Merge branch 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6] / kernel / perf_counter.c
1 /*
2  * Performance counter 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/hardirq.h>
24 #include <linux/rculist.h>
25 #include <linux/uaccess.h>
26 #include <linux/syscalls.h>
27 #include <linux/anon_inodes.h>
28 #include <linux/kernel_stat.h>
29 #include <linux/perf_counter.h>
30
31 #include <asm/irq_regs.h>
32
33 /*
34  * Each CPU has a list of per CPU counters:
35  */
36 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
37
38 int perf_max_counters __read_mostly = 1;
39 static int perf_reserved_percpu __read_mostly;
40 static int perf_overcommit __read_mostly = 1;
41
42 static atomic_t nr_counters __read_mostly;
43 static atomic_t nr_mmap_counters __read_mostly;
44 static atomic_t nr_comm_counters __read_mostly;
45
46 /*
47  * perf counter paranoia level:
48  *  0 - not paranoid
49  *  1 - disallow cpu counters to unpriv
50  *  2 - disallow kernel profiling to unpriv
51  */
52 int sysctl_perf_counter_paranoid __read_mostly;
53
54 static inline bool perf_paranoid_cpu(void)
55 {
56         return sysctl_perf_counter_paranoid > 0;
57 }
58
59 static inline bool perf_paranoid_kernel(void)
60 {
61         return sysctl_perf_counter_paranoid > 1;
62 }
63
64 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
65
66 /*
67  * max perf counter sample rate
68  */
69 int sysctl_perf_counter_sample_rate __read_mostly = 100000;
70
71 static atomic64_t perf_counter_id;
72
73 /*
74  * Lock for (sysadmin-configurable) counter reservations:
75  */
76 static DEFINE_SPINLOCK(perf_resource_lock);
77
78 /*
79  * Architecture provided APIs - weak aliases:
80  */
81 extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
82 {
83         return NULL;
84 }
85
86 void __weak hw_perf_disable(void)               { barrier(); }
87 void __weak hw_perf_enable(void)                { barrier(); }
88
89 void __weak hw_perf_counter_setup(int cpu)      { barrier(); }
90
91 int __weak
92 hw_perf_group_sched_in(struct perf_counter *group_leader,
93                struct perf_cpu_context *cpuctx,
94                struct perf_counter_context *ctx, int cpu)
95 {
96         return 0;
97 }
98
99 void __weak perf_counter_print_debug(void)      { }
100
101 static DEFINE_PER_CPU(int, disable_count);
102
103 void __perf_disable(void)
104 {
105         __get_cpu_var(disable_count)++;
106 }
107
108 bool __perf_enable(void)
109 {
110         return !--__get_cpu_var(disable_count);
111 }
112
113 void perf_disable(void)
114 {
115         __perf_disable();
116         hw_perf_disable();
117 }
118
119 void perf_enable(void)
120 {
121         if (__perf_enable())
122                 hw_perf_enable();
123 }
124
125 static void get_ctx(struct perf_counter_context *ctx)
126 {
127         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
128 }
129
130 static void free_ctx(struct rcu_head *head)
131 {
132         struct perf_counter_context *ctx;
133
134         ctx = container_of(head, struct perf_counter_context, rcu_head);
135         kfree(ctx);
136 }
137
138 static void put_ctx(struct perf_counter_context *ctx)
139 {
140         if (atomic_dec_and_test(&ctx->refcount)) {
141                 if (ctx->parent_ctx)
142                         put_ctx(ctx->parent_ctx);
143                 if (ctx->task)
144                         put_task_struct(ctx->task);
145                 call_rcu(&ctx->rcu_head, free_ctx);
146         }
147 }
148
149 /*
150  * Get the perf_counter_context for a task and lock it.
151  * This has to cope with with the fact that until it is locked,
152  * the context could get moved to another task.
153  */
154 static struct perf_counter_context *
155 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
156 {
157         struct perf_counter_context *ctx;
158
159         rcu_read_lock();
160  retry:
161         ctx = rcu_dereference(task->perf_counter_ctxp);
162         if (ctx) {
163                 /*
164                  * If this context is a clone of another, it might
165                  * get swapped for another underneath us by
166                  * perf_counter_task_sched_out, though the
167                  * rcu_read_lock() protects us from any context
168                  * getting freed.  Lock the context and check if it
169                  * got swapped before we could get the lock, and retry
170                  * if so.  If we locked the right context, then it
171                  * can't get swapped on us any more.
172                  */
173                 spin_lock_irqsave(&ctx->lock, *flags);
174                 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
175                         spin_unlock_irqrestore(&ctx->lock, *flags);
176                         goto retry;
177                 }
178
179                 if (!atomic_inc_not_zero(&ctx->refcount)) {
180                         spin_unlock_irqrestore(&ctx->lock, *flags);
181                         ctx = NULL;
182                 }
183         }
184         rcu_read_unlock();
185         return ctx;
186 }
187
188 /*
189  * Get the context for a task and increment its pin_count so it
190  * can't get swapped to another task.  This also increments its
191  * reference count so that the context can't get freed.
192  */
193 static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
194 {
195         struct perf_counter_context *ctx;
196         unsigned long flags;
197
198         ctx = perf_lock_task_context(task, &flags);
199         if (ctx) {
200                 ++ctx->pin_count;
201                 spin_unlock_irqrestore(&ctx->lock, flags);
202         }
203         return ctx;
204 }
205
206 static void perf_unpin_context(struct perf_counter_context *ctx)
207 {
208         unsigned long flags;
209
210         spin_lock_irqsave(&ctx->lock, flags);
211         --ctx->pin_count;
212         spin_unlock_irqrestore(&ctx->lock, flags);
213         put_ctx(ctx);
214 }
215
216 /*
217  * Add a counter from the lists for its context.
218  * Must be called with ctx->mutex and ctx->lock held.
219  */
220 static void
221 list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
222 {
223         struct perf_counter *group_leader = counter->group_leader;
224
225         /*
226          * Depending on whether it is a standalone or sibling counter,
227          * add it straight to the context's counter list, or to the group
228          * leader's sibling list:
229          */
230         if (group_leader == counter)
231                 list_add_tail(&counter->list_entry, &ctx->counter_list);
232         else {
233                 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
234                 group_leader->nr_siblings++;
235         }
236
237         list_add_rcu(&counter->event_entry, &ctx->event_list);
238         ctx->nr_counters++;
239         if (counter->attr.inherit_stat)
240                 ctx->nr_stat++;
241 }
242
243 /*
244  * Remove a counter from the lists for its context.
245  * Must be called with ctx->mutex and ctx->lock held.
246  */
247 static void
248 list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
249 {
250         struct perf_counter *sibling, *tmp;
251
252         if (list_empty(&counter->list_entry))
253                 return;
254         ctx->nr_counters--;
255         if (counter->attr.inherit_stat)
256                 ctx->nr_stat--;
257
258         list_del_init(&counter->list_entry);
259         list_del_rcu(&counter->event_entry);
260
261         if (counter->group_leader != counter)
262                 counter->group_leader->nr_siblings--;
263
264         /*
265          * If this was a group counter with sibling counters then
266          * upgrade the siblings to singleton counters by adding them
267          * to the context list directly:
268          */
269         list_for_each_entry_safe(sibling, tmp,
270                                  &counter->sibling_list, list_entry) {
271
272                 list_move_tail(&sibling->list_entry, &ctx->counter_list);
273                 sibling->group_leader = sibling;
274         }
275 }
276
277 static void
278 counter_sched_out(struct perf_counter *counter,
279                   struct perf_cpu_context *cpuctx,
280                   struct perf_counter_context *ctx)
281 {
282         if (counter->state != PERF_COUNTER_STATE_ACTIVE)
283                 return;
284
285         counter->state = PERF_COUNTER_STATE_INACTIVE;
286         counter->tstamp_stopped = ctx->time;
287         counter->pmu->disable(counter);
288         counter->oncpu = -1;
289
290         if (!is_software_counter(counter))
291                 cpuctx->active_oncpu--;
292         ctx->nr_active--;
293         if (counter->attr.exclusive || !cpuctx->active_oncpu)
294                 cpuctx->exclusive = 0;
295 }
296
297 static void
298 group_sched_out(struct perf_counter *group_counter,
299                 struct perf_cpu_context *cpuctx,
300                 struct perf_counter_context *ctx)
301 {
302         struct perf_counter *counter;
303
304         if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
305                 return;
306
307         counter_sched_out(group_counter, cpuctx, ctx);
308
309         /*
310          * Schedule out siblings (if any):
311          */
312         list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
313                 counter_sched_out(counter, cpuctx, ctx);
314
315         if (group_counter->attr.exclusive)
316                 cpuctx->exclusive = 0;
317 }
318
319 /*
320  * Cross CPU call to remove a performance counter
321  *
322  * We disable the counter on the hardware level first. After that we
323  * remove it from the context list.
324  */
325 static void __perf_counter_remove_from_context(void *info)
326 {
327         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
328         struct perf_counter *counter = info;
329         struct perf_counter_context *ctx = counter->ctx;
330
331         /*
332          * If this is a task context, we need to check whether it is
333          * the current task context of this cpu. If not it has been
334          * scheduled out before the smp call arrived.
335          */
336         if (ctx->task && cpuctx->task_ctx != ctx)
337                 return;
338
339         spin_lock(&ctx->lock);
340         /*
341          * Protect the list operation against NMI by disabling the
342          * counters on a global level.
343          */
344         perf_disable();
345
346         counter_sched_out(counter, cpuctx, ctx);
347
348         list_del_counter(counter, ctx);
349
350         if (!ctx->task) {
351                 /*
352                  * Allow more per task counters with respect to the
353                  * reservation:
354                  */
355                 cpuctx->max_pertask =
356                         min(perf_max_counters - ctx->nr_counters,
357                             perf_max_counters - perf_reserved_percpu);
358         }
359
360         perf_enable();
361         spin_unlock(&ctx->lock);
362 }
363
364
365 /*
366  * Remove the counter from a task's (or a CPU's) list of counters.
367  *
368  * Must be called with ctx->mutex held.
369  *
370  * CPU counters are removed with a smp call. For task counters we only
371  * call when the task is on a CPU.
372  *
373  * If counter->ctx is a cloned context, callers must make sure that
374  * every task struct that counter->ctx->task could possibly point to
375  * remains valid.  This is OK when called from perf_release since
376  * that only calls us on the top-level context, which can't be a clone.
377  * When called from perf_counter_exit_task, it's OK because the
378  * context has been detached from its task.
379  */
380 static void perf_counter_remove_from_context(struct perf_counter *counter)
381 {
382         struct perf_counter_context *ctx = counter->ctx;
383         struct task_struct *task = ctx->task;
384
385         if (!task) {
386                 /*
387                  * Per cpu counters are removed via an smp call and
388                  * the removal is always sucessful.
389                  */
390                 smp_call_function_single(counter->cpu,
391                                          __perf_counter_remove_from_context,
392                                          counter, 1);
393                 return;
394         }
395
396 retry:
397         task_oncpu_function_call(task, __perf_counter_remove_from_context,
398                                  counter);
399
400         spin_lock_irq(&ctx->lock);
401         /*
402          * If the context is active we need to retry the smp call.
403          */
404         if (ctx->nr_active && !list_empty(&counter->list_entry)) {
405                 spin_unlock_irq(&ctx->lock);
406                 goto retry;
407         }
408
409         /*
410          * The lock prevents that this context is scheduled in so we
411          * can remove the counter safely, if the call above did not
412          * succeed.
413          */
414         if (!list_empty(&counter->list_entry)) {
415                 list_del_counter(counter, ctx);
416         }
417         spin_unlock_irq(&ctx->lock);
418 }
419
420 static inline u64 perf_clock(void)
421 {
422         return cpu_clock(smp_processor_id());
423 }
424
425 /*
426  * Update the record of the current time in a context.
427  */
428 static void update_context_time(struct perf_counter_context *ctx)
429 {
430         u64 now = perf_clock();
431
432         ctx->time += now - ctx->timestamp;
433         ctx->timestamp = now;
434 }
435
436 /*
437  * Update the total_time_enabled and total_time_running fields for a counter.
438  */
439 static void update_counter_times(struct perf_counter *counter)
440 {
441         struct perf_counter_context *ctx = counter->ctx;
442         u64 run_end;
443
444         if (counter->state < PERF_COUNTER_STATE_INACTIVE)
445                 return;
446
447         counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
448
449         if (counter->state == PERF_COUNTER_STATE_INACTIVE)
450                 run_end = counter->tstamp_stopped;
451         else
452                 run_end = ctx->time;
453
454         counter->total_time_running = run_end - counter->tstamp_running;
455 }
456
457 /*
458  * Update total_time_enabled and total_time_running for all counters in a group.
459  */
460 static void update_group_times(struct perf_counter *leader)
461 {
462         struct perf_counter *counter;
463
464         update_counter_times(leader);
465         list_for_each_entry(counter, &leader->sibling_list, list_entry)
466                 update_counter_times(counter);
467 }
468
469 /*
470  * Cross CPU call to disable a performance counter
471  */
472 static void __perf_counter_disable(void *info)
473 {
474         struct perf_counter *counter = info;
475         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
476         struct perf_counter_context *ctx = counter->ctx;
477
478         /*
479          * If this is a per-task counter, need to check whether this
480          * counter's task is the current task on this cpu.
481          */
482         if (ctx->task && cpuctx->task_ctx != ctx)
483                 return;
484
485         spin_lock(&ctx->lock);
486
487         /*
488          * If the counter is on, turn it off.
489          * If it is in error state, leave it in error state.
490          */
491         if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
492                 update_context_time(ctx);
493                 update_counter_times(counter);
494                 if (counter == counter->group_leader)
495                         group_sched_out(counter, cpuctx, ctx);
496                 else
497                         counter_sched_out(counter, cpuctx, ctx);
498                 counter->state = PERF_COUNTER_STATE_OFF;
499         }
500
501         spin_unlock(&ctx->lock);
502 }
503
504 /*
505  * Disable a counter.
506  *
507  * If counter->ctx is a cloned context, callers must make sure that
508  * every task struct that counter->ctx->task could possibly point to
509  * remains valid.  This condition is satisifed when called through
510  * perf_counter_for_each_child or perf_counter_for_each because they
511  * hold the top-level counter's child_mutex, so any descendant that
512  * goes to exit will block in sync_child_counter.
513  * When called from perf_pending_counter it's OK because counter->ctx
514  * is the current context on this CPU and preemption is disabled,
515  * hence we can't get into perf_counter_task_sched_out for this context.
516  */
517 static void perf_counter_disable(struct perf_counter *counter)
518 {
519         struct perf_counter_context *ctx = counter->ctx;
520         struct task_struct *task = ctx->task;
521
522         if (!task) {
523                 /*
524                  * Disable the counter on the cpu that it's on
525                  */
526                 smp_call_function_single(counter->cpu, __perf_counter_disable,
527                                          counter, 1);
528                 return;
529         }
530
531  retry:
532         task_oncpu_function_call(task, __perf_counter_disable, counter);
533
534         spin_lock_irq(&ctx->lock);
535         /*
536          * If the counter is still active, we need to retry the cross-call.
537          */
538         if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
539                 spin_unlock_irq(&ctx->lock);
540                 goto retry;
541         }
542
543         /*
544          * Since we have the lock this context can't be scheduled
545          * in, so we can change the state safely.
546          */
547         if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
548                 update_counter_times(counter);
549                 counter->state = PERF_COUNTER_STATE_OFF;
550         }
551
552         spin_unlock_irq(&ctx->lock);
553 }
554
555 static int
556 counter_sched_in(struct perf_counter *counter,
557                  struct perf_cpu_context *cpuctx,
558                  struct perf_counter_context *ctx,
559                  int cpu)
560 {
561         if (counter->state <= PERF_COUNTER_STATE_OFF)
562                 return 0;
563
564         counter->state = PERF_COUNTER_STATE_ACTIVE;
565         counter->oncpu = cpu;   /* TODO: put 'cpu' into cpuctx->cpu */
566         /*
567          * The new state must be visible before we turn it on in the hardware:
568          */
569         smp_wmb();
570
571         if (counter->pmu->enable(counter)) {
572                 counter->state = PERF_COUNTER_STATE_INACTIVE;
573                 counter->oncpu = -1;
574                 return -EAGAIN;
575         }
576
577         counter->tstamp_running += ctx->time - counter->tstamp_stopped;
578
579         if (!is_software_counter(counter))
580                 cpuctx->active_oncpu++;
581         ctx->nr_active++;
582
583         if (counter->attr.exclusive)
584                 cpuctx->exclusive = 1;
585
586         return 0;
587 }
588
589 static int
590 group_sched_in(struct perf_counter *group_counter,
591                struct perf_cpu_context *cpuctx,
592                struct perf_counter_context *ctx,
593                int cpu)
594 {
595         struct perf_counter *counter, *partial_group;
596         int ret;
597
598         if (group_counter->state == PERF_COUNTER_STATE_OFF)
599                 return 0;
600
601         ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
602         if (ret)
603                 return ret < 0 ? ret : 0;
604
605         if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
606                 return -EAGAIN;
607
608         /*
609          * Schedule in siblings as one group (if any):
610          */
611         list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
612                 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
613                         partial_group = counter;
614                         goto group_error;
615                 }
616         }
617
618         return 0;
619
620 group_error:
621         /*
622          * Groups can be scheduled in as one unit only, so undo any
623          * partial group before returning:
624          */
625         list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
626                 if (counter == partial_group)
627                         break;
628                 counter_sched_out(counter, cpuctx, ctx);
629         }
630         counter_sched_out(group_counter, cpuctx, ctx);
631
632         return -EAGAIN;
633 }
634
635 /*
636  * Return 1 for a group consisting entirely of software counters,
637  * 0 if the group contains any hardware counters.
