Loading...
1/*
2 * Performance events core code:
3 *
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 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/idr.h>
17#include <linux/file.h>
18#include <linux/poll.h>
19#include <linux/slab.h>
20#include <linux/hash.h>
21#include <linux/tick.h>
22#include <linux/sysfs.h>
23#include <linux/dcache.h>
24#include <linux/percpu.h>
25#include <linux/ptrace.h>
26#include <linux/reboot.h>
27#include <linux/vmstat.h>
28#include <linux/device.h>
29#include <linux/export.h>
30#include <linux/vmalloc.h>
31#include <linux/hardirq.h>
32#include <linux/rculist.h>
33#include <linux/uaccess.h>
34#include <linux/syscalls.h>
35#include <linux/anon_inodes.h>
36#include <linux/kernel_stat.h>
37#include <linux/perf_event.h>
38#include <linux/ftrace_event.h>
39#include <linux/hw_breakpoint.h>
40#include <linux/mm_types.h>
41#include <linux/cgroup.h>
42
43#include "internal.h"
44
45#include <asm/irq_regs.h>
46
47struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
50 void *info;
51 int ret;
52};
53
54static void remote_function(void *data)
55{
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
58
59 if (p) {
60 tfc->ret = -EAGAIN;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
62 return;
63 }
64
65 tfc->ret = tfc->func(tfc->info);
66}
67
68/**
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
73 *
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
76 *
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
80 */
81static int
82task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
83{
84 struct remote_function_call data = {
85 .p = p,
86 .func = func,
87 .info = info,
88 .ret = -ESRCH, /* No such (running) process */
89 };
90
91 if (task_curr(p))
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
93
94 return data.ret;
95}
96
97/**
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
101 *
102 * Calls the function @func on the remote cpu.
103 *
104 * returns: @func return value or -ENXIO when the cpu is offline
105 */
106static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
107{
108 struct remote_function_call data = {
109 .p = NULL,
110 .func = func,
111 .info = info,
112 .ret = -ENXIO, /* No such CPU */
113 };
114
115 smp_call_function_single(cpu, remote_function, &data, 1);
116
117 return data.ret;
118}
119
120#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP |\
123 PERF_FLAG_FD_CLOEXEC)
124
125/*
126 * branch priv levels that need permission checks
127 */
128#define PERF_SAMPLE_BRANCH_PERM_PLM \
129 (PERF_SAMPLE_BRANCH_KERNEL |\
130 PERF_SAMPLE_BRANCH_HV)
131
132enum event_type_t {
133 EVENT_FLEXIBLE = 0x1,
134 EVENT_PINNED = 0x2,
135 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
136};
137
138/*
139 * perf_sched_events : >0 events exist
140 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 */
142struct static_key_deferred perf_sched_events __read_mostly;
143static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
144static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
145
146static atomic_t nr_mmap_events __read_mostly;
147static atomic_t nr_comm_events __read_mostly;
148static atomic_t nr_task_events __read_mostly;
149static atomic_t nr_freq_events __read_mostly;
150
151static LIST_HEAD(pmus);
152static DEFINE_MUTEX(pmus_lock);
153static struct srcu_struct pmus_srcu;
154
155/*
156 * perf event paranoia level:
157 * -1 - not paranoid at all
158 * 0 - disallow raw tracepoint access for unpriv
159 * 1 - disallow cpu events for unpriv
160 * 2 - disallow kernel profiling for unpriv
161 */
162int sysctl_perf_event_paranoid __read_mostly = 1;
163
164/* Minimum for 512 kiB + 1 user control page */
165int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
166
167/*
168 * max perf event sample rate
169 */
170#define DEFAULT_MAX_SAMPLE_RATE 100000
171#define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
172#define DEFAULT_CPU_TIME_MAX_PERCENT 25
173
174int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
175
176static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
177static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
178
179static int perf_sample_allowed_ns __read_mostly =
180 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
181
182void update_perf_cpu_limits(void)
183{
184 u64 tmp = perf_sample_period_ns;
185
186 tmp *= sysctl_perf_cpu_time_max_percent;
187 do_div(tmp, 100);
188 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
189}
190
191static int perf_rotate_context(struct perf_cpu_context *cpuctx);
192
193int perf_proc_update_handler(struct ctl_table *table, int write,
194 void __user *buffer, size_t *lenp,
195 loff_t *ppos)
196{
197 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
198
199 if (ret || !write)
200 return ret;
201
202 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
203 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
204 update_perf_cpu_limits();
205
206 return 0;
207}
208
209int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
210
211int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
212 void __user *buffer, size_t *lenp,
213 loff_t *ppos)
214{
215 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
216
217 if (ret || !write)
218 return ret;
219
220 update_perf_cpu_limits();
221
222 return 0;
223}
224
225/*
226 * perf samples are done in some very critical code paths (NMIs).
227 * If they take too much CPU time, the system can lock up and not
228 * get any real work done. This will drop the sample rate when
229 * we detect that events are taking too long.
230 */
231#define NR_ACCUMULATED_SAMPLES 128
232static DEFINE_PER_CPU(u64, running_sample_length);
233
234static void perf_duration_warn(struct irq_work *w)
235{
236 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
237 u64 avg_local_sample_len;
238 u64 local_samples_len;
239
240 local_samples_len = __get_cpu_var(running_sample_length);
241 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
242
243 printk_ratelimited(KERN_WARNING
244 "perf interrupt took too long (%lld > %lld), lowering "
245 "kernel.perf_event_max_sample_rate to %d\n",
246 avg_local_sample_len, allowed_ns >> 1,
247 sysctl_perf_event_sample_rate);
248}
249
250static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
251
252void perf_sample_event_took(u64 sample_len_ns)
253{
254 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
255 u64 avg_local_sample_len;
256 u64 local_samples_len;
257
258 if (allowed_ns == 0)
259 return;
260
261 /* decay the counter by 1 average sample */
262 local_samples_len = __get_cpu_var(running_sample_length);
263 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
264 local_samples_len += sample_len_ns;
265 __get_cpu_var(running_sample_length) = local_samples_len;
266
267 /*
268 * note: this will be biased artifically low until we have
269 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
270 * from having to maintain a count.
271 */
272 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
273
274 if (avg_local_sample_len <= allowed_ns)
275 return;
276
277 if (max_samples_per_tick <= 1)
278 return;
279
280 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
281 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
282 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
283
284 update_perf_cpu_limits();
285
286 if (!irq_work_queue(&perf_duration_work)) {
287 early_printk("perf interrupt took too long (%lld > %lld), lowering "
288 "kernel.perf_event_max_sample_rate to %d\n",
289 avg_local_sample_len, allowed_ns >> 1,
290 sysctl_perf_event_sample_rate);
291 }
292}
293
294static atomic64_t perf_event_id;
295
296static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
297 enum event_type_t event_type);
298
299static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
300 enum event_type_t event_type,
301 struct task_struct *task);
302
303static void update_context_time(struct perf_event_context *ctx);
304static u64 perf_event_time(struct perf_event *event);
305
306void __weak perf_event_print_debug(void) { }
307
308extern __weak const char *perf_pmu_name(void)
309{
310 return "pmu";
311}
312
313static inline u64 perf_clock(void)
314{
315 return local_clock();
316}
317
318static inline struct perf_cpu_context *
319__get_cpu_context(struct perf_event_context *ctx)
320{
321 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
322}
323
324static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
325 struct perf_event_context *ctx)
326{
327 raw_spin_lock(&cpuctx->ctx.lock);
328 if (ctx)
329 raw_spin_lock(&ctx->lock);
330}
331
332static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
333 struct perf_event_context *ctx)
334{
335 if (ctx)
336 raw_spin_unlock(&ctx->lock);
337 raw_spin_unlock(&cpuctx->ctx.lock);
338}
339
340#ifdef CONFIG_CGROUP_PERF
341
342/*
343 * perf_cgroup_info keeps track of time_enabled for a cgroup.
344 * This is a per-cpu dynamically allocated data structure.
345 */
346struct perf_cgroup_info {
347 u64 time;
348 u64 timestamp;
349};
350
351struct perf_cgroup {
352 struct cgroup_subsys_state css;
353 struct perf_cgroup_info __percpu *info;
354};
355
356/*
357 * Must ensure cgroup is pinned (css_get) before calling
358 * this function. In other words, we cannot call this function
359 * if there is no cgroup event for the current CPU context.
360 */
361static inline struct perf_cgroup *
362perf_cgroup_from_task(struct task_struct *task)
363{
364 return container_of(task_css(task, perf_event_cgrp_id),
365 struct perf_cgroup, css);
366}
367
368static inline bool
369perf_cgroup_match(struct perf_event *event)
370{
371 struct perf_event_context *ctx = event->ctx;
372 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
373
374 /* @event doesn't care about cgroup */
375 if (!event->cgrp)
376 return true;
377
378 /* wants specific cgroup scope but @cpuctx isn't associated with any */
379 if (!cpuctx->cgrp)
380 return false;
381
382 /*
383 * Cgroup scoping is recursive. An event enabled for a cgroup is
384 * also enabled for all its descendant cgroups. If @cpuctx's
385 * cgroup is a descendant of @event's (the test covers identity
386 * case), it's a match.
387 */
388 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
389 event->cgrp->css.cgroup);
390}
391
392static inline void perf_put_cgroup(struct perf_event *event)
393{
394 css_put(&event->cgrp->css);
395}
396
397static inline void perf_detach_cgroup(struct perf_event *event)
398{
399 perf_put_cgroup(event);
400 event->cgrp = NULL;
401}
402
403static inline int is_cgroup_event(struct perf_event *event)
404{
405 return event->cgrp != NULL;
406}
407
408static inline u64 perf_cgroup_event_time(struct perf_event *event)
409{
410 struct perf_cgroup_info *t;
411
412 t = per_cpu_ptr(event->cgrp->info, event->cpu);
413 return t->time;
414}
415
416static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
417{
418 struct perf_cgroup_info *info;
419 u64 now;
420
421 now = perf_clock();
422
423 info = this_cpu_ptr(cgrp->info);
424
425 info->time += now - info->timestamp;
426 info->timestamp = now;
427}
428
429static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
430{
431 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
432 if (cgrp_out)
433 __update_cgrp_time(cgrp_out);
434}
435
436static inline void update_cgrp_time_from_event(struct perf_event *event)
437{
438 struct perf_cgroup *cgrp;
439
440 /*
441 * ensure we access cgroup data only when needed and
442 * when we know the cgroup is pinned (css_get)
443 */
444 if (!is_cgroup_event(event))
445 return;
446
447 cgrp = perf_cgroup_from_task(current);
448 /*
449 * Do not update time when cgroup is not active
450 */
451 if (cgrp == event->cgrp)
452 __update_cgrp_time(event->cgrp);
453}
454
455static inline void
456perf_cgroup_set_timestamp(struct task_struct *task,
457 struct perf_event_context *ctx)
458{
459 struct perf_cgroup *cgrp;
460 struct perf_cgroup_info *info;
461
462 /*
463 * ctx->lock held by caller
464 * ensure we do not access cgroup data
465 * unless we have the cgroup pinned (css_get)
466 */
467 if (!task || !ctx->nr_cgroups)
468 return;
469
470 cgrp = perf_cgroup_from_task(task);
471 info = this_cpu_ptr(cgrp->info);
472 info->timestamp = ctx->timestamp;
473}
474
475#define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
476#define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
477
478/*
479 * reschedule events based on the cgroup constraint of task.
480 *
481 * mode SWOUT : schedule out everything
482 * mode SWIN : schedule in based on cgroup for next
483 */
484void perf_cgroup_switch(struct task_struct *task, int mode)
485{
486 struct perf_cpu_context *cpuctx;
487 struct pmu *pmu;
488 unsigned long flags;
489
490 /*
491 * disable interrupts to avoid geting nr_cgroup
492 * changes via __perf_event_disable(). Also
493 * avoids preemption.
494 */
495 local_irq_save(flags);
496
497 /*
498 * we reschedule only in the presence of cgroup
499 * constrained events.
500 */
501 rcu_read_lock();
502
503 list_for_each_entry_rcu(pmu, &pmus, entry) {
504 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
505 if (cpuctx->unique_pmu != pmu)
506 continue; /* ensure we process each cpuctx once */
507
508 /*
509 * perf_cgroup_events says at least one
510 * context on this CPU has cgroup events.
511 *
512 * ctx->nr_cgroups reports the number of cgroup
513 * events for a context.
514 */
515 if (cpuctx->ctx.nr_cgroups > 0) {
516 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
517 perf_pmu_disable(cpuctx->ctx.pmu);
518
519 if (mode & PERF_CGROUP_SWOUT) {
520 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
521 /*
522 * must not be done before ctxswout due
523 * to event_filter_match() in event_sched_out()
524 */
525 cpuctx->cgrp = NULL;
526 }
527
528 if (mode & PERF_CGROUP_SWIN) {
529 WARN_ON_ONCE(cpuctx->cgrp);
530 /*
531 * set cgrp before ctxsw in to allow
532 * event_filter_match() to not have to pass
533 * task around
534 */
535 cpuctx->cgrp = perf_cgroup_from_task(task);
536 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
537 }
538 perf_pmu_enable(cpuctx->ctx.pmu);
539 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
540 }
541 }
542
543 rcu_read_unlock();
544
545 local_irq_restore(flags);
546}
547
548static inline void perf_cgroup_sched_out(struct task_struct *task,
549 struct task_struct *next)
550{
551 struct perf_cgroup *cgrp1;
552 struct perf_cgroup *cgrp2 = NULL;
553
554 /*
555 * we come here when we know perf_cgroup_events > 0
556 */
557 cgrp1 = perf_cgroup_from_task(task);
558
559 /*
560 * next is NULL when called from perf_event_enable_on_exec()
561 * that will systematically cause a cgroup_switch()
562 */
563 if (next)
564 cgrp2 = perf_cgroup_from_task(next);
565
566 /*
567 * only schedule out current cgroup events if we know
568 * that we are switching to a different cgroup. Otherwise,
569 * do no touch the cgroup events.
570 */
571 if (cgrp1 != cgrp2)
572 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
573}
574
575static inline void perf_cgroup_sched_in(struct task_struct *prev,
576 struct task_struct *task)
577{
578 struct perf_cgroup *cgrp1;
579 struct perf_cgroup *cgrp2 = NULL;
580
581 /*
582 * we come here when we know perf_cgroup_events > 0
583 */
584 cgrp1 = perf_cgroup_from_task(task);
585
586 /* prev can never be NULL */
587 cgrp2 = perf_cgroup_from_task(prev);
588
589 /*
590 * only need to schedule in cgroup events if we are changing
591 * cgroup during ctxsw. Cgroup events were not scheduled
592 * out of ctxsw out if that was not the case.
593 */
594 if (cgrp1 != cgrp2)
595 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
596}
597
598static inline int perf_cgroup_connect(int fd, struct perf_event *event,
599 struct perf_event_attr *attr,
600 struct perf_event *group_leader)
601{
602 struct perf_cgroup *cgrp;
603 struct cgroup_subsys_state *css;
604 struct fd f = fdget(fd);
605 int ret = 0;
606
607 if (!f.file)
608 return -EBADF;
609
610 css = css_tryget_from_dir(f.file->f_dentry, &perf_event_cgrp_subsys);
611 if (IS_ERR(css)) {
612 ret = PTR_ERR(css);
613 goto out;
614 }
615
616 cgrp = container_of(css, struct perf_cgroup, css);
617 event->cgrp = cgrp;
618
619 /*
620 * all events in a group must monitor
621 * the same cgroup because a task belongs
622 * to only one perf cgroup at a time
623 */
624 if (group_leader && group_leader->cgrp != cgrp) {
625 perf_detach_cgroup(event);
626 ret = -EINVAL;
627 }
628out:
629 fdput(f);
630 return ret;
631}
632
633static inline void
634perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
635{
636 struct perf_cgroup_info *t;
637 t = per_cpu_ptr(event->cgrp->info, event->cpu);
638 event->shadow_ctx_time = now - t->timestamp;
639}
640
641static inline void
642perf_cgroup_defer_enabled(struct perf_event *event)
643{
644 /*
645 * when the current task's perf cgroup does not match
646 * the event's, we need to remember to call the
647 * perf_mark_enable() function the first time a task with
648 * a matching perf cgroup is scheduled in.
649 */
650 if (is_cgroup_event(event) && !perf_cgroup_match(event))
651 event->cgrp_defer_enabled = 1;
652}
653
654static inline void
655perf_cgroup_mark_enabled(struct perf_event *event,
656 struct perf_event_context *ctx)
657{
658 struct perf_event *sub;
659 u64 tstamp = perf_event_time(event);
660
661 if (!event->cgrp_defer_enabled)
662 return;
663
664 event->cgrp_defer_enabled = 0;
665
666 event->tstamp_enabled = tstamp - event->total_time_enabled;
667 list_for_each_entry(sub, &event->sibling_list, group_entry) {
668 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
669 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
670 sub->cgrp_defer_enabled = 0;
671 }
672 }
673}
674#else /* !CONFIG_CGROUP_PERF */
675
676static inline bool
677perf_cgroup_match(struct perf_event *event)
678{
679 return true;
680}
681
682static inline void perf_detach_cgroup(struct perf_event *event)
683{}
684
685static inline int is_cgroup_event(struct perf_event *event)
686{
687 return 0;
688}
689
690static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
691{
692 return 0;
693}
694
695static inline void update_cgrp_time_from_event(struct perf_event *event)
696{
697}
698
699static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
700{
701}
702
703static inline void perf_cgroup_sched_out(struct task_struct *task,
704 struct task_struct *next)
705{
706}
707
708static inline void perf_cgroup_sched_in(struct task_struct *prev,
709 struct task_struct *task)
710{
711}
712
713static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
714 struct perf_event_attr *attr,
715 struct perf_event *group_leader)
716{
717 return -EINVAL;
718}
719
720static inline void
721perf_cgroup_set_timestamp(struct task_struct *task,
722 struct perf_event_context *ctx)
723{
724}
725
726void
727perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
728{
729}
730
731static inline void
732perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
733{
734}
735
736static inline u64 perf_cgroup_event_time(struct perf_event *event)
737{
738 return 0;
739}
740
741static inline void
742perf_cgroup_defer_enabled(struct perf_event *event)
743{
744}
745
746static inline void
747perf_cgroup_mark_enabled(struct perf_event *event,
748 struct perf_event_context *ctx)
749{
750}
751#endif
752
753/*
754 * set default to be dependent on timer tick just
755 * like original code
756 */
757#define PERF_CPU_HRTIMER (1000 / HZ)
758/*
759 * function must be called with interrupts disbled
760 */
761static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
762{
763 struct perf_cpu_context *cpuctx;
764 enum hrtimer_restart ret = HRTIMER_NORESTART;
765 int rotations = 0;
766
767 WARN_ON(!irqs_disabled());
768
769 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
770
771 rotations = perf_rotate_context(cpuctx);
772
773 /*
774 * arm timer if needed
775 */
776 if (rotations) {
777 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
778 ret = HRTIMER_RESTART;
779 }
780
781 return ret;
782}
783
784/* CPU is going down */
785void perf_cpu_hrtimer_cancel(int cpu)
786{
787 struct perf_cpu_context *cpuctx;
788 struct pmu *pmu;
789 unsigned long flags;
790
791 if (WARN_ON(cpu != smp_processor_id()))
792 return;
793
794 local_irq_save(flags);
795
796 rcu_read_lock();
797
798 list_for_each_entry_rcu(pmu, &pmus, entry) {
799 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
800
801 if (pmu->task_ctx_nr == perf_sw_context)
802 continue;
803
804 hrtimer_cancel(&cpuctx->hrtimer);
805 }
806
807 rcu_read_unlock();
808
809 local_irq_restore(flags);
810}
811
812static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
813{
814 struct hrtimer *hr = &cpuctx->hrtimer;
815 struct pmu *pmu = cpuctx->ctx.pmu;
816 int timer;
817
818 /* no multiplexing needed for SW PMU */
819 if (pmu->task_ctx_nr == perf_sw_context)
820 return;
821
822 /*
823 * check default is sane, if not set then force to
824 * default interval (1/tick)
825 */
826 timer = pmu->hrtimer_interval_ms;
827 if (timer < 1)
828 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
829
830 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
831
832 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
833 hr->function = perf_cpu_hrtimer_handler;
834}
835
836static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
837{
838 struct hrtimer *hr = &cpuctx->hrtimer;
839 struct pmu *pmu = cpuctx->ctx.pmu;
840
841 /* not for SW PMU */
842 if (pmu->task_ctx_nr == perf_sw_context)
843 return;
844
845 if (hrtimer_active(hr))
846 return;
847
848 if (!hrtimer_callback_running(hr))
849 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
850 0, HRTIMER_MODE_REL_PINNED, 0);
851}
852
853void perf_pmu_disable(struct pmu *pmu)
854{
855 int *count = this_cpu_ptr(pmu->pmu_disable_count);
856 if (!(*count)++)
857 pmu->pmu_disable(pmu);
858}
859
860void perf_pmu_enable(struct pmu *pmu)
861{
862 int *count = this_cpu_ptr(pmu->pmu_disable_count);
863 if (!--(*count))
864 pmu->pmu_enable(pmu);
865}
866
867static DEFINE_PER_CPU(struct list_head, rotation_list);
868
869/*
870 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
871 * because they're strictly cpu affine and rotate_start is called with IRQs
872 * disabled, while rotate_context is called from IRQ context.
873 */
874static void perf_pmu_rotate_start(struct pmu *pmu)
875{
876 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
877 struct list_head *head = &__get_cpu_var(rotation_list);
878
879 WARN_ON(!irqs_disabled());
880
881 if (list_empty(&cpuctx->rotation_list))
882 list_add(&cpuctx->rotation_list, head);
883}
884
885static void get_ctx(struct perf_event_context *ctx)
886{
887 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
888}
889
890static void put_ctx(struct perf_event_context *ctx)
891{
892 if (atomic_dec_and_test(&ctx->refcount)) {
893 if (ctx->parent_ctx)
894 put_ctx(ctx->parent_ctx);
895 if (ctx->task)
896 put_task_struct(ctx->task);
897 kfree_rcu(ctx, rcu_head);
898 }
899}
900
901static void unclone_ctx(struct perf_event_context *ctx)
902{
903 if (ctx->parent_ctx) {
904 put_ctx(ctx->parent_ctx);
905 ctx->parent_ctx = NULL;
906 }
907 ctx->generation++;
908}
909
910static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
911{
912 /*
913 * only top level events have the pid namespace they were created in
914 */
915 if (event->parent)
916 event = event->parent;
917
918 return task_tgid_nr_ns(p, event->ns);
919}
920
921static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
922{
923 /*
924 * only top level events have the pid namespace they were created in
925 */
926 if (event->parent)
927 event = event->parent;
928
929 return task_pid_nr_ns(p, event->ns);
930}
931
932/*
933 * If we inherit events we want to return the parent event id
934 * to userspace.
935 */
936static u64 primary_event_id(struct perf_event *event)
937{
938 u64 id = event->id;
939
940 if (event->parent)
941 id = event->parent->id;
942
943 return id;
944}
945
946/*
947 * Get the perf_event_context for a task and lock it.
948 * This has to cope with with the fact that until it is locked,
949 * the context could get moved to another task.
950 */
951static struct perf_event_context *
952perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
953{
954 struct perf_event_context *ctx;
955
956retry:
957 /*
958 * One of the few rules of preemptible RCU is that one cannot do
959 * rcu_read_unlock() while holding a scheduler (or nested) lock when
960 * part of the read side critical section was preemptible -- see
961 * rcu_read_unlock_special().
962 *
963 * Since ctx->lock nests under rq->lock we must ensure the entire read
964 * side critical section is non-preemptible.
965 */
966 preempt_disable();
967 rcu_read_lock();
968 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
969 if (ctx) {
970 /*
971 * If this context is a clone of another, it might
972 * get swapped for another underneath us by
973 * perf_event_task_sched_out, though the
974 * rcu_read_lock() protects us from any context
975 * getting freed. Lock the context and check if it
976 * got swapped before we could get the lock, and retry
977 * if so. If we locked the right context, then it
978 * can't get swapped on us any more.
979 */
980 raw_spin_lock_irqsave(&ctx->lock, *flags);
981 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
982 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
983 rcu_read_unlock();
984 preempt_enable();
985 goto retry;
986 }
987
988 if (!atomic_inc_not_zero(&ctx->refcount)) {
989 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
990 ctx = NULL;
991 }
992 }
993 rcu_read_unlock();
994 preempt_enable();
995 return ctx;
996}
997
998/*
999 * Get the context for a task and increment its pin_count so it
1000 * can't get swapped to another task. This also increments its
1001 * reference count so that the context can't get freed.
1002 */
1003static struct perf_event_context *
1004perf_pin_task_context(struct task_struct *task, int ctxn)
1005{
1006 struct perf_event_context *ctx;
1007 unsigned long flags;
1008
1009 ctx = perf_lock_task_context(task, ctxn, &flags);
1010 if (ctx) {
1011 ++ctx->pin_count;
1012 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1013 }
1014 return ctx;
1015}
1016
1017static void perf_unpin_context(struct perf_event_context *ctx)
1018{
1019 unsigned long flags;
1020
1021 raw_spin_lock_irqsave(&ctx->lock, flags);
1022 --ctx->pin_count;
1023 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1024}
1025
1026/*
1027 * Update the record of the current time in a context.
1028 */
1029static void update_context_time(struct perf_event_context *ctx)
1030{
1031 u64 now = perf_clock();
1032
1033 ctx->time += now - ctx->timestamp;
1034 ctx->timestamp = now;
1035}
1036
1037static u64 perf_event_time(struct perf_event *event)
1038{
1039 struct perf_event_context *ctx = event->ctx;
1040
1041 if (is_cgroup_event(event))
1042 return perf_cgroup_event_time(event);
1043
1044 return ctx ? ctx->time : 0;
1045}
1046
1047/*
1048 * Update the total_time_enabled and total_time_running fields for a event.
1049 * The caller of this function needs to hold the ctx->lock.
1050 */
1051static void update_event_times(struct perf_event *event)
1052{
1053 struct perf_event_context *ctx = event->ctx;
1054 u64 run_end;
1055
1056 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1057 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1058 return;
1059 /*
1060 * in cgroup mode, time_enabled represents
1061 * the time the event was enabled AND active
1062 * tasks were in the monitored cgroup. This is
1063 * independent of the activity of the context as
1064 * there may be a mix of cgroup and non-cgroup events.
1065 *
1066 * That is why we treat cgroup events differently
1067 * here.
1068 */
1069 if (is_cgroup_event(event))
1070 run_end = perf_cgroup_event_time(event);
1071 else if (ctx->is_active)
1072 run_end = ctx->time;
1073 else
1074 run_end = event->tstamp_stopped;
1075
1076 event->total_time_enabled = run_end - event->tstamp_enabled;
1077
1078 if (event->state == PERF_EVENT_STATE_INACTIVE)
1079 run_end = event->tstamp_stopped;
1080 else
1081 run_end = perf_event_time(event);
1082
1083 event->total_time_running = run_end - event->tstamp_running;
1084
1085}
1086
1087/*
1088 * Update total_time_enabled and total_time_running for all events in a group.
1089 */
1090static void update_group_times(struct perf_event *leader)
1091{
1092 struct perf_event *event;
1093
1094 update_event_times(leader);
1095 list_for_each_entry(event, &leader->sibling_list, group_entry)
1096 update_event_times(event);
1097}
1098
1099static struct list_head *
1100ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1101{
1102 if (event->attr.pinned)
1103 return &ctx->pinned_groups;
1104 else
1105 return &ctx->flexible_groups;
1106}
1107
1108/*
1109 * Add a event from the lists for its context.
1110 * Must be called with ctx->mutex and ctx->lock held.
1111 */
1112static void
1113list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1114{
1115 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1116 event->attach_state |= PERF_ATTACH_CONTEXT;
1117
1118 /*
1119 * If we're a stand alone event or group leader, we go to the context
1120 * list, group events are kept attached to the group so that
1121 * perf_group_detach can, at all times, locate all siblings.
1122 */
1123 if (event->group_leader == event) {
1124 struct list_head *list;
1125
1126 if (is_software_event(event))
1127 event->group_flags |= PERF_GROUP_SOFTWARE;
1128
1129 list = ctx_group_list(event, ctx);
1130 list_add_tail(&event->group_entry, list);
1131 }
1132
1133 if (is_cgroup_event(event))
1134 ctx->nr_cgroups++;
1135
1136 if (has_branch_stack(event))
1137 ctx->nr_branch_stack++;
1138
1139 list_add_rcu(&event->event_entry, &ctx->event_list);
1140 if (!ctx->nr_events)
1141 perf_pmu_rotate_start(ctx->pmu);
1142 ctx->nr_events++;
1143 if (event->attr.inherit_stat)
1144 ctx->nr_stat++;
1145
1146 ctx->generation++;
1147}
1148
1149/*
1150 * Initialize event state based on the perf_event_attr::disabled.
1151 */
1152static inline void perf_event__state_init(struct perf_event *event)
1153{
1154 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1155 PERF_EVENT_STATE_INACTIVE;
1156}
1157
1158/*
1159 * Called at perf_event creation and when events are attached/detached from a
1160 * group.
1161 */
1162static void perf_event__read_size(struct perf_event *event)
1163{
1164 int entry = sizeof(u64); /* value */
1165 int size = 0;
1166 int nr = 1;
1167
1168 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1169 size += sizeof(u64);
1170
1171 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1172 size += sizeof(u64);
1173
1174 if (event->attr.read_format & PERF_FORMAT_ID)
1175 entry += sizeof(u64);
1176
1177 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1178 nr += event->group_leader->nr_siblings;
1179 size += sizeof(u64);
1180 }
1181
1182 size += entry * nr;
1183 event->read_size = size;
1184}
1185
1186static void perf_event__header_size(struct perf_event *event)
1187{
1188 struct perf_sample_data *data;
1189 u64 sample_type = event->attr.sample_type;
1190 u16 size = 0;
1191
1192 perf_event__read_size(event);
1193
1194 if (sample_type & PERF_SAMPLE_IP)
1195 size += sizeof(data->ip);
1196
1197 if (sample_type & PERF_SAMPLE_ADDR)
1198 size += sizeof(data->addr);
1199
1200 if (sample_type & PERF_SAMPLE_PERIOD)
1201 size += sizeof(data->period);
1202
1203 if (sample_type & PERF_SAMPLE_WEIGHT)
1204 size += sizeof(data->weight);
1205
1206 if (sample_type & PERF_SAMPLE_READ)
1207 size += event->read_size;
1208
1209 if (sample_type & PERF_SAMPLE_DATA_SRC)
1210 size += sizeof(data->data_src.val);
1211
1212 if (sample_type & PERF_SAMPLE_TRANSACTION)
1213 size += sizeof(data->txn);
1214
1215 event->header_size = size;
1216}
1217
1218static void perf_event__id_header_size(struct perf_event *event)
1219{
1220 struct perf_sample_data *data;
1221 u64 sample_type = event->attr.sample_type;
1222 u16 size = 0;
1223
1224 if (sample_type & PERF_SAMPLE_TID)
1225 size += sizeof(data->tid_entry);
1226
1227 if (sample_type & PERF_SAMPLE_TIME)
1228 size += sizeof(data->time);
1229
1230 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1231 size += sizeof(data->id);
1232
1233 if (sample_type & PERF_SAMPLE_ID)
1234 size += sizeof(data->id);
1235
1236 if (sample_type & PERF_SAMPLE_STREAM_ID)
1237 size += sizeof(data->stream_id);
1238
1239 if (sample_type & PERF_SAMPLE_CPU)
1240 size += sizeof(data->cpu_entry);
1241
1242 event->id_header_size = size;
1243}
1244
1245static void perf_group_attach(struct perf_event *event)
1246{
1247 struct perf_event *group_leader = event->group_leader, *pos;
1248
1249 /*
1250 * We can have double attach due to group movement in perf_event_open.
1251 */
1252 if (event->attach_state & PERF_ATTACH_GROUP)
1253 return;
1254
1255 event->attach_state |= PERF_ATTACH_GROUP;
1256
1257 if (group_leader == event)
1258 return;
1259
1260 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1261 !is_software_event(event))
1262 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1263
1264 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1265 group_leader->nr_siblings++;
1266
1267 perf_event__header_size(group_leader);
1268
1269 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1270 perf_event__header_size(pos);
1271}
1272
1273/*
1274 * Remove a event from the lists for its context.
1275 * Must be called with ctx->mutex and ctx->lock held.
1276 */
1277static void
1278list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1279{
1280 struct perf_cpu_context *cpuctx;
1281 /*
1282 * We can have double detach due to exit/hot-unplug + close.
1283 */
1284 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1285 return;
1286
1287 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1288
1289 if (is_cgroup_event(event)) {
1290 ctx->nr_cgroups--;
1291 cpuctx = __get_cpu_context(ctx);
1292 /*
1293 * if there are no more cgroup events
1294 * then cler cgrp to avoid stale pointer
1295 * in update_cgrp_time_from_cpuctx()
1296 */
1297 if (!ctx->nr_cgroups)
1298 cpuctx->cgrp = NULL;
1299 }
1300
1301 if (has_branch_stack(event))
1302 ctx->nr_branch_stack--;
1303
1304 ctx->nr_events--;
1305 if (event->attr.inherit_stat)
1306 ctx->nr_stat--;
1307
1308 list_del_rcu(&event->event_entry);
1309
1310 if (event->group_leader == event)
1311 list_del_init(&event->group_entry);
1312
1313 update_group_times(event);
1314
1315 /*
1316 * If event was in error state, then keep it
1317 * that way, otherwise bogus counts will be
1318 * returned on read(). The only way to get out
1319 * of error state is by explicit re-enabling
1320 * of the event
1321 */
1322 if (event->state > PERF_EVENT_STATE_OFF)
1323 event->state = PERF_EVENT_STATE_OFF;
1324
1325 ctx->generation++;
1326}
1327
1328static void perf_group_detach(struct perf_event *event)
1329{
1330 struct perf_event *sibling, *tmp;
1331 struct list_head *list = NULL;
1332
1333 /*
1334 * We can have double detach due to exit/hot-unplug + close.
1335 */
1336 if (!(event->attach_state & PERF_ATTACH_GROUP))
1337 return;
1338
1339 event->attach_state &= ~PERF_ATTACH_GROUP;
1340
1341 /*
1342 * If this is a sibling, remove it from its group.
1343 */
1344 if (event->group_leader != event) {
1345 list_del_init(&event->group_entry);
1346 event->group_leader->nr_siblings--;
1347 goto out;
1348 }
1349
1350 if (!list_empty(&event->group_entry))
1351 list = &event->group_entry;
1352
1353 /*
1354 * If this was a group event with sibling events then
1355 * upgrade the siblings to singleton events by adding them
1356 * to whatever list we are on.
1357 */
1358 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1359 if (list)
1360 list_move_tail(&sibling->group_entry, list);
1361 sibling->group_leader = sibling;
1362
1363 /* Inherit group flags from the previous leader */
1364 sibling->group_flags = event->group_flags;
1365 }
1366
1367out:
1368 perf_event__header_size(event->group_leader);
1369
1370 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1371 perf_event__header_size(tmp);
1372}
1373
1374static inline int
1375event_filter_match(struct perf_event *event)
1376{
1377 return (event->cpu == -1 || event->cpu == smp_processor_id())
1378 && perf_cgroup_match(event);
1379}
1380
1381static void
1382event_sched_out(struct perf_event *event,
1383 struct perf_cpu_context *cpuctx,
1384 struct perf_event_context *ctx)
1385{
1386 u64 tstamp = perf_event_time(event);
1387 u64 delta;
1388 /*
1389 * An event which could not be activated because of
1390 * filter mismatch still needs to have its timings
1391 * maintained, otherwise bogus information is return
1392 * via read() for time_enabled, time_running:
1393 */
1394 if (event->state == PERF_EVENT_STATE_INACTIVE
1395 && !event_filter_match(event)) {
1396 delta = tstamp - event->tstamp_stopped;
1397 event->tstamp_running += delta;
1398 event->tstamp_stopped = tstamp;
1399 }
1400
1401 if (event->state != PERF_EVENT_STATE_ACTIVE)
1402 return;
1403
1404 perf_pmu_disable(event->pmu);
1405
1406 event->state = PERF_EVENT_STATE_INACTIVE;
1407 if (event->pending_disable) {
1408 event->pending_disable = 0;
1409 event->state = PERF_EVENT_STATE_OFF;
1410 }
1411 event->tstamp_stopped = tstamp;
1412 event->pmu->del(event, 0);
1413 event->oncpu = -1;
1414
1415 if (!is_software_event(event))
1416 cpuctx->active_oncpu--;
1417 ctx->nr_active--;
1418 if (event->attr.freq && event->attr.sample_freq)
1419 ctx->nr_freq--;
1420 if (event->attr.exclusive || !cpuctx->active_oncpu)
1421 cpuctx->exclusive = 0;
1422
1423 perf_pmu_enable(event->pmu);
1424}
1425
1426static void
1427group_sched_out(struct perf_event *group_event,
1428 struct perf_cpu_context *cpuctx,
1429 struct perf_event_context *ctx)
1430{
1431 struct perf_event *event;
1432 int state = group_event->state;
1433
1434 event_sched_out(group_event, cpuctx, ctx);
1435
1436 /*
1437 * Schedule out siblings (if any):
1438 */
1439 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1440 event_sched_out(event, cpuctx, ctx);
1441
1442 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1443 cpuctx->exclusive = 0;
1444}
1445
1446struct remove_event {
1447 struct perf_event *event;
1448 bool detach_group;
1449};
1450
1451/*
1452 * Cross CPU call to remove a performance event
1453 *
1454 * We disable the event on the hardware level first. After that we
1455 * remove it from the context list.
1456 */
1457static int __perf_remove_from_context(void *info)
1458{
1459 struct remove_event *re = info;
1460 struct perf_event *event = re->event;
1461 struct perf_event_context *ctx = event->ctx;
1462 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1463
1464 raw_spin_lock(&ctx->lock);
1465 event_sched_out(event, cpuctx, ctx);
1466 if (re->detach_group)
1467 perf_group_detach(event);
1468 list_del_event(event, ctx);
1469 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1470 ctx->is_active = 0;
1471 cpuctx->task_ctx = NULL;
1472 }
1473 raw_spin_unlock(&ctx->lock);
1474
1475 return 0;
1476}
1477
1478
1479/*
1480 * Remove the event from a task's (or a CPU's) list of events.
1481 *
1482 * CPU events are removed with a smp call. For task events we only
1483 * call when the task is on a CPU.
1484 *
1485 * If event->ctx is a cloned context, callers must make sure that
1486 * every task struct that event->ctx->task could possibly point to
1487 * remains valid. This is OK when called from perf_release since
1488 * that only calls us on the top-level context, which can't be a clone.
1489 * When called from perf_event_exit_task, it's OK because the
1490 * context has been detached from its task.
1491 */
1492static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1493{
1494 struct perf_event_context *ctx = event->ctx;
1495 struct task_struct *task = ctx->task;
1496 struct remove_event re = {
1497 .event = event,
1498 .detach_group = detach_group,
1499 };
1500
1501 lockdep_assert_held(&ctx->mutex);
1502
1503 if (!task) {
1504 /*
1505 * Per cpu events are removed via an smp call and
1506 * the removal is always successful.
1507 */
1508 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1509 return;
1510 }
1511
1512retry:
1513 if (!task_function_call(task, __perf_remove_from_context, &re))
1514 return;
1515
1516 raw_spin_lock_irq(&ctx->lock);
1517 /*
1518 * If we failed to find a running task, but find the context active now
1519 * that we've acquired the ctx->lock, retry.
1520 */
1521 if (ctx->is_active) {
1522 raw_spin_unlock_irq(&ctx->lock);
1523 goto retry;
1524 }
1525
1526 /*
1527 * Since the task isn't running, its safe to remove the event, us
1528 * holding the ctx->lock ensures the task won't get scheduled in.
1529 */
1530 if (detach_group)
1531 perf_group_detach(event);
1532 list_del_event(event, ctx);
1533 raw_spin_unlock_irq(&ctx->lock);
1534}
1535
1536/*
1537 * Cross CPU call to disable a performance event
1538 */
1539int __perf_event_disable(void *info)
1540{
1541 struct perf_event *event = info;
1542 struct perf_event_context *ctx = event->ctx;
1543 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1544
1545 /*
1546 * If this is a per-task event, need to check whether this
1547 * event's task is the current task on this cpu.
1548 *
1549 * Can trigger due to concurrent perf_event_context_sched_out()
1550 * flipping contexts around.
1551 */
1552 if (ctx->task && cpuctx->task_ctx != ctx)
1553 return -EINVAL;
1554
1555 raw_spin_lock(&ctx->lock);
1556
1557 /*
1558 * If the event is on, turn it off.
1559 * If it is in error state, leave it in error state.
1560 */
1561 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1562 update_context_time(ctx);
1563 update_cgrp_time_from_event(event);
1564 update_group_times(event);
1565 if (event == event->group_leader)
1566 group_sched_out(event, cpuctx, ctx);
1567 else
1568 event_sched_out(event, cpuctx, ctx);
1569 event->state = PERF_EVENT_STATE_OFF;
1570 }
1571
1572 raw_spin_unlock(&ctx->lock);
1573
1574 return 0;
1575}
1576
1577/*
1578 * Disable a event.
1579 *
1580 * If event->ctx is a cloned context, callers must make sure that
1581 * every task struct that event->ctx->task could possibly point to
1582 * remains valid. This condition is satisifed when called through
1583 * perf_event_for_each_child or perf_event_for_each because they
1584 * hold the top-level event's child_mutex, so any descendant that
1585 * goes to exit will block in sync_child_event.
1586 * When called from perf_pending_event it's OK because event->ctx
1587 * is the current context on this CPU and preemption is disabled,
1588 * hence we can't get into perf_event_task_sched_out for this context.
1589 */
1590void perf_event_disable(struct perf_event *event)
1591{
1592 struct perf_event_context *ctx = event->ctx;
1593 struct task_struct *task = ctx->task;
1594
1595 if (!task) {
1596 /*
1597 * Disable the event on the cpu that it's on
1598 */
1599 cpu_function_call(event->cpu, __perf_event_disable, event);
1600 return;
1601 }
1602
1603retry:
1604 if (!task_function_call(task, __perf_event_disable, event))
1605 return;
1606
1607 raw_spin_lock_irq(&ctx->lock);
1608 /*
1609 * If the event is still active, we need to retry the cross-call.
1610 */
1611 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1612 raw_spin_unlock_irq(&ctx->lock);
1613 /*
1614 * Reload the task pointer, it might have been changed by
1615 * a concurrent perf_event_context_sched_out().
1616 */
1617 task = ctx->task;
1618 goto retry;
1619 }
1620
1621 /*
1622 * Since we have the lock this context can't be scheduled
1623 * in, so we can change the state safely.
1624 */
1625 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1626 update_group_times(event);
1627 event->state = PERF_EVENT_STATE_OFF;
1628 }
1629 raw_spin_unlock_irq(&ctx->lock);
1630}
1631EXPORT_SYMBOL_GPL(perf_event_disable);
1632
1633static void perf_set_shadow_time(struct perf_event *event,
1634 struct perf_event_context *ctx,
1635 u64 tstamp)
1636{
1637 /*
1638 * use the correct time source for the time snapshot
1639 *
1640 * We could get by without this by leveraging the
1641 * fact that to get to this function, the caller
1642 * has most likely already called update_context_time()
1643 * and update_cgrp_time_xx() and thus both timestamp
1644 * are identical (or very close). Given that tstamp is,
1645 * already adjusted for cgroup, we could say that:
1646 * tstamp - ctx->timestamp
1647 * is equivalent to
1648 * tstamp - cgrp->timestamp.
1649 *
1650 * Then, in perf_output_read(), the calculation would
1651 * work with no changes because:
1652 * - event is guaranteed scheduled in
1653 * - no scheduled out in between
1654 * - thus the timestamp would be the same
1655 *
1656 * But this is a bit hairy.
1657 *
1658 * So instead, we have an explicit cgroup call to remain
1659 * within the time time source all along. We believe it
1660 * is cleaner and simpler to understand.
1661 */
1662 if (is_cgroup_event(event))
1663 perf_cgroup_set_shadow_time(event, tstamp);
1664 else
1665 event->shadow_ctx_time = tstamp - ctx->timestamp;
1666}
1667
1668#define MAX_INTERRUPTS (~0ULL)
1669
1670static void perf_log_throttle(struct perf_event *event, int enable);
1671
1672static int
1673event_sched_in(struct perf_event *event,
1674 struct perf_cpu_context *cpuctx,
1675 struct perf_event_context *ctx)
1676{
1677 u64 tstamp = perf_event_time(event);
1678 int ret = 0;
1679
1680 if (event->state <= PERF_EVENT_STATE_OFF)
1681 return 0;
1682
1683 event->state = PERF_EVENT_STATE_ACTIVE;
1684 event->oncpu = smp_processor_id();
1685
1686 /*
1687 * Unthrottle events, since we scheduled we might have missed several
1688 * ticks already, also for a heavily scheduling task there is little
1689 * guarantee it'll get a tick in a timely manner.
1690 */
1691 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1692 perf_log_throttle(event, 1);
1693 event->hw.interrupts = 0;
1694 }
1695
1696 /*
1697 * The new state must be visible before we turn it on in the hardware:
1698 */
1699 smp_wmb();
1700
1701 perf_pmu_disable(event->pmu);
1702
1703 if (event->pmu->add(event, PERF_EF_START)) {
1704 event->state = PERF_EVENT_STATE_INACTIVE;
1705 event->oncpu = -1;
1706 ret = -EAGAIN;
1707 goto out;
1708 }
1709
1710 event->tstamp_running += tstamp - event->tstamp_stopped;
1711
1712 perf_set_shadow_time(event, ctx, tstamp);
1713
1714 if (!is_software_event(event))
1715 cpuctx->active_oncpu++;
1716 ctx->nr_active++;
1717 if (event->attr.freq && event->attr.sample_freq)
1718 ctx->nr_freq++;
1719
1720 if (event->attr.exclusive)
1721 cpuctx->exclusive = 1;
1722
1723out:
1724 perf_pmu_enable(event->pmu);
1725
1726 return ret;
1727}
1728
1729static int
1730group_sched_in(struct perf_event *group_event,
1731 struct perf_cpu_context *cpuctx,
1732 struct perf_event_context *ctx)
1733{
1734 struct perf_event *event, *partial_group = NULL;
1735 struct pmu *pmu = ctx->pmu;
1736 u64 now = ctx->time;
1737 bool simulate = false;
1738
1739 if (group_event->state == PERF_EVENT_STATE_OFF)
1740 return 0;
1741
1742 pmu->start_txn(pmu);
1743
1744 if (event_sched_in(group_event, cpuctx, ctx)) {
1745 pmu->cancel_txn(pmu);
1746 perf_cpu_hrtimer_restart(cpuctx);
1747 return -EAGAIN;
1748 }
1749
1750 /*
1751 * Schedule in siblings as one group (if any):
1752 */
1753 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1754 if (event_sched_in(event, cpuctx, ctx)) {
1755 partial_group = event;
1756 goto group_error;
1757 }
1758 }
1759
1760 if (!pmu->commit_txn(pmu))
1761 return 0;
1762
1763group_error:
1764 /*
1765 * Groups can be scheduled in as one unit only, so undo any
1766 * partial group before returning:
1767 * The events up to the failed event are scheduled out normally,
1768 * tstamp_stopped will be updated.
1769 *
1770 * The failed events and the remaining siblings need to have
1771 * their timings updated as if they had gone thru event_sched_in()
1772 * and event_sched_out(). This is required to get consistent timings
1773 * across the group. This also takes care of the case where the group
1774 * could never be scheduled by ensuring tstamp_stopped is set to mark
1775 * the time the event was actually stopped, such that time delta
1776 * calculation in update_event_times() is correct.
1777 */
1778 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1779 if (event == partial_group)
1780 simulate = true;
1781
1782 if (simulate) {
1783 event->tstamp_running += now - event->tstamp_stopped;
1784 event->tstamp_stopped = now;
1785 } else {
1786 event_sched_out(event, cpuctx, ctx);
1787 }
1788 }
1789 event_sched_out(group_event, cpuctx, ctx);
1790
1791 pmu->cancel_txn(pmu);
1792
1793 perf_cpu_hrtimer_restart(cpuctx);
1794
1795 return -EAGAIN;
1796}
1797
1798/*
1799 * Work out whether we can put this event group on the CPU now.
1800 */
1801static int group_can_go_on(struct perf_event *event,
1802 struct perf_cpu_context *cpuctx,
1803 int can_add_hw)
1804{
1805 /*
1806 * Groups consisting entirely of software events can always go on.
1807 */
1808 if (event->group_flags & PERF_GROUP_SOFTWARE)
1809 return 1;
1810 /*
1811 * If an exclusive group is already on, no other hardware
1812 * events can go on.
1813 */
1814 if (cpuctx->exclusive)
1815 return 0;
1816 /*
1817 * If this group is exclusive and there are already
1818 * events on the CPU, it can't go on.
1819 */
1820 if (event->attr.exclusive && cpuctx->active_oncpu)
1821 return 0;
1822 /*
1823 * Otherwise, try to add it if all previous groups were able
1824 * to go on.
1825 */
1826 return can_add_hw;
1827}
1828
1829static void add_event_to_ctx(struct perf_event *event,
1830 struct perf_event_context *ctx)
1831{
1832 u64 tstamp = perf_event_time(event);
1833
1834 list_add_event(event, ctx);
1835 perf_group_attach(event);
1836 event->tstamp_enabled = tstamp;
1837 event->tstamp_running = tstamp;
1838 event->tstamp_stopped = tstamp;
1839}
1840
1841static void task_ctx_sched_out(struct perf_event_context *ctx);
1842static void
1843ctx_sched_in(struct perf_event_context *ctx,
1844 struct perf_cpu_context *cpuctx,
1845 enum event_type_t event_type,
1846 struct task_struct *task);
1847
1848static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1849 struct perf_event_context *ctx,
1850 struct task_struct *task)
1851{
1852 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1853 if (ctx)
1854 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1855 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1856 if (ctx)
1857 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1858}
1859
1860/*
1861 * Cross CPU call to install and enable a performance event
1862 *
1863 * Must be called with ctx->mutex held
1864 */
1865static int __perf_install_in_context(void *info)
1866{
1867 struct perf_event *event = info;
1868 struct perf_event_context *ctx = event->ctx;
1869 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1870 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1871 struct task_struct *task = current;
1872
1873 perf_ctx_lock(cpuctx, task_ctx);
1874 perf_pmu_disable(cpuctx->ctx.pmu);
1875
1876 /*
1877 * If there was an active task_ctx schedule it out.
1878 */
1879 if (task_ctx)
1880 task_ctx_sched_out(task_ctx);
1881
1882 /*
1883 * If the context we're installing events in is not the
1884 * active task_ctx, flip them.
1885 */
1886 if (ctx->task && task_ctx != ctx) {
1887 if (task_ctx)
1888 raw_spin_unlock(&task_ctx->lock);
1889 raw_spin_lock(&ctx->lock);
1890 task_ctx = ctx;
1891 }
1892
1893 if (task_ctx) {
1894 cpuctx->task_ctx = task_ctx;
1895 task = task_ctx->task;
1896 }
1897
1898 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1899
1900 update_context_time(ctx);
1901 /*
1902 * update cgrp time only if current cgrp
1903 * matches event->cgrp. Must be done before
1904 * calling add_event_to_ctx()
1905 */
1906 update_cgrp_time_from_event(event);
1907
1908 add_event_to_ctx(event, ctx);
1909
1910 /*
1911 * Schedule everything back in
1912 */
1913 perf_event_sched_in(cpuctx, task_ctx, task);
1914
1915 perf_pmu_enable(cpuctx->ctx.pmu);
1916 perf_ctx_unlock(cpuctx, task_ctx);
1917
1918 return 0;
1919}
1920
1921/*
1922 * Attach a performance event to a context
1923 *
1924 * First we add the event to the list with the hardware enable bit
1925 * in event->hw_config cleared.
1926 *
1927 * If the event is attached to a task which is on a CPU we use a smp
1928 * call to enable it in the task context. The task might have been
1929 * scheduled away, but we check this in the smp call again.
1930 */
1931static void
1932perf_install_in_context(struct perf_event_context *ctx,
1933 struct perf_event *event,
1934 int cpu)
1935{
1936 struct task_struct *task = ctx->task;
1937
1938 lockdep_assert_held(&ctx->mutex);
1939
1940 event->ctx = ctx;
1941 if (event->cpu != -1)
1942 event->cpu = cpu;
1943
1944 if (!task) {
1945 /*
1946 * Per cpu events are installed via an smp call and
1947 * the install is always successful.
1948 */
1949 cpu_function_call(cpu, __perf_install_in_context, event);
1950 return;
1951 }
1952
1953retry:
1954 if (!task_function_call(task, __perf_install_in_context, event))
1955 return;
1956
1957 raw_spin_lock_irq(&ctx->lock);
1958 /*
1959 * If we failed to find a running task, but find the context active now
1960 * that we've acquired the ctx->lock, retry.
1961 */
1962 if (ctx->is_active) {
1963 raw_spin_unlock_irq(&ctx->lock);
1964 goto retry;
1965 }
1966
1967 /*
1968 * Since the task isn't running, its safe to add the event, us holding
1969 * the ctx->lock ensures the task won't get scheduled in.
1970 */
1971 add_event_to_ctx(event, ctx);
1972 raw_spin_unlock_irq(&ctx->lock);
1973}
1974
1975/*
1976 * Put a event into inactive state and update time fields.
