex);
2006         event = event->group_leader;
2007
2008         perf_event_for_each_child(event, func);
2009         func(event);
2010         list_for_each_entry(sibling, &event->sibling_list, group_entry)
2011                 perf_event_for_each_child(event, func);
2012         mutex_unlock(&ctx->mutex);
2013 }
2014
2015 static int perf_event_period(struct perf_event *event, u64 __user *arg)
2016 {
2017         struct perf_event_context *ctx = event->ctx;
2018         unsigned long size;
2019         int ret = 0;
2020         u64 value;
2021
2022         if (!event->attr.sample_period)
2023                 return -EINVAL;
2024
2025         size = copy_from_user(&value, arg, sizeof(value));
2026         if (size != sizeof(value))
2027                 return -EFAULT;
2028
2029         if (!value)
2030                 return -EINVAL;
2031
2032         raw_spin_lock_irq(&ctx->lock);
2033         if (event->attr.freq) {
2034                 if (value > sysctl_perf_event_sample_rate) {
2035                         ret = -EINVAL;
2036                         goto unlock;
2037                 }
2038
2039                 event->attr.sample_freq = value;
2040         } else {
2041                 event->attr.sample_period = value;
2042                 event->hw.sample_period = value;
2043         }
2044 unlock:
2045         raw_spin_unlock_irq(&ctx->lock);
2046
2047         return ret;
2048 }
2049
2050 static int perf_event_set_output(struct perf_event *event, int output_fd);
2051 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
2052
2053 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2054 {
2055         struct perf_event *event = file->private_data;
2056         void (*func)(struct perf_event *);
2057         u32 flags = arg;
2058
2059         switch (cmd) {
2060         case PERF_EVENT_IOC_ENABLE:
2061                 func = perf_event_enable;
2062                 break;
2063         case PERF_EVENT_IOC_DISABLE:
2064                 func = perf_event_disable;
2065                 break;
2066         case PERF_EVENT_IOC_RESET:
2067                 func = perf_event_reset;
2068                 break;
2069
2070         case PERF_EVENT_IOC_REFRESH:
2071                 return perf_event_refresh(event, arg);
2072
2073         case PERF_EVENT_IOC_PERIOD:
2074                 return perf_event_period(event, (u64 __user *)arg);
2075
2076         case PERF_EVENT_IOC_SET_OUTPUT:
2077                 return perf_event_set_output(event, arg);
2078
2079         case PERF_EVENT_IOC_SET_FILTER:
2080                 return perf_event_set_filter(event, (void __user *)arg);
2081
2082         default:
2083                 return -ENOTTY;
2084         }
2085
2086         if (flags & PERF_IOC_FLAG_GROUP)
2087                 perf_event_for_each(event, func);
2088         else
2089                 perf_event_for_each_child(event, func);
2090
2091         return 0;
2092 }
2093
2094 int perf_event_task_enable(void)
2095 {
2096         struct perf_event *event;
2097
2098         mutex_lock(&current->perf_event_mutex);
2099         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2100                 perf_event_for_each_child(event, perf_event_enable);
2101         mutex_unlock(&current->perf_event_mutex);
2102
2103         return 0;
2104 }
2105
2106 int perf_event_task_disable(void)
2107 {
2108         struct perf_event *event;
2109
2110         mutex_lock(&current->perf_event_mutex);
2111         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2112                 perf_event_for_each_child(event, perf_event_disable);
2113         mutex_unlock(&current->perf_event_mutex);
2114
2115         return 0;
2116 }
2117
2118 #ifndef PERF_EVENT_INDEX_OFFSET
2119 # define PERF_EVENT_INDEX_OFFSET 0
2120 #endif
2121
2122 static int perf_event_index(struct perf_event *event)
2123 {
2124         if (event->state != PERF_EVENT_STATE_ACTIVE)
2125                 return 0;
2126
2127         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2128 }
2129
2130 /*
2131  * Callers need to ensure there can be no nesting of this function, otherwise
2132  * the seqlock logic goes bad. We can not serialize this because the arch
2133  * code calls this from NMI context.
2134  */
2135 void perf_event_update_userpage(struct perf_event *event)
2136 {
2137         struct perf_event_mmap_page *userpg;
2138         struct perf_mmap_data *data;
2139
2140         rcu_read_lock();
2141         data = rcu_dereference(event->data);
2142         if (!data)
2143                 goto unlock;
2144
2145         userpg = data->user_page;
2146
2147         /*
2148          * Disable preemption so as to not let the corresponding user-space
2149          * spin too long if we get preempted.
2150          */
2151         preempt_disable();
2152         ++userpg->lock;
2153         barrier();
2154         userpg->index = perf_event_index(event);
2155         userpg->offset = atomic64_read(&event->count);
2156         if (event->state == PERF_EVENT_STATE_ACTIVE)
2157                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2158
2159         userpg->time_enabled = event->total_time_enabled +
2160                         atomic64_read(&event->child_total_time_enabled);
2161
2162         userpg->time_running = event->total_time_running +
2163                         atomic64_read(&event->child_total_time_running);
2164
2165         barrier();
2166         ++userpg->lock;
2167         preempt_enable();
2168 unlock:
2169         rcu_read_unlock();
2170 }
2171
2172 static unsigned long perf_data_size(struct perf_mmap_data *data)
2173 {
2174         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2175 }
2176
2177 #ifndef CONFIG_PERF_USE_VMALLOC
2178
2179 /*
2180  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2181  */
2182
2183 static struct page *
2184 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2185 {
2186         if (pgoff > data->nr_pages)
2187                 return NULL;
2188
2189         if (pgoff == 0)
2190                 return virt_to_page(data->user_page);
2191
2192         return virt_to_page(data->data_pages[pgoff - 1]);
2193 }
2194
2195 static struct perf_mmap_data *
2196 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2197 {
2198         struct perf_mmap_data *data;
2199         unsigned long size;
2200         int i;
2201
2202         WARN_ON(atomic_read(&event->mmap_count));
2203
2204         size = sizeof(struct perf_mmap_data);
2205         size += nr_pages * sizeof(void *);
2206
2207         data = kzalloc(size, GFP_KERNEL);
2208         if (!data)
2209                 goto fail;
2210
2211         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2212         if (!data->user_page)
2213                 goto fail_user_page;
2214
2215         for (i = 0; i < nr_pages; i++) {
2216                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2217                 if (!data->data_pages[i])
2218                         goto fail_data_pages;
2219         }
2220
2221         data->data_order = 0;
2222         data->nr_pages = nr_pages;
2223
2224         return data;
2225
2226 fail_data_pages:
2227         for (i--; i >= 0; i--)
2228                 free_page((unsigned long)data->data_pages[i]);
2229
2230         free_page((unsigned long)data->user_page);
2231
2232 fail_user_page:
2233         kfree(data);
2234
2235 fail:
2236         return NULL;
2237 }
2238
2239 static void perf_mmap_free_page(unsigned long addr)
2240 {
2241         struct page *page = virt_to_page((void *)addr);
2242
2243         page->mapping = NULL;
2244         __free_page(page);
2245 }
2246
2247 static void perf_mmap_data_free(struct perf_mmap_data *data)
2248 {
2249         int i;
2250
2251         perf_mmap_free_page((unsigned long)data->user_page);
2252         for (i = 0; i < data->nr_pages; i++)
2253                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2254         kfree(data);
2255 }
2256
2257 #else
2258
2259 /*
2260  * Back perf_mmap() with vmalloc memory.
2261  *
2262  * Required for architectures that have d-cache aliasing issues.
2263  */
2264
2265 static struct page *
2266 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2267 {
2268         if (pgoff > (1UL << data->data_order))
2269                 return NULL;
2270
2271         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2272 }
2273
2274 static void perf_mmap_unmark_page(void *addr)
2275 {
2276         struct page *page = vmalloc_to_page(addr);
2277
2278         page->mapping = NULL;
2279 }
2280
2281 static void perf_mmap_data_free_work(struct work_struct *work)
2282 {
2283         struct perf_mmap_data *data;
2284         void *base;
2285         int i, nr;
2286
2287         data = container_of(work, struct perf_mmap_data, work);
2288         nr = 1 << data->data_order;
2289
2290         base = data->user_page;
2291         for (i = 0; i < nr + 1; i++)
2292                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2293
2294         vfree(base);
2295         kfree(data);
2296 }
2297
2298 static void perf_mmap_data_free(struct perf_mmap_data *data)
2299 {
2300         schedule_work(&data->work);
2301 }
2302
2303 static struct perf_mmap_data *
2304 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2305 {
2306         struct perf_mmap_data *data;
2307         unsigned long size;
2308         void *all_buf;
2309
2310         WARN_ON(atomic_read(&event->mmap_count));
2311
2312         size = sizeof(struct perf_mmap_data);
2313         size += sizeof(void *);
2314
2315         data = kzalloc(size, GFP_KERNEL);
2316         if (!data)
2317                 goto fail;
2318
2319         INIT_WORK(&data->work, perf_mmap_data_free_work);
2320
2321         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2322         if (!all_buf)
2323                 goto fail_all_buf;
2324
2325         data->user_page = all_buf;
2326         data->data_pages[0] = all_buf + PAGE_SIZE;
2327         data->data_order = ilog2(nr_pages);
2328         data->nr_pages = 1;
2329
2330         return data;
2331
2332 fail_all_buf:
2333         kfree(data);
2334
2335 fail:
2336         return NULL;
2337 }
2338
2339 #endif
2340
2341 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2342 {
2343         struct perf_event *event = vma->vm_file->private_data;
2344         struct perf_mmap_data *data;
2345         int ret = VM_FAULT_SIGBUS;
2346
2347         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2348                 if (vmf->pgoff == 0)
2349                         ret = 0;
2350                 return ret;
2351         }
2352
2353         rcu_read_lock();
2354         data = rcu_dereference(event->data);
2355         if (!data)
2356                 goto unlock;
2357
2358         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2359                 goto unlock;
2360
2361         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2362         if (!vmf->page)
2363                 goto unlock;
2364
2365         get_page(vmf->page);
2366         vmf->page->mapping = vma->vm_file->f_mapping;
2367         vmf->page->index   = vmf->pgoff;
2368
2369         ret = 0;
2370 unlock:
2371         rcu_read_unlock();
2372
2373         return ret;
2374 }
2375
2376 static void
2377 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2378 {
2379         long max_size = perf_data_size(data);
2380
2381         atomic_set(&data->lock, -1);
2382
2383         if (event->attr.watermark) {
2384                 data->watermark = min_t(long, max_size,
2385                                         event->attr.wakeup_watermark);
2386         }
2387
2388         if (!data->watermark)
2389                 data->watermark = max_size / 2;
2390
2391
2392         rcu_assign_pointer(event->data, data);
2393 }
2394
2395 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2396 {
2397         struct perf_mmap_data *data;
2398
2399         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2400         perf_mmap_data_free(data);
2401 }
2402
2403 static void perf_mmap_data_release(struct perf_event *event)
2404 {
2405         struct perf_mmap_data *data = event->data;
2406
2407         WARN_ON(atomic_read(&event->mmap_count));
2408
2409         rcu_assign_pointer(event->data, NULL);
2410         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2411 }
2412
2413 static void perf_mmap_open(struct vm_area_struct *vma)
2414 {
2415         struct perf_event *event = vma->vm_file->private_data;
2416
2417         atomic_inc(&event->mmap_count);
2418 }
2419
2420 static void perf_mmap_close(struct vm_area_struct *vma)
2421 {
2422         struct perf_event *event = vma->vm_file->private_data;
2423
2424         WARN_ON_ONCE(event->ctx->parent_ctx);
2425         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2426                 unsigned long size = perf_data_size(event->data);
2427                 struct user_struct *user = current_user();
2428
2429                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2430                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2431                 perf_mmap_data_release(event);
2432                 mutex_unlock(&event->mmap_mutex);
2433         }
2434 }
2435
2436 static const struct vm_operations_struct perf_mmap_vmops = {
2437         .open           = perf_mmap_open,
2438         .close          = perf_mmap_close,
2439         .fault          = perf_mmap_fault,
2440         .page_mkwrite   = perf_mmap_fault,
2441 };
2442
2443 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2444 {
2445         struct perf_event *event = file->private_data;
2446         unsigned long user_locked, user_lock_limit;
2447         struct user_struct *user = current_user();
2448         unsigned long locked, lock_limit;
2449         struct perf_mmap_data *data;
2450         unsigned long vma_size;
2451         unsigned long nr_pages;
2452         long user_extra, extra;
2453         int ret = 0;
2454
2455         if (!(vma->vm_flags & VM_SHARED))
2456                 return -EINVAL;
2457
2458         vma_size = vma->vm_end - vma->vm_start;
2459         nr_pages = (vma_size / PAGE_SIZE) - 1;
2460
2461         /*
2462          * If we have data pages ensure they're a power-of-two number, so we
2463          * can do bitmasks instead of modulo.
