2 * Performance events core code:
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
7 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 * For licensing details see kernel-base/COPYING
14 #include <linux/cpu.h>
15 #include <linux/smp.h>
16 #include <linux/idr.h>
17 #include <linux/file.h>
18 #include <linux/poll.h>
19 #include <linux/slab.h>
20 #include <linux/hash.h>
21 #include <linux/tick.h>
22 #include <linux/sysfs.h>
23 #include <linux/dcache.h>
24 #include <linux/percpu.h>
25 #include <linux/ptrace.h>
26 #include <linux/reboot.h>
27 #include <linux/vmstat.h>
28 #include <linux/device.h>
29 #include <linux/export.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hardirq.h>
32 #include <linux/rculist.h>
33 #include <linux/uaccess.h>
34 #include <linux/syscalls.h>
35 #include <linux/anon_inodes.h>
36 #include <linux/kernel_stat.h>
37 #include <linux/perf_event.h>
38 #include <linux/ftrace_event.h>
39 #include <linux/hw_breakpoint.h>
40 #include <linux/mm_types.h>
41 #include <linux/cgroup.h>
42 #include <linux/module.h>
43 #include <linux/mman.h>
44 #include <linux/compat.h>
48 #include <asm/irq_regs.h>
50 static struct workqueue_struct *perf_wq;
52 struct remote_function_call {
53 struct task_struct *p;
54 int (*func)(void *info);
59 static void remote_function(void *data)
61 struct remote_function_call *tfc = data;
62 struct task_struct *p = tfc->p;
66 if (task_cpu(p) != smp_processor_id() || !task_curr(p))
70 tfc->ret = tfc->func(tfc->info);
74 * task_function_call - call a function on the cpu on which a task runs
75 * @p: the task to evaluate
76 * @func: the function to be called
77 * @info: the function call argument
79 * Calls the function @func when the task is currently running. This might
80 * be on the current CPU, which just calls the function directly
82 * returns: @func return value, or
83 * -ESRCH - when the process isn't running
84 * -EAGAIN - when the process moved away
87 task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
89 struct remote_function_call data = {
93 .ret = -ESRCH, /* No such (running) process */
97 smp_call_function_single(task_cpu(p), remote_function, &data, 1);
103 * cpu_function_call - call a function on the cpu
104 * @func: the function to be called
105 * @info: the function call argument
107 * Calls the function @func on the remote cpu.
109 * returns: @func return value or -ENXIO when the cpu is offline
111 static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
113 struct remote_function_call data = {
117 .ret = -ENXIO, /* No such CPU */
120 smp_call_function_single(cpu, remote_function, &data, 1);
125 #define EVENT_OWNER_KERNEL ((void *) -1)
127 static bool is_kernel_event(struct perf_event *event)
129 return event->owner == EVENT_OWNER_KERNEL;
132 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
133 PERF_FLAG_FD_OUTPUT |\
134 PERF_FLAG_PID_CGROUP |\
135 PERF_FLAG_FD_CLOEXEC)
138 * branch priv levels that need permission checks
140 #define PERF_SAMPLE_BRANCH_PERM_PLM \
141 (PERF_SAMPLE_BRANCH_KERNEL |\
142 PERF_SAMPLE_BRANCH_HV)
145 EVENT_FLEXIBLE = 0x1,
147 EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
151 * perf_sched_events : >0 events exist
152 * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
154 struct static_key_deferred perf_sched_events __read_mostly;
155 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
156 static DEFINE_PER_CPU(atomic_t, perf_branch_stack_events);
158 static atomic_t nr_mmap_events __read_mostly;
159 static atomic_t nr_comm_events __read_mostly;
160 static atomic_t nr_task_events __read_mostly;
161 static atomic_t nr_freq_events __read_mostly;
163 static LIST_HEAD(pmus);
164 static DEFINE_MUTEX(pmus_lock);
165 static struct srcu_struct pmus_srcu;
168 * perf event paranoia level:
169 * -1 - not paranoid at all
170 * 0 - disallow raw tracepoint access for unpriv
171 * 1 - disallow cpu events for unpriv
172 * 2 - disallow kernel profiling for unpriv
174 int sysctl_perf_event_paranoid __read_mostly = 1;
176 /* Minimum for 512 kiB + 1 user control page */
177 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
180 * max perf event sample rate
182 #define DEFAULT_MAX_SAMPLE_RATE 100000
183 #define DEFAULT_SAMPLE_PERIOD_NS (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
184 #define DEFAULT_CPU_TIME_MAX_PERCENT 25
186 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
188 static int max_samples_per_tick __read_mostly = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
189 static int perf_sample_period_ns __read_mostly = DEFAULT_SAMPLE_PERIOD_NS;
191 static int perf_sample_allowed_ns __read_mostly =
192 DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
194 void update_perf_cpu_limits(void)
196 u64 tmp = perf_sample_period_ns;
198 tmp *= sysctl_perf_cpu_time_max_percent;
200 ACCESS_ONCE(perf_sample_allowed_ns) = tmp;
203 static int perf_rotate_context(struct perf_cpu_context *cpuctx);
205 int perf_proc_update_handler(struct ctl_table *table, int write,
206 void __user *buffer, size_t *lenp,
209 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
214 max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
215 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
216 update_perf_cpu_limits();
221 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
223 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
224 void __user *buffer, size_t *lenp,
227 int ret = proc_dointvec(table, write, buffer, lenp, ppos);
232 update_perf_cpu_limits();
238 * perf samples are done in some very critical code paths (NMIs).
239 * If they take too much CPU time, the system can lock up and not
240 * get any real work done. This will drop the sample rate when
241 * we detect that events are taking too long.
243 #define NR_ACCUMULATED_SAMPLES 128
244 static DEFINE_PER_CPU(u64, running_sample_length);
246 static void perf_duration_warn(struct irq_work *w)
248 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
249 u64 avg_local_sample_len;
250 u64 local_samples_len;
252 local_samples_len = __get_cpu_var(running_sample_length);
253 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
255 printk_ratelimited(KERN_WARNING
256 "perf interrupt took too long (%lld > %lld), lowering "
257 "kernel.perf_event_max_sample_rate to %d\n",
258 avg_local_sample_len, allowed_ns >> 1,
259 sysctl_perf_event_sample_rate);
262 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
264 void perf_sample_event_took(u64 sample_len_ns)
266 u64 allowed_ns = ACCESS_ONCE(perf_sample_allowed_ns);
267 u64 avg_local_sample_len;
268 u64 local_samples_len;
273 /* decay the counter by 1 average sample */
274 local_samples_len = __get_cpu_var(running_sample_length);
275 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
276 local_samples_len += sample_len_ns;
277 __get_cpu_var(running_sample_length) = local_samples_len;
280 * note: this will be biased artifically low until we have
281 * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
282 * from having to maintain a count.
284 avg_local_sample_len = local_samples_len/NR_ACCUMULATED_SAMPLES;
286 if (avg_local_sample_len <= allowed_ns)
289 if (max_samples_per_tick <= 1)
292 max_samples_per_tick = DIV_ROUND_UP(max_samples_per_tick, 2);
293 sysctl_perf_event_sample_rate = max_samples_per_tick * HZ;
294 perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
296 update_perf_cpu_limits();
298 if (!irq_work_queue(&perf_duration_work)) {
299 early_printk("perf interrupt took too long (%lld > %lld), lowering "
300 "kernel.perf_event_max_sample_rate to %d\n",
301 avg_local_sample_len, allowed_ns >> 1,
302 sysctl_perf_event_sample_rate);
306 static atomic64_t perf_event_id;
308 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
309 enum event_type_t event_type);
311 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
312 enum event_type_t event_type,
313 struct task_struct *task);
315 static void update_context_time(struct perf_event_context *ctx);
316 static u64 perf_event_time(struct perf_event *event);
318 void __weak perf_event_print_debug(void) { }
320 extern __weak const char *perf_pmu_name(void)
325 static inline u64 perf_clock(void)
327 return local_clock();
330 static inline struct perf_cpu_context *
331 __get_cpu_context(struct perf_event_context *ctx)
333 return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
336 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
337 struct perf_event_context *ctx)
339 raw_spin_lock(&cpuctx->ctx.lock);
341 raw_spin_lock(&ctx->lock);
344 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
345 struct perf_event_context *ctx)
348 raw_spin_unlock(&ctx->lock);
349 raw_spin_unlock(&cpuctx->ctx.lock);
352 #ifdef CONFIG_CGROUP_PERF
355 * perf_cgroup_info keeps track of time_enabled for a cgroup.
