Kernel

Kernel Git Source Tree

Root/kernel/perf_event.c

1	/*
2	* Performance events core code:
3	*
4	* Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5	* Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
6	* Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7	* Copyright � 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8	*
9	* For licensing details see kernel-base/COPYING
10	*/
11
12	#include <linux/fs.h>
13	#include <linux/mm.h>
14	#include <linux/cpu.h>
15	#include <linux/smp.h>
16	#include <linux/file.h>
17	#include <linux/poll.h>
18	#include <linux/slab.h>
19	#include <linux/sysfs.h>
20	#include <linux/dcache.h>
21	#include <linux/percpu.h>
22	#include <linux/ptrace.h>
23	#include <linux/vmstat.h>
24	#include <linux/vmalloc.h>
25	#include <linux/hardirq.h>
26	#include <linux/rculist.h>
27	#include <linux/uaccess.h>
28	#include <linux/syscalls.h>
29	#include <linux/anon_inodes.h>
30	#include <linux/kernel_stat.h>
31	#include <linux/perf_event.h>
32	#include <linux/ftrace_event.h>
33	#include <linux/hw_breakpoint.h>
34
35	#include <asm/irq_regs.h>
36
37	/*
38	* Each CPU has a list of per CPU events:
39	*/
40	static DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
41
42	int perf_max_events __read_mostly = 1;
43	static int perf_reserved_percpu __read_mostly;
44	static int perf_overcommit __read_mostly = 1;
45
46	static atomic_t nr_events __read_mostly;
47	static atomic_t nr_mmap_events __read_mostly;
48	static atomic_t nr_comm_events __read_mostly;
49	static atomic_t nr_task_events __read_mostly;
50
51	/*
52	* perf event paranoia level:
53	* -1 - not paranoid at all
54	* 0 - disallow raw tracepoint access for unpriv
55	* 1 - disallow cpu events for unpriv
56	* 2 - disallow kernel profiling for unpriv
57	*/
58	int sysctl_perf_event_paranoid __read_mostly = 1;
59
60	int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
61
62	/*
63	* max perf event sample rate
64	*/
65	int sysctl_perf_event_sample_rate __read_mostly = 100000;
66
67	static atomic64_t perf_event_id;
68
69	/*
70	* Lock for (sysadmin-configurable) event reservations:
71	*/
72	static DEFINE_SPINLOCK(perf_resource_lock);
73
74	/*
75	* Architecture provided APIs - weak aliases:
76	*/
77	extern __weak const struct pmu hw_perf_event_init(struct perf_event event)
78	{
79	return NULL;
80	}
81
82	void __weak hw_perf_disable(void) { barrier(); }
83	void __weak hw_perf_enable(void) { barrier(); }
84
85	int __weak
86	hw_perf_group_sched_in(struct perf_event *group_leader,
87	struct perf_cpu_context *cpuctx,
88	struct perf_event_context *ctx)
89	{
90	return 0;
91	}
92
93	void __weak perf_event_print_debug(void) { }
94
95	static DEFINE_PER_CPU(int, perf_disable_count);
96
97	void perf_disable(void)
98	{
99	if (!__get_cpu_var(perf_disable_count)++)
100	hw_perf_disable();
101	}
102
103	void perf_enable(void)
104	{
105	if (!--__get_cpu_var(perf_disable_count))
106	hw_perf_enable();
107	}
108
109	static void get_ctx(struct perf_event_context *ctx)
110	{
111	WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
112	}
113
114	static void free_ctx(struct rcu_head *head)
115	{
116	struct perf_event_context *ctx;
117
118	ctx = container_of(head, struct perf_event_context, rcu_head);
119	kfree(ctx);
120	}
121
122	static void put_ctx(struct perf_event_context *ctx)
123	{
124	if (atomic_dec_and_test(&ctx->refcount)) {
125	if (ctx->parent_ctx)
126	put_ctx(ctx->parent_ctx);
127	if (ctx->task)
128	put_task_struct(ctx->task);
129	call_rcu(&ctx->rcu_head, free_ctx);
130	}
131	}
132
133	static void unclone_ctx(struct perf_event_context *ctx)
134	{
135	if (ctx->parent_ctx) {
136	put_ctx(ctx->parent_ctx);
137	ctx->parent_ctx = NULL;
138	}
139	}
140
141	/*
142	* If we inherit events we want to return the parent event id
143	* to userspace.
144	*/
145	static u64 primary_event_id(struct perf_event *event)
146	{
147	u64 id = event->id;
148
149	if (event->parent)
150	id = event->parent->id;
151
152	return id;
153	}
154
155	/*
156	* Get the perf_event_context for a task and lock it.
157	* This has to cope with with the fact that until it is locked,
158	* the context could get moved to another task.
159	*/
160	static struct perf_event_context *
161	perf_lock_task_context(struct task_struct task, unsigned long flags)
162	{
163	struct perf_event_context *ctx;
164
165	rcu_read_lock();
166	retry:
167	ctx = rcu_dereference(task->perf_event_ctxp);
168	if (ctx) {
169	/*
170	* If this context is a clone of another, it might
171	* get swapped for another underneath us by
172	* perf_event_task_sched_out, though the
173	* rcu_read_lock() protects us from any context
174	* getting freed. Lock the context and check if it
175	* got swapped before we could get the lock, and retry
176	* if so. If we locked the right context, then it
177	* can't get swapped on us any more.
178	*/
179	raw_spin_lock_irqsave(&ctx->lock, *flags);
180	if (ctx != rcu_dereference(task->perf_event_ctxp)) {
181	raw_spin_unlock_irqrestore(&ctx->lock, *flags);
182	goto retry;
183	}
184
185	if (!atomic_inc_not_zero(&ctx->refcount)) {
186	raw_spin_unlock_irqrestore(&ctx->lock, *flags);
187	ctx = NULL;
188	}
189	}
190	rcu_read_unlock();
191	return ctx;
192	}
193
194	/*
195	* Get the context for a task and increment its pin_count so it
196	* can't get swapped to another task. This also increments its
197	* reference count so that the context can't get freed.
198	*/
199	static struct perf_event_context perf_pin_task_context(struct task_struct task)
200	{
201	struct perf_event_context *ctx;
202	unsigned long flags;
203
204	ctx = perf_lock_task_context(task, &flags);
205	if (ctx) {
206	++ctx->pin_count;
207	raw_spin_unlock_irqrestore(&ctx->lock, flags);
208	}
209	return ctx;
210	}
211
212	static void perf_unpin_context(struct perf_event_context *ctx)
213	{
214	unsigned long flags;
215
216	raw_spin_lock_irqsave(&ctx->lock, flags);
217	--ctx->pin_count;
218	raw_spin_unlock_irqrestore(&ctx->lock, flags);
219	put_ctx(ctx);
220	}
221
222	static inline u64 perf_clock(void)
223	{
224	return cpu_clock(raw_smp_processor_id());
225	}
226
227	/*
228	* Update the record of the current time in a context.
229	*/
230	static void update_context_time(struct perf_event_context *ctx)
231	{
232	u64 now = perf_clock();
233
234	ctx->time += now - ctx->timestamp;
235	ctx->timestamp = now;
236	}
237
238	/*
239	* Update the total_time_enabled and total_time_running fields for a event.
240	*/
241	static void update_event_times(struct perf_event *event)
242	{
243	struct perf_event_context *ctx = event->ctx;
244	u64 run_end;
245
246	if (event->state < PERF_EVENT_STATE_INACTIVE \|\|
247	event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
248	return;
249
250	if (ctx->is_active)
251	run_end = ctx->time;
252	else
253	run_end = event->tstamp_stopped;
254
255	event->total_time_enabled = run_end - event->tstamp_enabled;
256
257	if (event->state == PERF_EVENT_STATE_INACTIVE)
258	run_end = event->tstamp_stopped;
259	else
260	run_end = ctx->time;
261
262	event->total_time_running = run_end - event->tstamp_running;
263	}
264
265	static struct list_head *
266	ctx_group_list(struct perf_event event, struct perf_event_context ctx)
267	{
268	if (event->attr.pinned)
269	return &ctx->pinned_groups;
270	else
271	return &ctx->flexible_groups;
272	}
273
274	/*
275	* Add a event from the lists for its context.
276	* Must be called with ctx->mutex and ctx->lock held.
277	*/
278	static void
279	list_add_event(struct perf_event event, struct perf_event_context ctx)
280	{
281	struct perf_event *group_leader = event->group_leader;
282
283	/*
284	* Depending on whether it is a standalone or sibling event,
285	* add it straight to the context's event list, or to the group
286	* leader's sibling list:
287	*/
288	if (group_leader == event) {
289	struct list_head *list;
290
291	if (is_software_event(event))
292	event->group_flags \|= PERF_GROUP_SOFTWARE;
293
294	list = ctx_group_list(event, ctx);
295	list_add_tail(&event->group_entry, list);
296	} else {
297	if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
298	!is_software_event(event))
299	group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
300
301	list_add_tail(&event->group_entry, &group_leader->sibling_list);
302	group_leader->nr_siblings++;
303	}
304
305	list_add_rcu(&event->event_entry, &ctx->event_list);
306	ctx->nr_events++;
307	if (event->attr.inherit_stat)
308	ctx->nr_stat++;
309	}
310
311	/*
312	* Remove a event from the lists for its context.
313	* Must be called with ctx->mutex and ctx->lock held.
314	*/
315	static void
316	list_del_event(struct perf_event event, struct perf_event_context ctx)
317	{
318	struct perf_event sibling, tmp;
319
320	if (list_empty(&event->group_entry))
321	return;
322	ctx->nr_events--;
323	if (event->attr.inherit_stat)
324	ctx->nr_stat--;
325
326	list_del_init(&event->group_entry);
327	list_del_rcu(&event->event_entry);
328
329	if (event->group_leader != event)
330	event->group_leader->nr_siblings--;
331
332	update_event_times(event);
333
334	/*
335	* If event was in error state, then keep it
336	* that way, otherwise bogus counts will be
337	* returned on read(). The only way to get out
338	* of error state is by explicit re-enabling
339	* of the event
340	*/
341	if (event->state > PERF_EVENT_STATE_OFF)
342	event->state = PERF_EVENT_STATE_OFF;
343
344	/*
345	* If this was a group event with sibling events then
346	* upgrade the siblings to singleton events by adding them
347	* to the context list directly:
348	*/
349	list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
350	struct list_head *list;
351
352	list = ctx_group_list(event, ctx);
353	list_move_tail(&sibling->group_entry, list);
354	sibling->group_leader = sibling;
355
356	/* Inherit group flags from the previous leader */
357	sibling->group_flags = event->group_flags;
358	}
359	}
360
361	static void
362	event_sched_out(struct perf_event *event,
363	struct perf_cpu_context *cpuctx,
364	struct perf_event_context *ctx)
365	{
366	if (event->state != PERF_EVENT_STATE_ACTIVE)
367	return;
368
369	event->state = PERF_EVENT_STATE_INACTIVE;
370	if (event->pending_disable) {
371	event->pending_disable = 0;
372	event->state = PERF_EVENT_STATE_OFF;
373	}
374	event->tstamp_stopped = ctx->time;
375	event->pmu->disable(event);
376	event->oncpu = -1;
377
378	if (!is_software_event(event))
379	cpuctx->active_oncpu--;
380	ctx->nr_active--;
381	if (event->attr.exclusive \|\| !cpuctx->active_oncpu)
382	cpuctx->exclusive = 0;
383	}
384
385	static void
386	group_sched_out(struct perf_event *group_event,
387	struct perf_cpu_context *cpuctx,
388	struct perf_event_context *ctx)
389	{
390	struct perf_event *event;
391
392	if (group_event->state != PERF_EVENT_STATE_ACTIVE)
393	return;
394
395	event_sched_out(group_event, cpuctx, ctx);
396
397	/*
398	* Schedule out siblings (if any):
399	*/
400	list_for_each_entry(event, &group_event->sibling_list, group_entry)
401	event_sched_out(event, cpuctx, ctx);
402
403	if (group_event->attr.exclusive)
404	cpuctx->exclusive = 0;
405	}
406
407	/*
408	* Cross CPU call to remove a performance event
409	*
410	* We disable the event on the hardware level first. After that we
411	* remove it from the context list.
412	*/
413	static void __perf_event_remove_from_context(void *info)
414	{
415	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
416	struct perf_event *event = info;
417	struct perf_event_context *ctx = event->ctx;
418
419	/*
420	* If this is a task context, we need to check whether it is
421	* the current task context of this cpu. If not it has been
422	* scheduled out before the smp call arrived.
423	*/
424	if (ctx->task && cpuctx->task_ctx != ctx)
425	return;
426
427	raw_spin_lock(&ctx->lock);
428	/*
429	* Protect the list operation against NMI by disabling the
430	* events on a global level.
431	*/
432	perf_disable();
433
434	event_sched_out(event, cpuctx, ctx);
435
436	list_del_event(event, ctx);
437
438	if (!ctx->task) {
439	/*
440	* Allow more per task events with respect to the
441	* reservation:
442	*/
443	cpuctx->max_pertask =
444	min(perf_max_events - ctx->nr_events,
445	perf_max_events - perf_reserved_percpu);
446	}
447
448	perf_enable();
449	raw_spin_unlock(&ctx->lock);
450	}
451
452
453	/*
454	* Remove the event from a task's (or a CPU's) list of events.
455	*
456	* Must be called with ctx->mutex held.
457	*
458	* CPU events are removed with a smp call. For task events we only
459	* call when the task is on a CPU.
460	*
461	* If event->ctx is a cloned context, callers must make sure that
462	* every task struct that event->ctx->task could possibly point to
463	* remains valid. This is OK when called from perf_release since
464	* that only calls us on the top-level context, which can't be a clone.
465	* When called from perf_event_exit_task, it's OK because the
466	* context has been detached from its task.
467	*/
468	static void perf_event_remove_from_context(struct perf_event *event)
469	{
470	struct perf_event_context *ctx = event->ctx;
471	struct task_struct *task = ctx->task;
472
473	if (!task) {
474	/*
475	* Per cpu events are removed via an smp call and
476	* the removal is always successful.
477	*/
478	smp_call_function_single(event->cpu,
479	__perf_event_remove_from_context,
480	event, 1);
481	return;
482	}
483
484	retry:
485	task_oncpu_function_call(task, __perf_event_remove_from_context,
486	event);
487
488	raw_spin_lock_irq(&ctx->lock);
489	/*
490	* If the context is active we need to retry the smp call.
491	*/
492	if (ctx->nr_active && !list_empty(&event->group_entry)) {
493	raw_spin_unlock_irq(&ctx->lock);
494	goto retry;
495	}
496
497	/*
498	* The lock prevents that this context is scheduled in so we
499	* can remove the event safely, if the call above did not
500	* succeed.
501	*/
502	if (!list_empty(&event->group_entry))
503	list_del_event(event, ctx);
504	raw_spin_unlock_irq(&ctx->lock);
505	}
506
507	/*
508	* Update total_time_enabled and total_time_running for all events in a group.
509	*/
510	static void update_group_times(struct perf_event *leader)
511	{
512	struct perf_event *event;
513
514	update_event_times(leader);
515	list_for_each_entry(event, &leader->sibling_list, group_entry)
516	update_event_times(event);
517	}
518
519	/*
520	* Cross CPU call to disable a performance event
521	*/
522	static void __perf_event_disable(void *info)
523	{
524	struct perf_event *event = info;
525	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
526	struct perf_event_context *ctx = event->ctx;
527
528	/*
529	* If this is a per-task event, need to check whether this
530	* event's task is the current task on this cpu.
531	*/
532	if (ctx->task && cpuctx->task_ctx != ctx)
533	return;
534
535	raw_spin_lock(&ctx->lock);
536
537	/*
538	* If the event is on, turn it off.
539	* If it is in error state, leave it in error state.
540	*/
541	if (event->state >= PERF_EVENT_STATE_INACTIVE) {
542	update_context_time(ctx);
543	update_group_times(event);
544	if (event == event->group_leader)
545	group_sched_out(event, cpuctx, ctx);
546	else
547	event_sched_out(event, cpuctx, ctx);
548	event->state = PERF_EVENT_STATE_OFF;
549	}
550
551	raw_spin_unlock(&ctx->lock);
552	}
553
554	/*
555	* Disable a event.
556	*
557	* If event->ctx is a cloned context, callers must make sure that
558	* every task struct that event->ctx->task could possibly point to
559	* remains valid. This condition is satisifed when called through
560	* perf_event_for_each_child or perf_event_for_each because they
561	* hold the top-level event's child_mutex, so any descendant that
562	* goes to exit will block in sync_child_event.
563	* When called from perf_pending_event it's OK because event->ctx
564	* is the current context on this CPU and preemption is disabled,
565	* hence we can't get into perf_event_task_sched_out for this context.
566	*/
567	void perf_event_disable(struct perf_event *event)
568	{
569	struct perf_event_context *ctx = event->ctx;
570	struct task_struct *task = ctx->task;
571
572	if (!task) {
573	/*
574	* Disable the event on the cpu that it's on
575	*/
576	smp_call_function_single(event->cpu, __perf_event_disable,
577	event, 1);
578	return;
579	}
580
581	retry:
582	task_oncpu_function_call(task, __perf_event_disable, event);
583
584	raw_spin_lock_irq(&ctx->lock);
585	/*
586	* If the event is still active, we need to retry the cross-call.
587	*/
588	if (event->state == PERF_EVENT_STATE_ACTIVE) {
589	raw_spin_unlock_irq(&ctx->lock);
590	goto retry;
591	}
592
593	/*
594	* Since we have the lock this context can't be scheduled
595	* in, so we can change the state safely.
596	*/
597	if (event->state == PERF_EVENT_STATE_INACTIVE) {
598	update_group_times(event);
599	event->state = PERF_EVENT_STATE_OFF;
600	}
601
602	raw_spin_unlock_irq(&ctx->lock);
603	}
604
605	static int
606	event_sched_in(struct perf_event *event,
607	struct perf_cpu_context *cpuctx,
608	struct perf_event_context *ctx)
609	{
610	if (event->state <= PERF_EVENT_STATE_OFF)
611	return 0;
612
613	event->state = PERF_EVENT_STATE_ACTIVE;
614	event->oncpu = smp_processor_id();
615	/*
616	* The new state must be visible before we turn it on in the hardware:
617	*/
618	smp_wmb();
619
620	if (event->pmu->enable(event)) {
621	event->state = PERF_EVENT_STATE_INACTIVE;
622	event->oncpu = -1;
623	return -EAGAIN;
624	}
625
626	event->tstamp_running += ctx->time - event->tstamp_stopped;
627
628	if (!is_software_event(event))
629	cpuctx->active_oncpu++;
630	ctx->nr_active++;
631
632	if (event->attr.exclusive)
633	cpuctx->exclusive = 1;
634
635	return 0;
636	}
637
638	static int
639	group_sched_in(struct perf_event *group_event,
640	struct perf_cpu_context *cpuctx,
641	struct perf_event_context *ctx)
642	{
643	struct perf_event event, partial_group;
644	int ret;
645
646	if (group_event->state == PERF_EVENT_STATE_OFF)
647	return 0;
648
649	ret = hw_perf_group_sched_in(group_event, cpuctx, ctx);
650	if (ret)
651	return ret < 0 ? ret : 0;
652
653	if (event_sched_in(group_event, cpuctx, ctx))
654	return -EAGAIN;
655
656	/*
657	* Schedule in siblings as one group (if any):
658	*/
659	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
660	if (event_sched_in(event, cpuctx, ctx)) {
661	partial_group = event;
662	goto group_error;
663	}
664	}
665
666	return 0;
667
668	group_error:
669	/*
670	* Groups can be scheduled in as one unit only, so undo any
671	* partial group before returning:
672	*/
673	list_for_each_entry(event, &group_event->sibling_list, group_entry) {
674	if (event == partial_group)
675	break;
676	event_sched_out(event, cpuctx, ctx);
677	}
678	event_sched_out(group_event, cpuctx, ctx);
679
680	return -EAGAIN;
681	}
682
683	/*
684	* Work out whether we can put this event group on the CPU now.
685	*/
686	static int group_can_go_on(struct perf_event *event,
687	struct perf_cpu_context *cpuctx,
688	int can_add_hw)
689	{
690	/*
691	* Groups consisting entirely of software events can always go on.
692	*/
693	if (event->group_flags & PERF_GROUP_SOFTWARE)
694	return 1;
695	/*
696	* If an exclusive group is already on, no other hardware
697	* events can go on.
698	*/
699	if (cpuctx->exclusive)
700	return 0;
701	/*
702	* If this group is exclusive and there are already
703	* events on the CPU, it can't go on.
704	*/
705	if (event->attr.exclusive && cpuctx->active_oncpu)
706	return 0;
707	/*
708	* Otherwise, try to add it if all previous groups were able
709	* to go on.
