2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
45 #include <asm/irq_regs.h>
47 struct remote_function_call {
48 struct task_struct *p;
49 int (*func)(void *info);
54 static void remote_function(void *data)
56 struct remote_function_call *tfc = data;
57 struct task_struct *p = tfc->p;
61 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
65 tfc->ret = tfc->func(tfc->info);
69 * task_function_call - call a function on the cpu on which a task runs
70 * @p: the task to evaluate
71 * @func: the function to be called
72 * @info: the function call argument
74 * Calls the function @func when the task is currently running. This might
75 * be on the current CPU, which just calls the function directly
77 * returns: @func return value, or
78 * -ESRCH - when the process isn't running
79 * -EAGAIN - when the process moved away
82 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
84 struct remote_function_call data = {
88 .ret = -ESRCH, /* No such (running) process */
92 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
98 * cpu_function_call - call a function on the cpu
99 * @func: the function to be called
100 * @info: the function call argument
102 * Calls the function @func on the remote cpu.
104 * returns: @func return value or -ENXIO when the cpu is offline
106 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
108 struct remote_function_call data = {
112 .ret = -ENXIO, /* No such CPU */
115 smp_call_function_single(cpu, remote_function, &data, 1);
120 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
121 PERF_FLAG_FD_OUTPUT |\
122 PERF_FLAG_PID_CGROUP)
125 * branch priv levels that need permission checks
127 #define PERF_SAMPLE_BRANCH_PERM_PLM \
128 (PERF_SAMPLE_BRANCH_KERNEL |\
129 PERF_SAMPLE_BRANCH_HV)
132 EVENT_FLEXIBLE = 0x1,
134 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
138 * perf_sched_events : >0 events exist
139 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
141 struct static_key_deferred perf_sched_events __read_mostly;
142 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
143 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
145 static atomic_t nr_mmap_events __read_mostly;
146 static atomic_t nr_comm_events __read_mostly;
147 static atomic_t nr_task_events __read_mostly;
148 static atomic_t nr_freq_events __read_mostly;
150 static LIST_HEAD(pmus);
151 static DEFINE_MUTEX(pmus_lock);
152 static struct srcu_struct pmus_srcu;
155 * perf event paranoia level:
156 * -1 - not paranoid at all
157 * 0 - disallow raw tracepoint access for unpriv
158 * 1 - disallow cpu events for unpriv
159 * 2 - disallow kernel profiling for unpriv
161 int sysctl_perf_event_paranoid __read_mostly = 1;
163 /* Minimum for 512 kiB + 1 user control page */
164 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
167 * max perf event sample rate
169 #define DEFAULT_MAX_SAMPLE_RATE 100000
170 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
171 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
173 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
175 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
176 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
178 static int perf_sample_allowed_ns __read_mostly =
179 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
181 void update_perf_cpu_limits(void)
183 u64 tmp = perf_sample_period_ns;
185 tmp *= sysctl_perf_cpu_time_max_percent;
187 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
190 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
192 int perf_proc_update_handler(struct ctl_table *table, int write,
193 void __user *buffer, size_t *lenp,
196 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
201 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
202 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
203 update_perf_cpu_limits();
208 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
210 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
211 void __user *buffer, size_t *lenp,
214 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
219 update_perf_cpu_limits();
225 * perf samples are done in some very critical code paths (NMIs).
226 * If they take too much CPU time, the system can lock up and not
227 * get any real work done. This will drop the sample rate when
228 * we detect that events are taking too long.
230 #define NR_ACCUMULATED_SAMPLES 128
231 static DEFINE_PER_CPU(u64, running_sample_length);
233 void perf_sample_event_took(u64 sample_len_ns)
235 u64 avg_local_sample_len;
236 u64 local_samples_len;
237 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
242 /* decay the counter by 1 average sample */
243 local_samples_len = __get_cpu_var(running_sample_length);
244 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
245 local_samples_len += sample_len_ns;
246 __get_cpu_var(running_sample_length) = local_samples_len;
249 * note: this will be biased artifically low until we have
250 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
251 * from having to maintain a count.
253 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
255 if (avg_local_sample_len <= allowed_ns)
258 if (max_samples_per_tick <= 1)
261 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
262 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
263 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
265 printk_ratelimited(KERN_WARNING
266 "perf samples too long (%lld > %lld), lowering "
267 "kernel.perf_event_max_sample_rate to %d\n",
268 avg_local_sample_len, allowed_ns,
269 sysctl_perf_event_sample_rate);
271 update_perf_cpu_limits();
274 static atomic64_t perf_event_id;
276 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
277 enum event_type_t event_type);
279 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
280 enum event_type_t event_type,
281 struct task_struct *task);
283 static void update_context_time(struct perf_event_context *ctx);
284 static u64 perf_event_time(struct perf_event *event);
286 void __weak perf_event_print_debug(void) { }
288 extern __weak const char *perf_pmu_name(void)
293 static inline u64 perf_clock(void)
295 return local_clock();
298 static inline struct perf_cpu_context *
299 __get_cpu_context(struct perf_event_context *ctx)
301 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
304 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
305 struct perf_event_context *ctx)
307 raw_spin_lock(&cpuctx->ctx.lock);
309 raw_spin_lock(&ctx->lock);
312 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
313 struct perf_event_context *ctx)
316 raw_spin_unlock(&ctx->lock);
317 raw_spin_unlock(&cpuctx->ctx.lock);
320 #ifdef CONFIG_CGROUP_PERF
323 * perf_cgroup_info keeps track of time_enabled for a cgroup.
324 * This is a per-cpu dynamically allocated data structure.
326 struct perf_cgroup_info {
332 struct cgroup_subsys_state css;
333 struct perf_cgroup_info __percpu *info;
337 * Must ensure cgroup is pinned (css_get) before calling
338 * this function. In other words, we cannot call this function
339 * if there is no cgroup event for the current CPU context.
341 static inline struct perf_cgroup *
342 perf_cgroup_from_task(struct task_struct *task)
344 return container_of(task_css(task, perf_subsys_id),
345 struct perf_cgroup, css);
349 perf_cgroup_match(struct perf_event *event)
351 struct perf_event_context *ctx = event->ctx;
352 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
354 /* @event doesn't care about cgroup */
358 /* wants specific cgroup scope but @cpuctx isn't associated with any */
363 * Cgroup scoping is recursive. An event enabled for a cgroup is
364 * also enabled for all its descendant cgroups. If @cpuctx's
365 * cgroup is a descendant of @event's (the test covers identity
366 * case), it's a match.
368 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
369 event->cgrp->css.cgroup);
372 static inline bool perf_tryget_cgroup(struct perf_event *event)
374 return css_tryget(&event->cgrp->css);
377 static inline void perf_put_cgroup(struct perf_event *event)
379 css_put(&event->cgrp->css);
382 static inline void perf_detach_cgroup(struct perf_event *event)
384 perf_put_cgroup(event);
388 static inline int is_cgroup_event(struct perf_event *event)
390 return event->cgrp != NULL;
393 static inline u64 perf_cgroup_event_time(struct perf_event *event)
395 struct perf_cgroup_info *t;
397 t = per_cpu_ptr(event->cgrp->info, event->cpu);
401 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
403 struct perf_cgroup_info *info;
408 info = this_cpu_ptr(cgrp->info);
410 info->time += now - info->timestamp;
411 info->timestamp = now;
414 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
416 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
418 __update_cgrp_time(cgrp_out);
421 static inline void update_cgrp_time_from_event(struct perf_event *event)
423 struct perf_cgroup *cgrp;
426 * ensure we access cgroup data only when needed and
427 * when we know the cgroup is pinned (css_get)
429 if (!is_cgroup_event(event))
432 cgrp = perf_cgroup_from_task(current);
434 * Do not update time when cgroup is not active
436 if (cgrp == event->cgrp)
437 __update_cgrp_time(event->cgrp);
441 perf_cgroup_set_timestamp(struct task_struct *task,
442 struct perf_event_context *ctx)
444 struct perf_cgroup *cgrp;
445 struct perf_cgroup_info *info;
448 * ctx->lock held by caller
449 * ensure we do not access cgroup data
450 * unless we have the cgroup pinned (css_get)
452 if (!task || !ctx->nr_cgroups)
455 cgrp = perf_cgroup_from_task(task);
456 info = this_cpu_ptr(cgrp->info);
457 info->timestamp = ctx->timestamp;
460 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
461 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
464 * reschedule events based on the cgroup constraint of task.
466 * mode SWOUT : schedule out everything
467 * mode SWIN : schedule in based on cgroup for next
469 void perf_cgroup_switch(struct task_struct *task, int mode)
471 struct perf_cpu_context *cpuctx;
476 * disable interrupts to avoid geting nr_cgroup
477 * changes via __perf_event_disable(). Also
480 local_irq_save(flags);
483 * we reschedule only in the presence of cgroup
484 * constrained events.
488 list_for_each_entry_rcu(pmu, &pmus, entry) {
489 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
490 if (cpuctx->unique_pmu != pmu)
491 continue; /* ensure we process each cpuctx once */
494 * perf_cgroup_events says at least one
495 * context on this CPU has cgroup events.
497 * ctx->nr_cgroups reports the number of cgroup
498 * events for a context.
500 if (cpuctx->ctx.nr_cgroups > 0) {
501 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
502 perf_pmu_disable(cpuctx->ctx.pmu);
504 if (mode & PERF_CGROUP_SWOUT) {
505 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
507 * must not be done before ctxswout due
508 * to event_filter_match() in event_sched_out()
513 if (mode & PERF_CGROUP_SWIN) {
514 WARN_ON_ONCE(cpuctx->cgrp);
516 * set cgrp before ctxsw in to allow
517 * event_filter_match() to not have to pass
520 cpuctx->cgrp = perf_cgroup_from_task(task);
521 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
523 perf_pmu_enable(cpuctx->ctx.pmu);
524 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
530 local_irq_restore(flags);
533 static inline void perf_cgroup_sched_out(struct task_struct *task,
534 struct task_struct *next)
536 struct perf_cgroup *cgrp1;
537 struct perf_cgroup *cgrp2 = NULL;
540 * we come here when we know perf_cgroup_events > 0
542 cgrp1 = perf_cgroup_from_task(task);
545 * next is NULL when called from perf_event_enable_on_exec()
546 * that will systematically cause a cgroup_switch()
549 cgrp2 = perf_cgroup_from_task(next);
552 * only schedule out current cgroup events if we know
553 * that we are switching to a different cgroup. Otherwise,
554 * do no touch the cgroup events.
557 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
560 static inline void perf_cgroup_sched_in(struct task_struct *prev,
561 struct task_struct *task)
563 struct perf_cgroup *cgrp1;
564 struct perf_cgroup *cgrp2 = NULL;
567 * we come here when we know perf_cgroup_events > 0
569 cgrp1 = perf_cgroup_from_task(task);
571 /* prev can never be NULL */
572 cgrp2 = perf_cgroup_from_task(prev);
575 * only need to schedule in cgroup events if we are changing
576 * cgroup during ctxsw. Cgroup events were not scheduled
577 * out of ctxsw out if that was not the case.
580 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
583 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
584 struct perf_event_attr *attr,
585 struct perf_event *group_leader)
587 struct perf_cgroup *cgrp;
588 struct cgroup_subsys_state *css;
589 struct fd f = fdget(fd);
597 css = css_from_dir(f.file->f_dentry, &perf_subsys);
603 cgrp = container_of(css, struct perf_cgroup, css);
606 /* must be done before we fput() the file */
607 if (!perf_tryget_cgroup(event)) {
614 * all events in a group must monitor
615 * the same cgroup because a task belongs
616 * to only one perf cgroup at a time
618 if (group_leader && group_leader->cgrp != cgrp) {
619 perf_detach_cgroup(event);
629 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
631 struct perf_cgroup_info *t;
632 t = per_cpu_ptr(event->cgrp->info, event->cpu);
633 event->shadow_ctx_time = now - t->timestamp;
637 perf_cgroup_defer_enabled(struct perf_event *event)
640 * when the current task's perf cgroup does not match
641 * the event's, we need to remember to call the
642 * perf_mark_enable() function the first time a task with
643 * a matching perf cgroup is scheduled in.
645 if (is_cgroup_event(event) && !perf_cgroup_match(event))
646 event->cgrp_defer_enabled = 1;
650 perf_cgroup_mark_enabled(struct perf_event *event,
651 struct perf_event_context *ctx)
653 struct perf_event *sub;
654 u64 tstamp = perf_event_time(event);
656 if (!event->cgrp_defer_enabled)
659 event->cgrp_defer_enabled = 0;
661 event->tstamp_enabled = tstamp - event->total_time_enabled;
662 list_for_each_entry(sub, &event->sibling_list, group_entry) {
663 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
664 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
665 sub->cgrp_defer_enabled = 0;
669 #else /* !CONFIG_CGROUP_PERF */
672 perf_cgroup_match(struct perf_event *event)
677 static inline void perf_detach_cgroup(struct perf_event *event)
680 static inline int is_cgroup_event(struct perf_event *event)
685 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
690 static inline void update_cgrp_time_from_event(struct perf_event *event)
694 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
698 static inline void perf_cgroup_sched_out(struct task_struct *task,
699 struct task_struct *next)
703 static inline void perf_cgroup_sched_in(struct task_struct *prev,
704 struct task_struct *task)
708 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
709 struct perf_event_attr *attr,
710 struct perf_event *group_leader)
716 perf_cgroup_set_timestamp(struct task_struct *task,
717 struct perf_event_context *ctx)
722 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
727 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
731 static inline u64 perf_cgroup_event_time(struct perf_event *event)
737 perf_cgroup_defer_enabled(struct perf_event *event)
742 perf_cgroup_mark_enabled(struct perf_event *event,
743 struct perf_event_context *ctx)
749 * set default to be dependent on timer tick just
752 #define PERF_CPU_HRTIMER (1000 / HZ)
754 * function must be called with interrupts disbled
756 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
758 struct perf_cpu_context *cpuctx;
759 enum hrtimer_restart ret = HRTIMER_NORESTART;
762 WARN_ON(!irqs_disabled());
764 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
766 rotations = perf_rotate_context(cpuctx);
769 * arm timer if needed
772 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
773 ret = HRTIMER_RESTART;
779 /* CPU is going down */
780 void perf_cpu_hrtimer_cancel(int cpu)
782 struct perf_cpu_context *cpuctx;
786 if (WARN_ON(cpu != smp_processor_id()))
789 local_irq_save(flags);
793 list_for_each_entry_rcu(pmu, &pmus, entry) {
794 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
796 if (pmu->task_ctx_nr == perf_sw_context)
799 hrtimer_cancel(&cpuctx->hrtimer);
804 local_irq_restore(flags);
807 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
809 struct hrtimer *hr = &cpuctx->hrtimer;
810 struct pmu *pmu = cpuctx->ctx.pmu;
813 /* no multiplexing needed for SW PMU */
814 if (pmu->task_ctx_nr == perf_sw_context)
818 * check default is sane, if not set then force to
819 * default interval (1/tick)
821 timer = pmu->hrtimer_interval_ms;
823 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
825 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
827 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
828 hr->function = perf_cpu_hrtimer_handler;
831 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
833 struct hrtimer *hr = &cpuctx->hrtimer;
834 struct pmu *pmu = cpuctx->ctx.pmu;
837 if (pmu->task_ctx_nr == perf_sw_context)
840 if (hrtimer_active(hr))
843 if (!hrtimer_callback_running(hr))
844 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
845 0, HRTIMER_MODE_REL_PINNED, 0);
848 void perf_pmu_disable(struct pmu *pmu)
850 int *count = this_cpu_ptr(pmu->pmu_disable_count);
852 pmu->pmu_disable(pmu);
855 void perf_pmu_enable(struct pmu *pmu)
857 int *count = this_cpu_ptr(pmu->pmu_disable_count);
859 pmu->pmu_enable(pmu);
862 static DEFINE_PER_CPU(struct list_head, rotation_list);
865 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
866 * because they're strictly cpu affine and rotate_start is called with IRQs
867 * disabled, while rotate_context is called from IRQ context.
869 static void perf_pmu_rotate_start(struct pmu *pmu)
871 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
872 struct list_head *head = &__get_cpu_var(rotation_list);
874 WARN_ON(!irqs_disabled());
876 if (list_empty(&cpuctx->rotation_list))
877 list_add(&cpuctx->rotation_list, head);
880 static void get_ctx(struct perf_event_context *ctx)
882 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
885 static void put_ctx(struct perf_event_context *ctx)
887 if (atomic_dec_and_test(&ctx->refcount)) {
889 put_ctx(ctx->parent_ctx);
891 put_task_struct(ctx->task);
892 kfree_rcu(ctx, rcu_head);
896 static void unclone_ctx(struct perf_event_context *ctx)
898 if (ctx->parent_ctx) {
899 put_ctx(ctx->parent_ctx);
900 ctx->parent_ctx = NULL;
905 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
908 * only top level events have the pid namespace they were created in
911 event = event->parent;
913 return task_tgid_nr_ns(p, event->ns);
916 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
919 * only top level events have the pid namespace they were created in
922 event = event->parent;
924 return task_pid_nr_ns(p, event->ns);
928 * If we inherit events we want to return the parent event id
931 static u64 primary_event_id(struct perf_event *event)
936 id = event->parent->id;
942 * Get the perf_event_context for a task and lock it.
943 * This has to cope with with the fact that until it is locked,
944 * the context could get moved to another task.
946 static struct perf_event_context *
947 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
949 struct perf_event_context *ctx;
953 * One of the few rules of preemptible RCU is that one cannot do
954 * rcu_read_unlock() while holding a scheduler (or nested) lock when
955 * part of the read side critical section was preemptible -- see
956 * rcu_read_unlock_special().
958 * Since ctx->lock nests under rq->lock we must ensure the entire read
959 * side critical section is non-preemptible.
963 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
966 * If this context is a clone of another, it might
967 * get swapped for another underneath us by
968 * perf_event_task_sched_out, though the
969 * rcu_read_lock() protects us from any context
970 * getting freed. Lock the context and check if it
971 * got swapped before we could get the lock, and retry
972 * if so. If we locked the right context, then it
973 * can't get swapped on us any more.
975 raw_spin_lock_irqsave(&ctx->lock, *flags);
976 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
977 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
983 if (!atomic_inc_not_zero(&ctx->refcount)) {
984 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
994 * Get the context for a task and increment its pin_count so it
995 * can't get swapped to another task. This also increments its
996 * reference count so that the context can't get freed.
998 static struct perf_event_context *
999 perf_pin_task_context(struct task_struct *task, int ctxn)
1001 struct perf_event_context *ctx;
1002 unsigned long flags;
1004 ctx = perf_lock_task_context(task, ctxn, &flags);
1007 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1012 static void perf_unpin_context(struct perf_event_context *ctx)
1014 unsigned long flags;
1016 raw_spin_lock_irqsave(&ctx->lock, flags);
1018 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1022 * Update the record of the current time in a context.
1024 static void update_context_time(struct perf_event_context *ctx)
1026 u64 now = perf_clock();
1028 ctx->time += now - ctx->timestamp;
1029 ctx->timestamp = now;
1032 static u64 perf_event_time(struct perf_event *event)
1034 struct perf_event_context *ctx = event->ctx;
1036 if (is_cgroup_event(event))
1037 return perf_cgroup_event_time(event);
1039 return ctx ? ctx->time : 0;
1043 * Update the total_time_enabled and total_time_running fields for a event.
1044 * The caller of this function needs to hold the ctx->lock.
1046 static void update_event_times(struct perf_event *event)
1048 struct perf_event_context *ctx = event->ctx;
1051 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1052 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1055 * in cgroup mode, time_enabled represents
1056 * the time the event was enabled AND active
1057 * tasks were in the monitored cgroup. This is
1058 * independent of the activity of the context as
1059 * there may be a mix of cgroup and non-cgroup events.
1061 * That is why we treat cgroup events differently
1064 if (is_cgroup_event(event))
1065 run_end = perf_cgroup_event_time(event);
1066 else if (ctx->is_active)
1067 run_end = ctx->time;
1069 run_end = event->tstamp_stopped;
1071 event->total_time_enabled = run_end - event->tstamp_enabled;
1073 if (event->state == PERF_EVENT_STATE_INACTIVE)
1074 run_end = event->tstamp_stopped;
1076 run_end = perf_event_time(event);
1078 event->total_time_running = run_end - event->tstamp_running;
1083 * Update total_time_enabled and total_time_running for all events in a group.
1085 static void update_group_times(struct perf_event *leader)
1087 struct perf_event *event;
1089 update_event_times(leader);
1090 list_for_each_entry(event, &leader->sibling_list, group_entry)
1091 update_event_times(event);
1094 static struct list_head *
1095 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1097 if (event->attr.pinned)
1098 return &ctx->pinned_groups;
1100 return &ctx->flexible_groups;
1104 * Add a event from the lists for its context.
1105 * Must be called with ctx->mutex and ctx->lock held.
1108 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1110 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1111 event->attach_state |= PERF_ATTACH_CONTEXT;
1114 * If we're a stand alone event or group leader, we go to the context
1115 * list, group events are kept attached to the group so that
1116 * perf_group_detach can, at all times, locate all siblings.