638  */
639 static int is_software_only_group(struct perf_counter *leader)
640 {
641         struct perf_counter *counter;
642
643         if (!is_software_counter(leader))
644                 return 0;
645
646         list_for_each_entry(counter, &leader->sibling_list, list_entry)
647                 if (!is_software_counter(counter))
648                         return 0;
649
650         return 1;
651 }
652
653 /*
654  * Work out whether we can put this counter group on the CPU now.
655  */
656 static int group_can_go_on(struct perf_counter *counter,
657                            struct perf_cpu_context *cpuctx,
658                            int can_add_hw)
659 {
660         /*
661          * Groups consisting entirely of software counters can always go on.
662          */
663         if (is_software_only_group(counter))
664                 return 1;
665         /*
666          * If an exclusive group is already on, no other hardware
667          * counters can go on.
668          */
669         if (cpuctx->exclusive)
670                 return 0;
671         /*
672          * If this group is exclusive and there are already
673          * counters on the CPU, it can't go on.
674          */
675         if (counter->attr.exclusive && cpuctx->active_oncpu)
676                 return 0;
677         /*
678          * Otherwise, try to add it if all previous groups were able
679          * to go on.
680          */
681         return can_add_hw;
682 }
683
684 static void add_counter_to_ctx(struct perf_counter *counter,
685                                struct perf_counter_context *ctx)
686 {
687         list_add_counter(counter, ctx);
688         counter->tstamp_enabled = ctx->time;
689         counter->tstamp_running = ctx->time;
690         counter->tstamp_stopped = ctx->time;
691 }
692
693 /*
694  * Cross CPU call to install and enable a performance counter
695  *
696  * Must be called with ctx->mutex held
697  */
698 static void __perf_install_in_context(void *info)
699 {
700         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
701         struct perf_counter *counter = info;
702         struct perf_counter_context *ctx = counter->ctx;
703         struct perf_counter *leader = counter->group_leader;
704         int cpu = smp_processor_id();
705         int err;
706
707         /*
708          * If this is a task context, we need to check whether it is
709          * the current task context of this cpu. If not it has been
710          * scheduled out before the smp call arrived.
711          * Or possibly this is the right context but it isn't
712          * on this cpu because it had no counters.
713          */
714         if (ctx->task && cpuctx->task_ctx != ctx) {
715                 if (cpuctx->task_ctx || ctx->task != current)
716                         return;
717                 cpuctx->task_ctx = ctx;
718         }
719
720         spin_lock(&ctx->lock);
721         ctx->is_active = 1;
722         update_context_time(ctx);
723
724         /*
725          * Protect the list operation against NMI by disabling the
726          * counters on a global level. NOP for non NMI based counters.
727          */
728         perf_disable();
729
730         add_counter_to_ctx(counter, ctx);
731
732         /*
733          * Don't put the counter on if it is disabled or if
734          * it is in a group and the group isn't on.
735          */
736         if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
737             (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
738                 goto unlock;
739
740         /*
741          * An exclusive counter can't go on if there are already active
742          * hardware counters, and no hardware counter can go on if there
743          * is already an exclusive counter on.
744          */
745         if (!group_can_go_on(counter, cpuctx, 1))
746                 err = -EEXIST;
747         else
748                 err = counter_sched_in(counter, cpuctx, ctx, cpu);
749
750         if (err) {
751                 /*
752                  * This counter couldn't go on.  If it is in a group
753                  * then we have to pull the whole group off.
754                  * If the counter group is pinned then put it in error state.
755                  */
756                 if (leader != counter)
757                         group_sched_out(leader, cpuctx, ctx);
758                 if (leader->attr.pinned) {
759                         update_group_times(leader);
760                         leader->state = PERF_COUNTER_STATE_ERROR;
761                 }
762         }
763
764         if (!err && !ctx->task && cpuctx->max_pertask)
765                 cpuctx->max_pertask--;
766
767  unlock:
768         perf_enable();
769
770         spin_unlock(&ctx->lock);
771 }
772
773 /*
774  * Attach a performance counter to a context
775  *
776  * First we add the counter to the list with the hardware enable bit
777  * in counter->hw_config cleared.
778  *
779  * If the counter is attached to a task which is on a CPU we use a smp
780  * call to enable it in the task context. The task might have been
781  * scheduled away, but we check this in the smp call again.
782  *
783  * Must be called with ctx->mutex held.
784  */
785 static void
786 perf_install_in_context(struct perf_counter_context *ctx,
787                         struct perf_counter *counter,
788                         int cpu)
789 {
790         struct task_struct *task = ctx->task;
791
792         if (!task) {
793                 /*
794                  * Per cpu counters are installed via an smp call and
795                  * the install is always sucessful.
796                  */
797                 smp_call_function_single(cpu, __perf_install_in_context,
798                                          counter, 1);
799                 return;
800         }
801
802 retry:
803         task_oncpu_function_call(task, __perf_install_in_context,
804                                  counter);
805
806         spin_lock_irq(&ctx->lock);
807         /*
808          * we need to retry the smp call.
809          */
810         if (ctx->is_active && list_empty(&counter->list_entry)) {
811                 spin_unlock_irq(&ctx->lock);
812                 goto retry;
813         }
814
815         /*
816          * The lock prevents that this context is scheduled in so we
817          * can add the counter safely, if it the call above did not
818          * succeed.
819          */
820         if (list_empty(&counter->list_entry))
821                 add_counter_to_ctx(counter, ctx);
822         spin_unlock_irq(&ctx->lock);
823 }
824
825 /*
826  * Cross CPU call to enable a performance counter
827  */
828 static void __perf_counter_enable(void *info)
829 {
830         struct perf_counter *counter = info;
831         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
832         struct perf_counter_context *ctx = counter->ctx;
833         struct perf_counter *leader = counter->group_leader;
834         int err;
835
836         /*
837          * If this is a per-task counter, need to check whether this
838          * counter's task is the current task on this cpu.
839          */
840         if (ctx->task && cpuctx->task_ctx != ctx) {
841                 if (cpuctx->task_ctx || ctx->task != current)
842                         return;
843                 cpuctx->task_ctx = ctx;
844         }
845
846         spin_lock(&ctx->lock);
847         ctx->is_active = 1;
848         update_context_time(ctx);
849
850         if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
851                 goto unlock;
852         counter->state = PERF_COUNTER_STATE_INACTIVE;
853         counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
854
855         /*
856          * If the counter is in a group and isn't the group leader,
857          * then don't put it on unless the group is on.
858          */
859         if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
860                 goto unlock;
861
862         if (!group_can_go_on(counter, cpuctx, 1)) {
863                 err = -EEXIST;
864         } else {
865                 perf_disable();
866                 if (counter == leader)
867                         err = group_sched_in(counter, cpuctx, ctx,
868                                              smp_processor_id());
869                 else
870                         err = counter_sched_in(counter, cpuctx, ctx,
871                                                smp_processor_id());
872                 perf_enable();
873         }
874
875         if (err) {
876                 /*
877                  * If this counter can't go on and it's part of a
878                  * group, then the whole group has to come off.
879                  */
880                 if (leader != counter)
881                         group_sched_out(leader, cpuctx, ctx);
882                 if (leader->attr.pinned) {
883                         update_group_times(leader);
884                         leader->state = PERF_COUNTER_STATE_ERROR;
885                 }
886         }
887
888  unlock:
889         spin_unlock(&ctx->lock);
890 }
891
892 /*
893  * Enable a counter.
894  *
895  * If counter->ctx is a cloned context, callers must make sure that
896  * every task struct that counter->ctx->task could possibly point to
897  * remains valid.  This condition is satisfied when called through
898  * perf_counter_for_each_child or perf_counter_for_each as described
899  * for perf_counter_disable.
900  */
901 static void perf_counter_enable(struct perf_counter *counter)
902 {
903         struct perf_counter_context *ctx = counter->ctx;
904         struct task_struct *task = ctx->task;
905
906         if (!task) {
907                 /*
908                  * Enable the counter on the cpu that it's on
909                  */
910                 smp_call_function_single(counter->cpu, __perf_counter_enable,
911                                          counter, 1);
912                 return;
913         }
914
915         spin_lock_irq(&ctx->lock);
916         if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
917                 goto out;
918
919         /*
920          * If the counter is in error state, clear that first.
921          * That way, if we see the counter in error state below, we
922          * know that it has gone back into error state, as distinct
923          * from the task having been scheduled away before the
924          * cross-call arrived.
925          */
926         if (counter->state == PERF_COUNTER_STATE_ERROR)
927                 counter->state = PERF_COUNTER_STATE_OFF;
928
929  retry:
930         spin_unlock_irq(&ctx->lock);
931         task_oncpu_function_call(task, __perf_counter_enable, counter);
932
933         spin_lock_irq(&ctx->lock);
934
935         /*
936          * If the context is active and the counter is still off,
937          * we need to retry the cross-call.
938          */
939         if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
940                 goto retry;
941
942         /*
943          * Since we have the lock this context can't be scheduled
944          * in, so we can change the state safely.
945          */
946         if (counter->state == PERF_COUNTER_STATE_OFF) {
947                 counter->state = PERF_COUNTER_STATE_INACTIVE;
948                 counter->tstamp_enabled =
949                         ctx->time - counter->total_time_enabled;
950         }
951  out:
952         spin_unlock_irq(&ctx->lock);
953 }
954
955 static int perf_counter_refresh(struct perf_counter *counter, int refresh)
956 {
957         /*
958          * not supported on inherited counters
959          */
960         if (counter->attr.inherit)
961                 return -EINVAL;
962
963         atomic_add(refresh, &counter->event_limit);
964         perf_counter_enable(counter);
965
966         return 0;
967 }
968
969 void __perf_counter_sched_out(struct perf_counter_context *ctx,
970                               struct perf_cpu_context *cpuctx)
971 {
972         struct perf_counter *counter;
973
974         spin_lock(&ctx->lock);
975         ctx->is_active = 0;
976         if (likely(!ctx->nr_counters))
977                 goto out;
978         update_context_time(ctx);
979
980         perf_disable();
981         if (ctx->nr_active) {
982                 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
983                         if (counter != counter->group_leader)
984                                 counter_sched_out(counter, cpuctx, ctx);
985                         else
986                                 group_sched_out(counter, cpuctx, ctx);
987                 }
988         }
989         perf_enable();
990  out:
991         spin_unlock(&ctx->lock);
992 }
993
994 /*
995  * Test whether two contexts are equivalent, i.e. whether they
996  * have both been cloned from the same version of the same context
997  * and they both have the same number of enabled counters.
998  * If the number of enabled counters is the same, then the set
999  * of enabled counters should be the same, because these are both
1000  * inherited contexts, therefore we can't access individual counters
1001  * in them directly with an fd; we can only enable/disable all
1002  * counters via prctl, or enable/disable all counters in a family
1003  * via ioctl, which will have the same effect on both contexts.
1004  */
1005 static int context_equiv(struct perf_counter_context *ctx1,
1006                          struct perf_counter_context *ctx2)
1007 {
1008         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1009                 && ctx1->parent_gen == ctx2->parent_gen
1010                 && !ctx1->pin_count && !ctx2->pin_count;
1011 }
1012
1013 static void __perf_counter_read(void *counter);
1014
1015 static void __perf_counter_sync_stat(struct perf_counter *counter,
1016                                      struct perf_counter *next_counter)
1017 {
1018         u64 value;
1019
1020         if (!counter->attr.inherit_stat)
1021                 return;
1022
1023         /*
1024          * Update the counter value, we cannot use perf_counter_read()
1025          * because we're in the middle of a context switch and have IRQs
1026          * disabled, which upsets smp_call_function_single(), however
1027          * we know the counter must be on the current CPU, therefore we
1028          * don't need to use it.
1029          */
1030         switch (counter->state) {
1031         case PERF_COUNTER_STATE_ACTIVE:
1032                 __perf_counter_read(counter);
1033                 break;
1034
1035         case PERF_COUNTER_STATE_INACTIVE:
1036                 update_counter_times(counter);
1037                 break;
1038
1039         default:
1040                 break;
1041         }
1042
1043         /*
1044          * In order to keep per-task stats reliable we need to flip the counter
1045          * values when we flip the contexts.
1046          */
1047         value = atomic64_read(&next_counter->count);
1048         value = atomic64_xchg(&counter->count, value);
1049         atomic64_set(&next_counter->count, value);
1050
1051         swap(counter->total_time_enabled, next_counter->total_time_enabled);
1052         swap(counter->total_time_running, next_counter->total_time_running);
1053
1054         /*
1055          * Since we swizzled the values, update the user visible data too.
1056          */
1057         perf_counter_update_userpage(counter);
1058         perf_counter_update_userpage(next_counter);
1059 }
1060
1061 #define list_next_entry(pos, member) \
1062         list_entry(pos->member.next, typeof(*pos), member)
1063
1064 static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1065                                    struct perf_counter_context *next_ctx)
1066 {
1067         struct perf_counter *counter, *next_counter;
1068
1069         if (!ctx->nr_stat)
1070                 return;
1071
1072         counter = list_first_entry(&ctx->event_list,
1073                                    struct perf_counter, event_entry);
1074
1075         next_counter = list_first_entry(&next_ctx->event_list,
1076                                         struct perf_counter, event_entry);
1077
1078         while (&counter->event_entry != &ctx->event_list &&
1079                &next_counter->event_entry != &next_ctx->event_list) {
1080
1081                 __perf_counter_sync_stat(counter, next_counter);
1082
1083                 counter = list_next_entry(counter, event_entry);
1084                 next_counter = list_next_entry(counter, event_entry);
1085         }
1086 }
1087
1088 /*
1089  * Called from scheduler to remove the counters of the current task,
1090  * with interrupts disabled.
1091  *
1092  * We stop each counter and update the counter value in counter->count.
1093  *
1094  * This does not protect us against NMI, but disable()
1095  * sets the disabled bit in the control field of counter _before_
1096  * accessing the counter control register. If a NMI hits, then it will
1097  * not restart the counter.
1098  */
1099 void perf_counter_task_sched_out(struct task_struct *task,
1100                                  struct task_struct *next, int cpu)
1101 {
1102         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1103         struct perf_counter_context *ctx = task->perf_counter_ctxp;
1104         struct perf_counter_context *next_ctx;
1105         struct perf_counter_context *parent;
1106         struct pt_regs *regs;
1107         int do_switch = 1;
1108
1109         regs = task_pt_regs(task);
1110         perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1111
1112         if (likely(!ctx || !cpuctx->task_ctx))
1113                 return;
1114
1115         update_context_time(ctx);
1116
1117         rcu_read_lock();
1118         parent = rcu_dereference(ctx->parent_ctx);
1119         next_ctx = next->perf_counter_ctxp;
1120         if (parent && next_ctx &&
1121             rcu_dereference(next_ctx->parent_ctx) == parent) {
1122                 /*
1123                  * Looks like the two contexts are clones, so we might be
1124                  * able to optimize the context switch.  We lock both
1125                  * contexts and check that they are clones under the
1126                  * lock (including re-checking that neither has been
1127                  * uncloned in the meantime).  It doesn't matter which
1128                  * order we take the locks because no other cpu could
1129                  * be trying to lock both of these tasks.
1130                  */
1131                 spin_lock(&ctx->lock);
1132                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1133                 if (context_equiv(ctx, next_ctx)) {
1134                         /*
1135                          * XXX do we need a memory barrier of sorts
1136                          * wrt to rcu_dereference() of perf_counter_ctxp
1137                          */
1138                         task->perf_counter_ctxp = next_ctx;
1139                         next->perf_counter_ctxp = ctx;
1140                         ctx->task = next;
1141                         next_ctx->task = task;
1142                         do_switch = 0;
1143
1144                         perf_counter_sync_stat(ctx, next_ctx);
1145                 }
1146                 spin_unlock(&next_ctx->lock);
1147                 spin_unlock(&ctx->lock);
1148         }
1149         rcu_read_unlock();
1150
1151         if (do_switch) {
1152                 __perf_counter_sched_out(ctx, cpuctx);
1153                 cpuctx->task_ctx = NULL;
1154         }
1155 }
1156
1157 /*
1158  * Called with IRQs disabled
1159  */
1160 static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1161 {
1162         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1163
1164         if (!cpuctx->task_ctx)
1165                 return;
1166
1167         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1168                 return;
1169
1170         __perf_counter_sched_out(ctx, cpuctx);
1171         cpuctx->task_ctx = NULL;
1172 }
1173
1174 /*
1175  * Called with IRQs disabled
1176  */
1177 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1178 {
1179         __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1180 }
1181
1182 static void
1183 __perf_counter_sched_in(struct perf_counter_context *ctx,
1184                         struct perf_cpu_context *cpuctx, int cpu)
1185 {
1186         struct perf_counter *counter;
1187         int can_add_hw = 1;
1188
1189         spin_lock(&ctx->lock);
1190         ctx->is_active = 1;
1191         if (likely(!ctx->nr_counters))
1192                 goto out;
1193
1194         ctx->timestamp = perf_clock();
1195
1196         perf_disable();
1197
1198         /*
1199          * First go through the list and put on any pinned groups
1200          * in order to give them the best chance of going on.
