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>
42 #include <linux/module.h>
43 #include <linux/mman.h>
44 #include <linux/compat.h>
48 #include <asm/irq_regs.h>
50 struct remote_function_call {
51 struct task_struct *p;
52 int (*func)(void *info);
57 static void remote_function(void *data)
59 struct remote_function_call *tfc = data;
60 struct task_struct *p = tfc->p;
64 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
68 tfc->ret = tfc->func(tfc->info);
72 * task_function_call - call a function on the cpu on which a task runs
73 * @p: the task to evaluate
74 * @func: the function to be called
75 * @info: the function call argument
77 * Calls the function @func when the task is currently running. This might
78 * be on the current CPU, which just calls the function directly
80 * returns: @func return value, or
81 * -ESRCH - when the process isn't running
82 * -EAGAIN - when the process moved away
85 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
87 struct remote_function_call data = {
91 .ret = -ESRCH, /* No such (running) process */
95 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
101 * cpu_function_call - call a function on the cpu
102 * @func: the function to be called
103 * @info: the function call argument
105 * Calls the function @func on the remote cpu.
107 * returns: @func return value or -ENXIO when the cpu is offline
109 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
111 struct remote_function_call data = {
115 .ret = -ENXIO, /* No such CPU */
118 smp_call_function_single(cpu, remote_function, &data, 1);
123 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
124 PERF_FLAG_FD_OUTPUT |\
125 PERF_FLAG_PID_CGROUP |\
126 PERF_FLAG_FD_CLOEXEC)
129 * branch priv levels that need permission checks
131 #define PERF_SAMPLE_BRANCH_PERM_PLM \
132 (PERF_SAMPLE_BRANCH_KERNEL |\
133 PERF_SAMPLE_BRANCH_HV)
136 EVENT_FLEXIBLE = 0x1,
138 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
142 * perf_sched_events : >0 events exist
143 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
145 struct static_key_deferred perf_sched_events __read_mostly;
146 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
147 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
149 static atomic_t nr_mmap_events __read_mostly;
150 static atomic_t nr_comm_events __read_mostly;
151 static atomic_t nr_task_events __read_mostly;
152 static atomic_t nr_freq_events __read_mostly;
154 static LIST_HEAD(pmus);
155 static DEFINE_MUTEX(pmus_lock);
156 static struct srcu_struct pmus_srcu;
159 * perf event paranoia level:
160 * -1 - not paranoid at all
161 * 0 - disallow raw tracepoint access for unpriv
162 * 1 - disallow cpu events for unpriv
163 * 2 - disallow kernel profiling for unpriv
165 int sysctl_perf_event_paranoid __read_mostly = 1;
167 /* Minimum for 512 kiB + 1 user control page */
168 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
171 * max perf event sample rate
173 #define DEFAULT_MAX_SAMPLE_RATE 100000
174 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
175 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
177 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
179 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
180 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
182 static int perf_sample_allowed_ns __read_mostly =
183 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
185 void update_perf_cpu_limits(void)
187 u64 tmp = perf_sample_period_ns;
189 tmp *= sysctl_perf_cpu_time_max_percent;
191 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
194 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
196 int perf_proc_update_handler(struct ctl_table *table, int write,
197 void __user *buffer, size_t *lenp,
200 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
205 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
206 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
207 update_perf_cpu_limits();
212 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
214 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
215 void __user *buffer, size_t *lenp,
218 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
223 update_perf_cpu_limits();
229 * perf samples are done in some very critical code paths (NMIs).
230 * If they take too much CPU time, the system can lock up and not
231 * get any real work done. This will drop the sample rate when
232 * we detect that events are taking too long.
234 #define NR_ACCUMULATED_SAMPLES 128
235 static DEFINE_PER_CPU(u64, running_sample_length);
237 static void perf_duration_warn(struct irq_work *w)
239 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
240 u64 avg_local_sample_len;
241 u64 local_samples_len;
243 local_samples_len = __get_cpu_var(running_sample_length);
244 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
246 printk_ratelimited(KERN_WARNING
247 "perf interrupt took too long (%lld > %lld), lowering "
248 "kernel.perf_event_max_sample_rate to %d\n",
249 avg_local_sample_len, allowed_ns >> 1,
250 sysctl_perf_event_sample_rate);
253 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
255 void perf_sample_event_took(u64 sample_len_ns)
257 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
258 u64 avg_local_sample_len;
259 u64 local_samples_len;
264 /* decay the counter by 1 average sample */
265 local_samples_len = __get_cpu_var(running_sample_length);
266 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
267 local_samples_len += sample_len_ns;
268 __get_cpu_var(running_sample_length) = local_samples_len;
271 * note: this will be biased artifically low until we have
272 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
273 * from having to maintain a count.
275 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
277 if (avg_local_sample_len <= allowed_ns)
280 if (max_samples_per_tick <= 1)
283 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
284 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
285 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
287 update_perf_cpu_limits();
289 if (!irq_work_queue(&perf_duration_work)) {
290 early_printk("perf interrupt took too long (%lld > %lld), lowering "
291 "kernel.perf_event_max_sample_rate to %d\n",
292 avg_local_sample_len, allowed_ns >> 1,
293 sysctl_perf_event_sample_rate);
297 static atomic64_t perf_event_id;
299 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
300 enum event_type_t event_type);
302 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
303 enum event_type_t event_type,
304 struct task_struct *task);
306 static void update_context_time(struct perf_event_context *ctx);
307 static u64 perf_event_time(struct perf_event *event);
309 void __weak perf_event_print_debug(void) { }
311 extern __weak const char *perf_pmu_name(void)
316 static inline u64 perf_clock(void)
318 return local_clock();
321 static inline struct perf_cpu_context *
322 __get_cpu_context(struct perf_event_context *ctx)
324 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
327 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
328 struct perf_event_context *ctx)
330 raw_spin_lock(&cpuctx->ctx.lock);
332 raw_spin_lock(&ctx->lock);
335 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
336 struct perf_event_context *ctx)
339 raw_spin_unlock(&ctx->lock);
340 raw_spin_unlock(&cpuctx->ctx.lock);
343 #ifdef CONFIG_CGROUP_PERF
346 * perf_cgroup_info keeps track of time_enabled for a cgroup.
347 * This is a per-cpu dynamically allocated data structure.
349 struct perf_cgroup_info {
355 struct cgroup_subsys_state css;
356 struct perf_cgroup_info __percpu *info;
360 * Must ensure cgroup is pinned (css_get) before calling
361 * this function. In other words, we cannot call this function
362 * if there is no cgroup event for the current CPU context.
364 static inline struct perf_cgroup *
365 perf_cgroup_from_task(struct task_struct *task)
367 return container_of(task_css(task, perf_event_cgrp_id),
368 struct perf_cgroup, css);
372 perf_cgroup_match(struct perf_event *event)
374 struct perf_event_context *ctx = event->ctx;
375 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
377 /* @event doesn't care about cgroup */
381 /* wants specific cgroup scope but @cpuctx isn't associated with any */
386 * Cgroup scoping is recursive. An event enabled for a cgroup is
387 * also enabled for all its descendant cgroups. If @cpuctx's
388 * cgroup is a descendant of @event's (the test covers identity
389 * case), it's a match.
391 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
392 event->cgrp->css.cgroup);
395 static inline void perf_put_cgroup(struct perf_event *event)
397 css_put(&event->cgrp->css);
400 static inline void perf_detach_cgroup(struct perf_event *event)
402 perf_put_cgroup(event);
406 static inline int is_cgroup_event(struct perf_event *event)
408 return event->cgrp != NULL;
411 static inline u64 perf_cgroup_event_time(struct perf_event *event)
413 struct perf_cgroup_info *t;
415 t = per_cpu_ptr(event->cgrp->info, event->cpu);
419 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
421 struct perf_cgroup_info *info;
426 info = this_cpu_ptr(cgrp->info);
428 info->time += now - info->timestamp;
429 info->timestamp = now;
432 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
434 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
436 __update_cgrp_time(cgrp_out);
439 static inline void update_cgrp_time_from_event(struct perf_event *event)
441 struct perf_cgroup *cgrp;
444 * ensure we access cgroup data only when needed and
445 * when we know the cgroup is pinned (css_get)
447 if (!is_cgroup_event(event))
450 cgrp = perf_cgroup_from_task(current);
452 * Do not update time when cgroup is not active
454 if (cgrp == event->cgrp)
455 __update_cgrp_time(event->cgrp);
459 perf_cgroup_set_timestamp(struct task_struct *task,
460 struct perf_event_context *ctx)
462 struct perf_cgroup *cgrp;
463 struct perf_cgroup_info *info;
466 * ctx->lock held by caller
467 * ensure we do not access cgroup data
468 * unless we have the cgroup pinned (css_get)
470 if (!task || !ctx->nr_cgroups)
473 cgrp = perf_cgroup_from_task(task);
474 info = this_cpu_ptr(cgrp->info);
475 info->timestamp = ctx->timestamp;
478 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
479 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
482 * reschedule events based on the cgroup constraint of task.
484 * mode SWOUT : schedule out everything
485 * mode SWIN : schedule in based on cgroup for next
487 void perf_cgroup_switch(struct task_struct *task, int mode)
489 struct perf_cpu_context *cpuctx;
494 * disable interrupts to avoid geting nr_cgroup
495 * changes via __perf_event_disable(). Also
498 local_irq_save(flags);
501 * we reschedule only in the presence of cgroup
502 * constrained events.
506 list_for_each_entry_rcu(pmu, &pmus, entry) {
507 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
508 if (cpuctx->unique_pmu != pmu)
509 continue; /* ensure we process each cpuctx once */
512 * perf_cgroup_events says at least one
513 * context on this CPU has cgroup events.
515 * ctx->nr_cgroups reports the number of cgroup
516 * events for a context.
518 if (cpuctx->ctx.nr_cgroups > 0) {
519 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
520 perf_pmu_disable(cpuctx->ctx.pmu);
522 if (mode & PERF_CGROUP_SWOUT) {
523 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
525 * must not be done before ctxswout due
526 * to event_filter_match() in event_sched_out()
531 if (mode & PERF_CGROUP_SWIN) {
532 WARN_ON_ONCE(cpuctx->cgrp);
534 * set cgrp before ctxsw in to allow
535 * event_filter_match() to not have to pass
538 cpuctx->cgrp = perf_cgroup_from_task(task);
539 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
541 perf_pmu_enable(cpuctx->ctx.pmu);
542 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
548 local_irq_restore(flags);
551 static inline void perf_cgroup_sched_out(struct task_struct *task,
552 struct task_struct *next)
554 struct perf_cgroup *cgrp1;
555 struct perf_cgroup *cgrp2 = NULL;
558 * we come here when we know perf_cgroup_events > 0
560 cgrp1 = perf_cgroup_from_task(task);
563 * next is NULL when called from perf_event_enable_on_exec()
564 * that will systematically cause a cgroup_switch()
567 cgrp2 = perf_cgroup_from_task(next);
570 * only schedule out current cgroup events if we know
571 * that we are switching to a different cgroup. Otherwise,
572 * do no touch the cgroup events.
575 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
578 static inline void perf_cgroup_sched_in(struct task_struct *prev,
579 struct task_struct *task)
581 struct perf_cgroup *cgrp1;
582 struct perf_cgroup *cgrp2 = NULL;
585 * we come here when we know perf_cgroup_events > 0
587 cgrp1 = perf_cgroup_from_task(task);
589 /* prev can never be NULL */
590 cgrp2 = perf_cgroup_from_task(prev);
593 * only need to schedule in cgroup events if we are changing
594 * cgroup during ctxsw. Cgroup events were not scheduled
595 * out of ctxsw out if that was not the case.
598 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
601 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
602 struct perf_event_attr *attr,
603 struct perf_event *group_leader)
605 struct perf_cgroup *cgrp;
606 struct cgroup_subsys_state *css;
607 struct fd f = fdget(fd);
613 css = css_tryget_online_from_dir(f.file->f_dentry,
614 &perf_event_cgrp_subsys);
620 cgrp = container_of(css, struct perf_cgroup, css);
624 * all events in a group must monitor
625 * the same cgroup because a task belongs
626 * to only one perf cgroup at a time
628 if (group_leader && group_leader->cgrp != cgrp) {
629 perf_detach_cgroup(event);
638 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
640 struct perf_cgroup_info *t;
641 t = per_cpu_ptr(event->cgrp->info, event->cpu);
642 event->shadow_ctx_time = now - t->timestamp;
646 perf_cgroup_defer_enabled(struct perf_event *event)
649 * when the current task's perf cgroup does not match
650 * the event's, we need to remember to call the
651 * perf_mark_enable() function the first time a task with
652 * a matching perf cgroup is scheduled in.
654 if (is_cgroup_event(event) && !perf_cgroup_match(event))
655 event->cgrp_defer_enabled = 1;
659 perf_cgroup_mark_enabled(struct perf_event *event,
660 struct perf_event_context *ctx)
662 struct perf_event *sub;
663 u64 tstamp = perf_event_time(event);
665 if (!event->cgrp_defer_enabled)
668 event->cgrp_defer_enabled = 0;
670 event->tstamp_enabled = tstamp - event->total_time_enabled;
671 list_for_each_entry(sub, &event->sibling_list, group_entry) {
672 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
673 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
674 sub->cgrp_defer_enabled = 0;
678 #else /* !CONFIG_CGROUP_PERF */
681 perf_cgroup_match(struct perf_event *event)
686 static inline void perf_detach_cgroup(struct perf_event *event)
689 static inline int is_cgroup_event(struct perf_event *event)
694 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
699 static inline void update_cgrp_time_from_event(struct perf_event *event)
703 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
707 static inline void perf_cgroup_sched_out(struct task_struct *task,
708 struct task_struct *next)
712 static inline void perf_cgroup_sched_in(struct task_struct *prev,
713 struct task_struct *task)
717 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
718 struct perf_event_attr *attr,
719 struct perf_event *group_leader)
725 perf_cgroup_set_timestamp(struct task_struct *task,
726 struct perf_event_context *ctx)
731 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
736 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
740 static inline u64 perf_cgroup_event_time(struct perf_event *event)
746 perf_cgroup_defer_enabled(struct perf_event *event)
751 perf_cgroup_mark_enabled(struct perf_event *event,
752 struct perf_event_context *ctx)
758 * set default to be dependent on timer tick just
761 #define PERF_CPU_HRTIMER (1000 / HZ)
763 * function must be called with interrupts disbled
765 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
767 struct perf_cpu_context *cpuctx;
768 enum hrtimer_restart ret = HRTIMER_NORESTART;
771 WARN_ON(!irqs_disabled());
773 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
775 rotations = perf_rotate_context(cpuctx);
778 * arm timer if needed
781 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
782 ret = HRTIMER_RESTART;
788 /* CPU is going down */
789 void perf_cpu_hrtimer_cancel(int cpu)
791 struct perf_cpu_context *cpuctx;
795 if (WARN_ON(cpu != smp_processor_id()))
798 local_irq_save(flags);
802 list_for_each_entry_rcu(pmu, &pmus, entry) {
803 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
805 if (pmu->task_ctx_nr == perf_sw_context)
808 hrtimer_cancel(&cpuctx->hrtimer);
813 local_irq_restore(flags);
816 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
818 struct hrtimer *hr = &cpuctx->hrtimer;
819 struct pmu *pmu = cpuctx->ctx.pmu;
822 /* no multiplexing needed for SW PMU */
823 if (pmu->task_ctx_nr == perf_sw_context)
827 * check default is sane, if not set then force to
828 * default interval (1/tick)
830 timer = pmu->hrtimer_interval_ms;
832 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
834 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
836 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
837 hr->function = perf_cpu_hrtimer_handler;
840 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
842 struct hrtimer *hr = &cpuctx->hrtimer;
843 struct pmu *pmu = cpuctx->ctx.pmu;
846 if (pmu->task_ctx_nr == perf_sw_context)
849 if (hrtimer_active(hr))
852 if (!hrtimer_callback_running(hr))
853 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
854 0, HRTIMER_MODE_REL_PINNED, 0);
857 void perf_pmu_disable(struct pmu *pmu)
859 int *count = this_cpu_ptr(pmu->pmu_disable_count);
861 pmu->pmu_disable(pmu);
864 void perf_pmu_enable(struct pmu *pmu)
866 int *count = this_cpu_ptr(pmu->pmu_disable_count);
868 pmu->pmu_enable(pmu);
871 static DEFINE_PER_CPU(struct list_head, rotation_list);
874 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
875 * because they're strictly cpu affine and rotate_start is called with IRQs
876 * disabled, while rotate_context is called from IRQ context.
878 static void perf_pmu_rotate_start(struct pmu *pmu)
880 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
881 struct list_head *head = &__get_cpu_var(rotation_list);
883 WARN_ON(!irqs_disabled());
885 if (list_empty(&cpuctx->rotation_list))
886 list_add(&cpuctx->rotation_list, head);
889 static void get_ctx(struct perf_event_context *ctx)
891 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
894 static void put_ctx(struct perf_event_context *ctx)
896 if (atomic_dec_and_test(&ctx->refcount)) {
898 put_ctx(ctx->parent_ctx);
900 put_task_struct(ctx->task);
901 kfree_rcu(ctx, rcu_head);
905 static void unclone_ctx(struct perf_event_context *ctx)
907 if (ctx->parent_ctx) {
908 put_ctx(ctx->parent_ctx);
909 ctx->parent_ctx = NULL;
914 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
917 * only top level events have the pid namespace they were created in
920 event = event->parent;
922 return task_tgid_nr_ns(p, event->ns);
925 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
928 * only top level events have the pid namespace they were created in
931 event = event->parent;
933 return task_pid_nr_ns(p, event->ns);
937 * If we inherit events we want to return the parent event id
940 static u64 primary_event_id(struct perf_event *event)
945 id = event->parent->id;
951 * Get the perf_event_context for a task and lock it.
952 * This has to cope with with the fact that until it is locked,
953 * the context could get moved to another task.
955 static struct perf_event_context *
956 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
958 struct perf_event_context *ctx;
962 * One of the few rules of preemptible RCU is that one cannot do
963 * rcu_read_unlock() while holding a scheduler (or nested) lock when
964 * part of the read side critical section was preemptible -- see
965 * rcu_read_unlock_special().
967 * Since ctx->lock nests under rq->lock we must ensure the entire read
968 * side critical section is non-preemptible.
972 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
975 * If this context is a clone of another, it might
976 * get swapped for another underneath us by
977 * perf_event_task_sched_out, though the
978 * rcu_read_lock() protects us from any context
979 * getting freed. Lock the context and check if it
980 * got swapped before we could get the lock, and retry
981 * if so. If we locked the right context, then it
982 * can't get swapped on us any more.
984 raw_spin_lock_irqsave(&ctx->lock, *flags);
985 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
986 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
992 if (!atomic_inc_not_zero(&ctx->refcount)) {
993 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1003 * Get the context for a task and increment its pin_count so it
1004 * can't get swapped to another task. This also increments its
1005 * reference count so that the context can't get freed.
1007 static struct perf_event_context *
1008 perf_pin_task_context(struct task_struct *task, int ctxn)
1010 struct perf_event_context *ctx;
1011 unsigned long flags;
1013 ctx = perf_lock_task_context(task, ctxn, &flags);
1016 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1021 static void perf_unpin_context(struct perf_event_context *ctx)
1023 unsigned long flags;
1025 raw_spin_lock_irqsave(&ctx->lock, flags);
1027 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1031 * Update the record of the current time in a context.
1033 static void update_context_time(struct perf_event_context *ctx)
1035 u64 now = perf_clock();
1037 ctx->time += now - ctx->timestamp;
1038 ctx->timestamp = now;
1041 static u64 perf_event_time(struct perf_event *event)
1043 struct perf_event_context *ctx = event->ctx;
1045 if (is_cgroup_event(event))
1046 return perf_cgroup_event_time(event);
1048 return ctx ? ctx->time : 0;
1052 * Update the total_time_enabled and total_time_running fields for a event.
1053 * The caller of this function needs to hold the ctx->lock.
1055 static void update_event_times(struct perf_event *event)
1057 struct perf_event_context *ctx = event->ctx;
1060 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1061 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1064 * in cgroup mode, time_enabled represents
1065 * the time the event was enabled AND active
1066 * tasks were in the monitored cgroup. This is
1067 * independent of the activity of the context as
1068 * there may be a mix of cgroup and non-cgroup events.
1070 * That is why we treat cgroup events differently
1073 if (is_cgroup_event(event))
1074 run_end = perf_cgroup_event_time(event);
1075 else if (ctx->is_active)
1076 run_end = ctx->time;
1078 run_end = event->tstamp_stopped;
1080 event->total_time_enabled = run_end - event->tstamp_enabled;
1082 if (event->state == PERF_EVENT_STATE_INACTIVE)
1083 run_end = event->tstamp_stopped;
1085 run_end = perf_event_time(event);
1087 event->total_time_running = run_end - event->tstamp_running;
1092 * Update total_time_enabled and total_time_running for all events in a group.
1094 static void update_group_times(struct perf_event *leader)
1096 struct perf_event *event;
1098 update_event_times(leader);
1099 list_for_each_entry(event, &leader->sibling_list, group_entry)
1100 update_event_times(event);
1103 static struct list_head *
1104 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1106 if (event->attr.pinned)
1107 return &ctx->pinned_groups;
1109 return &ctx->flexible_groups;
1113 * Add a event from the lists for its context.
1114 * Must be called with ctx->mutex and ctx->lock held.
1117 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1119 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1120 event->attach_state |= PERF_ATTACH_CONTEXT;
1123 * If we're a stand alone event or group leader, we go to the context
1124 * list, group events are kept attached to the group so that
1125 * perf_group_detach can, at all times, locate all siblings.