710	*/
711	return can_add_hw;
712	}
713
714	static void add_event_to_ctx(struct perf_event *event,
715	struct perf_event_context *ctx)
716	{
717	list_add_event(event, ctx);
718	event->tstamp_enabled = ctx->time;
719	event->tstamp_running = ctx->time;
720	event->tstamp_stopped = ctx->time;
721	}
722
723	/*
724	* Cross CPU call to install and enable a performance event
725	*
726	* Must be called with ctx->mutex held
727	*/
728	static void __perf_install_in_context(void *info)
729	{
730	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
731	struct perf_event *event = info;
732	struct perf_event_context *ctx = event->ctx;
733	struct perf_event *leader = event->group_leader;
734	int err;
735
736	/*
737	* If this is a task context, we need to check whether it is
738	* the current task context of this cpu. If not it has been
739	* scheduled out before the smp call arrived.
740	* Or possibly this is the right context but it isn't
741	* on this cpu because it had no events.
742	*/
743	if (ctx->task && cpuctx->task_ctx != ctx) {
744	if (cpuctx->task_ctx \|\| ctx->task != current)
745	return;
746	cpuctx->task_ctx = ctx;
747	}
748
749	raw_spin_lock(&ctx->lock);
750	ctx->is_active = 1;
751	update_context_time(ctx);
752
753	/*
754	* Protect the list operation against NMI by disabling the
755	* events on a global level. NOP for non NMI based events.
756	*/
757	perf_disable();
758
759	add_event_to_ctx(event, ctx);
760
761	if (event->cpu != -1 && event->cpu != smp_processor_id())
762	goto unlock;
763
764	/*
765	* Don't put the event on if it is disabled or if
766	* it is in a group and the group isn't on.
767	*/
768	if (event->state != PERF_EVENT_STATE_INACTIVE \|\|
769	(leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
770	goto unlock;
771
772	/*
773	* An exclusive event can't go on if there are already active
774	* hardware events, and no hardware event can go on if there
775	* is already an exclusive event on.
776	*/
777	if (!group_can_go_on(event, cpuctx, 1))
778	err = -EEXIST;
779	else
780	err = event_sched_in(event, cpuctx, ctx);
781
782	if (err) {
783	/*
784	* This event couldn't go on. If it is in a group
785	* then we have to pull the whole group off.
786	* If the event group is pinned then put it in error state.
787	*/
788	if (leader != event)
789	group_sched_out(leader, cpuctx, ctx);
790	if (leader->attr.pinned) {
791	update_group_times(leader);
792	leader->state = PERF_EVENT_STATE_ERROR;
793	}
794	}
795
796	if (!err && !ctx->task && cpuctx->max_pertask)
797	cpuctx->max_pertask--;
798
799	unlock:
800	perf_enable();
801
802	raw_spin_unlock(&ctx->lock);
803	}
804
805	/*
806	* Attach a performance event to a context
807	*
808	* First we add the event to the list with the hardware enable bit
809	* in event->hw_config cleared.
810	*
811	* If the event is attached to a task which is on a CPU we use a smp
812	* call to enable it in the task context. The task might have been
813	* scheduled away, but we check this in the smp call again.
814	*
815	* Must be called with ctx->mutex held.
816	*/
817	static void
818	perf_install_in_context(struct perf_event_context *ctx,
819	struct perf_event *event,
820	int cpu)
821	{
822	struct task_struct *task = ctx->task;
823
824	if (!task) {
825	/*
826	* Per cpu events are installed via an smp call and
827	* the install is always successful.
828	*/
829	smp_call_function_single(cpu, __perf_install_in_context,
830	event, 1);
831	return;
832	}
833
834	retry:
835	task_oncpu_function_call(task, __perf_install_in_context,
836	event);
837
838	raw_spin_lock_irq(&ctx->lock);
839	/*
840	* we need to retry the smp call.
841	*/
842	if (ctx->is_active && list_empty(&event->group_entry)) {
843	raw_spin_unlock_irq(&ctx->lock);
844	goto retry;
845	}
846
847	/*
848	* The lock prevents that this context is scheduled in so we
849	* can add the event safely, if it the call above did not
850	* succeed.
851	*/
852	if (list_empty(&event->group_entry))
853	add_event_to_ctx(event, ctx);
854	raw_spin_unlock_irq(&ctx->lock);
855	}
856
857	/*
858	* Put a event into inactive state and update time fields.
859	* Enabling the leader of a group effectively enables all
860	* the group members that aren't explicitly disabled, so we
861	* have to update their ->tstamp_enabled also.
862	* Note: this works for group members as well as group leaders
863	* since the non-leader members' sibling_lists will be empty.
864	*/
865	static void __perf_event_mark_enabled(struct perf_event *event,
866	struct perf_event_context *ctx)
867	{
868	struct perf_event *sub;
869
870	event->state = PERF_EVENT_STATE_INACTIVE;
871	event->tstamp_enabled = ctx->time - event->total_time_enabled;
872	list_for_each_entry(sub, &event->sibling_list, group_entry)
873	if (sub->state >= PERF_EVENT_STATE_INACTIVE)
874	sub->tstamp_enabled =
875	ctx->time - sub->total_time_enabled;
876	}
877
878	/*
879	* Cross CPU call to enable a performance event
880	*/
881	static void __perf_event_enable(void *info)
882	{
883	struct perf_event *event = info;
884	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
885	struct perf_event_context *ctx = event->ctx;
886	struct perf_event *leader = event->group_leader;
887	int err;
888
889	/*
890	* If this is a per-task event, need to check whether this
891	* event's task is the current task on this cpu.
892	*/
893	if (ctx->task && cpuctx->task_ctx != ctx) {
894	if (cpuctx->task_ctx \|\| ctx->task != current)
895	return;
896	cpuctx->task_ctx = ctx;
897	}
898
899	raw_spin_lock(&ctx->lock);
900	ctx->is_active = 1;
901	update_context_time(ctx);
902
903	if (event->state >= PERF_EVENT_STATE_INACTIVE)
904	goto unlock;
905	__perf_event_mark_enabled(event, ctx);
906
907	if (event->cpu != -1 && event->cpu != smp_processor_id())
908	goto unlock;
909
910	/*
911	* If the event is in a group and isn't the group leader,
912	* then don't put it on unless the group is on.
913	*/
914	if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
915	goto unlock;
916
917	if (!group_can_go_on(event, cpuctx, 1)) {
918	err = -EEXIST;
919	} else {
920	perf_disable();
921	if (event == leader)
922	err = group_sched_in(event, cpuctx, ctx);
923	else
924	err = event_sched_in(event, cpuctx, ctx);
925	perf_enable();
926	}
927
928	if (err) {
929	/*
930	* If this event can't go on and it's part of a
931	* group, then the whole group has to come off.
932	*/
933	if (leader != event)
934	group_sched_out(leader, cpuctx, ctx);
935	if (leader->attr.pinned) {
936	update_group_times(leader);
937	leader->state = PERF_EVENT_STATE_ERROR;
938	}
939	}
940
941	unlock:
942	raw_spin_unlock(&ctx->lock);
943	}
944
945	/*
946	* Enable a event.
947	*
948	* If event->ctx is a cloned context, callers must make sure that
949	* every task struct that event->ctx->task could possibly point to
950	* remains valid. This condition is satisfied when called through
951	* perf_event_for_each_child or perf_event_for_each as described
952	* for perf_event_disable.
953	*/
954	void perf_event_enable(struct perf_event *event)
955	{
956	struct perf_event_context *ctx = event->ctx;
957	struct task_struct *task = ctx->task;
958
959	if (!task) {
960	/*
961	* Enable the event on the cpu that it's on
962	*/
963	smp_call_function_single(event->cpu, __perf_event_enable,
964	event, 1);
965	return;
966	}
967
968	raw_spin_lock_irq(&ctx->lock);
969	if (event->state >= PERF_EVENT_STATE_INACTIVE)
970	goto out;
971
972	/*
973	* If the event is in error state, clear that first.
974	* That way, if we see the event in error state below, we
975	* know that it has gone back into error state, as distinct
976	* from the task having been scheduled away before the
977	* cross-call arrived.
978	*/
979	if (event->state == PERF_EVENT_STATE_ERROR)
980	event->state = PERF_EVENT_STATE_OFF;
981
982	retry:
983	raw_spin_unlock_irq(&ctx->lock);
984	task_oncpu_function_call(task, __perf_event_enable, event);
985
986	raw_spin_lock_irq(&ctx->lock);
987
988	/*
989	* If the context is active and the event is still off,
990	* we need to retry the cross-call.
991	*/
992	if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
993	goto retry;
994
995	/*
996	* Since we have the lock this context can't be scheduled
997	* in, so we can change the state safely.
998	*/
999	if (event->state == PERF_EVENT_STATE_OFF)
1000	__perf_event_mark_enabled(event, ctx);
1001
1002	out:
1003	raw_spin_unlock_irq(&ctx->lock);
1004	}
1005
1006	static int perf_event_refresh(struct perf_event *event, int refresh)
1007	{
1008	/*
1009	* not supported on inherited events
1010	*/
1011	if (event->attr.inherit)
1012	return -EINVAL;
1013
1014	atomic_add(refresh, &event->event_limit);
1015	perf_event_enable(event);
1016
1017	return 0;
1018	}
1019
1020	enum event_type_t {
1021	EVENT_FLEXIBLE = 0x1,
1022	EVENT_PINNED = 0x2,
1023	EVENT_ALL = EVENT_FLEXIBLE \| EVENT_PINNED,
1024	};
1025
1026	static void ctx_sched_out(struct perf_event_context *ctx,
1027	struct perf_cpu_context *cpuctx,
1028	enum event_type_t event_type)
1029	{
1030	struct perf_event *event;
1031
1032	raw_spin_lock(&ctx->lock);
1033	ctx->is_active = 0;
1034	if (likely(!ctx->nr_events))
1035	goto out;
1036	update_context_time(ctx);
1037
1038	perf_disable();
1039	if (!ctx->nr_active)
1040	goto out_enable;
1041
1042	if (event_type & EVENT_PINNED)
1043	list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1044	group_sched_out(event, cpuctx, ctx);
1045
1046	if (event_type & EVENT_FLEXIBLE)
1047	list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1048	group_sched_out(event, cpuctx, ctx);
1049
1050	out_enable:
1051	perf_enable();
1052	out:
1053	raw_spin_unlock(&ctx->lock);
1054	}
1055
1056	/*
1057	* Test whether two contexts are equivalent, i.e. whether they
1058	* have both been cloned from the same version of the same context
1059	* and they both have the same number of enabled events.
1060	* If the number of enabled events is the same, then the set
1061	* of enabled events should be the same, because these are both
1062	* inherited contexts, therefore we can't access individual events
1063	* in them directly with an fd; we can only enable/disable all
1064	* events via prctl, or enable/disable all events in a family
1065	* via ioctl, which will have the same effect on both contexts.
1066	*/
1067	static int context_equiv(struct perf_event_context *ctx1,
1068	struct perf_event_context *ctx2)
1069	{
1070	return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1071	&& ctx1->parent_gen == ctx2->parent_gen
1072	&& !ctx1->pin_count && !ctx2->pin_count;
1073	}
1074
1075	static void __perf_event_sync_stat(struct perf_event *event,
1076	struct perf_event *next_event)
1077	{
1078	u64 value;
1079
1080	if (!event->attr.inherit_stat)
1081	return;
1082
1083	/*
1084	* Update the event value, we cannot use perf_event_read()
1085	* because we're in the middle of a context switch and have IRQs
1086	* disabled, which upsets smp_call_function_single(), however
1087	* we know the event must be on the current CPU, therefore we
1088	* don't need to use it.
1089	*/
1090	switch (event->state) {
1091	case PERF_EVENT_STATE_ACTIVE:
1092	event->pmu->read(event);
1093	/* fall-through */
1094
1095	case PERF_EVENT_STATE_INACTIVE:
1096	update_event_times(event);
1097	break;
1098
1099	default:
1100	break;
1101	}
1102
1103	/*
1104	* In order to keep per-task stats reliable we need to flip the event
1105	* values when we flip the contexts.
1106	*/
1107	value = atomic64_read(&next_event->count);
1108	value = atomic64_xchg(&event->count, value);
1109	atomic64_set(&next_event->count, value);
1110
1111	swap(event->total_time_enabled, next_event->total_time_enabled);
1112	swap(event->total_time_running, next_event->total_time_running);
1113
1114	/*
1115	* Since we swizzled the values, update the user visible data too.
1116	*/
1117	perf_event_update_userpage(event);
1118	perf_event_update_userpage(next_event);
1119	}
1120
1121	#define list_next_entry(pos, member) \
1122	list_entry(pos->member.next, typeof(*pos), member)
1123
1124	static void perf_event_sync_stat(struct perf_event_context *ctx,
1125	struct perf_event_context *next_ctx)
1126	{
1127	struct perf_event event, next_event;
1128
1129	if (!ctx->nr_stat)
1130	return;
1131
1132	update_context_time(ctx);
1133
1134	event = list_first_entry(&ctx->event_list,
1135	struct perf_event, event_entry);
1136
1137	next_event = list_first_entry(&next_ctx->event_list,
1138	struct perf_event, event_entry);
1139
1140	while (&event->event_entry != &ctx->event_list &&
1141	&next_event->event_entry != &next_ctx->event_list) {
1142
1143	__perf_event_sync_stat(event, next_event);
1144
1145	event = list_next_entry(event, event_entry);
1146	next_event = list_next_entry(next_event, event_entry);
1147	}
1148	}
1149
1150	/*
1151	* Called from scheduler to remove the events of the current task,
1152	* with interrupts disabled.
1153	*
1154	* We stop each event and update the event value in event->count.
1155	*
1156	* This does not protect us against NMI, but disable()
1157	* sets the disabled bit in the control field of event _before_
1158	* accessing the event control register. If a NMI hits, then it will
1159	* not restart the event.
1160	*/
1161	void perf_event_task_sched_out(struct task_struct *task,
1162	struct task_struct *next)
1163	{
1164	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1165	struct perf_event_context *ctx = task->perf_event_ctxp;
1166	struct perf_event_context *next_ctx;
1167	struct perf_event_context *parent;
1168	int do_switch = 1;
1169
1170	perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1171
1172	if (likely(!ctx \|\| !cpuctx->task_ctx))
1173	return;
1174
1175	rcu_read_lock();
1176	parent = rcu_dereference(ctx->parent_ctx);
1177	next_ctx = next->perf_event_ctxp;
1178	if (parent && next_ctx &&
1179	rcu_dereference(next_ctx->parent_ctx) == parent) {
1180	/*
1181	* Looks like the two contexts are clones, so we might be
1182	* able to optimize the context switch. We lock both
1183	* contexts and check that they are clones under the
1184	* lock (including re-checking that neither has been
1185	* uncloned in the meantime). It doesn't matter which
1186	* order we take the locks because no other cpu could
1187	* be trying to lock both of these tasks.
1188	*/
1189	raw_spin_lock(&ctx->lock);
1190	raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1191	if (context_equiv(ctx, next_ctx)) {
1192	/*
1193	* XXX do we need a memory barrier of sorts
1194	* wrt to rcu_dereference() of perf_event_ctxp
1195	*/
1196	task->perf_event_ctxp = next_ctx;
1197	next->perf_event_ctxp = ctx;
1198	ctx->task = next;
1199	next_ctx->task = task;
1200	do_switch = 0;
1201
1202	perf_event_sync_stat(ctx, next_ctx);
1203	}
1204	raw_spin_unlock(&next_ctx->lock);
1205	raw_spin_unlock(&ctx->lock);
1206	}
1207	rcu_read_unlock();
1208
1209	if (do_switch) {
1210	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
1211	cpuctx->task_ctx = NULL;
1212	}
1213	}
1214
1215	static void task_ctx_sched_out(struct perf_event_context *ctx,
1216	enum event_type_t event_type)
1217	{
1218	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1219
1220	if (!cpuctx->task_ctx)
1221	return;
1222
1223	if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1224	return;
1225
1226	ctx_sched_out(ctx, cpuctx, event_type);
1227	cpuctx->task_ctx = NULL;
1228	}
1229
1230	/*
1231	* Called with IRQs disabled
1232	*/
1233	static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1234	{
1235	task_ctx_sched_out(ctx, EVENT_ALL);
1236	}
1237
1238	/*
1239	* Called with IRQs disabled
1240	*/
1241	static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
1242	enum event_type_t event_type)
1243	{
1244	ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
1245	}
1246
1247	static void
1248	ctx_pinned_sched_in(struct perf_event_context *ctx,
1249	struct perf_cpu_context *cpuctx)
1250	{
1251	struct perf_event *event;
1252
1253	list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1254	if (event->state <= PERF_EVENT_STATE_OFF)
1255	continue;
1256	if (event->cpu != -1 && event->cpu != smp_processor_id())
1257	continue;
1258
1259	if (group_can_go_on(event, cpuctx, 1))
1260	group_sched_in(event, cpuctx, ctx);
1261
1262	/*
1263	* If this pinned group hasn't been scheduled,
1264	* put it in error state.
1265	*/
1266	if (event->state == PERF_EVENT_STATE_INACTIVE) {
1267	update_group_times(event);
1268	event->state = PERF_EVENT_STATE_ERROR;
1269	}
1270	}
1271	}
1272
1273	static void
1274	ctx_flexible_sched_in(struct perf_event_context *ctx,
1275	struct perf_cpu_context *cpuctx)
1276	{
1277	struct perf_event *event;
1278	int can_add_hw = 1;
1279
1280	list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1281	/* Ignore events in OFF or ERROR state */
1282	if (event->state <= PERF_EVENT_STATE_OFF)
1283	continue;
1284	/*
1285	* Listen to the 'cpu' scheduling filter constraint
1286	* of events:
1287	*/
1288	if (event->cpu != -1 && event->cpu != smp_processor_id())
1289	continue;
1290
1291	if (group_can_go_on(event, cpuctx, can_add_hw))
1292	if (group_sched_in(event, cpuctx, ctx))
1293	can_add_hw = 0;
1294	}
1295	}
1296
1297	static void
1298	ctx_sched_in(struct perf_event_context *ctx,
1299	struct perf_cpu_context *cpuctx,
1300	enum event_type_t event_type)
1301	{
1302	raw_spin_lock(&ctx->lock);
1303	ctx->is_active = 1;
1304	if (likely(!ctx->nr_events))
1305	goto out;
1306
1307	ctx->timestamp = perf_clock();
1308
1309	perf_disable();
1310
1311	/*
1312	* First go through the list and put on any pinned groups
1313	* in order to give them the best chance of going on.
1314	*/
1315	if (event_type & EVENT_PINNED)
1316	ctx_pinned_sched_in(ctx, cpuctx);
1317
1318	/* Then walk through the lower prio flexible groups */
1319	if (event_type & EVENT_FLEXIBLE)
1320	ctx_flexible_sched_in(ctx, cpuctx);
1321
1322	perf_enable();
1323	out:
1324	raw_spin_unlock(&ctx->lock);
1325	}
1326
1327	static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
1328	enum event_type_t event_type)
1329	{
1330	struct perf_event_context *ctx = &cpuctx->ctx;
1331
1332	ctx_sched_in(ctx, cpuctx, event_type);
1333	}
1334
1335	static void task_ctx_sched_in(struct task_struct *task,
1336	enum event_type_t event_type)
1337	{
1338	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1339	struct perf_event_context *ctx = task->perf_event_ctxp;
1340
1341	if (likely(!ctx))
1342	return;
1343	if (cpuctx->task_ctx == ctx)
1344	return;
1345	ctx_sched_in(ctx, cpuctx, event_type);
1346	cpuctx->task_ctx = ctx;
1347	}
1348	/*
1349	* Called from scheduler to add the events of the current task
1350	* with interrupts disabled.
1351	*
1352	* We restore the event value and then enable it.
1353	*
1354	* This does not protect us against NMI, but enable()
1355	* sets the enabled bit in the control field of event _before_
1356	* accessing the event control register. If a NMI hits, then it will
1357	* keep the event running.
1358	*/
1359	void perf_event_task_sched_in(struct task_struct *task)
1360	{
1361	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1362	struct perf_event_context *ctx = task->perf_event_ctxp;
1363
1364	if (likely(!ctx))
1365	return;
1366
1367	if (cpuctx->task_ctx == ctx)
1368	return;
1369
1370	/*
1371	* We want to keep the following priority order:
1372	* cpu pinned (that don't need to move), task pinned,
1373	* cpu flexible, task flexible.