1201          */
1202         list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1203                 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1204                     !counter->attr.pinned)
1205                         continue;
1206                 if (counter->cpu != -1 && counter->cpu != cpu)
1207                         continue;
1208
1209                 if (counter != counter->group_leader)
1210                         counter_sched_in(counter, cpuctx, ctx, cpu);
1211                 else {
1212                         if (group_can_go_on(counter, cpuctx, 1))
1213                                 group_sched_in(counter, cpuctx, ctx, cpu);
1214                 }
1215
1216                 /*
1217                  * If this pinned group hasn't been scheduled,
1218                  * put it in error state.
1219                  */
1220                 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1221                         update_group_times(counter);
1222                         counter->state = PERF_COUNTER_STATE_ERROR;
1223                 }
1224         }
1225
1226         list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1227                 /*
1228                  * Ignore counters in OFF or ERROR state, and
1229                  * ignore pinned counters since we did them already.
1230                  */
1231                 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1232                     counter->attr.pinned)
1233                         continue;
1234
1235                 /*
1236                  * Listen to the 'cpu' scheduling filter constraint
1237                  * of counters:
1238                  */
1239                 if (counter->cpu != -1 && counter->cpu != cpu)
1240                         continue;
1241
1242                 if (counter != counter->group_leader) {
1243                         if (counter_sched_in(counter, cpuctx, ctx, cpu))
1244                                 can_add_hw = 0;
1245                 } else {
1246                         if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1247                                 if (group_sched_in(counter, cpuctx, ctx, cpu))
1248                                         can_add_hw = 0;
1249                         }
1250                 }
1251         }
1252         perf_enable();
1253  out:
1254         spin_unlock(&ctx->lock);
1255 }
1256
1257 /*
1258  * Called from scheduler to add the counters of the current task
1259  * with interrupts disabled.
1260  *
1261  * We restore the counter value and then enable it.
1262  *
1263  * This does not protect us against NMI, but enable()
1264  * sets the enabled bit in the control field of counter _before_
1265  * accessing the counter control register. If a NMI hits, then it will
1266  * keep the counter running.
1267  */
1268 void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1269 {
1270         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1271         struct perf_counter_context *ctx = task->perf_counter_ctxp;
1272
1273         if (likely(!ctx))
1274                 return;
1275         if (cpuctx->task_ctx == ctx)
1276                 return;
1277         __perf_counter_sched_in(ctx, cpuctx, cpu);
1278         cpuctx->task_ctx = ctx;
1279 }
1280
1281 static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1282 {
1283         struct perf_counter_context *ctx = &cpuctx->ctx;
1284
1285         __perf_counter_sched_in(ctx, cpuctx, cpu);
1286 }
1287
1288 #define MAX_INTERRUPTS (~0ULL)
1289
1290 static void perf_log_throttle(struct perf_counter *counter, int enable);
1291 static void perf_log_period(struct perf_counter *counter, u64 period);
1292
1293 static void perf_adjust_period(struct perf_counter *counter, u64 events)
1294 {
1295         struct hw_perf_counter *hwc = &counter->hw;
1296         u64 period, sample_period;
1297         s64 delta;
1298
1299         events *= hwc->sample_period;
1300         period = div64_u64(events, counter->attr.sample_freq);
1301
1302         delta = (s64)(period - hwc->sample_period);
1303         delta = (delta + 7) / 8; /* low pass filter */
1304
1305         sample_period = hwc->sample_period + delta;
1306
1307         if (!sample_period)
1308                 sample_period = 1;
1309
1310         perf_log_period(counter, sample_period);
1311
1312         hwc->sample_period = sample_period;
1313 }
1314
1315 static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1316 {
1317         struct perf_counter *counter;
1318         struct hw_perf_counter *hwc;
1319         u64 interrupts, freq;
1320
1321         spin_lock(&ctx->lock);
1322         list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1323                 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1324                         continue;
1325
1326                 hwc = &counter->hw;
1327
1328                 interrupts = hwc->interrupts;
1329                 hwc->interrupts = 0;
1330
1331                 /*
1332                  * unthrottle counters on the tick
1333                  */
1334                 if (interrupts == MAX_INTERRUPTS) {
1335                         perf_log_throttle(counter, 1);
1336                         counter->pmu->unthrottle(counter);
1337                         interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1338                 }
1339
1340                 if (!counter->attr.freq || !counter->attr.sample_freq)
1341                         continue;
1342
1343                 /*
1344                  * if the specified freq < HZ then we need to skip ticks
1345                  */
1346                 if (counter->attr.sample_freq < HZ) {
1347                         freq = counter->attr.sample_freq;
1348
1349                         hwc->freq_count += freq;
1350                         hwc->freq_interrupts += interrupts;
1351
1352                         if (hwc->freq_count < HZ)
1353                                 continue;
1354
1355                         interrupts = hwc->freq_interrupts;
1356                         hwc->freq_interrupts = 0;
1357                         hwc->freq_count -= HZ;
1358                 } else
1359                         freq = HZ;
1360
1361                 perf_adjust_period(counter, freq * interrupts);
1362
1363                 /*
1364                  * In order to avoid being stalled by an (accidental) huge
1365                  * sample period, force reset the sample period if we didn't
1366                  * get any events in this freq period.
1367                  */
1368                 if (!interrupts) {
1369                         perf_disable();
1370                         counter->pmu->disable(counter);
1371                         atomic64_set(&hwc->period_left, 0);
1372                         counter->pmu->enable(counter);
1373                         perf_enable();
1374                 }
1375         }
1376         spin_unlock(&ctx->lock);
1377 }
1378
1379 /*
1380  * Round-robin a context's counters:
1381  */
1382 static void rotate_ctx(struct perf_counter_context *ctx)
1383 {
1384         struct perf_counter *counter;
1385
1386         if (!ctx->nr_counters)
1387                 return;
1388
1389         spin_lock(&ctx->lock);
1390         /*
1391          * Rotate the first entry last (works just fine for group counters too):
1392          */
1393         perf_disable();
1394         list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1395                 list_move_tail(&counter->list_entry, &ctx->counter_list);
1396                 break;
1397         }
1398         perf_enable();
1399
1400         spin_unlock(&ctx->lock);
1401 }
1402
1403 void perf_counter_task_tick(struct task_struct *curr, int cpu)
1404 {
1405         struct perf_cpu_context *cpuctx;
1406         struct perf_counter_context *ctx;
1407
1408         if (!atomic_read(&nr_counters))
1409                 return;
1410
1411         cpuctx = &per_cpu(perf_cpu_context, cpu);
1412         ctx = curr->perf_counter_ctxp;
1413
1414         perf_ctx_adjust_freq(&cpuctx->ctx);
1415         if (ctx)
1416                 perf_ctx_adjust_freq(ctx);
1417
1418         perf_counter_cpu_sched_out(cpuctx);
1419         if (ctx)
1420                 __perf_counter_task_sched_out(ctx);
1421
1422         rotate_ctx(&cpuctx->ctx);
1423         if (ctx)
1424                 rotate_ctx(ctx);
1425
1426         perf_counter_cpu_sched_in(cpuctx, cpu);
1427         if (ctx)
1428                 perf_counter_task_sched_in(curr, cpu);
1429 }
1430
1431 /*
1432  * Enable all of a task's counters that have been marked enable-on-exec.
1433  * This expects task == current.
1434  */
1435 static void perf_counter_enable_on_exec(struct task_struct *task)
1436 {
1437         struct perf_counter_context *ctx;
1438         struct perf_counter *counter;
1439         unsigned long flags;
1440         int enabled = 0;
1441
1442         local_irq_save(flags);
1443         ctx = task->perf_counter_ctxp;
1444         if (!ctx || !ctx->nr_counters)
1445                 goto out;
1446
1447         __perf_counter_task_sched_out(ctx);
1448
1449         spin_lock(&ctx->lock);
1450
1451         list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1452                 if (!counter->attr.enable_on_exec)
1453                         continue;
1454                 counter->attr.enable_on_exec = 0;
1455                 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
1456                         continue;
1457                 counter->state = PERF_COUNTER_STATE_INACTIVE;
1458                 counter->tstamp_enabled =
1459                         ctx->time - counter->total_time_enabled;
1460                 enabled = 1;
1461         }
1462
1463         /*
1464          * Unclone this context if we enabled any counter.
1465          */
1466         if (enabled && ctx->parent_ctx) {
1467                 put_ctx(ctx->parent_ctx);
1468                 ctx->parent_ctx = NULL;
1469         }
1470
1471         spin_unlock(&ctx->lock);
1472
1473         perf_counter_task_sched_in(task, smp_processor_id());
1474  out:
1475         local_irq_restore(flags);
1476 }
1477
1478 /*
1479  * Cross CPU call to read the hardware counter
1480  */
1481 static void __perf_counter_read(void *info)
1482 {
1483         struct perf_counter *counter = info;
1484         struct perf_counter_context *ctx = counter->ctx;
1485         unsigned long flags;
1486
1487         local_irq_save(flags);
1488         if (ctx->is_active)
1489                 update_context_time(ctx);
1490         counter->pmu->read(counter);
1491         update_counter_times(counter);
1492         local_irq_restore(flags);
1493 }
1494
1495 static u64 perf_counter_read(struct perf_counter *counter)
1496 {
1497         /*
1498          * If counter is enabled and currently active on a CPU, update the
1499          * value in the counter structure:
1500          */
1501         if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1502                 smp_call_function_single(counter->oncpu,
1503                                          __perf_counter_read, counter, 1);
1504         } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1505                 update_counter_times(counter);
1506         }
1507
1508         return atomic64_read(&counter->count);
1509 }
1510
1511 /*
1512  * Initialize the perf_counter context in a task_struct:
1513  */
1514 static void
1515 __perf_counter_init_context(struct perf_counter_context *ctx,
1516                             struct task_struct *task)
1517 {
1518         memset(ctx, 0, sizeof(*ctx));
1519         spin_lock_init(&ctx->lock);
1520         mutex_init(&ctx->mutex);
1521         INIT_LIST_HEAD(&ctx->counter_list);
1522         INIT_LIST_HEAD(&ctx->event_list);
1523         atomic_set(&ctx->refcount, 1);
1524         ctx->task = task;
1525 }
1526
1527 static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1528 {
1529         struct perf_counter_context *parent_ctx;
1530         struct perf_counter_context *ctx;
1531         struct perf_cpu_context *cpuctx;
1532         struct task_struct *task;
1533         unsigned long flags;
1534         int err;
1535
1536         /*
1537          * If cpu is not a wildcard then this is a percpu counter:
1538          */
1539         if (cpu != -1) {
1540                 /* Must be root to operate on a CPU counter: */
1541                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1542                         return ERR_PTR(-EACCES);
1543
1544                 if (cpu < 0 || cpu > num_possible_cpus())
1545                         return ERR_PTR(-EINVAL);
1546
1547                 /*
1548                  * We could be clever and allow to attach a counter to an
1549                  * offline CPU and activate it when the CPU comes up, but
1550                  * that's for later.
1551                  */
1552                 if (!cpu_isset(cpu, cpu_online_map))
1553                         return ERR_PTR(-ENODEV);
1554
1555                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1556                 ctx = &cpuctx->ctx;
1557                 get_ctx(ctx);
1558
1559                 return ctx;
1560         }
1561
1562         rcu_read_lock();
1563         if (!pid)
1564                 task = current;
1565         else
1566                 task = find_task_by_vpid(pid);
1567         if (task)
1568                 get_task_struct(task);
1569         rcu_read_unlock();
1570
1571         if (!task)
1572                 return ERR_PTR(-ESRCH);
1573
1574         /*
1575          * Can't attach counters to a dying task.
1576          */
1577         err = -ESRCH;
1578         if (task->flags & PF_EXITING)
1579                 goto errout;
1580
1581         /* Reuse ptrace permission checks for now. */
1582         err = -EACCES;
1583         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1584                 goto errout;
1585
1586  retry:
1587         ctx = perf_lock_task_context(task, &flags);
1588         if (ctx) {
1589                 parent_ctx = ctx->parent_ctx;
1590                 if (parent_ctx) {
1591                         put_ctx(parent_ctx);
1592                         ctx->parent_ctx = NULL;         /* no longer a clone */
1593                 }
1594                 spin_unlock_irqrestore(&ctx->lock, flags);
1595         }
1596
1597         if (!ctx) {
1598                 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1599                 err = -ENOMEM;
1600                 if (!ctx)
1601                         goto errout;
1602                 __perf_counter_init_context(ctx, task);
1603                 get_ctx(ctx);
1604                 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1605                         /*
1606                          * We raced with some other task; use
1607                          * the context they set.
1608                          */
1609                         kfree(ctx);
1610                         goto retry;
1611                 }
1612                 get_task_struct(task);
1613         }
1614
1615         put_task_struct(task);
1616         return ctx;
1617
1618  errout:
1619         put_task_struct(task);
1620         return ERR_PTR(err);
1621 }
1622
1623 static void free_counter_rcu(struct rcu_head *head)
1624 {
1625         struct perf_counter *counter;
1626
1627         counter = container_of(head, struct perf_counter, rcu_head);
1628         if (counter->ns)
1629                 put_pid_ns(counter->ns);
1630         kfree(counter);
1631 }
1632
1633 static void perf_pending_sync(struct perf_counter *counter);
1634
1635 static void free_counter(struct perf_counter *counter)
1636 {
1637         perf_pending_sync(counter);
1638
1639         if (!counter->parent) {
1640                 atomic_dec(&nr_counters);
1641                 if (counter->attr.mmap)
1642                         atomic_dec(&nr_mmap_counters);
1643                 if (counter->attr.comm)
1644                         atomic_dec(&nr_comm_counters);
1645         }
1646
1647         if (counter->destroy)
1648                 counter->destroy(counter);
1649
1650         put_ctx(counter->ctx);
1651         call_rcu(&counter->rcu_head, free_counter_rcu);
1652 }
1653
1654 /*
1655  * Called when the last reference to the file is gone.
1656  */
1657 static int perf_release(struct inode *inode, struct file *file)
1658 {
1659         struct perf_counter *counter = file->private_data;
1660         struct perf_counter_context *ctx = counter->ctx;
1661
1662         file->private_data = NULL;
1663
1664         WARN_ON_ONCE(ctx->parent_ctx);
1665         mutex_lock(&ctx->mutex);
1666         perf_counter_remove_from_context(counter);
1667         mutex_unlock(&ctx->mutex);
1668
1669         mutex_lock(&counter->owner->perf_counter_mutex);
1670         list_del_init(&counter->owner_entry);
1671         mutex_unlock(&counter->owner->perf_counter_mutex);
1672         put_task_struct(counter->owner);
1673
1674         free_counter(counter);
1675
1676         return 0;
1677 }
1678
1679 /*
1680  * Read the performance counter - simple non blocking version for now
1681  */
1682 static ssize_t
1683 perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1684 {
1685         u64 values[4];
1686         int n;
1687
1688         /*
1689          * Return end-of-file for a read on a counter that is in
1690          * error state (i.e. because it was pinned but it couldn't be
1691          * scheduled on to the CPU at some point).
1692          */
1693         if (counter->state == PERF_COUNTER_STATE_ERROR)
1694                 return 0;
1695
1696         WARN_ON_ONCE(counter->ctx->parent_ctx);
1697         mutex_lock(&counter->child_mutex);
1698         values[0] = perf_counter_read(counter);
1699         n = 1;
1700         if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1701                 values[n++] = counter->total_time_enabled +
1702                         atomic64_read(&counter->child_total_time_enabled);
1703         if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1704                 values[n++] = counter->total_time_running +
1705                         atomic64_read(&counter->child_total_time_running);
1706         if (counter->attr.read_format & PERF_FORMAT_ID)
1707                 values[n++] = counter->id;
1708         mutex_unlock(&counter->child_mutex);
1709
1710         if (count < n * sizeof(u64))
1711                 return -EINVAL;
1712         count = n * sizeof(u64);
1713
1714         if (copy_to_user(buf, values, count))
1715                 return -EFAULT;
1716
1717         return count;
1718 }
1719
1720 static ssize_t
1721 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1722 {
1723         struct perf_counter *counter = file->private_data;
1724
1725         return perf_read_hw(counter, buf, count);
1726 }
1727
1728 static unsigned int perf_poll(struct file *file, poll_table *wait)
1729 {
1730         struct perf_counter *counter = file->private_data;
1731         struct perf_mmap_data *data;
1732         unsigned int events = POLL_HUP;
1733
1734         rcu_read_lock();
1735         data = rcu_dereference(counter->data);
1736         if (data)
1737                 events = atomic_xchg(&data->poll, 0);
1738         rcu_read_unlock();
1739
1740         poll_wait(file, &counter->waitq, wait);
1741
1742         return events;
1743 }
1744
1745 static void perf_counter_reset(struct perf_counter *counter)
1746 {
1747         (void)perf_counter_read(counter);
1748         atomic64_set(&counter->count, 0);
1749         perf_counter_update_userpage(counter);
1750 }
1751
1752 /*
1753  * Holding the top-level counter's child_mutex means that any
1754  * descendant process that has inherited this counter will block
1755  * in sync_child_counter if it goes to exit, thus satisfying the
1756  * task existence requirements of perf_counter_enable/disable.