1118 if (event->group_leader == event) {
1119 struct list_head *list;
1121 if (is_software_event(event))
1122 event->group_flags |= PERF_GROUP_SOFTWARE;
1124 list = ctx_group_list(event, ctx);
1125 list_add_tail(&event->group_entry, list);
1128 if (is_cgroup_event(event))
1131 if (has_branch_stack(event))
1132 ctx->nr_branch_stack++;
1134 list_add_rcu(&event->event_entry, &ctx->event_list);
1135 if (!ctx->nr_events)
1136 perf_pmu_rotate_start(ctx->pmu);
1138 if (event->attr.inherit_stat)
1145 * Initialize event state based on the perf_event_attr::disabled.
1147 static inline void perf_event__state_init(struct perf_event *event)
1149 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1150 PERF_EVENT_STATE_INACTIVE;
1154 * Called at perf_event creation and when events are attached/detached from a
1157 static void perf_event__read_size(struct perf_event *event)
1159 int entry = sizeof(u64); /* value */
1163 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1164 size += sizeof(u64);
1166 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1167 size += sizeof(u64);
1169 if (event->attr.read_format & PERF_FORMAT_ID)
1170 entry += sizeof(u64);
1172 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1173 nr += event->group_leader->nr_siblings;
1174 size += sizeof(u64);
1178 event->read_size = size;
1181 static void perf_event__header_size(struct perf_event *event)
1183 struct perf_sample_data *data;
1184 u64 sample_type = event->attr.sample_type;
1187 perf_event__read_size(event);
1189 if (sample_type & PERF_SAMPLE_IP)
1190 size += sizeof(data->ip);
1192 if (sample_type & PERF_SAMPLE_ADDR)
1193 size += sizeof(data->addr);
1195 if (sample_type & PERF_SAMPLE_PERIOD)
1196 size += sizeof(data->period);
1198 if (sample_type & PERF_SAMPLE_WEIGHT)
1199 size += sizeof(data->weight);
1201 if (sample_type & PERF_SAMPLE_READ)
1202 size += event->read_size;
1204 if (sample_type & PERF_SAMPLE_DATA_SRC)
1205 size += sizeof(data->data_src.val);
1207 if (sample_type & PERF_SAMPLE_TRANSACTION)
1208 size += sizeof(data->txn);
1210 event->header_size = size;
1213 static void perf_event__id_header_size(struct perf_event *event)
1215 struct perf_sample_data *data;
1216 u64 sample_type = event->attr.sample_type;
1219 if (sample_type & PERF_SAMPLE_TID)
1220 size += sizeof(data->tid_entry);
1222 if (sample_type & PERF_SAMPLE_TIME)
1223 size += sizeof(data->time);
1225 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1226 size += sizeof(data->id);
1228 if (sample_type & PERF_SAMPLE_ID)
1229 size += sizeof(data->id);
1231 if (sample_type & PERF_SAMPLE_STREAM_ID)
1232 size += sizeof(data->stream_id);
1234 if (sample_type & PERF_SAMPLE_CPU)
1235 size += sizeof(data->cpu_entry);
1237 event->id_header_size = size;
1240 static void perf_group_attach(struct perf_event *event)
1242 struct perf_event *group_leader = event->group_leader, *pos;
1245 * We can have double attach due to group movement in perf_event_open.
1247 if (event->attach_state & PERF_ATTACH_GROUP)
1250 event->attach_state |= PERF_ATTACH_GROUP;
1252 if (group_leader == event)
1255 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1256 !is_software_event(event))
1257 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1259 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1260 group_leader->nr_siblings++;
1262 perf_event__header_size(group_leader);
1264 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1265 perf_event__header_size(pos);
1269 * Remove a event from the lists for its context.
1270 * Must be called with ctx->mutex and ctx->lock held.
1273 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1275 struct perf_cpu_context *cpuctx;
1277 * We can have double detach due to exit/hot-unplug + close.
1279 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1282 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1284 if (is_cgroup_event(event)) {
1286 cpuctx = __get_cpu_context(ctx);
1288 * if there are no more cgroup events
1289 * then cler cgrp to avoid stale pointer
1290 * in update_cgrp_time_from_cpuctx()
1292 if (!ctx->nr_cgroups)
1293 cpuctx->cgrp = NULL;
1296 if (has_branch_stack(event))
1297 ctx->nr_branch_stack--;
1300 if (event->attr.inherit_stat)
1303 list_del_rcu(&event->event_entry);
1305 if (event->group_leader == event)
1306 list_del_init(&event->group_entry);
1308 update_group_times(event);
1311 * If event was in error state, then keep it
1312 * that way, otherwise bogus counts will be
1313 * returned on read(). The only way to get out
1314 * of error state is by explicit re-enabling
1317 if (event->state > PERF_EVENT_STATE_OFF)
1318 event->state = PERF_EVENT_STATE_OFF;
1323 static void perf_group_detach(struct perf_event *event)
1325 struct perf_event *sibling, *tmp;
1326 struct list_head *list = NULL;
1329 * We can have double detach due to exit/hot-unplug + close.
1331 if (!(event->attach_state & PERF_ATTACH_GROUP))
1334 event->attach_state &= ~PERF_ATTACH_GROUP;
1337 * If this is a sibling, remove it from its group.
1339 if (event->group_leader != event) {
1340 list_del_init(&event->group_entry);
1341 event->group_leader->nr_siblings--;
1345 if (!list_empty(&event->group_entry))
1346 list = &event->group_entry;
1349 * If this was a group event with sibling events then
1350 * upgrade the siblings to singleton events by adding them
1351 * to whatever list we are on.
1353 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1355 list_move_tail(&sibling->group_entry, list);
1356 sibling->group_leader = sibling;
1358 /* Inherit group flags from the previous leader */
1359 sibling->group_flags = event->group_flags;
1363 perf_event__header_size(event->group_leader);
1365 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1366 perf_event__header_size(tmp);
1370 event_filter_match(struct perf_event *event)
1372 return (event->cpu == -1 || event->cpu == smp_processor_id())
1373 && perf_cgroup_match(event);
1377 event_sched_out(struct perf_event *event,
1378 struct perf_cpu_context *cpuctx,
1379 struct perf_event_context *ctx)
1381 u64 tstamp = perf_event_time(event);
1384 * An event which could not be activated because of
1385 * filter mismatch still needs to have its timings
1386 * maintained, otherwise bogus information is return
1387 * via read() for time_enabled, time_running:
1389 if (event->state == PERF_EVENT_STATE_INACTIVE
1390 && !event_filter_match(event)) {
1391 delta = tstamp - event->tstamp_stopped;
1392 event->tstamp_running += delta;
1393 event->tstamp_stopped = tstamp;
1396 if (event->state != PERF_EVENT_STATE_ACTIVE)
1399 event->state = PERF_EVENT_STATE_INACTIVE;
1400 if (event->pending_disable) {
1401 event->pending_disable = 0;
1402 event->state = PERF_EVENT_STATE_OFF;
1404 event->tstamp_stopped = tstamp;
1405 event->pmu->del(event, 0);
1408 if (!is_software_event(event))
1409 cpuctx->active_oncpu--;
1411 if (event->attr.freq && event->attr.sample_freq)
1413 if (event->attr.exclusive || !cpuctx->active_oncpu)
1414 cpuctx->exclusive = 0;
1418 group_sched_out(struct perf_event *group_event,
1419 struct perf_cpu_context *cpuctx,
1420 struct perf_event_context *ctx)
1422 struct perf_event *event;
1423 int state = group_event->state;
1425 event_sched_out(group_event, cpuctx, ctx);
1428 * Schedule out siblings (if any):
1430 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1431 event_sched_out(event, cpuctx, ctx);
1433 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1434 cpuctx->exclusive = 0;
1438 * Cross CPU call to remove a performance event
1440 * We disable the event on the hardware level first. After that we
1441 * remove it from the context list.
1443 static int __perf_remove_from_context(void *info)
1445 struct perf_event *event = info;
1446 struct perf_event_context *ctx = event->ctx;
1447 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1449 raw_spin_lock(&ctx->lock);
1450 event_sched_out(event, cpuctx, ctx);
1451 list_del_event(event, ctx);
1452 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1454 cpuctx->task_ctx = NULL;
1456 raw_spin_unlock(&ctx->lock);
1463 * Remove the event from a task's (or a CPU's) list of events.
1465 * CPU events are removed with a smp call. For task events we only
1466 * call when the task is on a CPU.
1468 * If event->ctx is a cloned context, callers must make sure that
1469 * every task struct that event->ctx->task could possibly point to
1470 * remains valid. This is OK when called from perf_release since
1471 * that only calls us on the top-level context, which can't be a clone.
1472 * When called from perf_event_exit_task, it's OK because the
1473 * context has been detached from its task.
1475 static void perf_remove_from_context(struct perf_event *event)
1477 struct perf_event_context *ctx = event->ctx;
1478 struct task_struct *task = ctx->task;
1480 lockdep_assert_held(&ctx->mutex);
1484 * Per cpu events are removed via an smp call and
1485 * the removal is always successful.
1487 cpu_function_call(event->cpu, __perf_remove_from_context, event);
1492 if (!task_function_call(task, __perf_remove_from_context, event))
1495 raw_spin_lock_irq(&ctx->lock);
1497 * If we failed to find a running task, but find the context active now
1498 * that we've acquired the ctx->lock, retry.
1500 if (ctx->is_active) {
1501 raw_spin_unlock_irq(&ctx->lock);
1506 * Since the task isn't running, its safe to remove the event, us
1507 * holding the ctx->lock ensures the task won't get scheduled in.
1509 list_del_event(event, ctx);
1510 raw_spin_unlock_irq(&ctx->lock);
1514 * Cross CPU call to disable a performance event
1516 int __perf_event_disable(void *info)
1518 struct perf_event *event = info;
1519 struct perf_event_context *ctx = event->ctx;
1520 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1523 * If this is a per-task event, need to check whether this
1524 * event's task is the current task on this cpu.
1526 * Can trigger due to concurrent perf_event_context_sched_out()
1527 * flipping contexts around.
1529 if (ctx->task && cpuctx->task_ctx != ctx)
1532 raw_spin_lock(&ctx->lock);
1535 * If the event is on, turn it off.
1536 * If it is in error state, leave it in error state.
1538 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1539 update_context_time(ctx);
1540 update_cgrp_time_from_event(event);
1541 update_group_times(event);
1542 if (event == event->group_leader)
1543 group_sched_out(event, cpuctx, ctx);
1545 event_sched_out(event, cpuctx, ctx);
1546 event->state = PERF_EVENT_STATE_OFF;
1549 raw_spin_unlock(&ctx->lock);
1557 * If event->ctx is a cloned context, callers must make sure that
1558 * every task struct that event->ctx->task could possibly point to
1559 * remains valid. This condition is satisifed when called through
1560 * perf_event_for_each_child or perf_event_for_each because they
1561 * hold the top-level event's child_mutex, so any descendant that
1562 * goes to exit will block in sync_child_event.
1563 * When called from perf_pending_event it's OK because event->ctx
1564 * is the current context on this CPU and preemption is disabled,
1565 * hence we can't get into perf_event_task_sched_out for this context.
1567 void perf_event_disable(struct perf_event *event)
1569 struct perf_event_context *ctx = event->ctx;
1570 struct task_struct *task = ctx->task;
1574 * Disable the event on the cpu that it's on
1576 cpu_function_call(event->cpu, __perf_event_disable, event);
1581 if (!task_function_call(task, __perf_event_disable, event))
1584 raw_spin_lock_irq(&ctx->lock);
1586 * If the event is still active, we need to retry the cross-call.
1588 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1589 raw_spin_unlock_irq(&ctx->lock);
1591 * Reload the task pointer, it might have been changed by
1592 * a concurrent perf_event_context_sched_out().
1599 * Since we have the lock this context can't be scheduled
1600 * in, so we can change the state safely.
1602 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1603 update_group_times(event);
1604 event->state = PERF_EVENT_STATE_OFF;
1606 raw_spin_unlock_irq(&ctx->lock);
1608 EXPORT_SYMBOL_GPL(perf_event_disable);
1610 static void perf_set_shadow_time(struct perf_event *event,
1611 struct perf_event_context *ctx,
1615 * use the correct time source for the time snapshot
1617 * We could get by without this by leveraging the
1618 * fact that to get to this function, the caller
1619 * has most likely already called update_context_time()
1620 * and update_cgrp_time_xx() and thus both timestamp
1621 * are identical (or very close). Given that tstamp is,
1622 * already adjusted for cgroup, we could say that:
1623 * tstamp - ctx->timestamp
1625 * tstamp - cgrp->timestamp.
1627 * Then, in perf_output_read(), the calculation would
1628 * work with no changes because:
1629 * - event is guaranteed scheduled in
1630 * - no scheduled out in between
1631 * - thus the timestamp would be the same
1633 * But this is a bit hairy.
1635 * So instead, we have an explicit cgroup call to remain
1636 * within the time time source all along. We believe it
1637 * is cleaner and simpler to understand.
1639 if (is_cgroup_event(event))
1640 perf_cgroup_set_shadow_time(event, tstamp);
1642 event->shadow_ctx_time = tstamp - ctx->timestamp;
1645 #define MAX_INTERRUPTS (~0ULL)
1647 static void perf_log_throttle(struct perf_event *event, int enable);
1650 event_sched_in(struct perf_event *event,
1651 struct perf_cpu_context *cpuctx,
1652 struct perf_event_context *ctx)
1654 u64 tstamp = perf_event_time(event);
1656 if (event->state <= PERF_EVENT_STATE_OFF)
1659 event->state = PERF_EVENT_STATE_ACTIVE;
1660 event->oncpu = smp_processor_id();
1663 * Unthrottle events, since we scheduled we might have missed several
1664 * ticks already, also for a heavily scheduling task there is little
1665 * guarantee it'll get a tick in a timely manner.
1667 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1668 perf_log_throttle(event, 1);
1669 event->hw.interrupts = 0;
1673 * The new state must be visible before we turn it on in the hardware:
1677 if (event->pmu->add(event, PERF_EF_START)) {
1678 event->state = PERF_EVENT_STATE_INACTIVE;
1683 event->tstamp_running += tstamp - event->tstamp_stopped;
1685 perf_set_shadow_time(event, ctx, tstamp);
1687 if (!is_software_event(event))
1688 cpuctx->active_oncpu++;
1690 if (event->attr.freq && event->attr.sample_freq)
1693 if (event->attr.exclusive)
1694 cpuctx->exclusive = 1;
1700 group_sched_in(struct perf_event *group_event,
1701 struct perf_cpu_context *cpuctx,
1702 struct perf_event_context *ctx)
1704 struct perf_event *event, *partial_group = NULL;
1705 struct pmu *pmu = group_event->pmu;
1706 u64 now = ctx->time;
1707 bool simulate = false;
1709 if (group_event->state == PERF_EVENT_STATE_OFF)
1712 pmu->start_txn(pmu);
1714 if (event_sched_in(group_event, cpuctx, ctx)) {
1715 pmu->cancel_txn(pmu);
1716 perf_cpu_hrtimer_restart(cpuctx);
1721 * Schedule in siblings as one group (if any):
1723 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1724 if (event_sched_in(event, cpuctx, ctx)) {
1725 partial_group = event;
1730 if (!pmu->commit_txn(pmu))
1735 * Groups can be scheduled in as one unit only, so undo any
1736 * partial group before returning:
1737 * The events up to the failed event are scheduled out normally,
1738 * tstamp_stopped will be updated.
1740 * The failed events and the remaining siblings need to have
1741 * their timings updated as if they had gone thru event_sched_in()
1742 * and event_sched_out(). This is required to get consistent timings
1743 * across the group. This also takes care of the case where the group
1744 * could never be scheduled by ensuring tstamp_stopped is set to mark
1745 * the time the event was actually stopped, such that time delta
1746 * calculation in update_event_times() is correct.
1748 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1749 if (event == partial_group)
1753 event->tstamp_running += now - event->tstamp_stopped;
1754 event->tstamp_stopped = now;
1756 event_sched_out(event, cpuctx, ctx);
1759 event_sched_out(group_event, cpuctx, ctx);
1761 pmu->cancel_txn(pmu);
1763 perf_cpu_hrtimer_restart(cpuctx);
1769 * Work out whether we can put this event group on the CPU now.
1771 static int group_can_go_on(struct perf_event *event,
1772 struct perf_cpu_context *cpuctx,
1776 * Groups consisting entirely of software events can always go on.
1778 if (event->group_flags & PERF_GROUP_SOFTWARE)
1781 * If an exclusive group is already on, no other hardware
1784 if (cpuctx->exclusive)
1787 * If this group is exclusive and there are already
1788 * events on the CPU, it can't go on.
1790 if (event->attr.exclusive && cpuctx->active_oncpu)
1793 * Otherwise, try to add it if all previous groups were able
1799 static void add_event_to_ctx(struct perf_event *event,
1800 struct perf_event_context *ctx)
1802 u64 tstamp = perf_event_time(event);
1804 list_add_event(event, ctx);
1805 perf_group_attach(event);
1806 event->tstamp_enabled = tstamp;
1807 event->tstamp_running = tstamp;
1808 event->tstamp_stopped = tstamp;
1811 static void task_ctx_sched_out(struct perf_event_context *ctx);
1813 ctx_sched_in(struct perf_event_context *ctx,
1814 struct perf_cpu_context *cpuctx,
1815 enum event_type_t event_type,
1816 struct task_struct *task);
1818 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1819 struct perf_event_context *ctx,
1820 struct task_struct *task)
1822 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1824 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1825 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1827 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1831 * Cross CPU call to install and enable a performance event
1833 * Must be called with ctx->mutex held
1835 static int __perf_install_in_context(void *info)
1837 struct perf_event *event = info;
1838 struct perf_event_context *ctx = event->ctx;
1839 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1840 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1841 struct task_struct *task = current;
1843 perf_ctx_lock(cpuctx, task_ctx);
1844 perf_pmu_disable(cpuctx->ctx.pmu);
1847 * If there was an active task_ctx schedule it out.
1850 task_ctx_sched_out(task_ctx);
1853 * If the context we're installing events in is not the
1854 * active task_ctx, flip them.
1856 if (ctx->task && task_ctx != ctx) {
1858 raw_spin_unlock(&task_ctx->lock);
1859 raw_spin_lock(&ctx->lock);
1864 cpuctx->task_ctx = task_ctx;
1865 task = task_ctx->task;
1868 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1870 update_context_time(ctx);
1872 * update cgrp time only if current cgrp
1873 * matches event->cgrp. Must be done before
1874 * calling add_event_to_ctx()
1876 update_cgrp_time_from_event(event);
1878 add_event_to_ctx(event, ctx);
1881 * Schedule everything back in
1883 perf_event_sched_in(cpuctx, task_ctx, task);
1885 perf_pmu_enable(cpuctx->ctx.pmu);
1886 perf_ctx_unlock(cpuctx, task_ctx);
1892 * Attach a performance event to a context
1894 * First we add the event to the list with the hardware enable bit
1895 * in event->hw_config cleared.
1897 * If the event is attached to a task which is on a CPU we use a smp
1898 * call to enable it in the task context. The task might have been
1899 * scheduled away, but we check this in the smp call again.
1902 perf_install_in_context(struct perf_event_context *ctx,
1903 struct perf_event *event,
1906 struct task_struct *task = ctx->task;
1908 lockdep_assert_held(&ctx->mutex);
1911 if (event->cpu != -1)
1916 * Per cpu events are installed via an smp call and
1917 * the install is always successful.
1919 cpu_function_call(cpu, __perf_install_in_context, event);
1924 if (!task_function_call(task, __perf_install_in_context, event))
1927 raw_spin_lock_irq(&ctx->lock);
1929 * If we failed to find a running task, but find the context active now
1930 * that we've acquired the ctx->lock, retry.
1932 if (ctx->is_active) {
1933 raw_spin_unlock_irq(&ctx->lock);
1938 * Since the task isn't running, its safe to add the event, us holding
1939 * the ctx->lock ensures the task won't get scheduled in.
1941 add_event_to_ctx(event, ctx);
1942 raw_spin_unlock_irq(&ctx->lock);
1946 * Put a event into inactive state and update time fields.
1947 * Enabling the leader of a group effectively enables all
1948 * the group members that aren't explicitly disabled, so we
1949 * have to update their ->tstamp_enabled also.
1950 * Note: this works for group members as well as group leaders
1951 * since the non-leader members' sibling_lists will be empty.
1953 static void __perf_event_mark_enabled(struct perf_event *event)
1955 struct perf_event *sub;
1956 u64 tstamp = perf_event_time(event);
1958 event->state = PERF_EVENT_STATE_INACTIVE;
1959 event->tstamp_enabled = tstamp - event->total_time_enabled;
1960 list_for_each_entry(sub, &event->sibling_list, group_entry) {
1961 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
1962 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
1967 * Cross CPU call to enable a performance event
1969 static int __perf_event_enable(void *info)
1971 struct perf_event *event = info;
1972 struct perf_event_context *ctx = event->ctx;
1973 struct perf_event *leader = event->group_leader;
1974 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1978 * There's a time window between 'ctx->is_active' check
1979 * in perf_event_enable function and this place having:
1981 * - ctx->lock unlocked
1983 * where the task could be killed and 'ctx' deactivated
1984 * by perf_event_exit_task.
1986 if (!ctx->is_active)
1989 raw_spin_lock(&ctx->lock);
1990 update_context_time(ctx);
1992 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1996 * set current task's cgroup time reference point
1998 perf_cgroup_set_timestamp(current, ctx);
2000 __perf_event_mark_enabled(event);
2002 if (!event_filter_match(event)) {
2003 if (is_cgroup_event(event))
2004 perf_cgroup_defer_enabled(event);
2009 * If the event is in a group and isn't the group leader,
2010 * then don't put it on unless the group is on.
