2 * Performance events core code:
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>
9 * For licensing details see kernel-base/COPYING
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>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
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
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
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
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)
126 * branch priv levels that need permission checks
128 #define PERF_SAMPLE_BRANCH_PERM_PLM \
129 (PERF_SAMPLE_BRANCH_KERNEL |\
130 PERF_SAMPLE_BRANCH_HV)
133 EVENT_FLEXIBLE = 0x1,
135 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
139 * perf_sched_events : >0 events exist
140 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
142 struct static_key_deferred perf_sched_events __read_mostly;
143 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
144 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
146 static atomic_t nr_mmap_events __read_mostly;
147 static atomic_t nr_comm_events __read_mostly;
148 static atomic_t nr_task_events __read_mostly;
149 static atomic_t nr_freq_events __read_mostly;
151 static LIST_HEAD(pmus);
152 static DEFINE_MUTEX(pmus_lock);
153 static struct srcu_struct pmus_srcu;
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
162 int sysctl_perf_event_paranoid __read_mostly = 1;
164 /* Minimum for 512 kiB + 1 user control page */
165 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
168 * max perf event sample rate
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
174 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
176 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
177 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
179 static int perf_sample_allowed_ns __read_mostly =
180 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
182 void update_perf_cpu_limits(void)
184 u64 tmp = perf_sample_period_ns;
186 tmp *= sysctl_perf_cpu_time_max_percent;
188 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
191 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
193 int perf_proc_update_handler(struct ctl_table *table, int write,
194 void __user *buffer, size_t *lenp,
197 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
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();
209 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
211 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
212 void __user *buffer, size_t *lenp,
215 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
220 update_perf_cpu_limits();
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.
231 #define NR_ACCUMULATED_SAMPLES 128
232 static DEFINE_PER_CPU(u64, running_sample_length);
234 static void perf_duration_warn(struct irq_work *w)
236 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
237 u64 avg_local_sample_len;
238 u64 local_samples_len;
240 local_samples_len = __get_cpu_var(running_sample_length);
241 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
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);
250 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
252 void perf_sample_event_took(u64 sample_len_ns)
254 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
255 u64 avg_local_sample_len;
256 u64 local_samples_len;
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;
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.
272 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
274 if (avg_local_sample_len <= allowed_ns)
277 if (max_samples_per_tick <= 1)
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;
284 update_perf_cpu_limits();
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);
294 static atomic64_t perf_event_id;
296 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
297 enum event_type_t event_type);
299 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
300 enum event_type_t event_type,
301 struct task_struct *task);
303 static void update_context_time(struct perf_event_context *ctx);
304 static u64 perf_event_time(struct perf_event *event);
306 void __weak perf_event_print_debug(void) { }
308 extern __weak const char *perf_pmu_name(void)
313 static inline u64 perf_clock(void)
315 return local_clock();
318 static inline struct perf_cpu_context *
319 __get_cpu_context(struct perf_event_context *ctx)
321 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
324 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
325 struct perf_event_context *ctx)
327 raw_spin_lock(&cpuctx->ctx.lock);
329 raw_spin_lock(&ctx->lock);
332 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
333 struct perf_event_context *ctx)
336 raw_spin_unlock(&ctx->lock);
337 raw_spin_unlock(&cpuctx->ctx.lock);
340 #ifdef CONFIG_CGROUP_PERF
343 * perf_cgroup_info keeps track of time_enabled for a cgroup.
344 * This is a per-cpu dynamically allocated data structure.
346 struct perf_cgroup_info {
352 struct cgroup_subsys_state css;
353 struct perf_cgroup_info __percpu *info;
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.
361 static inline struct perf_cgroup *
362 perf_cgroup_from_task(struct task_struct *task)
364 return container_of(task_css(task, perf_event_cgrp_id),
365 struct perf_cgroup, css);
369 perf_cgroup_match(struct perf_event *event)
371 struct perf_event_context *ctx = event->ctx;
372 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
374 /* @event doesn't care about cgroup */
378 /* wants specific cgroup scope but @cpuctx isn't associated with any */
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.
388 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
389 event->cgrp->css.cgroup);
392 static inline void perf_put_cgroup(struct perf_event *event)
394 css_put(&event->cgrp->css);
397 static inline void perf_detach_cgroup(struct perf_event *event)
399 perf_put_cgroup(event);
403 static inline int is_cgroup_event(struct perf_event *event)
405 return event->cgrp != NULL;
408 static inline u64 perf_cgroup_event_time(struct perf_event *event)
410 struct perf_cgroup_info *t;
412 t = per_cpu_ptr(event->cgrp->info, event->cpu);
416 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
418 struct perf_cgroup_info *info;
423 info = this_cpu_ptr(cgrp->info);
425 info->time += now - info->timestamp;
426 info->timestamp = now;
429 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
431 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
433 __update_cgrp_time(cgrp_out);
436 static inline void update_cgrp_time_from_event(struct perf_event *event)
438 struct perf_cgroup *cgrp;
441 * ensure we access cgroup data only when needed and
442 * when we know the cgroup is pinned (css_get)
444 if (!is_cgroup_event(event))
447 cgrp = perf_cgroup_from_task(current);
449 * Do not update time when cgroup is not active
451 if (cgrp == event->cgrp)
452 __update_cgrp_time(event->cgrp);
456 perf_cgroup_set_timestamp(struct task_struct *task,
457 struct perf_event_context *ctx)
459 struct perf_cgroup *cgrp;
460 struct perf_cgroup_info *info;
463 * ctx->lock held by caller
464 * ensure we do not access cgroup data
465 * unless we have the cgroup pinned (css_get)
467 if (!task || !ctx->nr_cgroups)
470 cgrp = perf_cgroup_from_task(task);
471 info = this_cpu_ptr(cgrp->info);
472 info->timestamp = ctx->timestamp;
475 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
476 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
479 * reschedule events based on the cgroup constraint of task.
481 * mode SWOUT : schedule out everything
482 * mode SWIN : schedule in based on cgroup for next
484 void perf_cgroup_switch(struct task_struct *task, int mode)
486 struct perf_cpu_context *cpuctx;
491 * disable interrupts to avoid geting nr_cgroup
492 * changes via __perf_event_disable(). Also
495 local_irq_save(flags);
498 * we reschedule only in the presence of cgroup
499 * constrained events.
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 */
509 * perf_cgroup_events says at least one
510 * context on this CPU has cgroup events.
512 * ctx->nr_cgroups reports the number of cgroup
513 * events for a context.
515 if (cpuctx->ctx.nr_cgroups > 0) {
516 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
517 perf_pmu_disable(cpuctx->ctx.pmu);
519 if (mode & PERF_CGROUP_SWOUT) {
520 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
522 * must not be done before ctxswout due
523 * to event_filter_match() in event_sched_out()
528 if (mode & PERF_CGROUP_SWIN) {
529 WARN_ON_ONCE(cpuctx->cgrp);
531 * set cgrp before ctxsw in to allow
532 * event_filter_match() to not have to pass
535 cpuctx->cgrp = perf_cgroup_from_task(task);
536 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
538 perf_pmu_enable(cpuctx->ctx.pmu);
539 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
545 local_irq_restore(flags);
548 static inline void perf_cgroup_sched_out(struct task_struct *task,
549 struct task_struct *next)
551 struct perf_cgroup *cgrp1;
552 struct perf_cgroup *cgrp2 = NULL;
555 * we come here when we know perf_cgroup_events > 0
557 cgrp1 = perf_cgroup_from_task(task);
560 * next is NULL when called from perf_event_enable_on_exec()
561 * that will systematically cause a cgroup_switch()
564 cgrp2 = perf_cgroup_from_task(next);
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.
572 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
575 static inline void perf_cgroup_sched_in(struct task_struct *prev,
576 struct task_struct *task)
578 struct perf_cgroup *cgrp1;
579 struct perf_cgroup *cgrp2 = NULL;
582 * we come here when we know perf_cgroup_events > 0
584 cgrp1 = perf_cgroup_from_task(task);
586 /* prev can never be NULL */
587 cgrp2 = perf_cgroup_from_task(prev);
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.
595 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
598 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
599 struct perf_event_attr *attr,
600 struct perf_event *group_leader)
602 struct perf_cgroup *cgrp;
603 struct cgroup_subsys_state *css;
604 struct fd f = fdget(fd);
610 css = css_tryget_from_dir(f.file->f_dentry, &perf_event_cgrp_subsys);
616 cgrp = container_of(css, struct perf_cgroup, css);
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
624 if (group_leader && group_leader->cgrp != cgrp) {
625 perf_detach_cgroup(event);
634 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
636 struct perf_cgroup_info *t;
637 t = per_cpu_ptr(event->cgrp->info, event->cpu);
638 event->shadow_ctx_time = now - t->timestamp;
642 perf_cgroup_defer_enabled(struct perf_event *event)
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.
650 if (is_cgroup_event(event) && !perf_cgroup_match(event))
651 event->cgrp_defer_enabled = 1;
655 perf_cgroup_mark_enabled(struct perf_event *event,
656 struct perf_event_context *ctx)
658 struct perf_event *sub;
659 u64 tstamp = perf_event_time(event);
661 if (!event->cgrp_defer_enabled)
664 event->cgrp_defer_enabled = 0;
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;
674 #else /* !CONFIG_CGROUP_PERF */
677 perf_cgroup_match(struct perf_event *event)
682 static inline void perf_detach_cgroup(struct perf_event *event)
685 static inline int is_cgroup_event(struct perf_event *event)
690 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
695 static inline void update_cgrp_time_from_event(struct perf_event *event)
699 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
703 static inline void perf_cgroup_sched_out(struct task_struct *task,
704 struct task_struct *next)
708 static inline void perf_cgroup_sched_in(struct task_struct *prev,
709 struct task_struct *task)
713 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
714 struct perf_event_attr *attr,
715 struct perf_event *group_leader)
721 perf_cgroup_set_timestamp(struct task_struct *task,
722 struct perf_event_context *ctx)
727 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
732 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
736 static inline u64 perf_cgroup_event_time(struct perf_event *event)
742 perf_cgroup_defer_enabled(struct perf_event *event)
747 perf_cgroup_mark_enabled(struct perf_event *event,
748 struct perf_event_context *ctx)
754 * set default to be dependent on timer tick just
757 #define PERF_CPU_HRTIMER (1000 / HZ)
759 * function must be called with interrupts disbled
761 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
763 struct perf_cpu_context *cpuctx;
764 enum hrtimer_restart ret = HRTIMER_NORESTART;
767 WARN_ON(!irqs_disabled());
769 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
771 rotations = perf_rotate_context(cpuctx);
774 * arm timer if needed
777 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
778 ret = HRTIMER_RESTART;
784 /* CPU is going down */
785 void perf_cpu_hrtimer_cancel(int cpu)
787 struct perf_cpu_context *cpuctx;
791 if (WARN_ON(cpu != smp_processor_id()))
794 local_irq_save(flags);
798 list_for_each_entry_rcu(pmu, &pmus, entry) {
799 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
801 if (pmu->task_ctx_nr == perf_sw_context)
804 hrtimer_cancel(&cpuctx->hrtimer);
809 local_irq_restore(flags);
812 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
814 struct hrtimer *hr = &cpuctx->hrtimer;
815 struct pmu *pmu = cpuctx->ctx.pmu;
818 /* no multiplexing needed for SW PMU */
819 if (pmu->task_ctx_nr == perf_sw_context)
823 * check default is sane, if not set then force to
824 * default interval (1/tick)
826 timer = pmu->hrtimer_interval_ms;
828 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
830 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
832 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
833 hr->function = perf_cpu_hrtimer_handler;
836 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
838 struct hrtimer *hr = &cpuctx->hrtimer;
839 struct pmu *pmu = cpuctx->ctx.pmu;
842 if (pmu->task_ctx_nr == perf_sw_context)
845 if (hrtimer_active(hr))
848 if (!hrtimer_callback_running(hr))
849 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
850 0, HRTIMER_MODE_REL_PINNED, 0);
853 void perf_pmu_disable(struct pmu *pmu)
855 int *count = this_cpu_ptr(pmu->pmu_disable_count);
857 pmu->pmu_disable(pmu);
860 void perf_pmu_enable(struct pmu *pmu)
862 int *count = this_cpu_ptr(pmu->pmu_disable_count);
864 pmu->pmu_enable(pmu);
867 static DEFINE_PER_CPU(struct list_head, rotation_list);
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.
874 static void perf_pmu_rotate_start(struct pmu *pmu)
876 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
877 struct list_head *head = &__get_cpu_var(rotation_list);
879 WARN_ON(!irqs_disabled());
881 if (list_empty(&cpuctx->rotation_list))
882 list_add(&cpuctx->rotation_list, head);
885 static void get_ctx(struct perf_event_context *ctx)
887 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
890 static void put_ctx(struct perf_event_context *ctx)
892 if (atomic_dec_and_test(&ctx->refcount)) {
894 put_ctx(ctx->parent_ctx);
896 put_task_struct(ctx->task);
897 kfree_rcu(ctx, rcu_head);
901 static void unclone_ctx(struct perf_event_context *ctx)
903 if (ctx->parent_ctx) {
904 put_ctx(ctx->parent_ctx);
905 ctx->parent_ctx = NULL;
910 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
913 * only top level events have the pid namespace they were created in
916 event = event->parent;
918 return task_tgid_nr_ns(p, event->ns);
921 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
924 * only top level events have the pid namespace they were created in
927 event = event->parent;
929 return task_pid_nr_ns(p, event->ns);
933 * If we inherit events we want to return the parent event id
936 static u64 primary_event_id(struct perf_event *event)
941 id = event->parent->id;
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.
951 static struct perf_event_context *
952 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
954 struct perf_event_context *ctx;
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().
963 * Since ctx->lock nests under rq->lock we must ensure the entire read
964 * side critical section is non-preemptible.
968 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
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.
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);
988 if (!atomic_inc_not_zero(&ctx->refcount)) {
989 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
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.
1003 static struct perf_event_context *
1004 perf_pin_task_context(struct task_struct *task, int ctxn)
1006 struct perf_event_context *ctx;
1007 unsigned long flags;
1009 ctx = perf_lock_task_context(task, ctxn, &flags);
1012 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1017 static void perf_unpin_context(struct perf_event_context *ctx)
1019 unsigned long flags;
1021 raw_spin_lock_irqsave(&ctx->lock, flags);
1023 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1027 * Update the record of the current time in a context.
1029 static void update_context_time(struct perf_event_context *ctx)
1031 u64 now = perf_clock();
1033 ctx->time += now - ctx->timestamp;
1034 ctx->timestamp = now;
1037 static u64 perf_event_time(struct perf_event *event)
1039 struct perf_event_context *ctx = event->ctx;
1041 if (is_cgroup_event(event))
1042 return perf_cgroup_event_time(event);
1044 return ctx ? ctx->time : 0;
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.
1051 static void update_event_times(struct perf_event *event)
1053 struct perf_event_context *ctx = event->ctx;
1056 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1057 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
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.
1066 * That is why we treat cgroup events differently
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;
1074 run_end = event->tstamp_stopped;
1076 event->total_time_enabled = run_end - event->tstamp_enabled;
1078 if (event->state == PERF_EVENT_STATE_INACTIVE)
1079 run_end = event->tstamp_stopped;
1081 run_end = perf_event_time(event);
1083 event->total_time_running = run_end - event->tstamp_running;
1088 * Update total_time_enabled and total_time_running for all events in a group.
1090 static void update_group_times(struct perf_event *leader)
1092 struct perf_event *event;
1094 update_event_times(leader);
1095 list_for_each_entry(event, &leader->sibling_list, group_entry)
1096 update_event_times(event);
1099 static struct list_head *
1100 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1102 if (event->attr.pinned)
1103 return &ctx->pinned_groups;
1105 return &ctx->flexible_groups;
1109 * Add a event from the lists for its context.
1110 * Must be called with ctx->mutex and ctx->lock held.
1113 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1115 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1116 event->attach_state |= PERF_ATTACH_CONTEXT;
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.
1123 if (event->group_leader == event) {
1124 struct list_head *list;
1126 if (is_software_event(event))
1127 event->group_flags |= PERF_GROUP_SOFTWARE;
1129 list = ctx_group_list(event, ctx);
1130 list_add_tail(&event->group_entry, list);
1133 if (is_cgroup_event(event))
1136 if (has_branch_stack(event))
1137 ctx->nr_branch_stack++;
1139 list_add_rcu(&event->event_entry, &ctx->event_list);
1140 if (!ctx->nr_events)
1141 perf_pmu_rotate_start(ctx->pmu);
1143 if (event->attr.inherit_stat)
1150 * Initialize event state based on the perf_event_attr::disabled.
1152 static inline void perf_event__state_init(struct perf_event *event)
1154 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1155 PERF_EVENT_STATE_INACTIVE;
1159 * Called at perf_event creation and when events are attached/detached from a
1162 static void perf_event__read_size(struct perf_event *event)
1164 int entry = sizeof(u64); /* value */
1168 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1169 size += sizeof(u64);
1171 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1172 size += sizeof(u64);
1174 if (event->attr.read_format & PERF_FORMAT_ID)
1175 entry += sizeof(u64);
1177 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1178 nr += event->group_leader->nr_siblings;
1179 size += sizeof(u64);
1183 event->read_size = size;
1186 static void perf_event__header_size(struct perf_event *event)
1188 struct perf_sample_data *data;
1189 u64 sample_type = event->attr.sample_type;
1192 perf_event__read_size(event);
1194 if (sample_type & PERF_SAMPLE_IP)
1195 size += sizeof(data->ip);
1197 if (sample_type & PERF_SAMPLE_ADDR)
1198 size += sizeof(data->addr);
1200 if (sample_type & PERF_SAMPLE_PERIOD)
1201 size += sizeof(data->period);
1203 if (sample_type & PERF_SAMPLE_WEIGHT)
1204 size += sizeof(data->weight);
1206 if (sample_type & PERF_SAMPLE_READ)
1207 size += event->read_size;
1209 if (sample_type & PERF_SAMPLE_DATA_SRC)
1210 size += sizeof(data->data_src.val);
1212 if (sample_type & PERF_SAMPLE_TRANSACTION)
1213 size += sizeof(data->txn);
1215 event->header_size = size;
1218 static void perf_event__id_header_size(struct perf_event *event)
1220 struct perf_sample_data *data;
1221 u64 sample_type = event->attr.sample_type;
1224 if (sample_type & PERF_SAMPLE_TID)
1225 size += sizeof(data->tid_entry);
1227 if (sample_type & PERF_SAMPLE_TIME)
1228 size += sizeof(data->time);
1230 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1231 size += sizeof(data->id);
1233 if (sample_type & PERF_SAMPLE_ID)
1234 size += sizeof(data->id);
1236 if (sample_type & PERF_SAMPLE_STREAM_ID)
1237 size += sizeof(data->stream_id);
1239 if (sample_type & PERF_SAMPLE_CPU)
1240 size += sizeof(data->cpu_entry);
1242 event->id_header_size = size;
1245 static void perf_group_attach(struct perf_event *event)
1247 struct perf_event *group_leader = event->group_leader, *pos;
1250 * We can have double attach due to group movement in perf_event_open.
1252 if (event->attach_state & PERF_ATTACH_GROUP)
1255 event->attach_state |= PERF_ATTACH_GROUP;
1257 if (group_leader == event)
1260 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1261 !is_software_event(event))
1262 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1264 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1265 group_leader->nr_siblings++;
1267 perf_event__header_size(group_leader);
1269 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1270 perf_event__header_size(pos);
1274 * Remove a event from the lists for its context.