1977 * Enabling the leader of a group effectively enables all
1978 * the group members that aren't explicitly disabled, so we
1979 * have to update their ->tstamp_enabled also.
1980 * Note: this works for group members as well as group leaders
1981 * since the non-leader members' sibling_lists will be empty.
1982 */
1983static void __perf_event_mark_enabled(struct perf_event *event)
1984{
1985 struct perf_event *sub;
1986 u64 tstamp = perf_event_time(event);
1987
1988 event->state = PERF_EVENT_STATE_INACTIVE;
1989 event->tstamp_enabled = tstamp - event->total_time_enabled;
1990 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1991 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1992 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1993 }
1994}
1995
1996/*
1997 * Cross CPU call to enable a performance event
1998 */
1999static int __perf_event_enable(void *info)
2000{
2001 struct perf_event *event = info;
2002 struct perf_event_context *ctx = event->ctx;
2003 struct perf_event *leader = event->group_leader;
2004 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2005 int err;
2006
2007 /*
2008 * There's a time window between 'ctx->is_active' check
2009 * in perf_event_enable function and this place having:
2010 * - IRQs on
2011 * - ctx->lock unlocked
2012 *
2013 * where the task could be killed and 'ctx' deactivated
2014 * by perf_event_exit_task.
2015 */
2016 if (!ctx->is_active)
2017 return -EINVAL;
2018
2019 raw_spin_lock(&ctx->lock);
2020 update_context_time(ctx);
2021
2022 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2023 goto unlock;
2024
2025 /*
2026 * set current task's cgroup time reference point
2027 */
2028 perf_cgroup_set_timestamp(current, ctx);
2029
2030 __perf_event_mark_enabled(event);
2031
2032 if (!event_filter_match(event)) {
2033 if (is_cgroup_event(event))
2034 perf_cgroup_defer_enabled(event);
2035 goto unlock;
2036 }
2037
2038 /*
2039 * If the event is in a group and isn't the group leader,
2040 * then don't put it on unless the group is on.
2041 */
2042 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2043 goto unlock;
2044
2045 if (!group_can_go_on(event, cpuctx, 1)) {
2046 err = -EEXIST;
2047 } else {
2048 if (event == leader)
2049 err = group_sched_in(event, cpuctx, ctx);
2050 else
2051 err = event_sched_in(event, cpuctx, ctx);
2052 }
2053
2054 if (err) {
2055 /*
2056 * If this event can't go on and it's part of a
2057 * group, then the whole group has to come off.
2058 */
2059 if (leader != event) {
2060 group_sched_out(leader, cpuctx, ctx);
2061 perf_cpu_hrtimer_restart(cpuctx);
2062 }
2063 if (leader->attr.pinned) {
2064 update_group_times(leader);
2065 leader->state = PERF_EVENT_STATE_ERROR;
2066 }
2067 }
2068
2069unlock:
2070 raw_spin_unlock(&ctx->lock);
2071
2072 return 0;
2073}
2074
2075/*
2076 * Enable a event.
2077 *
2078 * If event->ctx is a cloned context, callers must make sure that
2079 * every task struct that event->ctx->task could possibly point to
2080 * remains valid. This condition is satisfied when called through
2081 * perf_event_for_each_child or perf_event_for_each as described
2082 * for perf_event_disable.
2083 */
2084void perf_event_enable(struct perf_event *event)
2085{
2086 struct perf_event_context *ctx = event->ctx;
2087 struct task_struct *task = ctx->task;
2088
2089 if (!task) {
2090 /*
2091 * Enable the event on the cpu that it's on
2092 */
2093 cpu_function_call(event->cpu, __perf_event_enable, event);
2094 return;
2095 }
2096
2097 raw_spin_lock_irq(&ctx->lock);
2098 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2099 goto out;
2100
2101 /*
2102 * If the event is in error state, clear that first.
2103 * That way, if we see the event in error state below, we
2104 * know that it has gone back into error state, as distinct
2105 * from the task having been scheduled away before the
2106 * cross-call arrived.
2107 */
2108 if (event->state == PERF_EVENT_STATE_ERROR)
2109 event->state = PERF_EVENT_STATE_OFF;
2110
2111retry:
2112 if (!ctx->is_active) {
2113 __perf_event_mark_enabled(event);
2114 goto out;
2115 }
2116
2117 raw_spin_unlock_irq(&ctx->lock);
2118
2119 if (!task_function_call(task, __perf_event_enable, event))
2120 return;
2121
2122 raw_spin_lock_irq(&ctx->lock);
2123
2124 /*
2125 * If the context is active and the event is still off,
2126 * we need to retry the cross-call.
2127 */
2128 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2129 /*
2130 * task could have been flipped by a concurrent
2131 * perf_event_context_sched_out()
2132 */
2133 task = ctx->task;
2134 goto retry;
2135 }
2136
2137out:
2138 raw_spin_unlock_irq(&ctx->lock);
2139}
2140EXPORT_SYMBOL_GPL(perf_event_enable);
2141
2142int perf_event_refresh(struct perf_event *event, int refresh)
2143{
2144 /*
2145 * not supported on inherited events
2146 */
2147 if (event->attr.inherit || !is_sampling_event(event))
2148 return -EINVAL;
2149
2150 atomic_add(refresh, &event->event_limit);
2151 perf_event_enable(event);
2152
2153 return 0;
2154}
2155EXPORT_SYMBOL_GPL(perf_event_refresh);
2156
2157static void ctx_sched_out(struct perf_event_context *ctx,
2158 struct perf_cpu_context *cpuctx,
2159 enum event_type_t event_type)
2160{
2161 struct perf_event *event;
2162 int is_active = ctx->is_active;
2163
2164 ctx->is_active &= ~event_type;
2165 if (likely(!ctx->nr_events))
2166 return;
2167
2168 update_context_time(ctx);
2169 update_cgrp_time_from_cpuctx(cpuctx);
2170 if (!ctx->nr_active)
2171 return;
2172
2173 perf_pmu_disable(ctx->pmu);
2174 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2175 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2176 group_sched_out(event, cpuctx, ctx);
2177 }
2178
2179 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2180 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2181 group_sched_out(event, cpuctx, ctx);
2182 }
2183 perf_pmu_enable(ctx->pmu);
2184}
2185
2186/*
2187 * Test whether two contexts are equivalent, i.e. whether they have both been
2188 * cloned from the same version of the same context.
2189 *
2190 * Equivalence is measured using a generation number in the context that is
2191 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2192 * and list_del_event().
2193 */
2194static int context_equiv(struct perf_event_context *ctx1,
2195 struct perf_event_context *ctx2)
2196{
2197 /* Pinning disables the swap optimization */
2198 if (ctx1->pin_count || ctx2->pin_count)
2199 return 0;
2200
2201 /* If ctx1 is the parent of ctx2 */
2202 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2203 return 1;
2204
2205 /* If ctx2 is the parent of ctx1 */
2206 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2207 return 1;
2208
2209 /*
2210 * If ctx1 and ctx2 have the same parent; we flatten the parent
2211 * hierarchy, see perf_event_init_context().
2212 */
2213 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2214 ctx1->parent_gen == ctx2->parent_gen)
2215 return 1;
2216
2217 /* Unmatched */
2218 return 0;
2219}
2220
2221static void __perf_event_sync_stat(struct perf_event *event,
2222 struct perf_event *next_event)
2223{
2224 u64 value;
2225
2226 if (!event->attr.inherit_stat)
2227 return;
2228
2229 /*
2230 * Update the event value, we cannot use perf_event_read()
2231 * because we're in the middle of a context switch and have IRQs
2232 * disabled, which upsets smp_call_function_single(), however
2233 * we know the event must be on the current CPU, therefore we
2234 * don't need to use it.
2235 */
2236 switch (event->state) {
2237 case PERF_EVENT_STATE_ACTIVE:
2238 event->pmu->read(event);
2239 /* fall-through */
2240
2241 case PERF_EVENT_STATE_INACTIVE:
2242 update_event_times(event);
2243 break;
2244
2245 default:
2246 break;
2247 }
2248
2249 /*
2250 * In order to keep per-task stats reliable we need to flip the event
2251 * values when we flip the contexts.
2252 */
2253 value = local64_read(&next_event->count);
2254 value = local64_xchg(&event->count, value);
2255 local64_set(&next_event->count, value);
2256
2257 swap(event->total_time_enabled, next_event->total_time_enabled);
2258 swap(event->total_time_running, next_event->total_time_running);
2259
2260 /*
2261 * Since we swizzled the values, update the user visible data too.
2262 */
2263 perf_event_update_userpage(event);
2264 perf_event_update_userpage(next_event);
2265}
2266
2267static void perf_event_sync_stat(struct perf_event_context *ctx,
2268 struct perf_event_context *next_ctx)
2269{
2270 struct perf_event *event, *next_event;
2271
2272 if (!ctx->nr_stat)
2273 return;
2274
2275 update_context_time(ctx);
2276
2277 event = list_first_entry(&ctx->event_list,
2278 struct perf_event, event_entry);
2279
2280 next_event = list_first_entry(&next_ctx->event_list,
2281 struct perf_event, event_entry);
2282
2283 while (&event->event_entry != &ctx->event_list &&
2284 &next_event->event_entry != &next_ctx->event_list) {
2285
2286 __perf_event_sync_stat(event, next_event);
2287
2288 event = list_next_entry(event, event_entry);
2289 next_event = list_next_entry(next_event, event_entry);
2290 }
2291}
2292
2293static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2294 struct task_struct *next)
2295{
2296 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2297 struct perf_event_context *next_ctx;
2298 struct perf_event_context *parent, *next_parent;
2299 struct perf_cpu_context *cpuctx;
2300 int do_switch = 1;
2301
2302 if (likely(!ctx))
2303 return;
2304
2305 cpuctx = __get_cpu_context(ctx);
2306 if (!cpuctx->task_ctx)
2307 return;
2308
2309 rcu_read_lock();
2310 next_ctx = next->perf_event_ctxp[ctxn];
2311 if (!next_ctx)
2312 goto unlock;
2313
2314 parent = rcu_dereference(ctx->parent_ctx);
2315 next_parent = rcu_dereference(next_ctx->parent_ctx);
2316
2317 /* If neither context have a parent context; they cannot be clones. */
2318 if (!parent && !next_parent)
2319 goto unlock;
2320
2321 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2322 /*
2323 * Looks like the two contexts are clones, so we might be
2324 * able to optimize the context switch. We lock both
2325 * contexts and check that they are clones under the
2326 * lock (including re-checking that neither has been
2327 * uncloned in the meantime). It doesn't matter which
2328 * order we take the locks because no other cpu could
2329 * be trying to lock both of these tasks.
2330 */
2331 raw_spin_lock(&ctx->lock);
2332 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2333 if (context_equiv(ctx, next_ctx)) {
2334 /*
2335 * XXX do we need a memory barrier of sorts
2336 * wrt to rcu_dereference() of perf_event_ctxp
2337 */
2338 task->perf_event_ctxp[ctxn] = next_ctx;
2339 next->perf_event_ctxp[ctxn] = ctx;
2340 ctx->task = next;
2341 next_ctx->task = task;
2342 do_switch = 0;
2343
2344 perf_event_sync_stat(ctx, next_ctx);
2345 }
2346 raw_spin_unlock(&next_ctx->lock);
2347 raw_spin_unlock(&ctx->lock);
2348 }
2349unlock:
2350 rcu_read_unlock();
2351
2352 if (do_switch) {
2353 raw_spin_lock(&ctx->lock);
2354 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2355 cpuctx->task_ctx = NULL;
2356 raw_spin_unlock(&ctx->lock);
2357 }
2358}
2359
2360#define for_each_task_context_nr(ctxn) \
2361 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2362
2363/*
2364 * Called from scheduler to remove the events of the current task,
2365 * with interrupts disabled.
2366 *
2367 * We stop each event and update the event value in event->count.
2368 *
2369 * This does not protect us against NMI, but disable()
2370 * sets the disabled bit in the control field of event _before_
2371 * accessing the event control register. If a NMI hits, then it will
2372 * not restart the event.
2373 */
2374void __perf_event_task_sched_out(struct task_struct *task,
2375 struct task_struct *next)
2376{
2377 int ctxn;
2378
2379 for_each_task_context_nr(ctxn)
2380 perf_event_context_sched_out(task, ctxn, next);
2381
2382 /*
2383 * if cgroup events exist on this CPU, then we need
2384 * to check if we have to switch out PMU state.
2385 * cgroup event are system-wide mode only
2386 */
2387 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2388 perf_cgroup_sched_out(task, next);
2389}
2390
2391static void task_ctx_sched_out(struct perf_event_context *ctx)
2392{
2393 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2394
2395 if (!cpuctx->task_ctx)
2396 return;
2397
2398 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2399 return;
2400
2401 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2402 cpuctx->task_ctx = NULL;
2403}
2404
2405/*
2406 * Called with IRQs disabled
2407 */
2408static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2409 enum event_type_t event_type)
2410{
2411 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2412}
2413
2414static void
2415ctx_pinned_sched_in(struct perf_event_context *ctx,
2416 struct perf_cpu_context *cpuctx)
2417{
2418 struct perf_event *event;
2419
2420 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2421 if (event->state <= PERF_EVENT_STATE_OFF)
2422 continue;
2423 if (!event_filter_match(event))
2424 continue;
2425
2426 /* may need to reset tstamp_enabled */
2427 if (is_cgroup_event(event))
2428 perf_cgroup_mark_enabled(event, ctx);
2429
2430 if (group_can_go_on(event, cpuctx, 1))
2431 group_sched_in(event, cpuctx, ctx);
2432
2433 /*
2434 * If this pinned group hasn't been scheduled,
2435 * put it in error state.
2436 */
2437 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2438 update_group_times(event);
2439 event->state = PERF_EVENT_STATE_ERROR;
2440 }
2441 }
2442}
2443
2444static void
2445ctx_flexible_sched_in(struct perf_event_context *ctx,
2446 struct perf_cpu_context *cpuctx)
2447{
2448 struct perf_event *event;
2449 int can_add_hw = 1;
2450
2451 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2452 /* Ignore events in OFF or ERROR state */
2453 if (event->state <= PERF_EVENT_STATE_OFF)
2454 continue;
2455 /*
2456 * Listen to the 'cpu' scheduling filter constraint
2457 * of events:
2458 */
2459 if (!event_filter_match(event))
2460 continue;
2461
2462 /* may need to reset tstamp_enabled */
2463 if (is_cgroup_event(event))
2464 perf_cgroup_mark_enabled(event, ctx);
2465
2466 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2467 if (group_sched_in(event, cpuctx, ctx))
2468 can_add_hw = 0;
2469 }
2470 }
2471}
2472
2473static void
2474ctx_sched_in(struct perf_event_context *ctx,
2475 struct perf_cpu_context *cpuctx,
2476 enum event_type_t event_type,
2477 struct task_struct *task)
2478{
2479 u64 now;
2480 int is_active = ctx->is_active;
2481
2482 ctx->is_active |= event_type;
2483 if (likely(!ctx->nr_events))
2484 return;
2485
2486 now = perf_clock();
2487 ctx->timestamp = now;
2488 perf_cgroup_set_timestamp(task, ctx);
2489 /*
2490 * First go through the list and put on any pinned groups
2491 * in order to give them the best chance of going on.
2492 */
2493 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2494 ctx_pinned_sched_in(ctx, cpuctx);
2495
2496 /* Then walk through the lower prio flexible groups */
2497 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2498 ctx_flexible_sched_in(ctx, cpuctx);
2499}
2500
2501static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2502 enum event_type_t event_type,
2503 struct task_struct *task)
2504{
2505 struct perf_event_context *ctx = &cpuctx->ctx;
2506
2507 ctx_sched_in(ctx, cpuctx, event_type, task);
2508}
2509
2510static void perf_event_context_sched_in(struct perf_event_context *ctx,
2511 struct task_struct *task)
2512{
2513 struct perf_cpu_context *cpuctx;
2514
2515 cpuctx = __get_cpu_context(ctx);
2516 if (cpuctx->task_ctx == ctx)
2517 return;
2518
2519 perf_ctx_lock(cpuctx, ctx);
2520 perf_pmu_disable(ctx->pmu);
2521 /*
2522 * We want to keep the following priority order:
2523 * cpu pinned (that don't need to move), task pinned,
2524 * cpu flexible, task flexible.
2525 */
2526 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2527
2528 if (ctx->nr_events)
2529 cpuctx->task_ctx = ctx;
2530
2531 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2532
2533 perf_pmu_enable(ctx->pmu);
2534 perf_ctx_unlock(cpuctx, ctx);
2535
2536 /*
2537 * Since these rotations are per-cpu, we need to ensure the
2538 * cpu-context we got scheduled on is actually rotating.
2539 */
2540 perf_pmu_rotate_start(ctx->pmu);
2541}
2542
2543/*
2544 * When sampling the branck stack in system-wide, it may be necessary
2545 * to flush the stack on context switch. This happens when the branch
2546 * stack does not tag its entries with the pid of the current task.
2547 * Otherwise it becomes impossible to associate a branch entry with a
2548 * task. This ambiguity is more likely to appear when the branch stack
2549 * supports priv level filtering and the user sets it to monitor only
2550 * at the user level (which could be a useful measurement in system-wide
2551 * mode). In that case, the risk is high of having a branch stack with
2552 * branch from multiple tasks. Flushing may mean dropping the existing
2553 * entries or stashing them somewhere in the PMU specific code layer.
2554 *
2555 * This function provides the context switch callback to the lower code
2556 * layer. It is invoked ONLY when there is at least one system-wide context
2557 * with at least one active event using taken branch sampling.
2558 */
2559static void perf_branch_stack_sched_in(struct task_struct *prev,
2560 struct task_struct *task)
2561{
2562 struct perf_cpu_context *cpuctx;
2563 struct pmu *pmu;
2564 unsigned long flags;
2565
2566 /* no need to flush branch stack if not changing task */
2567 if (prev == task)
2568 return;
2569
2570 local_irq_save(flags);
2571
2572 rcu_read_lock();
2573
2574 list_for_each_entry_rcu(pmu, &pmus, entry) {
2575 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2576
2577 /*
2578 * check if the context has at least one
2579 * event using PERF_SAMPLE_BRANCH_STACK
2580 */
2581 if (cpuctx->ctx.nr_branch_stack > 0
2582 && pmu->flush_branch_stack) {
2583
2584 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2585
2586 perf_pmu_disable(pmu);
2587
2588 pmu->flush_branch_stack();
2589
2590 perf_pmu_enable(pmu);
2591
2592 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2593 }
2594 }
2595
2596 rcu_read_unlock();
2597
2598 local_irq_restore(flags);
2599}
2600
2601/*
2602 * Called from scheduler to add the events of the current task
2603 * with interrupts disabled.
2604 *
2605 * We restore the event value and then enable it.
2606 *
2607 * This does not protect us against NMI, but enable()
2608 * sets the enabled bit in the control field of event _before_
2609 * accessing the event control register. If a NMI hits, then it will
2610 * keep the event running.
2611 */
2612void __perf_event_task_sched_in(struct task_struct *prev,
2613 struct task_struct *task)
2614{
2615 struct perf_event_context *ctx;
2616 int ctxn;
2617
2618 for_each_task_context_nr(ctxn) {
2619 ctx = task->perf_event_ctxp[ctxn];
2620 if (likely(!ctx))
2621 continue;
2622
2623 perf_event_context_sched_in(ctx, task);
2624 }
2625 /*
2626 * if cgroup events exist on this CPU, then we need
2627 * to check if we have to switch in PMU state.
2628 * cgroup event are system-wide mode only
2629 */
2630 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2631 perf_cgroup_sched_in(prev, task);
2632
2633 /* check for system-wide branch_stack events */
2634 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2635 perf_branch_stack_sched_in(prev, task);
2636}
2637
2638static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2639{
2640 u64 frequency = event->attr.sample_freq;
2641 u64 sec = NSEC_PER_SEC;
2642 u64 divisor, dividend;
2643
2644 int count_fls, nsec_fls, frequency_fls, sec_fls;
2645
2646 count_fls = fls64(count);
2647 nsec_fls = fls64(nsec);
2648 frequency_fls = fls64(frequency);
2649 sec_fls = 30;
2650
2651 /*
2652 * We got @count in @nsec, with a target of sample_freq HZ
2653 * the target period becomes:
2654 *
2655 * @count * 10^9
2656 * period = -------------------
2657 * @nsec * sample_freq
2658 *
2659 */
2660
2661 /*
2662 * Reduce accuracy by one bit such that @a and @b converge
2663 * to a similar magnitude.
2664 */
2665#define REDUCE_FLS(a, b) \
2666do { \
2667 if (a##_fls > b##_fls) { \
2668 a >>= 1; \
2669 a##_fls--; \
2670 } else { \
2671 b >>= 1; \
2672 b##_fls--; \
2673 } \
2674} while (0)
2675
2676 /*
2677 * Reduce accuracy until either term fits in a u64, then proceed with
2678 * the other, so that finally we can do a u64/u64 division.
2679 */
2680 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2681 REDUCE_FLS(nsec, frequency);
2682 REDUCE_FLS(sec, count);
2683 }
2684
2685 if (count_fls + sec_fls > 64) {
2686 divisor = nsec * frequency;
2687
2688 while (count_fls + sec_fls > 64) {
2689 REDUCE_FLS(count, sec);
2690 divisor >>= 1;
2691 }
2692
2693 dividend = count * sec;
2694 } else {
2695 dividend = count * sec;
2696
2697 while (nsec_fls + frequency_fls > 64) {
2698 REDUCE_FLS(nsec, frequency);
2699 dividend >>= 1;
2700 }
2701
2702 divisor = nsec * frequency;
2703 }
2704
2705 if (!divisor)
2706 return dividend;
2707
2708 return div64_u64(dividend, divisor);
2709}
2710
2711static DEFINE_PER_CPU(int, perf_throttled_count);
2712static DEFINE_PER_CPU(u64, perf_throttled_seq);
2713
2714static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2715{
2716 struct hw_perf_event *hwc = &event->hw;
2717 s64 period, sample_period;
2718 s64 delta;
2719
2720 period = perf_calculate_period(event, nsec, count);
2721
2722 delta = (s64)(period - hwc->sample_period);
2723 delta = (delta + 7) / 8; /* low pass filter */
2724
2725 sample_period = hwc->sample_period + delta;
2726
2727 if (!sample_period)
2728 sample_period = 1;
2729
2730 hwc->sample_period = sample_period;
2731
2732 if (local64_read(&hwc->period_left) > 8*sample_period) {
2733 if (disable)
2734 event->pmu->stop(event, PERF_EF_UPDATE);
2735
2736 local64_set(&hwc->period_left, 0);
2737
2738 if (disable)
2739 event->pmu->start(event, PERF_EF_RELOAD);
2740 }
2741}
2742
2743/*
2744 * combine freq adjustment with unthrottling to avoid two passes over the
2745 * events. At the same time, make sure, having freq events does not change
2746 * the rate of unthrottling as that would introduce bias.
2747 */
2748static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2749 int needs_unthr)
2750{
2751 struct perf_event *event;
2752 struct hw_perf_event *hwc;
2753 u64 now, period = TICK_NSEC;
2754 s64 delta;
2755
2756 /*
2757 * only need to iterate over all events iff:
2758 * - context have events in frequency mode (needs freq adjust)
2759 * - there are events to unthrottle on this cpu
2760 */
2761 if (!(ctx->nr_freq || needs_unthr))
2762 return;
2763
2764 raw_spin_lock(&ctx->lock);
2765 perf_pmu_disable(ctx->pmu);
2766
2767 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2768 if (event->state != PERF_EVENT_STATE_ACTIVE)
2769 continue;
2770
2771 if (!event_filter_match(event))
2772 continue;
2773
2774 perf_pmu_disable(event->pmu);
2775
2776 hwc = &event->hw;
2777
2778 if (hwc->interrupts == MAX_INTERRUPTS) {
2779 hwc->interrupts = 0;
2780 perf_log_throttle(event, 1);
2781 event->pmu->start(event, 0);
2782 }
2783
2784 if (!event->attr.freq || !event->attr.sample_freq)
2785 goto next;
2786
2787 /*
2788 * stop the event and update event->count
2789 */
2790 event->pmu->stop(event, PERF_EF_UPDATE);
2791
2792 now = local64_read(&event->count);
2793 delta = now - hwc->freq_count_stamp;
2794 hwc->freq_count_stamp = now;
2795
2796 /*
2797 * restart the event
2798 * reload only if value has changed
2799 * we have stopped the event so tell that
2800 * to perf_adjust_period() to avoid stopping it
2801 * twice.
2802 */
2803 if (delta > 0)
2804 perf_adjust_period(event, period, delta, false);
2805
2806 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2807 next:
2808 perf_pmu_enable(event->pmu);
2809 }
2810
2811 perf_pmu_enable(ctx->pmu);
2812 raw_spin_unlock(&ctx->lock);
2813}
2814
2815/*
2816 * Round-robin a context's events:
2817 */
2818static void rotate_ctx(struct perf_event_context *ctx)
2819{
2820 /*
2821 * Rotate the first entry last of non-pinned groups. Rotation might be
2822 * disabled by the inheritance code.
2823 */
2824 if (!ctx->rotate_disable)
2825 list_rotate_left(&ctx->flexible_groups);
2826}
2827
2828/*
2829 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2830 * because they're strictly cpu affine and rotate_start is called with IRQs
2831 * disabled, while rotate_context is called from IRQ context.
2832 */
2833static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2834{
2835 struct perf_event_context *ctx = NULL;
2836 int rotate = 0, remove = 1;
2837
2838 if (cpuctx->ctx.nr_events) {
2839 remove = 0;
2840 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2841 rotate = 1;
2842 }
2843
2844 ctx = cpuctx->task_ctx;
2845 if (ctx && ctx->nr_events) {
2846 remove = 0;
2847 if (ctx->nr_events != ctx->nr_active)
2848 rotate = 1;
2849 }
2850
2851 if (!rotate)
2852 goto done;
2853
2854 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2855 perf_pmu_disable(cpuctx->ctx.pmu);
2856
2857 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2858 if (ctx)
2859 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2860
2861 rotate_ctx(&cpuctx->ctx);
2862 if (ctx)
2863 rotate_ctx(ctx);
2864
2865 perf_event_sched_in(cpuctx, ctx, current);
2866
2867 perf_pmu_enable(cpuctx->ctx.pmu);
2868 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2869done:
2870 if (remove)
2871 list_del_init(&cpuctx->rotation_list);
2872
2873 return rotate;
2874}
2875
2876#ifdef CONFIG_NO_HZ_FULL
2877bool perf_event_can_stop_tick(void)
2878{
2879 if (atomic_read(&nr_freq_events) ||
2880 __this_cpu_read(perf_throttled_count))
2881 return false;
2882 else
2883 return true;
2884}
2885#endif
2886
2887void perf_event_task_tick(void)
2888{
2889 struct list_head *head = &__get_cpu_var(rotation_list);
2890 struct perf_cpu_context *cpuctx, *tmp;
2891 struct perf_event_context *ctx;
2892 int throttled;
2893
2894 WARN_ON(!irqs_disabled());
2895
2896 __this_cpu_inc(perf_throttled_seq);
2897 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2898
2899 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2900 ctx = &cpuctx->ctx;
2901 perf_adjust_freq_unthr_context(ctx, throttled);
2902
2903 ctx = cpuctx->task_ctx;
2904 if (ctx)
2905 perf_adjust_freq_unthr_context(ctx, throttled);
2906 }
2907}
2908
2909static int event_enable_on_exec(struct perf_event *event,
2910 struct perf_event_context *ctx)
2911{
2912 if (!event->attr.enable_on_exec)
2913 return 0;
2914
2915 event->attr.enable_on_exec = 0;
2916 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2917 return 0;
2918
2919 __perf_event_mark_enabled(event);
2920
2921 return 1;
2922}
2923
2924/*
2925 * Enable all of a task's events that have been marked enable-on-exec.
2926 * This expects task == current.
2927 */
2928static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2929{
2930 struct perf_event *event;
2931 unsigned long flags;
2932 int enabled = 0;
2933 int ret;
2934
2935 local_irq_save(flags);
2936 if (!ctx || !ctx->nr_events)
2937 goto out;
2938
2939 /*
2940 * We must ctxsw out cgroup events to avoid conflict
2941 * when invoking perf_task_event_sched_in() later on
2942 * in this function. Otherwise we end up trying to
2943 * ctxswin cgroup events which are already scheduled
2944 * in.
2945 */
2946 perf_cgroup_sched_out(current, NULL);
2947
2948 raw_spin_lock(&ctx->lock);
2949 task_ctx_sched_out(ctx);
2950
2951 list_for_each_entry(event, &ctx->event_list, event_entry) {
2952 ret = event_enable_on_exec(event, ctx);
2953 if (ret)
2954 enabled = 1;
2955 }
2956
2957 /*
2958 * Unclone this context if we enabled any event.
2959 */
2960 if (enabled)
2961 unclone_ctx(ctx);
2962
2963 raw_spin_unlock(&ctx->lock);
2964
2965 /*
2966 * Also calls ctxswin for cgroup events, if any:
2967 */
2968 perf_event_context_sched_in(ctx, ctx->task);
2969out:
2970 local_irq_restore(flags);
2971}
2972
2973/*
2974 * Cross CPU call to read the hardware event
2975 */
2976static void __perf_event_read(void *info)
2977{
2978 struct perf_event *event = info;
2979 struct perf_event_context *ctx = event->ctx;
2980 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2981
2982 /*
2983 * If this is a task context, we need to check whether it is
2984 * the current task context of this cpu. If not it has been
2985 * scheduled out before the smp call arrived. In that case
2986 * event->count would have been updated to a recent sample
2987 * when the event was scheduled out.
2988 */
2989 if (ctx->task && cpuctx->task_ctx != ctx)
2990 return;
2991
2992 raw_spin_lock(&ctx->lock);
2993 if (ctx->is_active) {
2994 update_context_time(ctx);
2995 update_cgrp_time_from_event(event);
2996 }
2997 update_event_times(event);
2998 if (event->state == PERF_EVENT_STATE_ACTIVE)
2999 event->pmu->read(event);
3000 raw_spin_unlock(&ctx->lock);
3001}
3002
3003static inline u64 perf_event_count(struct perf_event *event)
3004{
3005 return local64_read(&event->count) + atomic64_read(&event->child_count);
3006}
3007
3008static u64 perf_event_read(struct perf_event *event)
3009{
3010 /*
3011 * If event is enabled and currently active on a CPU, update the
3012 * value in the event structure:
3013 */
3014 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3015 smp_call_function_single(event->oncpu,
3016 __perf_event_read, event, 1);
3017 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3018 struct perf_event_context *ctx = event->ctx;
3019 unsigned long flags;
3020
3021 raw_spin_lock_irqsave(&ctx->lock, flags);
3022 /*
3023 * may read while context is not active
3024 * (e.g., thread is blocked), in that case
3025 * we cannot update context time
3026 */
3027 if (ctx->is_active) {
3028 update_context_time(ctx);
3029 update_cgrp_time_from_event(event);
3030 }
3031 update_event_times(event);
3032 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3033 }
3034
3035 return perf_event_count(event);
3036}
3037
3038/*
3039 * Initialize the perf_event context in a task_struct:
3040 */
3041static void __perf_event_init_context(struct perf_event_context *ctx)
3042{
3043 raw_spin_lock_init(&ctx->lock);
3044 mutex_init(&ctx->mutex);
3045 INIT_LIST_HEAD(&ctx->pinned_groups);
3046 INIT_LIST_HEAD(&ctx->flexible_groups);
3047 INIT_LIST_HEAD(&ctx->event_list);
3048 atomic_set(&ctx->refcount, 1);
3049}
3050
3051static struct perf_event_context *
3052alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3053{
3054 struct perf_event_context *ctx;
3055
3056 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3057 if (!ctx)
3058 return NULL;
3059
3060 __perf_event_init_context(ctx);
3061 if (task) {
3062 ctx->task = task;
3063 get_task_struct(task);
3064 }
3065 ctx->pmu = pmu;
3066
3067 return ctx;
3068}
3069
3070static struct task_struct *
3071find_lively_task_by_vpid(pid_t vpid)
3072{
3073 struct task_struct *task;
3074 int err;
3075
3076 rcu_read_lock();
3077 if (!vpid)
3078 task = current;
3079 else
3080 task = find_task_by_vpid(vpid);
3081 if (task)
3082 get_task_struct(task);
3083 rcu_read_unlock();
3084
3085 if (!task)
3086 return ERR_PTR(-ESRCH);
3087
3088 /* Reuse ptrace permission checks for now. */
3089 err = -EACCES;
3090 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3091 goto errout;
3092
3093 return task;
3094errout:
3095 put_task_struct(task);
3096 return ERR_PTR(err);
3097
3098}
3099
3100/*
3101 * Returns a matching context with refcount and pincount.
3102 */
3103static struct perf_event_context *
3104find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3105{
3106 struct perf_event_context *ctx;
3107 struct perf_cpu_context *cpuctx;
3108 unsigned long flags;
3109 int ctxn, err;
3110
3111 if (!task) {
3112 /* Must be root to operate on a CPU event: */
3113 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3114 return ERR_PTR(-EACCES);
3115
3116 /*
3117 * We could be clever and allow to attach a event to an
3118 * offline CPU and activate it when the CPU comes up, but
3119 * that's for later.
3120 */
3121 if (!cpu_online(cpu))
3122 return ERR_PTR(-ENODEV);
3123
3124 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3125 ctx = &cpuctx->ctx;
3126 get_ctx(ctx);
3127 ++ctx->pin_count;
3128
3129 return ctx;
3130 }
3131
3132 err = -EINVAL;
3133 ctxn = pmu->task_ctx_nr;
3134 if (ctxn < 0)
3135 goto errout;
3136
3137retry:
3138 ctx = perf_lock_task_context(task, ctxn, &flags);
3139 if (ctx) {
3140 unclone_ctx(ctx);
3141 ++ctx->pin_count;
3142 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3143 } else {
3144 ctx = alloc_perf_context(pmu, task);
3145 err = -ENOMEM;
3146 if (!ctx)
3147 goto errout;
3148
3149 err = 0;
3150 mutex_lock(&task->perf_event_mutex);
3151 /*
3152 * If it has already passed perf_event_exit_task().
3153 * we must see PF_EXITING, it takes this mutex too.
3154 */
3155 if (task->flags & PF_EXITING)
3156 err = -ESRCH;
3157 else if (task->perf_event_ctxp[ctxn])
3158 err = -EAGAIN;
3159 else {
3160 get_ctx(ctx);
3161 ++ctx->pin_count;
3162 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3163 }
3164 mutex_unlock(&task->perf_event_mutex);
3165
3166 if (unlikely(err)) {
3167 put_ctx(ctx);
3168
3169 if (err == -EAGAIN)
3170 goto retry;
3171 goto errout;
3172 }
3173 }
3174
3175 return ctx;
3176
3177errout:
3178 return ERR_PTR(err);
3179}
3180
3181static void perf_event_free_filter(struct perf_event *event);
3182
3183static void free_event_rcu(struct rcu_head *head)
3184{
3185 struct perf_event *event;
3186
3187 event = container_of(head, struct perf_event, rcu_head);
3188 if (event->ns)
3189 put_pid_ns(event->ns);
3190 perf_event_free_filter(event);
3191 kfree(event);
3192}
3193
3194static void ring_buffer_put(struct ring_buffer *rb);
3195static void ring_buffer_attach(struct perf_event *event,
3196 struct ring_buffer *rb);
3197
3198static void unaccount_event_cpu(struct perf_event *event, int cpu)
3199{
3200 if (event->parent)
3201 return;
3202
3203 if (has_branch_stack(event)) {
3204 if (!(event->attach_state & PERF_ATTACH_TASK))
3205 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3206 }
3207 if (is_cgroup_event(event))
3208 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3209}
3210
3211static void unaccount_event(struct perf_event *event)
3212{
3213 if (event->parent)
3214 return;
3215
3216 if (event->attach_state & PERF_ATTACH_TASK)
3217 static_key_slow_dec_deferred(&perf_sched_events);
3218 if (event->attr.mmap || event->attr.mmap_data)
3219 atomic_dec(&nr_mmap_events);
3220 if (event->attr.comm)
3221 atomic_dec(&nr_comm_events);
3222 if (event->attr.task)
3223 atomic_dec(&nr_task_events);
3224 if (event->attr.freq)
3225 atomic_dec(&nr_freq_events);
3226 if (is_cgroup_event(event))
3227 static_key_slow_dec_deferred(&perf_sched_events);
3228 if (has_branch_stack(event))
3229 static_key_slow_dec_deferred(&perf_sched_events);
3230
3231 unaccount_event_cpu(event, event->cpu);
3232}
3233
3234static void __free_event(struct perf_event *event)
3235{
3236 if (!event->parent) {
3237 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3238 put_callchain_buffers();
3239 }
3240
3241 if (event->destroy)
3242 event->destroy(event);
3243
3244 if (event->ctx)
3245 put_ctx(event->ctx);
3246
3247 call_rcu(&event->rcu_head, free_event_rcu);
3248}
3249static void free_event(struct perf_event *event)
3250{
3251 irq_work_sync(&event->pending);
3252
3253 unaccount_event(event);
3254
3255 if (event->rb) {
3256 /*
3257 * Can happen when we close an event with re-directed output.
3258 *
3259 * Since we have a 0 refcount, perf_mmap_close() will skip
3260 * over us; possibly making our ring_buffer_put() the last.
3261 */
3262 mutex_lock(&event->mmap_mutex);
3263 ring_buffer_attach(event, NULL);
3264 mutex_unlock(&event->mmap_mutex);
3265 }
3266
3267 if (is_cgroup_event(event))
3268 perf_detach_cgroup(event);
3269
3270
3271 __free_event(event);
3272}
3273
3274int perf_event_release_kernel(struct perf_event *event)
3275{
3276 struct perf_event_context *ctx = event->ctx;
3277
3278 WARN_ON_ONCE(ctx->parent_ctx);
3279 /*
3280 * There are two ways this annotation is useful:
3281 *
3282 * 1) there is a lock recursion from perf_event_exit_task
3283 * see the comment there.
3284 *
3285 * 2) there is a lock-inversion with mmap_sem through
3286 * perf_event_read_group(), which takes faults while
3287 * holding ctx->mutex, however this is called after
3288 * the last filedesc died, so there is no possibility
3289 * to trigger the AB-BA case.
3290 */
3291 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3292 perf_remove_from_context(event, true);
3293 mutex_unlock(&ctx->mutex);
3294
3295 free_event(event);
3296
3297 return 0;
3298}
3299EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3300
3301/*
3302 * Called when the last reference to the file is gone.
3303 */
3304static void put_event(struct perf_event *event)
3305{
3306 struct task_struct *owner;
3307
3308 if (!atomic_long_dec_and_test(&event->refcount))
3309 return;
3310
3311 rcu_read_lock();
3312 owner = ACCESS_ONCE(event->owner);
3313 /*
3314 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3315 * !owner it means the list deletion is complete and we can indeed
3316 * free this event, otherwise we need to serialize on
3317 * owner->perf_event_mutex.
3318 */
3319 smp_read_barrier_depends();
3320 if (owner) {
3321 /*
3322 * Since delayed_put_task_struct() also drops the last
3323 * task reference we can safely take a new reference
3324 * while holding the rcu_read_lock().
3325 */
3326 get_task_struct(owner);
3327 }
3328 rcu_read_unlock();
3329
3330 if (owner) {
3331 mutex_lock(&owner->perf_event_mutex);
3332 /*
3333 * We have to re-check the event->owner field, if it is cleared
3334 * we raced with perf_event_exit_task(), acquiring the mutex
3335 * ensured they're done, and we can proceed with freeing the
3336 * event.
3337 */
3338 if (event->owner)
3339 list_del_init(&event->owner_entry);
3340 mutex_unlock(&owner->perf_event_mutex);
3341 put_task_struct(owner);
3342 }
3343
3344 perf_event_release_kernel(event);
3345}
3346
3347static int perf_release(struct inode *inode, struct file *file)
3348{
3349 put_event(file->private_data);
3350 return 0;
3351}
3352
3353u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3354{
3355 struct perf_event *child;
3356 u64 total = 0;
3357
3358 *enabled = 0;
3359 *running = 0;
3360
3361 mutex_lock(&event->child_mutex);
3362 total += perf_event_read(event);
3363 *enabled += event->total_time_enabled +
3364 atomic64_read(&event->child_total_time_enabled);
3365 *running += event->total_time_running +
3366 atomic64_read(&event->child_total_time_running);
3367
3368 list_for_each_entry(child, &event->child_list, child_list) {
3369 total += perf_event_read(child);
3370 *enabled += child->total_time_enabled;
3371 *running += child->total_time_running;
3372 }
3373 mutex_unlock(&event->child_mutex);
3374
3375 return total;
3376}
3377EXPORT_SYMBOL_GPL(perf_event_read_value);
3378
3379static int perf_event_read_group(struct perf_event *event,
3380 u64 read_format, char __user *buf)
3381{
3382 struct perf_event *leader = event->group_leader, *sub;
3383 int n = 0, size = 0, ret = -EFAULT;
3384 struct perf_event_context *ctx = leader->ctx;
3385 u64 values[5];
3386 u64 count, enabled, running;
3387
3388 mutex_lock(&ctx->mutex);
3389 count = perf_event_read_value(leader, &enabled, &running);
3390
3391 values[n++] = 1 + leader->nr_siblings;
3392 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3393 values[n++] = enabled;
3394 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3395 values[n++] = running;
3396 values[n++] = count;
3397 if (read_format & PERF_FORMAT_ID)
3398 values[n++] = primary_event_id(leader);
3399
3400 size = n * sizeof(u64);
3401
3402 if (copy_to_user(buf, values, size))
3403 goto unlock;
3404
3405 ret = size;
3406
3407 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3408 n = 0;
3409
3410 values[n++] = perf_event_read_value(sub, &enabled, &running);
3411 if (read_format & PERF_FORMAT_ID)
3412 values[n++] = primary_event_id(sub);
3413
3414 size = n * sizeof(u64);
3415
3416 if (copy_to_user(buf + ret, values, size)) {
3417 ret = -EFAULT;
3418 goto unlock;
3419 }
3420
3421 ret += size;
3422 }
3423unlock:
3424 mutex_unlock(&ctx->mutex);
3425
3426 return ret;
3427}
3428
3429static int perf_event_read_one(struct perf_event *event,
3430 u64 read_format, char __user *buf)
3431{
3432 u64 enabled, running;
3433 u64 values[4];
3434 int n = 0;
3435
3436 values[n++] = perf_event_read_value(event, &enabled, &running);
3437 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3438 values[n++] = enabled;
3439 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3440 values[n++] = running;
3441 if (read_format & PERF_FORMAT_ID)
3442 values[n++] = primary_event_id(event);
3443
3444 if (copy_to_user(buf, values, n * sizeof(u64)))
3445 return -EFAULT;
3446
3447 return n * sizeof(u64);
3448}
3449
3450/*
3451 * Read the performance event - simple non blocking version for now
3452 */
3453static ssize_t
3454perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3455{
3456 u64 read_format = event->attr.read_format;
3457 int ret;
3458
3459 /*
3460 * Return end-of-file for a read on a event that is in
3461 * error state (i.e. because it was pinned but it couldn't be
3462 * scheduled on to the CPU at some point).
3463 */
3464 if (event->state == PERF_EVENT_STATE_ERROR)
3465 return 0;
3466
3467 if (count < event->read_size)
3468 return -ENOSPC;
3469
3470 WARN_ON_ONCE(event->ctx->parent_ctx);
3471 if (read_format & PERF_FORMAT_GROUP)
3472 ret = perf_event_read_group(event, read_format, buf);
3473 else
3474 ret = perf_event_read_one(event, read_format, buf);
3475
3476 return ret;
3477}
3478
3479static ssize_t
3480perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3481{
3482 struct perf_event *event = file->private_data;
3483
3484 return perf_read_hw(event, buf, count);
3485}
3486
3487static unsigned int perf_poll(struct file *file, poll_table *wait)
3488{
3489 struct perf_event *event = file->private_data;
3490 struct ring_buffer *rb;
3491 unsigned int events = POLL_HUP;
3492
3493 /*
3494 * Pin the event->rb by taking event->mmap_mutex; otherwise
3495 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3496 */
3497 mutex_lock(&event->mmap_mutex);
3498 rb = event->rb;
3499 if (rb)
3500 events = atomic_xchg(&rb->poll, 0);
3501 mutex_unlock(&event->mmap_mutex);
3502
3503 poll_wait(file, &event->waitq, wait);
3504
3505 return events;
3506}
3507
3508static void perf_event_reset(struct perf_event *event)
3509{
3510 (void)perf_event_read(event);
3511 local64_set(&event->count, 0);
3512 perf_event_update_userpage(event);
3513}
3514
3515/*
3516 * Holding the top-level event's child_mutex means that any
3517 * descendant process that has inherited this event will block
3518 * in sync_child_event if it goes to exit, thus satisfying the
3519 * task existence requirements of perf_event_enable/disable.
3520 */
3521static void perf_event_for_each_child(struct perf_event *event,
3522 void (*func)(struct perf_event *))
3523{
3524 struct perf_event *child;
3525
3526 WARN_ON_ONCE(event->ctx->parent_ctx);
3527 mutex_lock(&event->child_mutex);
3528 func(event);
3529 list_for_each_entry(child, &event->child_list, child_list)
3530 func(child);
3531 mutex_unlock(&event->child_mutex);
3532}
3533
3534static void perf_event_for_each(struct perf_event *event,
3535 void (*func)(struct perf_event *))
3536{
3537 struct perf_event_context *ctx = event->ctx;
3538 struct perf_event *sibling;
3539
3540 WARN_ON_ONCE(ctx->parent_ctx);
3541 mutex_lock(&ctx->mutex);
3542 event = event->group_leader;
3543
3544 perf_event_for_each_child(event, func);
3545 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3546 perf_event_for_each_child(sibling, func);
3547 mutex_unlock(&ctx->mutex);
3548}
3549
3550static int perf_event_period(struct perf_event *event, u64 __user *arg)
3551{
3552 struct perf_event_context *ctx = event->ctx;
3553 int ret = 0, active;
3554 u64 value;
3555
3556 if (!is_sampling_event(event))
3557 return -EINVAL;
3558
3559 if (copy_from_user(&value, arg, sizeof(value)))
3560 return -EFAULT;
3561
3562 if (!value)
3563 return -EINVAL;
3564
3565 raw_spin_lock_irq(&ctx->lock);
3566 if (event->attr.freq) {
3567 if (value > sysctl_perf_event_sample_rate) {
3568 ret = -EINVAL;
3569 goto unlock;
3570 }
3571
3572 event->attr.sample_freq = value;
3573 } else {
3574 event->attr.sample_period = value;
3575 event->hw.sample_period = value;
3576 }
3577
3578 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3579 if (active) {
3580 perf_pmu_disable(ctx->pmu);
3581 event->pmu->stop(event, PERF_EF_UPDATE);
3582 }
3583
3584 local64_set(&event->hw.period_left, 0);
3585
3586 if (active) {
3587 event->pmu->start(event, PERF_EF_RELOAD);
3588 perf_pmu_enable(ctx->pmu);
3589 }
3590
3591unlock:
3592 raw_spin_unlock_irq(&ctx->lock);
3593
3594 return ret;
3595}
3596
3597static const struct file_operations perf_fops;
3598
3599static inline int perf_fget_light(int fd, struct fd *p)
3600{
3601 struct fd f = fdget(fd);
3602 if (!f.file)
3603 return -EBADF;
3604
3605 if (f.file->f_op != &perf_fops) {
3606 fdput(f);
3607 return -EBADF;
3608 }
3609 *p = f;
3610 return 0;
3611}
3612
3613static int perf_event_set_output(struct perf_event *event,
3614 struct perf_event *output_event);
3615static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3616
3617static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3618{
3619 struct perf_event *event = file->private_data;
3620 void (*func)(struct perf_event *);
3621 u32 flags = arg;
3622
3623 switch (cmd) {
3624 case PERF_EVENT_IOC_ENABLE:
3625 func = perf_event_enable;
3626 break;
3627 case PERF_EVENT_IOC_DISABLE:
3628 func = perf_event_disable;
3629 break;
3630 case PERF_EVENT_IOC_RESET:
3631 func = perf_event_reset;
3632 break;
3633
3634 case PERF_EVENT_IOC_REFRESH:
3635 return perf_event_refresh(event, arg);
3636
3637 case PERF_EVENT_IOC_PERIOD:
3638 return perf_event_period(event, (u64 __user *)arg);
3639
3640 case PERF_EVENT_IOC_ID:
3641 {
3642 u64 id = primary_event_id(event);
3643
3644 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3645 return -EFAULT;
3646 return 0;
3647 }
3648
3649 case PERF_EVENT_IOC_SET_OUTPUT:
3650 {
3651 int ret;
3652 if (arg != -1) {
3653 struct perf_event *output_event;
3654 struct fd output;
3655 ret = perf_fget_light(arg, &output);
3656 if (ret)
3657 return ret;
3658 output_event = output.file->private_data;
3659 ret = perf_event_set_output(event, output_event);
3660 fdput(output);
3661 } else {
3662 ret = perf_event_set_output(event, NULL);
3663 }
3664 return ret;
3665 }
3666
3667 case PERF_EVENT_IOC_SET_FILTER:
3668 return perf_event_set_filter(event, (void __user *)arg);
3669
3670 default:
3671 return -ENOTTY;
3672 }
3673
3674 if (flags & PERF_IOC_FLAG_GROUP)
3675 perf_event_for_each(event, func);
3676 else
3677 perf_event_for_each_child(event, func);
3678
3679 return 0;
3680}
3681
3682int perf_event_task_enable(void)
3683{
3684 struct perf_event *event;
3685
3686 mutex_lock(¤t->perf_event_mutex);
3687 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3688 perf_event_for_each_child(event, perf_event_enable);
3689 mutex_unlock(¤t->perf_event_mutex);
3690
3691 return 0;
3692}
3693
3694int perf_event_task_disable(void)
3695{
3696 struct perf_event *event;
3697
3698 mutex_lock(¤t->perf_event_mutex);
3699 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3700 perf_event_for_each_child(event, perf_event_disable);
3701 mutex_unlock(¤t->perf_event_mutex);
3702
3703 return 0;
3704}
3705
3706static int perf_event_index(struct perf_event *event)
3707{
3708 if (event->hw.state & PERF_HES_STOPPED)
3709 return 0;
3710
3711 if (event->state != PERF_EVENT_STATE_ACTIVE)
3712 return 0;
3713
3714 return event->pmu->event_idx(event);
3715}
3716
3717static void calc_timer_values(struct perf_event *event,
3718 u64 *now,
3719 u64 *enabled,
3720 u64 *running)
3721{
3722 u64 ctx_time;
3723
3724 *now = perf_clock();
3725 ctx_time = event->shadow_ctx_time + *now;
3726 *enabled = ctx_time - event->tstamp_enabled;
3727 *running = ctx_time - event->tstamp_running;
3728}
3729
3730static void perf_event_init_userpage(struct perf_event *event)
3731{
3732 struct perf_event_mmap_page *userpg;
3733 struct ring_buffer *rb;
3734
3735 rcu_read_lock();
3736 rb = rcu_dereference(event->rb);
3737 if (!rb)
3738 goto unlock;
3739
3740 userpg = rb->user_page;
3741
3742 /* Allow new userspace to detect that bit 0 is deprecated */
3743 userpg->cap_bit0_is_deprecated = 1;
3744 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3745
3746unlock:
3747 rcu_read_unlock();
3748}
3749
3750void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3751{
3752}
3753
3754/*
3755 * Callers need to ensure there can be no nesting of this function, otherwise
3756 * the seqlock logic goes bad. We can not serialize this because the arch
3757 * code calls this from NMI context.
3758 */
3759void perf_event_update_userpage(struct perf_event *event)
3760{
3761 struct perf_event_mmap_page *userpg;
3762 struct ring_buffer *rb;
3763 u64 enabled, running, now;
3764
3765 rcu_read_lock();
3766 rb = rcu_dereference(event->rb);
3767 if (!rb)
3768 goto unlock;
3769
3770 /*
3771 * compute total_time_enabled, total_time_running
3772 * based on snapshot values taken when the event
3773 * was last scheduled in.
3774 *
3775 * we cannot simply called update_context_time()
3776 * because of locking issue as we can be called in
3777 * NMI context
3778 */
3779 calc_timer_values(event, &now, &enabled, &running);
3780
3781 userpg = rb->user_page;
3782 /*
3783 * Disable preemption so as to not let the corresponding user-space
3784 * spin too long if we get preempted.
3785 */
3786 preempt_disable();
3787 ++userpg->lock;
3788 barrier();
3789 userpg->index = perf_event_index(event);
3790 userpg->offset = perf_event_count(event);
3791 if (userpg->index)
3792 userpg->offset -= local64_read(&event->hw.prev_count);
3793
3794 userpg->time_enabled = enabled +
3795 atomic64_read(&event->child_total_time_enabled);
3796
3797 userpg->time_running = running +
3798 atomic64_read(&event->child_total_time_running);
3799
3800 arch_perf_update_userpage(userpg, now);
3801
3802 barrier();
3803 ++userpg->lock;
3804 preempt_enable();
3805unlock:
3806 rcu_read_unlock();
3807}
3808
3809static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3810{
3811 struct perf_event *event = vma->vm_file->private_data;
3812 struct ring_buffer *rb;
3813 int ret = VM_FAULT_SIGBUS;
3814
3815 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3816 if (vmf->pgoff == 0)
3817 ret = 0;
3818 return ret;
3819 }
3820
3821 rcu_read_lock();
3822 rb = rcu_dereference(event->rb);
3823 if (!rb)
3824 goto unlock;
3825
3826 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3827 goto unlock;
3828
3829 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3830 if (!vmf->page)
3831 goto unlock;
3832
3833 get_page(vmf->page);
3834 vmf->page->mapping = vma->vm_file->f_mapping;
3835 vmf->page->index = vmf->pgoff;
3836
3837 ret = 0;
3838unlock:
3839 rcu_read_unlock();
3840
3841 return ret;
3842}
3843
3844static void ring_buffer_attach(struct perf_event *event,
3845 struct ring_buffer *rb)
3846{
3847 struct ring_buffer *old_rb = NULL;
3848 unsigned long flags;
3849
3850 if (event->rb) {
3851 /*
3852 * Should be impossible, we set this when removing
3853 * event->rb_entry and wait/clear when adding event->rb_entry.