2012 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2015 if (!group_can_go_on(event, cpuctx, 1)) {
2018 if (event == leader)
2019 err = group_sched_in(event, cpuctx, ctx);
2021 err = event_sched_in(event, cpuctx, ctx);
2026 * If this event can't go on and it's part of a
2027 * group, then the whole group has to come off.
2029 if (leader != event) {
2030 group_sched_out(leader, cpuctx, ctx);
2031 perf_cpu_hrtimer_restart(cpuctx);
2033 if (leader->attr.pinned) {
2034 update_group_times(leader);
2035 leader->state = PERF_EVENT_STATE_ERROR;
2040 raw_spin_unlock(&ctx->lock);
2048 * If event->ctx is a cloned context, callers must make sure that
2049 * every task struct that event->ctx->task could possibly point to
2050 * remains valid. This condition is satisfied when called through
2051 * perf_event_for_each_child or perf_event_for_each as described
2052 * for perf_event_disable.
2054 void perf_event_enable(struct perf_event *event)
2056 struct perf_event_context *ctx = event->ctx;
2057 struct task_struct *task = ctx->task;
2061 * Enable the event on the cpu that it's on
2063 cpu_function_call(event->cpu, __perf_event_enable, event);
2067 raw_spin_lock_irq(&ctx->lock);
2068 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2072 * If the event is in error state, clear that first.
2073 * That way, if we see the event in error state below, we
2074 * know that it has gone back into error state, as distinct
2075 * from the task having been scheduled away before the
2076 * cross-call arrived.
2078 if (event->state == PERF_EVENT_STATE_ERROR)
2079 event->state = PERF_EVENT_STATE_OFF;
2082 if (!ctx->is_active) {
2083 __perf_event_mark_enabled(event);
2087 raw_spin_unlock_irq(&ctx->lock);
2089 if (!task_function_call(task, __perf_event_enable, event))
2092 raw_spin_lock_irq(&ctx->lock);
2095 * If the context is active and the event is still off,
2096 * we need to retry the cross-call.
2098 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2100 * task could have been flipped by a concurrent
2101 * perf_event_context_sched_out()
2108 raw_spin_unlock_irq(&ctx->lock);
2110 EXPORT_SYMBOL_GPL(perf_event_enable);
2112 int perf_event_refresh(struct perf_event *event, int refresh)
2115 * not supported on inherited events
2117 if (event->attr.inherit || !is_sampling_event(event))
2120 atomic_add(refresh, &event->event_limit);
2121 perf_event_enable(event);
2125 EXPORT_SYMBOL_GPL(perf_event_refresh);
2127 static void ctx_sched_out(struct perf_event_context *ctx,
2128 struct perf_cpu_context *cpuctx,
2129 enum event_type_t event_type)
2131 struct perf_event *event;
2132 int is_active = ctx->is_active;
2134 ctx->is_active &= ~event_type;
2135 if (likely(!ctx->nr_events))
2138 update_context_time(ctx);
2139 update_cgrp_time_from_cpuctx(cpuctx);
2140 if (!ctx->nr_active)
2143 perf_pmu_disable(ctx->pmu);
2144 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2145 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2146 group_sched_out(event, cpuctx, ctx);
2149 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2150 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2151 group_sched_out(event, cpuctx, ctx);
2153 perf_pmu_enable(ctx->pmu);
2157 * Test whether two contexts are equivalent, i.e. whether they have both been
2158 * cloned from the same version of the same context.
2160 * Equivalence is measured using a generation number in the context that is
2161 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2162 * and list_del_event().
2164 static int context_equiv(struct perf_event_context *ctx1,
2165 struct perf_event_context *ctx2)
2167 /* Pinning disables the swap optimization */
2168 if (ctx1->pin_count || ctx2->pin_count)
2171 /* If ctx1 is the parent of ctx2 */
2172 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2175 /* If ctx2 is the parent of ctx1 */
2176 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2180 * If ctx1 and ctx2 have the same parent; we flatten the parent
2181 * hierarchy, see perf_event_init_context().
2183 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2184 ctx1->parent_gen == ctx2->parent_gen)
2191 static void __perf_event_sync_stat(struct perf_event *event,
2192 struct perf_event *next_event)
2196 if (!event->attr.inherit_stat)
2200 * Update the event value, we cannot use perf_event_read()
2201 * because we're in the middle of a context switch and have IRQs
2202 * disabled, which upsets smp_call_function_single(), however
2203 * we know the event must be on the current CPU, therefore we
2204 * don't need to use it.
2206 switch (event->state) {
2207 case PERF_EVENT_STATE_ACTIVE:
2208 event->pmu->read(event);
2211 case PERF_EVENT_STATE_INACTIVE:
2212 update_event_times(event);
2220 * In order to keep per-task stats reliable we need to flip the event
2221 * values when we flip the contexts.
2223 value = local64_read(&next_event->count);
2224 value = local64_xchg(&event->count, value);
2225 local64_set(&next_event->count, value);
2227 swap(event->total_time_enabled, next_event->total_time_enabled);
2228 swap(event->total_time_running, next_event->total_time_running);
2231 * Since we swizzled the values, update the user visible data too.
2233 perf_event_update_userpage(event);
2234 perf_event_update_userpage(next_event);
2237 #define list_next_entry(pos, member) \
2238 list_entry(pos->member.next, typeof(*pos), member)
2240 static void perf_event_sync_stat(struct perf_event_context *ctx,
2241 struct perf_event_context *next_ctx)
2243 struct perf_event *event, *next_event;
2248 update_context_time(ctx);
2250 event = list_first_entry(&ctx->event_list,
2251 struct perf_event, event_entry);
2253 next_event = list_first_entry(&next_ctx->event_list,
2254 struct perf_event, event_entry);
2256 while (&event->event_entry != &ctx->event_list &&
2257 &next_event->event_entry != &next_ctx->event_list) {
2259 __perf_event_sync_stat(event, next_event);
2261 event = list_next_entry(event, event_entry);
2262 next_event = list_next_entry(next_event, event_entry);
2266 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2267 struct task_struct *next)
2269 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2270 struct perf_event_context *next_ctx;
2271 struct perf_event_context *parent, *next_parent;
2272 struct perf_cpu_context *cpuctx;
2278 cpuctx = __get_cpu_context(ctx);
2279 if (!cpuctx->task_ctx)
2283 next_ctx = next->perf_event_ctxp[ctxn];
2287 parent = rcu_dereference(ctx->parent_ctx);
2288 next_parent = rcu_dereference(next_ctx->parent_ctx);
2290 /* If neither context have a parent context; they cannot be clones. */
2291 if (!parent && !next_parent)
2294 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2296 * Looks like the two contexts are clones, so we might be
2297 * able to optimize the context switch. We lock both
2298 * contexts and check that they are clones under the
2299 * lock (including re-checking that neither has been
2300 * uncloned in the meantime). It doesn't matter which
2301 * order we take the locks because no other cpu could
2302 * be trying to lock both of these tasks.
2304 raw_spin_lock(&ctx->lock);
2305 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2306 if (context_equiv(ctx, next_ctx)) {
2308 * XXX do we need a memory barrier of sorts
2309 * wrt to rcu_dereference() of perf_event_ctxp
2311 task->perf_event_ctxp[ctxn] = next_ctx;
2312 next->perf_event_ctxp[ctxn] = ctx;
2314 next_ctx->task = task;
2317 perf_event_sync_stat(ctx, next_ctx);
2319 raw_spin_unlock(&next_ctx->lock);
2320 raw_spin_unlock(&ctx->lock);
2326 raw_spin_lock(&ctx->lock);
2327 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2328 cpuctx->task_ctx = NULL;
2329 raw_spin_unlock(&ctx->lock);
2333 #define for_each_task_context_nr(ctxn) \
2334 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2337 * Called from scheduler to remove the events of the current task,
2338 * with interrupts disabled.
2340 * We stop each event and update the event value in event->count.
2342 * This does not protect us against NMI, but disable()
2343 * sets the disabled bit in the control field of event _before_
2344 * accessing the event control register. If a NMI hits, then it will
2345 * not restart the event.
2347 void __perf_event_task_sched_out(struct task_struct *task,
2348 struct task_struct *next)
2352 for_each_task_context_nr(ctxn)
2353 perf_event_context_sched_out(task, ctxn, next);
2356 * if cgroup events exist on this CPU, then we need
2357 * to check if we have to switch out PMU state.
2358 * cgroup event are system-wide mode only
2360 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2361 perf_cgroup_sched_out(task, next);
2364 static void task_ctx_sched_out(struct perf_event_context *ctx)
2366 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2368 if (!cpuctx->task_ctx)
2371 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2374 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2375 cpuctx->task_ctx = NULL;
2379 * Called with IRQs disabled
2381 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2382 enum event_type_t event_type)
2384 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2388 ctx_pinned_sched_in(struct perf_event_context *ctx,
2389 struct perf_cpu_context *cpuctx)
2391 struct perf_event *event;
2393 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2394 if (event->state <= PERF_EVENT_STATE_OFF)
2396 if (!event_filter_match(event))
2399 /* may need to reset tstamp_enabled */
2400 if (is_cgroup_event(event))
2401 perf_cgroup_mark_enabled(event, ctx);
2403 if (group_can_go_on(event, cpuctx, 1))
2404 group_sched_in(event, cpuctx, ctx);
2407 * If this pinned group hasn't been scheduled,
2408 * put it in error state.
2410 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2411 update_group_times(event);
2412 event->state = PERF_EVENT_STATE_ERROR;
2418 ctx_flexible_sched_in(struct perf_event_context *ctx,
2419 struct perf_cpu_context *cpuctx)
2421 struct perf_event *event;
2424 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2425 /* Ignore events in OFF or ERROR state */
2426 if (event->state <= PERF_EVENT_STATE_OFF)
2429 * Listen to the 'cpu' scheduling filter constraint
2432 if (!event_filter_match(event))
2435 /* may need to reset tstamp_enabled */
2436 if (is_cgroup_event(event))
2437 perf_cgroup_mark_enabled(event, ctx);
2439 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2440 if (group_sched_in(event, cpuctx, ctx))
2447 ctx_sched_in(struct perf_event_context *ctx,
2448 struct perf_cpu_context *cpuctx,
2449 enum event_type_t event_type,
2450 struct task_struct *task)
2453 int is_active = ctx->is_active;
2455 ctx->is_active |= event_type;
2456 if (likely(!ctx->nr_events))
2460 ctx->timestamp = now;
2461 perf_cgroup_set_timestamp(task, ctx);
2463 * First go through the list and put on any pinned groups
2464 * in order to give them the best chance of going on.
2466 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2467 ctx_pinned_sched_in(ctx, cpuctx);
2469 /* Then walk through the lower prio flexible groups */
2470 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2471 ctx_flexible_sched_in(ctx, cpuctx);
2474 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2475 enum event_type_t event_type,
2476 struct task_struct *task)
2478 struct perf_event_context *ctx = &cpuctx->ctx;
2480 ctx_sched_in(ctx, cpuctx, event_type, task);
2483 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2484 struct task_struct *task)
2486 struct perf_cpu_context *cpuctx;
2488 cpuctx = __get_cpu_context(ctx);
2489 if (cpuctx->task_ctx == ctx)
2492 perf_ctx_lock(cpuctx, ctx);
2493 perf_pmu_disable(ctx->pmu);
2495 * We want to keep the following priority order:
2496 * cpu pinned (that don't need to move), task pinned,
2497 * cpu flexible, task flexible.
2499 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2502 cpuctx->task_ctx = ctx;
2504 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2506 perf_pmu_enable(ctx->pmu);
2507 perf_ctx_unlock(cpuctx, ctx);
2510 * Since these rotations are per-cpu, we need to ensure the
2511 * cpu-context we got scheduled on is actually rotating.
2513 perf_pmu_rotate_start(ctx->pmu);
2517 * When sampling the branck stack in system-wide, it may be necessary
2518 * to flush the stack on context switch. This happens when the branch
2519 * stack does not tag its entries with the pid of the current task.
2520 * Otherwise it becomes impossible to associate a branch entry with a
2521 * task. This ambiguity is more likely to appear when the branch stack
2522 * supports priv level filtering and the user sets it to monitor only
2523 * at the user level (which could be a useful measurement in system-wide
2524 * mode). In that case, the risk is high of having a branch stack with
2525 * branch from multiple tasks. Flushing may mean dropping the existing
2526 * entries or stashing them somewhere in the PMU specific code layer.
2528 * This function provides the context switch callback to the lower code
2529 * layer. It is invoked ONLY when there is at least one system-wide context
2530 * with at least one active event using taken branch sampling.
2532 static void perf_branch_stack_sched_in(struct task_struct *prev,
2533 struct task_struct *task)
2535 struct perf_cpu_context *cpuctx;
2537 unsigned long flags;
2539 /* no need to flush branch stack if not changing task */
2543 local_irq_save(flags);
2547 list_for_each_entry_rcu(pmu, &pmus, entry) {
2548 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2551 * check if the context has at least one
2552 * event using PERF_SAMPLE_BRANCH_STACK
2554 if (cpuctx->ctx.nr_branch_stack > 0
2555 && pmu->flush_branch_stack) {
2557 pmu = cpuctx->ctx.pmu;
2559 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2561 perf_pmu_disable(pmu);
2563 pmu->flush_branch_stack();
2565 perf_pmu_enable(pmu);
2567 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2573 local_irq_restore(flags);
2577 * Called from scheduler to add the events of the current task
2578 * with interrupts disabled.
2580 * We restore the event value and then enable it.
2582 * This does not protect us against NMI, but enable()
2583 * sets the enabled bit in the control field of event _before_
2584 * accessing the event control register. If a NMI hits, then it will
2585 * keep the event running.
2587 void __perf_event_task_sched_in(struct task_struct *prev,
2588 struct task_struct *task)
2590 struct perf_event_context *ctx;
2593 for_each_task_context_nr(ctxn) {
2594 ctx = task->perf_event_ctxp[ctxn];
2598 perf_event_context_sched_in(ctx, task);
2601 * if cgroup events exist on this CPU, then we need
2602 * to check if we have to switch in PMU state.
2603 * cgroup event are system-wide mode only
2605 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2606 perf_cgroup_sched_in(prev, task);
2608 /* check for system-wide branch_stack events */
2609 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2610 perf_branch_stack_sched_in(prev, task);
2613 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2615 u64 frequency = event->attr.sample_freq;
2616 u64 sec = NSEC_PER_SEC;
2617 u64 divisor, dividend;
2619 int count_fls, nsec_fls, frequency_fls, sec_fls;
2621 count_fls = fls64(count);
2622 nsec_fls = fls64(nsec);
2623 frequency_fls = fls64(frequency);
2627 * We got @count in @nsec, with a target of sample_freq HZ
2628 * the target period becomes:
2631 * period = -------------------
2632 * @nsec * sample_freq
2637 * Reduce accuracy by one bit such that @a and @b converge
2638 * to a similar magnitude.
2640 #define REDUCE_FLS(a, b) \
2642 if (a##_fls > b##_fls) { \
2652 * Reduce accuracy until either term fits in a u64, then proceed with
2653 * the other, so that finally we can do a u64/u64 division.
2655 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2656 REDUCE_FLS(nsec, frequency);
2657 REDUCE_FLS(sec, count);
2660 if (count_fls + sec_fls > 64) {
2661 divisor = nsec * frequency;
2663 while (count_fls + sec_fls > 64) {
2664 REDUCE_FLS(count, sec);
2668 dividend = count * sec;
2670 dividend = count * sec;
2672 while (nsec_fls + frequency_fls > 64) {
2673 REDUCE_FLS(nsec, frequency);
2677 divisor = nsec * frequency;
2683 return div64_u64(dividend, divisor);
2686 static DEFINE_PER_CPU(int, perf_throttled_count);
2687 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2689 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2691 struct hw_perf_event *hwc = &event->hw;
2692 s64 period, sample_period;
2695 period = perf_calculate_period(event, nsec, count);
2697 delta = (s64)(period - hwc->sample_period);
2698 delta = (delta + 7) / 8; /* low pass filter */
2700 sample_period = hwc->sample_period + delta;
2705 hwc->sample_period = sample_period;
2707 if (local64_read(&hwc->period_left) > 8*sample_period) {
2709 event->pmu->stop(event, PERF_EF_UPDATE);
2711 local64_set(&hwc->period_left, 0);
2714 event->pmu->start(event, PERF_EF_RELOAD);
2719 * combine freq adjustment with unthrottling to avoid two passes over the
2720 * events. At the same time, make sure, having freq events does not change
2721 * the rate of unthrottling as that would introduce bias.
2723 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2726 struct perf_event *event;
2727 struct hw_perf_event *hwc;
2728 u64 now, period = TICK_NSEC;
2732 * only need to iterate over all events iff:
2733 * - context have events in frequency mode (needs freq adjust)
2734 * - there are events to unthrottle on this cpu
2736 if (!(ctx->nr_freq || needs_unthr))
2739 raw_spin_lock(&ctx->lock);
2740 perf_pmu_disable(ctx->pmu);
2742 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2743 if (event->state != PERF_EVENT_STATE_ACTIVE)
2746 if (!event_filter_match(event))
2751 if (hwc->interrupts == MAX_INTERRUPTS) {
2752 hwc->interrupts = 0;
2753 perf_log_throttle(event, 1);
2754 event->pmu->start(event, 0);
2757 if (!event->attr.freq || !event->attr.sample_freq)
2761 * stop the event and update event->count
2763 event->pmu->stop(event, PERF_EF_UPDATE);
2765 now = local64_read(&event->count);
2766 delta = now - hwc->freq_count_stamp;
2767 hwc->freq_count_stamp = now;
2771 * reload only if value has changed
2772 * we have stopped the event so tell that
2773 * to perf_adjust_period() to avoid stopping it
2777 perf_adjust_period(event, period, delta, false);
2779 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2782 perf_pmu_enable(ctx->pmu);
2783 raw_spin_unlock(&ctx->lock);
2787 * Round-robin a context's events:
2789 static void rotate_ctx(struct perf_event_context *ctx)
2792 * Rotate the first entry last of non-pinned groups. Rotation might be
2793 * disabled by the inheritance code.
2795 if (!ctx->rotate_disable)
2796 list_rotate_left(&ctx->flexible_groups);
2800 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2801 * because they're strictly cpu affine and rotate_start is called with IRQs
2802 * disabled, while rotate_context is called from IRQ context.
2804 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2806 struct perf_event_context *ctx = NULL;
2807 int rotate = 0, remove = 1;
2809 if (cpuctx->ctx.nr_events) {
2811 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2815 ctx = cpuctx->task_ctx;
2816 if (ctx && ctx->nr_events) {
2818 if (ctx->nr_events != ctx->nr_active)
2825 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2826 perf_pmu_disable(cpuctx->ctx.pmu);
2828 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2830 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2832 rotate_ctx(&cpuctx->ctx);
2836 perf_event_sched_in(cpuctx, ctx, current);
2838 perf_pmu_enable(cpuctx->ctx.pmu);
2839 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2842 list_del_init(&cpuctx->rotation_list);
2847 #ifdef CONFIG_NO_HZ_FULL
2848 bool perf_event_can_stop_tick(void)
2850 if (atomic_read(&nr_freq_events) ||
2851 __this_cpu_read(perf_throttled_count))
2858 void perf_event_task_tick(void)
2860 struct list_head *head = &__get_cpu_var(rotation_list);
2861 struct perf_cpu_context *cpuctx, *tmp;
2862 struct perf_event_context *ctx;
2865 WARN_ON(!irqs_disabled());
2867 __this_cpu_inc(perf_throttled_seq);
2868 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2870 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2872 perf_adjust_freq_unthr_context(ctx, throttled);
2874 ctx = cpuctx->task_ctx;
2876 perf_adjust_freq_unthr_context(ctx, throttled);
2880 static int event_enable_on_exec(struct perf_event *event,
2881 struct perf_event_context *ctx)
2883 if (!event->attr.enable_on_exec)
2886 event->attr.enable_on_exec = 0;
2887 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2890 __perf_event_mark_enabled(event);
2896 * Enable all of a task's events that have been marked enable-on-exec.
2897 * This expects task == current.
2899 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2901 struct perf_event *event;
2902 unsigned long flags;
2906 local_irq_save(flags);
2907 if (!ctx || !ctx->nr_events)
2911 * We must ctxsw out cgroup events to avoid conflict
2912 * when invoking perf_task_event_sched_in() later on
2913 * in this function. Otherwise we end up trying to
2914 * ctxswin cgroup events which are already scheduled
2917 perf_cgroup_sched_out(current, NULL);
2919 raw_spin_lock(&ctx->lock);
2920 task_ctx_sched_out(ctx);
2922 list_for_each_entry(event, &ctx->event_list, event_entry) {
2923 ret = event_enable_on_exec(event, ctx);
2929 * Unclone this context if we enabled any event.
2934 raw_spin_unlock(&ctx->lock);
2937 * Also calls ctxswin for cgroup events, if any:
2939 perf_event_context_sched_in(ctx, ctx->task);
2941 local_irq_restore(flags);
2945 * Cross CPU call to read the hardware event
2947 static void __perf_event_read(void *info)
2949 struct perf_event *event = info;
2950 struct perf_event_context *ctx = event->ctx;
2951 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2954 * If this is a task context, we need to check whether it is
2955 * the current task context of this cpu. If not it has been
2956 * scheduled out before the smp call arrived. In that case
2957 * event->count would have been updated to a recent sample
2958 * when the event was scheduled out.