1275 * Must be called with ctx->mutex and ctx->lock held.
1278 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1280 struct perf_cpu_context *cpuctx;
1282 * We can have double detach due to exit/hot-unplug + close.
1284 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1287 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1289 if (is_cgroup_event(event)) {
1291 cpuctx = __get_cpu_context(ctx);
1293 * if there are no more cgroup events
1294 * then cler cgrp to avoid stale pointer
1295 * in update_cgrp_time_from_cpuctx()
1297 if (!ctx->nr_cgroups)
1298 cpuctx->cgrp = NULL;
1301 if (has_branch_stack(event))
1302 ctx->nr_branch_stack--;
1305 if (event->attr.inherit_stat)
1308 list_del_rcu(&event->event_entry);
1310 if (event->group_leader == event)
1311 list_del_init(&event->group_entry);
1313 update_group_times(event);
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
1322 if (event->state > PERF_EVENT_STATE_OFF)
1323 event->state = PERF_EVENT_STATE_OFF;
1328 static void perf_group_detach(struct perf_event *event)
1330 struct perf_event *sibling, *tmp;
1331 struct list_head *list = NULL;
1334 * We can have double detach due to exit/hot-unplug + close.
1336 if (!(event->attach_state & PERF_ATTACH_GROUP))
1339 event->attach_state &= ~PERF_ATTACH_GROUP;
1342 * If this is a sibling, remove it from its group.
1344 if (event->group_leader != event) {
1345 list_del_init(&event->group_entry);
1346 event->group_leader->nr_siblings--;
1350 if (!list_empty(&event->group_entry))
1351 list = &event->group_entry;
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.
1358 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1360 list_move_tail(&sibling->group_entry, list);
1361 sibling->group_leader = sibling;
1363 /* Inherit group flags from the previous leader */
1364 sibling->group_flags = event->group_flags;
1368 perf_event__header_size(event->group_leader);
1370 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1371 perf_event__header_size(tmp);
1375 event_filter_match(struct perf_event *event)
1377 return (event->cpu == -1 || event->cpu == smp_processor_id())
1378 && perf_cgroup_match(event);
1382 event_sched_out(struct perf_event *event,
1383 struct perf_cpu_context *cpuctx,
1384 struct perf_event_context *ctx)
1386 u64 tstamp = perf_event_time(event);
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:
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;
1401 if (event->state != PERF_EVENT_STATE_ACTIVE)
1404 perf_pmu_disable(event->pmu);
1406 event->state = PERF_EVENT_STATE_INACTIVE;
1407 if (event->pending_disable) {
1408 event->pending_disable = 0;
1409 event->state = PERF_EVENT_STATE_OFF;
1411 event->tstamp_stopped = tstamp;
1412 event->pmu->del(event, 0);
1415 if (!is_software_event(event))
1416 cpuctx->active_oncpu--;
1418 if (event->attr.freq && event->attr.sample_freq)
1420 if (event->attr.exclusive || !cpuctx->active_oncpu)
1421 cpuctx->exclusive = 0;
1423 perf_pmu_enable(event->pmu);
1427 group_sched_out(struct perf_event *group_event,
1428 struct perf_cpu_context *cpuctx,
1429 struct perf_event_context *ctx)
1431 struct perf_event *event;
1432 int state = group_event->state;
1434 event_sched_out(group_event, cpuctx, ctx);
1437 * Schedule out siblings (if any):
1439 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1440 event_sched_out(event, cpuctx, ctx);
1442 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1443 cpuctx->exclusive = 0;
1446 struct remove_event {
1447 struct perf_event *event;
1452 * Cross CPU call to remove a performance event
1454 * We disable the event on the hardware level first. After that we
1455 * remove it from the context list.
1457 static int __perf_remove_from_context(void *info)
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);
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) {
1471 cpuctx->task_ctx = NULL;
1473 raw_spin_unlock(&ctx->lock);
1480 * Remove the event from a task's (or a CPU's) list of events.
1482 * CPU events are removed with a smp call. For task events we only
1483 * call when the task is on a CPU.
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.
1492 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1494 struct perf_event_context *ctx = event->ctx;
1495 struct task_struct *task = ctx->task;
1496 struct remove_event re = {
1498 .detach_group = detach_group,
1501 lockdep_assert_held(&ctx->mutex);
1505 * Per cpu events are removed via an smp call and
1506 * the removal is always successful.
1508 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1513 if (!task_function_call(task, __perf_remove_from_context, &re))
1516 raw_spin_lock_irq(&ctx->lock);
1518 * If we failed to find a running task, but find the context active now
1519 * that we've acquired the ctx->lock, retry.
1521 if (ctx->is_active) {
1522 raw_spin_unlock_irq(&ctx->lock);
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.
1531 perf_group_detach(event);
1532 list_del_event(event, ctx);
1533 raw_spin_unlock_irq(&ctx->lock);
1537 * Cross CPU call to disable a performance event
1539 int __perf_event_disable(void *info)
1541 struct perf_event *event = info;
1542 struct perf_event_context *ctx = event->ctx;
1543 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1546 * If this is a per-task event, need to check whether this
1547 * event's task is the current task on this cpu.
1549 * Can trigger due to concurrent perf_event_context_sched_out()
1550 * flipping contexts around.
1552 if (ctx->task && cpuctx->task_ctx != ctx)
1555 raw_spin_lock(&ctx->lock);
1558 * If the event is on, turn it off.
1559 * If it is in error state, leave it in error state.
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);
1568 event_sched_out(event, cpuctx, ctx);
1569 event->state = PERF_EVENT_STATE_OFF;
1572 raw_spin_unlock(&ctx->lock);
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.
1590 void perf_event_disable(struct perf_event *event)
1592 struct perf_event_context *ctx = event->ctx;
1593 struct task_struct *task = ctx->task;
1597 * Disable the event on the cpu that it's on
1599 cpu_function_call(event->cpu, __perf_event_disable, event);
1604 if (!task_function_call(task, __perf_event_disable, event))
1607 raw_spin_lock_irq(&ctx->lock);
1609 * If the event is still active, we need to retry the cross-call.
1611 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1612 raw_spin_unlock_irq(&ctx->lock);
1614 * Reload the task pointer, it might have been changed by
1615 * a concurrent perf_event_context_sched_out().
1622 * Since we have the lock this context can't be scheduled
1623 * in, so we can change the state safely.
1625 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1626 update_group_times(event);
1627 event->state = PERF_EVENT_STATE_OFF;
1629 raw_spin_unlock_irq(&ctx->lock);
1631 EXPORT_SYMBOL_GPL(perf_event_disable);
1633 static void perf_set_shadow_time(struct perf_event *event,
1634 struct perf_event_context *ctx,
1638 * use the correct time source for the time snapshot
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
1648 * tstamp - cgrp->timestamp.
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
1656 * But this is a bit hairy.
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.
1662 if (is_cgroup_event(event))
1663 perf_cgroup_set_shadow_time(event, tstamp);
1665 event->shadow_ctx_time = tstamp - ctx->timestamp;
1668 #define MAX_INTERRUPTS (~0ULL)
1670 static void perf_log_throttle(struct perf_event *event, int enable);
1673 event_sched_in(struct perf_event *event,
1674 struct perf_cpu_context *cpuctx,
1675 struct perf_event_context *ctx)
1677 u64 tstamp = perf_event_time(event);
1680 if (event->state <= PERF_EVENT_STATE_OFF)
1683 event->state = PERF_EVENT_STATE_ACTIVE;
1684 event->oncpu = smp_processor_id();
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.
1691 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1692 perf_log_throttle(event, 1);
1693 event->hw.interrupts = 0;
1697 * The new state must be visible before we turn it on in the hardware:
1701 perf_pmu_disable(event->pmu);
1703 if (event->pmu->add(event, PERF_EF_START)) {
1704 event->state = PERF_EVENT_STATE_INACTIVE;
1710 event->tstamp_running += tstamp - event->tstamp_stopped;
1712 perf_set_shadow_time(event, ctx, tstamp);
1714 if (!is_software_event(event))
1715 cpuctx->active_oncpu++;
1717 if (event->attr.freq && event->attr.sample_freq)
1720 if (event->attr.exclusive)
1721 cpuctx->exclusive = 1;
1724 perf_pmu_enable(event->pmu);
1730 group_sched_in(struct perf_event *group_event,
1731 struct perf_cpu_context *cpuctx,
1732 struct perf_event_context *ctx)
1734 struct perf_event *event, *partial_group = NULL;
1735 struct pmu *pmu = ctx->pmu;
1736 u64 now = ctx->time;
1737 bool simulate = false;
1739 if (group_event->state == PERF_EVENT_STATE_OFF)
1742 pmu->start_txn(pmu);
1744 if (event_sched_in(group_event, cpuctx, ctx)) {
1745 pmu->cancel_txn(pmu);
1746 perf_cpu_hrtimer_restart(cpuctx);
1751 * Schedule in siblings as one group (if any):
1753 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1754 if (event_sched_in(event, cpuctx, ctx)) {
1755 partial_group = event;
1760 if (!pmu->commit_txn(pmu))
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.
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.
1778 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1779 if (event == partial_group)
1783 event->tstamp_running += now - event->tstamp_stopped;
1784 event->tstamp_stopped = now;
1786 event_sched_out(event, cpuctx, ctx);
1789 event_sched_out(group_event, cpuctx, ctx);
1791 pmu->cancel_txn(pmu);
1793 perf_cpu_hrtimer_restart(cpuctx);
1799 * Work out whether we can put this event group on the CPU now.
1801 static int group_can_go_on(struct perf_event *event,
1802 struct perf_cpu_context *cpuctx,
1806 * Groups consisting entirely of software events can always go on.
1808 if (event->group_flags & PERF_GROUP_SOFTWARE)
1811 * If an exclusive group is already on, no other hardware
1814 if (cpuctx->exclusive)
1817 * If this group is exclusive and there are already
1818 * events on the CPU, it can't go on.
1820 if (event->attr.exclusive && cpuctx->active_oncpu)
1823 * Otherwise, try to add it if all previous groups were able
1829 static void add_event_to_ctx(struct perf_event *event,
1830 struct perf_event_context *ctx)
1832 u64 tstamp = perf_event_time(event);
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;
1841 static void task_ctx_sched_out(struct perf_event_context *ctx);
1843 ctx_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);
1848 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1849 struct perf_event_context *ctx,
1850 struct task_struct *task)
1852 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1854 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1855 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1857 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1861 * Cross CPU call to install and enable a performance event
1863 * Must be called with ctx->mutex held
1865 static int __perf_install_in_context(void *info)
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;
1873 perf_ctx_lock(cpuctx, task_ctx);
1874 perf_pmu_disable(cpuctx->ctx.pmu);
1877 * If there was an active task_ctx schedule it out.
1880 task_ctx_sched_out(task_ctx);
1883 * If the context we're installing events in is not the
1884 * active task_ctx, flip them.
1886 if (ctx->task && task_ctx != ctx) {
1888 raw_spin_unlock(&task_ctx->lock);
1889 raw_spin_lock(&ctx->lock);
1894 cpuctx->task_ctx = task_ctx;
1895 task = task_ctx->task;
1898 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1900 update_context_time(ctx);
1902 * update cgrp time only if current cgrp
1903 * matches event->cgrp. Must be done before
1904 * calling add_event_to_ctx()
1906 update_cgrp_time_from_event(event);
1908 add_event_to_ctx(event, ctx);
1911 * Schedule everything back in
1913 perf_event_sched_in(cpuctx, task_ctx, task);
1915 perf_pmu_enable(cpuctx->ctx.pmu);
1916 perf_ctx_unlock(cpuctx, task_ctx);
1922 * Attach a performance event to a context
1924 * First we add the event to the list with the hardware enable bit
1925 * in event->hw_config cleared.
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.
1932 perf_install_in_context(struct perf_event_context *ctx,
1933 struct perf_event *event,
1936 struct task_struct *task = ctx->task;
1938 lockdep_assert_held(&ctx->mutex);
1941 if (event->cpu != -1)
1946 * Per cpu events are installed via an smp call and
1947 * the install is always successful.
1949 cpu_function_call(cpu, __perf_install_in_context, event);
1954 if (!task_function_call(task, __perf_install_in_context, event))
1957 raw_spin_lock_irq(&ctx->lock);
1959 * If we failed to find a running task, but find the context active now
1960 * that we've acquired the ctx->lock, retry.
1962 if (ctx->is_active) {
1963 raw_spin_unlock_irq(&ctx->lock);
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.
1971 add_event_to_ctx(event, ctx);
1972 raw_spin_unlock_irq(&ctx->lock);
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.
1983 static void __perf_event_mark_enabled(struct perf_event *event)
1985 struct perf_event *sub;
1986 u64 tstamp = perf_event_time(event);
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;
1997 * Cross CPU call to enable a performance event
1999 static int __perf_event_enable(void *info)
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);
2008 * There's a time window between 'ctx->is_active' check
2009 * in perf_event_enable function and this place having:
2011 * - ctx->lock unlocked
2013 * where the task could be killed and 'ctx' deactivated
2014 * by perf_event_exit_task.
2016 if (!ctx->is_active)
2019 raw_spin_lock(&ctx->lock);
2020 update_context_time(ctx);
2022 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2026 * set current task's cgroup time reference point
2028 perf_cgroup_set_timestamp(current, ctx);
2030 __perf_event_mark_enabled(event);
2032 if (!event_filter_match(event)) {
2033 if (is_cgroup_event(event))
2034 perf_cgroup_defer_enabled(event);
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.
2042 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2045 if (!group_can_go_on(event, cpuctx, 1)) {
2048 if (event == leader)
2049 err = group_sched_in(event, cpuctx, ctx);
2051 err = event_sched_in(event, cpuctx, ctx);
2056 * If this event can't go on and it's part of a
2057 * group, then the whole group has to come off.
2059 if (leader != event) {
2060 group_sched_out(leader, cpuctx, ctx);
2061 perf_cpu_hrtimer_restart(cpuctx);
2063 if (leader->attr.pinned) {
2064 update_group_times(leader);
2065 leader->state = PERF_EVENT_STATE_ERROR;
2070 raw_spin_unlock(&ctx->lock);
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.
2084 void perf_event_enable(struct perf_event *event)
2086 struct perf_event_context *ctx = event->ctx;
2087 struct task_struct *task = ctx->task;
2091 * Enable the event on the cpu that it's on
2093 cpu_function_call(event->cpu, __perf_event_enable, event);
2097 raw_spin_lock_irq(&ctx->lock);
2098 if (event->state >= PERF_EVENT_STATE_INACTIVE)
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.
2108 if (event->state == PERF_EVENT_STATE_ERROR)
2109 event->state = PERF_EVENT_STATE_OFF;
2112 if (!ctx->is_active) {
2113 __perf_event_mark_enabled(event);
2117 raw_spin_unlock_irq(&ctx->lock);
2119 if (!task_function_call(task, __perf_event_enable, event))
2122 raw_spin_lock_irq(&ctx->lock);
2125 * If the context is active and the event is still off,
2126 * we need to retry the cross-call.
2128 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2130 * task could have been flipped by a concurrent
2131 * perf_event_context_sched_out()
2138 raw_spin_unlock_irq(&ctx->lock);
2140 EXPORT_SYMBOL_GPL(perf_event_enable);
2142 int perf_event_refresh(struct perf_event *event, int refresh)
2145 * not supported on inherited events
2147 if (event->attr.inherit || !is_sampling_event(event))
2150 atomic_add(refresh, &event->event_limit);
2151 perf_event_enable(event);
2155 EXPORT_SYMBOL_GPL(perf_event_refresh);
2157 static void ctx_sched_out(struct perf_event_context *ctx,
2158 struct perf_cpu_context *cpuctx,
2159 enum event_type_t event_type)
2161 struct perf_event *event;
2162 int is_active = ctx->is_active;
2164 ctx->is_active &= ~event_type;
2165 if (likely(!ctx->nr_events))
2168 update_context_time(ctx);
2169 update_cgrp_time_from_cpuctx(cpuctx);
2170 if (!ctx->nr_active)
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);
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);
2183 perf_pmu_enable(ctx->pmu);
2187 * Test whether two contexts are equivalent, i.e. whether they have both been
2188 * cloned from the same version of the same context.
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().
2194 static int context_equiv(struct perf_event_context *ctx1,
2195 struct perf_event_context *ctx2)
2197 /* Pinning disables the swap optimization */
2198 if (ctx1->pin_count || ctx2->pin_count)
2201 /* If ctx1 is the parent of ctx2 */
2202 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2205 /* If ctx2 is the parent of ctx1 */
2206 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2210 * If ctx1 and ctx2 have the same parent; we flatten the parent
2211 * hierarchy, see perf_event_init_context().
2213 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2214 ctx1->parent_gen == ctx2->parent_gen)
2221 static void __perf_event_sync_stat(struct perf_event *event,
2222 struct perf_event *next_event)
2226 if (!event->attr.inherit_stat)
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.
2236 switch (event->state) {
2237 case PERF_EVENT_STATE_ACTIVE:
2238 event->pmu->read(event);
2241 case PERF_EVENT_STATE_INACTIVE:
2242 update_event_times(event);
2250 * In order to keep per-task stats reliable we need to flip the event
2251 * values when we flip the contexts.
2253 value = local64_read(&next_event->count);
2254 value = local64_xchg(&event->count, value);
2255 local64_set(&next_event->count, value);
2257 swap(event->total_time_enabled, next_event->total_time_enabled);
2258 swap(event->total_time_running, next_event->total_time_running);
2261 * Since we swizzled the values, update the user visible data too.
2263 perf_event_update_userpage(event);
2264 perf_event_update_userpage(next_event);
2267 static void perf_event_sync_stat(struct perf_event_context *ctx,
2268 struct perf_event_context *next_ctx)
2270 struct perf_event *event, *next_event;
2275 update_context_time(ctx);
2277 event = list_first_entry(&ctx->event_list,
2278 struct perf_event, event_entry);
2280 next_event = list_first_entry(&next_ctx->event_list,
2281 struct perf_event, event_entry);
2283 while (&event->event_entry != &ctx->event_list &&
2284 &next_event->event_entry != &next_ctx->event_list) {
2286 __perf_event_sync_stat(event, next_event);
2288 event = list_next_entry(event, event_entry);
2289 next_event = list_next_entry(next_event, event_entry);
2293 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2294 struct task_struct *next)
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;
2305 cpuctx = __get_cpu_context(ctx);
2306 if (!cpuctx->task_ctx)
2310 next_ctx = next->perf_event_ctxp[ctxn];
2314 parent = rcu_dereference(ctx->parent_ctx);
2315 next_parent = rcu_dereference(next_ctx->parent_ctx);
2317 /* If neither context have a parent context; they cannot be clones. */
2318 if (!parent && !next_parent)
2321 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
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.
2331 raw_spin_lock(&ctx->lock);
2332 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2333 if (context_equiv(ctx, next_ctx)) {
2335 * XXX do we need a memory barrier of sorts
2336 * wrt to rcu_dereference() of perf_event_ctxp
2338 task->perf_event_ctxp[ctxn] = next_ctx;
2339 next->perf_event_ctxp[ctxn] = ctx;
2341 next_ctx->task = task;
2344 perf_event_sync_stat(ctx, next_ctx);
2346 raw_spin_unlock(&next_ctx->lock);
2347 raw_spin_unlock(&ctx->lock);
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);
2360 #define for_each_task_context_nr(ctxn) \
2361 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2364 * Called from scheduler to remove the events of the current task,
2365 * with interrupts disabled.