356 * This is a per-cpu dynamically allocated data structure.
358 struct perf_cgroup_info {
364 struct cgroup_subsys_state css;
365 struct perf_cgroup_info __percpu *info;
369 * Must ensure cgroup is pinned (css_get) before calling
370 * this function. In other words, we cannot call this function
371 * if there is no cgroup event for the current CPU context.
373 static inline struct perf_cgroup *
374 perf_cgroup_from_task(struct task_struct *task)
376 return container_of(task_css(task, perf_event_cgrp_id),
377 struct perf_cgroup, css);
381 perf_cgroup_match(struct perf_event *event)
383 struct perf_event_context *ctx = event->ctx;
384 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
386 /* @event doesn't care about cgroup */
390 /* wants specific cgroup scope but @cpuctx isn't associated with any */
395 * Cgroup scoping is recursive. An event enabled for a cgroup is
396 * also enabled for all its descendant cgroups. If @cpuctx's
397 * cgroup is a descendant of @event's (the test covers identity
398 * case), it's a match.
400 return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
401 event->cgrp->css.cgroup);
404 static inline void perf_put_cgroup(struct perf_event *event)
406 css_put(&event->cgrp->css);
409 static inline void perf_detach_cgroup(struct perf_event *event)
411 perf_put_cgroup(event);
415 static inline int is_cgroup_event(struct perf_event *event)
417 return event->cgrp != NULL;
420 static inline u64 perf_cgroup_event_time(struct perf_event *event)
422 struct perf_cgroup_info *t;
424 t = per_cpu_ptr(event->cgrp->info, event->cpu);
428 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
430 struct perf_cgroup_info *info;
435 info = this_cpu_ptr(cgrp->info);
437 info->time += now - info->timestamp;
438 info->timestamp = now;
441 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
443 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
445 __update_cgrp_time(cgrp_out);
448 static inline void update_cgrp_time_from_event(struct perf_event *event)
450 struct perf_cgroup *cgrp;
453 * ensure we access cgroup data only when needed and
454 * when we know the cgroup is pinned (css_get)
456 if (!is_cgroup_event(event))
459 cgrp = perf_cgroup_from_task(current);
461 * Do not update time when cgroup is not active
463 if (cgrp == event->cgrp)
464 __update_cgrp_time(event->cgrp);
468 perf_cgroup_set_timestamp(struct task_struct *task,
469 struct perf_event_context *ctx)
471 struct perf_cgroup *cgrp;
472 struct perf_cgroup_info *info;
475 * ctx->lock held by caller
476 * ensure we do not access cgroup data
477 * unless we have the cgroup pinned (css_get)
479 if (!task || !ctx->nr_cgroups)
482 cgrp = perf_cgroup_from_task(task);
483 info = this_cpu_ptr(cgrp->info);
484 info->timestamp = ctx->timestamp;
487 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
488 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
491 * reschedule events based on the cgroup constraint of task.
493 * mode SWOUT : schedule out everything
494 * mode SWIN : schedule in based on cgroup for next
496 void perf_cgroup_switch(struct task_struct *task, int mode)
498 struct perf_cpu_context *cpuctx;
503 * disable interrupts to avoid geting nr_cgroup
504 * changes via __perf_event_disable(). Also
507 local_irq_save(flags);
510 * we reschedule only in the presence of cgroup
511 * constrained events.
515 list_for_each_entry_rcu(pmu, &pmus, entry) {
516 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
517 if (cpuctx->unique_pmu != pmu)
518 continue; /* ensure we process each cpuctx once */
521 * perf_cgroup_events says at least one
522 * context on this CPU has cgroup events.
524 * ctx->nr_cgroups reports the number of cgroup
525 * events for a context.
527 if (cpuctx->ctx.nr_cgroups > 0) {
528 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
529 perf_pmu_disable(cpuctx->ctx.pmu);
531 if (mode & PERF_CGROUP_SWOUT) {
532 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
534 * must not be done before ctxswout due
535 * to event_filter_match() in event_sched_out()
540 if (mode & PERF_CGROUP_SWIN) {
541 WARN_ON_ONCE(cpuctx->cgrp);
543 * set cgrp before ctxsw in to allow
544 * event_filter_match() to not have to pass
547 cpuctx->cgrp = perf_cgroup_from_task(task);
548 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
550 perf_pmu_enable(cpuctx->ctx.pmu);
551 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
557 local_irq_restore(flags);
560 static inline void perf_cgroup_sched_out(struct task_struct *task,
561 struct task_struct *next)
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);
572 * next is NULL when called from perf_event_enable_on_exec()
573 * that will systematically cause a cgroup_switch()
576 cgrp2 = perf_cgroup_from_task(next);
579 * only schedule out current cgroup events if we know
580 * that we are switching to a different cgroup. Otherwise,
581 * do no touch the cgroup events.
584 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
587 static inline void perf_cgroup_sched_in(struct task_struct *prev,
588 struct task_struct *task)
590 struct perf_cgroup *cgrp1;
591 struct perf_cgroup *cgrp2 = NULL;
594 * we come here when we know perf_cgroup_events > 0
596 cgrp1 = perf_cgroup_from_task(task);
598 /* prev can never be NULL */
599 cgrp2 = perf_cgroup_from_task(prev);
602 * only need to schedule in cgroup events if we are changing
603 * cgroup during ctxsw. Cgroup events were not scheduled
604 * out of ctxsw out if that was not the case.
607 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
610 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
611 struct perf_event_attr *attr,
612 struct perf_event *group_leader)
614 struct perf_cgroup *cgrp;
615 struct cgroup_subsys_state *css;
616 struct fd f = fdget(fd);
622 css = css_tryget_online_from_dir(f.file->f_dentry,
623 &perf_event_cgrp_subsys);
629 cgrp = container_of(css, struct perf_cgroup, css);
633 * all events in a group must monitor
634 * the same cgroup because a task belongs
635 * to only one perf cgroup at a time
637 if (group_leader && group_leader->cgrp != cgrp) {
638 perf_detach_cgroup(event);
647 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
649 struct perf_cgroup_info *t;
650 t = per_cpu_ptr(event->cgrp->info, event->cpu);
651 event->shadow_ctx_time = now - t->timestamp;
655 perf_cgroup_defer_enabled(struct perf_event *event)
658 * when the current task's perf cgroup does not match
659 * the event's, we need to remember to call the
660 * perf_mark_enable() function the first time a task with
661 * a matching perf cgroup is scheduled in.