3854 */
3855 WARN_ON_ONCE(event->rcu_pending);
3856
3857 old_rb = event->rb;
3858 event->rcu_batches = get_state_synchronize_rcu();
3859 event->rcu_pending = 1;
3860
3861 spin_lock_irqsave(&old_rb->event_lock, flags);
3862 list_del_rcu(&event->rb_entry);
3863 spin_unlock_irqrestore(&old_rb->event_lock, flags);
3864 }
3865
3866 if (event->rcu_pending && rb) {
3867 cond_synchronize_rcu(event->rcu_batches);
3868 event->rcu_pending = 0;
3869 }
3870
3871 if (rb) {
3872 spin_lock_irqsave(&rb->event_lock, flags);
3873 list_add_rcu(&event->rb_entry, &rb->event_list);
3874 spin_unlock_irqrestore(&rb->event_lock, flags);
3875 }
3876
3877 rcu_assign_pointer(event->rb, rb);
3878
3879 if (old_rb) {
3880 ring_buffer_put(old_rb);
3881 /*
3882 * Since we detached before setting the new rb, so that we
3883 * could attach the new rb, we could have missed a wakeup.
3884 * Provide it now.
3885 */
3886 wake_up_all(&event->waitq);
3887 }
3888}
3889
3890static void ring_buffer_wakeup(struct perf_event *event)
3891{
3892 struct ring_buffer *rb;
3893
3894 rcu_read_lock();
3895 rb = rcu_dereference(event->rb);
3896 if (rb) {
3897 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3898 wake_up_all(&event->waitq);
3899 }
3900 rcu_read_unlock();
3901}
3902
3903static void rb_free_rcu(struct rcu_head *rcu_head)
3904{
3905 struct ring_buffer *rb;
3906
3907 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3908 rb_free(rb);
3909}
3910
3911static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3912{
3913 struct ring_buffer *rb;
3914
3915 rcu_read_lock();
3916 rb = rcu_dereference(event->rb);
3917 if (rb) {
3918 if (!atomic_inc_not_zero(&rb->refcount))
3919 rb = NULL;
3920 }
3921 rcu_read_unlock();
3922
3923 return rb;
3924}
3925
3926static void ring_buffer_put(struct ring_buffer *rb)
3927{
3928 if (!atomic_dec_and_test(&rb->refcount))
3929 return;
3930
3931 WARN_ON_ONCE(!list_empty(&rb->event_list));
3932
3933 call_rcu(&rb->rcu_head, rb_free_rcu);
3934}
3935
3936static void perf_mmap_open(struct vm_area_struct *vma)
3937{
3938 struct perf_event *event = vma->vm_file->private_data;
3939
3940 atomic_inc(&event->mmap_count);
3941 atomic_inc(&event->rb->mmap_count);
3942}
3943
3944/*
3945 * A buffer can be mmap()ed multiple times; either directly through the same
3946 * event, or through other events by use of perf_event_set_output().
3947 *
3948 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3949 * the buffer here, where we still have a VM context. This means we need
3950 * to detach all events redirecting to us.
3951 */
3952static void perf_mmap_close(struct vm_area_struct *vma)
3953{
3954 struct perf_event *event = vma->vm_file->private_data;
3955
3956 struct ring_buffer *rb = ring_buffer_get(event);
3957 struct user_struct *mmap_user = rb->mmap_user;
3958 int mmap_locked = rb->mmap_locked;
3959 unsigned long size = perf_data_size(rb);
3960
3961 atomic_dec(&rb->mmap_count);
3962
3963 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3964 goto out_put;
3965
3966 ring_buffer_attach(event, NULL);
3967 mutex_unlock(&event->mmap_mutex);
3968
3969 /* If there's still other mmap()s of this buffer, we're done. */
3970 if (atomic_read(&rb->mmap_count))
3971 goto out_put;
3972
3973 /*
3974 * No other mmap()s, detach from all other events that might redirect
3975 * into the now unreachable buffer. Somewhat complicated by the
3976 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3977 */
3978again:
3979 rcu_read_lock();
3980 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3981 if (!atomic_long_inc_not_zero(&event->refcount)) {
3982 /*
3983 * This event is en-route to free_event() which will
3984 * detach it and remove it from the list.
3985 */
3986 continue;
3987 }
3988 rcu_read_unlock();
3989
3990 mutex_lock(&event->mmap_mutex);
3991 /*
3992 * Check we didn't race with perf_event_set_output() which can
3993 * swizzle the rb from under us while we were waiting to
3994 * acquire mmap_mutex.
3995 *
3996 * If we find a different rb; ignore this event, a next
3997 * iteration will no longer find it on the list. We have to
3998 * still restart the iteration to make sure we're not now
3999 * iterating the wrong list.
4000 */
4001 if (event->rb == rb)
4002 ring_buffer_attach(event, NULL);
4003
4004 mutex_unlock(&event->mmap_mutex);
4005 put_event(event);
4006
4007 /*
4008 * Restart the iteration; either we're on the wrong list or
4009 * destroyed its integrity by doing a deletion.
4010 */
4011 goto again;
4012 }
4013 rcu_read_unlock();
4014
4015 /*
4016 * It could be there's still a few 0-ref events on the list; they'll
4017 * get cleaned up by free_event() -- they'll also still have their
4018 * ref on the rb and will free it whenever they are done with it.
4019 *
4020 * Aside from that, this buffer is 'fully' detached and unmapped,
4021 * undo the VM accounting.
4022 */
4023
4024 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4025 vma->vm_mm->pinned_vm -= mmap_locked;
4026 free_uid(mmap_user);
4027
4028out_put:
4029 ring_buffer_put(rb); /* could be last */
4030}
4031
4032static const struct vm_operations_struct perf_mmap_vmops = {
4033 .open = perf_mmap_open,
4034 .close = perf_mmap_close,
4035 .fault = perf_mmap_fault,
4036 .page_mkwrite = perf_mmap_fault,
4037};
4038
4039static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4040{
4041 struct perf_event *event = file->private_data;
4042 unsigned long user_locked, user_lock_limit;
4043 struct user_struct *user = current_user();
4044 unsigned long locked, lock_limit;
4045 struct ring_buffer *rb;
4046 unsigned long vma_size;
4047 unsigned long nr_pages;
4048 long user_extra, extra;
4049 int ret = 0, flags = 0;
4050
4051 /*
4052 * Don't allow mmap() of inherited per-task counters. This would
4053 * create a performance issue due to all children writing to the
4054 * same rb.
4055 */
4056 if (event->cpu == -1 && event->attr.inherit)
4057 return -EINVAL;
4058
4059 if (!(vma->vm_flags & VM_SHARED))
4060 return -EINVAL;
4061
4062 vma_size = vma->vm_end - vma->vm_start;
4063 nr_pages = (vma_size / PAGE_SIZE) - 1;
4064
4065 /*
4066 * If we have rb pages ensure they're a power-of-two number, so we
4067 * can do bitmasks instead of modulo.
4068 */
4069 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4070 return -EINVAL;
4071
4072 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4073 return -EINVAL;
4074
4075 if (vma->vm_pgoff != 0)
4076 return -EINVAL;
4077
4078 WARN_ON_ONCE(event->ctx->parent_ctx);
4079again:
4080 mutex_lock(&event->mmap_mutex);
4081 if (event->rb) {
4082 if (event->rb->nr_pages != nr_pages) {
4083 ret = -EINVAL;
4084 goto unlock;
4085 }
4086
4087 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4088 /*
4089 * Raced against perf_mmap_close() through
4090 * perf_event_set_output(). Try again, hope for better
4091 * luck.
4092 */
4093 mutex_unlock(&event->mmap_mutex);
4094 goto again;
4095 }
4096
4097 goto unlock;
4098 }
4099
4100 user_extra = nr_pages + 1;
4101 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4102
4103 /*
4104 * Increase the limit linearly with more CPUs:
4105 */
4106 user_lock_limit *= num_online_cpus();
4107
4108 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4109
4110 extra = 0;
4111 if (user_locked > user_lock_limit)
4112 extra = user_locked - user_lock_limit;
4113
4114 lock_limit = rlimit(RLIMIT_MEMLOCK);
4115 lock_limit >>= PAGE_SHIFT;
4116 locked = vma->vm_mm->pinned_vm + extra;
4117
4118 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4119 !capable(CAP_IPC_LOCK)) {
4120 ret = -EPERM;
4121 goto unlock;
4122 }
4123
4124 WARN_ON(event->rb);
4125
4126 if (vma->vm_flags & VM_WRITE)
4127 flags |= RING_BUFFER_WRITABLE;
4128
4129 rb = rb_alloc(nr_pages,
4130 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4131 event->cpu, flags);
4132
4133 if (!rb) {
4134 ret = -ENOMEM;
4135 goto unlock;
4136 }
4137
4138 atomic_set(&rb->mmap_count, 1);
4139 rb->mmap_locked = extra;
4140 rb->mmap_user = get_current_user();
4141
4142 atomic_long_add(user_extra, &user->locked_vm);
4143 vma->vm_mm->pinned_vm += extra;
4144
4145 ring_buffer_attach(event, rb);
4146
4147 perf_event_init_userpage(event);
4148 perf_event_update_userpage(event);
4149
4150unlock:
4151 if (!ret)
4152 atomic_inc(&event->mmap_count);
4153 mutex_unlock(&event->mmap_mutex);
4154
4155 /*
4156 * Since pinned accounting is per vm we cannot allow fork() to copy our
4157 * vma.
4158 */
4159 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4160 vma->vm_ops = &perf_mmap_vmops;
4161
4162 return ret;
4163}
4164
4165static int perf_fasync(int fd, struct file *filp, int on)
4166{
4167 struct inode *inode = file_inode(filp);
4168 struct perf_event *event = filp->private_data;
4169 int retval;
4170
4171 mutex_lock(&inode->i_mutex);
4172 retval = fasync_helper(fd, filp, on, &event->fasync);
4173 mutex_unlock(&inode->i_mutex);
4174
4175 if (retval < 0)
4176 return retval;
4177
4178 return 0;
4179}
4180
4181static const struct file_operations perf_fops = {
4182 .llseek = no_llseek,
4183 .release = perf_release,
4184 .read = perf_read,
4185 .poll = perf_poll,
4186 .unlocked_ioctl = perf_ioctl,
4187 .compat_ioctl = perf_ioctl,
4188 .mmap = perf_mmap,
4189 .fasync = perf_fasync,
4190};
4191
4192/*
4193 * Perf event wakeup
4194 *
4195 * If there's data, ensure we set the poll() state and publish everything
4196 * to user-space before waking everybody up.
4197 */
4198
4199void perf_event_wakeup(struct perf_event *event)
4200{
4201 ring_buffer_wakeup(event);
4202
4203 if (event->pending_kill) {
4204 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4205 event->pending_kill = 0;
4206 }
4207}
4208
4209static void perf_pending_event(struct irq_work *entry)
4210{
4211 struct perf_event *event = container_of(entry,
4212 struct perf_event, pending);
4213
4214 if (event->pending_disable) {
4215 event->pending_disable = 0;
4216 __perf_event_disable(event);
4217 }
4218
4219 if (event->pending_wakeup) {
4220 event->pending_wakeup = 0;
4221 perf_event_wakeup(event);
4222 }
4223}
4224
4225/*
4226 * We assume there is only KVM supporting the callbacks.
4227 * Later on, we might change it to a list if there is
4228 * another virtualization implementation supporting the callbacks.
4229 */
4230struct perf_guest_info_callbacks *perf_guest_cbs;
4231
4232int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4233{
4234 perf_guest_cbs = cbs;
4235 return 0;
4236}
4237EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4238
4239int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4240{
4241 perf_guest_cbs = NULL;
4242 return 0;
4243}
4244EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4245
4246static void
4247perf_output_sample_regs(struct perf_output_handle *handle,
4248 struct pt_regs *regs, u64 mask)
4249{
4250 int bit;
4251
4252 for_each_set_bit(bit, (const unsigned long *) &mask,
4253 sizeof(mask) * BITS_PER_BYTE) {
4254 u64 val;
4255
4256 val = perf_reg_value(regs, bit);
4257 perf_output_put(handle, val);
4258 }
4259}
4260
4261static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4262 struct pt_regs *regs)
4263{
4264 if (!user_mode(regs)) {
4265 if (current->mm)
4266 regs = task_pt_regs(current);
4267 else
4268 regs = NULL;
4269 }
4270
4271 if (regs) {
4272 regs_user->regs = regs;
4273 regs_user->abi = perf_reg_abi(current);
4274 }
4275}
4276
4277/*
4278 * Get remaining task size from user stack pointer.
4279 *
4280 * It'd be better to take stack vma map and limit this more
4281 * precisly, but there's no way to get it safely under interrupt,
4282 * so using TASK_SIZE as limit.
4283 */
4284static u64 perf_ustack_task_size(struct pt_regs *regs)
4285{
4286 unsigned long addr = perf_user_stack_pointer(regs);
4287
4288 if (!addr || addr >= TASK_SIZE)
4289 return 0;
4290
4291 return TASK_SIZE - addr;
4292}
4293
4294static u16
4295perf_sample_ustack_size(u16 stack_size, u16 header_size,
4296 struct pt_regs *regs)
4297{
4298 u64 task_size;
4299
4300 /* No regs, no stack pointer, no dump. */
4301 if (!regs)
4302 return 0;
4303
4304 /*
4305 * Check if we fit in with the requested stack size into the:
4306 * - TASK_SIZE
4307 * If we don't, we limit the size to the TASK_SIZE.
4308 *
4309 * - remaining sample size
4310 * If we don't, we customize the stack size to
4311 * fit in to the remaining sample size.
4312 */
4313
4314 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4315 stack_size = min(stack_size, (u16) task_size);
4316
4317 /* Current header size plus static size and dynamic size. */
4318 header_size += 2 * sizeof(u64);
4319
4320 /* Do we fit in with the current stack dump size? */
4321 if ((u16) (header_size + stack_size) < header_size) {
4322 /*
4323 * If we overflow the maximum size for the sample,
4324 * we customize the stack dump size to fit in.
4325 */
4326 stack_size = USHRT_MAX - header_size - sizeof(u64);
4327 stack_size = round_up(stack_size, sizeof(u64));
4328 }
4329
4330 return stack_size;
4331}
4332
4333static void
4334perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4335 struct pt_regs *regs)
4336{
4337 /* Case of a kernel thread, nothing to dump */
4338 if (!regs) {
4339 u64 size = 0;
4340 perf_output_put(handle, size);
4341 } else {
4342 unsigned long sp;
4343 unsigned int rem;
4344 u64 dyn_size;
4345
4346 /*
4347 * We dump:
4348 * static size
4349 * - the size requested by user or the best one we can fit
4350 * in to the sample max size
4351 * data
4352 * - user stack dump data
4353 * dynamic size
4354 * - the actual dumped size
4355 */
4356
4357 /* Static size. */
4358 perf_output_put(handle, dump_size);
4359
4360 /* Data. */
4361 sp = perf_user_stack_pointer(regs);
4362 rem = __output_copy_user(handle, (void *) sp, dump_size);
4363 dyn_size = dump_size - rem;
4364
4365 perf_output_skip(handle, rem);
4366
4367 /* Dynamic size. */
4368 perf_output_put(handle, dyn_size);
4369 }
4370}
4371
4372static void __perf_event_header__init_id(struct perf_event_header *header,
4373 struct perf_sample_data *data,
4374 struct perf_event *event)
4375{
4376 u64 sample_type = event->attr.sample_type;
4377
4378 data->type = sample_type;
4379 header->size += event->id_header_size;
4380
4381 if (sample_type & PERF_SAMPLE_TID) {
4382 /* namespace issues */
4383 data->tid_entry.pid = perf_event_pid(event, current);
4384 data->tid_entry.tid = perf_event_tid(event, current);
4385 }
4386
4387 if (sample_type & PERF_SAMPLE_TIME)
4388 data->time = perf_clock();
4389
4390 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4391 data->id = primary_event_id(event);
4392
4393 if (sample_type & PERF_SAMPLE_STREAM_ID)
4394 data->stream_id = event->id;
4395
4396 if (sample_type & PERF_SAMPLE_CPU) {
4397 data->cpu_entry.cpu = raw_smp_processor_id();
4398 data->cpu_entry.reserved = 0;
4399 }
4400}
4401
4402void perf_event_header__init_id(struct perf_event_header *header,
4403 struct perf_sample_data *data,
4404 struct perf_event *event)
4405{
4406 if (event->attr.sample_id_all)
4407 __perf_event_header__init_id(header, data, event);
4408}
4409
4410static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4411 struct perf_sample_data *data)
4412{
4413 u64 sample_type = data->type;
4414
4415 if (sample_type & PERF_SAMPLE_TID)
4416 perf_output_put(handle, data->tid_entry);
4417
4418 if (sample_type & PERF_SAMPLE_TIME)
4419 perf_output_put(handle, data->time);
4420
4421 if (sample_type & PERF_SAMPLE_ID)
4422 perf_output_put(handle, data->id);
4423
4424 if (sample_type & PERF_SAMPLE_STREAM_ID)
4425 perf_output_put(handle, data->stream_id);
4426
4427 if (sample_type & PERF_SAMPLE_CPU)
4428 perf_output_put(handle, data->cpu_entry);
4429
4430 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4431 perf_output_put(handle, data->id);
4432}
4433
4434void perf_event__output_id_sample(struct perf_event *event,
4435 struct perf_output_handle *handle,
4436 struct perf_sample_data *sample)
4437{
4438 if (event->attr.sample_id_all)
4439 __perf_event__output_id_sample(handle, sample);
4440}
4441
4442static void perf_output_read_one(struct perf_output_handle *handle,
4443 struct perf_event *event,
4444 u64 enabled, u64 running)
4445{
4446 u64 read_format = event->attr.read_format;
4447 u64 values[4];
4448 int n = 0;
4449
4450 values[n++] = perf_event_count(event);
4451 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4452 values[n++] = enabled +
4453 atomic64_read(&event->child_total_time_enabled);
4454 }
4455 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4456 values[n++] = running +
4457 atomic64_read(&event->child_total_time_running);
4458 }
4459 if (read_format & PERF_FORMAT_ID)
4460 values[n++] = primary_event_id(event);
4461
4462 __output_copy(handle, values, n * sizeof(u64));
4463}
4464
4465/*
4466 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4467 */
4468static void perf_output_read_group(struct perf_output_handle *handle,
4469 struct perf_event *event,
4470 u64 enabled, u64 running)
4471{
4472 struct perf_event *leader = event->group_leader, *sub;
4473 u64 read_format = event->attr.read_format;
4474 u64 values[5];
4475 int n = 0;
4476
4477 values[n++] = 1 + leader->nr_siblings;
4478
4479 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4480 values[n++] = enabled;
4481
4482 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4483 values[n++] = running;
4484
4485 if (leader != event)
4486 leader->pmu->read(leader);
4487
4488 values[n++] = perf_event_count(leader);
4489 if (read_format & PERF_FORMAT_ID)
4490 values[n++] = primary_event_id(leader);
4491
4492 __output_copy(handle, values, n * sizeof(u64));
4493
4494 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4495 n = 0;
4496
4497 if ((sub != event) &&
4498 (sub->state == PERF_EVENT_STATE_ACTIVE))
4499 sub->pmu->read(sub);
4500
4501 values[n++] = perf_event_count(sub);
4502 if (read_format & PERF_FORMAT_ID)
4503 values[n++] = primary_event_id(sub);
4504
4505 __output_copy(handle, values, n * sizeof(u64));
4506 }
4507}
4508
4509#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4510 PERF_FORMAT_TOTAL_TIME_RUNNING)
4511
4512static void perf_output_read(struct perf_output_handle *handle,
4513 struct perf_event *event)
4514{
4515 u64 enabled = 0, running = 0, now;
4516 u64 read_format = event->attr.read_format;
4517
4518 /*
4519 * compute total_time_enabled, total_time_running
4520 * based on snapshot values taken when the event
4521 * was last scheduled in.
4522 *
4523 * we cannot simply called update_context_time()
4524 * because of locking issue as we are called in
4525 * NMI context
4526 */
4527 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4528 calc_timer_values(event, &now, &enabled, &running);
4529
4530 if (event->attr.read_format & PERF_FORMAT_GROUP)
4531 perf_output_read_group(handle, event, enabled, running);
4532 else
4533 perf_output_read_one(handle, event, enabled, running);
4534}
4535
4536void perf_output_sample(struct perf_output_handle *handle,
4537 struct perf_event_header *header,
4538 struct perf_sample_data *data,
4539 struct perf_event *event)
4540{
4541 u64 sample_type = data->type;
4542
4543 perf_output_put(handle, *header);
4544
4545 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4546 perf_output_put(handle, data->id);
4547
4548 if (sample_type & PERF_SAMPLE_IP)
4549 perf_output_put(handle, data->ip);
4550
4551 if (sample_type & PERF_SAMPLE_TID)
4552 perf_output_put(handle, data->tid_entry);
4553
4554 if (sample_type & PERF_SAMPLE_TIME)
4555 perf_output_put(handle, data->time);
4556
4557 if (sample_type & PERF_SAMPLE_ADDR)
4558 perf_output_put(handle, data->addr);
4559
4560 if (sample_type & PERF_SAMPLE_ID)
4561 perf_output_put(handle, data->id);
4562
4563 if (sample_type & PERF_SAMPLE_STREAM_ID)
4564 perf_output_put(handle, data->stream_id);
4565
4566 if (sample_type & PERF_SAMPLE_CPU)
4567 perf_output_put(handle, data->cpu_entry);
4568
4569 if (sample_type & PERF_SAMPLE_PERIOD)
4570 perf_output_put(handle, data->period);
4571
4572 if (sample_type & PERF_SAMPLE_READ)
4573 perf_output_read(handle, event);
4574
4575 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4576 if (data->callchain) {
4577 int size = 1;
4578
4579 if (data->callchain)
4580 size += data->callchain->nr;
4581
4582 size *= sizeof(u64);
4583
4584 __output_copy(handle, data->callchain, size);
4585 } else {
4586 u64 nr = 0;
4587 perf_output_put(handle, nr);
4588 }
4589 }
4590
4591 if (sample_type & PERF_SAMPLE_RAW) {
4592 if (data->raw) {
4593 perf_output_put(handle, data->raw->size);
4594 __output_copy(handle, data->raw->data,
4595 data->raw->size);
4596 } else {
4597 struct {
4598 u32 size;
4599 u32 data;
4600 } raw = {
4601 .size = sizeof(u32),
4602 .data = 0,
4603 };
4604 perf_output_put(handle, raw);
4605 }
4606 }
4607
4608 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4609 if (data->br_stack) {
4610 size_t size;
4611
4612 size = data->br_stack->nr
4613 * sizeof(struct perf_branch_entry);
4614
4615 perf_output_put(handle, data->br_stack->nr);
4616 perf_output_copy(handle, data->br_stack->entries, size);
4617 } else {
4618 /*
4619 * we always store at least the value of nr
4620 */
4621 u64 nr = 0;
4622 perf_output_put(handle, nr);
4623 }
4624 }
4625
4626 if (sample_type & PERF_SAMPLE_REGS_USER) {
4627 u64 abi = data->regs_user.abi;
4628
4629 /*
4630 * If there are no regs to dump, notice it through
4631 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4632 */
4633 perf_output_put(handle, abi);
4634
4635 if (abi) {
4636 u64 mask = event->attr.sample_regs_user;
4637 perf_output_sample_regs(handle,
4638 data->regs_user.regs,
4639 mask);
4640 }
4641 }
4642
4643 if (sample_type & PERF_SAMPLE_STACK_USER) {
4644 perf_output_sample_ustack(handle,
4645 data->stack_user_size,
4646 data->regs_user.regs);
4647 }
4648
4649 if (sample_type & PERF_SAMPLE_WEIGHT)
4650 perf_output_put(handle, data->weight);
4651
4652 if (sample_type & PERF_SAMPLE_DATA_SRC)
4653 perf_output_put(handle, data->data_src.val);
4654
4655 if (sample_type & PERF_SAMPLE_TRANSACTION)
4656 perf_output_put(handle, data->txn);
4657
4658 if (!event->attr.watermark) {
4659 int wakeup_events = event->attr.wakeup_events;
4660
4661 if (wakeup_events) {
4662 struct ring_buffer *rb = handle->rb;
4663 int events = local_inc_return(&rb->events);
4664
4665 if (events >= wakeup_events) {
4666 local_sub(wakeup_events, &rb->events);
4667 local_inc(&rb->wakeup);
4668 }
4669 }
4670 }
4671}
4672
4673void perf_prepare_sample(struct perf_event_header *header,
4674 struct perf_sample_data *data,
4675 struct perf_event *event,
4676 struct pt_regs *regs)
4677{
4678 u64 sample_type = event->attr.sample_type;
4679
4680 header->type = PERF_RECORD_SAMPLE;
4681 header->size = sizeof(*header) + event->header_size;
4682
4683 header->misc = 0;
4684 header->misc |= perf_misc_flags(regs);
4685
4686 __perf_event_header__init_id(header, data, event);
4687
4688 if (sample_type & PERF_SAMPLE_IP)
4689 data->ip = perf_instruction_pointer(regs);
4690
4691 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4692 int size = 1;
4693
4694 data->callchain = perf_callchain(event, regs);
4695
4696 if (data->callchain)
4697 size += data->callchain->nr;
4698
4699 header->size += size * sizeof(u64);
4700 }
4701
4702 if (sample_type & PERF_SAMPLE_RAW) {
4703 int size = sizeof(u32);
4704
4705 if (data->raw)
4706 size += data->raw->size;
4707 else
4708 size += sizeof(u32);
4709
4710 WARN_ON_ONCE(size & (sizeof(u64)-1));
4711 header->size += size;
4712 }
4713
4714 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4715 int size = sizeof(u64); /* nr */
4716 if (data->br_stack) {
4717 size += data->br_stack->nr
4718 * sizeof(struct perf_branch_entry);
4719 }
4720 header->size += size;
4721 }
4722
4723 if (sample_type & PERF_SAMPLE_REGS_USER) {
4724 /* regs dump ABI info */
4725 int size = sizeof(u64);
4726
4727 perf_sample_regs_user(&data->regs_user, regs);
4728
4729 if (data->regs_user.regs) {
4730 u64 mask = event->attr.sample_regs_user;
4731 size += hweight64(mask) * sizeof(u64);
4732 }
4733
4734 header->size += size;
4735 }
4736
4737 if (sample_type & PERF_SAMPLE_STACK_USER) {
4738 /*
4739 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4740 * processed as the last one or have additional check added
4741 * in case new sample type is added, because we could eat
4742 * up the rest of the sample size.
4743 */
4744 struct perf_regs_user *uregs = &data->regs_user;
4745 u16 stack_size = event->attr.sample_stack_user;
4746 u16 size = sizeof(u64);
4747
4748 if (!uregs->abi)
4749 perf_sample_regs_user(uregs, regs);
4750
4751 stack_size = perf_sample_ustack_size(stack_size, header->size,
4752 uregs->regs);
4753
4754 /*
4755 * If there is something to dump, add space for the dump
4756 * itself and for the field that tells the dynamic size,
4757 * which is how many have been actually dumped.
4758 */
4759 if (stack_size)
4760 size += sizeof(u64) + stack_size;
4761
4762 data->stack_user_size = stack_size;
4763 header->size += size;
4764 }
4765}
4766
4767static void perf_event_output(struct perf_event *event,
4768 struct perf_sample_data *data,
4769 struct pt_regs *regs)
4770{
4771 struct perf_output_handle handle;
4772 struct perf_event_header header;
4773
4774 /* protect the callchain buffers */
4775 rcu_read_lock();
4776
4777 perf_prepare_sample(&header, data, event, regs);
4778
4779 if (perf_output_begin(&handle, event, header.size))
4780 goto exit;
4781
4782 perf_output_sample(&handle, &header, data, event);
4783
4784 perf_output_end(&handle);
4785
4786exit:
4787 rcu_read_unlock();
4788}
4789
4790/*
4791 * read event_id
4792 */
4793
4794struct perf_read_event {
4795 struct perf_event_header header;
4796
4797 u32 pid;
4798 u32 tid;
4799};
4800
4801static void
4802perf_event_read_event(struct perf_event *event,
4803 struct task_struct *task)
4804{
4805 struct perf_output_handle handle;
4806 struct perf_sample_data sample;
4807 struct perf_read_event read_event = {
4808 .header = {
4809 .type = PERF_RECORD_READ,
4810 .misc = 0,
4811 .size = sizeof(read_event) + event->read_size,
4812 },
4813 .pid = perf_event_pid(event, task),
4814 .tid = perf_event_tid(event, task),
4815 };
4816 int ret;
4817
4818 perf_event_header__init_id(&read_event.header, &sample, event);
4819 ret = perf_output_begin(&handle, event, read_event.header.size);
4820 if (ret)
4821 return;
4822
4823 perf_output_put(&handle, read_event);
4824 perf_output_read(&handle, event);
4825 perf_event__output_id_sample(event, &handle, &sample);
4826
4827 perf_output_end(&handle);
4828}
4829
4830typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4831
4832static void
4833perf_event_aux_ctx(struct perf_event_context *ctx,
4834 perf_event_aux_output_cb output,
4835 void *data)
4836{
4837 struct perf_event *event;
4838
4839 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4840 if (event->state < PERF_EVENT_STATE_INACTIVE)
4841 continue;
4842 if (!event_filter_match(event))
4843 continue;
4844 output(event, data);
4845 }
4846}
4847
4848static void
4849perf_event_aux(perf_event_aux_output_cb output, void *data,
4850 struct perf_event_context *task_ctx)
4851{
4852 struct perf_cpu_context *cpuctx;
4853 struct perf_event_context *ctx;
4854 struct pmu *pmu;
4855 int ctxn;
4856
4857 rcu_read_lock();
4858 list_for_each_entry_rcu(pmu, &pmus, entry) {
4859 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4860 if (cpuctx->unique_pmu != pmu)
4861 goto next;
4862 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4863 if (task_ctx)
4864 goto next;
4865 ctxn = pmu->task_ctx_nr;
4866 if (ctxn < 0)
4867 goto next;
4868 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4869 if (ctx)
4870 perf_event_aux_ctx(ctx, output, data);
4871next:
4872 put_cpu_ptr(pmu->pmu_cpu_context);
4873 }
4874
4875 if (task_ctx) {
4876 preempt_disable();
4877 perf_event_aux_ctx(task_ctx, output, data);
4878 preempt_enable();
4879 }
4880 rcu_read_unlock();
4881}
4882
4883/*
4884 * task tracking -- fork/exit
4885 *
4886 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4887 */
4888
4889struct perf_task_event {
4890 struct task_struct *task;
4891 struct perf_event_context *task_ctx;
4892
4893 struct {
4894 struct perf_event_header header;
4895
4896 u32 pid;
4897 u32 ppid;
4898 u32 tid;
4899 u32 ptid;
4900 u64 time;
4901 } event_id;
4902};
4903
4904static int perf_event_task_match(struct perf_event *event)
4905{
4906 return event->attr.comm || event->attr.mmap ||
4907 event->attr.mmap2 || event->attr.mmap_data ||
4908 event->attr.task;
4909}
4910
4911static void perf_event_task_output(struct perf_event *event,
4912 void *data)
4913{
4914 struct perf_task_event *task_event = data;
4915 struct perf_output_handle handle;
4916 struct perf_sample_data sample;
4917 struct task_struct *task = task_event->task;
4918 int ret, size = task_event->event_id.header.size;
4919
4920 if (!perf_event_task_match(event))
4921 return;
4922
4923 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4924
4925 ret = perf_output_begin(&handle, event,
4926 task_event->event_id.header.size);
4927 if (ret)
4928 goto out;
4929
4930 task_event->event_id.pid = perf_event_pid(event, task);
4931 task_event->event_id.ppid = perf_event_pid(event, current);
4932
4933 task_event->event_id.tid = perf_event_tid(event, task);
4934 task_event->event_id.ptid = perf_event_tid(event, current);
4935
4936 perf_output_put(&handle, task_event->event_id);
4937
4938 perf_event__output_id_sample(event, &handle, &sample);
4939
4940 perf_output_end(&handle);
4941out:
4942 task_event->event_id.header.size = size;
4943}
4944
4945static void perf_event_task(struct task_struct *task,
4946 struct perf_event_context *task_ctx,
4947 int new)
4948{
4949 struct perf_task_event task_event;
4950
4951 if (!atomic_read(&nr_comm_events) &&
4952 !atomic_read(&nr_mmap_events) &&
4953 !atomic_read(&nr_task_events))
4954 return;
4955
4956 task_event = (struct perf_task_event){
4957 .task = task,
4958 .task_ctx = task_ctx,
4959 .event_id = {
4960 .header = {
4961 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4962 .misc = 0,
4963 .size = sizeof(task_event.event_id),
4964 },
4965 /* .pid */
4966 /* .ppid */
4967 /* .tid */
4968 /* .ptid */
4969 .time = perf_clock(),
4970 },
4971 };
4972
4973 perf_event_aux(perf_event_task_output,
4974 &task_event,
4975 task_ctx);
4976}
4977
4978void perf_event_fork(struct task_struct *task)
4979{
4980 perf_event_task(task, NULL, 1);
4981}
4982
4983/*
4984 * comm tracking
4985 */
4986
4987struct perf_comm_event {
4988 struct task_struct *task;
4989 char *comm;
4990 int comm_size;
4991
4992 struct {
4993 struct perf_event_header header;
4994
4995 u32 pid;
4996 u32 tid;
4997 } event_id;
4998};
4999
5000static int perf_event_comm_match(struct perf_event *event)
5001{
5002 return event->attr.comm;
5003}
5004
5005static void perf_event_comm_output(struct perf_event *event,
5006 void *data)
5007{
5008 struct perf_comm_event *comm_event = data;
5009 struct perf_output_handle handle;
5010 struct perf_sample_data sample;
5011 int size = comm_event->event_id.header.size;
5012 int ret;
5013
5014 if (!perf_event_comm_match(event))
5015 return;
5016
5017 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5018 ret = perf_output_begin(&handle, event,
5019 comm_event->event_id.header.size);
5020
5021 if (ret)
5022 goto out;
5023
5024 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5025 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5026
5027 perf_output_put(&handle, comm_event->event_id);
5028 __output_copy(&handle, comm_event->comm,
5029 comm_event->comm_size);
5030
5031 perf_event__output_id_sample(event, &handle, &sample);
5032
5033 perf_output_end(&handle);
5034out:
5035 comm_event->event_id.header.size = size;
5036}
5037
5038static void perf_event_comm_event(struct perf_comm_event *comm_event)
5039{
5040 char comm[TASK_COMM_LEN];
5041 unsigned int size;
5042
5043 memset(comm, 0, sizeof(comm));
5044 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5045 size = ALIGN(strlen(comm)+1, sizeof(u64));
5046
5047 comm_event->comm = comm;
5048 comm_event->comm_size = size;
5049
5050 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5051
5052 perf_event_aux(perf_event_comm_output,
5053 comm_event,
5054 NULL);
5055}
5056
5057void perf_event_comm(struct task_struct *task)
5058{
5059 struct perf_comm_event comm_event;
5060 struct perf_event_context *ctx;
5061 int ctxn;
5062
5063 rcu_read_lock();
5064 for_each_task_context_nr(ctxn) {
5065 ctx = task->perf_event_ctxp[ctxn];
5066 if (!ctx)
5067 continue;
5068
5069 perf_event_enable_on_exec(ctx);
5070 }
5071 rcu_read_unlock();
5072
5073 if (!atomic_read(&nr_comm_events))
5074 return;
5075
5076 comm_event = (struct perf_comm_event){
5077 .task = task,
5078 /* .comm */
5079 /* .comm_size */
5080 .event_id = {
5081 .header = {
5082 .type = PERF_RECORD_COMM,
5083 .misc = 0,
5084 /* .size */
5085 },
5086 /* .pid */
5087 /* .tid */
5088 },
5089 };
5090
5091 perf_event_comm_event(&comm_event);
5092}
5093
5094/*
5095 * mmap tracking
5096 */
5097
5098struct perf_mmap_event {
5099 struct vm_area_struct *vma;
5100
5101 const char *file_name;
5102 int file_size;
5103 int maj, min;
5104 u64 ino;
5105 u64 ino_generation;
5106
5107 struct {
5108 struct perf_event_header header;
5109
5110 u32 pid;
5111 u32 tid;
5112 u64 start;
5113 u64 len;
5114 u64 pgoff;
5115 } event_id;
5116};
5117
5118static int perf_event_mmap_match(struct perf_event *event,
5119 void *data)
5120{
5121 struct perf_mmap_event *mmap_event = data;
5122 struct vm_area_struct *vma = mmap_event->vma;
5123 int executable = vma->vm_flags & VM_EXEC;
5124
5125 return (!executable && event->attr.mmap_data) ||
5126 (executable && (event->attr.mmap || event->attr.mmap2));
5127}
5128
5129static void perf_event_mmap_output(struct perf_event *event,
5130 void *data)
5131{
5132 struct perf_mmap_event *mmap_event = data;
5133 struct perf_output_handle handle;
5134 struct perf_sample_data sample;
5135 int size = mmap_event->event_id.header.size;
5136 int ret;
5137
5138 if (!perf_event_mmap_match(event, data))
5139 return;
5140
5141 if (event->attr.mmap2) {
5142 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5143 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5144 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5145 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5146 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5147 }
5148
5149 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5150 ret = perf_output_begin(&handle, event,
5151 mmap_event->event_id.header.size);
5152 if (ret)
5153 goto out;
5154
5155 mmap_event->event_id.pid = perf_event_pid(event, current);
5156 mmap_event->event_id.tid = perf_event_tid(event, current);
5157
5158 perf_output_put(&handle, mmap_event->event_id);
5159
5160 if (event->attr.mmap2) {
5161 perf_output_put(&handle, mmap_event->maj);
5162 perf_output_put(&handle, mmap_event->min);
5163 perf_output_put(&handle, mmap_event->ino);
5164 perf_output_put(&handle, mmap_event->ino_generation);
5165 }
5166
5167 __output_copy(&handle, mmap_event->file_name,
5168 mmap_event->file_size);
5169
5170 perf_event__output_id_sample(event, &handle, &sample);
5171
5172 perf_output_end(&handle);
5173out:
5174 mmap_event->event_id.header.size = size;
5175}
5176
5177static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5178{
5179 struct vm_area_struct *vma = mmap_event->vma;
5180 struct file *file = vma->vm_file;
5181 int maj = 0, min = 0;
5182 u64 ino = 0, gen = 0;
5183 unsigned int size;
5184 char tmp[16];
5185 char *buf = NULL;
5186 char *name;
5187
5188 if (file) {
5189 struct inode *inode;
5190 dev_t dev;
5191
5192 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5193 if (!buf) {
5194 name = "//enomem";
5195 goto cpy_name;
5196 }
5197 /*
5198 * d_path() works from the end of the rb backwards, so we
5199 * need to add enough zero bytes after the string to handle
5200 * the 64bit alignment we do later.
5201 */
5202 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5203 if (IS_ERR(name)) {
5204 name = "//toolong";
5205 goto cpy_name;
5206 }
5207 inode = file_inode(vma->vm_file);
5208 dev = inode->i_sb->s_dev;
5209 ino = inode->i_ino;
5210 gen = inode->i_generation;
5211 maj = MAJOR(dev);
5212 min = MINOR(dev);
5213 goto got_name;
5214 } else {
5215 name = (char *)arch_vma_name(vma);
5216 if (name)
5217 goto cpy_name;
5218
5219 if (vma->vm_start <= vma->vm_mm->start_brk &&
5220 vma->vm_end >= vma->vm_mm->brk) {
5221 name = "[heap]";
5222 goto cpy_name;
5223 }
5224 if (vma->vm_start <= vma->vm_mm->start_stack &&
5225 vma->vm_end >= vma->vm_mm->start_stack) {
5226 name = "[stack]";
5227 goto cpy_name;
5228 }
5229
5230 name = "//anon";
5231 goto cpy_name;
5232 }
5233
5234cpy_name:
5235 strlcpy(tmp, name, sizeof(tmp));
5236 name = tmp;
5237got_name:
5238 /*
5239 * Since our buffer works in 8 byte units we need to align our string
5240 * size to a multiple of 8. However, we must guarantee the tail end is
5241 * zero'd out to avoid leaking random bits to userspace.
5242 */
5243 size = strlen(name)+1;
5244 while (!IS_ALIGNED(size, sizeof(u64)))
5245 name[size++] = '\0';
5246
5247 mmap_event->file_name = name;
5248 mmap_event->file_size = size;
5249 mmap_event->maj = maj;
5250 mmap_event->min = min;
5251 mmap_event->ino = ino;
5252 mmap_event->ino_generation = gen;
5253
5254 if (!(vma->vm_flags & VM_EXEC))
5255 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5256
5257 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5258
5259 perf_event_aux(perf_event_mmap_output,
5260 mmap_event,
5261 NULL);
5262
5263 kfree(buf);
5264}
5265
5266void perf_event_mmap(struct vm_area_struct *vma)
5267{
5268 struct perf_mmap_event mmap_event;
5269
5270 if (!atomic_read(&nr_mmap_events))
5271 return;
5272
5273 mmap_event = (struct perf_mmap_event){
5274 .vma = vma,
5275 /* .file_name */
5276 /* .file_size */
5277 .event_id = {
5278 .header = {
5279 .type = PERF_RECORD_MMAP,
5280 .misc = PERF_RECORD_MISC_USER,
5281 /* .size */
5282 },
5283 /* .pid */
5284 /* .tid */
5285 .start = vma->vm_start,
5286 .len = vma->vm_end - vma->vm_start,
5287 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5288 },
5289 /* .maj (attr_mmap2 only) */
5290 /* .min (attr_mmap2 only) */
5291 /* .ino (attr_mmap2 only) */
5292 /* .ino_generation (attr_mmap2 only) */
5293 };
5294
5295 perf_event_mmap_event(&mmap_event);
5296}
5297
5298/*
5299 * IRQ throttle logging
5300 */
5301
5302static void perf_log_throttle(struct perf_event *event, int enable)
5303{
5304 struct perf_output_handle handle;
5305 struct perf_sample_data sample;
5306 int ret;
5307
5308 struct {
5309 struct perf_event_header header;
5310 u64 time;
5311 u64 id;
5312 u64 stream_id;
5313 } throttle_event = {
5314 .header = {
5315 .type = PERF_RECORD_THROTTLE,
5316 .misc = 0,
5317 .size = sizeof(throttle_event),
5318 },
5319 .time = perf_clock(),
5320 .id = primary_event_id(event),
5321 .stream_id = event->id,
5322 };
5323
5324 if (enable)
5325 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5326
5327 perf_event_header__init_id(&throttle_event.header, &sample, event);
5328
5329 ret = perf_output_begin(&handle, event,
5330 throttle_event.header.size);
5331 if (ret)
5332 return;
5333
5334 perf_output_put(&handle, throttle_event);
5335 perf_event__output_id_sample(event, &handle, &sample);
5336 perf_output_end(&handle);
5337}
5338
5339/*
5340 * Generic event overflow handling, sampling.
5341 */
5342
5343static int __perf_event_overflow(struct perf_event *event,
5344 int throttle, struct perf_sample_data *data,
5345 struct pt_regs *regs)
5346{
5347 int events = atomic_read(&event->event_limit);
5348 struct hw_perf_event *hwc = &event->hw;
5349 u64 seq;
5350 int ret = 0;
5351
5352 /*
5353 * Non-sampling counters might still use the PMI to fold short
5354 * hardware counters, ignore those.
5355 */
5356 if (unlikely(!is_sampling_event(event)))
5357 return 0;
5358
5359 seq = __this_cpu_read(perf_throttled_seq);
5360 if (seq != hwc->interrupts_seq) {
5361 hwc->interrupts_seq = seq;
5362 hwc->interrupts = 1;
5363 } else {
5364 hwc->interrupts++;
5365 if (unlikely(throttle
5366 && hwc->interrupts >= max_samples_per_tick)) {
5367 __this_cpu_inc(perf_throttled_count);
5368 hwc->interrupts = MAX_INTERRUPTS;
5369 perf_log_throttle(event, 0);
5370 tick_nohz_full_kick();
5371 ret = 1;
5372 }
5373 }
5374
5375 if (event->attr.freq) {
5376 u64 now = perf_clock();
5377 s64 delta = now - hwc->freq_time_stamp;
5378
5379 hwc->freq_time_stamp = now;
5380
5381 if (delta > 0 && delta < 2*TICK_NSEC)
5382 perf_adjust_period(event, delta, hwc->last_period, true);
5383 }
5384
5385 /*
5386 * XXX event_limit might not quite work as expected on inherited
5387 * events
5388 */
5389
5390 event->pending_kill = POLL_IN;
5391 if (events && atomic_dec_and_test(&event->event_limit)) {
5392 ret = 1;
5393 event->pending_kill = POLL_HUP;
5394 event->pending_disable = 1;
5395 irq_work_queue(&event->pending);
5396 }
5397
5398 if (event->overflow_handler)
5399 event->overflow_handler(event, data, regs);
5400 else
5401 perf_event_output(event, data, regs);
5402
5403 if (event->fasync && event->pending_kill) {
5404 event->pending_wakeup = 1;
5405 irq_work_queue(&event->pending);
5406 }
5407
5408 return ret;
5409}
5410
5411int perf_event_overflow(struct perf_event *event,
5412 struct perf_sample_data *data,
5413 struct pt_regs *regs)
5414{
5415 return __perf_event_overflow(event, 1, data, regs);
5416}
5417
5418/*
5419 * Generic software event infrastructure
5420 */
5421
5422struct swevent_htable {
5423 struct swevent_hlist *swevent_hlist;
5424 struct mutex hlist_mutex;
5425 int hlist_refcount;
5426
5427 /* Recursion avoidance in each contexts */
5428 int recursion[PERF_NR_CONTEXTS];
5429
5430 /* Keeps track of cpu being initialized/exited */
5431 bool online;
5432};
5433
5434static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5435
5436/*
5437 * We directly increment event->count and keep a second value in
5438 * event->hw.period_left to count intervals. This period event
5439 * is kept in the range [-sample_period, 0] so that we can use the
5440 * sign as trigger.
5441 */
5442
5443u64 perf_swevent_set_period(struct perf_event *event)
5444{
5445 struct hw_perf_event *hwc = &event->hw;
5446 u64 period = hwc->last_period;
5447 u64 nr, offset;
5448 s64 old, val;
5449
5450 hwc->last_period = hwc->sample_period;
5451
5452again:
5453 old = val = local64_read(&hwc->period_left);
5454 if (val < 0)
5455 return 0;
5456
5457 nr = div64_u64(period + val, period);
5458 offset = nr * period;
5459 val -= offset;
5460 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5461 goto again;
5462
5463 return nr;
5464}
5465
5466static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5467 struct perf_sample_data *data,
5468 struct pt_regs *regs)
5469{
5470 struct hw_perf_event *hwc = &event->hw;
5471 int throttle = 0;
5472
5473 if (!overflow)
5474 overflow = perf_swevent_set_period(event);
5475
5476 if (hwc->interrupts == MAX_INTERRUPTS)
5477 return;
5478
5479 for (; overflow; overflow--) {
5480 if (__perf_event_overflow(event, throttle,
5481 data, regs)) {
5482 /*
5483 * We inhibit the overflow from happening when
5484 * hwc->interrupts == MAX_INTERRUPTS.
5485 */
5486 break;
5487 }
5488 throttle = 1;
5489 }
5490}
5491
5492static void perf_swevent_event(struct perf_event *event, u64 nr,
5493 struct perf_sample_data *data,
5494 struct pt_regs *regs)
5495{
5496 struct hw_perf_event *hwc = &event->hw;
5497
5498 local64_add(nr, &event->count);
5499
5500 if (!regs)
5501 return;
5502
5503 if (!is_sampling_event(event))
5504 return;
5505
5506 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5507 data->period = nr;
5508 return perf_swevent_overflow(event, 1, data, regs);
5509 } else
5510 data->period = event->hw.last_period;
5511
5512 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5513 return perf_swevent_overflow(event, 1, data, regs);
5514
5515 if (local64_add_negative(nr, &hwc->period_left))
5516 return;
5517
5518 perf_swevent_overflow(event, 0, data, regs);
5519}
5520
5521static int perf_exclude_event(struct perf_event *event,
5522 struct pt_regs *regs)
5523{
5524 if (event->hw.state & PERF_HES_STOPPED)
5525 return 1;
5526
5527 if (regs) {
5528 if (event->attr.exclude_user && user_mode(regs))
5529 return 1;
5530
5531 if (event->attr.exclude_kernel && !user_mode(regs))
5532 return 1;
5533 }
5534
5535 return 0;
5536}
5537
5538static int perf_swevent_match(struct perf_event *event,
5539 enum perf_type_id type,
5540 u32 event_id,
5541 struct perf_sample_data *data,
5542 struct pt_regs *regs)
5543{
5544 if (event->attr.type != type)
5545 return 0;
5546
5547 if (event->attr.config != event_id)
5548 return 0;
5549
5550 if (perf_exclude_event(event, regs))
5551 return 0;
5552
5553 return 1;
5554}
5555
5556static inline u64 swevent_hash(u64 type, u32 event_id)
5557{
5558 u64 val = event_id | (type << 32);
5559
5560 return hash_64(val, SWEVENT_HLIST_BITS);
5561}
5562
5563static inline struct hlist_head *
5564__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5565{
5566 u64 hash = swevent_hash(type, event_id);
5567
5568 return &hlist->heads[hash];
5569}
5570
5571/* For the read side: events when they trigger */
5572static inline struct hlist_head *
5573find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5574{
5575 struct swevent_hlist *hlist;
5576
5577 hlist = rcu_dereference(swhash->swevent_hlist);
5578 if (!hlist)
5579 return NULL;
5580
5581 return __find_swevent_head(hlist, type, event_id);
5582}
5583
5584/* For the event head insertion and removal in the hlist */
5585static inline struct hlist_head *
5586find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5587{
5588 struct swevent_hlist *hlist;
5589 u32 event_id = event->attr.config;
5590 u64 type = event->attr.type;
5591
5592 /*
5593 * Event scheduling is always serialized against hlist allocation
5594 * and release. Which makes the protected version suitable here.
5595 * The context lock guarantees that.
5596 */
5597 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5598 lockdep_is_held(&event->ctx->lock));
5599 if (!hlist)
5600 return NULL;
5601
5602 return __find_swevent_head(hlist, type, event_id);
5603}
5604
5605static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5606 u64 nr,
5607 struct perf_sample_data *data,
5608 struct pt_regs *regs)
5609{
5610 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5611 struct perf_event *event;
5612 struct hlist_head *head;
5613
5614 rcu_read_lock();
5615 head = find_swevent_head_rcu(swhash, type, event_id);
5616 if (!head)
5617 goto end;
5618
5619 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5620 if (perf_swevent_match(event, type, event_id, data, regs))
5621 perf_swevent_event(event, nr, data, regs);
5622 }
5623end:
5624 rcu_read_unlock();
5625}
5626
5627int perf_swevent_get_recursion_context(void)
5628{
5629 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5630
5631 return get_recursion_context(swhash->recursion);
5632}
5633EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5634
5635inline void perf_swevent_put_recursion_context(int rctx)
5636{
5637 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5638
5639 put_recursion_context(swhash->recursion, rctx);
5640}
5641
5642void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5643{
5644 struct perf_sample_data data;
5645 int rctx;
5646
5647 preempt_disable_notrace();
5648 rctx = perf_swevent_get_recursion_context();
5649 if (rctx < 0)
5650 return;
5651
5652 perf_sample_data_init(&data, addr, 0);
5653
5654 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5655
5656 perf_swevent_put_recursion_context(rctx);
5657 preempt_enable_notrace();
5658}
5659
5660static void perf_swevent_read(struct perf_event *event)
5661{
5662}
5663
5664static int perf_swevent_add(struct perf_event *event, int flags)
5665{
5666 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5667 struct hw_perf_event *hwc = &event->hw;
5668 struct hlist_head *head;
5669
5670 if (is_sampling_event(event)) {
5671 hwc->last_period = hwc->sample_period;
5672 perf_swevent_set_period(event);
5673 }
5674
5675 hwc->state = !(flags & PERF_EF_START);
5676
5677 head = find_swevent_head(swhash, event);
5678 if (!head) {
5679 /*
5680 * We can race with cpu hotplug code. Do not
5681 * WARN if the cpu just got unplugged.