2367 * We stop each event and update the event value in event->count.
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.
2374 void __perf_event_task_sched_out(struct task_struct *task,
2375 struct task_struct *next)
2379 for_each_task_context_nr(ctxn)
2380 perf_event_context_sched_out(task, ctxn, next);
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
2387 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2388 perf_cgroup_sched_out(task, next);
2391 static void task_ctx_sched_out(struct perf_event_context *ctx)
2393 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2395 if (!cpuctx->task_ctx)
2398 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2401 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2402 cpuctx->task_ctx = NULL;
2406 * Called with IRQs disabled
2408 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2409 enum event_type_t event_type)
2411 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2415 ctx_pinned_sched_in(struct perf_event_context *ctx,
2416 struct perf_cpu_context *cpuctx)
2418 struct perf_event *event;
2420 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2421 if (event->state <= PERF_EVENT_STATE_OFF)
2423 if (!event_filter_match(event))
2426 /* may need to reset tstamp_enabled */
2427 if (is_cgroup_event(event))
2428 perf_cgroup_mark_enabled(event, ctx);
2430 if (group_can_go_on(event, cpuctx, 1))
2431 group_sched_in(event, cpuctx, ctx);
2434 * If this pinned group hasn't been scheduled,
2435 * put it in error state.
2437 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2438 update_group_times(event);
2439 event->state = PERF_EVENT_STATE_ERROR;
2445 ctx_flexible_sched_in(struct perf_event_context *ctx,
2446 struct perf_cpu_context *cpuctx)
2448 struct perf_event *event;
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)
2456 * Listen to the 'cpu' scheduling filter constraint
2459 if (!event_filter_match(event))
2462 /* may need to reset tstamp_enabled */
2463 if (is_cgroup_event(event))
2464 perf_cgroup_mark_enabled(event, ctx);
2466 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2467 if (group_sched_in(event, cpuctx, ctx))
2474 ctx_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)
2480 int is_active = ctx->is_active;
2482 ctx->is_active |= event_type;
2483 if (likely(!ctx->nr_events))
2487 ctx->timestamp = now;
2488 perf_cgroup_set_timestamp(task, ctx);
2490 * First go through the list and put on any pinned groups
2491 * in order to give them the best chance of going on.
2493 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2494 ctx_pinned_sched_in(ctx, cpuctx);
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);
2501 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2502 enum event_type_t event_type,
2503 struct task_struct *task)
2505 struct perf_event_context *ctx = &cpuctx->ctx;
2507 ctx_sched_in(ctx, cpuctx, event_type, task);
2510 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2511 struct task_struct *task)
2513 struct perf_cpu_context *cpuctx;
2515 cpuctx = __get_cpu_context(ctx);
2516 if (cpuctx->task_ctx == ctx)
2519 perf_ctx_lock(cpuctx, ctx);
2520 perf_pmu_disable(ctx->pmu);
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.
2526 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2529 cpuctx->task_ctx = ctx;
2531 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2533 perf_pmu_enable(ctx->pmu);
2534 perf_ctx_unlock(cpuctx, ctx);
2537 * Since these rotations are per-cpu, we need to ensure the
2538 * cpu-context we got scheduled on is actually rotating.
2540 perf_pmu_rotate_start(ctx->pmu);
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.
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.
2559 static void perf_branch_stack_sched_in(struct task_struct *prev,
2560 struct task_struct *task)
2562 struct perf_cpu_context *cpuctx;
2564 unsigned long flags;
2566 /* no need to flush branch stack if not changing task */
2570 local_irq_save(flags);
2574 list_for_each_entry_rcu(pmu, &pmus, entry) {
2575 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2578 * check if the context has at least one
2579 * event using PERF_SAMPLE_BRANCH_STACK
2581 if (cpuctx->ctx.nr_branch_stack > 0
2582 && pmu->flush_branch_stack) {
2584 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2586 perf_pmu_disable(pmu);
2588 pmu->flush_branch_stack();
2590 perf_pmu_enable(pmu);
2592 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2598 local_irq_restore(flags);
2602 * Called from scheduler to add the events of the current task
2603 * with interrupts disabled.
2605 * We restore the event value and then enable it.
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.
2612 void __perf_event_task_sched_in(struct task_struct *prev,
2613 struct task_struct *task)
2615 struct perf_event_context *ctx;
2618 for_each_task_context_nr(ctxn) {
2619 ctx = task->perf_event_ctxp[ctxn];
2623 perf_event_context_sched_in(ctx, task);
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
2630 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2631 perf_cgroup_sched_in(prev, task);
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);
2638 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2640 u64 frequency = event->attr.sample_freq;
2641 u64 sec = NSEC_PER_SEC;
2642 u64 divisor, dividend;
2644 int count_fls, nsec_fls, frequency_fls, sec_fls;
2646 count_fls = fls64(count);
2647 nsec_fls = fls64(nsec);
2648 frequency_fls = fls64(frequency);
2652 * We got @count in @nsec, with a target of sample_freq HZ
2653 * the target period becomes:
2656 * period = -------------------
2657 * @nsec * sample_freq
2662 * Reduce accuracy by one bit such that @a and @b converge
2663 * to a similar magnitude.
2665 #define REDUCE_FLS(a, b) \
2667 if (a##_fls > b##_fls) { \
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.
2680 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2681 REDUCE_FLS(nsec, frequency);
2682 REDUCE_FLS(sec, count);
2685 if (count_fls + sec_fls > 64) {
2686 divisor = nsec * frequency;
2688 while (count_fls + sec_fls > 64) {
2689 REDUCE_FLS(count, sec);
2693 dividend = count * sec;
2695 dividend = count * sec;
2697 while (nsec_fls + frequency_fls > 64) {
2698 REDUCE_FLS(nsec, frequency);
2702 divisor = nsec * frequency;
2708 return div64_u64(dividend, divisor);
2711 static DEFINE_PER_CPU(int, perf_throttled_count);
2712 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2714 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2716 struct hw_perf_event *hwc = &event->hw;
2717 s64 period, sample_period;
2720 period = perf_calculate_period(event, nsec, count);
2722 delta = (s64)(period - hwc->sample_period);
2723 delta = (delta + 7) / 8; /* low pass filter */
2725 sample_period = hwc->sample_period + delta;
2730 hwc->sample_period = sample_period;
2732 if (local64_read(&hwc->period_left) > 8*sample_period) {
2734 event->pmu->stop(event, PERF_EF_UPDATE);
2736 local64_set(&hwc->period_left, 0);
2739 event->pmu->start(event, PERF_EF_RELOAD);
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.
2748 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2751 struct perf_event *event;
2752 struct hw_perf_event *hwc;
2753 u64 now, period = TICK_NSEC;
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
2761 if (!(ctx->nr_freq || needs_unthr))
2764 raw_spin_lock(&ctx->lock);
2765 perf_pmu_disable(ctx->pmu);
2767 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2768 if (event->state != PERF_EVENT_STATE_ACTIVE)
2771 if (!event_filter_match(event))
2774 perf_pmu_disable(event->pmu);
2778 if (hwc->interrupts == MAX_INTERRUPTS) {
2779 hwc->interrupts = 0;
2780 perf_log_throttle(event, 1);
2781 event->pmu->start(event, 0);
2784 if (!event->attr.freq || !event->attr.sample_freq)
2788 * stop the event and update event->count
2790 event->pmu->stop(event, PERF_EF_UPDATE);
2792 now = local64_read(&event->count);
2793 delta = now - hwc->freq_count_stamp;
2794 hwc->freq_count_stamp = now;
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
2804 perf_adjust_period(event, period, delta, false);
2806 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2808 perf_pmu_enable(event->pmu);
2811 perf_pmu_enable(ctx->pmu);
2812 raw_spin_unlock(&ctx->lock);
2816 * Round-robin a context's events:
2818 static void rotate_ctx(struct perf_event_context *ctx)
2821 * Rotate the first entry last of non-pinned groups. Rotation might be
2822 * disabled by the inheritance code.
2824 if (!ctx->rotate_disable)
2825 list_rotate_left(&ctx->flexible_groups);
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.
2833 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2835 struct perf_event_context *ctx = NULL;
2836 int rotate = 0, remove = 1;
2838 if (cpuctx->ctx.nr_events) {
2840 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2844 ctx = cpuctx->task_ctx;
2845 if (ctx && ctx->nr_events) {
2847 if (ctx->nr_events != ctx->nr_active)
2854 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2855 perf_pmu_disable(cpuctx->ctx.pmu);
2857 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2859 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2861 rotate_ctx(&cpuctx->ctx);
2865 perf_event_sched_in(cpuctx, ctx, current);
2867 perf_pmu_enable(cpuctx->ctx.pmu);
2868 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2871 list_del_init(&cpuctx->rotation_list);
2876 #ifdef CONFIG_NO_HZ_FULL
2877 bool perf_event_can_stop_tick(void)
2879 if (atomic_read(&nr_freq_events) ||
2880 __this_cpu_read(perf_throttled_count))
2887 void perf_event_task_tick(void)
2889 struct list_head *head = &__get_cpu_var(rotation_list);
2890 struct perf_cpu_context *cpuctx, *tmp;
2891 struct perf_event_context *ctx;
2894 WARN_ON(!irqs_disabled());
2896 __this_cpu_inc(perf_throttled_seq);
2897 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2899 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2901 perf_adjust_freq_unthr_context(ctx, throttled);
2903 ctx = cpuctx->task_ctx;
2905 perf_adjust_freq_unthr_context(ctx, throttled);
2909 static int event_enable_on_exec(struct perf_event *event,
2910 struct perf_event_context *ctx)
2912 if (!event->attr.enable_on_exec)
2915 event->attr.enable_on_exec = 0;
2916 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2919 __perf_event_mark_enabled(event);
2925 * Enable all of a task's events that have been marked enable-on-exec.
2926 * This expects task == current.
2928 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2930 struct perf_event *event;
2931 unsigned long flags;
2935 local_irq_save(flags);
2936 if (!ctx || !ctx->nr_events)
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
2946 perf_cgroup_sched_out(current, NULL);
2948 raw_spin_lock(&ctx->lock);
2949 task_ctx_sched_out(ctx);
2951 list_for_each_entry(event, &ctx->event_list, event_entry) {
2952 ret = event_enable_on_exec(event, ctx);
2958 * Unclone this context if we enabled any event.
2963 raw_spin_unlock(&ctx->lock);
2966 * Also calls ctxswin for cgroup events, if any:
2968 perf_event_context_sched_in(ctx, ctx->task);
2970 local_irq_restore(flags);
2974 * Cross CPU call to read the hardware event
2976 static void __perf_event_read(void *info)
2978 struct perf_event *event = info;
2979 struct perf_event_context *ctx = event->ctx;
2980 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
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.
2989 if (ctx->task && cpuctx->task_ctx != ctx)
2992 raw_spin_lock(&ctx->lock);
2993 if (ctx->is_active) {
2994 update_context_time(ctx);
2995 update_cgrp_time_from_event(event);
2997 update_event_times(event);
2998 if (event->state == PERF_EVENT_STATE_ACTIVE)
2999 event->pmu->read(event);
3000 raw_spin_unlock(&ctx->lock);
3003 static inline u64 perf_event_count(struct perf_event *event)
3005 return local64_read(&event->count) + atomic64_read(&event->child_count);
3008 static u64 perf_event_read(struct perf_event *event)
3011 * If event is enabled and currently active on a CPU, update the
3012 * value in the event structure:
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;
3021 raw_spin_lock_irqsave(&ctx->lock, flags);
3023 * may read while context is not active
3024 * (e.g., thread is blocked), in that case
3025 * we cannot update context time
3027 if (ctx->is_active) {
3028 update_context_time(ctx);
3029 update_cgrp_time_from_event(event);
3031 update_event_times(event);
3032 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3035 return perf_event_count(event);
3039 * Initialize the perf_event context in a task_struct:
3041 static void __perf_event_init_context(struct perf_event_context *ctx)
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);
3051 static struct perf_event_context *
3052 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3054 struct perf_event_context *ctx;
3056 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3060 __perf_event_init_context(ctx);
3063 get_task_struct(task);
3070 static struct task_struct *
3071 find_lively_task_by_vpid(pid_t vpid)
3073 struct task_struct *task;
3080 task = find_task_by_vpid(vpid);
3082 get_task_struct(task);
3086 return ERR_PTR(-ESRCH);
3088 /* Reuse ptrace permission checks for now. */
3090 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3095 put_task_struct(task);
3096 return ERR_PTR(err);
3101 * Returns a matching context with refcount and pincount.
3103 static struct perf_event_context *
3104 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3106 struct perf_event_context *ctx;
3107 struct perf_cpu_context *cpuctx;
3108 unsigned long flags;
3112 /* Must be root to operate on a CPU event: */
3113 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3114 return ERR_PTR(-EACCES);
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
3121 if (!cpu_online(cpu))
3122 return ERR_PTR(-ENODEV);
3124 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3133 ctxn = pmu->task_ctx_nr;
3138 ctx = perf_lock_task_context(task, ctxn, &flags);
3142 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3144 ctx = alloc_perf_context(pmu, task);
3150 mutex_lock(&task->perf_event_mutex);
3152 * If it has already passed perf_event_exit_task().
3153 * we must see PF_EXITING, it takes this mutex too.
3155 if (task->flags & PF_EXITING)
3157 else if (task->perf_event_ctxp[ctxn])
3162 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3164 mutex_unlock(&task->perf_event_mutex);
3166 if (unlikely(err)) {
3178 return ERR_PTR(err);
3181 static void perf_event_free_filter(struct perf_event *event);
3183 static void free_event_rcu(struct rcu_head *head)
3185 struct perf_event *event;
3187 event = container_of(head, struct perf_event, rcu_head);
3189 put_pid_ns(event->ns);
3190 perf_event_free_filter(event);
3194 static void ring_buffer_put(struct ring_buffer *rb);
3195 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3197 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3202 if (has_branch_stack(event)) {
3203 if (!(event->attach_state & PERF_ATTACH_TASK))
3204 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3206 if (is_cgroup_event(event))
3207 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3210 static void unaccount_event(struct perf_event *event)
3215 if (event->attach_state & PERF_ATTACH_TASK)
3216 static_key_slow_dec_deferred(&perf_sched_events);
3217 if (event->attr.mmap || event->attr.mmap_data)
3218 atomic_dec(&nr_mmap_events);
3219 if (event->attr.comm)
3220 atomic_dec(&nr_comm_events);
3221 if (event->attr.task)
3222 atomic_dec(&nr_task_events);
3223 if (event->attr.freq)
3224 atomic_dec(&nr_freq_events);
3225 if (is_cgroup_event(event))
3226 static_key_slow_dec_deferred(&perf_sched_events);
3227 if (has_branch_stack(event))
3228 static_key_slow_dec_deferred(&perf_sched_events);
3230 unaccount_event_cpu(event, event->cpu);
3233 static void __free_event(struct perf_event *event)
3235 if (!event->parent) {
3236 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3237 put_callchain_buffers();
3241 event->destroy(event);
3244 put_ctx(event->ctx);
3246 call_rcu(&event->rcu_head, free_event_rcu);
3248 static void free_event(struct perf_event *event)
3250 irq_work_sync(&event->pending);
3252 unaccount_event(event);
3255 struct ring_buffer *rb;
3258 * Can happen when we close an event with re-directed output.
3260 * Since we have a 0 refcount, perf_mmap_close() will skip
3261 * over us; possibly making our ring_buffer_put() the last.
3263 mutex_lock(&event->mmap_mutex);
3266 rcu_assign_pointer(event->rb, NULL);
3267 ring_buffer_detach(event, rb);
3268 ring_buffer_put(rb); /* could be last */
3270 mutex_unlock(&event->mmap_mutex);
3273 if (is_cgroup_event(event))
3274 perf_detach_cgroup(event);
3277 __free_event(event);
3280 int perf_event_release_kernel(struct perf_event *event)
3282 struct perf_event_context *ctx = event->ctx;
3284 WARN_ON_ONCE(ctx->parent_ctx);
3286 * There are two ways this annotation is useful:
3288 * 1) there is a lock recursion from perf_event_exit_task
3289 * see the comment there.
3291 * 2) there is a lock-inversion with mmap_sem through
3292 * perf_event_read_group(), which takes faults while
3293 * holding ctx->mutex, however this is called after
3294 * the last filedesc died, so there is no possibility
3295 * to trigger the AB-BA case.
3297 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3298 perf_remove_from_context(event, true);
3299 mutex_unlock(&ctx->mutex);
3305 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3308 * Called when the last reference to the file is gone.
3310 static void put_event(struct perf_event *event)
3312 struct task_struct *owner;
3314 if (!atomic_long_dec_and_test(&event->refcount))
3318 owner = ACCESS_ONCE(event->owner);
3320 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3321 * !owner it means the list deletion is complete and we can indeed
3322 * free this event, otherwise we need to serialize on
3323 * owner->perf_event_mutex.
3325 smp_read_barrier_depends();
3328 * Since delayed_put_task_struct() also drops the last
3329 * task reference we can safely take a new reference
3330 * while holding the rcu_read_lock().
3332 get_task_struct(owner);
3337 mutex_lock(&owner->perf_event_mutex);
3339 * We have to re-check the event->owner field, if it is cleared
3340 * we raced with perf_event_exit_task(), acquiring the mutex
3341 * ensured they're done, and we can proceed with freeing the
3345 list_del_init(&event->owner_entry);
3346 mutex_unlock(&owner->perf_event_mutex);
3347 put_task_struct(owner);
3350 perf_event_release_kernel(event);
3353 static int perf_release(struct inode *inode, struct file *file)
3355 put_event(file->private_data);
3359 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3361 struct perf_event *child;
3367 mutex_lock(&event->child_mutex);
3368 total += perf_event_read(event);
3369 *enabled += event->total_time_enabled +
3370 atomic64_read(&event->child_total_time_enabled);
3371 *running += event->total_time_running +
3372 atomic64_read(&event->child_total_time_running);
3374 list_for_each_entry(child, &event->child_list, child_list) {
3375 total += perf_event_read(child);
3376 *enabled += child->total_time_enabled;
3377 *running += child->total_time_running;
3379 mutex_unlock(&event->child_mutex);
3383 EXPORT_SYMBOL_GPL(perf_event_read_value);
3385 static int perf_event_read_group(struct perf_event *event,
3386 u64 read_format, char __user *buf)
3388 struct perf_event *leader = event->group_leader, *sub;
3389 int n = 0, size = 0, ret = -EFAULT;
3390 struct perf_event_context *ctx = leader->ctx;
3392 u64 count, enabled, running;
3394 mutex_lock(&ctx->mutex);
3395 count = perf_event_read_value(leader, &enabled, &running);
3397 values[n++] = 1 + leader->nr_siblings;
3398 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3399 values[n++] = enabled;
3400 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3401 values[n++] = running;
3402 values[n++] = count;
3403 if (read_format & PERF_FORMAT_ID)
3404 values[n++] = primary_event_id(leader);
3406 size = n * sizeof(u64);
3408 if (copy_to_user(buf, values, size))
3413 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3416 values[n++] = perf_event_read_value(sub, &enabled, &running);
3417 if (read_format & PERF_FORMAT_ID)
3418 values[n++] = primary_event_id(sub);
3420 size = n * sizeof(u64);
3422 if (copy_to_user(buf + ret, values, size)) {
3430 mutex_unlock(&ctx->mutex);
3435 static int perf_event_read_one(struct perf_event *event,
3436 u64 read_format, char __user *buf)
3438 u64 enabled, running;
3442 values[n++] = perf_event_read_value(event, &enabled, &running);
3443 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3444 values[n++] = enabled;
3445 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3446 values[n++] = running;
3447 if (read_format & PERF_FORMAT_ID)
3448 values[n++] = primary_event_id(event);
3450 if (copy_to_user(buf, values, n * sizeof(u64)))
3453 return n * sizeof(u64);
3457 * Read the performance event - simple non blocking version for now
3460 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3462 u64 read_format = event->attr.read_format;
3466 * Return end-of-file for a read on a event that is in
3467 * error state (i.e. because it was pinned but it couldn't be
3468 * scheduled on to the CPU at some point).