663 if (is_cgroup_event(event) && !perf_cgroup_match(event))
664 event->cgrp_defer_enabled = 1;
668 perf_cgroup_mark_enabled(struct perf_event *event,
669 struct perf_event_context *ctx)
671 struct perf_event *sub;
672 u64 tstamp = perf_event_time(event);
674 if (!event->cgrp_defer_enabled)
677 event->cgrp_defer_enabled = 0;
679 event->tstamp_enabled = tstamp - event->total_time_enabled;
680 list_for_each_entry(sub, &event->sibling_list, group_entry) {
681 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
682 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
683 sub->cgrp_defer_enabled = 0;
687 #else /* !CONFIG_CGROUP_PERF */
690 perf_cgroup_match(struct perf_event *event)
695 static inline void perf_detach_cgroup(struct perf_event *event)
698 static inline int is_cgroup_event(struct perf_event *event)
703 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
708 static inline void update_cgrp_time_from_event(struct perf_event *event)
712 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
716 static inline void perf_cgroup_sched_out(struct task_struct *task,
717 struct task_struct *next)
721 static inline void perf_cgroup_sched_in(struct task_struct *prev,
722 struct task_struct *task)
726 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
727 struct perf_event_attr *attr,
728 struct perf_event *group_leader)
734 perf_cgroup_set_timestamp(struct task_struct *task,
735 struct perf_event_context *ctx)
740 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
745 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
749 static inline u64 perf_cgroup_event_time(struct perf_event *event)
755 perf_cgroup_defer_enabled(struct perf_event *event)
760 perf_cgroup_mark_enabled(struct perf_event *event,
761 struct perf_event_context *ctx)
767 * set default to be dependent on timer tick just
770 #define PERF_CPU_HRTIMER (1000 / HZ)
772 * function must be called with interrupts disbled
774 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
776 struct perf_cpu_context *cpuctx;
777 enum hrtimer_restart ret = HRTIMER_NORESTART;
780 WARN_ON(!irqs_disabled());
782 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
784 rotations = perf_rotate_context(cpuctx);
787 * arm timer if needed
790 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
791 ret = HRTIMER_RESTART;
797 /* CPU is going down */
798 void perf_cpu_hrtimer_cancel(int cpu)
800 struct perf_cpu_context *cpuctx;
804 if (WARN_ON(cpu != smp_processor_id()))
807 local_irq_save(flags);
811 list_for_each_entry_rcu(pmu, &pmus, entry) {
812 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
814 if (pmu->task_ctx_nr == perf_sw_context)
817 hrtimer_cancel(&cpuctx->hrtimer);
822 local_irq_restore(flags);
825 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
827 struct hrtimer *hr = &cpuctx->hrtimer;
828 struct pmu *pmu = cpuctx->ctx.pmu;
831 /* no multiplexing needed for SW PMU */
832 if (pmu->task_ctx_nr == perf_sw_context)
836 * check default is sane, if not set then force to
837 * default interval (1/tick)
839 timer = pmu->hrtimer_interval_ms;
841 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
843 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
845 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
846 hr->function = perf_cpu_hrtimer_handler;
849 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
851 struct hrtimer *hr = &cpuctx->hrtimer;
852 struct pmu *pmu = cpuctx->ctx.pmu;
855 if (pmu->task_ctx_nr == perf_sw_context)
858 if (hrtimer_active(hr))
861 if (!hrtimer_callback_running(hr))
862 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
863 0, HRTIMER_MODE_REL_PINNED, 0);
866 void perf_pmu_disable(struct pmu *pmu)
868 int *count = this_cpu_ptr(pmu->pmu_disable_count);
870 pmu->pmu_disable(pmu);
873 void perf_pmu_enable(struct pmu *pmu)
875 int *count = this_cpu_ptr(pmu->pmu_disable_count);
877 pmu->pmu_enable(pmu);
880 static DEFINE_PER_CPU(struct list_head, rotation_list);
883 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
884 * because they're strictly cpu affine and rotate_start is called with IRQs
885 * disabled, while rotate_context is called from IRQ context.
887 static void perf_pmu_rotate_start(struct pmu *pmu)
889 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
890 struct list_head *head = &__get_cpu_var(rotation_list);
892 WARN_ON(!irqs_disabled());
894 if (list_empty(&cpuctx->rotation_list))
895 list_add(&cpuctx->rotation_list, head);
898 static void get_ctx(struct perf_event_context *ctx)
900 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
903 static void put_ctx(struct perf_event_context *ctx)
905 if (atomic_dec_and_test(&ctx->refcount)) {
907 put_ctx(ctx->parent_ctx);
909 put_task_struct(ctx->task);
910 kfree_rcu(ctx, rcu_head);
914 static void unclone_ctx(struct perf_event_context *ctx)
916 if (ctx->parent_ctx) {
917 put_ctx(ctx->parent_ctx);
918 ctx->parent_ctx = NULL;
923 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
926 * only top level events have the pid namespace they were created in
929 event = event->parent;
931 return task_tgid_nr_ns(p, event->ns);
934 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
937 * only top level events have the pid namespace they were created in
940 event = event->parent;
942 return task_pid_nr_ns(p, event->ns);
946 * If we inherit events we want to return the parent event id
949 static u64 primary_event_id(struct perf_event *event)
954 id = event->parent->id;
960 * Get the perf_event_context for a task and lock it.
961 * This has to cope with with the fact that until it is locked,
962 * the context could get moved to another task.
964 static struct perf_event_context *
965 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
967 struct perf_event_context *ctx;
971 * One of the few rules of preemptible RCU is that one cannot do
972 * rcu_read_unlock() while holding a scheduler (or nested) lock when
973 * part of the read side critical section was preemptible -- see
974 * rcu_read_unlock_special().
976 * Since ctx->lock nests under rq->lock we must ensure the entire read
977 * side critical section is non-preemptible.
981 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
984 * If this context is a clone of another, it might
985 * get swapped for another underneath us by
986 * perf_event_task_sched_out, though the
987 * rcu_read_lock() protects us from any context
988 * getting freed. Lock the context and check if it
989 * got swapped before we could get the lock, and retry
990 * if so. If we locked the right context, then it
991 * can't get swapped on us any more.
993 raw_spin_lock_irqsave(&ctx->lock, *flags);
994 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
995 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1001 if (!atomic_inc_not_zero(&ctx->refcount)) {
1002 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1012 * Get the context for a task and increment its pin_count so it
1013 * can't get swapped to another task. This also increments its
1014 * reference count so that the context can't get freed.
1016 static struct perf_event_context *
1017 perf_pin_task_context(struct task_struct *task, int ctxn)
1019 struct perf_event_context *ctx;
1020 unsigned long flags;
1022 ctx = perf_lock_task_context(task, ctxn, &flags);
1025 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1030 static void perf_unpin_context(struct perf_event_context *ctx)
1032 unsigned long flags;
1034 raw_spin_lock_irqsave(&ctx->lock, flags);
1036 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1040 * Update the record of the current time in a context.
1042 static void update_context_time(struct perf_event_context *ctx)
1044 u64 now = perf_clock();
1046 ctx->time += now - ctx->timestamp;
1047 ctx->timestamp = now;
1050 static u64 perf_event_time(struct perf_event *event)
1052 struct perf_event_context *ctx = event->ctx;
1054 if (is_cgroup_event(event))
1055 return perf_cgroup_event_time(event);
1057 return ctx ? ctx->time : 0;
1061 * Update the total_time_enabled and total_time_running fields for a event.
1062 * The caller of this function needs to hold the ctx->lock.
1064 static void update_event_times(struct perf_event *event)
1066 struct perf_event_context *ctx = event->ctx;
1069 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1070 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1073 * in cgroup mode, time_enabled represents
1074 * the time the event was enabled AND active
1075 * tasks were in the monitored cgroup. This is
1076 * independent of the activity of the context as
1077 * there may be a mix of cgroup and non-cgroup events.
1079 * That is why we treat cgroup events differently
1082 if (is_cgroup_event(event))
1083 run_end = perf_cgroup_event_time(event);
1084 else if (ctx->is_active)
1085 run_end = ctx->time;
1087 run_end = event->tstamp_stopped;
1089 event->total_time_enabled = run_end - event->tstamp_enabled;
1091 if (event->state == PERF_EVENT_STATE_INACTIVE)
1092 run_end = event->tstamp_stopped;
1094 run_end = perf_event_time(event);
1096 event->total_time_running = run_end - event->tstamp_running;
1101 * Update total_time_enabled and total_time_running for all events in a group.
1103 static void update_group_times(struct perf_event *leader)
1105 struct perf_event *event;
1107 update_event_times(leader);
1108 list_for_each_entry(event, &leader->sibling_list, group_entry)
1109 update_event_times(event);
1112 static struct list_head *
1113 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1115 if (event->attr.pinned)
1116 return &ctx->pinned_groups;
1118 return &ctx->flexible_groups;
1122 * Add a event from the lists for its context.
1123 * Must be called with ctx->mutex and ctx->lock held.
1126 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1128 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1129 event->attach_state |= PERF_ATTACH_CONTEXT;
1132 * If we're a stand alone event or group leader, we go to the context
1133 * list, group events are kept attached to the group so that
1134 * perf_group_detach can, at all times, locate all siblings.