1757  */
1758 static void perf_counter_for_each_child(struct perf_counter *counter,
1759                                         void (*func)(struct perf_counter *))
1760 {
1761         struct perf_counter *child;
1762
1763         WARN_ON_ONCE(counter->ctx->parent_ctx);
1764         mutex_lock(&counter->child_mutex);
1765         func(counter);
1766         list_for_each_entry(child, &counter->child_list, child_list)
1767                 func(child);
1768         mutex_unlock(&counter->child_mutex);
1769 }
1770
1771 static void perf_counter_for_each(struct perf_counter *counter,
1772                                   void (*func)(struct perf_counter *))
1773 {
1774         struct perf_counter_context *ctx = counter->ctx;
1775         struct perf_counter *sibling;
1776
1777         WARN_ON_ONCE(ctx->parent_ctx);
1778         mutex_lock(&ctx->mutex);
1779         counter = counter->group_leader;
1780
1781         perf_counter_for_each_child(counter, func);
1782         func(counter);
1783         list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1784                 perf_counter_for_each_child(counter, func);
1785         mutex_unlock(&ctx->mutex);
1786 }
1787
1788 static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1789 {
1790         struct perf_counter_context *ctx = counter->ctx;
1791         unsigned long size;
1792         int ret = 0;
1793         u64 value;
1794
1795         if (!counter->attr.sample_period)
1796                 return -EINVAL;
1797
1798         size = copy_from_user(&value, arg, sizeof(value));
1799         if (size != sizeof(value))
1800                 return -EFAULT;
1801
1802         if (!value)
1803                 return -EINVAL;
1804
1805         spin_lock_irq(&ctx->lock);
1806         if (counter->attr.freq) {
1807                 if (value > sysctl_perf_counter_sample_rate) {
1808                         ret = -EINVAL;
1809                         goto unlock;
1810                 }
1811
1812                 counter->attr.sample_freq = value;
1813         } else {
1814                 perf_log_period(counter, value);
1815
1816                 counter->attr.sample_period = value;
1817                 counter->hw.sample_period = value;
1818         }
1819 unlock:
1820         spin_unlock_irq(&ctx->lock);
1821
1822         return ret;
1823 }
1824
1825 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1826 {
1827         struct perf_counter *counter = file->private_data;
1828         void (*func)(struct perf_counter *);
1829         u32 flags = arg;
1830
1831         switch (cmd) {
1832         case PERF_COUNTER_IOC_ENABLE:
1833                 func = perf_counter_enable;
1834                 break;
1835         case PERF_COUNTER_IOC_DISABLE:
1836                 func = perf_counter_disable;
1837                 break;
1838         case PERF_COUNTER_IOC_RESET:
1839                 func = perf_counter_reset;
1840                 break;
1841
1842         case PERF_COUNTER_IOC_REFRESH:
1843                 return perf_counter_refresh(counter, arg);
1844
1845         case PERF_COUNTER_IOC_PERIOD:
1846                 return perf_counter_period(counter, (u64 __user *)arg);
1847
1848         default:
1849                 return -ENOTTY;
1850         }
1851
1852         if (flags & PERF_IOC_FLAG_GROUP)
1853                 perf_counter_for_each(counter, func);
1854         else
1855                 perf_counter_for_each_child(counter, func);
1856
1857         return 0;
1858 }
1859
1860 int perf_counter_task_enable(void)
1861 {
1862         struct perf_counter *counter;
1863
1864         mutex_lock(&current->perf_counter_mutex);
1865         list_for_each_entry(counter, &current->perf_counter_list, owner_entry)
1866                 perf_counter_for_each_child(counter, perf_counter_enable);
1867         mutex_unlock(&current->perf_counter_mutex);
1868
1869         return 0;
1870 }
1871
1872 int perf_counter_task_disable(void)
1873 {
1874         struct perf_counter *counter;
1875
1876         mutex_lock(&current->perf_counter_mutex);
1877         list_for_each_entry(counter, &current->perf_counter_list, owner_entry)
1878                 perf_counter_for_each_child(counter, perf_counter_disable);
1879         mutex_unlock(&current->perf_counter_mutex);
1880
1881         return 0;
1882 }
1883
1884 static int perf_counter_index(struct perf_counter *counter)
1885 {
1886         if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1887                 return 0;
1888
1889         return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
1890 }
1891
1892 /*
1893  * Callers need to ensure there can be no nesting of this function, otherwise
1894  * the seqlock logic goes bad. We can not serialize this because the arch
1895  * code calls this from NMI context.
1896  */
1897 void perf_counter_update_userpage(struct perf_counter *counter)
1898 {
1899         struct perf_counter_mmap_page *userpg;
1900         struct perf_mmap_data *data;
1901
1902         rcu_read_lock();
1903         data = rcu_dereference(counter->data);
1904         if (!data)
1905                 goto unlock;
1906
1907         userpg = data->user_page;
1908
1909         /*
1910          * Disable preemption so as to not let the corresponding user-space
1911          * spin too long if we get preempted.
1912          */
1913         preempt_disable();
1914         ++userpg->lock;
1915         barrier();
1916         userpg->index = perf_counter_index(counter);
1917         userpg->offset = atomic64_read(&counter->count);
1918         if (counter->state == PERF_COUNTER_STATE_ACTIVE)
1919                 userpg->offset -= atomic64_read(&counter->hw.prev_count);
1920
1921         userpg->time_enabled = counter->total_time_enabled +
1922                         atomic64_read(&counter->child_total_time_enabled);
1923
1924         userpg->time_running = counter->total_time_running +
1925                         atomic64_read(&counter->child_total_time_running);
1926
1927         barrier();
1928         ++userpg->lock;
1929         preempt_enable();
1930 unlock:
1931         rcu_read_unlock();
1932 }
1933
1934 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1935 {
1936         struct perf_counter *counter = vma->vm_file->private_data;
1937         struct perf_mmap_data *data;
1938         int ret = VM_FAULT_SIGBUS;
1939
1940         if (vmf->flags & FAULT_FLAG_MKWRITE) {
1941                 if (vmf->pgoff == 0)
1942                         ret = 0;
1943                 return ret;
1944         }
1945
1946         rcu_read_lock();
1947         data = rcu_dereference(counter->data);
1948         if (!data)
1949                 goto unlock;
1950
1951         if (vmf->pgoff == 0) {
1952                 vmf->page = virt_to_page(data->user_page);
1953         } else {
1954                 int nr = vmf->pgoff - 1;
1955
1956                 if ((unsigned)nr > data->nr_pages)
1957                         goto unlock;
1958
1959                 if (vmf->flags & FAULT_FLAG_WRITE)
1960                         goto unlock;
1961
1962                 vmf->page = virt_to_page(data->data_pages[nr]);
1963         }
1964
1965         get_page(vmf->page);
1966         vmf->page->mapping = vma->vm_file->f_mapping;
1967         vmf->page->index   = vmf->pgoff;
1968
1969         ret = 0;
1970 unlock:
1971         rcu_read_unlock();
1972
1973         return ret;
1974 }
1975
1976 static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
1977 {
1978         struct perf_mmap_data *data;
1979         unsigned long size;
1980         int i;
1981
1982         WARN_ON(atomic_read(&counter->mmap_count));
1983
1984         size = sizeof(struct perf_mmap_data);
1985         size += nr_pages * sizeof(void *);
1986
1987         data = kzalloc(size, GFP_KERNEL);
1988         if (!data)
1989                 goto fail;
1990
1991         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
1992         if (!data->user_page)
1993                 goto fail_user_page;
1994
1995         for (i = 0; i < nr_pages; i++) {
1996                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
1997                 if (!data->data_pages[i])
1998                         goto fail_data_pages;
1999         }
2000
2001         data->nr_pages = nr_pages;
2002         atomic_set(&data->lock, -1);
2003
2004         rcu_assign_pointer(counter->data, data);
2005
2006         return 0;
2007
2008 fail_data_pages:
2009         for (i--; i >= 0; i--)
2010                 free_page((unsigned long)data->data_pages[i]);
2011
2012         free_page((unsigned long)data->user_page);
2013
2014 fail_user_page:
2015         kfree(data);
2016
2017 fail:
2018         return -ENOMEM;
2019 }
2020
2021 static void perf_mmap_free_page(unsigned long addr)
2022 {
2023         struct page *page = virt_to_page((void *)addr);
2024
2025         page->mapping = NULL;
2026         __free_page(page);
2027 }
2028
2029 static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2030 {
2031         struct perf_mmap_data *data;
2032         int i;
2033
2034         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2035
2036         perf_mmap_free_page((unsigned long)data->user_page);
2037         for (i = 0; i < data->nr_pages; i++)
2038                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2039
2040         kfree(data);
2041 }
2042
2043 static void perf_mmap_data_free(struct perf_counter *counter)
2044 {
2045         struct perf_mmap_data *data = counter->data;
2046
2047         WARN_ON(atomic_read(&counter->mmap_count));
2048
2049         rcu_assign_pointer(counter->data, NULL);
2050         call_rcu(&data->rcu_head, __perf_mmap_data_free);
2051 }
2052
2053 static void perf_mmap_open(struct vm_area_struct *vma)
2054 {
2055         struct perf_counter *counter = vma->vm_file->private_data;
2056
2057         atomic_inc(&counter->mmap_count);
2058 }
2059
2060 static void perf_mmap_close(struct vm_area_struct *vma)
2061 {
2062         struct perf_counter *counter = vma->vm_file->private_data;
2063
2064         WARN_ON_ONCE(counter->ctx->parent_ctx);
2065         if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2066                 struct user_struct *user = current_user();
2067
2068                 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2069                 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2070                 perf_mmap_data_free(counter);
2071                 mutex_unlock(&counter->mmap_mutex);
2072         }
2073 }
2074
2075 static struct vm_operations_struct perf_mmap_vmops = {
2076         .open           = perf_mmap_open,
2077         .close          = perf_mmap_close,
2078         .fault          = perf_mmap_fault,
2079         .page_mkwrite   = perf_mmap_fault,
2080 };
2081
2082 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2083 {
2084         struct perf_counter *counter = file->private_data;
2085         unsigned long user_locked, user_lock_limit;
2086         struct user_struct *user = current_user();
2087         unsigned long locked, lock_limit;
2088         unsigned long vma_size;
2089         unsigned long nr_pages;
2090         long user_extra, extra;
2091         int ret = 0;
2092
2093         if (!(vma->vm_flags & VM_SHARED))
2094                 return -EINVAL;
2095
2096         vma_size = vma->vm_end - vma->vm_start;
2097         nr_pages = (vma_size / PAGE_SIZE) - 1;
2098
2099         /*
2100          * If we have data pages ensure they're a power-of-two number, so we
2101          * can do bitmasks instead of modulo.
2102          */
2103         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2104                 return -EINVAL;
2105
2106         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2107                 return -EINVAL;
2108
2109         if (vma->vm_pgoff != 0)
2110                 return -EINVAL;
2111
2112         WARN_ON_ONCE(counter->ctx->parent_ctx);
2113         mutex_lock(&counter->mmap_mutex);
2114         if (atomic_inc_not_zero(&counter->mmap_count)) {
2115                 if (nr_pages != counter->data->nr_pages)
2116                         ret = -EINVAL;
2117                 goto unlock;
2118         }
2119
2120         user_extra = nr_pages + 1;
2121         user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2122
2123         /*
2124          * Increase the limit linearly with more CPUs:
2125          */
2126         user_lock_limit *= num_online_cpus();
2127
2128         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2129
2130         extra = 0;
2131         if (user_locked > user_lock_limit)
2132                 extra = user_locked - user_lock_limit;
2133
2134         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2135         lock_limit >>= PAGE_SHIFT;
2136         locked = vma->vm_mm->locked_vm + extra;
2137
2138         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2139                 ret = -EPERM;
2140                 goto unlock;
2141         }
2142
2143         WARN_ON(counter->data);
2144         ret = perf_mmap_data_alloc(counter, nr_pages);
2145         if (ret)
2146                 goto unlock;
2147
2148         atomic_set(&counter->mmap_count, 1);
2149         atomic_long_add(user_extra, &user->locked_vm);
2150         vma->vm_mm->locked_vm += extra;
2151         counter->data->nr_locked = extra;
2152         if (vma->vm_flags & VM_WRITE)
2153                 counter->data->writable = 1;
2154
2155 unlock:
2156         mutex_unlock(&counter->mmap_mutex);
2157
2158         vma->vm_flags |= VM_RESERVED;
2159         vma->vm_ops = &perf_mmap_vmops;
2160
2161         return ret;
2162 }
2163
2164 static int perf_fasync(int fd, struct file *filp, int on)
2165 {
2166         struct inode *inode = filp->f_path.dentry->d_inode;
2167         struct perf_counter *counter = filp->private_data;
2168         int retval;
2169
2170         mutex_lock(&inode->i_mutex);
2171         retval = fasync_helper(fd, filp, on, &counter->fasync);
2172         mutex_unlock(&inode->i_mutex);
2173
2174         if (retval < 0)
2175                 return retval;
2176
2177         return 0;
2178 }
2179
2180 static const struct file_operations perf_fops = {
2181         .release                = perf_release,
2182         .read                   = perf_read,
2183         .poll                   = perf_poll,
2184         .unlocked_ioctl         = perf_ioctl,
2185         .compat_ioctl           = perf_ioctl,
2186         .mmap                   = perf_mmap,
2187         .fasync                 = perf_fasync,
2188 };
2189
2190 /*
2191  * Perf counter wakeup
2192  *
2193  * If there's data, ensure we set the poll() state and publish everything
2194  * to user-space before waking everybody up.
2195  */
2196
2197 void perf_counter_wakeup(struct perf_counter *counter)
2198 {
2199         wake_up_all(&counter->waitq);
2200
2201         if (counter->pending_kill) {
2202                 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2203                 counter->pending_kill = 0;
2204         }
2205 }
2206
2207 /*
2208  * Pending wakeups
2209  *
2210  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2211  *
2212  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2213  * single linked list and use cmpxchg() to add entries lockless.
2214  */
2215
2216 static void perf_pending_counter(struct perf_pending_entry *entry)
2217 {
2218         struct perf_counter *counter = container_of(entry,
2219                         struct perf_counter, pending);
2220
2221         if (counter->pending_disable) {
2222                 counter->pending_disable = 0;
2223                 perf_counter_disable(counter);
2224         }
2225
2226         if (counter->pending_wakeup) {
2227                 counter->pending_wakeup = 0;
2228                 perf_counter_wakeup(counter);
2229         }
2230 }
2231
2232 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2233
2234 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2235         PENDING_TAIL,
2236 };
2237
2238 static void perf_pending_queue(struct perf_pending_entry *entry,
2239                                void (*func)(struct perf_pending_entry *))
2240 {
2241         struct perf_pending_entry **head;
2242
2243         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2244                 return;
2245
2246         entry->func = func;
2247
2248         head = &get_cpu_var(perf_pending_head);
2249
2250         do {
2251                 entry->next = *head;
2252         } while (cmpxchg(head, entry->next, entry) != entry->next);
2253
2254         set_perf_counter_pending();
2255
2256         put_cpu_var(perf_pending_head);
2257 }
2258
2259 static int __perf_pending_run(void)
2260 {
2261         struct perf_pending_entry *list;
2262         int nr = 0;
2263
2264         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2265         while (list != PENDING_TAIL) {
2266                 void (*func)(struct perf_pending_entry *);
2267                 struct perf_pending_entry *entry = list;
2268
2269                 list = list->next;
2270
2271                 func = entry->func;
2272                 entry->next = NULL;
2273                 /*
2274                  * Ensure we observe the unqueue before we issue the wakeup,
2275                  * so that we won't be waiting forever.
2276                  * -- see perf_not_pending().
2277                  */
2278                 smp_wmb();
2279
2280                 func(entry);
2281                 nr++;
2282         }
2283
2284         return nr;
2285 }
2286
2287 static inline int perf_not_pending(struct perf_counter *counter)
2288 {
2289         /*
2290          * If we flush on whatever cpu we run, there is a chance we don't
2291          * need to wait.