1127 if (event->group_leader == event) {
1128 struct list_head *list;
1130 if (is_software_event(event))
1131 event->group_flags |= PERF_GROUP_SOFTWARE;
1133 list = ctx_group_list(event, ctx);
1134 list_add_tail(&event->group_entry, list);
1137 if (is_cgroup_event(event))
1140 if (has_branch_stack(event))
1141 ctx->nr_branch_stack++;
1143 list_add_rcu(&event->event_entry, &ctx->event_list);
1144 if (!ctx->nr_events)
1145 perf_pmu_rotate_start(ctx->pmu);
1147 if (event->attr.inherit_stat)
1154 * Initialize event state based on the perf_event_attr::disabled.
1156 static inline void perf_event__state_init(struct perf_event *event)
1158 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1159 PERF_EVENT_STATE_INACTIVE;
1163 * Called at perf_event creation and when events are attached/detached from a
1166 static void perf_event__read_size(struct perf_event *event)
1168 int entry = sizeof(u64); /* value */
1172 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1173 size += sizeof(u64);
1175 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1176 size += sizeof(u64);
1178 if (event->attr.read_format & PERF_FORMAT_ID)
1179 entry += sizeof(u64);
1181 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1182 nr += event->group_leader->nr_siblings;
1183 size += sizeof(u64);
1187 event->read_size = size;
1190 static void perf_event__header_size(struct perf_event *event)
1192 struct perf_sample_data *data;
1193 u64 sample_type = event->attr.sample_type;
1196 perf_event__read_size(event);
1198 if (sample_type & PERF_SAMPLE_IP)
1199 size += sizeof(data->ip);
1201 if (sample_type & PERF_SAMPLE_ADDR)
1202 size += sizeof(data->addr);
1204 if (sample_type & PERF_SAMPLE_PERIOD)
1205 size += sizeof(data->period);
1207 if (sample_type & PERF_SAMPLE_WEIGHT)
1208 size += sizeof(data->weight);
1210 if (sample_type & PERF_SAMPLE_READ)
1211 size += event->read_size;
1213 if (sample_type & PERF_SAMPLE_DATA_SRC)
1214 size += sizeof(data->data_src.val);
1216 if (sample_type & PERF_SAMPLE_TRANSACTION)
1217 size += sizeof(data->txn);
1219 event->header_size = size;
1222 static void perf_event__id_header_size(struct perf_event *event)
1224 struct perf_sample_data *data;
1225 u64 sample_type = event->attr.sample_type;
1228 if (sample_type & PERF_SAMPLE_TID)
1229 size += sizeof(data->tid_entry);
1231 if (sample_type & PERF_SAMPLE_TIME)
1232 size += sizeof(data->time);
1234 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1235 size += sizeof(data->id);
1237 if (sample_type & PERF_SAMPLE_ID)
1238 size += sizeof(data->id);
1240 if (sample_type & PERF_SAMPLE_STREAM_ID)
1241 size += sizeof(data->stream_id);
1243 if (sample_type & PERF_SAMPLE_CPU)
1244 size += sizeof(data->cpu_entry);
1246 event->id_header_size = size;
1249 static void perf_group_attach(struct perf_event *event)
1251 struct perf_event *group_leader = event->group_leader, *pos;
1254 * We can have double attach due to group movement in perf_event_open.
1256 if (event->attach_state & PERF_ATTACH_GROUP)
1259 event->attach_state |= PERF_ATTACH_GROUP;
1261 if (group_leader == event)
1264 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1265 !is_software_event(event))
1266 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1268 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1269 group_leader->nr_siblings++;
1271 perf_event__header_size(group_leader);
1273 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1274 perf_event__header_size(pos);
1278 * Remove a event from the lists for its context.
1279 * Must be called with ctx->mutex and ctx->lock held.
1282 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1284 struct perf_cpu_context *cpuctx;
1286 * We can have double detach due to exit/hot-unplug + close.
1288 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1291 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1293 if (is_cgroup_event(event)) {
1295 cpuctx = __get_cpu_context(ctx);
1297 * if there are no more cgroup events
1298 * then cler cgrp to avoid stale pointer
1299 * in update_cgrp_time_from_cpuctx()
1301 if (!ctx->nr_cgroups)
1302 cpuctx->cgrp = NULL;
1305 if (has_branch_stack(event))
1306 ctx->nr_branch_stack--;
1309 if (event->attr.inherit_stat)
1312 list_del_rcu(&event->event_entry);
1314 if (event->group_leader == event)
1315 list_del_init(&event->group_entry);
1317 update_group_times(event);
1320 * If event was in error state, then keep it
1321 * that way, otherwise bogus counts will be
1322 * returned on read(). The only way to get out
1323 * of error state is by explicit re-enabling
1326 if (event->state > PERF_EVENT_STATE_OFF)
1327 event->state = PERF_EVENT_STATE_OFF;
1332 static void perf_group_detach(struct perf_event *event)
1334 struct perf_event *sibling, *tmp;
1335 struct list_head *list = NULL;
1338 * We can have double detach due to exit/hot-unplug + close.
1340 if (!(event->attach_state & PERF_ATTACH_GROUP))
1343 event->attach_state &= ~PERF_ATTACH_GROUP;
1346 * If this is a sibling, remove it from its group.
1348 if (event->group_leader != event) {
1349 list_del_init(&event->group_entry);
1350 event->group_leader->nr_siblings--;
1354 if (!list_empty(&event->group_entry))
1355 list = &event->group_entry;
1358 * If this was a group event with sibling events then
1359 * upgrade the siblings to singleton events by adding them
1360 * to whatever list we are on.
1362 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1364 list_move_tail(&sibling->group_entry, list);
1365 sibling->group_leader = sibling;
1367 /* Inherit group flags from the previous leader */
1368 sibling->group_flags = event->group_flags;
1372 perf_event__header_size(event->group_leader);
1374 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1375 perf_event__header_size(tmp);
1379 event_filter_match(struct perf_event *event)
1381 return (event->cpu == -1 || event->cpu == smp_processor_id())
1382 && perf_cgroup_match(event);
1386 event_sched_out(struct perf_event *event,
1387 struct perf_cpu_context *cpuctx,
1388 struct perf_event_context *ctx)
1390 u64 tstamp = perf_event_time(event);
1393 * An event which could not be activated because of
1394 * filter mismatch still needs to have its timings
1395 * maintained, otherwise bogus information is return
1396 * via read() for time_enabled, time_running:
1398 if (event->state == PERF_EVENT_STATE_INACTIVE
1399 && !event_filter_match(event)) {
1400 delta = tstamp - event->tstamp_stopped;
1401 event->tstamp_running += delta;
1402 event->tstamp_stopped = tstamp;
1405 if (event->state != PERF_EVENT_STATE_ACTIVE)
1408 perf_pmu_disable(event->pmu);
1410 event->state = PERF_EVENT_STATE_INACTIVE;
1411 if (event->pending_disable) {
1412 event->pending_disable = 0;
1413 event->state = PERF_EVENT_STATE_OFF;
1415 event->tstamp_stopped = tstamp;
1416 event->pmu->del(event, 0);
1419 if (!is_software_event(event))
1420 cpuctx->active_oncpu--;
1422 if (event->attr.freq && event->attr.sample_freq)
1424 if (event->attr.exclusive || !cpuctx->active_oncpu)
1425 cpuctx->exclusive = 0;
1427 perf_pmu_enable(event->pmu);
1431 group_sched_out(struct perf_event *group_event,
1432 struct perf_cpu_context *cpuctx,
1433 struct perf_event_context *ctx)
1435 struct perf_event *event;
1436 int state = group_event->state;
1438 event_sched_out(group_event, cpuctx, ctx);
1441 * Schedule out siblings (if any):
1443 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1444 event_sched_out(event, cpuctx, ctx);
1446 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1447 cpuctx->exclusive = 0;
1450 struct remove_event {
1451 struct perf_event *event;
1456 * Cross CPU call to remove a performance event
1458 * We disable the event on the hardware level first. After that we
1459 * remove it from the context list.
1461 static int __perf_remove_from_context(void *info)
1463 struct remove_event *re = info;
1464 struct perf_event *event = re->event;
1465 struct perf_event_context *ctx = event->ctx;
1466 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1468 raw_spin_lock(&ctx->lock);
1469 event_sched_out(event, cpuctx, ctx);
1470 if (re->detach_group)
1471 perf_group_detach(event);
1472 list_del_event(event, ctx);
1473 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1475 cpuctx->task_ctx = NULL;
1477 raw_spin_unlock(&ctx->lock);
1484 * Remove the event from a task's (or a CPU's) list of events.
1486 * CPU events are removed with a smp call. For task events we only
1487 * call when the task is on a CPU.
1489 * If event->ctx is a cloned context, callers must make sure that
1490 * every task struct that event->ctx->task could possibly point to
1491 * remains valid. This is OK when called from perf_release since
1492 * that only calls us on the top-level context, which can't be a clone.
1493 * When called from perf_event_exit_task, it's OK because the
1494 * context has been detached from its task.
1496 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1498 struct perf_event_context *ctx = event->ctx;
1499 struct task_struct *task = ctx->task;
1500 struct remove_event re = {
1502 .detach_group = detach_group,
1505 lockdep_assert_held(&ctx->mutex);
1509 * Per cpu events are removed via an smp call and
1510 * the removal is always successful.
1512 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1517 if (!task_function_call(task, __perf_remove_from_context, &re))
1520 raw_spin_lock_irq(&ctx->lock);
1522 * If we failed to find a running task, but find the context active now
1523 * that we've acquired the ctx->lock, retry.
1525 if (ctx->is_active) {
1526 raw_spin_unlock_irq(&ctx->lock);
1531 * Since the task isn't running, its safe to remove the event, us
1532 * holding the ctx->lock ensures the task won't get scheduled in.
1535 perf_group_detach(event);
1536 list_del_event(event, ctx);
1537 raw_spin_unlock_irq(&ctx->lock);
1541 * Cross CPU call to disable a performance event
1543 int __perf_event_disable(void *info)
1545 struct perf_event *event = info;
1546 struct perf_event_context *ctx = event->ctx;
1547 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1550 * If this is a per-task event, need to check whether this
1551 * event's task is the current task on this cpu.
1553 * Can trigger due to concurrent perf_event_context_sched_out()
1554 * flipping contexts around.
1556 if (ctx->task && cpuctx->task_ctx != ctx)
1559 raw_spin_lock(&ctx->lock);
1562 * If the event is on, turn it off.
1563 * If it is in error state, leave it in error state.
1565 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1566 update_context_time(ctx);
1567 update_cgrp_time_from_event(event);
1568 update_group_times(event);
1569 if (event == event->group_leader)
1570 group_sched_out(event, cpuctx, ctx);
1572 event_sched_out(event, cpuctx, ctx);
1573 event->state = PERF_EVENT_STATE_OFF;
1576 raw_spin_unlock(&ctx->lock);
1584 * If event->ctx is a cloned context, callers must make sure that
1585 * every task struct that event->ctx->task could possibly point to
1586 * remains valid. This condition is satisifed when called through
1587 * perf_event_for_each_child or perf_event_for_each because they
1588 * hold the top-level event's child_mutex, so any descendant that
1589 * goes to exit will block in sync_child_event.
1590 * When called from perf_pending_event it's OK because event->ctx
1591 * is the current context on this CPU and preemption is disabled,
1592 * hence we can't get into perf_event_task_sched_out for this context.
1594 void perf_event_disable(struct perf_event *event)
1596 struct perf_event_context *ctx = event->ctx;
1597 struct task_struct *task = ctx->task;
1601 * Disable the event on the cpu that it's on
1603 cpu_function_call(event->cpu, __perf_event_disable, event);
1608 if (!task_function_call(task, __perf_event_disable, event))
1611 raw_spin_lock_irq(&ctx->lock);
1613 * If the event is still active, we need to retry the cross-call.
1615 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1616 raw_spin_unlock_irq(&ctx->lock);
1618 * Reload the task pointer, it might have been changed by
1619 * a concurrent perf_event_context_sched_out().
1626 * Since we have the lock this context can't be scheduled
1627 * in, so we can change the state safely.
1629 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1630 update_group_times(event);
1631 event->state = PERF_EVENT_STATE_OFF;
1633 raw_spin_unlock_irq(&ctx->lock);
1635 EXPORT_SYMBOL_GPL(perf_event_disable);
1637 static void perf_set_shadow_time(struct perf_event *event,
1638 struct perf_event_context *ctx,
1642 * use the correct time source for the time snapshot
1644 * We could get by without this by leveraging the
1645 * fact that to get to this function, the caller
1646 * has most likely already called update_context_time()
1647 * and update_cgrp_time_xx() and thus both timestamp
1648 * are identical (or very close). Given that tstamp is,
1649 * already adjusted for cgroup, we could say that:
1650 * tstamp - ctx->timestamp
1652 * tstamp - cgrp->timestamp.
1654 * Then, in perf_output_read(), the calculation would
1655 * work with no changes because:
1656 * - event is guaranteed scheduled in
1657 * - no scheduled out in between
1658 * - thus the timestamp would be the same
1660 * But this is a bit hairy.
1662 * So instead, we have an explicit cgroup call to remain
1663 * within the time time source all along. We believe it
1664 * is cleaner and simpler to understand.
1666 if (is_cgroup_event(event))
1667 perf_cgroup_set_shadow_time(event, tstamp);
1669 event->shadow_ctx_time = tstamp - ctx->timestamp;
1672 #define MAX_INTERRUPTS (~0ULL)
1674 static void perf_log_throttle(struct perf_event *event, int enable);
1677 event_sched_in(struct perf_event *event,
1678 struct perf_cpu_context *cpuctx,
1679 struct perf_event_context *ctx)
1681 u64 tstamp = perf_event_time(event);
1684 lockdep_assert_held(&ctx->lock);
1686 if (event->state <= PERF_EVENT_STATE_OFF)
1689 event->state = PERF_EVENT_STATE_ACTIVE;
1690 event->oncpu = smp_processor_id();
1693 * Unthrottle events, since we scheduled we might have missed several
1694 * ticks already, also for a heavily scheduling task there is little
1695 * guarantee it'll get a tick in a timely manner.
1697 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1698 perf_log_throttle(event, 1);
1699 event->hw.interrupts = 0;
1703 * The new state must be visible before we turn it on in the hardware:
1707 perf_pmu_disable(event->pmu);
1709 if (event->pmu->add(event, PERF_EF_START)) {
1710 event->state = PERF_EVENT_STATE_INACTIVE;
1716 event->tstamp_running += tstamp - event->tstamp_stopped;
1718 perf_set_shadow_time(event, ctx, tstamp);
1720 if (!is_software_event(event))
1721 cpuctx->active_oncpu++;
1723 if (event->attr.freq && event->attr.sample_freq)
1726 if (event->attr.exclusive)
1727 cpuctx->exclusive = 1;
1730 perf_pmu_enable(event->pmu);
1736 group_sched_in(struct perf_event *group_event,
1737 struct perf_cpu_context *cpuctx,
1738 struct perf_event_context *ctx)
1740 struct perf_event *event, *partial_group = NULL;
1741 struct pmu *pmu = ctx->pmu;
1742 u64 now = ctx->time;
1743 bool simulate = false;
1745 if (group_event->state == PERF_EVENT_STATE_OFF)
1748 pmu->start_txn(pmu);
1750 if (event_sched_in(group_event, cpuctx, ctx)) {
1751 pmu->cancel_txn(pmu);
1752 perf_cpu_hrtimer_restart(cpuctx);
1757 * Schedule in siblings as one group (if any):
1759 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1760 if (event_sched_in(event, cpuctx, ctx)) {
1761 partial_group = event;
1766 if (!pmu->commit_txn(pmu))
1771 * Groups can be scheduled in as one unit only, so undo any
1772 * partial group before returning:
1773 * The events up to the failed event are scheduled out normally,
1774 * tstamp_stopped will be updated.
1776 * The failed events and the remaining siblings need to have
1777 * their timings updated as if they had gone thru event_sched_in()
1778 * and event_sched_out(). This is required to get consistent timings
1779 * across the group. This also takes care of the case where the group
1780 * could never be scheduled by ensuring tstamp_stopped is set to mark
1781 * the time the event was actually stopped, such that time delta
1782 * calculation in update_event_times() is correct.
1784 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1785 if (event == partial_group)
1789 event->tstamp_running += now - event->tstamp_stopped;
1790 event->tstamp_stopped = now;
1792 event_sched_out(event, cpuctx, ctx);
1795 event_sched_out(group_event, cpuctx, ctx);
1797 pmu->cancel_txn(pmu);
1799 perf_cpu_hrtimer_restart(cpuctx);
1805 * Work out whether we can put this event group on the CPU now.
1807 static int group_can_go_on(struct perf_event *event,
1808 struct perf_cpu_context *cpuctx,
1812 * Groups consisting entirely of software events can always go on.
1814 if (event->group_flags & PERF_GROUP_SOFTWARE)
1817 * If an exclusive group is already on, no other hardware
1820 if (cpuctx->exclusive)
1823 * If this group is exclusive and there are already
1824 * events on the CPU, it can't go on.
1826 if (event->attr.exclusive && cpuctx->active_oncpu)
1829 * Otherwise, try to add it if all previous groups were able
1835 static void add_event_to_ctx(struct perf_event *event,
1836 struct perf_event_context *ctx)
1838 u64 tstamp = perf_event_time(event);
1840 list_add_event(event, ctx);
1841 perf_group_attach(event);
1842 event->tstamp_enabled = tstamp;
1843 event->tstamp_running = tstamp;
1844 event->tstamp_stopped = tstamp;
1847 static void task_ctx_sched_out(struct perf_event_context *ctx);
1849 ctx_sched_in(struct perf_event_context *ctx,
1850 struct perf_cpu_context *cpuctx,
1851 enum event_type_t event_type,
1852 struct task_struct *task);
1854 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1855 struct perf_event_context *ctx,
1856 struct task_struct *task)
1858 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1860 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1861 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1863 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1867 * Cross CPU call to install and enable a performance event
1869 * Must be called with ctx->mutex held
1871 static int __perf_install_in_context(void *info)
1873 struct perf_event *event = info;
1874 struct perf_event_context *ctx = event->ctx;
1875 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1876 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1877 struct task_struct *task = current;
1879 perf_ctx_lock(cpuctx, task_ctx);
1880 perf_pmu_disable(cpuctx->ctx.pmu);
1883 * If there was an active task_ctx schedule it out.
1886 task_ctx_sched_out(task_ctx);
1889 * If the context we're installing events in is not the
1890 * active task_ctx, flip them.
1892 if (ctx->task && task_ctx != ctx) {
1894 raw_spin_unlock(&task_ctx->lock);
1895 raw_spin_lock(&ctx->lock);
1900 cpuctx->task_ctx = task_ctx;
1901 task = task_ctx->task;
1904 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1906 update_context_time(ctx);
1908 * update cgrp time only if current cgrp
1909 * matches event->cgrp. Must be done before
1910 * calling add_event_to_ctx()
1912 update_cgrp_time_from_event(event);
1914 add_event_to_ctx(event, ctx);
1917 * Schedule everything back in
1919 perf_event_sched_in(cpuctx, task_ctx, task);
1921 perf_pmu_enable(cpuctx->ctx.pmu);
1922 perf_ctx_unlock(cpuctx, task_ctx);
1928 * Attach a performance event to a context
1930 * First we add the event to the list with the hardware enable bit
1931 * in event->hw_config cleared.
1933 * If the event is attached to a task which is on a CPU we use a smp
1934 * call to enable it in the task context. The task might have been
1935 * scheduled away, but we check this in the smp call again.
1938 perf_install_in_context(struct perf_event_context *ctx,
1939 struct perf_event *event,
1942 struct task_struct *task = ctx->task;
1944 lockdep_assert_held(&ctx->mutex);
1947 if (event->cpu != -1)
1952 * Per cpu events are installed via an smp call and
1953 * the install is always successful.
1955 cpu_function_call(cpu, __perf_install_in_context, event);
1960 if (!task_function_call(task, __perf_install_in_context, event))
1963 raw_spin_lock_irq(&ctx->lock);
1965 * If we failed to find a running task, but find the context active now
1966 * that we've acquired the ctx->lock, retry.
1968 if (ctx->is_active) {
1969 raw_spin_unlock_irq(&ctx->lock);
1974 * Since the task isn't running, its safe to add the event, us holding
1975 * the ctx->lock ensures the task won't get scheduled in.
1977 add_event_to_ctx(event, ctx);
1978 raw_spin_unlock_irq(&ctx->lock);
1982 * Put a event into inactive state and update time fields.
1983 * Enabling the leader of a group effectively enables all
1984 * the group members that aren't explicitly disabled, so we
1985 * have to update their ->tstamp_enabled also.
1986 * Note: this works for group members as well as group leaders
1987 * since the non-leader members' sibling_lists will be empty.
1989 static void __perf_event_mark_enabled(struct perf_event *event)
1991 struct perf_event *sub;
1992 u64 tstamp = perf_event_time(event);
1994 event->state = PERF_EVENT_STATE_INACTIVE;
1995 event->tstamp_enabled = tstamp - event->total_time_enabled;
1996 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1997 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1998 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2003 * Cross CPU call to enable a performance event
2005 static int __perf_event_enable(void *info)
2007 struct perf_event *event = info;
2008 struct perf_event_context *ctx = event->ctx;
2009 struct perf_event *leader = event->group_leader;
2010 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2014 * There's a time window between 'ctx->is_active' check
2015 * in perf_event_enable function and this place having:
2017 * - ctx->lock unlocked
2019 * where the task could be killed and 'ctx' deactivated
2020 * by perf_event_exit_task.