2960 if (ctx->task && cpuctx->task_ctx != ctx)
2963 raw_spin_lock(&ctx->lock);
2964 if (ctx->is_active) {
2965 update_context_time(ctx);
2966 update_cgrp_time_from_event(event);
2968 update_event_times(event);
2969 if (event->state == PERF_EVENT_STATE_ACTIVE)
2970 event->pmu->read(event);
2971 raw_spin_unlock(&ctx->lock);
2974 static inline u64 perf_event_count(struct perf_event *event)
2976 return local64_read(&event->count) + atomic64_read(&event->child_count);
2979 static u64 perf_event_read(struct perf_event *event)
2982 * If event is enabled and currently active on a CPU, update the
2983 * value in the event structure:
2985 if (event->state == PERF_EVENT_STATE_ACTIVE) {
2986 smp_call_function_single(event->oncpu,
2987 __perf_event_read, event, 1);
2988 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
2989 struct perf_event_context *ctx = event->ctx;
2990 unsigned long flags;
2992 raw_spin_lock_irqsave(&ctx->lock, flags);
2994 * may read while context is not active
2995 * (e.g., thread is blocked), in that case
2996 * we cannot update context time
2998 if (ctx->is_active) {
2999 update_context_time(ctx);
3000 update_cgrp_time_from_event(event);
3002 update_event_times(event);
3003 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3006 return perf_event_count(event);
3010 * Initialize the perf_event context in a task_struct:
3012 static void __perf_event_init_context(struct perf_event_context *ctx)
3014 raw_spin_lock_init(&ctx->lock);
3015 mutex_init(&ctx->mutex);
3016 INIT_LIST_HEAD(&ctx->pinned_groups);
3017 INIT_LIST_HEAD(&ctx->flexible_groups);
3018 INIT_LIST_HEAD(&ctx->event_list);
3019 atomic_set(&ctx->refcount, 1);
3022 static struct perf_event_context *
3023 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3025 struct perf_event_context *ctx;
3027 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3031 __perf_event_init_context(ctx);
3034 get_task_struct(task);
3041 static struct task_struct *
3042 find_lively_task_by_vpid(pid_t vpid)
3044 struct task_struct *task;
3051 task = find_task_by_vpid(vpid);
3053 get_task_struct(task);
3057 return ERR_PTR(-ESRCH);
3059 /* Reuse ptrace permission checks for now. */
3061 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3066 put_task_struct(task);
3067 return ERR_PTR(err);
3072 * Returns a matching context with refcount and pincount.
3074 static struct perf_event_context *
3075 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3077 struct perf_event_context *ctx;
3078 struct perf_cpu_context *cpuctx;
3079 unsigned long flags;
3083 /* Must be root to operate on a CPU event: */
3084 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3085 return ERR_PTR(-EACCES);
3088 * We could be clever and allow to attach a event to an
3089 * offline CPU and activate it when the CPU comes up, but
3092 if (!cpu_online(cpu))
3093 return ERR_PTR(-ENODEV);
3095 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3104 ctxn = pmu->task_ctx_nr;
3109 ctx = perf_lock_task_context(task, ctxn, &flags);
3113 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3115 ctx = alloc_perf_context(pmu, task);
3121 mutex_lock(&task->perf_event_mutex);
3123 * If it has already passed perf_event_exit_task().
3124 * we must see PF_EXITING, it takes this mutex too.
3126 if (task->flags & PF_EXITING)
3128 else if (task->perf_event_ctxp[ctxn])
3133 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3135 mutex_unlock(&task->perf_event_mutex);
3137 if (unlikely(err)) {
3149 return ERR_PTR(err);
3152 static void perf_event_free_filter(struct perf_event *event);
3154 static void free_event_rcu(struct rcu_head *head)
3156 struct perf_event *event;
3158 event = container_of(head, struct perf_event, rcu_head);
3160 put_pid_ns(event->ns);
3161 perf_event_free_filter(event);
3165 static void ring_buffer_put(struct ring_buffer *rb);
3166 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb);
3168 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3173 if (has_branch_stack(event)) {
3174 if (!(event->attach_state & PERF_ATTACH_TASK))
3175 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3177 if (is_cgroup_event(event))
3178 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3181 static void unaccount_event(struct perf_event *event)
3186 if (event->attach_state & PERF_ATTACH_TASK)
3187 static_key_slow_dec_deferred(&perf_sched_events);
3188 if (event->attr.mmap || event->attr.mmap_data)
3189 atomic_dec(&nr_mmap_events);
3190 if (event->attr.comm)
3191 atomic_dec(&nr_comm_events);
3192 if (event->attr.task)
3193 atomic_dec(&nr_task_events);
3194 if (event->attr.freq)
3195 atomic_dec(&nr_freq_events);
3196 if (is_cgroup_event(event))
3197 static_key_slow_dec_deferred(&perf_sched_events);
3198 if (has_branch_stack(event))
3199 static_key_slow_dec_deferred(&perf_sched_events);
3201 unaccount_event_cpu(event, event->cpu);
3204 static void __free_event(struct perf_event *event)
3206 if (!event->parent) {
3207 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3208 put_callchain_buffers();
3212 event->destroy(event);
3215 put_ctx(event->ctx);
3217 call_rcu(&event->rcu_head, free_event_rcu);
3219 static void free_event(struct perf_event *event)
3221 irq_work_sync(&event->pending);
3223 unaccount_event(event);
3226 struct ring_buffer *rb;
3229 * Can happen when we close an event with re-directed output.
3231 * Since we have a 0 refcount, perf_mmap_close() will skip
3232 * over us; possibly making our ring_buffer_put() the last.
3234 mutex_lock(&event->mmap_mutex);
3237 rcu_assign_pointer(event->rb, NULL);
3238 ring_buffer_detach(event, rb);
3239 ring_buffer_put(rb); /* could be last */
3241 mutex_unlock(&event->mmap_mutex);
3244 if (is_cgroup_event(event))
3245 perf_detach_cgroup(event);
3248 __free_event(event);
3251 int perf_event_release_kernel(struct perf_event *event)
3253 struct perf_event_context *ctx = event->ctx;
3255 WARN_ON_ONCE(ctx->parent_ctx);
3257 * There are two ways this annotation is useful:
3259 * 1) there is a lock recursion from perf_event_exit_task
3260 * see the comment there.
3262 * 2) there is a lock-inversion with mmap_sem through
3263 * perf_event_read_group(), which takes faults while
3264 * holding ctx->mutex, however this is called after
3265 * the last filedesc died, so there is no possibility
3266 * to trigger the AB-BA case.
3268 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3269 raw_spin_lock_irq(&ctx->lock);
3270 perf_group_detach(event);
3271 raw_spin_unlock_irq(&ctx->lock);
3272 perf_remove_from_context(event);
3273 mutex_unlock(&ctx->mutex);
3279 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3282 * Called when the last reference to the file is gone.
3284 static void put_event(struct perf_event *event)
3286 struct task_struct *owner;
3288 if (!atomic_long_dec_and_test(&event->refcount))
3292 owner = ACCESS_ONCE(event->owner);
3294 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3295 * !owner it means the list deletion is complete and we can indeed
3296 * free this event, otherwise we need to serialize on
3297 * owner->perf_event_mutex.
3299 smp_read_barrier_depends();
3302 * Since delayed_put_task_struct() also drops the last
3303 * task reference we can safely take a new reference
3304 * while holding the rcu_read_lock().
3306 get_task_struct(owner);
3311 mutex_lock(&owner->perf_event_mutex);
3313 * We have to re-check the event->owner field, if it is cleared
3314 * we raced with perf_event_exit_task(), acquiring the mutex
3315 * ensured they're done, and we can proceed with freeing the
3319 list_del_init(&event->owner_entry);
3320 mutex_unlock(&owner->perf_event_mutex);
3321 put_task_struct(owner);
3324 perf_event_release_kernel(event);
3327 static int perf_release(struct inode *inode, struct file *file)
3329 put_event(file->private_data);
3333 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3335 struct perf_event *child;
3341 mutex_lock(&event->child_mutex);
3342 total += perf_event_read(event);
3343 *enabled += event->total_time_enabled +
3344 atomic64_read(&event->child_total_time_enabled);
3345 *running += event->total_time_running +
3346 atomic64_read(&event->child_total_time_running);
3348 list_for_each_entry(child, &event->child_list, child_list) {
3349 total += perf_event_read(child);
3350 *enabled += child->total_time_enabled;
3351 *running += child->total_time_running;
3353 mutex_unlock(&event->child_mutex);
3357 EXPORT_SYMBOL_GPL(perf_event_read_value);
3359 static int perf_event_read_group(struct perf_event *event,
3360 u64 read_format, char __user *buf)
3362 struct perf_event *leader = event->group_leader, *sub;
3363 int n = 0, size = 0, ret = -EFAULT;
3364 struct perf_event_context *ctx = leader->ctx;
3366 u64 count, enabled, running;
3368 mutex_lock(&ctx->mutex);
3369 count = perf_event_read_value(leader, &enabled, &running);
3371 values[n++] = 1 + leader->nr_siblings;
3372 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3373 values[n++] = enabled;
3374 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3375 values[n++] = running;
3376 values[n++] = count;
3377 if (read_format & PERF_FORMAT_ID)
3378 values[n++] = primary_event_id(leader);
3380 size = n * sizeof(u64);
3382 if (copy_to_user(buf, values, size))
3387 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3390 values[n++] = perf_event_read_value(sub, &enabled, &running);
3391 if (read_format & PERF_FORMAT_ID)
3392 values[n++] = primary_event_id(sub);
3394 size = n * sizeof(u64);
3396 if (copy_to_user(buf + ret, values, size)) {
3404 mutex_unlock(&ctx->mutex);
3409 static int perf_event_read_one(struct perf_event *event,
3410 u64 read_format, char __user *buf)
3412 u64 enabled, running;
3416 values[n++] = perf_event_read_value(event, &enabled, &running);
3417 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3418 values[n++] = enabled;
3419 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3420 values[n++] = running;
3421 if (read_format & PERF_FORMAT_ID)
3422 values[n++] = primary_event_id(event);
3424 if (copy_to_user(buf, values, n * sizeof(u64)))
3427 return n * sizeof(u64);
3431 * Read the performance event - simple non blocking version for now
3434 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3436 u64 read_format = event->attr.read_format;
3440 * Return end-of-file for a read on a event that is in
3441 * error state (i.e. because it was pinned but it couldn't be
3442 * scheduled on to the CPU at some point).
3444 if (event->state == PERF_EVENT_STATE_ERROR)
3447 if (count < event->read_size)
3450 WARN_ON_ONCE(event->ctx->parent_ctx);
3451 if (read_format & PERF_FORMAT_GROUP)
3452 ret = perf_event_read_group(event, read_format, buf);
3454 ret = perf_event_read_one(event, read_format, buf);
3460 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3462 struct perf_event *event = file->private_data;
3464 return perf_read_hw(event, buf, count);
3467 static unsigned int perf_poll(struct file *file, poll_table *wait)
3469 struct perf_event *event = file->private_data;
3470 struct ring_buffer *rb;
3471 unsigned int events = POLL_HUP;
3474 * Pin the event->rb by taking event->mmap_mutex; otherwise
3475 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3477 mutex_lock(&event->mmap_mutex);
3480 events = atomic_xchg(&rb->poll, 0);
3481 mutex_unlock(&event->mmap_mutex);
3483 poll_wait(file, &event->waitq, wait);
3488 static void perf_event_reset(struct perf_event *event)
3490 (void)perf_event_read(event);
3491 local64_set(&event->count, 0);
3492 perf_event_update_userpage(event);
3496 * Holding the top-level event's child_mutex means that any
3497 * descendant process that has inherited this event will block
3498 * in sync_child_event if it goes to exit, thus satisfying the
3499 * task existence requirements of perf_event_enable/disable.
3501 static void perf_event_for_each_child(struct perf_event *event,
3502 void (*func)(struct perf_event *))
3504 struct perf_event *child;
3506 WARN_ON_ONCE(event->ctx->parent_ctx);
3507 mutex_lock(&event->child_mutex);
3509 list_for_each_entry(child, &event->child_list, child_list)
3511 mutex_unlock(&event->child_mutex);
3514 static void perf_event_for_each(struct perf_event *event,
3515 void (*func)(struct perf_event *))
3517 struct perf_event_context *ctx = event->ctx;
3518 struct perf_event *sibling;
3520 WARN_ON_ONCE(ctx->parent_ctx);
3521 mutex_lock(&ctx->mutex);
3522 event = event->group_leader;
3524 perf_event_for_each_child(event, func);
3525 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3526 perf_event_for_each_child(sibling, func);
3527 mutex_unlock(&ctx->mutex);
3530 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3532 struct perf_event_context *ctx = event->ctx;
3536 if (!is_sampling_event(event))
3539 if (copy_from_user(&value, arg, sizeof(value)))
3545 raw_spin_lock_irq(&ctx->lock);
3546 if (event->attr.freq) {
3547 if (value > sysctl_perf_event_sample_rate) {
3552 event->attr.sample_freq = value;
3554 event->attr.sample_period = value;
3555 event->hw.sample_period = value;
3558 raw_spin_unlock_irq(&ctx->lock);
3563 static const struct file_operations perf_fops;
3565 static inline int perf_fget_light(int fd, struct fd *p)
3567 struct fd f = fdget(fd);
3571 if (f.file->f_op != &perf_fops) {
3579 static int perf_event_set_output(struct perf_event *event,
3580 struct perf_event *output_event);
3581 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3583 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3585 struct perf_event *event = file->private_data;
3586 void (*func)(struct perf_event *);
3590 case PERF_EVENT_IOC_ENABLE:
3591 func = perf_event_enable;
3593 case PERF_EVENT_IOC_DISABLE:
3594 func = perf_event_disable;
3596 case PERF_EVENT_IOC_RESET:
3597 func = perf_event_reset;
3600 case PERF_EVENT_IOC_REFRESH:
3601 return perf_event_refresh(event, arg);
3603 case PERF_EVENT_IOC_PERIOD:
3604 return perf_event_period(event, (u64 __user *)arg);
3606 case PERF_EVENT_IOC_ID:
3608 u64 id = primary_event_id(event);
3610 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3615 case PERF_EVENT_IOC_SET_OUTPUT:
3619 struct perf_event *output_event;
3621 ret = perf_fget_light(arg, &output);
3624 output_event = output.file->private_data;
3625 ret = perf_event_set_output(event, output_event);
3628 ret = perf_event_set_output(event, NULL);
3633 case PERF_EVENT_IOC_SET_FILTER:
3634 return perf_event_set_filter(event, (void __user *)arg);
3640 if (flags & PERF_IOC_FLAG_GROUP)
3641 perf_event_for_each(event, func);
3643 perf_event_for_each_child(event, func);
3648 int perf_event_task_enable(void)
3650 struct perf_event *event;
3652 mutex_lock(¤t->perf_event_mutex);
3653 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3654 perf_event_for_each_child(event, perf_event_enable);
3655 mutex_unlock(¤t->perf_event_mutex);
3660 int perf_event_task_disable(void)
3662 struct perf_event *event;
3664 mutex_lock(¤t->perf_event_mutex);
3665 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3666 perf_event_for_each_child(event, perf_event_disable);
3667 mutex_unlock(¤t->perf_event_mutex);
3672 static int perf_event_index(struct perf_event *event)
3674 if (event->hw.state & PERF_HES_STOPPED)
3677 if (event->state != PERF_EVENT_STATE_ACTIVE)
3680 return event->pmu->event_idx(event);
3683 static void calc_timer_values(struct perf_event *event,
3690 *now = perf_clock();
3691 ctx_time = event->shadow_ctx_time + *now;
3692 *enabled = ctx_time - event->tstamp_enabled;
3693 *running = ctx_time - event->tstamp_running;
3696 static void perf_event_init_userpage(struct perf_event *event)
3698 struct perf_event_mmap_page *userpg;
3699 struct ring_buffer *rb;
3702 rb = rcu_dereference(event->rb);
3706 userpg = rb->user_page;
3708 /* Allow new userspace to detect that bit 0 is deprecated */
3709 userpg->cap_bit0_is_deprecated = 1;
3710 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3716 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3721 * Callers need to ensure there can be no nesting of this function, otherwise
3722 * the seqlock logic goes bad. We can not serialize this because the arch
3723 * code calls this from NMI context.
3725 void perf_event_update_userpage(struct perf_event *event)
3727 struct perf_event_mmap_page *userpg;
3728 struct ring_buffer *rb;
3729 u64 enabled, running, now;
3732 rb = rcu_dereference(event->rb);
3737 * compute total_time_enabled, total_time_running
3738 * based on snapshot values taken when the event
3739 * was last scheduled in.
3741 * we cannot simply called update_context_time()
3742 * because of locking issue as we can be called in
3745 calc_timer_values(event, &now, &enabled, &running);
3747 userpg = rb->user_page;
3749 * Disable preemption so as to not let the corresponding user-space
3750 * spin too long if we get preempted.
3755 userpg->index = perf_event_index(event);
3756 userpg->offset = perf_event_count(event);
3758 userpg->offset -= local64_read(&event->hw.prev_count);
3760 userpg->time_enabled = enabled +
3761 atomic64_read(&event->child_total_time_enabled);
3763 userpg->time_running = running +
3764 atomic64_read(&event->child_total_time_running);
3766 arch_perf_update_userpage(userpg, now);
3775 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3777 struct perf_event *event = vma->vm_file->private_data;
3778 struct ring_buffer *rb;
3779 int ret = VM_FAULT_SIGBUS;
3781 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3782 if (vmf->pgoff == 0)
3788 rb = rcu_dereference(event->rb);
3792 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3795 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3799 get_page(vmf->page);
3800 vmf->page->mapping = vma->vm_file->f_mapping;
3801 vmf->page->index = vmf->pgoff;
3810 static void ring_buffer_attach(struct perf_event *event,
3811 struct ring_buffer *rb)
3813 unsigned long flags;
3815 if (!list_empty(&event->rb_entry))
3818 spin_lock_irqsave(&rb->event_lock, flags);
3819 if (list_empty(&event->rb_entry))
3820 list_add(&event->rb_entry, &rb->event_list);
3821 spin_unlock_irqrestore(&rb->event_lock, flags);
3824 static void ring_buffer_detach(struct perf_event *event, struct ring_buffer *rb)
3826 unsigned long flags;
3828 if (list_empty(&event->rb_entry))
3831 spin_lock_irqsave(&rb->event_lock, flags);
3832 list_del_init(&event->rb_entry);
3833 wake_up_all(&event->waitq);
3834 spin_unlock_irqrestore(&rb->event_lock, flags);
3837 static void ring_buffer_wakeup(struct perf_event *event)
3839 struct ring_buffer *rb;
3842 rb = rcu_dereference(event->rb);
3844 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
3845 wake_up_all(&event->waitq);
3850 static void rb_free_rcu(struct rcu_head *rcu_head)
3852 struct ring_buffer *rb;
3854 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
3858 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
3860 struct ring_buffer *rb;
3863 rb = rcu_dereference(event->rb);
3865 if (!atomic_inc_not_zero(&rb->refcount))
3873 static void ring_buffer_put(struct ring_buffer *rb)
3875 if (!atomic_dec_and_test(&rb->refcount))
3878 WARN_ON_ONCE(!list_empty(&rb->event_list));
3880 call_rcu(&rb->rcu_head, rb_free_rcu);
3883 static void perf_mmap_open(struct vm_area_struct *vma)
3885 struct perf_event *event = vma->vm_file->private_data;
3887 atomic_inc(&event->mmap_count);
3888 atomic_inc(&event->rb->mmap_count);
3892 * A buffer can be mmap()ed multiple times; either directly through the same
3893 * event, or through other events by use of perf_event_set_output().
3895 * In order to undo the VM accounting done by perf_mmap() we need to destroy
3896 * the buffer here, where we still have a VM context. This means we need
3897 * to detach all events redirecting to us.
3899 static void perf_mmap_close(struct vm_area_struct *vma)
3901 struct perf_event *event = vma->vm_file->private_data;
3903 struct ring_buffer *rb = event->rb;
3904 struct user_struct *mmap_user = rb->mmap_user;
3905 int mmap_locked = rb->mmap_locked;
3906 unsigned long size = perf_data_size(rb);
3908 atomic_dec(&rb->mmap_count);
3910 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
3913 /* Detach current event from the buffer. */
3914 rcu_assign_pointer(event->rb, NULL);
3915 ring_buffer_detach(event, rb);
3916 mutex_unlock(&event->mmap_mutex);
3918 /* If there's still other mmap()s of this buffer, we're done. */
3919 if (atomic_read(&rb->mmap_count)) {
3920 ring_buffer_put(rb); /* can't be last */
3925 * No other mmap()s, detach from all other events that might redirect
3926 * into the now unreachable buffer. Somewhat complicated by the
3927 * fact that rb::event_lock otherwise nests inside mmap_mutex.
3931 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
3932 if (!atomic_long_inc_not_zero(&event->refcount)) {
3934 * This event is en-route to free_event() which will
3935 * detach it and remove it from the list.
3941 mutex_lock(&event->mmap_mutex);
3943 * Check we didn't race with perf_event_set_output() which can
3944 * swizzle the rb from under us while we were waiting to
3945 * acquire mmap_mutex.
3947 * If we find a different rb; ignore this event, a next
3948 * iteration will no longer find it on the list. We have to
3949 * still restart the iteration to make sure we're not now
3950 * iterating the wrong list.