3470 if (event->state == PERF_EVENT_STATE_ERROR)
3473 if (count < event->read_size)
3476 WARN_ON_ONCE(event->ctx->parent_ctx);
3477 if (read_format & PERF_FORMAT_GROUP)
3478 ret = perf_event_read_group(event, read_format, buf);
3480 ret = perf_event_read_one(event, read_format, buf);
3486 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3488 struct perf_event *event = file->private_data;
3490 return perf_read_hw(event, buf, count);
3493 static unsigned int perf_poll(struct file *file, poll_table *wait)
3495 struct perf_event *event = file->private_data;
3496 struct ring_buffer *rb;
3497 unsigned int events = POLL_HUP;
3500 * Pin the event->rb by taking event->mmap_mutex; otherwise
3501 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3503 mutex_lock(&event->mmap_mutex);
3506 events = atomic_xchg(&rb->poll, 0);
3507 mutex_unlock(&event->mmap_mutex);
3509 poll_wait(file, &event->waitq, wait);
3514 static void perf_event_reset(struct perf_event *event)
3516 (void)perf_event_read(event);
3517 local64_set(&event->count, 0);
3518 perf_event_update_userpage(event);
3522 * Holding the top-level event's child_mutex means that any
3523 * descendant process that has inherited this event will block
3524 * in sync_child_event if it goes to exit, thus satisfying the
3525 * task existence requirements of perf_event_enable/disable.
3527 static void perf_event_for_each_child(struct perf_event *event,
3528 void (*func)(struct perf_event *))
3530 struct perf_event *child;
3532 WARN_ON_ONCE(event->ctx->parent_ctx);
3533 mutex_lock(&event->child_mutex);
3535 list_for_each_entry(child, &event->child_list, child_list)
3537 mutex_unlock(&event->child_mutex);
3540 static void perf_event_for_each(struct perf_event *event,
3541 void (*func)(struct perf_event *))
3543 struct perf_event_context *ctx = event->ctx;
3544 struct perf_event *sibling;
3546 WARN_ON_ONCE(ctx->parent_ctx);
3547 mutex_lock(&ctx->mutex);
3548 event = event->group_leader;
3550 perf_event_for_each_child(event, func);
3551 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3552 perf_event_for_each_child(sibling, func);
3553 mutex_unlock(&ctx->mutex);
3556 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3558 struct perf_event_context *ctx = event->ctx;
3559 int ret = 0, active;
3562 if (!is_sampling_event(event))
3565 if (copy_from_user(&value, arg, sizeof(value)))
3571 raw_spin_lock_irq(&ctx->lock);
3572 if (event->attr.freq) {
3573 if (value > sysctl_perf_event_sample_rate) {
3578 event->attr.sample_freq = value;
3580 event->attr.sample_period = value;
3581 event->hw.sample_period = value;
3584 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3586 perf_pmu_disable(ctx->pmu);
3587 event->pmu->stop(event, PERF_EF_UPDATE);
3590 local64_set(&event->hw.period_left, 0);
3593 event->pmu->start(event, PERF_EF_RELOAD);
3594 perf_pmu_enable(ctx->pmu);
3598 raw_spin_unlock_irq(&ctx->lock);
3603 static const struct file_operations perf_fops;
3605 static inline int perf_fget_light(int fd, struct fd *p)
3607 struct fd f = fdget(fd);
3611 if (f.file->f_op != &perf_fops) {
3619 static int perf_event_set_output(struct perf_event *event,
3620 struct perf_event *output_event);
3621 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3623 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3625 struct perf_event *event = file->private_data;
3626 void (*func)(struct perf_event *);
3630 case PERF_EVENT_IOC_ENABLE:
3631 func = perf_event_enable;
3633 case PERF_EVENT_IOC_DISABLE:
3634 func = perf_event_disable;
3636 case PERF_EVENT_IOC_RESET:
3637 func = perf_event_reset;
3640 case PERF_EVENT_IOC_REFRESH:
3641 return perf_event_refresh(event, arg);
3643 case PERF_EVENT_IOC_PERIOD:
3644 return perf_event_period(event, (u64 __user *)arg);
3646 case PERF_EVENT_IOC_ID:
3648 u64 id = primary_event_id(event);
3650 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3655 case PERF_EVENT_IOC_SET_OUTPUT:
3659 struct perf_event *output_event;
3661 ret = perf_fget_light(arg, &output);
3664 output_event = output.file->private_data;
3665 ret = perf_event_set_output(event, output_event);
3668 ret = perf_event_set_output(event, NULL);
3673 case PERF_EVENT_IOC_SET_FILTER:
3674 return perf_event_set_filter(event, (void __user *)arg);
3680 if (flags & PERF_IOC_FLAG_GROUP)
3681 perf_event_for_each(event, func);
3683 perf_event_for_each_child(event, func);
3688 int perf_event_task_enable(void)
3690 struct perf_event *event;
3692 mutex_lock(¤t->perf_event_mutex);
3693 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3694 perf_event_for_each_child(event, perf_event_enable);
3695 mutex_unlock(¤t->perf_event_mutex);
3700 int perf_event_task_disable(void)
3702 struct perf_event *event;
3704 mutex_lock(¤t->perf_event_mutex);
3705 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3706 perf_event_for_each_child(event, perf_event_disable);
3707 mutex_unlock(¤t->perf_event_mutex);
3712 static int perf_event_index(struct perf_event *event)
3714 if (event->hw.state & PERF_HES_STOPPED)
3717 if (event->state != PERF_EVENT_STATE_ACTIVE)
3720 return event->pmu->event_idx(event);
3723 static void calc_timer_values(struct perf_event *event,
3730 *now = perf_clock();
3731 ctx_time = event->shadow_ctx_time + *now;
3732 *enabled = ctx_time - event->tstamp_enabled;
3733 *running = ctx_time - event->tstamp_running;
3736 static void perf_event_init_userpage(struct perf_event *event)
3738 struct perf_event_mmap_page *userpg;
3739 struct ring_buffer *rb;
3742 rb = rcu_dereference(event->rb);
3746 userpg = rb->user_page;
3748 /* Allow new userspace to detect that bit 0 is deprecated */
3749 userpg->cap_bit0_is_deprecated = 1;
3750 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3756 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3761 * Callers need to ensure there can be no nesting of this function, otherwise
3762 * the seqlock logic goes bad. We can not serialize this because the arch
3763 * code calls this from NMI context.
3765 void perf_event_update_userpage(struct perf_event *event)
3767 struct perf_event_mmap_page *userpg;
3768 struct ring_buffer *rb;
3769 u64 enabled, running, now;
3772 rb = rcu_dereference(event->rb);
3777 * compute total_time_enabled, total_time_running
3778 * based on snapshot values taken when the event
3779 * was last scheduled in.
3781 * we cannot simply called update_context_time()
3782 * because of locking issue as we can be called in
3785 calc_timer_values(event, &now, &enabled, &running);
3787 userpg = rb->user_page;
3789 * Disable preemption so as to not let the corresponding user-space
3790 * spin too long if we get preempted.
3795 userpg->index = perf_event_index(event);
3796 userpg->offset = perf_event_count(event);
3798 userpg->offset -= local64_read(&event->hw.prev_count);
3800 userpg->time_enabled = enabled +
3801 atomic64_read(&event->child_total_time_enabled);
3803 userpg->time_running = running +
3804 atomic64_read(&event->child_total_time_running);
3806 arch_perf_update_userpage(userpg, now);
3815 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3817 struct perf_event *event = vma->vm_file->private_data;
3818 struct ring_buffer *rb;
3819 int ret = VM_FAULT_SIGBUS;
3821 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3822 if (vmf->pgoff == 0)
3828 rb = rcu_dereference(event->rb);
3832 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3835 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3839 get_page(vmf->page);
3840 vmf->page->mapping = vma->vm_file->f_mapping;
3841 vmf->page->index = vmf->pgoff;
3850 static void ring_buffer_attach(struct perf_event *event,
3851 struct ring_buffer *rb)
3853 unsigned long flags;
3855 if (!list_empty(&event->rb_entry))
3858 spin_lock_irqsave(&rb->event_lock, flags);
3859 if (list_empty(&event->rb_entry))
3860 list_add(&event->rb_entry, &rb->event_list);
3861 spin_unlock_irqrestore(&rb->event_lock, flags);
3864 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3866 unsigned long flags;
3868 if (list_empty(&event->rb_entry))
3871 spin_lock_irqsave(&rb->event_lock, flags);
3872 list_del_init(&event->rb_entry);
3873 wake_up_all(&event->waitq);
3874 spin_unlock_irqrestore(&rb->event_lock, flags);
3877 static void ring_buffer_wakeup(struct perf_event *event)
3879 struct ring_buffer *rb;
3882 rb = rcu_dereference(event->rb);
3884 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3885 wake_up_all(&event->waitq);
3890 static void rb_free_rcu(struct rcu_head *rcu_head)
3892 struct ring_buffer *rb;
3894 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3898 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3900 struct ring_buffer *rb;
3903 rb = rcu_dereference(event->rb);
3905 if (!atomic_inc_not_zero(&rb->refcount))
3913 static void ring_buffer_put(struct ring_buffer *rb)
3915 if (!atomic_dec_and_test(&rb->refcount))
3918 WARN_ON_ONCE(!list_empty(&rb->event_list));
3920 call_rcu(&rb->rcu_head, rb_free_rcu);
3923 static void perf_mmap_open(struct vm_area_struct *vma)
3925 struct perf_event *event = vma->vm_file->private_data;
3927 atomic_inc(&event->mmap_count);
3928 atomic_inc(&event->rb->mmap_count);
3932 * A buffer can be mmap()ed multiple times; either directly through the same
3933 * event, or through other events by use of perf_event_set_output().
3935 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3936 * the buffer here, where we still have a VM context. This means we need
3937 * to detach all events redirecting to us.
3939 static void perf_mmap_close(struct vm_area_struct *vma)
3941 struct perf_event *event = vma->vm_file->private_data;
3943 struct ring_buffer *rb = event->rb;
3944 struct user_struct *mmap_user = rb->mmap_user;
3945 int mmap_locked = rb->mmap_locked;
3946 unsigned long size = perf_data_size(rb);
3948 atomic_dec(&rb->mmap_count);
3950 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3953 /* Detach current event from the buffer. */
3954 rcu_assign_pointer(event->rb, NULL);
3955 ring_buffer_detach(event, rb);
3956 mutex_unlock(&event->mmap_mutex);
3958 /* If there's still other mmap()s of this buffer, we're done. */
3959 if (atomic_read(&rb->mmap_count)) {
3960 ring_buffer_put(rb); /* can't be last */
3965 * No other mmap()s, detach from all other events that might redirect
3966 * into the now unreachable buffer. Somewhat complicated by the
3967 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3971 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3972 if (!atomic_long_inc_not_zero(&event->refcount)) {
3974 * This event is en-route to free_event() which will
3975 * detach it and remove it from the list.
3981 mutex_lock(&event->mmap_mutex);
3983 * Check we didn't race with perf_event_set_output() which can
3984 * swizzle the rb from under us while we were waiting to
3985 * acquire mmap_mutex.
3987 * If we find a different rb; ignore this event, a next
3988 * iteration will no longer find it on the list. We have to
3989 * still restart the iteration to make sure we're not now
3990 * iterating the wrong list.
3992 if (event->rb == rb) {
3993 rcu_assign_pointer(event->rb, NULL);
3994 ring_buffer_detach(event, rb);
3995 ring_buffer_put(rb); /* can't be last, we still have one */
3997 mutex_unlock(&event->mmap_mutex);
4001 * Restart the iteration; either we're on the wrong list or
4002 * destroyed its integrity by doing a deletion.
4009 * It could be there's still a few 0-ref events on the list; they'll
4010 * get cleaned up by free_event() -- they'll also still have their
4011 * ref on the rb and will free it whenever they are done with it.
4013 * Aside from that, this buffer is 'fully' detached and unmapped,
4014 * undo the VM accounting.
4017 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4018 vma->vm_mm->pinned_vm -= mmap_locked;
4019 free_uid(mmap_user);
4021 ring_buffer_put(rb); /* could be last */
4024 static const struct vm_operations_struct perf_mmap_vmops = {
4025 .open = perf_mmap_open,
4026 .close = perf_mmap_close,
4027 .fault = perf_mmap_fault,
4028 .page_mkwrite = perf_mmap_fault,
4031 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4033 struct perf_event *event = file->private_data;
4034 unsigned long user_locked, user_lock_limit;
4035 struct user_struct *user = current_user();
4036 unsigned long locked, lock_limit;
4037 struct ring_buffer *rb;
4038 unsigned long vma_size;
4039 unsigned long nr_pages;
4040 long user_extra, extra;
4041 int ret = 0, flags = 0;
4044 * Don't allow mmap() of inherited per-task counters. This would
4045 * create a performance issue due to all children writing to the
4048 if (event->cpu == -1 && event->attr.inherit)
4051 if (!(vma->vm_flags & VM_SHARED))
4054 vma_size = vma->vm_end - vma->vm_start;
4055 nr_pages = (vma_size / PAGE_SIZE) - 1;
4058 * If we have rb pages ensure they're a power-of-two number, so we
4059 * can do bitmasks instead of modulo.
4061 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4064 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4067 if (vma->vm_pgoff != 0)
4070 WARN_ON_ONCE(event->ctx->parent_ctx);
4072 mutex_lock(&event->mmap_mutex);
4074 if (event->rb->nr_pages != nr_pages) {
4079 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4081 * Raced against perf_mmap_close() through
4082 * perf_event_set_output(). Try again, hope for better
4085 mutex_unlock(&event->mmap_mutex);
4092 user_extra = nr_pages + 1;
4093 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4096 * Increase the limit linearly with more CPUs:
4098 user_lock_limit *= num_online_cpus();
4100 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4103 if (user_locked > user_lock_limit)
4104 extra = user_locked - user_lock_limit;
4106 lock_limit = rlimit(RLIMIT_MEMLOCK);
4107 lock_limit >>= PAGE_SHIFT;
4108 locked = vma->vm_mm->pinned_vm + extra;
4110 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4111 !capable(CAP_IPC_LOCK)) {
4118 if (vma->vm_flags & VM_WRITE)
4119 flags |= RING_BUFFER_WRITABLE;
4121 rb = rb_alloc(nr_pages,
4122 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4130 atomic_set(&rb->mmap_count, 1);
4131 rb->mmap_locked = extra;
4132 rb->mmap_user = get_current_user();
4134 atomic_long_add(user_extra, &user->locked_vm);
4135 vma->vm_mm->pinned_vm += extra;
4137 ring_buffer_attach(event, rb);
4138 rcu_assign_pointer(event->rb, rb);
4140 perf_event_init_userpage(event);
4141 perf_event_update_userpage(event);
4145 atomic_inc(&event->mmap_count);
4146 mutex_unlock(&event->mmap_mutex);
4149 * Since pinned accounting is per vm we cannot allow fork() to copy our
4152 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4153 vma->vm_ops = &perf_mmap_vmops;
4158 static int perf_fasync(int fd, struct file *filp, int on)
4160 struct inode *inode = file_inode(filp);
4161 struct perf_event *event = filp->private_data;
4164 mutex_lock(&inode->i_mutex);
4165 retval = fasync_helper(fd, filp, on, &event->fasync);
4166 mutex_unlock(&inode->i_mutex);
4174 static const struct file_operations perf_fops = {
4175 .llseek = no_llseek,
4176 .release = perf_release,
4179 .unlocked_ioctl = perf_ioctl,
4180 .compat_ioctl = perf_ioctl,
4182 .fasync = perf_fasync,
4188 * If there's data, ensure we set the poll() state and publish everything
4189 * to user-space before waking everybody up.
4192 void perf_event_wakeup(struct perf_event *event)
4194 ring_buffer_wakeup(event);
4196 if (event->pending_kill) {
4197 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4198 event->pending_kill = 0;
4202 static void perf_pending_event(struct irq_work *entry)
4204 struct perf_event *event = container_of(entry,
4205 struct perf_event, pending);
4207 if (event->pending_disable) {
4208 event->pending_disable = 0;
4209 __perf_event_disable(event);
4212 if (event->pending_wakeup) {
4213 event->pending_wakeup = 0;
4214 perf_event_wakeup(event);
4219 * We assume there is only KVM supporting the callbacks.
4220 * Later on, we might change it to a list if there is
4221 * another virtualization implementation supporting the callbacks.
4223 struct perf_guest_info_callbacks *perf_guest_cbs;
4225 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4227 perf_guest_cbs = cbs;
4230 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4232 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4234 perf_guest_cbs = NULL;
4237 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4240 perf_output_sample_regs(struct perf_output_handle *handle,
4241 struct pt_regs *regs, u64 mask)
4245 for_each_set_bit(bit, (const unsigned long *) &mask,
4246 sizeof(mask) * BITS_PER_BYTE) {
4249 val = perf_reg_value(regs, bit);
4250 perf_output_put(handle, val);
4254 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4255 struct pt_regs *regs)
4257 if (!user_mode(regs)) {
4259 regs = task_pt_regs(current);
4265 regs_user->regs = regs;
4266 regs_user->abi = perf_reg_abi(current);
4271 * Get remaining task size from user stack pointer.
4273 * It'd be better to take stack vma map and limit this more
4274 * precisly, but there's no way to get it safely under interrupt,
4275 * so using TASK_SIZE as limit.
4277 static u64 perf_ustack_task_size(struct pt_regs *regs)
4279 unsigned long addr = perf_user_stack_pointer(regs);
4281 if (!addr || addr >= TASK_SIZE)
4284 return TASK_SIZE - addr;
4288 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4289 struct pt_regs *regs)
4293 /* No regs, no stack pointer, no dump. */
4298 * Check if we fit in with the requested stack size into the:
4300 * If we don't, we limit the size to the TASK_SIZE.
4302 * - remaining sample size
4303 * If we don't, we customize the stack size to
4304 * fit in to the remaining sample size.
4307 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4308 stack_size = min(stack_size, (u16) task_size);
4310 /* Current header size plus static size and dynamic size. */
4311 header_size += 2 * sizeof(u64);
4313 /* Do we fit in with the current stack dump size? */
4314 if ((u16) (header_size + stack_size) < header_size) {
4316 * If we overflow the maximum size for the sample,
4317 * we customize the stack dump size to fit in.