2464          */
2465         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2466                 return -EINVAL;
2467
2468         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2469                 return -EINVAL;
2470
2471         if (vma->vm_pgoff != 0)
2472                 return -EINVAL;
2473
2474         WARN_ON_ONCE(event->ctx->parent_ctx);
2475         mutex_lock(&event->mmap_mutex);
2476         if (event->output) {
2477                 ret = -EINVAL;
2478                 goto unlock;
2479         }
2480
2481         if (atomic_inc_not_zero(&event->mmap_count)) {
2482                 if (nr_pages != event->data->nr_pages)
2483                         ret = -EINVAL;
2484                 goto unlock;
2485         }
2486
2487         user_extra = nr_pages + 1;
2488         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2489
2490         /*
2491          * Increase the limit linearly with more CPUs:
2492          */
2493         user_lock_limit *= num_online_cpus();
2494
2495         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2496
2497         extra = 0;
2498         if (user_locked > user_lock_limit)
2499                 extra = user_locked - user_lock_limit;
2500
2501         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2502         lock_limit >>= PAGE_SHIFT;
2503         locked = vma->vm_mm->locked_vm + extra;
2504
2505         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2506                 !capable(CAP_IPC_LOCK)) {
2507                 ret = -EPERM;
2508                 goto unlock;
2509         }
2510
2511         WARN_ON(event->data);
2512
2513         data = perf_mmap_data_alloc(event, nr_pages);
2514         ret = -ENOMEM;
2515         if (!data)
2516                 goto unlock;
2517
2518         ret = 0;
2519         perf_mmap_data_init(event, data);
2520
2521         atomic_set(&event->mmap_count, 1);
2522         atomic_long_add(user_extra, &user->locked_vm);
2523         vma->vm_mm->locked_vm += extra;
2524         event->data->nr_locked = extra;
2525         if (vma->vm_flags & VM_WRITE)
2526                 event->data->writable = 1;
2527
2528 unlock:
2529         mutex_unlock(&event->mmap_mutex);
2530
2531         vma->vm_flags |= VM_RESERVED;
2532         vma->vm_ops = &perf_mmap_vmops;
2533
2534         return ret;
2535 }
2536
2537 static int perf_fasync(int fd, struct file *filp, int on)
2538 {
2539         struct inode *inode = filp->f_path.dentry->d_inode;
2540         struct perf_event *event = filp->private_data;
2541         int retval;
2542
2543         mutex_lock(&inode->i_mutex);
2544         retval = fasync_helper(fd, filp, on, &event->fasync);
2545         mutex_unlock(&inode->i_mutex);
2546
2547         if (retval < 0)
2548                 return retval;
2549
2550         return 0;
2551 }
2552
2553 static const struct file_operations perf_fops = {
2554         .release                = perf_release,
2555         .read                   = perf_read,
2556         .poll                   = perf_poll,
2557         .unlocked_ioctl         = perf_ioctl,
2558         .compat_ioctl           = perf_ioctl,
2559         .mmap                   = perf_mmap,
2560         .fasync                 = perf_fasync,
2561 };
2562
2563 /*
2564  * Perf event wakeup
2565  *
2566  * If there's data, ensure we set the poll() state and publish everything
2567  * to user-space before waking everybody up.
2568  */
2569
2570 void perf_event_wakeup(struct perf_event *event)
2571 {
2572         wake_up_all(&event->waitq);
2573
2574         if (event->pending_kill) {
2575                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2576                 event->pending_kill = 0;
2577         }
2578 }
2579
2580 /*
2581  * Pending wakeups
2582  *
2583  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2584  *
2585  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2586  * single linked list and use cmpxchg() to add entries lockless.
2587  */
2588
2589 static void perf_pending_event(struct perf_pending_entry *entry)
2590 {
2591         struct perf_event *event = container_of(entry,
2592                         struct perf_event, pending);
2593
2594         if (event->pending_disable) {
2595                 event->pending_disable = 0;
2596                 __perf_event_disable(event);
2597         }
2598
2599         if (event->pending_wakeup) {
2600                 event->pending_wakeup = 0;
2601                 perf_event_wakeup(event);
2602         }
2603 }
2604
2605 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2606
2607 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2608         PENDING_TAIL,
2609 };
2610
2611 static void perf_pending_queue(struct perf_pending_entry *entry,
2612                                void (*func)(struct perf_pending_entry *))
2613 {
2614         struct perf_pending_entry **head;
2615
2616         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2617                 return;
2618
2619         entry->func = func;
2620
2621         head = &get_cpu_var(perf_pending_head);
2622
2623         do {
2624                 entry->next = *head;
2625         } while (cmpxchg(head, entry->next, entry) != entry->next);
2626
2627         set_perf_event_pending();
2628
2629         put_cpu_var(perf_pending_head);
2630 }
2631
2632 static int __perf_pending_run(void)
2633 {
2634         struct perf_pending_entry *list;
2635         int nr = 0;
2636
2637         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2638         while (list != PENDING_TAIL) {
2639                 void (*func)(struct perf_pending_entry *);
2640                 struct perf_pending_entry *entry = list;
2641
2642                 list = list->next;
2643
2644                 func = entry->func;
2645                 entry->next = NULL;
2646                 /*
2647                  * Ensure we observe the unqueue before we issue the wakeup,
2648                  * so that we won't be waiting forever.
2649                  * -- see perf_not_pending().
2650                  */
2651                 smp_wmb();
2652
2653                 func(entry);
2654                 nr++;
2655         }
2656
2657         return nr;
2658 }
2659
2660 static inline int perf_not_pending(struct perf_event *event)
2661 {
2662         /*
2663          * If we flush on whatever cpu we run, there is a chance we don't
2664          * need to wait.
2665          */
2666         get_cpu();
2667         __perf_pending_run();
2668         put_cpu();
2669
2670         /*
2671          * Ensure we see the proper queue state before going to sleep
2672          * so that we do not miss the wakeup. -- see perf_pending_handle()
2673          */
2674         smp_rmb();
2675         return event->pending.next == NULL;
2676 }
2677
2678 static void perf_pending_sync(struct perf_event *event)
2679 {
2680         wait_event(event->waitq, perf_not_pending(event));
2681 }
2682
2683 void perf_event_do_pending(void)
2684 {
2685         __perf_pending_run();
2686 }
2687
2688 /*
2689  * Callchain support -- arch specific
2690  */
2691
2692 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2693 {
2694         return NULL;
2695 }
2696
2697 /*
2698  * Output
2699  */
2700 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2701                               unsigned long offset, unsigned long head)
2702 {
2703         unsigned long mask;
2704
2705         if (!data->writable)
2706                 return true;
2707
2708         mask = perf_data_size(data) - 1;
2709
2710         offset = (offset - tail) & mask;
2711         head   = (head   - tail) & mask;
2712
2713         if ((int)(head - offset) < 0)
2714                 return false;
2715
2716         return true;
2717 }
2718
2719 static void perf_output_wakeup(struct perf_output_handle *handle)
2720 {
2721         atomic_set(&handle->data->poll, POLL_IN);
2722
2723         if (handle->nmi) {
2724                 handle->event->pending_wakeup = 1;
2725                 perf_pending_queue(&handle->event->pending,
2726                                    perf_pending_event);
2727         } else
2728                 perf_event_wakeup(handle->event);
2729 }
2730
2731 /*
2732  * Curious locking construct.
2733  *
2734  * We need to ensure a later event_id doesn't publish a head when a former
2735  * event_id isn't done writing. However since we need to deal with NMIs we
2736  * cannot fully serialize things.
2737  *
2738  * What we do is serialize between CPUs so we only have to deal with NMI
2739  * nesting on a single CPU.
2740  *
2741  * We only publish the head (and generate a wakeup) when the outer-most
2742  * event_id completes.
2743  */
2744 static void perf_output_lock(struct perf_output_handle *handle)
2745 {
2746         struct perf_mmap_data *data = handle->data;
2747         int cur, cpu = get_cpu();
2748
2749         handle->locked = 0;
2750
2751         for (;;) {
2752                 cur = atomic_cmpxchg(&data->lock, -1, cpu);
2753                 if (cur == -1) {
2754                         handle->locked = 1;
2755                         break;
2756                 }
2757                 if (cur == cpu)
2758                         break;
2759
2760                 cpu_relax();
2761         }
2762 }
2763
2764 static void perf_output_unlock(struct perf_output_handle *handle)
2765 {
2766         struct perf_mmap_data *data = handle->data;
2767         unsigned long head;
2768         int cpu;
2769
2770         data->done_head = data->head;
2771
2772         if (!handle->locked)
2773                 goto out;
2774
2775 again:
2776         /*
2777          * The xchg implies a full barrier that ensures all writes are done
2778          * before we publish the new head, matched by a rmb() in userspace when
2779          * reading this position.
2780          */
2781         while ((head = atomic_long_xchg(&data->done_head, 0)))
2782                 data->user_page->data_head = head;
2783
2784         /*
2785          * NMI can happen here, which means we can miss a done_head update.
2786          */
2787
2788         cpu = atomic_xchg(&data->lock, -1);
2789         WARN_ON_ONCE(cpu != smp_processor_id());
2790
2791         /*
2792          * Therefore we have to validate we did not indeed do so.
2793          */
2794         if (unlikely(atomic_long_read(&data->done_head))) {
2795                 /*
2796                  * Since we had it locked, we can lock it again.
2797                  */
2798                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2799                         cpu_relax();
2800
2801                 goto again;
2802         }
2803
2804         if (atomic_xchg(&data->wakeup, 0))
2805                 perf_output_wakeup(handle);
2806 out:
2807         put_cpu();
2808 }
2809
2810 void perf_output_copy(struct perf_output_handle *handle,
2811                       const void *buf, unsigned int len)
2812 {
2813         unsigned int pages_mask;
2814         unsigned long offset;
2815         unsigned int size;
2816         void **pages;
2817
2818         offset          = handle->offset;
2819         pages_mask      = handle->data->nr_pages - 1;
2820         pages           = handle->data->data_pages;
2821
2822         do {
2823                 unsigned long page_offset;
2824                 unsigned long page_size;
2825                 int nr;
2826
2827                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2828                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2829                 page_offset = offset & (page_size - 1);
2830                 size        = min_t(unsigned int, page_size - page_offset, len);
2831
2832                 memcpy(pages[nr] + page_offset, buf, size);
2833
2834                 len         -= size;
2835                 buf         += size;
2836                 offset      += size;
2837         } while (len);
2838
2839         handle->offset = offset;
2840
2841         /*
2842          * Check we didn't copy past our reservation window, taking the
2843          * possible unsigned int wrap into account.
2844          */
2845         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2846 }
2847
2848 int perf_output_begin(struct perf_output_handle *handle,
2849                       struct perf_event *event, unsigned int size,
2850                       int nmi, int sample)
2851 {
2852         struct perf_event *output_event;
2853         struct perf_mmap_data *data;
2854         unsigned long tail, offset, head;
2855         int have_lost;
2856         struct {
2857                 struct perf_event_header header;
2858                 u64                      id;
2859                 u64                      lost;
2860         } lost_event;
2861
2862         rcu_read_lock();
2863         /*
2864          * For inherited events we send all the output towards the parent.
2865          */
2866         if (event->parent)
2867                 event = event->parent;
2868
2869         output_event = rcu_dereference(event->output);
2870         if (output_event)
2871                 event = output_event;
2872
2873         data = rcu_dereference(event->data);
2874         if (!data)
2875                 goto out;
2876
2877         handle->data    = data;
2878         handle->event   = event;
2879         handle->nmi     = nmi;
2880         handle->sample  = sample;
2881
2882         if (!data->nr_pages)
2883                 goto fail;
2884
2885         have_lost = atomic_read(&data->lost);
2886         if (have_lost)
2887                 size += sizeof(lost_event);
2888
2889         perf_output_lock(handle);
2890
2891         do {
2892                 /*
2893                  * Userspace could choose to issue a mb() before updating the
2894                  * tail pointer. So that all reads will be completed before the
2895                  * write is issued.