1136 if (event->group_leader == event) {
1137 struct list_head *list;
1139 if (is_software_event(event))
1140 event->group_flags |= PERF_GROUP_SOFTWARE;
1142 list = ctx_group_list(event, ctx);
1143 list_add_tail(&event->group_entry, list);
1146 if (is_cgroup_event(event))
1149 if (has_branch_stack(event))
1150 ctx->nr_branch_stack++;
1152 list_add_rcu(&event->event_entry, &ctx->event_list);
1153 if (!ctx->nr_events)
1154 perf_pmu_rotate_start(ctx->pmu);
1156 if (event->attr.inherit_stat)
1163 * Initialize event state based on the perf_event_attr::disabled.
1165 static inline void perf_event__state_init(struct perf_event *event)
1167 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1168 PERF_EVENT_STATE_INACTIVE;
1172 * Called at perf_event creation and when events are attached/detached from a
1175 static void perf_event__read_size(struct perf_event *event)
1177 int entry = sizeof(u64); /* value */
1181 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1182 size += sizeof(u64);
1184 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1185 size += sizeof(u64);
1187 if (event->attr.read_format & PERF_FORMAT_ID)
1188 entry += sizeof(u64);
1190 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1191 nr += event->group_leader->nr_siblings;
1192 size += sizeof(u64);
1196 event->read_size = size;
1199 static void perf_event__header_size(struct perf_event *event)
1201 struct perf_sample_data *data;
1202 u64 sample_type = event->attr.sample_type;
1205 perf_event__read_size(event);
1207 if (sample_type & PERF_SAMPLE_IP)
1208 size += sizeof(data->ip);
1210 if (sample_type & PERF_SAMPLE_ADDR)
1211 size += sizeof(data->addr);
1213 if (sample_type & PERF_SAMPLE_PERIOD)
1214 size += sizeof(data->period);
1216 if (sample_type & PERF_SAMPLE_WEIGHT)
1217 size += sizeof(data->weight);
1219 if (sample_type & PERF_SAMPLE_READ)
1220 size += event->read_size;
1222 if (sample_type & PERF_SAMPLE_DATA_SRC)
1223 size += sizeof(data->data_src.val);
1225 if (sample_type & PERF_SAMPLE_TRANSACTION)
1226 size += sizeof(data->txn);
1228 event->header_size = size;
1231 static void perf_event__id_header_size(struct perf_event *event)
1233 struct perf_sample_data *data;
1234 u64 sample_type = event->attr.sample_type;
1237 if (sample_type & PERF_SAMPLE_TID)
1238 size += sizeof(data->tid_entry);
1240 if (sample_type & PERF_SAMPLE_TIME)
1241 size += sizeof(data->time);
1243 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1244 size += sizeof(data->id);
1246 if (sample_type & PERF_SAMPLE_ID)
1247 size += sizeof(data->id);
1249 if (sample_type & PERF_SAMPLE_STREAM_ID)
1250 size += sizeof(data->stream_id);
1252 if (sample_type & PERF_SAMPLE_CPU)
1253 size += sizeof(data->cpu_entry);
1255 event->id_header_size = size;
1258 static void perf_group_attach(struct perf_event *event)
1260 struct perf_event *group_leader = event->group_leader, *pos;
1263 * We can have double attach due to group movement in perf_event_open.
1265 if (event->attach_state & PERF_ATTACH_GROUP)
1268 event->attach_state |= PERF_ATTACH_GROUP;
1270 if (group_leader == event)
1273 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1274 !is_software_event(event))
1275 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1277 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1278 group_leader->nr_siblings++;
1280 perf_event__header_size(group_leader);
1282 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1283 perf_event__header_size(pos);
1287 * Remove a event from the lists for its context.
1288 * Must be called with ctx->mutex and ctx->lock held.
1291 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1293 struct perf_cpu_context *cpuctx;
1295 * We can have double detach due to exit/hot-unplug + close.
1297 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1300 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1302 if (is_cgroup_event(event)) {
1304 cpuctx = __get_cpu_context(ctx);
1306 * if there are no more cgroup events
1307 * then cler cgrp to avoid stale pointer
1308 * in update_cgrp_time_from_cpuctx()
1310 if (!ctx->nr_cgroups)
1311 cpuctx->cgrp = NULL;
1314 if (has_branch_stack(event))
1315 ctx->nr_branch_stack--;
1318 if (event->attr.inherit_stat)
1321 list_del_rcu(&event->event_entry);
1323 if (event->group_leader == event)
1324 list_del_init(&event->group_entry);
1326 update_group_times(event);
1329 * If event was in error state, then keep it
1330 * that way, otherwise bogus counts will be
1331 * returned on read(). The only way to get out
1332 * of error state is by explicit re-enabling
1335 if (event->state > PERF_EVENT_STATE_OFF)
1336 event->state = PERF_EVENT_STATE_OFF;
1341 static void perf_group_detach(struct perf_event *event)
1343 struct perf_event *sibling, *tmp;
1344 struct list_head *list = NULL;
1347 * We can have double detach due to exit/hot-unplug + close.
1349 if (!(event->attach_state & PERF_ATTACH_GROUP))
1352 event->attach_state &= ~PERF_ATTACH_GROUP;
1355 * If this is a sibling, remove it from its group.
1357 if (event->group_leader != event) {
1358 list_del_init(&event->group_entry);
1359 event->group_leader->nr_siblings--;
1363 if (!list_empty(&event->group_entry))
1364 list = &event->group_entry;
1367 * If this was a group event with sibling events then
1368 * upgrade the siblings to singleton events by adding them
1369 * to whatever list we are on.
1371 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1373 list_move_tail(&sibling->group_entry, list);
1374 sibling->group_leader = sibling;
1376 /* Inherit group flags from the previous leader */
1377 sibling->group_flags = event->group_flags;
1381 perf_event__header_size(event->group_leader);
1383 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1384 perf_event__header_size(tmp);
1388 * User event without the task.
1390 static bool is_orphaned_event(struct perf_event *event)
1392 return event && !is_kernel_event(event) && !event->owner;
1396 * Event has a parent but parent's task finished and it's
1397 * alive only because of children holding refference.
1399 static bool is_orphaned_child(struct perf_event *event)
1401 return is_orphaned_event(event->parent);
1404 static void orphans_remove_work(struct work_struct *work);
1406 static void schedule_orphans_remove(struct perf_event_context *ctx)
1408 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1411 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1413 ctx->orphans_remove_sched = true;
1417 static int __init perf_workqueue_init(void)
1419 perf_wq = create_singlethread_workqueue("perf");
1420 WARN(!perf_wq, "failed to create perf workqueue\n");
1421 return perf_wq ? 0 : -1;
1424 core_initcall(perf_workqueue_init);
1427 event_filter_match(struct perf_event *event)
1429 return (event->cpu == -1 || event->cpu == smp_processor_id())
1430 && perf_cgroup_match(event);
1434 event_sched_out(struct perf_event *event,
1435 struct perf_cpu_context *cpuctx,
1436 struct perf_event_context *ctx)
1438 u64 tstamp = perf_event_time(event);
1441 * An event which could not be activated because of
1442 * filter mismatch still needs to have its timings
1443 * maintained, otherwise bogus information is return
1444 * via read() for time_enabled, time_running:
1446 if (event->state == PERF_EVENT_STATE_INACTIVE
1447 && !event_filter_match(event)) {
1448 delta = tstamp - event->tstamp_stopped;
1449 event->tstamp_running += delta;
1450 event->tstamp_stopped = tstamp;
1453 if (event->state != PERF_EVENT_STATE_ACTIVE)
1456 perf_pmu_disable(event->pmu);
1458 event->state = PERF_EVENT_STATE_INACTIVE;
1459 if (event->pending_disable) {
1460 event->pending_disable = 0;
1461 event->state = PERF_EVENT_STATE_OFF;
1463 event->tstamp_stopped = tstamp;
1464 event->pmu->del(event, 0);
1467 if (!is_software_event(event))
1468 cpuctx->active_oncpu--;
1470 if (event->attr.freq && event->attr.sample_freq)
1472 if (event->attr.exclusive || !cpuctx->active_oncpu)
1473 cpuctx->exclusive = 0;
1475 if (is_orphaned_child(event))
1476 schedule_orphans_remove(ctx);
1478 perf_pmu_enable(event->pmu);
1482 group_sched_out(struct perf_event *group_event,
1483 struct perf_cpu_context *cpuctx,
1484 struct perf_event_context *ctx)
1486 struct perf_event *event;
1487 int state = group_event->state;
1489 event_sched_out(group_event, cpuctx, ctx);
1492 * Schedule out siblings (if any):
1494 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1495 event_sched_out(event, cpuctx, ctx);
1497 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1498 cpuctx->exclusive = 0;
1501 struct remove_event {
1502 struct perf_event *event;
1507 * Cross CPU call to remove a performance event
1509 * We disable the event on the hardware level first. After that we
1510 * remove it from the context list.