4319 stack_size = USHRT_MAX - header_size - sizeof(u64);
4320 stack_size = round_up(stack_size, sizeof(u64));
4327 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4328 struct pt_regs *regs)
4330 /* Case of a kernel thread, nothing to dump */
4333 perf_output_put(handle, size);
4342 * - the size requested by user or the best one we can fit
4343 * in to the sample max size
4345 * - user stack dump data
4347 * - the actual dumped size
4351 perf_output_put(handle, dump_size);
4354 sp = perf_user_stack_pointer(regs);
4355 rem = __output_copy_user(handle, (void *) sp, dump_size);
4356 dyn_size = dump_size - rem;
4358 perf_output_skip(handle, rem);
4361 perf_output_put(handle, dyn_size);
4365 static void __perf_event_header__init_id(struct perf_event_header *header,
4366 struct perf_sample_data *data,
4367 struct perf_event *event)
4369 u64 sample_type = event->attr.sample_type;
4371 data->type = sample_type;
4372 header->size += event->id_header_size;
4374 if (sample_type & PERF_SAMPLE_TID) {
4375 /* namespace issues */
4376 data->tid_entry.pid = perf_event_pid(event, current);
4377 data->tid_entry.tid = perf_event_tid(event, current);
4380 if (sample_type & PERF_SAMPLE_TIME)
4381 data->time = perf_clock();
4383 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4384 data->id = primary_event_id(event);
4386 if (sample_type & PERF_SAMPLE_STREAM_ID)
4387 data->stream_id = event->id;
4389 if (sample_type & PERF_SAMPLE_CPU) {
4390 data->cpu_entry.cpu = raw_smp_processor_id();
4391 data->cpu_entry.reserved = 0;
4395 void perf_event_header__init_id(struct perf_event_header *header,
4396 struct perf_sample_data *data,
4397 struct perf_event *event)
4399 if (event->attr.sample_id_all)
4400 __perf_event_header__init_id(header, data, event);
4403 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4404 struct perf_sample_data *data)
4406 u64 sample_type = data->type;
4408 if (sample_type & PERF_SAMPLE_TID)
4409 perf_output_put(handle, data->tid_entry);
4411 if (sample_type & PERF_SAMPLE_TIME)
4412 perf_output_put(handle, data->time);
4414 if (sample_type & PERF_SAMPLE_ID)
4415 perf_output_put(handle, data->id);
4417 if (sample_type & PERF_SAMPLE_STREAM_ID)
4418 perf_output_put(handle, data->stream_id);
4420 if (sample_type & PERF_SAMPLE_CPU)
4421 perf_output_put(handle, data->cpu_entry);
4423 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4424 perf_output_put(handle, data->id);
4427 void perf_event__output_id_sample(struct perf_event *event,
4428 struct perf_output_handle *handle,
4429 struct perf_sample_data *sample)
4431 if (event->attr.sample_id_all)
4432 __perf_event__output_id_sample(handle, sample);
4435 static void perf_output_read_one(struct perf_output_handle *handle,
4436 struct perf_event *event,
4437 u64 enabled, u64 running)
4439 u64 read_format = event->attr.read_format;
4443 values[n++] = perf_event_count(event);
4444 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4445 values[n++] = enabled +
4446 atomic64_read(&event->child_total_time_enabled);
4448 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4449 values[n++] = running +
4450 atomic64_read(&event->child_total_time_running);
4452 if (read_format & PERF_FORMAT_ID)
4453 values[n++] = primary_event_id(event);
4455 __output_copy(handle, values, n * sizeof(u64));
4459 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4461 static void perf_output_read_group(struct perf_output_handle *handle,
4462 struct perf_event *event,
4463 u64 enabled, u64 running)
4465 struct perf_event *leader = event->group_leader, *sub;
4466 u64 read_format = event->attr.read_format;
4470 values[n++] = 1 + leader->nr_siblings;
4472 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4473 values[n++] = enabled;
4475 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4476 values[n++] = running;
4478 if (leader != event)
4479 leader->pmu->read(leader);
4481 values[n++] = perf_event_count(leader);
4482 if (read_format & PERF_FORMAT_ID)
4483 values[n++] = primary_event_id(leader);
4485 __output_copy(handle, values, n * sizeof(u64));
4487 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4490 if ((sub != event) &&
4491 (sub->state == PERF_EVENT_STATE_ACTIVE))
4492 sub->pmu->read(sub);
4494 values[n++] = perf_event_count(sub);
4495 if (read_format & PERF_FORMAT_ID)
4496 values[n++] = primary_event_id(sub);
4498 __output_copy(handle, values, n * sizeof(u64));
4502 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4503 PERF_FORMAT_TOTAL_TIME_RUNNING)
4505 static void perf_output_read(struct perf_output_handle *handle,
4506 struct perf_event *event)
4508 u64 enabled = 0, running = 0, now;
4509 u64 read_format = event->attr.read_format;
4512 * compute total_time_enabled, total_time_running
4513 * based on snapshot values taken when the event
4514 * was last scheduled in.
4516 * we cannot simply called update_context_time()
4517 * because of locking issue as we are called in
4520 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4521 calc_timer_values(event, &now, &enabled, &running);
4523 if (event->attr.read_format & PERF_FORMAT_GROUP)
4524 perf_output_read_group(handle, event, enabled, running);
4526 perf_output_read_one(handle, event, enabled, running);
4529 void perf_output_sample(struct perf_output_handle *handle,
4530 struct perf_event_header *header,
4531 struct perf_sample_data *data,
4532 struct perf_event *event)
4534 u64 sample_type = data->type;
4536 perf_output_put(handle, *header);
4538 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4539 perf_output_put(handle, data->id);
4541 if (sample_type & PERF_SAMPLE_IP)
4542 perf_output_put(handle, data->ip);
4544 if (sample_type & PERF_SAMPLE_TID)
4545 perf_output_put(handle, data->tid_entry);
4547 if (sample_type & PERF_SAMPLE_TIME)
4548 perf_output_put(handle, data->time);
4550 if (sample_type & PERF_SAMPLE_ADDR)
4551 perf_output_put(handle, data->addr);
4553 if (sample_type & PERF_SAMPLE_ID)
4554 perf_output_put(handle, data->id);
4556 if (sample_type & PERF_SAMPLE_STREAM_ID)
4557 perf_output_put(handle, data->stream_id);
4559 if (sample_type & PERF_SAMPLE_CPU)
4560 perf_output_put(handle, data->cpu_entry);
4562 if (sample_type & PERF_SAMPLE_PERIOD)
4563 perf_output_put(handle, data->period);
4565 if (sample_type & PERF_SAMPLE_READ)
4566 perf_output_read(handle, event);
4568 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4569 if (data->callchain) {
4572 if (data->callchain)
4573 size += data->callchain->nr;
4575 size *= sizeof(u64);
4577 __output_copy(handle, data->callchain, size);
4580 perf_output_put(handle, nr);
4584 if (sample_type & PERF_SAMPLE_RAW) {
4586 perf_output_put(handle, data->raw->size);
4587 __output_copy(handle, data->raw->data,
4594 .size = sizeof(u32),
4597 perf_output_put(handle, raw);
4601 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4602 if (data->br_stack) {
4605 size = data->br_stack->nr
4606 * sizeof(struct perf_branch_entry);
4608 perf_output_put(handle, data->br_stack->nr);
4609 perf_output_copy(handle, data->br_stack->entries, size);
4612 * we always store at least the value of nr
4615 perf_output_put(handle, nr);
4619 if (sample_type & PERF_SAMPLE_REGS_USER) {
4620 u64 abi = data->regs_user.abi;
4623 * If there are no regs to dump, notice it through
4624 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4626 perf_output_put(handle, abi);
4629 u64 mask = event->attr.sample_regs_user;
4630 perf_output_sample_regs(handle,
4631 data->regs_user.regs,
4636 if (sample_type & PERF_SAMPLE_STACK_USER) {
4637 perf_output_sample_ustack(handle,
4638 data->stack_user_size,
4639 data->regs_user.regs);
4642 if (sample_type & PERF_SAMPLE_WEIGHT)
4643 perf_output_put(handle, data->weight);
4645 if (sample_type & PERF_SAMPLE_DATA_SRC)
4646 perf_output_put(handle, data->data_src.val);
4648 if (sample_type & PERF_SAMPLE_TRANSACTION)
4649 perf_output_put(handle, data->txn);
4651 if (!event->attr.watermark) {
4652 int wakeup_events = event->attr.wakeup_events;
4654 if (wakeup_events) {
4655 struct ring_buffer *rb = handle->rb;
4656 int events = local_inc_return(&rb->events);
4658 if (events >= wakeup_events) {
4659 local_sub(wakeup_events, &rb->events);
4660 local_inc(&rb->wakeup);
4666 void perf_prepare_sample(struct perf_event_header *header,
4667 struct perf_sample_data *data,
4668 struct perf_event *event,
4669 struct pt_regs *regs)
4671 u64 sample_type = event->attr.sample_type;
4673 header->type = PERF_RECORD_SAMPLE;
4674 header->size = sizeof(*header) + event->header_size;
4677 header->misc |= perf_misc_flags(regs);
4679 __perf_event_header__init_id(header, data, event);
4681 if (sample_type & PERF_SAMPLE_IP)
4682 data->ip = perf_instruction_pointer(regs);
4684 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4687 data->callchain = perf_callchain(event, regs);
4689 if (data->callchain)
4690 size += data->callchain->nr;
4692 header->size += size * sizeof(u64);
4695 if (sample_type & PERF_SAMPLE_RAW) {
4696 int size = sizeof(u32);
4699 size += data->raw->size;
4701 size += sizeof(u32);
4703 WARN_ON_ONCE(size & (sizeof(u64)-1));
4704 header->size += size;
4707 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4708 int size = sizeof(u64); /* nr */
4709 if (data->br_stack) {
4710 size += data->br_stack->nr
4711 * sizeof(struct perf_branch_entry);
4713 header->size += size;
4716 if (sample_type & PERF_SAMPLE_REGS_USER) {
4717 /* regs dump ABI info */
4718 int size = sizeof(u64);
4720 perf_sample_regs_user(&data->regs_user, regs);
4722 if (data->regs_user.regs) {
4723 u64 mask = event->attr.sample_regs_user;
4724 size += hweight64(mask) * sizeof(u64);
4727 header->size += size;
4730 if (sample_type & PERF_SAMPLE_STACK_USER) {
4732 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4733 * processed as the last one or have additional check added
4734 * in case new sample type is added, because we could eat
4735 * up the rest of the sample size.
4737 struct perf_regs_user *uregs = &data->regs_user;
4738 u16 stack_size = event->attr.sample_stack_user;
4739 u16 size = sizeof(u64);
4742 perf_sample_regs_user(uregs, regs);
4744 stack_size = perf_sample_ustack_size(stack_size, header->size,
4748 * If there is something to dump, add space for the dump
4749 * itself and for the field that tells the dynamic size,
4750 * which is how many have been actually dumped.
4753 size += sizeof(u64) + stack_size;
4755 data->stack_user_size = stack_size;
4756 header->size += size;
4760 static void perf_event_output(struct perf_event *event,
4761 struct perf_sample_data *data,
4762 struct pt_regs *regs)
4764 struct perf_output_handle handle;
4765 struct perf_event_header header;
4767 /* protect the callchain buffers */
4770 perf_prepare_sample(&header, data, event, regs);
4772 if (perf_output_begin(&handle, event, header.size))
4775 perf_output_sample(&handle, &header, data, event);
4777 perf_output_end(&handle);
4787 struct perf_read_event {
4788 struct perf_event_header header;
4795 perf_event_read_event(struct perf_event *event,
4796 struct task_struct *task)
4798 struct perf_output_handle handle;
4799 struct perf_sample_data sample;
4800 struct perf_read_event read_event = {
4802 .type = PERF_RECORD_READ,
4804 .size = sizeof(read_event) + event->read_size,
4806 .pid = perf_event_pid(event, task),
4807 .tid = perf_event_tid(event, task),
4811 perf_event_header__init_id(&read_event.header, &sample, event);
4812 ret = perf_output_begin(&handle, event, read_event.header.size);
4816 perf_output_put(&handle, read_event);
4817 perf_output_read(&handle, event);
4818 perf_event__output_id_sample(event, &handle, &sample);
4820 perf_output_end(&handle);
4823 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4826 perf_event_aux_ctx(struct perf_event_context *ctx,
4827 perf_event_aux_output_cb output,
4830 struct perf_event *event;
4832 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4833 if (event->state < PERF_EVENT_STATE_INACTIVE)
4835 if (!event_filter_match(event))
4837 output(event, data);
4842 perf_event_aux(perf_event_aux_output_cb output, void *data,
4843 struct perf_event_context *task_ctx)
4845 struct perf_cpu_context *cpuctx;
4846 struct perf_event_context *ctx;
4851 list_for_each_entry_rcu(pmu, &pmus, entry) {
4852 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4853 if (cpuctx->unique_pmu != pmu)
4855 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4858 ctxn = pmu->task_ctx_nr;
4861 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4863 perf_event_aux_ctx(ctx, output, data);
4865 put_cpu_ptr(pmu->pmu_cpu_context);
4870 perf_event_aux_ctx(task_ctx, output, data);
4877 * task tracking -- fork/exit
4879 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4882 struct perf_task_event {
4883 struct task_struct *task;
4884 struct perf_event_context *task_ctx;
4887 struct perf_event_header header;
4897 static int perf_event_task_match(struct perf_event *event)
4899 return event->attr.comm || event->attr.mmap ||
4900 event->attr.mmap2 || event->attr.mmap_data ||
4904 static void perf_event_task_output(struct perf_event *event,
4907 struct perf_task_event *task_event = data;
4908 struct perf_output_handle handle;
4909 struct perf_sample_data sample;
4910 struct task_struct *task = task_event->task;
4911 int ret, size = task_event->event_id.header.size;
4913 if (!perf_event_task_match(event))
4916 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4918 ret = perf_output_begin(&handle, event,
4919 task_event->event_id.header.size);
4923 task_event->event_id.pid = perf_event_pid(event, task);
4924 task_event->event_id.ppid = perf_event_pid(event, current);
4926 task_event->event_id.tid = perf_event_tid(event, task);
4927 task_event->event_id.ptid = perf_event_tid(event, current);
4929 perf_output_put(&handle, task_event->event_id);
4931 perf_event__output_id_sample(event, &handle, &sample);
4933 perf_output_end(&handle);
4935 task_event->event_id.header.size = size;
4938 static void perf_event_task(struct task_struct *task,
4939 struct perf_event_context *task_ctx,
4942 struct perf_task_event task_event;
4944 if (!atomic_read(&nr_comm_events) &&
4945 !atomic_read(&nr_mmap_events) &&
4946 !atomic_read(&nr_task_events))
4949 task_event = (struct perf_task_event){
4951 .task_ctx = task_ctx,
4954 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4956 .size = sizeof(task_event.event_id),
4962 .time = perf_clock(),
4966 perf_event_aux(perf_event_task_output,
4971 void perf_event_fork(struct task_struct *task)
4973 perf_event_task(task, NULL, 1);
4980 struct perf_comm_event {
4981 struct task_struct *task;
4986 struct perf_event_header header;
4993 static int perf_event_comm_match(struct perf_event *event)
4995 return event->attr.comm;
4998 static void perf_event_comm_output(struct perf_event *event,
5001 struct perf_comm_event *comm_event = data;
5002 struct perf_output_handle handle;
5003 struct perf_sample_data sample;
5004 int size = comm_event->event_id.header.size;
5007 if (!perf_event_comm_match(event))
5010 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5011 ret = perf_output_begin(&handle, event,
5012 comm_event->event_id.header.size);
5017 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5018 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5020 perf_output_put(&handle, comm_event->event_id);
5021 __output_copy(&handle, comm_event->comm,
5022 comm_event->comm_size);
5024 perf_event__output_id_sample(event, &handle, &sample);
5026 perf_output_end(&handle);
5028 comm_event->event_id.header.size = size;
5031 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5033 char comm[TASK_COMM_LEN];
5036 memset(comm, 0, sizeof(comm));
5037 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5038 size = ALIGN(strlen(comm)+1, sizeof(u64));
5040 comm_event->comm = comm;
5041 comm_event->comm_size = size;
5043 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5045 perf_event_aux(perf_event_comm_output,
5050 void perf_event_comm(struct task_struct *task)
5052 struct perf_comm_event comm_event;
5053 struct perf_event_context *ctx;
5057 for_each_task_context_nr(ctxn) {
5058 ctx = task->perf_event_ctxp[ctxn];
5062 perf_event_enable_on_exec(ctx);
5066 if (!atomic_read(&nr_comm_events))
5069 comm_event = (struct perf_comm_event){
5075 .type = PERF_RECORD_COMM,
5084 perf_event_comm_event(&comm_event);
5091 struct perf_mmap_event {
5092 struct vm_area_struct *vma;
5094 const char *file_name;
5101 struct perf_event_header header;
5111 static int perf_event_mmap_match(struct perf_event *event,
5114 struct perf_mmap_event *mmap_event = data;
5115 struct vm_area_struct *vma = mmap_event->vma;
5116 int executable = vma->vm_flags & VM_EXEC;
5118 return (!executable && event->attr.mmap_data) ||
5119 (executable && (event->attr.mmap || event->attr.mmap2));
5122 static void perf_event_mmap_output(struct perf_event *event,
5125 struct perf_mmap_event *mmap_event = data;
5126 struct perf_output_handle handle;
5127 struct perf_sample_data sample;
5128 int size = mmap_event->event_id.header.size;
5131 if (!perf_event_mmap_match(event, data))
5134 if (event->attr.mmap2) {
5135 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5136 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5137 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5138 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5139 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5142 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5143 ret = perf_output_begin(&handle, event,
5144 mmap_event->event_id.header.size);
5148 mmap_event->event_id.pid = perf_event_pid(event, current);
5149 mmap_event->event_id.tid = perf_event_tid(event, current);
5151 perf_output_put(&handle, mmap_event->event_id);
5153 if (event->attr.mmap2) {
5154 perf_output_put(&handle, mmap_event->maj);
5155 perf_output_put(&handle, mmap_event->min);
5156 perf_output_put(&handle, mmap_event->ino);
5157 perf_output_put(&handle, mmap_event->ino_generation);
5160 __output_copy(&handle, mmap_event->file_name,
5161 mmap_event->file_size);
5163 perf_event__output_id_sample(event, &handle, &sample);
5165 perf_output_end(&handle);
5167 mmap_event->event_id.header.size = size;
5170 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5172 struct vm_area_struct *vma = mmap_event->vma;
5173 struct file *file = vma->vm_file;
5174 int maj = 0, min = 0;
5175 u64 ino = 0, gen = 0;
5182 struct inode *inode;
5185 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5191 * d_path() works from the end of the rb backwards, so we
5192 * need to add enough zero bytes after the string to handle
5193 * the 64bit alignment we do later.
5195 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5200 inode = file_inode(vma->vm_file);
5201 dev = inode->i_sb->s_dev;
5203 gen = inode->i_generation;
5208 name = (char *)arch_vma_name(vma);
5212 if (vma->vm_start <= vma->vm_mm->start_brk &&
5213 vma->vm_end >= vma->vm_mm->brk) {
5217 if (vma->vm_start <= vma->vm_mm->start_stack &&
5218 vma->vm_end >= vma->vm_mm->start_stack) {
5228 strlcpy(tmp, name, sizeof(tmp));
5232 * Since our buffer works in 8 byte units we need to align our string
5233 * size to a multiple of 8. However, we must guarantee the tail end is
5234 * zero'd out to avoid leaking random bits to userspace.