2022 if (!ctx->is_active)
2025 raw_spin_lock(&ctx->lock);
2026 update_context_time(ctx);
2028 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2032 * set current task's cgroup time reference point
2034 perf_cgroup_set_timestamp(current, ctx);
2036 __perf_event_mark_enabled(event);
2038 if (!event_filter_match(event)) {
2039 if (is_cgroup_event(event))
2040 perf_cgroup_defer_enabled(event);
2045 * If the event is in a group and isn't the group leader,
2046 * then don't put it on unless the group is on.
2048 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2051 if (!group_can_go_on(event, cpuctx, 1)) {
2054 if (event == leader)
2055 err = group_sched_in(event, cpuctx, ctx);
2057 err = event_sched_in(event, cpuctx, ctx);
2062 * If this event can't go on and it's part of a
2063 * group, then the whole group has to come off.
2065 if (leader != event) {
2066 group_sched_out(leader, cpuctx, ctx);
2067 perf_cpu_hrtimer_restart(cpuctx);
2069 if (leader->attr.pinned) {
2070 update_group_times(leader);
2071 leader->state = PERF_EVENT_STATE_ERROR;
2076 raw_spin_unlock(&ctx->lock);
2084 * If event->ctx is a cloned context, callers must make sure that
2085 * every task struct that event->ctx->task could possibly point to
2086 * remains valid. This condition is satisfied when called through
2087 * perf_event_for_each_child or perf_event_for_each as described
2088 * for perf_event_disable.
2090 void perf_event_enable(struct perf_event *event)
2092 struct perf_event_context *ctx = event->ctx;
2093 struct task_struct *task = ctx->task;
2097 * Enable the event on the cpu that it's on
2099 cpu_function_call(event->cpu, __perf_event_enable, event);
2103 raw_spin_lock_irq(&ctx->lock);
2104 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2108 * If the event is in error state, clear that first.
2109 * That way, if we see the event in error state below, we
2110 * know that it has gone back into error state, as distinct
2111 * from the task having been scheduled away before the
2112 * cross-call arrived.
2114 if (event->state == PERF_EVENT_STATE_ERROR)
2115 event->state = PERF_EVENT_STATE_OFF;
2118 if (!ctx->is_active) {
2119 __perf_event_mark_enabled(event);
2123 raw_spin_unlock_irq(&ctx->lock);
2125 if (!task_function_call(task, __perf_event_enable, event))
2128 raw_spin_lock_irq(&ctx->lock);
2131 * If the context is active and the event is still off,
2132 * we need to retry the cross-call.
2134 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2136 * task could have been flipped by a concurrent
2137 * perf_event_context_sched_out()
2144 raw_spin_unlock_irq(&ctx->lock);
2146 EXPORT_SYMBOL_GPL(perf_event_enable);
2148 int perf_event_refresh(struct perf_event *event, int refresh)
2151 * not supported on inherited events
2153 if (event->attr.inherit || !is_sampling_event(event))
2156 atomic_add(refresh, &event->event_limit);
2157 perf_event_enable(event);
2161 EXPORT_SYMBOL_GPL(perf_event_refresh);
2163 static void ctx_sched_out(struct perf_event_context *ctx,
2164 struct perf_cpu_context *cpuctx,
2165 enum event_type_t event_type)
2167 struct perf_event *event;
2168 int is_active = ctx->is_active;
2170 ctx->is_active &= ~event_type;
2171 if (likely(!ctx->nr_events))
2174 update_context_time(ctx);
2175 update_cgrp_time_from_cpuctx(cpuctx);
2176 if (!ctx->nr_active)
2179 perf_pmu_disable(ctx->pmu);
2180 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2181 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2182 group_sched_out(event, cpuctx, ctx);
2185 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2186 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2187 group_sched_out(event, cpuctx, ctx);
2189 perf_pmu_enable(ctx->pmu);
2193 * Test whether two contexts are equivalent, i.e. whether they have both been
2194 * cloned from the same version of the same context.
2196 * Equivalence is measured using a generation number in the context that is
2197 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2198 * and list_del_event().
2200 static int context_equiv(struct perf_event_context *ctx1,
2201 struct perf_event_context *ctx2)
2203 /* Pinning disables the swap optimization */
2204 if (ctx1->pin_count || ctx2->pin_count)
2207 /* If ctx1 is the parent of ctx2 */
2208 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2211 /* If ctx2 is the parent of ctx1 */
2212 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2216 * If ctx1 and ctx2 have the same parent; we flatten the parent
2217 * hierarchy, see perf_event_init_context().
2219 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2220 ctx1->parent_gen == ctx2->parent_gen)
2227 static void __perf_event_sync_stat(struct perf_event *event,
2228 struct perf_event *next_event)
2232 if (!event->attr.inherit_stat)
2236 * Update the event value, we cannot use perf_event_read()
2237 * because we're in the middle of a context switch and have IRQs
2238 * disabled, which upsets smp_call_function_single(), however
2239 * we know the event must be on the current CPU, therefore we
2240 * don't need to use it.
2242 switch (event->state) {
2243 case PERF_EVENT_STATE_ACTIVE:
2244 event->pmu->read(event);
2247 case PERF_EVENT_STATE_INACTIVE:
2248 update_event_times(event);
2256 * In order to keep per-task stats reliable we need to flip the event
2257 * values when we flip the contexts.
2259 value = local64_read(&next_event->count);
2260 value = local64_xchg(&event->count, value);
2261 local64_set(&next_event->count, value);
2263 swap(event->total_time_enabled, next_event->total_time_enabled);
2264 swap(event->total_time_running, next_event->total_time_running);
2267 * Since we swizzled the values, update the user visible data too.
2269 perf_event_update_userpage(event);
2270 perf_event_update_userpage(next_event);
2273 static void perf_event_sync_stat(struct perf_event_context *ctx,
2274 struct perf_event_context *next_ctx)
2276 struct perf_event *event, *next_event;
2281 update_context_time(ctx);
2283 event = list_first_entry(&ctx->event_list,
2284 struct perf_event, event_entry);
2286 next_event = list_first_entry(&next_ctx->event_list,
2287 struct perf_event, event_entry);
2289 while (&event->event_entry != &ctx->event_list &&
2290 &next_event->event_entry != &next_ctx->event_list) {
2292 __perf_event_sync_stat(event, next_event);
2294 event = list_next_entry(event, event_entry);
2295 next_event = list_next_entry(next_event, event_entry);
2299 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2300 struct task_struct *next)
2302 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2303 struct perf_event_context *next_ctx;
2304 struct perf_event_context *parent, *next_parent;
2305 struct perf_cpu_context *cpuctx;
2311 cpuctx = __get_cpu_context(ctx);
2312 if (!cpuctx->task_ctx)
2316 next_ctx = next->perf_event_ctxp[ctxn];
2320 parent = rcu_dereference(ctx->parent_ctx);
2321 next_parent = rcu_dereference(next_ctx->parent_ctx);
2323 /* If neither context have a parent context; they cannot be clones. */
2324 if (!parent || !next_parent)
2327 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2329 * Looks like the two contexts are clones, so we might be
2330 * able to optimize the context switch. We lock both
2331 * contexts and check that they are clones under the
2332 * lock (including re-checking that neither has been
2333 * uncloned in the meantime). It doesn't matter which
2334 * order we take the locks because no other cpu could
2335 * be trying to lock both of these tasks.
2337 raw_spin_lock(&ctx->lock);
2338 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2339 if (context_equiv(ctx, next_ctx)) {
2341 * XXX do we need a memory barrier of sorts
2342 * wrt to rcu_dereference() of perf_event_ctxp
2344 task->perf_event_ctxp[ctxn] = next_ctx;
2345 next->perf_event_ctxp[ctxn] = ctx;
2347 next_ctx->task = task;
2350 perf_event_sync_stat(ctx, next_ctx);
2352 raw_spin_unlock(&next_ctx->lock);
2353 raw_spin_unlock(&ctx->lock);
2359 raw_spin_lock(&ctx->lock);
2360 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2361 cpuctx->task_ctx = NULL;
2362 raw_spin_unlock(&ctx->lock);
2366 #define for_each_task_context_nr(ctxn) \
2367 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2370 * Called from scheduler to remove the events of the current task,
2371 * with interrupts disabled.
2373 * We stop each event and update the event value in event->count.
2375 * This does not protect us against NMI, but disable()
2376 * sets the disabled bit in the control field of event _before_
2377 * accessing the event control register. If a NMI hits, then it will
2378 * not restart the event.
2380 void __perf_event_task_sched_out(struct task_struct *task,
2381 struct task_struct *next)
2385 for_each_task_context_nr(ctxn)
2386 perf_event_context_sched_out(task, ctxn, next);
2389 * if cgroup events exist on this CPU, then we need
2390 * to check if we have to switch out PMU state.
2391 * cgroup event are system-wide mode only
2393 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2394 perf_cgroup_sched_out(task, next);
2397 static void task_ctx_sched_out(struct perf_event_context *ctx)
2399 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2401 if (!cpuctx->task_ctx)
2404 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2407 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2408 cpuctx->task_ctx = NULL;
2412 * Called with IRQs disabled
2414 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2415 enum event_type_t event_type)
2417 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2421 ctx_pinned_sched_in(struct perf_event_context *ctx,
2422 struct perf_cpu_context *cpuctx)
2424 struct perf_event *event;
2426 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2427 if (event->state <= PERF_EVENT_STATE_OFF)
2429 if (!event_filter_match(event))
2432 /* may need to reset tstamp_enabled */
2433 if (is_cgroup_event(event))
2434 perf_cgroup_mark_enabled(event, ctx);
2436 if (group_can_go_on(event, cpuctx, 1))
2437 group_sched_in(event, cpuctx, ctx);
2440 * If this pinned group hasn't been scheduled,
2441 * put it in error state.
2443 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2444 update_group_times(event);
2445 event->state = PERF_EVENT_STATE_ERROR;
2451 ctx_flexible_sched_in(struct perf_event_context *ctx,
2452 struct perf_cpu_context *cpuctx)
2454 struct perf_event *event;
2457 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2458 /* Ignore events in OFF or ERROR state */
2459 if (event->state <= PERF_EVENT_STATE_OFF)
2462 * Listen to the 'cpu' scheduling filter constraint
2465 if (!event_filter_match(event))
2468 /* may need to reset tstamp_enabled */
2469 if (is_cgroup_event(event))
2470 perf_cgroup_mark_enabled(event, ctx);
2472 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2473 if (group_sched_in(event, cpuctx, ctx))
2480 ctx_sched_in(struct perf_event_context *ctx,
2481 struct perf_cpu_context *cpuctx,
2482 enum event_type_t event_type,
2483 struct task_struct *task)
2486 int is_active = ctx->is_active;
2488 ctx->is_active |= event_type;
2489 if (likely(!ctx->nr_events))
2493 ctx->timestamp = now;
2494 perf_cgroup_set_timestamp(task, ctx);
2496 * First go through the list and put on any pinned groups
2497 * in order to give them the best chance of going on.
2499 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2500 ctx_pinned_sched_in(ctx, cpuctx);
2502 /* Then walk through the lower prio flexible groups */
2503 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2504 ctx_flexible_sched_in(ctx, cpuctx);
2507 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2508 enum event_type_t event_type,
2509 struct task_struct *task)
2511 struct perf_event_context *ctx = &cpuctx->ctx;
2513 ctx_sched_in(ctx, cpuctx, event_type, task);
2516 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2517 struct task_struct *task)
2519 struct perf_cpu_context *cpuctx;
2521 cpuctx = __get_cpu_context(ctx);
2522 if (cpuctx->task_ctx == ctx)
2525 perf_ctx_lock(cpuctx, ctx);
2526 perf_pmu_disable(ctx->pmu);
2528 * We want to keep the following priority order:
2529 * cpu pinned (that don't need to move), task pinned,
2530 * cpu flexible, task flexible.
2532 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2535 cpuctx->task_ctx = ctx;
2537 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2539 perf_pmu_enable(ctx->pmu);
2540 perf_ctx_unlock(cpuctx, ctx);
2543 * Since these rotations are per-cpu, we need to ensure the
2544 * cpu-context we got scheduled on is actually rotating.
2546 perf_pmu_rotate_start(ctx->pmu);
2550 * When sampling the branck stack in system-wide, it may be necessary
2551 * to flush the stack on context switch. This happens when the branch
2552 * stack does not tag its entries with the pid of the current task.
2553 * Otherwise it becomes impossible to associate a branch entry with a
2554 * task. This ambiguity is more likely to appear when the branch stack
2555 * supports priv level filtering and the user sets it to monitor only
2556 * at the user level (which could be a useful measurement in system-wide
2557 * mode). In that case, the risk is high of having a branch stack with
2558 * branch from multiple tasks. Flushing may mean dropping the existing
2559 * entries or stashing them somewhere in the PMU specific code layer.
2561 * This function provides the context switch callback to the lower code
2562 * layer. It is invoked ONLY when there is at least one system-wide context
2563 * with at least one active event using taken branch sampling.
2565 static void perf_branch_stack_sched_in(struct task_struct *prev,
2566 struct task_struct *task)
2568 struct perf_cpu_context *cpuctx;
2570 unsigned long flags;
2572 /* no need to flush branch stack if not changing task */
2576 local_irq_save(flags);
2580 list_for_each_entry_rcu(pmu, &pmus, entry) {
2581 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2584 * check if the context has at least one
2585 * event using PERF_SAMPLE_BRANCH_STACK
2587 if (cpuctx->ctx.nr_branch_stack > 0
2588 && pmu->flush_branch_stack) {
2590 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2592 perf_pmu_disable(pmu);
2594 pmu->flush_branch_stack();
2596 perf_pmu_enable(pmu);
2598 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2604 local_irq_restore(flags);
2608 * Called from scheduler to add the events of the current task
2609 * with interrupts disabled.
2611 * We restore the event value and then enable it.
2613 * This does not protect us against NMI, but enable()
2614 * sets the enabled bit in the control field of event _before_
2615 * accessing the event control register. If a NMI hits, then it will
2616 * keep the event running.
2618 void __perf_event_task_sched_in(struct task_struct *prev,
2619 struct task_struct *task)
2621 struct perf_event_context *ctx;
2624 for_each_task_context_nr(ctxn) {
2625 ctx = task->perf_event_ctxp[ctxn];
2629 perf_event_context_sched_in(ctx, task);
2632 * if cgroup events exist on this CPU, then we need
2633 * to check if we have to switch in PMU state.
2634 * cgroup event are system-wide mode only
2636 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2637 perf_cgroup_sched_in(prev, task);
2639 /* check for system-wide branch_stack events */
2640 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2641 perf_branch_stack_sched_in(prev, task);
2644 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2646 u64 frequency = event->attr.sample_freq;
2647 u64 sec = NSEC_PER_SEC;
2648 u64 divisor, dividend;
2650 int count_fls, nsec_fls, frequency_fls, sec_fls;
2652 count_fls = fls64(count);
2653 nsec_fls = fls64(nsec);
2654 frequency_fls = fls64(frequency);
2658 * We got @count in @nsec, with a target of sample_freq HZ
2659 * the target period becomes:
2662 * period = -------------------
2663 * @nsec * sample_freq
2668 * Reduce accuracy by one bit such that @a and @b converge
2669 * to a similar magnitude.
2671 #define REDUCE_FLS(a, b) \
2673 if (a##_fls > b##_fls) { \
2683 * Reduce accuracy until either term fits in a u64, then proceed with
2684 * the other, so that finally we can do a u64/u64 division.
2686 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2687 REDUCE_FLS(nsec, frequency);
2688 REDUCE_FLS(sec, count);
2691 if (count_fls + sec_fls > 64) {
2692 divisor = nsec * frequency;
2694 while (count_fls + sec_fls > 64) {
2695 REDUCE_FLS(count, sec);
2699 dividend = count * sec;
2701 dividend = count * sec;
2703 while (nsec_fls + frequency_fls > 64) {
2704 REDUCE_FLS(nsec, frequency);
2708 divisor = nsec * frequency;
2714 return div64_u64(dividend, divisor);
2717 static DEFINE_PER_CPU(int, perf_throttled_count);
2718 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2720 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2722 struct hw_perf_event *hwc = &event->hw;
2723 s64 period, sample_period;
2726 period = perf_calculate_period(event, nsec, count);
2728 delta = (s64)(period - hwc->sample_period);
2729 delta = (delta + 7) / 8; /* low pass filter */
2731 sample_period = hwc->sample_period + delta;
2736 hwc->sample_period = sample_period;
2738 if (local64_read(&hwc->period_left) > 8*sample_period) {
2740 event->pmu->stop(event, PERF_EF_UPDATE);
2742 local64_set(&hwc->period_left, 0);
2745 event->pmu->start(event, PERF_EF_RELOAD);
2750 * combine freq adjustment with unthrottling to avoid two passes over the
2751 * events. At the same time, make sure, having freq events does not change
2752 * the rate of unthrottling as that would introduce bias.
2754 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2757 struct perf_event *event;
2758 struct hw_perf_event *hwc;
2759 u64 now, period = TICK_NSEC;
2763 * only need to iterate over all events iff:
2764 * - context have events in frequency mode (needs freq adjust)
2765 * - there are events to unthrottle on this cpu
2767 if (!(ctx->nr_freq || needs_unthr))
2770 raw_spin_lock(&ctx->lock);
2771 perf_pmu_disable(ctx->pmu);
2773 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2774 if (event->state != PERF_EVENT_STATE_ACTIVE)
2777 if (!event_filter_match(event))
2780 perf_pmu_disable(event->pmu);
2784 if (hwc->interrupts == MAX_INTERRUPTS) {
2785 hwc->interrupts = 0;
2786 perf_log_throttle(event, 1);
2787 event->pmu->start(event, 0);
2790 if (!event->attr.freq || !event->attr.sample_freq)
2794 * stop the event and update event->count
2796 event->pmu->stop(event, PERF_EF_UPDATE);
2798 now = local64_read(&event->count);
2799 delta = now - hwc->freq_count_stamp;
2800 hwc->freq_count_stamp = now;
2804 * reload only if value has changed
2805 * we have stopped the event so tell that
2806 * to perf_adjust_period() to avoid stopping it
2810 perf_adjust_period(event, period, delta, false);
2812 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2814 perf_pmu_enable(event->pmu);
2817 perf_pmu_enable(ctx->pmu);
2818 raw_spin_unlock(&ctx->lock);
2822 * Round-robin a context's events:
2824 static void rotate_ctx(struct perf_event_context *ctx)
2827 * Rotate the first entry last of non-pinned groups. Rotation might be
2828 * disabled by the inheritance code.
2830 if (!ctx->rotate_disable)
2831 list_rotate_left(&ctx->flexible_groups);
2835 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2836 * because they're strictly cpu affine and rotate_start is called with IRQs
2837 * disabled, while rotate_context is called from IRQ context.
2839 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2841 struct perf_event_context *ctx = NULL;
2842 int rotate = 0, remove = 1;
2844 if (cpuctx->ctx.nr_events) {
2846 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2850 ctx = cpuctx->task_ctx;
2851 if (ctx && ctx->nr_events) {
2853 if (ctx->nr_events != ctx->nr_active)
2860 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2861 perf_pmu_disable(cpuctx->ctx.pmu);
2863 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2865 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2867 rotate_ctx(&cpuctx->ctx);
2871 perf_event_sched_in(cpuctx, ctx, current);
2873 perf_pmu_enable(cpuctx->ctx.pmu);
2874 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2877 list_del_init(&cpuctx->rotation_list);
2882 #ifdef CONFIG_NO_HZ_FULL
2883 bool perf_event_can_stop_tick(void)
2885 if (atomic_read(&nr_freq_events) ||
2886 __this_cpu_read(perf_throttled_count))
2893 void perf_event_task_tick(void)
2895 struct list_head *head = &__get_cpu_var(rotation_list);
2896 struct perf_cpu_context *cpuctx, *tmp;
2897 struct perf_event_context *ctx;
2900 WARN_ON(!irqs_disabled());
2902 __this_cpu_inc(perf_throttled_seq);
2903 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2905 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2907 perf_adjust_freq_unthr_context(ctx, throttled);
2909 ctx = cpuctx->task_ctx;
2911 perf_adjust_freq_unthr_context(ctx, throttled);
2915 static int event_enable_on_exec(struct perf_event *event,
2916 struct perf_event_context *ctx)
2918 if (!event->attr.enable_on_exec)
2921 event->attr.enable_on_exec = 0;
2922 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2925 __perf_event_mark_enabled(event);
2931 * Enable all of a task's events that have been marked enable-on-exec.
2932 * This expects task == current.
2934 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2936 struct perf_event *event;
2937 unsigned long flags;
2941 local_irq_save(flags);
2942 if (!ctx || !ctx->nr_events)
2946 * We must ctxsw out cgroup events to avoid conflict
2947 * when invoking perf_task_event_sched_in() later on
2948 * in this function. Otherwise we end up trying to
2949 * ctxswin cgroup events which are already scheduled
2952 perf_cgroup_sched_out(current, NULL);
2954 raw_spin_lock(&ctx->lock);
2955 task_ctx_sched_out(ctx);
2957 list_for_each_entry(event, &ctx->event_list, event_entry) {
2958 ret = event_enable_on_exec(event, ctx);
2964 * Unclone this context if we enabled any event.
2969 raw_spin_unlock(&ctx->lock);
2972 * Also calls ctxswin for cgroup events, if any:
2974 perf_event_context_sched_in(ctx, ctx->task);
2976 local_irq_restore(flags);
2979 void perf_event_exec(void)
2981 struct perf_event_context *ctx;
2985 for_each_task_context_nr(ctxn) {
2986 ctx = current->perf_event_ctxp[ctxn];
2990 perf_event_enable_on_exec(ctx);
2996 * Cross CPU call to read the hardware event
2998 static void __perf_event_read(void *info)
3000 struct perf_event *event = info;
3001 struct perf_event_context *ctx = event->ctx;
3002 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3005 * If this is a task context, we need to check whether it is
3006 * the current task context of this cpu. If not it has been
3007 * scheduled out before the smp call arrived. In that case
3008 * event->count would have been updated to a recent sample
3009 * when the event was scheduled out.