1374	*/
1375	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1376
1377	ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
1378	cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1379	ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
1380
1381	cpuctx->task_ctx = ctx;
1382	}
1383
1384	#define MAX_INTERRUPTS (~0ULL)
1385
1386	static void perf_log_throttle(struct perf_event *event, int enable);
1387
1388	static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
1389	{
1390	u64 frequency = event->attr.sample_freq;
1391	u64 sec = NSEC_PER_SEC;
1392	u64 divisor, dividend;
1393
1394	int count_fls, nsec_fls, frequency_fls, sec_fls;
1395
1396	count_fls = fls64(count);
1397	nsec_fls = fls64(nsec);
1398	frequency_fls = fls64(frequency);
1399	sec_fls = 30;
1400
1401	/*
1402	* We got @count in @nsec, with a target of sample_freq HZ
1403	* the target period becomes:
1404	*
1405	* @count * 10^9
1406	* period = -------------------
1407	* @nsec * sample_freq
1408	*
1409	*/
1410
1411	/*
1412	* Reduce accuracy by one bit such that @a and @b converge
1413	* to a similar magnitude.
1414	*/
1415	#define REDUCE_FLS(a, b) \
1416	do { \
1417	if (a##_fls > b##_fls) { \
1418	a >>= 1; \
1419	a##_fls--; \
1420	} else { \
1421	b >>= 1; \
1422	b##_fls--; \
1423	} \
1424	} while (0)
1425
1426	/*
1427	* Reduce accuracy until either term fits in a u64, then proceed with
1428	* the other, so that finally we can do a u64/u64 division.
1429	*/
1430	while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
1431	REDUCE_FLS(nsec, frequency);
1432	REDUCE_FLS(sec, count);
1433	}
1434
1435	if (count_fls + sec_fls > 64) {
1436	divisor = nsec * frequency;
1437
1438	while (count_fls + sec_fls > 64) {
1439	REDUCE_FLS(count, sec);
1440	divisor >>= 1;
1441	}
1442
1443	dividend = count * sec;
1444	} else {
1445	dividend = count * sec;
1446
1447	while (nsec_fls + frequency_fls > 64) {
1448	REDUCE_FLS(nsec, frequency);
1449	dividend >>= 1;
1450	}
1451
1452	divisor = nsec * frequency;
1453	}
1454
1455	return div64_u64(dividend, divisor);
1456	}
1457
1458	static void perf_event_stop(struct perf_event *event)
1459	{
1460	if (!event->pmu->stop)
1461	return event->pmu->disable(event);
1462
1463	return event->pmu->stop(event);
1464	}
1465
1466	static int perf_event_start(struct perf_event *event)
1467	{
1468	if (!event->pmu->start)
1469	return event->pmu->enable(event);
1470
1471	return event->pmu->start(event);
1472	}
1473
1474	static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
1475	{
1476	struct hw_perf_event *hwc = &event->hw;
1477	u64 period, sample_period;
1478	s64 delta;
1479
1480	period = perf_calculate_period(event, nsec, count);
1481
1482	delta = (s64)(period - hwc->sample_period);
1483	delta = (delta + 7) / 8; /* low pass filter */
1484
1485	sample_period = hwc->sample_period + delta;
1486
1487	if (!sample_period)
1488	sample_period = 1;
1489
1490	hwc->sample_period = sample_period;
1491
1492	if (atomic64_read(&hwc->period_left) > 8*sample_period) {
1493	perf_disable();
1494	perf_event_stop(event);
1495	atomic64_set(&hwc->period_left, 0);
1496	perf_event_start(event);
1497	perf_enable();
1498	}
1499	}
1500
1501	static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1502	{
1503	struct perf_event *event;
1504	struct hw_perf_event *hwc;
1505	u64 interrupts, now;
1506	s64 delta;
1507
1508	raw_spin_lock(&ctx->lock);
1509	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1510	if (event->state != PERF_EVENT_STATE_ACTIVE)
1511	continue;
1512
1513	if (event->cpu != -1 && event->cpu != smp_processor_id())
1514	continue;
1515
1516	hwc = &event->hw;
1517
1518	interrupts = hwc->interrupts;
1519	hwc->interrupts = 0;
1520
1521	/*
1522	* unthrottle events on the tick
1523	*/
1524	if (interrupts == MAX_INTERRUPTS) {
1525	perf_log_throttle(event, 1);
1526	perf_disable();
1527	event->pmu->unthrottle(event);
1528	perf_enable();
1529	}
1530
1531	if (!event->attr.freq \|\| !event->attr.sample_freq)
1532	continue;
1533
1534	perf_disable();
1535	event->pmu->read(event);
1536	now = atomic64_read(&event->count);
1537	delta = now - hwc->freq_count_stamp;
1538	hwc->freq_count_stamp = now;
1539
1540	if (delta > 0)
1541	perf_adjust_period(event, TICK_NSEC, delta);
1542	perf_enable();
1543	}
1544	raw_spin_unlock(&ctx->lock);
1545	}
1546
1547	/*
1548	* Round-robin a context's events:
1549	*/
1550	static void rotate_ctx(struct perf_event_context *ctx)
1551	{
1552	raw_spin_lock(&ctx->lock);
1553
1554	/* Rotate the first entry last of non-pinned groups */
1555	list_rotate_left(&ctx->flexible_groups);
1556
1557	raw_spin_unlock(&ctx->lock);
1558	}
1559
1560	void perf_event_task_tick(struct task_struct *curr)
1561	{
1562	struct perf_cpu_context *cpuctx;
1563	struct perf_event_context *ctx;
1564	int rotate = 0;
1565
1566	if (!atomic_read(&nr_events))
1567	return;
1568
1569	cpuctx = &__get_cpu_var(perf_cpu_context);
1570	if (cpuctx->ctx.nr_events &&
1571	cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
1572	rotate = 1;
1573
1574	ctx = curr->perf_event_ctxp;
1575	if (ctx && ctx->nr_events && ctx->nr_events != ctx->nr_active)
1576	rotate = 1;
1577
1578	perf_ctx_adjust_freq(&cpuctx->ctx);
1579	if (ctx)
1580	perf_ctx_adjust_freq(ctx);
1581
1582	if (!rotate)
1583	return;
1584
1585	perf_disable();
1586	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
1587	if (ctx)
1588	task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
1589
1590	rotate_ctx(&cpuctx->ctx);
1591	if (ctx)
1592	rotate_ctx(ctx);
1593
1594	cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
1595	if (ctx)
1596	task_ctx_sched_in(curr, EVENT_FLEXIBLE);
1597	perf_enable();
1598	}
1599
1600	static int event_enable_on_exec(struct perf_event *event,
1601	struct perf_event_context *ctx)
1602	{
1603	if (!event->attr.enable_on_exec)
1604	return 0;
1605
1606	event->attr.enable_on_exec = 0;
1607	if (event->state >= PERF_EVENT_STATE_INACTIVE)
1608	return 0;
1609
1610	__perf_event_mark_enabled(event, ctx);
1611
1612	return 1;
1613	}
1614
1615	/*
1616	* Enable all of a task's events that have been marked enable-on-exec.
1617	* This expects task == current.
1618	*/
1619	static void perf_event_enable_on_exec(struct task_struct *task)
1620	{
1621	struct perf_event_context *ctx;
1622	struct perf_event *event;
1623	unsigned long flags;
1624	int enabled = 0;
1625	int ret;
1626
1627	local_irq_save(flags);
1628	ctx = task->perf_event_ctxp;
1629	if (!ctx \|\| !ctx->nr_events)
1630	goto out;
1631
1632	__perf_event_task_sched_out(ctx);
1633
1634	raw_spin_lock(&ctx->lock);
1635
1636	list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
1637	ret = event_enable_on_exec(event, ctx);
1638	if (ret)
1639	enabled = 1;
1640	}
1641
1642	list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
1643	ret = event_enable_on_exec(event, ctx);
1644	if (ret)
1645	enabled = 1;
1646	}
1647
1648	/*
1649	* Unclone this context if we enabled any event.
1650	*/
1651	if (enabled)
1652	unclone_ctx(ctx);
1653
1654	raw_spin_unlock(&ctx->lock);
1655
1656	perf_event_task_sched_in(task);
1657	out:
1658	local_irq_restore(flags);
1659	}
1660
1661	/*
1662	* Cross CPU call to read the hardware event
1663	*/
1664	static void __perf_event_read(void *info)
1665	{
1666	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1667	struct perf_event *event = info;
1668	struct perf_event_context *ctx = event->ctx;
1669
1670	/*
1671	* If this is a task context, we need to check whether it is
1672	* the current task context of this cpu. If not it has been
1673	* scheduled out before the smp call arrived. In that case
1674	* event->count would have been updated to a recent sample
1675	* when the event was scheduled out.
1676	*/
1677	if (ctx->task && cpuctx->task_ctx != ctx)
1678	return;
1679
1680	raw_spin_lock(&ctx->lock);
1681	update_context_time(ctx);
1682	update_event_times(event);
1683	raw_spin_unlock(&ctx->lock);
1684
1685	event->pmu->read(event);
1686	}
1687
1688	static u64 perf_event_read(struct perf_event *event)
1689	{
1690	/*
1691	* If event is enabled and currently active on a CPU, update the
1692	* value in the event structure:
1693	*/
1694	if (event->state == PERF_EVENT_STATE_ACTIVE) {
1695	smp_call_function_single(event->oncpu,
1696	__perf_event_read, event, 1);
1697	} else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1698	struct perf_event_context *ctx = event->ctx;
1699	unsigned long flags;
1700
1701	raw_spin_lock_irqsave(&ctx->lock, flags);
1702	update_context_time(ctx);
1703	update_event_times(event);
1704	raw_spin_unlock_irqrestore(&ctx->lock, flags);
1705	}
1706
1707	return atomic64_read(&event->count);
1708	}
1709
1710	/*
1711	* Initialize the perf_event context in a task_struct:
1712	*/
1713	static void
1714	__perf_event_init_context(struct perf_event_context *ctx,
1715	struct task_struct *task)
1716	{
1717	raw_spin_lock_init(&ctx->lock);
1718	mutex_init(&ctx->mutex);
1719	INIT_LIST_HEAD(&ctx->pinned_groups);
1720	INIT_LIST_HEAD(&ctx->flexible_groups);
1721	INIT_LIST_HEAD(&ctx->event_list);
1722	atomic_set(&ctx->refcount, 1);
1723	ctx->task = task;
1724	}
1725
1726	static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1727	{
1728	struct perf_event_context *ctx;
1729	struct perf_cpu_context *cpuctx;
1730	struct task_struct *task;
1731	unsigned long flags;
1732	int err;
1733
1734	if (pid == -1 && cpu != -1) {
1735	/* Must be root to operate on a CPU event: */
1736	if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1737	return ERR_PTR(-EACCES);
1738
1739	if (cpu < 0 \|\| cpu >= nr_cpumask_bits)
1740	return ERR_PTR(-EINVAL);
1741
1742	/*
1743	* We could be clever and allow to attach a event to an
1744	* offline CPU and activate it when the CPU comes up, but
1745	* that's for later.
1746	*/
1747	if (!cpu_online(cpu))
1748	return ERR_PTR(-ENODEV);
1749
1750	cpuctx = &per_cpu(perf_cpu_context, cpu);
1751	ctx = &cpuctx->ctx;
1752	get_ctx(ctx);
1753
1754	return ctx;
1755	}
1756
1757	rcu_read_lock();
1758	if (!pid)
1759	task = current;
1760	else
1761	task = find_task_by_vpid(pid);
1762	if (task)
1763	get_task_struct(task);
1764	rcu_read_unlock();
1765
1766	if (!task)
1767	return ERR_PTR(-ESRCH);
1768
1769	/*
1770	* Can't attach events to a dying task.
1771	*/
1772	err = -ESRCH;
1773	if (task->flags & PF_EXITING)
1774	goto errout;
1775
1776	/* Reuse ptrace permission checks for now. */
1777	err = -EACCES;
1778	if (!ptrace_may_access(task, PTRACE_MODE_READ))
1779	goto errout;
1780
1781	retry:
1782	ctx = perf_lock_task_context(task, &flags);
1783	if (ctx) {
1784	unclone_ctx(ctx);
1785	raw_spin_unlock_irqrestore(&ctx->lock, flags);
1786	}
1787
1788	if (!ctx) {
1789	ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1790	err = -ENOMEM;
1791	if (!ctx)
1792	goto errout;
1793	__perf_event_init_context(ctx, task);
1794	get_ctx(ctx);
1795	if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1796	/*
1797	* We raced with some other task; use
1798	* the context they set.
1799	*/
1800	kfree(ctx);
1801	goto retry;
1802	}
1803	get_task_struct(task);
1804	}
1805
1806	put_task_struct(task);
1807	return ctx;
1808
1809	errout:
1810	put_task_struct(task);
1811	return ERR_PTR(err);
1812	}
1813
1814	static void perf_event_free_filter(struct perf_event *event);
1815
1816	static void free_event_rcu(struct rcu_head *head)
1817	{
1818	struct perf_event *event;
1819
1820	event = container_of(head, struct perf_event, rcu_head);
1821	if (event->ns)
1822	put_pid_ns(event->ns);
1823	perf_event_free_filter(event);
1824	kfree(event);
1825	}
1826
1827	static void perf_pending_sync(struct perf_event *event);
1828
1829	static void free_event(struct perf_event *event)
1830	{
1831	perf_pending_sync(event);
1832
1833	if (!event->parent) {
1834	atomic_dec(&nr_events);
1835	if (event->attr.mmap)
1836	atomic_dec(&nr_mmap_events);
1837	if (event->attr.comm)
1838	atomic_dec(&nr_comm_events);
1839	if (event->attr.task)
1840	atomic_dec(&nr_task_events);
1841	}
1842
1843	if (event->output) {
1844	fput(event->output->filp);
1845	event->output = NULL;
1846	}
1847
1848	if (event->destroy)
1849	event->destroy(event);
1850
1851	put_ctx(event->ctx);
1852	call_rcu(&event->rcu_head, free_event_rcu);
1853	}
1854
1855	int perf_event_release_kernel(struct perf_event *event)
1856	{
1857	struct perf_event_context *ctx = event->ctx;
1858
1859	WARN_ON_ONCE(ctx->parent_ctx);
1860	mutex_lock(&ctx->mutex);
1861	perf_event_remove_from_context(event);
1862	mutex_unlock(&ctx->mutex);
1863
1864	mutex_lock(&event->owner->perf_event_mutex);
1865	list_del_init(&event->owner_entry);
1866	mutex_unlock(&event->owner->perf_event_mutex);
1867	put_task_struct(event->owner);
1868
1869	free_event(event);
1870
1871	return 0;
1872	}
1873	EXPORT_SYMBOL_GPL(perf_event_release_kernel);
1874
1875	/*
1876	* Called when the last reference to the file is gone.
1877	*/
1878	static int perf_release(struct inode inode, struct file file)
1879	{
1880	struct perf_event *event = file->private_data;
1881
1882	file->private_data = NULL;
1883
1884	return perf_event_release_kernel(event);
1885	}
1886
1887	static int perf_event_read_size(struct perf_event *event)
1888	{
1889	int entry = sizeof(u64); /* value */
1890	int size = 0;
1891	int nr = 1;
1892
1893	if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1894	size += sizeof(u64);
1895
1896	if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1897	size += sizeof(u64);
1898
1899	if (event->attr.read_format & PERF_FORMAT_ID)
1900	entry += sizeof(u64);
1901
1902	if (event->attr.read_format & PERF_FORMAT_GROUP) {
1903	nr += event->group_leader->nr_siblings;
1904	size += sizeof(u64);
1905	}
1906
1907	size += entry * nr;
1908
1909	return size;
1910	}
1911
1912	u64 perf_event_read_value(struct perf_event event, u64 enabled, u64 *running)
1913	{
1914	struct perf_event *child;
1915	u64 total = 0;
1916
1917	*enabled = 0;
1918	*running = 0;
1919
1920	mutex_lock(&event->child_mutex);
1921	total += perf_event_read(event);
1922	*enabled += event->total_time_enabled +
1923	atomic64_read(&event->child_total_time_enabled);
1924	*running += event->total_time_running +
1925	atomic64_read(&event->child_total_time_running);
1926
1927	list_for_each_entry(child, &event->child_list, child_list) {
1928	total += perf_event_read(child);
1929	*enabled += child->total_time_enabled;
1930	*running += child->total_time_running;
1931	}
1932	mutex_unlock(&event->child_mutex);
1933
1934	return total;
1935	}
1936	EXPORT_SYMBOL_GPL(perf_event_read_value);
1937
1938	static int perf_event_read_group(struct perf_event *event,
1939	u64 read_format, char __user *buf)
1940	{
1941	struct perf_event leader = event->group_leader, sub;
1942	int n = 0, size = 0, ret = -EFAULT;
1943	struct perf_event_context *ctx = leader->ctx;
1944	u64 values[5];
1945	u64 count, enabled, running;
1946
1947	mutex_lock(&ctx->mutex);
1948	count = perf_event_read_value(leader, &enabled, &running);
1949
1950	values[n++] = 1 + leader->nr_siblings;
1951	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1952	values[n++] = enabled;
1953	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1954	values[n++] = running;
1955	values[n++] = count;
1956	if (read_format & PERF_FORMAT_ID)
1957	values[n++] = primary_event_id(leader);
1958
1959	size = n * sizeof(u64);
1960
1961	if (copy_to_user(buf, values, size))
1962	goto unlock;
1963
1964	ret = size;
1965
1966	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1967	n = 0;
1968
1969	values[n++] = perf_event_read_value(sub, &enabled, &running);
1970	if (read_format & PERF_FORMAT_ID)
1971	values[n++] = primary_event_id(sub);
1972
1973	size = n * sizeof(u64);
1974
1975	if (copy_to_user(buf + ret, values, size)) {
1976	ret = -EFAULT;
1977	goto unlock;
1978	}
1979
1980	ret += size;
1981	}
1982	unlock:
1983	mutex_unlock(&ctx->mutex);
1984
1985	return ret;
1986	}
1987
1988	static int perf_event_read_one(struct perf_event *event,
1989	u64 read_format, char __user *buf)
1990	{
1991	u64 enabled, running;
1992	u64 values[4];
1993	int n = 0;
1994
1995	values[n++] = perf_event_read_value(event, &enabled, &running);
1996	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1997	values[n++] = enabled;
1998	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1999	values[n++] = running;
2000	if (read_format & PERF_FORMAT_ID)
2001	values[n++] = primary_event_id(event);
2002
2003	if (copy_to_user(buf, values, n * sizeof(u64)))
2004	return -EFAULT;
2005
2006	return n * sizeof(u64);
2007	}
2008
2009	/*
2010	* Read the performance event - simple non blocking version for now
2011	*/
2012	static ssize_t
2013	perf_read_hw(struct perf_event event, char __user buf, size_t count)
2014	{
2015	u64 read_format = event->attr.read_format;
2016	int ret;
2017
2018	/*
2019	* Return end-of-file for a read on a event that is in
2020	* error state (i.e. because it was pinned but it couldn't be
2021	* scheduled on to the CPU at some point).
2022	*/
2023	if (event->state == PERF_EVENT_STATE_ERROR)
2024	return 0;
2025
2026	if (count < perf_event_read_size(event))
2027	return -ENOSPC;
2028
2029	WARN_ON_ONCE(event->ctx->parent_ctx);
2030	if (read_format & PERF_FORMAT_GROUP)
2031	ret = perf_event_read_group(event, read_format, buf);
2032	else
2033	ret = perf_event_read_one(event, read_format, buf);
2034
2035	return ret;
2036	}
2037
2038	static ssize_t
2039	perf_read(struct file file, char __user buf, size_t count, loff_t *ppos)
2040	{
2041	struct perf_event *event = file->private_data;
2042
2043	return perf_read_hw(event, buf, count);
2044	}
2045
2046	static unsigned int perf_poll(struct file file, poll_table wait)
2047	{
2048	struct perf_event *event = file->private_data;
2049	struct perf_mmap_data *data;
2050	unsigned int events = POLL_HUP;
2051
2052	rcu_read_lock();
2053	data = rcu_dereference(event->data);
2054	if (data)
2055	events = atomic_xchg(&data->poll, 0);
2056	rcu_read_unlock();
2057
2058	poll_wait(file, &event->waitq, wait);
2059
2060	return events;
2061	}
2062
2063	static void perf_event_reset(struct perf_event *event)
2064	{
2065	(void)perf_event_read(event);
2066	atomic64_set(&event->count, 0);
2067	perf_event_update_userpage(event);
2068	}
2069
2070	/*
2071	* Holding the top-level event's child_mutex means that any
2072	* descendant process that has inherited this event will block
2073	* in sync_child_event if it goes to exit, thus satisfying the
2074	* task existence requirements of perf_event_enable/disable.