5682 */
5683 WARN_ON_ONCE(swhash->online);
5684 return -EINVAL;
5685 }
5686
5687 hlist_add_head_rcu(&event->hlist_entry, head);
5688
5689 return 0;
5690}
5691
5692static void perf_swevent_del(struct perf_event *event, int flags)
5693{
5694 hlist_del_rcu(&event->hlist_entry);
5695}
5696
5697static void perf_swevent_start(struct perf_event *event, int flags)
5698{
5699 event->hw.state = 0;
5700}
5701
5702static void perf_swevent_stop(struct perf_event *event, int flags)
5703{
5704 event->hw.state = PERF_HES_STOPPED;
5705}
5706
5707/* Deref the hlist from the update side */
5708static inline struct swevent_hlist *
5709swevent_hlist_deref(struct swevent_htable *swhash)
5710{
5711 return rcu_dereference_protected(swhash->swevent_hlist,
5712 lockdep_is_held(&swhash->hlist_mutex));
5713}
5714
5715static void swevent_hlist_release(struct swevent_htable *swhash)
5716{
5717 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5718
5719 if (!hlist)
5720 return;
5721
5722 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5723 kfree_rcu(hlist, rcu_head);
5724}
5725
5726static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5727{
5728 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5729
5730 mutex_lock(&swhash->hlist_mutex);
5731
5732 if (!--swhash->hlist_refcount)
5733 swevent_hlist_release(swhash);
5734
5735 mutex_unlock(&swhash->hlist_mutex);
5736}
5737
5738static void swevent_hlist_put(struct perf_event *event)
5739{
5740 int cpu;
5741
5742 for_each_possible_cpu(cpu)
5743 swevent_hlist_put_cpu(event, cpu);
5744}
5745
5746static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5747{
5748 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5749 int err = 0;
5750
5751 mutex_lock(&swhash->hlist_mutex);
5752
5753 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5754 struct swevent_hlist *hlist;
5755
5756 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5757 if (!hlist) {
5758 err = -ENOMEM;
5759 goto exit;
5760 }
5761 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5762 }
5763 swhash->hlist_refcount++;
5764exit:
5765 mutex_unlock(&swhash->hlist_mutex);
5766
5767 return err;
5768}
5769
5770static int swevent_hlist_get(struct perf_event *event)
5771{
5772 int err;
5773 int cpu, failed_cpu;
5774
5775 get_online_cpus();
5776 for_each_possible_cpu(cpu) {
5777 err = swevent_hlist_get_cpu(event, cpu);
5778 if (err) {
5779 failed_cpu = cpu;
5780 goto fail;
5781 }
5782 }
5783 put_online_cpus();
5784
5785 return 0;
5786fail:
5787 for_each_possible_cpu(cpu) {
5788 if (cpu == failed_cpu)
5789 break;
5790 swevent_hlist_put_cpu(event, cpu);
5791 }
5792
5793 put_online_cpus();
5794 return err;
5795}
5796
5797struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5798
5799static void sw_perf_event_destroy(struct perf_event *event)
5800{
5801 u64 event_id = event->attr.config;
5802
5803 WARN_ON(event->parent);
5804
5805 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5806 swevent_hlist_put(event);
5807}
5808
5809static int perf_swevent_init(struct perf_event *event)
5810{
5811 u64 event_id = event->attr.config;
5812
5813 if (event->attr.type != PERF_TYPE_SOFTWARE)
5814 return -ENOENT;
5815
5816 /*
5817 * no branch sampling for software events
5818 */
5819 if (has_branch_stack(event))
5820 return -EOPNOTSUPP;
5821
5822 switch (event_id) {
5823 case PERF_COUNT_SW_CPU_CLOCK:
5824 case PERF_COUNT_SW_TASK_CLOCK:
5825 return -ENOENT;
5826
5827 default:
5828 break;
5829 }
5830
5831 if (event_id >= PERF_COUNT_SW_MAX)
5832 return -ENOENT;
5833
5834 if (!event->parent) {
5835 int err;
5836
5837 err = swevent_hlist_get(event);
5838 if (err)
5839 return err;
5840
5841 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5842 event->destroy = sw_perf_event_destroy;
5843 }
5844
5845 return 0;
5846}
5847
5848static int perf_swevent_event_idx(struct perf_event *event)
5849{
5850 return 0;
5851}
5852
5853static struct pmu perf_swevent = {
5854 .task_ctx_nr = perf_sw_context,
5855
5856 .event_init = perf_swevent_init,
5857 .add = perf_swevent_add,
5858 .del = perf_swevent_del,
5859 .start = perf_swevent_start,
5860 .stop = perf_swevent_stop,
5861 .read = perf_swevent_read,
5862
5863 .event_idx = perf_swevent_event_idx,
5864};
5865
5866#ifdef CONFIG_EVENT_TRACING
5867
5868static int perf_tp_filter_match(struct perf_event *event,
5869 struct perf_sample_data *data)
5870{
5871 void *record = data->raw->data;
5872
5873 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5874 return 1;
5875 return 0;
5876}
5877
5878static int perf_tp_event_match(struct perf_event *event,
5879 struct perf_sample_data *data,
5880 struct pt_regs *regs)
5881{
5882 if (event->hw.state & PERF_HES_STOPPED)
5883 return 0;
5884 /*
5885 * All tracepoints are from kernel-space.
5886 */
5887 if (event->attr.exclude_kernel)
5888 return 0;
5889
5890 if (!perf_tp_filter_match(event, data))
5891 return 0;
5892
5893 return 1;
5894}
5895
5896void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5897 struct pt_regs *regs, struct hlist_head *head, int rctx,
5898 struct task_struct *task)
5899{
5900 struct perf_sample_data data;
5901 struct perf_event *event;
5902
5903 struct perf_raw_record raw = {
5904 .size = entry_size,
5905 .data = record,
5906 };
5907
5908 perf_sample_data_init(&data, addr, 0);
5909 data.raw = &raw;
5910
5911 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5912 if (perf_tp_event_match(event, &data, regs))
5913 perf_swevent_event(event, count, &data, regs);
5914 }
5915
5916 /*
5917 * If we got specified a target task, also iterate its context and
5918 * deliver this event there too.
5919 */
5920 if (task && task != current) {
5921 struct perf_event_context *ctx;
5922 struct trace_entry *entry = record;
5923
5924 rcu_read_lock();
5925 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5926 if (!ctx)
5927 goto unlock;
5928
5929 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5930 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5931 continue;
5932 if (event->attr.config != entry->type)
5933 continue;
5934 if (perf_tp_event_match(event, &data, regs))
5935 perf_swevent_event(event, count, &data, regs);
5936 }
5937unlock:
5938 rcu_read_unlock();
5939 }
5940
5941 perf_swevent_put_recursion_context(rctx);
5942}
5943EXPORT_SYMBOL_GPL(perf_tp_event);
5944
5945static void tp_perf_event_destroy(struct perf_event *event)
5946{
5947 perf_trace_destroy(event);
5948}
5949
5950static int perf_tp_event_init(struct perf_event *event)
5951{
5952 int err;
5953
5954 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5955 return -ENOENT;
5956
5957 /*
5958 * no branch sampling for tracepoint events
5959 */
5960 if (has_branch_stack(event))
5961 return -EOPNOTSUPP;
5962
5963 err = perf_trace_init(event);
5964 if (err)
5965 return err;
5966
5967 event->destroy = tp_perf_event_destroy;
5968
5969 return 0;
5970}
5971
5972static struct pmu perf_tracepoint = {
5973 .task_ctx_nr = perf_sw_context,
5974
5975 .event_init = perf_tp_event_init,
5976 .add = perf_trace_add,
5977 .del = perf_trace_del,
5978 .start = perf_swevent_start,
5979 .stop = perf_swevent_stop,
5980 .read = perf_swevent_read,
5981
5982 .event_idx = perf_swevent_event_idx,
5983};
5984
5985static inline void perf_tp_register(void)
5986{
5987 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5988}
5989
5990static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5991{
5992 char *filter_str;
5993 int ret;
5994
5995 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5996 return -EINVAL;
5997
5998 filter_str = strndup_user(arg, PAGE_SIZE);
5999 if (IS_ERR(filter_str))
6000 return PTR_ERR(filter_str);
6001
6002 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6003
6004 kfree(filter_str);
6005 return ret;
6006}
6007
6008static void perf_event_free_filter(struct perf_event *event)
6009{
6010 ftrace_profile_free_filter(event);
6011}
6012
6013#else
6014
6015static inline void perf_tp_register(void)
6016{
6017}
6018
6019static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6020{
6021 return -ENOENT;
6022}
6023
6024static void perf_event_free_filter(struct perf_event *event)
6025{
6026}
6027
6028#endif /* CONFIG_EVENT_TRACING */
6029
6030#ifdef CONFIG_HAVE_HW_BREAKPOINT
6031void perf_bp_event(struct perf_event *bp, void *data)
6032{
6033 struct perf_sample_data sample;
6034 struct pt_regs *regs = data;
6035
6036 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6037
6038 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6039 perf_swevent_event(bp, 1, &sample, regs);
6040}
6041#endif
6042
6043/*
6044 * hrtimer based swevent callback
6045 */
6046
6047static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6048{
6049 enum hrtimer_restart ret = HRTIMER_RESTART;
6050 struct perf_sample_data data;
6051 struct pt_regs *regs;
6052 struct perf_event *event;
6053 u64 period;
6054
6055 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6056
6057 if (event->state != PERF_EVENT_STATE_ACTIVE)
6058 return HRTIMER_NORESTART;
6059
6060 event->pmu->read(event);
6061
6062 perf_sample_data_init(&data, 0, event->hw.last_period);
6063 regs = get_irq_regs();
6064
6065 if (regs && !perf_exclude_event(event, regs)) {
6066 if (!(event->attr.exclude_idle && is_idle_task(current)))
6067 if (__perf_event_overflow(event, 1, &data, regs))
6068 ret = HRTIMER_NORESTART;
6069 }
6070
6071 period = max_t(u64, 10000, event->hw.sample_period);
6072 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6073
6074 return ret;
6075}
6076
6077static void perf_swevent_start_hrtimer(struct perf_event *event)
6078{
6079 struct hw_perf_event *hwc = &event->hw;
6080 s64 period;
6081
6082 if (!is_sampling_event(event))
6083 return;
6084
6085 period = local64_read(&hwc->period_left);
6086 if (period) {
6087 if (period < 0)
6088 period = 10000;
6089
6090 local64_set(&hwc->period_left, 0);
6091 } else {
6092 period = max_t(u64, 10000, hwc->sample_period);
6093 }
6094 __hrtimer_start_range_ns(&hwc->hrtimer,
6095 ns_to_ktime(period), 0,
6096 HRTIMER_MODE_REL_PINNED, 0);
6097}
6098
6099static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6100{
6101 struct hw_perf_event *hwc = &event->hw;
6102
6103 if (is_sampling_event(event)) {
6104 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6105 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6106
6107 hrtimer_cancel(&hwc->hrtimer);
6108 }
6109}
6110
6111static void perf_swevent_init_hrtimer(struct perf_event *event)
6112{
6113 struct hw_perf_event *hwc = &event->hw;
6114
6115 if (!is_sampling_event(event))
6116 return;
6117
6118 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6119 hwc->hrtimer.function = perf_swevent_hrtimer;
6120
6121 /*
6122 * Since hrtimers have a fixed rate, we can do a static freq->period
6123 * mapping and avoid the whole period adjust feedback stuff.
6124 */
6125 if (event->attr.freq) {
6126 long freq = event->attr.sample_freq;
6127
6128 event->attr.sample_period = NSEC_PER_SEC / freq;
6129 hwc->sample_period = event->attr.sample_period;
6130 local64_set(&hwc->period_left, hwc->sample_period);
6131 hwc->last_period = hwc->sample_period;
6132 event->attr.freq = 0;
6133 }
6134}
6135
6136/*
6137 * Software event: cpu wall time clock
6138 */
6139
6140static void cpu_clock_event_update(struct perf_event *event)
6141{
6142 s64 prev;
6143 u64 now;
6144
6145 now = local_clock();
6146 prev = local64_xchg(&event->hw.prev_count, now);
6147 local64_add(now - prev, &event->count);
6148}
6149
6150static void cpu_clock_event_start(struct perf_event *event, int flags)
6151{
6152 local64_set(&event->hw.prev_count, local_clock());
6153 perf_swevent_start_hrtimer(event);
6154}
6155
6156static void cpu_clock_event_stop(struct perf_event *event, int flags)
6157{
6158 perf_swevent_cancel_hrtimer(event);
6159 cpu_clock_event_update(event);
6160}
6161
6162static int cpu_clock_event_add(struct perf_event *event, int flags)
6163{
6164 if (flags & PERF_EF_START)
6165 cpu_clock_event_start(event, flags);
6166
6167 return 0;
6168}
6169
6170static void cpu_clock_event_del(struct perf_event *event, int flags)
6171{
6172 cpu_clock_event_stop(event, flags);
6173}
6174
6175static void cpu_clock_event_read(struct perf_event *event)
6176{
6177 cpu_clock_event_update(event);
6178}
6179
6180static int cpu_clock_event_init(struct perf_event *event)
6181{
6182 if (event->attr.type != PERF_TYPE_SOFTWARE)
6183 return -ENOENT;
6184
6185 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6186 return -ENOENT;
6187
6188 /*
6189 * no branch sampling for software events
6190 */
6191 if (has_branch_stack(event))
6192 return -EOPNOTSUPP;
6193
6194 perf_swevent_init_hrtimer(event);
6195
6196 return 0;
6197}
6198
6199static struct pmu perf_cpu_clock = {
6200 .task_ctx_nr = perf_sw_context,
6201
6202 .event_init = cpu_clock_event_init,
6203 .add = cpu_clock_event_add,
6204 .del = cpu_clock_event_del,
6205 .start = cpu_clock_event_start,
6206 .stop = cpu_clock_event_stop,
6207 .read = cpu_clock_event_read,
6208
6209 .event_idx = perf_swevent_event_idx,
6210};
6211
6212/*
6213 * Software event: task time clock
6214 */
6215
6216static void task_clock_event_update(struct perf_event *event, u64 now)
6217{
6218 u64 prev;
6219 s64 delta;
6220
6221 prev = local64_xchg(&event->hw.prev_count, now);
6222 delta = now - prev;
6223 local64_add(delta, &event->count);
6224}
6225
6226static void task_clock_event_start(struct perf_event *event, int flags)
6227{
6228 local64_set(&event->hw.prev_count, event->ctx->time);
6229 perf_swevent_start_hrtimer(event);
6230}
6231
6232static void task_clock_event_stop(struct perf_event *event, int flags)
6233{
6234 perf_swevent_cancel_hrtimer(event);
6235 task_clock_event_update(event, event->ctx->time);
6236}
6237
6238static int task_clock_event_add(struct perf_event *event, int flags)
6239{
6240 if (flags & PERF_EF_START)
6241 task_clock_event_start(event, flags);
6242
6243 return 0;
6244}
6245
6246static void task_clock_event_del(struct perf_event *event, int flags)
6247{
6248 task_clock_event_stop(event, PERF_EF_UPDATE);
6249}
6250
6251static void task_clock_event_read(struct perf_event *event)
6252{
6253 u64 now = perf_clock();
6254 u64 delta = now - event->ctx->timestamp;
6255 u64 time = event->ctx->time + delta;
6256
6257 task_clock_event_update(event, time);
6258}
6259
6260static int task_clock_event_init(struct perf_event *event)
6261{
6262 if (event->attr.type != PERF_TYPE_SOFTWARE)
6263 return -ENOENT;
6264
6265 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6266 return -ENOENT;
6267
6268 /*
6269 * no branch sampling for software events
6270 */
6271 if (has_branch_stack(event))
6272 return -EOPNOTSUPP;
6273
6274 perf_swevent_init_hrtimer(event);
6275
6276 return 0;
6277}
6278
6279static struct pmu perf_task_clock = {
6280 .task_ctx_nr = perf_sw_context,
6281
6282 .event_init = task_clock_event_init,
6283 .add = task_clock_event_add,
6284 .del = task_clock_event_del,
6285 .start = task_clock_event_start,
6286 .stop = task_clock_event_stop,
6287 .read = task_clock_event_read,
6288
6289 .event_idx = perf_swevent_event_idx,
6290};
6291
6292static void perf_pmu_nop_void(struct pmu *pmu)
6293{
6294}
6295
6296static int perf_pmu_nop_int(struct pmu *pmu)
6297{
6298 return 0;
6299}
6300
6301static void perf_pmu_start_txn(struct pmu *pmu)
6302{
6303 perf_pmu_disable(pmu);
6304}
6305
6306static int perf_pmu_commit_txn(struct pmu *pmu)
6307{
6308 perf_pmu_enable(pmu);
6309 return 0;
6310}
6311
6312static void perf_pmu_cancel_txn(struct pmu *pmu)
6313{
6314 perf_pmu_enable(pmu);
6315}
6316
6317static int perf_event_idx_default(struct perf_event *event)
6318{
6319 return event->hw.idx + 1;
6320}
6321
6322/*
6323 * Ensures all contexts with the same task_ctx_nr have the same
6324 * pmu_cpu_context too.
6325 */
6326static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6327{
6328 struct pmu *pmu;
6329
6330 if (ctxn < 0)
6331 return NULL;
6332
6333 list_for_each_entry(pmu, &pmus, entry) {
6334 if (pmu->task_ctx_nr == ctxn)
6335 return pmu->pmu_cpu_context;
6336 }
6337
6338 return NULL;
6339}
6340
6341static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6342{
6343 int cpu;
6344
6345 for_each_possible_cpu(cpu) {
6346 struct perf_cpu_context *cpuctx;
6347
6348 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6349
6350 if (cpuctx->unique_pmu == old_pmu)
6351 cpuctx->unique_pmu = pmu;
6352 }
6353}
6354
6355static void free_pmu_context(struct pmu *pmu)
6356{
6357 struct pmu *i;
6358
6359 mutex_lock(&pmus_lock);
6360 /*
6361 * Like a real lame refcount.
6362 */
6363 list_for_each_entry(i, &pmus, entry) {
6364 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6365 update_pmu_context(i, pmu);
6366 goto out;
6367 }
6368 }
6369
6370 free_percpu(pmu->pmu_cpu_context);
6371out:
6372 mutex_unlock(&pmus_lock);
6373}
6374static struct idr pmu_idr;
6375
6376static ssize_t
6377type_show(struct device *dev, struct device_attribute *attr, char *page)
6378{
6379 struct pmu *pmu = dev_get_drvdata(dev);
6380
6381 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6382}
6383static DEVICE_ATTR_RO(type);
6384
6385static ssize_t
6386perf_event_mux_interval_ms_show(struct device *dev,
6387 struct device_attribute *attr,
6388 char *page)
6389{
6390 struct pmu *pmu = dev_get_drvdata(dev);
6391
6392 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6393}
6394
6395static ssize_t
6396perf_event_mux_interval_ms_store(struct device *dev,
6397 struct device_attribute *attr,
6398 const char *buf, size_t count)
6399{
6400 struct pmu *pmu = dev_get_drvdata(dev);
6401 int timer, cpu, ret;
6402
6403 ret = kstrtoint(buf, 0, &timer);
6404 if (ret)
6405 return ret;
6406
6407 if (timer < 1)
6408 return -EINVAL;
6409
6410 /* same value, noting to do */
6411 if (timer == pmu->hrtimer_interval_ms)
6412 return count;
6413
6414 pmu->hrtimer_interval_ms = timer;
6415
6416 /* update all cpuctx for this PMU */
6417 for_each_possible_cpu(cpu) {
6418 struct perf_cpu_context *cpuctx;
6419 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6420 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6421
6422 if (hrtimer_active(&cpuctx->hrtimer))
6423 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6424 }
6425
6426 return count;
6427}
6428static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6429
6430static struct attribute *pmu_dev_attrs[] = {
6431 &dev_attr_type.attr,
6432 &dev_attr_perf_event_mux_interval_ms.attr,
6433 NULL,
6434};
6435ATTRIBUTE_GROUPS(pmu_dev);
6436
6437static int pmu_bus_running;
6438static struct bus_type pmu_bus = {
6439 .name = "event_source",
6440 .dev_groups = pmu_dev_groups,
6441};
6442
6443static void pmu_dev_release(struct device *dev)
6444{
6445 kfree(dev);
6446}
6447
6448static int pmu_dev_alloc(struct pmu *pmu)
6449{
6450 int ret = -ENOMEM;
6451
6452 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6453 if (!pmu->dev)
6454 goto out;
6455
6456 pmu->dev->groups = pmu->attr_groups;
6457 device_initialize(pmu->dev);
6458 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6459 if (ret)
6460 goto free_dev;
6461
6462 dev_set_drvdata(pmu->dev, pmu);
6463 pmu->dev->bus = &pmu_bus;
6464 pmu->dev->release = pmu_dev_release;
6465 ret = device_add(pmu->dev);
6466 if (ret)
6467 goto free_dev;
6468
6469out:
6470 return ret;
6471
6472free_dev:
6473 put_device(pmu->dev);
6474 goto out;
6475}
6476
6477static struct lock_class_key cpuctx_mutex;
6478static struct lock_class_key cpuctx_lock;
6479
6480int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6481{
6482 int cpu, ret;
6483
6484 mutex_lock(&pmus_lock);
6485 ret = -ENOMEM;
6486 pmu->pmu_disable_count = alloc_percpu(int);
6487 if (!pmu->pmu_disable_count)
6488 goto unlock;
6489
6490 pmu->type = -1;
6491 if (!name)
6492 goto skip_type;
6493 pmu->name = name;
6494
6495 if (type < 0) {
6496 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6497 if (type < 0) {
6498 ret = type;
6499 goto free_pdc;
6500 }
6501 }
6502 pmu->type = type;
6503
6504 if (pmu_bus_running) {
6505 ret = pmu_dev_alloc(pmu);
6506 if (ret)
6507 goto free_idr;
6508 }
6509
6510skip_type:
6511 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6512 if (pmu->pmu_cpu_context)
6513 goto got_cpu_context;
6514
6515 ret = -ENOMEM;
6516 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6517 if (!pmu->pmu_cpu_context)
6518 goto free_dev;
6519
6520 for_each_possible_cpu(cpu) {
6521 struct perf_cpu_context *cpuctx;
6522
6523 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6524 __perf_event_init_context(&cpuctx->ctx);
6525 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6526 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6527 cpuctx->ctx.type = cpu_context;
6528 cpuctx->ctx.pmu = pmu;
6529
6530 __perf_cpu_hrtimer_init(cpuctx, cpu);
6531
6532 INIT_LIST_HEAD(&cpuctx->rotation_list);
6533 cpuctx->unique_pmu = pmu;
6534 }
6535
6536got_cpu_context:
6537 if (!pmu->start_txn) {
6538 if (pmu->pmu_enable) {
6539 /*
6540 * If we have pmu_enable/pmu_disable calls, install
6541 * transaction stubs that use that to try and batch
6542 * hardware accesses.
6543 */
6544 pmu->start_txn = perf_pmu_start_txn;
6545 pmu->commit_txn = perf_pmu_commit_txn;
6546 pmu->cancel_txn = perf_pmu_cancel_txn;
6547 } else {
6548 pmu->start_txn = perf_pmu_nop_void;
6549 pmu->commit_txn = perf_pmu_nop_int;
6550 pmu->cancel_txn = perf_pmu_nop_void;
6551 }
6552 }
6553
6554 if (!pmu->pmu_enable) {
6555 pmu->pmu_enable = perf_pmu_nop_void;
6556 pmu->pmu_disable = perf_pmu_nop_void;
6557 }
6558
6559 if (!pmu->event_idx)
6560 pmu->event_idx = perf_event_idx_default;
6561
6562 list_add_rcu(&pmu->entry, &pmus);
6563 ret = 0;
6564unlock:
6565 mutex_unlock(&pmus_lock);
6566
6567 return ret;
6568
6569free_dev:
6570 device_del(pmu->dev);
6571 put_device(pmu->dev);
6572
6573free_idr:
6574 if (pmu->type >= PERF_TYPE_MAX)
6575 idr_remove(&pmu_idr, pmu->type);
6576
6577free_pdc:
6578 free_percpu(pmu->pmu_disable_count);
6579 goto unlock;
6580}
6581
6582void perf_pmu_unregister(struct pmu *pmu)
6583{
6584 mutex_lock(&pmus_lock);
6585 list_del_rcu(&pmu->entry);
6586 mutex_unlock(&pmus_lock);
6587
6588 /*
6589 * We dereference the pmu list under both SRCU and regular RCU, so
6590 * synchronize against both of those.
6591 */
6592 synchronize_srcu(&pmus_srcu);
6593 synchronize_rcu();
6594
6595 free_percpu(pmu->pmu_disable_count);
6596 if (pmu->type >= PERF_TYPE_MAX)
6597 idr_remove(&pmu_idr, pmu->type);
6598 device_del(pmu->dev);
6599 put_device(pmu->dev);
6600 free_pmu_context(pmu);
6601}
6602
6603struct pmu *perf_init_event(struct perf_event *event)
6604{
6605 struct pmu *pmu = NULL;
6606 int idx;
6607 int ret;
6608
6609 idx = srcu_read_lock(&pmus_srcu);
6610
6611 rcu_read_lock();
6612 pmu = idr_find(&pmu_idr, event->attr.type);
6613 rcu_read_unlock();
6614 if (pmu) {
6615 event->pmu = pmu;
6616 ret = pmu->event_init(event);
6617 if (ret)
6618 pmu = ERR_PTR(ret);
6619 goto unlock;
6620 }
6621
6622 list_for_each_entry_rcu(pmu, &pmus, entry) {
6623 event->pmu = pmu;
6624 ret = pmu->event_init(event);
6625 if (!ret)
6626 goto unlock;
6627
6628 if (ret != -ENOENT) {
6629 pmu = ERR_PTR(ret);
6630 goto unlock;
6631 }
6632 }
6633 pmu = ERR_PTR(-ENOENT);
6634unlock:
6635 srcu_read_unlock(&pmus_srcu, idx);
6636
6637 return pmu;
6638}
6639
6640static void account_event_cpu(struct perf_event *event, int cpu)
6641{
6642 if (event->parent)
6643 return;
6644
6645 if (has_branch_stack(event)) {
6646 if (!(event->attach_state & PERF_ATTACH_TASK))
6647 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6648 }
6649 if (is_cgroup_event(event))
6650 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6651}
6652
6653static void account_event(struct perf_event *event)
6654{
6655 if (event->parent)
6656 return;
6657
6658 if (event->attach_state & PERF_ATTACH_TASK)
6659 static_key_slow_inc(&perf_sched_events.key);
6660 if (event->attr.mmap || event->attr.mmap_data)
6661 atomic_inc(&nr_mmap_events);
6662 if (event->attr.comm)
6663 atomic_inc(&nr_comm_events);
6664 if (event->attr.task)
6665 atomic_inc(&nr_task_events);
6666 if (event->attr.freq) {
6667 if (atomic_inc_return(&nr_freq_events) == 1)
6668 tick_nohz_full_kick_all();
6669 }
6670 if (has_branch_stack(event))
6671 static_key_slow_inc(&perf_sched_events.key);
6672 if (is_cgroup_event(event))
6673 static_key_slow_inc(&perf_sched_events.key);
6674
6675 account_event_cpu(event, event->cpu);
6676}
6677
6678/*
6679 * Allocate and initialize a event structure
6680 */
6681static struct perf_event *
6682perf_event_alloc(struct perf_event_attr *attr, int cpu,
6683 struct task_struct *task,
6684 struct perf_event *group_leader,
6685 struct perf_event *parent_event,
6686 perf_overflow_handler_t overflow_handler,
6687 void *context)
6688{
6689 struct pmu *pmu;
6690 struct perf_event *event;
6691 struct hw_perf_event *hwc;
6692 long err = -EINVAL;
6693
6694 if ((unsigned)cpu >= nr_cpu_ids) {
6695 if (!task || cpu != -1)
6696 return ERR_PTR(-EINVAL);
6697 }
6698
6699 event = kzalloc(sizeof(*event), GFP_KERNEL);
6700 if (!event)
6701 return ERR_PTR(-ENOMEM);
6702
6703 /*
6704 * Single events are their own group leaders, with an
6705 * empty sibling list:
6706 */
6707 if (!group_leader)
6708 group_leader = event;
6709
6710 mutex_init(&event->child_mutex);
6711 INIT_LIST_HEAD(&event->child_list);
6712
6713 INIT_LIST_HEAD(&event->group_entry);
6714 INIT_LIST_HEAD(&event->event_entry);
6715 INIT_LIST_HEAD(&event->sibling_list);
6716 INIT_LIST_HEAD(&event->rb_entry);
6717 INIT_LIST_HEAD(&event->active_entry);
6718 INIT_HLIST_NODE(&event->hlist_entry);
6719
6720
6721 init_waitqueue_head(&event->waitq);
6722 init_irq_work(&event->pending, perf_pending_event);
6723
6724 mutex_init(&event->mmap_mutex);
6725
6726 atomic_long_set(&event->refcount, 1);
6727 event->cpu = cpu;
6728 event->attr = *attr;
6729 event->group_leader = group_leader;
6730 event->pmu = NULL;
6731 event->oncpu = -1;
6732
6733 event->parent = parent_event;
6734
6735 event->ns = get_pid_ns(task_active_pid_ns(current));
6736 event->id = atomic64_inc_return(&perf_event_id);
6737
6738 event->state = PERF_EVENT_STATE_INACTIVE;
6739
6740 if (task) {
6741 event->attach_state = PERF_ATTACH_TASK;
6742
6743 if (attr->type == PERF_TYPE_TRACEPOINT)
6744 event->hw.tp_target = task;
6745#ifdef CONFIG_HAVE_HW_BREAKPOINT
6746 /*
6747 * hw_breakpoint is a bit difficult here..
6748 */
6749 else if (attr->type == PERF_TYPE_BREAKPOINT)
6750 event->hw.bp_target = task;
6751#endif
6752 }
6753
6754 if (!overflow_handler && parent_event) {
6755 overflow_handler = parent_event->overflow_handler;
6756 context = parent_event->overflow_handler_context;
6757 }
6758
6759 event->overflow_handler = overflow_handler;
6760 event->overflow_handler_context = context;
6761
6762 perf_event__state_init(event);
6763
6764 pmu = NULL;
6765
6766 hwc = &event->hw;
6767 hwc->sample_period = attr->sample_period;
6768 if (attr->freq && attr->sample_freq)
6769 hwc->sample_period = 1;
6770 hwc->last_period = hwc->sample_period;
6771
6772 local64_set(&hwc->period_left, hwc->sample_period);
6773
6774 /*
6775 * we currently do not support PERF_FORMAT_GROUP on inherited events
6776 */
6777 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6778 goto err_ns;
6779
6780 pmu = perf_init_event(event);
6781 if (!pmu)
6782 goto err_ns;
6783 else if (IS_ERR(pmu)) {
6784 err = PTR_ERR(pmu);
6785 goto err_ns;
6786 }
6787
6788 if (!event->parent) {
6789 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6790 err = get_callchain_buffers();
6791 if (err)
6792 goto err_pmu;
6793 }
6794 }
6795
6796 return event;
6797
6798err_pmu:
6799 if (event->destroy)
6800 event->destroy(event);
6801err_ns:
6802 if (event->ns)
6803 put_pid_ns(event->ns);
6804 kfree(event);
6805
6806 return ERR_PTR(err);
6807}
6808
6809static int perf_copy_attr(struct perf_event_attr __user *uattr,
6810 struct perf_event_attr *attr)
6811{
6812 u32 size;
6813 int ret;
6814
6815 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6816 return -EFAULT;
6817
6818 /*
6819 * zero the full structure, so that a short copy will be nice.
6820 */
6821 memset(attr, 0, sizeof(*attr));
6822
6823 ret = get_user(size, &uattr->size);
6824 if (ret)
6825 return ret;
6826
6827 if (size > PAGE_SIZE) /* silly large */
6828 goto err_size;
6829
6830 if (!size) /* abi compat */
6831 size = PERF_ATTR_SIZE_VER0;
6832
6833 if (size < PERF_ATTR_SIZE_VER0)
6834 goto err_size;
6835
6836 /*
6837 * If we're handed a bigger struct than we know of,
6838 * ensure all the unknown bits are 0 - i.e. new
6839 * user-space does not rely on any kernel feature
6840 * extensions we dont know about yet.
6841 */
6842 if (size > sizeof(*attr)) {
6843 unsigned char __user *addr;
6844 unsigned char __user *end;
6845 unsigned char val;
6846
6847 addr = (void __user *)uattr + sizeof(*attr);
6848 end = (void __user *)uattr + size;
6849
6850 for (; addr < end; addr++) {
6851 ret = get_user(val, addr);
6852 if (ret)
6853 return ret;
6854 if (val)
6855 goto err_size;
6856 }
6857 size = sizeof(*attr);
6858 }
6859
6860 ret = copy_from_user(attr, uattr, size);
6861 if (ret)
6862 return -EFAULT;
6863
6864 /* disabled for now */
6865 if (attr->mmap2)
6866 return -EINVAL;
6867
6868 if (attr->__reserved_1)
6869 return -EINVAL;
6870
6871 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6872 return -EINVAL;
6873
6874 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6875 return -EINVAL;
6876
6877 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6878 u64 mask = attr->branch_sample_type;
6879
6880 /* only using defined bits */
6881 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6882 return -EINVAL;
6883
6884 /* at least one branch bit must be set */
6885 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6886 return -EINVAL;
6887
6888 /* propagate priv level, when not set for branch */
6889 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6890
6891 /* exclude_kernel checked on syscall entry */
6892 if (!attr->exclude_kernel)
6893 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6894
6895 if (!attr->exclude_user)
6896 mask |= PERF_SAMPLE_BRANCH_USER;
6897
6898 if (!attr->exclude_hv)
6899 mask |= PERF_SAMPLE_BRANCH_HV;
6900 /*
6901 * adjust user setting (for HW filter setup)
6902 */
6903 attr->branch_sample_type = mask;
6904 }
6905 /* privileged levels capture (kernel, hv): check permissions */
6906 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6907 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6908 return -EACCES;
6909 }
6910
6911 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6912 ret = perf_reg_validate(attr->sample_regs_user);
6913 if (ret)
6914 return ret;
6915 }
6916
6917 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6918 if (!arch_perf_have_user_stack_dump())
6919 return -ENOSYS;
6920
6921 /*
6922 * We have __u32 type for the size, but so far
6923 * we can only use __u16 as maximum due to the
6924 * __u16 sample size limit.
6925 */
6926 if (attr->sample_stack_user >= USHRT_MAX)
6927 ret = -EINVAL;
6928 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6929 ret = -EINVAL;
6930 }
6931
6932out:
6933 return ret;
6934
6935err_size:
6936 put_user(sizeof(*attr), &uattr->size);
6937 ret = -E2BIG;
6938 goto out;
6939}
6940
6941static int
6942perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6943{
6944 struct ring_buffer *rb = NULL;
6945 int ret = -EINVAL;
6946
6947 if (!output_event)
6948 goto set;
6949
6950 /* don't allow circular references */
6951 if (event == output_event)
6952 goto out;
6953
6954 /*
6955 * Don't allow cross-cpu buffers
6956 */
6957 if (output_event->cpu != event->cpu)
6958 goto out;
6959
6960 /*
6961 * If its not a per-cpu rb, it must be the same task.
6962 */
6963 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6964 goto out;
6965
6966set:
6967 mutex_lock(&event->mmap_mutex);
6968 /* Can't redirect output if we've got an active mmap() */
6969 if (atomic_read(&event->mmap_count))
6970 goto unlock;
6971
6972 if (output_event) {
6973 /* get the rb we want to redirect to */
6974 rb = ring_buffer_get(output_event);
6975 if (!rb)
6976 goto unlock;
6977 }
6978
6979 ring_buffer_attach(event, rb);
6980
6981 ret = 0;
6982unlock:
6983 mutex_unlock(&event->mmap_mutex);
6984
6985out:
6986 return ret;
6987}
6988
6989/**
6990 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6991 *
6992 * @attr_uptr: event_id type attributes for monitoring/sampling
6993 * @pid: target pid
6994 * @cpu: target cpu
6995 * @group_fd: group leader event fd
6996 */
6997SYSCALL_DEFINE5(perf_event_open,
6998 struct perf_event_attr __user *, attr_uptr,
6999 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7000{
7001 struct perf_event *group_leader = NULL, *output_event = NULL;
7002 struct perf_event *event, *sibling;
7003 struct perf_event_attr attr;
7004 struct perf_event_context *ctx;
7005 struct file *event_file = NULL;
7006 struct fd group = {NULL, 0};
7007 struct task_struct *task = NULL;
7008 struct pmu *pmu;
7009 int event_fd;
7010 int move_group = 0;
7011 int err;
7012 int f_flags = O_RDWR;
7013
7014 /* for future expandability... */
7015 if (flags & ~PERF_FLAG_ALL)
7016 return -EINVAL;
7017
7018 err = perf_copy_attr(attr_uptr, &attr);
7019 if (err)
7020 return err;
7021
7022 if (!attr.exclude_kernel) {
7023 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7024 return -EACCES;
7025 }
7026
7027 if (attr.freq) {
7028 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7029 return -EINVAL;
7030 } else {
7031 if (attr.sample_period & (1ULL << 63))
7032 return -EINVAL;
7033 }
7034
7035 /*
7036 * In cgroup mode, the pid argument is used to pass the fd
7037 * opened to the cgroup directory in cgroupfs. The cpu argument
7038 * designates the cpu on which to monitor threads from that
7039 * cgroup.
7040 */
7041 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7042 return -EINVAL;
7043
7044 if (flags & PERF_FLAG_FD_CLOEXEC)
7045 f_flags |= O_CLOEXEC;
7046
7047 event_fd = get_unused_fd_flags(f_flags);
7048 if (event_fd < 0)
7049 return event_fd;
7050
7051 if (group_fd != -1) {
7052 err = perf_fget_light(group_fd, &group);
7053 if (err)
7054 goto err_fd;
7055 group_leader = group.file->private_data;
7056 if (flags & PERF_FLAG_FD_OUTPUT)
7057 output_event = group_leader;
7058 if (flags & PERF_FLAG_FD_NO_GROUP)
7059 group_leader = NULL;
7060 }
7061
7062 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7063 task = find_lively_task_by_vpid(pid);
7064 if (IS_ERR(task)) {
7065 err = PTR_ERR(task);
7066 goto err_group_fd;
7067 }
7068 }
7069
7070 get_online_cpus();
7071
7072 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7073 NULL, NULL);
7074 if (IS_ERR(event)) {
7075 err = PTR_ERR(event);
7076 goto err_task;
7077 }
7078
7079 if (flags & PERF_FLAG_PID_CGROUP) {
7080 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7081 if (err) {
7082 __free_event(event);
7083 goto err_task;
7084 }
7085 }
7086
7087 account_event(event);
7088
7089 /*
7090 * Special case software events and allow them to be part of
7091 * any hardware group.
7092 */
7093 pmu = event->pmu;
7094
7095 if (group_leader &&
7096 (is_software_event(event) != is_software_event(group_leader))) {
7097 if (is_software_event(event)) {
7098 /*
7099 * If event and group_leader are not both a software
7100 * event, and event is, then group leader is not.
7101 *
7102 * Allow the addition of software events to !software
7103 * groups, this is safe because software events never
7104 * fail to schedule.
7105 */
7106 pmu = group_leader->pmu;
7107 } else if (is_software_event(group_leader) &&
7108 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7109 /*
7110 * In case the group is a pure software group, and we
7111 * try to add a hardware event, move the whole group to
7112 * the hardware context.
7113 */
7114 move_group = 1;
7115 }
7116 }
7117
7118 /*
7119 * Get the target context (task or percpu):
7120 */
7121 ctx = find_get_context(pmu, task, event->cpu);
7122 if (IS_ERR(ctx)) {
7123 err = PTR_ERR(ctx);
7124 goto err_alloc;
7125 }
7126
7127 if (task) {
7128 put_task_struct(task);
7129 task = NULL;
7130 }
7131
7132 /*
7133 * Look up the group leader (we will attach this event to it):
7134 */
7135 if (group_leader) {
7136 err = -EINVAL;
7137
7138 /*
7139 * Do not allow a recursive hierarchy (this new sibling
7140 * becoming part of another group-sibling):
7141 */
7142 if (group_leader->group_leader != group_leader)
7143 goto err_context;
7144 /*
7145 * Do not allow to attach to a group in a different
7146 * task or CPU context:
7147 */
7148 if (move_group) {
7149 if (group_leader->ctx->type != ctx->type)
7150 goto err_context;
7151 } else {
7152 if (group_leader->ctx != ctx)
7153 goto err_context;
7154 }
7155
7156 /*
7157 * Only a group leader can be exclusive or pinned
7158 */
7159 if (attr.exclusive || attr.pinned)
7160 goto err_context;
7161 }
7162
7163 if (output_event) {
7164 err = perf_event_set_output(event, output_event);
7165 if (err)
7166 goto err_context;
7167 }
7168
7169 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7170 f_flags);
7171 if (IS_ERR(event_file)) {
7172 err = PTR_ERR(event_file);
7173 goto err_context;
7174 }
7175
7176 if (move_group) {
7177 struct perf_event_context *gctx = group_leader->ctx;
7178
7179 mutex_lock(&gctx->mutex);
7180 perf_remove_from_context(group_leader, false);
7181
7182 /*
7183 * Removing from the context ends up with disabled
7184 * event. What we want here is event in the initial
7185 * startup state, ready to be add into new context.
7186 */
7187 perf_event__state_init(group_leader);
7188 list_for_each_entry(sibling, &group_leader->sibling_list,
7189 group_entry) {
7190 perf_remove_from_context(sibling, false);
7191 perf_event__state_init(sibling);
7192 put_ctx(gctx);
7193 }
7194 mutex_unlock(&gctx->mutex);
7195 put_ctx(gctx);
7196 }
7197
7198 WARN_ON_ONCE(ctx->parent_ctx);
7199 mutex_lock(&ctx->mutex);
7200
7201 if (move_group) {
7202 synchronize_rcu();
7203 perf_install_in_context(ctx, group_leader, event->cpu);
7204 get_ctx(ctx);
7205 list_for_each_entry(sibling, &group_leader->sibling_list,
7206 group_entry) {
7207 perf_install_in_context(ctx, sibling, event->cpu);
7208 get_ctx(ctx);
7209 }
7210 }
7211
7212 perf_install_in_context(ctx, event, event->cpu);
7213 perf_unpin_context(ctx);
7214 mutex_unlock(&ctx->mutex);
7215
7216 put_online_cpus();
7217
7218 event->owner = current;
7219
7220 mutex_lock(¤t->perf_event_mutex);
7221 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7222 mutex_unlock(¤t->perf_event_mutex);
7223
7224 /*
7225 * Precalculate sample_data sizes
7226 */
7227 perf_event__header_size(event);
7228 perf_event__id_header_size(event);
7229
7230 /*
7231 * Drop the reference on the group_event after placing the
7232 * new event on the sibling_list. This ensures destruction
7233 * of the group leader will find the pointer to itself in
7234 * perf_group_detach().
7235 */
7236 fdput(group);
7237 fd_install(event_fd, event_file);
7238 return event_fd;
7239
7240err_context:
7241 perf_unpin_context(ctx);
7242 put_ctx(ctx);
7243err_alloc:
7244 free_event(event);
7245err_task:
7246 put_online_cpus();
7247 if (task)
7248 put_task_struct(task);
7249err_group_fd:
7250 fdput(group);
7251err_fd:
7252 put_unused_fd(event_fd);
7253 return err;
7254}
7255
7256/**
7257 * perf_event_create_kernel_counter
7258 *
7259 * @attr: attributes of the counter to create
7260 * @cpu: cpu in which the counter is bound
7261 * @task: task to profile (NULL for percpu)
7262 */
7263struct perf_event *
7264perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7265 struct task_struct *task,
7266 perf_overflow_handler_t overflow_handler,
7267 void *context)
7268{
7269 struct perf_event_context *ctx;
7270 struct perf_event *event;
7271 int err;
7272
7273 /*
7274 * Get the target context (task or percpu):
7275 */
7276
7277 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7278 overflow_handler, context);
7279 if (IS_ERR(event)) {
7280 err = PTR_ERR(event);
7281 goto err;
7282 }
7283
7284 account_event(event);
7285
7286 ctx = find_get_context(event->pmu, task, cpu);
7287 if (IS_ERR(ctx)) {
7288 err = PTR_ERR(ctx);
7289 goto err_free;
7290 }
7291
7292 WARN_ON_ONCE(ctx->parent_ctx);
7293 mutex_lock(&ctx->mutex);
7294 perf_install_in_context(ctx, event, cpu);
7295 perf_unpin_context(ctx);
7296 mutex_unlock(&ctx->mutex);
7297
7298 return event;
7299
7300err_free:
7301 free_event(event);
7302err:
7303 return ERR_PTR(err);
7304}
7305EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7306
7307void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7308{
7309 struct perf_event_context *src_ctx;
7310 struct perf_event_context *dst_ctx;
7311 struct perf_event *event, *tmp;
7312 LIST_HEAD(events);
7313
7314 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7315 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7316
7317 mutex_lock(&src_ctx->mutex);
7318 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7319 event_entry) {
7320 perf_remove_from_context(event, false);
7321 unaccount_event_cpu(event, src_cpu);
7322 put_ctx(src_ctx);
7323 list_add(&event->migrate_entry, &events);
7324 }
7325 mutex_unlock(&src_ctx->mutex);
7326
7327 synchronize_rcu();
7328
7329 mutex_lock(&dst_ctx->mutex);
7330 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7331 list_del(&event->migrate_entry);
7332 if (event->state >= PERF_EVENT_STATE_OFF)
7333 event->state = PERF_EVENT_STATE_INACTIVE;
7334 account_event_cpu(event, dst_cpu);
7335 perf_install_in_context(dst_ctx, event, dst_cpu);
7336 get_ctx(dst_ctx);
7337 }
7338 mutex_unlock(&dst_ctx->mutex);
7339}
7340EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7341
7342static void sync_child_event(struct perf_event *child_event,
7343 struct task_struct *child)
7344{
7345 struct perf_event *parent_event = child_event->parent;
7346 u64 child_val;
7347
7348 if (child_event->attr.inherit_stat)
7349 perf_event_read_event(child_event, child);
7350
7351 child_val = perf_event_count(child_event);
7352
7353 /*
7354 * Add back the child's count to the parent's count:
7355 */
7356 atomic64_add(child_val, &parent_event->child_count);
7357 atomic64_add(child_event->total_time_enabled,
7358 &parent_event->child_total_time_enabled);
7359 atomic64_add(child_event->total_time_running,
7360 &parent_event->child_total_time_running);
7361
7362 /*
7363 * Remove this event from the parent's list
7364 */
7365 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7366 mutex_lock(&parent_event->child_mutex);
7367 list_del_init(&child_event->child_list);
7368 mutex_unlock(&parent_event->child_mutex);
7369
7370 /*
7371 * Release the parent event, if this was the last
7372 * reference to it.
7373 */
7374 put_event(parent_event);
7375}
7376
7377static void
7378__perf_event_exit_task(struct perf_event *child_event,
7379 struct perf_event_context *child_ctx,
7380 struct task_struct *child)
7381{
7382 perf_remove_from_context(child_event, !!child_event->parent);
7383
7384 /*
7385 * It can happen that the parent exits first, and has events
7386 * that are still around due to the child reference. These
7387 * events need to be zapped.
7388 */
7389 if (child_event->parent) {
7390 sync_child_event(child_event, child);
7391 free_event(child_event);
7392 }
7393}
7394
7395static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7396{
7397 struct perf_event *child_event, *tmp;
7398 struct perf_event_context *child_ctx;
7399 unsigned long flags;
7400
7401 if (likely(!child->perf_event_ctxp[ctxn])) {
7402 perf_event_task(child, NULL, 0);
7403 return;
7404 }
7405
7406 local_irq_save(flags);
7407 /*
7408 * We can't reschedule here because interrupts are disabled,
7409 * and either child is current or it is a task that can't be
7410 * scheduled, so we are now safe from rescheduling changing
7411 * our context.
7412 */
7413 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7414
7415 /*
7416 * Take the context lock here so that if find_get_context is
7417 * reading child->perf_event_ctxp, we wait until it has
7418 * incremented the context's refcount before we do put_ctx below.
7419 */
7420 raw_spin_lock(&child_ctx->lock);
7421 task_ctx_sched_out(child_ctx);
7422 child->perf_event_ctxp[ctxn] = NULL;
7423 /*
7424 * If this context is a clone; unclone it so it can't get
7425 * swapped to another process while we're removing all
7426 * the events from it.
7427 */
7428 unclone_ctx(child_ctx);
7429 update_context_time(child_ctx);
7430 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7431
7432 /*
7433 * Report the task dead after unscheduling the events so that we
7434 * won't get any samples after PERF_RECORD_EXIT. We can however still
7435 * get a few PERF_RECORD_READ events.
7436 */
7437 perf_event_task(child, child_ctx, 0);
7438
7439 /*
7440 * We can recurse on the same lock type through:
7441 *
7442 * __perf_event_exit_task()
7443 * sync_child_event()
7444 * put_event()
7445 * mutex_lock(&ctx->mutex)
7446 *
7447 * But since its the parent context it won't be the same instance.
7448 */
7449 mutex_lock(&child_ctx->mutex);
7450
7451again:
7452 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7453 group_entry)
7454 __perf_event_exit_task(child_event, child_ctx, child);
7455
7456 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7457 group_entry)
7458 __perf_event_exit_task(child_event, child_ctx, child);
7459
7460 /*
7461 * If the last event was a group event, it will have appended all
7462 * its siblings to the list, but we obtained 'tmp' before that which
7463 * will still point to the list head terminating the iteration.
7464 */
7465 if (!list_empty(&child_ctx->pinned_groups) ||
7466 !list_empty(&child_ctx->flexible_groups))
7467 goto again;
7468
7469 mutex_unlock(&child_ctx->mutex);
7470
7471 put_ctx(child_ctx);
7472}
7473
7474/*
7475 * When a child task exits, feed back event values to parent events.
7476 */
7477void perf_event_exit_task(struct task_struct *child)
7478{
7479 struct perf_event *event, *tmp;
7480 int ctxn;
7481
7482 mutex_lock(&child->perf_event_mutex);
7483 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7484 owner_entry) {
7485 list_del_init(&event->owner_entry);
7486
7487 /*
7488 * Ensure the list deletion is visible before we clear
7489 * the owner, closes a race against perf_release() where
7490 * we need to serialize on the owner->perf_event_mutex.
7491 */
7492 smp_wmb();
7493 event->owner = NULL;
7494 }
7495 mutex_unlock(&child->perf_event_mutex);
7496
7497 for_each_task_context_nr(ctxn)
7498 perf_event_exit_task_context(child, ctxn);
7499}
7500
7501static void perf_free_event(struct perf_event *event,
7502 struct perf_event_context *ctx)
7503{
7504 struct perf_event *parent = event->parent;
7505
7506 if (WARN_ON_ONCE(!parent))
7507 return;
7508
7509 mutex_lock(&parent->child_mutex);
7510 list_del_init(&event->child_list);
7511 mutex_unlock(&parent->child_mutex);
7512
7513 put_event(parent);
7514
7515 perf_group_detach(event);
7516 list_del_event(event, ctx);
7517 free_event(event);
7518}
7519
7520/*
7521 * free an unexposed, unused context as created by inheritance by
7522 * perf_event_init_task below, used by fork() in case of fail.
7523 */
7524void perf_event_free_task(struct task_struct *task)
7525{
7526 struct perf_event_context *ctx;
7527 struct perf_event *event, *tmp;
7528 int ctxn;
7529
7530 for_each_task_context_nr(ctxn) {
7531 ctx = task->perf_event_ctxp[ctxn];
7532 if (!ctx)
7533 continue;
7534
7535 mutex_lock(&ctx->mutex);
7536again:
7537 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7538 group_entry)
7539 perf_free_event(event, ctx);
7540
7541 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7542 group_entry)
7543 perf_free_event(event, ctx);
7544
7545 if (!list_empty(&ctx->pinned_groups) ||
7546 !list_empty(&ctx->flexible_groups))
7547 goto again;
7548
7549 mutex_unlock(&ctx->mutex);
7550
7551 put_ctx(ctx);
7552 }
7553}
7554
7555void perf_event_delayed_put(struct task_struct *task)
7556{
7557 int ctxn;
7558
7559 for_each_task_context_nr(ctxn)
7560 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7561}
7562
7563/*
7564 * inherit a event from parent task to child task:
7565 */
7566static struct perf_event *
7567inherit_event(struct perf_event *parent_event,
7568 struct task_struct *parent,
7569 struct perf_event_context *parent_ctx,
7570 struct task_struct *child,
7571 struct perf_event *group_leader,
7572 struct perf_event_context *child_ctx)
7573{
7574 struct perf_event *child_event;
7575 unsigned long flags;
7576
7577 /*
7578 * Instead of creating recursive hierarchies of events,
7579 * we link inherited events back to the original parent,
7580 * which has a filp for sure, which we use as the reference
7581 * count:
7582 */
7583 if (parent_event->parent)
7584 parent_event = parent_event->parent;
7585
7586 child_event = perf_event_alloc(&parent_event->attr,
7587 parent_event->cpu,
7588 child,
7589 group_leader, parent_event,
7590 NULL, NULL);
7591 if (IS_ERR(child_event))
7592 return child_event;
7593
7594 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7595 free_event(child_event);
7596 return NULL;
7597 }
7598
7599 get_ctx(child_ctx);
7600
7601 /*
7602 * Make the child state follow the state of the parent event,
7603 * not its attr.disabled bit. We hold the parent's mutex,
7604 * so we won't race with perf_event_{en, dis}able_family.