5236 size = strlen(name)+1;
5237 while (!IS_ALIGNED(size, sizeof(u64)))
5238 name[size++] = '\0';
5240 mmap_event->file_name = name;
5241 mmap_event->file_size = size;
5242 mmap_event->maj = maj;
5243 mmap_event->min = min;
5244 mmap_event->ino = ino;
5245 mmap_event->ino_generation = gen;
5247 if (!(vma->vm_flags & VM_EXEC))
5248 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5250 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5252 perf_event_aux(perf_event_mmap_output,
5259 void perf_event_mmap(struct vm_area_struct *vma)
5261 struct perf_mmap_event mmap_event;
5263 if (!atomic_read(&nr_mmap_events))
5266 mmap_event = (struct perf_mmap_event){
5272 .type = PERF_RECORD_MMAP,
5273 .misc = PERF_RECORD_MISC_USER,
5278 .start = vma->vm_start,
5279 .len = vma->vm_end - vma->vm_start,
5280 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5282 /* .maj (attr_mmap2 only) */
5283 /* .min (attr_mmap2 only) */
5284 /* .ino (attr_mmap2 only) */
5285 /* .ino_generation (attr_mmap2 only) */
5288 perf_event_mmap_event(&mmap_event);
5292 * IRQ throttle logging
5295 static void perf_log_throttle(struct perf_event *event, int enable)
5297 struct perf_output_handle handle;
5298 struct perf_sample_data sample;
5302 struct perf_event_header header;
5306 } throttle_event = {
5308 .type = PERF_RECORD_THROTTLE,
5310 .size = sizeof(throttle_event),
5312 .time = perf_clock(),
5313 .id = primary_event_id(event),
5314 .stream_id = event->id,
5318 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5320 perf_event_header__init_id(&throttle_event.header, &sample, event);
5322 ret = perf_output_begin(&handle, event,
5323 throttle_event.header.size);
5327 perf_output_put(&handle, throttle_event);
5328 perf_event__output_id_sample(event, &handle, &sample);
5329 perf_output_end(&handle);
5333 * Generic event overflow handling, sampling.
5336 static int __perf_event_overflow(struct perf_event *event,
5337 int throttle, struct perf_sample_data *data,
5338 struct pt_regs *regs)
5340 int events = atomic_read(&event->event_limit);
5341 struct hw_perf_event *hwc = &event->hw;
5346 * Non-sampling counters might still use the PMI to fold short
5347 * hardware counters, ignore those.
5349 if (unlikely(!is_sampling_event(event)))
5352 seq = __this_cpu_read(perf_throttled_seq);
5353 if (seq != hwc->interrupts_seq) {
5354 hwc->interrupts_seq = seq;
5355 hwc->interrupts = 1;
5358 if (unlikely(throttle
5359 && hwc->interrupts >= max_samples_per_tick)) {
5360 __this_cpu_inc(perf_throttled_count);
5361 hwc->interrupts = MAX_INTERRUPTS;
5362 perf_log_throttle(event, 0);
5363 tick_nohz_full_kick();
5368 if (event->attr.freq) {
5369 u64 now = perf_clock();
5370 s64 delta = now - hwc->freq_time_stamp;
5372 hwc->freq_time_stamp = now;
5374 if (delta > 0 && delta < 2*TICK_NSEC)
5375 perf_adjust_period(event, delta, hwc->last_period, true);
5379 * XXX event_limit might not quite work as expected on inherited
5383 event->pending_kill = POLL_IN;
5384 if (events && atomic_dec_and_test(&event->event_limit)) {
5386 event->pending_kill = POLL_HUP;
5387 event->pending_disable = 1;
5388 irq_work_queue(&event->pending);
5391 if (event->overflow_handler)
5392 event->overflow_handler(event, data, regs);
5394 perf_event_output(event, data, regs);
5396 if (event->fasync && event->pending_kill) {
5397 event->pending_wakeup = 1;
5398 irq_work_queue(&event->pending);
5404 int perf_event_overflow(struct perf_event *event,
5405 struct perf_sample_data *data,
5406 struct pt_regs *regs)
5408 return __perf_event_overflow(event, 1, data, regs);
5412 * Generic software event infrastructure
5415 struct swevent_htable {
5416 struct swevent_hlist *swevent_hlist;
5417 struct mutex hlist_mutex;
5420 /* Recursion avoidance in each contexts */
5421 int recursion[PERF_NR_CONTEXTS];
5424 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5427 * We directly increment event->count and keep a second value in
5428 * event->hw.period_left to count intervals. This period event
5429 * is kept in the range [-sample_period, 0] so that we can use the
5433 u64 perf_swevent_set_period(struct perf_event *event)
5435 struct hw_perf_event *hwc = &event->hw;
5436 u64 period = hwc->last_period;
5440 hwc->last_period = hwc->sample_period;
5443 old = val = local64_read(&hwc->period_left);
5447 nr = div64_u64(period + val, period);
5448 offset = nr * period;
5450 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5456 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5457 struct perf_sample_data *data,
5458 struct pt_regs *regs)
5460 struct hw_perf_event *hwc = &event->hw;
5464 overflow = perf_swevent_set_period(event);
5466 if (hwc->interrupts == MAX_INTERRUPTS)
5469 for (; overflow; overflow--) {
5470 if (__perf_event_overflow(event, throttle,
5473 * We inhibit the overflow from happening when
5474 * hwc->interrupts == MAX_INTERRUPTS.
5482 static void perf_swevent_event(struct perf_event *event, u64 nr,
5483 struct perf_sample_data *data,
5484 struct pt_regs *regs)
5486 struct hw_perf_event *hwc = &event->hw;
5488 local64_add(nr, &event->count);
5493 if (!is_sampling_event(event))
5496 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5498 return perf_swevent_overflow(event, 1, data, regs);
5500 data->period = event->hw.last_period;
5502 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5503 return perf_swevent_overflow(event, 1, data, regs);
5505 if (local64_add_negative(nr, &hwc->period_left))
5508 perf_swevent_overflow(event, 0, data, regs);
5511 static int perf_exclude_event(struct perf_event *event,
5512 struct pt_regs *regs)
5514 if (event->hw.state & PERF_HES_STOPPED)
5518 if (event->attr.exclude_user && user_mode(regs))
5521 if (event->attr.exclude_kernel && !user_mode(regs))
5528 static int perf_swevent_match(struct perf_event *event,
5529 enum perf_type_id type,
5531 struct perf_sample_data *data,
5532 struct pt_regs *regs)
5534 if (event->attr.type != type)
5537 if (event->attr.config != event_id)
5540 if (perf_exclude_event(event, regs))
5546 static inline u64 swevent_hash(u64 type, u32 event_id)
5548 u64 val = event_id | (type << 32);
5550 return hash_64(val, SWEVENT_HLIST_BITS);
5553 static inline struct hlist_head *
5554 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5556 u64 hash = swevent_hash(type, event_id);
5558 return &hlist->heads[hash];
5561 /* For the read side: events when they trigger */
5562 static inline struct hlist_head *
5563 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5565 struct swevent_hlist *hlist;
5567 hlist = rcu_dereference(swhash->swevent_hlist);
5571 return __find_swevent_head(hlist, type, event_id);
5574 /* For the event head insertion and removal in the hlist */
5575 static inline struct hlist_head *
5576 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5578 struct swevent_hlist *hlist;
5579 u32 event_id = event->attr.config;
5580 u64 type = event->attr.type;
5583 * Event scheduling is always serialized against hlist allocation
5584 * and release. Which makes the protected version suitable here.
5585 * The context lock guarantees that.
5587 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5588 lockdep_is_held(&event->ctx->lock));
5592 return __find_swevent_head(hlist, type, event_id);
5595 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5597 struct perf_sample_data *data,
5598 struct pt_regs *regs)
5600 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5601 struct perf_event *event;
5602 struct hlist_head *head;
5605 head = find_swevent_head_rcu(swhash, type, event_id);
5609 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5610 if (perf_swevent_match(event, type, event_id, data, regs))
5611 perf_swevent_event(event, nr, data, regs);
5617 int perf_swevent_get_recursion_context(void)
5619 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5621 return get_recursion_context(swhash->recursion);
5623 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5625 inline void perf_swevent_put_recursion_context(int rctx)
5627 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5629 put_recursion_context(swhash->recursion, rctx);
5632 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5634 struct perf_sample_data data;
5637 preempt_disable_notrace();
5638 rctx = perf_swevent_get_recursion_context();
5642 perf_sample_data_init(&data, addr, 0);
5644 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5646 perf_swevent_put_recursion_context(rctx);
5647 preempt_enable_notrace();
5650 static void perf_swevent_read(struct perf_event *event)
5654 static int perf_swevent_add(struct perf_event *event, int flags)
5656 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5657 struct hw_perf_event *hwc = &event->hw;
5658 struct hlist_head *head;
5660 if (is_sampling_event(event)) {
5661 hwc->last_period = hwc->sample_period;
5662 perf_swevent_set_period(event);
5665 hwc->state = !(flags & PERF_EF_START);
5667 head = find_swevent_head(swhash, event);
5668 if (WARN_ON_ONCE(!head))
5671 hlist_add_head_rcu(&event->hlist_entry, head);
5676 static void perf_swevent_del(struct perf_event *event, int flags)
5678 hlist_del_rcu(&event->hlist_entry);
5681 static void perf_swevent_start(struct perf_event *event, int flags)
5683 event->hw.state = 0;
5686 static void perf_swevent_stop(struct perf_event *event, int flags)
5688 event->hw.state = PERF_HES_STOPPED;
5691 /* Deref the hlist from the update side */
5692 static inline struct swevent_hlist *
5693 swevent_hlist_deref(struct swevent_htable *swhash)
5695 return rcu_dereference_protected(swhash->swevent_hlist,
5696 lockdep_is_held(&swhash->hlist_mutex));
5699 static void swevent_hlist_release(struct swevent_htable *swhash)
5701 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5706 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5707 kfree_rcu(hlist, rcu_head);
5710 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5712 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5714 mutex_lock(&swhash->hlist_mutex);
5716 if (!--swhash->hlist_refcount)
5717 swevent_hlist_release(swhash);
5719 mutex_unlock(&swhash->hlist_mutex);
5722 static void swevent_hlist_put(struct perf_event *event)
5726 for_each_possible_cpu(cpu)
5727 swevent_hlist_put_cpu(event, cpu);
5730 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5732 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5735 mutex_lock(&swhash->hlist_mutex);
5737 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5738 struct swevent_hlist *hlist;
5740 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5745 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5747 swhash->hlist_refcount++;
5749 mutex_unlock(&swhash->hlist_mutex);
5754 static int swevent_hlist_get(struct perf_event *event)
5757 int cpu, failed_cpu;
5760 for_each_possible_cpu(cpu) {
5761 err = swevent_hlist_get_cpu(event, cpu);
5771 for_each_possible_cpu(cpu) {
5772 if (cpu == failed_cpu)
5774 swevent_hlist_put_cpu(event, cpu);
5781 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5783 static void sw_perf_event_destroy(struct perf_event *event)
5785 u64 event_id = event->attr.config;
5787 WARN_ON(event->parent);
5789 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5790 swevent_hlist_put(event);
5793 static int perf_swevent_init(struct perf_event *event)
5795 u64 event_id = event->attr.config;
5797 if (event->attr.type != PERF_TYPE_SOFTWARE)
5801 * no branch sampling for software events
5803 if (has_branch_stack(event))
5807 case PERF_COUNT_SW_CPU_CLOCK:
5808 case PERF_COUNT_SW_TASK_CLOCK:
5815 if (event_id >= PERF_COUNT_SW_MAX)
5818 if (!event->parent) {
5821 err = swevent_hlist_get(event);
5825 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5826 event->destroy = sw_perf_event_destroy;
5832 static int perf_swevent_event_idx(struct perf_event *event)
5837 static struct pmu perf_swevent = {
5838 .task_ctx_nr = perf_sw_context,
5840 .event_init = perf_swevent_init,
5841 .add = perf_swevent_add,
5842 .del = perf_swevent_del,
5843 .start = perf_swevent_start,
5844 .stop = perf_swevent_stop,
5845 .read = perf_swevent_read,
5847 .event_idx = perf_swevent_event_idx,
5850 #ifdef CONFIG_EVENT_TRACING
5852 static int perf_tp_filter_match(struct perf_event *event,
5853 struct perf_sample_data *data)
5855 void *record = data->raw->data;
5857 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5862 static int perf_tp_event_match(struct perf_event *event,
5863 struct perf_sample_data *data,
5864 struct pt_regs *regs)
5866 if (event->hw.state & PERF_HES_STOPPED)
5869 * All tracepoints are from kernel-space.
5871 if (event->attr.exclude_kernel)
5874 if (!perf_tp_filter_match(event, data))
5880 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5881 struct pt_regs *regs, struct hlist_head *head, int rctx,
5882 struct task_struct *task)
5884 struct perf_sample_data data;
5885 struct perf_event *event;
5887 struct perf_raw_record raw = {
5892 perf_sample_data_init(&data, addr, 0);
5895 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5896 if (perf_tp_event_match(event, &data, regs))
5897 perf_swevent_event(event, count, &data, regs);
5901 * If we got specified a target task, also iterate its context and
5902 * deliver this event there too.
5904 if (task && task != current) {
5905 struct perf_event_context *ctx;
5906 struct trace_entry *entry = record;
5909 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5913 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5914 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5916 if (event->attr.config != entry->type)
5918 if (perf_tp_event_match(event, &data, regs))
5919 perf_swevent_event(event, count, &data, regs);
5925 perf_swevent_put_recursion_context(rctx);
5927 EXPORT_SYMBOL_GPL(perf_tp_event);
5929 static void tp_perf_event_destroy(struct perf_event *event)
5931 perf_trace_destroy(event);
5934 static int perf_tp_event_init(struct perf_event *event)
5938 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5942 * no branch sampling for tracepoint events
5944 if (has_branch_stack(event))
5947 err = perf_trace_init(event);
5951 event->destroy = tp_perf_event_destroy;
5956 static struct pmu perf_tracepoint = {
5957 .task_ctx_nr = perf_sw_context,
5959 .event_init = perf_tp_event_init,
5960 .add = perf_trace_add,
5961 .del = perf_trace_del,
5962 .start = perf_swevent_start,
5963 .stop = perf_swevent_stop,
5964 .read = perf_swevent_read,
5966 .event_idx = perf_swevent_event_idx,
5969 static inline void perf_tp_register(void)
5971 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5974 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5979 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5982 filter_str = strndup_user(arg, PAGE_SIZE);
5983 if (IS_ERR(filter_str))
5984 return PTR_ERR(filter_str);
5986 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5992 static void perf_event_free_filter(struct perf_event *event)
5994 ftrace_profile_free_filter(event);
5999 static inline void perf_tp_register(void)
6003 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6008 static void perf_event_free_filter(struct perf_event *event)
6012 #endif /* CONFIG_EVENT_TRACING */
6014 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6015 void perf_bp_event(struct perf_event *bp, void *data)
6017 struct perf_sample_data sample;
6018 struct pt_regs *regs = data;
6020 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6022 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6023 perf_swevent_event(bp, 1, &sample, regs);
6028 * hrtimer based swevent callback
6031 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6033 enum hrtimer_restart ret = HRTIMER_RESTART;
6034 struct perf_sample_data data;
6035 struct pt_regs *regs;
6036 struct perf_event *event;
6039 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6041 if (event->state != PERF_EVENT_STATE_ACTIVE)
6042 return HRTIMER_NORESTART;
6044 event->pmu->read(event);
6046 perf_sample_data_init(&data, 0, event->hw.last_period);
6047 regs = get_irq_regs();
6049 if (regs && !perf_exclude_event(event, regs)) {
6050 if (!(event->attr.exclude_idle && is_idle_task(current)))
6051 if (__perf_event_overflow(event, 1, &data, regs))
6052 ret = HRTIMER_NORESTART;
6055 period = max_t(u64, 10000, event->hw.sample_period);
6056 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6061 static void perf_swevent_start_hrtimer(struct perf_event *event)
6063 struct hw_perf_event *hwc = &event->hw;
6066 if (!is_sampling_event(event))
6069 period = local64_read(&hwc->period_left);
6074 local64_set(&hwc->period_left, 0);
6076 period = max_t(u64, 10000, hwc->sample_period);
6078 __hrtimer_start_range_ns(&hwc->hrtimer,
6079 ns_to_ktime(period), 0,
6080 HRTIMER_MODE_REL_PINNED, 0);
6083 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6085 struct hw_perf_event *hwc = &event->hw;
6087 if (is_sampling_event(event)) {
6088 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6089 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6091 hrtimer_cancel(&hwc->hrtimer);
6095 static void perf_swevent_init_hrtimer(struct perf_event *event)
6097 struct hw_perf_event *hwc = &event->hw;
6099 if (!is_sampling_event(event))
6102 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6103 hwc->hrtimer.function = perf_swevent_hrtimer;
6106 * Since hrtimers have a fixed rate, we can do a static freq->period
6107 * mapping and avoid the whole period adjust feedback stuff.