2292          */
2293         get_cpu();
2294         __perf_pending_run();
2295         put_cpu();
2296
2297         /*
2298          * Ensure we see the proper queue state before going to sleep
2299          * so that we do not miss the wakeup. -- see perf_pending_handle()
2300          */
2301         smp_rmb();
2302         return counter->pending.next == NULL;
2303 }
2304
2305 static void perf_pending_sync(struct perf_counter *counter)
2306 {
2307         wait_event(counter->waitq, perf_not_pending(counter));
2308 }
2309
2310 void perf_counter_do_pending(void)
2311 {
2312         __perf_pending_run();
2313 }
2314
2315 /*
2316  * Callchain support -- arch specific
2317  */
2318
2319 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2320 {
2321         return NULL;
2322 }
2323
2324 /*
2325  * Output
2326  */
2327
2328 struct perf_output_handle {
2329         struct perf_counter     *counter;
2330         struct perf_mmap_data   *data;
2331         unsigned long           head;
2332         unsigned long           offset;
2333         int                     nmi;
2334         int                     sample;
2335         int                     locked;
2336         unsigned long           flags;
2337 };
2338
2339 static bool perf_output_space(struct perf_mmap_data *data,
2340                               unsigned int offset, unsigned int head)
2341 {
2342         unsigned long tail;
2343         unsigned long mask;
2344
2345         if (!data->writable)
2346                 return true;
2347
2348         mask = (data->nr_pages << PAGE_SHIFT) - 1;
2349         /*
2350          * Userspace could choose to issue a mb() before updating the tail
2351          * pointer. So that all reads will be completed before the write is
2352          * issued.
2353          */
2354         tail = ACCESS_ONCE(data->user_page->data_tail);
2355         smp_rmb();
2356
2357         offset = (offset - tail) & mask;
2358         head   = (head   - tail) & mask;
2359
2360         if ((int)(head - offset) < 0)
2361                 return false;
2362
2363         return true;
2364 }
2365
2366 static void perf_output_wakeup(struct perf_output_handle *handle)
2367 {
2368         atomic_set(&handle->data->poll, POLL_IN);
2369
2370         if (handle->nmi) {
2371                 handle->counter->pending_wakeup = 1;
2372                 perf_pending_queue(&handle->counter->pending,
2373                                    perf_pending_counter);
2374         } else
2375                 perf_counter_wakeup(handle->counter);
2376 }
2377
2378 /*
2379  * Curious locking construct.
2380  *
2381  * We need to ensure a later event doesn't publish a head when a former
2382  * event isn't done writing. However since we need to deal with NMIs we
2383  * cannot fully serialize things.
2384  *
2385  * What we do is serialize between CPUs so we only have to deal with NMI
2386  * nesting on a single CPU.
2387  *
2388  * We only publish the head (and generate a wakeup) when the outer-most
2389  * event completes.
2390  */
2391 static void perf_output_lock(struct perf_output_handle *handle)
2392 {
2393         struct perf_mmap_data *data = handle->data;
2394         int cpu;
2395
2396         handle->locked = 0;
2397
2398         local_irq_save(handle->flags);
2399         cpu = smp_processor_id();
2400
2401         if (in_nmi() && atomic_read(&data->lock) == cpu)
2402                 return;
2403
2404         while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2405                 cpu_relax();
2406
2407         handle->locked = 1;
2408 }
2409
2410 static void perf_output_unlock(struct perf_output_handle *handle)
2411 {
2412         struct perf_mmap_data *data = handle->data;
2413         unsigned long head;
2414         int cpu;
2415
2416         data->done_head = data->head;
2417
2418         if (!handle->locked)
2419                 goto out;
2420
2421 again:
2422         /*
2423          * The xchg implies a full barrier that ensures all writes are done
2424          * before we publish the new head, matched by a rmb() in userspace when
2425          * reading this position.
2426          */
2427         while ((head = atomic_long_xchg(&data->done_head, 0)))
2428                 data->user_page->data_head = head;
2429
2430         /*
2431          * NMI can happen here, which means we can miss a done_head update.
2432          */
2433
2434         cpu = atomic_xchg(&data->lock, -1);
2435         WARN_ON_ONCE(cpu != smp_processor_id());
2436
2437         /*
2438          * Therefore we have to validate we did not indeed do so.
2439          */
2440         if (unlikely(atomic_long_read(&data->done_head))) {
2441                 /*
2442                  * Since we had it locked, we can lock it again.
2443                  */
2444                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2445                         cpu_relax();
2446
2447                 goto again;
2448         }
2449
2450         if (atomic_xchg(&data->wakeup, 0))
2451                 perf_output_wakeup(handle);
2452 out:
2453         local_irq_restore(handle->flags);
2454 }
2455
2456 static void perf_output_copy(struct perf_output_handle *handle,
2457                              const void *buf, unsigned int len)
2458 {
2459         unsigned int pages_mask;
2460         unsigned int offset;
2461         unsigned int size;
2462         void **pages;
2463
2464         offset          = handle->offset;
2465         pages_mask      = handle->data->nr_pages - 1;
2466         pages           = handle->data->data_pages;
2467
2468         do {
2469                 unsigned int page_offset;
2470                 int nr;
2471
2472                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2473                 page_offset = offset & (PAGE_SIZE - 1);
2474                 size        = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2475
2476                 memcpy(pages[nr] + page_offset, buf, size);
2477
2478                 len         -= size;
2479                 buf         += size;
2480                 offset      += size;
2481         } while (len);
2482
2483         handle->offset = offset;
2484
2485         /*
2486          * Check we didn't copy past our reservation window, taking the
2487          * possible unsigned int wrap into account.
2488          */
2489         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2490 }
2491
2492 #define perf_output_put(handle, x) \
2493         perf_output_copy((handle), &(x), sizeof(x))
2494
2495 static int perf_output_begin(struct perf_output_handle *handle,
2496                              struct perf_counter *counter, unsigned int size,
2497                              int nmi, int sample)
2498 {
2499         struct perf_mmap_data *data;
2500         unsigned int offset, head;
2501         int have_lost;
2502         struct {
2503                 struct perf_event_header header;
2504                 u64                      id;
2505                 u64                      lost;
2506         } lost_event;
2507
2508         /*
2509          * For inherited counters we send all the output towards the parent.
2510          */
2511         if (counter->parent)
2512                 counter = counter->parent;
2513
2514         rcu_read_lock();
2515         data = rcu_dereference(counter->data);
2516         if (!data)
2517                 goto out;
2518
2519         handle->data    = data;
2520         handle->counter = counter;
2521         handle->nmi     = nmi;
2522         handle->sample  = sample;
2523
2524         if (!data->nr_pages)
2525                 goto fail;
2526
2527         have_lost = atomic_read(&data->lost);
2528         if (have_lost)
2529                 size += sizeof(lost_event);
2530
2531         perf_output_lock(handle);
2532
2533         do {
2534                 offset = head = atomic_long_read(&data->head);
2535                 head += size;
2536                 if (unlikely(!perf_output_space(data, offset, head)))
2537                         goto fail;
2538         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2539
2540         handle->offset  = offset;
2541         handle->head    = head;
2542
2543         if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2544                 atomic_set(&data->wakeup, 1);
2545
2546         if (have_lost) {
2547                 lost_event.header.type = PERF_EVENT_LOST;
2548                 lost_event.header.misc = 0;
2549                 lost_event.header.size = sizeof(lost_event);
2550                 lost_event.id          = counter->id;
2551                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2552
2553                 perf_output_put(handle, lost_event);
2554         }
2555
2556         return 0;
2557
2558 fail:
2559         atomic_inc(&data->lost);
2560         perf_output_unlock(handle);
2561 out:
2562         rcu_read_unlock();
2563
2564         return -ENOSPC;
2565 }
2566
2567 static void perf_output_end(struct perf_output_handle *handle)
2568 {
2569         struct perf_counter *counter = handle->counter;
2570         struct perf_mmap_data *data = handle->data;
2571
2572         int wakeup_events = counter->attr.wakeup_events;
2573
2574         if (handle->sample && wakeup_events) {
2575                 int events = atomic_inc_return(&data->events);
2576                 if (events >= wakeup_events) {
2577                         atomic_sub(wakeup_events, &data->events);
2578                         atomic_set(&data->wakeup, 1);
2579                 }
2580         }
2581
2582         perf_output_unlock(handle);
2583         rcu_read_unlock();
2584 }
2585
2586 static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2587 {
2588         /*
2589          * only top level counters have the pid namespace they were created in
2590          */
2591         if (counter->parent)
2592                 counter = counter->parent;
2593
2594         return task_tgid_nr_ns(p, counter->ns);
2595 }
2596
2597 static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2598 {
2599         /*
2600          * only top level counters have the pid namespace they were created in
2601          */
2602         if (counter->parent)
2603                 counter = counter->parent;
2604
2605         return task_pid_nr_ns(p, counter->ns);
2606 }
2607
2608 static void perf_counter_output(struct perf_counter *counter, int nmi,
2609                                 struct perf_sample_data *data)
2610 {
2611         int ret;
2612         u64 sample_type = counter->attr.sample_type;
2613         struct perf_output_handle handle;
2614         struct perf_event_header header;
2615         u64 ip;
2616         struct {
2617                 u32 pid, tid;
2618         } tid_entry;
2619         struct {
2620                 u64 id;
2621                 u64 counter;
2622         } group_entry;
2623         struct perf_callchain_entry *callchain = NULL;
2624         int callchain_size = 0;
2625         u64 time;
2626         struct {
2627                 u32 cpu, reserved;
2628         } cpu_entry;
2629
2630         header.type = PERF_EVENT_SAMPLE;
2631         header.size = sizeof(header);
2632
2633         header.misc = 0;
2634         header.misc |= perf_misc_flags(data->regs);
2635
2636         if (sample_type & PERF_SAMPLE_IP) {
2637                 ip = perf_instruction_pointer(data->regs);
2638                 header.size += sizeof(ip);
2639         }
2640
2641         if (sample_type & PERF_SAMPLE_TID) {
2642                 /* namespace issues */
2643                 tid_entry.pid = perf_counter_pid(counter, current);
2644                 tid_entry.tid = perf_counter_tid(counter, current);
2645
2646                 header.size += sizeof(tid_entry);
2647         }
2648
2649         if (sample_type & PERF_SAMPLE_TIME) {
2650                 /*
2651                  * Maybe do better on x86 and provide cpu_clock_nmi()
2652                  */
2653                 time = sched_clock();
2654
2655                 header.size += sizeof(u64);
2656         }
2657
2658         if (sample_type & PERF_SAMPLE_ADDR)
2659                 header.size += sizeof(u64);
2660
2661         if (sample_type & PERF_SAMPLE_ID)
2662                 header.size += sizeof(u64);
2663
2664         if (sample_type & PERF_SAMPLE_CPU) {
2665                 header.size += sizeof(cpu_entry);
2666
2667                 cpu_entry.cpu = raw_smp_processor_id();
2668                 cpu_entry.reserved = 0;
2669         }
2670
2671         if (sample_type & PERF_SAMPLE_PERIOD)
2672                 header.size += sizeof(u64);
2673
2674         if (sample_type & PERF_SAMPLE_GROUP) {
2675                 header.size += sizeof(u64) +
2676                         counter->nr_siblings * sizeof(group_entry);
2677         }
2678
2679         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2680                 callchain = perf_callchain(data->regs);
2681
2682                 if (callchain) {
2683                         callchain_size = (1 + callchain->nr) * sizeof(u64);
2684                         header.size += callchain_size;
2685                 } else
2686                         header.size += sizeof(u64);
2687         }
2688
2689         ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2690         if (ret)
2691                 return;
2692
2693         perf_output_put(&handle, header);
2694
2695         if (sample_type & PERF_SAMPLE_IP)
2696                 perf_output_put(&handle, ip);
2697
2698         if (sample_type & PERF_SAMPLE_TID)
2699                 perf_output_put(&handle, tid_entry);
2700
2701         if (sample_type & PERF_SAMPLE_TIME)
2702                 perf_output_put(&handle, time);
2703
2704         if (sample_type & PERF_SAMPLE_ADDR)
2705                 perf_output_put(&handle, data->addr);
2706
2707         if (sample_type & PERF_SAMPLE_ID)
2708                 perf_output_put(&handle, counter->id);
2709
2710         if (sample_type & PERF_SAMPLE_CPU)
2711                 perf_output_put(&handle, cpu_entry);
2712
2713         if (sample_type & PERF_SAMPLE_PERIOD)
2714                 perf_output_put(&handle, data->period);
2715
2716         /*
2717          * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
2718          */
2719         if (sample_type & PERF_SAMPLE_GROUP) {
2720                 struct perf_counter *leader, *sub;
2721                 u64 nr = counter->nr_siblings;
2722
2723                 perf_output_put(&handle, nr);
2724
2725                 leader = counter->group_leader;
2726                 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2727                         if (sub != counter)
2728                                 sub->pmu->read(sub);
2729
2730                         group_entry.id = sub->id;
2731                         group_entry.counter = atomic64_read(&sub->count);
2732
2733                         perf_output_put(&handle, group_entry);
2734                 }
2735         }
2736
2737         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2738                 if (callchain)
2739                         perf_output_copy(&handle, callchain, callchain_size);
2740                 else {
2741                         u64 nr = 0;
2742                         perf_output_put(&handle, nr);
2743                 }
2744         }
2745
2746         perf_output_end(&handle);
2747 }
2748
2749 /*
2750  * read event
2751  */
2752
2753 struct perf_read_event {
2754         struct perf_event_header        header;
2755
2756         u32                             pid;
2757         u32                             tid;
2758         u64                             value;
2759         u64                             format[3];
2760 };
2761
2762 static void
2763 perf_counter_read_event(struct perf_counter *counter,
2764                         struct task_struct *task)
2765 {
2766         struct perf_output_handle handle;
2767         struct perf_read_event event = {
2768                 .header = {
2769                         .type = PERF_EVENT_READ,
2770                         .misc = 0,
2771                         .size = sizeof(event) - sizeof(event.format),
2772                 },
2773                 .pid = perf_counter_pid(counter, task),
2774                 .tid = perf_counter_tid(counter, task),
2775                 .value = atomic64_read(&counter->count),
2776         };
2777         int ret, i = 0;
2778
2779         if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2780                 event.header.size += sizeof(u64);
2781                 event.format[i++] = counter->total_time_enabled;
2782         }
2783
2784         if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2785                 event.header.size += sizeof(u64);
2786                 event.format[i++] = counter->total_time_running;
2787         }
2788
2789         if (counter->attr.read_format & PERF_FORMAT_ID) {
2790                 u64 id;
2791
2792                 event.header.size += sizeof(u64);
2793                 if (counter->parent)
2794                         id = counter->parent->id;
2795                 else
2796                         id = counter->id;
2797
2798                 event.format[i++] = id;
2799         }
2800
2801         ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
2802         if (ret)
2803                 return;
2804
2805         perf_output_copy(&handle, &event, event.header.size);
2806         perf_output_end(&handle);
2807 }
2808
2809 /*
2810  * fork tracking
2811  */
2812
2813 struct perf_fork_event {
2814         struct task_struct      *task;
2815
2816         struct {
2817                 struct perf_event_header        header;
2818
2819                 u32                             pid;
2820                 u32                             ppid;
2821         } event;
2822 };
2823
2824 static void perf_counter_fork_output(struct perf_counter *counter,
2825                                      struct perf_fork_event *fork_event)
2826 {
2827         struct perf_output_handle handle;
2828         int size = fork_event->event.header.size;
2829         struct task_struct *task = fork_event->task;
2830         int ret = perf_output_begin(&handle, counter, size, 0, 0);
2831
2832         if (ret)
2833                 return;
2834
2835         fork_event->event.pid = perf_counter_pid(counter, task);
2836         fork_event->event.ppid = perf_counter_pid(counter, task->real_parent);
2837
2838         perf_output_put(&handle, fork_event->event);
2839         perf_output_end(&handle);
2840 }
2841
2842 static int perf_counter_fork_match(struct perf_counter *counter)
2843 {
2844         if (counter->attr.comm || counter->attr.mmap)
2845                 return 1;
2846
2847         return 0;
2848 }
2849
2850 static void perf_counter_fork_ctx(struct perf_counter_context *ctx,
2851                                   struct perf_fork_event *fork_event)
2852 {
2853         struct perf_counter *counter;
2854
2855         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2856                 return;
2857
2858         rcu_read_lock();
2859         list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2860                 if (perf_counter_fork_match(counter))
2861                         perf_counter_fork_output(counter, fork_event);
2862         }
2863         rcu_read_unlock();
2864 }
2865
2866 static void perf_counter_fork_event(struct perf_fork_event *fork_event)
2867 {
2868         struct perf_cpu_context *cpuctx;
2869         struct perf_counter_context *ctx;
2870
2871         cpuctx = &get_cpu_var(perf_cpu_context);
2872         perf_counter_fork_ctx(&cpuctx->ctx, fork_event);
2873         put_cpu_var(perf_cpu_context);
2874
2875         rcu_read_lock();
2876         /*
2877          * doesn't really matter which of the child contexts the
2878          * events ends up in.