2075	*/
2076	static void perf_event_for_each_child(struct perf_event *event,
2077	void (func)(struct perf_event ))
2078	{
2079	struct perf_event *child;
2080
2081	WARN_ON_ONCE(event->ctx->parent_ctx);
2082	mutex_lock(&event->child_mutex);
2083	func(event);
2084	list_for_each_entry(child, &event->child_list, child_list)
2085	func(child);
2086	mutex_unlock(&event->child_mutex);
2087	}
2088
2089	static void perf_event_for_each(struct perf_event *event,
2090	void (func)(struct perf_event ))
2091	{
2092	struct perf_event_context *ctx = event->ctx;
2093	struct perf_event *sibling;
2094
2095	WARN_ON_ONCE(ctx->parent_ctx);
2096	mutex_lock(&ctx->mutex);
2097	event = event->group_leader;
2098
2099	perf_event_for_each_child(event, func);
2100	func(event);
2101	list_for_each_entry(sibling, &event->sibling_list, group_entry)
2102	perf_event_for_each_child(event, func);
2103	mutex_unlock(&ctx->mutex);
2104	}
2105
2106	static int perf_event_period(struct perf_event event, u64 __user arg)
2107	{
2108	struct perf_event_context *ctx = event->ctx;
2109	unsigned long size;
2110	int ret = 0;
2111	u64 value;
2112
2113	if (!event->attr.sample_period)
2114	return -EINVAL;
2115
2116	size = copy_from_user(&value, arg, sizeof(value));
2117	if (size != sizeof(value))
2118	return -EFAULT;
2119
2120	if (!value)
2121	return -EINVAL;
2122
2123	raw_spin_lock_irq(&ctx->lock);
2124	if (event->attr.freq) {
2125	if (value > sysctl_perf_event_sample_rate) {
2126	ret = -EINVAL;
2127	goto unlock;
2128	}
2129
2130	event->attr.sample_freq = value;
2131	} else {
2132	event->attr.sample_period = value;
2133	event->hw.sample_period = value;
2134	}
2135	unlock:
2136	raw_spin_unlock_irq(&ctx->lock);
2137
2138	return ret;
2139	}
2140
2141	static int perf_event_set_output(struct perf_event *event, int output_fd);
2142	static int perf_event_set_filter(struct perf_event event, void __user arg);
2143
2144	static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
2145	{
2146	struct perf_event *event = file->private_data;
2147	void (func)(struct perf_event );
2148	u32 flags = arg;
2149
2150	switch (cmd) {
2151	case PERF_EVENT_IOC_ENABLE:
2152	func = perf_event_enable;
2153	break;
2154	case PERF_EVENT_IOC_DISABLE:
2155	func = perf_event_disable;
2156	break;
2157	case PERF_EVENT_IOC_RESET:
2158	func = perf_event_reset;
2159	break;
2160
2161	case PERF_EVENT_IOC_REFRESH:
2162	return perf_event_refresh(event, arg);
2163
2164	case PERF_EVENT_IOC_PERIOD:
2165	return perf_event_period(event, (u64 __user *)arg);
2166
2167	case PERF_EVENT_IOC_SET_OUTPUT:
2168	return perf_event_set_output(event, arg);
2169
2170	case PERF_EVENT_IOC_SET_FILTER:
2171	return perf_event_set_filter(event, (void __user *)arg);
2172
2173	default:
2174	return -ENOTTY;
2175	}
2176
2177	if (flags & PERF_IOC_FLAG_GROUP)
2178	perf_event_for_each(event, func);
2179	else
2180	perf_event_for_each_child(event, func);
2181
2182	return 0;
2183	}
2184
2185	int perf_event_task_enable(void)
2186	{
2187	struct perf_event *event;
2188
2189	mutex_lock(&current->perf_event_mutex);
2190	list_for_each_entry(event, &current->perf_event_list, owner_entry)
2191	perf_event_for_each_child(event, perf_event_enable);
2192	mutex_unlock(&current->perf_event_mutex);
2193
2194	return 0;
2195	}
2196
2197	int perf_event_task_disable(void)
2198	{
2199	struct perf_event *event;
2200
2201	mutex_lock(&current->perf_event_mutex);
2202	list_for_each_entry(event, &current->perf_event_list, owner_entry)
2203	perf_event_for_each_child(event, perf_event_disable);
2204	mutex_unlock(&current->perf_event_mutex);
2205
2206	return 0;
2207	}
2208
2209	#ifndef PERF_EVENT_INDEX_OFFSET
2210	# define PERF_EVENT_INDEX_OFFSET 0
2211	#endif
2212
2213	static int perf_event_index(struct perf_event *event)
2214	{
2215	if (event->state != PERF_EVENT_STATE_ACTIVE)
2216	return 0;
2217
2218	return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2219	}
2220
2221	/*
2222	* Callers need to ensure there can be no nesting of this function, otherwise
2223	* the seqlock logic goes bad. We can not serialize this because the arch
2224	* code calls this from NMI context.
2225	*/
2226	void perf_event_update_userpage(struct perf_event *event)
2227	{
2228	struct perf_event_mmap_page *userpg;
2229	struct perf_mmap_data *data;
2230
2231	rcu_read_lock();
2232	data = rcu_dereference(event->data);
2233	if (!data)
2234	goto unlock;
2235
2236	userpg = data->user_page;
2237
2238	/*
2239	* Disable preemption so as to not let the corresponding user-space
2240	* spin too long if we get preempted.
2241	*/
2242	preempt_disable();
2243	++userpg->lock;
2244	barrier();
2245	userpg->index = perf_event_index(event);
2246	userpg->offset = atomic64_read(&event->count);
2247	if (event->state == PERF_EVENT_STATE_ACTIVE)
2248	userpg->offset -= atomic64_read(&event->hw.prev_count);
2249
2250	userpg->time_enabled = event->total_time_enabled +
2251	atomic64_read(&event->child_total_time_enabled);
2252
2253	userpg->time_running = event->total_time_running +
2254	atomic64_read(&event->child_total_time_running);
2255
2256	barrier();
2257	++userpg->lock;
2258	preempt_enable();
2259	unlock:
2260	rcu_read_unlock();
2261	}
2262
2263	static unsigned long perf_data_size(struct perf_mmap_data *data)
2264	{
2265	return data->nr_pages << (PAGE_SHIFT + data->data_order);
2266	}
2267
2268	#ifndef CONFIG_PERF_USE_VMALLOC
2269
2270	/*
2271	* Back perf_mmap() with regular GFP_KERNEL-0 pages.
2272	*/
2273
2274	static struct page *
2275	perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2276	{
2277	if (pgoff > data->nr_pages)
2278	return NULL;
2279
2280	if (pgoff == 0)
2281	return virt_to_page(data->user_page);
2282
2283	return virt_to_page(data->data_pages[pgoff - 1]);
2284	}
2285
2286	static struct perf_mmap_data *
2287	perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2288	{
2289	struct perf_mmap_data *data;
2290	unsigned long size;
2291	int i;
2292
2293	WARN_ON(atomic_read(&event->mmap_count));
2294
2295	size = sizeof(struct perf_mmap_data);
2296	size += nr_pages * sizeof(void *);
2297
2298	data = kzalloc(size, GFP_KERNEL);
2299	if (!data)
2300	goto fail;
2301
2302	data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2303	if (!data->user_page)
2304	goto fail_user_page;
2305
2306	for (i = 0; i < nr_pages; i++) {
2307	data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2308	if (!data->data_pages[i])
2309	goto fail_data_pages;
2310	}
2311
2312	data->data_order = 0;
2313	data->nr_pages = nr_pages;
2314
2315	return data;
2316
2317	fail_data_pages:
2318	for (i--; i >= 0; i--)
2319	free_page((unsigned long)data->data_pages[i]);
2320
2321	free_page((unsigned long)data->user_page);
2322
2323	fail_user_page:
2324	kfree(data);
2325
2326	fail:
2327	return NULL;
2328	}
2329
2330	static void perf_mmap_free_page(unsigned long addr)
2331	{
2332	struct page page = virt_to_page((void )addr);
2333
2334	page->mapping = NULL;
2335	__free_page(page);
2336	}
2337
2338	static void perf_mmap_data_free(struct perf_mmap_data *data)
2339	{
2340	int i;
2341
2342	perf_mmap_free_page((unsigned long)data->user_page);
2343	for (i = 0; i < data->nr_pages; i++)
2344	perf_mmap_free_page((unsigned long)data->data_pages[i]);
2345	kfree(data);
2346	}
2347
2348	#else
2349
2350	/*
2351	* Back perf_mmap() with vmalloc memory.
2352	*
2353	* Required for architectures that have d-cache aliasing issues.
2354	*/
2355
2356	static struct page *
2357	perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2358	{
2359	if (pgoff > (1UL << data->data_order))
2360	return NULL;
2361
2362	return vmalloc_to_page((void )data->user_page + pgoff PAGE_SIZE);
2363	}
2364
2365	static void perf_mmap_unmark_page(void *addr)
2366	{
2367	struct page *page = vmalloc_to_page(addr);
2368
2369	page->mapping = NULL;
2370	}
2371
2372	static void perf_mmap_data_free_work(struct work_struct *work)
2373	{
2374	struct perf_mmap_data *data;
2375	void *base;
2376	int i, nr;
2377
2378	data = container_of(work, struct perf_mmap_data, work);
2379	nr = 1 << data->data_order;
2380
2381	base = data->user_page;
2382	for (i = 0; i < nr + 1; i++)
2383	perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2384
2385	vfree(base);
2386	kfree(data);
2387	}
2388
2389	static void perf_mmap_data_free(struct perf_mmap_data *data)
2390	{
2391	schedule_work(&data->work);
2392	}
2393
2394	static struct perf_mmap_data *
2395	perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2396	{
2397	struct perf_mmap_data *data;
2398	unsigned long size;
2399	void *all_buf;
2400
2401	WARN_ON(atomic_read(&event->mmap_count));
2402
2403	size = sizeof(struct perf_mmap_data);
2404	size += sizeof(void *);
2405
2406	data = kzalloc(size, GFP_KERNEL);
2407	if (!data)
2408	goto fail;
2409
2410	INIT_WORK(&data->work, perf_mmap_data_free_work);
2411
2412	all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2413	if (!all_buf)
2414	goto fail_all_buf;
2415
2416	data->user_page = all_buf;
2417	data->data_pages[0] = all_buf + PAGE_SIZE;
2418	data->data_order = ilog2(nr_pages);
2419	data->nr_pages = 1;
2420
2421	return data;
2422
2423	fail_all_buf:
2424	kfree(data);
2425
2426	fail:
2427	return NULL;
2428	}
2429
2430	#endif
2431
2432	static int perf_mmap_fault(struct vm_area_struct vma, struct vm_fault vmf)
2433	{
2434	struct perf_event *event = vma->vm_file->private_data;
2435	struct perf_mmap_data *data;
2436	int ret = VM_FAULT_SIGBUS;
2437
2438	if (vmf->flags & FAULT_FLAG_MKWRITE) {
2439	if (vmf->pgoff == 0)
2440	ret = 0;
2441	return ret;
2442	}
2443
2444	rcu_read_lock();
2445	data = rcu_dereference(event->data);
2446	if (!data)
2447	goto unlock;
2448
2449	if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2450	goto unlock;
2451
2452	vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2453	if (!vmf->page)
2454	goto unlock;
2455
2456	get_page(vmf->page);
2457	vmf->page->mapping = vma->vm_file->f_mapping;
2458	vmf->page->index = vmf->pgoff;
2459
2460	ret = 0;
2461	unlock:
2462	rcu_read_unlock();
2463
2464	return ret;
2465	}
2466
2467	static void
2468	perf_mmap_data_init(struct perf_event event, struct perf_mmap_data data)
2469	{
2470	long max_size = perf_data_size(data);
2471
2472	atomic_set(&data->lock, -1);
2473
2474	if (event->attr.watermark) {
2475	data->watermark = min_t(long, max_size,
2476	event->attr.wakeup_watermark);
2477	}
2478
2479	if (!data->watermark)
2480	data->watermark = max_size / 2;
2481
2482
2483	rcu_assign_pointer(event->data, data);
2484	}
2485
2486	static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2487	{
2488	struct perf_mmap_data *data;
2489
2490	data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2491	perf_mmap_data_free(data);
2492	}
2493
2494	static void perf_mmap_data_release(struct perf_event *event)
2495	{
2496	struct perf_mmap_data *data = event->data;
2497
2498	WARN_ON(atomic_read(&event->mmap_count));
2499
2500	rcu_assign_pointer(event->data, NULL);
2501	call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2502	}
2503
2504	static void perf_mmap_open(struct vm_area_struct *vma)
2505	{
2506	struct perf_event *event = vma->vm_file->private_data;
2507
2508	atomic_inc(&event->mmap_count);
2509	}
2510
2511	static void perf_mmap_close(struct vm_area_struct *vma)
2512	{
2513	struct perf_event *event = vma->vm_file->private_data;
2514
2515	WARN_ON_ONCE(event->ctx->parent_ctx);
2516	if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2517	unsigned long size = perf_data_size(event->data);
2518	struct user_struct *user = current_user();
2519
2520	atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2521	vma->vm_mm->locked_vm -= event->data->nr_locked;
2522	perf_mmap_data_release(event);
2523	mutex_unlock(&event->mmap_mutex);
2524	}
2525	}
2526
2527	static const struct vm_operations_struct perf_mmap_vmops = {
2528	.open = perf_mmap_open,
2529	.close = perf_mmap_close,
2530	.fault = perf_mmap_fault,
2531	.page_mkwrite = perf_mmap_fault,
2532	};
2533
2534	static int perf_mmap(struct file file, struct vm_area_struct vma)
2535	{
2536	struct perf_event *event = file->private_data;
2537	unsigned long user_locked, user_lock_limit;
2538	struct user_struct *user = current_user();
2539	unsigned long locked, lock_limit;
2540	struct perf_mmap_data *data;
2541	unsigned long vma_size;
2542	unsigned long nr_pages;
2543	long user_extra, extra;
2544	int ret = 0;
2545
2546	if (!(vma->vm_flags & VM_SHARED))
2547	return -EINVAL;
2548
2549	vma_size = vma->vm_end - vma->vm_start;
2550	nr_pages = (vma_size / PAGE_SIZE) - 1;
2551
2552	/*
2553	* If we have data pages ensure they're a power-of-two number, so we
2554	* can do bitmasks instead of modulo.
2555	*/
2556	if (nr_pages != 0 && !is_power_of_2(nr_pages))
2557	return -EINVAL;
2558
2559	if (vma_size != PAGE_SIZE * (1 + nr_pages))
2560	return -EINVAL;
2561
2562	if (vma->vm_pgoff != 0)
2563	return -EINVAL;
2564
2565	WARN_ON_ONCE(event->ctx->parent_ctx);
2566	mutex_lock(&event->mmap_mutex);
2567	if (event->output) {
2568	ret = -EINVAL;
2569	goto unlock;
2570	}
2571
2572	if (atomic_inc_not_zero(&event->mmap_count)) {
2573	if (nr_pages != event->data->nr_pages)
2574	ret = -EINVAL;
2575	goto unlock;
2576	}
2577
2578	user_extra = nr_pages + 1;
2579	user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2580
2581	/*
2582	* Increase the limit linearly with more CPUs:
2583	*/
2584	user_lock_limit *= num_online_cpus();
2585
2586	user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2587
2588	extra = 0;
2589	if (user_locked > user_lock_limit)
2590	extra = user_locked - user_lock_limit;
2591
2592	lock_limit = rlimit(RLIMIT_MEMLOCK);
2593	lock_limit >>= PAGE_SHIFT;
2594	locked = vma->vm_mm->locked_vm + extra;
2595
2596	if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2597	!capable(CAP_IPC_LOCK)) {
2598	ret = -EPERM;
2599	goto unlock;
2600	}
2601
2602	WARN_ON(event->data);
2603
2604	data = perf_mmap_data_alloc(event, nr_pages);
2605	ret = -ENOMEM;
2606	if (!data)
2607	goto unlock;
2608
2609	ret = 0;
2610	perf_mmap_data_init(event, data);
2611
2612	atomic_set(&event->mmap_count, 1);
2613	atomic_long_add(user_extra, &user->locked_vm);
2614	vma->vm_mm->locked_vm += extra;
2615	event->data->nr_locked = extra;
2616	if (vma->vm_flags & VM_WRITE)
2617	event->data->writable = 1;
2618
2619	unlock:
2620	mutex_unlock(&event->mmap_mutex);
2621
2622	vma->vm_flags \|= VM_RESERVED;
2623	vma->vm_ops = &perf_mmap_vmops;
2624
2625	return ret;
2626	}
2627
2628	static int perf_fasync(int fd, struct file *filp, int on)
2629	{
2630	struct inode *inode = filp->f_path.dentry->d_inode;
2631	struct perf_event *event = filp->private_data;
2632	int retval;
2633
2634	mutex_lock(&inode->i_mutex);
2635	retval = fasync_helper(fd, filp, on, &event->fasync);
2636	mutex_unlock(&inode->i_mutex);
2637
2638	if (retval < 0)
2639	return retval;
2640
2641	return 0;
2642	}
2643
2644	static const struct file_operations perf_fops = {
2645	.release = perf_release,
2646	.read = perf_read,
2647	.poll = perf_poll,
2648	.unlocked_ioctl = perf_ioctl,
2649	.compat_ioctl = perf_ioctl,
2650	.mmap = perf_mmap,
2651	.fasync = perf_fasync,
2652	};
2653
2654	/*
2655	* Perf event wakeup
2656	*
2657	* If there's data, ensure we set the poll() state and publish everything
2658	* to user-space before waking everybody up.
2659	*/
2660
2661	void perf_event_wakeup(struct perf_event *event)
2662	{
2663	wake_up_all(&event->waitq);
2664
2665	if (event->pending_kill) {
2666	kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2667	event->pending_kill = 0;
2668	}
2669	}
2670
2671	/*
2672	* Pending wakeups
2673	*
2674	* Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2675	*
2676	* The NMI bit means we cannot possibly take locks. Therefore, maintain a
2677	* single linked list and use cmpxchg() to add entries lockless.
2678	*/
2679
2680	static void perf_pending_event(struct perf_pending_entry *entry)
2681	{
2682	struct perf_event *event = container_of(entry,
2683	struct perf_event, pending);
2684
2685	if (event->pending_disable) {
2686	event->pending_disable = 0;
2687	__perf_event_disable(event);
2688	}
2689
2690	if (event->pending_wakeup) {
2691	event->pending_wakeup = 0;
2692	perf_event_wakeup(event);
2693	}
2694	}
2695
2696	#define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2697
2698	static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2699	PENDING_TAIL,
2700	};
2701
2702	static void perf_pending_queue(struct perf_pending_entry *entry,
2703	void (func)(struct perf_pending_entry ))
2704	{
2705	struct perf_pending_entry **head;
2706
2707	if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2708	return;
2709
2710	entry->func = func;
2711
2712	head = &get_cpu_var(perf_pending_head);
2713
2714	do {
2715	entry->next = *head;
2716	} while (cmpxchg(head, entry->next, entry) != entry->next);
2717
2718	set_perf_event_pending();
2719
2720	put_cpu_var(perf_pending_head);
2721	}
2722
2723	static int __perf_pending_run(void)
2724	{
2725	struct perf_pending_entry *list;
2726	int nr = 0;
2727
2728	list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2729	while (list != PENDING_TAIL) {
2730	void (func)(struct perf_pending_entry );
2731	struct perf_pending_entry *entry = list;
2732
2733	list = list->next;
2734
2735	func = entry->func;
2736	entry->next = NULL;
2737	/*
2738	* Ensure we observe the unqueue before we issue the wakeup,
2739	* so that we won't be waiting forever.
2740	* -- see perf_not_pending().
2741	*/
2742	smp_wmb();
2743
2744	func(entry);
2745	nr++;
2746	}
2747
2748	return nr;
2749	}
2750
2751	static inline int perf_not_pending(struct perf_event *event)
2752	{
2753	/*
2754	* If we flush on whatever cpu we run, there is a chance we don't
2755	* need to wait.