7605 */
7606 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7607 child_event->state = PERF_EVENT_STATE_INACTIVE;
7608 else
7609 child_event->state = PERF_EVENT_STATE_OFF;
7610
7611 if (parent_event->attr.freq) {
7612 u64 sample_period = parent_event->hw.sample_period;
7613 struct hw_perf_event *hwc = &child_event->hw;
7614
7615 hwc->sample_period = sample_period;
7616 hwc->last_period = sample_period;
7617
7618 local64_set(&hwc->period_left, sample_period);
7619 }
7620
7621 child_event->ctx = child_ctx;
7622 child_event->overflow_handler = parent_event->overflow_handler;
7623 child_event->overflow_handler_context
7624 = parent_event->overflow_handler_context;
7625
7626 /*
7627 * Precalculate sample_data sizes
7628 */
7629 perf_event__header_size(child_event);
7630 perf_event__id_header_size(child_event);
7631
7632 /*
7633 * Link it up in the child's context:
7634 */
7635 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7636 add_event_to_ctx(child_event, child_ctx);
7637 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7638
7639 /*
7640 * Link this into the parent event's child list
7641 */
7642 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7643 mutex_lock(&parent_event->child_mutex);
7644 list_add_tail(&child_event->child_list, &parent_event->child_list);
7645 mutex_unlock(&parent_event->child_mutex);
7646
7647 return child_event;
7648}
7649
7650static int inherit_group(struct perf_event *parent_event,
7651 struct task_struct *parent,
7652 struct perf_event_context *parent_ctx,
7653 struct task_struct *child,
7654 struct perf_event_context *child_ctx)
7655{
7656 struct perf_event *leader;
7657 struct perf_event *sub;
7658 struct perf_event *child_ctr;
7659
7660 leader = inherit_event(parent_event, parent, parent_ctx,
7661 child, NULL, child_ctx);
7662 if (IS_ERR(leader))
7663 return PTR_ERR(leader);
7664 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7665 child_ctr = inherit_event(sub, parent, parent_ctx,
7666 child, leader, child_ctx);
7667 if (IS_ERR(child_ctr))
7668 return PTR_ERR(child_ctr);
7669 }
7670 return 0;
7671}
7672
7673static int
7674inherit_task_group(struct perf_event *event, struct task_struct *parent,
7675 struct perf_event_context *parent_ctx,
7676 struct task_struct *child, int ctxn,
7677 int *inherited_all)
7678{
7679 int ret;
7680 struct perf_event_context *child_ctx;
7681
7682 if (!event->attr.inherit) {
7683 *inherited_all = 0;
7684 return 0;
7685 }
7686
7687 child_ctx = child->perf_event_ctxp[ctxn];
7688 if (!child_ctx) {
7689 /*
7690 * This is executed from the parent task context, so
7691 * inherit events that have been marked for cloning.
7692 * First allocate and initialize a context for the
7693 * child.
7694 */
7695
7696 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7697 if (!child_ctx)
7698 return -ENOMEM;
7699
7700 child->perf_event_ctxp[ctxn] = child_ctx;
7701 }
7702
7703 ret = inherit_group(event, parent, parent_ctx,
7704 child, child_ctx);
7705
7706 if (ret)
7707 *inherited_all = 0;
7708
7709 return ret;
7710}
7711
7712/*
7713 * Initialize the perf_event context in task_struct
7714 */
7715int perf_event_init_context(struct task_struct *child, int ctxn)
7716{
7717 struct perf_event_context *child_ctx, *parent_ctx;
7718 struct perf_event_context *cloned_ctx;
7719 struct perf_event *event;
7720 struct task_struct *parent = current;
7721 int inherited_all = 1;
7722 unsigned long flags;
7723 int ret = 0;
7724
7725 if (likely(!parent->perf_event_ctxp[ctxn]))
7726 return 0;
7727
7728 /*
7729 * If the parent's context is a clone, pin it so it won't get
7730 * swapped under us.
7731 */
7732 parent_ctx = perf_pin_task_context(parent, ctxn);
7733 if (!parent_ctx)
7734 return 0;
7735
7736 /*
7737 * No need to check if parent_ctx != NULL here; since we saw
7738 * it non-NULL earlier, the only reason for it to become NULL
7739 * is if we exit, and since we're currently in the middle of
7740 * a fork we can't be exiting at the same time.
7741 */
7742
7743 /*
7744 * Lock the parent list. No need to lock the child - not PID
7745 * hashed yet and not running, so nobody can access it.
7746 */
7747 mutex_lock(&parent_ctx->mutex);
7748
7749 /*
7750 * We dont have to disable NMIs - we are only looking at
7751 * the list, not manipulating it:
7752 */
7753 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7754 ret = inherit_task_group(event, parent, parent_ctx,
7755 child, ctxn, &inherited_all);
7756 if (ret)
7757 break;
7758 }
7759
7760 /*
7761 * We can't hold ctx->lock when iterating the ->flexible_group list due
7762 * to allocations, but we need to prevent rotation because
7763 * rotate_ctx() will change the list from interrupt context.
7764 */
7765 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7766 parent_ctx->rotate_disable = 1;
7767 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7768
7769 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7770 ret = inherit_task_group(event, parent, parent_ctx,
7771 child, ctxn, &inherited_all);
7772 if (ret)
7773 break;
7774 }
7775
7776 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7777 parent_ctx->rotate_disable = 0;
7778
7779 child_ctx = child->perf_event_ctxp[ctxn];
7780
7781 if (child_ctx && inherited_all) {
7782 /*
7783 * Mark the child context as a clone of the parent
7784 * context, or of whatever the parent is a clone of.
7785 *
7786 * Note that if the parent is a clone, the holding of
7787 * parent_ctx->lock avoids it from being uncloned.
7788 */
7789 cloned_ctx = parent_ctx->parent_ctx;
7790 if (cloned_ctx) {
7791 child_ctx->parent_ctx = cloned_ctx;
7792 child_ctx->parent_gen = parent_ctx->parent_gen;
7793 } else {
7794 child_ctx->parent_ctx = parent_ctx;
7795 child_ctx->parent_gen = parent_ctx->generation;
7796 }
7797 get_ctx(child_ctx->parent_ctx);
7798 }
7799
7800 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7801 mutex_unlock(&parent_ctx->mutex);
7802
7803 perf_unpin_context(parent_ctx);
7804 put_ctx(parent_ctx);
7805
7806 return ret;
7807}
7808
7809/*
7810 * Initialize the perf_event context in task_struct
7811 */
7812int perf_event_init_task(struct task_struct *child)
7813{
7814 int ctxn, ret;
7815
7816 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7817 mutex_init(&child->perf_event_mutex);
7818 INIT_LIST_HEAD(&child->perf_event_list);
7819
7820 for_each_task_context_nr(ctxn) {
7821 ret = perf_event_init_context(child, ctxn);
7822 if (ret)
7823 return ret;
7824 }
7825
7826 return 0;
7827}
7828
7829static void __init perf_event_init_all_cpus(void)
7830{
7831 struct swevent_htable *swhash;
7832 int cpu;
7833
7834 for_each_possible_cpu(cpu) {
7835 swhash = &per_cpu(swevent_htable, cpu);
7836 mutex_init(&swhash->hlist_mutex);
7837 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7838 }
7839}
7840
7841static void perf_event_init_cpu(int cpu)
7842{
7843 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7844
7845 mutex_lock(&swhash->hlist_mutex);
7846 swhash->online = true;
7847 if (swhash->hlist_refcount > 0) {
7848 struct swevent_hlist *hlist;
7849
7850 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7851 WARN_ON(!hlist);
7852 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7853 }
7854 mutex_unlock(&swhash->hlist_mutex);
7855}
7856
7857#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7858static void perf_pmu_rotate_stop(struct pmu *pmu)
7859{
7860 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7861
7862 WARN_ON(!irqs_disabled());
7863
7864 list_del_init(&cpuctx->rotation_list);
7865}
7866
7867static void __perf_event_exit_context(void *__info)
7868{
7869 struct remove_event re = { .detach_group = false };
7870 struct perf_event_context *ctx = __info;
7871
7872 perf_pmu_rotate_stop(ctx->pmu);
7873
7874 rcu_read_lock();
7875 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
7876 __perf_remove_from_context(&re);
7877 rcu_read_unlock();
7878}
7879
7880static void perf_event_exit_cpu_context(int cpu)
7881{
7882 struct perf_event_context *ctx;
7883 struct pmu *pmu;
7884 int idx;
7885
7886 idx = srcu_read_lock(&pmus_srcu);
7887 list_for_each_entry_rcu(pmu, &pmus, entry) {
7888 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7889
7890 mutex_lock(&ctx->mutex);
7891 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7892 mutex_unlock(&ctx->mutex);
7893 }
7894 srcu_read_unlock(&pmus_srcu, idx);
7895}
7896
7897static void perf_event_exit_cpu(int cpu)
7898{
7899 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7900
7901 perf_event_exit_cpu_context(cpu);
7902
7903 mutex_lock(&swhash->hlist_mutex);
7904 swhash->online = false;
7905 swevent_hlist_release(swhash);
7906 mutex_unlock(&swhash->hlist_mutex);
7907}
7908#else
7909static inline void perf_event_exit_cpu(int cpu) { }
7910#endif
7911
7912static int
7913perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7914{
7915 int cpu;
7916
7917 for_each_online_cpu(cpu)
7918 perf_event_exit_cpu(cpu);
7919
7920 return NOTIFY_OK;
7921}
7922
7923/*
7924 * Run the perf reboot notifier at the very last possible moment so that
7925 * the generic watchdog code runs as long as possible.
7926 */
7927static struct notifier_block perf_reboot_notifier = {
7928 .notifier_call = perf_reboot,
7929 .priority = INT_MIN,
7930};
7931
7932static int
7933perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7934{
7935 unsigned int cpu = (long)hcpu;
7936
7937 switch (action & ~CPU_TASKS_FROZEN) {
7938
7939 case CPU_UP_PREPARE:
7940 case CPU_DOWN_FAILED:
7941 perf_event_init_cpu(cpu);
7942 break;
7943
7944 case CPU_UP_CANCELED:
7945 case CPU_DOWN_PREPARE:
7946 perf_event_exit_cpu(cpu);
7947 break;
7948 default:
7949 break;
7950 }
7951
7952 return NOTIFY_OK;
7953}
7954
7955void __init perf_event_init(void)
7956{
7957 int ret;
7958
7959 idr_init(&pmu_idr);
7960
7961 perf_event_init_all_cpus();
7962 init_srcu_struct(&pmus_srcu);
7963 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7964 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7965 perf_pmu_register(&perf_task_clock, NULL, -1);
7966 perf_tp_register();
7967 perf_cpu_notifier(perf_cpu_notify);
7968 register_reboot_notifier(&perf_reboot_notifier);
7969
7970 ret = init_hw_breakpoint();
7971 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7972
7973 /* do not patch jump label more than once per second */
7974 jump_label_rate_limit(&perf_sched_events, HZ);
7975
7976 /*
7977 * Build time assertion that we keep the data_head at the intended
7978 * location. IOW, validation we got the __reserved[] size right.
7979 */
7980 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7981 != 1024);
7982}
7983
7984static int __init perf_event_sysfs_init(void)
7985{
7986 struct pmu *pmu;
7987 int ret;
7988
7989 mutex_lock(&pmus_lock);
7990
7991 ret = bus_register(&pmu_bus);
7992 if (ret)
7993 goto unlock;
7994
7995 list_for_each_entry(pmu, &pmus, entry) {
7996 if (!pmu->name || pmu->type < 0)
7997 continue;
7998
7999 ret = pmu_dev_alloc(pmu);
8000 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8001 }
8002 pmu_bus_running = 1;
8003 ret = 0;
8004
8005unlock:
8006 mutex_unlock(&pmus_lock);
8007
8008 return ret;
8009}
8010device_initcall(perf_event_sysfs_init);
8011
8012#ifdef CONFIG_CGROUP_PERF
8013static struct cgroup_subsys_state *
8014perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8015{
8016 struct perf_cgroup *jc;
8017
8018 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8019 if (!jc)
8020 return ERR_PTR(-ENOMEM);
8021
8022 jc->info = alloc_percpu(struct perf_cgroup_info);
8023 if (!jc->info) {
8024 kfree(jc);
8025 return ERR_PTR(-ENOMEM);
8026 }
8027
8028 return &jc->css;
8029}
8030
8031static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8032{
8033 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8034
8035 free_percpu(jc->info);
8036 kfree(jc);
8037}
8038
8039static int __perf_cgroup_move(void *info)
8040{
8041 struct task_struct *task = info;
8042 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8043 return 0;
8044}
8045
8046static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8047 struct cgroup_taskset *tset)
8048{
8049 struct task_struct *task;
8050
8051 cgroup_taskset_for_each(task, tset)
8052 task_function_call(task, __perf_cgroup_move, task);
8053}
8054
8055static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8056 struct cgroup_subsys_state *old_css,
8057 struct task_struct *task)
8058{
8059 /*
8060 * cgroup_exit() is called in the copy_process() failure path.
8061 * Ignore this case since the task hasn't ran yet, this avoids
8062 * trying to poke a half freed task state from generic code.
8063 */
8064 if (!(task->flags & PF_EXITING))
8065 return;
8066
8067 task_function_call(task, __perf_cgroup_move, task);
8068}
8069
8070struct cgroup_subsys perf_event_cgrp_subsys = {
8071 .css_alloc = perf_cgroup_css_alloc,
8072 .css_free = perf_cgroup_css_free,
8073 .exit = perf_cgroup_exit,
8074 .attach = perf_cgroup_attach,
8075};
8076#endif /* CONFIG_CGROUP_PERF */
1/*
2 * Performance events core code:
3 *
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 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/idr.h>
17#include <linux/file.h>
18#include <linux/poll.h>
19#include <linux/slab.h>
20#include <linux/hash.h>
21#include <linux/sysfs.h>
22#include <linux/dcache.h>
23#include <linux/percpu.h>
24#include <linux/ptrace.h>
25#include <linux/reboot.h>
26#include <linux/vmstat.h>
27#include <linux/device.h>
28#include <linux/vmalloc.h>
29#include <linux/hardirq.h>
30#include <linux/rculist.h>
31#include <linux/uaccess.h>
32#include <linux/syscalls.h>
33#include <linux/anon_inodes.h>
34#include <linux/kernel_stat.h>
35#include <linux/perf_event.h>
36#include <linux/ftrace_event.h>
37#include <linux/hw_breakpoint.h>
38
39#include "internal.h"
40
41#include <asm/irq_regs.h>
42
43struct remote_function_call {
44 struct task_struct *p;
45 int (*func)(void *info);
46 void *info;
47 int ret;
48};
49
50static void remote_function(void *data)
51{
52 struct remote_function_call *tfc = data;
53 struct task_struct *p = tfc->p;
54
55 if (p) {
56 tfc->ret = -EAGAIN;
57 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
58 return;
59 }
60
61 tfc->ret = tfc->func(tfc->info);
62}
63
64/**
65 * task_function_call - call a function on the cpu on which a task runs
66 * @p: the task to evaluate
67 * @func: the function to be called
68 * @info: the function call argument
69 *
70 * Calls the function @func when the task is currently running. This might
71 * be on the current CPU, which just calls the function directly
72 *
73 * returns: @func return value, or
74 * -ESRCH - when the process isn't running
75 * -EAGAIN - when the process moved away
76 */
77static int
78task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
79{
80 struct remote_function_call data = {
81 .p = p,
82 .func = func,
83 .info = info,
84 .ret = -ESRCH, /* No such (running) process */
85 };
86
87 if (task_curr(p))
88 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
89
90 return data.ret;
91}
92
93/**
94 * cpu_function_call - call a function on the cpu
95 * @func: the function to be called
96 * @info: the function call argument
97 *
98 * Calls the function @func on the remote cpu.
99 *
100 * returns: @func return value or -ENXIO when the cpu is offline
101 */
102static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
103{
104 struct remote_function_call data = {
105 .p = NULL,
106 .func = func,
107 .info = info,
108 .ret = -ENXIO, /* No such CPU */
109 };
110
111 smp_call_function_single(cpu, remote_function, &data, 1);
112
113 return data.ret;
114}
115
116#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
117 PERF_FLAG_FD_OUTPUT |\
118 PERF_FLAG_PID_CGROUP)
119
120enum event_type_t {
121 EVENT_FLEXIBLE = 0x1,
122 EVENT_PINNED = 0x2,
123 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
124};
125
126/*
127 * perf_sched_events : >0 events exist
128 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
129 */
130struct jump_label_key perf_sched_events __read_mostly;
131static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
132
133static atomic_t nr_mmap_events __read_mostly;
134static atomic_t nr_comm_events __read_mostly;
135static atomic_t nr_task_events __read_mostly;
136
137static LIST_HEAD(pmus);
138static DEFINE_MUTEX(pmus_lock);
139static struct srcu_struct pmus_srcu;
140
141/*
142 * perf event paranoia level:
143 * -1 - not paranoid at all
144 * 0 - disallow raw tracepoint access for unpriv
145 * 1 - disallow cpu events for unpriv
146 * 2 - disallow kernel profiling for unpriv
147 */
148int sysctl_perf_event_paranoid __read_mostly = 1;
149
150/* Minimum for 512 kiB + 1 user control page */
151int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
152
153/*
154 * max perf event sample rate
155 */
156#define DEFAULT_MAX_SAMPLE_RATE 100000
157int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
158static int max_samples_per_tick __read_mostly =
159 DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
160
161int perf_proc_update_handler(struct ctl_table *table, int write,
162 void __user *buffer, size_t *lenp,
163 loff_t *ppos)
164{
165 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
166
167 if (ret || !write)
168 return ret;
169
170 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
171
172 return 0;
173}
174
175static atomic64_t perf_event_id;
176
177static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
178 enum event_type_t event_type);
179
180static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
181 enum event_type_t event_type,
182 struct task_struct *task);
183
184static void update_context_time(struct perf_event_context *ctx);
185static u64 perf_event_time(struct perf_event *event);
186
187void __weak perf_event_print_debug(void) { }
188
189extern __weak const char *perf_pmu_name(void)
190{
191 return "pmu";
192}
193
194static inline u64 perf_clock(void)
195{
196 return local_clock();
197}
198
199static inline struct perf_cpu_context *
200__get_cpu_context(struct perf_event_context *ctx)
201{
202 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
203}
204
205static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
206 struct perf_event_context *ctx)
207{
208 raw_spin_lock(&cpuctx->ctx.lock);
209 if (ctx)
210 raw_spin_lock(&ctx->lock);
211}
212
213static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
214 struct perf_event_context *ctx)
215{
216 if (ctx)
217 raw_spin_unlock(&ctx->lock);
218 raw_spin_unlock(&cpuctx->ctx.lock);
219}
220
221#ifdef CONFIG_CGROUP_PERF
222
223/*
224 * Must ensure cgroup is pinned (css_get) before calling
225 * this function. In other words, we cannot call this function
226 * if there is no cgroup event for the current CPU context.
227 */
228static inline struct perf_cgroup *
229perf_cgroup_from_task(struct task_struct *task)
230{
231 return container_of(task_subsys_state(task, perf_subsys_id),
232 struct perf_cgroup, css);
233}
234
235static inline bool
236perf_cgroup_match(struct perf_event *event)
237{
238 struct perf_event_context *ctx = event->ctx;
239 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
240
241 return !event->cgrp || event->cgrp == cpuctx->cgrp;
242}
243
244static inline void perf_get_cgroup(struct perf_event *event)
245{
246 css_get(&event->cgrp->css);
247}
248
249static inline void perf_put_cgroup(struct perf_event *event)
250{
251 css_put(&event->cgrp->css);
252}
253
254static inline void perf_detach_cgroup(struct perf_event *event)
255{
256 perf_put_cgroup(event);
257 event->cgrp = NULL;
258}
259
260static inline int is_cgroup_event(struct perf_event *event)
261{
262 return event->cgrp != NULL;
263}
264
265static inline u64 perf_cgroup_event_time(struct perf_event *event)
266{
267 struct perf_cgroup_info *t;
268
269 t = per_cpu_ptr(event->cgrp->info, event->cpu);
270 return t->time;
271}
272
273static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
274{
275 struct perf_cgroup_info *info;
276 u64 now;
277
278 now = perf_clock();
279
280 info = this_cpu_ptr(cgrp->info);
281
282 info->time += now - info->timestamp;
283 info->timestamp = now;
284}
285
286static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
287{
288 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
289 if (cgrp_out)
290 __update_cgrp_time(cgrp_out);
291}
292
293static inline void update_cgrp_time_from_event(struct perf_event *event)
294{
295 struct perf_cgroup *cgrp;
296
297 /*
298 * ensure we access cgroup data only when needed and
299 * when we know the cgroup is pinned (css_get)
300 */
301 if (!is_cgroup_event(event))
302 return;
303
304 cgrp = perf_cgroup_from_task(current);
305 /*
306 * Do not update time when cgroup is not active
307 */
308 if (cgrp == event->cgrp)
309 __update_cgrp_time(event->cgrp);
310}
311
312static inline void
313perf_cgroup_set_timestamp(struct task_struct *task,
314 struct perf_event_context *ctx)
315{
316 struct perf_cgroup *cgrp;
317 struct perf_cgroup_info *info;
318
319 /*
320 * ctx->lock held by caller
321 * ensure we do not access cgroup data
322 * unless we have the cgroup pinned (css_get)
323 */
324 if (!task || !ctx->nr_cgroups)
325 return;
326
327 cgrp = perf_cgroup_from_task(task);
328 info = this_cpu_ptr(cgrp->info);
329 info->timestamp = ctx->timestamp;
330}
331
332#define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
333#define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
334
335/*
336 * reschedule events based on the cgroup constraint of task.
337 *
338 * mode SWOUT : schedule out everything
339 * mode SWIN : schedule in based on cgroup for next
340 */
341void perf_cgroup_switch(struct task_struct *task, int mode)
342{
343 struct perf_cpu_context *cpuctx;
344 struct pmu *pmu;
345 unsigned long flags;
346
347 /*
348 * disable interrupts to avoid geting nr_cgroup
349 * changes via __perf_event_disable(). Also
350 * avoids preemption.
351 */
352 local_irq_save(flags);
353
354 /*
355 * we reschedule only in the presence of cgroup
356 * constrained events.
357 */
358 rcu_read_lock();
359
360 list_for_each_entry_rcu(pmu, &pmus, entry) {
361 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
362
363 /*
364 * perf_cgroup_events says at least one
365 * context on this CPU has cgroup events.
366 *
367 * ctx->nr_cgroups reports the number of cgroup
368 * events for a context.
369 */
370 if (cpuctx->ctx.nr_cgroups > 0) {
371 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
372 perf_pmu_disable(cpuctx->ctx.pmu);
373
374 if (mode & PERF_CGROUP_SWOUT) {
375 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
376 /*
377 * must not be done before ctxswout due
378 * to event_filter_match() in event_sched_out()
379 */
380 cpuctx->cgrp = NULL;
381 }
382
383 if (mode & PERF_CGROUP_SWIN) {
384 WARN_ON_ONCE(cpuctx->cgrp);
385 /* set cgrp before ctxsw in to
386 * allow event_filter_match() to not
387 * have to pass task around
388 */
389 cpuctx->cgrp = perf_cgroup_from_task(task);
390 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
391 }
392 perf_pmu_enable(cpuctx->ctx.pmu);
393 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
394 }
395 }
396
397 rcu_read_unlock();
398
399 local_irq_restore(flags);
400}
401
402static inline void perf_cgroup_sched_out(struct task_struct *task,
403 struct task_struct *next)
404{
405 struct perf_cgroup *cgrp1;
406 struct perf_cgroup *cgrp2 = NULL;
407
408 /*
409 * we come here when we know perf_cgroup_events > 0
410 */
411 cgrp1 = perf_cgroup_from_task(task);
412
413 /*
414 * next is NULL when called from perf_event_enable_on_exec()
415 * that will systematically cause a cgroup_switch()
416 */
417 if (next)
418 cgrp2 = perf_cgroup_from_task(next);
419
420 /*
421 * only schedule out current cgroup events if we know
422 * that we are switching to a different cgroup. Otherwise,
423 * do no touch the cgroup events.
424 */
425 if (cgrp1 != cgrp2)
426 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
427}
428
429static inline void perf_cgroup_sched_in(struct task_struct *prev,
430 struct task_struct *task)
431{
432 struct perf_cgroup *cgrp1;
433 struct perf_cgroup *cgrp2 = NULL;
434
435 /*
436 * we come here when we know perf_cgroup_events > 0
437 */
438 cgrp1 = perf_cgroup_from_task(task);
439
440 /* prev can never be NULL */
441 cgrp2 = perf_cgroup_from_task(prev);
442
443 /*
444 * only need to schedule in cgroup events if we are changing
445 * cgroup during ctxsw. Cgroup events were not scheduled
446 * out of ctxsw out if that was not the case.
447 */
448 if (cgrp1 != cgrp2)
449 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
450}
451
452static inline int perf_cgroup_connect(int fd, struct perf_event *event,
453 struct perf_event_attr *attr,
454 struct perf_event *group_leader)
455{
456 struct perf_cgroup *cgrp;
457 struct cgroup_subsys_state *css;
458 struct file *file;
459 int ret = 0, fput_needed;
460
461 file = fget_light(fd, &fput_needed);
462 if (!file)
463 return -EBADF;
464
465 css = cgroup_css_from_dir(file, perf_subsys_id);
466 if (IS_ERR(css)) {
467 ret = PTR_ERR(css);
468 goto out;
469 }
470
471 cgrp = container_of(css, struct perf_cgroup, css);
472 event->cgrp = cgrp;
473
474 /* must be done before we fput() the file */
475 perf_get_cgroup(event);
476
477 /*
478 * all events in a group must monitor
479 * the same cgroup because a task belongs
480 * to only one perf cgroup at a time
481 */
482 if (group_leader && group_leader->cgrp != cgrp) {
483 perf_detach_cgroup(event);
484 ret = -EINVAL;
485 }
486out:
487 fput_light(file, fput_needed);
488 return ret;
489}
490
491static inline void
492perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
493{
494 struct perf_cgroup_info *t;
495 t = per_cpu_ptr(event->cgrp->info, event->cpu);
496 event->shadow_ctx_time = now - t->timestamp;
497}
498
499static inline void
500perf_cgroup_defer_enabled(struct perf_event *event)
501{
502 /*
503 * when the current task's perf cgroup does not match
504 * the event's, we need to remember to call the
505 * perf_mark_enable() function the first time a task with
506 * a matching perf cgroup is scheduled in.
507 */
508 if (is_cgroup_event(event) && !perf_cgroup_match(event))
509 event->cgrp_defer_enabled = 1;
510}
511
512static inline void
513perf_cgroup_mark_enabled(struct perf_event *event,
514 struct perf_event_context *ctx)
515{
516 struct perf_event *sub;
517 u64 tstamp = perf_event_time(event);
518
519 if (!event->cgrp_defer_enabled)
520 return;
521
522 event->cgrp_defer_enabled = 0;
523
524 event->tstamp_enabled = tstamp - event->total_time_enabled;
525 list_for_each_entry(sub, &event->sibling_list, group_entry) {
526 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
527 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
528 sub->cgrp_defer_enabled = 0;
529 }
530 }
531}
532#else /* !CONFIG_CGROUP_PERF */
533
534static inline bool
535perf_cgroup_match(struct perf_event *event)
536{
537 return true;
538}
539
540static inline void perf_detach_cgroup(struct perf_event *event)
541{}
542
543static inline int is_cgroup_event(struct perf_event *event)
544{
545 return 0;
546}
547
548static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
549{
550 return 0;
551}
552
553static inline void update_cgrp_time_from_event(struct perf_event *event)
554{
555}
556
557static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
558{
559}
560
561static inline void perf_cgroup_sched_out(struct task_struct *task,
562 struct task_struct *next)
563{
564}
565
566static inline void perf_cgroup_sched_in(struct task_struct *prev,
567 struct task_struct *task)
568{
569}
570
571static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
572 struct perf_event_attr *attr,
573 struct perf_event *group_leader)
574{
575 return -EINVAL;
576}
577
578static inline void
579perf_cgroup_set_timestamp(struct task_struct *task,
580 struct perf_event_context *ctx)
581{
582}
583
584void
585perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
586{
587}
588
589static inline void
590perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
591{
592}
593
594static inline u64 perf_cgroup_event_time(struct perf_event *event)
595{
596 return 0;
597}
598
599static inline void
600perf_cgroup_defer_enabled(struct perf_event *event)
601{
602}
603
604static inline void
605perf_cgroup_mark_enabled(struct perf_event *event,
606 struct perf_event_context *ctx)
607{
608}
609#endif
610
611void perf_pmu_disable(struct pmu *pmu)
612{
613 int *count = this_cpu_ptr(pmu->pmu_disable_count);
614 if (!(*count)++)
615 pmu->pmu_disable(pmu);
616}
617
618void perf_pmu_enable(struct pmu *pmu)
619{
620 int *count = this_cpu_ptr(pmu->pmu_disable_count);
621 if (!--(*count))
622 pmu->pmu_enable(pmu);
623}
624
625static DEFINE_PER_CPU(struct list_head, rotation_list);
626
627/*
628 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
629 * because they're strictly cpu affine and rotate_start is called with IRQs
630 * disabled, while rotate_context is called from IRQ context.
631 */
632static void perf_pmu_rotate_start(struct pmu *pmu)
633{
634 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
635 struct list_head *head = &__get_cpu_var(rotation_list);
636
637 WARN_ON(!irqs_disabled());
638
639 if (list_empty(&cpuctx->rotation_list))
640 list_add(&cpuctx->rotation_list, head);
641}
642
643static void get_ctx(struct perf_event_context *ctx)
644{
645 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
646}
647
648static void put_ctx(struct perf_event_context *ctx)
649{
650 if (atomic_dec_and_test(&ctx->refcount)) {
651 if (ctx->parent_ctx)
652 put_ctx(ctx->parent_ctx);
653 if (ctx->task)
654 put_task_struct(ctx->task);
655 kfree_rcu(ctx, rcu_head);
656 }
657}
658
659static void unclone_ctx(struct perf_event_context *ctx)
660{
661 if (ctx->parent_ctx) {
662 put_ctx(ctx->parent_ctx);
663 ctx->parent_ctx = NULL;
664 }
665}
666
667static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
668{
669 /*
670 * only top level events have the pid namespace they were created in
671 */
672 if (event->parent)
673 event = event->parent;
674
675 return task_tgid_nr_ns(p, event->ns);
676}
677
678static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
679{
680 /*
681 * only top level events have the pid namespace they were created in
682 */
683 if (event->parent)
684 event = event->parent;
685
686 return task_pid_nr_ns(p, event->ns);
687}
688
689/*
690 * If we inherit events we want to return the parent event id
691 * to userspace.
692 */
693static u64 primary_event_id(struct perf_event *event)
694{
695 u64 id = event->id;
696
697 if (event->parent)
698 id = event->parent->id;
699
700 return id;
701}
702
703/*
704 * Get the perf_event_context for a task and lock it.
705 * This has to cope with with the fact that until it is locked,
706 * the context could get moved to another task.
707 */
708static struct perf_event_context *
709perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
710{
711 struct perf_event_context *ctx;
712
713 rcu_read_lock();
714retry:
715 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
716 if (ctx) {
717 /*
718 * If this context is a clone of another, it might
719 * get swapped for another underneath us by
720 * perf_event_task_sched_out, though the
721 * rcu_read_lock() protects us from any context
722 * getting freed. Lock the context and check if it
723 * got swapped before we could get the lock, and retry
724 * if so. If we locked the right context, then it
725 * can't get swapped on us any more.
726 */
727 raw_spin_lock_irqsave(&ctx->lock, *flags);
728 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
729 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
730 goto retry;
731 }
732
733 if (!atomic_inc_not_zero(&ctx->refcount)) {
734 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
735 ctx = NULL;
736 }
737 }
738 rcu_read_unlock();
739 return ctx;
740}
741
742/*
743 * Get the context for a task and increment its pin_count so it
744 * can't get swapped to another task. This also increments its
745 * reference count so that the context can't get freed.
746 */
747static struct perf_event_context *
748perf_pin_task_context(struct task_struct *task, int ctxn)
749{
750 struct perf_event_context *ctx;
751 unsigned long flags;
752
753 ctx = perf_lock_task_context(task, ctxn, &flags);
754 if (ctx) {
755 ++ctx->pin_count;
756 raw_spin_unlock_irqrestore(&ctx->lock, flags);
757 }
758 return ctx;
759}
760
761static void perf_unpin_context(struct perf_event_context *ctx)
762{
763 unsigned long flags;
764
765 raw_spin_lock_irqsave(&ctx->lock, flags);
766 --ctx->pin_count;
767 raw_spin_unlock_irqrestore(&ctx->lock, flags);
768}
769
770/*
771 * Update the record of the current time in a context.
772 */
773static void update_context_time(struct perf_event_context *ctx)
774{
775 u64 now = perf_clock();
776
777 ctx->time += now - ctx->timestamp;
778 ctx->timestamp = now;
779}
780
781static u64 perf_event_time(struct perf_event *event)
782{
783 struct perf_event_context *ctx = event->ctx;
784
785 if (is_cgroup_event(event))
786 return perf_cgroup_event_time(event);
787
788 return ctx ? ctx->time : 0;
789}
790
791/*
792 * Update the total_time_enabled and total_time_running fields for a event.
793 * The caller of this function needs to hold the ctx->lock.
794 */
795static void update_event_times(struct perf_event *event)
796{
797 struct perf_event_context *ctx = event->ctx;
798 u64 run_end;
799
800 if (event->state < PERF_EVENT_STATE_INACTIVE ||
801 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
802 return;
803 /*
804 * in cgroup mode, time_enabled represents
805 * the time the event was enabled AND active
806 * tasks were in the monitored cgroup. This is
807 * independent of the activity of the context as
808 * there may be a mix of cgroup and non-cgroup events.
809 *
810 * That is why we treat cgroup events differently
811 * here.
812 */
813 if (is_cgroup_event(event))
814 run_end = perf_event_time(event);
815 else if (ctx->is_active)
816 run_end = ctx->time;
817 else
818 run_end = event->tstamp_stopped;
819
820 event->total_time_enabled = run_end - event->tstamp_enabled;
821
822 if (event->state == PERF_EVENT_STATE_INACTIVE)
823 run_end = event->tstamp_stopped;
824 else
825 run_end = perf_event_time(event);
826
827 event->total_time_running = run_end - event->tstamp_running;
828
829}
830
831/*
832 * Update total_time_enabled and total_time_running for all events in a group.
833 */
834static void update_group_times(struct perf_event *leader)
835{
836 struct perf_event *event;
837
838 update_event_times(leader);
839 list_for_each_entry(event, &leader->sibling_list, group_entry)
840 update_event_times(event);
841}
842
843static struct list_head *
844ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
845{
846 if (event->attr.pinned)
847 return &ctx->pinned_groups;
848 else
849 return &ctx->flexible_groups;
850}
851
852/*
853 * Add a event from the lists for its context.
854 * Must be called with ctx->mutex and ctx->lock held.
855 */
856static void
857list_add_event(struct perf_event *event, struct perf_event_context *ctx)
858{
859 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
860 event->attach_state |= PERF_ATTACH_CONTEXT;
861
862 /*
863 * If we're a stand alone event or group leader, we go to the context
864 * list, group events are kept attached to the group so that
865 * perf_group_detach can, at all times, locate all siblings.
866 */
867 if (event->group_leader == event) {
868 struct list_head *list;
869
870 if (is_software_event(event))
871 event->group_flags |= PERF_GROUP_SOFTWARE;
872
873 list = ctx_group_list(event, ctx);
874 list_add_tail(&event->group_entry, list);
875 }
876
877 if (is_cgroup_event(event))
878 ctx->nr_cgroups++;
879
880 list_add_rcu(&event->event_entry, &ctx->event_list);
881 if (!ctx->nr_events)
882 perf_pmu_rotate_start(ctx->pmu);
883 ctx->nr_events++;
884 if (event->attr.inherit_stat)
885 ctx->nr_stat++;
886}
887
888/*
889 * Called at perf_event creation and when events are attached/detached from a
890 * group.
891 */
892static void perf_event__read_size(struct perf_event *event)
893{
894 int entry = sizeof(u64); /* value */
895 int size = 0;
896 int nr = 1;
897
898 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
899 size += sizeof(u64);
900
901 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
902 size += sizeof(u64);
903
904 if (event->attr.read_format & PERF_FORMAT_ID)
905 entry += sizeof(u64);
906
907 if (event->attr.read_format & PERF_FORMAT_GROUP) {
908 nr += event->group_leader->nr_siblings;
909 size += sizeof(u64);
910 }
911
912 size += entry * nr;
913 event->read_size = size;
914}
915
916static void perf_event__header_size(struct perf_event *event)
917{
918 struct perf_sample_data *data;
919 u64 sample_type = event->attr.sample_type;
920 u16 size = 0;
921
922 perf_event__read_size(event);
923
924 if (sample_type & PERF_SAMPLE_IP)
925 size += sizeof(data->ip);
926
927 if (sample_type & PERF_SAMPLE_ADDR)
928 size += sizeof(data->addr);
929
930 if (sample_type & PERF_SAMPLE_PERIOD)
931 size += sizeof(data->period);
932
933 if (sample_type & PERF_SAMPLE_READ)
934 size += event->read_size;
935
936 event->header_size = size;
937}
938
939static void perf_event__id_header_size(struct perf_event *event)
940{
941 struct perf_sample_data *data;
942 u64 sample_type = event->attr.sample_type;
943 u16 size = 0;
944
945 if (sample_type & PERF_SAMPLE_TID)
946 size += sizeof(data->tid_entry);
947
948 if (sample_type & PERF_SAMPLE_TIME)
949 size += sizeof(data->time);
950
951 if (sample_type & PERF_SAMPLE_ID)
952 size += sizeof(data->id);
953
954 if (sample_type & PERF_SAMPLE_STREAM_ID)
955 size += sizeof(data->stream_id);
956
957 if (sample_type & PERF_SAMPLE_CPU)
958 size += sizeof(data->cpu_entry);
959
960 event->id_header_size = size;
961}
962
963static void perf_group_attach(struct perf_event *event)
964{
965 struct perf_event *group_leader = event->group_leader, *pos;
966
967 /*
968 * We can have double attach due to group movement in perf_event_open.
969 */
970 if (event->attach_state & PERF_ATTACH_GROUP)
971 return;
972
973 event->attach_state |= PERF_ATTACH_GROUP;
974
975 if (group_leader == event)
976 return;
977
978 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
979 !is_software_event(event))
980 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
981
982 list_add_tail(&event->group_entry, &group_leader->sibling_list);
983 group_leader->nr_siblings++;
984
985 perf_event__header_size(group_leader);
986
987 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
988 perf_event__header_size(pos);
989}
990
991/*
992 * Remove a event from the lists for its context.
993 * Must be called with ctx->mutex and ctx->lock held.
994 */
995static void
996list_del_event(struct perf_event *event, struct perf_event_context *ctx)
997{
998 struct perf_cpu_context *cpuctx;
999 /*
1000 * We can have double detach due to exit/hot-unplug + close.
1001 */
1002 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1003 return;
1004
1005 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1006
1007 if (is_cgroup_event(event)) {
1008 ctx->nr_cgroups--;
1009 cpuctx = __get_cpu_context(ctx);
1010 /*
1011 * if there are no more cgroup events
1012 * then cler cgrp to avoid stale pointer
1013 * in update_cgrp_time_from_cpuctx()
1014 */
1015 if (!ctx->nr_cgroups)
1016 cpuctx->cgrp = NULL;
1017 }
1018
1019 ctx->nr_events--;
1020 if (event->attr.inherit_stat)
1021 ctx->nr_stat--;
1022
1023 list_del_rcu(&event->event_entry);
1024
1025 if (event->group_leader == event)
1026 list_del_init(&event->group_entry);
1027
1028 update_group_times(event);
1029
1030 /*
1031 * If event was in error state, then keep it
1032 * that way, otherwise bogus counts will be
1033 * returned on read(). The only way to get out
1034 * of error state is by explicit re-enabling
1035 * of the event
1036 */
1037 if (event->state > PERF_EVENT_STATE_OFF)
1038 event->state = PERF_EVENT_STATE_OFF;
1039}
1040
1041static void perf_group_detach(struct perf_event *event)
1042{
1043 struct perf_event *sibling, *tmp;
1044 struct list_head *list = NULL;
1045
1046 /*
1047 * We can have double detach due to exit/hot-unplug + close.
1048 */
1049 if (!(event->attach_state & PERF_ATTACH_GROUP))
1050 return;
1051
1052 event->attach_state &= ~PERF_ATTACH_GROUP;
1053
1054 /*
1055 * If this is a sibling, remove it from its group.
1056 */
1057 if (event->group_leader != event) {
1058 list_del_init(&event->group_entry);
1059 event->group_leader->nr_siblings--;
1060 goto out;
1061 }
1062
1063 if (!list_empty(&event->group_entry))
1064 list = &event->group_entry;
1065
1066 /*
1067 * If this was a group event with sibling events then
1068 * upgrade the siblings to singleton events by adding them
1069 * to whatever list we are on.
1070 */
1071 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1072 if (list)
1073 list_move_tail(&sibling->group_entry, list);
1074 sibling->group_leader = sibling;
1075
1076 /* Inherit group flags from the previous leader */
1077 sibling->group_flags = event->group_flags;
1078 }
1079
1080out:
1081 perf_event__header_size(event->group_leader);
1082
1083 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1084 perf_event__header_size(tmp);
1085}
1086
1087static inline int
1088event_filter_match(struct perf_event *event)
1089{
1090 return (event->cpu == -1 || event->cpu == smp_processor_id())
1091 && perf_cgroup_match(event);
1092}
1093
1094static void
1095event_sched_out(struct perf_event *event,
1096 struct perf_cpu_context *cpuctx,
1097 struct perf_event_context *ctx)
1098{
1099 u64 tstamp = perf_event_time(event);
1100 u64 delta;
1101 /*
1102 * An event which could not be activated because of
1103 * filter mismatch still needs to have its timings
1104 * maintained, otherwise bogus information is return
1105 * via read() for time_enabled, time_running:
1106 */
1107 if (event->state == PERF_EVENT_STATE_INACTIVE
1108 && !event_filter_match(event)) {
1109 delta = tstamp - event->tstamp_stopped;
1110 event->tstamp_running += delta;
1111 event->tstamp_stopped = tstamp;
1112 }
1113
1114 if (event->state != PERF_EVENT_STATE_ACTIVE)
1115 return;
1116
1117 event->state = PERF_EVENT_STATE_INACTIVE;
1118 if (event->pending_disable) {
1119 event->pending_disable = 0;
1120 event->state = PERF_EVENT_STATE_OFF;
1121 }
1122 event->tstamp_stopped = tstamp;
1123 event->pmu->del(event, 0);
1124 event->oncpu = -1;
1125
1126 if (!is_software_event(event))
1127 cpuctx->active_oncpu--;
1128 ctx->nr_active--;
1129 if (event->attr.exclusive || !cpuctx->active_oncpu)
1130 cpuctx->exclusive = 0;
1131}
1132
1133static void
1134group_sched_out(struct perf_event *group_event,
1135 struct perf_cpu_context *cpuctx,
1136 struct perf_event_context *ctx)
1137{
1138 struct perf_event *event;
1139 int state = group_event->state;
1140
1141 event_sched_out(group_event, cpuctx, ctx);
1142
1143 /*
1144 * Schedule out siblings (if any):
1145 */
1146 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1147 event_sched_out(event, cpuctx, ctx);
1148
1149 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1150 cpuctx->exclusive = 0;
1151}
1152
1153/*
1154 * Cross CPU call to remove a performance event
1155 *
1156 * We disable the event on the hardware level first. After that we
1157 * remove it from the context list.
1158 */
1159static int __perf_remove_from_context(void *info)
1160{
1161 struct perf_event *event = info;
1162 struct perf_event_context *ctx = event->ctx;
1163 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1164
1165 raw_spin_lock(&ctx->lock);
1166 event_sched_out(event, cpuctx, ctx);
1167 list_del_event(event, ctx);
1168 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1169 ctx->is_active = 0;
1170 cpuctx->task_ctx = NULL;
1171 }
1172 raw_spin_unlock(&ctx->lock);
1173
1174 return 0;
1175}
1176
1177
1178/*
1179 * Remove the event from a task's (or a CPU's) list of events.
1180 *
1181 * CPU events are removed with a smp call. For task events we only
1182 * call when the task is on a CPU.
1183 *
1184 * If event->ctx is a cloned context, callers must make sure that
1185 * every task struct that event->ctx->task could possibly point to
1186 * remains valid. This is OK when called from perf_release since
1187 * that only calls us on the top-level context, which can't be a clone.
1188 * When called from perf_event_exit_task, it's OK because the
1189 * context has been detached from its task.
1190 */
1191static void perf_remove_from_context(struct perf_event *event)
1192{
1193 struct perf_event_context *ctx = event->ctx;
1194 struct task_struct *task = ctx->task;
1195
1196 lockdep_assert_held(&ctx->mutex);
1197
1198 if (!task) {
1199 /*
1200 * Per cpu events are removed via an smp call and
1201 * the removal is always successful.
1202 */
1203 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1204 return;
1205 }
1206
1207retry:
1208 if (!task_function_call(task, __perf_remove_from_context, event))
1209 return;
1210
1211 raw_spin_lock_irq(&ctx->lock);
1212 /*
1213 * If we failed to find a running task, but find the context active now
1214 * that we've acquired the ctx->lock, retry.
1215 */
1216 if (ctx->is_active) {
1217 raw_spin_unlock_irq(&ctx->lock);
1218 goto retry;
1219 }
1220
1221 /*
1222 * Since the task isn't running, its safe to remove the event, us
1223 * holding the ctx->lock ensures the task won't get scheduled in.
1224 */
1225 list_del_event(event, ctx);
1226 raw_spin_unlock_irq(&ctx->lock);
1227}
1228
1229/*
1230 * Cross CPU call to disable a performance event
1231 */
1232static int __perf_event_disable(void *info)
1233{
1234 struct perf_event *event = info;
1235 struct perf_event_context *ctx = event->ctx;
1236 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1237
1238 /*
1239 * If this is a per-task event, need to check whether this
1240 * event's task is the current task on this cpu.
1241 *
1242 * Can trigger due to concurrent perf_event_context_sched_out()
1243 * flipping contexts around.
1244 */
1245 if (ctx->task && cpuctx->task_ctx != ctx)
1246 return -EINVAL;
1247
1248 raw_spin_lock(&ctx->lock);
1249
1250 /*
1251 * If the event is on, turn it off.
1252 * If it is in error state, leave it in error state.
1253 */
1254 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1255 update_context_time(ctx);
1256 update_cgrp_time_from_event(event);
1257 update_group_times(event);
1258 if (event == event->group_leader)
1259 group_sched_out(event, cpuctx, ctx);
1260 else
1261 event_sched_out(event, cpuctx, ctx);
1262 event->state = PERF_EVENT_STATE_OFF;
1263 }
1264
1265 raw_spin_unlock(&ctx->lock);
1266
1267 return 0;
1268}
1269
1270/*
1271 * Disable a event.
1272 *
1273 * If event->ctx is a cloned context, callers must make sure that
1274 * every task struct that event->ctx->task could possibly point to
1275 * remains valid. This condition is satisifed when called through
1276 * perf_event_for_each_child or perf_event_for_each because they
1277 * hold the top-level event's child_mutex, so any descendant that
1278 * goes to exit will block in sync_child_event.
1279 * When called from perf_pending_event it's OK because event->ctx
1280 * is the current context on this CPU and preemption is disabled,
1281 * hence we can't get into perf_event_task_sched_out for this context.
1282 */
1283void perf_event_disable(struct perf_event *event)
1284{
1285 struct perf_event_context *ctx = event->ctx;
1286 struct task_struct *task = ctx->task;
1287
1288 if (!task) {
1289 /*
1290 * Disable the event on the cpu that it's on
1291 */
1292 cpu_function_call(event->cpu, __perf_event_disable, event);
1293 return;
1294 }
1295
1296retry:
1297 if (!task_function_call(task, __perf_event_disable, event))
1298 return;
1299
1300 raw_spin_lock_irq(&ctx->lock);
1301 /*
1302 * If the event is still active, we need to retry the cross-call.
1303 */
1304 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1305 raw_spin_unlock_irq(&ctx->lock);
1306 /*
1307 * Reload the task pointer, it might have been changed by
1308 * a concurrent perf_event_context_sched_out().
1309 */
1310 task = ctx->task;
1311 goto retry;
1312 }
1313
1314 /*
1315 * Since we have the lock this context can't be scheduled
1316 * in, so we can change the state safely.
1317 */
1318 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1319 update_group_times(event);
1320 event->state = PERF_EVENT_STATE_OFF;
1321 }
1322 raw_spin_unlock_irq(&ctx->lock);
1323}
1324
1325static void perf_set_shadow_time(struct perf_event *event,
1326 struct perf_event_context *ctx,
1327 u64 tstamp)
1328{
1329 /*
1330 * use the correct time source for the time snapshot
1331 *
1332 * We could get by without this by leveraging the
1333 * fact that to get to this function, the caller
1334 * has most likely already called update_context_time()
1335 * and update_cgrp_time_xx() and thus both timestamp
1336 * are identical (or very close). Given that tstamp is,
1337 * already adjusted for cgroup, we could say that:
1338 * tstamp - ctx->timestamp
1339 * is equivalent to
1340 * tstamp - cgrp->timestamp.
1341 *
1342 * Then, in perf_output_read(), the calculation would
1343 * work with no changes because:
1344 * - event is guaranteed scheduled in
1345 * - no scheduled out in between
1346 * - thus the timestamp would be the same
1347 *
1348 * But this is a bit hairy.
1349 *
1350 * So instead, we have an explicit cgroup call to remain
1351 * within the time time source all along. We believe it
1352 * is cleaner and simpler to understand.
1353 */
1354 if (is_cgroup_event(event))
1355 perf_cgroup_set_shadow_time(event, tstamp);
1356 else
1357 event->shadow_ctx_time = tstamp - ctx->timestamp;
1358}
1359
1360#define MAX_INTERRUPTS (~0ULL)
1361
1362static void perf_log_throttle(struct perf_event *event, int enable);
1363
1364static int
1365event_sched_in(struct perf_event *event,
1366 struct perf_cpu_context *cpuctx,
1367 struct perf_event_context *ctx)
1368{
1369 u64 tstamp = perf_event_time(event);
1370
1371 if (event->state <= PERF_EVENT_STATE_OFF)
1372 return 0;
1373
1374 event->state = PERF_EVENT_STATE_ACTIVE;
1375 event->oncpu = smp_processor_id();
1376
1377 /*
1378 * Unthrottle events, since we scheduled we might have missed several
1379 * ticks already, also for a heavily scheduling task there is little
1380 * guarantee it'll get a tick in a timely manner.
1381 */
1382 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1383 perf_log_throttle(event, 1);
1384 event->hw.interrupts = 0;
1385 }
1386
1387 /*
1388 * The new state must be visible before we turn it on in the hardware:
1389 */
1390 smp_wmb();
1391
1392 if (event->pmu->add(event, PERF_EF_START)) {
1393 event->state = PERF_EVENT_STATE_INACTIVE;
1394 event->oncpu = -1;
1395 return -EAGAIN;
1396 }
1397
1398 event->tstamp_running += tstamp - event->tstamp_stopped;
1399
1400 perf_set_shadow_time(event, ctx, tstamp);
1401
1402 if (!is_software_event(event))
1403 cpuctx->active_oncpu++;
1404 ctx->nr_active++;
1405
1406 if (event->attr.exclusive)
1407 cpuctx->exclusive = 1;
1408
1409 return 0;
1410}
1411
1412static int
1413group_sched_in(struct perf_event *group_event,
1414 struct perf_cpu_context *cpuctx,
1415 struct perf_event_context *ctx)
1416{
1417 struct perf_event *event, *partial_group = NULL;
1418 struct pmu *pmu = group_event->pmu;
1419 u64 now = ctx->time;
1420 bool simulate = false;
1421
1422 if (group_event->state == PERF_EVENT_STATE_OFF)
1423 return 0;
1424
1425 pmu->start_txn(pmu);
1426
1427 if (event_sched_in(group_event, cpuctx, ctx)) {
1428 pmu->cancel_txn(pmu);
1429 return -EAGAIN;
1430 }
1431
1432 /*
1433 * Schedule in siblings as one group (if any):
1434 */
1435 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1436 if (event_sched_in(event, cpuctx, ctx)) {
1437 partial_group = event;
1438 goto group_error;
1439 }
1440 }
1441
1442 if (!pmu->commit_txn(pmu))
1443 return 0;
1444
1445group_error:
1446 /*
1447 * Groups can be scheduled in as one unit only, so undo any
1448 * partial group before returning:
1449 * The events up to the failed event are scheduled out normally,
1450 * tstamp_stopped will be updated.
1451 *
1452 * The failed events and the remaining siblings need to have
1453 * their timings updated as if they had gone thru event_sched_in()
1454 * and event_sched_out(). This is required to get consistent timings
1455 * across the group. This also takes care of the case where the group
1456 * could never be scheduled by ensuring tstamp_stopped is set to mark
1457 * the time the event was actually stopped, such that time delta
1458 * calculation in update_event_times() is correct.
1459 */
1460 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1461 if (event == partial_group)
1462 simulate = true;
1463
1464 if (simulate) {
1465 event->tstamp_running += now - event->tstamp_stopped;
1466 event->tstamp_stopped = now;
1467 } else {
1468 event_sched_out(event, cpuctx, ctx);
1469 }
1470 }
1471 event_sched_out(group_event, cpuctx, ctx);
1472
1473 pmu->cancel_txn(pmu);
1474
1475 return -EAGAIN;
1476}
1477
1478/*
1479 * Work out whether we can put this event group on the CPU now.