2896                  */
2897                 tail = ACCESS_ONCE(data->user_page->data_tail);
2898                 smp_rmb();
2899                 offset = head = atomic_long_read(&data->head);
2900                 head += size;
2901                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2902                         goto fail;
2903         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2904
2905         handle->offset  = offset;
2906         handle->head    = head;
2907
2908         if (head - tail > data->watermark)
2909                 atomic_set(&data->wakeup, 1);
2910
2911         if (have_lost) {
2912                 lost_event.header.type = PERF_RECORD_LOST;
2913                 lost_event.header.misc = 0;
2914                 lost_event.header.size = sizeof(lost_event);
2915                 lost_event.id          = event->id;
2916                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2917
2918                 perf_output_put(handle, lost_event);
2919         }
2920
2921         return 0;
2922
2923 fail:
2924         atomic_inc(&data->lost);
2925         perf_output_unlock(handle);
2926 out:
2927         rcu_read_unlock();
2928
2929         return -ENOSPC;
2930 }
2931
2932 void perf_output_end(struct perf_output_handle *handle)
2933 {
2934         struct perf_event *event = handle->event;
2935         struct perf_mmap_data *data = handle->data;
2936
2937         int wakeup_events = event->attr.wakeup_events;
2938
2939         if (handle->sample && wakeup_events) {
2940                 int events = atomic_inc_return(&data->events);
2941                 if (events >= wakeup_events) {
2942                         atomic_sub(wakeup_events, &data->events);
2943                         atomic_set(&data->wakeup, 1);
2944                 }
2945         }
2946
2947         perf_output_unlock(handle);
2948         rcu_read_unlock();
2949 }
2950
2951 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2952 {
2953         /*
2954          * only top level events have the pid namespace they were created in
2955          */
2956         if (event->parent)
2957                 event = event->parent;
2958
2959         return task_tgid_nr_ns(p, event->ns);
2960 }
2961
2962 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2963 {
2964         /*
2965          * only top level events have the pid namespace they were created in
2966          */
2967         if (event->parent)
2968                 event = event->parent;
2969
2970         return task_pid_nr_ns(p, event->ns);
2971 }
2972
2973 static void perf_output_read_one(struct perf_output_handle *handle,
2974                                  struct perf_event *event)
2975 {
2976         u64 read_format = event->attr.read_format;
2977         u64 values[4];
2978         int n = 0;
2979
2980         values[n++] = atomic64_read(&event->count);
2981         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2982                 values[n++] = event->total_time_enabled +
2983                         atomic64_read(&event->child_total_time_enabled);
2984         }
2985         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2986                 values[n++] = event->total_time_running +
2987                         atomic64_read(&event->child_total_time_running);
2988         }
2989         if (read_format & PERF_FORMAT_ID)
2990                 values[n++] = primary_event_id(event);
2991
2992         perf_output_copy(handle, values, n * sizeof(u64));
2993 }
2994
2995 /*
2996  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2997  */
2998 static void perf_output_read_group(struct perf_output_handle *handle,
2999                             struct perf_event *event)
3000 {
3001         struct perf_event *leader = event->group_leader, *sub;
3002         u64 read_format = event->attr.read_format;
3003         u64 values[5];
3004         int n = 0;
3005
3006         values[n++] = 1 + leader->nr_siblings;
3007
3008         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3009                 values[n++] = leader->total_time_enabled;
3010
3011         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3012                 values[n++] = leader->total_time_running;
3013
3014         if (leader != event)
3015                 leader->pmu->read(leader);
3016
3017         values[n++] = atomic64_read(&leader->count);
3018         if (read_format & PERF_FORMAT_ID)
3019                 values[n++] = primary_event_id(leader);
3020
3021         perf_output_copy(handle, values, n * sizeof(u64));
3022
3023         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3024                 n = 0;
3025
3026                 if (sub != event)
3027                         sub->pmu->read(sub);
3028
3029                 values[n++] = atomic64_read(&sub->count);
3030                 if (read_format & PERF_FORMAT_ID)
3031                         values[n++] = primary_event_id(sub);
3032
3033                 perf_output_copy(handle, values, n * sizeof(u64));
3034         }
3035 }
3036
3037 static void perf_output_read(struct perf_output_handle *handle,
3038                              struct perf_event *event)
3039 {
3040         if (event->attr.read_format & PERF_FORMAT_GROUP)
3041                 perf_output_read_group(handle, event);
3042         else
3043                 perf_output_read_one(handle, event);
3044 }
3045
3046 void perf_output_sample(struct perf_output_handle *handle,
3047                         struct perf_event_header *header,
3048                         struct perf_sample_data *data,
3049                         struct perf_event *event)
3050 {
3051         u64 sample_type = data->type;
3052
3053         perf_output_put(handle, *header);
3054
3055         if (sample_type & PERF_SAMPLE_IP)
3056                 perf_output_put(handle, data->ip);
3057
3058         if (sample_type & PERF_SAMPLE_TID)
3059                 perf_output_put(handle, data->tid_entry);
3060
3061         if (sample_type & PERF_SAMPLE_TIME)
3062                 perf_output_put(handle, data->time);
3063
3064         if (sample_type & PERF_SAMPLE_ADDR)
3065                 perf_output_put(handle, data->addr);
3066
3067         if (sample_type & PERF_SAMPLE_ID)
3068                 perf_output_put(handle, data->id);
3069
3070         if (sample_type & PERF_SAMPLE_STREAM_ID)
3071                 perf_output_put(handle, data->stream_id);
3072
3073         if (sample_type & PERF_SAMPLE_CPU)
3074                 perf_output_put(handle, data->cpu_entry);
3075
3076         if (sample_type & PERF_SAMPLE_PERIOD)
3077                 perf_output_put(handle, data->period);
3078
3079         if (sample_type & PERF_SAMPLE_READ)
3080                 perf_output_read(handle, event);
3081
3082         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3083                 if (data->callchain) {
3084                         int size = 1;
3085
3086                         if (data->callchain)
3087                                 size += data->callchain->nr;
3088
3089                         size *= sizeof(u64);
3090
3091                         perf_output_copy(handle, data->callchain, size);
3092                 } else {
3093                         u64 nr = 0;
3094                         perf_output_put(handle, nr);
3095                 }
3096         }
3097
3098         if (sample_type & PERF_SAMPLE_RAW) {
3099                 if (data->raw) {
3100                         perf_output_put(handle, data->raw->size);
3101                         perf_output_copy(handle, data->raw->data,
3102                                          data->raw->size);
3103                 } else {
3104                         struct {
3105                                 u32     size;
3106                                 u32     data;
3107                         } raw = {
3108                                 .size = sizeof(u32),
3109                                 .data = 0,
3110                         };
3111                         perf_output_put(handle, raw);
3112                 }
3113         }
3114 }
3115
3116 void perf_prepare_sample(struct perf_event_header *header,
3117                          struct perf_sample_data *data,
3118                          struct perf_event *event,
3119                          struct pt_regs *regs)
3120 {
3121         u64 sample_type = event->attr.sample_type;
3122
3123         data->type = sample_type;
3124
3125         header->type = PERF_RECORD_SAMPLE;
3126         header->size = sizeof(*header);
3127
3128         header->misc = 0;
3129         header->misc |= perf_misc_flags(regs);
3130
3131         if (sample_type & PERF_SAMPLE_IP) {
3132                 data->ip = perf_instruction_pointer(regs);
3133
3134                 header->size += sizeof(data->ip);
3135         }
3136
3137         if (sample_type & PERF_SAMPLE_TID) {
3138                 /* namespace issues */
3139                 data->tid_entry.pid = perf_event_pid(event, current);
3140                 data->tid_entry.tid = perf_event_tid(event, current);
3141
3142                 header->size += sizeof(data->tid_entry);
3143         }
3144
3145         if (sample_type & PERF_SAMPLE_TIME) {
3146                 data->time = perf_clock();
3147
3148                 header->size += sizeof(data->time);
3149         }
3150
3151         if (sample_type & PERF_SAMPLE_ADDR)
3152                 header->size += sizeof(data->addr);
3153
3154         if (sample_type & PERF_SAMPLE_ID) {
3155                 data->id = primary_event_id(event);
3156
3157                 header->size += sizeof(data->id);
3158         }
3159
3160         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3161                 data->stream_id = event->id;
3162
3163                 header->size += sizeof(data->stream_id);
3164         }
3165
3166         if (sample_type & PERF_SAMPLE_CPU) {
3167                 data->cpu_entry.cpu             = raw_smp_processor_id();
3168                 data->cpu_entry.reserved        = 0;
3169
3170                 header->size += sizeof(data->cpu_entry);
3171         }
3172
3173         if (sample_type & PERF_SAMPLE_PERIOD)
3174                 header->size += sizeof(data->period);
3175
3176         if (sample_type & PERF_SAMPLE_READ)
3177                 header->size += perf_event_read_size(event);
3178
3179         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3180                 int size = 1;
3181
3182                 data->callchain = perf_callchain(regs);
3183
3184                 if (data->callchain)
3185                         size += data->callchain->nr;
3186
3187                 header->size += size * sizeof(u64);
3188         }
3189
3190         if (sample_type & PERF_SAMPLE_RAW) {
3191                 int size = sizeof(u32);
3192
3193                 if (data->raw)
3194                         size += data->raw->size;
3195                 else
3196                         size += sizeof(u32);
3197
3198                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3199                 header->size += size;
3200         }
3201 }
3202
3203 static void perf_event_output(struct perf_event *event, int nmi,
3204                                 struct perf_sample_data *data,
3205                                 struct pt_regs *regs)
3206 {
3207         struct perf_output_handle handle;
3208         struct perf_event_header header;
3209
3210         perf_prepare_sample(&header, data, event, regs);
3211
3212         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3213                 return;
3214
3215         perf_output_sample(&handle, &header, data, event);
3216
3217         perf_output_end(&handle);
3218 }
3219
3220 /*
3221  * read event_id
3222  */
3223
3224 struct perf_read_event {
3225         struct perf_event_header        header;
3226
3227         u32                             pid;
3228         u32                             tid;
3229 };
3230
3231 static void
3232 perf_event_read_event(struct perf_event *event,
3233                         struct task_struct *task)
3234 {
3235         struct perf_output_handle handle;
3236         struct perf_read_event read_event = {
3237                 .header = {
3238                         .type = PERF_RECORD_READ,
3239                         .misc = 0,
3240                         .size = sizeof(read_event) + perf_event_read_size(event),
3241                 },
3242                 .pid = perf_event_pid(event, task),
3243                 .tid = perf_event_tid(event, task),
3244         };
3245         int ret;
3246
3247         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3248         if (ret)
3249                 return;
3250
3251         perf_output_put(&handle, read_event);
3252         perf_output_read(&handle, event);
3253
3254         perf_output_end(&handle);
3255 }
3256
3257 /*
3258  * task tracking -- fork/exit
3259  *
3260  * enabled by: attr.comm | attr.mmap | attr.task
3261  */
3262
3263 struct perf_task_event {
3264         struct task_struct              *task;
3265         struct perf_event_context       *task_ctx;
3266
3267         struct {
3268                 struct perf_event_header        header;
3269
3270                 u32                             pid;
3271                 u32                             ppid;
3272                 u32                             tid;
3273                 u32                             ptid;
3274                 u64                             time;
3275         } event_id;
3276 };
3277
3278 static void perf_event_task_output(struct perf_event *event,
3279                                      struct perf_task_event *task_event)
3280 {
3281         struct perf_output_handle handle;
3282         int size;
3283         struct task_struct *task = task_event->task;
3284         int ret;
3285
3286         size  = task_event->event_id.header.size;
3287         ret = perf_output_begin(&handle, event, size, 0, 0);
3288
3289         if (ret)
3290                 return;
3291
3292         task_event->event_id.pid = perf_event_pid(event, task);
3293         task_event->event_id.ppid = perf_event_pid(event, current);
3294
3295         task_event->event_id.tid = perf_event_tid(event, task);
3296         task_event->event_id.ptid = perf_event_tid(event, current);
3297
3298         task_event->event_id.time = perf_clock();
3299
3300         perf_output_put(&handle, task_event->event_id);
3301
3302         perf_output_end(&handle);
3303 }
3304
3305 static int perf_event_task_match(struct perf_event *event)
3306 {
3307         if (event->cpu != -1 && event->cpu != smp_processor_id())
3308                 return 0;
3309
3310         if (event->attr.comm || event->attr.mmap || event->attr.task)
3311                 return 1;
3312
3313         return 0;
3314 }
3315
3316 static void perf_event_task_ctx(struct perf_event_context *ctx,
3317                                   struct perf_task_event *task_event)
3318 {
3319         struct perf_event *event;
3320
3321         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3322                 if (perf_event_task_match(event))
3323                         perf_event_task_output(event, task_event);
3324         }
3325 }
3326
3327 static void perf_event_task_event(struct perf_task_event *task_event)
3328 {
3329         struct perf_cpu_context *cpuctx;
3330         struct perf_event_context *ctx = task_event->task_ctx;
3331
3332         rcu_read_lock();
3333         cpuctx = &get_cpu_var(perf_cpu_context);