2756	*/
2757	get_cpu();
2758	__perf_pending_run();
2759	put_cpu();
2760
2761	/*
2762	* Ensure we see the proper queue state before going to sleep
2763	* so that we do not miss the wakeup. -- see perf_pending_handle()
2764	*/
2765	smp_rmb();
2766	return event->pending.next == NULL;
2767	}
2768
2769	static void perf_pending_sync(struct perf_event *event)
2770	{
2771	wait_event(event->waitq, perf_not_pending(event));
2772	}
2773
2774	void perf_event_do_pending(void)
2775	{
2776	__perf_pending_run();
2777	}
2778
2779	/*
2780	* Callchain support -- arch specific
2781	*/
2782
2783	__weak struct perf_callchain_entry perf_callchain(struct pt_regs regs)
2784	{
2785	return NULL;
2786	}
2787
2788	__weak
2789	void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip)
2790	{
2791	}
2792
2793
2794	/*
2795	* Output
2796	*/
2797	static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2798	unsigned long offset, unsigned long head)
2799	{
2800	unsigned long mask;
2801
2802	if (!data->writable)
2803	return true;
2804
2805	mask = perf_data_size(data) - 1;
2806
2807	offset = (offset - tail) & mask;
2808	head = (head - tail) & mask;
2809
2810	if ((int)(head - offset) < 0)
2811	return false;
2812
2813	return true;
2814	}
2815
2816	static void perf_output_wakeup(struct perf_output_handle *handle)
2817	{
2818	atomic_set(&handle->data->poll, POLL_IN);
2819
2820	if (handle->nmi) {
2821	handle->event->pending_wakeup = 1;
2822	perf_pending_queue(&handle->event->pending,
2823	perf_pending_event);
2824	} else
2825	perf_event_wakeup(handle->event);
2826	}
2827
2828	/*
2829	* Curious locking construct.
2830	*
2831	* We need to ensure a later event_id doesn't publish a head when a former
2832	* event_id isn't done writing. However since we need to deal with NMIs we
2833	* cannot fully serialize things.
2834	*
2835	* What we do is serialize between CPUs so we only have to deal with NMI
2836	* nesting on a single CPU.
2837	*
2838	* We only publish the head (and generate a wakeup) when the outer-most
2839	* event_id completes.
2840	*/
2841	static void perf_output_lock(struct perf_output_handle *handle)
2842	{
2843	struct perf_mmap_data *data = handle->data;
2844	int cur, cpu = get_cpu();
2845
2846	handle->locked = 0;
2847
2848	for (;;) {
2849	cur = atomic_cmpxchg(&data->lock, -1, cpu);
2850	if (cur == -1) {
2851	handle->locked = 1;
2852	break;
2853	}
2854	if (cur == cpu)
2855	break;
2856
2857	cpu_relax();
2858	}
2859	}
2860
2861	static void perf_output_unlock(struct perf_output_handle *handle)
2862	{
2863	struct perf_mmap_data *data = handle->data;
2864	unsigned long head;
2865	int cpu;
2866
2867	data->done_head = data->head;
2868
2869	if (!handle->locked)
2870	goto out;
2871
2872	again:
2873	/*
2874	* The xchg implies a full barrier that ensures all writes are done
2875	* before we publish the new head, matched by a rmb() in userspace when
2876	* reading this position.
2877	*/
2878	while ((head = atomic_long_xchg(&data->done_head, 0)))
2879	data->user_page->data_head = head;
2880
2881	/*
2882	* NMI can happen here, which means we can miss a done_head update.
2883	*/
2884
2885	cpu = atomic_xchg(&data->lock, -1);
2886	WARN_ON_ONCE(cpu != smp_processor_id());
2887
2888	/*
2889	* Therefore we have to validate we did not indeed do so.
2890	*/
2891	if (unlikely(atomic_long_read(&data->done_head))) {
2892	/*
2893	* Since we had it locked, we can lock it again.
2894	*/
2895	while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2896	cpu_relax();
2897
2898	goto again;
2899	}
2900
2901	if (atomic_xchg(&data->wakeup, 0))
2902	perf_output_wakeup(handle);
2903	out:
2904	put_cpu();
2905	}
2906
2907	void perf_output_copy(struct perf_output_handle *handle,
2908	const void *buf, unsigned int len)
2909	{
2910	unsigned int pages_mask;
2911	unsigned long offset;
2912	unsigned int size;
2913	void **pages;
2914
2915	offset = handle->offset;
2916	pages_mask = handle->data->nr_pages - 1;
2917	pages = handle->data->data_pages;
2918
2919	do {
2920	unsigned long page_offset;
2921	unsigned long page_size;
2922	int nr;
2923
2924	nr = (offset >> PAGE_SHIFT) & pages_mask;
2925	page_size = 1UL << (handle->data->data_order + PAGE_SHIFT);
2926	page_offset = offset & (page_size - 1);
2927	size = min_t(unsigned int, page_size - page_offset, len);
2928
2929	memcpy(pages[nr] + page_offset, buf, size);
2930
2931	len -= size;
2932	buf += size;
2933	offset += size;
2934	} while (len);
2935
2936	handle->offset = offset;
2937
2938	/*
2939	* Check we didn't copy past our reservation window, taking the
2940	* possible unsigned int wrap into account.
2941	*/
2942	WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2943	}
2944
2945	int perf_output_begin(struct perf_output_handle *handle,
2946	struct perf_event *event, unsigned int size,
2947	int nmi, int sample)
2948	{
2949	struct perf_event *output_event;
2950	struct perf_mmap_data *data;
2951	unsigned long tail, offset, head;
2952	int have_lost;
2953	struct {
2954	struct perf_event_header header;
2955	u64 id;
2956	u64 lost;
2957	} lost_event;
2958
2959	rcu_read_lock();
2960	/*
2961	* For inherited events we send all the output towards the parent.
2962	*/
2963	if (event->parent)
2964	event = event->parent;
2965
2966	output_event = rcu_dereference(event->output);
2967	if (output_event)
2968	event = output_event;
2969
2970	data = rcu_dereference(event->data);
2971	if (!data)
2972	goto out;
2973
2974	handle->data = data;
2975	handle->event = event;
2976	handle->nmi = nmi;
2977	handle->sample = sample;
2978
2979	if (!data->nr_pages)
2980	goto fail;
2981
2982	have_lost = atomic_read(&data->lost);
2983	if (have_lost)
2984	size += sizeof(lost_event);
2985
2986	perf_output_lock(handle);
2987
2988	do {
2989	/*
2990	* Userspace could choose to issue a mb() before updating the
2991	* tail pointer. So that all reads will be completed before the
2992	* write is issued.
2993	*/
2994	tail = ACCESS_ONCE(data->user_page->data_tail);
2995	smp_rmb();
2996	offset = head = atomic_long_read(&data->head);
2997	head += size;
2998	if (unlikely(!perf_output_space(data, tail, offset, head)))
2999	goto fail;
3000	} while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
3001
3002	handle->offset = offset;
3003	handle->head = head;
3004
3005	if (head - tail > data->watermark)
3006	atomic_set(&data->wakeup, 1);
3007
3008	if (have_lost) {
3009	lost_event.header.type = PERF_RECORD_LOST;
3010	lost_event.header.misc = 0;
3011	lost_event.header.size = sizeof(lost_event);
3012	lost_event.id = event->id;
3013	lost_event.lost = atomic_xchg(&data->lost, 0);
3014
3015	perf_output_put(handle, lost_event);
3016	}
3017
3018	return 0;
3019
3020	fail:
3021	atomic_inc(&data->lost);
3022	perf_output_unlock(handle);
3023	out:
3024	rcu_read_unlock();
3025
3026	return -ENOSPC;
3027	}
3028
3029	void perf_output_end(struct perf_output_handle *handle)
3030	{
3031	struct perf_event *event = handle->event;
3032	struct perf_mmap_data *data = handle->data;
3033
3034	int wakeup_events = event->attr.wakeup_events;
3035
3036	if (handle->sample && wakeup_events) {
3037	int events = atomic_inc_return(&data->events);
3038	if (events >= wakeup_events) {
3039	atomic_sub(wakeup_events, &data->events);
3040	atomic_set(&data->wakeup, 1);
3041	}
3042	}
3043
3044	perf_output_unlock(handle);
3045	rcu_read_unlock();
3046	}
3047
3048	static u32 perf_event_pid(struct perf_event event, struct task_struct p)
3049	{
3050	/*
3051	* only top level events have the pid namespace they were created in
3052	*/
3053	if (event->parent)
3054	event = event->parent;
3055
3056	return task_tgid_nr_ns(p, event->ns);
3057	}
3058
3059	static u32 perf_event_tid(struct perf_event event, struct task_struct p)
3060	{
3061	/*
3062	* only top level events have the pid namespace they were created in
3063	*/
3064	if (event->parent)
3065	event = event->parent;
3066
3067	return task_pid_nr_ns(p, event->ns);
3068	}
3069
3070	static void perf_output_read_one(struct perf_output_handle *handle,
3071	struct perf_event *event)
3072	{
3073	u64 read_format = event->attr.read_format;
3074	u64 values[4];
3075	int n = 0;
3076
3077	values[n++] = atomic64_read(&event->count);
3078	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3079	values[n++] = event->total_time_enabled +
3080	atomic64_read(&event->child_total_time_enabled);
3081	}
3082	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3083	values[n++] = event->total_time_running +
3084	atomic64_read(&event->child_total_time_running);
3085	}
3086	if (read_format & PERF_FORMAT_ID)
3087	values[n++] = primary_event_id(event);
3088
3089	perf_output_copy(handle, values, n * sizeof(u64));
3090	}
3091
3092	/*
3093	* XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3094	*/
3095	static void perf_output_read_group(struct perf_output_handle *handle,
3096	struct perf_event *event)
3097	{
3098	struct perf_event leader = event->group_leader, sub;
3099	u64 read_format = event->attr.read_format;
3100	u64 values[5];
3101	int n = 0;
3102
3103	values[n++] = 1 + leader->nr_siblings;
3104
3105	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3106	values[n++] = leader->total_time_enabled;
3107
3108	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3109	values[n++] = leader->total_time_running;
3110
3111	if (leader != event)
3112	leader->pmu->read(leader);
3113
3114	values[n++] = atomic64_read(&leader->count);
3115	if (read_format & PERF_FORMAT_ID)
3116	values[n++] = primary_event_id(leader);
3117
3118	perf_output_copy(handle, values, n * sizeof(u64));
3119
3120	list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3121	n = 0;
3122
3123	if (sub != event)
3124	sub->pmu->read(sub);
3125
3126	values[n++] = atomic64_read(&sub->count);
3127	if (read_format & PERF_FORMAT_ID)
3128	values[n++] = primary_event_id(sub);
3129
3130	perf_output_copy(handle, values, n * sizeof(u64));
3131	}
3132	}
3133
3134	static void perf_output_read(struct perf_output_handle *handle,
3135	struct perf_event *event)
3136	{
3137	if (event->attr.read_format & PERF_FORMAT_GROUP)
3138	perf_output_read_group(handle, event);
3139	else
3140	perf_output_read_one(handle, event);
3141	}
3142
3143	void perf_output_sample(struct perf_output_handle *handle,
3144	struct perf_event_header *header,
3145	struct perf_sample_data *data,
3146	struct perf_event *event)
3147	{
3148	u64 sample_type = data->type;
3149
3150	perf_output_put(handle, *header);
3151
3152	if (sample_type & PERF_SAMPLE_IP)
3153	perf_output_put(handle, data->ip);
3154
3155	if (sample_type & PERF_SAMPLE_TID)
3156	perf_output_put(handle, data->tid_entry);
3157
3158	if (sample_type & PERF_SAMPLE_TIME)
3159	perf_output_put(handle, data->time);
3160
3161	if (sample_type & PERF_SAMPLE_ADDR)
3162	perf_output_put(handle, data->addr);
3163
3164	if (sample_type & PERF_SAMPLE_ID)
3165	perf_output_put(handle, data->id);
3166
3167	if (sample_type & PERF_SAMPLE_STREAM_ID)
3168	perf_output_put(handle, data->stream_id);
3169
3170	if (sample_type & PERF_SAMPLE_CPU)
3171	perf_output_put(handle, data->cpu_entry);
3172
3173	if (sample_type & PERF_SAMPLE_PERIOD)
3174	perf_output_put(handle, data->period);
3175
3176	if (sample_type & PERF_SAMPLE_READ)
3177	perf_output_read(handle, event);
3178
3179	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3180	if (data->callchain) {
3181	int size = 1;
3182
3183	if (data->callchain)
3184	size += data->callchain->nr;
3185
3186	size *= sizeof(u64);
3187
3188	perf_output_copy(handle, data->callchain, size);
3189	} else {
3190	u64 nr = 0;
3191	perf_output_put(handle, nr);
3192	}
3193	}
3194
3195	if (sample_type & PERF_SAMPLE_RAW) {
3196	if (data->raw) {
3197	perf_output_put(handle, data->raw->size);
3198	perf_output_copy(handle, data->raw->data,
3199	data->raw->size);
3200	} else {
3201	struct {
3202	u32 size;
3203	u32 data;
3204	} raw = {
3205	.size = sizeof(u32),
3206	.data = 0,
3207	};
3208	perf_output_put(handle, raw);
3209	}
3210	}
3211	}
3212
3213	void perf_prepare_sample(struct perf_event_header *header,
3214	struct perf_sample_data *data,
3215	struct perf_event *event,
3216	struct pt_regs *regs)
3217	{
3218	u64 sample_type = event->attr.sample_type;
3219
3220	data->type = sample_type;
3221
3222	header->type = PERF_RECORD_SAMPLE;
3223	header->size = sizeof(*header);
3224
3225	header->misc = 0;
3226	header->misc \|= perf_misc_flags(regs);
3227
3228	if (sample_type & PERF_SAMPLE_IP) {
3229	data->ip = perf_instruction_pointer(regs);
3230
3231	header->size += sizeof(data->ip);
3232	}
3233
3234	if (sample_type & PERF_SAMPLE_TID) {
3235	/* namespace issues */
3236	data->tid_entry.pid = perf_event_pid(event, current);
3237	data->tid_entry.tid = perf_event_tid(event, current);
3238
3239	header->size += sizeof(data->tid_entry);
3240	}
3241
3242	if (sample_type & PERF_SAMPLE_TIME) {
3243	data->time = perf_clock();
3244
3245	header->size += sizeof(data->time);
3246	}
3247
3248	if (sample_type & PERF_SAMPLE_ADDR)
3249	header->size += sizeof(data->addr);
3250
3251	if (sample_type & PERF_SAMPLE_ID) {
3252	data->id = primary_event_id(event);
3253
3254	header->size += sizeof(data->id);
3255	}
3256
3257	if (sample_type & PERF_SAMPLE_STREAM_ID) {
3258	data->stream_id = event->id;
3259
3260	header->size += sizeof(data->stream_id);
3261	}
3262
3263	if (sample_type & PERF_SAMPLE_CPU) {
3264	data->cpu_entry.cpu = raw_smp_processor_id();
3265	data->cpu_entry.reserved = 0;
3266
3267	header->size += sizeof(data->cpu_entry);
3268	}
3269
3270	if (sample_type & PERF_SAMPLE_PERIOD)
3271	header->size += sizeof(data->period);
3272
3273	if (sample_type & PERF_SAMPLE_READ)
3274	header->size += perf_event_read_size(event);
3275
3276	if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3277	int size = 1;
3278
3279	data->callchain = perf_callchain(regs);
3280
3281	if (data->callchain)
3282	size += data->callchain->nr;
3283
3284	header->size += size * sizeof(u64);
3285	}
3286
3287	if (sample_type & PERF_SAMPLE_RAW) {
3288	int size = sizeof(u32);
3289
3290	if (data->raw)
3291	size += data->raw->size;
3292	else
3293	size += sizeof(u32);
3294
3295	WARN_ON_ONCE(size & (sizeof(u64)-1));
3296	header->size += size;
3297	}
3298	}
3299
3300	static void perf_event_output(struct perf_event *event, int nmi,
3301	struct perf_sample_data *data,
3302	struct pt_regs *regs)
3303	{
3304	struct perf_output_handle handle;
3305	struct perf_event_header header;
3306
3307	perf_prepare_sample(&header, data, event, regs);
3308
3309	if (perf_output_begin(&handle, event, header.size, nmi, 1))
3310	return;
3311
3312	perf_output_sample(&handle, &header, data, event);
3313
3314	perf_output_end(&handle);
3315	}
3316
3317	/*
3318	* read event_id
3319	*/
3320
3321	struct perf_read_event {
3322	struct perf_event_header header;
3323
3324	u32 pid;
3325	u32 tid;
3326	};
3327
3328	static void
3329	perf_event_read_event(struct perf_event *event,
3330	struct task_struct *task)
3331	{
3332	struct perf_output_handle handle;
3333	struct perf_read_event read_event = {
3334	.header = {
3335	.type = PERF_RECORD_READ,
3336	.misc = 0,
3337	.size = sizeof(read_event) + perf_event_read_size(event),
3338	},
3339	.pid = perf_event_pid(event, task),
3340	.tid = perf_event_tid(event, task),
3341	};
3342	int ret;
3343
3344	ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3345	if (ret)
3346	return;
3347
3348	perf_output_put(&handle, read_event);
3349	perf_output_read(&handle, event);
3350
3351	perf_output_end(&handle);
3352	}
3353
3354	/*
3355	* task tracking -- fork/exit
3356	*
3357	* enabled by: attr.comm \| attr.mmap \| attr.task
3358	*/
3359
3360	struct perf_task_event {
3361	struct task_struct *task;
3362	struct perf_event_context *task_ctx;
3363
3364	struct {
3365	struct perf_event_header header;
3366
3367	u32 pid;
3368	u32 ppid;
3369	u32 tid;
3370	u32 ptid;
3371	u64 time;
3372	} event_id;
3373	};
3374
3375	static void perf_event_task_output(struct perf_event *event,
3376	struct perf_task_event *task_event)
3377	{
3378	struct perf_output_handle handle;
3379	struct task_struct *task = task_event->task;
3380	unsigned long flags;
3381	int size, ret;
3382
3383	/*
3384	* If this CPU attempts to acquire an rq lock held by a CPU spinning
3385	* in perf_output_lock() from interrupt context, it's game over.