2879          */
2880         ctx = rcu_dereference(current->perf_counter_ctxp);
2881         if (ctx)
2882                 perf_counter_fork_ctx(ctx, fork_event);
2883         rcu_read_unlock();
2884 }
2885
2886 void perf_counter_fork(struct task_struct *task)
2887 {
2888         struct perf_fork_event fork_event;
2889
2890         if (!atomic_read(&nr_comm_counters) &&
2891             !atomic_read(&nr_mmap_counters))
2892                 return;
2893
2894         fork_event = (struct perf_fork_event){
2895                 .task   = task,
2896                 .event  = {
2897                         .header = {
2898                                 .type = PERF_EVENT_FORK,
2899                                 .size = sizeof(fork_event.event),
2900                         },
2901                 },
2902         };
2903
2904         perf_counter_fork_event(&fork_event);
2905 }
2906
2907 /*
2908  * comm tracking
2909  */
2910
2911 struct perf_comm_event {
2912         struct task_struct      *task;
2913         char                    *comm;
2914         int                     comm_size;
2915
2916         struct {
2917                 struct perf_event_header        header;
2918
2919                 u32                             pid;
2920                 u32                             tid;
2921         } event;
2922 };
2923
2924 static void perf_counter_comm_output(struct perf_counter *counter,
2925                                      struct perf_comm_event *comm_event)
2926 {
2927         struct perf_output_handle handle;
2928         int size = comm_event->event.header.size;
2929         int ret = perf_output_begin(&handle, counter, size, 0, 0);
2930
2931         if (ret)
2932                 return;
2933
2934         comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
2935         comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
2936
2937         perf_output_put(&handle, comm_event->event);
2938         perf_output_copy(&handle, comm_event->comm,
2939                                    comm_event->comm_size);
2940         perf_output_end(&handle);
2941 }
2942
2943 static int perf_counter_comm_match(struct perf_counter *counter)
2944 {
2945         if (counter->attr.comm)
2946                 return 1;
2947
2948         return 0;
2949 }
2950
2951 static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
2952                                   struct perf_comm_event *comm_event)
2953 {
2954         struct perf_counter *counter;
2955
2956         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
2957                 return;
2958
2959         rcu_read_lock();
2960         list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
2961                 if (perf_counter_comm_match(counter))
2962                         perf_counter_comm_output(counter, comm_event);
2963         }
2964         rcu_read_unlock();
2965 }
2966
2967 static void perf_counter_comm_event(struct perf_comm_event *comm_event)
2968 {
2969         struct perf_cpu_context *cpuctx;
2970         struct perf_counter_context *ctx;
2971         unsigned int size;
2972         char *comm = comm_event->task->comm;
2973
2974         size = ALIGN(strlen(comm)+1, sizeof(u64));
2975
2976         comm_event->comm = comm;
2977         comm_event->comm_size = size;
2978
2979         comm_event->event.header.size = sizeof(comm_event->event) + size;
2980
2981         cpuctx = &get_cpu_var(perf_cpu_context);
2982         perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
2983         put_cpu_var(perf_cpu_context);
2984
2985         rcu_read_lock();
2986         /*
2987          * doesn't really matter which of the child contexts the
2988          * events ends up in.
2989          */
2990         ctx = rcu_dereference(current->perf_counter_ctxp);
2991         if (ctx)
2992                 perf_counter_comm_ctx(ctx, comm_event);
2993         rcu_read_unlock();
2994 }
2995
2996 void perf_counter_comm(struct task_struct *task)
2997 {
2998         struct perf_comm_event comm_event;
2999
3000         if (task->perf_counter_ctxp)
3001                 perf_counter_enable_on_exec(task);
3002
3003         if (!atomic_read(&nr_comm_counters))
3004                 return;
3005
3006         comm_event = (struct perf_comm_event){
3007                 .task   = task,
3008                 .event  = {
3009                         .header = { .type = PERF_EVENT_COMM, },
3010                 },
3011         };
3012
3013         perf_counter_comm_event(&comm_event);
3014 }
3015
3016 /*
3017  * mmap tracking
3018  */
3019
3020 struct perf_mmap_event {
3021         struct vm_area_struct   *vma;
3022
3023         const char              *file_name;
3024         int                     file_size;
3025
3026         struct {
3027                 struct perf_event_header        header;
3028
3029                 u32                             pid;
3030                 u32                             tid;
3031                 u64                             start;
3032                 u64                             len;
3033                 u64                             pgoff;
3034         } event;
3035 };
3036
3037 static void perf_counter_mmap_output(struct perf_counter *counter,
3038                                      struct perf_mmap_event *mmap_event)
3039 {
3040         struct perf_output_handle handle;
3041         int size = mmap_event->event.header.size;
3042         int ret = perf_output_begin(&handle, counter, size, 0, 0);
3043
3044         if (ret)
3045                 return;
3046
3047         mmap_event->event.pid = perf_counter_pid(counter, current);
3048         mmap_event->event.tid = perf_counter_tid(counter, current);
3049
3050         perf_output_put(&handle, mmap_event->event);
3051         perf_output_copy(&handle, mmap_event->file_name,
3052                                    mmap_event->file_size);
3053         perf_output_end(&handle);
3054 }
3055
3056 static int perf_counter_mmap_match(struct perf_counter *counter,
3057                                    struct perf_mmap_event *mmap_event)
3058 {
3059         if (counter->attr.mmap)
3060                 return 1;
3061
3062         return 0;
3063 }
3064
3065 static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3066                                   struct perf_mmap_event *mmap_event)
3067 {
3068         struct perf_counter *counter;
3069
3070         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3071                 return;
3072
3073         rcu_read_lock();
3074         list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3075                 if (perf_counter_mmap_match(counter, mmap_event))
3076                         perf_counter_mmap_output(counter, mmap_event);
3077         }
3078         rcu_read_unlock();
3079 }
3080
3081 static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3082 {
3083         struct perf_cpu_context *cpuctx;
3084         struct perf_counter_context *ctx;
3085         struct vm_area_struct *vma = mmap_event->vma;
3086         struct file *file = vma->vm_file;
3087         unsigned int size;
3088         char tmp[16];
3089         char *buf = NULL;
3090         const char *name;
3091
3092         if (file) {
3093                 buf = kzalloc(PATH_MAX, GFP_KERNEL);
3094                 if (!buf) {
3095                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3096                         goto got_name;
3097                 }
3098                 name = d_path(&file->f_path, buf, PATH_MAX);
3099                 if (IS_ERR(name)) {
3100                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3101                         goto got_name;
3102                 }
3103         } else {
3104                 name = arch_vma_name(mmap_event->vma);
3105                 if (name)
3106                         goto got_name;
3107
3108                 if (!vma->vm_mm) {
3109                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3110                         goto got_name;
3111                 }
3112
3113                 name = strncpy(tmp, "//anon", sizeof(tmp));
3114                 goto got_name;
3115         }
3116
3117 got_name:
3118         size = ALIGN(strlen(name)+1, sizeof(u64));
3119
3120         mmap_event->file_name = name;
3121         mmap_event->file_size = size;
3122
3123         mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3124
3125         cpuctx = &get_cpu_var(perf_cpu_context);
3126         perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3127         put_cpu_var(perf_cpu_context);
3128
3129         rcu_read_lock();
3130         /*
3131          * doesn't really matter which of the child contexts the
3132          * events ends up in.
3133          */
3134         ctx = rcu_dereference(current->perf_counter_ctxp);
3135         if (ctx)
3136                 perf_counter_mmap_ctx(ctx, mmap_event);
3137         rcu_read_unlock();
3138
3139         kfree(buf);
3140 }
3141
3142 void __perf_counter_mmap(struct vm_area_struct *vma)
3143 {
3144         struct perf_mmap_event mmap_event;
3145
3146         if (!atomic_read(&nr_mmap_counters))
3147                 return;
3148
3149         mmap_event = (struct perf_mmap_event){
3150                 .vma    = vma,
3151                 .event  = {
3152                         .header = { .type = PERF_EVENT_MMAP, },
3153                         .start  = vma->vm_start,
3154                         .len    = vma->vm_end - vma->vm_start,
3155                         .pgoff  = vma->vm_pgoff,
3156                 },
3157         };
3158
3159         perf_counter_mmap_event(&mmap_event);
3160 }
3161
3162 /*
3163  * Log sample_period changes so that analyzing tools can re-normalize the
3164  * event flow.
3165  */
3166
3167 struct freq_event {
3168         struct perf_event_header        header;
3169         u64                             time;
3170         u64                             id;
3171         u64                             period;
3172 };
3173
3174 static void perf_log_period(struct perf_counter *counter, u64 period)
3175 {
3176         struct perf_output_handle handle;
3177         struct freq_event event;
3178         int ret;
3179
3180         if (counter->hw.sample_period == period)
3181                 return;
3182
3183         if (counter->attr.sample_type & PERF_SAMPLE_PERIOD)
3184                 return;
3185
3186         event = (struct freq_event) {
3187                 .header = {
3188                         .type = PERF_EVENT_PERIOD,
3189                         .misc = 0,
3190                         .size = sizeof(event),
3191                 },
3192                 .time = sched_clock(),
3193                 .id = counter->id,
3194                 .period = period,
3195         };
3196
3197         ret = perf_output_begin(&handle, counter, sizeof(event), 1, 0);
3198         if (ret)
3199                 return;
3200
3201         perf_output_put(&handle, event);
3202         perf_output_end(&handle);
3203 }
3204
3205 /*
3206  * IRQ throttle logging
3207  */
3208
3209 static void perf_log_throttle(struct perf_counter *counter, int enable)
3210 {
3211         struct perf_output_handle handle;
3212         int ret;
3213
3214         struct {
3215                 struct perf_event_header        header;
3216                 u64                             time;
3217                 u64                             id;
3218         } throttle_event = {
3219                 .header = {
3220                         .type = PERF_EVENT_THROTTLE + 1,
3221                         .misc = 0,
3222                         .size = sizeof(throttle_event),
3223                 },
3224                 .time   = sched_clock(),
3225                 .id     = counter->id,
3226         };
3227
3228         ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3229         if (ret)
3230                 return;
3231
3232         perf_output_put(&handle, throttle_event);
3233         perf_output_end(&handle);
3234 }
3235
3236 /*
3237  * Generic counter overflow handling, sampling.
3238  */
3239
3240 int perf_counter_overflow(struct perf_counter *counter, int nmi,
3241                           struct perf_sample_data *data)
3242 {
3243         int events = atomic_read(&counter->event_limit);
3244         int throttle = counter->pmu->unthrottle != NULL;
3245         struct hw_perf_counter *hwc = &counter->hw;
3246         int ret = 0;
3247
3248         if (!throttle) {
3249                 hwc->interrupts++;
3250         } else {
3251                 if (hwc->interrupts != MAX_INTERRUPTS) {
3252                         hwc->interrupts++;
3253                         if (HZ * hwc->interrupts >
3254                                         (u64)sysctl_perf_counter_sample_rate) {
3255                                 hwc->interrupts = MAX_INTERRUPTS;
3256                                 perf_log_throttle(counter, 0);
3257                                 ret = 1;
3258                         }
3259                 } else {
3260                         /*
3261                          * Keep re-disabling counters even though on the previous
3262                          * pass we disabled it - just in case we raced with a
3263                          * sched-in and the counter got enabled again:
3264                          */
3265                         ret = 1;
3266                 }
3267         }
3268
3269         if (counter->attr.freq) {
3270                 u64 now = sched_clock();
3271                 s64 delta = now - hwc->freq_stamp;
3272
3273                 hwc->freq_stamp = now;
3274
3275                 if (delta > 0 && delta < TICK_NSEC)
3276                         perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3277         }
3278
3279         /*
3280          * XXX event_limit might not quite work as expected on inherited
3281          * counters
3282          */
3283
3284         counter->pending_kill = POLL_IN;
3285         if (events && atomic_dec_and_test(&counter->event_limit)) {
3286                 ret = 1;
3287                 counter->pending_kill = POLL_HUP;
3288                 if (nmi) {
3289                         counter->pending_disable = 1;
3290                         perf_pending_queue(&counter->pending,
3291                                            perf_pending_counter);
3292                 } else
3293                         perf_counter_disable(counter);
3294         }
3295
3296         perf_counter_output(counter, nmi, data);
3297         return ret;
3298 }
3299
3300 /*
3301  * Generic software counter infrastructure
3302  */
3303
3304 static void perf_swcounter_update(struct perf_counter *counter)
3305 {
3306         struct hw_perf_counter *hwc = &counter->hw;
3307         u64 prev, now;
3308         s64 delta;
3309
3310 again:
3311         prev = atomic64_read(&hwc->prev_count);
3312         now = atomic64_read(&hwc->count);
3313         if (atomic64_cmpxchg(&hwc->prev_count, prev, now) != prev)
3314                 goto again;
3315
3316         delta = now - prev;
3317
3318         atomic64_add(delta, &counter->count);
3319         atomic64_sub(delta, &hwc->period_left);
3320 }
3321
3322 static void perf_swcounter_set_period(struct perf_counter *counter)
3323 {
3324         struct hw_perf_counter *hwc = &counter->hw;
3325         s64 left = atomic64_read(&hwc->period_left);
3326         s64 period = hwc->sample_period;
3327
3328         if (unlikely(left <= -period)) {
3329                 left = period;
3330                 atomic64_set(&hwc->period_left, left);
3331                 hwc->last_period = period;
3332         }
3333
3334         if (unlikely(left <= 0)) {
3335                 left += period;
3336                 atomic64_add(period, &hwc->period_left);
3337                 hwc->last_period = period;
3338         }
3339
3340         atomic64_set(&hwc->prev_count, -left);
3341         atomic64_set(&hwc->count, -left);
3342 }
3343
3344 static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3345 {
3346         enum hrtimer_restart ret = HRTIMER_RESTART;
3347         struct perf_sample_data data;
3348         struct perf_counter *counter;
3349         u64 period;
3350
3351         counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3352         counter->pmu->read(counter);
3353
3354         data.addr = 0;
3355         data.regs = get_irq_regs();
3356         /*
3357          * In case we exclude kernel IPs or are somehow not in interrupt
3358          * context, provide the next best thing, the user IP.
3359          */
3360         if ((counter->attr.exclude_kernel || !data.regs) &&
3361                         !counter->attr.exclude_user)
3362                 data.regs = task_pt_regs(current);
3363
3364         if (data.regs) {
3365                 if (perf_counter_overflow(counter, 0, &data))
3366                         ret = HRTIMER_NORESTART;
3367         }
3368
3369         period = max_t(u64, 10000, counter->hw.sample_period);
3370         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3371
3372         return ret;
3373 }
3374
3375 static void perf_swcounter_overflow(struct perf_counter *counter,
3376                                     int nmi, struct perf_sample_data *data)
3377 {
3378         data->period = counter->hw.last_period;
3379
3380         perf_swcounter_update(counter);
3381         perf_swcounter_set_period(counter);
3382         if (perf_counter_overflow(counter, nmi, data))
3383                 /* soft-disable the counter */
3384                 ;
3385 }
3386
3387 static int perf_swcounter_is_counting(struct perf_counter *counter)
3388 {
3389         struct perf_counter_context *ctx;
3390         unsigned long flags;
3391         int count;
3392
3393         if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3394                 return 1;
3395
3396         if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3397                 return 0;
3398
3399         /*
3400          * If the counter is inactive, it could be just because
3401          * its task is scheduled out, or because it's in a group
3402          * which could not go on the PMU.  We want to count in
3403          * the first case but not the second.  If the context is
3404          * currently active then an inactive software counter must
3405          * be the second case.  If it's not currently active then
3406          * we need to know whether the counter was active when the
3407          * context was last active, which we can determine by
3408          * comparing counter->tstamp_stopped with ctx->time.
3409          *
3410          * We are within an RCU read-side critical section,
3411          * which protects the existence of *ctx.
3412          */
3413         ctx = counter->ctx;
3414         spin_lock_irqsave(&ctx->lock, flags);
3415         count = 1;
3416         /* Re-check state now we have the lock */
3417         if (counter->state < PERF_COUNTER_STATE_INACTIVE ||
3418             counter->ctx->is_active ||
3419             counter->tstamp_stopped < ctx->time)
3420                 count = 0;
3421         spin_unlock_irqrestore(&ctx->lock, flags);
3422         return count;
3423 }
3424
3425 static int perf_swcounter_match(struct perf_counter *counter,
3426                                 enum perf_type_id type,
3427                                 u32 event, struct pt_regs *regs)
3428 {
3429         if (!perf_swcounter_is_counting(counter))
3430                 return 0;
3431
3432         if (counter->attr.type != type)
3433                 return 0;
3434         if (counter->attr.config != event)
3435                 return 0;
3436
3437         if (regs) {
3438                 if (counter->attr.exclude_user && user_mode(regs))
3439                         return 0;
3440
3441                 if (counter->attr.exclude_kernel && !user_mode(regs))
3442                         return 0;
3443         }
3444
3445         return 1;
3446 }
3447
3448 static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3449                                int nmi, struct perf_sample_data *data)
3450 {
3451         int neg = atomic64_add_negative(nr, &counter->hw.count);
3452
3453         if (counter->hw.sample_period && !neg && data->regs)
3454                 perf_swcounter_overflow(counter, nmi, data);
3455 }
3456
3457 static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3458                                      enum perf_type_id type,
3459                                      u32 event, u64 nr, int nmi,
3460                                      struct perf_sample_data *data)
3461 {
3462         struct perf_counter *counter;
3463
3464         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3465                 return;
3466
3467         rcu_read_lock();
3468         list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3469                 if (perf_swcounter_match(counter, type, event, data->regs))
3470                         perf_swcounter_add(counter, nr, nmi, data);
3471         }
3472         rcu_read_unlock();
3473 }
3474
3475 static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3476 {
3477         if (in_nmi())
3478                 return &cpuctx->recursion[3];
3479
3480         if (in_irq())
3481                 return &cpuctx->recursion[2];
3482
3483         if (in_softirq())
3484                 return &cpuctx->recursion[1];
3485
3486         return &cpuctx->recursion[0];
3487 }
3488
3489 static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3490                                     u64 nr, int nmi,
3491                                     struct perf_sample_data *data)
3492 {
3493         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3494         int *recursion = perf_swcounter_recursion_context(cpuctx);
3495         struct perf_counter_context *ctx;
3496
3497         if (*recursion)
3498                 goto out;
3499
3500         (*recursion)++;
3501         barrier();
3502
3503         perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3504                                  nr, nmi, data);
3505         rcu_read_lock();
3506         /*
3507          * doesn't really matter which of the child contexts the
3508          * events ends up in.
