aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/perf_counter.c
diff options
context:
space:
mode:
Diffstat (limited to 'kernel/perf_counter.c')
-rw-r--r--kernel/perf_counter.c4861
1 files changed, 0 insertions, 4861 deletions
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
deleted file mode 100644
index d7cbc579fc80..000000000000
--- a/kernel/perf_counter.c
+++ /dev/null
@@ -1,4861 +0,0 @@
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 */
36DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
37
38int perf_max_counters __read_mostly = 1;
39static int perf_reserved_percpu __read_mostly;
40static int perf_overcommit __read_mostly = 1;
41
42static atomic_t nr_counters __read_mostly;
43static atomic_t nr_mmap_counters __read_mostly;
44static atomic_t nr_comm_counters __read_mostly;
45static atomic_t nr_task_counters __read_mostly;
46
47/*
48 * perf counter paranoia level:
49 * 0 - not paranoid
50 * 1 - disallow cpu counters to unpriv
51 * 2 - disallow kernel profiling to unpriv
52 */
53int sysctl_perf_counter_paranoid __read_mostly = 1;
54
55static inline bool perf_paranoid_cpu(void)
56{
57 return sysctl_perf_counter_paranoid > 0;
58}
59
60static inline bool perf_paranoid_kernel(void)
61{
62 return sysctl_perf_counter_paranoid > 1;
63}
64
65int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
66
67/*
68 * max perf counter sample rate
69 */
70int sysctl_perf_counter_sample_rate __read_mostly = 100000;
71
72static atomic64_t perf_counter_id;
73
74/*
75 * Lock for (sysadmin-configurable) counter reservations:
76 */
77static DEFINE_SPINLOCK(perf_resource_lock);
78
79/*
80 * Architecture provided APIs - weak aliases:
81 */
82extern __weak const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
83{
84 return NULL;
85}
86
87void __weak hw_perf_disable(void) { barrier(); }
88void __weak hw_perf_enable(void) { barrier(); }
89
90void __weak hw_perf_counter_setup(int cpu) { barrier(); }
91void __weak hw_perf_counter_setup_online(int cpu) { barrier(); }
92
93int __weak
94hw_perf_group_sched_in(struct perf_counter *group_leader,
95 struct perf_cpu_context *cpuctx,
96 struct perf_counter_context *ctx, int cpu)
97{
98 return 0;
99}
100
101void __weak perf_counter_print_debug(void) { }
102
103static DEFINE_PER_CPU(int, disable_count);
104
105void __perf_disable(void)
106{
107 __get_cpu_var(disable_count)++;
108}
109
110bool __perf_enable(void)
111{
112 return !--__get_cpu_var(disable_count);
113}
114
115void perf_disable(void)
116{
117 __perf_disable();
118 hw_perf_disable();
119}
120
121void perf_enable(void)
122{
123 if (__perf_enable())
124 hw_perf_enable();
125}
126
127static void get_ctx(struct perf_counter_context *ctx)
128{
129 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
130}
131
132static void free_ctx(struct rcu_head *head)
133{
134 struct perf_counter_context *ctx;
135
136 ctx = container_of(head, struct perf_counter_context, rcu_head);
137 kfree(ctx);
138}
139
140static void put_ctx(struct perf_counter_context *ctx)
141{
142 if (atomic_dec_and_test(&ctx->refcount)) {
143 if (ctx->parent_ctx)
144 put_ctx(ctx->parent_ctx);
145 if (ctx->task)
146 put_task_struct(ctx->task);
147 call_rcu(&ctx->rcu_head, free_ctx);
148 }
149}
150
151static void unclone_ctx(struct perf_counter_context *ctx)
152{
153 if (ctx->parent_ctx) {
154 put_ctx(ctx->parent_ctx);
155 ctx->parent_ctx = NULL;
156 }
157}
158
159/*
160 * If we inherit counters we want to return the parent counter id
161 * to userspace.
162 */
163static u64 primary_counter_id(struct perf_counter *counter)
164{
165 u64 id = counter->id;
166
167 if (counter->parent)
168 id = counter->parent->id;
169
170 return id;
171}
172
173/*
174 * Get the perf_counter_context for a task and lock it.
175 * This has to cope with with the fact that until it is locked,
176 * the context could get moved to another task.
177 */
178static struct perf_counter_context *
179perf_lock_task_context(struct task_struct *task, unsigned long *flags)
180{
181 struct perf_counter_context *ctx;
182
183 rcu_read_lock();
184 retry:
185 ctx = rcu_dereference(task->perf_counter_ctxp);
186 if (ctx) {
187 /*
188 * If this context is a clone of another, it might
189 * get swapped for another underneath us by
190 * perf_counter_task_sched_out, though the
191 * rcu_read_lock() protects us from any context
192 * getting freed. Lock the context and check if it
193 * got swapped before we could get the lock, and retry
194 * if so. If we locked the right context, then it
195 * can't get swapped on us any more.
196 */
197 spin_lock_irqsave(&ctx->lock, *flags);
198 if (ctx != rcu_dereference(task->perf_counter_ctxp)) {
199 spin_unlock_irqrestore(&ctx->lock, *flags);
200 goto retry;
201 }
202
203 if (!atomic_inc_not_zero(&ctx->refcount)) {
204 spin_unlock_irqrestore(&ctx->lock, *flags);
205 ctx = NULL;
206 }
207 }
208 rcu_read_unlock();
209 return ctx;
210}
211
212/*
213 * Get the context for a task and increment its pin_count so it
214 * can't get swapped to another task. This also increments its
215 * reference count so that the context can't get freed.
216 */
217static struct perf_counter_context *perf_pin_task_context(struct task_struct *task)
218{
219 struct perf_counter_context *ctx;
220 unsigned long flags;
221
222 ctx = perf_lock_task_context(task, &flags);
223 if (ctx) {
224 ++ctx->pin_count;
225 spin_unlock_irqrestore(&ctx->lock, flags);
226 }
227 return ctx;
228}
229
230static void perf_unpin_context(struct perf_counter_context *ctx)
231{
232 unsigned long flags;
233
234 spin_lock_irqsave(&ctx->lock, flags);
235 --ctx->pin_count;
236 spin_unlock_irqrestore(&ctx->lock, flags);
237 put_ctx(ctx);
238}
239
240/*
241 * Add a counter from the lists for its context.
242 * Must be called with ctx->mutex and ctx->lock held.
243 */
244static void
245list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
246{
247 struct perf_counter *group_leader = counter->group_leader;
248
249 /*
250 * Depending on whether it is a standalone or sibling counter,
251 * add it straight to the context's counter list, or to the group
252 * leader's sibling list:
253 */
254 if (group_leader == counter)
255 list_add_tail(&counter->list_entry, &ctx->counter_list);
256 else {
257 list_add_tail(&counter->list_entry, &group_leader->sibling_list);
258 group_leader->nr_siblings++;
259 }
260
261 list_add_rcu(&counter->event_entry, &ctx->event_list);
262 ctx->nr_counters++;
263 if (counter->attr.inherit_stat)
264 ctx->nr_stat++;
265}
266
267/*
268 * Remove a counter from the lists for its context.
269 * Must be called with ctx->mutex and ctx->lock held.
270 */
271static void
272list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
273{
274 struct perf_counter *sibling, *tmp;
275
276 if (list_empty(&counter->list_entry))
277 return;
278 ctx->nr_counters--;
279 if (counter->attr.inherit_stat)
280 ctx->nr_stat--;
281
282 list_del_init(&counter->list_entry);
283 list_del_rcu(&counter->event_entry);
284
285 if (counter->group_leader != counter)
286 counter->group_leader->nr_siblings--;
287
288 /*
289 * If this was a group counter with sibling counters then
290 * upgrade the siblings to singleton counters by adding them
291 * to the context list directly:
292 */
293 list_for_each_entry_safe(sibling, tmp,
294 &counter->sibling_list, list_entry) {
295
296 list_move_tail(&sibling->list_entry, &ctx->counter_list);
297 sibling->group_leader = sibling;
298 }
299}
300
301static void
302counter_sched_out(struct perf_counter *counter,
303 struct perf_cpu_context *cpuctx,
304 struct perf_counter_context *ctx)
305{
306 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
307 return;
308
309 counter->state = PERF_COUNTER_STATE_INACTIVE;
310 if (counter->pending_disable) {
311 counter->pending_disable = 0;
312 counter->state = PERF_COUNTER_STATE_OFF;
313 }
314 counter->tstamp_stopped = ctx->time;
315 counter->pmu->disable(counter);
316 counter->oncpu = -1;
317
318 if (!is_software_counter(counter))
319 cpuctx->active_oncpu--;
320 ctx->nr_active--;
321 if (counter->attr.exclusive || !cpuctx->active_oncpu)
322 cpuctx->exclusive = 0;
323}
324
325static void
326group_sched_out(struct perf_counter *group_counter,
327 struct perf_cpu_context *cpuctx,
328 struct perf_counter_context *ctx)
329{
330 struct perf_counter *counter;
331
332 if (group_counter->state != PERF_COUNTER_STATE_ACTIVE)
333 return;
334
335 counter_sched_out(group_counter, cpuctx, ctx);
336
337 /*
338 * Schedule out siblings (if any):
339 */
340 list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
341 counter_sched_out(counter, cpuctx, ctx);
342
343 if (group_counter->attr.exclusive)
344 cpuctx->exclusive = 0;
345}
346
347/*
348 * Cross CPU call to remove a performance counter
349 *
350 * We disable the counter on the hardware level first. After that we
351 * remove it from the context list.
352 */
353static void __perf_counter_remove_from_context(void *info)
354{
355 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
356 struct perf_counter *counter = info;
357 struct perf_counter_context *ctx = counter->ctx;
358
359 /*
360 * If this is a task context, we need to check whether it is
361 * the current task context of this cpu. If not it has been
362 * scheduled out before the smp call arrived.
363 */
364 if (ctx->task && cpuctx->task_ctx != ctx)
365 return;
366
367 spin_lock(&ctx->lock);
368 /*
369 * Protect the list operation against NMI by disabling the
370 * counters on a global level.
371 */
372 perf_disable();
373
374 counter_sched_out(counter, cpuctx, ctx);
375
376 list_del_counter(counter, ctx);
377
378 if (!ctx->task) {
379 /*
380 * Allow more per task counters with respect to the
381 * reservation:
382 */
383 cpuctx->max_pertask =
384 min(perf_max_counters - ctx->nr_counters,
385 perf_max_counters - perf_reserved_percpu);
386 }
387
388 perf_enable();
389 spin_unlock(&ctx->lock);
390}
391
392
393/*
394 * Remove the counter from a task's (or a CPU's) list of counters.
395 *
396 * Must be called with ctx->mutex held.
397 *
398 * CPU counters are removed with a smp call. For task counters we only
399 * call when the task is on a CPU.
400 *
401 * If counter->ctx is a cloned context, callers must make sure that
402 * every task struct that counter->ctx->task could possibly point to
403 * remains valid. This is OK when called from perf_release since
404 * that only calls us on the top-level context, which can't be a clone.
405 * When called from perf_counter_exit_task, it's OK because the
406 * context has been detached from its task.
407 */
408static void perf_counter_remove_from_context(struct perf_counter *counter)
409{
410 struct perf_counter_context *ctx = counter->ctx;
411 struct task_struct *task = ctx->task;
412
413 if (!task) {
414 /*
415 * Per cpu counters are removed via an smp call and
416 * the removal is always sucessful.
417 */
418 smp_call_function_single(counter->cpu,
419 __perf_counter_remove_from_context,
420 counter, 1);
421 return;
422 }
423
424retry:
425 task_oncpu_function_call(task, __perf_counter_remove_from_context,
426 counter);
427
428 spin_lock_irq(&ctx->lock);
429 /*
430 * If the context is active we need to retry the smp call.
431 */
432 if (ctx->nr_active && !list_empty(&counter->list_entry)) {
433 spin_unlock_irq(&ctx->lock);
434 goto retry;
435 }
436
437 /*
438 * The lock prevents that this context is scheduled in so we
439 * can remove the counter safely, if the call above did not
440 * succeed.
441 */
442 if (!list_empty(&counter->list_entry)) {
443 list_del_counter(counter, ctx);
444 }
445 spin_unlock_irq(&ctx->lock);
446}
447
448static inline u64 perf_clock(void)
449{
450 return cpu_clock(smp_processor_id());
451}
452
453/*
454 * Update the record of the current time in a context.
455 */
456static void update_context_time(struct perf_counter_context *ctx)
457{
458 u64 now = perf_clock();
459
460 ctx->time += now - ctx->timestamp;
461 ctx->timestamp = now;
462}
463
464/*
465 * Update the total_time_enabled and total_time_running fields for a counter.
466 */
467static void update_counter_times(struct perf_counter *counter)
468{
469 struct perf_counter_context *ctx = counter->ctx;
470 u64 run_end;
471
472 if (counter->state < PERF_COUNTER_STATE_INACTIVE)
473 return;
474
475 counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
476
477 if (counter->state == PERF_COUNTER_STATE_INACTIVE)
478 run_end = counter->tstamp_stopped;
479 else
480 run_end = ctx->time;
481
482 counter->total_time_running = run_end - counter->tstamp_running;
483}
484
485/*
486 * Update total_time_enabled and total_time_running for all counters in a group.
487 */
488static void update_group_times(struct perf_counter *leader)
489{
490 struct perf_counter *counter;
491
492 update_counter_times(leader);
493 list_for_each_entry(counter, &leader->sibling_list, list_entry)
494 update_counter_times(counter);
495}
496
497/*
498 * Cross CPU call to disable a performance counter
499 */
500static void __perf_counter_disable(void *info)
501{
502 struct perf_counter *counter = info;
503 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
504 struct perf_counter_context *ctx = counter->ctx;
505
506 /*
507 * If this is a per-task counter, need to check whether this
508 * counter's task is the current task on this cpu.
509 */
510 if (ctx->task && cpuctx->task_ctx != ctx)
511 return;
512
513 spin_lock(&ctx->lock);
514
515 /*
516 * If the counter is on, turn it off.
517 * If it is in error state, leave it in error state.
518 */
519 if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
520 update_context_time(ctx);
521 update_counter_times(counter);
522 if (counter == counter->group_leader)
523 group_sched_out(counter, cpuctx, ctx);
524 else
525 counter_sched_out(counter, cpuctx, ctx);
526 counter->state = PERF_COUNTER_STATE_OFF;
527 }
528
529 spin_unlock(&ctx->lock);
530}
531
532/*
533 * Disable a counter.
534 *
535 * If counter->ctx is a cloned context, callers must make sure that
536 * every task struct that counter->ctx->task could possibly point to
537 * remains valid. This condition is satisifed when called through
538 * perf_counter_for_each_child or perf_counter_for_each because they
539 * hold the top-level counter's child_mutex, so any descendant that
540 * goes to exit will block in sync_child_counter.
541 * When called from perf_pending_counter it's OK because counter->ctx
542 * is the current context on this CPU and preemption is disabled,
543 * hence we can't get into perf_counter_task_sched_out for this context.
544 */
545static void perf_counter_disable(struct perf_counter *counter)
546{
547 struct perf_counter_context *ctx = counter->ctx;
548 struct task_struct *task = ctx->task;
549
550 if (!task) {
551 /*
552 * Disable the counter on the cpu that it's on
553 */
554 smp_call_function_single(counter->cpu, __perf_counter_disable,
555 counter, 1);
556 return;
557 }
558
559 retry:
560 task_oncpu_function_call(task, __perf_counter_disable, counter);
561
562 spin_lock_irq(&ctx->lock);
563 /*
564 * If the counter is still active, we need to retry the cross-call.
565 */
566 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
567 spin_unlock_irq(&ctx->lock);
568 goto retry;
569 }
570
571 /*
572 * Since we have the lock this context can't be scheduled
573 * in, so we can change the state safely.
574 */
575 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
576 update_counter_times(counter);
577 counter->state = PERF_COUNTER_STATE_OFF;
578 }
579
580 spin_unlock_irq(&ctx->lock);
581}
582
583static int
584counter_sched_in(struct perf_counter *counter,
585 struct perf_cpu_context *cpuctx,
586 struct perf_counter_context *ctx,
587 int cpu)
588{
589 if (counter->state <= PERF_COUNTER_STATE_OFF)
590 return 0;
591
592 counter->state = PERF_COUNTER_STATE_ACTIVE;
593 counter->oncpu = cpu; /* TODO: put 'cpu' into cpuctx->cpu */
594 /*
595 * The new state must be visible before we turn it on in the hardware:
596 */
597 smp_wmb();
598
599 if (counter->pmu->enable(counter)) {
600 counter->state = PERF_COUNTER_STATE_INACTIVE;
601 counter->oncpu = -1;
602 return -EAGAIN;
603 }
604
605 counter->tstamp_running += ctx->time - counter->tstamp_stopped;
606
607 if (!is_software_counter(counter))
608 cpuctx->active_oncpu++;
609 ctx->nr_active++;
610
611 if (counter->attr.exclusive)
612 cpuctx->exclusive = 1;
613
614 return 0;
615}
616
617static int
618group_sched_in(struct perf_counter *group_counter,
619 struct perf_cpu_context *cpuctx,
620 struct perf_counter_context *ctx,
621 int cpu)
622{
623 struct perf_counter *counter, *partial_group;
624 int ret;
625
626 if (group_counter->state == PERF_COUNTER_STATE_OFF)
627 return 0;
628
629 ret = hw_perf_group_sched_in(group_counter, cpuctx, ctx, cpu);
630 if (ret)
631 return ret < 0 ? ret : 0;
632
633 if (counter_sched_in(group_counter, cpuctx, ctx, cpu))
634 return -EAGAIN;
635
636 /*
637 * Schedule in siblings as one group (if any):
638 */
639 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
640 if (counter_sched_in(counter, cpuctx, ctx, cpu)) {
641 partial_group = counter;
642 goto group_error;
643 }
644 }
645
646 return 0;
647
648group_error:
649 /*
650 * Groups can be scheduled in as one unit only, so undo any
651 * partial group before returning:
652 */
653 list_for_each_entry(counter, &group_counter->sibling_list, list_entry) {
654 if (counter == partial_group)
655 break;
656 counter_sched_out(counter, cpuctx, ctx);
657 }
658 counter_sched_out(group_counter, cpuctx, ctx);
659
660 return -EAGAIN;
661}
662
663/*
664 * Return 1 for a group consisting entirely of software counters,
665 * 0 if the group contains any hardware counters.