3334         perf_event_task_ctx(&cpuctx->ctx, task_event);
3335         if (!ctx)
3336                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3337         if (ctx)
3338                 perf_event_task_ctx(ctx, task_event);
3339         put_cpu_var(perf_cpu_context);
3340         rcu_read_unlock();
3341 }
3342
3343 static void perf_event_task(struct task_struct *task,
3344                               struct perf_event_context *task_ctx,
3345                               int new)
3346 {
3347         struct perf_task_event task_event;
3348
3349         if (!atomic_read(&nr_comm_events) &&
3350             !atomic_read(&nr_mmap_events) &&
3351             !atomic_read(&nr_task_events))
3352                 return;
3353
3354         task_event = (struct perf_task_event){
3355                 .task     = task,
3356                 .task_ctx = task_ctx,
3357                 .event_id    = {
3358                         .header = {
3359                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3360                                 .misc = 0,
3361                                 .size = sizeof(task_event.event_id),
3362                         },
3363                         /* .pid  */
3364                         /* .ppid */
3365                         /* .tid  */
3366                         /* .ptid */
3367                 },
3368         };
3369
3370         perf_event_task_event(&task_event);
3371 }
3372
3373 void perf_event_fork(struct task_struct *task)
3374 {
3375         perf_event_task(task, NULL, 1);
3376 }
3377
3378 /*
3379  * comm tracking
3380  */
3381
3382 struct perf_comm_event {
3383         struct task_struct      *task;
3384         char                    *comm;
3385         int                     comm_size;
3386
3387         struct {
3388                 struct perf_event_header        header;
3389
3390                 u32                             pid;
3391                 u32                             tid;
3392         } event_id;
3393 };
3394
3395 static void perf_event_comm_output(struct perf_event *event,
3396                                      struct perf_comm_event *comm_event)
3397 {
3398         struct perf_output_handle handle;
3399         int size = comm_event->event_id.header.size;
3400         int ret = perf_output_begin(&handle, event, size, 0, 0);
3401
3402         if (ret)
3403                 return;
3404
3405         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3406         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3407
3408         perf_output_put(&handle, comm_event->event_id);
3409         perf_output_copy(&handle, comm_event->comm,
3410                                    comm_event->comm_size);
3411         perf_output_end(&handle);
3412 }
3413
3414 static int perf_event_comm_match(struct perf_event *event)
3415 {
3416         if (event->cpu != -1 && event->cpu != smp_processor_id())
3417                 return 0;
3418
3419         if (event->attr.comm)
3420                 return 1;
3421
3422         return 0;
3423 }
3424
3425 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3426                                   struct perf_comm_event *comm_event)
3427 {
3428         struct perf_event *event;
3429
3430         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3431                 if (perf_event_comm_match(event))
3432                         perf_event_comm_output(event, comm_event);
3433         }
3434 }
3435
3436 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3437 {
3438         struct perf_cpu_context *cpuctx;
3439         struct perf_event_context *ctx;
3440         unsigned int size;
3441         char comm[TASK_COMM_LEN];
3442
3443         memset(comm, 0, sizeof(comm));
3444         strlcpy(comm, comm_event->task->comm, sizeof(comm));
3445         size = ALIGN(strlen(comm)+1, sizeof(u64));
3446
3447         comm_event->comm = comm;
3448         comm_event->comm_size = size;
3449
3450         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3451
3452         rcu_read_lock();
3453         cpuctx = &get_cpu_var(perf_cpu_context);
3454         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3455         ctx = rcu_dereference(current->perf_event_ctxp);
3456         if (ctx)
3457                 perf_event_comm_ctx(ctx, comm_event);
3458         put_cpu_var(perf_cpu_context);
3459         rcu_read_unlock();
3460 }
3461
3462 void perf_event_comm(struct task_struct *task)
3463 {
3464         struct perf_comm_event comm_event;
3465
3466         if (task->perf_event_ctxp)
3467                 perf_event_enable_on_exec(task);
3468
3469         if (!atomic_read(&nr_comm_events))
3470                 return;
3471
3472         comm_event = (struct perf_comm_event){
3473                 .task   = task,
3474                 /* .comm      */
3475                 /* .comm_size */
3476                 .event_id  = {
3477                         .header = {
3478                                 .type = PERF_RECORD_COMM,
3479                                 .misc = 0,
3480                                 /* .size */
3481                         },
3482                         /* .pid */
3483                         /* .tid */
3484                 },
3485         };
3486
3487         perf_event_comm_event(&comm_event);
3488 }
3489
3490 /*
3491  * mmap tracking
3492  */
3493
3494 struct perf_mmap_event {
3495         struct vm_area_struct   *vma;
3496
3497         const char              *file_name;
3498         int                     file_size;
3499
3500         struct {
3501                 struct perf_event_header        header;
3502
3503                 u32                             pid;
3504                 u32                             tid;
3505                 u64                             start;
3506                 u64                             len;
3507                 u64                             pgoff;
3508         } event_id;
3509 };
3510
3511 static void perf_event_mmap_output(struct perf_event *event,
3512                                      struct perf_mmap_event *mmap_event)
3513 {
3514         struct perf_output_handle handle;
3515         int size = mmap_event->event_id.header.size;
3516         int ret = perf_output_begin(&handle, event, size, 0, 0);
3517
3518         if (ret)
3519                 return;
3520
3521         mmap_event->event_id.pid = perf_event_pid(event, current);
3522         mmap_event->event_id.tid = perf_event_tid(event, current);
3523
3524         perf_output_put(&handle, mmap_event->event_id);
3525         perf_output_copy(&handle, mmap_event->file_name,
3526                                    mmap_event->file_size);
3527         perf_output_end(&handle);
3528 }
3529
3530 static int perf_event_mmap_match(struct perf_event *event,
3531                                    struct perf_mmap_event *mmap_event)
3532 {
3533         if (event->cpu != -1 && event->cpu != smp_processor_id())
3534                 return 0;
3535
3536         if (event->attr.mmap)
3537                 return 1;
3538
3539         return 0;
3540 }
3541
3542 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3543                                   struct perf_mmap_event *mmap_event)
3544 {
3545         struct perf_event *event;
3546
3547         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3548                 if (perf_event_mmap_match(event, mmap_event))
3549                         perf_event_mmap_output(event, mmap_event);
3550         }
3551 }
3552
3553 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3554 {
3555         struct perf_cpu_context *cpuctx;
3556         struct perf_event_context *ctx;
3557         struct vm_area_struct *vma = mmap_event->vma;
3558         struct file *file = vma->vm_file;
3559         unsigned int size;
3560         char tmp[16];
3561         char *buf = NULL;
3562         const char *name;
3563
3564         memset(tmp, 0, sizeof(tmp));
3565
3566         if (file) {
3567                 /*
3568                  * d_path works from the end of the buffer backwards, so we
3569                  * need to add enough zero bytes after the string to handle
3570                  * the 64bit alignment we do later.
3571                  */
3572                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3573                 if (!buf) {
3574                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3575                         goto got_name;
3576                 }
3577                 name = d_path(&file->f_path, buf, PATH_MAX);
3578                 if (IS_ERR(name)) {
3579                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3580                         goto got_name;
3581                 }
3582         } else {
3583                 if (arch_vma_name(mmap_event->vma)) {
3584                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3585                                        sizeof(tmp));
3586                         goto got_name;
3587                 }
3588
3589                 if (!vma->vm_mm) {
3590                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3591                         goto got_name;
3592                 }
3593
3594                 name = strncpy(tmp, "//anon", sizeof(tmp));
3595                 goto got_name;
3596         }
3597
3598 got_name:
3599         size = ALIGN(strlen(name)+1, sizeof(u64));
3600
3601         mmap_event->file_name = name;
3602         mmap_event->file_size = size;
3603
3604         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3605
3606         rcu_read_lock();
3607         cpuctx = &get_cpu_var(perf_cpu_context);
3608         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3609         ctx = rcu_dereference(current->perf_event_ctxp);
3610         if (ctx)
3611                 perf_event_mmap_ctx(ctx, mmap_event);
3612         put_cpu_var(perf_cpu_context);
3613         rcu_read_unlock();
3614
3615         kfree(buf);
3616 }
3617
3618 void __perf_event_mmap(struct vm_area_struct *vma)
3619 {
3620         struct perf_mmap_event mmap_event;
3621
3622         if (!atomic_read(&nr_mmap_events))
3623                 return;
3624
3625         mmap_event = (struct perf_mmap_event){
3626                 .vma    = vma,
3627                 /* .file_name */
3628                 /* .file_size */
3629                 .event_id  = {
3630                         .header = {
3631                                 .type = PERF_RECORD_MMAP,
3632                                 .misc = 0,
3633                                 /* .size */
3634                         },
3635                         /* .pid */
3636                         /* .tid */
3637                         .start  = vma->vm_start,
3638                         .len    = vma->vm_end - vma->vm_start,
3639                         .pgoff  = vma->vm_pgoff,
3640                 },
3641         };
3642
3643         perf_event_mmap_event(&mmap_event);
3644 }
3645
3646 /*
3647  * IRQ throttle logging
3648  */
3649
3650 static void perf_log_throttle(struct perf_event *event, int enable)
3651 {
3652         struct perf_output_handle handle;
3653         int ret;
3654
3655         struct {
3656                 struct perf_event_header        header;
3657                 u64                             time;
3658                 u64                             id;
3659                 u64                             stream_id;
3660         } throttle_event = {
3661                 .header = {
3662                         .type = PERF_RECORD_THROTTLE,
3663                         .misc = 0,
3664                         .size = sizeof(throttle_event),
3665                 },
3666                 .time           = perf_clock(),
3667                 .id             = primary_event_id(event),
3668                 .stream_id      = event->id,
3669         };
3670
3671         if (enable)
3672                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3673
3674         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3675         if (ret)
3676                 return;
3677
3678         perf_output_put(&handle, throttle_event);
3679         perf_output_end(&handle);
3680 }
3681
3682 /*
3683  * Generic event overflow handling, sampling.
3684  */
3685
3686 static int __perf_event_overflow(struct perf_event *event, int nmi,
3687                                    int throttle, struct perf_sample_data *data,
3688                                    struct pt_regs *regs)
3689 {
3690         int events = atomic_read(&event->event_limit);
3691         struct hw_perf_event *hwc = &event->hw;
3692         int ret = 0;
3693
3694         throttle = (throttle && event->pmu->unthrottle != NULL);
3695
3696         if (!throttle) {
3697                 hwc->interrupts++;
3698         } else {
3699                 if (hwc->interrupts != MAX_INTERRUPTS) {
3700                         hwc->interrupts++;
3701                         if (HZ * hwc->interrupts >
3702                                         (u64)sysctl_perf_event_sample_rate) {
3703                                 hwc->interrupts = MAX_INTERRUPTS;
3704                                 perf_log_throttle(event, 0);
3705                                 ret = 1;
3706                         }
3707                 } else {
3708                         /*
3709                          * Keep re-disabling events even though on the previous
3710                          * pass we disabled it - just in case we raced with a
3711                          * sched-in and the event got enabled again:
3712                          */
3713                         ret = 1;
3714                 }
3715         }
3716
3717         if (event->attr.freq) {
3718                 u64 now = perf_clock();
3719                 s64 delta = now - hwc->freq_stamp;
3720
3721                 hwc->freq_stamp = now;
3722
3723                 if (delta > 0 && delta < TICK_NSEC)
3724                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3725         }
3726
3727         /*
3728          * XXX event_limit might not quite work as expected on inherited
3729          * events
3730          */
3731
3732         event->pending_kill = POLL_IN;
3733         if (events && atomic_dec_and_test(&event->event_limit)) {
3734                 ret = 1;
3735                 event->pending_kill = POLL_HUP;
3736                 if (nmi) {
3737                         event->pending_disable = 1;
3738                         perf_pending_queue(&event->pending,
3739                                            perf_pending_event);
3740                 } else
3741                         perf_event_disable(event);
3742         }
3743
3744         if (event->overflow_handler)
3745                 event->overflow_handler(event, nmi, data, regs);
3746         else
3747                 perf_event_output(event, nmi, data, regs);
3748
3749         return ret;
3750 }
3751
3752 int perf_event_overflow(struct perf_event *event, int nmi,
3753                           struct perf_sample_data *data,
3754                           struct pt_regs *regs)
3755 {
3756         return __perf_event_overflow(event, nmi, 1, data, regs);
3757 }
3758
3759 /*
3760  * Generic software event infrastructure
3761  */
3762
3763 /*
3764  * We directly increment event->count and keep a second value in
3765  * event->hw.period_left to count intervals. This period event
3766  * is kept in the range [-sample_period, 0] so that we can use the
3767  * sign as trigger.