3011 if (ctx->task && cpuctx->task_ctx != ctx)
3014 raw_spin_lock(&ctx->lock);
3015 if (ctx->is_active) {
3016 update_context_time(ctx);
3017 update_cgrp_time_from_event(event);
3019 update_event_times(event);
3020 if (event->state == PERF_EVENT_STATE_ACTIVE)
3021 event->pmu->read(event);
3022 raw_spin_unlock(&ctx->lock);
3025 static inline u64 perf_event_count(struct perf_event *event)
3027 return local64_read(&event->count) + atomic64_read(&event->child_count);
3030 static u64 perf_event_read(struct perf_event *event)
3033 * If event is enabled and currently active on a CPU, update the
3034 * value in the event structure:
3036 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3037 smp_call_function_single(event->oncpu,
3038 __perf_event_read, event, 1);
3039 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3040 struct perf_event_context *ctx = event->ctx;
3041 unsigned long flags;
3043 raw_spin_lock_irqsave(&ctx->lock, flags);
3045 * may read while context is not active
3046 * (e.g., thread is blocked), in that case
3047 * we cannot update context time
3049 if (ctx->is_active) {
3050 update_context_time(ctx);
3051 update_cgrp_time_from_event(event);
3053 update_event_times(event);
3054 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3057 return perf_event_count(event);
3061 * Initialize the perf_event context in a task_struct:
3063 static void __perf_event_init_context(struct perf_event_context *ctx)
3065 raw_spin_lock_init(&ctx->lock);
3066 mutex_init(&ctx->mutex);
3067 INIT_LIST_HEAD(&ctx->pinned_groups);
3068 INIT_LIST_HEAD(&ctx->flexible_groups);
3069 INIT_LIST_HEAD(&ctx->event_list);
3070 atomic_set(&ctx->refcount, 1);
3073 static struct perf_event_context *
3074 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3076 struct perf_event_context *ctx;
3078 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3082 __perf_event_init_context(ctx);
3085 get_task_struct(task);
3092 static struct task_struct *
3093 find_lively_task_by_vpid(pid_t vpid)
3095 struct task_struct *task;
3102 task = find_task_by_vpid(vpid);
3104 get_task_struct(task);
3108 return ERR_PTR(-ESRCH);
3110 /* Reuse ptrace permission checks for now. */
3112 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3117 put_task_struct(task);
3118 return ERR_PTR(err);
3123 * Returns a matching context with refcount and pincount.
3125 static struct perf_event_context *
3126 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3128 struct perf_event_context *ctx;
3129 struct perf_cpu_context *cpuctx;
3130 unsigned long flags;
3134 /* Must be root to operate on a CPU event: */
3135 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3136 return ERR_PTR(-EACCES);
3139 * We could be clever and allow to attach a event to an
3140 * offline CPU and activate it when the CPU comes up, but
3143 if (!cpu_online(cpu))
3144 return ERR_PTR(-ENODEV);
3146 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3155 ctxn = pmu->task_ctx_nr;
3160 ctx = perf_lock_task_context(task, ctxn, &flags);
3164 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3166 ctx = alloc_perf_context(pmu, task);
3172 mutex_lock(&task->perf_event_mutex);
3174 * If it has already passed perf_event_exit_task().
3175 * we must see PF_EXITING, it takes this mutex too.
3177 if (task->flags & PF_EXITING)
3179 else if (task->perf_event_ctxp[ctxn])
3184 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3186 mutex_unlock(&task->perf_event_mutex);
3188 if (unlikely(err)) {
3200 return ERR_PTR(err);
3203 static void perf_event_free_filter(struct perf_event *event);
3205 static void free_event_rcu(struct rcu_head *head)
3207 struct perf_event *event;
3209 event = container_of(head, struct perf_event, rcu_head);
3211 put_pid_ns(event->ns);
3212 perf_event_free_filter(event);
3216 static void ring_buffer_put(struct ring_buffer *rb);
3217 static void ring_buffer_attach(struct perf_event *event,
3218 struct ring_buffer *rb);
3220 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3225 if (has_branch_stack(event)) {
3226 if (!(event->attach_state & PERF_ATTACH_TASK))
3227 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3229 if (is_cgroup_event(event))
3230 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3233 static void unaccount_event(struct perf_event *event)
3238 if (event->attach_state & PERF_ATTACH_TASK)
3239 static_key_slow_dec_deferred(&perf_sched_events);
3240 if (event->attr.mmap || event->attr.mmap_data)
3241 atomic_dec(&nr_mmap_events);
3242 if (event->attr.comm)
3243 atomic_dec(&nr_comm_events);
3244 if (event->attr.task)
3245 atomic_dec(&nr_task_events);
3246 if (event->attr.freq)
3247 atomic_dec(&nr_freq_events);
3248 if (is_cgroup_event(event))
3249 static_key_slow_dec_deferred(&perf_sched_events);
3250 if (has_branch_stack(event))
3251 static_key_slow_dec_deferred(&perf_sched_events);
3253 unaccount_event_cpu(event, event->cpu);
3256 static void __free_event(struct perf_event *event)
3258 if (!event->parent) {
3259 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3260 put_callchain_buffers();
3264 event->destroy(event);
3267 put_ctx(event->ctx);
3270 module_put(event->pmu->module);
3272 call_rcu(&event->rcu_head, free_event_rcu);
3275 static void _free_event(struct perf_event *event)
3277 irq_work_sync(&event->pending);
3279 unaccount_event(event);
3283 * Can happen when we close an event with re-directed output.
3285 * Since we have a 0 refcount, perf_mmap_close() will skip
3286 * over us; possibly making our ring_buffer_put() the last.
3288 mutex_lock(&event->mmap_mutex);
3289 ring_buffer_attach(event, NULL);
3290 mutex_unlock(&event->mmap_mutex);
3293 if (is_cgroup_event(event))
3294 perf_detach_cgroup(event);
3296 __free_event(event);
3300 * Used to free events which have a known refcount of 1, such as in error paths
3301 * where the event isn't exposed yet and inherited events.
3303 static void free_event(struct perf_event *event)
3305 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3306 "unexpected event refcount: %ld; ptr=%p\n",
3307 atomic_long_read(&event->refcount), event)) {
3308 /* leak to avoid use-after-free */
3316 * Called when the last reference to the file is gone.
3318 static void put_event(struct perf_event *event)
3320 struct perf_event_context *ctx = event->ctx;
3321 struct task_struct *owner;
3323 if (!atomic_long_dec_and_test(&event->refcount))
3327 owner = ACCESS_ONCE(event->owner);
3329 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3330 * !owner it means the list deletion is complete and we can indeed
3331 * free this event, otherwise we need to serialize on
3332 * owner->perf_event_mutex.
3334 smp_read_barrier_depends();
3337 * Since delayed_put_task_struct() also drops the last
3338 * task reference we can safely take a new reference
3339 * while holding the rcu_read_lock().
3341 get_task_struct(owner);
3346 mutex_lock(&owner->perf_event_mutex);
3348 * We have to re-check the event->owner field, if it is cleared
3349 * we raced with perf_event_exit_task(), acquiring the mutex
3350 * ensured they're done, and we can proceed with freeing the
3354 list_del_init(&event->owner_entry);
3355 mutex_unlock(&owner->perf_event_mutex);
3356 put_task_struct(owner);
3359 WARN_ON_ONCE(ctx->parent_ctx);
3361 * There are two ways this annotation is useful:
3363 * 1) there is a lock recursion from perf_event_exit_task
3364 * see the comment there.
3366 * 2) there is a lock-inversion with mmap_sem through
3367 * perf_event_read_group(), which takes faults while
3368 * holding ctx->mutex, however this is called after
3369 * the last filedesc died, so there is no possibility
3370 * to trigger the AB-BA case.
3372 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3373 perf_remove_from_context(event, true);
3374 mutex_unlock(&ctx->mutex);
3379 int perf_event_release_kernel(struct perf_event *event)
3384 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3386 static int perf_release(struct inode *inode, struct file *file)
3388 put_event(file->private_data);
3392 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3394 struct perf_event *child;
3400 mutex_lock(&event->child_mutex);
3401 total += perf_event_read(event);
3402 *enabled += event->total_time_enabled +
3403 atomic64_read(&event->child_total_time_enabled);
3404 *running += event->total_time_running +
3405 atomic64_read(&event->child_total_time_running);
3407 list_for_each_entry(child, &event->child_list, child_list) {
3408 total += perf_event_read(child);
3409 *enabled += child->total_time_enabled;
3410 *running += child->total_time_running;
3412 mutex_unlock(&event->child_mutex);
3416 EXPORT_SYMBOL_GPL(perf_event_read_value);
3418 static int perf_event_read_group(struct perf_event *event,
3419 u64 read_format, char __user *buf)
3421 struct perf_event *leader = event->group_leader, *sub;
3422 int n = 0, size = 0, ret = -EFAULT;
3423 struct perf_event_context *ctx = leader->ctx;
3425 u64 count, enabled, running;
3427 mutex_lock(&ctx->mutex);
3428 count = perf_event_read_value(leader, &enabled, &running);
3430 values[n++] = 1 + leader->nr_siblings;
3431 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3432 values[n++] = enabled;
3433 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3434 values[n++] = running;
3435 values[n++] = count;
3436 if (read_format & PERF_FORMAT_ID)
3437 values[n++] = primary_event_id(leader);
3439 size = n * sizeof(u64);
3441 if (copy_to_user(buf, values, size))
3446 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3449 values[n++] = perf_event_read_value(sub, &enabled, &running);
3450 if (read_format & PERF_FORMAT_ID)
3451 values[n++] = primary_event_id(sub);
3453 size = n * sizeof(u64);
3455 if (copy_to_user(buf + ret, values, size)) {
3463 mutex_unlock(&ctx->mutex);
3468 static int perf_event_read_one(struct perf_event *event,
3469 u64 read_format, char __user *buf)
3471 u64 enabled, running;
3475 values[n++] = perf_event_read_value(event, &enabled, &running);
3476 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3477 values[n++] = enabled;
3478 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3479 values[n++] = running;
3480 if (read_format & PERF_FORMAT_ID)
3481 values[n++] = primary_event_id(event);
3483 if (copy_to_user(buf, values, n * sizeof(u64)))
3486 return n * sizeof(u64);
3490 * Read the performance event - simple non blocking version for now
3493 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3495 u64 read_format = event->attr.read_format;
3499 * Return end-of-file for a read on a event that is in
3500 * error state (i.e. because it was pinned but it couldn't be
3501 * scheduled on to the CPU at some point).
3503 if (event->state == PERF_EVENT_STATE_ERROR)
3506 if (count < event->read_size)
3509 WARN_ON_ONCE(event->ctx->parent_ctx);
3510 if (read_format & PERF_FORMAT_GROUP)
3511 ret = perf_event_read_group(event, read_format, buf);
3513 ret = perf_event_read_one(event, read_format, buf);
3519 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3521 struct perf_event *event = file->private_data;
3523 return perf_read_hw(event, buf, count);
3526 static unsigned int perf_poll(struct file *file, poll_table *wait)
3528 struct perf_event *event = file->private_data;
3529 struct ring_buffer *rb;
3530 unsigned int events = POLL_HUP;
3533 * Pin the event->rb by taking event->mmap_mutex; otherwise
3534 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3536 mutex_lock(&event->mmap_mutex);
3539 events = atomic_xchg(&rb->poll, 0);
3540 mutex_unlock(&event->mmap_mutex);
3542 poll_wait(file, &event->waitq, wait);
3547 static void perf_event_reset(struct perf_event *event)
3549 (void)perf_event_read(event);
3550 local64_set(&event->count, 0);
3551 perf_event_update_userpage(event);
3555 * Holding the top-level event's child_mutex means that any
3556 * descendant process that has inherited this event will block
3557 * in sync_child_event if it goes to exit, thus satisfying the
3558 * task existence requirements of perf_event_enable/disable.
3560 static void perf_event_for_each_child(struct perf_event *event,
3561 void (*func)(struct perf_event *))
3563 struct perf_event *child;
3565 WARN_ON_ONCE(event->ctx->parent_ctx);
3566 mutex_lock(&event->child_mutex);
3568 list_for_each_entry(child, &event->child_list, child_list)
3570 mutex_unlock(&event->child_mutex);
3573 static void perf_event_for_each(struct perf_event *event,
3574 void (*func)(struct perf_event *))
3576 struct perf_event_context *ctx = event->ctx;
3577 struct perf_event *sibling;
3579 WARN_ON_ONCE(ctx->parent_ctx);
3580 mutex_lock(&ctx->mutex);
3581 event = event->group_leader;
3583 perf_event_for_each_child(event, func);
3584 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3585 perf_event_for_each_child(sibling, func);
3586 mutex_unlock(&ctx->mutex);
3589 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3591 struct perf_event_context *ctx = event->ctx;
3592 int ret = 0, active;
3595 if (!is_sampling_event(event))
3598 if (copy_from_user(&value, arg, sizeof(value)))
3604 raw_spin_lock_irq(&ctx->lock);
3605 if (event->attr.freq) {
3606 if (value > sysctl_perf_event_sample_rate) {
3611 event->attr.sample_freq = value;
3613 event->attr.sample_period = value;
3614 event->hw.sample_period = value;
3617 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3619 perf_pmu_disable(ctx->pmu);
3620 event->pmu->stop(event, PERF_EF_UPDATE);
3623 local64_set(&event->hw.period_left, 0);
3626 event->pmu->start(event, PERF_EF_RELOAD);
3627 perf_pmu_enable(ctx->pmu);
3631 raw_spin_unlock_irq(&ctx->lock);
3636 static const struct file_operations perf_fops;
3638 static inline int perf_fget_light(int fd, struct fd *p)
3640 struct fd f = fdget(fd);
3644 if (f.file->f_op != &perf_fops) {
3652 static int perf_event_set_output(struct perf_event *event,
3653 struct perf_event *output_event);
3654 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3656 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3658 struct perf_event *event = file->private_data;
3659 void (*func)(struct perf_event *);
3663 case PERF_EVENT_IOC_ENABLE:
3664 func = perf_event_enable;
3666 case PERF_EVENT_IOC_DISABLE:
3667 func = perf_event_disable;
3669 case PERF_EVENT_IOC_RESET:
3670 func = perf_event_reset;
3673 case PERF_EVENT_IOC_REFRESH:
3674 return perf_event_refresh(event, arg);
3676 case PERF_EVENT_IOC_PERIOD:
3677 return perf_event_period(event, (u64 __user *)arg);
3679 case PERF_EVENT_IOC_ID:
3681 u64 id = primary_event_id(event);
3683 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3688 case PERF_EVENT_IOC_SET_OUTPUT:
3692 struct perf_event *output_event;
3694 ret = perf_fget_light(arg, &output);
3697 output_event = output.file->private_data;
3698 ret = perf_event_set_output(event, output_event);
3701 ret = perf_event_set_output(event, NULL);
3706 case PERF_EVENT_IOC_SET_FILTER:
3707 return perf_event_set_filter(event, (void __user *)arg);
3713 if (flags & PERF_IOC_FLAG_GROUP)
3714 perf_event_for_each(event, func);
3716 perf_event_for_each_child(event, func);
3721 #ifdef CONFIG_COMPAT
3722 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
3725 switch (_IOC_NR(cmd)) {
3726 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
3727 case _IOC_NR(PERF_EVENT_IOC_ID):
3728 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
3729 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
3730 cmd &= ~IOCSIZE_MASK;
3731 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
3735 return perf_ioctl(file, cmd, arg);
3738 # define perf_compat_ioctl NULL
3741 int perf_event_task_enable(void)
3743 struct perf_event *event;
3745 mutex_lock(¤t->perf_event_mutex);
3746 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3747 perf_event_for_each_child(event, perf_event_enable);
3748 mutex_unlock(¤t->perf_event_mutex);
3753 int perf_event_task_disable(void)
3755 struct perf_event *event;
3757 mutex_lock(¤t->perf_event_mutex);
3758 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3759 perf_event_for_each_child(event, perf_event_disable);
3760 mutex_unlock(¤t->perf_event_mutex);
3765 static int perf_event_index(struct perf_event *event)
3767 if (event->hw.state & PERF_HES_STOPPED)
3770 if (event->state != PERF_EVENT_STATE_ACTIVE)
3773 return event->pmu->event_idx(event);
3776 static void calc_timer_values(struct perf_event *event,
3783 *now = perf_clock();
3784 ctx_time = event->shadow_ctx_time + *now;
3785 *enabled = ctx_time - event->tstamp_enabled;
3786 *running = ctx_time - event->tstamp_running;
3789 static void perf_event_init_userpage(struct perf_event *event)
3791 struct perf_event_mmap_page *userpg;
3792 struct ring_buffer *rb;
3795 rb = rcu_dereference(event->rb);
3799 userpg = rb->user_page;
3801 /* Allow new userspace to detect that bit 0 is deprecated */
3802 userpg->cap_bit0_is_deprecated = 1;
3803 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3809 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3814 * Callers need to ensure there can be no nesting of this function, otherwise
3815 * the seqlock logic goes bad. We can not serialize this because the arch
3816 * code calls this from NMI context.
3818 void perf_event_update_userpage(struct perf_event *event)
3820 struct perf_event_mmap_page *userpg;
3821 struct ring_buffer *rb;
3822 u64 enabled, running, now;
3825 rb = rcu_dereference(event->rb);
3830 * compute total_time_enabled, total_time_running
3831 * based on snapshot values taken when the event
3832 * was last scheduled in.
3834 * we cannot simply called update_context_time()
3835 * because of locking issue as we can be called in
3838 calc_timer_values(event, &now, &enabled, &running);
3840 userpg = rb->user_page;
3842 * Disable preemption so as to not let the corresponding user-space
3843 * spin too long if we get preempted.
3848 userpg->index = perf_event_index(event);
3849 userpg->offset = perf_event_count(event);
3851 userpg->offset -= local64_read(&event->hw.prev_count);
3853 userpg->time_enabled = enabled +
3854 atomic64_read(&event->child_total_time_enabled);
3856 userpg->time_running = running +
3857 atomic64_read(&event->child_total_time_running);
3859 arch_perf_update_userpage(userpg, now);
3868 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3870 struct perf_event *event = vma->vm_file->private_data;
3871 struct ring_buffer *rb;
3872 int ret = VM_FAULT_SIGBUS;
3874 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3875 if (vmf->pgoff == 0)
3881 rb = rcu_dereference(event->rb);
3885 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3888 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3892 get_page(vmf->page);
3893 vmf->page->mapping = vma->vm_file->f_mapping;
3894 vmf->page->index = vmf->pgoff;
3903 static void ring_buffer_attach(struct perf_event *event,
3904 struct ring_buffer *rb)
3906 struct ring_buffer *old_rb = NULL;
3907 unsigned long flags;
3911 * Should be impossible, we set this when removing
3912 * event->rb_entry and wait/clear when adding event->rb_entry.
3914 WARN_ON_ONCE(event->rcu_pending);
3917 event->rcu_batches = get_state_synchronize_rcu();
3918 event->rcu_pending = 1;
3920 spin_lock_irqsave(&old_rb->event_lock, flags);
3921 list_del_rcu(&event->rb_entry);
3922 spin_unlock_irqrestore(&old_rb->event_lock, flags);
3925 if (event->rcu_pending && rb) {
3926 cond_synchronize_rcu(event->rcu_batches);
3927 event->rcu_pending = 0;
3931 spin_lock_irqsave(&rb->event_lock, flags);
3932 list_add_rcu(&event->rb_entry, &rb->event_list);
3933 spin_unlock_irqrestore(&rb->event_lock, flags);
3936 rcu_assign_pointer(event->rb, rb);
3939 ring_buffer_put(old_rb);
3941 * Since we detached before setting the new rb, so that we
3942 * could attach the new rb, we could have missed a wakeup.
3945 wake_up_all(&event->waitq);
3949 static void ring_buffer_wakeup(struct perf_event *event)
3951 struct ring_buffer *rb;
3954 rb = rcu_dereference(event->rb);
3956 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3957 wake_up_all(&event->waitq);
3962 static void rb_free_rcu(struct rcu_head *rcu_head)
3964 struct ring_buffer *rb;
3966 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3970 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3972 struct ring_buffer *rb;
3975 rb = rcu_dereference(event->rb);
3977 if (!atomic_inc_not_zero(&rb->refcount))
3985 static void ring_buffer_put(struct ring_buffer *rb)
3987 if (!atomic_dec_and_test(&rb->refcount))
3990 WARN_ON_ONCE(!list_empty(&rb->event_list));
3992 call_rcu(&rb->rcu_head, rb_free_rcu);
3995 static void perf_mmap_open(struct vm_area_struct *vma)
3997 struct perf_event *event = vma->vm_file->private_data;
3999 atomic_inc(&event->mmap_count);
4000 atomic_inc(&event->rb->mmap_count);
4004 * A buffer can be mmap()ed multiple times; either directly through the same
4005 * event, or through other events by use of perf_event_set_output().
4007 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4008 * the buffer here, where we still have a VM context. This means we need
4009 * to detach all events redirecting to us.
4011 static void perf_mmap_close(struct vm_area_struct *vma)
4013 struct perf_event *event = vma->vm_file->private_data;
4015 struct ring_buffer *rb = ring_buffer_get(event);
4016 struct user_struct *mmap_user = rb->mmap_user;
4017 int mmap_locked = rb->mmap_locked;
4018 unsigned long size = perf_data_size(rb);
4020 atomic_dec(&rb->mmap_count);
4022 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4025 ring_buffer_attach(event, NULL);
4026 mutex_unlock(&event->mmap_mutex);
4028 /* If there's still other mmap()s of this buffer, we're done. */
4029 if (atomic_read(&rb->mmap_count))
4033 * No other mmap()s, detach from all other events that might redirect
4034 * into the now unreachable buffer. Somewhat complicated by the
4035 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4039 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4040 if (!atomic_long_inc_not_zero(&event->refcount)) {
4042 * This event is en-route to free_event() which will
4043 * detach it and remove it from the list.