666 */
667static int is_software_only_group(struct perf_counter *leader)
668{
669 struct perf_counter *counter;
670
671 if (!is_software_counter(leader))
672 return 0;
673
674 list_for_each_entry(counter, &leader->sibling_list, list_entry)
675 if (!is_software_counter(counter))
676 return 0;
677
678 return 1;
679}
680
681/*
682 * Work out whether we can put this counter group on the CPU now.
683 */
684static int group_can_go_on(struct perf_counter *counter,
685 struct perf_cpu_context *cpuctx,
686 int can_add_hw)
687{
688 /*
689 * Groups consisting entirely of software counters can always go on.
690 */
691 if (is_software_only_group(counter))
692 return 1;
693 /*
694 * If an exclusive group is already on, no other hardware
695 * counters can go on.
696 */
697 if (cpuctx->exclusive)
698 return 0;
699 /*
700 * If this group is exclusive and there are already
701 * counters on the CPU, it can't go on.
702 */
703 if (counter->attr.exclusive && cpuctx->active_oncpu)
704 return 0;
705 /*
706 * Otherwise, try to add it if all previous groups were able
707 * to go on.
708 */
709 return can_add_hw;
710}
711
712static void add_counter_to_ctx(struct perf_counter *counter,
713 struct perf_counter_context *ctx)
714{
715 list_add_counter(counter, ctx);
716 counter->tstamp_enabled = ctx->time;
717 counter->tstamp_running = ctx->time;
718 counter->tstamp_stopped = ctx->time;
719}
720
721/*
722 * Cross CPU call to install and enable a performance counter
723 *
724 * Must be called with ctx->mutex held
725 */
726static void __perf_install_in_context(void *info)
727{
728 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
729 struct perf_counter *counter = info;
730 struct perf_counter_context *ctx = counter->ctx;
731 struct perf_counter *leader = counter->group_leader;
732 int cpu = smp_processor_id();
733 int err;
734
735 /*
736 * If this is a task context, we need to check whether it is
737 * the current task context of this cpu. If not it has been
738 * scheduled out before the smp call arrived.
739 * Or possibly this is the right context but it isn't
740 * on this cpu because it had no counters.
741 */
742 if (ctx->task && cpuctx->task_ctx != ctx) {
743 if (cpuctx->task_ctx || ctx->task != current)
744 return;
745 cpuctx->task_ctx = ctx;
746 }
747
748 spin_lock(&ctx->lock);
749 ctx->is_active = 1;
750 update_context_time(ctx);
751
752 /*
753 * Protect the list operation against NMI by disabling the
754 * counters on a global level. NOP for non NMI based counters.
755 */
756 perf_disable();
757
758 add_counter_to_ctx(counter, ctx);
759
760 /*
761 * Don't put the counter on if it is disabled or if
762 * it is in a group and the group isn't on.
763 */
764 if (counter->state != PERF_COUNTER_STATE_INACTIVE ||
765 (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE))
766 goto unlock;
767
768 /*
769 * An exclusive counter can't go on if there are already active
770 * hardware counters, and no hardware counter can go on if there
771 * is already an exclusive counter on.
772 */
773 if (!group_can_go_on(counter, cpuctx, 1))
774 err = -EEXIST;
775 else
776 err = counter_sched_in(counter, cpuctx, ctx, cpu);
777
778 if (err) {
779 /*
780 * This counter couldn't go on. If it is in a group
781 * then we have to pull the whole group off.
782 * If the counter group is pinned then put it in error state.
783 */
784 if (leader != counter)
785 group_sched_out(leader, cpuctx, ctx);
786 if (leader->attr.pinned) {
787 update_group_times(leader);
788 leader->state = PERF_COUNTER_STATE_ERROR;
789 }
790 }
791
792 if (!err && !ctx->task && cpuctx->max_pertask)
793 cpuctx->max_pertask--;
794
795 unlock:
796 perf_enable();
797
798 spin_unlock(&ctx->lock);
799}
800
801/*
802 * Attach a performance counter to a context
803 *
804 * First we add the counter to the list with the hardware enable bit
805 * in counter->hw_config cleared.
806 *
807 * If the counter is attached to a task which is on a CPU we use a smp
808 * call to enable it in the task context. The task might have been
809 * scheduled away, but we check this in the smp call again.
810 *
811 * Must be called with ctx->mutex held.
812 */
813static void
814perf_install_in_context(struct perf_counter_context *ctx,
815 struct perf_counter *counter,
816 int cpu)
817{
818 struct task_struct *task = ctx->task;
819
820 if (!task) {
821 /*
822 * Per cpu counters are installed via an smp call and
823 * the install is always sucessful.
824 */
825 smp_call_function_single(cpu, __perf_install_in_context,
826 counter, 1);
827 return;
828 }
829
830retry:
831 task_oncpu_function_call(task, __perf_install_in_context,
832 counter);
833
834 spin_lock_irq(&ctx->lock);
835 /*
836 * we need to retry the smp call.
837 */
838 if (ctx->is_active && list_empty(&counter->list_entry)) {
839 spin_unlock_irq(&ctx->lock);
840 goto retry;
841 }
842
843 /*
844 * The lock prevents that this context is scheduled in so we
845 * can add the counter safely, if it the call above did not
846 * succeed.
847 */
848 if (list_empty(&counter->list_entry))
849 add_counter_to_ctx(counter, ctx);
850 spin_unlock_irq(&ctx->lock);
851}
852
853/*
854 * Cross CPU call to enable a performance counter
855 */
856static void __perf_counter_enable(void *info)
857{
858 struct perf_counter *counter = info;
859 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
860 struct perf_counter_context *ctx = counter->ctx;
861 struct perf_counter *leader = counter->group_leader;
862 int err;
863
864 /*
865 * If this is a per-task counter, need to check whether this
866 * counter's task is the current task on this cpu.
867 */
868 if (ctx->task && cpuctx->task_ctx != ctx) {
869 if (cpuctx->task_ctx || ctx->task != current)
870 return;
871 cpuctx->task_ctx = ctx;
872 }
873
874 spin_lock(&ctx->lock);
875 ctx->is_active = 1;
876 update_context_time(ctx);
877
878 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
879 goto unlock;
880 counter->state = PERF_COUNTER_STATE_INACTIVE;
881 counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
882
883 /*
884 * If the counter is in a group and isn't the group leader,
885 * then don't put it on unless the group is on.
886 */
887 if (leader != counter && leader->state != PERF_COUNTER_STATE_ACTIVE)
888 goto unlock;
889
890 if (!group_can_go_on(counter, cpuctx, 1)) {
891 err = -EEXIST;
892 } else {
893 perf_disable();
894 if (counter == leader)
895 err = group_sched_in(counter, cpuctx, ctx,
896 smp_processor_id());
897 else
898 err = counter_sched_in(counter, cpuctx, ctx,
899 smp_processor_id());
900 perf_enable();
901 }
902
903 if (err) {
904 /*
905 * If this counter can't go on and it's part of a
906 * group, then the whole group has to come off.
907 */
908 if (leader != counter)
909 group_sched_out(leader, cpuctx, ctx);
910 if (leader->attr.pinned) {
911 update_group_times(leader);
912 leader->state = PERF_COUNTER_STATE_ERROR;
913 }
914 }
915
916 unlock:
917 spin_unlock(&ctx->lock);
918}
919
920/*
921 * Enable a counter.
922 *
923 * If counter->ctx is a cloned context, callers must make sure that
924 * every task struct that counter->ctx->task could possibly point to
925 * remains valid. This condition is satisfied when called through
926 * perf_counter_for_each_child or perf_counter_for_each as described
927 * for perf_counter_disable.
928 */
929static void perf_counter_enable(struct perf_counter *counter)
930{
931 struct perf_counter_context *ctx = counter->ctx;
932 struct task_struct *task = ctx->task;
933
934 if (!task) {
935 /*
936 * Enable the counter on the cpu that it's on
937 */
938 smp_call_function_single(counter->cpu, __perf_counter_enable,
939 counter, 1);
940 return;
941 }
942
943 spin_lock_irq(&ctx->lock);
944 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
945 goto out;
946
947 /*
948 * If the counter is in error state, clear that first.
949 * That way, if we see the counter in error state below, we
950 * know that it has gone back into error state, as distinct
951 * from the task having been scheduled away before the
952 * cross-call arrived.
953 */
954 if (counter->state == PERF_COUNTER_STATE_ERROR)
955 counter->state = PERF_COUNTER_STATE_OFF;
956
957 retry:
958 spin_unlock_irq(&ctx->lock);
959 task_oncpu_function_call(task, __perf_counter_enable, counter);
960
961 spin_lock_irq(&ctx->lock);
962
963 /*
964 * If the context is active and the counter is still off,
965 * we need to retry the cross-call.
966 */
967 if (ctx->is_active && counter->state == PERF_COUNTER_STATE_OFF)
968 goto retry;
969
970 /*
971 * Since we have the lock this context can't be scheduled
972 * in, so we can change the state safely.
973 */
974 if (counter->state == PERF_COUNTER_STATE_OFF) {
975 counter->state = PERF_COUNTER_STATE_INACTIVE;
976 counter->tstamp_enabled =
977 ctx->time - counter->total_time_enabled;
978 }
979 out:
980 spin_unlock_irq(&ctx->lock);
981}
982
983static int perf_counter_refresh(struct perf_counter *counter, int refresh)
984{
985 /*
986 * not supported on inherited counters
987 */
988 if (counter->attr.inherit)
989 return -EINVAL;
990
991 atomic_add(refresh, &counter->event_limit);
992 perf_counter_enable(counter);
993
994 return 0;
995}
996
997void __perf_counter_sched_out(struct perf_counter_context *ctx,
998 struct perf_cpu_context *cpuctx)
999{
1000 struct perf_counter *counter;
1001
1002 spin_lock(&ctx->lock);
1003 ctx->is_active = 0;
1004 if (likely(!ctx->nr_counters))
1005 goto out;
1006 update_context_time(ctx);
1007
1008 perf_disable();
1009 if (ctx->nr_active) {
1010 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1011 if (counter != counter->group_leader)
1012 counter_sched_out(counter, cpuctx, ctx);
1013 else
1014 group_sched_out(counter, cpuctx, ctx);
1015 }
1016 }
1017 perf_enable();
1018 out:
1019 spin_unlock(&ctx->lock);
1020}
1021
1022/*
1023 * Test whether two contexts are equivalent, i.e. whether they
1024 * have both been cloned from the same version of the same context
1025 * and they both have the same number of enabled counters.
1026 * If the number of enabled counters is the same, then the set
1027 * of enabled counters should be the same, because these are both
1028 * inherited contexts, therefore we can't access individual counters
1029 * in them directly with an fd; we can only enable/disable all
1030 * counters via prctl, or enable/disable all counters in a family
1031 * via ioctl, which will have the same effect on both contexts.
1032 */
1033static int context_equiv(struct perf_counter_context *ctx1,
1034 struct perf_counter_context *ctx2)
1035{
1036 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1037 && ctx1->parent_gen == ctx2->parent_gen
1038 && !ctx1->pin_count && !ctx2->pin_count;
1039}
1040
1041static void __perf_counter_read(void *counter);
1042
1043static void __perf_counter_sync_stat(struct perf_counter *counter,
1044 struct perf_counter *next_counter)
1045{
1046 u64 value;
1047
1048 if (!counter->attr.inherit_stat)
1049 return;
1050
1051 /*
1052 * Update the counter value, we cannot use perf_counter_read()
1053 * because we're in the middle of a context switch and have IRQs
1054 * disabled, which upsets smp_call_function_single(), however
1055 * we know the counter must be on the current CPU, therefore we
1056 * don't need to use it.
1057 */
1058 switch (counter->state) {
1059 case PERF_COUNTER_STATE_ACTIVE:
1060 __perf_counter_read(counter);
1061 break;
1062
1063 case PERF_COUNTER_STATE_INACTIVE:
1064 update_counter_times(counter);
1065 break;
1066
1067 default:
1068 break;
1069 }
1070
1071 /*
1072 * In order to keep per-task stats reliable we need to flip the counter
1073 * values when we flip the contexts.
1074 */
1075 value = atomic64_read(&next_counter->count);
1076 value = atomic64_xchg(&counter->count, value);
1077 atomic64_set(&next_counter->count, value);
1078
1079 swap(counter->total_time_enabled, next_counter->total_time_enabled);
1080 swap(counter->total_time_running, next_counter->total_time_running);
1081
1082 /*
1083 * Since we swizzled the values, update the user visible data too.
1084 */
1085 perf_counter_update_userpage(counter);
1086 perf_counter_update_userpage(next_counter);
1087}
1088
1089#define list_next_entry(pos, member) \
1090 list_entry(pos->member.next, typeof(*pos), member)
1091
1092static void perf_counter_sync_stat(struct perf_counter_context *ctx,
1093 struct perf_counter_context *next_ctx)
1094{
1095 struct perf_counter *counter, *next_counter;
1096
1097 if (!ctx->nr_stat)
1098 return;
1099
1100 counter = list_first_entry(&ctx->event_list,
1101 struct perf_counter, event_entry);
1102
1103 next_counter = list_first_entry(&next_ctx->event_list,
1104 struct perf_counter, event_entry);
1105
1106 while (&counter->event_entry != &ctx->event_list &&
1107 &next_counter->event_entry != &next_ctx->event_list) {
1108
1109 __perf_counter_sync_stat(counter, next_counter);
1110
1111 counter = list_next_entry(counter, event_entry);
1112 next_counter = list_next_entry(next_counter, event_entry);
1113 }
1114}
1115
1116/*
1117 * Called from scheduler to remove the counters of the current task,
1118 * with interrupts disabled.
1119 *
1120 * We stop each counter and update the counter value in counter->count.
1121 *
1122 * This does not protect us against NMI, but disable()
1123 * sets the disabled bit in the control field of counter _before_
1124 * accessing the counter control register. If a NMI hits, then it will
1125 * not restart the counter.
1126 */
1127void perf_counter_task_sched_out(struct task_struct *task,
1128 struct task_struct *next, int cpu)
1129{
1130 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1131 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1132 struct perf_counter_context *next_ctx;
1133 struct perf_counter_context *parent;
1134 struct pt_regs *regs;
1135 int do_switch = 1;
1136
1137 regs = task_pt_regs(task);
1138 perf_swcounter_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1139
1140 if (likely(!ctx || !cpuctx->task_ctx))
1141 return;
1142
1143 update_context_time(ctx);
1144
1145 rcu_read_lock();
1146 parent = rcu_dereference(ctx->parent_ctx);
1147 next_ctx = next->perf_counter_ctxp;
1148 if (parent && next_ctx &&
1149 rcu_dereference(next_ctx->parent_ctx) == parent) {
1150 /*
1151 * Looks like the two contexts are clones, so we might be
1152 * able to optimize the context switch. We lock both
1153 * contexts and check that they are clones under the
1154 * lock (including re-checking that neither has been
1155 * uncloned in the meantime). It doesn't matter which
1156 * order we take the locks because no other cpu could
1157 * be trying to lock both of these tasks.
1158 */
1159 spin_lock(&ctx->lock);
1160 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1161 if (context_equiv(ctx, next_ctx)) {
1162 /*
1163 * XXX do we need a memory barrier of sorts
1164 * wrt to rcu_dereference() of perf_counter_ctxp
1165 */
1166 task->perf_counter_ctxp = next_ctx;
1167 next->perf_counter_ctxp = ctx;
1168 ctx->task = next;
1169 next_ctx->task = task;
1170 do_switch = 0;
1171
1172 perf_counter_sync_stat(ctx, next_ctx);
1173 }
1174 spin_unlock(&next_ctx->lock);
1175 spin_unlock(&ctx->lock);
1176 }
1177 rcu_read_unlock();
1178
1179 if (do_switch) {
1180 __perf_counter_sched_out(ctx, cpuctx);
1181 cpuctx->task_ctx = NULL;
1182 }
1183}
1184
1185/*
1186 * Called with IRQs disabled
1187 */
1188static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
1189{
1190 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1191
1192 if (!cpuctx->task_ctx)
1193 return;
1194
1195 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1196 return;
1197
1198 __perf_counter_sched_out(ctx, cpuctx);
1199 cpuctx->task_ctx = NULL;
1200}
1201
1202/*
1203 * Called with IRQs disabled
1204 */
1205static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
1206{
1207 __perf_counter_sched_out(&cpuctx->ctx, cpuctx);
1208}
1209
1210static void
1211__perf_counter_sched_in(struct perf_counter_context *ctx,
1212 struct perf_cpu_context *cpuctx, int cpu)
1213{
1214 struct perf_counter *counter;
1215 int can_add_hw = 1;
1216
1217 spin_lock(&ctx->lock);
1218 ctx->is_active = 1;
1219 if (likely(!ctx->nr_counters))
1220 goto out;
1221
1222 ctx->timestamp = perf_clock();
1223
1224 perf_disable();
1225
1226 /*
1227 * First go through the list and put on any pinned groups
1228 * in order to give them the best chance of going on.
1229 */
1230 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1231 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1232 !counter->attr.pinned)
1233 continue;
1234 if (counter->cpu != -1 && counter->cpu != cpu)
1235 continue;
1236
1237 if (counter != counter->group_leader)
1238 counter_sched_in(counter, cpuctx, ctx, cpu);
1239 else {
1240 if (group_can_go_on(counter, cpuctx, 1))
1241 group_sched_in(counter, cpuctx, ctx, cpu);
1242 }
1243
1244 /*
1245 * If this pinned group hasn't been scheduled,
1246 * put it in error state.
1247 */
1248 if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1249 update_group_times(counter);
1250 counter->state = PERF_COUNTER_STATE_ERROR;
1251 }
1252 }
1253
1254 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1255 /*
1256 * Ignore counters in OFF or ERROR state, and
1257 * ignore pinned counters since we did them already.