3386	*/
3387	local_irq_save(flags);
3388
3389	size = task_event->event_id.header.size;
3390	ret = perf_output_begin(&handle, event, size, 0, 0);
3391
3392	if (ret) {
3393	local_irq_restore(flags);
3394	return;
3395	}
3396
3397	task_event->event_id.pid = perf_event_pid(event, task);
3398	task_event->event_id.ppid = perf_event_pid(event, current);
3399
3400	task_event->event_id.tid = perf_event_tid(event, task);
3401	task_event->event_id.ptid = perf_event_tid(event, current);
3402
3403	perf_output_put(&handle, task_event->event_id);
3404
3405	perf_output_end(&handle);
3406	local_irq_restore(flags);
3407	}
3408
3409	static int perf_event_task_match(struct perf_event *event)
3410	{
3411	if (event->state < PERF_EVENT_STATE_INACTIVE)
3412	return 0;
3413
3414	if (event->cpu != -1 && event->cpu != smp_processor_id())
3415	return 0;
3416
3417	if (event->attr.comm \|\| event->attr.mmap \|\| event->attr.task)
3418	return 1;
3419
3420	return 0;
3421	}
3422
3423	static void perf_event_task_ctx(struct perf_event_context *ctx,
3424	struct perf_task_event *task_event)
3425	{
3426	struct perf_event *event;
3427
3428	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3429	if (perf_event_task_match(event))
3430	perf_event_task_output(event, task_event);
3431	}
3432	}
3433
3434	static void perf_event_task_event(struct perf_task_event *task_event)
3435	{
3436	struct perf_cpu_context *cpuctx;
3437	struct perf_event_context *ctx = task_event->task_ctx;
3438
3439	rcu_read_lock();
3440	cpuctx = &get_cpu_var(perf_cpu_context);
3441	perf_event_task_ctx(&cpuctx->ctx, task_event);
3442	if (!ctx)
3443	ctx = rcu_dereference(current->perf_event_ctxp);
3444	if (ctx)
3445	perf_event_task_ctx(ctx, task_event);
3446	put_cpu_var(perf_cpu_context);
3447	rcu_read_unlock();
3448	}
3449
3450	static void perf_event_task(struct task_struct *task,
3451	struct perf_event_context *task_ctx,
3452	int new)
3453	{
3454	struct perf_task_event task_event;
3455
3456	if (!atomic_read(&nr_comm_events) &&
3457	!atomic_read(&nr_mmap_events) &&
3458	!atomic_read(&nr_task_events))
3459	return;
3460
3461	task_event = (struct perf_task_event){
3462	.task = task,
3463	.task_ctx = task_ctx,
3464	.event_id = {
3465	.header = {
3466	.type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3467	.misc = 0,
3468	.size = sizeof(task_event.event_id),
3469	},
3470	/* .pid */
3471	/* .ppid */
3472	/* .tid */
3473	/* .ptid */
3474	.time = perf_clock(),
3475	},
3476	};
3477
3478	perf_event_task_event(&task_event);
3479	}
3480
3481	void perf_event_fork(struct task_struct *task)
3482	{
3483	perf_event_task(task, NULL, 1);
3484	}
3485
3486	/*
3487	* comm tracking
3488	*/
3489
3490	struct perf_comm_event {
3491	struct task_struct *task;
3492	char *comm;
3493	int comm_size;
3494
3495	struct {
3496	struct perf_event_header header;
3497
3498	u32 pid;
3499	u32 tid;
3500	} event_id;
3501	};
3502
3503	static void perf_event_comm_output(struct perf_event *event,
3504	struct perf_comm_event *comm_event)
3505	{
3506	struct perf_output_handle handle;
3507	int size = comm_event->event_id.header.size;
3508	int ret = perf_output_begin(&handle, event, size, 0, 0);
3509
3510	if (ret)
3511	return;
3512
3513	comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3514	comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3515
3516	perf_output_put(&handle, comm_event->event_id);
3517	perf_output_copy(&handle, comm_event->comm,
3518	comm_event->comm_size);
3519	perf_output_end(&handle);
3520	}
3521
3522	static int perf_event_comm_match(struct perf_event *event)
3523	{
3524	if (event->state < PERF_EVENT_STATE_INACTIVE)
3525	return 0;
3526
3527	if (event->cpu != -1 && event->cpu != smp_processor_id())
3528	return 0;
3529
3530	if (event->attr.comm)
3531	return 1;
3532
3533	return 0;
3534	}
3535
3536	static void perf_event_comm_ctx(struct perf_event_context *ctx,
3537	struct perf_comm_event *comm_event)
3538	{
3539	struct perf_event *event;
3540
3541	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3542	if (perf_event_comm_match(event))
3543	perf_event_comm_output(event, comm_event);
3544	}
3545	}
3546
3547	static void perf_event_comm_event(struct perf_comm_event *comm_event)
3548	{
3549	struct perf_cpu_context *cpuctx;
3550	struct perf_event_context *ctx;
3551	unsigned int size;
3552	char comm[TASK_COMM_LEN];
3553
3554	memset(comm, 0, sizeof(comm));
3555	strlcpy(comm, comm_event->task->comm, sizeof(comm));
3556	size = ALIGN(strlen(comm)+1, sizeof(u64));
3557
3558	comm_event->comm = comm;
3559	comm_event->comm_size = size;
3560
3561	comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3562
3563	rcu_read_lock();
3564	cpuctx = &get_cpu_var(perf_cpu_context);
3565	perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3566	ctx = rcu_dereference(current->perf_event_ctxp);
3567	if (ctx)
3568	perf_event_comm_ctx(ctx, comm_event);
3569	put_cpu_var(perf_cpu_context);
3570	rcu_read_unlock();
3571	}
3572
3573	void perf_event_comm(struct task_struct *task)
3574	{
3575	struct perf_comm_event comm_event;
3576
3577	if (task->perf_event_ctxp)
3578	perf_event_enable_on_exec(task);
3579
3580	if (!atomic_read(&nr_comm_events))
3581	return;
3582
3583	comm_event = (struct perf_comm_event){
3584	.task = task,
3585	/* .comm */
3586	/* .comm_size */
3587	.event_id = {
3588	.header = {
3589	.type = PERF_RECORD_COMM,
3590	.misc = 0,
3591	/* .size */
3592	},
3593	/* .pid */
3594	/* .tid */
3595	},
3596	};
3597
3598	perf_event_comm_event(&comm_event);
3599	}
3600
3601	/*
3602	* mmap tracking
3603	*/
3604
3605	struct perf_mmap_event {
3606	struct vm_area_struct *vma;
3607
3608	const char *file_name;
3609	int file_size;
3610
3611	struct {
3612	struct perf_event_header header;
3613
3614	u32 pid;
3615	u32 tid;
3616	u64 start;
3617	u64 len;
3618	u64 pgoff;
3619	} event_id;
3620	};
3621
3622	static void perf_event_mmap_output(struct perf_event *event,
3623	struct perf_mmap_event *mmap_event)
3624	{
3625	struct perf_output_handle handle;
3626	int size = mmap_event->event_id.header.size;
3627	int ret = perf_output_begin(&handle, event, size, 0, 0);
3628
3629	if (ret)
3630	return;
3631
3632	mmap_event->event_id.pid = perf_event_pid(event, current);
3633	mmap_event->event_id.tid = perf_event_tid(event, current);
3634
3635	perf_output_put(&handle, mmap_event->event_id);
3636	perf_output_copy(&handle, mmap_event->file_name,
3637	mmap_event->file_size);
3638	perf_output_end(&handle);
3639	}
3640
3641	static int perf_event_mmap_match(struct perf_event *event,
3642	struct perf_mmap_event *mmap_event)
3643	{
3644	if (event->state < PERF_EVENT_STATE_INACTIVE)
3645	return 0;
3646
3647	if (event->cpu != -1 && event->cpu != smp_processor_id())
3648	return 0;
3649
3650	if (event->attr.mmap)
3651	return 1;
3652
3653	return 0;
3654	}
3655
3656	static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3657	struct perf_mmap_event *mmap_event)
3658	{
3659	struct perf_event *event;
3660
3661	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3662	if (perf_event_mmap_match(event, mmap_event))
3663	perf_event_mmap_output(event, mmap_event);
3664	}
3665	}
3666
3667	static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3668	{
3669	struct perf_cpu_context *cpuctx;
3670	struct perf_event_context *ctx;
3671	struct vm_area_struct *vma = mmap_event->vma;
3672	struct file *file = vma->vm_file;
3673	unsigned int size;
3674	char tmp[16];
3675	char *buf = NULL;
3676	const char *name;
3677
3678	memset(tmp, 0, sizeof(tmp));
3679
3680	if (file) {
3681	/*
3682	* d_path works from the end of the buffer backwards, so we
3683	* need to add enough zero bytes after the string to handle
3684	* the 64bit alignment we do later.
3685	*/
3686	buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3687	if (!buf) {
3688	name = strncpy(tmp, "//enomem", sizeof(tmp));
3689	goto got_name;
3690	}
3691	name = d_path(&file->f_path, buf, PATH_MAX);
3692	if (IS_ERR(name)) {
3693	name = strncpy(tmp, "//toolong", sizeof(tmp));
3694	goto got_name;
3695	}
3696	} else {
3697	if (arch_vma_name(mmap_event->vma)) {
3698	name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3699	sizeof(tmp));
3700	goto got_name;
3701	}
3702
3703	if (!vma->vm_mm) {
3704	name = strncpy(tmp, "[vdso]", sizeof(tmp));
3705	goto got_name;
3706	}
3707
3708	name = strncpy(tmp, "//anon", sizeof(tmp));
3709	goto got_name;
3710	}
3711
3712	got_name:
3713	size = ALIGN(strlen(name)+1, sizeof(u64));
3714
3715	mmap_event->file_name = name;
3716	mmap_event->file_size = size;
3717
3718	mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3719
3720	rcu_read_lock();
3721	cpuctx = &get_cpu_var(perf_cpu_context);
3722	perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3723	ctx = rcu_dereference(current->perf_event_ctxp);
3724	if (ctx)
3725	perf_event_mmap_ctx(ctx, mmap_event);
3726	put_cpu_var(perf_cpu_context);
3727	rcu_read_unlock();
3728
3729	kfree(buf);
3730	}
3731
3732	void __perf_event_mmap(struct vm_area_struct *vma)
3733	{
3734	struct perf_mmap_event mmap_event;
3735
3736	if (!atomic_read(&nr_mmap_events))
3737	return;
3738
3739	mmap_event = (struct perf_mmap_event){
3740	.vma = vma,
3741	/* .file_name */
3742	/* .file_size */
3743	.event_id = {
3744	.header = {
3745	.type = PERF_RECORD_MMAP,
3746	.misc = 0,
3747	/* .size */
3748	},
3749	/* .pid */
3750	/* .tid */
3751	.start = vma->vm_start,
3752	.len = vma->vm_end - vma->vm_start,
3753	.pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
3754	},
3755	};
3756
3757	perf_event_mmap_event(&mmap_event);
3758	}
3759
3760	/*
3761	* IRQ throttle logging
3762	*/
3763
3764	static void perf_log_throttle(struct perf_event *event, int enable)
3765	{
3766	struct perf_output_handle handle;
3767	int ret;
3768
3769	struct {
3770	struct perf_event_header header;
3771	u64 time;
3772	u64 id;
3773	u64 stream_id;
3774	} throttle_event = {
3775	.header = {
3776	.type = PERF_RECORD_THROTTLE,
3777	.misc = 0,
3778	.size = sizeof(throttle_event),
3779	},
3780	.time = perf_clock(),
3781	.id = primary_event_id(event),
3782	.stream_id = event->id,
3783	};
3784
3785	if (enable)
3786	throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3787
3788	ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3789	if (ret)
3790	return;
3791
3792	perf_output_put(&handle, throttle_event);
3793	perf_output_end(&handle);
3794	}
3795
3796	/*
3797	* Generic event overflow handling, sampling.
3798	*/
3799
3800	static int __perf_event_overflow(struct perf_event *event, int nmi,
3801	int throttle, struct perf_sample_data *data,
3802	struct pt_regs *regs)
3803	{
3804	int events = atomic_read(&event->event_limit);
3805	struct hw_perf_event *hwc = &event->hw;
3806	int ret = 0;
3807
3808	throttle = (throttle && event->pmu->unthrottle != NULL);
3809
3810	if (!throttle) {
3811	hwc->interrupts++;
3812	} else {
3813	if (hwc->interrupts != MAX_INTERRUPTS) {
3814	hwc->interrupts++;
3815	if (HZ * hwc->interrupts >
3816	(u64)sysctl_perf_event_sample_rate) {
3817	hwc->interrupts = MAX_INTERRUPTS;
3818	perf_log_throttle(event, 0);
3819	ret = 1;
3820	}
3821	} else {
3822	/*
3823	* Keep re-disabling events even though on the previous
3824	* pass we disabled it - just in case we raced with a
3825	* sched-in and the event got enabled again:
3826	*/
3827	ret = 1;
3828	}
3829	}
3830
3831	if (event->attr.freq) {
3832	u64 now = perf_clock();
3833	s64 delta = now - hwc->freq_time_stamp;
3834
3835	hwc->freq_time_stamp = now;
3836
3837	if (delta > 0 && delta < 2*TICK_NSEC)
3838	perf_adjust_period(event, delta, hwc->last_period);
3839	}
3840
3841	/*
3842	* XXX event_limit might not quite work as expected on inherited
3843	* events
3844	*/
3845
3846	event->pending_kill = POLL_IN;
3847	if (events && atomic_dec_and_test(&event->event_limit)) {
3848	ret = 1;
3849	event->pending_kill = POLL_HUP;
3850	if (nmi) {
3851	event->pending_disable = 1;
3852	perf_pending_queue(&event->pending,
3853	perf_pending_event);
3854	} else
3855	perf_event_disable(event);
3856	}
3857
3858	if (event->overflow_handler)
3859	event->overflow_handler(event, nmi, data, regs);
3860	else
3861	perf_event_output(event, nmi, data, regs);
3862
3863	return ret;
3864	}
3865
3866	int perf_event_overflow(struct perf_event *event, int nmi,
3867	struct perf_sample_data *data,
3868	struct pt_regs *regs)
3869	{
3870	return __perf_event_overflow(event, nmi, 1, data, regs);
3871	}
3872
3873	/*
3874	* Generic software event infrastructure
3875	*/
3876
3877	/*
3878	* We directly increment event->count and keep a second value in
3879	* event->hw.period_left to count intervals. This period event
3880	* is kept in the range [-sample_period, 0] so that we can use the
3881	* sign as trigger.
3882	*/
3883
3884	static u64 perf_swevent_set_period(struct perf_event *event)
3885	{
3886	struct hw_perf_event *hwc = &event->hw;
3887	u64 period = hwc->last_period;
3888	u64 nr, offset;
3889	s64 old, val;
3890
3891	hwc->last_period = hwc->sample_period;
3892
3893	again:
3894	old = val = atomic64_read(&hwc->period_left);
3895	if (val < 0)
3896	return 0;
3897
3898	nr = div64_u64(period + val, period);
3899	offset = nr * period;
3900	val -= offset;
3901	if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3902	goto again;
3903
3904	return nr;
3905	}
3906
3907	static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
3908	int nmi, struct perf_sample_data *data,
3909	struct pt_regs *regs)
3910	{
3911	struct hw_perf_event *hwc = &event->hw;
3912	int throttle = 0;
3913
3914	data->period = event->hw.last_period;
3915	if (!overflow)
3916	overflow = perf_swevent_set_period(event);
3917
3918	if (hwc->interrupts == MAX_INTERRUPTS)
3919	return;
3920
3921	for (; overflow; overflow--) {
3922	if (__perf_event_overflow(event, nmi, throttle,
3923	data, regs)) {
3924	/*
3925	* We inhibit the overflow from happening when
3926	* hwc->interrupts == MAX_INTERRUPTS.
3927	*/
3928	break;
3929	}
3930	throttle = 1;
3931	}
3932	}
3933
3934	static void perf_swevent_unthrottle(struct perf_event *event)
3935	{
3936	/*
3937	* Nothing to do, we already reset hwc->interrupts.
3938	*/
3939	}
3940
3941	static void perf_swevent_add(struct perf_event *event, u64 nr,
3942	int nmi, struct perf_sample_data *data,
3943	struct pt_regs *regs)
3944	{
3945	struct hw_perf_event *hwc = &event->hw;
3946
3947	atomic64_add(nr, &event->count);
3948
3949	if (!regs)
3950	return;
3951
3952	if (!hwc->sample_period)
3953	return;
3954
3955	if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
3956	return perf_swevent_overflow(event, 1, nmi, data, regs);
3957
3958	if (atomic64_add_negative(nr, &hwc->period_left))
3959	return;
3960
3961	perf_swevent_overflow(event, 0, nmi, data, regs);
3962	}
3963
3964	static int perf_swevent_is_counting(struct perf_event *event)
3965	{
3966	/*
3967	* The event is active, we're good!
3968	*/
3969	if (event->state == PERF_EVENT_STATE_ACTIVE)
3970	return 1;
3971
3972	/*
3973	* The event is off/error, not counting.
3974	*/
3975	if (event->state != PERF_EVENT_STATE_INACTIVE)
3976	return 0;
3977
3978	/*
3979	* The event is inactive, if the context is active
3980	* we're part of a group that didn't make it on the 'pmu',
3981	* not counting.
3982	*/
3983	if (event->ctx->is_active)
3984	return 0;
3985
3986	/*
3987	* We're inactive and the context is too, this means the
3988	* task is scheduled out, we're counting events that happen
3989	* to us, like migration events.
3990	*/
3991	return 1;
3992	}
3993
3994	static int perf_tp_event_match(struct perf_event *event,
3995	struct perf_sample_data *data);
3996
3997	static int perf_exclude_event(struct perf_event *event,
3998	struct pt_regs *regs)
3999	{
4000	if (regs) {
4001	if (event->attr.exclude_user && user_mode(regs))
4002	return 1;
4003
4004	if (event->attr.exclude_kernel && !user_mode(regs))
4005	return 1;
4006	}
4007
4008	return 0;
4009	}
4010
4011	static int perf_swevent_match(struct perf_event *event,
4012	enum perf_type_id type,
4013	u32 event_id,
4014	struct perf_sample_data *data,
4015	struct pt_regs *regs)
4016	{
4017	if (event->cpu != -1 && event->cpu != smp_processor_id())
4018	return 0;
4019
4020	if (!perf_swevent_is_counting(event))
4021	return 0;
4022
4023	if (event->attr.type != type)
4024	return 0;
4025
4026	if (event->attr.config != event_id)
4027	return 0;
4028
4029	if (perf_exclude_event(event, regs))
4030	return 0;
4031
4032	if (event->attr.type == PERF_TYPE_TRACEPOINT &&
4033	!perf_tp_event_match(event, data))
4034	return 0;
4035
4036	return 1;
4037	}
4038
4039	static void perf_swevent_ctx_event(struct perf_event_context *ctx,
4040	enum perf_type_id type,
4041	u32 event_id, u64 nr, int nmi,
4042	struct perf_sample_data *data,
4043	struct pt_regs *regs)
4044	{
4045	struct perf_event *event;
4046
4047	list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4048	if (perf_swevent_match(event, type, event_id, data, regs))
4049	perf_swevent_add(event, nr, nmi, data, regs);
4050	}
4051	}
4052
4053	int perf_swevent_get_recursion_context(void)
4054	{
4055	struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
4056	int rctx;
4057
4058	if (in_nmi())
4059	rctx = 3;
4060	else if (in_irq())
4061	rctx = 2;
4062	else if (in_softirq())
4063	rctx = 1;
4064	else
4065	rctx = 0;
4066
4067	if (cpuctx->recursion[rctx]) {
4068	put_cpu_var(perf_cpu_context);
4069	return -1;
4070	}
4071
4072	cpuctx->recursion[rctx]++;
4073	barrier();
4074
4075	return rctx;
4076	}
4077	EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4078
4079	void perf_swevent_put_recursion_context(int rctx)
4080	{
4081	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
4082	barrier();
4083	cpuctx->recursion[rctx]--;
4084	put_cpu_var(perf_cpu_context);
4085	}
4086	EXPORT_SYMBOL_GPL(perf_swevent_put_recursion_context);
4087
4088	static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4089	u64 nr, int nmi,
4090	struct perf_sample_data *data,
4091	struct pt_regs *regs)
4092	{
4093	struct perf_cpu_context *cpuctx;
4094	struct perf_event_context *ctx;
4095
4096	cpuctx = &__get_cpu_var(perf_cpu_context);
4097	rcu_read_lock();
4098	perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
4099	nr, nmi, data, regs);
4100	/*
4101	* doesn't really matter which of the child contexts the
4102	* events ends up in.
4103	*/
4104	ctx = rcu_dereference(current->perf_event_ctxp);
4105	if (ctx)
4106	perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
4107	rcu_read_unlock();
4108	}
4109
4110	void __perf_sw_event(u32 event_id, u64 nr, int nmi,
4111	struct pt_regs *regs, u64 addr)
4112	{
4113	struct perf_sample_data data;
4114	int rctx;
4115
4116	rctx = perf_swevent_get_recursion_context();
4117	if (rctx < 0)
4118	return;
4119
4120	perf_sample_data_init(&data, addr);
4121
4122	do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
4123
4124	perf_swevent_put_recursion_context(rctx);
4125	}
4126
4127	static void perf_swevent_read(struct perf_event *event)
4128	{
4129	}
4130
4131	static int perf_swevent_enable(struct perf_event *event)
4132	{
4133	struct hw_perf_event *hwc = &event->hw;
4134
4135	if (hwc->sample_period) {
4136	hwc->last_period = hwc->sample_period;
4137	perf_swevent_set_period(event);
4138	}
4139	return 0;
4140	}
4141
4142	static void perf_swevent_disable(struct perf_event *event)
4143	{
4144	}
4145
4146	static const struct pmu perf_ops_generic = {
4147	.enable = perf_swevent_enable,
4148	.disable = perf_swevent_disable,
4149	.read = perf_swevent_read,
4150	.unthrottle = perf_swevent_unthrottle,
4151	};
4152
4153	/*
4154	* hrtimer based swevent callback
4155	*/
4156
4157	static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
4158	{
4159	enum hrtimer_restart ret = HRTIMER_RESTART;
4160	struct perf_sample_data data;
4161	struct pt_regs *regs;
4162	struct perf_event *event;
4163	u64 period;
4164
4165	event = container_of(hrtimer, struct perf_event, hw.hrtimer);
4166	event->pmu->read(event);
4167
4168	perf_sample_data_init(&data, 0);
4169	data.period = event->hw.last_period;
4170	regs = get_irq_regs();
4171	/*
4172	* In case we exclude kernel IPs or are somehow not in interrupt
4173	* context, provide the next best thing, the user IP.