4049 mutex_lock(&event->mmap_mutex);
4051 * Check we didn't race with perf_event_set_output() which can
4052 * swizzle the rb from under us while we were waiting to
4053 * acquire mmap_mutex.
4055 * If we find a different rb; ignore this event, a next
4056 * iteration will no longer find it on the list. We have to
4057 * still restart the iteration to make sure we're not now
4058 * iterating the wrong list.
4060 if (event->rb == rb)
4061 ring_buffer_attach(event, NULL);
4063 mutex_unlock(&event->mmap_mutex);
4067 * Restart the iteration; either we're on the wrong list or
4068 * destroyed its integrity by doing a deletion.
4075 * It could be there's still a few 0-ref events on the list; they'll
4076 * get cleaned up by free_event() -- they'll also still have their
4077 * ref on the rb and will free it whenever they are done with it.
4079 * Aside from that, this buffer is 'fully' detached and unmapped,
4080 * undo the VM accounting.
4083 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4084 vma->vm_mm->pinned_vm -= mmap_locked;
4085 free_uid(mmap_user);
4088 ring_buffer_put(rb); /* could be last */
4091 static const struct vm_operations_struct perf_mmap_vmops = {
4092 .open = perf_mmap_open,
4093 .close = perf_mmap_close,
4094 .fault = perf_mmap_fault,
4095 .page_mkwrite = perf_mmap_fault,
4098 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4100 struct perf_event *event = file->private_data;
4101 unsigned long user_locked, user_lock_limit;
4102 struct user_struct *user = current_user();
4103 unsigned long locked, lock_limit;
4104 struct ring_buffer *rb;
4105 unsigned long vma_size;
4106 unsigned long nr_pages;
4107 long user_extra, extra;
4108 int ret = 0, flags = 0;
4111 * Don't allow mmap() of inherited per-task counters. This would
4112 * create a performance issue due to all children writing to the
4115 if (event->cpu == -1 && event->attr.inherit)
4118 if (!(vma->vm_flags & VM_SHARED))
4121 vma_size = vma->vm_end - vma->vm_start;
4122 nr_pages = (vma_size / PAGE_SIZE) - 1;
4125 * If we have rb pages ensure they're a power-of-two number, so we
4126 * can do bitmasks instead of modulo.
4128 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4131 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4134 if (vma->vm_pgoff != 0)
4137 WARN_ON_ONCE(event->ctx->parent_ctx);
4139 mutex_lock(&event->mmap_mutex);
4141 if (event->rb->nr_pages != nr_pages) {
4146 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4148 * Raced against perf_mmap_close() through
4149 * perf_event_set_output(). Try again, hope for better
4152 mutex_unlock(&event->mmap_mutex);
4159 user_extra = nr_pages + 1;
4160 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4163 * Increase the limit linearly with more CPUs:
4165 user_lock_limit *= num_online_cpus();
4167 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4170 if (user_locked > user_lock_limit)
4171 extra = user_locked - user_lock_limit;
4173 lock_limit = rlimit(RLIMIT_MEMLOCK);
4174 lock_limit >>= PAGE_SHIFT;
4175 locked = vma->vm_mm->pinned_vm + extra;
4177 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4178 !capable(CAP_IPC_LOCK)) {
4185 if (vma->vm_flags & VM_WRITE)
4186 flags |= RING_BUFFER_WRITABLE;
4188 rb = rb_alloc(nr_pages,
4189 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4197 atomic_set(&rb->mmap_count, 1);
4198 rb->mmap_locked = extra;
4199 rb->mmap_user = get_current_user();
4201 atomic_long_add(user_extra, &user->locked_vm);
4202 vma->vm_mm->pinned_vm += extra;
4204 ring_buffer_attach(event, rb);
4206 perf_event_init_userpage(event);
4207 perf_event_update_userpage(event);
4211 atomic_inc(&event->mmap_count);
4212 mutex_unlock(&event->mmap_mutex);
4215 * Since pinned accounting is per vm we cannot allow fork() to copy our
4218 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4219 vma->vm_ops = &perf_mmap_vmops;
4224 static int perf_fasync(int fd, struct file *filp, int on)
4226 struct inode *inode = file_inode(filp);
4227 struct perf_event *event = filp->private_data;
4230 mutex_lock(&inode->i_mutex);
4231 retval = fasync_helper(fd, filp, on, &event->fasync);
4232 mutex_unlock(&inode->i_mutex);
4240 static const struct file_operations perf_fops = {
4241 .llseek = no_llseek,
4242 .release = perf_release,
4245 .unlocked_ioctl = perf_ioctl,
4246 .compat_ioctl = perf_compat_ioctl,
4248 .fasync = perf_fasync,
4254 * If there's data, ensure we set the poll() state and publish everything
4255 * to user-space before waking everybody up.
4258 void perf_event_wakeup(struct perf_event *event)
4260 ring_buffer_wakeup(event);
4262 if (event->pending_kill) {
4263 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4264 event->pending_kill = 0;
4268 static void perf_pending_event(struct irq_work *entry)
4270 struct perf_event *event = container_of(entry,
4271 struct perf_event, pending);
4273 if (event->pending_disable) {
4274 event->pending_disable = 0;
4275 __perf_event_disable(event);
4278 if (event->pending_wakeup) {
4279 event->pending_wakeup = 0;
4280 perf_event_wakeup(event);
4285 * We assume there is only KVM supporting the callbacks.
4286 * Later on, we might change it to a list if there is
4287 * another virtualization implementation supporting the callbacks.
4289 struct perf_guest_info_callbacks *perf_guest_cbs;
4291 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4293 perf_guest_cbs = cbs;
4296 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4298 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4300 perf_guest_cbs = NULL;
4303 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4306 perf_output_sample_regs(struct perf_output_handle *handle,
4307 struct pt_regs *regs, u64 mask)
4311 for_each_set_bit(bit, (const unsigned long *) &mask,
4312 sizeof(mask) * BITS_PER_BYTE) {
4315 val = perf_reg_value(regs, bit);
4316 perf_output_put(handle, val);
4320 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4321 struct pt_regs *regs)
4323 if (!user_mode(regs)) {
4325 regs = task_pt_regs(current);
4331 regs_user->regs = regs;
4332 regs_user->abi = perf_reg_abi(current);
4337 * Get remaining task size from user stack pointer.
4339 * It'd be better to take stack vma map and limit this more
4340 * precisly, but there's no way to get it safely under interrupt,
4341 * so using TASK_SIZE as limit.
4343 static u64 perf_ustack_task_size(struct pt_regs *regs)
4345 unsigned long addr = perf_user_stack_pointer(regs);
4347 if (!addr || addr >= TASK_SIZE)
4350 return TASK_SIZE - addr;
4354 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4355 struct pt_regs *regs)
4359 /* No regs, no stack pointer, no dump. */
4364 * Check if we fit in with the requested stack size into the:
4366 * If we don't, we limit the size to the TASK_SIZE.
4368 * - remaining sample size
4369 * If we don't, we customize the stack size to
4370 * fit in to the remaining sample size.
4373 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4374 stack_size = min(stack_size, (u16) task_size);
4376 /* Current header size plus static size and dynamic size. */
4377 header_size += 2 * sizeof(u64);
4379 /* Do we fit in with the current stack dump size? */
4380 if ((u16) (header_size + stack_size) < header_size) {
4382 * If we overflow the maximum size for the sample,
4383 * we customize the stack dump size to fit in.
4385 stack_size = USHRT_MAX - header_size - sizeof(u64);
4386 stack_size = round_up(stack_size, sizeof(u64));
4393 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4394 struct pt_regs *regs)
4396 /* Case of a kernel thread, nothing to dump */
4399 perf_output_put(handle, size);
4408 * - the size requested by user or the best one we can fit
4409 * in to the sample max size
4411 * - user stack dump data
4413 * - the actual dumped size
4417 perf_output_put(handle, dump_size);
4420 sp = perf_user_stack_pointer(regs);
4421 rem = __output_copy_user(handle, (void *) sp, dump_size);
4422 dyn_size = dump_size - rem;
4424 perf_output_skip(handle, rem);
4427 perf_output_put(handle, dyn_size);
4431 static void __perf_event_header__init_id(struct perf_event_header *header,
4432 struct perf_sample_data *data,
4433 struct perf_event *event)
4435 u64 sample_type = event->attr.sample_type;
4437 data->type = sample_type;
4438 header->size += event->id_header_size;
4440 if (sample_type & PERF_SAMPLE_TID) {
4441 /* namespace issues */
4442 data->tid_entry.pid = perf_event_pid(event, current);
4443 data->tid_entry.tid = perf_event_tid(event, current);
4446 if (sample_type & PERF_SAMPLE_TIME)
4447 data->time = perf_clock();
4449 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4450 data->id = primary_event_id(event);
4452 if (sample_type & PERF_SAMPLE_STREAM_ID)
4453 data->stream_id = event->id;
4455 if (sample_type & PERF_SAMPLE_CPU) {
4456 data->cpu_entry.cpu = raw_smp_processor_id();
4457 data->cpu_entry.reserved = 0;
4461 void perf_event_header__init_id(struct perf_event_header *header,
4462 struct perf_sample_data *data,
4463 struct perf_event *event)
4465 if (event->attr.sample_id_all)
4466 __perf_event_header__init_id(header, data, event);
4469 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4470 struct perf_sample_data *data)
4472 u64 sample_type = data->type;
4474 if (sample_type & PERF_SAMPLE_TID)
4475 perf_output_put(handle, data->tid_entry);
4477 if (sample_type & PERF_SAMPLE_TIME)
4478 perf_output_put(handle, data->time);
4480 if (sample_type & PERF_SAMPLE_ID)
4481 perf_output_put(handle, data->id);
4483 if (sample_type & PERF_SAMPLE_STREAM_ID)
4484 perf_output_put(handle, data->stream_id);
4486 if (sample_type & PERF_SAMPLE_CPU)
4487 perf_output_put(handle, data->cpu_entry);
4489 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4490 perf_output_put(handle, data->id);
4493 void perf_event__output_id_sample(struct perf_event *event,
4494 struct perf_output_handle *handle,
4495 struct perf_sample_data *sample)
4497 if (event->attr.sample_id_all)
4498 __perf_event__output_id_sample(handle, sample);
4501 static void perf_output_read_one(struct perf_output_handle *handle,
4502 struct perf_event *event,
4503 u64 enabled, u64 running)
4505 u64 read_format = event->attr.read_format;
4509 values[n++] = perf_event_count(event);
4510 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4511 values[n++] = enabled +
4512 atomic64_read(&event->child_total_time_enabled);
4514 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4515 values[n++] = running +
4516 atomic64_read(&event->child_total_time_running);
4518 if (read_format & PERF_FORMAT_ID)
4519 values[n++] = primary_event_id(event);
4521 __output_copy(handle, values, n * sizeof(u64));
4525 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4527 static void perf_output_read_group(struct perf_output_handle *handle,
4528 struct perf_event *event,
4529 u64 enabled, u64 running)
4531 struct perf_event *leader = event->group_leader, *sub;
4532 u64 read_format = event->attr.read_format;
4536 values[n++] = 1 + leader->nr_siblings;
4538 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4539 values[n++] = enabled;
4541 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4542 values[n++] = running;
4544 if (leader != event)
4545 leader->pmu->read(leader);
4547 values[n++] = perf_event_count(leader);
4548 if (read_format & PERF_FORMAT_ID)
4549 values[n++] = primary_event_id(leader);
4551 __output_copy(handle, values, n * sizeof(u64));
4553 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4556 if ((sub != event) &&
4557 (sub->state == PERF_EVENT_STATE_ACTIVE))
4558 sub->pmu->read(sub);
4560 values[n++] = perf_event_count(sub);
4561 if (read_format & PERF_FORMAT_ID)
4562 values[n++] = primary_event_id(sub);
4564 __output_copy(handle, values, n * sizeof(u64));
4568 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4569 PERF_FORMAT_TOTAL_TIME_RUNNING)
4571 static void perf_output_read(struct perf_output_handle *handle,
4572 struct perf_event *event)
4574 u64 enabled = 0, running = 0, now;
4575 u64 read_format = event->attr.read_format;
4578 * compute total_time_enabled, total_time_running
4579 * based on snapshot values taken when the event
4580 * was last scheduled in.
4582 * we cannot simply called update_context_time()
4583 * because of locking issue as we are called in
4586 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4587 calc_timer_values(event, &now, &enabled, &running);
4589 if (event->attr.read_format & PERF_FORMAT_GROUP)
4590 perf_output_read_group(handle, event, enabled, running);
4592 perf_output_read_one(handle, event, enabled, running);
4595 void perf_output_sample(struct perf_output_handle *handle,
4596 struct perf_event_header *header,
4597 struct perf_sample_data *data,
4598 struct perf_event *event)
4600 u64 sample_type = data->type;
4602 perf_output_put(handle, *header);
4604 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4605 perf_output_put(handle, data->id);
4607 if (sample_type & PERF_SAMPLE_IP)
4608 perf_output_put(handle, data->ip);
4610 if (sample_type & PERF_SAMPLE_TID)
4611 perf_output_put(handle, data->tid_entry);
4613 if (sample_type & PERF_SAMPLE_TIME)
4614 perf_output_put(handle, data->time);
4616 if (sample_type & PERF_SAMPLE_ADDR)
4617 perf_output_put(handle, data->addr);
4619 if (sample_type & PERF_SAMPLE_ID)
4620 perf_output_put(handle, data->id);
4622 if (sample_type & PERF_SAMPLE_STREAM_ID)
4623 perf_output_put(handle, data->stream_id);
4625 if (sample_type & PERF_SAMPLE_CPU)
4626 perf_output_put(handle, data->cpu_entry);
4628 if (sample_type & PERF_SAMPLE_PERIOD)
4629 perf_output_put(handle, data->period);
4631 if (sample_type & PERF_SAMPLE_READ)
4632 perf_output_read(handle, event);
4634 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4635 if (data->callchain) {
4638 if (data->callchain)
4639 size += data->callchain->nr;
4641 size *= sizeof(u64);
4643 __output_copy(handle, data->callchain, size);
4646 perf_output_put(handle, nr);
4650 if (sample_type & PERF_SAMPLE_RAW) {
4652 perf_output_put(handle, data->raw->size);
4653 __output_copy(handle, data->raw->data,
4660 .size = sizeof(u32),
4663 perf_output_put(handle, raw);
4667 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4668 if (data->br_stack) {
4671 size = data->br_stack->nr
4672 * sizeof(struct perf_branch_entry);
4674 perf_output_put(handle, data->br_stack->nr);
4675 perf_output_copy(handle, data->br_stack->entries, size);
4678 * we always store at least the value of nr
4681 perf_output_put(handle, nr);
4685 if (sample_type & PERF_SAMPLE_REGS_USER) {
4686 u64 abi = data->regs_user.abi;
4689 * If there are no regs to dump, notice it through
4690 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4692 perf_output_put(handle, abi);
4695 u64 mask = event->attr.sample_regs_user;
4696 perf_output_sample_regs(handle,
4697 data->regs_user.regs,
4702 if (sample_type & PERF_SAMPLE_STACK_USER) {
4703 perf_output_sample_ustack(handle,
4704 data->stack_user_size,
4705 data->regs_user.regs);
4708 if (sample_type & PERF_SAMPLE_WEIGHT)
4709 perf_output_put(handle, data->weight);
4711 if (sample_type & PERF_SAMPLE_DATA_SRC)
4712 perf_output_put(handle, data->data_src.val);
4714 if (sample_type & PERF_SAMPLE_TRANSACTION)
4715 perf_output_put(handle, data->txn);
4717 if (!event->attr.watermark) {
4718 int wakeup_events = event->attr.wakeup_events;
4720 if (wakeup_events) {
4721 struct ring_buffer *rb = handle->rb;
4722 int events = local_inc_return(&rb->events);
4724 if (events >= wakeup_events) {
4725 local_sub(wakeup_events, &rb->events);
4726 local_inc(&rb->wakeup);
4732 void perf_prepare_sample(struct perf_event_header *header,
4733 struct perf_sample_data *data,
4734 struct perf_event *event,
4735 struct pt_regs *regs)
4737 u64 sample_type = event->attr.sample_type;
4739 header->type = PERF_RECORD_SAMPLE;
4740 header->size = sizeof(*header) + event->header_size;
4743 header->misc |= perf_misc_flags(regs);
4745 __perf_event_header__init_id(header, data, event);
4747 if (sample_type & PERF_SAMPLE_IP)
4748 data->ip = perf_instruction_pointer(regs);
4750 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4753 data->callchain = perf_callchain(event, regs);
4755 if (data->callchain)
4756 size += data->callchain->nr;
4758 header->size += size * sizeof(u64);
4761 if (sample_type & PERF_SAMPLE_RAW) {
4762 int size = sizeof(u32);
4765 size += data->raw->size;
4767 size += sizeof(u32);
4769 WARN_ON_ONCE(size & (sizeof(u64)-1));
4770 header->size += size;
4773 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4774 int size = sizeof(u64); /* nr */
4775 if (data->br_stack) {
4776 size += data->br_stack->nr
4777 * sizeof(struct perf_branch_entry);
4779 header->size += size;
4782 if (sample_type & PERF_SAMPLE_REGS_USER) {
4783 /* regs dump ABI info */
4784 int size = sizeof(u64);
4786 perf_sample_regs_user(&data->regs_user, regs);
4788 if (data->regs_user.regs) {
4789 u64 mask = event->attr.sample_regs_user;
4790 size += hweight64(mask) * sizeof(u64);
4793 header->size += size;
4796 if (sample_type & PERF_SAMPLE_STACK_USER) {
4798 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4799 * processed as the last one or have additional check added
4800 * in case new sample type is added, because we could eat
4801 * up the rest of the sample size.
4803 struct perf_regs_user *uregs = &data->regs_user;
4804 u16 stack_size = event->attr.sample_stack_user;
4805 u16 size = sizeof(u64);
4808 perf_sample_regs_user(uregs, regs);
4810 stack_size = perf_sample_ustack_size(stack_size, header->size,
4814 * If there is something to dump, add space for the dump
4815 * itself and for the field that tells the dynamic size,
4816 * which is how many have been actually dumped.