1258 */
1259 if (counter->state <= PERF_COUNTER_STATE_OFF ||
1260 counter->attr.pinned)
1261 continue;
1262
1263 /*
1264 * Listen to the 'cpu' scheduling filter constraint
1265 * of counters:
1266 */
1267 if (counter->cpu != -1 && counter->cpu != cpu)
1268 continue;
1269
1270 if (counter != counter->group_leader) {
1271 if (counter_sched_in(counter, cpuctx, ctx, cpu))
1272 can_add_hw = 0;
1273 } else {
1274 if (group_can_go_on(counter, cpuctx, can_add_hw)) {
1275 if (group_sched_in(counter, cpuctx, ctx, cpu))
1276 can_add_hw = 0;
1277 }
1278 }
1279 }
1280 perf_enable();
1281 out:
1282 spin_unlock(&ctx->lock);
1283}
1284
1285/*
1286 * Called from scheduler to add the counters of the current task
1287 * with interrupts disabled.
1288 *
1289 * We restore the counter value and then enable it.
1290 *
1291 * This does not protect us against NMI, but enable()
1292 * sets the enabled bit in the control field of counter _before_
1293 * accessing the counter control register. If a NMI hits, then it will
1294 * keep the counter running.
1295 */
1296void perf_counter_task_sched_in(struct task_struct *task, int cpu)
1297{
1298 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1299 struct perf_counter_context *ctx = task->perf_counter_ctxp;
1300
1301 if (likely(!ctx))
1302 return;
1303 if (cpuctx->task_ctx == ctx)
1304 return;
1305 __perf_counter_sched_in(ctx, cpuctx, cpu);
1306 cpuctx->task_ctx = ctx;
1307}
1308
1309static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1310{
1311 struct perf_counter_context *ctx = &cpuctx->ctx;
1312
1313 __perf_counter_sched_in(ctx, cpuctx, cpu);
1314}
1315
1316#define MAX_INTERRUPTS (~0ULL)
1317
1318static void perf_log_throttle(struct perf_counter *counter, int enable);
1319
1320static void perf_adjust_period(struct perf_counter *counter, u64 events)
1321{
1322 struct hw_perf_counter *hwc = &counter->hw;
1323 u64 period, sample_period;
1324 s64 delta;
1325
1326 events *= hwc->sample_period;
1327 period = div64_u64(events, counter->attr.sample_freq);
1328
1329 delta = (s64)(period - hwc->sample_period);
1330 delta = (delta + 7) / 8; /* low pass filter */
1331
1332 sample_period = hwc->sample_period + delta;
1333
1334 if (!sample_period)
1335 sample_period = 1;
1336
1337 hwc->sample_period = sample_period;
1338}
1339
1340static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
1341{
1342 struct perf_counter *counter;
1343 struct hw_perf_counter *hwc;
1344 u64 interrupts, freq;
1345
1346 spin_lock(&ctx->lock);
1347 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1348 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
1349 continue;
1350
1351 hwc = &counter->hw;
1352
1353 interrupts = hwc->interrupts;
1354 hwc->interrupts = 0;
1355
1356 /*
1357 * unthrottle counters on the tick
1358 */
1359 if (interrupts == MAX_INTERRUPTS) {
1360 perf_log_throttle(counter, 1);
1361 counter->pmu->unthrottle(counter);
1362 interrupts = 2*sysctl_perf_counter_sample_rate/HZ;
1363 }
1364
1365 if (!counter->attr.freq || !counter->attr.sample_freq)
1366 continue;
1367
1368 /*
1369 * if the specified freq < HZ then we need to skip ticks
1370 */
1371 if (counter->attr.sample_freq < HZ) {
1372 freq = counter->attr.sample_freq;
1373
1374 hwc->freq_count += freq;
1375 hwc->freq_interrupts += interrupts;
1376
1377 if (hwc->freq_count < HZ)
1378 continue;
1379
1380 interrupts = hwc->freq_interrupts;
1381 hwc->freq_interrupts = 0;
1382 hwc->freq_count -= HZ;
1383 } else
1384 freq = HZ;
1385
1386 perf_adjust_period(counter, freq * interrupts);
1387
1388 /*
1389 * In order to avoid being stalled by an (accidental) huge
1390 * sample period, force reset the sample period if we didn't
1391 * get any events in this freq period.
1392 */
1393 if (!interrupts) {
1394 perf_disable();
1395 counter->pmu->disable(counter);
1396 atomic64_set(&hwc->period_left, 0);
1397 counter->pmu->enable(counter);
1398 perf_enable();
1399 }
1400 }
1401 spin_unlock(&ctx->lock);
1402}
1403
1404/*
1405 * Round-robin a context's counters:
1406 */
1407static void rotate_ctx(struct perf_counter_context *ctx)
1408{
1409 struct perf_counter *counter;
1410
1411 if (!ctx->nr_counters)
1412 return;
1413
1414 spin_lock(&ctx->lock);
1415 /*
1416 * Rotate the first entry last (works just fine for group counters too):
1417 */
1418 perf_disable();
1419 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1420 list_move_tail(&counter->list_entry, &ctx->counter_list);
1421 break;
1422 }
1423 perf_enable();
1424
1425 spin_unlock(&ctx->lock);
1426}
1427
1428void perf_counter_task_tick(struct task_struct *curr, int cpu)
1429{
1430 struct perf_cpu_context *cpuctx;
1431 struct perf_counter_context *ctx;
1432
1433 if (!atomic_read(&nr_counters))
1434 return;
1435
1436 cpuctx = &per_cpu(perf_cpu_context, cpu);
1437 ctx = curr->perf_counter_ctxp;
1438
1439 perf_ctx_adjust_freq(&cpuctx->ctx);
1440 if (ctx)
1441 perf_ctx_adjust_freq(ctx);
1442
1443 perf_counter_cpu_sched_out(cpuctx);
1444 if (ctx)
1445 __perf_counter_task_sched_out(ctx);
1446
1447 rotate_ctx(&cpuctx->ctx);
1448 if (ctx)
1449 rotate_ctx(ctx);
1450
1451 perf_counter_cpu_sched_in(cpuctx, cpu);
1452 if (ctx)
1453 perf_counter_task_sched_in(curr, cpu);
1454}
1455
1456/*
1457 * Enable all of a task's counters that have been marked enable-on-exec.
1458 * This expects task == current.
1459 */
1460static void perf_counter_enable_on_exec(struct task_struct *task)
1461{
1462 struct perf_counter_context *ctx;
1463 struct perf_counter *counter;
1464 unsigned long flags;
1465 int enabled = 0;
1466
1467 local_irq_save(flags);
1468 ctx = task->perf_counter_ctxp;
1469 if (!ctx || !ctx->nr_counters)
1470 goto out;
1471
1472 __perf_counter_task_sched_out(ctx);
1473
1474 spin_lock(&ctx->lock);
1475
1476 list_for_each_entry(counter, &ctx->counter_list, list_entry) {
1477 if (!counter->attr.enable_on_exec)
1478 continue;
1479 counter->attr.enable_on_exec = 0;
1480 if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
1481 continue;
1482 counter->state = PERF_COUNTER_STATE_INACTIVE;
1483 counter->tstamp_enabled =
1484 ctx->time - counter->total_time_enabled;
1485 enabled = 1;
1486 }
1487
1488 /*
1489 * Unclone this context if we enabled any counter.
1490 */
1491 if (enabled)
1492 unclone_ctx(ctx);
1493
1494 spin_unlock(&ctx->lock);
1495
1496 perf_counter_task_sched_in(task, smp_processor_id());
1497 out:
1498 local_irq_restore(flags);
1499}
1500
1501/*
1502 * Cross CPU call to read the hardware counter
1503 */
1504static void __perf_counter_read(void *info)
1505{
1506 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1507 struct perf_counter *counter = info;
1508 struct perf_counter_context *ctx = counter->ctx;
1509 unsigned long flags;
1510
1511 /*
1512 * If this is a task context, we need to check whether it is
1513 * the current task context of this cpu. If not it has been
1514 * scheduled out before the smp call arrived. In that case
1515 * counter->count would have been updated to a recent sample
1516 * when the counter was scheduled out.
1517 */
1518 if (ctx->task && cpuctx->task_ctx != ctx)
1519 return;
1520
1521 local_irq_save(flags);
1522 if (ctx->is_active)
1523 update_context_time(ctx);
1524 counter->pmu->read(counter);
1525 update_counter_times(counter);
1526 local_irq_restore(flags);
1527}
1528
1529static u64 perf_counter_read(struct perf_counter *counter)
1530{
1531 /*
1532 * If counter is enabled and currently active on a CPU, update the
1533 * value in the counter structure:
1534 */
1535 if (counter->state == PERF_COUNTER_STATE_ACTIVE) {
1536 smp_call_function_single(counter->oncpu,
1537 __perf_counter_read, counter, 1);
1538 } else if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
1539 update_counter_times(counter);
1540 }
1541
1542 return atomic64_read(&counter->count);
1543}
1544
1545/*
1546 * Initialize the perf_counter context in a task_struct:
1547 */
1548static void
1549__perf_counter_init_context(struct perf_counter_context *ctx,
1550 struct task_struct *task)
1551{
1552 memset(ctx, 0, sizeof(*ctx));
1553 spin_lock_init(&ctx->lock);
1554 mutex_init(&ctx->mutex);
1555 INIT_LIST_HEAD(&ctx->counter_list);
1556 INIT_LIST_HEAD(&ctx->event_list);
1557 atomic_set(&ctx->refcount, 1);
1558 ctx->task = task;
1559}
1560
1561static struct perf_counter_context *find_get_context(pid_t pid, int cpu)
1562{
1563 struct perf_counter_context *ctx;
1564 struct perf_cpu_context *cpuctx;
1565 struct task_struct *task;
1566 unsigned long flags;
1567 int err;
1568
1569 /*
1570 * If cpu is not a wildcard then this is a percpu counter:
1571 */
1572 if (cpu != -1) {
1573 /* Must be root to operate on a CPU counter: */
1574 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1575 return ERR_PTR(-EACCES);
1576
1577 if (cpu < 0 || cpu > num_possible_cpus())
1578 return ERR_PTR(-EINVAL);
1579
1580 /*
1581 * We could be clever and allow to attach a counter to an
1582 * offline CPU and activate it when the CPU comes up, but
1583 * that's for later.
1584 */
1585 if (!cpu_isset(cpu, cpu_online_map))
1586 return ERR_PTR(-ENODEV);
1587
1588 cpuctx = &per_cpu(perf_cpu_context, cpu);
1589 ctx = &cpuctx->ctx;
1590 get_ctx(ctx);
1591
1592 return ctx;
1593 }
1594
1595 rcu_read_lock();
1596 if (!pid)
1597 task = current;
1598 else
1599 task = find_task_by_vpid(pid);
1600 if (task)
1601 get_task_struct(task);
1602 rcu_read_unlock();
1603
1604 if (!task)
1605 return ERR_PTR(-ESRCH);
1606
1607 /*
1608 * Can't attach counters to a dying task.
1609 */
1610 err = -ESRCH;
1611 if (task->flags & PF_EXITING)
1612 goto errout;
1613
1614 /* Reuse ptrace permission checks for now. */
1615 err = -EACCES;
1616 if (!ptrace_may_access(task, PTRACE_MODE_READ))
1617 goto errout;
1618
1619 retry:
1620 ctx = perf_lock_task_context(task, &flags);
1621 if (ctx) {
1622 unclone_ctx(ctx);
1623 spin_unlock_irqrestore(&ctx->lock, flags);
1624 }
1625
1626 if (!ctx) {
1627 ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
1628 err = -ENOMEM;
1629 if (!ctx)
1630 goto errout;
1631 __perf_counter_init_context(ctx, task);
1632 get_ctx(ctx);
1633 if (cmpxchg(&task->perf_counter_ctxp, NULL, ctx)) {
1634 /*
1635 * We raced with some other task; use
1636 * the context they set.
1637 */
1638 kfree(ctx);
1639 goto retry;
1640 }
1641 get_task_struct(task);
1642 }
1643
1644 put_task_struct(task);
1645 return ctx;
1646
1647 errout:
1648 put_task_struct(task);
1649 return ERR_PTR(err);
1650}
1651
1652static void free_counter_rcu(struct rcu_head *head)
1653{
1654 struct perf_counter *counter;
1655
1656 counter = container_of(head, struct perf_counter, rcu_head);
1657 if (counter->ns)
1658 put_pid_ns(counter->ns);
1659 kfree(counter);
1660}
1661
1662static void perf_pending_sync(struct perf_counter *counter);
1663
1664static void free_counter(struct perf_counter *counter)
1665{
1666 perf_pending_sync(counter);
1667
1668 if (!counter->parent) {
1669 atomic_dec(&nr_counters);
1670 if (counter->attr.mmap)
1671 atomic_dec(&nr_mmap_counters);
1672 if (counter->attr.comm)
1673 atomic_dec(&nr_comm_counters);
1674 if (counter->attr.task)
1675 atomic_dec(&nr_task_counters);
1676 }
1677
1678 if (counter->destroy)
1679 counter->destroy(counter);
1680
1681 put_ctx(counter->ctx);
1682 call_rcu(&counter->rcu_head, free_counter_rcu);
1683}
1684
1685/*
1686 * Called when the last reference to the file is gone.
1687 */
1688static int perf_release(struct inode *inode, struct file *file)
1689{
1690 struct perf_counter *counter = file->private_data;
1691 struct perf_counter_context *ctx = counter->ctx;
1692
1693 file->private_data = NULL;
1694
1695 WARN_ON_ONCE(ctx->parent_ctx);
1696 mutex_lock(&ctx->mutex);
1697 perf_counter_remove_from_context(counter);
1698 mutex_unlock(&ctx->mutex);
1699
1700 mutex_lock(&counter->owner->perf_counter_mutex);
1701 list_del_init(&counter->owner_entry);
1702 mutex_unlock(&counter->owner->perf_counter_mutex);
1703 put_task_struct(counter->owner);
1704
1705 free_counter(counter);
1706
1707 return 0;
1708}
1709
1710static int perf_counter_read_size(struct perf_counter *counter)
1711{
1712 int entry = sizeof(u64); /* value */
1713 int size = 0;
1714 int nr = 1;
1715
1716 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1717 size += sizeof(u64);
1718
1719 if (counter->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1720 size += sizeof(u64);
1721
1722 if (counter->attr.read_format & PERF_FORMAT_ID)
1723 entry += sizeof(u64);
1724
1725 if (counter->attr.read_format & PERF_FORMAT_GROUP) {
1726 nr += counter->group_leader->nr_siblings;
1727 size += sizeof(u64);
1728 }
1729
1730 size += entry * nr;
1731
1732 return size;
1733}
1734
1735static u64 perf_counter_read_value(struct perf_counter *counter)
1736{
1737 struct perf_counter *child;
1738 u64 total = 0;
1739
1740 total += perf_counter_read(counter);
1741 list_for_each_entry(child, &counter->child_list, child_list)
1742 total += perf_counter_read(child);
1743
1744 return total;
1745}
1746
1747static int perf_counter_read_entry(struct perf_counter *counter,
1748 u64 read_format, char __user *buf)
1749{
1750 int n = 0, count = 0;
1751 u64 values[2];
1752
1753 values[n++] = perf_counter_read_value(counter);
1754 if (read_format & PERF_FORMAT_ID)
1755 values[n++] = primary_counter_id(counter);
1756
1757 count = n * sizeof(u64);
1758
1759 if (copy_to_user(buf, values, count))
1760 return -EFAULT;
1761
1762 return count;
1763}
1764
1765static int perf_counter_read_group(struct perf_counter *counter,
1766 u64 read_format, char __user *buf)
1767{
1768 struct perf_counter *leader = counter->group_leader, *sub;
1769 int n = 0, size = 0, err = -EFAULT;
1770 u64 values[3];
1771
1772 values[n++] = 1 + leader->nr_siblings;
1773 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1774 values[n++] = leader->total_time_enabled +
1775 atomic64_read(&leader->child_total_time_enabled);
1776 }
1777 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1778 values[n++] = leader->total_time_running +
1779 atomic64_read(&leader->child_total_time_running);
1780 }
1781
1782 size = n * sizeof(u64);
1783
1784 if (copy_to_user(buf, values, size))
1785 return -EFAULT;
1786
1787 err = perf_counter_read_entry(leader, read_format, buf + size);
1788 if (err < 0)
1789 return err;
1790
1791 size += err;
1792
1793 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
1794 err = perf_counter_read_entry(sub, read_format,
1795 buf + size);
1796 if (err < 0)
1797 return err;
1798
1799 size += err;
1800 }
1801
1802 return size;
1803}
1804
1805static int perf_counter_read_one(struct perf_counter *counter,
1806 u64 read_format, char __user *buf)
1807{
1808 u64 values[4];
1809 int n = 0;
1810
1811 values[n++] = perf_counter_read_value(counter);
1812 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1813 values[n++] = counter->total_time_enabled +
1814 atomic64_read(&counter->child_total_time_enabled);
1815 }
1816 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1817 values[n++] = counter->total_time_running +
1818 atomic64_read(&counter->child_total_time_running);
1819 }
1820 if (read_format & PERF_FORMAT_ID)
1821 values[n++] = primary_counter_id(counter);
1822
1823 if (copy_to_user(buf, values, n * sizeof(u64)))
1824 return -EFAULT;
1825
1826 return n * sizeof(u64);
1827}
1828
1829/*
1830 * Read the performance counter - simple non blocking version for now
1831 */
1832static ssize_t
1833perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
1834{
1835 u64 read_format = counter->attr.read_format;
1836 int ret;
1837
1838 /*
1839 * Return end-of-file for a read on a counter that is in
1840 * error state (i.e. because it was pinned but it couldn't be
1841 * scheduled on to the CPU at some point).
1842 */
1843 if (counter->state == PERF_COUNTER_STATE_ERROR)
1844 return 0;
1845
1846 if (count < perf_counter_read_size(counter))
1847 return -ENOSPC;
1848
1849 WARN_ON_ONCE(counter->ctx->parent_ctx);
1850 mutex_lock(&counter->child_mutex);
1851 if (read_format & PERF_FORMAT_GROUP)
1852 ret = perf_counter_read_group(counter, read_format, buf);
1853 else
1854 ret = perf_counter_read_one(counter, read_format, buf);
1855 mutex_unlock(&counter->child_mutex);
1856
1857 return ret;
1858}
1859
1860static ssize_t
1861perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1862{
1863 struct perf_counter *counter = file->private_data;
1864
1865 return perf_read_hw(counter, buf, count);
1866}
1867
1868static unsigned int perf_poll(struct file *file, poll_table *wait)
1869{
1870 struct perf_counter *counter = file->private_data;
1871 struct perf_mmap_data *data;
1872 unsigned int events = POLL_HUP;
1873
1874 rcu_read_lock();
1875 data = rcu_dereference(counter->data);
1876 if (data)
1877 events = atomic_xchg(&data->poll, 0);
1878 rcu_read_unlock();
1879
1880 poll_wait(file, &counter->waitq, wait);
1881
1882 return events;
1883}
1884
1885static void perf_counter_reset(struct perf_counter *counter)
1886{
1887 (void)perf_counter_read(counter);
1888 atomic64_set(&counter->count, 0);
1889 perf_counter_update_userpage(counter);
1890}
1891
1892/*
1893 * Holding the top-level counter's child_mutex means that any
1894 * descendant process that has inherited this counter will block
1895 * in sync_child_counter if it goes to exit, thus satisfying the
1896 * task existence requirements of perf_counter_enable/disable.