4174	*/
4175	if ((event->attr.exclude_kernel \|\| !regs) &&
4176	!event->attr.exclude_user)
4177	regs = task_pt_regs(current);
4178
4179	if (regs) {
4180	if (!(event->attr.exclude_idle && current->pid == 0))
4181	if (perf_event_overflow(event, 0, &data, regs))
4182	ret = HRTIMER_NORESTART;
4183	}
4184
4185	period = max_t(u64, 10000, event->hw.sample_period);
4186	hrtimer_forward_now(hrtimer, ns_to_ktime(period));
4187
4188	return ret;
4189	}
4190
4191	static void perf_swevent_start_hrtimer(struct perf_event *event)
4192	{
4193	struct hw_perf_event *hwc = &event->hw;
4194
4195	hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4196	hwc->hrtimer.function = perf_swevent_hrtimer;
4197	if (hwc->sample_period) {
4198	u64 period;
4199
4200	if (hwc->remaining) {
4201	if (hwc->remaining < 0)
4202	period = 10000;
4203	else
4204	period = hwc->remaining;
4205	hwc->remaining = 0;
4206	} else {
4207	period = max_t(u64, 10000, hwc->sample_period);
4208	}
4209	__hrtimer_start_range_ns(&hwc->hrtimer,
4210	ns_to_ktime(period), 0,
4211	HRTIMER_MODE_REL, 0);
4212	}
4213	}
4214
4215	static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4216	{
4217	struct hw_perf_event *hwc = &event->hw;
4218
4219	if (hwc->sample_period) {
4220	ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4221	hwc->remaining = ktime_to_ns(remaining);
4222
4223	hrtimer_cancel(&hwc->hrtimer);
4224	}
4225	}
4226
4227	/*
4228	* Software event: cpu wall time clock
4229	*/
4230
4231	static void cpu_clock_perf_event_update(struct perf_event *event)
4232	{
4233	int cpu = raw_smp_processor_id();
4234	s64 prev;
4235	u64 now;
4236
4237	now = cpu_clock(cpu);
4238	prev = atomic64_xchg(&event->hw.prev_count, now);
4239	atomic64_add(now - prev, &event->count);
4240	}
4241
4242	static int cpu_clock_perf_event_enable(struct perf_event *event)
4243	{
4244	struct hw_perf_event *hwc = &event->hw;
4245	int cpu = raw_smp_processor_id();
4246
4247	atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4248	perf_swevent_start_hrtimer(event);
4249
4250	return 0;
4251	}
4252
4253	static void cpu_clock_perf_event_disable(struct perf_event *event)
4254	{
4255	perf_swevent_cancel_hrtimer(event);
4256	cpu_clock_perf_event_update(event);
4257	}
4258
4259	static void cpu_clock_perf_event_read(struct perf_event *event)
4260	{
4261	cpu_clock_perf_event_update(event);
4262	}
4263
4264	static const struct pmu perf_ops_cpu_clock = {
4265	.enable = cpu_clock_perf_event_enable,
4266	.disable = cpu_clock_perf_event_disable,
4267	.read = cpu_clock_perf_event_read,
4268	};
4269
4270	/*
4271	* Software event: task time clock
4272	*/
4273
4274	static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4275	{
4276	u64 prev;
4277	s64 delta;
4278
4279	prev = atomic64_xchg(&event->hw.prev_count, now);
4280	delta = now - prev;
4281	atomic64_add(delta, &event->count);
4282	}
4283
4284	static int task_clock_perf_event_enable(struct perf_event *event)
4285	{
4286	struct hw_perf_event *hwc = &event->hw;
4287	u64 now;
4288
4289	now = event->ctx->time;
4290
4291	atomic64_set(&hwc->prev_count, now);
4292
4293	perf_swevent_start_hrtimer(event);
4294
4295	return 0;
4296	}
4297
4298	static void task_clock_perf_event_disable(struct perf_event *event)
4299	{
4300	perf_swevent_cancel_hrtimer(event);
4301	task_clock_perf_event_update(event, event->ctx->time);
4302
4303	}
4304
4305	static void task_clock_perf_event_read(struct perf_event *event)
4306	{
4307	u64 time;
4308
4309	if (!in_nmi()) {
4310	update_context_time(event->ctx);
4311	time = event->ctx->time;
4312	} else {
4313	u64 now = perf_clock();
4314	u64 delta = now - event->ctx->timestamp;
4315	time = event->ctx->time + delta;
4316	}
4317
4318	task_clock_perf_event_update(event, time);
4319	}
4320
4321	static const struct pmu perf_ops_task_clock = {
4322	.enable = task_clock_perf_event_enable,
4323	.disable = task_clock_perf_event_disable,
4324	.read = task_clock_perf_event_read,
4325	};
4326
4327	#ifdef CONFIG_EVENT_TRACING
4328
4329	void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4330	int entry_size, struct pt_regs *regs)
4331	{
4332	struct perf_sample_data data;
4333	struct perf_raw_record raw = {
4334	.size = entry_size,
4335	.data = record,
4336	};
4337
4338	perf_sample_data_init(&data, addr);
4339	data.raw = &raw;
4340
4341	/* Trace events already protected against recursion */
4342	do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4343	&data, regs);
4344	}
4345	EXPORT_SYMBOL_GPL(perf_tp_event);
4346
4347	static int perf_tp_event_match(struct perf_event *event,
4348	struct perf_sample_data *data)
4349	{
4350	void *record = data->raw->data;
4351
4352	if (likely(!event->filter) \|\| filter_match_preds(event->filter, record))
4353	return 1;
4354	return 0;
4355	}
4356
4357	static void tp_perf_event_destroy(struct perf_event *event)
4358	{
4359	perf_trace_disable(event->attr.config);
4360	}
4361
4362	static const struct pmu tp_perf_event_init(struct perf_event event)
4363	{
4364	/*
4365	* Raw tracepoint data is a severe data leak, only allow root to
4366	* have these.
4367	*/
4368	if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4369	perf_paranoid_tracepoint_raw() &&
4370	!capable(CAP_SYS_ADMIN))
4371	return ERR_PTR(-EPERM);
4372
4373	if (perf_trace_enable(event->attr.config))
4374	return NULL;
4375
4376	event->destroy = tp_perf_event_destroy;
4377
4378	return &perf_ops_generic;
4379	}
4380
4381	static int perf_event_set_filter(struct perf_event event, void __user arg)
4382	{
4383	char *filter_str;
4384	int ret;
4385
4386	if (event->attr.type != PERF_TYPE_TRACEPOINT)
4387	return -EINVAL;
4388
4389	filter_str = strndup_user(arg, PAGE_SIZE);
4390	if (IS_ERR(filter_str))
4391	return PTR_ERR(filter_str);
4392
4393	ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
4394
4395	kfree(filter_str);
4396	return ret;
4397	}
4398
4399	static void perf_event_free_filter(struct perf_event *event)
4400	{
4401	ftrace_profile_free_filter(event);
4402	}
4403
4404	#else
4405
4406	static int perf_tp_event_match(struct perf_event *event,
4407	struct perf_sample_data *data)
4408	{
4409	return 1;
4410	}
4411
4412	static const struct pmu tp_perf_event_init(struct perf_event event)
4413	{
4414	return NULL;
4415	}
4416
4417	static int perf_event_set_filter(struct perf_event event, void __user arg)
4418	{
4419	return -ENOENT;
4420	}
4421
4422	static void perf_event_free_filter(struct perf_event *event)
4423	{
4424	}
4425
4426	#endif /* CONFIG_EVENT_TRACING */
4427
4428	#ifdef CONFIG_HAVE_HW_BREAKPOINT
4429	static void bp_perf_event_destroy(struct perf_event *event)
4430	{
4431	release_bp_slot(event);
4432	}
4433
4434	static const struct pmu bp_perf_event_init(struct perf_event bp)
4435	{
4436	int err;
4437
4438	err = register_perf_hw_breakpoint(bp);
4439	if (err)
4440	return ERR_PTR(err);
4441
4442	bp->destroy = bp_perf_event_destroy;
4443
4444	return &perf_ops_bp;
4445	}
4446
4447	void perf_bp_event(struct perf_event bp, void data)
4448	{
4449	struct perf_sample_data sample;
4450	struct pt_regs *regs = data;
4451
4452	perf_sample_data_init(&sample, bp->attr.bp_addr);
4453
4454	if (!perf_exclude_event(bp, regs))
4455	perf_swevent_add(bp, 1, 1, &sample, regs);
4456	}
4457	#else
4458	static const struct pmu bp_perf_event_init(struct perf_event bp)
4459	{
4460	return NULL;
4461	}
4462
4463	void perf_bp_event(struct perf_event bp, void regs)
4464	{
4465	}
4466	#endif
4467
4468	atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4469
4470	static void sw_perf_event_destroy(struct perf_event *event)
4471	{
4472	u64 event_id = event->attr.config;
4473
4474	WARN_ON(event->parent);
4475
4476	atomic_dec(&perf_swevent_enabled[event_id]);
4477	}
4478
4479	static const struct pmu sw_perf_event_init(struct perf_event event)
4480	{
4481	const struct pmu *pmu = NULL;
4482	u64 event_id = event->attr.config;
4483
4484	/*
4485	* Software events (currently) can't in general distinguish
4486	* between user, kernel and hypervisor events.
4487	* However, context switches and cpu migrations are considered
4488	* to be kernel events, and page faults are never hypervisor
4489	* events.
4490	*/
4491	switch (event_id) {
4492	case PERF_COUNT_SW_CPU_CLOCK:
4493	pmu = &perf_ops_cpu_clock;
4494
4495	break;
4496	case PERF_COUNT_SW_TASK_CLOCK:
4497	/*
4498	* If the user instantiates this as a per-cpu event,
4499	* use the cpu_clock event instead.
4500	*/
4501	if (event->ctx->task)
4502	pmu = &perf_ops_task_clock;
4503	else
4504	pmu = &perf_ops_cpu_clock;
4505
4506	break;
4507	case PERF_COUNT_SW_PAGE_FAULTS:
4508	case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4509	case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4510	case PERF_COUNT_SW_CONTEXT_SWITCHES:
4511	case PERF_COUNT_SW_CPU_MIGRATIONS:
4512	case PERF_COUNT_SW_ALIGNMENT_FAULTS:
4513	case PERF_COUNT_SW_EMULATION_FAULTS:
4514	if (!event->parent) {
4515	atomic_inc(&perf_swevent_enabled[event_id]);
4516	event->destroy = sw_perf_event_destroy;
4517	}
4518	pmu = &perf_ops_generic;
4519	break;
4520	}
4521
4522	return pmu;
4523	}
4524
4525	/*
4526	* Allocate and initialize a event structure
4527	*/
4528	static struct perf_event *
4529	perf_event_alloc(struct perf_event_attr *attr,
4530	int cpu,
4531	struct perf_event_context *ctx,
4532	struct perf_event *group_leader,
4533	struct perf_event *parent_event,
4534	perf_overflow_handler_t overflow_handler,
4535	gfp_t gfpflags)
4536	{
4537	const struct pmu *pmu;
4538	struct perf_event *event;
4539	struct hw_perf_event *hwc;
4540	long err;
4541
4542	event = kzalloc(sizeof(*event), gfpflags);
4543	if (!event)
4544	return ERR_PTR(-ENOMEM);
4545
4546	/*
4547	* Single events are their own group leaders, with an
4548	* empty sibling list:
4549	*/
4550	if (!group_leader)
4551	group_leader = event;
4552
4553	mutex_init(&event->child_mutex);
4554	INIT_LIST_HEAD(&event->child_list);
4555
4556	INIT_LIST_HEAD(&event->group_entry);
4557	INIT_LIST_HEAD(&event->event_entry);
4558	INIT_LIST_HEAD(&event->sibling_list);
4559	init_waitqueue_head(&event->waitq);
4560
4561	mutex_init(&event->mmap_mutex);
4562
4563	event->cpu = cpu;
4564	event->attr = *attr;
4565	event->group_leader = group_leader;
4566	event->pmu = NULL;
4567	event->ctx = ctx;
4568	event->oncpu = -1;
4569
4570	event->parent = parent_event;
4571
4572	event->ns = get_pid_ns(current->nsproxy->pid_ns);
4573	event->id = atomic64_inc_return(&perf_event_id);
4574
4575	event->state = PERF_EVENT_STATE_INACTIVE;
4576
4577	if (!overflow_handler && parent_event)
4578	overflow_handler = parent_event->overflow_handler;
4579
4580	event->overflow_handler = overflow_handler;
4581
4582	if (attr->disabled)
4583	event->state = PERF_EVENT_STATE_OFF;
4584
4585	pmu = NULL;
4586
4587	hwc = &event->hw;
4588	hwc->sample_period = attr->sample_period;
4589	if (attr->freq && attr->sample_freq)
4590	hwc->sample_period = 1;
4591	hwc->last_period = hwc->sample_period;
4592
4593	atomic64_set(&hwc->period_left, hwc->sample_period);
4594
4595	/*
4596	* we currently do not support PERF_FORMAT_GROUP on inherited events
4597	*/
4598	if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4599	goto done;
4600
4601	switch (attr->type) {
4602	case PERF_TYPE_RAW:
4603	case PERF_TYPE_HARDWARE:
4604	case PERF_TYPE_HW_CACHE:
4605	pmu = hw_perf_event_init(event);
4606	break;
4607
4608	case PERF_TYPE_SOFTWARE:
4609	pmu = sw_perf_event_init(event);
4610	break;
4611
4612	case PERF_TYPE_TRACEPOINT:
4613	pmu = tp_perf_event_init(event);
4614	break;
4615
4616	case PERF_TYPE_BREAKPOINT:
4617	pmu = bp_perf_event_init(event);
4618	break;
4619
4620
4621	default:
4622	break;
4623	}
4624	done:
4625	err = 0;
4626	if (!pmu)
4627	err = -EINVAL;
4628	else if (IS_ERR(pmu))
4629	err = PTR_ERR(pmu);
4630
4631	if (err) {
4632	if (event->ns)
4633	put_pid_ns(event->ns);
4634	kfree(event);
4635	return ERR_PTR(err);
4636	}
4637
4638	event->pmu = pmu;
4639
4640	if (!event->parent) {
4641	atomic_inc(&nr_events);
4642	if (event->attr.mmap)
4643	atomic_inc(&nr_mmap_events);
4644	if (event->attr.comm)
4645	atomic_inc(&nr_comm_events);
4646	if (event->attr.task)
4647	atomic_inc(&nr_task_events);
4648	}
4649
4650	return event;
4651	}
4652
4653	static int perf_copy_attr(struct perf_event_attr __user *uattr,
4654	struct perf_event_attr *attr)
4655	{
4656	u32 size;
4657	int ret;
4658
4659	if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4660	return -EFAULT;
4661
4662	/*
4663	* zero the full structure, so that a short copy will be nice.
4664	*/
4665	memset(attr, 0, sizeof(*attr));
4666
4667	ret = get_user(size, &uattr->size);
4668	if (ret)
4669	return ret;
4670
4671	if (size > PAGE_SIZE) /* silly large */
4672	goto err_size;
4673
4674	if (!size) /* abi compat */
4675	size = PERF_ATTR_SIZE_VER0;
4676
4677	if (size < PERF_ATTR_SIZE_VER0)
4678	goto err_size;
4679
4680	/*
4681	* If we're handed a bigger struct than we know of,
4682	* ensure all the unknown bits are 0 - i.e. new
4683	* user-space does not rely on any kernel feature
4684	* extensions we dont know about yet.
4685	*/
4686	if (size > sizeof(*attr)) {
4687	unsigned char __user *addr;
4688	unsigned char __user *end;
4689	unsigned char val;
4690
4691	addr = (void __user )uattr + sizeof(attr);
4692	end = (void __user *)uattr + size;
4693
4694	for (; addr < end; addr++) {
4695	ret = get_user(val, addr);
4696	if (ret)
4697	return ret;
4698	if (val)
4699	goto err_size;
4700	}
4701	size = sizeof(*attr);
4702	}
4703
4704	ret = copy_from_user(attr, uattr, size);
4705	if (ret)
4706	return -EFAULT;
4707
4708	/*
4709	* If the type exists, the corresponding creation will verify
4710	* the attr->config.
4711	*/
4712	if (attr->type >= PERF_TYPE_MAX)
4713	return -EINVAL;
4714
4715	if (attr->__reserved_1)
4716	return -EINVAL;
4717
4718	if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4719	return -EINVAL;
4720
4721	if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4722	return -EINVAL;
4723
4724	out:
4725	return ret;
4726
4727	err_size:
4728	put_user(sizeof(*attr), &uattr->size);
4729	ret = -E2BIG;
4730	goto out;
4731	}
4732
4733	static int perf_event_set_output(struct perf_event *event, int output_fd)
4734	{
4735	struct perf_event *output_event = NULL;
4736	struct file *output_file = NULL;
4737	struct perf_event *old_output;
4738	int fput_needed = 0;
4739	int ret = -EINVAL;
4740
4741	if (!output_fd)
4742	goto set;
4743
4744	output_file = fget_light(output_fd, &fput_needed);
4745	if (!output_file)
4746	return -EBADF;
4747
4748	if (output_file->f_op != &perf_fops)
4749	goto out;
4750
4751	output_event = output_file->private_data;
4752
4753	/* Don't chain output fds */
4754	if (output_event->output)
4755	goto out;
4756
4757	/* Don't set an output fd when we already have an output channel */
4758	if (event->data)
4759	goto out;
4760
4761	atomic_long_inc(&output_file->f_count);
4762
4763	set:
4764	mutex_lock(&event->mmap_mutex);
4765	old_output = event->output;
4766	rcu_assign_pointer(event->output, output_event);
4767	mutex_unlock(&event->mmap_mutex);
4768
4769	if (old_output) {
4770	/*
4771	* we need to make sure no existing perf_output_*()
4772	* is still referencing this event.
4773	*/
4774	synchronize_rcu();
4775	fput(old_output->filp);
4776	}
4777
4778	ret = 0;
4779	out:
4780	fput_light(output_file, fput_needed);
4781	return ret;
4782	}
4783
4784	/**
4785	* sys_perf_event_open - open a performance event, associate it to a task/cpu
4786	*
4787	* @attr_uptr: event_id type attributes for monitoring/sampling
4788	* @pid: target pid
4789	* @cpu: target cpu
4790	* @group_fd: group leader event fd
4791	*/
4792	SYSCALL_DEFINE5(perf_event_open,
4793	struct perf_event_attr __user *, attr_uptr,
4794	pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4795	{
4796	struct perf_event event, group_leader;
4797	struct perf_event_attr attr;
4798	struct perf_event_context *ctx;
4799	struct file *event_file = NULL;
4800	struct file *group_file = NULL;
4801	int fput_needed = 0;
4802	int fput_needed2 = 0;
4803	int err;
4804
4805	/* for future expandability... */
4806	if (flags & ~(PERF_FLAG_FD_NO_GROUP \| PERF_FLAG_FD_OUTPUT))
4807	return -EINVAL;
4808
4809	err = perf_copy_attr(attr_uptr, &attr);
4810	if (err)
4811	return err;
4812
4813	if (!attr.exclude_kernel) {
4814	if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4815	return -EACCES;
4816	}
4817
4818	if (attr.freq) {
4819	if (attr.sample_freq > sysctl_perf_event_sample_rate)
4820	return -EINVAL;
4821	}
4822
4823	/*
4824	* Get the target context (task or percpu):
4825	*/
4826	ctx = find_get_context(pid, cpu);
4827	if (IS_ERR(ctx))
4828	return PTR_ERR(ctx);
4829
4830	/*
4831	* Look up the group leader (we will attach this event to it):
4832	*/
4833	group_leader = NULL;
4834	if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4835	err = -EINVAL;
4836	group_file = fget_light(group_fd, &fput_needed);
4837	if (!group_file)
4838	goto err_put_context;
4839	if (group_file->f_op != &perf_fops)
4840	goto err_put_context;
4841
4842	group_leader = group_file->private_data;
4843	/*
4844	* Do not allow a recursive hierarchy (this new sibling
4845	* becoming part of another group-sibling):
4846	*/
4847	if (group_leader->group_leader != group_leader)
4848	goto err_put_context;
4849	/*
4850	* Do not allow to attach to a group in a different
4851	* task or CPU context:
4852	*/
4853	if (group_leader->ctx != ctx)
4854	goto err_put_context;
4855	/*
4856	* Only a group leader can be exclusive or pinned
4857	*/
4858	if (attr.exclusive \|\| attr.pinned)
4859	goto err_put_context;
4860	}
4861
4862	event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4863	NULL, NULL, GFP_KERNEL);
4864	err = PTR_ERR(event);
4865	if (IS_ERR(event))
4866	goto err_put_context;
4867
4868	err = anon_inode_getfd("[perf_event]", &perf_fops, event, O_RDWR);
4869	if (err < 0)
4870	goto err_free_put_context;
4871
4872	event_file = fget_light(err, &fput_needed2);
4873	if (!event_file)
4874	goto err_free_put_context;
4875
4876	if (flags & PERF_FLAG_FD_OUTPUT) {
4877	err = perf_event_set_output(event, group_fd);
4878	if (err)
4879	goto err_fput_free_put_context;
4880	}
4881
4882	event->filp = event_file;
4883	WARN_ON_ONCE(ctx->parent_ctx);
4884	mutex_lock(&ctx->mutex);
4885	perf_install_in_context(ctx, event, cpu);
4886	++ctx->generation;
4887	mutex_unlock(&ctx->mutex);
4888
4889	event->owner = current;
4890	get_task_struct(current);
4891	mutex_lock(&current->perf_event_mutex);
4892	list_add_tail(&event->owner_entry, &current->perf_event_list);
4893	mutex_unlock(&current->perf_event_mutex);
4894
4895	err_fput_free_put_context:
4896	fput_light(event_file, fput_needed2);
4897
4898	err_free_put_context:
4899	if (err < 0)
4900	kfree(event);
4901
4902	err_put_context:
4903	if (err < 0)
4904	put_ctx(ctx);
4905
4906	fput_light(group_file, fput_needed);
4907
4908	return err;
4909	}
4910
4911	/**
4912	* perf_event_create_kernel_counter
4913	*
4914	* @attr: attributes of the counter to create
4915	* @cpu: cpu in which the counter is bound
4916	* @pid: task to profile
4917	*/
4918	struct perf_event *
4919	perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
4920	pid_t pid,
4921	perf_overflow_handler_t overflow_handler)
4922	{
4923	struct perf_event *event;
4924	struct perf_event_context *ctx;
4925	int err;
4926
4927	/*
4928	* Get the target context (task or percpu):
4929	*/
4930
4931	ctx = find_get_context(pid, cpu);
4932	if (IS_ERR(ctx)) {
4933	err = PTR_ERR(ctx);
4934	goto err_exit;
4935	}
4936
4937	event = perf_event_alloc(attr, cpu, ctx, NULL,
4938	NULL, overflow_handler, GFP_KERNEL);
4939	if (IS_ERR(event)) {
4940	err = PTR_ERR(event);
4941	goto err_put_context;
4942	}
4943
4944	event->filp = NULL;
4945	WARN_ON_ONCE(ctx->parent_ctx);
4946	mutex_lock(&ctx->mutex);
4947	perf_install_in_context(ctx, event, cpu);
4948	++ctx->generation;
4949	mutex_unlock(&ctx->mutex);
4950
4951	event->owner = current;
4952	get_task_struct(current);
4953	mutex_lock(&current->perf_event_mutex);
4954	list_add_tail(&event->owner_entry, &current->perf_event_list);
4955	mutex_unlock(&current->perf_event_mutex);
4956
4957	return event;
4958
4959	err_put_context:
4960	put_ctx(ctx);
4961	err_exit:
4962	return ERR_PTR(err);
4963	}
4964	EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
4965
4966	/*
4967	* inherit a event from parent task to child task:
4968	*/
4969	static struct perf_event *
4970	inherit_event(struct perf_event *parent_event,
4971	struct task_struct *parent,
4972	struct perf_event_context *parent_ctx,
4973	struct task_struct *child,
4974	struct perf_event *group_leader,
4975	struct perf_event_context *child_ctx)
4976	{
4977	struct perf_event *child_event;
4978
4979	/*
4980	* Instead of creating recursive hierarchies of events,
4981	* we link inherited events back to the original parent,
4982	* which has a filp for sure, which we use as the reference
4983	* count:
4984	*/
4985	if (parent_event->parent)
4986	parent_event = parent_event->parent;
4987
4988	child_event = perf_event_alloc(&parent_event->attr,
4989	parent_event->cpu, child_ctx,
4990	group_leader, parent_event,
4991	NULL, GFP_KERNEL);
4992	if (IS_ERR(child_event))
4993	return child_event;
4994	get_ctx(child_ctx);
4995
4996	/*
4997	* Make the child state follow the state of the parent event,
4998	* not its attr.disabled bit. We hold the parent's mutex,
4999	* so we won't race with perf_event_{en, dis}able_family.