1480 */
1481static int group_can_go_on(struct perf_event *event,
1482 struct perf_cpu_context *cpuctx,
1483 int can_add_hw)
1484{
1485 /*
1486 * Groups consisting entirely of software events can always go on.
1487 */
1488 if (event->group_flags & PERF_GROUP_SOFTWARE)
1489 return 1;
1490 /*
1491 * If an exclusive group is already on, no other hardware
1492 * events can go on.
1493 */
1494 if (cpuctx->exclusive)
1495 return 0;
1496 /*
1497 * If this group is exclusive and there are already
1498 * events on the CPU, it can't go on.
1499 */
1500 if (event->attr.exclusive && cpuctx->active_oncpu)
1501 return 0;
1502 /*
1503 * Otherwise, try to add it if all previous groups were able
1504 * to go on.
1505 */
1506 return can_add_hw;
1507}
1508
1509static void add_event_to_ctx(struct perf_event *event,
1510 struct perf_event_context *ctx)
1511{
1512 u64 tstamp = perf_event_time(event);
1513
1514 list_add_event(event, ctx);
1515 perf_group_attach(event);
1516 event->tstamp_enabled = tstamp;
1517 event->tstamp_running = tstamp;
1518 event->tstamp_stopped = tstamp;
1519}
1520
1521static void task_ctx_sched_out(struct perf_event_context *ctx);
1522static void
1523ctx_sched_in(struct perf_event_context *ctx,
1524 struct perf_cpu_context *cpuctx,
1525 enum event_type_t event_type,
1526 struct task_struct *task);
1527
1528static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1529 struct perf_event_context *ctx,
1530 struct task_struct *task)
1531{
1532 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1533 if (ctx)
1534 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1535 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1536 if (ctx)
1537 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1538}
1539
1540/*
1541 * Cross CPU call to install and enable a performance event
1542 *
1543 * Must be called with ctx->mutex held
1544 */
1545static int __perf_install_in_context(void *info)
1546{
1547 struct perf_event *event = info;
1548 struct perf_event_context *ctx = event->ctx;
1549 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1550 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1551 struct task_struct *task = current;
1552
1553 perf_ctx_lock(cpuctx, task_ctx);
1554 perf_pmu_disable(cpuctx->ctx.pmu);
1555
1556 /*
1557 * If there was an active task_ctx schedule it out.
1558 */
1559 if (task_ctx)
1560 task_ctx_sched_out(task_ctx);
1561
1562 /*
1563 * If the context we're installing events in is not the
1564 * active task_ctx, flip them.
1565 */
1566 if (ctx->task && task_ctx != ctx) {
1567 if (task_ctx)
1568 raw_spin_unlock(&task_ctx->lock);
1569 raw_spin_lock(&ctx->lock);
1570 task_ctx = ctx;
1571 }
1572
1573 if (task_ctx) {
1574 cpuctx->task_ctx = task_ctx;
1575 task = task_ctx->task;
1576 }
1577
1578 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1579
1580 update_context_time(ctx);
1581 /*
1582 * update cgrp time only if current cgrp
1583 * matches event->cgrp. Must be done before
1584 * calling add_event_to_ctx()
1585 */
1586 update_cgrp_time_from_event(event);
1587
1588 add_event_to_ctx(event, ctx);
1589
1590 /*
1591 * Schedule everything back in
1592 */
1593 perf_event_sched_in(cpuctx, task_ctx, task);
1594
1595 perf_pmu_enable(cpuctx->ctx.pmu);
1596 perf_ctx_unlock(cpuctx, task_ctx);
1597
1598 return 0;
1599}
1600
1601/*
1602 * Attach a performance event to a context
1603 *
1604 * First we add the event to the list with the hardware enable bit
1605 * in event->hw_config cleared.
1606 *
1607 * If the event is attached to a task which is on a CPU we use a smp
1608 * call to enable it in the task context. The task might have been
1609 * scheduled away, but we check this in the smp call again.
1610 */
1611static void
1612perf_install_in_context(struct perf_event_context *ctx,
1613 struct perf_event *event,
1614 int cpu)
1615{
1616 struct task_struct *task = ctx->task;
1617
1618 lockdep_assert_held(&ctx->mutex);
1619
1620 event->ctx = ctx;
1621
1622 if (!task) {
1623 /*
1624 * Per cpu events are installed via an smp call and
1625 * the install is always successful.
1626 */
1627 cpu_function_call(cpu, __perf_install_in_context, event);
1628 return;
1629 }
1630
1631retry:
1632 if (!task_function_call(task, __perf_install_in_context, event))
1633 return;
1634
1635 raw_spin_lock_irq(&ctx->lock);
1636 /*
1637 * If we failed to find a running task, but find the context active now
1638 * that we've acquired the ctx->lock, retry.
1639 */
1640 if (ctx->is_active) {
1641 raw_spin_unlock_irq(&ctx->lock);
1642 goto retry;
1643 }
1644
1645 /*
1646 * Since the task isn't running, its safe to add the event, us holding
1647 * the ctx->lock ensures the task won't get scheduled in.
1648 */
1649 add_event_to_ctx(event, ctx);
1650 raw_spin_unlock_irq(&ctx->lock);
1651}
1652
1653/*
1654 * Put a event into inactive state and update time fields.
1655 * Enabling the leader of a group effectively enables all
1656 * the group members that aren't explicitly disabled, so we
1657 * have to update their ->tstamp_enabled also.
1658 * Note: this works for group members as well as group leaders
1659 * since the non-leader members' sibling_lists will be empty.
1660 */
1661static void __perf_event_mark_enabled(struct perf_event *event,
1662 struct perf_event_context *ctx)
1663{
1664 struct perf_event *sub;
1665 u64 tstamp = perf_event_time(event);
1666
1667 event->state = PERF_EVENT_STATE_INACTIVE;
1668 event->tstamp_enabled = tstamp - event->total_time_enabled;
1669 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1670 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1671 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1672 }
1673}
1674
1675/*
1676 * Cross CPU call to enable a performance event
1677 */
1678static int __perf_event_enable(void *info)
1679{
1680 struct perf_event *event = info;
1681 struct perf_event_context *ctx = event->ctx;
1682 struct perf_event *leader = event->group_leader;
1683 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1684 int err;
1685
1686 if (WARN_ON_ONCE(!ctx->is_active))
1687 return -EINVAL;
1688
1689 raw_spin_lock(&ctx->lock);
1690 update_context_time(ctx);
1691
1692 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1693 goto unlock;
1694
1695 /*
1696 * set current task's cgroup time reference point
1697 */
1698 perf_cgroup_set_timestamp(current, ctx);
1699
1700 __perf_event_mark_enabled(event, ctx);
1701
1702 if (!event_filter_match(event)) {
1703 if (is_cgroup_event(event))
1704 perf_cgroup_defer_enabled(event);
1705 goto unlock;
1706 }
1707
1708 /*
1709 * If the event is in a group and isn't the group leader,
1710 * then don't put it on unless the group is on.
1711 */
1712 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
1713 goto unlock;
1714
1715 if (!group_can_go_on(event, cpuctx, 1)) {
1716 err = -EEXIST;
1717 } else {
1718 if (event == leader)
1719 err = group_sched_in(event, cpuctx, ctx);
1720 else
1721 err = event_sched_in(event, cpuctx, ctx);
1722 }
1723
1724 if (err) {
1725 /*
1726 * If this event can't go on and it's part of a
1727 * group, then the whole group has to come off.
1728 */
1729 if (leader != event)
1730 group_sched_out(leader, cpuctx, ctx);
1731 if (leader->attr.pinned) {
1732 update_group_times(leader);
1733 leader->state = PERF_EVENT_STATE_ERROR;
1734 }
1735 }
1736
1737unlock:
1738 raw_spin_unlock(&ctx->lock);
1739
1740 return 0;
1741}
1742
1743/*
1744 * Enable a event.
1745 *
1746 * If event->ctx is a cloned context, callers must make sure that
1747 * every task struct that event->ctx->task could possibly point to
1748 * remains valid. This condition is satisfied when called through
1749 * perf_event_for_each_child or perf_event_for_each as described
1750 * for perf_event_disable.
1751 */
1752void perf_event_enable(struct perf_event *event)
1753{
1754 struct perf_event_context *ctx = event->ctx;
1755 struct task_struct *task = ctx->task;
1756
1757 if (!task) {
1758 /*
1759 * Enable the event on the cpu that it's on
1760 */
1761 cpu_function_call(event->cpu, __perf_event_enable, event);
1762 return;
1763 }
1764
1765 raw_spin_lock_irq(&ctx->lock);
1766 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1767 goto out;
1768
1769 /*
1770 * If the event is in error state, clear that first.
1771 * That way, if we see the event in error state below, we
1772 * know that it has gone back into error state, as distinct
1773 * from the task having been scheduled away before the
1774 * cross-call arrived.
1775 */
1776 if (event->state == PERF_EVENT_STATE_ERROR)
1777 event->state = PERF_EVENT_STATE_OFF;
1778
1779retry:
1780 if (!ctx->is_active) {
1781 __perf_event_mark_enabled(event, ctx);
1782 goto out;
1783 }
1784
1785 raw_spin_unlock_irq(&ctx->lock);
1786
1787 if (!task_function_call(task, __perf_event_enable, event))
1788 return;
1789
1790 raw_spin_lock_irq(&ctx->lock);
1791
1792 /*
1793 * If the context is active and the event is still off,
1794 * we need to retry the cross-call.
1795 */
1796 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
1797 /*
1798 * task could have been flipped by a concurrent
1799 * perf_event_context_sched_out()
1800 */
1801 task = ctx->task;
1802 goto retry;
1803 }
1804
1805out:
1806 raw_spin_unlock_irq(&ctx->lock);
1807}
1808
1809int perf_event_refresh(struct perf_event *event, int refresh)
1810{
1811 /*
1812 * not supported on inherited events
1813 */
1814 if (event->attr.inherit || !is_sampling_event(event))
1815 return -EINVAL;
1816
1817 atomic_add(refresh, &event->event_limit);
1818 perf_event_enable(event);
1819
1820 return 0;
1821}
1822EXPORT_SYMBOL_GPL(perf_event_refresh);
1823
1824static void ctx_sched_out(struct perf_event_context *ctx,
1825 struct perf_cpu_context *cpuctx,
1826 enum event_type_t event_type)
1827{
1828 struct perf_event *event;
1829 int is_active = ctx->is_active;
1830
1831 ctx->is_active &= ~event_type;
1832 if (likely(!ctx->nr_events))
1833 return;
1834
1835 update_context_time(ctx);
1836 update_cgrp_time_from_cpuctx(cpuctx);
1837 if (!ctx->nr_active)
1838 return;
1839
1840 perf_pmu_disable(ctx->pmu);
1841 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
1842 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
1843 group_sched_out(event, cpuctx, ctx);
1844 }
1845
1846 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
1847 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
1848 group_sched_out(event, cpuctx, ctx);
1849 }
1850 perf_pmu_enable(ctx->pmu);
1851}
1852
1853/*
1854 * Test whether two contexts are equivalent, i.e. whether they
1855 * have both been cloned from the same version of the same context
1856 * and they both have the same number of enabled events.
1857 * If the number of enabled events is the same, then the set
1858 * of enabled events should be the same, because these are both
1859 * inherited contexts, therefore we can't access individual events
1860 * in them directly with an fd; we can only enable/disable all
1861 * events via prctl, or enable/disable all events in a family
1862 * via ioctl, which will have the same effect on both contexts.
1863 */
1864static int context_equiv(struct perf_event_context *ctx1,
1865 struct perf_event_context *ctx2)
1866{
1867 return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1868 && ctx1->parent_gen == ctx2->parent_gen
1869 && !ctx1->pin_count && !ctx2->pin_count;
1870}
1871
1872static void __perf_event_sync_stat(struct perf_event *event,
1873 struct perf_event *next_event)
1874{
1875 u64 value;
1876
1877 if (!event->attr.inherit_stat)
1878 return;
1879
1880 /*
1881 * Update the event value, we cannot use perf_event_read()
1882 * because we're in the middle of a context switch and have IRQs
1883 * disabled, which upsets smp_call_function_single(), however
1884 * we know the event must be on the current CPU, therefore we
1885 * don't need to use it.
1886 */
1887 switch (event->state) {
1888 case PERF_EVENT_STATE_ACTIVE:
1889 event->pmu->read(event);
1890 /* fall-through */
1891
1892 case PERF_EVENT_STATE_INACTIVE:
1893 update_event_times(event);
1894 break;
1895
1896 default:
1897 break;
1898 }
1899
1900 /*
1901 * In order to keep per-task stats reliable we need to flip the event
1902 * values when we flip the contexts.
1903 */
1904 value = local64_read(&next_event->count);
1905 value = local64_xchg(&event->count, value);
1906 local64_set(&next_event->count, value);
1907
1908 swap(event->total_time_enabled, next_event->total_time_enabled);
1909 swap(event->total_time_running, next_event->total_time_running);
1910
1911 /*
1912 * Since we swizzled the values, update the user visible data too.
1913 */
1914 perf_event_update_userpage(event);
1915 perf_event_update_userpage(next_event);
1916}
1917
1918#define list_next_entry(pos, member) \
1919 list_entry(pos->member.next, typeof(*pos), member)
1920
1921static void perf_event_sync_stat(struct perf_event_context *ctx,
1922 struct perf_event_context *next_ctx)
1923{
1924 struct perf_event *event, *next_event;
1925
1926 if (!ctx->nr_stat)
1927 return;
1928
1929 update_context_time(ctx);
1930
1931 event = list_first_entry(&ctx->event_list,
1932 struct perf_event, event_entry);
1933
1934 next_event = list_first_entry(&next_ctx->event_list,
1935 struct perf_event, event_entry);
1936
1937 while (&event->event_entry != &ctx->event_list &&
1938 &next_event->event_entry != &next_ctx->event_list) {
1939
1940 __perf_event_sync_stat(event, next_event);
1941
1942 event = list_next_entry(event, event_entry);
1943 next_event = list_next_entry(next_event, event_entry);
1944 }
1945}
1946
1947static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
1948 struct task_struct *next)
1949{
1950 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
1951 struct perf_event_context *next_ctx;
1952 struct perf_event_context *parent;
1953 struct perf_cpu_context *cpuctx;
1954 int do_switch = 1;
1955
1956 if (likely(!ctx))
1957 return;
1958
1959 cpuctx = __get_cpu_context(ctx);
1960 if (!cpuctx->task_ctx)
1961 return;
1962
1963 rcu_read_lock();
1964 parent = rcu_dereference(ctx->parent_ctx);
1965 next_ctx = next->perf_event_ctxp[ctxn];
1966 if (parent && next_ctx &&
1967 rcu_dereference(next_ctx->parent_ctx) == parent) {
1968 /*
1969 * Looks like the two contexts are clones, so we might be
1970 * able to optimize the context switch. We lock both
1971 * contexts and check that they are clones under the
1972 * lock (including re-checking that neither has been
1973 * uncloned in the meantime). It doesn't matter which
1974 * order we take the locks because no other cpu could
1975 * be trying to lock both of these tasks.
1976 */
1977 raw_spin_lock(&ctx->lock);
1978 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1979 if (context_equiv(ctx, next_ctx)) {
1980 /*
1981 * XXX do we need a memory barrier of sorts
1982 * wrt to rcu_dereference() of perf_event_ctxp
1983 */
1984 task->perf_event_ctxp[ctxn] = next_ctx;
1985 next->perf_event_ctxp[ctxn] = ctx;
1986 ctx->task = next;
1987 next_ctx->task = task;
1988 do_switch = 0;
1989
1990 perf_event_sync_stat(ctx, next_ctx);
1991 }
1992 raw_spin_unlock(&next_ctx->lock);
1993 raw_spin_unlock(&ctx->lock);
1994 }
1995 rcu_read_unlock();
1996
1997 if (do_switch) {
1998 raw_spin_lock(&ctx->lock);
1999 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2000 cpuctx->task_ctx = NULL;
2001 raw_spin_unlock(&ctx->lock);
2002 }
2003}
2004
2005#define for_each_task_context_nr(ctxn) \
2006 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2007
2008/*
2009 * Called from scheduler to remove the events of the current task,
2010 * with interrupts disabled.
2011 *
2012 * We stop each event and update the event value in event->count.
2013 *
2014 * This does not protect us against NMI, but disable()
2015 * sets the disabled bit in the control field of event _before_
2016 * accessing the event control register. If a NMI hits, then it will
2017 * not restart the event.
2018 */
2019void __perf_event_task_sched_out(struct task_struct *task,
2020 struct task_struct *next)
2021{
2022 int ctxn;
2023
2024 for_each_task_context_nr(ctxn)
2025 perf_event_context_sched_out(task, ctxn, next);
2026
2027 /*
2028 * if cgroup events exist on this CPU, then we need
2029 * to check if we have to switch out PMU state.
2030 * cgroup event are system-wide mode only
2031 */
2032 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2033 perf_cgroup_sched_out(task, next);
2034}
2035
2036static void task_ctx_sched_out(struct perf_event_context *ctx)
2037{
2038 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2039
2040 if (!cpuctx->task_ctx)
2041 return;
2042
2043 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2044 return;
2045
2046 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2047 cpuctx->task_ctx = NULL;
2048}
2049
2050/*
2051 * Called with IRQs disabled
2052 */
2053static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2054 enum event_type_t event_type)
2055{
2056 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2057}
2058
2059static void
2060ctx_pinned_sched_in(struct perf_event_context *ctx,
2061 struct perf_cpu_context *cpuctx)
2062{
2063 struct perf_event *event;
2064
2065 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2066 if (event->state <= PERF_EVENT_STATE_OFF)
2067 continue;
2068 if (!event_filter_match(event))
2069 continue;
2070
2071 /* may need to reset tstamp_enabled */
2072 if (is_cgroup_event(event))
2073 perf_cgroup_mark_enabled(event, ctx);
2074
2075 if (group_can_go_on(event, cpuctx, 1))
2076 group_sched_in(event, cpuctx, ctx);
2077
2078 /*
2079 * If this pinned group hasn't been scheduled,
2080 * put it in error state.
2081 */
2082 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2083 update_group_times(event);
2084 event->state = PERF_EVENT_STATE_ERROR;
2085 }
2086 }
2087}
2088
2089static void
2090ctx_flexible_sched_in(struct perf_event_context *ctx,
2091 struct perf_cpu_context *cpuctx)
2092{
2093 struct perf_event *event;
2094 int can_add_hw = 1;
2095
2096 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2097 /* Ignore events in OFF or ERROR state */
2098 if (event->state <= PERF_EVENT_STATE_OFF)
2099 continue;
2100 /*
2101 * Listen to the 'cpu' scheduling filter constraint
2102 * of events:
2103 */
2104 if (!event_filter_match(event))
2105 continue;
2106
2107 /* may need to reset tstamp_enabled */
2108 if (is_cgroup_event(event))
2109 perf_cgroup_mark_enabled(event, ctx);
2110
2111 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2112 if (group_sched_in(event, cpuctx, ctx))
2113 can_add_hw = 0;
2114 }
2115 }
2116}
2117
2118static void
2119ctx_sched_in(struct perf_event_context *ctx,
2120 struct perf_cpu_context *cpuctx,
2121 enum event_type_t event_type,
2122 struct task_struct *task)
2123{
2124 u64 now;
2125 int is_active = ctx->is_active;
2126
2127 ctx->is_active |= event_type;
2128 if (likely(!ctx->nr_events))
2129 return;
2130
2131 now = perf_clock();
2132 ctx->timestamp = now;
2133 perf_cgroup_set_timestamp(task, ctx);
2134 /*
2135 * First go through the list and put on any pinned groups
2136 * in order to give them the best chance of going on.
2137 */
2138 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2139 ctx_pinned_sched_in(ctx, cpuctx);
2140
2141 /* Then walk through the lower prio flexible groups */
2142 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2143 ctx_flexible_sched_in(ctx, cpuctx);
2144}
2145
2146static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2147 enum event_type_t event_type,
2148 struct task_struct *task)
2149{
2150 struct perf_event_context *ctx = &cpuctx->ctx;
2151
2152 ctx_sched_in(ctx, cpuctx, event_type, task);
2153}
2154
2155static void perf_event_context_sched_in(struct perf_event_context *ctx,
2156 struct task_struct *task)
2157{
2158 struct perf_cpu_context *cpuctx;
2159
2160 cpuctx = __get_cpu_context(ctx);
2161 if (cpuctx->task_ctx == ctx)
2162 return;
2163
2164 perf_ctx_lock(cpuctx, ctx);
2165 perf_pmu_disable(ctx->pmu);
2166 /*
2167 * We want to keep the following priority order:
2168 * cpu pinned (that don't need to move), task pinned,
2169 * cpu flexible, task flexible.
2170 */
2171 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2172
2173 perf_event_sched_in(cpuctx, ctx, task);
2174
2175 cpuctx->task_ctx = ctx;
2176
2177 perf_pmu_enable(ctx->pmu);
2178 perf_ctx_unlock(cpuctx, ctx);
2179
2180 /*
2181 * Since these rotations are per-cpu, we need to ensure the
2182 * cpu-context we got scheduled on is actually rotating.
2183 */
2184 perf_pmu_rotate_start(ctx->pmu);
2185}
2186
2187/*
2188 * Called from scheduler to add the events of the current task
2189 * with interrupts disabled.
2190 *
2191 * We restore the event value and then enable it.
2192 *
2193 * This does not protect us against NMI, but enable()
2194 * sets the enabled bit in the control field of event _before_
2195 * accessing the event control register. If a NMI hits, then it will
2196 * keep the event running.
2197 */
2198void __perf_event_task_sched_in(struct task_struct *prev,
2199 struct task_struct *task)
2200{
2201 struct perf_event_context *ctx;
2202 int ctxn;
2203
2204 for_each_task_context_nr(ctxn) {
2205 ctx = task->perf_event_ctxp[ctxn];
2206 if (likely(!ctx))
2207 continue;
2208
2209 perf_event_context_sched_in(ctx, task);
2210 }
2211 /*
2212 * if cgroup events exist on this CPU, then we need
2213 * to check if we have to switch in PMU state.
2214 * cgroup event are system-wide mode only
2215 */
2216 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2217 perf_cgroup_sched_in(prev, task);
2218}
2219
2220static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2221{
2222 u64 frequency = event->attr.sample_freq;
2223 u64 sec = NSEC_PER_SEC;
2224 u64 divisor, dividend;
2225
2226 int count_fls, nsec_fls, frequency_fls, sec_fls;
2227
2228 count_fls = fls64(count);
2229 nsec_fls = fls64(nsec);
2230 frequency_fls = fls64(frequency);
2231 sec_fls = 30;
2232
2233 /*
2234 * We got @count in @nsec, with a target of sample_freq HZ
2235 * the target period becomes:
2236 *
2237 * @count * 10^9
2238 * period = -------------------
2239 * @nsec * sample_freq
2240 *
2241 */
2242
2243 /*
2244 * Reduce accuracy by one bit such that @a and @b converge
2245 * to a similar magnitude.
2246 */
2247#define REDUCE_FLS(a, b) \
2248do { \
2249 if (a##_fls > b##_fls) { \
2250 a >>= 1; \
2251 a##_fls--; \
2252 } else { \
2253 b >>= 1; \
2254 b##_fls--; \
2255 } \
2256} while (0)
2257
2258 /*
2259 * Reduce accuracy until either term fits in a u64, then proceed with
2260 * the other, so that finally we can do a u64/u64 division.
2261 */
2262 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2263 REDUCE_FLS(nsec, frequency);
2264 REDUCE_FLS(sec, count);
2265 }
2266
2267 if (count_fls + sec_fls > 64) {
2268 divisor = nsec * frequency;
2269
2270 while (count_fls + sec_fls > 64) {
2271 REDUCE_FLS(count, sec);
2272 divisor >>= 1;
2273 }
2274
2275 dividend = count * sec;
2276 } else {
2277 dividend = count * sec;
2278
2279 while (nsec_fls + frequency_fls > 64) {
2280 REDUCE_FLS(nsec, frequency);
2281 dividend >>= 1;
2282 }
2283
2284 divisor = nsec * frequency;
2285 }
2286
2287 if (!divisor)
2288 return dividend;
2289
2290 return div64_u64(dividend, divisor);
2291}
2292
2293static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
2294{
2295 struct hw_perf_event *hwc = &event->hw;
2296 s64 period, sample_period;
2297 s64 delta;
2298
2299 period = perf_calculate_period(event, nsec, count);
2300
2301 delta = (s64)(period - hwc->sample_period);
2302 delta = (delta + 7) / 8; /* low pass filter */
2303
2304 sample_period = hwc->sample_period + delta;
2305
2306 if (!sample_period)
2307 sample_period = 1;
2308
2309 hwc->sample_period = sample_period;
2310
2311 if (local64_read(&hwc->period_left) > 8*sample_period) {
2312 event->pmu->stop(event, PERF_EF_UPDATE);
2313 local64_set(&hwc->period_left, 0);
2314 event->pmu->start(event, PERF_EF_RELOAD);
2315 }
2316}
2317
2318static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
2319{
2320 struct perf_event *event;
2321 struct hw_perf_event *hwc;
2322 u64 interrupts, now;
2323 s64 delta;
2324
2325 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2326 if (event->state != PERF_EVENT_STATE_ACTIVE)
2327 continue;
2328
2329 if (!event_filter_match(event))
2330 continue;
2331
2332 hwc = &event->hw;
2333
2334 interrupts = hwc->interrupts;
2335 hwc->interrupts = 0;
2336
2337 /*
2338 * unthrottle events on the tick
2339 */
2340 if (interrupts == MAX_INTERRUPTS) {
2341 perf_log_throttle(event, 1);
2342 event->pmu->start(event, 0);
2343 }
2344
2345 if (!event->attr.freq || !event->attr.sample_freq)
2346 continue;
2347
2348 event->pmu->read(event);
2349 now = local64_read(&event->count);
2350 delta = now - hwc->freq_count_stamp;
2351 hwc->freq_count_stamp = now;
2352
2353 if (delta > 0)
2354 perf_adjust_period(event, period, delta);
2355 }
2356}
2357
2358/*
2359 * Round-robin a context's events:
2360 */
2361static void rotate_ctx(struct perf_event_context *ctx)
2362{
2363 /*
2364 * Rotate the first entry last of non-pinned groups. Rotation might be
2365 * disabled by the inheritance code.
2366 */
2367 if (!ctx->rotate_disable)
2368 list_rotate_left(&ctx->flexible_groups);
2369}
2370
2371/*
2372 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2373 * because they're strictly cpu affine and rotate_start is called with IRQs
2374 * disabled, while rotate_context is called from IRQ context.
2375 */
2376static void perf_rotate_context(struct perf_cpu_context *cpuctx)
2377{
2378 u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
2379 struct perf_event_context *ctx = NULL;
2380 int rotate = 0, remove = 1;
2381
2382 if (cpuctx->ctx.nr_events) {
2383 remove = 0;
2384 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2385 rotate = 1;
2386 }
2387
2388 ctx = cpuctx->task_ctx;
2389 if (ctx && ctx->nr_events) {
2390 remove = 0;
2391 if (ctx->nr_events != ctx->nr_active)
2392 rotate = 1;
2393 }
2394
2395 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2396 perf_pmu_disable(cpuctx->ctx.pmu);
2397 perf_ctx_adjust_freq(&cpuctx->ctx, interval);
2398 if (ctx)
2399 perf_ctx_adjust_freq(ctx, interval);
2400
2401 if (!rotate)
2402 goto done;
2403
2404 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2405 if (ctx)
2406 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2407
2408 rotate_ctx(&cpuctx->ctx);
2409 if (ctx)
2410 rotate_ctx(ctx);
2411
2412 perf_event_sched_in(cpuctx, ctx, current);
2413
2414done:
2415 if (remove)
2416 list_del_init(&cpuctx->rotation_list);
2417
2418 perf_pmu_enable(cpuctx->ctx.pmu);
2419 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2420}
2421
2422void perf_event_task_tick(void)
2423{
2424 struct list_head *head = &__get_cpu_var(rotation_list);
2425 struct perf_cpu_context *cpuctx, *tmp;
2426
2427 WARN_ON(!irqs_disabled());
2428
2429 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2430 if (cpuctx->jiffies_interval == 1 ||
2431 !(jiffies % cpuctx->jiffies_interval))
2432 perf_rotate_context(cpuctx);
2433 }
2434}
2435
2436static int event_enable_on_exec(struct perf_event *event,
2437 struct perf_event_context *ctx)
2438{
2439 if (!event->attr.enable_on_exec)
2440 return 0;
2441
2442 event->attr.enable_on_exec = 0;
2443 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2444 return 0;
2445
2446 __perf_event_mark_enabled(event, ctx);
2447
2448 return 1;
2449}
2450
2451/*
2452 * Enable all of a task's events that have been marked enable-on-exec.
2453 * This expects task == current.
2454 */
2455static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2456{
2457 struct perf_event *event;
2458 unsigned long flags;
2459 int enabled = 0;
2460 int ret;
2461
2462 local_irq_save(flags);
2463 if (!ctx || !ctx->nr_events)
2464 goto out;
2465
2466 /*
2467 * We must ctxsw out cgroup events to avoid conflict
2468 * when invoking perf_task_event_sched_in() later on
2469 * in this function. Otherwise we end up trying to
2470 * ctxswin cgroup events which are already scheduled
2471 * in.
2472 */
2473 perf_cgroup_sched_out(current, NULL);
2474
2475 raw_spin_lock(&ctx->lock);
2476 task_ctx_sched_out(ctx);
2477
2478 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2479 ret = event_enable_on_exec(event, ctx);
2480 if (ret)
2481 enabled = 1;
2482 }
2483
2484 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2485 ret = event_enable_on_exec(event, ctx);
2486 if (ret)
2487 enabled = 1;
2488 }
2489
2490 /*
2491 * Unclone this context if we enabled any event.
2492 */
2493 if (enabled)
2494 unclone_ctx(ctx);
2495
2496 raw_spin_unlock(&ctx->lock);
2497
2498 /*
2499 * Also calls ctxswin for cgroup events, if any:
2500 */
2501 perf_event_context_sched_in(ctx, ctx->task);
2502out:
2503 local_irq_restore(flags);
2504}
2505
2506/*
2507 * Cross CPU call to read the hardware event
2508 */
2509static void __perf_event_read(void *info)
2510{
2511 struct perf_event *event = info;
2512 struct perf_event_context *ctx = event->ctx;
2513 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2514
2515 /*
2516 * If this is a task context, we need to check whether it is
2517 * the current task context of this cpu. If not it has been
2518 * scheduled out before the smp call arrived. In that case
2519 * event->count would have been updated to a recent sample
2520 * when the event was scheduled out.
2521 */
2522 if (ctx->task && cpuctx->task_ctx != ctx)
2523 return;
2524
2525 raw_spin_lock(&ctx->lock);
2526 if (ctx->is_active) {
2527 update_context_time(ctx);
2528 update_cgrp_time_from_event(event);
2529 }
2530 update_event_times(event);
2531 if (event->state == PERF_EVENT_STATE_ACTIVE)
2532 event->pmu->read(event);
2533 raw_spin_unlock(&ctx->lock);
2534}
2535
2536static inline u64 perf_event_count(struct perf_event *event)
2537{
2538 return local64_read(&event->count) + atomic64_read(&event->child_count);
2539}
2540
2541static u64 perf_event_read(struct perf_event *event)
2542{
2543 /*
2544 * If event is enabled and currently active on a CPU, update the
2545 * value in the event structure:
2546 */
2547 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2548 smp_call_function_single(event->oncpu,
2549 __perf_event_read, event, 1);
2550 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2551 struct perf_event_context *ctx = event->ctx;
2552 unsigned long flags;
2553
2554 raw_spin_lock_irqsave(&ctx->lock, flags);
2555 /*
2556 * may read while context is not active
2557 * (e.g., thread is blocked), in that case
2558 * we cannot update context time
2559 */
2560 if (ctx->is_active) {
2561 update_context_time(ctx);
2562 update_cgrp_time_from_event(event);
2563 }
2564 update_event_times(event);
2565 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2566 }
2567
2568 return perf_event_count(event);
2569}
2570
2571/*
2572 * Callchain support
2573 */
2574
2575struct callchain_cpus_entries {
2576 struct rcu_head rcu_head;
2577 struct perf_callchain_entry *cpu_entries[0];
2578};
2579
2580static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
2581static atomic_t nr_callchain_events;
2582static DEFINE_MUTEX(callchain_mutex);
2583struct callchain_cpus_entries *callchain_cpus_entries;
2584
2585
2586__weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
2587 struct pt_regs *regs)
2588{
2589}
2590
2591__weak void perf_callchain_user(struct perf_callchain_entry *entry,
2592 struct pt_regs *regs)
2593{
2594}
2595
2596static void release_callchain_buffers_rcu(struct rcu_head *head)
2597{
2598 struct callchain_cpus_entries *entries;
2599 int cpu;
2600
2601 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
2602
2603 for_each_possible_cpu(cpu)
2604 kfree(entries->cpu_entries[cpu]);
2605
2606 kfree(entries);
2607}
2608
2609static void release_callchain_buffers(void)
2610{
2611 struct callchain_cpus_entries *entries;
2612
2613 entries = callchain_cpus_entries;
2614 rcu_assign_pointer(callchain_cpus_entries, NULL);
2615 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
2616}
2617
2618static int alloc_callchain_buffers(void)
2619{
2620 int cpu;
2621 int size;
2622 struct callchain_cpus_entries *entries;
2623
2624 /*
2625 * We can't use the percpu allocation API for data that can be
2626 * accessed from NMI. Use a temporary manual per cpu allocation
2627 * until that gets sorted out.
2628 */
2629 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
2630
2631 entries = kzalloc(size, GFP_KERNEL);
2632 if (!entries)
2633 return -ENOMEM;
2634
2635 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
2636
2637 for_each_possible_cpu(cpu) {
2638 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
2639 cpu_to_node(cpu));
2640 if (!entries->cpu_entries[cpu])
2641 goto fail;
2642 }
2643
2644 rcu_assign_pointer(callchain_cpus_entries, entries);
2645
2646 return 0;
2647
2648fail:
2649 for_each_possible_cpu(cpu)
2650 kfree(entries->cpu_entries[cpu]);
2651 kfree(entries);
2652
2653 return -ENOMEM;
2654}
2655
2656static int get_callchain_buffers(void)
2657{
2658 int err = 0;
2659 int count;
2660
2661 mutex_lock(&callchain_mutex);
2662
2663 count = atomic_inc_return(&nr_callchain_events);
2664 if (WARN_ON_ONCE(count < 1)) {
2665 err = -EINVAL;
2666 goto exit;
2667 }
2668
2669 if (count > 1) {
2670 /* If the allocation failed, give up */
2671 if (!callchain_cpus_entries)
2672 err = -ENOMEM;
2673 goto exit;
2674 }
2675
2676 err = alloc_callchain_buffers();
2677 if (err)
2678 release_callchain_buffers();
2679exit:
2680 mutex_unlock(&callchain_mutex);
2681
2682 return err;
2683}
2684
2685static void put_callchain_buffers(void)
2686{
2687 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
2688 release_callchain_buffers();
2689 mutex_unlock(&callchain_mutex);
2690 }
2691}
2692
2693static int get_recursion_context(int *recursion)
2694{
2695 int rctx;
2696
2697 if (in_nmi())
2698 rctx = 3;
2699 else if (in_irq())
2700 rctx = 2;
2701 else if (in_softirq())
2702 rctx = 1;
2703 else
2704 rctx = 0;
2705
2706 if (recursion[rctx])
2707 return -1;
2708
2709 recursion[rctx]++;
2710 barrier();
2711
2712 return rctx;
2713}
2714
2715static inline void put_recursion_context(int *recursion, int rctx)
2716{
2717 barrier();
2718 recursion[rctx]--;
2719}
2720
2721static struct perf_callchain_entry *get_callchain_entry(int *rctx)
2722{
2723 int cpu;
2724 struct callchain_cpus_entries *entries;
2725
2726 *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
2727 if (*rctx == -1)
2728 return NULL;
2729
2730 entries = rcu_dereference(callchain_cpus_entries);
2731 if (!entries)
2732 return NULL;
2733
2734 cpu = smp_processor_id();
2735
2736 return &entries->cpu_entries[cpu][*rctx];
2737}
2738
2739static void
2740put_callchain_entry(int rctx)
2741{
2742 put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
2743}
2744
2745static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2746{
2747 int rctx;
2748 struct perf_callchain_entry *entry;
2749
2750
2751 entry = get_callchain_entry(&rctx);
2752 if (rctx == -1)
2753 return NULL;
2754
2755 if (!entry)
2756 goto exit_put;
2757
2758 entry->nr = 0;
2759
2760 if (!user_mode(regs)) {
2761 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
2762 perf_callchain_kernel(entry, regs);
2763 if (current->mm)
2764 regs = task_pt_regs(current);
2765 else
2766 regs = NULL;
2767 }
2768
2769 if (regs) {
2770 perf_callchain_store(entry, PERF_CONTEXT_USER);
2771 perf_callchain_user(entry, regs);
2772 }
2773
2774exit_put:
2775 put_callchain_entry(rctx);
2776
2777 return entry;
2778}
2779
2780/*
2781 * Initialize the perf_event context in a task_struct:
2782 */
2783static void __perf_event_init_context(struct perf_event_context *ctx)
2784{
2785 raw_spin_lock_init(&ctx->lock);
2786 mutex_init(&ctx->mutex);
2787 INIT_LIST_HEAD(&ctx->pinned_groups);
2788 INIT_LIST_HEAD(&ctx->flexible_groups);
2789 INIT_LIST_HEAD(&ctx->event_list);
2790 atomic_set(&ctx->refcount, 1);
2791}
2792
2793static struct perf_event_context *
2794alloc_perf_context(struct pmu *pmu, struct task_struct *task)
2795{
2796 struct perf_event_context *ctx;
2797
2798 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
2799 if (!ctx)
2800 return NULL;
2801
2802 __perf_event_init_context(ctx);
2803 if (task) {
2804 ctx->task = task;
2805 get_task_struct(task);
2806 }
2807 ctx->pmu = pmu;
2808
2809 return ctx;
2810}
2811
2812static struct task_struct *
2813find_lively_task_by_vpid(pid_t vpid)
2814{
2815 struct task_struct *task;
2816 int err;
2817
2818 rcu_read_lock();
2819 if (!vpid)
2820 task = current;
2821 else
2822 task = find_task_by_vpid(vpid);
2823 if (task)
2824 get_task_struct(task);
2825 rcu_read_unlock();
2826
2827 if (!task)
2828 return ERR_PTR(-ESRCH);
2829
2830 /* Reuse ptrace permission checks for now. */
2831 err = -EACCES;
2832 if (!ptrace_may_access(task, PTRACE_MODE_READ))
2833 goto errout;
2834
2835 return task;
2836errout:
2837 put_task_struct(task);
2838 return ERR_PTR(err);
2839
2840}
2841
2842/*
2843 * Returns a matching context with refcount and pincount.
2844 */
2845static struct perf_event_context *
2846find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
2847{
2848 struct perf_event_context *ctx;
2849 struct perf_cpu_context *cpuctx;
2850 unsigned long flags;
2851 int ctxn, err;
2852
2853 if (!task) {
2854 /* Must be root to operate on a CPU event: */
2855 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
2856 return ERR_PTR(-EACCES);
2857
2858 /*
2859 * We could be clever and allow to attach a event to an
2860 * offline CPU and activate it when the CPU comes up, but
2861 * that's for later.
2862 */
2863 if (!cpu_online(cpu))
2864 return ERR_PTR(-ENODEV);
2865
2866 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
2867 ctx = &cpuctx->ctx;
2868 get_ctx(ctx);
2869 ++ctx->pin_count;
2870
2871 return ctx;
2872 }
2873
2874 err = -EINVAL;
2875 ctxn = pmu->task_ctx_nr;
2876 if (ctxn < 0)
2877 goto errout;
2878
2879retry:
2880 ctx = perf_lock_task_context(task, ctxn, &flags);
2881 if (ctx) {
2882 unclone_ctx(ctx);
2883 ++ctx->pin_count;
2884 raw_spin_unlock_irqrestore(&ctx->lock, flags);
2885 } else {
2886 ctx = alloc_perf_context(pmu, task);
2887 err = -ENOMEM;
2888 if (!ctx)
2889 goto errout;
2890
2891 err = 0;
2892 mutex_lock(&task->perf_event_mutex);
2893 /*
2894 * If it has already passed perf_event_exit_task().
2895 * we must see PF_EXITING, it takes this mutex too.
2896 */
2897 if (task->flags & PF_EXITING)
2898 err = -ESRCH;
2899 else if (task->perf_event_ctxp[ctxn])
2900 err = -EAGAIN;
2901 else {
2902 get_ctx(ctx);
2903 ++ctx->pin_count;
2904 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
2905 }
2906 mutex_unlock(&task->perf_event_mutex);
2907
2908 if (unlikely(err)) {
2909 put_ctx(ctx);
2910
2911 if (err == -EAGAIN)
2912 goto retry;
2913 goto errout;
2914 }
2915 }
2916
2917 return ctx;
2918
2919errout:
2920 return ERR_PTR(err);
2921}
2922
2923static void perf_event_free_filter(struct perf_event *event);
2924
2925static void free_event_rcu(struct rcu_head *head)
2926{
2927 struct perf_event *event;
2928
2929 event = container_of(head, struct perf_event, rcu_head);
2930 if (event->ns)
2931 put_pid_ns(event->ns);
2932 perf_event_free_filter(event);
2933 kfree(event);
2934}
2935
2936static void ring_buffer_put(struct ring_buffer *rb);
2937
2938static void free_event(struct perf_event *event)
2939{
2940 irq_work_sync(&event->pending);
2941
2942 if (!event->parent) {
2943 if (event->attach_state & PERF_ATTACH_TASK)
2944 jump_label_dec(&perf_sched_events);
2945 if (event->attr.mmap || event->attr.mmap_data)
2946 atomic_dec(&nr_mmap_events);
2947 if (event->attr.comm)
2948 atomic_dec(&nr_comm_events);
2949 if (event->attr.task)
2950 atomic_dec(&nr_task_events);
2951 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
2952 put_callchain_buffers();
2953 if (is_cgroup_event(event)) {
2954 atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
2955 jump_label_dec(&perf_sched_events);
2956 }
2957 }
2958
2959 if (event->rb) {
2960 ring_buffer_put(event->rb);
2961 event->rb = NULL;
2962 }
2963
2964 if (is_cgroup_event(event))
2965 perf_detach_cgroup(event);
2966
2967 if (event->destroy)
2968 event->destroy(event);
2969
2970 if (event->ctx)
2971 put_ctx(event->ctx);
2972
2973 call_rcu(&event->rcu_head, free_event_rcu);
2974}
2975
2976int perf_event_release_kernel(struct perf_event *event)
2977{
2978 struct perf_event_context *ctx = event->ctx;
2979
2980 WARN_ON_ONCE(ctx->parent_ctx);
2981 /*
2982 * There are two ways this annotation is useful:
2983 *
2984 * 1) there is a lock recursion from perf_event_exit_task
2985 * see the comment there.
2986 *
2987 * 2) there is a lock-inversion with mmap_sem through
2988 * perf_event_read_group(), which takes faults while
2989 * holding ctx->mutex, however this is called after
2990 * the last filedesc died, so there is no possibility
2991 * to trigger the AB-BA case.
2992 */
2993 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
2994 raw_spin_lock_irq(&ctx->lock);
2995 perf_group_detach(event);
2996 raw_spin_unlock_irq(&ctx->lock);
2997 perf_remove_from_context(event);
2998 mutex_unlock(&ctx->mutex);
2999
3000 free_event(event);
3001
3002 return 0;
3003}
3004EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3005
3006/*
3007 * Called when the last reference to the file is gone.
3008 */
3009static int perf_release(struct inode *inode, struct file *file)
3010{
3011 struct perf_event *event = file->private_data;
3012 struct task_struct *owner;
3013
3014 file->private_data = NULL;
3015
3016 rcu_read_lock();
3017 owner = ACCESS_ONCE(event->owner);
3018 /*
3019 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3020 * !owner it means the list deletion is complete and we can indeed
3021 * free this event, otherwise we need to serialize on
3022 * owner->perf_event_mutex.
3023 */
3024 smp_read_barrier_depends();
3025 if (owner) {
3026 /*
3027 * Since delayed_put_task_struct() also drops the last
3028 * task reference we can safely take a new reference
3029 * while holding the rcu_read_lock().
3030 */
3031 get_task_struct(owner);
3032 }
3033 rcu_read_unlock();
3034
3035 if (owner) {
3036 mutex_lock(&owner->perf_event_mutex);
3037 /*
3038 * We have to re-check the event->owner field, if it is cleared
3039 * we raced with perf_event_exit_task(), acquiring the mutex
3040 * ensured they're done, and we can proceed with freeing the
3041 * event.
3042 */
3043 if (event->owner)
3044 list_del_init(&event->owner_entry);
3045 mutex_unlock(&owner->perf_event_mutex);
3046 put_task_struct(owner);
3047 }
3048
3049 return perf_event_release_kernel(event);
3050}
3051
3052u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3053{
3054 struct perf_event *child;
3055 u64 total = 0;
3056
3057 *enabled = 0;
3058 *running = 0;
3059
3060 mutex_lock(&event->child_mutex);
3061 total += perf_event_read(event);
3062 *enabled += event->total_time_enabled +
3063 atomic64_read(&event->child_total_time_enabled);
3064 *running += event->total_time_running +
3065 atomic64_read(&event->child_total_time_running);
3066
3067 list_for_each_entry(child, &event->child_list, child_list) {
3068 total += perf_event_read(child);
3069 *enabled += child->total_time_enabled;
3070 *running += child->total_time_running;
3071 }
3072 mutex_unlock(&event->child_mutex);
3073
3074 return total;
3075}
3076EXPORT_SYMBOL_GPL(perf_event_read_value);
3077
3078static int perf_event_read_group(struct perf_event *event,
3079 u64 read_format, char __user *buf)
3080{
3081 struct perf_event *leader = event->group_leader, *sub;
3082 int n = 0, size = 0, ret = -EFAULT;
3083 struct perf_event_context *ctx = leader->ctx;
3084 u64 values[5];
3085 u64 count, enabled, running;
3086
3087 mutex_lock(&ctx->mutex);
3088 count = perf_event_read_value(leader, &enabled, &running);
3089
3090 values[n++] = 1 + leader->nr_siblings;
3091 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3092 values[n++] = enabled;
3093 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3094 values[n++] = running;
3095 values[n++] = count;
3096 if (read_format & PERF_FORMAT_ID)
3097 values[n++] = primary_event_id(leader);
3098
3099 size = n * sizeof(u64);
3100
3101 if (copy_to_user(buf, values, size))
3102 goto unlock;
3103
3104 ret = size;
3105
3106 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3107 n = 0;
3108
3109 values[n++] = perf_event_read_value(sub, &enabled, &running);
3110 if (read_format & PERF_FORMAT_ID)
3111 values[n++] = primary_event_id(sub);
3112
3113 size = n * sizeof(u64);
3114
3115 if (copy_to_user(buf + ret, values, size)) {
3116 ret = -EFAULT;
3117 goto unlock;
3118 }
3119
3120 ret += size;
3121 }
3122unlock:
3123 mutex_unlock(&ctx->mutex);
3124
3125 return ret;
3126}
3127
3128static int perf_event_read_one(struct perf_event *event,
3129 u64 read_format, char __user *buf)
3130{
3131 u64 enabled, running;
3132 u64 values[4];
3133 int n = 0;
3134
3135 values[n++] = perf_event_read_value(event, &enabled, &running);
3136 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3137 values[n++] = enabled;
3138 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3139 values[n++] = running;
3140 if (read_format & PERF_FORMAT_ID)
3141 values[n++] = primary_event_id(event);
3142
3143 if (copy_to_user(buf, values, n * sizeof(u64)))
3144 return -EFAULT;
3145
3146 return n * sizeof(u64);
3147}
3148
3149/*
3150 * Read the performance event - simple non blocking version for now
3151 */
3152static ssize_t
3153perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3154{
3155 u64 read_format = event->attr.read_format;
3156 int ret;
3157
3158 /*
3159 * Return end-of-file for a read on a event that is in
3160 * error state (i.e. because it was pinned but it couldn't be
3161 * scheduled on to the CPU at some point).
3162 */
3163 if (event->state == PERF_EVENT_STATE_ERROR)
3164 return 0;
3165
3166 if (count < event->read_size)
3167 return -ENOSPC;
3168
3169 WARN_ON_ONCE(event->ctx->parent_ctx);
3170 if (read_format & PERF_FORMAT_GROUP)
3171 ret = perf_event_read_group(event, read_format, buf);
3172 else
3173 ret = perf_event_read_one(event, read_format, buf);
3174
3175 return ret;
3176}
3177
3178static ssize_t
3179perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3180{
3181 struct perf_event *event = file->private_data;
3182
3183 return perf_read_hw(event, buf, count);
3184}
3185
3186static unsigned int perf_poll(struct file *file, poll_table *wait)
3187{
3188 struct perf_event *event = file->private_data;
3189 struct ring_buffer *rb;
3190 unsigned int events = POLL_HUP;
3191
3192 rcu_read_lock();
3193 rb = rcu_dereference(event->rb);
3194 if (rb)
3195 events = atomic_xchg(&rb->poll, 0);
3196 rcu_read_unlock();
3197
3198 poll_wait(file, &event->waitq, wait);
3199
3200 return events;
3201}
3202
3203static void perf_event_reset(struct perf_event *event)
3204{
3205 (void)perf_event_read(event);
3206 local64_set(&event->count, 0);
3207 perf_event_update_userpage(event);
3208}
3209
3210/*
3211 * Holding the top-level event's child_mutex means that any
3212 * descendant process that has inherited this event will block
3213 * in sync_child_event if it goes to exit, thus satisfying the
3214 * task existence requirements of perf_event_enable/disable.
3215 */
3216static void perf_event_for_each_child(struct perf_event *event,
3217 void (*func)(struct perf_event *))
3218{
3219 struct perf_event *child;
3220
3221 WARN_ON_ONCE(event->ctx->parent_ctx);
3222 mutex_lock(&event->child_mutex);
3223 func(event);
3224 list_for_each_entry(child, &event->child_list, child_list)
3225 func(child);
3226 mutex_unlock(&event->child_mutex);
3227}
3228
3229static void perf_event_for_each(struct perf_event *event,
3230 void (*func)(struct perf_event *))
3231{
3232 struct perf_event_context *ctx = event->ctx;
3233 struct perf_event *sibling;
3234
3235 WARN_ON_ONCE(ctx->parent_ctx);
3236 mutex_lock(&ctx->mutex);
3237 event = event->group_leader;
3238
3239 perf_event_for_each_child(event, func);
3240 func(event);
3241 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3242 perf_event_for_each_child(event, func);
3243 mutex_unlock(&ctx->mutex);
3244}
3245
3246static int perf_event_period(struct perf_event *event, u64 __user *arg)
3247{
3248 struct perf_event_context *ctx = event->ctx;
3249 int ret = 0;
3250 u64 value;
3251
3252 if (!is_sampling_event(event))
3253 return -EINVAL;
3254
3255 if (copy_from_user(&value, arg, sizeof(value)))
3256 return -EFAULT;
3257
3258 if (!value)
3259 return -EINVAL;
3260
3261 raw_spin_lock_irq(&ctx->lock);
3262 if (event->attr.freq) {
3263 if (value > sysctl_perf_event_sample_rate) {
3264 ret = -EINVAL;
3265 goto unlock;
3266 }
3267
3268 event->attr.sample_freq = value;
3269 } else {
3270 event->attr.sample_period = value;
3271 event->hw.sample_period = value;
3272 }
3273unlock:
3274 raw_spin_unlock_irq(&ctx->lock);
3275
3276 return ret;
3277}
3278
3279static const struct file_operations perf_fops;
3280
3281static struct perf_event *perf_fget_light(int fd, int *fput_needed)
3282{
3283 struct file *file;
3284
3285 file = fget_light(fd, fput_needed);
3286 if (!file)
3287 return ERR_PTR(-EBADF);
3288
3289 if (file->f_op != &perf_fops) {
3290 fput_light(file, *fput_needed);
3291 *fput_needed = 0;
3292 return ERR_PTR(-EBADF);
3293 }
3294
3295 return file->private_data;
3296}
3297
3298static int perf_event_set_output(struct perf_event *event,
3299 struct perf_event *output_event);
3300static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3301
3302static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3303{
3304 struct perf_event *event = file->private_data;
3305 void (*func)(struct perf_event *);
3306 u32 flags = arg;
3307
3308 switch (cmd) {
3309 case PERF_EVENT_IOC_ENABLE:
3310 func = perf_event_enable;
3311 break;
3312 case PERF_EVENT_IOC_DISABLE:
3313 func = perf_event_disable;
3314 break;
3315 case PERF_EVENT_IOC_RESET:
3316 func = perf_event_reset;
3317 break;
3318
3319 case PERF_EVENT_IOC_REFRESH:
3320 return perf_event_refresh(event, arg);
3321
3322 case PERF_EVENT_IOC_PERIOD:
3323 return perf_event_period(event, (u64 __user *)arg);
3324
3325 case PERF_EVENT_IOC_SET_OUTPUT:
3326 {
3327 struct perf_event *output_event = NULL;
3328 int fput_needed = 0;
3329 int ret;
3330
3331 if (arg != -1) {
3332 output_event = perf_fget_light(arg, &fput_needed);
3333 if (IS_ERR(output_event))
3334 return PTR_ERR(output_event);
3335 }
3336
3337 ret = perf_event_set_output(event, output_event);
3338 if (output_event)
3339 fput_light(output_event->filp, fput_needed);
3340
3341 return ret;
3342 }
3343
3344 case PERF_EVENT_IOC_SET_FILTER:
3345 return perf_event_set_filter(event, (void __user *)arg);
3346
3347 default:
3348 return -ENOTTY;
3349 }
3350
3351 if (flags & PERF_IOC_FLAG_GROUP)
3352 perf_event_for_each(event, func);
3353 else
3354 perf_event_for_each_child(event, func);
3355
3356 return 0;
3357}
3358
3359int perf_event_task_enable(void)
3360{
3361 struct perf_event *event;
3362
3363 mutex_lock(¤t->perf_event_mutex);
3364 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3365 perf_event_for_each_child(event, perf_event_enable);
3366 mutex_unlock(¤t->perf_event_mutex);
3367
3368 return 0;
3369}
3370
3371int perf_event_task_disable(void)
3372{
3373 struct perf_event *event;
3374
3375 mutex_lock(¤t->perf_event_mutex);
3376 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3377 perf_event_for_each_child(event, perf_event_disable);
3378 mutex_unlock(¤t->perf_event_mutex);
3379
3380 return 0;
3381}
3382
3383#ifndef PERF_EVENT_INDEX_OFFSET
3384# define PERF_EVENT_INDEX_OFFSET 0
3385#endif
3386
3387static int perf_event_index(struct perf_event *event)
3388{
3389 if (event->hw.state & PERF_HES_STOPPED)
3390 return 0;
3391
3392 if (event->state != PERF_EVENT_STATE_ACTIVE)
3393 return 0;
3394
3395 return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
3396}
3397
3398static void calc_timer_values(struct perf_event *event,
3399 u64 *enabled,
3400 u64 *running)
3401{
3402 u64 now, ctx_time;
3403
3404 now = perf_clock();
3405 ctx_time = event->shadow_ctx_time + now;
3406 *enabled = ctx_time - event->tstamp_enabled;
3407 *running = ctx_time - event->tstamp_running;
3408}
3409
3410/*
3411 * Callers need to ensure there can be no nesting of this function, otherwise
3412 * the seqlock logic goes bad. We can not serialize this because the arch
3413 * code calls this from NMI context.