1897 */
1898static void perf_counter_for_each_child(struct perf_counter *counter,
1899 void (*func)(struct perf_counter *))
1900{
1901 struct perf_counter *child;
1902
1903 WARN_ON_ONCE(counter->ctx->parent_ctx);
1904 mutex_lock(&counter->child_mutex);
1905 func(counter);
1906 list_for_each_entry(child, &counter->child_list, child_list)
1907 func(child);
1908 mutex_unlock(&counter->child_mutex);
1909}
1910
1911static void perf_counter_for_each(struct perf_counter *counter,
1912 void (*func)(struct perf_counter *))
1913{
1914 struct perf_counter_context *ctx = counter->ctx;
1915 struct perf_counter *sibling;
1916
1917 WARN_ON_ONCE(ctx->parent_ctx);
1918 mutex_lock(&ctx->mutex);
1919 counter = counter->group_leader;
1920
1921 perf_counter_for_each_child(counter, func);
1922 func(counter);
1923 list_for_each_entry(sibling, &counter->sibling_list, list_entry)
1924 perf_counter_for_each_child(counter, func);
1925 mutex_unlock(&ctx->mutex);
1926}
1927
1928static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
1929{
1930 struct perf_counter_context *ctx = counter->ctx;
1931 unsigned long size;
1932 int ret = 0;
1933 u64 value;
1934
1935 if (!counter->attr.sample_period)
1936 return -EINVAL;
1937
1938 size = copy_from_user(&value, arg, sizeof(value));
1939 if (size != sizeof(value))
1940 return -EFAULT;
1941
1942 if (!value)
1943 return -EINVAL;
1944
1945 spin_lock_irq(&ctx->lock);
1946 if (counter->attr.freq) {
1947 if (value > sysctl_perf_counter_sample_rate) {
1948 ret = -EINVAL;
1949 goto unlock;
1950 }
1951
1952 counter->attr.sample_freq = value;
1953 } else {
1954 counter->attr.sample_period = value;
1955 counter->hw.sample_period = value;
1956 }
1957unlock:
1958 spin_unlock_irq(&ctx->lock);
1959
1960 return ret;
1961}
1962
1963static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1964{
1965 struct perf_counter *counter = file->private_data;
1966 void (*func)(struct perf_counter *);
1967 u32 flags = arg;
1968
1969 switch (cmd) {
1970 case PERF_COUNTER_IOC_ENABLE:
1971 func = perf_counter_enable;
1972 break;
1973 case PERF_COUNTER_IOC_DISABLE:
1974 func = perf_counter_disable;
1975 break;
1976 case PERF_COUNTER_IOC_RESET:
1977 func = perf_counter_reset;
1978 break;
1979
1980 case PERF_COUNTER_IOC_REFRESH:
1981 return perf_counter_refresh(counter, arg);
1982
1983 case PERF_COUNTER_IOC_PERIOD:
1984 return perf_counter_period(counter, (u64 __user *)arg);
1985
1986 default:
1987 return -ENOTTY;
1988 }
1989
1990 if (flags & PERF_IOC_FLAG_GROUP)
1991 perf_counter_for_each(counter, func);
1992 else
1993 perf_counter_for_each_child(counter, func);
1994
1995 return 0;
1996}
1997
1998int perf_counter_task_enable(void)
1999{
2000 struct perf_counter *counter;
2001
2002 mutex_lock(&current->perf_counter_mutex);
2003 list_for_each_entry(counter, &current->perf_counter_list, owner_entry)
2004 perf_counter_for_each_child(counter, perf_counter_enable);
2005 mutex_unlock(&current->perf_counter_mutex);
2006
2007 return 0;
2008}
2009
2010int perf_counter_task_disable(void)
2011{
2012 struct perf_counter *counter;
2013
2014 mutex_lock(&current->perf_counter_mutex);
2015 list_for_each_entry(counter, &current->perf_counter_list, owner_entry)
2016 perf_counter_for_each_child(counter, perf_counter_disable);
2017 mutex_unlock(&current->perf_counter_mutex);
2018
2019 return 0;
2020}
2021
2022#ifndef PERF_COUNTER_INDEX_OFFSET
2023# define PERF_COUNTER_INDEX_OFFSET 0
2024#endif
2025
2026static int perf_counter_index(struct perf_counter *counter)
2027{
2028 if (counter->state != PERF_COUNTER_STATE_ACTIVE)
2029 return 0;
2030
2031 return counter->hw.idx + 1 - PERF_COUNTER_INDEX_OFFSET;
2032}
2033
2034/*
2035 * Callers need to ensure there can be no nesting of this function, otherwise
2036 * the seqlock logic goes bad. We can not serialize this because the arch
2037 * code calls this from NMI context.
2038 */
2039void perf_counter_update_userpage(struct perf_counter *counter)
2040{
2041 struct perf_counter_mmap_page *userpg;
2042 struct perf_mmap_data *data;
2043
2044 rcu_read_lock();
2045 data = rcu_dereference(counter->data);
2046 if (!data)
2047 goto unlock;
2048
2049 userpg = data->user_page;
2050
2051 /*
2052 * Disable preemption so as to not let the corresponding user-space
2053 * spin too long if we get preempted.
2054 */
2055 preempt_disable();
2056 ++userpg->lock;
2057 barrier();
2058 userpg->index = perf_counter_index(counter);
2059 userpg->offset = atomic64_read(&counter->count);
2060 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
2061 userpg->offset -= atomic64_read(&counter->hw.prev_count);
2062
2063 userpg->time_enabled = counter->total_time_enabled +
2064 atomic64_read(&counter->child_total_time_enabled);
2065
2066 userpg->time_running = counter->total_time_running +
2067 atomic64_read(&counter->child_total_time_running);
2068
2069 barrier();
2070 ++userpg->lock;
2071 preempt_enable();
2072unlock:
2073 rcu_read_unlock();
2074}
2075
2076static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2077{
2078 struct perf_counter *counter = vma->vm_file->private_data;
2079 struct perf_mmap_data *data;
2080 int ret = VM_FAULT_SIGBUS;
2081
2082 if (vmf->flags & FAULT_FLAG_MKWRITE) {
2083 if (vmf->pgoff == 0)
2084 ret = 0;
2085 return ret;
2086 }
2087
2088 rcu_read_lock();
2089 data = rcu_dereference(counter->data);
2090 if (!data)
2091 goto unlock;
2092
2093 if (vmf->pgoff == 0) {
2094 vmf->page = virt_to_page(data->user_page);
2095 } else {
2096 int nr = vmf->pgoff - 1;
2097
2098 if ((unsigned)nr > data->nr_pages)
2099 goto unlock;
2100
2101 if (vmf->flags & FAULT_FLAG_WRITE)
2102 goto unlock;
2103
2104 vmf->page = virt_to_page(data->data_pages[nr]);
2105 }
2106
2107 get_page(vmf->page);
2108 vmf->page->mapping = vma->vm_file->f_mapping;
2109 vmf->page->index = vmf->pgoff;
2110
2111 ret = 0;
2112unlock:
2113 rcu_read_unlock();
2114
2115 return ret;
2116}
2117
2118static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
2119{
2120 struct perf_mmap_data *data;
2121 unsigned long size;
2122 int i;
2123
2124 WARN_ON(atomic_read(&counter->mmap_count));
2125
2126 size = sizeof(struct perf_mmap_data);
2127 size += nr_pages * sizeof(void *);
2128
2129 data = kzalloc(size, GFP_KERNEL);
2130 if (!data)
2131 goto fail;
2132
2133 data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2134 if (!data->user_page)
2135 goto fail_user_page;
2136
2137 for (i = 0; i < nr_pages; i++) {
2138 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2139 if (!data->data_pages[i])
2140 goto fail_data_pages;
2141 }
2142
2143 data->nr_pages = nr_pages;
2144 atomic_set(&data->lock, -1);
2145
2146 rcu_assign_pointer(counter->data, data);
2147
2148 return 0;
2149
2150fail_data_pages:
2151 for (i--; i >= 0; i--)
2152 free_page((unsigned long)data->data_pages[i]);
2153
2154 free_page((unsigned long)data->user_page);
2155
2156fail_user_page:
2157 kfree(data);
2158
2159fail:
2160 return -ENOMEM;
2161}
2162
2163static void perf_mmap_free_page(unsigned long addr)
2164{
2165 struct page *page = virt_to_page((void *)addr);
2166
2167 page->mapping = NULL;
2168 __free_page(page);
2169}
2170
2171static void __perf_mmap_data_free(struct rcu_head *rcu_head)
2172{
2173 struct perf_mmap_data *data;
2174 int i;
2175
2176 data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2177
2178 perf_mmap_free_page((unsigned long)data->user_page);
2179 for (i = 0; i < data->nr_pages; i++)
2180 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2181
2182 kfree(data);
2183}
2184
2185static void perf_mmap_data_free(struct perf_counter *counter)
2186{
2187 struct perf_mmap_data *data = counter->data;
2188
2189 WARN_ON(atomic_read(&counter->mmap_count));
2190
2191 rcu_assign_pointer(counter->data, NULL);
2192 call_rcu(&data->rcu_head, __perf_mmap_data_free);
2193}
2194
2195static void perf_mmap_open(struct vm_area_struct *vma)
2196{
2197 struct perf_counter *counter = vma->vm_file->private_data;
2198
2199 atomic_inc(&counter->mmap_count);
2200}
2201
2202static void perf_mmap_close(struct vm_area_struct *vma)
2203{
2204 struct perf_counter *counter = vma->vm_file->private_data;
2205
2206 WARN_ON_ONCE(counter->ctx->parent_ctx);
2207 if (atomic_dec_and_mutex_lock(&counter->mmap_count, &counter->mmap_mutex)) {
2208 struct user_struct *user = current_user();
2209
2210 atomic_long_sub(counter->data->nr_pages + 1, &user->locked_vm);
2211 vma->vm_mm->locked_vm -= counter->data->nr_locked;
2212 perf_mmap_data_free(counter);
2213 mutex_unlock(&counter->mmap_mutex);
2214 }
2215}
2216
2217static struct vm_operations_struct perf_mmap_vmops = {
2218 .open = perf_mmap_open,
2219 .close = perf_mmap_close,
2220 .fault = perf_mmap_fault,
2221 .page_mkwrite = perf_mmap_fault,
2222};
2223
2224static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2225{
2226 struct perf_counter *counter = file->private_data;
2227 unsigned long user_locked, user_lock_limit;
2228 struct user_struct *user = current_user();
2229 unsigned long locked, lock_limit;
2230 unsigned long vma_size;
2231 unsigned long nr_pages;
2232 long user_extra, extra;
2233 int ret = 0;
2234
2235 if (!(vma->vm_flags & VM_SHARED))
2236 return -EINVAL;
2237
2238 vma_size = vma->vm_end - vma->vm_start;
2239 nr_pages = (vma_size / PAGE_SIZE) - 1;
2240
2241 /*
2242 * If we have data pages ensure they're a power-of-two number, so we
2243 * can do bitmasks instead of modulo.
2244 */
2245 if (nr_pages != 0 && !is_power_of_2(nr_pages))
2246 return -EINVAL;
2247
2248 if (vma_size != PAGE_SIZE * (1 + nr_pages))
2249 return -EINVAL;
2250
2251 if (vma->vm_pgoff != 0)
2252 return -EINVAL;
2253
2254 WARN_ON_ONCE(counter->ctx->parent_ctx);
2255 mutex_lock(&counter->mmap_mutex);
2256 if (atomic_inc_not_zero(&counter->mmap_count)) {
2257 if (nr_pages != counter->data->nr_pages)
2258 ret = -EINVAL;
2259 goto unlock;
2260 }
2261
2262 user_extra = nr_pages + 1;
2263 user_lock_limit = sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
2264
2265 /*
2266 * Increase the limit linearly with more CPUs:
2267 */
2268 user_lock_limit *= num_online_cpus();
2269
2270 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2271
2272 extra = 0;
2273 if (user_locked > user_lock_limit)
2274 extra = user_locked - user_lock_limit;
2275
2276 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2277 lock_limit >>= PAGE_SHIFT;
2278 locked = vma->vm_mm->locked_vm + extra;
2279
2280 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
2281 ret = -EPERM;
2282 goto unlock;
2283 }
2284
2285 WARN_ON(counter->data);
2286 ret = perf_mmap_data_alloc(counter, nr_pages);
2287 if (ret)
2288 goto unlock;
2289
2290 atomic_set(&counter->mmap_count, 1);
2291 atomic_long_add(user_extra, &user->locked_vm);
2292 vma->vm_mm->locked_vm += extra;
2293 counter->data->nr_locked = extra;
2294 if (vma->vm_flags & VM_WRITE)
2295 counter->data->writable = 1;
2296
2297unlock:
2298 mutex_unlock(&counter->mmap_mutex);
2299
2300 vma->vm_flags |= VM_RESERVED;
2301 vma->vm_ops = &perf_mmap_vmops;
2302
2303 return ret;
2304}
2305
2306static int perf_fasync(int fd, struct file *filp, int on)
2307{
2308 struct inode *inode = filp->f_path.dentry->d_inode;
2309 struct perf_counter *counter = filp->private_data;
2310 int retval;
2311
2312 mutex_lock(&inode->i_mutex);
2313 retval = fasync_helper(fd, filp, on, &counter->fasync);
2314 mutex_unlock(&inode->i_mutex);
2315
2316 if (retval < 0)
2317 return retval;
2318
2319 return 0;
2320}
2321
2322static const struct file_operations perf_fops = {
2323 .release = perf_release,
2324 .read = perf_read,
2325 .poll = perf_poll,
2326 .unlocked_ioctl = perf_ioctl,
2327 .compat_ioctl = perf_ioctl,
2328 .mmap = perf_mmap,
2329 .fasync = perf_fasync,
2330};
2331
2332/*
2333 * Perf counter wakeup
2334 *
2335 * If there's data, ensure we set the poll() state and publish everything
2336 * to user-space before waking everybody up.
2337 */
2338
2339void perf_counter_wakeup(struct perf_counter *counter)
2340{
2341 wake_up_all(&counter->waitq);
2342
2343 if (counter->pending_kill) {
2344 kill_fasync(&counter->fasync, SIGIO, counter->pending_kill);
2345 counter->pending_kill = 0;
2346 }
2347}
2348
2349/*
2350 * Pending wakeups
2351 *
2352 * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2353 *
2354 * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2355 * single linked list and use cmpxchg() to add entries lockless.
2356 */
2357
2358static void perf_pending_counter(struct perf_pending_entry *entry)
2359{
2360 struct perf_counter *counter = container_of(entry,
2361 struct perf_counter, pending);
2362
2363 if (counter->pending_disable) {
2364 counter->pending_disable = 0;
2365 __perf_counter_disable(counter);
2366 }
2367
2368 if (counter->pending_wakeup) {
2369 counter->pending_wakeup = 0;
2370 perf_counter_wakeup(counter);
2371 }
2372}
2373
2374#define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2375
2376static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2377 PENDING_TAIL,
2378};
2379
2380static void perf_pending_queue(struct perf_pending_entry *entry,
2381 void (*func)(struct perf_pending_entry *))
2382{
2383 struct perf_pending_entry **head;
2384
2385 if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2386 return;
2387
2388 entry->func = func;
2389
2390 head = &get_cpu_var(perf_pending_head);
2391
2392 do {
2393 entry->next = *head;
2394 } while (cmpxchg(head, entry->next, entry) != entry->next);
2395
2396 set_perf_counter_pending();
2397
2398 put_cpu_var(perf_pending_head);
2399}
2400
2401static int __perf_pending_run(void)
2402{
2403 struct perf_pending_entry *list;
2404 int nr = 0;
2405
2406 list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2407 while (list != PENDING_TAIL) {
2408 void (*func)(struct perf_pending_entry *);
2409 struct perf_pending_entry *entry = list;
2410
2411 list = list->next;
2412
2413 func = entry->func;
2414 entry->next = NULL;
2415 /*
2416 * Ensure we observe the unqueue before we issue the wakeup,
2417 * so that we won't be waiting forever.
2418 * -- see perf_not_pending().
2419 */
2420 smp_wmb();
2421
2422 func(entry);
2423 nr++;
2424 }
2425
2426 return nr;
2427}
2428
2429static inline int perf_not_pending(struct perf_counter *counter)
2430{
2431 /*
2432 * If we flush on whatever cpu we run, there is a chance we don't
2433 * need to wait.
2434 */
2435 get_cpu();
2436 __perf_pending_run();
2437 put_cpu();
2438
2439 /*
2440 * Ensure we see the proper queue state before going to sleep
2441 * so that we do not miss the wakeup. -- see perf_pending_handle()
2442 */
2443 smp_rmb();
2444 return counter->pending.next == NULL;
2445}
2446
2447static void perf_pending_sync(struct perf_counter *counter)
2448{
2449 wait_event(counter->waitq, perf_not_pending(counter));
2450}
2451
2452void perf_counter_do_pending(void)
2453{
2454 __perf_pending_run();
2455}
2456
2457/*
2458 * Callchain support -- arch specific
2459 */
2460
2461__weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2462{
2463 return NULL;
2464}
2465
2466/*
2467 * Output
2468 */
2469
2470struct perf_output_handle {
2471 struct perf_counter *counter;
2472 struct perf_mmap_data *data;
2473 unsigned long head;
2474 unsigned long offset;
2475 int nmi;
2476 int sample;
2477 int locked;
2478 unsigned long flags;
2479};
2480
2481static bool perf_output_space(struct perf_mmap_data *data,
2482 unsigned int offset, unsigned int head)
2483{
2484 unsigned long tail;
2485 unsigned long mask;
2486
2487 if (!data->writable)
2488 return true;
2489
2490 mask = (data->nr_pages << PAGE_SHIFT) - 1;
2491 /*
2492 * Userspace could choose to issue a mb() before updating the tail
2493 * pointer. So that all reads will be completed before the write is
2494 * issued.