3768  */
3769
3770 static u64 perf_swevent_set_period(struct perf_event *event)
3771 {
3772         struct hw_perf_event *hwc = &event->hw;
3773         u64 period = hwc->last_period;
3774         u64 nr, offset;
3775         s64 old, val;
3776
3777         hwc->last_period = hwc->sample_period;
3778
3779 again:
3780         old = val = atomic64_read(&hwc->period_left);
3781         if (val < 0)
3782                 return 0;
3783
3784         nr = div64_u64(period + val, period);
3785         offset = nr * period;
3786         val -= offset;
3787         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3788                 goto again;
3789
3790         return nr;
3791 }
3792
3793 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3794                                     int nmi, struct perf_sample_data *data,
3795                                     struct pt_regs *regs)
3796 {
3797         struct hw_perf_event *hwc = &event->hw;
3798         int throttle = 0;
3799
3800         data->period = event->hw.last_period;
3801         if (!overflow)
3802                 overflow = perf_swevent_set_period(event);
3803
3804         if (hwc->interrupts == MAX_INTERRUPTS)
3805                 return;
3806
3807         for (; overflow; overflow--) {
3808                 if (__perf_event_overflow(event, nmi, throttle,
3809                                             data, regs)) {
3810                         /*
3811                          * We inhibit the overflow from happening when
3812                          * hwc->interrupts == MAX_INTERRUPTS.
3813                          */
3814                         break;
3815                 }
3816                 throttle = 1;
3817         }
3818 }
3819
3820 static void perf_swevent_unthrottle(struct perf_event *event)
3821 {
3822         /*
3823          * Nothing to do, we already reset hwc->interrupts.
3824          */
3825 }
3826
3827 static void perf_swevent_add(struct perf_event *event, u64 nr,
3828                                int nmi, struct perf_sample_data *data,
3829                                struct pt_regs *regs)
3830 {
3831         struct hw_perf_event *hwc = &event->hw;
3832
3833         atomic64_add(nr, &event->count);
3834
3835         if (!regs)
3836                 return;
3837
3838         if (!hwc->sample_period)
3839                 return;
3840
3841         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3842                 return perf_swevent_overflow(event, 1, nmi, data, regs);
3843
3844         if (atomic64_add_negative(nr, &hwc->period_left))
3845                 return;
3846
3847         perf_swevent_overflow(event, 0, nmi, data, regs);
3848 }
3849
3850 static int perf_swevent_is_counting(struct perf_event *event)
3851 {
3852         /*
3853          * The event is active, we're good!
3854          */
3855         if (event->state == PERF_EVENT_STATE_ACTIVE)
3856                 return 1;
3857
3858         /*
3859          * The event is off/error, not counting.
3860          */
3861         if (event->state != PERF_EVENT_STATE_INACTIVE)
3862                 return 0;
3863
3864         /*
3865          * The event is inactive, if the context is active
3866          * we're part of a group that didn't make it on the 'pmu',
3867          * not counting.
3868          */
3869         if (event->ctx->is_active)
3870                 return 0;
3871
3872         /*
3873          * We're inactive and the context is too, this means the
3874          * task is scheduled out, we're counting events that happen
3875          * to us, like migration events.
3876          */
3877         return 1;
3878 }
3879
3880 static int perf_tp_event_match(struct perf_event *event,
3881                                 struct perf_sample_data *data);
3882
3883 static int perf_exclude_event(struct perf_event *event,
3884                               struct pt_regs *regs)
3885 {
3886         if (regs) {
3887                 if (event->attr.exclude_user && user_mode(regs))
3888                         return 1;
3889
3890                 if (event->attr.exclude_kernel && !user_mode(regs))
3891                         return 1;
3892         }
3893
3894         return 0;
3895 }
3896
3897 static int perf_swevent_match(struct perf_event *event,
3898                                 enum perf_type_id type,
3899                                 u32 event_id,
3900                                 struct perf_sample_data *data,
3901                                 struct pt_regs *regs)
3902 {
3903         if (event->cpu != -1 && event->cpu != smp_processor_id())
3904                 return 0;
3905
3906         if (!perf_swevent_is_counting(event))
3907                 return 0;
3908
3909         if (event->attr.type != type)
3910                 return 0;
3911
3912         if (event->attr.config != event_id)
3913                 return 0;
3914
3915         if (perf_exclude_event(event, regs))
3916                 return 0;
3917
3918         if (event->attr.type == PERF_TYPE_TRACEPOINT &&
3919             !perf_tp_event_match(event, data))
3920                 return 0;
3921
3922         return 1;
3923 }
3924
3925 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3926                                      enum perf_type_id type,
3927                                      u32 event_id, u64 nr, int nmi,
3928                                      struct perf_sample_data *data,
3929                                      struct pt_regs *regs)
3930 {
3931         struct perf_event *event;
3932
3933         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3934                 if (perf_swevent_match(event, type, event_id, data, regs))
3935                         perf_swevent_add(event, nr, nmi, data, regs);
3936         }
3937 }
3938
3939 int perf_swevent_get_recursion_context(void)
3940 {
3941         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3942         int rctx;
3943
3944         if (in_nmi())
3945                 rctx = 3;
3946         else if (in_irq())
3947                 rctx = 2;
3948         else if (in_softirq())
3949                 rctx = 1;
3950         else
3951                 rctx = 0;
3952
3953         if (cpuctx->recursion[rctx]) {
3954                 put_cpu_var(perf_cpu_context);
3955                 return -1;
3956         }
3957
3958         cpuctx->recursion[rctx]++;
3959         barrier();
3960
3961         return rctx;
3962 }
3963 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
3964
3965 void perf_swevent_put_recursion_context(int rctx)
3966 {
3967         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
3968         barrier();
3969         cpuctx->recursion[rctx]--;
3970         put_cpu_var(perf_cpu_context);
3971 }
3972 EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
3973
3974 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3975                                     u64 nr, int nmi,
3976                                     struct perf_sample_data *data,
3977                                     struct pt_regs *regs)
3978 {
3979         struct perf_cpu_context *cpuctx;
3980         struct perf_event_context *ctx;
3981
3982         cpuctx = &__get_cpu_var(perf_cpu_context);
3983         rcu_read_lock();
3984         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3985                                  nr, nmi, data, regs);
3986         /*
3987          * doesn't really matter which of the child contexts the
3988          * events ends up in.
3989          */
3990         ctx = rcu_dereference(current->perf_event_ctxp);
3991         if (ctx)
3992                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3993         rcu_read_unlock();
3994 }
3995
3996 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3997                             struct pt_regs *regs, u64 addr)
3998 {
3999         struct perf_sample_data data;
4000         int rctx;
4001
4002         rctx = perf_swevent_get_recursion_context();
4003         if (rctx < 0)
4004                 return;
4005
4006         data.addr = addr;
4007         data.raw  = NULL;
4008
4009         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4010
4011         perf_swevent_put_recursion_context(rctx);
4012 }
4013
4014 static void perf_swevent_read(struct perf_event *event)
4015 {
4016 }
4017
4018 static int perf_swevent_enable(struct perf_event *event)
4019 {
4020         struct hw_perf_event *hwc = &event->hw;
4021
4022         if (hwc->sample_period) {
4023                 hwc->last_period = hwc->sample_period;
4024                 perf_swevent_set_period(event);
4025         }
4026         return 0;
4027 }
4028
4029 static void perf_swevent_disable(struct perf_event *event)
4030 {
4031 }
4032
4033 static const struct pmu perf_ops_generic = {
4034         .enable         = perf_swevent_enable,
4035         .disable        = perf_swevent_disable,
4036         .read           = perf_swevent_read,
4037         .unthrottle     = perf_swevent_unthrottle,
4038 };
4039
4040 /*
4041  * hrtimer based swevent callback
4042  */
4043
4044 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4045 {
4046         enum hrtimer_restart ret = HRTIMER_RESTART;
4047         struct perf_sample_data data;
4048         struct pt_regs *regs;
4049         struct perf_event *event;
4050         u64 period;
4051
4052         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
4053         event->pmu->read(event);
4054
4055         data.addr = 0;
4056         data.raw = NULL;
4057         data.period = event->hw.last_period;
4058         regs = get_irq_regs();
4059         /*
4060          * In case we exclude kernel IPs or are somehow not in interrupt
4061          * context, provide the next best thing, the user IP.
4062          */
4063         if ((event->attr.exclude_kernel || !regs) &&
4064                         !event->attr.exclude_user)
4065                 regs = task_pt_regs(current);
4066
4067         if (regs) {
4068                 if (!(event->attr.exclude_idle && current->pid == 0))
4069                         if (perf_event_overflow(event, 0, &data, regs))
4070                                 ret = HRTIMER_NORESTART;
4071         }
4072
4073         period = max_t(u64, 10000, event->hw.sample_period);
4074         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4075
4076         return ret;
4077 }
4078
4079 static void perf_swevent_start_hrtimer(struct perf_event *event)
4080 {
4081         struct hw_perf_event *hwc = &event->hw;
4082
4083         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4084         hwc->hrtimer.function = perf_swevent_hrtimer;
4085         if (hwc->sample_period) {
4086                 u64 period;
4087
4088                 if (hwc->remaining) {
4089                         if (hwc->remaining < 0)
4090                                 period = 10000;
4091                         else
4092                                 period = hwc->remaining;
4093                         hwc->remaining = 0;
4094                 } else {
4095                         period = max_t(u64, 10000, hwc->sample_period);
4096                 }
4097                 __hrtimer_start_range_ns(&hwc->hrtimer,
4098                                 ns_to_ktime(period), 0,
4099                                 HRTIMER_MODE_REL, 0);
4100         }
4101 }
4102
4103 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4104 {
4105         struct hw_perf_event *hwc = &event->hw;
4106
4107         if (hwc->sample_period) {
4108                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4109                 hwc->remaining = ktime_to_ns(remaining);
4110
4111                 hrtimer_cancel(&hwc->hrtimer);
4112         }
4113 }
4114
4115 /*
4116  * Software event: cpu wall time clock
4117  */
4118
4119 static void cpu_clock_perf_event_update(struct perf_event *event)
4120 {
4121         int cpu = raw_smp_processor_id();
4122         s64 prev;
4123         u64 now;
4124
4125         now = cpu_clock(cpu);
4126         prev = atomic64_xchg(&event->hw.prev_count, now);
4127         atomic64_add(now - prev, &event->count);
4128 }
4129
4130 static int cpu_clock_perf_event_enable(struct perf_event *event)
4131 {
4132         struct hw_perf_event *hwc = &event->hw;
4133         int cpu = raw_smp_processor_id();
4134
4135         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4136         perf_swevent_start_hrtimer(event);
4137
4138         return 0;
4139 }
4140
4141 static void cpu_clock_perf_event_disable(struct perf_event *event)
4142 {
4143         perf_swevent_cancel_hrtimer(event);
4144         cpu_clock_perf_event_update(event);
4145 }
4146
4147 static void cpu_clock_perf_event_read(struct perf_event *event)
4148 {
4149         cpu_clock_perf_event_update(event);
4150 }
4151
4152 static const struct pmu perf_ops_cpu_clock = {
4153         .enable         = cpu_clock_perf_event_enable,
4154         .disable        = cpu_clock_perf_event_disable,
4155         .read           = cpu_clock_perf_event_read,
4156 };
4157
4158 /*
4159  * Software event: task time clock
4160  */
4161
4162 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4163 {
4164         u64 prev;
4165         s64 delta;
4166
4167         prev = atomic64_xchg(&event->hw.prev_count, now);
4168         delta = now - prev;
4169         atomic64_add(delta, &event->count);
4170 }
4171
4172 static int task_clock_perf_event_enable(struct perf_event *event)
4173 {
4174         struct hw_perf_event *hwc = &event->hw;
4175         u64 now;
4176
4177         now = event->ctx->time;
4178
4179         atomic64_set(&hwc->prev_count, now);
4180
4181         perf_swevent_start_hrtimer(event);
4182
4183         return 0;
4184 }
4185
4186 static void task_clock_perf_event_disable(struct perf_event *event)
4187 {
4188         perf_swevent_cancel_hrtimer(event);
4189         task_clock_perf_event_update(event, event->ctx->time);
4190
4191 }
4192
4193 static void task_clock_perf_event_read(struct perf_event *event)
4194 {
4195         u64 time;
4196
4197         if (!in_nmi()) {
4198                 update_context_time(event->ctx);
4199                 time = event->ctx->time;
4200         } else {
4201                 u64 now = perf_clock();
4202                 u64 delta = now - event->ctx->timestamp;
4203                 time = event->ctx->time + delta;
4204         }
4205
4206         task_clock_perf_event_update(event, time);
4207 }
4208
4209 static const struct pmu perf_ops_task_clock = {
4210         .enable         = task_clock_perf_event_enable,
4211         .disable        = task_clock_perf_event_disable,
4212         .read           = task_clock_perf_event_read,
4213 };
4214
4215 #ifdef CONFIG_EVENT_TRACING
4216
4217 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4218                           int entry_size)
4219 {
4220         struct perf_raw_record raw = {
4221                 .size = entry_size,
4222                 .data = record,
4223         };
4224
4225         struct perf_sample_data data = {
4226                 .addr = addr,
4227                 .raw = &raw,
4228         };
4229
4230         struct pt_regs *regs = get_irq_regs();
4231
4232         if (!regs)
4233                 regs = task_pt_regs(current);
4234
4235         /* Trace events already protected against recursion */
4236         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4237                                 &data, regs);
4238 }
4239 EXPORT_SYMBOL_GPL(perf_tp_event);
4240
4241 static int perf_tp_event_match(struct perf_event *event,
4242                                 struct perf_sample_data *data)
4243 {
4244         void *record = data->raw->data;
4245
4246         if (likely(!event->filter) || filter_match_preds(event->filter, record))
4247                 return 1;
4248         return 0;
4249 }
4250
4251 static void tp_perf_event_destroy(struct perf_event *event)
4252 {
4253         ftrace_profile_disable(event->attr.config);
4254 }
4255
4256 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4257 {
4258         /*
4259          * Raw tracepoint data is a severe data leak, only allow root to
4260          * have these.