3952 if (event->rb == rb) {
3953 rcu_assign_pointer(event->rb, NULL);
3954 ring_buffer_detach(event, rb);
3955 ring_buffer_put(rb); /* can't be last, we still have one */
3957 mutex_unlock(&event->mmap_mutex);
3961 * Restart the iteration; either we're on the wrong list or
3962 * destroyed its integrity by doing a deletion.
3969 * It could be there's still a few 0-ref events on the list; they'll
3970 * get cleaned up by free_event() -- they'll also still have their
3971 * ref on the rb and will free it whenever they are done with it.
3973 * Aside from that, this buffer is 'fully' detached and unmapped,
3974 * undo the VM accounting.
3977 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
3978 vma->vm_mm->pinned_vm -= mmap_locked;
3979 free_uid(mmap_user);
3981 ring_buffer_put(rb); /* could be last */
3984 static const struct vm_operations_struct perf_mmap_vmops = {
3985 .open = perf_mmap_open,
3986 .close = perf_mmap_close,
3987 .fault = perf_mmap_fault,
3988 .page_mkwrite = perf_mmap_fault,
3991 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
3993 struct perf_event *event = file->private_data;
3994 unsigned long user_locked, user_lock_limit;
3995 struct user_struct *user = current_user();
3996 unsigned long locked, lock_limit;
3997 struct ring_buffer *rb;
3998 unsigned long vma_size;
3999 unsigned long nr_pages;
4000 long user_extra, extra;
4001 int ret = 0, flags = 0;
4004 * Don't allow mmap() of inherited per-task counters. This would
4005 * create a performance issue due to all children writing to the
4008 if (event->cpu == -1 && event->attr.inherit)
4011 if (!(vma->vm_flags & VM_SHARED))
4014 vma_size = vma->vm_end - vma->vm_start;
4015 nr_pages = (vma_size / PAGE_SIZE) - 1;
4018 * If we have rb pages ensure they're a power-of-two number, so we
4019 * can do bitmasks instead of modulo.
4021 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4024 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4027 if (vma->vm_pgoff != 0)
4030 WARN_ON_ONCE(event->ctx->parent_ctx);
4032 mutex_lock(&event->mmap_mutex);
4034 if (event->rb->nr_pages != nr_pages) {
4039 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4041 * Raced against perf_mmap_close() through
4042 * perf_event_set_output(). Try again, hope for better
4045 mutex_unlock(&event->mmap_mutex);
4052 user_extra = nr_pages + 1;
4053 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4056 * Increase the limit linearly with more CPUs:
4058 user_lock_limit *= num_online_cpus();
4060 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4063 if (user_locked > user_lock_limit)
4064 extra = user_locked - user_lock_limit;
4066 lock_limit = rlimit(RLIMIT_MEMLOCK);
4067 lock_limit >>= PAGE_SHIFT;
4068 locked = vma->vm_mm->pinned_vm + extra;
4070 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4071 !capable(CAP_IPC_LOCK)) {
4078 if (vma->vm_flags & VM_WRITE)
4079 flags |= RING_BUFFER_WRITABLE;
4081 rb = rb_alloc(nr_pages,
4082 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4090 atomic_set(&rb->mmap_count, 1);
4091 rb->mmap_locked = extra;
4092 rb->mmap_user = get_current_user();
4094 atomic_long_add(user_extra, &user->locked_vm);
4095 vma->vm_mm->pinned_vm += extra;
4097 ring_buffer_attach(event, rb);
4098 rcu_assign_pointer(event->rb, rb);
4100 perf_event_init_userpage(event);
4101 perf_event_update_userpage(event);
4105 atomic_inc(&event->mmap_count);
4106 mutex_unlock(&event->mmap_mutex);
4109 * Since pinned accounting is per vm we cannot allow fork() to copy our
4112 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4113 vma->vm_ops = &perf_mmap_vmops;
4118 static int perf_fasync(int fd, struct file *filp, int on)
4120 struct inode *inode = file_inode(filp);
4121 struct perf_event *event = filp->private_data;
4124 mutex_lock(&inode->i_mutex);
4125 retval = fasync_helper(fd, filp, on, &event->fasync);
4126 mutex_unlock(&inode->i_mutex);
4134 static const struct file_operations perf_fops = {
4135 .llseek = no_llseek,
4136 .release = perf_release,
4139 .unlocked_ioctl = perf_ioctl,
4140 .compat_ioctl = perf_ioctl,
4142 .fasync = perf_fasync,
4148 * If there's data, ensure we set the poll() state and publish everything
4149 * to user-space before waking everybody up.
4152 void perf_event_wakeup(struct perf_event *event)
4154 ring_buffer_wakeup(event);
4156 if (event->pending_kill) {
4157 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4158 event->pending_kill = 0;
4162 static void perf_pending_event(struct irq_work *entry)
4164 struct perf_event *event = container_of(entry,
4165 struct perf_event, pending);
4167 if (event->pending_disable) {
4168 event->pending_disable = 0;
4169 __perf_event_disable(event);
4172 if (event->pending_wakeup) {
4173 event->pending_wakeup = 0;
4174 perf_event_wakeup(event);
4179 * We assume there is only KVM supporting the callbacks.
4180 * Later on, we might change it to a list if there is
4181 * another virtualization implementation supporting the callbacks.
4183 struct perf_guest_info_callbacks *perf_guest_cbs;
4185 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4187 perf_guest_cbs = cbs;
4190 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4192 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4194 perf_guest_cbs = NULL;
4197 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4200 perf_output_sample_regs(struct perf_output_handle *handle,
4201 struct pt_regs *regs, u64 mask)
4205 for_each_set_bit(bit, (const unsigned long *) &mask,
4206 sizeof(mask) * BITS_PER_BYTE) {
4209 val = perf_reg_value(regs, bit);
4210 perf_output_put(handle, val);
4214 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4215 struct pt_regs *regs)
4217 if (!user_mode(regs)) {
4219 regs = task_pt_regs(current);
4225 regs_user->regs = regs;
4226 regs_user->abi = perf_reg_abi(current);
4231 * Get remaining task size from user stack pointer.
4233 * It'd be better to take stack vma map and limit this more
4234 * precisly, but there's no way to get it safely under interrupt,
4235 * so using TASK_SIZE as limit.
4237 static u64 perf_ustack_task_size(struct pt_regs *regs)
4239 unsigned long addr = perf_user_stack_pointer(regs);
4241 if (!addr || addr >= TASK_SIZE)
4244 return TASK_SIZE - addr;
4248 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4249 struct pt_regs *regs)
4253 /* No regs, no stack pointer, no dump. */
4258 * Check if we fit in with the requested stack size into the:
4260 * If we don't, we limit the size to the TASK_SIZE.
4262 * - remaining sample size
4263 * If we don't, we customize the stack size to
4264 * fit in to the remaining sample size.
4267 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4268 stack_size = min(stack_size, (u16) task_size);
4270 /* Current header size plus static size and dynamic size. */
4271 header_size += 2 * sizeof(u64);
4273 /* Do we fit in with the current stack dump size? */
4274 if ((u16) (header_size + stack_size) < header_size) {
4276 * If we overflow the maximum size for the sample,
4277 * we customize the stack dump size to fit in.
4279 stack_size = USHRT_MAX - header_size - sizeof(u64);
4280 stack_size = round_up(stack_size, sizeof(u64));
4287 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4288 struct pt_regs *regs)
4290 /* Case of a kernel thread, nothing to dump */
4293 perf_output_put(handle, size);
4302 * - the size requested by user or the best one we can fit
4303 * in to the sample max size
4305 * - user stack dump data
4307 * - the actual dumped size
4311 perf_output_put(handle, dump_size);
4314 sp = perf_user_stack_pointer(regs);
4315 rem = __output_copy_user(handle, (void *) sp, dump_size);
4316 dyn_size = dump_size - rem;
4318 perf_output_skip(handle, rem);
4321 perf_output_put(handle, dyn_size);
4325 static void __perf_event_header__init_id(struct perf_event_header *header,
4326 struct perf_sample_data *data,
4327 struct perf_event *event)
4329 u64 sample_type = event->attr.sample_type;
4331 data->type = sample_type;
4332 header->size += event->id_header_size;
4334 if (sample_type & PERF_SAMPLE_TID) {
4335 /* namespace issues */
4336 data->tid_entry.pid = perf_event_pid(event, current);
4337 data->tid_entry.tid = perf_event_tid(event, current);
4340 if (sample_type & PERF_SAMPLE_TIME)
4341 data->time = perf_clock();
4343 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4344 data->id = primary_event_id(event);
4346 if (sample_type & PERF_SAMPLE_STREAM_ID)
4347 data->stream_id = event->id;
4349 if (sample_type & PERF_SAMPLE_CPU) {
4350 data->cpu_entry.cpu = raw_smp_processor_id();
4351 data->cpu_entry.reserved = 0;
4355 void perf_event_header__init_id(struct perf_event_header *header,
4356 struct perf_sample_data *data,
4357 struct perf_event *event)
4359 if (event->attr.sample_id_all)
4360 __perf_event_header__init_id(header, data, event);
4363 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4364 struct perf_sample_data *data)
4366 u64 sample_type = data->type;
4368 if (sample_type & PERF_SAMPLE_TID)
4369 perf_output_put(handle, data->tid_entry);
4371 if (sample_type & PERF_SAMPLE_TIME)
4372 perf_output_put(handle, data->time);
4374 if (sample_type & PERF_SAMPLE_ID)
4375 perf_output_put(handle, data->id);
4377 if (sample_type & PERF_SAMPLE_STREAM_ID)
4378 perf_output_put(handle, data->stream_id);
4380 if (sample_type & PERF_SAMPLE_CPU)
4381 perf_output_put(handle, data->cpu_entry);
4383 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4384 perf_output_put(handle, data->id);
4387 void perf_event__output_id_sample(struct perf_event *event,
4388 struct perf_output_handle *handle,
4389 struct perf_sample_data *sample)
4391 if (event->attr.sample_id_all)
4392 __perf_event__output_id_sample(handle, sample);
4395 static void perf_output_read_one(struct perf_output_handle *handle,
4396 struct perf_event *event,
4397 u64 enabled, u64 running)
4399 u64 read_format = event->attr.read_format;
4403 values[n++] = perf_event_count(event);
4404 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4405 values[n++] = enabled +
4406 atomic64_read(&event->child_total_time_enabled);
4408 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4409 values[n++] = running +
4410 atomic64_read(&event->child_total_time_running);
4412 if (read_format & PERF_FORMAT_ID)
4413 values[n++] = primary_event_id(event);
4415 __output_copy(handle, values, n * sizeof(u64));
4419 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4421 static void perf_output_read_group(struct perf_output_handle *handle,
4422 struct perf_event *event,
4423 u64 enabled, u64 running)
4425 struct perf_event *leader = event->group_leader, *sub;
4426 u64 read_format = event->attr.read_format;
4430 values[n++] = 1 + leader->nr_siblings;
4432 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4433 values[n++] = enabled;
4435 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4436 values[n++] = running;
4438 if (leader != event)
4439 leader->pmu->read(leader);
4441 values[n++] = perf_event_count(leader);
4442 if (read_format & PERF_FORMAT_ID)
4443 values[n++] = primary_event_id(leader);
4445 __output_copy(handle, values, n * sizeof(u64));
4447 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4450 if ((sub != event) &&
4451 (sub->state == PERF_EVENT_STATE_ACTIVE))
4452 sub->pmu->read(sub);
4454 values[n++] = perf_event_count(sub);
4455 if (read_format & PERF_FORMAT_ID)
4456 values[n++] = primary_event_id(sub);
4458 __output_copy(handle, values, n * sizeof(u64));
4462 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4463 PERF_FORMAT_TOTAL_TIME_RUNNING)
4465 static void perf_output_read(struct perf_output_handle *handle,
4466 struct perf_event *event)
4468 u64 enabled = 0, running = 0, now;
4469 u64 read_format = event->attr.read_format;
4472 * compute total_time_enabled, total_time_running
4473 * based on snapshot values taken when the event
4474 * was last scheduled in.
4476 * we cannot simply called update_context_time()
4477 * because of locking issue as we are called in
4480 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4481 calc_timer_values(event, &now, &enabled, &running);
4483 if (event->attr.read_format & PERF_FORMAT_GROUP)
4484 perf_output_read_group(handle, event, enabled, running);
4486 perf_output_read_one(handle, event, enabled, running);
4489 void perf_output_sample(struct perf_output_handle *handle,
4490 struct perf_event_header *header,
4491 struct perf_sample_data *data,
4492 struct perf_event *event)
4494 u64 sample_type = data->type;
4496 perf_output_put(handle, *header);
4498 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4499 perf_output_put(handle, data->id);
4501 if (sample_type & PERF_SAMPLE_IP)
4502 perf_output_put(handle, data->ip);
4504 if (sample_type & PERF_SAMPLE_TID)
4505 perf_output_put(handle, data->tid_entry);
4507 if (sample_type & PERF_SAMPLE_TIME)
4508 perf_output_put(handle, data->time);
4510 if (sample_type & PERF_SAMPLE_ADDR)
4511 perf_output_put(handle, data->addr);
4513 if (sample_type & PERF_SAMPLE_ID)
4514 perf_output_put(handle, data->id);
4516 if (sample_type & PERF_SAMPLE_STREAM_ID)
4517 perf_output_put(handle, data->stream_id);
4519 if (sample_type & PERF_SAMPLE_CPU)
4520 perf_output_put(handle, data->cpu_entry);
4522 if (sample_type & PERF_SAMPLE_PERIOD)
4523 perf_output_put(handle, data->period);
4525 if (sample_type & PERF_SAMPLE_READ)
4526 perf_output_read(handle, event);
4528 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4529 if (data->callchain) {
4532 if (data->callchain)
4533 size += data->callchain->nr;
4535 size *= sizeof(u64);
4537 __output_copy(handle, data->callchain, size);
4540 perf_output_put(handle, nr);
4544 if (sample_type & PERF_SAMPLE_RAW) {
4546 perf_output_put(handle, data->raw->size);
4547 __output_copy(handle, data->raw->data,
4554 .size = sizeof(u32),
4557 perf_output_put(handle, raw);
4561 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4562 if (data->br_stack) {
4565 size = data->br_stack->nr
4566 * sizeof(struct perf_branch_entry);
4568 perf_output_put(handle, data->br_stack->nr);
4569 perf_output_copy(handle, data->br_stack->entries, size);
4572 * we always store at least the value of nr
4575 perf_output_put(handle, nr);
4579 if (sample_type & PERF_SAMPLE_REGS_USER) {
4580 u64 abi = data->regs_user.abi;
4583 * If there are no regs to dump, notice it through
4584 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4586 perf_output_put(handle, abi);
4589 u64 mask = event->attr.sample_regs_user;
4590 perf_output_sample_regs(handle,
4591 data->regs_user.regs,
4596 if (sample_type & PERF_SAMPLE_STACK_USER) {
4597 perf_output_sample_ustack(handle,
4598 data->stack_user_size,
4599 data->regs_user.regs);
4602 if (sample_type & PERF_SAMPLE_WEIGHT)
4603 perf_output_put(handle, data->weight);
4605 if (sample_type & PERF_SAMPLE_DATA_SRC)
4606 perf_output_put(handle, data->data_src.val);
4608 if (sample_type & PERF_SAMPLE_TRANSACTION)
4609 perf_output_put(handle, data->txn);
4611 if (!event->attr.watermark) {
4612 int wakeup_events = event->attr.wakeup_events;
4614 if (wakeup_events) {
4615 struct ring_buffer *rb = handle->rb;
4616 int events = local_inc_return(&rb->events);
4618 if (events >= wakeup_events) {
4619 local_sub(wakeup_events, &rb->events);
4620 local_inc(&rb->wakeup);
4626 void perf_prepare_sample(struct perf_event_header *header,
4627 struct perf_sample_data *data,
4628 struct perf_event *event,
4629 struct pt_regs *regs)
4631 u64 sample_type = event->attr.sample_type;
4633 header->type = PERF_RECORD_SAMPLE;
4634 header->size = sizeof(*header) + event->header_size;
4637 header->misc |= perf_misc_flags(regs);
4639 __perf_event_header__init_id(header, data, event);
4641 if (sample_type & PERF_SAMPLE_IP)
4642 data->ip = perf_instruction_pointer(regs);
4644 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4647 data->callchain = perf_callchain(event, regs);
4649 if (data->callchain)
4650 size += data->callchain->nr;
4652 header->size += size * sizeof(u64);
4655 if (sample_type & PERF_SAMPLE_RAW) {
4656 int size = sizeof(u32);
4659 size += data->raw->size;
4661 size += sizeof(u32);
4663 WARN_ON_ONCE(size & (sizeof(u64)-1));
4664 header->size += size;
4667 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4668 int size = sizeof(u64); /* nr */
4669 if (data->br_stack) {
4670 size += data->br_stack->nr
4671 * sizeof(struct perf_branch_entry);
4673 header->size += size;
4676 if (sample_type & PERF_SAMPLE_REGS_USER) {
4677 /* regs dump ABI info */
4678 int size = sizeof(u64);
4680 perf_sample_regs_user(&data->regs_user, regs);
4682 if (data->regs_user.regs) {
4683 u64 mask = event->attr.sample_regs_user;
4684 size += hweight64(mask) * sizeof(u64);
4687 header->size += size;
4690 if (sample_type & PERF_SAMPLE_STACK_USER) {
4692 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4693 * processed as the last one or have additional check added
4694 * in case new sample type is added, because we could eat
4695 * up the rest of the sample size.
4697 struct perf_regs_user *uregs = &data->regs_user;
4698 u16 stack_size = event->attr.sample_stack_user;
4699 u16 size = sizeof(u64);
4702 perf_sample_regs_user(uregs, regs);
4704 stack_size = perf_sample_ustack_size(stack_size, header->size,
4708 * If there is something to dump, add space for the dump
4709 * itself and for the field that tells the dynamic size,
4710 * which is how many have been actually dumped.