2495 */
2496 tail = ACCESS_ONCE(data->user_page->data_tail);
2497 smp_rmb();
2498
2499 offset = (offset - tail) & mask;
2500 head = (head - tail) & mask;
2501
2502 if ((int)(head - offset) < 0)
2503 return false;
2504
2505 return true;
2506}
2507
2508static void perf_output_wakeup(struct perf_output_handle *handle)
2509{
2510 atomic_set(&handle->data->poll, POLL_IN);
2511
2512 if (handle->nmi) {
2513 handle->counter->pending_wakeup = 1;
2514 perf_pending_queue(&handle->counter->pending,
2515 perf_pending_counter);
2516 } else
2517 perf_counter_wakeup(handle->counter);
2518}
2519
2520/*
2521 * Curious locking construct.
2522 *
2523 * We need to ensure a later event doesn't publish a head when a former
2524 * event isn't done writing. However since we need to deal with NMIs we
2525 * cannot fully serialize things.
2526 *
2527 * What we do is serialize between CPUs so we only have to deal with NMI
2528 * nesting on a single CPU.
2529 *
2530 * We only publish the head (and generate a wakeup) when the outer-most
2531 * event completes.
2532 */
2533static void perf_output_lock(struct perf_output_handle *handle)
2534{
2535 struct perf_mmap_data *data = handle->data;
2536 int cpu;
2537
2538 handle->locked = 0;
2539
2540 local_irq_save(handle->flags);
2541 cpu = smp_processor_id();
2542
2543 if (in_nmi() && atomic_read(&data->lock) == cpu)
2544 return;
2545
2546 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2547 cpu_relax();
2548
2549 handle->locked = 1;
2550}
2551
2552static void perf_output_unlock(struct perf_output_handle *handle)
2553{
2554 struct perf_mmap_data *data = handle->data;
2555 unsigned long head;
2556 int cpu;
2557
2558 data->done_head = data->head;
2559
2560 if (!handle->locked)
2561 goto out;
2562
2563again:
2564 /*
2565 * The xchg implies a full barrier that ensures all writes are done
2566 * before we publish the new head, matched by a rmb() in userspace when
2567 * reading this position.
2568 */
2569 while ((head = atomic_long_xchg(&data->done_head, 0)))
2570 data->user_page->data_head = head;
2571
2572 /*
2573 * NMI can happen here, which means we can miss a done_head update.
2574 */
2575
2576 cpu = atomic_xchg(&data->lock, -1);
2577 WARN_ON_ONCE(cpu != smp_processor_id());
2578
2579 /*
2580 * Therefore we have to validate we did not indeed do so.
2581 */
2582 if (unlikely(atomic_long_read(&data->done_head))) {
2583 /*
2584 * Since we had it locked, we can lock it again.
2585 */
2586 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2587 cpu_relax();
2588
2589 goto again;
2590 }
2591
2592 if (atomic_xchg(&data->wakeup, 0))
2593 perf_output_wakeup(handle);
2594out:
2595 local_irq_restore(handle->flags);
2596}
2597
2598static void perf_output_copy(struct perf_output_handle *handle,
2599 const void *buf, unsigned int len)
2600{
2601 unsigned int pages_mask;
2602 unsigned int offset;
2603 unsigned int size;
2604 void **pages;
2605
2606 offset = handle->offset;
2607 pages_mask = handle->data->nr_pages - 1;
2608 pages = handle->data->data_pages;
2609
2610 do {
2611 unsigned int page_offset;
2612 int nr;
2613
2614 nr = (offset >> PAGE_SHIFT) & pages_mask;
2615 page_offset = offset & (PAGE_SIZE - 1);
2616 size = min_t(unsigned int, PAGE_SIZE - page_offset, len);
2617
2618 memcpy(pages[nr] + page_offset, buf, size);
2619
2620 len -= size;
2621 buf += size;
2622 offset += size;
2623 } while (len);
2624
2625 handle->offset = offset;
2626
2627 /*
2628 * Check we didn't copy past our reservation window, taking the
2629 * possible unsigned int wrap into account.
2630 */
2631 WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2632}
2633
2634#define perf_output_put(handle, x) \
2635 perf_output_copy((handle), &(x), sizeof(x))
2636
2637static int perf_output_begin(struct perf_output_handle *handle,
2638 struct perf_counter *counter, unsigned int size,
2639 int nmi, int sample)
2640{
2641 struct perf_mmap_data *data;
2642 unsigned int offset, head;
2643 int have_lost;
2644 struct {
2645 struct perf_event_header header;
2646 u64 id;
2647 u64 lost;
2648 } lost_event;
2649
2650 /*
2651 * For inherited counters we send all the output towards the parent.
2652 */
2653 if (counter->parent)
2654 counter = counter->parent;
2655
2656 rcu_read_lock();
2657 data = rcu_dereference(counter->data);
2658 if (!data)
2659 goto out;
2660
2661 handle->data = data;
2662 handle->counter = counter;
2663 handle->nmi = nmi;
2664 handle->sample = sample;
2665
2666 if (!data->nr_pages)
2667 goto fail;
2668
2669 have_lost = atomic_read(&data->lost);
2670 if (have_lost)
2671 size += sizeof(lost_event);
2672
2673 perf_output_lock(handle);
2674
2675 do {
2676 offset = head = atomic_long_read(&data->head);
2677 head += size;
2678 if (unlikely(!perf_output_space(data, offset, head)))
2679 goto fail;
2680 } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2681
2682 handle->offset = offset;
2683 handle->head = head;
2684
2685 if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
2686 atomic_set(&data->wakeup, 1);
2687
2688 if (have_lost) {
2689 lost_event.header.type = PERF_EVENT_LOST;
2690 lost_event.header.misc = 0;
2691 lost_event.header.size = sizeof(lost_event);
2692 lost_event.id = counter->id;
2693 lost_event.lost = atomic_xchg(&data->lost, 0);
2694
2695 perf_output_put(handle, lost_event);
2696 }
2697
2698 return 0;
2699
2700fail:
2701 atomic_inc(&data->lost);
2702 perf_output_unlock(handle);
2703out:
2704 rcu_read_unlock();
2705
2706 return -ENOSPC;
2707}
2708
2709static void perf_output_end(struct perf_output_handle *handle)
2710{
2711 struct perf_counter *counter = handle->counter;
2712 struct perf_mmap_data *data = handle->data;
2713
2714 int wakeup_events = counter->attr.wakeup_events;
2715
2716 if (handle->sample && wakeup_events) {
2717 int events = atomic_inc_return(&data->events);
2718 if (events >= wakeup_events) {
2719 atomic_sub(wakeup_events, &data->events);
2720 atomic_set(&data->wakeup, 1);
2721 }
2722 }
2723
2724 perf_output_unlock(handle);
2725 rcu_read_unlock();
2726}
2727
2728static u32 perf_counter_pid(struct perf_counter *counter, struct task_struct *p)
2729{
2730 /*
2731 * only top level counters have the pid namespace they were created in
2732 */
2733 if (counter->parent)
2734 counter = counter->parent;
2735
2736 return task_tgid_nr_ns(p, counter->ns);
2737}
2738
2739static u32 perf_counter_tid(struct perf_counter *counter, struct task_struct *p)
2740{
2741 /*
2742 * only top level counters have the pid namespace they were created in
2743 */
2744 if (counter->parent)
2745 counter = counter->parent;
2746
2747 return task_pid_nr_ns(p, counter->ns);
2748}
2749
2750static void perf_output_read_one(struct perf_output_handle *handle,
2751 struct perf_counter *counter)
2752{
2753 u64 read_format = counter->attr.read_format;
2754 u64 values[4];
2755 int n = 0;
2756
2757 values[n++] = atomic64_read(&counter->count);
2758 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2759 values[n++] = counter->total_time_enabled +
2760 atomic64_read(&counter->child_total_time_enabled);
2761 }
2762 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2763 values[n++] = counter->total_time_running +
2764 atomic64_read(&counter->child_total_time_running);
2765 }
2766 if (read_format & PERF_FORMAT_ID)
2767 values[n++] = primary_counter_id(counter);
2768
2769 perf_output_copy(handle, values, n * sizeof(u64));
2770}
2771
2772/*
2773 * XXX PERF_FORMAT_GROUP vs inherited counters seems difficult.
2774 */
2775static void perf_output_read_group(struct perf_output_handle *handle,
2776 struct perf_counter *counter)
2777{
2778 struct perf_counter *leader = counter->group_leader, *sub;
2779 u64 read_format = counter->attr.read_format;
2780 u64 values[5];
2781 int n = 0;
2782
2783 values[n++] = 1 + leader->nr_siblings;
2784
2785 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2786 values[n++] = leader->total_time_enabled;
2787
2788 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2789 values[n++] = leader->total_time_running;
2790
2791 if (leader != counter)
2792 leader->pmu->read(leader);
2793
2794 values[n++] = atomic64_read(&leader->count);
2795 if (read_format & PERF_FORMAT_ID)
2796 values[n++] = primary_counter_id(leader);
2797
2798 perf_output_copy(handle, values, n * sizeof(u64));
2799
2800 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
2801 n = 0;
2802
2803 if (sub != counter)
2804 sub->pmu->read(sub);
2805
2806 values[n++] = atomic64_read(&sub->count);
2807 if (read_format & PERF_FORMAT_ID)
2808 values[n++] = primary_counter_id(sub);
2809
2810 perf_output_copy(handle, values, n * sizeof(u64));
2811 }
2812}
2813
2814static void perf_output_read(struct perf_output_handle *handle,
2815 struct perf_counter *counter)
2816{
2817 if (counter->attr.read_format & PERF_FORMAT_GROUP)
2818 perf_output_read_group(handle, counter);
2819 else
2820 perf_output_read_one(handle, counter);
2821}
2822
2823void perf_counter_output(struct perf_counter *counter, int nmi,
2824 struct perf_sample_data *data)
2825{
2826 int ret;
2827 u64 sample_type = counter->attr.sample_type;
2828 struct perf_output_handle handle;
2829 struct perf_event_header header;
2830 u64 ip;
2831 struct {
2832 u32 pid, tid;
2833 } tid_entry;
2834 struct perf_callchain_entry *callchain = NULL;
2835 int callchain_size = 0;
2836 u64 time;
2837 struct {
2838 u32 cpu, reserved;
2839 } cpu_entry;
2840
2841 header.type = PERF_EVENT_SAMPLE;
2842 header.size = sizeof(header);
2843
2844 header.misc = 0;
2845 header.misc |= perf_misc_flags(data->regs);
2846
2847 if (sample_type & PERF_SAMPLE_IP) {
2848 ip = perf_instruction_pointer(data->regs);
2849 header.size += sizeof(ip);
2850 }
2851
2852 if (sample_type & PERF_SAMPLE_TID) {
2853 /* namespace issues */
2854 tid_entry.pid = perf_counter_pid(counter, current);
2855 tid_entry.tid = perf_counter_tid(counter, current);
2856
2857 header.size += sizeof(tid_entry);
2858 }
2859
2860 if (sample_type & PERF_SAMPLE_TIME) {
2861 /*
2862 * Maybe do better on x86 and provide cpu_clock_nmi()
2863 */
2864 time = sched_clock();
2865
2866 header.size += sizeof(u64);
2867 }
2868
2869 if (sample_type & PERF_SAMPLE_ADDR)
2870 header.size += sizeof(u64);
2871
2872 if (sample_type & PERF_SAMPLE_ID)
2873 header.size += sizeof(u64);
2874
2875 if (sample_type & PERF_SAMPLE_STREAM_ID)
2876 header.size += sizeof(u64);
2877
2878 if (sample_type & PERF_SAMPLE_CPU) {
2879 header.size += sizeof(cpu_entry);
2880
2881 cpu_entry.cpu = raw_smp_processor_id();
2882 cpu_entry.reserved = 0;
2883 }
2884
2885 if (sample_type & PERF_SAMPLE_PERIOD)
2886 header.size += sizeof(u64);
2887
2888 if (sample_type & PERF_SAMPLE_READ)
2889 header.size += perf_counter_read_size(counter);
2890
2891 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2892 callchain = perf_callchain(data->regs);
2893
2894 if (callchain) {
2895 callchain_size = (1 + callchain->nr) * sizeof(u64);
2896 header.size += callchain_size;
2897 } else
2898 header.size += sizeof(u64);
2899 }
2900
2901 if (sample_type & PERF_SAMPLE_RAW) {
2902 int size = sizeof(u32);
2903
2904 if (data->raw)
2905 size += data->raw->size;
2906 else
2907 size += sizeof(u32);
2908
2909 WARN_ON_ONCE(size & (sizeof(u64)-1));
2910 header.size += size;
2911 }
2912
2913 ret = perf_output_begin(&handle, counter, header.size, nmi, 1);
2914 if (ret)
2915 return;
2916
2917 perf_output_put(&handle, header);
2918
2919 if (sample_type & PERF_SAMPLE_IP)
2920 perf_output_put(&handle, ip);
2921
2922 if (sample_type & PERF_SAMPLE_TID)
2923 perf_output_put(&handle, tid_entry);
2924
2925 if (sample_type & PERF_SAMPLE_TIME)
2926 perf_output_put(&handle, time);
2927
2928 if (sample_type & PERF_SAMPLE_ADDR)
2929 perf_output_put(&handle, data->addr);
2930
2931 if (sample_type & PERF_SAMPLE_ID) {
2932 u64 id = primary_counter_id(counter);
2933
2934 perf_output_put(&handle, id);
2935 }
2936
2937 if (sample_type & PERF_SAMPLE_STREAM_ID)
2938 perf_output_put(&handle, counter->id);
2939
2940 if (sample_type & PERF_SAMPLE_CPU)
2941 perf_output_put(&handle, cpu_entry);
2942
2943 if (sample_type & PERF_SAMPLE_PERIOD)
2944 perf_output_put(&handle, data->period);
2945
2946 if (sample_type & PERF_SAMPLE_READ)
2947 perf_output_read(&handle, counter);
2948
2949 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
2950 if (callchain)
2951 perf_output_copy(&handle, callchain, callchain_size);
2952 else {
2953 u64 nr = 0;
2954 perf_output_put(&handle, nr);
2955 }
2956 }
2957
2958 if (sample_type & PERF_SAMPLE_RAW) {
2959 if (data->raw) {
2960 perf_output_put(&handle, data->raw->size);
2961 perf_output_copy(&handle, data->raw->data, data->raw->size);
2962 } else {
2963 struct {
2964 u32 size;
2965 u32 data;
2966 } raw = {
2967 .size = sizeof(u32),
2968 .data = 0,
2969 };
2970 perf_output_put(&handle, raw);
2971 }
2972 }
2973
2974 perf_output_end(&handle);
2975}
2976
2977/*
2978 * read event
2979 */
2980
2981struct perf_read_event {
2982 struct perf_event_header header;
2983
2984 u32 pid;
2985 u32 tid;
2986};
2987
2988static void
2989perf_counter_read_event(struct perf_counter *counter,
2990 struct task_struct *task)
2991{
2992 struct perf_output_handle handle;
2993 struct perf_read_event event = {
2994 .header = {
2995 .type = PERF_EVENT_READ,
2996 .misc = 0,
2997 .size = sizeof(event) + perf_counter_read_size(counter),
2998 },
2999 .pid = perf_counter_pid(counter, task),
3000 .tid = perf_counter_tid(counter, task),
3001 };
3002 int ret;
3003
3004 ret = perf_output_begin(&handle, counter, event.header.size, 0, 0);
3005 if (ret)
3006 return;
3007
3008 perf_output_put(&handle, event);
3009 perf_output_read(&handle, counter);
3010
3011 perf_output_end(&handle);
3012}
3013
3014/*
3015 * task tracking -- fork/exit
3016 *
3017 * enabled by: attr.comm | attr.mmap | attr.task
3018 */
3019
3020struct perf_task_event {
3021 struct task_struct *task;
3022 struct perf_counter_context *task_ctx;
3023
3024 struct {
3025 struct perf_event_header header;
3026
3027 u32 pid;
3028 u32 ppid;
3029 u32 tid;
3030 u32 ptid;
3031 } event;
3032};
3033
3034static void perf_counter_task_output(struct perf_counter *counter,
3035 struct perf_task_event *task_event)
3036{
3037 struct perf_output_handle handle;
3038 int size = task_event->event.header.size;
3039 struct task_struct *task = task_event->task;
3040 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3041
3042 if (ret)
3043 return;
3044
3045 task_event->event.pid = perf_counter_pid(counter, task);
3046 task_event->event.ppid = perf_counter_pid(counter, current);
3047
3048 task_event->event.tid = perf_counter_tid(counter, task);
3049 task_event->event.ptid = perf_counter_tid(counter, current);
3050
3051 perf_output_put(&handle, task_event->event);
3052 perf_output_end(&handle);
3053}
3054
3055static int perf_counter_task_match(struct perf_counter *counter)
3056{
3057 if (counter->attr.comm || counter->attr.mmap || counter->attr.task)
3058 return 1;
3059
3060 return 0;
3061}
3062
3063static void perf_counter_task_ctx(struct perf_counter_context *ctx,
3064 struct perf_task_event *task_event)
3065{
3066 struct perf_counter *counter;
3067
3068 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3069 return;
3070
3071 rcu_read_lock();
3072 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3073 if (perf_counter_task_match(counter))
3074 perf_counter_task_output(counter, task_event);
3075 }
3076 rcu_read_unlock();
3077}
3078
3079static void perf_counter_task_event(struct perf_task_event *task_event)
3080{
3081 struct perf_cpu_context *cpuctx;
3082 struct perf_counter_context *ctx = task_event->task_ctx;
3083
3084 cpuctx = &get_cpu_var(perf_cpu_context);
3085 perf_counter_task_ctx(&cpuctx->ctx, task_event);
3086 put_cpu_var(perf_cpu_context);
3087
3088 rcu_read_lock();
3089 if (!ctx)
3090 ctx = rcu_dereference(task_event->task->perf_counter_ctxp);
3091 if (ctx)
3092 perf_counter_task_ctx(ctx, task_event);
3093 rcu_read_unlock();
3094}
3095
3096static void perf_counter_task(struct task_struct *task,
3097 struct perf_counter_context *task_ctx,
3098 int new)
3099{
3100 struct perf_task_event task_event;
3101
3102 if (!atomic_read(&nr_comm_counters) &&
3103 !atomic_read(&nr_mmap_counters) &&
3104 !atomic_read(&nr_task_counters))
3105 return;
3106
3107 task_event = (struct perf_task_event){
3108 .task = task,
3109 .task_ctx = task_ctx,
3110 .event = {
3111 .header = {
3112 .type = new ? PERF_EVENT_FORK : PERF_EVENT_EXIT,
3113 .misc = 0,
3114 .size = sizeof(task_event.event),
3115 },
3116 /* .pid */
3117 /* .ppid */
3118 /* .tid */
3119 /* .ptid */
3120 },
3121 };
3122
3123 perf_counter_task_event(&task_event);
3124}
3125
3126void perf_counter_fork(struct task_struct *task)
3127{
3128 perf_counter_task(task, NULL, 1);
3129}
3130
3131/*
3132 * comm tracking
3133 */
3134
3135struct perf_comm_event {
3136 struct task_struct *task;
3137 char *comm;
3138 int comm_size;
3139
3140 struct {
3141 struct perf_event_header header;
3142
3143 u32 pid;
3144 u32 tid;
3145 } event;
3146};
3147
3148static void perf_counter_comm_output(struct perf_counter *counter,
3149 struct perf_comm_event *comm_event)
3150{
3151 struct perf_output_handle handle;
3152 int size = comm_event->event.header.size;
3153 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3154
3155 if (ret)
3156 return;
3157
3158 comm_event->event.pid = perf_counter_pid(counter, comm_event->task);
3159 comm_event->event.tid = perf_counter_tid(counter, comm_event->task);
3160
3161 perf_output_put(&handle, comm_event->event);
3162 perf_output_copy(&handle, comm_event->comm,
3163 comm_event->comm_size);
3164 perf_output_end(&handle);
3165}
3166
3167static int perf_counter_comm_match(struct perf_counter *counter)
3168{
3169 if (counter->attr.comm)
3170 return 1;
3171
3172 return 0;
3173}
3174
3175static void perf_counter_comm_ctx(struct perf_counter_context *ctx,
3176 struct perf_comm_event *comm_event)
3177{
3178 struct perf_counter *counter;
3179
3180 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3181 return;
3182
3183 rcu_read_lock();
3184 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3185 if (perf_counter_comm_match(counter))
3186 perf_counter_comm_output(counter, comm_event);
3187 }
3188 rcu_read_unlock();
3189}
3190
3191static void perf_counter_comm_event(struct perf_comm_event *comm_event)
3192{
3193 struct perf_cpu_context *cpuctx;
3194 struct perf_counter_context *ctx;
3195 unsigned int size;
3196 char comm[TASK_COMM_LEN];
3197
3198 memset(comm, 0, sizeof(comm));
3199 strncpy(comm, comm_event->task->comm, sizeof(comm));
3200 size = ALIGN(strlen(comm)+1, sizeof(u64));
3201
3202 comm_event->comm = comm;
3203 comm_event->comm_size = size;
3204
3205 comm_event->event.header.size = sizeof(comm_event->event) + size;
3206
3207 cpuctx = &get_cpu_var(perf_cpu_context);
3208 perf_counter_comm_ctx(&cpuctx->ctx, comm_event);
3209 put_cpu_var(perf_cpu_context);
3210
3211 rcu_read_lock();
3212 /*
3213 * doesn't really matter which of the child contexts the
3214 * events ends up in.