4261          */
4262         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4263                         perf_paranoid_tracepoint_raw() &&
4264                         !capable(CAP_SYS_ADMIN))
4265                 return ERR_PTR(-EPERM);
4266
4267         if (ftrace_profile_enable(event->attr.config))
4268                 return NULL;
4269
4270         event->destroy = tp_perf_event_destroy;
4271
4272         return &perf_ops_generic;
4273 }
4274
4275 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4276 {
4277         char *filter_str;
4278         int ret;
4279
4280         if (event->attr.type != PERF_TYPE_TRACEPOINT)
4281                 return -EINVAL;
4282
4283         filter_str = strndup_user(arg, PAGE_SIZE);
4284         if (IS_ERR(filter_str))
4285                 return PTR_ERR(filter_str);
4286
4287         ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4288
4289         kfree(filter_str);
4290         return ret;
4291 }
4292
4293 static void perf_event_free_filter(struct perf_event *event)
4294 {
4295         ftrace_profile_free_filter(event);
4296 }
4297
4298 #else
4299
4300 static int perf_tp_event_match(struct perf_event *event,
4301                                 struct perf_sample_data *data)
4302 {
4303         return 1;
4304 }
4305
4306 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4307 {
4308         return NULL;
4309 }
4310
4311 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
4312 {
4313         return -ENOENT;
4314 }
4315
4316 static void perf_event_free_filter(struct perf_event *event)
4317 {
4318 }
4319
4320 #endif /* CONFIG_EVENT_TRACING */
4321
4322 #ifdef CONFIG_HAVE_HW_BREAKPOINT
4323 static void bp_perf_event_destroy(struct perf_event *event)
4324 {
4325         release_bp_slot(event);
4326 }
4327
4328 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4329 {
4330         int err;
4331
4332         err = register_perf_hw_breakpoint(bp);
4333         if (err)
4334                 return ERR_PTR(err);
4335
4336         bp->destroy = bp_perf_event_destroy;
4337
4338         return &perf_ops_bp;
4339 }
4340
4341 void perf_bp_event(struct perf_event *bp, void *data)
4342 {
4343         struct perf_sample_data sample;
4344         struct pt_regs *regs = data;
4345
4346         sample.raw = NULL;
4347         sample.addr = bp->attr.bp_addr;
4348
4349         if (!perf_exclude_event(bp, regs))
4350                 perf_swevent_add(bp, 1, 1, &sample, regs);
4351 }
4352 #else
4353 static const struct pmu *bp_perf_event_init(struct perf_event *bp)
4354 {
4355         return NULL;
4356 }
4357
4358 void perf_bp_event(struct perf_event *bp, void *regs)
4359 {
4360 }
4361 #endif
4362
4363 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4364
4365 static void sw_perf_event_destroy(struct perf_event *event)
4366 {
4367         u64 event_id = event->attr.config;
4368
4369         WARN_ON(event->parent);
4370
4371         atomic_dec(&perf_swevent_enabled[event_id]);
4372 }
4373
4374 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4375 {
4376         const struct pmu *pmu = NULL;
4377         u64 event_id = event->attr.config;
4378
4379         /*
4380          * Software events (currently) can't in general distinguish
4381          * between user, kernel and hypervisor events.
4382          * However, context switches and cpu migrations are considered
4383          * to be kernel events, and page faults are never hypervisor
4384          * events.
4385          */
4386         switch (event_id) {
4387         case PERF_COUNT_SW_CPU_CLOCK:
4388                 pmu = &perf_ops_cpu_clock;
4389
4390                 break;
4391         case PERF_COUNT_SW_TASK_CLOCK:
4392                 /*
4393                  * If the user instantiates this as a per-cpu event,
4394                  * use the cpu_clock event instead.
4395                  */
4396                 if (event->ctx->task)
4397                         pmu = &perf_ops_task_clock;
4398                 else
4399                         pmu = &perf_ops_cpu_clock;
4400
4401                 break;
4402         case PERF_COUNT_SW_PAGE_FAULTS:
4403         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4404         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4405         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4406         case PERF_COUNT_SW_CPU_MIGRATIONS:
4407         case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4408         case PERF_COUNT_SW_EMULATION_FAULTS:
4409                 if (!event->parent) {
4410                         atomic_inc(&perf_swevent_enabled[event_id]);
4411                         event->destroy = sw_perf_event_destroy;
4412                 }
4413                 pmu = &perf_ops_generic;
4414                 break;
4415         }
4416
4417         return pmu;
4418 }
4419
4420 /*
4421  * Allocate and initialize a event structure
4422  */
4423 static struct perf_event *
4424 perf_event_alloc(struct perf_event_attr *attr,
4425                    int cpu,
4426                    struct perf_event_context *ctx,
4427                    struct perf_event *group_leader,
4428                    struct perf_event *parent_event,
4429                    perf_overflow_handler_t overflow_handler,
4430                    gfp_t gfpflags)
4431 {
4432         const struct pmu *pmu;
4433         struct perf_event *event;
4434         struct hw_perf_event *hwc;
4435         long err;
4436
4437         event = kzalloc(sizeof(*event), gfpflags);
4438         if (!event)
4439                 return ERR_PTR(-ENOMEM);
4440
4441         /*
4442          * Single events are their own group leaders, with an
4443          * empty sibling list:
4444          */
4445         if (!group_leader)
4446                 group_leader = event;
4447
4448         mutex_init(&event->child_mutex);
4449         INIT_LIST_HEAD(&event->child_list);
4450
4451         INIT_LIST_HEAD(&event->group_entry);
4452         INIT_LIST_HEAD(&event->event_entry);
4453         INIT_LIST_HEAD(&event->sibling_list);
4454         init_waitqueue_head(&event->waitq);
4455
4456         mutex_init(&event->mmap_mutex);
4457
4458         event->cpu              = cpu;
4459         event->attr             = *attr;
4460         event->group_leader     = group_leader;
4461         event->pmu              = NULL;
4462         event->ctx              = ctx;
4463         event->oncpu            = -1;
4464
4465         event->parent           = parent_event;
4466
4467         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4468         event->id               = atomic64_inc_return(&perf_event_id);
4469
4470         event->state            = PERF_EVENT_STATE_INACTIVE;
4471
4472         if (!overflow_handler && parent_event)
4473                 overflow_handler = parent_event->overflow_handler;
4474         
4475         event->overflow_handler = overflow_handler;
4476
4477         if (attr->disabled)
4478                 event->state = PERF_EVENT_STATE_OFF;
4479
4480         pmu = NULL;
4481
4482         hwc = &event->hw;
4483         hwc->sample_period = attr->sample_period;
4484         if (attr->freq && attr->sample_freq)
4485                 hwc->sample_period = 1;
4486         hwc->last_period = hwc->sample_period;
4487
4488         atomic64_set(&hwc->period_left, hwc->sample_period);
4489
4490         /*
4491          * we currently do not support PERF_FORMAT_GROUP on inherited events
4492          */
4493         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4494                 goto done;
4495
4496         switch (attr->type) {
4497         case PERF_TYPE_RAW:
4498         case PERF_TYPE_HARDWARE:
4499         case PERF_TYPE_HW_CACHE:
4500                 pmu = hw_perf_event_init(event);
4501                 break;
4502
4503         case PERF_TYPE_SOFTWARE:
4504                 pmu = sw_perf_event_init(event);
4505                 break;
4506
4507         case PERF_TYPE_TRACEPOINT:
4508                 pmu = tp_perf_event_init(event);
4509                 break;
4510
4511         case PERF_TYPE_BREAKPOINT:
4512                 pmu = bp_perf_event_init(event);
4513                 break;
4514
4515
4516         default:
4517                 break;
4518         }
4519 done:
4520         err = 0;
4521         if (!pmu)
4522                 err = -EINVAL;
4523         else if (IS_ERR(pmu))
4524                 err = PTR_ERR(pmu);
4525
4526         if (err) {
4527                 if (event->ns)
4528                         put_pid_ns(event->ns);
4529                 kfree(event);
4530                 return ERR_PTR(err);
4531         }
4532
4533         event->pmu = pmu;
4534
4535         if (!event->parent) {
4536                 atomic_inc(&nr_events);
4537                 if (event->attr.mmap)
4538                         atomic_inc(&nr_mmap_events);
4539                 if (event->attr.comm)
4540                         atomic_inc(&nr_comm_events);
4541                 if (event->attr.task)
4542                         atomic_inc(&nr_task_events);
4543         }
4544
4545         return event;
4546 }
4547
4548 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4549                           struct perf_event_attr *attr)
4550 {
4551         u32 size;
4552         int ret;
4553
4554         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4555                 return -EFAULT;
4556
4557         /*
4558          * zero the full structure, so that a short copy will be nice.
4559          */
4560         memset(attr, 0, sizeof(*attr));
4561
4562         ret = get_user(size, &uattr->size);
4563         if (ret)
4564                 return ret;
4565
4566         if (size > PAGE_SIZE)   /* silly large */
4567                 goto err_size;
4568
4569         if (!size)              /* abi compat */
4570                 size = PERF_ATTR_SIZE_VER0;
4571
4572         if (size < PERF_ATTR_SIZE_VER0)
4573                 goto err_size;
4574
4575         /*
4576          * If we're handed a bigger struct than we know of,
4577          * ensure all the unknown bits are 0 - i.e. new
4578          * user-space does not rely on any kernel feature
4579          * extensions we dont know about yet.
4580          */
4581         if (size > sizeof(*attr)) {
4582                 unsigned char __user *addr;
4583                 unsigned char __user *end;
4584                 unsigned char val;
4585
4586                 addr = (void __user *)uattr + sizeof(*attr);
4587                 end  = (void __user *)uattr + size;
4588
4589                 for (; addr < end; addr++) {
4590                         ret = get_user(val, addr);
4591                         if (ret)
4592                                 return ret;
4593                         if (val)
4594                                 goto err_size;
4595                 }
4596                 size = sizeof(*attr);
4597         }
4598
4599         ret = copy_from_user(attr, uattr, size);
4600         if (ret)
4601                 return -EFAULT;
4602
4603         /*
4604          * If the type exists, the corresponding creation will verify
4605          * the attr->config.