4713 size += sizeof(u64) + stack_size;
4715 data->stack_user_size = stack_size;
4716 header->size += size;
4720 static void perf_event_output(struct perf_event *event,
4721 struct perf_sample_data *data,
4722 struct pt_regs *regs)
4724 struct perf_output_handle handle;
4725 struct perf_event_header header;
4727 /* protect the callchain buffers */
4730 perf_prepare_sample(&header, data, event, regs);
4732 if (perf_output_begin(&handle, event, header.size))
4735 perf_output_sample(&handle, &header, data, event);
4737 perf_output_end(&handle);
4747 struct perf_read_event {
4748 struct perf_event_header header;
4755 perf_event_read_event(struct perf_event *event,
4756 struct task_struct *task)
4758 struct perf_output_handle handle;
4759 struct perf_sample_data sample;
4760 struct perf_read_event read_event = {
4762 .type = PERF_RECORD_READ,
4764 .size = sizeof(read_event) + event->read_size,
4766 .pid = perf_event_pid(event, task),
4767 .tid = perf_event_tid(event, task),
4771 perf_event_header__init_id(&read_event.header, &sample, event);
4772 ret = perf_output_begin(&handle, event, read_event.header.size);
4776 perf_output_put(&handle, read_event);
4777 perf_output_read(&handle, event);
4778 perf_event__output_id_sample(event, &handle, &sample);
4780 perf_output_end(&handle);
4783 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4786 perf_event_aux_ctx(struct perf_event_context *ctx,
4787 perf_event_aux_output_cb output,
4790 struct perf_event *event;
4792 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4793 if (event->state < PERF_EVENT_STATE_INACTIVE)
4795 if (!event_filter_match(event))
4797 output(event, data);
4802 perf_event_aux(perf_event_aux_output_cb output, void *data,
4803 struct perf_event_context *task_ctx)
4805 struct perf_cpu_context *cpuctx;
4806 struct perf_event_context *ctx;
4811 list_for_each_entry_rcu(pmu, &pmus, entry) {
4812 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
4813 if (cpuctx->unique_pmu != pmu)
4815 perf_event_aux_ctx(&cpuctx->ctx, output, data);
4818 ctxn = pmu->task_ctx_nr;
4821 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
4823 perf_event_aux_ctx(ctx, output, data);
4825 put_cpu_ptr(pmu->pmu_cpu_context);
4830 perf_event_aux_ctx(task_ctx, output, data);
4837 * task tracking -- fork/exit
4839 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
4842 struct perf_task_event {
4843 struct task_struct *task;
4844 struct perf_event_context *task_ctx;
4847 struct perf_event_header header;
4857 static int perf_event_task_match(struct perf_event *event)
4859 return event->attr.comm || event->attr.mmap ||
4860 event->attr.mmap2 || event->attr.mmap_data ||
4864 static void perf_event_task_output(struct perf_event *event,
4867 struct perf_task_event *task_event = data;
4868 struct perf_output_handle handle;
4869 struct perf_sample_data sample;
4870 struct task_struct *task = task_event->task;
4871 int ret, size = task_event->event_id.header.size;
4873 if (!perf_event_task_match(event))
4876 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
4878 ret = perf_output_begin(&handle, event,
4879 task_event->event_id.header.size);
4883 task_event->event_id.pid = perf_event_pid(event, task);
4884 task_event->event_id.ppid = perf_event_pid(event, current);
4886 task_event->event_id.tid = perf_event_tid(event, task);
4887 task_event->event_id.ptid = perf_event_tid(event, current);
4889 perf_output_put(&handle, task_event->event_id);
4891 perf_event__output_id_sample(event, &handle, &sample);
4893 perf_output_end(&handle);
4895 task_event->event_id.header.size = size;
4898 static void perf_event_task(struct task_struct *task,
4899 struct perf_event_context *task_ctx,
4902 struct perf_task_event task_event;
4904 if (!atomic_read(&nr_comm_events) &&
4905 !atomic_read(&nr_mmap_events) &&
4906 !atomic_read(&nr_task_events))
4909 task_event = (struct perf_task_event){
4911 .task_ctx = task_ctx,
4914 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
4916 .size = sizeof(task_event.event_id),
4922 .time = perf_clock(),
4926 perf_event_aux(perf_event_task_output,
4931 void perf_event_fork(struct task_struct *task)
4933 perf_event_task(task, NULL, 1);
4940 struct perf_comm_event {
4941 struct task_struct *task;
4946 struct perf_event_header header;
4953 static int perf_event_comm_match(struct perf_event *event)
4955 return event->attr.comm;
4958 static void perf_event_comm_output(struct perf_event *event,
4961 struct perf_comm_event *comm_event = data;
4962 struct perf_output_handle handle;
4963 struct perf_sample_data sample;
4964 int size = comm_event->event_id.header.size;
4967 if (!perf_event_comm_match(event))
4970 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
4971 ret = perf_output_begin(&handle, event,
4972 comm_event->event_id.header.size);
4977 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
4978 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
4980 perf_output_put(&handle, comm_event->event_id);
4981 __output_copy(&handle, comm_event->comm,
4982 comm_event->comm_size);
4984 perf_event__output_id_sample(event, &handle, &sample);
4986 perf_output_end(&handle);
4988 comm_event->event_id.header.size = size;
4991 static void perf_event_comm_event(struct perf_comm_event *comm_event)
4993 char comm[TASK_COMM_LEN];
4996 memset(comm, 0, sizeof(comm));
4997 strlcpy(comm, comm_event->task->comm, sizeof(comm));
4998 size = ALIGN(strlen(comm)+1, sizeof(u64));
5000 comm_event->comm = comm;
5001 comm_event->comm_size = size;
5003 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5005 perf_event_aux(perf_event_comm_output,
5010 void perf_event_comm(struct task_struct *task)
5012 struct perf_comm_event comm_event;
5013 struct perf_event_context *ctx;
5017 for_each_task_context_nr(ctxn) {
5018 ctx = task->perf_event_ctxp[ctxn];
5022 perf_event_enable_on_exec(ctx);
5026 if (!atomic_read(&nr_comm_events))
5029 comm_event = (struct perf_comm_event){
5035 .type = PERF_RECORD_COMM,
5044 perf_event_comm_event(&comm_event);
5051 struct perf_mmap_event {
5052 struct vm_area_struct *vma;
5054 const char *file_name;
5061 struct perf_event_header header;
5071 static int perf_event_mmap_match(struct perf_event *event,
5074 struct perf_mmap_event *mmap_event = data;
5075 struct vm_area_struct *vma = mmap_event->vma;
5076 int executable = vma->vm_flags & VM_EXEC;
5078 return (!executable && event->attr.mmap_data) ||
5079 (executable && (event->attr.mmap || event->attr.mmap2));
5082 static void perf_event_mmap_output(struct perf_event *event,
5085 struct perf_mmap_event *mmap_event = data;
5086 struct perf_output_handle handle;
5087 struct perf_sample_data sample;
5088 int size = mmap_event->event_id.header.size;
5091 if (!perf_event_mmap_match(event, data))
5094 if (event->attr.mmap2) {
5095 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5096 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5097 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5098 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5099 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5102 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5103 ret = perf_output_begin(&handle, event,
5104 mmap_event->event_id.header.size);
5108 mmap_event->event_id.pid = perf_event_pid(event, current);
5109 mmap_event->event_id.tid = perf_event_tid(event, current);
5111 perf_output_put(&handle, mmap_event->event_id);
5113 if (event->attr.mmap2) {
5114 perf_output_put(&handle, mmap_event->maj);
5115 perf_output_put(&handle, mmap_event->min);
5116 perf_output_put(&handle, mmap_event->ino);
5117 perf_output_put(&handle, mmap_event->ino_generation);
5120 __output_copy(&handle, mmap_event->file_name,
5121 mmap_event->file_size);
5123 perf_event__output_id_sample(event, &handle, &sample);
5125 perf_output_end(&handle);
5127 mmap_event->event_id.header.size = size;
5130 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5132 struct vm_area_struct *vma = mmap_event->vma;
5133 struct file *file = vma->vm_file;
5134 int maj = 0, min = 0;
5135 u64 ino = 0, gen = 0;
5142 struct inode *inode;
5145 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5151 * d_path() works from the end of the rb backwards, so we
5152 * need to add enough zero bytes after the string to handle
5153 * the 64bit alignment we do later.
5155 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5160 inode = file_inode(vma->vm_file);
5161 dev = inode->i_sb->s_dev;
5163 gen = inode->i_generation;
5168 name = (char *)arch_vma_name(vma);
5172 if (vma->vm_start <= vma->vm_mm->start_brk &&
5173 vma->vm_end >= vma->vm_mm->brk) {
5177 if (vma->vm_start <= vma->vm_mm->start_stack &&
5178 vma->vm_end >= vma->vm_mm->start_stack) {
5188 strlcpy(tmp, name, sizeof(tmp));
5192 * Since our buffer works in 8 byte units we need to align our string
5193 * size to a multiple of 8. However, we must guarantee the tail end is
5194 * zero'd out to avoid leaking random bits to userspace.
5196 size = strlen(name)+1;
5197 while (!IS_ALIGNED(size, sizeof(u64)))
5198 name[size++] = '\0';
5200 mmap_event->file_name = name;
5201 mmap_event->file_size = size;
5202 mmap_event->maj = maj;
5203 mmap_event->min = min;
5204 mmap_event->ino = ino;
5205 mmap_event->ino_generation = gen;
5207 if (!(vma->vm_flags & VM_EXEC))
5208 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5210 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5212 perf_event_aux(perf_event_mmap_output,
5219 void perf_event_mmap(struct vm_area_struct *vma)
5221 struct perf_mmap_event mmap_event;
5223 if (!atomic_read(&nr_mmap_events))
5226 mmap_event = (struct perf_mmap_event){
5232 .type = PERF_RECORD_MMAP,
5233 .misc = PERF_RECORD_MISC_USER,
5238 .start = vma->vm_start,
5239 .len = vma->vm_end - vma->vm_start,
5240 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5242 /* .maj (attr_mmap2 only) */
5243 /* .min (attr_mmap2 only) */
5244 /* .ino (attr_mmap2 only) */
5245 /* .ino_generation (attr_mmap2 only) */
5248 perf_event_mmap_event(&mmap_event);
5252 * IRQ throttle logging
5255 static void perf_log_throttle(struct perf_event *event, int enable)
5257 struct perf_output_handle handle;
5258 struct perf_sample_data sample;
5262 struct perf_event_header header;
5266 } throttle_event = {
5268 .type = PERF_RECORD_THROTTLE,
5270 .size = sizeof(throttle_event),
5272 .time = perf_clock(),
5273 .id = primary_event_id(event),
5274 .stream_id = event->id,
5278 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5280 perf_event_header__init_id(&throttle_event.header, &sample, event);
5282 ret = perf_output_begin(&handle, event,
5283 throttle_event.header.size);
5287 perf_output_put(&handle, throttle_event);
5288 perf_event__output_id_sample(event, &handle, &sample);
5289 perf_output_end(&handle);
5293 * Generic event overflow handling, sampling.
5296 static int __perf_event_overflow(struct perf_event *event,
5297 int throttle, struct perf_sample_data *data,
5298 struct pt_regs *regs)
5300 int events = atomic_read(&event->event_limit);
5301 struct hw_perf_event *hwc = &event->hw;
5306 * Non-sampling counters might still use the PMI to fold short
5307 * hardware counters, ignore those.
5309 if (unlikely(!is_sampling_event(event)))
5312 seq = __this_cpu_read(perf_throttled_seq);
5313 if (seq != hwc->interrupts_seq) {
5314 hwc->interrupts_seq = seq;
5315 hwc->interrupts = 1;
5318 if (unlikely(throttle
5319 && hwc->interrupts >= max_samples_per_tick)) {
5320 __this_cpu_inc(perf_throttled_count);
5321 hwc->interrupts = MAX_INTERRUPTS;
5322 perf_log_throttle(event, 0);
5323 tick_nohz_full_kick();
5328 if (event->attr.freq) {
5329 u64 now = perf_clock();
5330 s64 delta = now - hwc->freq_time_stamp;
5332 hwc->freq_time_stamp = now;
5334 if (delta > 0 && delta < 2*TICK_NSEC)
5335 perf_adjust_period(event, delta, hwc->last_period, true);
5339 * XXX event_limit might not quite work as expected on inherited
5343 event->pending_kill = POLL_IN;
5344 if (events && atomic_dec_and_test(&event->event_limit)) {
5346 event->pending_kill = POLL_HUP;
5347 event->pending_disable = 1;
5348 irq_work_queue(&event->pending);
5351 if (event->overflow_handler)
5352 event->overflow_handler(event, data, regs);
5354 perf_event_output(event, data, regs);
5356 if (event->fasync && event->pending_kill) {
5357 event->pending_wakeup = 1;
5358 irq_work_queue(&event->pending);
5364 int perf_event_overflow(struct perf_event *event,
5365 struct perf_sample_data *data,
5366 struct pt_regs *regs)
5368 return __perf_event_overflow(event, 1, data, regs);
5372 * Generic software event infrastructure
5375 struct swevent_htable {
5376 struct swevent_hlist *swevent_hlist;
5377 struct mutex hlist_mutex;
5380 /* Recursion avoidance in each contexts */
5381 int recursion[PERF_NR_CONTEXTS];
5384 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5387 * We directly increment event->count and keep a second value in
5388 * event->hw.period_left to count intervals. This period event
5389 * is kept in the range [-sample_period, 0] so that we can use the
5393 u64 perf_swevent_set_period(struct perf_event *event)
5395 struct hw_perf_event *hwc = &event->hw;
5396 u64 period = hwc->last_period;
5400 hwc->last_period = hwc->sample_period;
5403 old = val = local64_read(&hwc->period_left);
5407 nr = div64_u64(period + val, period);
5408 offset = nr * period;
5410 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5416 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5417 struct perf_sample_data *data,
5418 struct pt_regs *regs)
5420 struct hw_perf_event *hwc = &event->hw;
5424 overflow = perf_swevent_set_period(event);
5426 if (hwc->interrupts == MAX_INTERRUPTS)
5429 for (; overflow; overflow--) {
5430 if (__perf_event_overflow(event, throttle,
5433 * We inhibit the overflow from happening when
5434 * hwc->interrupts == MAX_INTERRUPTS.
5442 static void perf_swevent_event(struct perf_event *event, u64 nr,
5443 struct perf_sample_data *data,
5444 struct pt_regs *regs)
5446 struct hw_perf_event *hwc = &event->hw;
5448 local64_add(nr, &event->count);
5453 if (!is_sampling_event(event))
5456 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5458 return perf_swevent_overflow(event, 1, data, regs);
5460 data->period = event->hw.last_period;
5462 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5463 return perf_swevent_overflow(event, 1, data, regs);
5465 if (local64_add_negative(nr, &hwc->period_left))
5468 perf_swevent_overflow(event, 0, data, regs);
5471 static int perf_exclude_event(struct perf_event *event,
5472 struct pt_regs *regs)
5474 if (event->hw.state & PERF_HES_STOPPED)
5478 if (event->attr.exclude_user && user_mode(regs))
5481 if (event->attr.exclude_kernel && !user_mode(regs))
5488 static int perf_swevent_match(struct perf_event *event,
5489 enum perf_type_id type,
5491 struct perf_sample_data *data,
5492 struct pt_regs *regs)
5494 if (event->attr.type != type)
5497 if (event->attr.config != event_id)
5500 if (perf_exclude_event(event, regs))
5506 static inline u64 swevent_hash(u64 type, u32 event_id)
5508 u64 val = event_id | (type << 32);
5510 return hash_64(val, SWEVENT_HLIST_BITS);
5513 static inline struct hlist_head *
5514 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5516 u64 hash = swevent_hash(type, event_id);
5518 return &hlist->heads[hash];
5521 /* For the read side: events when they trigger */
5522 static inline struct hlist_head *
5523 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5525 struct swevent_hlist *hlist;
5527 hlist = rcu_dereference(swhash->swevent_hlist);
5531 return __find_swevent_head(hlist, type, event_id);
5534 /* For the event head insertion and removal in the hlist */
5535 static inline struct hlist_head *
5536 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5538 struct swevent_hlist *hlist;
5539 u32 event_id = event->attr.config;
5540 u64 type = event->attr.type;
5543 * Event scheduling is always serialized against hlist allocation
5544 * and release. Which makes the protected version suitable here.
5545 * The context lock guarantees that.
5547 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5548 lockdep_is_held(&event->ctx->lock));
5552 return __find_swevent_head(hlist, type, event_id);
5555 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5557 struct perf_sample_data *data,
5558 struct pt_regs *regs)
5560 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5561 struct perf_event *event;
5562 struct hlist_head *head;
5565 head = find_swevent_head_rcu(swhash, type, event_id);
5569 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5570 if (perf_swevent_match(event, type, event_id, data, regs))
5571 perf_swevent_event(event, nr, data, regs);
5577 int perf_swevent_get_recursion_context(void)
5579 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5581 return get_recursion_context(swhash->recursion);
5583 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5585 inline void perf_swevent_put_recursion_context(int rctx)
5587 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5589 put_recursion_context(swhash->recursion, rctx);
5592 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5594 struct perf_sample_data data;
5597 preempt_disable_notrace();
5598 rctx = perf_swevent_get_recursion_context();
5602 perf_sample_data_init(&data, addr, 0);
5604 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5606 perf_swevent_put_recursion_context(rctx);
5607 preempt_enable_notrace();
5610 static void perf_swevent_read(struct perf_event *event)
5614 static int perf_swevent_add(struct perf_event *event, int flags)
5616 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5617 struct hw_perf_event *hwc = &event->hw;
5618 struct hlist_head *head;
5620 if (is_sampling_event(event)) {
5621 hwc->last_period = hwc->sample_period;
5622 perf_swevent_set_period(event);
5625 hwc->state = !(flags & PERF_EF_START);
5627 head = find_swevent_head(swhash, event);
5628 if (WARN_ON_ONCE(!head))
5631 hlist_add_head_rcu(&event->hlist_entry, head);
5636 static void perf_swevent_del(struct perf_event *event, int flags)
5638 hlist_del_rcu(&event->hlist_entry);
5641 static void perf_swevent_start(struct perf_event *event, int flags)
5643 event->hw.state = 0;
5646 static void perf_swevent_stop(struct perf_event *event, int flags)
5648 event->hw.state = PERF_HES_STOPPED;
5651 /* Deref the hlist from the update side */
5652 static inline struct swevent_hlist *
5653 swevent_hlist_deref(struct swevent_htable *swhash)
5655 return rcu_dereference_protected(swhash->swevent_hlist,
5656 lockdep_is_held(&swhash->hlist_mutex));
5659 static void swevent_hlist_release(struct swevent_htable *swhash)
5661 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5666 rcu_assign_pointer(swhash->swevent_hlist, NULL);
5667 kfree_rcu(hlist, rcu_head);
5670 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5672 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5674 mutex_lock(&swhash->hlist_mutex);
5676 if (!--swhash->hlist_refcount)
5677 swevent_hlist_release(swhash);
5679 mutex_unlock(&swhash->hlist_mutex);
5682 static void swevent_hlist_put(struct perf_event *event)
5686 if (event->cpu != -1) {
5687 swevent_hlist_put_cpu(event, event->cpu);
5691 for_each_possible_cpu(cpu)
5692 swevent_hlist_put_cpu(event, cpu);
5695 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5697 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5700 mutex_lock(&swhash->hlist_mutex);
5702 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5703 struct swevent_hlist *hlist;
5705 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5710 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5712 swhash->hlist_refcount++;
5714 mutex_unlock(&swhash->hlist_mutex);
5719 static int swevent_hlist_get(struct perf_event *event)
5722 int cpu, failed_cpu;
5724 if (event->cpu != -1)
5725 return swevent_hlist_get_cpu(event, event->cpu);
5728 for_each_possible_cpu(cpu) {
5729 err = swevent_hlist_get_cpu(event, cpu);
5739 for_each_possible_cpu(cpu) {
5740 if (cpu == failed_cpu)
5742 swevent_hlist_put_cpu(event, cpu);
5749 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5751 static void sw_perf_event_destroy(struct perf_event *event)
5753 u64 event_id = event->attr.config;
5755 WARN_ON(event->parent);
5757 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5758 swevent_hlist_put(event);
5761 static int perf_swevent_init(struct perf_event *event)
5763 u64 event_id = event->attr.config;
5765 if (event->attr.type != PERF_TYPE_SOFTWARE)
5769 * no branch sampling for software events
5771 if (has_branch_stack(event))
5775 case PERF_COUNT_SW_CPU_CLOCK:
5776 case PERF_COUNT_SW_TASK_CLOCK:
5783 if (event_id >= PERF_COUNT_SW_MAX)
5786 if (!event->parent) {
5789 err = swevent_hlist_get(event);
5793 static_key_slow_inc(&perf_swevent_enabled[event_id]);
5794 event->destroy = sw_perf_event_destroy;
5800 static int perf_swevent_event_idx(struct perf_event *event)
5805 static struct pmu perf_swevent = {
5806 .task_ctx_nr = perf_sw_context,
5808 .event_init = perf_swevent_init,
5809 .add = perf_swevent_add,
5810 .del = perf_swevent_del,
5811 .start = perf_swevent_start,
5812 .stop = perf_swevent_stop,
5813 .read = perf_swevent_read,
5815 .event_idx = perf_swevent_event_idx,
5818 #ifdef CONFIG_EVENT_TRACING
5820 static int perf_tp_filter_match(struct perf_event *event,
5821 struct perf_sample_data *data)
5823 void *record = data->raw->data;
5825 if (likely(!event->filter) || filter_match_preds(event->filter, record))
5830 static int perf_tp_event_match(struct perf_event *event,
5831 struct perf_sample_data *data,
5832 struct pt_regs *regs)
5834 if (event->hw.state & PERF_HES_STOPPED)
5837 * All tracepoints are from kernel-space.
5839 if (event->attr.exclude_kernel)
5842 if (!perf_tp_filter_match(event, data))
5848 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
5849 struct pt_regs *regs, struct hlist_head *head, int rctx,
5850 struct task_struct *task)
5852 struct perf_sample_data data;
5853 struct perf_event *event;
5855 struct perf_raw_record raw = {
5860 perf_sample_data_init(&data, addr, 0);
5863 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5864 if (perf_tp_event_match(event, &data, regs))
5865 perf_swevent_event(event, count, &data, regs);
5869 * If we got specified a target task, also iterate its context and
5870 * deliver this event there too.
5872 if (task && task != current) {
5873 struct perf_event_context *ctx;
5874 struct trace_entry *entry = record;
5877 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
5881 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5882 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5884 if (event->attr.config != entry->type)
5886 if (perf_tp_event_match(event, &data, regs))
5887 perf_swevent_event(event, count, &data, regs);
5893 perf_swevent_put_recursion_context(rctx);
5895 EXPORT_SYMBOL_GPL(perf_tp_event);
5897 static void tp_perf_event_destroy(struct perf_event *event)
5899 perf_trace_destroy(event);
5902 static int perf_tp_event_init(struct perf_event *event)
5906 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5910 * no branch sampling for tracepoint events
5912 if (has_branch_stack(event))
5915 err = perf_trace_init(event);
5919 event->destroy = tp_perf_event_destroy;
5924 static struct pmu perf_tracepoint = {
5925 .task_ctx_nr = perf_sw_context,
5927 .event_init = perf_tp_event_init,
5928 .add = perf_trace_add,
5929 .del = perf_trace_del,
5930 .start = perf_swevent_start,
5931 .stop = perf_swevent_stop,
5932 .read = perf_swevent_read,
5934 .event_idx = perf_swevent_event_idx,
5937 static inline void perf_tp_register(void)
5939 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
5942 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5947 if (event->attr.type != PERF_TYPE_TRACEPOINT)
5950 filter_str = strndup_user(arg, PAGE_SIZE);
5951 if (IS_ERR(filter_str))
5952 return PTR_ERR(filter_str);
5954 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
5960 static void perf_event_free_filter(struct perf_event *event)
5962 ftrace_profile_free_filter(event);
5967 static inline void perf_tp_register(void)
5971 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
5976 static void perf_event_free_filter(struct perf_event *event)
5980 #endif /* CONFIG_EVENT_TRACING */
5982 #ifdef CONFIG_HAVE_HW_BREAKPOINT
5983 void perf_bp_event(struct perf_event *bp, void *data)
5985 struct perf_sample_data sample;
5986 struct pt_regs *regs = data;
5988 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
5990 if (!bp->hw.state && !perf_exclude_event(bp, regs))
5991 perf_swevent_event(bp, 1, &sample, regs);
5996 * hrtimer based swevent callback
5999 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6001 enum hrtimer_restart ret = HRTIMER_RESTART;
6002 struct perf_sample_data data;
6003 struct pt_regs *regs;
6004 struct perf_event *event;
6007 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6009 if (event->state != PERF_EVENT_STATE_ACTIVE)
6010 return HRTIMER_NORESTART;
6012 event->pmu->read(event);
6014 perf_sample_data_init(&data, 0, event->hw.last_period);
6015 regs = get_irq_regs();
6017 if (regs && !perf_exclude_event(event, regs)) {
6018 if (!(event->attr.exclude_idle && is_idle_task(current)))
6019 if (__perf_event_overflow(event, 1, &data, regs))
6020 ret = HRTIMER_NORESTART;
6023 period = max_t(u64, 10000, event->hw.sample_period);
6024 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6029 static void perf_swevent_start_hrtimer(struct perf_event *event)
6031 struct hw_perf_event *hwc = &event->hw;
6034 if (!is_sampling_event(event))
6037 period = local64_read(&hwc->period_left);
6042 local64_set(&hwc->period_left, 0);
6044 period = max_t(u64, 10000, hwc->sample_period);
6046 __hrtimer_start_range_ns(&hwc->hrtimer,
6047 ns_to_ktime(period), 0,
6048 HRTIMER_MODE_REL_PINNED, 0);
6051 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6053 struct hw_perf_event *hwc = &event->hw;
6055 if (is_sampling_event(event)) {
6056 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6057 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6059 hrtimer_cancel(&hwc->hrtimer);
6063 static void perf_swevent_init_hrtimer(struct perf_event *event)
6065 struct hw_perf_event *hwc = &event->hw;
6067 if (!is_sampling_event(event))
6070 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6071 hwc->hrtimer.function = perf_swevent_hrtimer;
6074 * Since hrtimers have a fixed rate, we can do a static freq->period
6075 * mapping and avoid the whole period adjust feedback stuff.