6109 if (event->attr.freq) {
6110 long freq = event->attr.sample_freq;
6112 event->attr.sample_period = NSEC_PER_SEC / freq;
6113 hwc->sample_period = event->attr.sample_period;
6114 local64_set(&hwc->period_left, hwc->sample_period);
6115 hwc->last_period = hwc->sample_period;
6116 event->attr.freq = 0;
6121 * Software event: cpu wall time clock
6124 static void cpu_clock_event_update(struct perf_event *event)
6129 now = local_clock();
6130 prev = local64_xchg(&event->hw.prev_count, now);
6131 local64_add(now - prev, &event->count);
6134 static void cpu_clock_event_start(struct perf_event *event, int flags)
6136 local64_set(&event->hw.prev_count, local_clock());
6137 perf_swevent_start_hrtimer(event);
6140 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6142 perf_swevent_cancel_hrtimer(event);
6143 cpu_clock_event_update(event);
6146 static int cpu_clock_event_add(struct perf_event *event, int flags)
6148 if (flags & PERF_EF_START)
6149 cpu_clock_event_start(event, flags);
6154 static void cpu_clock_event_del(struct perf_event *event, int flags)
6156 cpu_clock_event_stop(event, flags);
6159 static void cpu_clock_event_read(struct perf_event *event)
6161 cpu_clock_event_update(event);
6164 static int cpu_clock_event_init(struct perf_event *event)
6166 if (event->attr.type != PERF_TYPE_SOFTWARE)
6169 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6173 * no branch sampling for software events
6175 if (has_branch_stack(event))
6178 perf_swevent_init_hrtimer(event);
6183 static struct pmu perf_cpu_clock = {
6184 .task_ctx_nr = perf_sw_context,
6186 .event_init = cpu_clock_event_init,
6187 .add = cpu_clock_event_add,
6188 .del = cpu_clock_event_del,
6189 .start = cpu_clock_event_start,
6190 .stop = cpu_clock_event_stop,
6191 .read = cpu_clock_event_read,
6193 .event_idx = perf_swevent_event_idx,
6197 * Software event: task time clock
6200 static void task_clock_event_update(struct perf_event *event, u64 now)
6205 prev = local64_xchg(&event->hw.prev_count, now);
6207 local64_add(delta, &event->count);
6210 static void task_clock_event_start(struct perf_event *event, int flags)
6212 local64_set(&event->hw.prev_count, event->ctx->time);
6213 perf_swevent_start_hrtimer(event);
6216 static void task_clock_event_stop(struct perf_event *event, int flags)
6218 perf_swevent_cancel_hrtimer(event);
6219 task_clock_event_update(event, event->ctx->time);
6222 static int task_clock_event_add(struct perf_event *event, int flags)
6224 if (flags & PERF_EF_START)
6225 task_clock_event_start(event, flags);
6230 static void task_clock_event_del(struct perf_event *event, int flags)
6232 task_clock_event_stop(event, PERF_EF_UPDATE);
6235 static void task_clock_event_read(struct perf_event *event)
6237 u64 now = perf_clock();
6238 u64 delta = now - event->ctx->timestamp;
6239 u64 time = event->ctx->time + delta;
6241 task_clock_event_update(event, time);
6244 static int task_clock_event_init(struct perf_event *event)
6246 if (event->attr.type != PERF_TYPE_SOFTWARE)
6249 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6253 * no branch sampling for software events
6255 if (has_branch_stack(event))
6258 perf_swevent_init_hrtimer(event);
6263 static struct pmu perf_task_clock = {
6264 .task_ctx_nr = perf_sw_context,
6266 .event_init = task_clock_event_init,
6267 .add = task_clock_event_add,
6268 .del = task_clock_event_del,
6269 .start = task_clock_event_start,
6270 .stop = task_clock_event_stop,
6271 .read = task_clock_event_read,
6273 .event_idx = perf_swevent_event_idx,
6276 static void perf_pmu_nop_void(struct pmu *pmu)
6280 static int perf_pmu_nop_int(struct pmu *pmu)
6285 static void perf_pmu_start_txn(struct pmu *pmu)
6287 perf_pmu_disable(pmu);
6290 static int perf_pmu_commit_txn(struct pmu *pmu)
6292 perf_pmu_enable(pmu);
6296 static void perf_pmu_cancel_txn(struct pmu *pmu)
6298 perf_pmu_enable(pmu);
6301 static int perf_event_idx_default(struct perf_event *event)
6303 return event->hw.idx + 1;
6307 * Ensures all contexts with the same task_ctx_nr have the same
6308 * pmu_cpu_context too.
6310 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6317 list_for_each_entry(pmu, &pmus, entry) {
6318 if (pmu->task_ctx_nr == ctxn)
6319 return pmu->pmu_cpu_context;
6325 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6329 for_each_possible_cpu(cpu) {
6330 struct perf_cpu_context *cpuctx;
6332 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6334 if (cpuctx->unique_pmu == old_pmu)
6335 cpuctx->unique_pmu = pmu;
6339 static void free_pmu_context(struct pmu *pmu)
6343 mutex_lock(&pmus_lock);
6345 * Like a real lame refcount.
6347 list_for_each_entry(i, &pmus, entry) {
6348 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6349 update_pmu_context(i, pmu);
6354 free_percpu(pmu->pmu_cpu_context);
6356 mutex_unlock(&pmus_lock);
6358 static struct idr pmu_idr;
6361 type_show(struct device *dev, struct device_attribute *attr, char *page)
6363 struct pmu *pmu = dev_get_drvdata(dev);
6365 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6367 static DEVICE_ATTR_RO(type);
6370 perf_event_mux_interval_ms_show(struct device *dev,
6371 struct device_attribute *attr,
6374 struct pmu *pmu = dev_get_drvdata(dev);
6376 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6380 perf_event_mux_interval_ms_store(struct device *dev,
6381 struct device_attribute *attr,
6382 const char *buf, size_t count)
6384 struct pmu *pmu = dev_get_drvdata(dev);
6385 int timer, cpu, ret;
6387 ret = kstrtoint(buf, 0, &timer);
6394 /* same value, noting to do */
6395 if (timer == pmu->hrtimer_interval_ms)
6398 pmu->hrtimer_interval_ms = timer;
6400 /* update all cpuctx for this PMU */
6401 for_each_possible_cpu(cpu) {
6402 struct perf_cpu_context *cpuctx;
6403 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6404 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6406 if (hrtimer_active(&cpuctx->hrtimer))
6407 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6412 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6414 static struct attribute *pmu_dev_attrs[] = {
6415 &dev_attr_type.attr,
6416 &dev_attr_perf_event_mux_interval_ms.attr,
6419 ATTRIBUTE_GROUPS(pmu_dev);
6421 static int pmu_bus_running;
6422 static struct bus_type pmu_bus = {
6423 .name = "event_source",
6424 .dev_groups = pmu_dev_groups,
6427 static void pmu_dev_release(struct device *dev)
6432 static int pmu_dev_alloc(struct pmu *pmu)
6436 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6440 pmu->dev->groups = pmu->attr_groups;
6441 device_initialize(pmu->dev);
6442 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6446 dev_set_drvdata(pmu->dev, pmu);
6447 pmu->dev->bus = &pmu_bus;
6448 pmu->dev->release = pmu_dev_release;
6449 ret = device_add(pmu->dev);
6457 put_device(pmu->dev);
6461 static struct lock_class_key cpuctx_mutex;
6462 static struct lock_class_key cpuctx_lock;
6464 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6468 mutex_lock(&pmus_lock);
6470 pmu->pmu_disable_count = alloc_percpu(int);
6471 if (!pmu->pmu_disable_count)
6480 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6488 if (pmu_bus_running) {
6489 ret = pmu_dev_alloc(pmu);
6495 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6496 if (pmu->pmu_cpu_context)
6497 goto got_cpu_context;
6500 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6501 if (!pmu->pmu_cpu_context)
6504 for_each_possible_cpu(cpu) {
6505 struct perf_cpu_context *cpuctx;
6507 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6508 __perf_event_init_context(&cpuctx->ctx);
6509 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6510 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6511 cpuctx->ctx.type = cpu_context;
6512 cpuctx->ctx.pmu = pmu;
6514 __perf_cpu_hrtimer_init(cpuctx, cpu);
6516 INIT_LIST_HEAD(&cpuctx->rotation_list);
6517 cpuctx->unique_pmu = pmu;
6521 if (!pmu->start_txn) {
6522 if (pmu->pmu_enable) {
6524 * If we have pmu_enable/pmu_disable calls, install
6525 * transaction stubs that use that to try and batch
6526 * hardware accesses.
6528 pmu->start_txn = perf_pmu_start_txn;
6529 pmu->commit_txn = perf_pmu_commit_txn;
6530 pmu->cancel_txn = perf_pmu_cancel_txn;
6532 pmu->start_txn = perf_pmu_nop_void;
6533 pmu->commit_txn = perf_pmu_nop_int;
6534 pmu->cancel_txn = perf_pmu_nop_void;
6538 if (!pmu->pmu_enable) {
6539 pmu->pmu_enable = perf_pmu_nop_void;
6540 pmu->pmu_disable = perf_pmu_nop_void;
6543 if (!pmu->event_idx)
6544 pmu->event_idx = perf_event_idx_default;
6546 list_add_rcu(&pmu->entry, &pmus);
6549 mutex_unlock(&pmus_lock);
6554 device_del(pmu->dev);
6555 put_device(pmu->dev);
6558 if (pmu->type >= PERF_TYPE_MAX)
6559 idr_remove(&pmu_idr, pmu->type);
6562 free_percpu(pmu->pmu_disable_count);
6566 void perf_pmu_unregister(struct pmu *pmu)
6568 mutex_lock(&pmus_lock);
6569 list_del_rcu(&pmu->entry);
6570 mutex_unlock(&pmus_lock);
6573 * We dereference the pmu list under both SRCU and regular RCU, so
6574 * synchronize against both of those.
6576 synchronize_srcu(&pmus_srcu);
6579 free_percpu(pmu->pmu_disable_count);
6580 if (pmu->type >= PERF_TYPE_MAX)
6581 idr_remove(&pmu_idr, pmu->type);
6582 device_del(pmu->dev);
6583 put_device(pmu->dev);
6584 free_pmu_context(pmu);
6587 struct pmu *perf_init_event(struct perf_event *event)
6589 struct pmu *pmu = NULL;
6593 idx = srcu_read_lock(&pmus_srcu);
6596 pmu = idr_find(&pmu_idr, event->attr.type);
6600 ret = pmu->event_init(event);
6606 list_for_each_entry_rcu(pmu, &pmus, entry) {
6608 ret = pmu->event_init(event);
6612 if (ret != -ENOENT) {
6617 pmu = ERR_PTR(-ENOENT);
6619 srcu_read_unlock(&pmus_srcu, idx);
6624 static void account_event_cpu(struct perf_event *event, int cpu)
6629 if (has_branch_stack(event)) {
6630 if (!(event->attach_state & PERF_ATTACH_TASK))
6631 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6633 if (is_cgroup_event(event))
6634 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6637 static void account_event(struct perf_event *event)
6642 if (event->attach_state & PERF_ATTACH_TASK)
6643 static_key_slow_inc(&perf_sched_events.key);
6644 if (event->attr.mmap || event->attr.mmap_data)
6645 atomic_inc(&nr_mmap_events);
6646 if (event->attr.comm)
6647 atomic_inc(&nr_comm_events);
6648 if (event->attr.task)
6649 atomic_inc(&nr_task_events);
6650 if (event->attr.freq) {
6651 if (atomic_inc_return(&nr_freq_events) == 1)
6652 tick_nohz_full_kick_all();
6654 if (has_branch_stack(event))
6655 static_key_slow_inc(&perf_sched_events.key);
6656 if (is_cgroup_event(event))
6657 static_key_slow_inc(&perf_sched_events.key);
6659 account_event_cpu(event, event->cpu);
6663 * Allocate and initialize a event structure
6665 static struct perf_event *
6666 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6667 struct task_struct *task,
6668 struct perf_event *group_leader,
6669 struct perf_event *parent_event,
6670 perf_overflow_handler_t overflow_handler,
6674 struct perf_event *event;
6675 struct hw_perf_event *hwc;
6678 if ((unsigned)cpu >= nr_cpu_ids) {
6679 if (!task || cpu != -1)
6680 return ERR_PTR(-EINVAL);
6683 event = kzalloc(sizeof(*event), GFP_KERNEL);
6685 return ERR_PTR(-ENOMEM);
6688 * Single events are their own group leaders, with an
6689 * empty sibling list:
6692 group_leader = event;
6694 mutex_init(&event->child_mutex);
6695 INIT_LIST_HEAD(&event->child_list);
6697 INIT_LIST_HEAD(&event->group_entry);
6698 INIT_LIST_HEAD(&event->event_entry);
6699 INIT_LIST_HEAD(&event->sibling_list);
6700 INIT_LIST_HEAD(&event->rb_entry);
6701 INIT_LIST_HEAD(&event->active_entry);
6702 INIT_HLIST_NODE(&event->hlist_entry);
6705 init_waitqueue_head(&event->waitq);
6706 init_irq_work(&event->pending, perf_pending_event);
6708 mutex_init(&event->mmap_mutex);
6710 atomic_long_set(&event->refcount, 1);
6712 event->attr = *attr;
6713 event->group_leader = group_leader;
6717 event->parent = parent_event;
6719 event->ns = get_pid_ns(task_active_pid_ns(current));
6720 event->id = atomic64_inc_return(&perf_event_id);
6722 event->state = PERF_EVENT_STATE_INACTIVE;
6725 event->attach_state = PERF_ATTACH_TASK;
6727 if (attr->type == PERF_TYPE_TRACEPOINT)
6728 event->hw.tp_target = task;
6729 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6731 * hw_breakpoint is a bit difficult here..
6733 else if (attr->type == PERF_TYPE_BREAKPOINT)
6734 event->hw.bp_target = task;
6738 if (!overflow_handler && parent_event) {
6739 overflow_handler = parent_event->overflow_handler;
6740 context = parent_event->overflow_handler_context;
6743 event->overflow_handler = overflow_handler;
6744 event->overflow_handler_context = context;
6746 perf_event__state_init(event);
6751 hwc->sample_period = attr->sample_period;
6752 if (attr->freq && attr->sample_freq)
6753 hwc->sample_period = 1;
6754 hwc->last_period = hwc->sample_period;
6756 local64_set(&hwc->period_left, hwc->sample_period);
6759 * we currently do not support PERF_FORMAT_GROUP on inherited events
6761 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6764 pmu = perf_init_event(event);
6767 else if (IS_ERR(pmu)) {
6772 if (!event->parent) {
6773 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6774 err = get_callchain_buffers();
6784 event->destroy(event);
6787 put_pid_ns(event->ns);
6790 return ERR_PTR(err);
6793 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6794 struct perf_event_attr *attr)
6799 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6803 * zero the full structure, so that a short copy will be nice.
6805 memset(attr, 0, sizeof(*attr));
6807 ret = get_user(size, &uattr->size);
6811 if (size > PAGE_SIZE) /* silly large */
6814 if (!size) /* abi compat */
6815 size = PERF_ATTR_SIZE_VER0;
6817 if (size < PERF_ATTR_SIZE_VER0)
6821 * If we're handed a bigger struct than we know of,
6822 * ensure all the unknown bits are 0 - i.e. new
6823 * user-space does not rely on any kernel feature
6824 * extensions we dont know about yet.
6826 if (size > sizeof(*attr)) {
6827 unsigned char __user *addr;
6828 unsigned char __user *end;
6831 addr = (void __user *)uattr + sizeof(*attr);
6832 end = (void __user *)uattr + size;
6834 for (; addr < end; addr++) {
6835 ret = get_user(val, addr);
6841 size = sizeof(*attr);
6844 ret = copy_from_user(attr, uattr, size);
6848 /* disabled for now */
6852 if (attr->__reserved_1)
6855 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6858 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6861 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6862 u64 mask = attr->branch_sample_type;
6864 /* only using defined bits */
6865 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6868 /* at least one branch bit must be set */
6869 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6872 /* propagate priv level, when not set for branch */
6873 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6875 /* exclude_kernel checked on syscall entry */
6876 if (!attr->exclude_kernel)
6877 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6879 if (!attr->exclude_user)
6880 mask |= PERF_SAMPLE_BRANCH_USER;
6882 if (!attr->exclude_hv)
6883 mask |= PERF_SAMPLE_BRANCH_HV;
6885 * adjust user setting (for HW filter setup)
6887 attr->branch_sample_type = mask;
6889 /* privileged levels capture (kernel, hv): check permissions */
6890 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6891 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6895 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6896 ret = perf_reg_validate(attr->sample_regs_user);
6901 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6902 if (!arch_perf_have_user_stack_dump())
6906 * We have __u32 type for the size, but so far
6907 * we can only use __u16 as maximum due to the
6908 * __u16 sample size limit.
6910 if (attr->sample_stack_user >= USHRT_MAX)
6912 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6920 put_user(sizeof(*attr), &uattr->size);
6926 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6928 struct ring_buffer *rb = NULL, *old_rb = NULL;
6934 /* don't allow circular references */
6935 if (event == output_event)
6939 * Don't allow cross-cpu buffers
6941 if (output_event->cpu != event->cpu)
6945 * If its not a per-cpu rb, it must be the same task.
6947 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6951 mutex_lock(&event->mmap_mutex);
6952 /* Can't redirect output if we've got an active mmap() */
6953 if (atomic_read(&event->mmap_count))
6959 /* get the rb we want to redirect to */
6960 rb = ring_buffer_get(output_event);
6966 ring_buffer_detach(event, old_rb);
6969 ring_buffer_attach(event, rb);
6971 rcu_assign_pointer(event->rb, rb);
6974 ring_buffer_put(old_rb);
6976 * Since we detached before setting the new rb, so that we
6977 * could attach the new rb, we could have missed a wakeup.
6980 wake_up_all(&event->waitq);
6985 mutex_unlock(&event->mmap_mutex);
6992 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6994 * @attr_uptr: event_id type attributes for monitoring/sampling
6997 * @group_fd: group leader event fd
6999 SYSCALL_DEFINE5(perf_event_open,
7000 struct perf_event_attr __user *, attr_uptr,
7001 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7003 struct perf_event *group_leader = NULL, *output_event = NULL;
7004 struct perf_event *event, *sibling;
7005 struct perf_event_attr attr;
7006 struct perf_event_context *ctx;
7007 struct file *event_file = NULL;
7008 struct fd group = {NULL, 0};
7009 struct task_struct *task = NULL;
7014 int f_flags = O_RDWR;
7016 /* for future expandability... */
7017 if (flags & ~PERF_FLAG_ALL)
7020 err = perf_copy_attr(attr_uptr, &attr);
7024 if (!attr.exclude_kernel) {
7025 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7030 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7035 * In cgroup mode, the pid argument is used to pass the fd
7036 * opened to the cgroup directory in cgroupfs. The cpu argument
7037 * designates the cpu on which to monitor threads from that
7040 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7043 if (flags & PERF_FLAG_FD_CLOEXEC)
7044 f_flags |= O_CLOEXEC;
7046 event_fd = get_unused_fd_flags(f_flags);
7050 if (group_fd != -1) {
7051 err = perf_fget_light(group_fd, &group);
7054 group_leader = group.file->private_data;
7055 if (flags & PERF_FLAG_FD_OUTPUT)
7056 output_event = group_leader;
7057 if (flags & PERF_FLAG_FD_NO_GROUP)
7058 group_leader = NULL;
7061 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7062 task = find_lively_task_by_vpid(pid);
7064 err = PTR_ERR(task);
7071 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7073 if (IS_ERR(event)) {
7074 err = PTR_ERR(event);
7078 if (flags & PERF_FLAG_PID_CGROUP) {
7079 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7081 __free_event(event);
7086 account_event(event);
7089 * Special case software events and allow them to be part of
7090 * any hardware group.
7095 (is_software_event(event) != is_software_event(group_leader))) {
7096 if (is_software_event(event)) {
7098 * If event and group_leader are not both a software
7099 * event, and event is, then group leader is not.
7101 * Allow the addition of software events to !software
7102 * groups, this is safe because software events never
7105 pmu = group_leader->pmu;
7106 } else if (is_software_event(group_leader) &&
7107 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7109 * In case the group is a pure software group, and we
7110 * try to add a hardware event, move the whole group to
7111 * the hardware context.
7118 * Get the target context (task or percpu):
7120 ctx = find_get_context(pmu, task, event->cpu);
7127 put_task_struct(task);
7132 * Look up the group leader (we will attach this event to it):
7138 * Do not allow a recursive hierarchy (this new sibling
7139 * becoming part of another group-sibling):
7141 if (group_leader->group_leader != group_leader)
7144 * Do not allow to attach to a group in a different
7145 * task or CPU context:
7148 if (group_leader->ctx->type != ctx->type)
7151 if (group_leader->ctx != ctx)
7156 * Only a group leader can be exclusive or pinned
7158 if (attr.exclusive || attr.pinned)
7163 err = perf_event_set_output(event, output_event);
7168 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7170 if (IS_ERR(event_file)) {
7171 err = PTR_ERR(event_file);
7176 struct perf_event_context *gctx = group_leader->ctx;
7178 mutex_lock(&gctx->mutex);
7179 perf_remove_from_context(group_leader, false);
7182 * Removing from the context ends up with disabled
7183 * event. What we want here is event in the initial
7184 * startup state, ready to be add into new context.