4606          */
4607         if (attr->type >= PERF_TYPE_MAX)
4608                 return -EINVAL;
4609
4610         if (attr->__reserved_1 || attr->__reserved_2)
4611                 return -EINVAL;
4612
4613         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4614                 return -EINVAL;
4615
4616         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4617                 return -EINVAL;
4618
4619 out:
4620         return ret;
4621
4622 err_size:
4623         put_user(sizeof(*attr), &uattr->size);
4624         ret = -E2BIG;
4625         goto out;
4626 }
4627
4628 static int perf_event_set_output(struct perf_event *event, int output_fd)
4629 {
4630         struct perf_event *output_event = NULL;
4631         struct file *output_file = NULL;
4632         struct perf_event *old_output;
4633         int fput_needed = 0;
4634         int ret = -EINVAL;
4635
4636         if (!output_fd)
4637                 goto set;
4638
4639         output_file = fget_light(output_fd, &fput_needed);
4640         if (!output_file)
4641                 return -EBADF;
4642
4643         if (output_file->f_op != &perf_fops)
4644                 goto out;
4645
4646         output_event = output_file->private_data;
4647
4648         /* Don't chain output fds */
4649         if (output_event->output)
4650                 goto out;
4651
4652         /* Don't set an output fd when we already have an output channel */
4653         if (event->data)
4654                 goto out;
4655
4656         atomic_long_inc(&output_file->f_count);
4657
4658 set:
4659         mutex_lock(&event->mmap_mutex);
4660         old_output = event->output;
4661         rcu_assign_pointer(event->output, output_event);
4662         mutex_unlock(&event->mmap_mutex);
4663
4664         if (old_output) {
4665                 /*
4666                  * we need to make sure no existing perf_output_*()
4667                  * is still referencing this event.
4668                  */
4669                 synchronize_rcu();
4670                 fput(old_output->filp);
4671         }
4672
4673         ret = 0;
4674 out:
4675         fput_light(output_file, fput_needed);
4676         return ret;
4677 }
4678
4679 /**
4680  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4681  *
4682  * @attr_uptr:  event_id type attributes for monitoring/sampling
4683  * @pid:                target pid
4684  * @cpu:                target cpu
4685  * @group_fd:           group leader event fd
4686  */
4687 SYSCALL_DEFINE5(perf_event_open,
4688                 struct perf_event_attr __user *, attr_uptr,
4689                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4690 {
4691         struct perf_event *event, *group_leader;
4692         struct perf_event_attr attr;
4693         struct perf_event_context *ctx;
4694         struct file *event_file = NULL;
4695         struct file *group_file = NULL;
4696         int fput_needed = 0;
4697         int fput_needed2 = 0;
4698         int err;
4699
4700         /* for future expandability... */
4701         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4702                 return -EINVAL;
4703
4704         err = perf_copy_attr(attr_uptr, &attr);
4705         if (err)
4706                 return err;
4707
4708         if (!attr.exclude_kernel) {
4709                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4710                         return -EACCES;
4711         }
4712
4713         if (attr.freq) {
4714                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4715                         return -EINVAL;
4716         }
4717
4718         /*
4719          * Get the target context (task or percpu):
4720          */
4721         ctx = find_get_context(pid, cpu);
4722         if (IS_ERR(ctx))
4723                 return PTR_ERR(ctx);
4724
4725         /*
4726          * Look up the group leader (we will attach this event to it):
4727          */
4728         group_leader = NULL;
4729         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4730                 err = -EINVAL;
4731                 group_file = fget_light(group_fd, &fput_needed);
4732                 if (!group_file)
4733                         goto err_put_context;
4734                 if (group_file->f_op != &perf_fops)
4735                         goto err_put_context;
4736
4737                 group_leader = group_file->private_data;
4738                 /*
4739                  * Do not allow a recursive hierarchy (this new sibling
4740                  * becoming part of another group-sibling):
4741                  */
4742                 if (group_leader->group_leader != group_leader)
4743                         goto err_put_context;
4744                 /*
4745                  * Do not allow to attach to a group in a different
4746                  * task or CPU context:
4747                  */
4748                 if (group_leader->ctx != ctx)
4749                         goto err_put_context;
4750                 /*
4751                  * Only a group leader can be exclusive or pinned
4752                  */
4753                 if (attr.exclusive || attr.pinned)
4754                         goto err_put_context;
4755         }
4756
4757         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4758                                      NULL, NULL, GFP_KERNEL);
4759         err = PTR_ERR(event);
4760         if (IS_ERR(event))
4761                 goto err_put_context;
4762
4763         err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4764         if (err < 0)
4765                 goto err_free_put_context;
4766
4767         event_file = fget_light(err, &fput_needed2);
4768         if (!event_file)
4769                 goto err_free_put_context;
4770
4771         if (flags & PERF_FLAG_FD_OUTPUT) {
4772                 err = perf_event_set_output(event, group_fd);
4773                 if (err)
4774                         goto err_fput_free_put_context;
4775         }
4776
4777         event->filp = event_file;
4778         WARN_ON_ONCE(ctx->parent_ctx);
4779         mutex_lock(&ctx->mutex);
4780         perf_install_in_context(ctx, event, cpu);
4781         ++ctx->generation;
4782         mutex_unlock(&ctx->mutex);
4783
4784         event->owner = current;
4785         get_task_struct(current);
4786         mutex_lock(&current->perf_event_mutex);
4787         list_add_tail(&event->owner_entry, &current->perf_event_list);
4788         mutex_unlock(&current->perf_event_mutex);
4789
4790 err_fput_free_put_context:
4791         fput_light(event_file, fput_needed2);
4792
4793 err_free_put_context:
4794         if (err < 0)
4795                 kfree(event);
4796
4797 err_put_context:
4798         if (err < 0)
4799                 put_ctx(ctx);
4800
4801         fput_light(group_file, fput_needed);
4802
4803         return err;
4804 }
4805
4806 /**
4807  * perf_event_create_kernel_counter
4808  *
4809  * @attr: attributes of the counter to create
4810  * @cpu: cpu in which the counter is bound
4811  * @pid: task to profile
4812  */
4813 struct perf_event *
4814 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4815                                  pid_t pid,
4816                                  perf_overflow_handler_t overflow_handler)
4817 {
4818         struct perf_event *event;
4819         struct perf_event_context *ctx;
4820         int err;
4821
4822         /*
4823          * Get the target context (task or percpu):
4824          */
4825
4826         ctx = find_get_context(pid, cpu);
4827         if (IS_ERR(ctx)) {
4828                 err = PTR_ERR(ctx);
4829                 goto err_exit;
4830         }
4831
4832         event = perf_event_alloc(attr, cpu, ctx, NULL,
4833                                  NULL, overflow_handler, GFP_KERNEL);
4834         if (IS_ERR(event)) {
4835                 err = PTR_ERR(event);
4836                 goto err_put_context;
4837         }
4838
4839         event->filp = NULL;
4840         WARN_ON_ONCE(ctx->parent_ctx);
4841         mutex_lock(&ctx->mutex);
4842         perf_install_in_context(ctx, event, cpu);
4843         ++ctx->generation;
4844         mutex_unlock(&ctx->mutex);
4845
4846         event->owner = current;
4847         get_task_struct(current);
4848         mutex_lock(&current->perf_event_mutex);
4849         list_add_tail(&event->owner_entry, &current->perf_event_list);
4850         mutex_unlock(&current->perf_event_mutex);
4851
4852         return event;
4853
4854  err_put_context:
4855         put_ctx(ctx);
4856  err_exit:
4857         return ERR_PTR(err);
4858 }
4859 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4860
4861 /*
4862  * inherit a event from parent task to child task:
4863  */
4864 static struct perf_event *
4865 inherit_event(struct perf_event *parent_event,
4866               struct task_struct *parent,
4867               struct perf_event_context *parent_ctx,
4868               struct task_struct *child,
4869               struct perf_event *group_leader,
4870               struct perf_event_context *child_ctx)
4871 {
4872         struct perf_event *child_event;
4873
4874         /*
4875          * Instead of creating recursive hierarchies of events,
4876          * we link inherited events back to the original parent,
4877          * which has a filp for sure, which we use as the reference
4878          * count:
4879          */
4880         if (parent_event->parent)
4881                 parent_event = parent_event->parent;
4882
4883         child_event = perf_event_alloc(&parent_event->attr,
4884                                            parent_event->cpu, child_ctx,
4885                                            group_leader, parent_event,
4886                                            NULL, GFP_KERNEL);
4887         if (IS_ERR(child_event))
4888                 return child_event;
4889         get_ctx(child_ctx);
4890
4891         /*
4892          * Make the child state follow the state of the parent event,
4893          * not its attr.disabled bit.  We hold the parent's mutex,
4894          * so we won't race with perf_event_{en, dis}able_family.
4895          */
4896         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4897                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4898         else
4899                 child_event->state = PERF_EVENT_STATE_OFF;
4900
4901         if (parent_event->attr.freq)
4902                 child_event->hw.sample_period = parent_event->hw.sample_period;
4903
4904         child_event->overflow_handler = parent_event->overflow_handler;
4905
4906         /*
4907          * Link it up in the child's context:
4908          */
4909         add_event_to_ctx(child_event, child_ctx);
4910
4911         /*
4912          * Get a reference to the parent filp - we will fput it
4913          * when the child event exits. This is safe to do because
4914          * we are in the parent and we know that the filp still
4915          * exists and has a nonzero count:
4916          */
4917         atomic_long_inc(&parent_event->filp->f_count);
4918
4919         /*
4920          * Link this into the parent event's child list
4921          */
4922         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4923         mutex_lock(&parent_event->child_mutex);
4924         list_add_tail(&child_event->child_list, &parent_event->child_list);
4925         mutex_unlock(&parent_event->child_mutex);
4926
4927         return child_event;
4928 }
4929
4930 static int inherit_group(struct perf_event *parent_event,
4931               struct task_struct *parent,
4932               struct perf_event_context *parent_ctx,
4933               struct task_struct *child,
4934               struct perf_event_context *child_ctx)
4935 {
4936         struct perf_event *leader;
4937         struct perf_event *sub;
4938         struct perf_event *child_ctr;
4939
4940         leader = inherit_event(parent_event, parent, parent_ctx,
4941                                  child, NULL, child_ctx);
4942         if (IS_ERR(leader))
4943                 return PTR_ERR(leader);
4944         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4945                 child_ctr = inherit_event(sub, parent, parent_ctx,
4946                                             child, leader, child_ctx);
4947                 if (IS_ERR(child_ctr))
4948                         return PTR_ERR(child_ctr);
4949         }
4950         return 0;
4951 }
4952
4953 static void sync_child_event(struct perf_event *child_event,
4954                                struct task_struct *child)
4955 {
4956         struct perf_event *parent_event = child_event->parent;
4957         u64 child_val;
4958
4959         if (child_event->attr.inherit_stat)
4960                 perf_event_read_event(child_event, child);
4961
4962         child_val = atomic64_read(&child_event->count);
4963
4964         /*
4965          * Add back the child's count to the parent's count:
4966          */
4967         atomic64_add(child_val, &parent_event->count);
4968         atomic64_add(child_event->total_time_enabled,
4969                      &parent_event->child_total_time_enabled);
4970         atomic64_add(child_event->total_time_running,
4971                      &parent_event->child_total_time_running);
4972
4973         /*
4974          * Remove this event from the parent's list
4975          */
4976         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4977         mutex_lock(&parent_event->child_mutex);
4978         list_del_init(&child_event->child_list);
4979         mutex_unlock(&parent_event->child_mutex);
4980
4981         /*
4982          * Release the parent event, if this was the last
4983          * reference to it.
4984          */
4985         fput(parent_event->filp);
4986 }
4987
4988 static void
4989 __perf_event_exit_task(struct perf_event *child_event,
4990                          struct perf_event_context *child_ctx,
4991                          struct task_struct *child)
4992 {
4993         struct perf_event *parent_event;
4994
4995         perf_event_remove_from_context(child_event);
4996
4997         parent_event = child_event->parent;
4998         /*
4999          * It can happen that parent exits first, and has events
5000          * that are still around due to the child reference. These
5001          * events need to be zapped - but otherwise linger.
5002          */
5003         if (parent_event) {
5004                 sync_child_event(child_event, child);
5005                 free_event(child_event);
5006         }
5007 }
5008
5009 /*
5010  * When a child task exits, feed back event values to parent events.
5011  */
5012 void perf_event_exit_task(struct task_struct *child)
5013 {
5014         struct perf_event *child_event, *tmp;
5015         struct perf_event_context *child_ctx;
5016         unsigned long flags;
5017
5018         if (likely(!child->perf_event_ctxp)) {
5019                 perf_event_task(child, NULL, 0);
5020                 return;
5021         }
5022
5023         local_irq_save(flags);
5024         /*
5025          * We can't reschedule here because interrupts are disabled,
5026          * and either child is current or it is a task that can't be
5027          * scheduled, so we are now safe from rescheduling changing
5028          * our context.