3509          */
3510         ctx = rcu_dereference(current->perf_counter_ctxp);
3511         if (ctx)
3512                 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3513         rcu_read_unlock();
3514
3515         barrier();
3516         (*recursion)--;
3517
3518 out:
3519         put_cpu_var(perf_cpu_context);
3520 }
3521
3522 void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3523                             struct pt_regs *regs, u64 addr)
3524 {
3525         struct perf_sample_data data = {
3526                 .regs = regs,
3527                 .addr = addr,
3528         };
3529
3530         do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3531 }
3532
3533 static void perf_swcounter_read(struct perf_counter *counter)
3534 {
3535         perf_swcounter_update(counter);
3536 }
3537
3538 static int perf_swcounter_enable(struct perf_counter *counter)
3539 {
3540         perf_swcounter_set_period(counter);
3541         return 0;
3542 }
3543
3544 static void perf_swcounter_disable(struct perf_counter *counter)
3545 {
3546         perf_swcounter_update(counter);
3547 }
3548
3549 static const struct pmu perf_ops_generic = {
3550         .enable         = perf_swcounter_enable,
3551         .disable        = perf_swcounter_disable,
3552         .read           = perf_swcounter_read,
3553 };
3554
3555 /*
3556  * Software counter: cpu wall time clock
3557  */
3558
3559 static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3560 {
3561         int cpu = raw_smp_processor_id();
3562         s64 prev;
3563         u64 now;
3564
3565         now = cpu_clock(cpu);
3566         prev = atomic64_read(&counter->hw.prev_count);
3567         atomic64_set(&counter->hw.prev_count, now);
3568         atomic64_add(now - prev, &counter->count);
3569 }
3570
3571 static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3572 {
3573         struct hw_perf_counter *hwc = &counter->hw;
3574         int cpu = raw_smp_processor_id();
3575
3576         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3577         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3578         hwc->hrtimer.function = perf_swcounter_hrtimer;
3579         if (hwc->sample_period) {
3580                 u64 period = max_t(u64, 10000, hwc->sample_period);
3581                 __hrtimer_start_range_ns(&hwc->hrtimer,
3582                                 ns_to_ktime(period), 0,
3583                                 HRTIMER_MODE_REL, 0);
3584         }
3585
3586         return 0;
3587 }
3588
3589 static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3590 {
3591         if (counter->hw.sample_period)
3592                 hrtimer_cancel(&counter->hw.hrtimer);
3593         cpu_clock_perf_counter_update(counter);
3594 }
3595
3596 static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3597 {
3598         cpu_clock_perf_counter_update(counter);
3599 }
3600
3601 static const struct pmu perf_ops_cpu_clock = {
3602         .enable         = cpu_clock_perf_counter_enable,
3603         .disable        = cpu_clock_perf_counter_disable,
3604         .read           = cpu_clock_perf_counter_read,
3605 };
3606
3607 /*
3608  * Software counter: task time clock
3609  */
3610
3611 static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3612 {
3613         u64 prev;
3614         s64 delta;
3615
3616         prev = atomic64_xchg(&counter->hw.prev_count, now);
3617         delta = now - prev;
3618         atomic64_add(delta, &counter->count);
3619 }
3620
3621 static int task_clock_perf_counter_enable(struct perf_counter *counter)
3622 {
3623         struct hw_perf_counter *hwc = &counter->hw;
3624         u64 now;
3625
3626         now = counter->ctx->time;
3627
3628         atomic64_set(&hwc->prev_count, now);
3629         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3630         hwc->hrtimer.function = perf_swcounter_hrtimer;
3631         if (hwc->sample_period) {
3632                 u64 period = max_t(u64, 10000, hwc->sample_period);
3633                 __hrtimer_start_range_ns(&hwc->hrtimer,
3634                                 ns_to_ktime(period), 0,
3635                                 HRTIMER_MODE_REL, 0);
3636         }
3637
3638         return 0;
3639 }
3640
3641 static void task_clock_perf_counter_disable(struct perf_counter *counter)
3642 {
3643         if (counter->hw.sample_period)
3644                 hrtimer_cancel(&counter->hw.hrtimer);
3645         task_clock_perf_counter_update(counter, counter->ctx->time);
3646
3647 }
3648
3649 static void task_clock_perf_counter_read(struct perf_counter *counter)
3650 {
3651         u64 time;
3652
3653         if (!in_nmi()) {
3654                 update_context_time(counter->ctx);
3655                 time = counter->ctx->time;
3656         } else {
3657                 u64 now = perf_clock();
3658                 u64 delta = now - counter->ctx->timestamp;
3659                 time = counter->ctx->time + delta;
3660         }
3661
3662         task_clock_perf_counter_update(counter, time);
3663 }
3664
3665 static const struct pmu perf_ops_task_clock = {
3666         .enable         = task_clock_perf_counter_enable,
3667         .disable        = task_clock_perf_counter_disable,
3668         .read           = task_clock_perf_counter_read,
3669 };
3670
3671 #ifdef CONFIG_EVENT_PROFILE
3672 void perf_tpcounter_event(int event_id)
3673 {
3674         struct perf_sample_data data = {
3675                 .regs = get_irq_regs();
3676                 .addr = 0,
3677         };
3678
3679         if (!data.regs)
3680                 data.regs = task_pt_regs(current);
3681
3682         do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, 1, 1, &data);
3683 }
3684 EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3685
3686 extern int ftrace_profile_enable(int);
3687 extern void ftrace_profile_disable(int);
3688
3689 static void tp_perf_counter_destroy(struct perf_counter *counter)
3690 {
3691         ftrace_profile_disable(perf_event_id(&counter->attr));
3692 }
3693
3694 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3695 {
3696         int event_id = perf_event_id(&counter->attr);
3697         int ret;
3698
3699         ret = ftrace_profile_enable(event_id);
3700         if (ret)
3701                 return NULL;
3702
3703         counter->destroy = tp_perf_counter_destroy;
3704
3705         return &perf_ops_generic;
3706 }
3707 #else
3708 static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3709 {
3710         return NULL;
3711 }
3712 #endif
3713
3714 atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3715
3716 static void sw_perf_counter_destroy(struct perf_counter *counter)
3717 {
3718         u64 event = counter->attr.config;
3719
3720         WARN_ON(counter->parent);
3721
3722         atomic_dec(&perf_swcounter_enabled[event]);
3723 }
3724
3725 static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3726 {
3727         const struct pmu *pmu = NULL;
3728         u64 event = counter->attr.config;
3729
3730         /*
3731          * Software counters (currently) can't in general distinguish
3732          * between user, kernel and hypervisor events.
3733          * However, context switches and cpu migrations are considered
3734          * to be kernel events, and page faults are never hypervisor
3735          * events.
3736          */
3737         switch (event) {
3738         case PERF_COUNT_SW_CPU_CLOCK:
3739                 pmu = &perf_ops_cpu_clock;
3740
3741                 break;
3742         case PERF_COUNT_SW_TASK_CLOCK:
3743                 /*
3744                  * If the user instantiates this as a per-cpu counter,
3745                  * use the cpu_clock counter instead.
3746                  */
3747                 if (counter->ctx->task)
3748                         pmu = &perf_ops_task_clock;
3749                 else
3750                         pmu = &perf_ops_cpu_clock;
3751
3752                 break;
3753         case PERF_COUNT_SW_PAGE_FAULTS:
3754         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
3755         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
3756         case PERF_COUNT_SW_CONTEXT_SWITCHES:
3757         case PERF_COUNT_SW_CPU_MIGRATIONS:
3758                 if (!counter->parent) {
3759                         atomic_inc(&perf_swcounter_enabled[event]);
3760                         counter->destroy = sw_perf_counter_destroy;
3761                 }
3762                 pmu = &perf_ops_generic;
3763                 break;
3764         }
3765
3766         return pmu;
3767 }
3768
3769 /*
3770  * Allocate and initialize a counter structure
3771  */
3772 static struct perf_counter *
3773 perf_counter_alloc(struct perf_counter_attr *attr,
3774                    int cpu,
3775                    struct perf_counter_context *ctx,
3776                    struct perf_counter *group_leader,
3777                    struct perf_counter *parent_counter,
3778                    gfp_t gfpflags)
3779 {
3780         const struct pmu *pmu;
3781         struct perf_counter *counter;
3782         struct hw_perf_counter *hwc;
3783         long err;
3784
3785         counter = kzalloc(sizeof(*counter), gfpflags);
3786         if (!counter)
3787                 return ERR_PTR(-ENOMEM);
3788
3789         /*
3790          * Single counters are their own group leaders, with an
3791          * empty sibling list:
3792          */
3793         if (!group_leader)
3794                 group_leader = counter;
3795
3796         mutex_init(&counter->child_mutex);
3797         INIT_LIST_HEAD(&counter->child_list);
3798
3799         INIT_LIST_HEAD(&counter->list_entry);
3800         INIT_LIST_HEAD(&counter->event_entry);
3801         INIT_LIST_HEAD(&counter->sibling_list);
3802         init_waitqueue_head(&counter->waitq);
3803
3804         mutex_init(&counter->mmap_mutex);
3805
3806         counter->cpu            = cpu;
3807         counter->attr           = *attr;
3808         counter->group_leader   = group_leader;
3809         counter->pmu            = NULL;
3810         counter->ctx            = ctx;
3811         counter->oncpu          = -1;
3812
3813         counter->parent         = parent_counter;
3814
3815         counter->ns             = get_pid_ns(current->nsproxy->pid_ns);
3816         counter->id             = atomic64_inc_return(&perf_counter_id);
3817
3818         counter->state          = PERF_COUNTER_STATE_INACTIVE;
3819
3820         if (attr->disabled)
3821                 counter->state = PERF_COUNTER_STATE_OFF;
3822
3823         pmu = NULL;
3824
3825         hwc = &counter->hw;
3826         hwc->sample_period = attr->sample_period;
3827         if (attr->freq && attr->sample_freq)
3828                 hwc->sample_period = 1;
3829
3830         atomic64_set(&hwc->period_left, hwc->sample_period);
3831
3832         /*
3833          * we currently do not support PERF_SAMPLE_GROUP on inherited counters
3834          */
3835         if (attr->inherit && (attr->sample_type & PERF_SAMPLE_GROUP))
3836                 goto done;
3837
3838         switch (attr->type) {
3839         case PERF_TYPE_RAW:
3840         case PERF_TYPE_HARDWARE:
3841         case PERF_TYPE_HW_CACHE:
3842                 pmu = hw_perf_counter_init(counter);
3843                 break;
3844
3845         case PERF_TYPE_SOFTWARE:
3846                 pmu = sw_perf_counter_init(counter);
3847                 break;
3848
3849         case PERF_TYPE_TRACEPOINT:
3850                 pmu = tp_perf_counter_init(counter);
3851                 break;
3852
3853         default:
3854                 break;
3855         }
3856 done:
3857         err = 0;
3858         if (!pmu)
3859                 err = -EINVAL;
3860         else if (IS_ERR(pmu))
3861                 err = PTR_ERR(pmu);
3862
3863         if (err) {
3864                 if (counter->ns)
3865                         put_pid_ns(counter->ns);
3866                 kfree(counter);
3867                 return ERR_PTR(err);
3868         }
3869
3870         counter->pmu = pmu;
3871
3872         if (!counter->parent) {
3873                 atomic_inc(&nr_counters);
3874                 if (counter->attr.mmap)
3875                         atomic_inc(&nr_mmap_counters);
3876                 if (counter->attr.comm)
3877                         atomic_inc(&nr_comm_counters);
3878         }
3879
3880         return counter;
3881 }
3882
3883 static int perf_copy_attr(struct perf_counter_attr __user *uattr,
3884                           struct perf_counter_attr *attr)
3885 {
3886         int ret;
3887         u32 size;
3888
3889         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
3890                 return -EFAULT;
3891
3892         /*
3893          * zero the full structure, so that a short copy will be nice.
3894          */
3895         memset(attr, 0, sizeof(*attr));
3896
3897         ret = get_user(size, &uattr->size);
3898         if (ret)
3899                 return ret;
3900
3901         if (size > PAGE_SIZE)   /* silly large */
3902                 goto err_size;
3903
3904         if (!size)              /* abi compat */
3905                 size = PERF_ATTR_SIZE_VER0;
3906
3907         if (size < PERF_ATTR_SIZE_VER0)
3908                 goto err_size;
3909
3910         /*
3911          * If we're handed a bigger struct than we know of,
3912          * ensure all the unknown bits are 0.
3913          */
3914         if (size > sizeof(*attr)) {
3915                 unsigned long val;
3916                 unsigned long __user *addr;
3917                 unsigned long __user *end;
3918
3919                 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
3920                                 sizeof(unsigned long));
3921                 end  = PTR_ALIGN((void __user *)uattr + size,
3922                                 sizeof(unsigned long));
3923
3924                 for (; addr < end; addr += sizeof(unsigned long)) {
3925                         ret = get_user(val, addr);
3926                         if (ret)
3927                                 return ret;
3928                         if (val)
3929                                 goto err_size;
3930                 }
3931         }
3932
3933         ret = copy_from_user(attr, uattr, size);
3934         if (ret)
3935                 return -EFAULT;
3936
3937         /*
3938          * If the type exists, the corresponding creation will verify
3939          * the attr->config.
3940          */
3941         if (attr->type >= PERF_TYPE_MAX)
3942                 return -EINVAL;
3943
3944         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
3945                 return -EINVAL;
3946
3947         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
3948                 return -EINVAL;
3949
3950         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
3951                 return -EINVAL;
3952
3953 out:
3954         return ret;
3955
3956 err_size:
3957         put_user(sizeof(*attr), &uattr->size);
3958         ret = -E2BIG;
3959         goto out;
3960 }
3961
3962 /**
3963  * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
3964  *
3965  * @attr_uptr:  event type attributes for monitoring/sampling
3966  * @pid:                target pid
3967  * @cpu:                target cpu
3968  * @group_fd:           group leader counter fd
3969  */
3970 SYSCALL_DEFINE5(perf_counter_open,
3971                 struct perf_counter_attr __user *, attr_uptr,
3972                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
3973 {
3974         struct perf_counter *counter, *group_leader;
3975         struct perf_counter_attr attr;
3976         struct perf_counter_context *ctx;
3977         struct file *counter_file = NULL;
3978         struct file *group_file = NULL;
3979         int fput_needed = 0;
3980         int fput_needed2 = 0;
3981         int ret;
3982
3983         /* for future expandability... */
3984         if (flags)
3985                 return -EINVAL;
3986
3987         ret = perf_copy_attr(attr_uptr, &attr);
3988         if (ret)
3989                 return ret;
3990
3991         if (!attr.exclude_kernel) {
3992                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
3993                         return -EACCES;
3994         }
3995
3996         if (attr.freq) {
3997                 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
3998                         return -EINVAL;
3999         }
4000
4001         /*
4002          * Get the target context (task or percpu):
4003          */
4004         ctx = find_get_context(pid, cpu);
4005         if (IS_ERR(ctx))
4006                 return PTR_ERR(ctx);
4007
4008         /*
4009          * Look up the group leader (we will attach this counter to it):
4010          */
4011         group_leader = NULL;
4012         if (group_fd != -1) {
4013                 ret = -EINVAL;
4014                 group_file = fget_light(group_fd, &fput_needed);
4015                 if (!group_file)
4016                         goto err_put_context;
4017                 if (group_file->f_op != &perf_fops)
4018                         goto err_put_context;
4019
4020                 group_leader = group_file->private_data;
4021                 /*
4022                  * Do not allow a recursive hierarchy (this new sibling
4023                  * becoming part of another group-sibling):
4024                  */
4025                 if (group_leader->group_leader != group_leader)
4026                         goto err_put_context;
4027                 /*
4028                  * Do not allow to attach to a group in a different
4029                  * task or CPU context:
4030                  */
4031                 if (group_leader->ctx != ctx)
4032                         goto err_put_context;
4033                 /*
4034                  * Only a group leader can be exclusive or pinned
4035                  */
4036                 if (attr.exclusive || attr.pinned)
4037                         goto err_put_context;
4038         }
4039
4040         counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
4041                                      NULL, GFP_KERNEL);
4042         ret = PTR_ERR(counter);
4043         if (IS_ERR(counter))
4044                 goto err_put_context;
4045
4046         ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
4047         if (ret < 0)
4048                 goto err_free_put_context;
4049
4050         counter_file = fget_light(ret, &fput_needed2);
4051         if (!counter_file)
4052                 goto err_free_put_context;
4053
4054         counter->filp = counter_file;
4055         WARN_ON_ONCE(ctx->parent_ctx);
4056         mutex_lock(&ctx->mutex);
4057         perf_install_in_context(ctx, counter, cpu);
4058         ++ctx->generation;
4059         mutex_unlock(&ctx->mutex);
4060
4061         counter->owner = current;
4062         get_task_struct(current);
4063         mutex_lock(&current->perf_counter_mutex);
4064         list_add_tail(&counter->owner_entry, &current->perf_counter_list);
4065         mutex_unlock(&current->perf_counter_mutex);
4066
4067         fput_light(counter_file, fput_needed2);
4068
4069 out_fput:
4070         fput_light(group_file, fput_needed);
4071
4072         return ret;
4073
4074 err_free_put_context:
4075         kfree(counter);
4076
4077 err_put_context:
4078         put_ctx(ctx);
4079
4080         goto out_fput;
4081 }
4082
4083 /*
4084  * inherit a counter from parent task to child task:
4085  */
4086 static struct perf_counter *
4087 inherit_counter(struct perf_counter *parent_counter,
4088               struct task_struct *parent,
4089               struct perf_counter_context *parent_ctx,
4090               struct task_struct *child,
4091               struct perf_counter *group_leader,
4092               struct perf_counter_context *child_ctx)
4093 {
4094         struct perf_counter *child_counter;
4095
4096         /*
4097          * Instead of creating recursive hierarchies of counters,
4098          * we link inherited counters back to the original parent,
4099          * which has a filp for sure, which we use as the reference
4100          * count:
4101          */
4102         if (parent_counter->parent)
4103                 parent_counter = parent_counter->parent;
4104
4105         child_counter = perf_counter_alloc(&parent_counter->attr,
4106                                            parent_counter->cpu, child_ctx,
4107                                            group_leader, parent_counter,
4108                                            GFP_KERNEL);
4109         if (IS_ERR(child_counter))
4110                 return child_counter;
4111         get_ctx(child_ctx);
4112
4113         /*
4114          * Make the child state follow the state of the parent counter,
4115          * not its attr.disabled bit.  We hold the parent's mutex,
4116          * so we won't race with perf_counter_{en, dis}able_family.