7186 perf_event__state_init(group_leader);
7187 list_for_each_entry(sibling, &group_leader->sibling_list,
7189 perf_remove_from_context(sibling, false);
7190 perf_event__state_init(sibling);
7193 mutex_unlock(&gctx->mutex);
7197 WARN_ON_ONCE(ctx->parent_ctx);
7198 mutex_lock(&ctx->mutex);
7202 perf_install_in_context(ctx, group_leader, event->cpu);
7204 list_for_each_entry(sibling, &group_leader->sibling_list,
7206 perf_install_in_context(ctx, sibling, event->cpu);
7211 perf_install_in_context(ctx, event, event->cpu);
7212 perf_unpin_context(ctx);
7213 mutex_unlock(&ctx->mutex);
7217 event->owner = current;
7219 mutex_lock(¤t->perf_event_mutex);
7220 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7221 mutex_unlock(¤t->perf_event_mutex);
7224 * Precalculate sample_data sizes
7226 perf_event__header_size(event);
7227 perf_event__id_header_size(event);
7230 * Drop the reference on the group_event after placing the
7231 * new event on the sibling_list. This ensures destruction
7232 * of the group leader will find the pointer to itself in
7233 * perf_group_detach().
7236 fd_install(event_fd, event_file);
7240 perf_unpin_context(ctx);
7247 put_task_struct(task);
7251 put_unused_fd(event_fd);
7256 * perf_event_create_kernel_counter
7258 * @attr: attributes of the counter to create
7259 * @cpu: cpu in which the counter is bound
7260 * @task: task to profile (NULL for percpu)
7263 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7264 struct task_struct *task,
7265 perf_overflow_handler_t overflow_handler,
7268 struct perf_event_context *ctx;
7269 struct perf_event *event;
7273 * Get the target context (task or percpu):
7276 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7277 overflow_handler, context);
7278 if (IS_ERR(event)) {
7279 err = PTR_ERR(event);
7283 account_event(event);
7285 ctx = find_get_context(event->pmu, task, cpu);
7291 WARN_ON_ONCE(ctx->parent_ctx);
7292 mutex_lock(&ctx->mutex);
7293 perf_install_in_context(ctx, event, cpu);
7294 perf_unpin_context(ctx);
7295 mutex_unlock(&ctx->mutex);
7302 return ERR_PTR(err);
7304 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7306 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7308 struct perf_event_context *src_ctx;
7309 struct perf_event_context *dst_ctx;
7310 struct perf_event *event, *tmp;
7313 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7314 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7316 mutex_lock(&src_ctx->mutex);
7317 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7319 perf_remove_from_context(event, false);
7320 unaccount_event_cpu(event, src_cpu);
7322 list_add(&event->migrate_entry, &events);
7324 mutex_unlock(&src_ctx->mutex);
7328 mutex_lock(&dst_ctx->mutex);
7329 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7330 list_del(&event->migrate_entry);
7331 if (event->state >= PERF_EVENT_STATE_OFF)
7332 event->state = PERF_EVENT_STATE_INACTIVE;
7333 account_event_cpu(event, dst_cpu);
7334 perf_install_in_context(dst_ctx, event, dst_cpu);
7337 mutex_unlock(&dst_ctx->mutex);
7339 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7341 static void sync_child_event(struct perf_event *child_event,
7342 struct task_struct *child)
7344 struct perf_event *parent_event = child_event->parent;
7347 if (child_event->attr.inherit_stat)
7348 perf_event_read_event(child_event, child);
7350 child_val = perf_event_count(child_event);
7353 * Add back the child's count to the parent's count:
7355 atomic64_add(child_val, &parent_event->child_count);
7356 atomic64_add(child_event->total_time_enabled,
7357 &parent_event->child_total_time_enabled);
7358 atomic64_add(child_event->total_time_running,
7359 &parent_event->child_total_time_running);
7362 * Remove this event from the parent's list
7364 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7365 mutex_lock(&parent_event->child_mutex);
7366 list_del_init(&child_event->child_list);
7367 mutex_unlock(&parent_event->child_mutex);
7370 * Release the parent event, if this was the last
7373 put_event(parent_event);
7377 __perf_event_exit_task(struct perf_event *child_event,
7378 struct perf_event_context *child_ctx,
7379 struct task_struct *child)
7381 perf_remove_from_context(child_event, !!child_event->parent);
7384 * It can happen that the parent exits first, and has events
7385 * that are still around due to the child reference. These
7386 * events need to be zapped.
7388 if (child_event->parent) {
7389 sync_child_event(child_event, child);
7390 free_event(child_event);
7394 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7396 struct perf_event *child_event, *tmp;
7397 struct perf_event_context *child_ctx;
7398 unsigned long flags;
7400 if (likely(!child->perf_event_ctxp[ctxn])) {
7401 perf_event_task(child, NULL, 0);
7405 local_irq_save(flags);
7407 * We can't reschedule here because interrupts are disabled,
7408 * and either child is current or it is a task that can't be
7409 * scheduled, so we are now safe from rescheduling changing
7412 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7415 * Take the context lock here so that if find_get_context is
7416 * reading child->perf_event_ctxp, we wait until it has
7417 * incremented the context's refcount before we do put_ctx below.
7419 raw_spin_lock(&child_ctx->lock);
7420 task_ctx_sched_out(child_ctx);
7421 child->perf_event_ctxp[ctxn] = NULL;
7423 * If this context is a clone; unclone it so it can't get
7424 * swapped to another process while we're removing all
7425 * the events from it.
7427 unclone_ctx(child_ctx);
7428 update_context_time(child_ctx);
7429 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7432 * Report the task dead after unscheduling the events so that we
7433 * won't get any samples after PERF_RECORD_EXIT. We can however still
7434 * get a few PERF_RECORD_READ events.
7436 perf_event_task(child, child_ctx, 0);
7439 * We can recurse on the same lock type through:
7441 * __perf_event_exit_task()
7442 * sync_child_event()
7444 * mutex_lock(&ctx->mutex)
7446 * But since its the parent context it won't be the same instance.
7448 mutex_lock(&child_ctx->mutex);
7451 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7453 __perf_event_exit_task(child_event, child_ctx, child);
7455 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7457 __perf_event_exit_task(child_event, child_ctx, child);
7460 * If the last event was a group event, it will have appended all
7461 * its siblings to the list, but we obtained 'tmp' before that which
7462 * will still point to the list head terminating the iteration.
7464 if (!list_empty(&child_ctx->pinned_groups) ||
7465 !list_empty(&child_ctx->flexible_groups))
7468 mutex_unlock(&child_ctx->mutex);
7474 * When a child task exits, feed back event values to parent events.
7476 void perf_event_exit_task(struct task_struct *child)
7478 struct perf_event *event, *tmp;
7481 mutex_lock(&child->perf_event_mutex);
7482 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7484 list_del_init(&event->owner_entry);
7487 * Ensure the list deletion is visible before we clear
7488 * the owner, closes a race against perf_release() where
7489 * we need to serialize on the owner->perf_event_mutex.
7492 event->owner = NULL;
7494 mutex_unlock(&child->perf_event_mutex);
7496 for_each_task_context_nr(ctxn)
7497 perf_event_exit_task_context(child, ctxn);
7500 static void perf_free_event(struct perf_event *event,
7501 struct perf_event_context *ctx)
7503 struct perf_event *parent = event->parent;
7505 if (WARN_ON_ONCE(!parent))
7508 mutex_lock(&parent->child_mutex);
7509 list_del_init(&event->child_list);
7510 mutex_unlock(&parent->child_mutex);
7514 perf_group_detach(event);
7515 list_del_event(event, ctx);
7520 * free an unexposed, unused context as created by inheritance by
7521 * perf_event_init_task below, used by fork() in case of fail.
7523 void perf_event_free_task(struct task_struct *task)
7525 struct perf_event_context *ctx;
7526 struct perf_event *event, *tmp;
7529 for_each_task_context_nr(ctxn) {
7530 ctx = task->perf_event_ctxp[ctxn];
7534 mutex_lock(&ctx->mutex);
7536 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7538 perf_free_event(event, ctx);
7540 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7542 perf_free_event(event, ctx);
7544 if (!list_empty(&ctx->pinned_groups) ||
7545 !list_empty(&ctx->flexible_groups))
7548 mutex_unlock(&ctx->mutex);
7554 void perf_event_delayed_put(struct task_struct *task)
7558 for_each_task_context_nr(ctxn)
7559 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7563 * inherit a event from parent task to child task:
7565 static struct perf_event *
7566 inherit_event(struct perf_event *parent_event,
7567 struct task_struct *parent,
7568 struct perf_event_context *parent_ctx,
7569 struct task_struct *child,
7570 struct perf_event *group_leader,
7571 struct perf_event_context *child_ctx)
7573 struct perf_event *child_event;
7574 unsigned long flags;
7577 * Instead of creating recursive hierarchies of events,
7578 * we link inherited events back to the original parent,
7579 * which has a filp for sure, which we use as the reference
7582 if (parent_event->parent)
7583 parent_event = parent_event->parent;
7585 child_event = perf_event_alloc(&parent_event->attr,
7588 group_leader, parent_event,
7590 if (IS_ERR(child_event))
7593 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7594 free_event(child_event);
7601 * Make the child state follow the state of the parent event,
7602 * not its attr.disabled bit. We hold the parent's mutex,
7603 * so we won't race with perf_event_{en, dis}able_family.
7605 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7606 child_event->state = PERF_EVENT_STATE_INACTIVE;
7608 child_event->state = PERF_EVENT_STATE_OFF;
7610 if (parent_event->attr.freq) {
7611 u64 sample_period = parent_event->hw.sample_period;
7612 struct hw_perf_event *hwc = &child_event->hw;
7614 hwc->sample_period = sample_period;
7615 hwc->last_period = sample_period;
7617 local64_set(&hwc->period_left, sample_period);
7620 child_event->ctx = child_ctx;
7621 child_event->overflow_handler = parent_event->overflow_handler;
7622 child_event->overflow_handler_context
7623 = parent_event->overflow_handler_context;
7626 * Precalculate sample_data sizes
7628 perf_event__header_size(child_event);
7629 perf_event__id_header_size(child_event);
7632 * Link it up in the child's context:
7634 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7635 add_event_to_ctx(child_event, child_ctx);
7636 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7639 * Link this into the parent event's child list
7641 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7642 mutex_lock(&parent_event->child_mutex);
7643 list_add_tail(&child_event->child_list, &parent_event->child_list);
7644 mutex_unlock(&parent_event->child_mutex);
7649 static int inherit_group(struct perf_event *parent_event,
7650 struct task_struct *parent,
7651 struct perf_event_context *parent_ctx,
7652 struct task_struct *child,
7653 struct perf_event_context *child_ctx)
7655 struct perf_event *leader;
7656 struct perf_event *sub;
7657 struct perf_event *child_ctr;
7659 leader = inherit_event(parent_event, parent, parent_ctx,
7660 child, NULL, child_ctx);
7662 return PTR_ERR(leader);
7663 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7664 child_ctr = inherit_event(sub, parent, parent_ctx,
7665 child, leader, child_ctx);
7666 if (IS_ERR(child_ctr))
7667 return PTR_ERR(child_ctr);
7673 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7674 struct perf_event_context *parent_ctx,
7675 struct task_struct *child, int ctxn,
7679 struct perf_event_context *child_ctx;
7681 if (!event->attr.inherit) {
7686 child_ctx = child->perf_event_ctxp[ctxn];
7689 * This is executed from the parent task context, so
7690 * inherit events that have been marked for cloning.
7691 * First allocate and initialize a context for the
7695 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7699 child->perf_event_ctxp[ctxn] = child_ctx;
7702 ret = inherit_group(event, parent, parent_ctx,
7712 * Initialize the perf_event context in task_struct
7714 int perf_event_init_context(struct task_struct *child, int ctxn)
7716 struct perf_event_context *child_ctx, *parent_ctx;
7717 struct perf_event_context *cloned_ctx;
7718 struct perf_event *event;
7719 struct task_struct *parent = current;
7720 int inherited_all = 1;
7721 unsigned long flags;
7724 if (likely(!parent->perf_event_ctxp[ctxn]))
7728 * If the parent's context is a clone, pin it so it won't get
7731 parent_ctx = perf_pin_task_context(parent, ctxn);
7736 * No need to check if parent_ctx != NULL here; since we saw
7737 * it non-NULL earlier, the only reason for it to become NULL
7738 * is if we exit, and since we're currently in the middle of
7739 * a fork we can't be exiting at the same time.
7743 * Lock the parent list. No need to lock the child - not PID
7744 * hashed yet and not running, so nobody can access it.
7746 mutex_lock(&parent_ctx->mutex);
7749 * We dont have to disable NMIs - we are only looking at
7750 * the list, not manipulating it:
7752 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7753 ret = inherit_task_group(event, parent, parent_ctx,
7754 child, ctxn, &inherited_all);
7760 * We can't hold ctx->lock when iterating the ->flexible_group list due
7761 * to allocations, but we need to prevent rotation because
7762 * rotate_ctx() will change the list from interrupt context.
7764 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7765 parent_ctx->rotate_disable = 1;
7766 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7768 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7769 ret = inherit_task_group(event, parent, parent_ctx,
7770 child, ctxn, &inherited_all);
7775 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7776 parent_ctx->rotate_disable = 0;
7778 child_ctx = child->perf_event_ctxp[ctxn];
7780 if (child_ctx && inherited_all) {
7782 * Mark the child context as a clone of the parent
7783 * context, or of whatever the parent is a clone of.
7785 * Note that if the parent is a clone, the holding of
7786 * parent_ctx->lock avoids it from being uncloned.
7788 cloned_ctx = parent_ctx->parent_ctx;
7790 child_ctx->parent_ctx = cloned_ctx;
7791 child_ctx->parent_gen = parent_ctx->parent_gen;
7793 child_ctx->parent_ctx = parent_ctx;
7794 child_ctx->parent_gen = parent_ctx->generation;
7796 get_ctx(child_ctx->parent_ctx);
7799 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7800 mutex_unlock(&parent_ctx->mutex);
7802 perf_unpin_context(parent_ctx);
7803 put_ctx(parent_ctx);
7809 * Initialize the perf_event context in task_struct
7811 int perf_event_init_task(struct task_struct *child)
7815 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7816 mutex_init(&child->perf_event_mutex);
7817 INIT_LIST_HEAD(&child->perf_event_list);
7819 for_each_task_context_nr(ctxn) {
7820 ret = perf_event_init_context(child, ctxn);
7828 static void __init perf_event_init_all_cpus(void)
7830 struct swevent_htable *swhash;
7833 for_each_possible_cpu(cpu) {
7834 swhash = &per_cpu(swevent_htable, cpu);
7835 mutex_init(&swhash->hlist_mutex);
7836 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7840 static void perf_event_init_cpu(int cpu)
7842 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7844 mutex_lock(&swhash->hlist_mutex);
7845 if (swhash->hlist_refcount > 0) {
7846 struct swevent_hlist *hlist;
7848 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7850 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7852 mutex_unlock(&swhash->hlist_mutex);
7855 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7856 static void perf_pmu_rotate_stop(struct pmu *pmu)
7858 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7860 WARN_ON(!irqs_disabled());
7862 list_del_init(&cpuctx->rotation_list);
7865 static void __perf_event_exit_context(void *__info)
7867 struct remove_event re = { .detach_group = false };
7868 struct perf_event_context *ctx = __info;
7870 perf_pmu_rotate_stop(ctx->pmu);
7873 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
7874 __perf_remove_from_context(&re);
7878 static void perf_event_exit_cpu_context(int cpu)
7880 struct perf_event_context *ctx;
7884 idx = srcu_read_lock(&pmus_srcu);
7885 list_for_each_entry_rcu(pmu, &pmus, entry) {
7886 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7888 mutex_lock(&ctx->mutex);
7889 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7890 mutex_unlock(&ctx->mutex);
7892 srcu_read_unlock(&pmus_srcu, idx);
7895 static void perf_event_exit_cpu(int cpu)
7897 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7899 perf_event_exit_cpu_context(cpu);
7901 mutex_lock(&swhash->hlist_mutex);
7902 swevent_hlist_release(swhash);
7903 mutex_unlock(&swhash->hlist_mutex);
7906 static inline void perf_event_exit_cpu(int cpu) { }
7910 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7914 for_each_online_cpu(cpu)
7915 perf_event_exit_cpu(cpu);
7921 * Run the perf reboot notifier at the very last possible moment so that
7922 * the generic watchdog code runs as long as possible.
7924 static struct notifier_block perf_reboot_notifier = {
7925 .notifier_call = perf_reboot,
7926 .priority = INT_MIN,
7930 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7932 unsigned int cpu = (long)hcpu;
7934 switch (action & ~CPU_TASKS_FROZEN) {
7936 case CPU_UP_PREPARE:
7937 case CPU_DOWN_FAILED:
7938 perf_event_init_cpu(cpu);
7941 case CPU_UP_CANCELED:
7942 case CPU_DOWN_PREPARE:
7943 perf_event_exit_cpu(cpu);
7952 void __init perf_event_init(void)
7958 perf_event_init_all_cpus();
7959 init_srcu_struct(&pmus_srcu);
7960 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7961 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7962 perf_pmu_register(&perf_task_clock, NULL, -1);
7964 perf_cpu_notifier(perf_cpu_notify);
7965 register_reboot_notifier(&perf_reboot_notifier);
7967 ret = init_hw_breakpoint();
7968 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7970 /* do not patch jump label more than once per second */
7971 jump_label_rate_limit(&perf_sched_events, HZ);
7974 * Build time assertion that we keep the data_head at the intended
7975 * location. IOW, validation we got the __reserved[] size right.
7977 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7981 static int __init perf_event_sysfs_init(void)
7986 mutex_lock(&pmus_lock);
7988 ret = bus_register(&pmu_bus);
7992 list_for_each_entry(pmu, &pmus, entry) {
7993 if (!pmu->name || pmu->type < 0)
7996 ret = pmu_dev_alloc(pmu);
7997 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7999 pmu_bus_running = 1;
8003 mutex_unlock(&pmus_lock);
8007 device_initcall(perf_event_sysfs_init);
8009 #ifdef CONFIG_CGROUP_PERF
8010 static struct cgroup_subsys_state *
8011 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8013 struct perf_cgroup *jc;
8015 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8017 return ERR_PTR(-ENOMEM);
8019 jc->info = alloc_percpu(struct perf_cgroup_info);
8022 return ERR_PTR(-ENOMEM);
8028 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8030 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8032 free_percpu(jc->info);
8036 static int __perf_cgroup_move(void *info)
8038 struct task_struct *task = info;
8039 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8043 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8044 struct cgroup_taskset *tset)
8046 struct task_struct *task;
8048 cgroup_taskset_for_each(task, tset)
8049 task_function_call(task, __perf_cgroup_move, task);
8052 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8053 struct cgroup_subsys_state *old_css,
8054 struct task_struct *task)
8057 * cgroup_exit() is called in the copy_process() failure path.
8058 * Ignore this case since the task hasn't ran yet, this avoids
8059 * trying to poke a half freed task state from generic code.
8061 if (!(task->flags & PF_EXITING))
8064 task_function_call(task, __perf_cgroup_move, task);
8067 struct cgroup_subsys perf_event_cgrp_subsys = {
8068 .css_alloc = perf_cgroup_css_alloc,
8069 .css_free = perf_cgroup_css_free,
8070 .exit = perf_cgroup_exit,
8071 .attach = perf_cgroup_attach,
8073 #endif /* CONFIG_CGROUP_PERF */