6077 if (event->attr.freq) {
6078 long freq = event->attr.sample_freq;
6080 event->attr.sample_period = NSEC_PER_SEC / freq;
6081 hwc->sample_period = event->attr.sample_period;
6082 local64_set(&hwc->period_left, hwc->sample_period);
6083 hwc->last_period = hwc->sample_period;
6084 event->attr.freq = 0;
6089 * Software event: cpu wall time clock
6092 static void cpu_clock_event_update(struct perf_event *event)
6097 now = local_clock();
6098 prev = local64_xchg(&event->hw.prev_count, now);
6099 local64_add(now - prev, &event->count);
6102 static void cpu_clock_event_start(struct perf_event *event, int flags)
6104 local64_set(&event->hw.prev_count, local_clock());
6105 perf_swevent_start_hrtimer(event);
6108 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6110 perf_swevent_cancel_hrtimer(event);
6111 cpu_clock_event_update(event);
6114 static int cpu_clock_event_add(struct perf_event *event, int flags)
6116 if (flags & PERF_EF_START)
6117 cpu_clock_event_start(event, flags);
6122 static void cpu_clock_event_del(struct perf_event *event, int flags)
6124 cpu_clock_event_stop(event, flags);
6127 static void cpu_clock_event_read(struct perf_event *event)
6129 cpu_clock_event_update(event);
6132 static int cpu_clock_event_init(struct perf_event *event)
6134 if (event->attr.type != PERF_TYPE_SOFTWARE)
6137 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6141 * no branch sampling for software events
6143 if (has_branch_stack(event))
6146 perf_swevent_init_hrtimer(event);
6151 static struct pmu perf_cpu_clock = {
6152 .task_ctx_nr = perf_sw_context,
6154 .event_init = cpu_clock_event_init,
6155 .add = cpu_clock_event_add,
6156 .del = cpu_clock_event_del,
6157 .start = cpu_clock_event_start,
6158 .stop = cpu_clock_event_stop,
6159 .read = cpu_clock_event_read,
6161 .event_idx = perf_swevent_event_idx,
6165 * Software event: task time clock
6168 static void task_clock_event_update(struct perf_event *event, u64 now)
6173 prev = local64_xchg(&event->hw.prev_count, now);
6175 local64_add(delta, &event->count);
6178 static void task_clock_event_start(struct perf_event *event, int flags)
6180 local64_set(&event->hw.prev_count, event->ctx->time);
6181 perf_swevent_start_hrtimer(event);
6184 static void task_clock_event_stop(struct perf_event *event, int flags)
6186 perf_swevent_cancel_hrtimer(event);
6187 task_clock_event_update(event, event->ctx->time);
6190 static int task_clock_event_add(struct perf_event *event, int flags)
6192 if (flags & PERF_EF_START)
6193 task_clock_event_start(event, flags);
6198 static void task_clock_event_del(struct perf_event *event, int flags)
6200 task_clock_event_stop(event, PERF_EF_UPDATE);
6203 static void task_clock_event_read(struct perf_event *event)
6205 u64 now = perf_clock();
6206 u64 delta = now - event->ctx->timestamp;
6207 u64 time = event->ctx->time + delta;
6209 task_clock_event_update(event, time);
6212 static int task_clock_event_init(struct perf_event *event)
6214 if (event->attr.type != PERF_TYPE_SOFTWARE)
6217 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6221 * no branch sampling for software events
6223 if (has_branch_stack(event))
6226 perf_swevent_init_hrtimer(event);
6231 static struct pmu perf_task_clock = {
6232 .task_ctx_nr = perf_sw_context,
6234 .event_init = task_clock_event_init,
6235 .add = task_clock_event_add,
6236 .del = task_clock_event_del,
6237 .start = task_clock_event_start,
6238 .stop = task_clock_event_stop,
6239 .read = task_clock_event_read,
6241 .event_idx = perf_swevent_event_idx,
6244 static void perf_pmu_nop_void(struct pmu *pmu)
6248 static int perf_pmu_nop_int(struct pmu *pmu)
6253 static void perf_pmu_start_txn(struct pmu *pmu)
6255 perf_pmu_disable(pmu);
6258 static int perf_pmu_commit_txn(struct pmu *pmu)
6260 perf_pmu_enable(pmu);
6264 static void perf_pmu_cancel_txn(struct pmu *pmu)
6266 perf_pmu_enable(pmu);
6269 static int perf_event_idx_default(struct perf_event *event)
6271 return event->hw.idx + 1;
6275 * Ensures all contexts with the same task_ctx_nr have the same
6276 * pmu_cpu_context too.
6278 static void *find_pmu_context(int ctxn)
6285 list_for_each_entry(pmu, &pmus, entry) {
6286 if (pmu->task_ctx_nr == ctxn)
6287 return pmu->pmu_cpu_context;
6293 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6297 for_each_possible_cpu(cpu) {
6298 struct perf_cpu_context *cpuctx;
6300 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6302 if (cpuctx->unique_pmu == old_pmu)
6303 cpuctx->unique_pmu = pmu;
6307 static void free_pmu_context(struct pmu *pmu)
6311 mutex_lock(&pmus_lock);
6313 * Like a real lame refcount.
6315 list_for_each_entry(i, &pmus, entry) {
6316 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6317 update_pmu_context(i, pmu);
6322 free_percpu(pmu->pmu_cpu_context);
6324 mutex_unlock(&pmus_lock);
6326 static struct idr pmu_idr;
6329 type_show(struct device *dev, struct device_attribute *attr, char *page)
6331 struct pmu *pmu = dev_get_drvdata(dev);
6333 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6335 static DEVICE_ATTR_RO(type);
6338 perf_event_mux_interval_ms_show(struct device *dev,
6339 struct device_attribute *attr,
6342 struct pmu *pmu = dev_get_drvdata(dev);
6344 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6348 perf_event_mux_interval_ms_store(struct device *dev,
6349 struct device_attribute *attr,
6350 const char *buf, size_t count)
6352 struct pmu *pmu = dev_get_drvdata(dev);
6353 int timer, cpu, ret;
6355 ret = kstrtoint(buf, 0, &timer);
6362 /* same value, noting to do */
6363 if (timer == pmu->hrtimer_interval_ms)
6366 pmu->hrtimer_interval_ms = timer;
6368 /* update all cpuctx for this PMU */
6369 for_each_possible_cpu(cpu) {
6370 struct perf_cpu_context *cpuctx;
6371 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6372 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6374 if (hrtimer_active(&cpuctx->hrtimer))
6375 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6380 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6382 static struct attribute *pmu_dev_attrs[] = {
6383 &dev_attr_type.attr,
6384 &dev_attr_perf_event_mux_interval_ms.attr,
6387 ATTRIBUTE_GROUPS(pmu_dev);
6389 static int pmu_bus_running;
6390 static struct bus_type pmu_bus = {
6391 .name = "event_source",
6392 .dev_groups = pmu_dev_groups,
6395 static void pmu_dev_release(struct device *dev)
6400 static int pmu_dev_alloc(struct pmu *pmu)
6404 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6408 pmu->dev->groups = pmu->attr_groups;
6409 device_initialize(pmu->dev);
6410 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6414 dev_set_drvdata(pmu->dev, pmu);
6415 pmu->dev->bus = &pmu_bus;
6416 pmu->dev->release = pmu_dev_release;
6417 ret = device_add(pmu->dev);
6425 put_device(pmu->dev);
6429 static struct lock_class_key cpuctx_mutex;
6430 static struct lock_class_key cpuctx_lock;
6432 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6436 mutex_lock(&pmus_lock);
6438 pmu->pmu_disable_count = alloc_percpu(int);
6439 if (!pmu->pmu_disable_count)
6448 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6456 if (pmu_bus_running) {
6457 ret = pmu_dev_alloc(pmu);
6463 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6464 if (pmu->pmu_cpu_context)
6465 goto got_cpu_context;
6468 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6469 if (!pmu->pmu_cpu_context)
6472 for_each_possible_cpu(cpu) {
6473 struct perf_cpu_context *cpuctx;
6475 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6476 __perf_event_init_context(&cpuctx->ctx);
6477 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6478 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6479 cpuctx->ctx.type = cpu_context;
6480 cpuctx->ctx.pmu = pmu;
6482 __perf_cpu_hrtimer_init(cpuctx, cpu);
6484 INIT_LIST_HEAD(&cpuctx->rotation_list);
6485 cpuctx->unique_pmu = pmu;
6489 if (!pmu->start_txn) {
6490 if (pmu->pmu_enable) {
6492 * If we have pmu_enable/pmu_disable calls, install
6493 * transaction stubs that use that to try and batch
6494 * hardware accesses.
6496 pmu->start_txn = perf_pmu_start_txn;
6497 pmu->commit_txn = perf_pmu_commit_txn;
6498 pmu->cancel_txn = perf_pmu_cancel_txn;
6500 pmu->start_txn = perf_pmu_nop_void;
6501 pmu->commit_txn = perf_pmu_nop_int;
6502 pmu->cancel_txn = perf_pmu_nop_void;
6506 if (!pmu->pmu_enable) {
6507 pmu->pmu_enable = perf_pmu_nop_void;
6508 pmu->pmu_disable = perf_pmu_nop_void;
6511 if (!pmu->event_idx)
6512 pmu->event_idx = perf_event_idx_default;
6514 list_add_rcu(&pmu->entry, &pmus);
6517 mutex_unlock(&pmus_lock);
6522 device_del(pmu->dev);
6523 put_device(pmu->dev);
6526 if (pmu->type >= PERF_TYPE_MAX)
6527 idr_remove(&pmu_idr, pmu->type);
6530 free_percpu(pmu->pmu_disable_count);
6534 void perf_pmu_unregister(struct pmu *pmu)
6536 mutex_lock(&pmus_lock);
6537 list_del_rcu(&pmu->entry);
6538 mutex_unlock(&pmus_lock);
6541 * We dereference the pmu list under both SRCU and regular RCU, so
6542 * synchronize against both of those.
6544 synchronize_srcu(&pmus_srcu);
6547 free_percpu(pmu->pmu_disable_count);
6548 if (pmu->type >= PERF_TYPE_MAX)
6549 idr_remove(&pmu_idr, pmu->type);
6550 device_del(pmu->dev);
6551 put_device(pmu->dev);
6552 free_pmu_context(pmu);
6555 struct pmu *perf_init_event(struct perf_event *event)
6557 struct pmu *pmu = NULL;
6561 idx = srcu_read_lock(&pmus_srcu);
6564 pmu = idr_find(&pmu_idr, event->attr.type);
6568 ret = pmu->event_init(event);
6574 list_for_each_entry_rcu(pmu, &pmus, entry) {
6576 ret = pmu->event_init(event);
6580 if (ret != -ENOENT) {
6585 pmu = ERR_PTR(-ENOENT);
6587 srcu_read_unlock(&pmus_srcu, idx);
6592 static void account_event_cpu(struct perf_event *event, int cpu)
6597 if (has_branch_stack(event)) {
6598 if (!(event->attach_state & PERF_ATTACH_TASK))
6599 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6601 if (is_cgroup_event(event))
6602 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6605 static void account_event(struct perf_event *event)
6610 if (event->attach_state & PERF_ATTACH_TASK)
6611 static_key_slow_inc(&perf_sched_events.key);
6612 if (event->attr.mmap || event->attr.mmap_data)
6613 atomic_inc(&nr_mmap_events);
6614 if (event->attr.comm)
6615 atomic_inc(&nr_comm_events);
6616 if (event->attr.task)
6617 atomic_inc(&nr_task_events);
6618 if (event->attr.freq) {
6619 if (atomic_inc_return(&nr_freq_events) == 1)
6620 tick_nohz_full_kick_all();
6622 if (has_branch_stack(event))
6623 static_key_slow_inc(&perf_sched_events.key);
6624 if (is_cgroup_event(event))
6625 static_key_slow_inc(&perf_sched_events.key);
6627 account_event_cpu(event, event->cpu);
6631 * Allocate and initialize a event structure
6633 static struct perf_event *
6634 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6635 struct task_struct *task,
6636 struct perf_event *group_leader,
6637 struct perf_event *parent_event,
6638 perf_overflow_handler_t overflow_handler,
6642 struct perf_event *event;
6643 struct hw_perf_event *hwc;
6646 if ((unsigned)cpu >= nr_cpu_ids) {
6647 if (!task || cpu != -1)
6648 return ERR_PTR(-EINVAL);
6651 event = kzalloc(sizeof(*event), GFP_KERNEL);
6653 return ERR_PTR(-ENOMEM);
6656 * Single events are their own group leaders, with an
6657 * empty sibling list:
6660 group_leader = event;
6662 mutex_init(&event->child_mutex);
6663 INIT_LIST_HEAD(&event->child_list);
6665 INIT_LIST_HEAD(&event->group_entry);
6666 INIT_LIST_HEAD(&event->event_entry);
6667 INIT_LIST_HEAD(&event->sibling_list);
6668 INIT_LIST_HEAD(&event->rb_entry);
6670 init_waitqueue_head(&event->waitq);
6671 init_irq_work(&event->pending, perf_pending_event);
6673 mutex_init(&event->mmap_mutex);
6675 atomic_long_set(&event->refcount, 1);
6677 event->attr = *attr;
6678 event->group_leader = group_leader;
6682 event->parent = parent_event;
6684 event->ns = get_pid_ns(task_active_pid_ns(current));
6685 event->id = atomic64_inc_return(&perf_event_id);
6687 event->state = PERF_EVENT_STATE_INACTIVE;
6690 event->attach_state = PERF_ATTACH_TASK;
6692 if (attr->type == PERF_TYPE_TRACEPOINT)
6693 event->hw.tp_target = task;
6694 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6696 * hw_breakpoint is a bit difficult here..
6698 else if (attr->type == PERF_TYPE_BREAKPOINT)
6699 event->hw.bp_target = task;
6703 if (!overflow_handler && parent_event) {
6704 overflow_handler = parent_event->overflow_handler;
6705 context = parent_event->overflow_handler_context;
6708 event->overflow_handler = overflow_handler;
6709 event->overflow_handler_context = context;
6711 perf_event__state_init(event);
6716 hwc->sample_period = attr->sample_period;
6717 if (attr->freq && attr->sample_freq)
6718 hwc->sample_period = 1;
6719 hwc->last_period = hwc->sample_period;
6721 local64_set(&hwc->period_left, hwc->sample_period);
6724 * we currently do not support PERF_FORMAT_GROUP on inherited events
6726 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6729 pmu = perf_init_event(event);
6732 else if (IS_ERR(pmu)) {
6737 if (!event->parent) {
6738 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6739 err = get_callchain_buffers();
6749 event->destroy(event);
6752 put_pid_ns(event->ns);
6755 return ERR_PTR(err);
6758 static int perf_copy_attr(struct perf_event_attr __user *uattr,
6759 struct perf_event_attr *attr)
6764 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
6768 * zero the full structure, so that a short copy will be nice.
6770 memset(attr, 0, sizeof(*attr));
6772 ret = get_user(size, &uattr->size);
6776 if (size > PAGE_SIZE) /* silly large */
6779 if (!size) /* abi compat */
6780 size = PERF_ATTR_SIZE_VER0;
6782 if (size < PERF_ATTR_SIZE_VER0)
6786 * If we're handed a bigger struct than we know of,
6787 * ensure all the unknown bits are 0 - i.e. new
6788 * user-space does not rely on any kernel feature
6789 * extensions we dont know about yet.
6791 if (size > sizeof(*attr)) {
6792 unsigned char __user *addr;
6793 unsigned char __user *end;
6796 addr = (void __user *)uattr + sizeof(*attr);
6797 end = (void __user *)uattr + size;
6799 for (; addr < end; addr++) {
6800 ret = get_user(val, addr);
6806 size = sizeof(*attr);
6809 ret = copy_from_user(attr, uattr, size);
6813 /* disabled for now */
6817 if (attr->__reserved_1)
6820 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
6823 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
6826 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
6827 u64 mask = attr->branch_sample_type;
6829 /* only using defined bits */
6830 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
6833 /* at least one branch bit must be set */
6834 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
6837 /* propagate priv level, when not set for branch */
6838 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
6840 /* exclude_kernel checked on syscall entry */
6841 if (!attr->exclude_kernel)
6842 mask |= PERF_SAMPLE_BRANCH_KERNEL;
6844 if (!attr->exclude_user)
6845 mask |= PERF_SAMPLE_BRANCH_USER;
6847 if (!attr->exclude_hv)
6848 mask |= PERF_SAMPLE_BRANCH_HV;
6850 * adjust user setting (for HW filter setup)
6852 attr->branch_sample_type = mask;
6854 /* privileged levels capture (kernel, hv): check permissions */
6855 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
6856 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6860 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
6861 ret = perf_reg_validate(attr->sample_regs_user);
6866 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
6867 if (!arch_perf_have_user_stack_dump())
6871 * We have __u32 type for the size, but so far
6872 * we can only use __u16 as maximum due to the
6873 * __u16 sample size limit.
6875 if (attr->sample_stack_user >= USHRT_MAX)
6877 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
6885 put_user(sizeof(*attr), &uattr->size);
6891 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
6893 struct ring_buffer *rb = NULL, *old_rb = NULL;
6899 /* don't allow circular references */
6900 if (event == output_event)
6904 * Don't allow cross-cpu buffers
6906 if (output_event->cpu != event->cpu)
6910 * If its not a per-cpu rb, it must be the same task.
6912 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
6916 mutex_lock(&event->mmap_mutex);
6917 /* Can't redirect output if we've got an active mmap() */
6918 if (atomic_read(&event->mmap_count))
6924 /* get the rb we want to redirect to */
6925 rb = ring_buffer_get(output_event);
6931 ring_buffer_detach(event, old_rb);
6934 ring_buffer_attach(event, rb);
6936 rcu_assign_pointer(event->rb, rb);
6939 ring_buffer_put(old_rb);
6941 * Since we detached before setting the new rb, so that we
6942 * could attach the new rb, we could have missed a wakeup.
6945 wake_up_all(&event->waitq);
6950 mutex_unlock(&event->mmap_mutex);
6957 * sys_perf_event_open - open a performance event, associate it to a task/cpu
6959 * @attr_uptr: event_id type attributes for monitoring/sampling
6962 * @group_fd: group leader event fd
6964 SYSCALL_DEFINE5(perf_event_open,
6965 struct perf_event_attr __user *, attr_uptr,
6966 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
6968 struct perf_event *group_leader = NULL, *output_event = NULL;
6969 struct perf_event *event, *sibling;
6970 struct perf_event_attr attr;
6971 struct perf_event_context *ctx;
6972 struct file *event_file = NULL;
6973 struct fd group = {NULL, 0};
6974 struct task_struct *task = NULL;
6980 /* for future expandability... */
6981 if (flags & ~PERF_FLAG_ALL)
6984 err = perf_copy_attr(attr_uptr, &attr);
6988 if (!attr.exclude_kernel) {
6989 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
6994 if (attr.sample_freq > sysctl_perf_event_sample_rate)
6999 * In cgroup mode, the pid argument is used to pass the fd
7000 * opened to the cgroup directory in cgroupfs. The cpu argument
7001 * designates the cpu on which to monitor threads from that
7004 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7007 event_fd = get_unused_fd();
7011 if (group_fd != -1) {
7012 err = perf_fget_light(group_fd, &group);
7015 group_leader = group.file->private_data;
7016 if (flags & PERF_FLAG_FD_OUTPUT)
7017 output_event = group_leader;
7018 if (flags & PERF_FLAG_FD_NO_GROUP)
7019 group_leader = NULL;
7022 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7023 task = find_lively_task_by_vpid(pid);
7025 err = PTR_ERR(task);
7032 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7034 if (IS_ERR(event)) {
7035 err = PTR_ERR(event);
7039 if (flags & PERF_FLAG_PID_CGROUP) {
7040 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7042 __free_event(event);
7047 account_event(event);
7050 * Special case software events and allow them to be part of
7051 * any hardware group.
7056 (is_software_event(event) != is_software_event(group_leader))) {
7057 if (is_software_event(event)) {
7059 * If event and group_leader are not both a software
7060 * event, and event is, then group leader is not.