1512 static int __perf_remove_from_context(void *info)
1514 struct remove_event *re = info;
1515 struct perf_event *event = re->event;
1516 struct perf_event_context *ctx = event->ctx;
1517 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1519 raw_spin_lock(&ctx->lock);
1520 event_sched_out(event, cpuctx, ctx);
1521 if (re->detach_group)
1522 perf_group_detach(event);
1523 list_del_event(event, ctx);
1524 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1526 cpuctx->task_ctx = NULL;
1528 raw_spin_unlock(&ctx->lock);
1535 * Remove the event from a task's (or a CPU's) list of events.
1537 * CPU events are removed with a smp call. For task events we only
1538 * call when the task is on a CPU.
1540 * If event->ctx is a cloned context, callers must make sure that
1541 * every task struct that event->ctx->task could possibly point to
1542 * remains valid. This is OK when called from perf_release since
1543 * that only calls us on the top-level context, which can't be a clone.
1544 * When called from perf_event_exit_task, it's OK because the
1545 * context has been detached from its task.
1547 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1549 struct perf_event_context *ctx = event->ctx;
1550 struct task_struct *task = ctx->task;
1551 struct remove_event re = {
1553 .detach_group = detach_group,
1556 lockdep_assert_held(&ctx->mutex);
1560 * Per cpu events are removed via an smp call and
1561 * the removal is always successful.
1563 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1568 if (!task_function_call(task, __perf_remove_from_context, &re))
1571 raw_spin_lock_irq(&ctx->lock);
1573 * If we failed to find a running task, but find the context active now
1574 * that we've acquired the ctx->lock, retry.
1576 if (ctx->is_active) {
1577 raw_spin_unlock_irq(&ctx->lock);
1582 * Since the task isn't running, its safe to remove the event, us
1583 * holding the ctx->lock ensures the task won't get scheduled in.
1586 perf_group_detach(event);
1587 list_del_event(event, ctx);
1588 raw_spin_unlock_irq(&ctx->lock);
1592 * Cross CPU call to disable a performance event
1594 int __perf_event_disable(void *info)
1596 struct perf_event *event = info;
1597 struct perf_event_context *ctx = event->ctx;
1598 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1601 * If this is a per-task event, need to check whether this
1602 * event's task is the current task on this cpu.
1604 * Can trigger due to concurrent perf_event_context_sched_out()
1605 * flipping contexts around.
1607 if (ctx->task && cpuctx->task_ctx != ctx)
1610 raw_spin_lock(&ctx->lock);
1613 * If the event is on, turn it off.
1614 * If it is in error state, leave it in error state.
1616 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1617 update_context_time(ctx);
1618 update_cgrp_time_from_event(event);
1619 update_group_times(event);
1620 if (event == event->group_leader)
1621 group_sched_out(event, cpuctx, ctx);
1623 event_sched_out(event, cpuctx, ctx);
1624 event->state = PERF_EVENT_STATE_OFF;
1627 raw_spin_unlock(&ctx->lock);
1635 * If event->ctx is a cloned context, callers must make sure that
1636 * every task struct that event->ctx->task could possibly point to
1637 * remains valid. This condition is satisifed when called through
1638 * perf_event_for_each_child or perf_event_for_each because they
1639 * hold the top-level event's child_mutex, so any descendant that
1640 * goes to exit will block in sync_child_event.
1641 * When called from perf_pending_event it's OK because event->ctx
1642 * is the current context on this CPU and preemption is disabled,
1643 * hence we can't get into perf_event_task_sched_out for this context.
1645 void perf_event_disable(struct perf_event *event)
1647 struct perf_event_context *ctx = event->ctx;
1648 struct task_struct *task = ctx->task;
1652 * Disable the event on the cpu that it's on
1654 cpu_function_call(event->cpu, __perf_event_disable, event);
1659 if (!task_function_call(task, __perf_event_disable, event))
1662 raw_spin_lock_irq(&ctx->lock);
1664 * If the event is still active, we need to retry the cross-call.
1666 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1667 raw_spin_unlock_irq(&ctx->lock);
1669 * Reload the task pointer, it might have been changed by
1670 * a concurrent perf_event_context_sched_out().
1677 * Since we have the lock this context can't be scheduled
1678 * in, so we can change the state safely.
1680 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1681 update_group_times(event);
1682 event->state = PERF_EVENT_STATE_OFF;
1684 raw_spin_unlock_irq(&ctx->lock);
1686 EXPORT_SYMBOL_GPL(perf_event_disable);
1688 static void perf_set_shadow_time(struct perf_event *event,
1689 struct perf_event_context *ctx,
1693 * use the correct time source for the time snapshot
1695 * We could get by without this by leveraging the
1696 * fact that to get to this function, the caller
1697 * has most likely already called update_context_time()
1698 * and update_cgrp_time_xx() and thus both timestamp
1699 * are identical (or very close). Given that tstamp is,
1700 * already adjusted for cgroup, we could say that:
1701 * tstamp - ctx->timestamp
1703 * tstamp - cgrp->timestamp.
1705 * Then, in perf_output_read(), the calculation would
1706 * work with no changes because:
1707 * - event is guaranteed scheduled in
1708 * - no scheduled out in between
1709 * - thus the timestamp would be the same
1711 * But this is a bit hairy.
1713 * So instead, we have an explicit cgroup call to remain
1714 * within the time time source all along. We believe it
1715 * is cleaner and simpler to understand.
1717 if (is_cgroup_event(event))
1718 perf_cgroup_set_shadow_time(event, tstamp);
1720 event->shadow_ctx_time = tstamp - ctx->timestamp;
1723 #define MAX_INTERRUPTS (~0ULL)
1725 static void perf_log_throttle(struct perf_event *event, int enable);
1728 event_sched_in(struct perf_event *event,
1729 struct perf_cpu_context *cpuctx,
1730 struct perf_event_context *ctx)
1732 u64 tstamp = perf_event_time(event);
1735 lockdep_assert_held(&ctx->lock);
1737 if (event->state <= PERF_EVENT_STATE_OFF)
1740 event->state = PERF_EVENT_STATE_ACTIVE;
1741 event->oncpu = smp_processor_id();
1744 * Unthrottle events, since we scheduled we might have missed several
1745 * ticks already, also for a heavily scheduling task there is little
1746 * guarantee it'll get a tick in a timely manner.
1748 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1749 perf_log_throttle(event, 1);
1750 event->hw.interrupts = 0;
1754 * The new state must be visible before we turn it on in the hardware:
1758 perf_pmu_disable(event->pmu);
1760 if (event->pmu->add(event, PERF_EF_START)) {
1761 event->state = PERF_EVENT_STATE_INACTIVE;
1767 event->tstamp_running += tstamp - event->tstamp_stopped;
1769 perf_set_shadow_time(event, ctx, tstamp);
1771 if (!is_software_event(event))
1772 cpuctx->active_oncpu++;
1774 if (event->attr.freq && event->attr.sample_freq)
1777 if (event->attr.exclusive)
1778 cpuctx->exclusive = 1;
1780 if (is_orphaned_child(event))
1781 schedule_orphans_remove(ctx);
1784 perf_pmu_enable(event->pmu);
1790 group_sched_in(struct perf_event *group_event,
1791 struct perf_cpu_context *cpuctx,
1792 struct perf_event_context *ctx)
1794 struct perf_event *event, *partial_group = NULL;
1795 struct pmu *pmu = ctx->pmu;
1796 u64 now = ctx->time;
1797 bool simulate = false;
1799 if (group_event->state == PERF_EVENT_STATE_OFF)
1802 pmu->start_txn(pmu);
1804 if (event_sched_in(group_event, cpuctx, ctx)) {
1805 pmu->cancel_txn(pmu);
1806 perf_cpu_hrtimer_restart(cpuctx);
1811 * Schedule in siblings as one group (if any):
1813 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1814 if (event_sched_in(event, cpuctx, ctx)) {
1815 partial_group = event;
1820 if (!pmu->commit_txn(pmu))
1825 * Groups can be scheduled in as one unit only, so undo any
1826 * partial group before returning:
1827 * The events up to the failed event are scheduled out normally,
1828 * tstamp_stopped will be updated.