5029          */
5030         child_ctx = child->perf_event_ctxp;
5031         __perf_event_task_sched_out(child_ctx);
5032
5033         /*
5034          * Take the context lock here so that if find_get_context is
5035          * reading child->perf_event_ctxp, we wait until it has
5036          * incremented the context's refcount before we do put_ctx below.
5037          */
5038         raw_spin_lock(&child_ctx->lock);
5039         child->perf_event_ctxp = NULL;
5040         /*
5041          * If this context is a clone; unclone it so it can't get
5042          * swapped to another process while we're removing all
5043          * the events from it.
5044          */
5045         unclone_ctx(child_ctx);
5046         update_context_time(child_ctx);
5047         raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5048
5049         /*
5050          * Report the task dead after unscheduling the events so that we
5051          * won't get any samples after PERF_RECORD_EXIT. We can however still
5052          * get a few PERF_RECORD_READ events.
5053          */
5054         perf_event_task(child, child_ctx, 0);
5055
5056         /*
5057          * We can recurse on the same lock type through:
5058          *
5059          *   __perf_event_exit_task()
5060          *     sync_child_event()
5061          *       fput(parent_event->filp)
5062          *         perf_release()
5063          *           mutex_lock(&ctx->mutex)
5064          *
5065          * But since its the parent context it won't be the same instance.
5066          */
5067         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5068
5069 again:
5070         list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5071                                  group_entry)
5072                 __perf_event_exit_task(child_event, child_ctx, child);
5073
5074         list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5075                                  group_entry)
5076                 __perf_event_exit_task(child_event, child_ctx, child);
5077
5078         /*
5079          * If the last event was a group event, it will have appended all
5080          * its siblings to the list, but we obtained 'tmp' before that which
5081          * will still point to the list head terminating the iteration.
5082          */
5083         if (!list_empty(&child_ctx->pinned_groups) ||
5084             !list_empty(&child_ctx->flexible_groups))
5085                 goto again;
5086
5087         mutex_unlock(&child_ctx->mutex);
5088
5089         put_ctx(child_ctx);
5090 }
5091
5092 static void perf_free_event(struct perf_event *event,
5093                             struct perf_event_context *ctx)
5094 {
5095         struct perf_event *parent = event->parent;
5096
5097         if (WARN_ON_ONCE(!parent))
5098                 return;
5099
5100         mutex_lock(&parent->child_mutex);
5101         list_del_init(&event->child_list);
5102         mutex_unlock(&parent->child_mutex);
5103
5104         fput(parent->filp);
5105
5106         list_del_event(event, ctx);
5107         free_event(event);
5108 }
5109
5110 /*
5111  * free an unexposed, unused context as created by inheritance by
5112  * init_task below, used by fork() in case of fail.
5113  */
5114 void perf_event_free_task(struct task_struct *task)
5115 {
5116         struct perf_event_context *ctx = task->perf_event_ctxp;
5117         struct perf_event *event, *tmp;
5118
5119         if (!ctx)
5120                 return;
5121
5122         mutex_lock(&ctx->mutex);
5123 again:
5124         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5125                 perf_free_event(event, ctx);
5126
5127         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5128                                  group_entry)
5129                 perf_free_event(event, ctx);
5130
5131         if (!list_empty(&ctx->pinned_groups) ||
5132             !list_empty(&ctx->flexible_groups))
5133                 goto again;
5134
5135         mutex_unlock(&ctx->mutex);
5136
5137         put_ctx(ctx);
5138 }
5139
5140 static int
5141 inherit_task_group(struct perf_event *event, struct task_struct *parent,
5142                    struct perf_event_context *parent_ctx,
5143                    struct task_struct *child,
5144                    int *inherited_all)
5145 {
5146         int ret;
5147         struct perf_event_context *child_ctx = child->perf_event_ctxp;
5148
5149         if (!event->attr.inherit) {
5150                 *inherited_all = 0;
5151                 return 0;
5152         }
5153
5154         if (!child_ctx) {
5155                 /*
5156                  * This is executed from the parent task context, so
5157                  * inherit events that have been marked for cloning.
5158                  * First allocate and initialize a context for the
5159                  * child.
5160                  */
5161
5162                 child_ctx = kzalloc(sizeof(struct perf_event_context),
5163                                     GFP_KERNEL);
5164                 if (!child_ctx)
5165                         return -ENOMEM;
5166
5167                 __perf_event_init_context(child_ctx, child);
5168                 child->perf_event_ctxp = child_ctx;
5169                 get_task_struct(child);
5170         }
5171
5172         ret = inherit_group(event, parent, parent_ctx,
5173                             child, child_ctx);
5174
5175         if (ret)
5176                 *inherited_all = 0;
5177
5178         return ret;
5179 }
5180
5181
5182 /*
5183  * Initialize the perf_event context in task_struct
5184  */
5185 int perf_event_init_task(struct task_struct *child)
5186 {
5187         struct perf_event_context *child_ctx, *parent_ctx;
5188         struct perf_event_context *cloned_ctx;
5189         struct perf_event *event;
5190         struct task_struct *parent = current;
5191         int inherited_all = 1;
5192         int ret = 0;
5193
5194         child->perf_event_ctxp = NULL;
5195
5196         mutex_init(&child->perf_event_mutex);
5197         INIT_LIST_HEAD(&child->perf_event_list);
5198
5199         if (likely(!parent->perf_event_ctxp))
5200                 return 0;
5201
5202         /*
5203          * If the parent's context is a clone, pin it so it won't get
5204          * swapped under us.
5205          */
5206         parent_ctx = perf_pin_task_context(parent);
5207
5208         /*
5209          * No need to check if parent_ctx != NULL here; since we saw
5210          * it non-NULL earlier, the only reason for it to become NULL
5211          * is if we exit, and since we're currently in the middle of
5212          * a fork we can't be exiting at the same time.
5213          */
5214
5215         /*
5216          * Lock the parent list. No need to lock the child - not PID
5217          * hashed yet and not running, so nobody can access it.
5218          */
5219         mutex_lock(&parent_ctx->mutex);
5220
5221         /*
5222          * We dont have to disable NMIs - we are only looking at
5223          * the list, not manipulating it:
5224          */
5225         list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5226                 ret = inherit_task_group(event, parent, parent_ctx, child,
5227                                          &inherited_all);
5228                 if (ret)
5229                         break;
5230         }
5231
5232         list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5233                 ret = inherit_task_group(event, parent, parent_ctx, child,
5234                                          &inherited_all);
5235                 if (ret)
5236                         break;
5237         }
5238
5239         child_ctx = child->perf_event_ctxp;
5240
5241         if (child_ctx && inherited_all) {
5242                 /*
5243                  * Mark the child context as a clone of the parent
5244                  * context, or of whatever the parent is a clone of.
5245                  * Note that if the parent is a clone, it could get
5246                  * uncloned at any point, but that doesn't matter
5247                  * because the list of events and the generation
5248                  * count can't have changed since we took the mutex.
5249                  */
5250                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5251                 if (cloned_ctx) {
5252                         child_ctx->parent_ctx = cloned_ctx;
5253                         child_ctx->parent_gen = parent_ctx->parent_gen;
5254                 } else {
5255                         child_ctx->parent_ctx = parent_ctx;
5256                         child_ctx->parent_gen = parent_ctx->generation;
5257                 }
5258                 get_ctx(child_ctx->parent_ctx);
5259         }
5260
5261         mutex_unlock(&parent_ctx->mutex);
5262
5263         perf_unpin_context(parent_ctx);
5264
5265         return ret;
5266 }
5267
5268 static void __cpuinit perf_event_init_cpu(int cpu)
5269 {
5270         struct perf_cpu_context *cpuctx;
5271
5272         cpuctx = &per_cpu(perf_cpu_context, cpu);
5273         __perf_event_init_context(&cpuctx->ctx, NULL);
5274
5275         spin_lock(&perf_resource_lock);
5276         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5277         spin_unlock(&perf_resource_lock);
5278
5279         hw_perf_event_setup(cpu);
5280 }
5281
5282 #ifdef CONFIG_HOTPLUG_CPU
5283 static void __perf_event_exit_cpu(void *info)
5284 {
5285         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5286         struct perf_event_context *ctx = &cpuctx->ctx;
5287         struct perf_event *event, *tmp;
5288
5289         list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5290                 __perf_event_remove_from_context(event);
5291         list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5292                 __perf_event_remove_from_context(event);
5293 }
5294 static void perf_event_exit_cpu(int cpu)
5295 {
5296         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5297         struct perf_event_context *ctx = &cpuctx->ctx;
5298
5299         mutex_lock(&ctx->mutex);
5300         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5301         mutex_unlock(&ctx->mutex);
5302 }
5303 #else
5304 static inline void perf_event_exit_cpu(int cpu) { }
5305 #endif
5306
5307 static int __cpuinit
5308 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5309 {
5310         unsigned int cpu = (long)hcpu;
5311
5312         switch (action) {
5313
5314         case CPU_UP_PREPARE:
5315         case CPU_UP_PREPARE_FROZEN:
5316                 perf_event_init_cpu(cpu);
5317                 break;
5318
5319         case CPU_ONLINE:
5320         case CPU_ONLINE_FROZEN:
5321                 hw_perf_event_setup_online(cpu);
5322                 break;
5323
5324         case CPU_DOWN_PREPARE:
5325         case CPU_DOWN_PREPARE_FROZEN:
5326                 perf_event_exit_cpu(cpu);
5327                 break;
5328
5329         default:
5330                 break;
5331         }
5332
5333         return NOTIFY_OK;
5334 }
5335
5336 /*
5337  * This has to have a higher priority than migration_notifier in sched.c.
5338  */
5339 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5340         .notifier_call          = perf_cpu_notify,
5341         .priority               = 20,
5342 };
5343
5344 void __init perf_event_init(void)
5345 {
5346         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5347                         (void *)(long)smp_processor_id());
5348         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5349                         (void *)(long)smp_processor_id());
5350         register_cpu_notifier(&perf_cpu_nb);
5351 }
5352
5353 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5354 {
5355         return sprintf(buf, "%d\n", perf_reserved_percpu);
5356 }
5357
5358 static ssize_t
5359 perf_set_reserve_percpu(struct sysdev_class *class,
5360                         const char *buf,
5361                         size_t count)
5362 {
5363         struct perf_cpu_context *cpuctx;
5364         unsigned long val;
5365         int err, cpu, mpt;
5366
5367         err = strict_strtoul(buf, 10, &val);
5368         if (err)
5369                 return err;
5370         if (val > perf_max_events)
5371                 return -EINVAL;
5372
5373         spin_lock(&perf_resource_lock);
5374         perf_reserved_percpu = val;
5375         for_each_online_cpu(cpu) {
5376                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5377                 raw_spin_lock_irq(&cpuctx->ctx.lock);
5378                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5379                           perf_max_events - perf_reserved_percpu);
5380                 cpuctx->max_pertask = mpt;
5381                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
5382         }
5383         spin_unlock(&perf_resource_lock);
5384
5385         return count;
5386 }
5387
5388 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5389 {
5390         return sprintf(buf, "%d\n", perf_overcommit);
5391 }
5392
5393 static ssize_t
5394 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5395 {
5396         unsigned long val;
5397         int err;
5398
5399         err = strict_strtoul(buf, 10, &val);
5400         if (err)
5401                 return err;
5402         if (val > 1)
5403                 return -EINVAL;
5404
5405         spin_lock(&perf_resource_lock);
5406         perf_overcommit = val;
5407         spin_unlock(&perf_resource_lock);
5408
5409         return count;
5410 }
5411
5412 static SYSDEV_CLASS_ATTR(
5413                                 reserve_percpu,
5414                                 0644,
5415                                 perf_show_reserve_percpu,
5416                                 perf_set_reserve_percpu
5417                         );
5418
5419 static SYSDEV_CLASS_ATTR(
5420                                 overcommit,
5421                                 0644,
5422                                 perf_show_overcommit,
5423                                 perf_set_overcommit
5424                         );
5425
5426 static struct attribute *perfclass_attrs[] = {
5427         &attr_reserve_percpu.attr,
5428         &attr_overcommit.attr,
5429         NULL
5430 };
5431
5432 static struct attribute_group perfclass_attr_group = {
5433         .attrs                  = perfclass_attrs,
5434         .name                   = "perf_events",
5435 };
5436
5437 static int __init perf_event_sysfs_init(void)
5438 {
5439         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5440                                   &perfclass_attr_group);
5441 }
5442 device_initcall(perf_event_sysfs_init);