5000	*/
5001	if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
5002	child_event->state = PERF_EVENT_STATE_INACTIVE;
5003	else
5004	child_event->state = PERF_EVENT_STATE_OFF;
5005
5006	if (parent_event->attr.freq) {
5007	u64 sample_period = parent_event->hw.sample_period;
5008	struct hw_perf_event *hwc = &child_event->hw;
5009
5010	hwc->sample_period = sample_period;
5011	hwc->last_period = sample_period;
5012
5013	atomic64_set(&hwc->period_left, sample_period);
5014	}
5015
5016	child_event->overflow_handler = parent_event->overflow_handler;
5017
5018	/*
5019	* Link it up in the child's context:
5020	*/
5021	add_event_to_ctx(child_event, child_ctx);
5022
5023	/*
5024	* Get a reference to the parent filp - we will fput it
5025	* when the child event exits. This is safe to do because
5026	* we are in the parent and we know that the filp still
5027	* exists and has a nonzero count:
5028	*/
5029	atomic_long_inc(&parent_event->filp->f_count);
5030
5031	/*
5032	* Link this into the parent event's child list
5033	*/
5034	WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5035	mutex_lock(&parent_event->child_mutex);
5036	list_add_tail(&child_event->child_list, &parent_event->child_list);
5037	mutex_unlock(&parent_event->child_mutex);
5038
5039	return child_event;
5040	}
5041
5042	static int inherit_group(struct perf_event *parent_event,
5043	struct task_struct *parent,
5044	struct perf_event_context *parent_ctx,
5045	struct task_struct *child,
5046	struct perf_event_context *child_ctx)
5047	{
5048	struct perf_event *leader;
5049	struct perf_event *sub;
5050	struct perf_event *child_ctr;
5051
5052	leader = inherit_event(parent_event, parent, parent_ctx,
5053	child, NULL, child_ctx);
5054	if (IS_ERR(leader))
5055	return PTR_ERR(leader);
5056	list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
5057	child_ctr = inherit_event(sub, parent, parent_ctx,
5058	child, leader, child_ctx);
5059	if (IS_ERR(child_ctr))
5060	return PTR_ERR(child_ctr);
5061	}
5062	return 0;
5063	}
5064
5065	static void sync_child_event(struct perf_event *child_event,
5066	struct task_struct *child)
5067	{
5068	struct perf_event *parent_event = child_event->parent;
5069	u64 child_val;
5070
5071	if (child_event->attr.inherit_stat)
5072	perf_event_read_event(child_event, child);
5073
5074	child_val = atomic64_read(&child_event->count);
5075
5076	/*
5077	* Add back the child's count to the parent's count:
5078	*/
5079	atomic64_add(child_val, &parent_event->count);
5080	atomic64_add(child_event->total_time_enabled,
5081	&parent_event->child_total_time_enabled);
5082	atomic64_add(child_event->total_time_running,
5083	&parent_event->child_total_time_running);
5084
5085	/*
5086	* Remove this event from the parent's list
5087	*/
5088	WARN_ON_ONCE(parent_event->ctx->parent_ctx);
5089	mutex_lock(&parent_event->child_mutex);
5090	list_del_init(&child_event->child_list);
5091	mutex_unlock(&parent_event->child_mutex);
5092
5093	/*
5094	* Release the parent event, if this was the last
5095	* reference to it.
5096	*/
5097	fput(parent_event->filp);
5098	}
5099
5100	static void
5101	__perf_event_exit_task(struct perf_event *child_event,
5102	struct perf_event_context *child_ctx,
5103	struct task_struct *child)
5104	{
5105	struct perf_event *parent_event;
5106
5107	perf_event_remove_from_context(child_event);
5108
5109	parent_event = child_event->parent;
5110	/*
5111	* It can happen that parent exits first, and has events
5112	* that are still around due to the child reference. These
5113	* events need to be zapped - but otherwise linger.
5114	*/
5115	if (parent_event) {
5116	sync_child_event(child_event, child);
5117	free_event(child_event);
5118	}
5119	}
5120
5121	/*
5122	* When a child task exits, feed back event values to parent events.
5123	*/
5124	void perf_event_exit_task(struct task_struct *child)
5125	{
5126	struct perf_event child_event, tmp;
5127	struct perf_event_context *child_ctx;
5128	unsigned long flags;
5129
5130	if (likely(!child->perf_event_ctxp)) {
5131	perf_event_task(child, NULL, 0);
5132	return;
5133	}
5134
5135	local_irq_save(flags);
5136	/*
5137	* We can't reschedule here because interrupts are disabled,
5138	* and either child is current or it is a task that can't be
5139	* scheduled, so we are now safe from rescheduling changing
5140	* our context.
5141	*/
5142	child_ctx = child->perf_event_ctxp;
5143	__perf_event_task_sched_out(child_ctx);
5144
5145	/*
5146	* Take the context lock here so that if find_get_context is
5147	* reading child->perf_event_ctxp, we wait until it has
5148	* incremented the context's refcount before we do put_ctx below.
5149	*/
5150	raw_spin_lock(&child_ctx->lock);
5151	child->perf_event_ctxp = NULL;
5152	/*
5153	* If this context is a clone; unclone it so it can't get
5154	* swapped to another process while we're removing all
5155	* the events from it.
5156	*/
5157	unclone_ctx(child_ctx);
5158	update_context_time(child_ctx);
5159	raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
5160
5161	/*
5162	* Report the task dead after unscheduling the events so that we
5163	* won't get any samples after PERF_RECORD_EXIT. We can however still
5164	* get a few PERF_RECORD_READ events.
5165	*/
5166	perf_event_task(child, child_ctx, 0);
5167
5168	/*
5169	* We can recurse on the same lock type through:
5170	*
5171	* __perf_event_exit_task()
5172	* sync_child_event()
5173	* fput(parent_event->filp)
5174	* perf_release()
5175	* mutex_lock(&ctx->mutex)
5176	*
5177	* But since its the parent context it won't be the same instance.
5178	*/
5179	mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
5180
5181	again:
5182	list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
5183	group_entry)
5184	__perf_event_exit_task(child_event, child_ctx, child);
5185
5186	list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
5187	group_entry)
5188	__perf_event_exit_task(child_event, child_ctx, child);
5189
5190	/*
5191	* If the last event was a group event, it will have appended all
5192	* its siblings to the list, but we obtained 'tmp' before that which
5193	* will still point to the list head terminating the iteration.
5194	*/
5195	if (!list_empty(&child_ctx->pinned_groups) \|\|
5196	!list_empty(&child_ctx->flexible_groups))
5197	goto again;
5198
5199	mutex_unlock(&child_ctx->mutex);
5200
5201	put_ctx(child_ctx);
5202	}
5203
5204	static void perf_free_event(struct perf_event *event,
5205	struct perf_event_context *ctx)
5206	{
5207	struct perf_event *parent = event->parent;
5208
5209	if (WARN_ON_ONCE(!parent))
5210	return;
5211
5212	mutex_lock(&parent->child_mutex);
5213	list_del_init(&event->child_list);
5214	mutex_unlock(&parent->child_mutex);
5215
5216	fput(parent->filp);
5217
5218	list_del_event(event, ctx);
5219	free_event(event);
5220	}
5221
5222	/*
5223	* free an unexposed, unused context as created by inheritance by
5224	* init_task below, used by fork() in case of fail.
5225	*/
5226	void perf_event_free_task(struct task_struct *task)
5227	{
5228	struct perf_event_context *ctx = task->perf_event_ctxp;
5229	struct perf_event event, tmp;
5230
5231	if (!ctx)
5232	return;
5233
5234	mutex_lock(&ctx->mutex);
5235	again:
5236	list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5237	perf_free_event(event, ctx);
5238
5239	list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
5240	group_entry)
5241	perf_free_event(event, ctx);
5242
5243	if (!list_empty(&ctx->pinned_groups) \|\|
5244	!list_empty(&ctx->flexible_groups))
5245	goto again;
5246
5247	mutex_unlock(&ctx->mutex);
5248
5249	put_ctx(ctx);
5250	}
5251
5252	static int
5253	inherit_task_group(struct perf_event event, struct task_struct parent,
5254	struct perf_event_context *parent_ctx,
5255	struct task_struct *child,
5256	int *inherited_all)
5257	{
5258	int ret;
5259	struct perf_event_context *child_ctx = child->perf_event_ctxp;
5260
5261	if (!event->attr.inherit) {
5262	*inherited_all = 0;
5263	return 0;
5264	}
5265
5266	if (!child_ctx) {
5267	/*
5268	* This is executed from the parent task context, so
5269	* inherit events that have been marked for cloning.
5270	* First allocate and initialize a context for the
5271	* child.
5272	*/
5273
5274	child_ctx = kzalloc(sizeof(struct perf_event_context),
5275	GFP_KERNEL);
5276	if (!child_ctx)
5277	return -ENOMEM;
5278
5279	__perf_event_init_context(child_ctx, child);
5280	child->perf_event_ctxp = child_ctx;
5281	get_task_struct(child);
5282	}
5283
5284	ret = inherit_group(event, parent, parent_ctx,
5285	child, child_ctx);
5286
5287	if (ret)
5288	*inherited_all = 0;
5289
5290	return ret;
5291	}
5292
5293
5294	/*
5295	* Initialize the perf_event context in task_struct
5296	*/
5297	int perf_event_init_task(struct task_struct *child)
5298	{
5299	struct perf_event_context child_ctx, parent_ctx;
5300	struct perf_event_context *cloned_ctx;
5301	struct perf_event *event;
5302	struct task_struct *parent = current;
5303	int inherited_all = 1;
5304	int ret = 0;
5305
5306	child->perf_event_ctxp = NULL;
5307
5308	mutex_init(&child->perf_event_mutex);
5309	INIT_LIST_HEAD(&child->perf_event_list);
5310
5311	if (likely(!parent->perf_event_ctxp))
5312	return 0;
5313
5314	/*
5315	* If the parent's context is a clone, pin it so it won't get
5316	* swapped under us.
5317	*/
5318	parent_ctx = perf_pin_task_context(parent);
5319
5320	/*
5321	* No need to check if parent_ctx != NULL here; since we saw
5322	* it non-NULL earlier, the only reason for it to become NULL
5323	* is if we exit, and since we're currently in the middle of
5324	* a fork we can't be exiting at the same time.
5325	*/
5326
5327	/*
5328	* Lock the parent list. No need to lock the child - not PID
5329	* hashed yet and not running, so nobody can access it.
5330	*/
5331	mutex_lock(&parent_ctx->mutex);
5332
5333	/*
5334	* We dont have to disable NMIs - we are only looking at
5335	* the list, not manipulating it:
5336	*/
5337	list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
5338	ret = inherit_task_group(event, parent, parent_ctx, child,
5339	&inherited_all);
5340	if (ret)
5341	break;
5342	}
5343
5344	list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
5345	ret = inherit_task_group(event, parent, parent_ctx, child,
5346	&inherited_all);
5347	if (ret)
5348	break;
5349	}
5350
5351	child_ctx = child->perf_event_ctxp;
5352
5353	if (child_ctx && inherited_all) {
5354	/*
5355	* Mark the child context as a clone of the parent
5356	* context, or of whatever the parent is a clone of.
5357	* Note that if the parent is a clone, it could get
5358	* uncloned at any point, but that doesn't matter
5359	* because the list of events and the generation
5360	* count can't have changed since we took the mutex.
5361	*/
5362	cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
5363	if (cloned_ctx) {
5364	child_ctx->parent_ctx = cloned_ctx;
5365	child_ctx->parent_gen = parent_ctx->parent_gen;
5366	} else {
5367	child_ctx->parent_ctx = parent_ctx;
5368	child_ctx->parent_gen = parent_ctx->generation;
5369	}
5370	get_ctx(child_ctx->parent_ctx);
5371	}
5372
5373	mutex_unlock(&parent_ctx->mutex);
5374
5375	perf_unpin_context(parent_ctx);
5376
5377	return ret;
5378	}
5379
5380	static void __init perf_event_init_all_cpus(void)
5381	{
5382	int cpu;
5383	struct perf_cpu_context *cpuctx;
5384
5385	for_each_possible_cpu(cpu) {
5386	cpuctx = &per_cpu(perf_cpu_context, cpu);
5387	__perf_event_init_context(&cpuctx->ctx, NULL);
5388	}
5389	}
5390
5391	static void __cpuinit perf_event_init_cpu(int cpu)
5392	{
5393	struct perf_cpu_context *cpuctx;
5394
5395	cpuctx = &per_cpu(perf_cpu_context, cpu);
5396
5397	spin_lock(&perf_resource_lock);
5398	cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5399	spin_unlock(&perf_resource_lock);
5400	}
5401
5402	#ifdef CONFIG_HOTPLUG_CPU
5403	static void __perf_event_exit_cpu(void *info)
5404	{
5405	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5406	struct perf_event_context *ctx = &cpuctx->ctx;
5407	struct perf_event event, tmp;
5408
5409	list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
5410	__perf_event_remove_from_context(event);
5411	list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
5412	__perf_event_remove_from_context(event);
5413	}
5414	static void perf_event_exit_cpu(int cpu)
5415	{
5416	struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5417	struct perf_event_context *ctx = &cpuctx->ctx;
5418
5419	mutex_lock(&ctx->mutex);
5420	smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5421	mutex_unlock(&ctx->mutex);
5422	}
5423	#else
5424	static inline void perf_event_exit_cpu(int cpu) { }
5425	#endif
5426
5427	static int __cpuinit
5428	perf_cpu_notify(struct notifier_block self, unsigned long action, void hcpu)
5429	{
5430	unsigned int cpu = (long)hcpu;
5431
5432	switch (action) {
5433
5434	case CPU_UP_PREPARE:
5435	case CPU_UP_PREPARE_FROZEN:
5436	perf_event_init_cpu(cpu);
5437	break;
5438
5439	case CPU_DOWN_PREPARE:
5440	case CPU_DOWN_PREPARE_FROZEN:
5441	perf_event_exit_cpu(cpu);
5442	break;
5443
5444	default:
5445	break;
5446	}
5447
5448	return NOTIFY_OK;
5449	}
5450
5451	/*
5452	* This has to have a higher priority than migration_notifier in sched.c.
5453	*/
5454	static struct notifier_block __cpuinitdata perf_cpu_nb = {
5455	.notifier_call = perf_cpu_notify,
5456	.priority = 20,
5457	};
5458
5459	void __init perf_event_init(void)
5460	{
5461	perf_event_init_all_cpus();
5462	perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5463	(void *)(long)smp_processor_id());
5464	perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5465	(void *)(long)smp_processor_id());
5466	register_cpu_notifier(&perf_cpu_nb);
5467	}
5468
5469	static ssize_t perf_show_reserve_percpu(struct sysdev_class *class,
5470	struct sysdev_class_attribute *attr,
5471	char *buf)
5472	{
5473	return sprintf(buf, "%d\n", perf_reserved_percpu);
5474	}
5475
5476	static ssize_t
5477	perf_set_reserve_percpu(struct sysdev_class *class,
5478	struct sysdev_class_attribute *attr,
5479	const char *buf,
5480	size_t count)
5481	{
5482	struct perf_cpu_context *cpuctx;
5483	unsigned long val;
5484	int err, cpu, mpt;
5485
5486	err = strict_strtoul(buf, 10, &val);
5487	if (err)
5488	return err;
5489	if (val > perf_max_events)
5490	return -EINVAL;
5491
5492	spin_lock(&perf_resource_lock);
5493	perf_reserved_percpu = val;
5494	for_each_online_cpu(cpu) {
5495	cpuctx = &per_cpu(perf_cpu_context, cpu);
5496	raw_spin_lock_irq(&cpuctx->ctx.lock);
5497	mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5498	perf_max_events - perf_reserved_percpu);
5499	cpuctx->max_pertask = mpt;
5500	raw_spin_unlock_irq(&cpuctx->ctx.lock);
5501	}
5502	spin_unlock(&perf_resource_lock);
5503
5504	return count;
5505	}
5506
5507	static ssize_t perf_show_overcommit(struct sysdev_class *class,
5508	struct sysdev_class_attribute *attr,
5509	char *buf)
5510	{
5511	return sprintf(buf, "%d\n", perf_overcommit);
5512	}
5513
5514	static ssize_t
5515	perf_set_overcommit(struct sysdev_class *class,
5516	struct sysdev_class_attribute *attr,
5517	const char *buf, size_t count)
5518	{
5519	unsigned long val;
5520	int err;
5521
5522	err = strict_strtoul(buf, 10, &val);
5523	if (err)
5524	return err;
5525	if (val > 1)
5526	return -EINVAL;
5527
5528	spin_lock(&perf_resource_lock);
5529	perf_overcommit = val;
5530	spin_unlock(&perf_resource_lock);
5531
5532	return count;
5533	}
5534
5535	static SYSDEV_CLASS_ATTR(
5536	reserve_percpu,
5537	0644,
5538	perf_show_reserve_percpu,
5539	perf_set_reserve_percpu
5540	);
5541
5542	static SYSDEV_CLASS_ATTR(
5543	overcommit,
5544	0644,
5545	perf_show_overcommit,
5546	perf_set_overcommit
5547	);
5548
5549	static struct attribute *perfclass_attrs[] = {
5550	&attr_reserve_percpu.attr,
5551	&attr_overcommit.attr,
5552	NULL
5553	};
5554
5555	static struct attribute_group perfclass_attr_group = {
5556	.attrs = perfclass_attrs,
5557	.name = "perf_events",
5558	};
5559
5560	static int __init perf_event_sysfs_init(void)
5561	{
5562	return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5563	&perfclass_attr_group);
5564	}
5565	device_initcall(perf_event_sysfs_init);
5566

Download this file

Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master

Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9

Kernel

Kernel Git Source Tree

Root/kernel/perf_event.c

interactive