3414 */
3415void perf_event_update_userpage(struct perf_event *event)
3416{
3417 struct perf_event_mmap_page *userpg;
3418 struct ring_buffer *rb;
3419 u64 enabled, running;
3420
3421 rcu_read_lock();
3422 /*
3423 * compute total_time_enabled, total_time_running
3424 * based on snapshot values taken when the event
3425 * was last scheduled in.
3426 *
3427 * we cannot simply called update_context_time()
3428 * because of locking issue as we can be called in
3429 * NMI context
3430 */
3431 calc_timer_values(event, &enabled, &running);
3432 rb = rcu_dereference(event->rb);
3433 if (!rb)
3434 goto unlock;
3435
3436 userpg = rb->user_page;
3437
3438 /*
3439 * Disable preemption so as to not let the corresponding user-space
3440 * spin too long if we get preempted.
3441 */
3442 preempt_disable();
3443 ++userpg->lock;
3444 barrier();
3445 userpg->index = perf_event_index(event);
3446 userpg->offset = perf_event_count(event);
3447 if (event->state == PERF_EVENT_STATE_ACTIVE)
3448 userpg->offset -= local64_read(&event->hw.prev_count);
3449
3450 userpg->time_enabled = enabled +
3451 atomic64_read(&event->child_total_time_enabled);
3452
3453 userpg->time_running = running +
3454 atomic64_read(&event->child_total_time_running);
3455
3456 barrier();
3457 ++userpg->lock;
3458 preempt_enable();
3459unlock:
3460 rcu_read_unlock();
3461}
3462
3463static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3464{
3465 struct perf_event *event = vma->vm_file->private_data;
3466 struct ring_buffer *rb;
3467 int ret = VM_FAULT_SIGBUS;
3468
3469 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3470 if (vmf->pgoff == 0)
3471 ret = 0;
3472 return ret;
3473 }
3474
3475 rcu_read_lock();
3476 rb = rcu_dereference(event->rb);
3477 if (!rb)
3478 goto unlock;
3479
3480 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3481 goto unlock;
3482
3483 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3484 if (!vmf->page)
3485 goto unlock;
3486
3487 get_page(vmf->page);
3488 vmf->page->mapping = vma->vm_file->f_mapping;
3489 vmf->page->index = vmf->pgoff;
3490
3491 ret = 0;
3492unlock:
3493 rcu_read_unlock();
3494
3495 return ret;
3496}
3497
3498static void rb_free_rcu(struct rcu_head *rcu_head)
3499{
3500 struct ring_buffer *rb;
3501
3502 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3503 rb_free(rb);
3504}
3505
3506static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3507{
3508 struct ring_buffer *rb;
3509
3510 rcu_read_lock();
3511 rb = rcu_dereference(event->rb);
3512 if (rb) {
3513 if (!atomic_inc_not_zero(&rb->refcount))
3514 rb = NULL;
3515 }
3516 rcu_read_unlock();
3517
3518 return rb;
3519}
3520
3521static void ring_buffer_put(struct ring_buffer *rb)
3522{
3523 if (!atomic_dec_and_test(&rb->refcount))
3524 return;
3525
3526 call_rcu(&rb->rcu_head, rb_free_rcu);
3527}
3528
3529static void perf_mmap_open(struct vm_area_struct *vma)
3530{
3531 struct perf_event *event = vma->vm_file->private_data;
3532
3533 atomic_inc(&event->mmap_count);
3534}
3535
3536static void perf_mmap_close(struct vm_area_struct *vma)
3537{
3538 struct perf_event *event = vma->vm_file->private_data;
3539
3540 if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
3541 unsigned long size = perf_data_size(event->rb);
3542 struct user_struct *user = event->mmap_user;
3543 struct ring_buffer *rb = event->rb;
3544
3545 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
3546 vma->vm_mm->locked_vm -= event->mmap_locked;
3547 rcu_assign_pointer(event->rb, NULL);
3548 mutex_unlock(&event->mmap_mutex);
3549
3550 ring_buffer_put(rb);
3551 free_uid(user);
3552 }
3553}
3554
3555static const struct vm_operations_struct perf_mmap_vmops = {
3556 .open = perf_mmap_open,
3557 .close = perf_mmap_close,
3558 .fault = perf_mmap_fault,
3559 .page_mkwrite = perf_mmap_fault,
3560};
3561
3562static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3563{
3564 struct perf_event *event = file->private_data;
3565 unsigned long user_locked, user_lock_limit;
3566 struct user_struct *user = current_user();
3567 unsigned long locked, lock_limit;
3568 struct ring_buffer *rb;
3569 unsigned long vma_size;
3570 unsigned long nr_pages;
3571 long user_extra, extra;
3572 int ret = 0, flags = 0;
3573
3574 /*
3575 * Don't allow mmap() of inherited per-task counters. This would
3576 * create a performance issue due to all children writing to the
3577 * same rb.
3578 */
3579 if (event->cpu == -1 && event->attr.inherit)
3580 return -EINVAL;
3581
3582 if (!(vma->vm_flags & VM_SHARED))
3583 return -EINVAL;
3584
3585 vma_size = vma->vm_end - vma->vm_start;
3586 nr_pages = (vma_size / PAGE_SIZE) - 1;
3587
3588 /*
3589 * If we have rb pages ensure they're a power-of-two number, so we
3590 * can do bitmasks instead of modulo.
3591 */
3592 if (nr_pages != 0 && !is_power_of_2(nr_pages))
3593 return -EINVAL;
3594
3595 if (vma_size != PAGE_SIZE * (1 + nr_pages))
3596 return -EINVAL;
3597
3598 if (vma->vm_pgoff != 0)
3599 return -EINVAL;
3600
3601 WARN_ON_ONCE(event->ctx->parent_ctx);
3602 mutex_lock(&event->mmap_mutex);
3603 if (event->rb) {
3604 if (event->rb->nr_pages == nr_pages)
3605 atomic_inc(&event->rb->refcount);
3606 else
3607 ret = -EINVAL;
3608 goto unlock;
3609 }
3610
3611 user_extra = nr_pages + 1;
3612 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
3613
3614 /*
3615 * Increase the limit linearly with more CPUs:
3616 */
3617 user_lock_limit *= num_online_cpus();
3618
3619 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
3620
3621 extra = 0;
3622 if (user_locked > user_lock_limit)
3623 extra = user_locked - user_lock_limit;
3624
3625 lock_limit = rlimit(RLIMIT_MEMLOCK);
3626 lock_limit >>= PAGE_SHIFT;
3627 locked = vma->vm_mm->locked_vm + extra;
3628
3629 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
3630 !capable(CAP_IPC_LOCK)) {
3631 ret = -EPERM;
3632 goto unlock;
3633 }
3634
3635 WARN_ON(event->rb);
3636
3637 if (vma->vm_flags & VM_WRITE)
3638 flags |= RING_BUFFER_WRITABLE;
3639
3640 rb = rb_alloc(nr_pages,
3641 event->attr.watermark ? event->attr.wakeup_watermark : 0,
3642 event->cpu, flags);
3643
3644 if (!rb) {
3645 ret = -ENOMEM;
3646 goto unlock;
3647 }
3648 rcu_assign_pointer(event->rb, rb);
3649
3650 atomic_long_add(user_extra, &user->locked_vm);
3651 event->mmap_locked = extra;
3652 event->mmap_user = get_current_user();
3653 vma->vm_mm->locked_vm += event->mmap_locked;
3654
3655unlock:
3656 if (!ret)
3657 atomic_inc(&event->mmap_count);
3658 mutex_unlock(&event->mmap_mutex);
3659
3660 vma->vm_flags |= VM_RESERVED;
3661 vma->vm_ops = &perf_mmap_vmops;
3662
3663 return ret;
3664}
3665
3666static int perf_fasync(int fd, struct file *filp, int on)
3667{
3668 struct inode *inode = filp->f_path.dentry->d_inode;
3669 struct perf_event *event = filp->private_data;
3670 int retval;
3671
3672 mutex_lock(&inode->i_mutex);
3673 retval = fasync_helper(fd, filp, on, &event->fasync);
3674 mutex_unlock(&inode->i_mutex);
3675
3676 if (retval < 0)
3677 return retval;
3678
3679 return 0;
3680}
3681
3682static const struct file_operations perf_fops = {
3683 .llseek = no_llseek,
3684 .release = perf_release,
3685 .read = perf_read,
3686 .poll = perf_poll,
3687 .unlocked_ioctl = perf_ioctl,
3688 .compat_ioctl = perf_ioctl,
3689 .mmap = perf_mmap,
3690 .fasync = perf_fasync,
3691};
3692
3693/*
3694 * Perf event wakeup
3695 *
3696 * If there's data, ensure we set the poll() state and publish everything
3697 * to user-space before waking everybody up.
3698 */
3699
3700void perf_event_wakeup(struct perf_event *event)
3701{
3702 wake_up_all(&event->waitq);
3703
3704 if (event->pending_kill) {
3705 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
3706 event->pending_kill = 0;
3707 }
3708}
3709
3710static void perf_pending_event(struct irq_work *entry)
3711{
3712 struct perf_event *event = container_of(entry,
3713 struct perf_event, pending);
3714
3715 if (event->pending_disable) {
3716 event->pending_disable = 0;
3717 __perf_event_disable(event);
3718 }
3719
3720 if (event->pending_wakeup) {
3721 event->pending_wakeup = 0;
3722 perf_event_wakeup(event);
3723 }
3724}
3725
3726/*
3727 * We assume there is only KVM supporting the callbacks.
3728 * Later on, we might change it to a list if there is
3729 * another virtualization implementation supporting the callbacks.
3730 */
3731struct perf_guest_info_callbacks *perf_guest_cbs;
3732
3733int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3734{
3735 perf_guest_cbs = cbs;
3736 return 0;
3737}
3738EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
3739
3740int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
3741{
3742 perf_guest_cbs = NULL;
3743 return 0;
3744}
3745EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
3746
3747static void __perf_event_header__init_id(struct perf_event_header *header,
3748 struct perf_sample_data *data,
3749 struct perf_event *event)
3750{
3751 u64 sample_type = event->attr.sample_type;
3752
3753 data->type = sample_type;
3754 header->size += event->id_header_size;
3755
3756 if (sample_type & PERF_SAMPLE_TID) {
3757 /* namespace issues */
3758 data->tid_entry.pid = perf_event_pid(event, current);
3759 data->tid_entry.tid = perf_event_tid(event, current);
3760 }
3761
3762 if (sample_type & PERF_SAMPLE_TIME)
3763 data->time = perf_clock();
3764
3765 if (sample_type & PERF_SAMPLE_ID)
3766 data->id = primary_event_id(event);
3767
3768 if (sample_type & PERF_SAMPLE_STREAM_ID)
3769 data->stream_id = event->id;
3770
3771 if (sample_type & PERF_SAMPLE_CPU) {
3772 data->cpu_entry.cpu = raw_smp_processor_id();
3773 data->cpu_entry.reserved = 0;
3774 }
3775}
3776
3777void perf_event_header__init_id(struct perf_event_header *header,
3778 struct perf_sample_data *data,
3779 struct perf_event *event)
3780{
3781 if (event->attr.sample_id_all)
3782 __perf_event_header__init_id(header, data, event);
3783}
3784
3785static void __perf_event__output_id_sample(struct perf_output_handle *handle,
3786 struct perf_sample_data *data)
3787{
3788 u64 sample_type = data->type;
3789
3790 if (sample_type & PERF_SAMPLE_TID)
3791 perf_output_put(handle, data->tid_entry);
3792
3793 if (sample_type & PERF_SAMPLE_TIME)
3794 perf_output_put(handle, data->time);
3795
3796 if (sample_type & PERF_SAMPLE_ID)
3797 perf_output_put(handle, data->id);
3798
3799 if (sample_type & PERF_SAMPLE_STREAM_ID)
3800 perf_output_put(handle, data->stream_id);
3801
3802 if (sample_type & PERF_SAMPLE_CPU)
3803 perf_output_put(handle, data->cpu_entry);
3804}
3805
3806void perf_event__output_id_sample(struct perf_event *event,
3807 struct perf_output_handle *handle,
3808 struct perf_sample_data *sample)
3809{
3810 if (event->attr.sample_id_all)
3811 __perf_event__output_id_sample(handle, sample);
3812}
3813
3814static void perf_output_read_one(struct perf_output_handle *handle,
3815 struct perf_event *event,
3816 u64 enabled, u64 running)
3817{
3818 u64 read_format = event->attr.read_format;
3819 u64 values[4];
3820 int n = 0;
3821
3822 values[n++] = perf_event_count(event);
3823 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
3824 values[n++] = enabled +
3825 atomic64_read(&event->child_total_time_enabled);
3826 }
3827 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
3828 values[n++] = running +
3829 atomic64_read(&event->child_total_time_running);
3830 }
3831 if (read_format & PERF_FORMAT_ID)
3832 values[n++] = primary_event_id(event);
3833
3834 __output_copy(handle, values, n * sizeof(u64));
3835}
3836
3837/*
3838 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
3839 */
3840static void perf_output_read_group(struct perf_output_handle *handle,
3841 struct perf_event *event,
3842 u64 enabled, u64 running)
3843{
3844 struct perf_event *leader = event->group_leader, *sub;
3845 u64 read_format = event->attr.read_format;
3846 u64 values[5];
3847 int n = 0;
3848
3849 values[n++] = 1 + leader->nr_siblings;
3850
3851 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3852 values[n++] = enabled;
3853
3854 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3855 values[n++] = running;
3856
3857 if (leader != event)
3858 leader->pmu->read(leader);
3859
3860 values[n++] = perf_event_count(leader);
3861 if (read_format & PERF_FORMAT_ID)
3862 values[n++] = primary_event_id(leader);
3863
3864 __output_copy(handle, values, n * sizeof(u64));
3865
3866 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3867 n = 0;
3868
3869 if (sub != event)
3870 sub->pmu->read(sub);
3871
3872 values[n++] = perf_event_count(sub);
3873 if (read_format & PERF_FORMAT_ID)
3874 values[n++] = primary_event_id(sub);
3875
3876 __output_copy(handle, values, n * sizeof(u64));
3877 }
3878}
3879
3880#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
3881 PERF_FORMAT_TOTAL_TIME_RUNNING)
3882
3883static void perf_output_read(struct perf_output_handle *handle,
3884 struct perf_event *event)
3885{
3886 u64 enabled = 0, running = 0;
3887 u64 read_format = event->attr.read_format;
3888
3889 /*
3890 * compute total_time_enabled, total_time_running
3891 * based on snapshot values taken when the event
3892 * was last scheduled in.
3893 *
3894 * we cannot simply called update_context_time()
3895 * because of locking issue as we are called in
3896 * NMI context
3897 */
3898 if (read_format & PERF_FORMAT_TOTAL_TIMES)
3899 calc_timer_values(event, &enabled, &running);
3900
3901 if (event->attr.read_format & PERF_FORMAT_GROUP)
3902 perf_output_read_group(handle, event, enabled, running);
3903 else
3904 perf_output_read_one(handle, event, enabled, running);
3905}
3906
3907void perf_output_sample(struct perf_output_handle *handle,
3908 struct perf_event_header *header,
3909 struct perf_sample_data *data,
3910 struct perf_event *event)
3911{
3912 u64 sample_type = data->type;
3913
3914 perf_output_put(handle, *header);
3915
3916 if (sample_type & PERF_SAMPLE_IP)
3917 perf_output_put(handle, data->ip);
3918
3919 if (sample_type & PERF_SAMPLE_TID)
3920 perf_output_put(handle, data->tid_entry);
3921
3922 if (sample_type & PERF_SAMPLE_TIME)
3923 perf_output_put(handle, data->time);
3924
3925 if (sample_type & PERF_SAMPLE_ADDR)
3926 perf_output_put(handle, data->addr);
3927
3928 if (sample_type & PERF_SAMPLE_ID)
3929 perf_output_put(handle, data->id);
3930
3931 if (sample_type & PERF_SAMPLE_STREAM_ID)
3932 perf_output_put(handle, data->stream_id);
3933
3934 if (sample_type & PERF_SAMPLE_CPU)
3935 perf_output_put(handle, data->cpu_entry);
3936
3937 if (sample_type & PERF_SAMPLE_PERIOD)
3938 perf_output_put(handle, data->period);
3939
3940 if (sample_type & PERF_SAMPLE_READ)
3941 perf_output_read(handle, event);
3942
3943 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3944 if (data->callchain) {
3945 int size = 1;
3946
3947 if (data->callchain)
3948 size += data->callchain->nr;
3949
3950 size *= sizeof(u64);
3951
3952 __output_copy(handle, data->callchain, size);
3953 } else {
3954 u64 nr = 0;
3955 perf_output_put(handle, nr);
3956 }
3957 }
3958
3959 if (sample_type & PERF_SAMPLE_RAW) {
3960 if (data->raw) {
3961 perf_output_put(handle, data->raw->size);
3962 __output_copy(handle, data->raw->data,
3963 data->raw->size);
3964 } else {
3965 struct {
3966 u32 size;
3967 u32 data;
3968 } raw = {
3969 .size = sizeof(u32),
3970 .data = 0,
3971 };
3972 perf_output_put(handle, raw);
3973 }
3974 }
3975
3976 if (!event->attr.watermark) {
3977 int wakeup_events = event->attr.wakeup_events;
3978
3979 if (wakeup_events) {
3980 struct ring_buffer *rb = handle->rb;
3981 int events = local_inc_return(&rb->events);
3982
3983 if (events >= wakeup_events) {
3984 local_sub(wakeup_events, &rb->events);
3985 local_inc(&rb->wakeup);
3986 }
3987 }
3988 }
3989}
3990
3991void perf_prepare_sample(struct perf_event_header *header,
3992 struct perf_sample_data *data,
3993 struct perf_event *event,
3994 struct pt_regs *regs)
3995{
3996 u64 sample_type = event->attr.sample_type;
3997
3998 header->type = PERF_RECORD_SAMPLE;
3999 header->size = sizeof(*header) + event->header_size;
4000
4001 header->misc = 0;
4002 header->misc |= perf_misc_flags(regs);
4003
4004 __perf_event_header__init_id(header, data, event);
4005
4006 if (sample_type & PERF_SAMPLE_IP)
4007 data->ip = perf_instruction_pointer(regs);
4008
4009 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4010 int size = 1;
4011
4012 data->callchain = perf_callchain(regs);
4013
4014 if (data->callchain)
4015 size += data->callchain->nr;
4016
4017 header->size += size * sizeof(u64);
4018 }
4019
4020 if (sample_type & PERF_SAMPLE_RAW) {
4021 int size = sizeof(u32);
4022
4023 if (data->raw)
4024 size += data->raw->size;
4025 else
4026 size += sizeof(u32);
4027
4028 WARN_ON_ONCE(size & (sizeof(u64)-1));
4029 header->size += size;
4030 }
4031}
4032
4033static void perf_event_output(struct perf_event *event,
4034 struct perf_sample_data *data,
4035 struct pt_regs *regs)
4036{
4037 struct perf_output_handle handle;
4038 struct perf_event_header header;
4039
4040 /* protect the callchain buffers */
4041 rcu_read_lock();
4042
4043 perf_prepare_sample(&header, data, event, regs);
4044
4045 if (perf_output_begin(&handle, event, header.size))
4046 goto exit;
4047
4048 perf_output_sample(&handle, &header, data, event);
4049
4050 perf_output_end(&handle);
4051
4052exit:
4053 rcu_read_unlock();
4054}
4055
4056/*
4057 * read event_id
4058 */
4059
4060struct perf_read_event {
4061 struct perf_event_header header;
4062
4063 u32 pid;
4064 u32 tid;
4065};
4066
4067static void
4068perf_event_read_event(struct perf_event *event,
4069 struct task_struct *task)
4070{
4071 struct perf_output_handle handle;
4072 struct perf_sample_data sample;
4073 struct perf_read_event read_event = {
4074 .header = {
4075 .type = PERF_RECORD_READ,
4076 .misc = 0,
4077 .size = sizeof(read_event) + event->read_size,
4078 },
4079 .pid = perf_event_pid(event, task),
4080 .tid = perf_event_tid(event, task),
4081 };
4082 int ret;
4083
4084 perf_event_header__init_id(&read_event.header, &sample, event);
4085 ret = perf_output_begin(&handle, event, read_event.header.size);
4086 if (ret)
4087 return;
4088
4089 perf_output_put(&handle, read_event);
4090 perf_output_read(&handle, event);
4091 perf_event__output_id_sample(event, &handle, &sample);
4092
4093 perf_output_end(&handle);
4094}
4095
4096/*
4097 * task tracking -- fork/exit
4098 *
4099 * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
4100 */
4101
4102struct perf_task_event {
4103 struct task_struct *task;
4104 struct perf_event_context *task_ctx;
4105
4106 struct {
4107 struct perf_event_header header;
4108
4109 u32 pid;
4110 u32 ppid;
4111 u32 tid;
4112 u32 ptid;
4113 u64 time;
4114 } event_id;
4115};
4116
4117static void perf_event_task_output(struct perf_event *event,
4118 struct perf_task_event *task_event)
4119{
4120 struct perf_output_handle handle;
4121 struct perf_sample_data sample;
4122 struct task_struct *task = task_event->task;
4123 int ret, size = task_event->event_id.header.size;
4124
4125 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4126
4127 ret = perf_output_begin(&handle, event,
4128 task_event->event_id.header.size);
4129 if (ret)
4130 goto out;
4131
4132 task_event->event_id.pid = perf_event_pid(event, task);
4133 task_event->event_id.ppid = perf_event_pid(event, current);
4134
4135 task_event->event_id.tid = perf_event_tid(event, task);
4136 task_event->event_id.ptid = perf_event_tid(event, current);
4137
4138 perf_output_put(&handle, task_event->event_id);
4139
4140 perf_event__output_id_sample(event, &handle, &sample);
4141
4142 perf_output_end(&handle);
4143out:
4144 task_event->event_id.header.size = size;
4145}
4146
4147static int perf_event_task_match(struct perf_event *event)
4148{
4149 if (event->state < PERF_EVENT_STATE_INACTIVE)
4150 return 0;
4151
4152 if (!event_filter_match(event))
4153 return 0;
4154
4155 if (event->attr.comm || event->attr.mmap ||
4156 event->attr.mmap_data || event->attr.task)
4157 return 1;
4158
4159 return 0;
4160}
4161
4162static void perf_event_task_ctx(struct perf_event_context *ctx,
4163 struct perf_task_event *task_event)
4164{
4165 struct perf_event *event;
4166
4167 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4168 if (perf_event_task_match(event))
4169 perf_event_task_output(event, task_event);
4170 }
4171}
4172
4173static void perf_event_task_event(struct perf_task_event *task_event)
4174{
4175 struct perf_cpu_context *cpuctx;
4176 struct perf_event_context *ctx;
4177 struct pmu *pmu;
4178 int ctxn;
4179
4180 rcu_read_lock();
4181 list_for_each_entry_rcu(pmu, &pmus, entry) {
4182 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4183 if (cpuctx->active_pmu != pmu)
4184 goto next;
4185 perf_event_task_ctx(&cpuctx->ctx, task_event);
4186
4187 ctx = task_event->task_ctx;
4188 if (!ctx) {
4189 ctxn = pmu->task_ctx_nr;
4190 if (ctxn < 0)
4191 goto next;
4192 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4193 }
4194 if (ctx)
4195 perf_event_task_ctx(ctx, task_event);
4196next:
4197 put_cpu_ptr(pmu->pmu_cpu_context);
4198 }
4199 rcu_read_unlock();
4200}
4201
4202static void perf_event_task(struct task_struct *task,
4203 struct perf_event_context *task_ctx,
4204 int new)
4205{
4206 struct perf_task_event task_event;
4207
4208 if (!atomic_read(&nr_comm_events) &&
4209 !atomic_read(&nr_mmap_events) &&
4210 !atomic_read(&nr_task_events))
4211 return;
4212
4213 task_event = (struct perf_task_event){
4214 .task = task,
4215 .task_ctx = task_ctx,
4216 .event_id = {
4217 .header = {
4218 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4219 .misc = 0,
4220 .size = sizeof(task_event.event_id),
4221 },
4222 /* .pid */
4223 /* .ppid */
4224 /* .tid */
4225 /* .ptid */
4226 .time = perf_clock(),
4227 },
4228 };
4229
4230 perf_event_task_event(&task_event);
4231}
4232
4233void perf_event_fork(struct task_struct *task)
4234{
4235 perf_event_task(task, NULL, 1);
4236}
4237
4238/*
4239 * comm tracking
4240 */
4241
4242struct perf_comm_event {
4243 struct task_struct *task;
4244 char *comm;
4245 int comm_size;
4246
4247 struct {
4248 struct perf_event_header header;
4249
4250 u32 pid;
4251 u32 tid;
4252 } event_id;
4253};
4254
4255static void perf_event_comm_output(struct perf_event *event,
4256 struct perf_comm_event *comm_event)
4257{
4258 struct perf_output_handle handle;
4259 struct perf_sample_data sample;
4260 int size = comm_event->event_id.header.size;
4261 int ret;
4262
4263 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4264 ret = perf_output_begin(&handle, event,
4265 comm_event->event_id.header.size);
4266
4267 if (ret)
4268 goto out;
4269
4270 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4271 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4272
4273 perf_output_put(&handle, comm_event->event_id);
4274 __output_copy(&handle, comm_event->comm,
4275 comm_event->comm_size);
4276
4277 perf_event__output_id_sample(event, &handle, &sample);
4278
4279 perf_output_end(&handle);
4280out:
4281 comm_event->event_id.header.size = size;
4282}
4283
4284static int perf_event_comm_match(struct perf_event *event)
4285{
4286 if (event->state < PERF_EVENT_STATE_INACTIVE)
4287 return 0;
4288
4289 if (!event_filter_match(event))
4290 return 0;
4291
4292 if (event->attr.comm)
4293 return 1;
4294
4295 return 0;
4296}
4297
4298static void perf_event_comm_ctx(struct perf_event_context *ctx,
4299 struct perf_comm_event *comm_event)
4300{
4301 struct perf_event *event;
4302
4303 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4304 if (perf_event_comm_match(event))
4305 perf_event_comm_output(event, comm_event);
4306 }
4307}
4308
4309static void perf_event_comm_event(struct perf_comm_event *comm_event)
4310{
4311 struct perf_cpu_context *cpuctx;
4312 struct perf_event_context *ctx;
4313 char comm[TASK_COMM_LEN];
4314 unsigned int size;
4315 struct pmu *pmu;
4316 int ctxn;
4317
4318 memset(comm, 0, sizeof(comm));
4319 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4320 size = ALIGN(strlen(comm)+1, sizeof(u64));
4321
4322 comm_event->comm = comm;
4323 comm_event->comm_size = size;
4324
4325 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
4326 rcu_read_lock();
4327 list_for_each_entry_rcu(pmu, &pmus, entry) {
4328 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4329 if (cpuctx->active_pmu != pmu)
4330 goto next;
4331 perf_event_comm_ctx(&cpuctx->ctx, comm_event);
4332
4333 ctxn = pmu->task_ctx_nr;
4334 if (ctxn < 0)
4335 goto next;
4336
4337 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4338 if (ctx)
4339 perf_event_comm_ctx(ctx, comm_event);
4340next:
4341 put_cpu_ptr(pmu->pmu_cpu_context);
4342 }
4343 rcu_read_unlock();
4344}
4345
4346void perf_event_comm(struct task_struct *task)
4347{
4348 struct perf_comm_event comm_event;
4349 struct perf_event_context *ctx;
4350 int ctxn;
4351
4352 for_each_task_context_nr(ctxn) {
4353 ctx = task->perf_event_ctxp[ctxn];
4354 if (!ctx)
4355 continue;
4356
4357 perf_event_enable_on_exec(ctx);
4358 }
4359
4360 if (!atomic_read(&nr_comm_events))
4361 return;
4362
4363 comm_event = (struct perf_comm_event){
4364 .task = task,
4365 /* .comm */
4366 /* .comm_size */
4367 .event_id = {
4368 .header = {
4369 .type = PERF_RECORD_COMM,
4370 .misc = 0,
4371 /* .size */
4372 },
4373 /* .pid */
4374 /* .tid */
4375 },
4376 };
4377
4378 perf_event_comm_event(&comm_event);
4379}
4380
4381/*
4382 * mmap tracking
4383 */
4384
4385struct perf_mmap_event {
4386 struct vm_area_struct *vma;
4387
4388 const char *file_name;
4389 int file_size;
4390
4391 struct {
4392 struct perf_event_header header;
4393
4394 u32 pid;
4395 u32 tid;
4396 u64 start;
4397 u64 len;
4398 u64 pgoff;
4399 } event_id;
4400};
4401
4402static void perf_event_mmap_output(struct perf_event *event,
4403 struct perf_mmap_event *mmap_event)
4404{
4405 struct perf_output_handle handle;
4406 struct perf_sample_data sample;
4407 int size = mmap_event->event_id.header.size;
4408 int ret;
4409
4410 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
4411 ret = perf_output_begin(&handle, event,
4412 mmap_event->event_id.header.size);
4413 if (ret)
4414 goto out;
4415
4416 mmap_event->event_id.pid = perf_event_pid(event, current);
4417 mmap_event->event_id.tid = perf_event_tid(event, current);
4418
4419 perf_output_put(&handle, mmap_event->event_id);
4420 __output_copy(&handle, mmap_event->file_name,
4421 mmap_event->file_size);
4422
4423 perf_event__output_id_sample(event, &handle, &sample);
4424
4425 perf_output_end(&handle);
4426out:
4427 mmap_event->event_id.header.size = size;
4428}
4429
4430static int perf_event_mmap_match(struct perf_event *event,
4431 struct perf_mmap_event *mmap_event,
4432 int executable)
4433{
4434 if (event->state < PERF_EVENT_STATE_INACTIVE)
4435 return 0;
4436
4437 if (!event_filter_match(event))
4438 return 0;
4439
4440 if ((!executable && event->attr.mmap_data) ||
4441 (executable && event->attr.mmap))
4442 return 1;
4443
4444 return 0;
4445}
4446
4447static void perf_event_mmap_ctx(struct perf_event_context *ctx,
4448 struct perf_mmap_event *mmap_event,
4449 int executable)
4450{
4451 struct perf_event *event;
4452
4453 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4454 if (perf_event_mmap_match(event, mmap_event, executable))
4455 perf_event_mmap_output(event, mmap_event);
4456 }
4457}
4458
4459static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
4460{
4461 struct perf_cpu_context *cpuctx;
4462 struct perf_event_context *ctx;
4463 struct vm_area_struct *vma = mmap_event->vma;
4464 struct file *file = vma->vm_file;
4465 unsigned int size;
4466 char tmp[16];
4467 char *buf = NULL;
4468 const char *name;
4469 struct pmu *pmu;
4470 int ctxn;
4471
4472 memset(tmp, 0, sizeof(tmp));
4473
4474 if (file) {
4475 /*
4476 * d_path works from the end of the rb backwards, so we
4477 * need to add enough zero bytes after the string to handle
4478 * the 64bit alignment we do later.
4479 */
4480 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
4481 if (!buf) {
4482 name = strncpy(tmp, "//enomem", sizeof(tmp));
4483 goto got_name;
4484 }
4485 name = d_path(&file->f_path, buf, PATH_MAX);
4486 if (IS_ERR(name)) {
4487 name = strncpy(tmp, "//toolong", sizeof(tmp));
4488 goto got_name;
4489 }
4490 } else {
4491 if (arch_vma_name(mmap_event->vma)) {
4492 name = strncpy(tmp, arch_vma_name(mmap_event->vma),
4493 sizeof(tmp));
4494 goto got_name;
4495 }
4496
4497 if (!vma->vm_mm) {
4498 name = strncpy(tmp, "[vdso]", sizeof(tmp));
4499 goto got_name;
4500 } else if (vma->vm_start <= vma->vm_mm->start_brk &&
4501 vma->vm_end >= vma->vm_mm->brk) {
4502 name = strncpy(tmp, "[heap]", sizeof(tmp));
4503 goto got_name;
4504 } else if (vma->vm_start <= vma->vm_mm->start_stack &&
4505 vma->vm_end >= vma->vm_mm->start_stack) {
4506 name = strncpy(tmp, "[stack]", sizeof(tmp));
4507 goto got_name;
4508 }
4509
4510 name = strncpy(tmp, "//anon", sizeof(tmp));
4511 goto got_name;
4512 }
4513
4514got_name:
4515 size = ALIGN(strlen(name)+1, sizeof(u64));
4516
4517 mmap_event->file_name = name;
4518 mmap_event->file_size = size;
4519
4520 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
4521
4522 rcu_read_lock();
4523 list_for_each_entry_rcu(pmu, &pmus, entry) {
4524 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4525 if (cpuctx->active_pmu != pmu)
4526 goto next;
4527 perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
4528 vma->vm_flags & VM_EXEC);
4529
4530 ctxn = pmu->task_ctx_nr;
4531 if (ctxn < 0)
4532 goto next;
4533
4534 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4535 if (ctx) {
4536 perf_event_mmap_ctx(ctx, mmap_event,
4537 vma->vm_flags & VM_EXEC);
4538 }
4539next:
4540 put_cpu_ptr(pmu->pmu_cpu_context);
4541 }
4542 rcu_read_unlock();
4543
4544 kfree(buf);
4545}
4546
4547void perf_event_mmap(struct vm_area_struct *vma)
4548{
4549 struct perf_mmap_event mmap_event;
4550
4551 if (!atomic_read(&nr_mmap_events))
4552 return;
4553
4554 mmap_event = (struct perf_mmap_event){
4555 .vma = vma,
4556 /* .file_name */
4557 /* .file_size */
4558 .event_id = {
4559 .header = {
4560 .type = PERF_RECORD_MMAP,
4561 .misc = PERF_RECORD_MISC_USER,
4562 /* .size */
4563 },
4564 /* .pid */
4565 /* .tid */
4566 .start = vma->vm_start,
4567 .len = vma->vm_end - vma->vm_start,
4568 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
4569 },
4570 };
4571
4572 perf_event_mmap_event(&mmap_event);
4573}
4574
4575/*
4576 * IRQ throttle logging
4577 */
4578
4579static void perf_log_throttle(struct perf_event *event, int enable)
4580{
4581 struct perf_output_handle handle;
4582 struct perf_sample_data sample;
4583 int ret;
4584
4585 struct {
4586 struct perf_event_header header;
4587 u64 time;
4588 u64 id;
4589 u64 stream_id;
4590 } throttle_event = {
4591 .header = {
4592 .type = PERF_RECORD_THROTTLE,
4593 .misc = 0,
4594 .size = sizeof(throttle_event),
4595 },
4596 .time = perf_clock(),
4597 .id = primary_event_id(event),
4598 .stream_id = event->id,
4599 };
4600
4601 if (enable)
4602 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
4603
4604 perf_event_header__init_id(&throttle_event.header, &sample, event);
4605
4606 ret = perf_output_begin(&handle, event,
4607 throttle_event.header.size);
4608 if (ret)
4609 return;
4610
4611 perf_output_put(&handle, throttle_event);
4612 perf_event__output_id_sample(event, &handle, &sample);
4613 perf_output_end(&handle);
4614}
4615
4616/*
4617 * Generic event overflow handling, sampling.
4618 */
4619
4620static int __perf_event_overflow(struct perf_event *event,
4621 int throttle, struct perf_sample_data *data,
4622 struct pt_regs *regs)
4623{
4624 int events = atomic_read(&event->event_limit);
4625 struct hw_perf_event *hwc = &event->hw;
4626 int ret = 0;
4627
4628 /*
4629 * Non-sampling counters might still use the PMI to fold short
4630 * hardware counters, ignore those.
4631 */
4632 if (unlikely(!is_sampling_event(event)))
4633 return 0;
4634
4635 if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
4636 if (throttle) {
4637 hwc->interrupts = MAX_INTERRUPTS;
4638 perf_log_throttle(event, 0);
4639 ret = 1;
4640 }
4641 } else
4642 hwc->interrupts++;
4643
4644 if (event->attr.freq) {
4645 u64 now = perf_clock();
4646 s64 delta = now - hwc->freq_time_stamp;
4647
4648 hwc->freq_time_stamp = now;
4649
4650 if (delta > 0 && delta < 2*TICK_NSEC)
4651 perf_adjust_period(event, delta, hwc->last_period);
4652 }
4653
4654 /*
4655 * XXX event_limit might not quite work as expected on inherited
4656 * events
4657 */
4658
4659 event->pending_kill = POLL_IN;
4660 if (events && atomic_dec_and_test(&event->event_limit)) {
4661 ret = 1;
4662 event->pending_kill = POLL_HUP;
4663 event->pending_disable = 1;
4664 irq_work_queue(&event->pending);
4665 }
4666
4667 if (event->overflow_handler)
4668 event->overflow_handler(event, data, regs);
4669 else
4670 perf_event_output(event, data, regs);
4671
4672 if (event->fasync && event->pending_kill) {
4673 event->pending_wakeup = 1;
4674 irq_work_queue(&event->pending);
4675 }
4676
4677 return ret;
4678}
4679
4680int perf_event_overflow(struct perf_event *event,
4681 struct perf_sample_data *data,
4682 struct pt_regs *regs)
4683{
4684 return __perf_event_overflow(event, 1, data, regs);
4685}
4686
4687/*
4688 * Generic software event infrastructure
4689 */
4690
4691struct swevent_htable {
4692 struct swevent_hlist *swevent_hlist;
4693 struct mutex hlist_mutex;
4694 int hlist_refcount;
4695
4696 /* Recursion avoidance in each contexts */
4697 int recursion[PERF_NR_CONTEXTS];
4698};
4699
4700static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
4701
4702/*
4703 * We directly increment event->count and keep a second value in
4704 * event->hw.period_left to count intervals. This period event
4705 * is kept in the range [-sample_period, 0] so that we can use the
4706 * sign as trigger.
4707 */
4708
4709static u64 perf_swevent_set_period(struct perf_event *event)
4710{
4711 struct hw_perf_event *hwc = &event->hw;
4712 u64 period = hwc->last_period;
4713 u64 nr, offset;
4714 s64 old, val;
4715
4716 hwc->last_period = hwc->sample_period;
4717
4718again:
4719 old = val = local64_read(&hwc->period_left);
4720 if (val < 0)
4721 return 0;
4722
4723 nr = div64_u64(period + val, period);
4724 offset = nr * period;
4725 val -= offset;
4726 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
4727 goto again;
4728
4729 return nr;
4730}
4731
4732static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
4733 struct perf_sample_data *data,
4734 struct pt_regs *regs)
4735{
4736 struct hw_perf_event *hwc = &event->hw;
4737 int throttle = 0;
4738
4739 data->period = event->hw.last_period;
4740 if (!overflow)
4741 overflow = perf_swevent_set_period(event);
4742
4743 if (hwc->interrupts == MAX_INTERRUPTS)
4744 return;
4745
4746 for (; overflow; overflow--) {
4747 if (__perf_event_overflow(event, throttle,
4748 data, regs)) {
4749 /*
4750 * We inhibit the overflow from happening when
4751 * hwc->interrupts == MAX_INTERRUPTS.
4752 */
4753 break;
4754 }
4755 throttle = 1;
4756 }
4757}
4758
4759static void perf_swevent_event(struct perf_event *event, u64 nr,
4760 struct perf_sample_data *data,
4761 struct pt_regs *regs)
4762{
4763 struct hw_perf_event *hwc = &event->hw;
4764
4765 local64_add(nr, &event->count);
4766
4767 if (!regs)
4768 return;
4769
4770 if (!is_sampling_event(event))
4771 return;
4772
4773 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
4774 return perf_swevent_overflow(event, 1, data, regs);
4775
4776 if (local64_add_negative(nr, &hwc->period_left))
4777 return;
4778
4779 perf_swevent_overflow(event, 0, data, regs);
4780}
4781
4782static int perf_exclude_event(struct perf_event *event,
4783 struct pt_regs *regs)
4784{
4785 if (event->hw.state & PERF_HES_STOPPED)
4786 return 1;
4787
4788 if (regs) {
4789 if (event->attr.exclude_user && user_mode(regs))
4790 return 1;
4791
4792 if (event->attr.exclude_kernel && !user_mode(regs))
4793 return 1;
4794 }
4795
4796 return 0;
4797}
4798
4799static int perf_swevent_match(struct perf_event *event,
4800 enum perf_type_id type,
4801 u32 event_id,
4802 struct perf_sample_data *data,
4803 struct pt_regs *regs)
4804{
4805 if (event->attr.type != type)
4806 return 0;
4807
4808 if (event->attr.config != event_id)
4809 return 0;
4810
4811 if (perf_exclude_event(event, regs))
4812 return 0;
4813
4814 return 1;
4815}
4816
4817static inline u64 swevent_hash(u64 type, u32 event_id)
4818{
4819 u64 val = event_id | (type << 32);
4820
4821 return hash_64(val, SWEVENT_HLIST_BITS);
4822}
4823
4824static inline struct hlist_head *
4825__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
4826{
4827 u64 hash = swevent_hash(type, event_id);
4828
4829 return &hlist->heads[hash];
4830}
4831
4832/* For the read side: events when they trigger */
4833static inline struct hlist_head *
4834find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
4835{
4836 struct swevent_hlist *hlist;
4837
4838 hlist = rcu_dereference(swhash->swevent_hlist);
4839 if (!hlist)
4840 return NULL;
4841
4842 return __find_swevent_head(hlist, type, event_id);
4843}
4844
4845/* For the event head insertion and removal in the hlist */
4846static inline struct hlist_head *
4847find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
4848{
4849 struct swevent_hlist *hlist;
4850 u32 event_id = event->attr.config;
4851 u64 type = event->attr.type;
4852
4853 /*
4854 * Event scheduling is always serialized against hlist allocation
4855 * and release. Which makes the protected version suitable here.
4856 * The context lock guarantees that.
4857 */
4858 hlist = rcu_dereference_protected(swhash->swevent_hlist,
4859 lockdep_is_held(&event->ctx->lock));
4860 if (!hlist)
4861 return NULL;
4862
4863 return __find_swevent_head(hlist, type, event_id);
4864}
4865
4866static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
4867 u64 nr,
4868 struct perf_sample_data *data,
4869 struct pt_regs *regs)
4870{
4871 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4872 struct perf_event *event;
4873 struct hlist_node *node;
4874 struct hlist_head *head;
4875
4876 rcu_read_lock();
4877 head = find_swevent_head_rcu(swhash, type, event_id);
4878 if (!head)
4879 goto end;
4880
4881 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
4882 if (perf_swevent_match(event, type, event_id, data, regs))
4883 perf_swevent_event(event, nr, data, regs);
4884 }
4885end:
4886 rcu_read_unlock();
4887}
4888
4889int perf_swevent_get_recursion_context(void)
4890{
4891 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4892
4893 return get_recursion_context(swhash->recursion);
4894}
4895EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
4896
4897inline void perf_swevent_put_recursion_context(int rctx)
4898{
4899 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4900
4901 put_recursion_context(swhash->recursion, rctx);
4902}
4903
4904void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
4905{
4906 struct perf_sample_data data;
4907 int rctx;
4908
4909 preempt_disable_notrace();
4910 rctx = perf_swevent_get_recursion_context();
4911 if (rctx < 0)
4912 return;
4913
4914 perf_sample_data_init(&data, addr);
4915
4916 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
4917
4918 perf_swevent_put_recursion_context(rctx);
4919 preempt_enable_notrace();
4920}
4921
4922static void perf_swevent_read(struct perf_event *event)
4923{
4924}
4925
4926static int perf_swevent_add(struct perf_event *event, int flags)
4927{
4928 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
4929 struct hw_perf_event *hwc = &event->hw;
4930 struct hlist_head *head;
4931
4932 if (is_sampling_event(event)) {
4933 hwc->last_period = hwc->sample_period;
4934 perf_swevent_set_period(event);
4935 }
4936
4937 hwc->state = !(flags & PERF_EF_START);
4938
4939 head = find_swevent_head(swhash, event);
4940 if (WARN_ON_ONCE(!head))
4941 return -EINVAL;
4942
4943 hlist_add_head_rcu(&event->hlist_entry, head);
4944
4945 return 0;
4946}
4947
4948static void perf_swevent_del(struct perf_event *event, int flags)
4949{
4950 hlist_del_rcu(&event->hlist_entry);
4951}
4952
4953static void perf_swevent_start(struct perf_event *event, int flags)
4954{
4955 event->hw.state = 0;
4956}
4957
4958static void perf_swevent_stop(struct perf_event *event, int flags)
4959{
4960 event->hw.state = PERF_HES_STOPPED;
4961}
4962
4963/* Deref the hlist from the update side */
4964static inline struct swevent_hlist *
4965swevent_hlist_deref(struct swevent_htable *swhash)
4966{
4967 return rcu_dereference_protected(swhash->swevent_hlist,
4968 lockdep_is_held(&swhash->hlist_mutex));
4969}
4970
4971static void swevent_hlist_release(struct swevent_htable *swhash)
4972{
4973 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
4974
4975 if (!hlist)
4976 return;
4977
4978 rcu_assign_pointer(swhash->swevent_hlist, NULL);
4979 kfree_rcu(hlist, rcu_head);
4980}
4981
4982static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
4983{
4984 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
4985
4986 mutex_lock(&swhash->hlist_mutex);
4987
4988 if (!--swhash->hlist_refcount)
4989 swevent_hlist_release(swhash);
4990
4991 mutex_unlock(&swhash->hlist_mutex);
4992}
4993
4994static void swevent_hlist_put(struct perf_event *event)
4995{
4996 int cpu;
4997
4998 if (event->cpu != -1) {
4999 swevent_hlist_put_cpu(event, event->cpu);
5000 return;
5001 }
5002
5003 for_each_possible_cpu(cpu)
5004 swevent_hlist_put_cpu(event, cpu);
5005}
5006
5007static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5008{
5009 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5010 int err = 0;
5011
5012 mutex_lock(&swhash->hlist_mutex);
5013
5014 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5015 struct swevent_hlist *hlist;
5016
5017 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5018 if (!hlist) {
5019 err = -ENOMEM;
5020 goto exit;
5021 }
5022 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5023 }
5024 swhash->hlist_refcount++;
5025exit:
5026 mutex_unlock(&swhash->hlist_mutex);
5027
5028 return err;
5029}
5030
5031static int swevent_hlist_get(struct perf_event *event)
5032{
5033 int err;
5034 int cpu, failed_cpu;
5035
5036 if (event->cpu != -1)
5037 return swevent_hlist_get_cpu(event, event->cpu);
5038
5039 get_online_cpus();
5040 for_each_possible_cpu(cpu) {
5041 err = swevent_hlist_get_cpu(event, cpu);
5042 if (err) {
5043 failed_cpu = cpu;
5044 goto fail;
5045 }
5046 }
5047 put_online_cpus();
5048
5049 return 0;
5050fail:
5051 for_each_possible_cpu(cpu) {
5052 if (cpu == failed_cpu)
5053 break;
5054 swevent_hlist_put_cpu(event, cpu);
5055 }
5056
5057 put_online_cpus();
5058 return err;
5059}
5060
5061struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5062
5063static void sw_perf_event_destroy(struct perf_event *event)
5064{
5065 u64 event_id = event->attr.config;
5066
5067 WARN_ON(event->parent);
5068
5069 jump_label_dec(&perf_swevent_enabled[event_id]);
5070 swevent_hlist_put(event);
5071}
5072
5073static int perf_swevent_init(struct perf_event *event)
5074{
5075 int event_id = event->attr.config;
5076
5077 if (event->attr.type != PERF_TYPE_SOFTWARE)
5078 return -ENOENT;
5079
5080 switch (event_id) {
5081 case PERF_COUNT_SW_CPU_CLOCK:
5082 case PERF_COUNT_SW_TASK_CLOCK:
5083 return -ENOENT;
5084
5085 default:
5086 break;
5087 }
5088
5089 if (event_id >= PERF_COUNT_SW_MAX)
5090 return -ENOENT;
5091
5092 if (!event->parent) {
5093 int err;
5094
5095 err = swevent_hlist_get(event);
5096 if (err)
5097 return err;
5098
5099 jump_label_inc(&perf_swevent_enabled[event_id]);
5100 event->destroy = sw_perf_event_destroy;
5101 }
5102
5103 return 0;
5104}
5105
5106static struct pmu perf_swevent = {
5107 .task_ctx_nr = perf_sw_context,
5108
5109 .event_init = perf_swevent_init,
5110 .add = perf_swevent_add,
5111 .del = perf_swevent_del,
5112 .start = perf_swevent_start,
5113 .stop = perf_swevent_stop,
5114 .read = perf_swevent_read,
5115};
5116
5117#ifdef CONFIG_EVENT_TRACING
5118
5119static int perf_tp_filter_match(struct perf_event *event,
5120 struct perf_sample_data *data)
5121{
5122 void *record = data->raw->data;
5123
5124 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5125 return 1;
5126 return 0;
5127}
5128
5129static int perf_tp_event_match(struct perf_event *event,
5130 struct perf_sample_data *data,
5131 struct pt_regs *regs)
5132{
5133 if (event->hw.state & PERF_HES_STOPPED)
5134 return 0;
5135 /*
5136 * All tracepoints are from kernel-space.