3215 */
3216 ctx = rcu_dereference(current->perf_counter_ctxp);
3217 if (ctx)
3218 perf_counter_comm_ctx(ctx, comm_event);
3219 rcu_read_unlock();
3220}
3221
3222void perf_counter_comm(struct task_struct *task)
3223{
3224 struct perf_comm_event comm_event;
3225
3226 if (task->perf_counter_ctxp)
3227 perf_counter_enable_on_exec(task);
3228
3229 if (!atomic_read(&nr_comm_counters))
3230 return;
3231
3232 comm_event = (struct perf_comm_event){
3233 .task = task,
3234 /* .comm */
3235 /* .comm_size */
3236 .event = {
3237 .header = {
3238 .type = PERF_EVENT_COMM,
3239 .misc = 0,
3240 /* .size */
3241 },
3242 /* .pid */
3243 /* .tid */
3244 },
3245 };
3246
3247 perf_counter_comm_event(&comm_event);
3248}
3249
3250/*
3251 * mmap tracking
3252 */
3253
3254struct perf_mmap_event {
3255 struct vm_area_struct *vma;
3256
3257 const char *file_name;
3258 int file_size;
3259
3260 struct {
3261 struct perf_event_header header;
3262
3263 u32 pid;
3264 u32 tid;
3265 u64 start;
3266 u64 len;
3267 u64 pgoff;
3268 } event;
3269};
3270
3271static void perf_counter_mmap_output(struct perf_counter *counter,
3272 struct perf_mmap_event *mmap_event)
3273{
3274 struct perf_output_handle handle;
3275 int size = mmap_event->event.header.size;
3276 int ret = perf_output_begin(&handle, counter, size, 0, 0);
3277
3278 if (ret)
3279 return;
3280
3281 mmap_event->event.pid = perf_counter_pid(counter, current);
3282 mmap_event->event.tid = perf_counter_tid(counter, current);
3283
3284 perf_output_put(&handle, mmap_event->event);
3285 perf_output_copy(&handle, mmap_event->file_name,
3286 mmap_event->file_size);
3287 perf_output_end(&handle);
3288}
3289
3290static int perf_counter_mmap_match(struct perf_counter *counter,
3291 struct perf_mmap_event *mmap_event)
3292{
3293 if (counter->attr.mmap)
3294 return 1;
3295
3296 return 0;
3297}
3298
3299static void perf_counter_mmap_ctx(struct perf_counter_context *ctx,
3300 struct perf_mmap_event *mmap_event)
3301{
3302 struct perf_counter *counter;
3303
3304 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3305 return;
3306
3307 rcu_read_lock();
3308 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3309 if (perf_counter_mmap_match(counter, mmap_event))
3310 perf_counter_mmap_output(counter, mmap_event);
3311 }
3312 rcu_read_unlock();
3313}
3314
3315static void perf_counter_mmap_event(struct perf_mmap_event *mmap_event)
3316{
3317 struct perf_cpu_context *cpuctx;
3318 struct perf_counter_context *ctx;
3319 struct vm_area_struct *vma = mmap_event->vma;
3320 struct file *file = vma->vm_file;
3321 unsigned int size;
3322 char tmp[16];
3323 char *buf = NULL;
3324 const char *name;
3325
3326 memset(tmp, 0, sizeof(tmp));
3327
3328 if (file) {
3329 /*
3330 * d_path works from the end of the buffer backwards, so we
3331 * need to add enough zero bytes after the string to handle
3332 * the 64bit alignment we do later.
3333 */
3334 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3335 if (!buf) {
3336 name = strncpy(tmp, "//enomem", sizeof(tmp));
3337 goto got_name;
3338 }
3339 name = d_path(&file->f_path, buf, PATH_MAX);
3340 if (IS_ERR(name)) {
3341 name = strncpy(tmp, "//toolong", sizeof(tmp));
3342 goto got_name;
3343 }
3344 } else {
3345 if (arch_vma_name(mmap_event->vma)) {
3346 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3347 sizeof(tmp));
3348 goto got_name;
3349 }
3350
3351 if (!vma->vm_mm) {
3352 name = strncpy(tmp, "[vdso]", sizeof(tmp));
3353 goto got_name;
3354 }
3355
3356 name = strncpy(tmp, "//anon", sizeof(tmp));
3357 goto got_name;
3358 }
3359
3360got_name:
3361 size = ALIGN(strlen(name)+1, sizeof(u64));
3362
3363 mmap_event->file_name = name;
3364 mmap_event->file_size = size;
3365
3366 mmap_event->event.header.size = sizeof(mmap_event->event) + size;
3367
3368 cpuctx = &get_cpu_var(perf_cpu_context);
3369 perf_counter_mmap_ctx(&cpuctx->ctx, mmap_event);
3370 put_cpu_var(perf_cpu_context);
3371
3372 rcu_read_lock();
3373 /*
3374 * doesn't really matter which of the child contexts the
3375 * events ends up in.
3376 */
3377 ctx = rcu_dereference(current->perf_counter_ctxp);
3378 if (ctx)
3379 perf_counter_mmap_ctx(ctx, mmap_event);
3380 rcu_read_unlock();
3381
3382 kfree(buf);
3383}
3384
3385void __perf_counter_mmap(struct vm_area_struct *vma)
3386{
3387 struct perf_mmap_event mmap_event;
3388
3389 if (!atomic_read(&nr_mmap_counters))
3390 return;
3391
3392 mmap_event = (struct perf_mmap_event){
3393 .vma = vma,
3394 /* .file_name */
3395 /* .file_size */
3396 .event = {
3397 .header = {
3398 .type = PERF_EVENT_MMAP,
3399 .misc = 0,
3400 /* .size */
3401 },
3402 /* .pid */
3403 /* .tid */
3404 .start = vma->vm_start,
3405 .len = vma->vm_end - vma->vm_start,
3406 .pgoff = vma->vm_pgoff,
3407 },
3408 };
3409
3410 perf_counter_mmap_event(&mmap_event);
3411}
3412
3413/*
3414 * IRQ throttle logging
3415 */
3416
3417static void perf_log_throttle(struct perf_counter *counter, int enable)
3418{
3419 struct perf_output_handle handle;
3420 int ret;
3421
3422 struct {
3423 struct perf_event_header header;
3424 u64 time;
3425 u64 id;
3426 u64 stream_id;
3427 } throttle_event = {
3428 .header = {
3429 .type = PERF_EVENT_THROTTLE,
3430 .misc = 0,
3431 .size = sizeof(throttle_event),
3432 },
3433 .time = sched_clock(),
3434 .id = primary_counter_id(counter),
3435 .stream_id = counter->id,
3436 };
3437
3438 if (enable)
3439 throttle_event.header.type = PERF_EVENT_UNTHROTTLE;
3440
3441 ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
3442 if (ret)
3443 return;
3444
3445 perf_output_put(&handle, throttle_event);
3446 perf_output_end(&handle);
3447}
3448
3449/*
3450 * Generic counter overflow handling, sampling.
3451 */
3452
3453int perf_counter_overflow(struct perf_counter *counter, int nmi,
3454 struct perf_sample_data *data)
3455{
3456 int events = atomic_read(&counter->event_limit);
3457 int throttle = counter->pmu->unthrottle != NULL;
3458 struct hw_perf_counter *hwc = &counter->hw;
3459 int ret = 0;
3460
3461 if (!throttle) {
3462 hwc->interrupts++;
3463 } else {
3464 if (hwc->interrupts != MAX_INTERRUPTS) {
3465 hwc->interrupts++;
3466 if (HZ * hwc->interrupts >
3467 (u64)sysctl_perf_counter_sample_rate) {
3468 hwc->interrupts = MAX_INTERRUPTS;
3469 perf_log_throttle(counter, 0);
3470 ret = 1;
3471 }
3472 } else {
3473 /*
3474 * Keep re-disabling counters even though on the previous
3475 * pass we disabled it - just in case we raced with a
3476 * sched-in and the counter got enabled again:
3477 */
3478 ret = 1;
3479 }
3480 }
3481
3482 if (counter->attr.freq) {
3483 u64 now = sched_clock();
3484 s64 delta = now - hwc->freq_stamp;
3485
3486 hwc->freq_stamp = now;
3487
3488 if (delta > 0 && delta < TICK_NSEC)
3489 perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
3490 }
3491
3492 /*
3493 * XXX event_limit might not quite work as expected on inherited
3494 * counters
3495 */
3496
3497 counter->pending_kill = POLL_IN;
3498 if (events && atomic_dec_and_test(&counter->event_limit)) {
3499 ret = 1;
3500 counter->pending_kill = POLL_HUP;
3501 if (nmi) {
3502 counter->pending_disable = 1;
3503 perf_pending_queue(&counter->pending,
3504 perf_pending_counter);
3505 } else
3506 perf_counter_disable(counter);
3507 }
3508
3509 perf_counter_output(counter, nmi, data);
3510 return ret;
3511}
3512
3513/*
3514 * Generic software counter infrastructure
3515 */
3516
3517/*
3518 * We directly increment counter->count and keep a second value in
3519 * counter->hw.period_left to count intervals. This period counter
3520 * is kept in the range [-sample_period, 0] so that we can use the
3521 * sign as trigger.
3522 */
3523
3524static u64 perf_swcounter_set_period(struct perf_counter *counter)
3525{
3526 struct hw_perf_counter *hwc = &counter->hw;
3527 u64 period = hwc->last_period;
3528 u64 nr, offset;
3529 s64 old, val;
3530
3531 hwc->last_period = hwc->sample_period;
3532
3533again:
3534 old = val = atomic64_read(&hwc->period_left);
3535 if (val < 0)
3536 return 0;
3537
3538 nr = div64_u64(period + val, period);
3539 offset = nr * period;
3540 val -= offset;
3541 if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3542 goto again;
3543
3544 return nr;
3545}
3546
3547static void perf_swcounter_overflow(struct perf_counter *counter,
3548 int nmi, struct perf_sample_data *data)
3549{
3550 struct hw_perf_counter *hwc = &counter->hw;
3551 u64 overflow;
3552
3553 data->period = counter->hw.last_period;
3554 overflow = perf_swcounter_set_period(counter);
3555
3556 if (hwc->interrupts == MAX_INTERRUPTS)
3557 return;
3558
3559 for (; overflow; overflow--) {
3560 if (perf_counter_overflow(counter, nmi, data)) {
3561 /*
3562 * We inhibit the overflow from happening when
3563 * hwc->interrupts == MAX_INTERRUPTS.
3564 */
3565 break;
3566 }
3567 }
3568}
3569
3570static void perf_swcounter_unthrottle(struct perf_counter *counter)
3571{
3572 /*
3573 * Nothing to do, we already reset hwc->interrupts.
3574 */
3575}
3576
3577static void perf_swcounter_add(struct perf_counter *counter, u64 nr,
3578 int nmi, struct perf_sample_data *data)
3579{
3580 struct hw_perf_counter *hwc = &counter->hw;
3581
3582 atomic64_add(nr, &counter->count);
3583
3584 if (!hwc->sample_period)
3585 return;
3586
3587 if (!data->regs)
3588 return;
3589
3590 if (!atomic64_add_negative(nr, &hwc->period_left))
3591 perf_swcounter_overflow(counter, nmi, data);
3592}
3593
3594static int perf_swcounter_is_counting(struct perf_counter *counter)
3595{
3596 /*
3597 * The counter is active, we're good!
3598 */
3599 if (counter->state == PERF_COUNTER_STATE_ACTIVE)
3600 return 1;
3601
3602 /*
3603 * The counter is off/error, not counting.
3604 */
3605 if (counter->state != PERF_COUNTER_STATE_INACTIVE)
3606 return 0;
3607
3608 /*
3609 * The counter is inactive, if the context is active
3610 * we're part of a group that didn't make it on the 'pmu',
3611 * not counting.
3612 */
3613 if (counter->ctx->is_active)
3614 return 0;
3615
3616 /*
3617 * We're inactive and the context is too, this means the
3618 * task is scheduled out, we're counting events that happen
3619 * to us, like migration events.
3620 */
3621 return 1;
3622}
3623
3624static int perf_swcounter_match(struct perf_counter *counter,
3625 enum perf_type_id type,
3626 u32 event, struct pt_regs *regs)
3627{
3628 if (!perf_swcounter_is_counting(counter))
3629 return 0;
3630
3631 if (counter->attr.type != type)
3632 return 0;
3633 if (counter->attr.config != event)
3634 return 0;
3635
3636 if (regs) {
3637 if (counter->attr.exclude_user && user_mode(regs))
3638 return 0;
3639
3640 if (counter->attr.exclude_kernel && !user_mode(regs))
3641 return 0;
3642 }
3643
3644 return 1;
3645}
3646
3647static void perf_swcounter_ctx_event(struct perf_counter_context *ctx,
3648 enum perf_type_id type,
3649 u32 event, u64 nr, int nmi,
3650 struct perf_sample_data *data)
3651{
3652 struct perf_counter *counter;
3653
3654 if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3655 return;
3656
3657 rcu_read_lock();
3658 list_for_each_entry_rcu(counter, &ctx->event_list, event_entry) {
3659 if (perf_swcounter_match(counter, type, event, data->regs))
3660 perf_swcounter_add(counter, nr, nmi, data);
3661 }
3662 rcu_read_unlock();
3663}
3664
3665static int *perf_swcounter_recursion_context(struct perf_cpu_context *cpuctx)
3666{
3667 if (in_nmi())
3668 return &cpuctx->recursion[3];
3669
3670 if (in_irq())
3671 return &cpuctx->recursion[2];
3672
3673 if (in_softirq())
3674 return &cpuctx->recursion[1];
3675
3676 return &cpuctx->recursion[0];
3677}
3678
3679static void do_perf_swcounter_event(enum perf_type_id type, u32 event,
3680 u64 nr, int nmi,
3681 struct perf_sample_data *data)
3682{
3683 struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3684 int *recursion = perf_swcounter_recursion_context(cpuctx);
3685 struct perf_counter_context *ctx;
3686
3687 if (*recursion)
3688 goto out;
3689
3690 (*recursion)++;
3691 barrier();
3692
3693 perf_swcounter_ctx_event(&cpuctx->ctx, type, event,
3694 nr, nmi, data);
3695 rcu_read_lock();
3696 /*
3697 * doesn't really matter which of the child contexts the
3698 * events ends up in.