1830 * The failed events and the remaining siblings need to have
1831 * their timings updated as if they had gone thru event_sched_in()
1832 * and event_sched_out(). This is required to get consistent timings
1833 * across the group. This also takes care of the case where the group
1834 * could never be scheduled by ensuring tstamp_stopped is set to mark
1835 * the time the event was actually stopped, such that time delta
1836 * calculation in update_event_times() is correct.
1838 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1839 if (event == partial_group)
1843 event->tstamp_running += now - event->tstamp_stopped;
1844 event->tstamp_stopped = now;
1846 event_sched_out(event, cpuctx, ctx);
1849 event_sched_out(group_event, cpuctx, ctx);
1851 pmu->cancel_txn(pmu);
1853 perf_cpu_hrtimer_restart(cpuctx);
1859 * Work out whether we can put this event group on the CPU now.
1861 static int group_can_go_on(struct perf_event *event,
1862 struct perf_cpu_context *cpuctx,
1866 * Groups consisting entirely of software events can always go on.
1868 if (event->group_flags & PERF_GROUP_SOFTWARE)
1871 * If an exclusive group is already on, no other hardware
1874 if (cpuctx->exclusive)
1877 * If this group is exclusive and there are already
1878 * events on the CPU, it can't go on.
1880 if (event->attr.exclusive && cpuctx->active_oncpu)
1883 * Otherwise, try to add it if all previous groups were able
1889 static void add_event_to_ctx(struct perf_event *event,
1890 struct perf_event_context *ctx)
1892 u64 tstamp = perf_event_time(event);
1894 list_add_event(event, ctx);
1895 perf_group_attach(event);
1896 event->tstamp_enabled = tstamp;
1897 event->tstamp_running = tstamp;
1898 event->tstamp_stopped = tstamp;
1901 static void task_ctx_sched_out(struct perf_event_context *ctx);
1903 ctx_sched_in(struct perf_event_context *ctx,
1904 struct perf_cpu_context *cpuctx,
1905 enum event_type_t event_type,
1906 struct task_struct *task);
1908 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
1909 struct perf_event_context *ctx,
1910 struct task_struct *task)
1912 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
1914 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
1915 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
1917 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
1921 * Cross CPU call to install and enable a performance event
1923 * Must be called with ctx->mutex held
1925 static int __perf_install_in_context(void *info)
1927 struct perf_event *event = info;
1928 struct perf_event_context *ctx = event->ctx;
1929 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1930 struct perf_event_context *task_ctx = cpuctx->task_ctx;
1931 struct task_struct *task = current;
1933 perf_ctx_lock(cpuctx, task_ctx);
1934 perf_pmu_disable(cpuctx->ctx.pmu);
1937 * If there was an active task_ctx schedule it out.
1940 task_ctx_sched_out(task_ctx);
1943 * If the context we're installing events in is not the
1944 * active task_ctx, flip them.
1946 if (ctx->task && task_ctx != ctx) {
1948 raw_spin_unlock(&task_ctx->lock);
1949 raw_spin_lock(&ctx->lock);
1954 cpuctx->task_ctx = task_ctx;
1955 task = task_ctx->task;
1958 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
1960 update_context_time(ctx);
1962 * update cgrp time only if current cgrp
1963 * matches event->cgrp. Must be done before
1964 * calling add_event_to_ctx()
1966 update_cgrp_time_from_event(event);
1968 add_event_to_ctx(event, ctx);
1971 * Schedule everything back in
1973 perf_event_sched_in(cpuctx, task_ctx, task);
1975 perf_pmu_enable(cpuctx->ctx.pmu);
1976 perf_ctx_unlock(cpuctx, task_ctx);
1982 * Attach a performance event to a context
1984 * First we add the event to the list with the hardware enable bit
1985 * in event->hw_config cleared.
1987 * If the event is attached to a task which is on a CPU we use a smp
1988 * call to enable it in the task context. The task might have been
1989 * scheduled away, but we check this in the smp call again.
1992 perf_install_in_context(struct perf_event_context *ctx,
1993 struct perf_event *event,
1996 struct task_struct *task = ctx->task;
1998 lockdep_assert_held(&ctx->mutex);
2001 if (event->cpu != -1)
2006 * Per cpu events are installed via an smp call and
2007 * the install is always successful.
2009 cpu_function_call(cpu, __perf_install_in_context, event);
2014 if (!task_function_call(task, __perf_install_in_context, event))
2017 raw_spin_lock_irq(&ctx->lock);
2019 * If we failed to find a running task, but find the context active now
2020 * that we've acquired the ctx->lock, retry.
2022 if (ctx->is_active) {
2023 raw_spin_unlock_irq(&ctx->lock);
2028 * Since the task isn't running, its safe to add the event, us holding
2029 * the ctx->lock ensures the task won't get scheduled in.
2031 add_event_to_ctx(event, ctx);
2032 raw_spin_unlock_irq(&ctx->lock);
2036 * Put a event into inactive state and update time fields.
2037 * Enabling the leader of a group effectively enables all
2038 * the group members that aren't explicitly disabled, so we
2039 * have to update their ->tstamp_enabled also.
2040 * Note: this works for group members as well as group leaders
2041 * since the non-leader members' sibling_lists will be empty.
2043 static void __perf_event_mark_enabled(struct perf_event *event)
2045 struct perf_event *sub;
2046 u64 tstamp = perf_event_time(event);
2048 event->state = PERF_EVENT_STATE_INACTIVE;
2049 event->tstamp_enabled = tstamp - event->total_time_enabled;
2050 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2051 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2052 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2057 * Cross CPU call to enable a performance event
2059 static int __perf_event_enable(void *info)
2061 struct perf_event *event = info;
2062 struct perf_event_context *ctx = event->ctx;
2063 struct perf_event *leader = event->group_leader;
2064 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2068 * There's a time window between 'ctx->is_active' check
2069 * in perf_event_enable function and this place having:
2071 * - ctx->lock unlocked
2073 * where the task could be killed and 'ctx' deactivated
2074 * by perf_event_exit_task.
2076 if (!ctx->is_active)
2079 raw_spin_lock(&ctx->lock);
2080 update_context_time(ctx);
2082 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2086 * set current task's cgroup time reference point
2088 perf_cgroup_set_timestamp(current, ctx);
2090 __perf_event_mark_enabled(event);
2092 if (!event_filter_match(event)) {
2093 if (is_cgroup_event(event))
2094 perf_cgroup_defer_enabled(event);
2099 * If the event is in a group and isn't the group leader,
2100 * then don't put it on unless the group is on.
2102 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2105 if (!group_can_go_on(event, cpuctx, 1)) {
2108 if (event == leader)
2109 err = group_sched_in(event, cpuctx, ctx);
2111 err = event_sched_in(event, cpuctx, ctx);
2116 * If this event can't go on and it's part of a
2117 * group, then the whole group has to come off.
2119 if (leader != event) {
2120 group_sched_out(leader, cpuctx, ctx);
2121 perf_cpu_hrtimer_restart(cpuctx);
2123 if (leader->attr.pinned) {
2124 update_group_times(leader);
2125 leader->state = PERF_EVENT_STATE_ERROR;
2130 raw_spin_unlock(&ctx->lock);
2138 * If event->ctx is a cloned context, callers must make sure that
2139 * every task struct that event->ctx->task could possibly point to
2140 * remains valid. This condition is satisfied when called through
2141 * perf_event_for_each_child or perf_event_for_each as described
2142 * for perf_event_disable.
2144 void perf_event_enable(struct perf_event *event)
2146 struct perf_event_context *ctx = event->ctx;
2147 struct task_struct *task = ctx->task;
2151 * Enable the event on the cpu that it's on
2153 cpu_function_call(event->cpu, __perf_event_enable, event);
2157 raw_spin_lock_irq(&ctx->lock);
2158 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2162 * If the event is in error state, clear that first.