5137 */
5138 if (event->attr.exclude_kernel)
5139 return 0;
5140
5141 if (!perf_tp_filter_match(event, data))
5142 return 0;
5143
5144 return 1;
5145}
5146
5147void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5148 struct pt_regs *regs, struct hlist_head *head, int rctx)
5149{
5150 struct perf_sample_data data;
5151 struct perf_event *event;
5152 struct hlist_node *node;
5153
5154 struct perf_raw_record raw = {
5155 .size = entry_size,
5156 .data = record,
5157 };
5158
5159 perf_sample_data_init(&data, addr);
5160 data.raw = &raw;
5161
5162 hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
5163 if (perf_tp_event_match(event, &data, regs))
5164 perf_swevent_event(event, count, &data, regs);
5165 }
5166
5167 perf_swevent_put_recursion_context(rctx);
5168}
5169EXPORT_SYMBOL_GPL(perf_tp_event);
5170
5171static void tp_perf_event_destroy(struct perf_event *event)
5172{
5173 perf_trace_destroy(event);
5174}
5175
5176static int perf_tp_event_init(struct perf_event *event)
5177{
5178 int err;
5179
5180 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5181 return -ENOENT;
5182
5183 err = perf_trace_init(event);
5184 if (err)
5185 return err;
5186
5187 event->destroy = tp_perf_event_destroy;
5188
5189 return 0;
5190}
5191
5192static struct pmu perf_tracepoint = {
5193 .task_ctx_nr = perf_sw_context,
5194
5195 .event_init = perf_tp_event_init,
5196 .add = perf_trace_add,
5197 .del = perf_trace_del,
5198 .start = perf_swevent_start,
5199 .stop = perf_swevent_stop,
5200 .read = perf_swevent_read,
5201};
5202
5203static inline void perf_tp_register(void)
5204{
5205 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5206}
5207
5208static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5209{
5210 char *filter_str;
5211 int ret;
5212
5213 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5214 return -EINVAL;
5215
5216 filter_str = strndup_user(arg, PAGE_SIZE);
5217 if (IS_ERR(filter_str))
5218 return PTR_ERR(filter_str);
5219
5220 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5221
5222 kfree(filter_str);
5223 return ret;
5224}
5225
5226static void perf_event_free_filter(struct perf_event *event)
5227{
5228 ftrace_profile_free_filter(event);
5229}
5230
5231#else
5232
5233static inline void perf_tp_register(void)
5234{
5235}
5236
5237static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5238{
5239 return -ENOENT;
5240}
5241
5242static void perf_event_free_filter(struct perf_event *event)
5243{
5244}
5245
5246#endif /* CONFIG_EVENT_TRACING */
5247
5248#ifdef CONFIG_HAVE_HW_BREAKPOINT
5249void perf_bp_event(struct perf_event *bp, void *data)
5250{
5251 struct perf_sample_data sample;
5252 struct pt_regs *regs = data;
5253
5254 perf_sample_data_init(&sample, bp->attr.bp_addr);
5255
5256 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5257 perf_swevent_event(bp, 1, &sample, regs);
5258}
5259#endif
5260
5261/*
5262 * hrtimer based swevent callback
5263 */
5264
5265static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
5266{
5267 enum hrtimer_restart ret = HRTIMER_RESTART;
5268 struct perf_sample_data data;
5269 struct pt_regs *regs;
5270 struct perf_event *event;
5271 u64 period;
5272
5273 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
5274
5275 if (event->state != PERF_EVENT_STATE_ACTIVE)
5276 return HRTIMER_NORESTART;
5277
5278 event->pmu->read(event);
5279
5280 perf_sample_data_init(&data, 0);
5281 data.period = event->hw.last_period;
5282 regs = get_irq_regs();
5283
5284 if (regs && !perf_exclude_event(event, regs)) {
5285 if (!(event->attr.exclude_idle && current->pid == 0))
5286 if (perf_event_overflow(event, &data, regs))
5287 ret = HRTIMER_NORESTART;
5288 }
5289
5290 period = max_t(u64, 10000, event->hw.sample_period);
5291 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
5292
5293 return ret;
5294}
5295
5296static void perf_swevent_start_hrtimer(struct perf_event *event)
5297{
5298 struct hw_perf_event *hwc = &event->hw;
5299 s64 period;
5300
5301 if (!is_sampling_event(event))
5302 return;
5303
5304 period = local64_read(&hwc->period_left);
5305 if (period) {
5306 if (period < 0)
5307 period = 10000;
5308
5309 local64_set(&hwc->period_left, 0);
5310 } else {
5311 period = max_t(u64, 10000, hwc->sample_period);
5312 }
5313 __hrtimer_start_range_ns(&hwc->hrtimer,
5314 ns_to_ktime(period), 0,
5315 HRTIMER_MODE_REL_PINNED, 0);
5316}
5317
5318static void perf_swevent_cancel_hrtimer(struct perf_event *event)
5319{
5320 struct hw_perf_event *hwc = &event->hw;
5321
5322 if (is_sampling_event(event)) {
5323 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
5324 local64_set(&hwc->period_left, ktime_to_ns(remaining));
5325
5326 hrtimer_cancel(&hwc->hrtimer);
5327 }
5328}
5329
5330static void perf_swevent_init_hrtimer(struct perf_event *event)
5331{
5332 struct hw_perf_event *hwc = &event->hw;
5333
5334 if (!is_sampling_event(event))
5335 return;
5336
5337 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5338 hwc->hrtimer.function = perf_swevent_hrtimer;
5339
5340 /*
5341 * Since hrtimers have a fixed rate, we can do a static freq->period
5342 * mapping and avoid the whole period adjust feedback stuff.
5343 */
5344 if (event->attr.freq) {
5345 long freq = event->attr.sample_freq;
5346
5347 event->attr.sample_period = NSEC_PER_SEC / freq;
5348 hwc->sample_period = event->attr.sample_period;
5349 local64_set(&hwc->period_left, hwc->sample_period);
5350 event->attr.freq = 0;
5351 }
5352}
5353
5354/*
5355 * Software event: cpu wall time clock
5356 */
5357
5358static void cpu_clock_event_update(struct perf_event *event)
5359{
5360 s64 prev;
5361 u64 now;
5362
5363 now = local_clock();
5364 prev = local64_xchg(&event->hw.prev_count, now);
5365 local64_add(now - prev, &event->count);
5366}
5367
5368static void cpu_clock_event_start(struct perf_event *event, int flags)
5369{
5370 local64_set(&event->hw.prev_count, local_clock());
5371 perf_swevent_start_hrtimer(event);
5372}
5373
5374static void cpu_clock_event_stop(struct perf_event *event, int flags)
5375{
5376 perf_swevent_cancel_hrtimer(event);
5377 cpu_clock_event_update(event);
5378}
5379
5380static int cpu_clock_event_add(struct perf_event *event, int flags)
5381{
5382 if (flags & PERF_EF_START)
5383 cpu_clock_event_start(event, flags);
5384
5385 return 0;
5386}
5387
5388static void cpu_clock_event_del(struct perf_event *event, int flags)
5389{
5390 cpu_clock_event_stop(event, flags);
5391}
5392
5393static void cpu_clock_event_read(struct perf_event *event)
5394{
5395 cpu_clock_event_update(event);
5396}
5397
5398static int cpu_clock_event_init(struct perf_event *event)
5399{
5400 if (event->attr.type != PERF_TYPE_SOFTWARE)
5401 return -ENOENT;
5402
5403 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
5404 return -ENOENT;
5405
5406 perf_swevent_init_hrtimer(event);
5407
5408 return 0;
5409}
5410
5411static struct pmu perf_cpu_clock = {
5412 .task_ctx_nr = perf_sw_context,
5413
5414 .event_init = cpu_clock_event_init,
5415 .add = cpu_clock_event_add,
5416 .del = cpu_clock_event_del,
5417 .start = cpu_clock_event_start,
5418 .stop = cpu_clock_event_stop,
5419 .read = cpu_clock_event_read,
5420};
5421
5422/*
5423 * Software event: task time clock
5424 */
5425
5426static void task_clock_event_update(struct perf_event *event, u64 now)
5427{
5428 u64 prev;
5429 s64 delta;
5430
5431 prev = local64_xchg(&event->hw.prev_count, now);
5432 delta = now - prev;
5433 local64_add(delta, &event->count);
5434}
5435
5436static void task_clock_event_start(struct perf_event *event, int flags)
5437{
5438 local64_set(&event->hw.prev_count, event->ctx->time);
5439 perf_swevent_start_hrtimer(event);
5440}
5441
5442static void task_clock_event_stop(struct perf_event *event, int flags)
5443{
5444 perf_swevent_cancel_hrtimer(event);
5445 task_clock_event_update(event, event->ctx->time);
5446}
5447
5448static int task_clock_event_add(struct perf_event *event, int flags)
5449{
5450 if (flags & PERF_EF_START)
5451 task_clock_event_start(event, flags);
5452
5453 return 0;
5454}
5455
5456static void task_clock_event_del(struct perf_event *event, int flags)
5457{
5458 task_clock_event_stop(event, PERF_EF_UPDATE);
5459}
5460
5461static void task_clock_event_read(struct perf_event *event)
5462{
5463 u64 now = perf_clock();
5464 u64 delta = now - event->ctx->timestamp;
5465 u64 time = event->ctx->time + delta;
5466
5467 task_clock_event_update(event, time);
5468}
5469
5470static int task_clock_event_init(struct perf_event *event)
5471{
5472 if (event->attr.type != PERF_TYPE_SOFTWARE)
5473 return -ENOENT;
5474
5475 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
5476 return -ENOENT;
5477
5478 perf_swevent_init_hrtimer(event);
5479
5480 return 0;
5481}
5482
5483static struct pmu perf_task_clock = {
5484 .task_ctx_nr = perf_sw_context,
5485
5486 .event_init = task_clock_event_init,
5487 .add = task_clock_event_add,
5488 .del = task_clock_event_del,
5489 .start = task_clock_event_start,
5490 .stop = task_clock_event_stop,
5491 .read = task_clock_event_read,
5492};
5493
5494static void perf_pmu_nop_void(struct pmu *pmu)
5495{
5496}
5497
5498static int perf_pmu_nop_int(struct pmu *pmu)
5499{
5500 return 0;
5501}
5502
5503static void perf_pmu_start_txn(struct pmu *pmu)
5504{
5505 perf_pmu_disable(pmu);
5506}
5507
5508static int perf_pmu_commit_txn(struct pmu *pmu)
5509{
5510 perf_pmu_enable(pmu);
5511 return 0;
5512}
5513
5514static void perf_pmu_cancel_txn(struct pmu *pmu)
5515{
5516 perf_pmu_enable(pmu);
5517}
5518
5519/*
5520 * Ensures all contexts with the same task_ctx_nr have the same
5521 * pmu_cpu_context too.
5522 */
5523static void *find_pmu_context(int ctxn)
5524{
5525 struct pmu *pmu;
5526
5527 if (ctxn < 0)
5528 return NULL;
5529
5530 list_for_each_entry(pmu, &pmus, entry) {
5531 if (pmu->task_ctx_nr == ctxn)
5532 return pmu->pmu_cpu_context;
5533 }
5534
5535 return NULL;
5536}
5537
5538static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
5539{
5540 int cpu;
5541
5542 for_each_possible_cpu(cpu) {
5543 struct perf_cpu_context *cpuctx;
5544
5545 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5546
5547 if (cpuctx->active_pmu == old_pmu)
5548 cpuctx->active_pmu = pmu;
5549 }
5550}
5551
5552static void free_pmu_context(struct pmu *pmu)
5553{
5554 struct pmu *i;
5555
5556 mutex_lock(&pmus_lock);
5557 /*
5558 * Like a real lame refcount.
5559 */
5560 list_for_each_entry(i, &pmus, entry) {
5561 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
5562 update_pmu_context(i, pmu);
5563 goto out;
5564 }
5565 }
5566
5567 free_percpu(pmu->pmu_cpu_context);
5568out:
5569 mutex_unlock(&pmus_lock);
5570}
5571static struct idr pmu_idr;
5572
5573static ssize_t
5574type_show(struct device *dev, struct device_attribute *attr, char *page)
5575{
5576 struct pmu *pmu = dev_get_drvdata(dev);
5577
5578 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
5579}
5580
5581static struct device_attribute pmu_dev_attrs[] = {
5582 __ATTR_RO(type),
5583 __ATTR_NULL,
5584};
5585
5586static int pmu_bus_running;
5587static struct bus_type pmu_bus = {
5588 .name = "event_source",
5589 .dev_attrs = pmu_dev_attrs,
5590};
5591
5592static void pmu_dev_release(struct device *dev)
5593{
5594 kfree(dev);
5595}
5596
5597static int pmu_dev_alloc(struct pmu *pmu)
5598{
5599 int ret = -ENOMEM;
5600
5601 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
5602 if (!pmu->dev)
5603 goto out;
5604
5605 device_initialize(pmu->dev);
5606 ret = dev_set_name(pmu->dev, "%s", pmu->name);
5607 if (ret)
5608 goto free_dev;
5609
5610 dev_set_drvdata(pmu->dev, pmu);
5611 pmu->dev->bus = &pmu_bus;
5612 pmu->dev->release = pmu_dev_release;
5613 ret = device_add(pmu->dev);
5614 if (ret)
5615 goto free_dev;
5616
5617out:
5618 return ret;
5619
5620free_dev:
5621 put_device(pmu->dev);
5622 goto out;
5623}
5624
5625static struct lock_class_key cpuctx_mutex;
5626static struct lock_class_key cpuctx_lock;
5627
5628int perf_pmu_register(struct pmu *pmu, char *name, int type)
5629{
5630 int cpu, ret;
5631
5632 mutex_lock(&pmus_lock);
5633 ret = -ENOMEM;
5634 pmu->pmu_disable_count = alloc_percpu(int);
5635 if (!pmu->pmu_disable_count)
5636 goto unlock;
5637
5638 pmu->type = -1;
5639 if (!name)
5640 goto skip_type;
5641 pmu->name = name;
5642
5643 if (type < 0) {
5644 int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
5645 if (!err)
5646 goto free_pdc;
5647
5648 err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
5649 if (err) {
5650 ret = err;
5651 goto free_pdc;
5652 }
5653 }
5654 pmu->type = type;
5655
5656 if (pmu_bus_running) {
5657 ret = pmu_dev_alloc(pmu);
5658 if (ret)
5659 goto free_idr;
5660 }
5661
5662skip_type:
5663 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
5664 if (pmu->pmu_cpu_context)
5665 goto got_cpu_context;
5666
5667 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
5668 if (!pmu->pmu_cpu_context)
5669 goto free_dev;
5670
5671 for_each_possible_cpu(cpu) {
5672 struct perf_cpu_context *cpuctx;
5673
5674 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
5675 __perf_event_init_context(&cpuctx->ctx);
5676 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
5677 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
5678 cpuctx->ctx.type = cpu_context;
5679 cpuctx->ctx.pmu = pmu;
5680 cpuctx->jiffies_interval = 1;
5681 INIT_LIST_HEAD(&cpuctx->rotation_list);
5682 cpuctx->active_pmu = pmu;
5683 }
5684
5685got_cpu_context:
5686 if (!pmu->start_txn) {
5687 if (pmu->pmu_enable) {
5688 /*
5689 * If we have pmu_enable/pmu_disable calls, install
5690 * transaction stubs that use that to try and batch
5691 * hardware accesses.
5692 */
5693 pmu->start_txn = perf_pmu_start_txn;
5694 pmu->commit_txn = perf_pmu_commit_txn;
5695 pmu->cancel_txn = perf_pmu_cancel_txn;
5696 } else {
5697 pmu->start_txn = perf_pmu_nop_void;
5698 pmu->commit_txn = perf_pmu_nop_int;
5699 pmu->cancel_txn = perf_pmu_nop_void;
5700 }
5701 }
5702
5703 if (!pmu->pmu_enable) {
5704 pmu->pmu_enable = perf_pmu_nop_void;
5705 pmu->pmu_disable = perf_pmu_nop_void;
5706 }
5707
5708 list_add_rcu(&pmu->entry, &pmus);
5709 ret = 0;
5710unlock:
5711 mutex_unlock(&pmus_lock);
5712
5713 return ret;
5714
5715free_dev:
5716 device_del(pmu->dev);
5717 put_device(pmu->dev);
5718
5719free_idr:
5720 if (pmu->type >= PERF_TYPE_MAX)
5721 idr_remove(&pmu_idr, pmu->type);
5722
5723free_pdc:
5724 free_percpu(pmu->pmu_disable_count);
5725 goto unlock;
5726}
5727
5728void perf_pmu_unregister(struct pmu *pmu)
5729{
5730 mutex_lock(&pmus_lock);
5731 list_del_rcu(&pmu->entry);
5732 mutex_unlock(&pmus_lock);
5733
5734 /*
5735 * We dereference the pmu list under both SRCU and regular RCU, so
5736 * synchronize against both of those.
5737 */
5738 synchronize_srcu(&pmus_srcu);
5739 synchronize_rcu();
5740
5741 free_percpu(pmu->pmu_disable_count);
5742 if (pmu->type >= PERF_TYPE_MAX)
5743 idr_remove(&pmu_idr, pmu->type);
5744 device_del(pmu->dev);
5745 put_device(pmu->dev);
5746 free_pmu_context(pmu);
5747}
5748
5749struct pmu *perf_init_event(struct perf_event *event)
5750{
5751 struct pmu *pmu = NULL;
5752 int idx;
5753 int ret;
5754
5755 idx = srcu_read_lock(&pmus_srcu);
5756
5757 rcu_read_lock();
5758 pmu = idr_find(&pmu_idr, event->attr.type);
5759 rcu_read_unlock();
5760 if (pmu) {
5761 ret = pmu->event_init(event);
5762 if (ret)
5763 pmu = ERR_PTR(ret);
5764 goto unlock;
5765 }
5766
5767 list_for_each_entry_rcu(pmu, &pmus, entry) {
5768 ret = pmu->event_init(event);
5769 if (!ret)
5770 goto unlock;
5771
5772 if (ret != -ENOENT) {
5773 pmu = ERR_PTR(ret);
5774 goto unlock;
5775 }
5776 }
5777 pmu = ERR_PTR(-ENOENT);
5778unlock:
5779 srcu_read_unlock(&pmus_srcu, idx);
5780
5781 return pmu;
5782}
5783
5784/*
5785 * Allocate and initialize a event structure
5786 */
5787static struct perf_event *
5788perf_event_alloc(struct perf_event_attr *attr, int cpu,
5789 struct task_struct *task,
5790 struct perf_event *group_leader,
5791 struct perf_event *parent_event,
5792 perf_overflow_handler_t overflow_handler,
5793 void *context)
5794{
5795 struct pmu *pmu;
5796 struct perf_event *event;
5797 struct hw_perf_event *hwc;
5798 long err;
5799
5800 if ((unsigned)cpu >= nr_cpu_ids) {
5801 if (!task || cpu != -1)
5802 return ERR_PTR(-EINVAL);
5803 }
5804
5805 event = kzalloc(sizeof(*event), GFP_KERNEL);
5806 if (!event)
5807 return ERR_PTR(-ENOMEM);
5808
5809 /*
5810 * Single events are their own group leaders, with an
5811 * empty sibling list:
5812 */
5813 if (!group_leader)
5814 group_leader = event;
5815
5816 mutex_init(&event->child_mutex);
5817 INIT_LIST_HEAD(&event->child_list);
5818
5819 INIT_LIST_HEAD(&event->group_entry);
5820 INIT_LIST_HEAD(&event->event_entry);
5821 INIT_LIST_HEAD(&event->sibling_list);
5822 init_waitqueue_head(&event->waitq);
5823 init_irq_work(&event->pending, perf_pending_event);
5824
5825 mutex_init(&event->mmap_mutex);
5826
5827 event->cpu = cpu;
5828 event->attr = *attr;
5829 event->group_leader = group_leader;
5830 event->pmu = NULL;
5831 event->oncpu = -1;
5832
5833 event->parent = parent_event;
5834
5835 event->ns = get_pid_ns(current->nsproxy->pid_ns);
5836 event->id = atomic64_inc_return(&perf_event_id);
5837
5838 event->state = PERF_EVENT_STATE_INACTIVE;
5839
5840 if (task) {
5841 event->attach_state = PERF_ATTACH_TASK;
5842#ifdef CONFIG_HAVE_HW_BREAKPOINT
5843 /*
5844 * hw_breakpoint is a bit difficult here..
5845 */
5846 if (attr->type == PERF_TYPE_BREAKPOINT)
5847 event->hw.bp_target = task;
5848#endif
5849 }
5850
5851 if (!overflow_handler && parent_event) {
5852 overflow_handler = parent_event->overflow_handler;
5853 context = parent_event->overflow_handler_context;
5854 }
5855
5856 event->overflow_handler = overflow_handler;
5857 event->overflow_handler_context = context;
5858
5859 if (attr->disabled)
5860 event->state = PERF_EVENT_STATE_OFF;
5861
5862 pmu = NULL;
5863
5864 hwc = &event->hw;
5865 hwc->sample_period = attr->sample_period;
5866 if (attr->freq && attr->sample_freq)
5867 hwc->sample_period = 1;
5868 hwc->last_period = hwc->sample_period;
5869
5870 local64_set(&hwc->period_left, hwc->sample_period);
5871
5872 /*
5873 * we currently do not support PERF_FORMAT_GROUP on inherited events
5874 */
5875 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
5876 goto done;
5877
5878 pmu = perf_init_event(event);
5879
5880done:
5881 err = 0;
5882 if (!pmu)
5883 err = -EINVAL;
5884 else if (IS_ERR(pmu))
5885 err = PTR_ERR(pmu);
5886
5887 if (err) {
5888 if (event->ns)
5889 put_pid_ns(event->ns);
5890 kfree(event);
5891 return ERR_PTR(err);
5892 }
5893
5894 event->pmu = pmu;
5895
5896 if (!event->parent) {
5897 if (event->attach_state & PERF_ATTACH_TASK)
5898 jump_label_inc(&perf_sched_events);
5899 if (event->attr.mmap || event->attr.mmap_data)
5900 atomic_inc(&nr_mmap_events);
5901 if (event->attr.comm)
5902 atomic_inc(&nr_comm_events);
5903 if (event->attr.task)
5904 atomic_inc(&nr_task_events);
5905 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
5906 err = get_callchain_buffers();
5907 if (err) {
5908 free_event(event);
5909 return ERR_PTR(err);
5910 }
5911 }
5912 }
5913
5914 return event;
5915}
5916
5917static int perf_copy_attr(struct perf_event_attr __user *uattr,
5918 struct perf_event_attr *attr)
5919{
5920 u32 size;
5921 int ret;
5922
5923 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
5924 return -EFAULT;
5925
5926 /*
5927 * zero the full structure, so that a short copy will be nice.
5928 */
5929 memset(attr, 0, sizeof(*attr));
5930
5931 ret = get_user(size, &uattr->size);
5932 if (ret)
5933 return ret;
5934
5935 if (size > PAGE_SIZE) /* silly large */
5936 goto err_size;
5937
5938 if (!size) /* abi compat */
5939 size = PERF_ATTR_SIZE_VER0;
5940
5941 if (size < PERF_ATTR_SIZE_VER0)
5942 goto err_size;
5943
5944 /*
5945 * If we're handed a bigger struct than we know of,
5946 * ensure all the unknown bits are 0 - i.e. new
5947 * user-space does not rely on any kernel feature
5948 * extensions we dont know about yet.
5949 */
5950 if (size > sizeof(*attr)) {
5951 unsigned char __user *addr;
5952 unsigned char __user *end;
5953 unsigned char val;
5954
5955 addr = (void __user *)uattr + sizeof(*attr);
5956 end = (void __user *)uattr + size;
5957
5958 for (; addr < end; addr++) {
5959 ret = get_user(val, addr);
5960 if (ret)
5961 return ret;
5962 if (val)
5963 goto err_size;
5964 }
5965 size = sizeof(*attr);
5966 }
5967
5968 ret = copy_from_user(attr, uattr, size);
5969 if (ret)
5970 return -EFAULT;
5971
5972 if (attr->__reserved_1)
5973 return -EINVAL;
5974
5975 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
5976 return -EINVAL;
5977
5978 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
5979 return -EINVAL;
5980
5981out:
5982 return ret;
5983
5984err_size:
5985 put_user(sizeof(*attr), &uattr->size);
5986 ret = -E2BIG;
5987 goto out;
5988}
5989
5990static int
5991perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
5992{
5993 struct ring_buffer *rb = NULL, *old_rb = NULL;
5994 int ret = -EINVAL;
5995
5996 if (!output_event)
5997 goto set;
5998
5999 /* don't allow circular references */
6000 if (event == output_event)
6001 goto out;
6002
6003 /*
6004 * Don't allow cross-cpu buffers
6005 */
6006 if (output_event->cpu != event->cpu)
6007 goto out;
6008
6009 /*
6010 * If its not a per-cpu rb, it must be the same task.
6011 */
6012 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6013 goto out;
6014
6015set:
6016 mutex_lock(&event->mmap_mutex);
6017 /* Can't redirect output if we've got an active mmap() */
6018 if (atomic_read(&event->mmap_count))
6019 goto unlock;
6020
6021 if (output_event) {
6022 /* get the rb we want to redirect to */
6023 rb = ring_buffer_get(output_event);
6024 if (!rb)
6025 goto unlock;
6026 }
6027
6028 old_rb = event->rb;
6029 rcu_assign_pointer(event->rb, rb);
6030 ret = 0;
6031unlock:
6032 mutex_unlock(&event->mmap_mutex);
6033
6034 if (old_rb)
6035 ring_buffer_put(old_rb);
6036out:
6037 return ret;
6038}
6039
6040/**
6041 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6042 *
6043 * @attr_uptr: event_id type attributes for monitoring/sampling
6044 * @pid: target pid
6045 * @cpu: target cpu
6046 * @group_fd: group leader event fd
6047 */
6048SYSCALL_DEFINE5(perf_event_open,
6049 struct perf_event_attr __user *, attr_uptr,
6050 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6051{
6052 struct perf_event *group_leader = NULL, *output_event = NULL;
6053 struct perf_event *event, *sibling;
6054 struct perf_event_attr attr;
6055 struct perf_event_context *ctx;
6056 struct file *event_file = NULL;
6057 struct file *group_file = NULL;
6058 struct task_struct *task = NULL;
6059 struct pmu *pmu;
6060 int event_fd;
6061 int move_group = 0;
6062 int fput_needed = 0;
6063 int err;
6064
6065 /* for future expandability... */
6066 if (flags & ~PERF_FLAG_ALL)
6067 return -EINVAL;
6068
6069 err = perf_copy_attr(attr_uptr, &attr);
6070 if (err)
6071 return err;
6072
6073 if (!attr.exclude_kernel) {
6074 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6075 return -EACCES;
6076 }
6077
6078 if (attr.freq) {
6079 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6080 return -EINVAL;
6081 }
6082
6083 /*
6084 * In cgroup mode, the pid argument is used to pass the fd
6085 * opened to the cgroup directory in cgroupfs. The cpu argument
6086 * designates the cpu on which to monitor threads from that
6087 * cgroup.
6088 */
6089 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
6090 return -EINVAL;
6091
6092 event_fd = get_unused_fd_flags(O_RDWR);
6093 if (event_fd < 0)
6094 return event_fd;
6095
6096 if (group_fd != -1) {
6097 group_leader = perf_fget_light(group_fd, &fput_needed);
6098 if (IS_ERR(group_leader)) {
6099 err = PTR_ERR(group_leader);
6100 goto err_fd;
6101 }
6102 group_file = group_leader->filp;
6103 if (flags & PERF_FLAG_FD_OUTPUT)
6104 output_event = group_leader;
6105 if (flags & PERF_FLAG_FD_NO_GROUP)
6106 group_leader = NULL;
6107 }
6108
6109 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
6110 task = find_lively_task_by_vpid(pid);
6111 if (IS_ERR(task)) {
6112 err = PTR_ERR(task);
6113 goto err_group_fd;
6114 }
6115 }
6116
6117 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
6118 NULL, NULL);
6119 if (IS_ERR(event)) {
6120 err = PTR_ERR(event);
6121 goto err_task;
6122 }
6123
6124 if (flags & PERF_FLAG_PID_CGROUP) {
6125 err = perf_cgroup_connect(pid, event, &attr, group_leader);
6126 if (err)
6127 goto err_alloc;
6128 /*
6129 * one more event:
6130 * - that has cgroup constraint on event->cpu
6131 * - that may need work on context switch
6132 */
6133 atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
6134 jump_label_inc(&perf_sched_events);
6135 }
6136
6137 /*
6138 * Special case software events and allow them to be part of
6139 * any hardware group.
6140 */
6141 pmu = event->pmu;
6142
6143 if (group_leader &&
6144 (is_software_event(event) != is_software_event(group_leader))) {
6145 if (is_software_event(event)) {
6146 /*
6147 * If event and group_leader are not both a software
6148 * event, and event is, then group leader is not.
6149 *
6150 * Allow the addition of software events to !software
6151 * groups, this is safe because software events never
6152 * fail to schedule.
6153 */
6154 pmu = group_leader->pmu;
6155 } else if (is_software_event(group_leader) &&
6156 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
6157 /*
6158 * In case the group is a pure software group, and we
6159 * try to add a hardware event, move the whole group to
6160 * the hardware context.
6161 */
6162 move_group = 1;
6163 }
6164 }
6165
6166 /*
6167 * Get the target context (task or percpu):
6168 */
6169 ctx = find_get_context(pmu, task, cpu);
6170 if (IS_ERR(ctx)) {
6171 err = PTR_ERR(ctx);
6172 goto err_alloc;
6173 }
6174
6175 if (task) {
6176 put_task_struct(task);
6177 task = NULL;
6178 }
6179
6180 /*
6181 * Look up the group leader (we will attach this event to it):
6182 */
6183 if (group_leader) {
6184 err = -EINVAL;
6185
6186 /*
6187 * Do not allow a recursive hierarchy (this new sibling
6188 * becoming part of another group-sibling):
6189 */
6190 if (group_leader->group_leader != group_leader)
6191 goto err_context;
6192 /*
6193 * Do not allow to attach to a group in a different
6194 * task or CPU context:
6195 */
6196 if (move_group) {
6197 if (group_leader->ctx->type != ctx->type)
6198 goto err_context;
6199 } else {
6200 if (group_leader->ctx != ctx)
6201 goto err_context;
6202 }
6203
6204 /*
6205 * Only a group leader can be exclusive or pinned
6206 */
6207 if (attr.exclusive || attr.pinned)
6208 goto err_context;
6209 }
6210
6211 if (output_event) {
6212 err = perf_event_set_output(event, output_event);
6213 if (err)
6214 goto err_context;
6215 }
6216
6217 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
6218 if (IS_ERR(event_file)) {
6219 err = PTR_ERR(event_file);
6220 goto err_context;
6221 }
6222
6223 if (move_group) {
6224 struct perf_event_context *gctx = group_leader->ctx;
6225
6226 mutex_lock(&gctx->mutex);
6227 perf_remove_from_context(group_leader);
6228 list_for_each_entry(sibling, &group_leader->sibling_list,
6229 group_entry) {
6230 perf_remove_from_context(sibling);
6231 put_ctx(gctx);
6232 }
6233 mutex_unlock(&gctx->mutex);
6234 put_ctx(gctx);
6235 }
6236
6237 event->filp = event_file;
6238 WARN_ON_ONCE(ctx->parent_ctx);
6239 mutex_lock(&ctx->mutex);
6240
6241 if (move_group) {
6242 perf_install_in_context(ctx, group_leader, cpu);
6243 get_ctx(ctx);
6244 list_for_each_entry(sibling, &group_leader->sibling_list,
6245 group_entry) {
6246 perf_install_in_context(ctx, sibling, cpu);
6247 get_ctx(ctx);
6248 }
6249 }
6250
6251 perf_install_in_context(ctx, event, cpu);
6252 ++ctx->generation;
6253 perf_unpin_context(ctx);
6254 mutex_unlock(&ctx->mutex);
6255
6256 event->owner = current;
6257
6258 mutex_lock(¤t->perf_event_mutex);
6259 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
6260 mutex_unlock(¤t->perf_event_mutex);
6261
6262 /*
6263 * Precalculate sample_data sizes
6264 */
6265 perf_event__header_size(event);
6266 perf_event__id_header_size(event);
6267
6268 /*
6269 * Drop the reference on the group_event after placing the
6270 * new event on the sibling_list. This ensures destruction
6271 * of the group leader will find the pointer to itself in
6272 * perf_group_detach().
6273 */
6274 fput_light(group_file, fput_needed);
6275 fd_install(event_fd, event_file);
6276 return event_fd;
6277
6278err_context:
6279 perf_unpin_context(ctx);
6280 put_ctx(ctx);
6281err_alloc:
6282 free_event(event);
6283err_task:
6284 if (task)
6285 put_task_struct(task);
6286err_group_fd:
6287 fput_light(group_file, fput_needed);
6288err_fd:
6289 put_unused_fd(event_fd);
6290 return err;
6291}
6292
6293/**
6294 * perf_event_create_kernel_counter
6295 *
6296 * @attr: attributes of the counter to create
6297 * @cpu: cpu in which the counter is bound
6298 * @task: task to profile (NULL for percpu)
6299 */
6300struct perf_event *
6301perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
6302 struct task_struct *task,
6303 perf_overflow_handler_t overflow_handler,
6304 void *context)
6305{
6306 struct perf_event_context *ctx;
6307 struct perf_event *event;
6308 int err;
6309
6310 /*
6311 * Get the target context (task or percpu):
6312 */
6313
6314 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
6315 overflow_handler, context);
6316 if (IS_ERR(event)) {
6317 err = PTR_ERR(event);
6318 goto err;
6319 }
6320
6321 ctx = find_get_context(event->pmu, task, cpu);
6322 if (IS_ERR(ctx)) {
6323 err = PTR_ERR(ctx);
6324 goto err_free;
6325 }
6326
6327 event->filp = NULL;
6328 WARN_ON_ONCE(ctx->parent_ctx);
6329 mutex_lock(&ctx->mutex);
6330 perf_install_in_context(ctx, event, cpu);
6331 ++ctx->generation;
6332 perf_unpin_context(ctx);
6333 mutex_unlock(&ctx->mutex);
6334
6335 return event;
6336
6337err_free:
6338 free_event(event);
6339err:
6340 return ERR_PTR(err);
6341}
6342EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
6343
6344static void sync_child_event(struct perf_event *child_event,
6345 struct task_struct *child)
6346{
6347 struct perf_event *parent_event = child_event->parent;
6348 u64 child_val;
6349
6350 if (child_event->attr.inherit_stat)
6351 perf_event_read_event(child_event, child);
6352
6353 child_val = perf_event_count(child_event);
6354
6355 /*
6356 * Add back the child's count to the parent's count:
6357 */
6358 atomic64_add(child_val, &parent_event->child_count);
6359 atomic64_add(child_event->total_time_enabled,
6360 &parent_event->child_total_time_enabled);
6361 atomic64_add(child_event->total_time_running,
6362 &parent_event->child_total_time_running);
6363
6364 /*
6365 * Remove this event from the parent's list
6366 */
6367 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6368 mutex_lock(&parent_event->child_mutex);
6369 list_del_init(&child_event->child_list);
6370 mutex_unlock(&parent_event->child_mutex);
6371
6372 /*
6373 * Release the parent event, if this was the last
6374 * reference to it.
6375 */
6376 fput(parent_event->filp);
6377}
6378
6379static void
6380__perf_event_exit_task(struct perf_event *child_event,
6381 struct perf_event_context *child_ctx,
6382 struct task_struct *child)
6383{
6384 if (child_event->parent) {
6385 raw_spin_lock_irq(&child_ctx->lock);
6386 perf_group_detach(child_event);
6387 raw_spin_unlock_irq(&child_ctx->lock);
6388 }
6389
6390 perf_remove_from_context(child_event);
6391
6392 /*
6393 * It can happen that the parent exits first, and has events
6394 * that are still around due to the child reference. These
6395 * events need to be zapped.
6396 */
6397 if (child_event->parent) {
6398 sync_child_event(child_event, child);
6399 free_event(child_event);
6400 }
6401}
6402
6403static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
6404{
6405 struct perf_event *child_event, *tmp;
6406 struct perf_event_context *child_ctx;
6407 unsigned long flags;
6408
6409 if (likely(!child->perf_event_ctxp[ctxn])) {
6410 perf_event_task(child, NULL, 0);
6411 return;
6412 }
6413
6414 local_irq_save(flags);
6415 /*
6416 * We can't reschedule here because interrupts are disabled,
6417 * and either child is current or it is a task that can't be
6418 * scheduled, so we are now safe from rescheduling changing
6419 * our context.
6420 */
6421 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
6422
6423 /*
6424 * Take the context lock here so that if find_get_context is
6425 * reading child->perf_event_ctxp, we wait until it has
6426 * incremented the context's refcount before we do put_ctx below.
6427 */
6428 raw_spin_lock(&child_ctx->lock);
6429 task_ctx_sched_out(child_ctx);
6430 child->perf_event_ctxp[ctxn] = NULL;
6431 /*
6432 * If this context is a clone; unclone it so it can't get
6433 * swapped to another process while we're removing all
6434 * the events from it.
6435 */
6436 unclone_ctx(child_ctx);
6437 update_context_time(child_ctx);
6438 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6439
6440 /*
6441 * Report the task dead after unscheduling the events so that we
6442 * won't get any samples after PERF_RECORD_EXIT. We can however still
6443 * get a few PERF_RECORD_READ events.
6444 */
6445 perf_event_task(child, child_ctx, 0);
6446
6447 /*
6448 * We can recurse on the same lock type through:
6449 *
6450 * __perf_event_exit_task()
6451 * sync_child_event()
6452 * fput(parent_event->filp)
6453 * perf_release()
6454 * mutex_lock(&ctx->mutex)
6455 *
6456 * But since its the parent context it won't be the same instance.
6457 */
6458 mutex_lock(&child_ctx->mutex);
6459
6460again:
6461 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
6462 group_entry)
6463 __perf_event_exit_task(child_event, child_ctx, child);
6464
6465 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
6466 group_entry)
6467 __perf_event_exit_task(child_event, child_ctx, child);
6468
6469 /*
6470 * If the last event was a group event, it will have appended all
6471 * its siblings to the list, but we obtained 'tmp' before that which
6472 * will still point to the list head terminating the iteration.
6473 */
6474 if (!list_empty(&child_ctx->pinned_groups) ||
6475 !list_empty(&child_ctx->flexible_groups))
6476 goto again;
6477
6478 mutex_unlock(&child_ctx->mutex);
6479
6480 put_ctx(child_ctx);
6481}
6482
6483/*
6484 * When a child task exits, feed back event values to parent events.
6485 */
6486void perf_event_exit_task(struct task_struct *child)
6487{
6488 struct perf_event *event, *tmp;
6489 int ctxn;
6490
6491 mutex_lock(&child->perf_event_mutex);
6492 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
6493 owner_entry) {
6494 list_del_init(&event->owner_entry);
6495
6496 /*
6497 * Ensure the list deletion is visible before we clear
6498 * the owner, closes a race against perf_release() where
6499 * we need to serialize on the owner->perf_event_mutex.
6500 */
6501 smp_wmb();
6502 event->owner = NULL;
6503 }
6504 mutex_unlock(&child->perf_event_mutex);
6505
6506 for_each_task_context_nr(ctxn)
6507 perf_event_exit_task_context(child, ctxn);
6508}
6509
6510static void perf_free_event(struct perf_event *event,
6511 struct perf_event_context *ctx)
6512{
6513 struct perf_event *parent = event->parent;
6514
6515 if (WARN_ON_ONCE(!parent))
6516 return;
6517
6518 mutex_lock(&parent->child_mutex);
6519 list_del_init(&event->child_list);
6520 mutex_unlock(&parent->child_mutex);
6521
6522 fput(parent->filp);
6523
6524 perf_group_detach(event);
6525 list_del_event(event, ctx);
6526 free_event(event);
6527}
6528
6529/*
6530 * free an unexposed, unused context as created by inheritance by
6531 * perf_event_init_task below, used by fork() in case of fail.
6532 */
6533void perf_event_free_task(struct task_struct *task)
6534{
6535 struct perf_event_context *ctx;
6536 struct perf_event *event, *tmp;
6537 int ctxn;
6538
6539 for_each_task_context_nr(ctxn) {
6540 ctx = task->perf_event_ctxp[ctxn];
6541 if (!ctx)
6542 continue;
6543
6544 mutex_lock(&ctx->mutex);
6545again:
6546 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
6547 group_entry)
6548 perf_free_event(event, ctx);
6549
6550 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
6551 group_entry)
6552 perf_free_event(event, ctx);
6553
6554 if (!list_empty(&ctx->pinned_groups) ||
6555 !list_empty(&ctx->flexible_groups))
6556 goto again;
6557
6558 mutex_unlock(&ctx->mutex);
6559
6560 put_ctx(ctx);
6561 }
6562}
6563
6564void perf_event_delayed_put(struct task_struct *task)
6565{
6566 int ctxn;
6567
6568 for_each_task_context_nr(ctxn)
6569 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
6570}
6571
6572/*
6573 * inherit a event from parent task to child task:
6574 */
6575static struct perf_event *
6576inherit_event(struct perf_event *parent_event,
6577 struct task_struct *parent,
6578 struct perf_event_context *parent_ctx,
6579 struct task_struct *child,
6580 struct perf_event *group_leader,
6581 struct perf_event_context *child_ctx)
6582{
6583 struct perf_event *child_event;
6584 unsigned long flags;
6585
6586 /*
6587 * Instead of creating recursive hierarchies of events,
6588 * we link inherited events back to the original parent,
6589 * which has a filp for sure, which we use as the reference
6590 * count:
6591 */
6592 if (parent_event->parent)
6593 parent_event = parent_event->parent;
6594
6595 child_event = perf_event_alloc(&parent_event->attr,
6596 parent_event->cpu,
6597 child,
6598 group_leader, parent_event,
6599 NULL, NULL);
6600 if (IS_ERR(child_event))
6601 return child_event;
6602 get_ctx(child_ctx);
6603
6604 /*
6605 * Make the child state follow the state of the parent event,
6606 * not its attr.disabled bit. We hold the parent's mutex,
6607 * so we won't race with perf_event_{en, dis}able_family.
6608 */
6609 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
6610 child_event->state = PERF_EVENT_STATE_INACTIVE;
6611 else
6612 child_event->state = PERF_EVENT_STATE_OFF;
6613
6614 if (parent_event->attr.freq) {
6615 u64 sample_period = parent_event->hw.sample_period;
6616 struct hw_perf_event *hwc = &child_event->hw;
6617
6618 hwc->sample_period = sample_period;
6619 hwc->last_period = sample_period;
6620
6621 local64_set(&hwc->period_left, sample_period);
6622 }
6623
6624 child_event->ctx = child_ctx;
6625 child_event->overflow_handler = parent_event->overflow_handler;
6626 child_event->overflow_handler_context
6627 = parent_event->overflow_handler_context;
6628
6629 /*
6630 * Precalculate sample_data sizes
6631 */
6632 perf_event__header_size(child_event);
6633 perf_event__id_header_size(child_event);
6634
6635 /*
6636 * Link it up in the child's context:
6637 */
6638 raw_spin_lock_irqsave(&child_ctx->lock, flags);
6639 add_event_to_ctx(child_event, child_ctx);
6640 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
6641
6642 /*
6643 * Get a reference to the parent filp - we will fput it
6644 * when the child event exits. This is safe to do because
6645 * we are in the parent and we know that the filp still
6646 * exists and has a nonzero count:
6647 */
6648 atomic_long_inc(&parent_event->filp->f_count);
6649
6650 /*
6651 * Link this into the parent event's child list
6652 */
6653 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
6654 mutex_lock(&parent_event->child_mutex);
6655 list_add_tail(&child_event->child_list, &parent_event->child_list);
6656 mutex_unlock(&parent_event->child_mutex);
6657
6658 return child_event;
6659}
6660
6661static int inherit_group(struct perf_event *parent_event,
6662 struct task_struct *parent,
6663 struct perf_event_context *parent_ctx,
6664 struct task_struct *child,
6665 struct perf_event_context *child_ctx)
6666{
6667 struct perf_event *leader;
6668 struct perf_event *sub;
6669 struct perf_event *child_ctr;
6670
6671 leader = inherit_event(parent_event, parent, parent_ctx,
6672 child, NULL, child_ctx);
6673 if (IS_ERR(leader))
6674 return PTR_ERR(leader);
6675 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
6676 child_ctr = inherit_event(sub, parent, parent_ctx,
6677 child, leader, child_ctx);
6678 if (IS_ERR(child_ctr))
6679 return PTR_ERR(child_ctr);
6680 }
6681 return 0;
6682}
6683
6684static int
6685inherit_task_group(struct perf_event *event, struct task_struct *parent,
6686 struct perf_event_context *parent_ctx,
6687 struct task_struct *child, int ctxn,
6688 int *inherited_all)
6689{
6690 int ret;
6691 struct perf_event_context *child_ctx;
6692
6693 if (!event->attr.inherit) {
6694 *inherited_all = 0;
6695 return 0;
6696 }
6697
6698 child_ctx = child->perf_event_ctxp[ctxn];
6699 if (!child_ctx) {
6700 /*
6701 * This is executed from the parent task context, so
6702 * inherit events that have been marked for cloning.
6703 * First allocate and initialize a context for the
6704 * child.
6705 */
6706
6707 child_ctx = alloc_perf_context(event->pmu, child);
6708 if (!child_ctx)
6709 return -ENOMEM;
6710
6711 child->perf_event_ctxp[ctxn] = child_ctx;
6712 }
6713
6714 ret = inherit_group(event, parent, parent_ctx,
6715 child, child_ctx);
6716
6717 if (ret)
6718 *inherited_all = 0;
6719
6720 return ret;
6721}
6722
6723/*
6724 * Initialize the perf_event context in task_struct
6725 */
6726int perf_event_init_context(struct task_struct *child, int ctxn)
6727{
6728 struct perf_event_context *child_ctx, *parent_ctx;
6729 struct perf_event_context *cloned_ctx;
6730 struct perf_event *event;
6731 struct task_struct *parent = current;
6732 int inherited_all = 1;
6733 unsigned long flags;
6734 int ret = 0;
6735
6736 if (likely(!parent->perf_event_ctxp[ctxn]))
6737 return 0;
6738
6739 /*
6740 * If the parent's context is a clone, pin it so it won't get
6741 * swapped under us.
6742 */
6743 parent_ctx = perf_pin_task_context(parent, ctxn);
6744
6745 /*
6746 * No need to check if parent_ctx != NULL here; since we saw
6747 * it non-NULL earlier, the only reason for it to become NULL
6748 * is if we exit, and since we're currently in the middle of
6749 * a fork we can't be exiting at the same time.
6750 */
6751
6752 /*
6753 * Lock the parent list. No need to lock the child - not PID
6754 * hashed yet and not running, so nobody can access it.
6755 */
6756 mutex_lock(&parent_ctx->mutex);
6757
6758 /*
6759 * We dont have to disable NMIs - we are only looking at
6760 * the list, not manipulating it:
6761 */
6762 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
6763 ret = inherit_task_group(event, parent, parent_ctx,
6764 child, ctxn, &inherited_all);
6765 if (ret)
6766 break;
6767 }
6768
6769 /*
6770 * We can't hold ctx->lock when iterating the ->flexible_group list due
6771 * to allocations, but we need to prevent rotation because
6772 * rotate_ctx() will change the list from interrupt context.
6773 */
6774 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6775 parent_ctx->rotate_disable = 1;
6776 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6777
6778 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
6779 ret = inherit_task_group(event, parent, parent_ctx,
6780 child, ctxn, &inherited_all);
6781 if (ret)
6782 break;
6783 }
6784
6785 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
6786 parent_ctx->rotate_disable = 0;
6787
6788 child_ctx = child->perf_event_ctxp[ctxn];
6789
6790 if (child_ctx && inherited_all) {
6791 /*
6792 * Mark the child context as a clone of the parent
6793 * context, or of whatever the parent is a clone of.
6794 *
6795 * Note that if the parent is a clone, the holding of
6796 * parent_ctx->lock avoids it from being uncloned.
6797 */
6798 cloned_ctx = parent_ctx->parent_ctx;
6799 if (cloned_ctx) {
6800 child_ctx->parent_ctx = cloned_ctx;
6801 child_ctx->parent_gen = parent_ctx->parent_gen;
6802 } else {
6803 child_ctx->parent_ctx = parent_ctx;
6804 child_ctx->parent_gen = parent_ctx->generation;
6805 }
6806 get_ctx(child_ctx->parent_ctx);
6807 }
6808
6809 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
6810 mutex_unlock(&parent_ctx->mutex);
6811
6812 perf_unpin_context(parent_ctx);
6813 put_ctx(parent_ctx);
6814
6815 return ret;
6816}
6817
6818/*
6819 * Initialize the perf_event context in task_struct
6820 */
6821int perf_event_init_task(struct task_struct *child)
6822{
6823 int ctxn, ret;
6824
6825 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
6826 mutex_init(&child->perf_event_mutex);
6827 INIT_LIST_HEAD(&child->perf_event_list);
6828
6829 for_each_task_context_nr(ctxn) {
6830 ret = perf_event_init_context(child, ctxn);
6831 if (ret)
6832 return ret;
6833 }
6834
6835 return 0;
6836}
6837
6838static void __init perf_event_init_all_cpus(void)
6839{
6840 struct swevent_htable *swhash;
6841 int cpu;
6842
6843 for_each_possible_cpu(cpu) {
6844 swhash = &per_cpu(swevent_htable, cpu);
6845 mutex_init(&swhash->hlist_mutex);
6846 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
6847 }
6848}
6849
6850static void __cpuinit perf_event_init_cpu(int cpu)
6851{
6852 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6853
6854 mutex_lock(&swhash->hlist_mutex);
6855 if (swhash->hlist_refcount > 0) {
6856 struct swevent_hlist *hlist;
6857
6858 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
6859 WARN_ON(!hlist);
6860 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6861 }
6862 mutex_unlock(&swhash->hlist_mutex);
6863}
6864
6865#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
6866static void perf_pmu_rotate_stop(struct pmu *pmu)
6867{
6868 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6869
6870 WARN_ON(!irqs_disabled());
6871
6872 list_del_init(&cpuctx->rotation_list);
6873}
6874
6875static void __perf_event_exit_context(void *__info)
6876{
6877 struct perf_event_context *ctx = __info;
6878 struct perf_event *event, *tmp;
6879
6880 perf_pmu_rotate_stop(ctx->pmu);
6881
6882 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
6883 __perf_remove_from_context(event);
6884 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
6885 __perf_remove_from_context(event);
6886}
6887
6888static void perf_event_exit_cpu_context(int cpu)
6889{
6890 struct perf_event_context *ctx;
6891 struct pmu *pmu;
6892 int idx;
6893
6894 idx = srcu_read_lock(&pmus_srcu);
6895 list_for_each_entry_rcu(pmu, &pmus, entry) {
6896 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
6897
6898 mutex_lock(&ctx->mutex);
6899 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
6900 mutex_unlock(&ctx->mutex);
6901 }
6902 srcu_read_unlock(&pmus_srcu, idx);
6903}
6904
6905static void perf_event_exit_cpu(int cpu)
6906{
6907 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6908
6909 mutex_lock(&swhash->hlist_mutex);
6910 swevent_hlist_release(swhash);
6911 mutex_unlock(&swhash->hlist_mutex);
6912
6913 perf_event_exit_cpu_context(cpu);
6914}
6915#else
6916static inline void perf_event_exit_cpu(int cpu) { }
6917#endif
6918
6919static int
6920perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
6921{
6922 int cpu;
6923
6924 for_each_online_cpu(cpu)
6925 perf_event_exit_cpu(cpu);
6926
6927 return NOTIFY_OK;
6928}
6929
6930/*
6931 * Run the perf reboot notifier at the very last possible moment so that
6932 * the generic watchdog code runs as long as possible.
6933 */
6934static struct notifier_block perf_reboot_notifier = {
6935 .notifier_call = perf_reboot,
6936 .priority = INT_MIN,
6937};
6938
6939static int __cpuinit
6940perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
6941{
6942 unsigned int cpu = (long)hcpu;
6943
6944 switch (action & ~CPU_TASKS_FROZEN) {
6945
6946 case CPU_UP_PREPARE:
6947 case CPU_DOWN_FAILED:
6948 perf_event_init_cpu(cpu);
6949 break;
6950
6951 case CPU_UP_CANCELED:
6952 case CPU_DOWN_PREPARE:
6953 perf_event_exit_cpu(cpu);
6954 break;
6955
6956 default:
6957 break;
6958 }
6959
6960 return NOTIFY_OK;
6961}
6962
6963void __init perf_event_init(void)
6964{
6965 int ret;
6966
6967 idr_init(&pmu_idr);
6968
6969 perf_event_init_all_cpus();
6970 init_srcu_struct(&pmus_srcu);
6971 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
6972 perf_pmu_register(&perf_cpu_clock, NULL, -1);
6973 perf_pmu_register(&perf_task_clock, NULL, -1);
6974 perf_tp_register();
6975 perf_cpu_notifier(perf_cpu_notify);
6976 register_reboot_notifier(&perf_reboot_notifier);
6977
6978 ret = init_hw_breakpoint();
6979 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
6980}
6981
6982static int __init perf_event_sysfs_init(void)
6983{
6984 struct pmu *pmu;
6985 int ret;
6986
6987 mutex_lock(&pmus_lock);
6988
6989 ret = bus_register(&pmu_bus);
6990 if (ret)
6991 goto unlock;
6992
6993 list_for_each_entry(pmu, &pmus, entry) {
6994 if (!pmu->name || pmu->type < 0)
6995 continue;
6996
6997 ret = pmu_dev_alloc(pmu);
6998 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
6999 }
7000 pmu_bus_running = 1;
7001 ret = 0;
7002
7003unlock:
7004 mutex_unlock(&pmus_lock);
7005
7006 return ret;
7007}
7008device_initcall(perf_event_sysfs_init);
7009
7010#ifdef CONFIG_CGROUP_PERF
7011static struct cgroup_subsys_state *perf_cgroup_create(
7012 struct cgroup_subsys *ss, struct cgroup *cont)
7013{
7014 struct perf_cgroup *jc;
7015
7016 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7017 if (!jc)
7018 return ERR_PTR(-ENOMEM);
7019
7020 jc->info = alloc_percpu(struct perf_cgroup_info);
7021 if (!jc->info) {
7022 kfree(jc);
7023 return ERR_PTR(-ENOMEM);
7024 }
7025
7026 return &jc->css;
7027}
7028
7029static void perf_cgroup_destroy(struct cgroup_subsys *ss,
7030 struct cgroup *cont)
7031{
7032 struct perf_cgroup *jc;
7033 jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
7034 struct perf_cgroup, css);
7035 free_percpu(jc->info);
7036 kfree(jc);
7037}
7038
7039static int __perf_cgroup_move(void *info)
7040{
7041 struct task_struct *task = info;
7042 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
7043 return 0;
7044}
7045
7046static void
7047perf_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *task)
7048{
7049 task_function_call(task, __perf_cgroup_move, task);
7050}
7051
7052static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7053 struct cgroup *old_cgrp, struct task_struct *task)
7054{
7055 /*
7056 * cgroup_exit() is called in the copy_process() failure path.
7057 * Ignore this case since the task hasn't ran yet, this avoids
7058 * trying to poke a half freed task state from generic code.
7059 */
7060 if (!(task->flags & PF_EXITING))
7061 return;
7062
7063 perf_cgroup_attach_task(cgrp, task);
7064}
7065
7066struct cgroup_subsys perf_subsys = {
7067 .name = "perf_event",
7068 .subsys_id = perf_subsys_id,
7069 .create = perf_cgroup_create,
7070 .destroy = perf_cgroup_destroy,
7071 .exit = perf_cgroup_exit,
7072 .attach_task = perf_cgroup_attach_task,
7073};
7074#endif /* CONFIG_CGROUP_PERF */