3699 */
3700 ctx = rcu_dereference(current->perf_counter_ctxp);
3701 if (ctx)
3702 perf_swcounter_ctx_event(ctx, type, event, nr, nmi, data);
3703 rcu_read_unlock();
3704
3705 barrier();
3706 (*recursion)--;
3707
3708out:
3709 put_cpu_var(perf_cpu_context);
3710}
3711
3712void __perf_swcounter_event(u32 event, u64 nr, int nmi,
3713 struct pt_regs *regs, u64 addr)
3714{
3715 struct perf_sample_data data = {
3716 .regs = regs,
3717 .addr = addr,
3718 };
3719
3720 do_perf_swcounter_event(PERF_TYPE_SOFTWARE, event, nr, nmi, &data);
3721}
3722
3723static void perf_swcounter_read(struct perf_counter *counter)
3724{
3725}
3726
3727static int perf_swcounter_enable(struct perf_counter *counter)
3728{
3729 struct hw_perf_counter *hwc = &counter->hw;
3730
3731 if (hwc->sample_period) {
3732 hwc->last_period = hwc->sample_period;
3733 perf_swcounter_set_period(counter);
3734 }
3735 return 0;
3736}
3737
3738static void perf_swcounter_disable(struct perf_counter *counter)
3739{
3740}
3741
3742static const struct pmu perf_ops_generic = {
3743 .enable = perf_swcounter_enable,
3744 .disable = perf_swcounter_disable,
3745 .read = perf_swcounter_read,
3746 .unthrottle = perf_swcounter_unthrottle,
3747};
3748
3749/*
3750 * hrtimer based swcounter callback
3751 */
3752
3753static enum hrtimer_restart perf_swcounter_hrtimer(struct hrtimer *hrtimer)
3754{
3755 enum hrtimer_restart ret = HRTIMER_RESTART;
3756 struct perf_sample_data data;
3757 struct perf_counter *counter;
3758 u64 period;
3759
3760 counter = container_of(hrtimer, struct perf_counter, hw.hrtimer);
3761 counter->pmu->read(counter);
3762
3763 data.addr = 0;
3764 data.regs = get_irq_regs();
3765 /*
3766 * In case we exclude kernel IPs or are somehow not in interrupt
3767 * context, provide the next best thing, the user IP.
3768 */
3769 if ((counter->attr.exclude_kernel || !data.regs) &&
3770 !counter->attr.exclude_user)
3771 data.regs = task_pt_regs(current);
3772
3773 if (data.regs) {
3774 if (perf_counter_overflow(counter, 0, &data))
3775 ret = HRTIMER_NORESTART;
3776 }
3777
3778 period = max_t(u64, 10000, counter->hw.sample_period);
3779 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3780
3781 return ret;
3782}
3783
3784/*
3785 * Software counter: cpu wall time clock
3786 */
3787
3788static void cpu_clock_perf_counter_update(struct perf_counter *counter)
3789{
3790 int cpu = raw_smp_processor_id();
3791 s64 prev;
3792 u64 now;
3793
3794 now = cpu_clock(cpu);
3795 prev = atomic64_read(&counter->hw.prev_count);
3796 atomic64_set(&counter->hw.prev_count, now);
3797 atomic64_add(now - prev, &counter->count);
3798}
3799
3800static int cpu_clock_perf_counter_enable(struct perf_counter *counter)
3801{
3802 struct hw_perf_counter *hwc = &counter->hw;
3803 int cpu = raw_smp_processor_id();
3804
3805 atomic64_set(&hwc->prev_count, cpu_clock(cpu));
3806 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3807 hwc->hrtimer.function = perf_swcounter_hrtimer;
3808 if (hwc->sample_period) {
3809 u64 period = max_t(u64, 10000, hwc->sample_period);
3810 __hrtimer_start_range_ns(&hwc->hrtimer,
3811 ns_to_ktime(period), 0,
3812 HRTIMER_MODE_REL, 0);
3813 }
3814
3815 return 0;
3816}
3817
3818static void cpu_clock_perf_counter_disable(struct perf_counter *counter)
3819{
3820 if (counter->hw.sample_period)
3821 hrtimer_cancel(&counter->hw.hrtimer);
3822 cpu_clock_perf_counter_update(counter);
3823}
3824
3825static void cpu_clock_perf_counter_read(struct perf_counter *counter)
3826{
3827 cpu_clock_perf_counter_update(counter);
3828}
3829
3830static const struct pmu perf_ops_cpu_clock = {
3831 .enable = cpu_clock_perf_counter_enable,
3832 .disable = cpu_clock_perf_counter_disable,
3833 .read = cpu_clock_perf_counter_read,
3834};
3835
3836/*
3837 * Software counter: task time clock
3838 */
3839
3840static void task_clock_perf_counter_update(struct perf_counter *counter, u64 now)
3841{
3842 u64 prev;
3843 s64 delta;
3844
3845 prev = atomic64_xchg(&counter->hw.prev_count, now);
3846 delta = now - prev;
3847 atomic64_add(delta, &counter->count);
3848}
3849
3850static int task_clock_perf_counter_enable(struct perf_counter *counter)
3851{
3852 struct hw_perf_counter *hwc = &counter->hw;
3853 u64 now;
3854
3855 now = counter->ctx->time;
3856
3857 atomic64_set(&hwc->prev_count, now);
3858 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3859 hwc->hrtimer.function = perf_swcounter_hrtimer;
3860 if (hwc->sample_period) {
3861 u64 period = max_t(u64, 10000, hwc->sample_period);
3862 __hrtimer_start_range_ns(&hwc->hrtimer,
3863 ns_to_ktime(period), 0,
3864 HRTIMER_MODE_REL, 0);
3865 }
3866
3867 return 0;
3868}
3869
3870static void task_clock_perf_counter_disable(struct perf_counter *counter)
3871{
3872 if (counter->hw.sample_period)
3873 hrtimer_cancel(&counter->hw.hrtimer);
3874 task_clock_perf_counter_update(counter, counter->ctx->time);
3875
3876}
3877
3878static void task_clock_perf_counter_read(struct perf_counter *counter)
3879{
3880 u64 time;
3881
3882 if (!in_nmi()) {
3883 update_context_time(counter->ctx);
3884 time = counter->ctx->time;
3885 } else {
3886 u64 now = perf_clock();
3887 u64 delta = now - counter->ctx->timestamp;
3888 time = counter->ctx->time + delta;
3889 }
3890
3891 task_clock_perf_counter_update(counter, time);
3892}
3893
3894static const struct pmu perf_ops_task_clock = {
3895 .enable = task_clock_perf_counter_enable,
3896 .disable = task_clock_perf_counter_disable,
3897 .read = task_clock_perf_counter_read,
3898};
3899
3900#ifdef CONFIG_EVENT_PROFILE
3901void perf_tpcounter_event(int event_id, u64 addr, u64 count, void *record,
3902 int entry_size)
3903{
3904 struct perf_raw_record raw = {
3905 .size = entry_size,
3906 .data = record,
3907 };
3908
3909 struct perf_sample_data data = {
3910 .regs = get_irq_regs(),
3911 .addr = addr,
3912 .raw = &raw,
3913 };
3914
3915 if (!data.regs)
3916 data.regs = task_pt_regs(current);
3917
3918 do_perf_swcounter_event(PERF_TYPE_TRACEPOINT, event_id, count, 1, &data);
3919}
3920EXPORT_SYMBOL_GPL(perf_tpcounter_event);
3921
3922extern int ftrace_profile_enable(int);
3923extern void ftrace_profile_disable(int);
3924
3925static void tp_perf_counter_destroy(struct perf_counter *counter)
3926{
3927 ftrace_profile_disable(counter->attr.config);
3928}
3929
3930static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3931{
3932 /*
3933 * Raw tracepoint data is a severe data leak, only allow root to
3934 * have these.
3935 */
3936 if ((counter->attr.sample_type & PERF_SAMPLE_RAW) &&
3937 !capable(CAP_SYS_ADMIN))
3938 return ERR_PTR(-EPERM);
3939
3940 if (ftrace_profile_enable(counter->attr.config))
3941 return NULL;
3942
3943 counter->destroy = tp_perf_counter_destroy;
3944
3945 return &perf_ops_generic;
3946}
3947#else
3948static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
3949{
3950 return NULL;
3951}
3952#endif
3953
3954atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
3955
3956static void sw_perf_counter_destroy(struct perf_counter *counter)
3957{
3958 u64 event = counter->attr.config;
3959
3960 WARN_ON(counter->parent);
3961
3962 atomic_dec(&perf_swcounter_enabled[event]);
3963}
3964
3965static const struct pmu *sw_perf_counter_init(struct perf_counter *counter)
3966{
3967 const struct pmu *pmu = NULL;
3968 u64 event = counter->attr.config;
3969
3970 /*
3971 * Software counters (currently) can't in general distinguish
3972 * between user, kernel and hypervisor events.
3973 * However, context switches and cpu migrations are considered
3974 * to be kernel events, and page faults are never hypervisor
3975 * events.
3976 */
3977 switch (event) {
3978 case PERF_COUNT_SW_CPU_CLOCK:
3979 pmu = &perf_ops_cpu_clock;
3980
3981 break;
3982 case PERF_COUNT_SW_TASK_CLOCK:
3983 /*
3984 * If the user instantiates this as a per-cpu counter,
3985 * use the cpu_clock counter instead.
3986 */
3987 if (counter->ctx->task)
3988 pmu = &perf_ops_task_clock;
3989 else
3990 pmu = &perf_ops_cpu_clock;
3991
3992 break;
3993 case PERF_COUNT_SW_PAGE_FAULTS:
3994 case PERF_COUNT_SW_PAGE_FAULTS_MIN:
3995 case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
3996 case PERF_COUNT_SW_CONTEXT_SWITCHES:
3997 case PERF_COUNT_SW_CPU_MIGRATIONS:
3998 if (!counter->parent) {
3999 atomic_inc(&perf_swcounter_enabled[event]);
4000 counter->destroy = sw_perf_counter_destroy;
4001 }
4002 pmu = &perf_ops_generic;
4003 break;
4004 }
4005
4006 return pmu;
4007}
4008
4009/*
4010 * Allocate and initialize a counter structure
4011 */
4012static struct perf_counter *
4013perf_counter_alloc(struct perf_counter_attr *attr,
4014 int cpu,
4015 struct perf_counter_context *ctx,
4016 struct perf_counter *group_leader,
4017 struct perf_counter *parent_counter,
4018 gfp_t gfpflags)
4019{
4020 const struct pmu *pmu;
4021 struct perf_counter *counter;
4022 struct hw_perf_counter *hwc;
4023 long err;
4024
4025 counter = kzalloc(sizeof(*counter), gfpflags);
4026 if (!counter)
4027 return ERR_PTR(-ENOMEM);
4028
4029 /*
4030 * Single counters are their own group leaders, with an
4031 * empty sibling list:
4032 */
4033 if (!group_leader)
4034 group_leader = counter;
4035
4036 mutex_init(&counter->child_mutex);
4037 INIT_LIST_HEAD(&counter->child_list);
4038
4039 INIT_LIST_HEAD(&counter->list_entry);
4040 INIT_LIST_HEAD(&counter->event_entry);
4041 INIT_LIST_HEAD(&counter->sibling_list);
4042 init_waitqueue_head(&counter->waitq);
4043
4044 mutex_init(&counter->mmap_mutex);
4045
4046 counter->cpu = cpu;
4047 counter->attr = *attr;
4048 counter->group_leader = group_leader;
4049 counter->pmu = NULL;
4050 counter->ctx = ctx;
4051 counter->oncpu = -1;
4052
4053 counter->parent = parent_counter;
4054
4055 counter->ns = get_pid_ns(current->nsproxy->pid_ns);
4056 counter->id = atomic64_inc_return(&perf_counter_id);
4057
4058 counter->state = PERF_COUNTER_STATE_INACTIVE;
4059
4060 if (attr->disabled)
4061 counter->state = PERF_COUNTER_STATE_OFF;
4062
4063 pmu = NULL;
4064
4065 hwc = &counter->hw;
4066 hwc->sample_period = attr->sample_period;
4067 if (attr->freq && attr->sample_freq)
4068 hwc->sample_period = 1;
4069 hwc->last_period = hwc->sample_period;
4070
4071 atomic64_set(&hwc->period_left, hwc->sample_period);
4072
4073 /*
4074 * we currently do not support PERF_FORMAT_GROUP on inherited counters
4075 */
4076 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4077 goto done;
4078
4079 switch (attr->type) {
4080 case PERF_TYPE_RAW:
4081 case PERF_TYPE_HARDWARE:
4082 case PERF_TYPE_HW_CACHE:
4083 pmu = hw_perf_counter_init(counter);
4084 break;
4085
4086 case PERF_TYPE_SOFTWARE:
4087 pmu = sw_perf_counter_init(counter);
4088 break;
4089
4090 case PERF_TYPE_TRACEPOINT:
4091 pmu = tp_perf_counter_init(counter);
4092 break;
4093
4094 default:
4095 break;
4096 }
4097done:
4098 err = 0;
4099 if (!pmu)
4100 err = -EINVAL;
4101 else if (IS_ERR(pmu))
4102 err = PTR_ERR(pmu);
4103
4104 if (err) {
4105 if (counter->ns)
4106 put_pid_ns(counter->ns);
4107 kfree(counter);
4108 return ERR_PTR(err);
4109 }
4110
4111 counter->pmu = pmu;
4112
4113 if (!counter->parent) {
4114 atomic_inc(&nr_counters);
4115 if (counter->attr.mmap)
4116 atomic_inc(&nr_mmap_counters);
4117 if (counter->attr.comm)
4118 atomic_inc(&nr_comm_counters);
4119 if (counter->attr.task)
4120 atomic_inc(&nr_task_counters);
4121 }
4122
4123 return counter;
4124}
4125
4126static int perf_copy_attr(struct perf_counter_attr __user *uattr,
4127 struct perf_counter_attr *attr)
4128{
4129 int ret;
4130 u32 size;
4131
4132 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4133 return -EFAULT;
4134
4135 /*
4136 * zero the full structure, so that a short copy will be nice.
4137 */
4138 memset(attr, 0, sizeof(*attr));
4139
4140 ret = get_user(size, &uattr->size);
4141 if (ret)
4142 return ret;
4143
4144 if (size > PAGE_SIZE) /* silly large */
4145 goto err_size;
4146
4147 if (!size) /* abi compat */
4148 size = PERF_ATTR_SIZE_VER0;
4149
4150 if (size < PERF_ATTR_SIZE_VER0)
4151 goto err_size;
4152
4153 /*
4154 * If we're handed a bigger struct than we know of,
4155 * ensure all the unknown bits are 0.
4156 */
4157 if (size > sizeof(*attr)) {
4158 unsigned long val;
4159 unsigned long __user *addr;
4160 unsigned long __user *end;
4161
4162 addr = PTR_ALIGN((void __user *)uattr + sizeof(*attr),
4163 sizeof(unsigned long));
4164 end = PTR_ALIGN((void __user *)uattr + size,
4165 sizeof(unsigned long));
4166
4167 for (; addr < end; addr += sizeof(unsigned long)) {
4168 ret = get_user(val, addr);
4169 if (ret)
4170 return ret;
4171 if (val)
4172 goto err_size;
4173 }
4174 }
4175
4176 ret = copy_from_user(attr, uattr, size);
4177 if (ret)
4178 return -EFAULT;
4179
4180 /*
4181 * If the type exists, the corresponding creation will verify
4182 * the attr->config.
4183 */
4184 if (attr->type >= PERF_TYPE_MAX)
4185 return -EINVAL;
4186
4187 if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4188 return -EINVAL;
4189
4190 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4191 return -EINVAL;
4192
4193 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4194 return -EINVAL;
4195
4196out:
4197 return ret;
4198
4199err_size:
4200 put_user(sizeof(*attr), &uattr->size);
4201 ret = -E2BIG;
4202 goto out;
4203}
4204
4205/**
4206 * sys_perf_counter_open - open a performance counter, associate it to a task/cpu
4207 *
4208 * @attr_uptr: event type attributes for monitoring/sampling
4209 * @pid: target pid
4210 * @cpu: target cpu
4211 * @group_fd: group leader counter fd
4212 */
4213SYSCALL_DEFINE5(perf_counter_open,
4214 struct perf_counter_attr __user *, attr_uptr,
4215 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4216{
4217 struct perf_counter *counter, *group_leader;
4218 struct perf_counter_attr attr;
4219 struct perf_counter_context *ctx;
4220 struct file *counter_file = NULL;
4221 struct file *group_file = NULL;
4222 int fput_needed = 0;
4223 int fput_needed2 = 0;
4224 int ret;
4225
4226 /* for future expandability... */
4227 if (flags)
4228 return -EINVAL;
4229
4230 ret = perf_copy_attr(attr_uptr, &attr);
4231 if (ret)
4232 return ret;
4233
4234 if (!attr.exclude_kernel) {
4235 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4236 return -EACCES;
4237 }
4238
4239 if (attr.freq) {
4240 if (attr.sample_freq > sysctl_perf_counter_sample_rate)
4241 return -EINVAL;
4242 }
4243
4244 /*
4245 * Get the target context (task or percpu):
4246 */
4247 ctx = find_get_context(pid, cpu);
4248 if (IS_ERR(ctx))
4249 return PTR_ERR(ctx);
4250
4251 /*
4252 * Look up the group leader (we will attach this counter to it):
4253 */
4254 group_leader = NULL;
4255 if (group_fd != -1) {
4256 ret = -EINVAL;
4257 group_file = fget_light(group_fd, &fput_needed);
4258 if (!group_file)
4259 goto err_put_context;
4260 if (group_file->f_op != &perf_fops)
4261 goto err_put_context;
4262
4263 group_leader = group_file->private_data;
4264 /*
4265 * Do not allow a recursive hierarchy (this new sibling
4266 * becoming part of another group-sibling):
4267 */
4268 if (group_leader->group_leader != group_leader)
4269 goto err_put_context;
4270 /*
4271 * Do not allow to attach to a group in a different
4272 * task or CPU context:
4273 */
4274 if (group_leader->ctx != ctx)
4275 goto err_put_context;
4276 /*
4277 * Only a group leader can be exclusive or pinned
4278 */
4279 if (attr.exclusive || attr.pinned)
4280 goto err_put_context;
4281 }
4282
4283 counter = perf_counter_alloc(&attr, cpu, ctx, group_leader,
4284 NULL, GFP_KERNEL);
4285 ret = PTR_ERR(counter);
4286 if (IS_ERR(counter))
4287 goto err_put_context;
4288
4289 ret = anon_inode_getfd("[perf_counter]", &perf_fops, counter, 0);
4290 if (ret < 0)
4291 goto err_free_put_context;
4292
4293 counter_file = fget_light(ret, &fput_needed2);
4294 if (!counter_file)
4295 goto err_free_put_context;
4296
4297 counter->filp = counter_file;
4298 WARN_ON_ONCE(ctx->parent_ctx);
4299 mutex_lock(&ctx->mutex);
4300 perf_install_in_context(ctx, counter, cpu);
4301 ++ctx->generation;
4302 mutex_unlock(&ctx->mutex);
4303
4304 counter->owner = current;
4305 get_task_struct(current);
4306 mutex_lock(&current->perf_counter_mutex);
4307 list_add_tail(&counter->owner_entry, &current->perf_counter_list);
4308 mutex_unlock(&current->perf_counter_mutex);
4309
4310 fput_light(counter_file, fput_needed2);
4311
4312out_fput:
4313 fput_light(group_file, fput_needed);
4314
4315 return ret;
4316
4317err_free_put_context:
4318 kfree(counter);
4319
4320err_put_context:
4321 put_ctx(ctx);
4322
4323 goto out_fput;
4324}
4325
4326/*
4327 * inherit a counter from parent task to child task:
4328 */
4329static struct perf_counter *
4330inherit_counter(struct perf_counter *parent_counter,
4331 struct task_struct *parent,
4332 struct perf_counter_context *parent_ctx,
4333 struct task_struct *child,
4334 struct perf_counter *group_leader,
4335 struct perf_counter_context *child_ctx)
4336{
4337 struct perf_counter *child_counter;
4338
4339 /*
4340 * Instead of creating recursive hierarchies of counters,
4341 * we link inherited counters back to the original parent,
4342 * which has a filp for sure, which we use as the reference
4343 * count:
4344 */
4345 if (parent_counter->parent)
4346 parent_counter = parent_counter->parent;
4347
4348 child_counter = perf_counter_alloc(&parent_counter->attr,
4349 parent_counter->cpu, child_ctx,
4350 group_leader, parent_counter,
4351 GFP_KERNEL);
4352 if (IS_ERR(child_counter))
4353 return child_counter;
4354 get_ctx(child_ctx);
4355
4356 /*
4357 * Make the child state follow the state of the parent counter,
4358 * not its attr.disabled bit. We hold the parent's mutex,
4359 * so we won't race with perf_counter_{en, dis}able_family.