2163 * That way, if we see the event in error state below, we
2164 * know that it has gone back into error state, as distinct
2165 * from the task having been scheduled away before the
2166 * cross-call arrived.
2168 if (event->state == PERF_EVENT_STATE_ERROR)
2169 event->state = PERF_EVENT_STATE_OFF;
2172 if (!ctx->is_active) {
2173 __perf_event_mark_enabled(event);
2177 raw_spin_unlock_irq(&ctx->lock);
2179 if (!task_function_call(task, __perf_event_enable, event))
2182 raw_spin_lock_irq(&ctx->lock);
2185 * If the context is active and the event is still off,
2186 * we need to retry the cross-call.
2188 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2190 * task could have been flipped by a concurrent
2191 * perf_event_context_sched_out()
2198 raw_spin_unlock_irq(&ctx->lock);
2200 EXPORT_SYMBOL_GPL(perf_event_enable);
2202 int perf_event_refresh(struct perf_event *event, int refresh)
2205 * not supported on inherited events
2207 if (event->attr.inherit || !is_sampling_event(event))
2210 atomic_add(refresh, &event->event_limit);
2211 perf_event_enable(event);
2215 EXPORT_SYMBOL_GPL(perf_event_refresh);
2217 static void ctx_sched_out(struct perf_event_context *ctx,
2218 struct perf_cpu_context *cpuctx,
2219 enum event_type_t event_type)
2221 struct perf_event *event;
2222 int is_active = ctx->is_active;
2224 ctx->is_active &= ~event_type;
2225 if (likely(!ctx->nr_events))
2228 update_context_time(ctx);
2229 update_cgrp_time_from_cpuctx(cpuctx);
2230 if (!ctx->nr_active)
2233 perf_pmu_disable(ctx->pmu);
2234 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2235 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2236 group_sched_out(event, cpuctx, ctx);
2239 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2240 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2241 group_sched_out(event, cpuctx, ctx);
2243 perf_pmu_enable(ctx->pmu);
2247 * Test whether two contexts are equivalent, i.e. whether they have both been
2248 * cloned from the same version of the same context.
2250 * Equivalence is measured using a generation number in the context that is
2251 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2252 * and list_del_event().
2254 static int context_equiv(struct perf_event_context *ctx1,
2255 struct perf_event_context *ctx2)
2257 /* Pinning disables the swap optimization */
2258 if (ctx1->pin_count || ctx2->pin_count)
2261 /* If ctx1 is the parent of ctx2 */
2262 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2265 /* If ctx2 is the parent of ctx1 */
2266 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2270 * If ctx1 and ctx2 have the same parent; we flatten the parent
2271 * hierarchy, see perf_event_init_context().
2273 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2274 ctx1->parent_gen == ctx2->parent_gen)
2281 static void __perf_event_sync_stat(struct perf_event *event,
2282 struct perf_event *next_event)
2286 if (!event->attr.inherit_stat)
2290 * Update the event value, we cannot use perf_event_read()
2291 * because we're in the middle of a context switch and have IRQs
2292 * disabled, which upsets smp_call_function_single(), however
2293 * we know the event must be on the current CPU, therefore we
2294 * don't need to use it.
2296 switch (event->state) {
2297 case PERF_EVENT_STATE_ACTIVE:
2298 event->pmu->read(event);
2301 case PERF_EVENT_STATE_INACTIVE:
2302 update_event_times(event);
2310 * In order to keep per-task stats reliable we need to flip the event
2311 * values when we flip the contexts.
2313 value = local64_read(&next_event->count);
2314 value = local64_xchg(&event->count, value);
2315 local64_set(&next_event->count, value);
2317 swap(event->total_time_enabled, next_event->total_time_enabled);
2318 swap(event->total_time_running, next_event->total_time_running);
2321 * Since we swizzled the values, update the user visible data too.
2323 perf_event_update_userpage(event);
2324 perf_event_update_userpage(next_event);
2327 static void perf_event_sync_stat(struct perf_event_context *ctx,
2328 struct perf_event_context *next_ctx)
2330 struct perf_event *event, *next_event;
2335 update_context_time(ctx);
2337 event = list_first_entry(&ctx->event_list,
2338 struct perf_event, event_entry);
2340 next_event = list_first_entry(&next_ctx->event_list,
2341 struct perf_event, event_entry);
2343 while (&event->event_entry != &ctx->event_list &&
2344 &next_event->event_entry != &next_ctx->event_list) {
2346 __perf_event_sync_stat(event, next_event);
2348 event = list_next_entry(event, event_entry);
2349 next_event = list_next_entry(next_event, event_entry);
2353 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2354 struct task_struct *next)
2356 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2357 struct perf_event_context *next_ctx;
2358 struct perf_event_context *parent, *next_parent;
2359 struct perf_cpu_context *cpuctx;
2365 cpuctx = __get_cpu_context(ctx);
2366 if (!cpuctx->task_ctx)
2370 next_ctx = next->perf_event_ctxp[ctxn];
2374 parent = rcu_dereference(ctx->parent_ctx);
2375 next_parent = rcu_dereference(next_ctx->parent_ctx);
2377 /* If neither context have a parent context; they cannot be clones. */
2378 if (!parent || !next_parent)
2381 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2383 * Looks like the two contexts are clones, so we might be
2384 * able to optimize the context switch. We lock both
2385 * contexts and check that they are clones under the
2386 * lock (including re-checking that neither has been
2387 * uncloned in the meantime). It doesn't matter which
2388 * order we take the locks because no other cpu could
2389 * be trying to lock both of these tasks.
2391 raw_spin_lock(&ctx->lock);
2392 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2393 if (context_equiv(ctx, next_ctx)) {
2395 * XXX do we need a memory barrier of sorts
2396 * wrt to rcu_dereference() of perf_event_ctxp
2398 task->perf_event_ctxp[ctxn] = next_ctx;
2399 next->perf_event_ctxp[ctxn] = ctx;
2401 next_ctx->task = task;
2404 perf_event_sync_stat(ctx, next_ctx);
2406 raw_spin_unlock(&next_ctx->lock);
2407 raw_spin_unlock(&ctx->lock);
2413 raw_spin_lock(&ctx->lock);
2414 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2415 cpuctx->task_ctx = NULL;
2416 raw_spin_unlock(&ctx->lock);
2420 #define for_each_task_context_nr(ctxn) \
2421 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2424 * Called from scheduler to remove the events of the current task,
2425 * with interrupts disabled.
2427 * We stop each event and update the event value in event->count.
2429 * This does not protect us against NMI, but disable()
2430 * sets the disabled bit in the control field of event _before_
2431 * accessing the event control register. If a NMI hits, then it will
2432 * not restart the event.
2434 void __perf_event_task_sched_out(struct task_struct *task,
2435 struct task_struct *next)
2439 for_each_task_context_nr(ctxn)
2440 perf_event_context_sched_out(task, ctxn, next);
2443 * if cgroup events exist on this CPU, then we need
2444 * to check if we have to switch out PMU state.
2445 * cgroup event are system-wide mode only
2447 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2448 perf_cgroup_sched_out(task, next);
2451 static void task_ctx_sched_out(struct perf_event_context *ctx)
2453 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2455 if (!cpuctx->task_ctx)
2458 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2461 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2462 cpuctx->task_ctx = NULL;
2466 * Called with IRQs disabled
2468 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2469 enum event_type_t event_type)
2471 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2475 ctx_pinned_sched_in(struct perf_event_context *ctx,
2476 struct perf_cpu_context *cpuctx)
2478 struct perf_event *event;
2480 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2481 if (event->state <= PERF_EVENT_STATE_OFF)
2483 if (!event_filter_match(event))
2486 /* may need to reset tstamp_enabled */
2487 if (is_cgroup_event(event))
2488 perf_cgroup_mark_enabled(event, ctx);
2490 if (group_can_go_on(event, cpuctx, 1))
2491 group_sched_in(event, cpuctx, ctx);
2494 * If this pinned group hasn't been scheduled,
2495 * put it in error state.