4819 size += sizeof(u64) + stack_size;
4821 data->stack_user_size = stack_size;
4822 header->size += size;
4826 static void perf_event_output(struct perf_event *event,
4827 struct perf_sample_data *data,
4828 struct pt_regs *regs)
4830 struct perf_output_handle handle;
4831 struct perf_event_header header;
4833 /* protect the callchain buffers */
4836 perf_prepare_sample(&header, data, event, regs);
4838 if (perf_output_begin(&handle, event, header.size))
4841 perf_output_sample(&handle, &header, data, event);
4843 perf_output_end(&handle);
4853 struct perf_read_event {
4854 struct perf_event_header header;
4861 perf_event_read_event(struct perf_event *event,
4862 struct task_struct *task)
4864 struct perf_output_handle handle;
4865 struct perf_sample_data sample;
4866 struct perf_read_event read_event = {
4868 .type = PERF_RECORD_READ,
4870 .size = sizeof(read_event) + event->read_size,
4872 .pid = perf_event_pid(event, task),
4873 .tid = perf_event_tid(event, task),
4877 perf_event_header__init_id(&read_event.header, &sample, event);
4878 ret = perf_output_begin(&handle, event, read_event.header.size);
4882 perf_output_put(&handle, read_event);
4883 perf_output_read(&handle, event);
4884 perf_event__output_id_sample(event, &handle, &sample);
4886 perf_output_end(&handle);
4889 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4892 perf_event_aux_ctx(struct perf_event_context *ctx,
4893 perf_event_aux_output_cb output,
4896 struct perf_event *event;
4898 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4899 if (event->state < PERF_EVENT_STATE_INACTIVE)
4901 if (!event_filter_match(event))
4903 output(event, data);
4908 perf_event_aux(perf_event_aux_output_cb output, void *data,
4909 struct perf_event_context *task_ctx)
4911 struct perf_cpu_context *cpuctx;
4912 struct perf_event_context *ctx;
4917 list_for_each_entry_rcu(pmu, &pmus, entry) {
4918 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4919 if (cpuctx->unique_pmu != pmu)
4921 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4924 ctxn = pmu->task_ctx_nr;
4927 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4929 perf_event_aux_ctx(ctx, output, data);
4931 put_cpu_ptr(pmu->pmu_cpu_context);
4936 perf_event_aux_ctx(task_ctx, output, data);
4943 * task tracking -- fork/exit
4945 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4948 struct perf_task_event {
4949 struct task_struct *task;
4950 struct perf_event_context *task_ctx;
4953 struct perf_event_header header;
4963 static int perf_event_task_match(struct perf_event *event)
4965 return event->attr.comm || event->attr.mmap ||
4966 event->attr.mmap2 || event->attr.mmap_data ||
4970 static void perf_event_task_output(struct perf_event *event,
4973 struct perf_task_event *task_event = data;
4974 struct perf_output_handle handle;
4975 struct perf_sample_data sample;
4976 struct task_struct *task = task_event->task;
4977 int ret, size = task_event->event_id.header.size;
4979 if (!perf_event_task_match(event))
4982 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4984 ret = perf_output_begin(&handle, event,
4985 task_event->event_id.header.size);
4989 task_event->event_id.pid = perf_event_pid(event, task);
4990 task_event->event_id.ppid = perf_event_pid(event, current);
4992 task_event->event_id.tid = perf_event_tid(event, task);
4993 task_event->event_id.ptid = perf_event_tid(event, current);
4995 perf_output_put(&handle, task_event->event_id);
4997 perf_event__output_id_sample(event, &handle, &sample);
4999 perf_output_end(&handle);
5001 task_event->event_id.header.size = size;
5004 static void perf_event_task(struct task_struct *task,
5005 struct perf_event_context *task_ctx,
5008 struct perf_task_event task_event;
5010 if (!atomic_read(&nr_comm_events) &&
5011 !atomic_read(&nr_mmap_events) &&
5012 !atomic_read(&nr_task_events))
5015 task_event = (struct perf_task_event){
5017 .task_ctx = task_ctx,
5020 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5022 .size = sizeof(task_event.event_id),
5028 .time = perf_clock(),
5032 perf_event_aux(perf_event_task_output,
5037 void perf_event_fork(struct task_struct *task)
5039 perf_event_task(task, NULL, 1);
5046 struct perf_comm_event {
5047 struct task_struct *task;
5052 struct perf_event_header header;
5059 static int perf_event_comm_match(struct perf_event *event)
5061 return event->attr.comm;
5064 static void perf_event_comm_output(struct perf_event *event,
5067 struct perf_comm_event *comm_event = data;
5068 struct perf_output_handle handle;
5069 struct perf_sample_data sample;
5070 int size = comm_event->event_id.header.size;
5073 if (!perf_event_comm_match(event))
5076 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5077 ret = perf_output_begin(&handle, event,
5078 comm_event->event_id.header.size);
5083 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5084 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5086 perf_output_put(&handle, comm_event->event_id);
5087 __output_copy(&handle, comm_event->comm,
5088 comm_event->comm_size);
5090 perf_event__output_id_sample(event, &handle, &sample);
5092 perf_output_end(&handle);
5094 comm_event->event_id.header.size = size;
5097 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5099 char comm[TASK_COMM_LEN];
5102 memset(comm, 0, sizeof(comm));
5103 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5104 size = ALIGN(strlen(comm)+1, sizeof(u64));
5106 comm_event->comm = comm;
5107 comm_event->comm_size = size;
5109 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5111 perf_event_aux(perf_event_comm_output,
5116 void perf_event_comm(struct task_struct *task, bool exec)
5118 struct perf_comm_event comm_event;
5120 if (!atomic_read(&nr_comm_events))
5123 comm_event = (struct perf_comm_event){
5129 .type = PERF_RECORD_COMM,
5130 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5138 perf_event_comm_event(&comm_event);
5145 struct perf_mmap_event {
5146 struct vm_area_struct *vma;
5148 const char *file_name;
5156 struct perf_event_header header;
5166 static int perf_event_mmap_match(struct perf_event *event,
5169 struct perf_mmap_event *mmap_event = data;
5170 struct vm_area_struct *vma = mmap_event->vma;
5171 int executable = vma->vm_flags & VM_EXEC;
5173 return (!executable && event->attr.mmap_data) ||
5174 (executable && (event->attr.mmap || event->attr.mmap2));
5177 static void perf_event_mmap_output(struct perf_event *event,
5180 struct perf_mmap_event *mmap_event = data;
5181 struct perf_output_handle handle;
5182 struct perf_sample_data sample;
5183 int size = mmap_event->event_id.header.size;
5186 if (!perf_event_mmap_match(event, data))
5189 if (event->attr.mmap2) {
5190 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5191 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5192 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5193 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5194 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5195 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5196 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5199 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5200 ret = perf_output_begin(&handle, event,
5201 mmap_event->event_id.header.size);
5205 mmap_event->event_id.pid = perf_event_pid(event, current);
5206 mmap_event->event_id.tid = perf_event_tid(event, current);
5208 perf_output_put(&handle, mmap_event->event_id);
5210 if (event->attr.mmap2) {
5211 perf_output_put(&handle, mmap_event->maj);
5212 perf_output_put(&handle, mmap_event->min);
5213 perf_output_put(&handle, mmap_event->ino);
5214 perf_output_put(&handle, mmap_event->ino_generation);
5215 perf_output_put(&handle, mmap_event->prot);
5216 perf_output_put(&handle, mmap_event->flags);
5219 __output_copy(&handle, mmap_event->file_name,
5220 mmap_event->file_size);
5222 perf_event__output_id_sample(event, &handle, &sample);
5224 perf_output_end(&handle);
5226 mmap_event->event_id.header.size = size;
5229 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5231 struct vm_area_struct *vma = mmap_event->vma;
5232 struct file *file = vma->vm_file;
5233 int maj = 0, min = 0;
5234 u64 ino = 0, gen = 0;
5235 u32 prot = 0, flags = 0;
5242 struct inode *inode;
5245 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5251 * d_path() works from the end of the rb backwards, so we
5252 * need to add enough zero bytes after the string to handle
5253 * the 64bit alignment we do later.
5255 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5260 inode = file_inode(vma->vm_file);
5261 dev = inode->i_sb->s_dev;
5263 gen = inode->i_generation;
5267 if (vma->vm_flags & VM_READ)
5269 if (vma->vm_flags & VM_WRITE)
5271 if (vma->vm_flags & VM_EXEC)
5274 if (vma->vm_flags & VM_MAYSHARE)
5277 flags = MAP_PRIVATE;
5279 if (vma->vm_flags & VM_DENYWRITE)
5280 flags |= MAP_DENYWRITE;
5281 if (vma->vm_flags & VM_MAYEXEC)
5282 flags |= MAP_EXECUTABLE;
5283 if (vma->vm_flags & VM_LOCKED)
5284 flags |= MAP_LOCKED;
5285 if (vma->vm_flags & VM_HUGETLB)
5286 flags |= MAP_HUGETLB;
5290 if (vma->vm_ops && vma->vm_ops->name) {
5291 name = (char *) vma->vm_ops->name(vma);
5296 name = (char *)arch_vma_name(vma);
5300 if (vma->vm_start <= vma->vm_mm->start_brk &&
5301 vma->vm_end >= vma->vm_mm->brk) {
5305 if (vma->vm_start <= vma->vm_mm->start_stack &&
5306 vma->vm_end >= vma->vm_mm->start_stack) {
5316 strlcpy(tmp, name, sizeof(tmp));
5320 * Since our buffer works in 8 byte units we need to align our string
5321 * size to a multiple of 8. However, we must guarantee the tail end is
5322 * zero'd out to avoid leaking random bits to userspace.
5324 size = strlen(name)+1;
5325 while (!IS_ALIGNED(size, sizeof(u64)))
5326 name[size++] = '\0';
5328 mmap_event->file_name = name;
5329 mmap_event->file_size = size;
5330 mmap_event->maj = maj;
5331 mmap_event->min = min;
5332 mmap_event->ino = ino;
5333 mmap_event->ino_generation = gen;
5334 mmap_event->prot = prot;
5335 mmap_event->flags = flags;
5337 if (!(vma->vm_flags & VM_EXEC))
5338 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5340 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5342 perf_event_aux(perf_event_mmap_output,
5349 void perf_event_mmap(struct vm_area_struct *vma)
5351 struct perf_mmap_event mmap_event;
5353 if (!atomic_read(&nr_mmap_events))
5356 mmap_event = (struct perf_mmap_event){
5362 .type = PERF_RECORD_MMAP,
5363 .misc = PERF_RECORD_MISC_USER,
5368 .start = vma->vm_start,
5369 .len = vma->vm_end - vma->vm_start,
5370 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5372 /* .maj (attr_mmap2 only) */
5373 /* .min (attr_mmap2 only) */
5374 /* .ino (attr_mmap2 only) */
5375 /* .ino_generation (attr_mmap2 only) */
5376 /* .prot (attr_mmap2 only) */
5377 /* .flags (attr_mmap2 only) */
5380 perf_event_mmap_event(&mmap_event);
5384 * IRQ throttle logging
5387 static void perf_log_throttle(struct perf_event *event, int enable)
5389 struct perf_output_handle handle;
5390 struct perf_sample_data sample;
5394 struct perf_event_header header;
5398 } throttle_event = {
5400 .type = PERF_RECORD_THROTTLE,
5402 .size = sizeof(throttle_event),
5404 .time = perf_clock(),
5405 .id = primary_event_id(event),
5406 .stream_id = event->id,
5410 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5412 perf_event_header__init_id(&throttle_event.header, &sample, event);
5414 ret = perf_output_begin(&handle, event,
5415 throttle_event.header.size);
5419 perf_output_put(&handle, throttle_event);
5420 perf_event__output_id_sample(event, &handle, &sample);
5421 perf_output_end(&handle);
5425 * Generic event overflow handling, sampling.
5428 static int __perf_event_overflow(struct perf_event *event,
5429 int throttle, struct perf_sample_data *data,
5430 struct pt_regs *regs)
5432 int events = atomic_read(&event->event_limit);
5433 struct hw_perf_event *hwc = &event->hw;
5438 * Non-sampling counters might still use the PMI to fold short
5439 * hardware counters, ignore those.
5441 if (unlikely(!is_sampling_event(event)))
5444 seq = __this_cpu_read(perf_throttled_seq);
5445 if (seq != hwc->interrupts_seq) {
5446 hwc->interrupts_seq = seq;
5447 hwc->interrupts = 1;
5450 if (unlikely(throttle
5451 && hwc->interrupts >= max_samples_per_tick)) {
5452 __this_cpu_inc(perf_throttled_count);
5453 hwc->interrupts = MAX_INTERRUPTS;
5454 perf_log_throttle(event, 0);
5455 tick_nohz_full_kick();
5460 if (event->attr.freq) {
5461 u64 now = perf_clock();
5462 s64 delta = now - hwc->freq_time_stamp;
5464 hwc->freq_time_stamp = now;
5466 if (delta > 0 && delta < 2*TICK_NSEC)
5467 perf_adjust_period(event, delta, hwc->last_period, true);
5471 * XXX event_limit might not quite work as expected on inherited
5475 event->pending_kill = POLL_IN;
5476 if (events && atomic_dec_and_test(&event->event_limit)) {
5478 event->pending_kill = POLL_HUP;
5479 event->pending_disable = 1;
5480 irq_work_queue(&event->pending);
5483 if (event->overflow_handler)
5484 event->overflow_handler(event, data, regs);
5486 perf_event_output(event, data, regs);
5488 if (event->fasync && event->pending_kill) {
5489 event->pending_wakeup = 1;
5490 irq_work_queue(&event->pending);
5496 int perf_event_overflow(struct perf_event *event,
5497 struct perf_sample_data *data,
5498 struct pt_regs *regs)
5500 return __perf_event_overflow(event, 1, data, regs);
5504 * Generic software event infrastructure
5507 struct swevent_htable {
5508 struct swevent_hlist *swevent_hlist;
5509 struct mutex hlist_mutex;
5512 /* Recursion avoidance in each contexts */
5513 int recursion[PERF_NR_CONTEXTS];
5515 /* Keeps track of cpu being initialized/exited */
5519 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5522 * We directly increment event->count and keep a second value in
5523 * event->hw.period_left to count intervals. This period event
5524 * is kept in the range [-sample_period, 0] so that we can use the
5528 u64 perf_swevent_set_period(struct perf_event *event)
5530 struct hw_perf_event *hwc = &event->hw;
5531 u64 period = hwc->last_period;
5535 hwc->last_period = hwc->sample_period;
5538 old = val = local64_read(&hwc->period_left);
5542 nr = div64_u64(period + val, period);
5543 offset = nr * period;
5545 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5551 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5552 struct perf_sample_data *data,
5553 struct pt_regs *regs)
5555 struct hw_perf_event *hwc = &event->hw;
5559 overflow = perf_swevent_set_period(event);
5561 if (hwc->interrupts == MAX_INTERRUPTS)
5564 for (; overflow; overflow--) {
5565 if (__perf_event_overflow(event, throttle,
5568 * We inhibit the overflow from happening when
5569 * hwc->interrupts == MAX_INTERRUPTS.
5577 static void perf_swevent_event(struct perf_event *event, u64 nr,
5578 struct perf_sample_data *data,
5579 struct pt_regs *regs)
5581 struct hw_perf_event *hwc = &event->hw;
5583 local64_add(nr, &event->count);
5588 if (!is_sampling_event(event))
5591 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5593 return perf_swevent_overflow(event, 1, data, regs);
5595 data->period = event->hw.last_period;
5597 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5598 return perf_swevent_overflow(event, 1, data, regs);
5600 if (local64_add_negative(nr, &hwc->period_left))
5603 perf_swevent_overflow(event, 0, data, regs);
5606 static int perf_exclude_event(struct perf_event *event,
5607 struct pt_regs *regs)
5609 if (event->hw.state & PERF_HES_STOPPED)
5613 if (event->attr.exclude_user && user_mode(regs))
5616 if (event->attr.exclude_kernel && !user_mode(regs))
5623 static int perf_swevent_match(struct perf_event *event,
5624 enum perf_type_id type,
5626 struct perf_sample_data *data,
5627 struct pt_regs *regs)
5629 if (event->attr.type != type)
5632 if (event->attr.config != event_id)
5635 if (perf_exclude_event(event, regs))
5641 static inline u64 swevent_hash(u64 type, u32 event_id)
5643 u64 val = event_id | (type << 32);
5645 return hash_64(val, SWEVENT_HLIST_BITS);
5648 static inline struct hlist_head *
5649 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5651 u64 hash = swevent_hash(type, event_id);
5653 return &hlist->heads[hash];
5656 /* For the read side: events when they trigger */
5657 static inline struct hlist_head *
5658 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5660 struct swevent_hlist *hlist;
5662 hlist = rcu_dereference(swhash->swevent_hlist);
5666 return __find_swevent_head(hlist, type, event_id);
5669 /* For the event head insertion and removal in the hlist */
5670 static inline struct hlist_head *
5671 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5673 struct swevent_hlist *hlist;
5674 u32 event_id = event->attr.config;
5675 u64 type = event->attr.type;
5678 * Event scheduling is always serialized against hlist allocation
5679 * and release. Which makes the protected version suitable here.
5680 * The context lock guarantees that.
5682 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5683 lockdep_is_held(&event->ctx->lock));
5687 return __find_swevent_head(hlist, type, event_id);
5690 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5692 struct perf_sample_data *data,
5693 struct pt_regs *regs)
5695 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5696 struct perf_event *event;
5697 struct hlist_head *head;
5700 head = find_swevent_head_rcu(swhash, type, event_id);
5704 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5705 if (perf_swevent_match(event, type, event_id, data, regs))
5706 perf_swevent_event(event, nr, data, regs);
5712 int perf_swevent_get_recursion_context(void)
5714 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5716 return get_recursion_context(swhash->recursion);
5718 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5720 inline void perf_swevent_put_recursion_context(int rctx)
5722 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5724 put_recursion_context(swhash->recursion, rctx);
5727 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5729 struct perf_sample_data data;
5732 preempt_disable_notrace();
5733 rctx = perf_swevent_get_recursion_context();
5737 perf_sample_data_init(&data, addr, 0);
5739 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5741 perf_swevent_put_recursion_context(rctx);
5742 preempt_enable_notrace();
5745 static void perf_swevent_read(struct perf_event *event)
5749 static int perf_swevent_add(struct perf_event *event, int flags)
5751 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5752 struct hw_perf_event *hwc = &event->hw;
5753 struct hlist_head *head;
5755 if (is_sampling_event(event)) {
5756 hwc->last_period = hwc->sample_period;
5757 perf_swevent_set_period(event);
5760 hwc->state = !(flags & PERF_EF_START);
5762 head = find_swevent_head(swhash, event);
5765 * We can race with cpu hotplug code. Do not
5766 * WARN if the cpu just got unplugged.
5768 WARN_ON_ONCE(swhash->online);
5772 hlist_add_head_rcu(&event->hlist_entry, head);
5777 static void perf_swevent_del(struct perf_event *event, int flags)
5779 hlist_del_rcu(&event->hlist_entry);
5782 static void perf_swevent_start(struct perf_event *event, int flags)
5784 event->hw.state = 0;
5787 static void perf_swevent_stop(struct perf_event *event, int flags)
5789 event->hw.state = PERF_HES_STOPPED;
5792 /* Deref the hlist from the update side */
5793 static inline struct swevent_hlist *
5794 swevent_hlist_deref(struct swevent_htable *swhash)
5796 return rcu_dereference_protected(swhash->swevent_hlist,
5797 lockdep_is_held(&swhash->hlist_mutex));
5800 static void swevent_hlist_release(struct swevent_htable *swhash)
5802 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5807 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5808 kfree_rcu(hlist, rcu_head);
5811 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5813 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5815 mutex_lock(&swhash->hlist_mutex);
5817 if (!--swhash->hlist_refcount)
5818 swevent_hlist_release(swhash);
5820 mutex_unlock(&swhash->hlist_mutex);
5823 static void swevent_hlist_put(struct perf_event *event)
5827 for_each_possible_cpu(cpu)
5828 swevent_hlist_put_cpu(event, cpu);
5831 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5833 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5836 mutex_lock(&swhash->hlist_mutex);
5838 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5839 struct swevent_hlist *hlist;
5841 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5846 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5848 swhash->hlist_refcount++;
5850 mutex_unlock(&swhash->hlist_mutex);
5855 static int swevent_hlist_get(struct perf_event *event)
5858 int cpu, failed_cpu;
5861 for_each_possible_cpu(cpu) {
5862 err = swevent_hlist_get_cpu(event, cpu);
5872 for_each_possible_cpu(cpu) {
5873 if (cpu == failed_cpu)
5875 swevent_hlist_put_cpu(event, cpu);
5882 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5884 static void sw_perf_event_destroy(struct perf_event *event)
5886 u64 event_id = event->attr.config;
5888 WARN_ON(event->parent);
5890 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5891 swevent_hlist_put(event);
5894 static int perf_swevent_init(struct perf_event *event)
5896 u64 event_id = event->attr.config;
5898 if (event->attr.type != PERF_TYPE_SOFTWARE)
5902 * no branch sampling for software events
5904 if (has_branch_stack(event))
5908 case PERF_COUNT_SW_CPU_CLOCK:
5909 case PERF_COUNT_SW_TASK_CLOCK:
5916 if (event_id >= PERF_COUNT_SW_MAX)
5919 if (!event->parent) {
5922 err = swevent_hlist_get(event);
5926 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5927 event->destroy = sw_perf_event_destroy;
5933 static int perf_swevent_event_idx(struct perf_event *event)
5938 static struct pmu perf_swevent = {
5939 .task_ctx_nr = perf_sw_context,
5941 .event_init = perf_swevent_init,
5942 .add = perf_swevent_add,
5943 .del = perf_swevent_del,
5944 .start = perf_swevent_start,
5945 .stop = perf_swevent_stop,
5946 .read = perf_swevent_read,
5948 .event_idx = perf_swevent_event_idx,
5951 #ifdef CONFIG_EVENT_TRACING
5953 static int perf_tp_filter_match(struct perf_event *event,
5954 struct perf_sample_data *data)
5956 void *record = data->raw->data;
5958 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5963 static int perf_tp_event_match(struct perf_event *event,
5964 struct perf_sample_data *data,
5965 struct pt_regs *regs)
5967 if (event->hw.state & PERF_HES_STOPPED)
5970 * All tracepoints are from kernel-space.
5972 if (event->attr.exclude_kernel)
5975 if (!perf_tp_filter_match(event, data))
5981 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5982 struct pt_regs *regs, struct hlist_head *head, int rctx,
5983 struct task_struct *task)
5985 struct perf_sample_data data;
5986 struct perf_event *event;
5988 struct perf_raw_record raw = {
5993 perf_sample_data_init(&data, addr, 0);
5996 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5997 if (perf_tp_event_match(event, &data, regs))
5998 perf_swevent_event(event, count, &data, regs);
6002 * If we got specified a target task, also iterate its context and
6003 * deliver this event there too.