4360 */
4361 if (parent_counter->state >= PERF_COUNTER_STATE_INACTIVE)
4362 child_counter->state = PERF_COUNTER_STATE_INACTIVE;
4363 else
4364 child_counter->state = PERF_COUNTER_STATE_OFF;
4365
4366 if (parent_counter->attr.freq)
4367 child_counter->hw.sample_period = parent_counter->hw.sample_period;
4368
4369 /*
4370 * Link it up in the child's context:
4371 */
4372 add_counter_to_ctx(child_counter, child_ctx);
4373
4374 /*
4375 * Get a reference to the parent filp - we will fput it
4376 * when the child counter exits. This is safe to do because
4377 * we are in the parent and we know that the filp still
4378 * exists and has a nonzero count:
4379 */
4380 atomic_long_inc(&parent_counter->filp->f_count);
4381
4382 /*
4383 * Link this into the parent counter's child list
4384 */
4385 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4386 mutex_lock(&parent_counter->child_mutex);
4387 list_add_tail(&child_counter->child_list, &parent_counter->child_list);
4388 mutex_unlock(&parent_counter->child_mutex);
4389
4390 return child_counter;
4391}
4392
4393static int inherit_group(struct perf_counter *parent_counter,
4394 struct task_struct *parent,
4395 struct perf_counter_context *parent_ctx,
4396 struct task_struct *child,
4397 struct perf_counter_context *child_ctx)
4398{
4399 struct perf_counter *leader;
4400 struct perf_counter *sub;
4401 struct perf_counter *child_ctr;
4402
4403 leader = inherit_counter(parent_counter, parent, parent_ctx,
4404 child, NULL, child_ctx);
4405 if (IS_ERR(leader))
4406 return PTR_ERR(leader);
4407 list_for_each_entry(sub, &parent_counter->sibling_list, list_entry) {
4408 child_ctr = inherit_counter(sub, parent, parent_ctx,
4409 child, leader, child_ctx);
4410 if (IS_ERR(child_ctr))
4411 return PTR_ERR(child_ctr);
4412 }
4413 return 0;
4414}
4415
4416static void sync_child_counter(struct perf_counter *child_counter,
4417 struct task_struct *child)
4418{
4419 struct perf_counter *parent_counter = child_counter->parent;
4420 u64 child_val;
4421
4422 if (child_counter->attr.inherit_stat)
4423 perf_counter_read_event(child_counter, child);
4424
4425 child_val = atomic64_read(&child_counter->count);
4426
4427 /*
4428 * Add back the child's count to the parent's count:
4429 */
4430 atomic64_add(child_val, &parent_counter->count);
4431 atomic64_add(child_counter->total_time_enabled,
4432 &parent_counter->child_total_time_enabled);
4433 atomic64_add(child_counter->total_time_running,
4434 &parent_counter->child_total_time_running);
4435
4436 /*
4437 * Remove this counter from the parent's list
4438 */
4439 WARN_ON_ONCE(parent_counter->ctx->parent_ctx);
4440 mutex_lock(&parent_counter->child_mutex);
4441 list_del_init(&child_counter->child_list);
4442 mutex_unlock(&parent_counter->child_mutex);
4443
4444 /*
4445 * Release the parent counter, if this was the last
4446 * reference to it.
4447 */
4448 fput(parent_counter->filp);
4449}
4450
4451static void
4452__perf_counter_exit_task(struct perf_counter *child_counter,
4453 struct perf_counter_context *child_ctx,
4454 struct task_struct *child)
4455{
4456 struct perf_counter *parent_counter;
4457
4458 update_counter_times(child_counter);
4459 perf_counter_remove_from_context(child_counter);
4460
4461 parent_counter = child_counter->parent;
4462 /*
4463 * It can happen that parent exits first, and has counters
4464 * that are still around due to the child reference. These
4465 * counters need to be zapped - but otherwise linger.
4466 */
4467 if (parent_counter) {
4468 sync_child_counter(child_counter, child);
4469 free_counter(child_counter);
4470 }
4471}
4472
4473/*
4474 * When a child task exits, feed back counter values to parent counters.
4475 */
4476void perf_counter_exit_task(struct task_struct *child)
4477{
4478 struct perf_counter *child_counter, *tmp;
4479 struct perf_counter_context *child_ctx;
4480 unsigned long flags;
4481
4482 if (likely(!child->perf_counter_ctxp)) {
4483 perf_counter_task(child, NULL, 0);
4484 return;
4485 }
4486
4487 local_irq_save(flags);
4488 /*
4489 * We can't reschedule here because interrupts are disabled,
4490 * and either child is current or it is a task that can't be
4491 * scheduled, so we are now safe from rescheduling changing
4492 * our context.
4493 */
4494 child_ctx = child->perf_counter_ctxp;
4495 __perf_counter_task_sched_out(child_ctx);
4496
4497 /*
4498 * Take the context lock here so that if find_get_context is
4499 * reading child->perf_counter_ctxp, we wait until it has
4500 * incremented the context's refcount before we do put_ctx below.
4501 */
4502 spin_lock(&child_ctx->lock);
4503 child->perf_counter_ctxp = NULL;
4504 /*
4505 * If this context is a clone; unclone it so it can't get
4506 * swapped to another process while we're removing all
4507 * the counters from it.
4508 */
4509 unclone_ctx(child_ctx);
4510 spin_unlock_irqrestore(&child_ctx->lock, flags);
4511
4512 /*
4513 * Report the task dead after unscheduling the counters so that we
4514 * won't get any samples after PERF_EVENT_EXIT. We can however still
4515 * get a few PERF_EVENT_READ events.
4516 */
4517 perf_counter_task(child, child_ctx, 0);
4518
4519 /*
4520 * We can recurse on the same lock type through:
4521 *
4522 * __perf_counter_exit_task()
4523 * sync_child_counter()
4524 * fput(parent_counter->filp)
4525 * perf_release()
4526 * mutex_lock(&ctx->mutex)
4527 *
4528 * But since its the parent context it won't be the same instance.
4529 */
4530 mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4531
4532again:
4533 list_for_each_entry_safe(child_counter, tmp, &child_ctx->counter_list,
4534 list_entry)
4535 __perf_counter_exit_task(child_counter, child_ctx, child);
4536
4537 /*
4538 * If the last counter was a group counter, it will have appended all
4539 * its siblings to the list, but we obtained 'tmp' before that which
4540 * will still point to the list head terminating the iteration.
4541 */
4542 if (!list_empty(&child_ctx->counter_list))
4543 goto again;
4544
4545 mutex_unlock(&child_ctx->mutex);
4546
4547 put_ctx(child_ctx);
4548}
4549
4550/*
4551 * free an unexposed, unused context as created by inheritance by
4552 * init_task below, used by fork() in case of fail.
4553 */
4554void perf_counter_free_task(struct task_struct *task)
4555{
4556 struct perf_counter_context *ctx = task->perf_counter_ctxp;
4557 struct perf_counter *counter, *tmp;
4558
4559 if (!ctx)
4560 return;
4561
4562 mutex_lock(&ctx->mutex);
4563again:
4564 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry) {
4565 struct perf_counter *parent = counter->parent;
4566
4567 if (WARN_ON_ONCE(!parent))
4568 continue;
4569
4570 mutex_lock(&parent->child_mutex);
4571 list_del_init(&counter->child_list);
4572 mutex_unlock(&parent->child_mutex);
4573
4574 fput(parent->filp);
4575
4576 list_del_counter(counter, ctx);
4577 free_counter(counter);
4578 }
4579
4580 if (!list_empty(&ctx->counter_list))
4581 goto again;
4582
4583 mutex_unlock(&ctx->mutex);
4584
4585 put_ctx(ctx);
4586}
4587
4588/*
4589 * Initialize the perf_counter context in task_struct
4590 */
4591int perf_counter_init_task(struct task_struct *child)
4592{
4593 struct perf_counter_context *child_ctx, *parent_ctx;
4594 struct perf_counter_context *cloned_ctx;
4595 struct perf_counter *counter;
4596 struct task_struct *parent = current;
4597 int inherited_all = 1;
4598 int ret = 0;
4599
4600 child->perf_counter_ctxp = NULL;
4601
4602 mutex_init(&child->perf_counter_mutex);
4603 INIT_LIST_HEAD(&child->perf_counter_list);
4604
4605 if (likely(!parent->perf_counter_ctxp))
4606 return 0;
4607
4608 /*
4609 * This is executed from the parent task context, so inherit
4610 * counters that have been marked for cloning.
4611 * First allocate and initialize a context for the child.
4612 */
4613
4614 child_ctx = kmalloc(sizeof(struct perf_counter_context), GFP_KERNEL);
4615 if (!child_ctx)
4616 return -ENOMEM;
4617
4618 __perf_counter_init_context(child_ctx, child);
4619 child->perf_counter_ctxp = child_ctx;
4620 get_task_struct(child);
4621
4622 /*
4623 * If the parent's context is a clone, pin it so it won't get
4624 * swapped under us.
4625 */
4626 parent_ctx = perf_pin_task_context(parent);
4627
4628 /*
4629 * No need to check if parent_ctx != NULL here; since we saw
4630 * it non-NULL earlier, the only reason for it to become NULL
4631 * is if we exit, and since we're currently in the middle of
4632 * a fork we can't be exiting at the same time.
4633 */
4634
4635 /*
4636 * Lock the parent list. No need to lock the child - not PID
4637 * hashed yet and not running, so nobody can access it.
4638 */
4639 mutex_lock(&parent_ctx->mutex);
4640
4641 /*
4642 * We dont have to disable NMIs - we are only looking at
4643 * the list, not manipulating it:
4644 */
4645 list_for_each_entry_rcu(counter, &parent_ctx->event_list, event_entry) {
4646 if (counter != counter->group_leader)
4647 continue;
4648
4649 if (!counter->attr.inherit) {
4650 inherited_all = 0;
4651 continue;
4652 }
4653
4654 ret = inherit_group(counter, parent, parent_ctx,
4655 child, child_ctx);
4656 if (ret) {
4657 inherited_all = 0;
4658 break;
4659 }
4660 }
4661
4662 if (inherited_all) {
4663 /*
4664 * Mark the child context as a clone of the parent
4665 * context, or of whatever the parent is a clone of.
4666 * Note that if the parent is a clone, it could get
4667 * uncloned at any point, but that doesn't matter
4668 * because the list of counters and the generation
4669 * count can't have changed since we took the mutex.
4670 */
4671 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4672 if (cloned_ctx) {
4673 child_ctx->parent_ctx = cloned_ctx;
4674 child_ctx->parent_gen = parent_ctx->parent_gen;
4675 } else {
4676 child_ctx->parent_ctx = parent_ctx;
4677 child_ctx->parent_gen = parent_ctx->generation;
4678 }
4679 get_ctx(child_ctx->parent_ctx);
4680 }
4681
4682 mutex_unlock(&parent_ctx->mutex);
4683
4684 perf_unpin_context(parent_ctx);
4685
4686 return ret;
4687}
4688
4689static void __cpuinit perf_counter_init_cpu(int cpu)
4690{
4691 struct perf_cpu_context *cpuctx;
4692
4693 cpuctx = &per_cpu(perf_cpu_context, cpu);
4694 __perf_counter_init_context(&cpuctx->ctx, NULL);
4695
4696 spin_lock(&perf_resource_lock);
4697 cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
4698 spin_unlock(&perf_resource_lock);
4699
4700 hw_perf_counter_setup(cpu);
4701}
4702
4703#ifdef CONFIG_HOTPLUG_CPU
4704static void __perf_counter_exit_cpu(void *info)
4705{
4706 struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4707 struct perf_counter_context *ctx = &cpuctx->ctx;
4708 struct perf_counter *counter, *tmp;
4709
4710 list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
4711 __perf_counter_remove_from_context(counter);
4712}
4713static void perf_counter_exit_cpu(int cpu)
4714{
4715 struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
4716 struct perf_counter_context *ctx = &cpuctx->ctx;
4717
4718 mutex_lock(&ctx->mutex);
4719 smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
4720 mutex_unlock(&ctx->mutex);
4721}
4722#else
4723static inline void perf_counter_exit_cpu(int cpu) { }
4724#endif
4725
4726static int __cpuinit
4727perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
4728{
4729 unsigned int cpu = (long)hcpu;
4730
4731 switch (action) {
4732
4733 case CPU_UP_PREPARE:
4734 case CPU_UP_PREPARE_FROZEN:
4735 perf_counter_init_cpu(cpu);
4736 break;
4737
4738 case CPU_ONLINE:
4739 case CPU_ONLINE_FROZEN:
4740 hw_perf_counter_setup_online(cpu);
4741 break;
4742
4743 case CPU_DOWN_PREPARE:
4744 case CPU_DOWN_PREPARE_FROZEN:
4745 perf_counter_exit_cpu(cpu);
4746 break;
4747
4748 default:
4749 break;
4750 }
4751
4752 return NOTIFY_OK;
4753}
4754
4755/*
4756 * This has to have a higher priority than migration_notifier in sched.c.
4757 */
4758static struct notifier_block __cpuinitdata perf_cpu_nb = {
4759 .notifier_call = perf_cpu_notify,
4760 .priority = 20,
4761};
4762
4763void __init perf_counter_init(void)
4764{
4765 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
4766 (void *)(long)smp_processor_id());
4767 perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
4768 (void *)(long)smp_processor_id());
4769 register_cpu_notifier(&perf_cpu_nb);
4770}
4771
4772static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
4773{
4774 return sprintf(buf, "%d\n", perf_reserved_percpu);
4775}
4776
4777static ssize_t
4778perf_set_reserve_percpu(struct sysdev_class *class,
4779 const char *buf,
4780 size_t count)
4781{
4782 struct perf_cpu_context *cpuctx;
4783 unsigned long val;
4784 int err, cpu, mpt;
4785
4786 err = strict_strtoul(buf, 10, &val);
4787 if (err)
4788 return err;
4789 if (val > perf_max_counters)
4790 return -EINVAL;
4791
4792 spin_lock(&perf_resource_lock);
4793 perf_reserved_percpu = val;
4794 for_each_online_cpu(cpu) {
4795 cpuctx = &per_cpu(perf_cpu_context, cpu);
4796 spin_lock_irq(&cpuctx->ctx.lock);
4797 mpt = min(perf_max_counters - cpuctx->ctx.nr_counters,
4798 perf_max_counters - perf_reserved_percpu);
4799 cpuctx->max_pertask = mpt;
4800 spin_unlock_irq(&cpuctx->ctx.lock);
4801 }
4802 spin_unlock(&perf_resource_lock);
4803
4804 return count;
4805}
4806
4807static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
4808{
4809 return sprintf(buf, "%d\n", perf_overcommit);
4810}
4811
4812static ssize_t
4813perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
4814{
4815 unsigned long val;
4816 int err;
4817
4818 err = strict_strtoul(buf, 10, &val);
4819 if (err)
4820 return err;
4821 if (val > 1)
4822 return -EINVAL;
4823
4824 spin_lock(&perf_resource_lock);
4825 perf_overcommit = val;
4826 spin_unlock(&perf_resource_lock);
4827
4828 return count;
4829}
4830
4831static SYSDEV_CLASS_ATTR(
4832 reserve_percpu,
4833 0644,
4834 perf_show_reserve_percpu,
4835 perf_set_reserve_percpu
4836 );
4837
4838static SYSDEV_CLASS_ATTR(
4839 overcommit,
4840 0644,
4841 perf_show_overcommit,
4842 perf_set_overcommit
4843 );
4844
4845static struct attribute *perfclass_attrs[] = {
4846 &attr_reserve_percpu.attr,
4847 &attr_overcommit.attr,
4848 NULL
4849};
4850
4851static struct attribute_group perfclass_attr_group = {
4852 .attrs = perfclass_attrs,
4853 .name = "perf_counters",
4854};
4855
4856static int __init perf_counter_sysfs_init(void)
4857{
4858 return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
4859 &perfclass_attr_group);
4860}
4861device_initcall(perf_counter_sysfs_init);
generated by cgit v1.2.2 (git 2.25.0) at 2024-12-26 19:08:34 -0500