4117          */
4118         if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4119                 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4120         else
4121                 child_counter->state = PERF_COUNTER_STATE_OFF;
4122
4123         if (parent_counter->attr.freq)
4124                 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4125
4126         /*
4127          * Link it up in the child's context:
4128          */
4129         add_counter_to_ctx(child_counter, child_ctx);
4130
4131         /*
4132          * Get a reference to the parent filp - we will fput it
4133          * when the child counter exits. This is safe to do because
4134          * we are in the parent and we know that the filp still
4135          * exists and has a nonzero count:
4136          */
4137         atomic_long_inc(&parent_counter->filp->f_count);
4138
4139         /*
4140          * Link this into the parent counter's child list
4141          */
4142         WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4143         mutex_lock(&parent_counter->child_mutex);
4144         list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4145         mutex_unlock(&parent_counter->child_mutex);
4146
4147         return child_counter;
4148 }
4149
4150 static int inherit_group(struct perf_counter *parent_counter,
4151               struct task_struct *parent,
4152               struct perf_counter_context *parent_ctx,
4153               struct task_struct *child,
4154               struct perf_counter_context *child_ctx)
4155 {
4156         struct perf_counter *leader;
4157         struct perf_counter *sub;
4158         struct perf_counter *child_ctr;
4159
4160         leader = inherit_counter(parent_counter, parent, parent_ctx,
4161                                  child, NULL, child_ctx);
4162         if (IS_ERR(leader))
4163                 return PTR_ERR(leader);
4164         list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4165                 child_ctr = inherit_counter(sub, parent, parent_ctx,
4166                                             child, leader, child_ctx);
4167                 if (IS_ERR(child_ctr))
4168                         return PTR_ERR(child_ctr);
4169         }
4170         return 0;
4171 }
4172
4173 static void sync_child_counter(struct perf_counter *child_counter,
4174                                struct task_struct *child)
4175 {
4176         struct perf_counter *parent_counter = child_counter->parent;
4177         u64 child_val;
4178
4179         if (child_counter->attr.inherit_stat)
4180                 perf_counter_read_event(child_counter, child);
4181
4182         child_val = atomic64_read(&child_counter->count);
4183
4184         /*
4185          * Add back the child's count to the parent's count:
4186          */
4187         atomic64_add(child_val, &parent_counter->count);
4188         atomic64_add(child_counter->total_time_enabled,
4189                      &parent_counter->child_total_time_enabled);
4190         atomic64_add(child_counter->total_time_running,
4191                      &parent_counter->child_total_time_running);
4192
4193         /*
4194          * Remove this counter from the parent's list
4195          */
4196         WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4197         mutex_lock(&parent_counter->child_mutex);
4198         list_del_init(&child_counter->child_list);
4199         mutex_unlock(&parent_counter->child_mutex);
4200
4201         /*
4202          * Release the parent counter, if this was the last
4203          * reference to it.
4204          */
4205         fput(parent_counter->filp);
4206 }
4207
4208 static void
4209 __perf_counter_exit_task(struct perf_counter *child_counter,
4210                          struct perf_counter_context *child_ctx,
4211                          struct task_struct *child)
4212 {
4213         struct perf_counter *parent_counter;
4214
4215         update_counter_times(child_counter);
4216         perf_counter_remove_from_context(child_counter);
4217
4218         parent_counter = child_counter->parent;
4219         /*
4220          * It can happen that parent exits first, and has counters
4221          * that are still around due to the child reference. These
4222          * counters need to be zapped - but otherwise linger.
4223          */
4224         if (parent_counter) {
4225                 sync_child_counter(child_counter, child);
4226                 free_counter(child_counter);
4227         }
4228 }
4229
4230 /*
4231  * When a child task exits, feed back counter values to parent counters.
4232  */
4233 void perf_counter_exit_task(struct task_struct *child)
4234 {
4235         struct perf_counter *child_counter, *tmp;
4236         struct perf_counter_context *child_ctx;
4237         unsigned long flags;
4238
4239         if (likely(!child->perf_counter_ctxp))
4240                 return;
4241
4242         local_irq_save(flags);
4243         /*
4244          * We can't reschedule here because interrupts are disabled,
4245          * and either child is current or it is a task that can't be
4246          * scheduled, so we are now safe from rescheduling changing
4247          * our context.
4248          */
4249         child_ctx = child->perf_counter_ctxp;
4250         __perf_counter_task_sched_out(child_ctx);
4251
4252         /*
4253          * Take the context lock here so that if find_get_context is
4254          * reading child->perf_counter_ctxp, we wait until it has
4255          * incremented the context's refcount before we do put_ctx below.
4256          */
4257         spin_lock(&child_ctx->lock);
4258         child->perf_counter_ctxp = NULL;
4259         if (child_ctx->parent_ctx) {
4260                 /*
4261                  * This context is a clone; unclone it so it can't get
4262                  * swapped to another process while we're removing all
4263                  * the counters from it.
4264                  */
4265                 put_ctx(child_ctx->parent_ctx);
4266                 child_ctx->parent_ctx = NULL;
4267         }
4268         spin_unlock(&child_ctx->lock);
4269         local_irq_restore(flags);
4270
4271         /*
4272          * We can recurse on the same lock type through:
4273          *
4274          *   __perf_counter_exit_task()
4275          *     sync_child_counter()
4276          *       fput(parent_counter->filp)
4277          *         perf_release()
4278          *           mutex_lock(&ctx->mutex)
4279          *
4280          * But since its the parent context it won't be the same instance.
4281          */
4282         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4283
4284 again:
4285         list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4286                                  list_entry)
4287                 __perf_counter_exit_task(child_counter, child_ctx, child);
4288
4289         /*
4290          * If the last counter was a group counter, it will have appended all
4291          * its siblings to the list, but we obtained 'tmp' before that which
4292          * will still point to the list head terminating the iteration.
4293          */
4294         if (!list_empty(&child_ctx->counter_list))
4295                 goto again;
4296
4297         mutex_unlock(&child_ctx->mutex);
4298
4299         put_ctx(child_ctx);
4300 }
4301
4302 /*
4303  * free an unexposed, unused context as created by inheritance by
4304  * init_task below, used by fork() in case of fail.
4305  */
4306 void perf_counter_free_task(struct task_struct *task)
4307 {
4308         struct perf_counter_context *ctx = task->perf_counter_ctxp;
4309         struct perf_counter *counter, *tmp;
4310
4311         if (!ctx)
4312                 return;
4313
4314         mutex_lock(&ctx->mutex);
4315 again:
4316         list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4317                 struct perf_counter *parent = counter->parent;
4318
4319                 if (WARN_ON_ONCE(!parent))
4320                         continue;
4321
4322                 mutex_lock(&parent->child_mutex);
4323                 list_del_init(&counter->child_list);
4324                 mutex_unlock(&parent->child_mutex);
4325
4326                 fput(parent->filp);
4327
4328                 list_del_counter(counter, ctx);
4329                 free_counter(counter);
4330         }
4331
4332         if (!list_empty(&ctx->counter_list))
4333                 goto again;
4334
4335         mutex_unlock(&ctx->mutex);
4336
4337         put_ctx(ctx);
4338 }
4339
4340 /*
4341  * Initialize the perf_counter context in task_struct
4342  */
4343 int perf_counter_init_task(struct task_struct *child)
4344 {
4345         struct perf_counter_context *child_ctx, *parent_ctx;
4346         struct perf_counter_context *cloned_ctx;
4347         struct perf_counter *counter;
4348         struct task_struct *parent = current;
4349         int inherited_all = 1;
4350         int ret = 0;
4351
4352         child->perf_counter_ctxp = NULL;
4353
4354         mutex_init(&child->perf_counter_mutex);
4355         INIT_LIST_HEAD(&child->perf_counter_list);
4356
4357         if (likely(!parent->perf_counter_ctxp))
4358                 return 0;
4359
4360         /*
4361          * This is executed from the parent task context, so inherit
4362          * counters that have been marked for cloning.
4363          * First allocate and initialize a context for the child.
4364          */
4365
4366         child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4367         if (!child_ctx)
4368                 return -ENOMEM;
4369
4370         __perf_counter_init_context(child_ctx, child);
4371         child->perf_counter_ctxp = child_ctx;
4372         get_task_struct(child);
4373
4374         /*
4375          * If the parent's context is a clone, pin it so it won't get
4376          * swapped under us.
4377          */
4378         parent_ctx = perf_pin_task_context(parent);
4379
4380         /*
4381          * No need to check if parent_ctx != NULL here; since we saw
4382          * it non-NULL earlier, the only reason for it to become NULL
4383          * is if we exit, and since we're currently in the middle of
4384          * a fork we can't be exiting at the same time.
4385          */
4386
4387         /*
4388          * Lock the parent list. No need to lock the child - not PID
4389          * hashed yet and not running, so nobody can access it.
4390          */
4391         mutex_lock(&parent_ctx->mutex);
4392
4393         /*
4394          * We dont have to disable NMIs - we are only looking at
4395          * the list, not manipulating it:
4396          */
4397         list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4398                 if (counter != counter->group_leader)
4399                         continue;
4400
4401                 if (!counter->attr.inherit) {
4402                         inherited_all = 0;
4403                         continue;
4404                 }
4405
4406                 ret = inherit_group(counter, parent, parent_ctx,
4407                                              child, child_ctx);
4408                 if (ret) {
4409                         inherited_all = 0;
4410                         break;
4411                 }
4412         }
4413
4414         if (inherited_all) {
4415                 /*
4416                  * Mark the child context as a clone of the parent
4417                  * context, or of whatever the parent is a clone of.
4418                  * Note that if the parent is a clone, it could get
4419                  * uncloned at any point, but that doesn't matter
4420                  * because the list of counters and the generation
4421                  * count can't have changed since we took the mutex.
4422                  */
4423                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4424                 if (cloned_ctx) {
4425                         child_ctx->parent_ctx = cloned_ctx;
4426                         child_ctx->parent_gen = parent_ctx->parent_gen;
4427                 } else {
4428                         child_ctx->parent_ctx = parent_ctx;
4429                         child_ctx->parent_gen = parent_ctx->generation;
4430                 }
4431                 get_ctx(child_ctx->parent_ctx);
4432         }
4433
4434         mutex_unlock(&parent_ctx->mutex);
4435
4436         perf_unpin_context(parent_ctx);
4437
4438         return ret;
4439 }
4440
4441 static void __cpuinit perf_counter_init_cpu(int cpu)
4442 {
4443         struct perf_cpu_context *cpuctx;
4444
4445         cpuctx = &per_cpu(perf_cpu_context, cpu);
4446         __perf_counter_init_context(&cpuctx->ctx, NULL);
4447
4448         spin_lock(&perf_resource_lock);
4449         cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4450         spin_unlock(&perf_resource_lock);
4451
4452         hw_perf_counter_setup(cpu);
4453 }
4454
4455 #ifdef CONFIG_HOTPLUG_CPU
4456 static void __perf_counter_exit_cpu(void *info)
4457 {
4458         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4459         struct perf_counter_context *ctx = &cpuctx->ctx;
4460         struct perf_counter *counter, *tmp;
4461
4462         list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4463                 __perf_counter_remove_from_context(counter);
4464 }
4465 static void perf_counter_exit_cpu(int cpu)
4466 {
4467         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4468         struct perf_counter_context *ctx = &cpuctx->ctx;
4469
4470         mutex_lock(&ctx->mutex);
4471         smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4472         mutex_unlock(&ctx->mutex);
4473 }
4474 #else
4475 static inline void perf_counter_exit_cpu(int cpu) { }
4476 #endif
4477
4478 static int __cpuinit
4479 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4480 {
4481         unsigned int cpu = (long)hcpu;
4482
4483         switch (action) {
4484
4485         case CPU_UP_PREPARE:
4486         case CPU_UP_PREPARE_FROZEN:
4487                 perf_counter_init_cpu(cpu);
4488                 break;
4489
4490         case CPU_DOWN_PREPARE:
4491         case CPU_DOWN_PREPARE_FROZEN:
4492                 perf_counter_exit_cpu(cpu);
4493                 break;
4494
4495         default:
4496                 break;
4497         }
4498
4499         return NOTIFY_OK;
4500 }
4501
4502 /*
4503  * This has to have a higher priority than migration_notifier in sched.c.
4504  */
4505 static struct notifier_block __cpuinitdata perf_cpu_nb = {
4506         .notifier_call          = perf_cpu_notify,
4507         .priority               = 20,
4508 };
4509
4510 void __init perf_counter_init(void)
4511 {
4512         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4513                         (void *)(long)smp_processor_id());
4514         register_cpu_notifier(&perf_cpu_nb);
4515 }
4516
4517 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4518 {
4519         return sprintf(buf, "%d\n", perf_reserved_percpu);
4520 }
4521
4522 static ssize_t
4523 perf_set_reserve_percpu(struct sysdev_class *class,
4524                         const char *buf,
4525                         size_t count)
4526 {
4527         struct perf_cpu_context *cpuctx;
4528         unsigned long val;
4529         int err, cpu, mpt;
4530
4531         err = strict_strtoul(buf, 10, &val);
4532         if (err)
4533                 return err;
4534         if (val > perf_max_counters)
4535                 return -EINVAL;
4536
4537         spin_lock(&perf_resource_lock);
4538         perf_reserved_percpu = val;
4539         for_each_online_cpu(cpu) {
4540                 cpuctx = &per_cpu(perf_cpu_context, cpu);
4541                 spin_lock_irq(&cpuctx->ctx.lock);
4542                 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4543                           perf_max_counters - perf_reserved_percpu);
4544                 cpuctx->max_pertask = mpt;
4545                 spin_unlock_irq(&cpuctx->ctx.lock);
4546         }
4547         spin_unlock(&perf_resource_lock);
4548
4549         return count;
4550 }
4551
4552 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4553 {
4554         return sprintf(buf, "%d\n", perf_overcommit);
4555 }
4556
4557 static ssize_t
4558 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4559 {
4560         unsigned long val;
4561         int err;
4562
4563         err = strict_strtoul(buf, 10, &val);
4564         if (err)
4565                 return err;
4566         if (val > 1)
4567                 return -EINVAL;
4568
4569         spin_lock(&perf_resource_lock);
4570         perf_overcommit = val;
4571         spin_unlock(&perf_resource_lock);
4572
4573         return count;
4574 }
4575
4576 static SYSDEV_CLASS_ATTR(
4577                                 reserve_percpu,
4578                                 0644,
4579                                 perf_show_reserve_percpu,
4580                                 perf_set_reserve_percpu
4581                         );
4582
4583 static SYSDEV_CLASS_ATTR(
4584                                 overcommit,
4585                                 0644,
4586                                 perf_show_overcommit,
4587                                 perf_set_overcommit
4588                         );
4589
4590 static struct attribute *perfclass_attrs[] = {
4591         &attr_reserve_percpu.attr,
4592         &attr_overcommit.attr,
4593         NULL
4594 };
4595
4596 static struct attribute_group perfclass_attr_group = {
4597         .attrs                  = perfclass_attrs,
4598         .name                   = "perf_counters",
4599 };
4600
4601 static int __init perf_counter_sysfs_init(void)
4602 {
4603         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4604                                   &perfclass_attr_group);
4605 }
4606 device_initcall(perf_counter_sysfs_init);