7062 * Allow the addition of software events to !software
7063 * groups, this is safe because software events never
7066 pmu = group_leader->pmu;
7067 } else if (is_software_event(group_leader) &&
7068 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7070 * In case the group is a pure software group, and we
7071 * try to add a hardware event, move the whole group to
7072 * the hardware context.
7079 * Get the target context (task or percpu):
7081 ctx = find_get_context(pmu, task, event->cpu);
7088 put_task_struct(task);
7093 * Look up the group leader (we will attach this event to it):
7099 * Do not allow a recursive hierarchy (this new sibling
7100 * becoming part of another group-sibling):
7102 if (group_leader->group_leader != group_leader)
7105 * Do not allow to attach to a group in a different
7106 * task or CPU context:
7109 if (group_leader->ctx->type != ctx->type)
7112 if (group_leader->ctx != ctx)
7117 * Only a group leader can be exclusive or pinned
7119 if (attr.exclusive || attr.pinned)
7124 err = perf_event_set_output(event, output_event);
7129 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
7130 if (IS_ERR(event_file)) {
7131 err = PTR_ERR(event_file);
7136 struct perf_event_context *gctx = group_leader->ctx;
7138 mutex_lock(&gctx->mutex);
7139 perf_remove_from_context(group_leader);
7142 * Removing from the context ends up with disabled
7143 * event. What we want here is event in the initial
7144 * startup state, ready to be add into new context.
7146 perf_event__state_init(group_leader);
7147 list_for_each_entry(sibling, &group_leader->sibling_list,
7149 perf_remove_from_context(sibling);
7150 perf_event__state_init(sibling);
7153 mutex_unlock(&gctx->mutex);
7157 WARN_ON_ONCE(ctx->parent_ctx);
7158 mutex_lock(&ctx->mutex);
7162 perf_install_in_context(ctx, group_leader, event->cpu);
7164 list_for_each_entry(sibling, &group_leader->sibling_list,
7166 perf_install_in_context(ctx, sibling, event->cpu);
7171 perf_install_in_context(ctx, event, event->cpu);
7172 perf_unpin_context(ctx);
7173 mutex_unlock(&ctx->mutex);
7177 event->owner = current;
7179 mutex_lock(¤t->perf_event_mutex);
7180 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7181 mutex_unlock(¤t->perf_event_mutex);
7184 * Precalculate sample_data sizes
7186 perf_event__header_size(event);
7187 perf_event__id_header_size(event);
7190 * Drop the reference on the group_event after placing the
7191 * new event on the sibling_list. This ensures destruction
7192 * of the group leader will find the pointer to itself in
7193 * perf_group_detach().
7196 fd_install(event_fd, event_file);
7200 perf_unpin_context(ctx);
7207 put_task_struct(task);
7211 put_unused_fd(event_fd);
7216 * perf_event_create_kernel_counter
7218 * @attr: attributes of the counter to create
7219 * @cpu: cpu in which the counter is bound
7220 * @task: task to profile (NULL for percpu)
7223 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7224 struct task_struct *task,
7225 perf_overflow_handler_t overflow_handler,
7228 struct perf_event_context *ctx;
7229 struct perf_event *event;
7233 * Get the target context (task or percpu):
7236 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7237 overflow_handler, context);
7238 if (IS_ERR(event)) {
7239 err = PTR_ERR(event);
7243 account_event(event);
7245 ctx = find_get_context(event->pmu, task, cpu);
7251 WARN_ON_ONCE(ctx->parent_ctx);
7252 mutex_lock(&ctx->mutex);
7253 perf_install_in_context(ctx, event, cpu);
7254 perf_unpin_context(ctx);
7255 mutex_unlock(&ctx->mutex);
7262 return ERR_PTR(err);
7264 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7266 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7268 struct perf_event_context *src_ctx;
7269 struct perf_event_context *dst_ctx;
7270 struct perf_event *event, *tmp;
7273 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7274 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7276 mutex_lock(&src_ctx->mutex);
7277 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7279 perf_remove_from_context(event);
7280 unaccount_event_cpu(event, src_cpu);
7282 list_add(&event->migrate_entry, &events);
7284 mutex_unlock(&src_ctx->mutex);
7288 mutex_lock(&dst_ctx->mutex);
7289 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7290 list_del(&event->migrate_entry);
7291 if (event->state >= PERF_EVENT_STATE_OFF)
7292 event->state = PERF_EVENT_STATE_INACTIVE;
7293 account_event_cpu(event, dst_cpu);
7294 perf_install_in_context(dst_ctx, event, dst_cpu);
7297 mutex_unlock(&dst_ctx->mutex);
7299 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7301 static void sync_child_event(struct perf_event *child_event,
7302 struct task_struct *child)
7304 struct perf_event *parent_event = child_event->parent;
7307 if (child_event->attr.inherit_stat)
7308 perf_event_read_event(child_event, child);
7310 child_val = perf_event_count(child_event);
7313 * Add back the child's count to the parent's count:
7315 atomic64_add(child_val, &parent_event->child_count);
7316 atomic64_add(child_event->total_time_enabled,
7317 &parent_event->child_total_time_enabled);
7318 atomic64_add(child_event->total_time_running,
7319 &parent_event->child_total_time_running);
7322 * Remove this event from the parent's list
7324 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7325 mutex_lock(&parent_event->child_mutex);
7326 list_del_init(&child_event->child_list);
7327 mutex_unlock(&parent_event->child_mutex);
7330 * Release the parent event, if this was the last
7333 put_event(parent_event);
7337 __perf_event_exit_task(struct perf_event *child_event,
7338 struct perf_event_context *child_ctx,
7339 struct task_struct *child)
7341 if (child_event->parent) {
7342 raw_spin_lock_irq(&child_ctx->lock);
7343 perf_group_detach(child_event);
7344 raw_spin_unlock_irq(&child_ctx->lock);
7347 perf_remove_from_context(child_event);
7350 * It can happen that the parent exits first, and has events
7351 * that are still around due to the child reference. These
7352 * events need to be zapped.
7354 if (child_event->parent) {
7355 sync_child_event(child_event, child);
7356 free_event(child_event);
7360 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7362 struct perf_event *child_event, *tmp;
7363 struct perf_event_context *child_ctx;
7364 unsigned long flags;
7366 if (likely(!child->perf_event_ctxp[ctxn])) {
7367 perf_event_task(child, NULL, 0);
7371 local_irq_save(flags);
7373 * We can't reschedule here because interrupts are disabled,
7374 * and either child is current or it is a task that can't be
7375 * scheduled, so we are now safe from rescheduling changing
7378 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7381 * Take the context lock here so that if find_get_context is
7382 * reading child->perf_event_ctxp, we wait until it has
7383 * incremented the context's refcount before we do put_ctx below.
7385 raw_spin_lock(&child_ctx->lock);
7386 task_ctx_sched_out(child_ctx);
7387 child->perf_event_ctxp[ctxn] = NULL;
7389 * If this context is a clone; unclone it so it can't get
7390 * swapped to another process while we're removing all
7391 * the events from it.
7393 unclone_ctx(child_ctx);
7394 update_context_time(child_ctx);
7395 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7398 * Report the task dead after unscheduling the events so that we
7399 * won't get any samples after PERF_RECORD_EXIT. We can however still
7400 * get a few PERF_RECORD_READ events.
7402 perf_event_task(child, child_ctx, 0);
7405 * We can recurse on the same lock type through:
7407 * __perf_event_exit_task()
7408 * sync_child_event()
7410 * mutex_lock(&ctx->mutex)
7412 * But since its the parent context it won't be the same instance.
7414 mutex_lock(&child_ctx->mutex);
7417 list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
7419 __perf_event_exit_task(child_event, child_ctx, child);
7421 list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
7423 __perf_event_exit_task(child_event, child_ctx, child);
7426 * If the last event was a group event, it will have appended all
7427 * its siblings to the list, but we obtained 'tmp' before that which
7428 * will still point to the list head terminating the iteration.
7430 if (!list_empty(&child_ctx->pinned_groups) ||
7431 !list_empty(&child_ctx->flexible_groups))
7434 mutex_unlock(&child_ctx->mutex);
7440 * When a child task exits, feed back event values to parent events.
7442 void perf_event_exit_task(struct task_struct *child)
7444 struct perf_event *event, *tmp;
7447 mutex_lock(&child->perf_event_mutex);
7448 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7450 list_del_init(&event->owner_entry);
7453 * Ensure the list deletion is visible before we clear
7454 * the owner, closes a race against perf_release() where
7455 * we need to serialize on the owner->perf_event_mutex.
7458 event->owner = NULL;
7460 mutex_unlock(&child->perf_event_mutex);
7462 for_each_task_context_nr(ctxn)
7463 perf_event_exit_task_context(child, ctxn);
7466 static void perf_free_event(struct perf_event *event,
7467 struct perf_event_context *ctx)
7469 struct perf_event *parent = event->parent;
7471 if (WARN_ON_ONCE(!parent))
7474 mutex_lock(&parent->child_mutex);
7475 list_del_init(&event->child_list);
7476 mutex_unlock(&parent->child_mutex);
7480 perf_group_detach(event);
7481 list_del_event(event, ctx);
7486 * free an unexposed, unused context as created by inheritance by
7487 * perf_event_init_task below, used by fork() in case of fail.
7489 void perf_event_free_task(struct task_struct *task)
7491 struct perf_event_context *ctx;
7492 struct perf_event *event, *tmp;
7495 for_each_task_context_nr(ctxn) {
7496 ctx = task->perf_event_ctxp[ctxn];
7500 mutex_lock(&ctx->mutex);
7502 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7504 perf_free_event(event, ctx);
7506 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7508 perf_free_event(event, ctx);
7510 if (!list_empty(&ctx->pinned_groups) ||
7511 !list_empty(&ctx->flexible_groups))
7514 mutex_unlock(&ctx->mutex);
7520 void perf_event_delayed_put(struct task_struct *task)
7524 for_each_task_context_nr(ctxn)
7525 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7529 * inherit a event from parent task to child task:
7531 static struct perf_event *
7532 inherit_event(struct perf_event *parent_event,
7533 struct task_struct *parent,
7534 struct perf_event_context *parent_ctx,
7535 struct task_struct *child,
7536 struct perf_event *group_leader,
7537 struct perf_event_context *child_ctx)
7539 struct perf_event *child_event;
7540 unsigned long flags;
7543 * Instead of creating recursive hierarchies of events,
7544 * we link inherited events back to the original parent,
7545 * which has a filp for sure, which we use as the reference
7548 if (parent_event->parent)
7549 parent_event = parent_event->parent;
7551 child_event = perf_event_alloc(&parent_event->attr,
7554 group_leader, parent_event,
7556 if (IS_ERR(child_event))
7559 if (!atomic_long_inc_not_zero(&parent_event->refcount)) {
7560 free_event(child_event);
7567 * Make the child state follow the state of the parent event,
7568 * not its attr.disabled bit. We hold the parent's mutex,
7569 * so we won't race with perf_event_{en, dis}able_family.
7571 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7572 child_event->state = PERF_EVENT_STATE_INACTIVE;
7574 child_event->state = PERF_EVENT_STATE_OFF;
7576 if (parent_event->attr.freq) {
7577 u64 sample_period = parent_event->hw.sample_period;
7578 struct hw_perf_event *hwc = &child_event->hw;
7580 hwc->sample_period = sample_period;
7581 hwc->last_period = sample_period;
7583 local64_set(&hwc->period_left, sample_period);
7586 child_event->ctx = child_ctx;
7587 child_event->overflow_handler = parent_event->overflow_handler;
7588 child_event->overflow_handler_context
7589 = parent_event->overflow_handler_context;
7592 * Precalculate sample_data sizes
7594 perf_event__header_size(child_event);
7595 perf_event__id_header_size(child_event);
7598 * Link it up in the child's context:
7600 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7601 add_event_to_ctx(child_event, child_ctx);
7602 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7605 * Link this into the parent event's child list
7607 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7608 mutex_lock(&parent_event->child_mutex);
7609 list_add_tail(&child_event->child_list, &parent_event->child_list);
7610 mutex_unlock(&parent_event->child_mutex);
7615 static int inherit_group(struct perf_event *parent_event,
7616 struct task_struct *parent,
7617 struct perf_event_context *parent_ctx,
7618 struct task_struct *child,
7619 struct perf_event_context *child_ctx)
7621 struct perf_event *leader;
7622 struct perf_event *sub;
7623 struct perf_event *child_ctr;
7625 leader = inherit_event(parent_event, parent, parent_ctx,
7626 child, NULL, child_ctx);
7628 return PTR_ERR(leader);
7629 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7630 child_ctr = inherit_event(sub, parent, parent_ctx,
7631 child, leader, child_ctx);
7632 if (IS_ERR(child_ctr))
7633 return PTR_ERR(child_ctr);
7639 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7640 struct perf_event_context *parent_ctx,
7641 struct task_struct *child, int ctxn,
7645 struct perf_event_context *child_ctx;
7647 if (!event->attr.inherit) {
7652 child_ctx = child->perf_event_ctxp[ctxn];
7655 * This is executed from the parent task context, so
7656 * inherit events that have been marked for cloning.
7657 * First allocate and initialize a context for the
7661 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7665 child->perf_event_ctxp[ctxn] = child_ctx;
7668 ret = inherit_group(event, parent, parent_ctx,
7678 * Initialize the perf_event context in task_struct
7680 int perf_event_init_context(struct task_struct *child, int ctxn)
7682 struct perf_event_context *child_ctx, *parent_ctx;
7683 struct perf_event_context *cloned_ctx;
7684 struct perf_event *event;
7685 struct task_struct *parent = current;
7686 int inherited_all = 1;
7687 unsigned long flags;
7690 if (likely(!parent->perf_event_ctxp[ctxn]))
7694 * If the parent's context is a clone, pin it so it won't get
7697 parent_ctx = perf_pin_task_context(parent, ctxn);
7700 * No need to check if parent_ctx != NULL here; since we saw
7701 * it non-NULL earlier, the only reason for it to become NULL
7702 * is if we exit, and since we're currently in the middle of
7703 * a fork we can't be exiting at the same time.
7707 * Lock the parent list. No need to lock the child - not PID
7708 * hashed yet and not running, so nobody can access it.
7710 mutex_lock(&parent_ctx->mutex);
7713 * We dont have to disable NMIs - we are only looking at
7714 * the list, not manipulating it:
7716 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7717 ret = inherit_task_group(event, parent, parent_ctx,
7718 child, ctxn, &inherited_all);
7724 * We can't hold ctx->lock when iterating the ->flexible_group list due
7725 * to allocations, but we need to prevent rotation because
7726 * rotate_ctx() will change the list from interrupt context.
7728 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7729 parent_ctx->rotate_disable = 1;
7730 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7732 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7733 ret = inherit_task_group(event, parent, parent_ctx,
7734 child, ctxn, &inherited_all);
7739 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7740 parent_ctx->rotate_disable = 0;
7742 child_ctx = child->perf_event_ctxp[ctxn];
7744 if (child_ctx && inherited_all) {
7746 * Mark the child context as a clone of the parent
7747 * context, or of whatever the parent is a clone of.
7749 * Note that if the parent is a clone, the holding of
7750 * parent_ctx->lock avoids it from being uncloned.
7752 cloned_ctx = parent_ctx->parent_ctx;
7754 child_ctx->parent_ctx = cloned_ctx;
7755 child_ctx->parent_gen = parent_ctx->parent_gen;
7757 child_ctx->parent_ctx = parent_ctx;
7758 child_ctx->parent_gen = parent_ctx->generation;
7760 get_ctx(child_ctx->parent_ctx);
7763 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7764 mutex_unlock(&parent_ctx->mutex);
7766 perf_unpin_context(parent_ctx);
7767 put_ctx(parent_ctx);
7773 * Initialize the perf_event context in task_struct
7775 int perf_event_init_task(struct task_struct *child)
7779 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
7780 mutex_init(&child->perf_event_mutex);
7781 INIT_LIST_HEAD(&child->perf_event_list);
7783 for_each_task_context_nr(ctxn) {
7784 ret = perf_event_init_context(child, ctxn);
7792 static void __init perf_event_init_all_cpus(void)
7794 struct swevent_htable *swhash;
7797 for_each_possible_cpu(cpu) {
7798 swhash = &per_cpu(swevent_htable, cpu);
7799 mutex_init(&swhash->hlist_mutex);
7800 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
7804 static void perf_event_init_cpu(int cpu)
7806 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7808 mutex_lock(&swhash->hlist_mutex);
7809 if (swhash->hlist_refcount > 0) {
7810 struct swevent_hlist *hlist;
7812 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
7814 rcu_assign_pointer(swhash->swevent_hlist, hlist);
7816 mutex_unlock(&swhash->hlist_mutex);
7819 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
7820 static void perf_pmu_rotate_stop(struct pmu *pmu)
7822 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
7824 WARN_ON(!irqs_disabled());
7826 list_del_init(&cpuctx->rotation_list);
7829 static void __perf_event_exit_context(void *__info)
7831 struct perf_event_context *ctx = __info;
7832 struct perf_event *event, *tmp;
7834 perf_pmu_rotate_stop(ctx->pmu);
7836 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
7837 __perf_remove_from_context(event);
7838 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
7839 __perf_remove_from_context(event);
7842 static void perf_event_exit_cpu_context(int cpu)
7844 struct perf_event_context *ctx;
7848 idx = srcu_read_lock(&pmus_srcu);
7849 list_for_each_entry_rcu(pmu, &pmus, entry) {
7850 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
7852 mutex_lock(&ctx->mutex);
7853 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
7854 mutex_unlock(&ctx->mutex);
7856 srcu_read_unlock(&pmus_srcu, idx);
7859 static void perf_event_exit_cpu(int cpu)
7861 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
7863 mutex_lock(&swhash->hlist_mutex);
7864 swevent_hlist_release(swhash);
7865 mutex_unlock(&swhash->hlist_mutex);
7867 perf_event_exit_cpu_context(cpu);
7870 static inline void perf_event_exit_cpu(int cpu) { }
7874 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
7878 for_each_online_cpu(cpu)
7879 perf_event_exit_cpu(cpu);
7885 * Run the perf reboot notifier at the very last possible moment so that
7886 * the generic watchdog code runs as long as possible.
7888 static struct notifier_block perf_reboot_notifier = {
7889 .notifier_call = perf_reboot,
7890 .priority = INT_MIN,
7894 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
7896 unsigned int cpu = (long)hcpu;
7898 switch (action & ~CPU_TASKS_FROZEN) {
7900 case CPU_UP_PREPARE:
7901 case CPU_DOWN_FAILED:
7902 perf_event_init_cpu(cpu);
7905 case CPU_UP_CANCELED:
7906 case CPU_DOWN_PREPARE:
7907 perf_event_exit_cpu(cpu);
7916 void __init perf_event_init(void)
7922 perf_event_init_all_cpus();
7923 init_srcu_struct(&pmus_srcu);
7924 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
7925 perf_pmu_register(&perf_cpu_clock, NULL, -1);
7926 perf_pmu_register(&perf_task_clock, NULL, -1);
7928 perf_cpu_notifier(perf_cpu_notify);
7929 register_reboot_notifier(&perf_reboot_notifier);
7931 ret = init_hw_breakpoint();
7932 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
7934 /* do not patch jump label more than once per second */
7935 jump_label_rate_limit(&perf_sched_events, HZ);
7938 * Build time assertion that we keep the data_head at the intended
7939 * location. IOW, validation we got the __reserved[] size right.
7941 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
7945 static int __init perf_event_sysfs_init(void)
7950 mutex_lock(&pmus_lock);
7952 ret = bus_register(&pmu_bus);
7956 list_for_each_entry(pmu, &pmus, entry) {
7957 if (!pmu->name || pmu->type < 0)
7960 ret = pmu_dev_alloc(pmu);
7961 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
7963 pmu_bus_running = 1;
7967 mutex_unlock(&pmus_lock);
7971 device_initcall(perf_event_sysfs_init);
7973 #ifdef CONFIG_CGROUP_PERF
7974 static struct cgroup_subsys_state *
7975 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7977 struct perf_cgroup *jc;
7979 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
7981 return ERR_PTR(-ENOMEM);
7983 jc->info = alloc_percpu(struct perf_cgroup_info);
7986 return ERR_PTR(-ENOMEM);
7992 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
7994 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
7996 free_percpu(jc->info);
8000 static int __perf_cgroup_move(void *info)
8002 struct task_struct *task = info;
8003 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8007 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8008 struct cgroup_taskset *tset)
8010 struct task_struct *task;
8012 cgroup_taskset_for_each(task, css, tset)
8013 task_function_call(task, __perf_cgroup_move, task);
8016 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8017 struct cgroup_subsys_state *old_css,
8018 struct task_struct *task)
8021 * cgroup_exit() is called in the copy_process() failure path.
8022 * Ignore this case since the task hasn't ran yet, this avoids
8023 * trying to poke a half freed task state from generic code.
8025 if (!(task->flags & PF_EXITING))
8028 task_function_call(task, __perf_cgroup_move, task);
8031 struct cgroup_subsys perf_subsys = {
8032 .name = "perf_event",
8033 .subsys_id = perf_subsys_id,
8034 .css_alloc = perf_cgroup_css_alloc,
8035 .css_free = perf_cgroup_css_free,
8036 .exit = perf_cgroup_exit,
8037 .attach = perf_cgroup_attach,
8039 #endif /* CONFIG_CGROUP_PERF */