6005 if (task && task != current) {
6006 struct perf_event_context *ctx;
6007 struct trace_entry *entry = record;
6010 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6014 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6015 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6017 if (event->attr.config != entry->type)
6019 if (perf_tp_event_match(event, &data, regs))
6020 perf_swevent_event(event, count, &data, regs);
6026 perf_swevent_put_recursion_context(rctx);
6028 EXPORT_SYMBOL_GPL(perf_tp_event);
6030 static void tp_perf_event_destroy(struct perf_event *event)
6032 perf_trace_destroy(event);
6035 static int perf_tp_event_init(struct perf_event *event)
6039 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6043 * no branch sampling for tracepoint events
6045 if (has_branch_stack(event))
6048 err = perf_trace_init(event);
6052 event->destroy = tp_perf_event_destroy;
6057 static struct pmu perf_tracepoint = {
6058 .task_ctx_nr = perf_sw_context,
6060 .event_init = perf_tp_event_init,
6061 .add = perf_trace_add,
6062 .del = perf_trace_del,
6063 .start = perf_swevent_start,
6064 .stop = perf_swevent_stop,
6065 .read = perf_swevent_read,
6067 .event_idx = perf_swevent_event_idx,
6070 static inline void perf_tp_register(void)
6072 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6075 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6080 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6083 filter_str = strndup_user(arg, PAGE_SIZE);
6084 if (IS_ERR(filter_str))
6085 return PTR_ERR(filter_str);
6087 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6093 static void perf_event_free_filter(struct perf_event *event)
6095 ftrace_profile_free_filter(event);
6100 static inline void perf_tp_register(void)
6104 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6109 static void perf_event_free_filter(struct perf_event *event)
6113 #endif /* CONFIG_EVENT_TRACING */
6115 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6116 void perf_bp_event(struct perf_event *bp, void *data)
6118 struct perf_sample_data sample;
6119 struct pt_regs *regs = data;
6121 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6123 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6124 perf_swevent_event(bp, 1, &sample, regs);
6129 * hrtimer based swevent callback
6132 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6134 enum hrtimer_restart ret = HRTIMER_RESTART;
6135 struct perf_sample_data data;
6136 struct pt_regs *regs;
6137 struct perf_event *event;
6140 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6142 if (event->state != PERF_EVENT_STATE_ACTIVE)
6143 return HRTIMER_NORESTART;
6145 event->pmu->read(event);
6147 perf_sample_data_init(&data, 0, event->hw.last_period);
6148 regs = get_irq_regs();
6150 if (regs && !perf_exclude_event(event, regs)) {
6151 if (!(event->attr.exclude_idle && is_idle_task(current)))
6152 if (__perf_event_overflow(event, 1, &data, regs))
6153 ret = HRTIMER_NORESTART;
6156 period = max_t(u64, 10000, event->hw.sample_period);
6157 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6162 static void perf_swevent_start_hrtimer(struct perf_event *event)
6164 struct hw_perf_event *hwc = &event->hw;
6167 if (!is_sampling_event(event))
6170 period = local64_read(&hwc->period_left);
6175 local64_set(&hwc->period_left, 0);
6177 period = max_t(u64, 10000, hwc->sample_period);
6179 __hrtimer_start_range_ns(&hwc->hrtimer,
6180 ns_to_ktime(period), 0,
6181 HRTIMER_MODE_REL_PINNED, 0);
6184 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6186 struct hw_perf_event *hwc = &event->hw;
6188 if (is_sampling_event(event)) {
6189 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6190 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6192 hrtimer_cancel(&hwc->hrtimer);
6196 static void perf_swevent_init_hrtimer(struct perf_event *event)
6198 struct hw_perf_event *hwc = &event->hw;
6200 if (!is_sampling_event(event))
6203 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6204 hwc->hrtimer.function = perf_swevent_hrtimer;
6207 * Since hrtimers have a fixed rate, we can do a static freq->period
6208 * mapping and avoid the whole period adjust feedback stuff.
6210 if (event->attr.freq) {
6211 long freq = event->attr.sample_freq;
6213 event->attr.sample_period = NSEC_PER_SEC / freq;
6214 hwc->sample_period = event->attr.sample_period;
6215 local64_set(&hwc->period_left, hwc->sample_period);
6216 hwc->last_period = hwc->sample_period;
6217 event->attr.freq = 0;
6222 * Software event: cpu wall time clock
6225 static void cpu_clock_event_update(struct perf_event *event)
6230 now = local_clock();
6231 prev = local64_xchg(&event->hw.prev_count, now);
6232 local64_add(now - prev, &event->count);
6235 static void cpu_clock_event_start(struct perf_event *event, int flags)
6237 local64_set(&event->hw.prev_count, local_clock());
6238 perf_swevent_start_hrtimer(event);
6241 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6243 perf_swevent_cancel_hrtimer(event);
6244 cpu_clock_event_update(event);
6247 static int cpu_clock_event_add(struct perf_event *event, int flags)
6249 if (flags & PERF_EF_START)
6250 cpu_clock_event_start(event, flags);
6255 static void cpu_clock_event_del(struct perf_event *event, int flags)
6257 cpu_clock_event_stop(event, flags);
6260 static void cpu_clock_event_read(struct perf_event *event)
6262 cpu_clock_event_update(event);
6265 static int cpu_clock_event_init(struct perf_event *event)
6267 if (event->attr.type != PERF_TYPE_SOFTWARE)
6270 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6274 * no branch sampling for software events
6276 if (has_branch_stack(event))
6279 perf_swevent_init_hrtimer(event);
6284 static struct pmu perf_cpu_clock = {
6285 .task_ctx_nr = perf_sw_context,
6287 .event_init = cpu_clock_event_init,
6288 .add = cpu_clock_event_add,
6289 .del = cpu_clock_event_del,
6290 .start = cpu_clock_event_start,
6291 .stop = cpu_clock_event_stop,
6292 .read = cpu_clock_event_read,
6294 .event_idx = perf_swevent_event_idx,
6298 * Software event: task time clock
6301 static void task_clock_event_update(struct perf_event *event, u64 now)
6306 prev = local64_xchg(&event->hw.prev_count, now);
6308 local64_add(delta, &event->count);
6311 static void task_clock_event_start(struct perf_event *event, int flags)
6313 local64_set(&event->hw.prev_count, event->ctx->time);
6314 perf_swevent_start_hrtimer(event);
6317 static void task_clock_event_stop(struct perf_event *event, int flags)
6319 perf_swevent_cancel_hrtimer(event);
6320 task_clock_event_update(event, event->ctx->time);
6323 static int task_clock_event_add(struct perf_event *event, int flags)
6325 if (flags & PERF_EF_START)
6326 task_clock_event_start(event, flags);
6331 static void task_clock_event_del(struct perf_event *event, int flags)
6333 task_clock_event_stop(event, PERF_EF_UPDATE);
6336 static void task_clock_event_read(struct perf_event *event)
6338 u64 now = perf_clock();
6339 u64 delta = now - event->ctx->timestamp;
6340 u64 time = event->ctx->time + delta;
6342 task_clock_event_update(event, time);
6345 static int task_clock_event_init(struct perf_event *event)
6347 if (event->attr.type != PERF_TYPE_SOFTWARE)
6350 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6354 * no branch sampling for software events
6356 if (has_branch_stack(event))
6359 perf_swevent_init_hrtimer(event);
6364 static struct pmu perf_task_clock = {
6365 .task_ctx_nr = perf_sw_context,
6367 .event_init = task_clock_event_init,
6368 .add = task_clock_event_add,
6369 .del = task_clock_event_del,
6370 .start = task_clock_event_start,
6371 .stop = task_clock_event_stop,
6372 .read = task_clock_event_read,
6374 .event_idx = perf_swevent_event_idx,
6377 static void perf_pmu_nop_void(struct pmu *pmu)
6381 static int perf_pmu_nop_int(struct pmu *pmu)
6386 static void perf_pmu_start_txn(struct pmu *pmu)
6388 perf_pmu_disable(pmu);
6391 static int perf_pmu_commit_txn(struct pmu *pmu)
6393 perf_pmu_enable(pmu);
6397 static void perf_pmu_cancel_txn(struct pmu *pmu)
6399 perf_pmu_enable(pmu);
6402 static int perf_event_idx_default(struct perf_event *event)
6404 return event->hw.idx + 1;
6408 * Ensures all contexts with the same task_ctx_nr have the same
6409 * pmu_cpu_context too.
6411 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6418 list_for_each_entry(pmu, &pmus, entry) {
6419 if (pmu->task_ctx_nr == ctxn)
6420 return pmu->pmu_cpu_context;
6426 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6430 for_each_possible_cpu(cpu) {
6431 struct perf_cpu_context *cpuctx;
6433 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6435 if (cpuctx->unique_pmu == old_pmu)
6436 cpuctx->unique_pmu = pmu;
6440 static void free_pmu_context(struct pmu *pmu)
6444 mutex_lock(&pmus_lock);
6446 * Like a real lame refcount.
6448 list_for_each_entry(i, &pmus, entry) {
6449 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6450 update_pmu_context(i, pmu);
6455 free_percpu(pmu->pmu_cpu_context);
6457 mutex_unlock(&pmus_lock);
6459 static struct idr pmu_idr;
6462 type_show(struct device *dev, struct device_attribute *attr, char *page)
6464 struct pmu *pmu = dev_get_drvdata(dev);
6466 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6468 static DEVICE_ATTR_RO(type);
6471 perf_event_mux_interval_ms_show(struct device *dev,
6472 struct device_attribute *attr,
6475 struct pmu *pmu = dev_get_drvdata(dev);
6477 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6481 perf_event_mux_interval_ms_store(struct device *dev,
6482 struct device_attribute *attr,
6483 const char *buf, size_t count)
6485 struct pmu *pmu = dev_get_drvdata(dev);
6486 int timer, cpu, ret;
6488 ret = kstrtoint(buf, 0, &timer);
6495 /* same value, noting to do */
6496 if (timer == pmu->hrtimer_interval_ms)
6499 pmu->hrtimer_interval_ms = timer;
6501 /* update all cpuctx for this PMU */
6502 for_each_possible_cpu(cpu) {
6503 struct perf_cpu_context *cpuctx;
6504 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6505 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6507 if (hrtimer_active(&cpuctx->hrtimer))
6508 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6513 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6515 static struct attribute *pmu_dev_attrs[] = {
6516 &dev_attr_type.attr,
6517 &dev_attr_perf_event_mux_interval_ms.attr,
6520 ATTRIBUTE_GROUPS(pmu_dev);
6522 static int pmu_bus_running;
6523 static struct bus_type pmu_bus = {
6524 .name = "event_source",
6525 .dev_groups = pmu_dev_groups,
6528 static void pmu_dev_release(struct device *dev)
6533 static int pmu_dev_alloc(struct pmu *pmu)
6537 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6541 pmu->dev->groups = pmu->attr_groups;
6542 device_initialize(pmu->dev);
6543 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6547 dev_set_drvdata(pmu->dev, pmu);
6548 pmu->dev->bus = &pmu_bus;
6549 pmu->dev->release = pmu_dev_release;
6550 ret = device_add(pmu->dev);
6558 put_device(pmu->dev);
6562 static struct lock_class_key cpuctx_mutex;
6563 static struct lock_class_key cpuctx_lock;
6565 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6569 mutex_lock(&pmus_lock);
6571 pmu->pmu_disable_count = alloc_percpu(int);
6572 if (!pmu->pmu_disable_count)
6581 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6589 if (pmu_bus_running) {
6590 ret = pmu_dev_alloc(pmu);
6596 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6597 if (pmu->pmu_cpu_context)
6598 goto got_cpu_context;
6601 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6602 if (!pmu->pmu_cpu_context)
6605 for_each_possible_cpu(cpu) {
6606 struct perf_cpu_context *cpuctx;
6608 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6609 __perf_event_init_context(&cpuctx->ctx);
6610 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6611 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6612 cpuctx->ctx.type = cpu_context;
6613 cpuctx->ctx.pmu = pmu;
6615 __perf_cpu_hrtimer_init(cpuctx, cpu);
6617 INIT_LIST_HEAD(&cpuctx->rotation_list);
6618 cpuctx->unique_pmu = pmu;
6622 if (!pmu->start_txn) {
6623 if (pmu->pmu_enable) {
6625 * If we have pmu_enable/pmu_disable calls, install
6626 * transaction stubs that use that to try and batch
6627 * hardware accesses.
6629 pmu->start_txn = perf_pmu_start_txn;
6630 pmu->commit_txn = perf_pmu_commit_txn;
6631 pmu->cancel_txn = perf_pmu_cancel_txn;
6633 pmu->start_txn = perf_pmu_nop_void;
6634 pmu->commit_txn = perf_pmu_nop_int;
6635 pmu->cancel_txn = perf_pmu_nop_void;
6639 if (!pmu->pmu_enable) {
6640 pmu->pmu_enable = perf_pmu_nop_void;
6641 pmu->pmu_disable = perf_pmu_nop_void;
6644 if (!pmu->event_idx)
6645 pmu->event_idx = perf_event_idx_default;
6647 list_add_rcu(&pmu->entry, &pmus);
6650 mutex_unlock(&pmus_lock);
6655 device_del(pmu->dev);
6656 put_device(pmu->dev);
6659 if (pmu->type >= PERF_TYPE_MAX)
6660 idr_remove(&pmu_idr, pmu->type);
6663 free_percpu(pmu->pmu_disable_count);
6666 EXPORT_SYMBOL_GPL(perf_pmu_register);
6668 void perf_pmu_unregister(struct pmu *pmu)
6670 mutex_lock(&pmus_lock);
6671 list_del_rcu(&pmu->entry);
6672 mutex_unlock(&pmus_lock);
6675 * We dereference the pmu list under both SRCU and regular RCU, so
6676 * synchronize against both of those.
6678 synchronize_srcu(&pmus_srcu);
6681 free_percpu(pmu->pmu_disable_count);
6682 if (pmu->type >= PERF_TYPE_MAX)
6683 idr_remove(&pmu_idr, pmu->type);
6684 device_del(pmu->dev);
6685 put_device(pmu->dev);
6686 free_pmu_context(pmu);
6688 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6690 struct pmu *perf_init_event(struct perf_event *event)
6692 struct pmu *pmu = NULL;
6696 idx = srcu_read_lock(&pmus_srcu);
6699 pmu = idr_find(&pmu_idr, event->attr.type);
6702 if (!try_module_get(pmu->module)) {
6703 pmu = ERR_PTR(-ENODEV);
6707 ret = pmu->event_init(event);
6713 list_for_each_entry_rcu(pmu, &pmus, entry) {
6714 if (!try_module_get(pmu->module)) {
6715 pmu = ERR_PTR(-ENODEV);
6719 ret = pmu->event_init(event);
6723 if (ret != -ENOENT) {
6728 pmu = ERR_PTR(-ENOENT);
6730 srcu_read_unlock(&pmus_srcu, idx);
6735 static void account_event_cpu(struct perf_event *event, int cpu)
6740 if (has_branch_stack(event)) {
6741 if (!(event->attach_state & PERF_ATTACH_TASK))
6742 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6744 if (is_cgroup_event(event))
6745 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6748 static void account_event(struct perf_event *event)
6753 if (event->attach_state & PERF_ATTACH_TASK)
6754 static_key_slow_inc(&perf_sched_events.key);
6755 if (event->attr.mmap || event->attr.mmap_data)
6756 atomic_inc(&nr_mmap_events);
6757 if (event->attr.comm)
6758 atomic_inc(&nr_comm_events);
6759 if (event->attr.task)
6760 atomic_inc(&nr_task_events);
6761 if (event->attr.freq) {
6762 if (atomic_inc_return(&nr_freq_events) == 1)
6763 tick_nohz_full_kick_all();
6765 if (has_branch_stack(event))
6766 static_key_slow_inc(&perf_sched_events.key);
6767 if (is_cgroup_event(event))
6768 static_key_slow_inc(&perf_sched_events.key);
6770 account_event_cpu(event, event->cpu);
6774 * Allocate and initialize a event structure
6776 static struct perf_event *
6777 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6778 struct task_struct *task,
6779 struct perf_event *group_leader,
6780 struct perf_event *parent_event,
6781 perf_overflow_handler_t overflow_handler,
6785 struct perf_event *event;
6786 struct hw_perf_event *hwc;
6789 if ((unsigned)cpu >= nr_cpu_ids) {
6790 if (!task || cpu != -1)
6791 return ERR_PTR(-EINVAL);
6794 event = kzalloc(sizeof(*event), GFP_KERNEL);
6796 return ERR_PTR(-ENOMEM);
6799 * Single events are their own group leaders, with an
6800 * empty sibling list:
6803 group_leader = event;
6805 mutex_init(&event->child_mutex);
6806 INIT_LIST_HEAD(&event->child_list);
6808 INIT_LIST_HEAD(&event->group_entry);
6809 INIT_LIST_HEAD(&event->event_entry);
6810 INIT_LIST_HEAD(&event->sibling_list);
6811 INIT_LIST_HEAD(&event->rb_entry);
6812 INIT_LIST_HEAD(&event->active_entry);
6813 INIT_HLIST_NODE(&event->hlist_entry);
6816 init_waitqueue_head(&event->waitq);
6817 init_irq_work(&event->pending, perf_pending_event);
6819 mutex_init(&event->mmap_mutex);
6821 atomic_long_set(&event->refcount, 1);
6823 event->attr = *attr;
6824 event->group_leader = group_leader;
6828 event->parent = parent_event;
6830 event->ns = get_pid_ns(task_active_pid_ns(current));
6831 event->id = atomic64_inc_return(&perf_event_id);
6833 event->state = PERF_EVENT_STATE_INACTIVE;
6836 event->attach_state = PERF_ATTACH_TASK;
6838 if (attr->type == PERF_TYPE_TRACEPOINT)
6839 event->hw.tp_target = task;
6840 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6842 * hw_breakpoint is a bit difficult here..
6844 else if (attr->type == PERF_TYPE_BREAKPOINT)
6845 event->hw.bp_target = task;
6849 if (!overflow_handler && parent_event) {
6850 overflow_handler = parent_event->overflow_handler;
6851 context = parent_event->overflow_handler_context;
6854 event->overflow_handler = overflow_handler;
6855 event->overflow_handler_context = context;
6857 perf_event__state_init(event);
6862 hwc->sample_period = attr->sample_period;
6863 if (attr->freq && attr->sample_freq)
6864 hwc->sample_period = 1;
6865 hwc->last_period = hwc->sample_period;
6867 local64_set(&hwc->period_left, hwc->sample_period);
6870 * we currently do not support PERF_FORMAT_GROUP on inherited events
6872 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6875 pmu = perf_init_event(event);
6878 else if (IS_ERR(pmu)) {
6883 if (!event->parent) {
6884 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6885 err = get_callchain_buffers();
6895 event->destroy(event);
6896 module_put(pmu->module);
6899 put_pid_ns(event->ns);
6902 return ERR_PTR(err);
6905 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6906 struct perf_event_attr *attr)
6911 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6915 * zero the full structure, so that a short copy will be nice.
6917 memset(attr, 0, sizeof(*attr));
6919 ret = get_user(size, &uattr->size);
6923 if (size > PAGE_SIZE) /* silly large */
6926 if (!size) /* abi compat */
6927 size = PERF_ATTR_SIZE_VER0;
6929 if (size < PERF_ATTR_SIZE_VER0)
6933 * If we're handed a bigger struct than we know of,
6934 * ensure all the unknown bits are 0 - i.e. new
6935 * user-space does not rely on any kernel feature
6936 * extensions we dont know about yet.
6938 if (size > sizeof(*attr)) {
6939 unsigned char __user *addr;
6940 unsigned char __user *end;
6943 addr = (void __user *)uattr + sizeof(*attr);
6944 end = (void __user *)uattr + size;
6946 for (; addr < end; addr++) {
6947 ret = get_user(val, addr);
6953 size = sizeof(*attr);
6956 ret = copy_from_user(attr, uattr, size);
6960 if (attr->__reserved_1)
6963 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6966 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6969 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6970 u64 mask = attr->branch_sample_type;
6972 /* only using defined bits */
6973 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6976 /* at least one branch bit must be set */
6977 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6980 /* propagate priv level, when not set for branch */
6981 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6983 /* exclude_kernel checked on syscall entry */
6984 if (!attr->exclude_kernel)
6985 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6987 if (!attr->exclude_user)
6988 mask |= PERF_SAMPLE_BRANCH_USER;
6990 if (!attr->exclude_hv)
6991 mask |= PERF_SAMPLE_BRANCH_HV;
6993 * adjust user setting (for HW filter setup)
6995 attr->branch_sample_type = mask;
6997 /* privileged levels capture (kernel, hv): check permissions */
6998 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6999 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7003 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7004 ret = perf_reg_validate(attr->sample_regs_user);
7009 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7010 if (!arch_perf_have_user_stack_dump())
7014 * We have __u32 type for the size, but so far
7015 * we can only use __u16 as maximum due to the
7016 * __u16 sample size limit.
7018 if (attr->sample_stack_user >= USHRT_MAX)
7020 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7028 put_user(sizeof(*attr), &uattr->size);
7034 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7036 struct ring_buffer *rb = NULL;
7042 /* don't allow circular references */
7043 if (event == output_event)
7047 * Don't allow cross-cpu buffers
7049 if (output_event->cpu != event->cpu)
7053 * If its not a per-cpu rb, it must be the same task.
7055 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7059 mutex_lock(&event->mmap_mutex);
7060 /* Can't redirect output if we've got an active mmap() */
7061 if (atomic_read(&event->mmap_count))
7065 /* get the rb we want to redirect to */
7066 rb = ring_buffer_get(output_event);
7071 ring_buffer_attach(event, rb);
7075 mutex_unlock(&event->mmap_mutex);
7082 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7084 * @attr_uptr: event_id type attributes for monitoring/sampling
7087 * @group_fd: group leader event fd
7089 SYSCALL_DEFINE5(perf_event_open,
7090 struct perf_event_attr __user *, attr_uptr,
7091 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7093 struct perf_event *group_leader = NULL, *output_event = NULL;
7094 struct perf_event *event, *sibling;
7095 struct perf_event_attr attr;
7096 struct perf_event_context *ctx;
7097 struct file *event_file = NULL;
7098 struct fd group = {NULL, 0};
7099 struct task_struct *task = NULL;
7104 int f_flags = O_RDWR;
7106 /* for future expandability... */
7107 if (flags & ~PERF_FLAG_ALL)
7110 err = perf_copy_attr(attr_uptr, &attr);
7114 if (!attr.exclude_kernel) {
7115 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7120 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7123 if (attr.sample_period & (1ULL << 63))
7128 * In cgroup mode, the pid argument is used to pass the fd
7129 * opened to the cgroup directory in cgroupfs. The cpu argument
7130 * designates the cpu on which to monitor threads from that
7133 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7136 if (flags & PERF_FLAG_FD_CLOEXEC)
7137 f_flags |= O_CLOEXEC;
7139 event_fd = get_unused_fd_flags(f_flags);
7143 if (group_fd != -1) {
7144 err = perf_fget_light(group_fd, &group);
7147 group_leader = group.file->private_data;
7148 if (flags & PERF_FLAG_FD_OUTPUT)
7149 output_event = group_leader;
7150 if (flags & PERF_FLAG_FD_NO_GROUP)
7151 group_leader = NULL;
7154 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7155 task = find_lively_task_by_vpid(pid);
7157 err = PTR_ERR(task);
7162 if (task && group_leader &&
7163 group_leader->attr.inherit != attr.inherit) {
7170 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7172 if (IS_ERR(event)) {
7173 err = PTR_ERR(event);
7177 if (flags & PERF_FLAG_PID_CGROUP) {
7178 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7180 __free_event(event);
7185 if (is_sampling_event(event)) {
7186 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7192 account_event(event);
7195 * Special case software events and allow them to be part of
7196 * any hardware group.
7201 (is_software_event(event) != is_software_event(group_leader))) {
7202 if (is_software_event(event)) {
7204 * If event and group_leader are not both a software
7205 * event, and event is, then group leader is not.
7207 * Allow the addition of software events to !software
7208 * groups, this is safe because software events never
7211 pmu = group_leader->pmu;
7212 } else if (is_software_event(group_leader) &&
7213 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7215 * In case the group is a pure software group, and we
7216 * try to add a hardware event, move the whole group to
7217 * the hardware context.
7224 * Get the target context (task or percpu):
7226 ctx = find_get_context(pmu, task, event->cpu);
7233 put_task_struct(task);
7238 * Look up the group leader (we will attach this event to it):
7244 * Do not allow a recursive hierarchy (this new sibling
7245 * becoming part of another group-sibling):
7247 if (group_leader->group_leader != group_leader)
7250 * Do not allow to attach to a group in a different
7251 * task or CPU context:
7254 if (group_leader->ctx->type != ctx->type)
7257 if (group_leader->ctx != ctx)
7262 * Only a group leader can be exclusive or pinned
7264 if (attr.exclusive || attr.pinned)
7269 err = perf_event_set_output(event, output_event);
7274 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7276 if (IS_ERR(event_file)) {
7277 err = PTR_ERR(event_file);
7282 struct perf_event_context *gctx = group_leader->ctx;
7284 mutex_lock(&gctx->mutex);
7285 perf_remove_from_context(group_leader, false);
7288 * Removing from the context ends up with disabled
7289 * event. What we want here is event in the initial
7290 * startup state, ready to be add into new context.
7292 perf_event__state_init(group_leader);
7293 list_for_each_entry(sibling, &group_leader->sibling_list,
7295 perf_remove_from_context(sibling, false);
7296 perf_event__state_init(sibling);
7299 mutex_unlock(&gctx->mutex);
7303 WARN_ON_ONCE(ctx->parent_ctx);
7304 mutex_lock(&ctx->mutex);
7308 perf_install_in_context(ctx, group_leader, event->cpu);
7310 list_for_each_entry(sibling, &group_leader->sibling_list,
7312 perf_install_in_context(ctx, sibling, event->cpu);
7317 perf_install_in_context(ctx, event, event->cpu);
7318 perf_unpin_context(ctx);
7319 mutex_unlock(&ctx->mutex);
7323 event->owner = current;
7325 mutex_lock(¤t->perf_event_mutex);
7326 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7327 mutex_unlock(¤t->perf_event_mutex);
7330 * Precalculate sample_data sizes
7332 perf_event__header_size(event);
7333 perf_event__id_header_size(event);
7336 * Drop the reference on the group_event after placing the
7337 * new event on the sibling_list. This ensures destruction
7338 * of the group leader will find the pointer to itself in
7339 * perf_group_detach().
7342 fd_install(event_fd, event_file);
7346 perf_unpin_context(ctx);
7354 put_task_struct(task);
7358 put_unused_fd(event_fd);
7363 * perf_event_create_kernel_counter
7365 * @attr: attributes of the counter to create
7366 * @cpu: cpu in which the counter is bound
7367 * @task: task to profile (NULL for percpu)
7370 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7371 struct task_struct *task,
7372 perf_overflow_handler_t overflow_handler,
7375 struct perf_event_context *ctx;
7376 struct perf_event *event;
7380 * Get the target context (task or percpu):
7383 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7384 overflow_handler, context);
7385 if (IS_ERR(event)) {
7386 err = PTR_ERR(event);
7390 account_event(event);
7392 ctx = find_get_context(event->pmu, task, cpu);
7398 WARN_ON_ONCE(ctx->parent_ctx);
7399 mutex_lock(&ctx->mutex);
7400 perf_install_in_context(ctx, event, cpu);
7401 perf_unpin_context(ctx);
7402 mutex_unlock(&ctx->mutex);
7409 return ERR_PTR(err);
7411 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7413 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7415 struct perf_event_context *src_ctx;
7416 struct perf_event_context *dst_ctx;
7417 struct perf_event *event, *tmp;
7420 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7421 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7423 mutex_lock(&src_ctx->mutex);
7424 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7426 perf_remove_from_context(event, false);
7427 unaccount_event_cpu(event, src_cpu);
7429 list_add(&event->migrate_entry, &events);
7431 mutex_unlock(&src_ctx->mutex);
7435 mutex_lock(&dst_ctx->mutex);
7436 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7437 list_del(&event->migrate_entry);
7438 if (event->state >= PERF_EVENT_STATE_OFF)
7439 event->state = PERF_EVENT_STATE_INACTIVE;
7440 account_event_cpu(event, dst_cpu);
7441 perf_install_in_context(dst_ctx, event, dst_cpu);
7444 mutex_unlock(&dst_ctx->mutex);
7446 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7448 static void sync_child_event(struct perf_event *child_event,
7449 struct task_struct *child)
7451 struct perf_event *parent_event = child_event->parent;
7454 if (child_event->attr.inherit_stat)
7455 perf_event_read_event(child_event, child);
7457 child_val = perf_event_count(child_event);
7460 * Add back the child's count to the parent's count:
7462 atomic64_add(child_val, &parent_event->child_count);
7463 atomic64_add(child_event->total_time_enabled,
7464 &parent_event->child_total_time_enabled);
7465 atomic64_add(child_event->total_time_running,
7466 &parent_event->child_total_time_running);
7469 * Remove this event from the parent's list
7471 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7472 mutex_lock(&parent_event->child_mutex);
7473 list_del_init(&child_event->child_list);
7474 mutex_unlock(&parent_event->child_mutex);
7477 * Release the parent event, if this was the last
7480 put_event(parent_event);
7484 __perf_event_exit_task(struct perf_event *child_event,
7485 struct perf_event_context *child_ctx,
7486 struct task_struct *child)
7489 * Do not destroy the 'original' grouping; because of the context
7490 * switch optimization the original events could've ended up in a
7491 * random child task.
7493 * If we were to destroy the original group, all group related
7494 * operations would cease to function properly after this random
7497 * Do destroy all inherited groups, we don't care about those
7498 * and being thorough is better.
7500 perf_remove_from_context(child_event, !!child_event->parent);
7503 * It can happen that the parent exits first, and has events
7504 * that are still around due to the child reference. These
7505 * events need to be zapped.
7507 if (child_event->parent) {
7508 sync_child_event(child_event, child);
7509 free_event(child_event);
7513 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7515 struct perf_event *child_event, *next;
7516 struct perf_event_context *child_ctx, *parent_ctx;
7517 unsigned long flags;
7519 if (likely(!child->perf_event_ctxp[ctxn])) {
7520 perf_event_task(child, NULL, 0);
7524 local_irq_save(flags);
7526 * We can't reschedule here because interrupts are disabled,
7527 * and either child is current or it is a task that can't be
7528 * scheduled, so we are now safe from rescheduling changing
7531 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7534 * Take the context lock here so that if find_get_context is
7535 * reading child->perf_event_ctxp, we wait until it has
7536 * incremented the context's refcount before we do put_ctx below.
7538 raw_spin_lock(&child_ctx->lock);
7539 task_ctx_sched_out(child_ctx);
7540 child->perf_event_ctxp[ctxn] = NULL;
7543 * In order to avoid freeing: child_ctx->parent_ctx->task
7544 * under perf_event_context::lock, grab another reference.
7546 parent_ctx = child_ctx->parent_ctx;
7548 get_ctx(parent_ctx);
7551 * If this context is a clone; unclone it so it can't get
7552 * swapped to another process while we're removing all
7553 * the events from it.
7555 unclone_ctx(child_ctx);
7556 update_context_time(child_ctx);
7557 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7560 * Now that we no longer hold perf_event_context::lock, drop
7561 * our extra child_ctx->parent_ctx reference.
7564 put_ctx(parent_ctx);
7567 * Report the task dead after unscheduling the events so that we
7568 * won't get any samples after PERF_RECORD_EXIT. We can however still
7569 * get a few PERF_RECORD_READ events.
7571 perf_event_task(child, child_ctx, 0);
7574 * We can recurse on the same lock type through:
7576 * __perf_event_exit_task()
7577 * sync_child_event()
7579 * mutex_lock(&ctx->mutex)
7581 * But since its the parent context it won't be the same instance.
7583 mutex_lock(&child_ctx->mutex);
7585 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7586 __perf_event_exit_task(child_event, child_ctx, child);
7588 mutex_unlock(&child_ctx->mutex);
7594 * When a child task exits, feed back event values to parent events.
7596 void perf_event_exit_task(struct task_struct *child)
7598 struct perf_event *event, *tmp;
7601 mutex_lock(&child->perf_event_mutex);
7602 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7604 list_del_init(&event->owner_entry);
7607 * Ensure the list deletion is visible before we clear
7608 * the owner, closes a race against perf_release() where
7609 * we need to serialize on the owner->perf_event_mutex.
7612 event->owner = NULL;
7614 mutex_unlock(&child->perf_event_mutex);
7616 for_each_task_context_nr(ctxn)
7617 perf_event_exit_task_context(child, ctxn);
7620 static void perf_free_event(struct perf_event *event,
7621 struct perf_event_context *ctx)
7623 struct perf_event *parent = event->parent;
7625 if (WARN_ON_ONCE(!parent))
7628 mutex_lock(&parent->child_mutex);
7629 list_del_init(&event->child_list);
7630 mutex_unlock(&parent->child_mutex);
7634 perf_group_detach(event);
7635 list_del_event(event, ctx);
7640 * free an unexposed, unused context as created by inheritance by
7641 * perf_event_init_task below, used by fork() in case of fail.
7643 void perf_event_free_task(struct task_struct *task)
7645 struct perf_event_context *ctx;
7646 struct perf_event *event, *tmp;
7649 for_each_task_context_nr(ctxn) {
7650 ctx = task->perf_event_ctxp[ctxn];
7654 mutex_lock(&ctx->mutex);
7656 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7658 perf_free_event(event, ctx);
7660 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7662 perf_free_event(event, ctx);
7664 if (!list_empty(&ctx->pinned_groups) ||
7665 !list_empty(&ctx->flexible_groups))
7668 mutex_unlock(&ctx->mutex);
7674 void perf_event_delayed_put(struct task_struct *task)
7678 for_each_task_context_nr(ctxn)
7679 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7683 * inherit a event from parent task to child task:
7685 static struct perf_event *
7686 inherit_event(struct perf_event *parent_event,
7687 struct task_struct *parent,
7688 struct perf_event_context *parent_ctx,
7689 struct task_struct *child,
7690 struct perf_event *group_leader,
7691 struct perf_event_context *child_ctx)
7693 struct perf_event *child_event;
7694 unsigned long flags;
7697 * Instead of creating recursive hierarchies of events,
7698 * we link inherited events back to the original parent,
7699 * which has a filp for sure, which we use as the reference
7702 if (parent_event->parent)
7703 parent_event = parent_event->parent;
7705 child_event = perf_event_alloc(&parent_event->attr,
7708 group_leader, parent_event,
7710 if (IS_ERR(child_event))
7713 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7714 free_event(child_event);
7721 * Make the child state follow the state of the parent event,
7722 * not its attr.disabled bit. We hold the parent's mutex,
7723 * so we won't race with perf_event_{en, dis}able_family.
7725 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7726 child_event->state = PERF_EVENT_STATE_INACTIVE;
7728 child_event->state = PERF_EVENT_STATE_OFF;
7730 if (parent_event->attr.freq) {
7731 u64 sample_period = parent_event->hw.sample_period;
7732 struct hw_perf_event *hwc = &child_event->hw;
7734 hwc->sample_period = sample_period;
7735 hwc->last_period = sample_period;
7737 local64_set(&hwc->period_left, sample_period);
7740 child_event->ctx = child_ctx;
7741 child_event->overflow_handler = parent_event->overflow_handler;
7742 child_event->overflow_handler_context
7743 = parent_event->overflow_handler_context;
7746 * Precalculate sample_data sizes
7748 perf_event__header_size(child_event);
7749 perf_event__id_header_size(child_event);
7752 * Link it up in the child's context:
7754 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7755 add_event_to_ctx(child_event, child_ctx);
7756 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7759 * Link this into the parent event's child list
7761 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7762 mutex_lock(&parent_event->child_mutex);
7763 list_add_tail(&child_event->child_list, &parent_event->child_list);
7764 mutex_unlock(&parent_event->child_mutex);
7769 static int inherit_group(struct perf_event *parent_event,
7770 struct task_struct *parent,
7771 struct perf_event_context *parent_ctx,
7772 struct task_struct *child,
7773 struct perf_event_context *child_ctx)
7775 struct perf_event *leader;
7776 struct perf_event *sub;
7777 struct perf_event *child_ctr;
7779 leader = inherit_event(parent_event, parent, parent_ctx,
7780 child, NULL, child_ctx);
7782 return PTR_ERR(leader);
7783 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7784 child_ctr = inherit_event(sub, parent, parent_ctx,
7785 child, leader, child_ctx);
7786 if (IS_ERR(child_ctr))
7787 return PTR_ERR(child_ctr);
7793 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7794 struct perf_event_context *parent_ctx,
7795 struct task_struct *child, int ctxn,
7799 struct perf_event_context *child_ctx;
7801 if (!event->attr.inherit) {
7806 child_ctx = child->perf_event_ctxp[ctxn];
7809 * This is executed from the parent task context, so
7810 * inherit events that have been marked for cloning.
7811 * First allocate and initialize a context for the
7815 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7819 child->perf_event_ctxp[ctxn] = child_ctx;
7822 ret = inherit_group(event, parent, parent_ctx,
7832 * Initialize the perf_event context in task_struct
7834 static int perf_event_init_context(struct task_struct *child, int ctxn)
7836 struct perf_event_context *child_ctx, *parent_ctx;
7837 struct perf_event_context *cloned_ctx;
7838 struct perf_event *event;
7839 struct task_struct *parent = current;
7840 int inherited_all = 1;
7841 unsigned long flags;
7844 if (likely(!parent->perf_event_ctxp[ctxn]))
7848 * If the parent's context is a clone, pin it so it won't get
7851 parent_ctx = perf_pin_task_context(parent, ctxn);
7856 * No need to check if parent_ctx != NULL here; since we saw
7857 * it non-NULL earlier, the only reason for it to become NULL
7858 * is if we exit, and since we're currently in the middle of
7859 * a fork we can't be exiting at the same time.
7863 * Lock the parent list. No need to lock the child - not PID
7864 * hashed yet and not running, so nobody can access it.
7866 mutex_lock(&parent_ctx->mutex);
7869 * We dont have to disable NMIs - we are only looking at
7870 * the list, not manipulating it:
7872 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7873 ret = inherit_task_group(event, parent, parent_ctx,
7874 child, ctxn, &inherited_all);
7880 * We can't hold ctx->lock when iterating the ->flexible_group list due
7881 * to allocations, but we need to prevent rotation because
7882 * rotate_ctx() will change the list from interrupt context.
7884 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7885 parent_ctx->rotate_disable = 1;
7886 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7888 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7889 ret = inherit_task_group(event, parent, parent_ctx,
7890 child, ctxn, &inherited_all);
7895 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7896 parent_ctx->rotate_disable = 0;
7898 child_ctx = child->perf_event_ctxp[ctxn];
7900 if (child_ctx && inherited_all) {
7902 * Mark the child context as a clone of the parent
7903 * context, or of whatever the parent is a clone of.
7905 * Note that if the parent is a clone, the holding of
7906 * parent_ctx->lock avoids it from being uncloned.
7908 cloned_ctx = parent_ctx->parent_ctx;
7910 child_ctx->parent_ctx = cloned_ctx;
7911 child_ctx->parent_gen = parent_ctx->parent_gen;
7913 child_ctx->parent_ctx = parent_ctx;
7914 child_ctx->parent_gen = parent_ctx->generation;
7916 get_ctx(child_ctx->parent_ctx);
7919 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7920 mutex_unlock(&parent_ctx->mutex);
7922 perf_unpin_context(parent_ctx);
7923 put_ctx(parent_ctx);
7929 * Initialize the perf_event context in task_struct
7931 int perf_event_init_task(struct task_struct *child)
7935 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7936 mutex_init(&child->perf_event_mutex);
7937 INIT_LIST_HEAD(&child->perf_event_list);
7939 for_each_task_context_nr(ctxn) {
7940 ret = perf_event_init_context(child, ctxn);
7948 static void __init perf_event_init_all_cpus(void)
7950 struct swevent_htable *swhash;
7953 for_each_possible_cpu(cpu) {
7954 swhash = &per_cpu(swevent_htable, cpu);
7955 mutex_init(&swhash->hlist_mutex);
7956 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7960 static void perf_event_init_cpu(int cpu)
7962 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7964 mutex_lock(&swhash->hlist_mutex);
7965 swhash->online = true;
7966 if (swhash->hlist_refcount > 0) {
7967 struct swevent_hlist *hlist;
7969 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7971 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7973 mutex_unlock(&swhash->hlist_mutex);
7976 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7977 static void perf_pmu_rotate_stop(struct pmu *pmu)
7979 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7981 WARN_ON(!irqs_disabled());
7983 list_del_init(&cpuctx->rotation_list);
7986 static void __perf_event_exit_context(void *__info)
7988 struct remove_event re = { .detach_group = false };
7989 struct perf_event_context *ctx = __info;
7991 perf_pmu_rotate_stop(ctx->pmu);
7994 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
7995 __perf_remove_from_context(&re);
7999 static void perf_event_exit_cpu_context(int cpu)
8001 struct perf_event_context *ctx;
8005 idx = srcu_read_lock(&pmus_srcu);
8006 list_for_each_entry_rcu(pmu, &pmus, entry) {
8007 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8009 mutex_lock(&ctx->mutex);
8010 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8011 mutex_unlock(&ctx->mutex);
8013 srcu_read_unlock(&pmus_srcu, idx);
8016 static void perf_event_exit_cpu(int cpu)
8018 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8020 perf_event_exit_cpu_context(cpu);
8022 mutex_lock(&swhash->hlist_mutex);
8023 swhash->online = false;
8024 swevent_hlist_release(swhash);
8025 mutex_unlock(&swhash->hlist_mutex);
8028 static inline void perf_event_exit_cpu(int cpu) { }
8032 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8036 for_each_online_cpu(cpu)
8037 perf_event_exit_cpu(cpu);
8043 * Run the perf reboot notifier at the very last possible moment so that
8044 * the generic watchdog code runs as long as possible.
8046 static struct notifier_block perf_reboot_notifier = {
8047 .notifier_call = perf_reboot,
8048 .priority = INT_MIN,
8052 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8054 unsigned int cpu = (long)hcpu;
8056 switch (action & ~CPU_TASKS_FROZEN) {
8058 case CPU_UP_PREPARE:
8059 case CPU_DOWN_FAILED:
8060 perf_event_init_cpu(cpu);
8063 case CPU_UP_CANCELED:
8064 case CPU_DOWN_PREPARE:
8065 perf_event_exit_cpu(cpu);
8074 void __init perf_event_init(void)
8080 perf_event_init_all_cpus();
8081 init_srcu_struct(&pmus_srcu);
8082 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8083 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8084 perf_pmu_register(&perf_task_clock, NULL, -1);
8086 perf_cpu_notifier(perf_cpu_notify);
8087 register_reboot_notifier(&perf_reboot_notifier);
8089 ret = init_hw_breakpoint();
8090 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8092 /* do not patch jump label more than once per second */
8093 jump_label_rate_limit(&perf_sched_events, HZ);
8096 * Build time assertion that we keep the data_head at the intended
8097 * location. IOW, validation we got the __reserved[] size right.
8099 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8103 static int __init perf_event_sysfs_init(void)
8108 mutex_lock(&pmus_lock);
8110 ret = bus_register(&pmu_bus);
8114 list_for_each_entry(pmu, &pmus, entry) {
8115 if (!pmu->name || pmu->type < 0)
8118 ret = pmu_dev_alloc(pmu);
8119 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8121 pmu_bus_running = 1;
8125 mutex_unlock(&pmus_lock);
8129 device_initcall(perf_event_sysfs_init);
8131 #ifdef CONFIG_CGROUP_PERF
8132 static struct cgroup_subsys_state *
8133 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8135 struct perf_cgroup *jc;
8137 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8139 return ERR_PTR(-ENOMEM);
8141 jc->info = alloc_percpu(struct perf_cgroup_info);
8144 return ERR_PTR(-ENOMEM);
8150 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8152 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8154 free_percpu(jc->info);
8158 static int __perf_cgroup_move(void *info)
8160 struct task_struct *task = info;
8161 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8165 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8166 struct cgroup_taskset *tset)
8168 struct task_struct *task;
8170 cgroup_taskset_for_each(task, tset)
8171 task_function_call(task, __perf_cgroup_move, task);
8174 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8175 struct cgroup_subsys_state *old_css,
8176 struct task_struct *task)
8179 * cgroup_exit() is called in the copy_process() failure path.
8180 * Ignore this case since the task hasn't ran yet, this avoids
8181 * trying to poke a half freed task state from generic code.
8183 if (!(task->flags & PF_EXITING))
8186 task_function_call(task, __perf_cgroup_move, task);
8189 struct cgroup_subsys perf_event_cgrp_subsys = {
8190 .css_alloc = perf_cgroup_css_alloc,
8191 .css_free = perf_cgroup_css_free,
8192 .exit = perf_cgroup_exit,
8193 .attach = perf_cgroup_attach,
8195 #endif /* CONFIG_CGROUP_PERF */