2497 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2498 update_group_times(event);
2499 event->state = PERF_EVENT_STATE_ERROR;
2505 ctx_flexible_sched_in(struct perf_event_context *ctx,
2506 struct perf_cpu_context *cpuctx)
2508 struct perf_event *event;
2511 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2512 /* Ignore events in OFF or ERROR state */
2513 if (event->state <= PERF_EVENT_STATE_OFF)
2516 * Listen to the 'cpu' scheduling filter constraint
2519 if (!event_filter_match(event))
2522 /* may need to reset tstamp_enabled */
2523 if (is_cgroup_event(event))
2524 perf_cgroup_mark_enabled(event, ctx);
2526 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2527 if (group_sched_in(event, cpuctx, ctx))
2534 ctx_sched_in(struct perf_event_context *ctx,
2535 struct perf_cpu_context *cpuctx,
2536 enum event_type_t event_type,
2537 struct task_struct *task)
2540 int is_active = ctx->is_active;
2542 ctx->is_active |= event_type;
2543 if (likely(!ctx->nr_events))
2547 ctx->timestamp = now;
2548 perf_cgroup_set_timestamp(task, ctx);
2550 * First go through the list and put on any pinned groups
2551 * in order to give them the best chance of going on.
2553 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2554 ctx_pinned_sched_in(ctx, cpuctx);
2556 /* Then walk through the lower prio flexible groups */
2557 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2558 ctx_flexible_sched_in(ctx, cpuctx);
2561 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2562 enum event_type_t event_type,
2563 struct task_struct *task)
2565 struct perf_event_context *ctx = &cpuctx->ctx;
2567 ctx_sched_in(ctx, cpuctx, event_type, task);
2570 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2571 struct task_struct *task)
2573 struct perf_cpu_context *cpuctx;
2575 cpuctx = __get_cpu_context(ctx);
2576 if (cpuctx->task_ctx == ctx)
2579 perf_ctx_lock(cpuctx, ctx);
2580 perf_pmu_disable(ctx->pmu);
2582 * We want to keep the following priority order:
2583 * cpu pinned (that don't need to move), task pinned,
2584 * cpu flexible, task flexible.
2586 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2589 cpuctx->task_ctx = ctx;
2591 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2593 perf_pmu_enable(ctx->pmu);
2594 perf_ctx_unlock(cpuctx, ctx);
2597 * Since these rotations are per-cpu, we need to ensure the
2598 * cpu-context we got scheduled on is actually rotating.
2600 perf_pmu_rotate_start(ctx->pmu);
2604 * When sampling the branck stack in system-wide, it may be necessary
2605 * to flush the stack on context switch. This happens when the branch
2606 * stack does not tag its entries with the pid of the current task.
2607 * Otherwise it becomes impossible to associate a branch entry with a
2608 * task. This ambiguity is more likely to appear when the branch stack
2609 * supports priv level filtering and the user sets it to monitor only
2610 * at the user level (which could be a useful measurement in system-wide
2611 * mode). In that case, the risk is high of having a branch stack with
2612 * branch from multiple tasks. Flushing may mean dropping the existing
2613 * entries or stashing them somewhere in the PMU specific code layer.
2615 * This function provides the context switch callback to the lower code
2616 * layer. It is invoked ONLY when there is at least one system-wide context
2617 * with at least one active event using taken branch sampling.
2619 static void perf_branch_stack_sched_in(struct task_struct *prev,
2620 struct task_struct *task)
2622 struct perf_cpu_context *cpuctx;
2624 unsigned long flags;
2626 /* no need to flush branch stack if not changing task */
2630 local_irq_save(flags);
2634 list_for_each_entry_rcu(pmu, &pmus, entry) {
2635 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2638 * check if the context has at least one
2639 * event using PERF_SAMPLE_BRANCH_STACK
2641 if (cpuctx->ctx.nr_branch_stack > 0
2642 && pmu->flush_branch_stack) {
2644 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2646 perf_pmu_disable(pmu);
2648 pmu->flush_branch_stack();
2650 perf_pmu_enable(pmu);
2652 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2658 local_irq_restore(flags);
2662 * Called from scheduler to add the events of the current task
2663 * with interrupts disabled.
2665 * We restore the event value and then enable it.
2667 * This does not protect us against NMI, but enable()
2668 * sets the enabled bit in the control field of event _before_
2669 * accessing the event control register. If a NMI hits, then it will
2670 * keep the event running.
2672 void __perf_event_task_sched_in(struct task_struct *prev,
2673 struct task_struct *task)
2675 struct perf_event_context *ctx;
2678 for_each_task_context_nr(ctxn) {
2679 ctx = task->perf_event_ctxp[ctxn];
2683 perf_event_context_sched_in(ctx, task);
2686 * if cgroup events exist on this CPU, then we need
2687 * to check if we have to switch in PMU state.
2688 * cgroup event are system-wide mode only
2690 if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
2691 perf_cgroup_sched_in(prev, task);
2693 /* check for system-wide branch_stack events */
2694 if (atomic_read(&__get_cpu_var(perf_branch_stack_events)))
2695 perf_branch_stack_sched_in(prev, task);
2698 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2700 u64 frequency = event->attr.sample_freq;
2701 u64 sec = NSEC_PER_SEC;
2702 u64 divisor, dividend;
2704 int count_fls, nsec_fls, frequency_fls, sec_fls;
2706 count_fls = fls64(count);
2707 nsec_fls = fls64(nsec);
2708 frequency_fls = fls64(frequency);
2712 * We got @count in @nsec, with a target of sample_freq HZ
2713 * the target period becomes:
2716 * period = -------------------
2717 * @nsec * sample_freq
2722 * Reduce accuracy by one bit such that @a and @b converge
2723 * to a similar magnitude.
2725 #define REDUCE_FLS(a, b) \
2727 if (a##_fls > b##_fls) { \
2737 * Reduce accuracy until either term fits in a u64, then proceed with
2738 * the other, so that finally we can do a u64/u64 division.
2740 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2741 REDUCE_FLS(nsec, frequency);
2742 REDUCE_FLS(sec, count);
2745 if (count_fls + sec_fls > 64) {
2746 divisor = nsec * frequency;
2748 while (count_fls + sec_fls > 64) {
2749 REDUCE_FLS(count, sec);
2753 dividend = count * sec;
2755 dividend = count * sec;
2757 while (nsec_fls + frequency_fls > 64) {
2758 REDUCE_FLS(nsec, frequency);
2762 divisor = nsec * frequency;
2768 return div64_u64(dividend, divisor);
2771 static DEFINE_PER_CPU(int, perf_throttled_count);
2772 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2774 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2776 struct hw_perf_event *hwc = &event->hw;
2777 s64 period, sample_period;
2780 period = perf_calculate_period(event, nsec, count);
2782 delta = (s64)(period - hwc->sample_period);
2783 delta = (delta + 7) / 8; /* low pass filter */
2785 sample_period = hwc->sample_period + delta;
2790 hwc->sample_period = sample_period;
2792 if (local64_read(&hwc->period_left) > 8*sample_period) {
2794 event->pmu->stop(event, PERF_EF_UPDATE);
2796 local64_set(&hwc->period_left, 0);
2799 event->pmu->start(event, PERF_EF_RELOAD);
2804 * combine freq adjustment with unthrottling to avoid two passes over the
2805 * events. At the same time, make sure, having freq events does not change
2806 * the rate of unthrottling as that would introduce bias.
2808 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2811 struct perf_event *event;
2812 struct hw_perf_event *hwc;
2813 u64 now, period = TICK_NSEC;
2817 * only need to iterate over all events iff:
2818 * - context have events in frequency mode (needs freq adjust)
2819 * - there are events to unthrottle on this cpu
2821 if (!(ctx->nr_freq || needs_unthr))
2824 raw_spin_lock(&ctx->lock);
2825 perf_pmu_disable(ctx->pmu);
2827 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2828 if (event->state != PERF_EVENT_STATE_ACTIVE)
2831 if (!event_filter_match(event))
2834 perf_pmu_disable(event->pmu);
2838 if (hwc->interrupts == MAX_INTERRUPTS) {
2839 hwc->interrupts = 0;
2840 perf_log_throttle(event, 1);
2841 event->pmu->start(event, 0);
2844 if (!event->attr.freq || !event->attr.sample_freq)
2848 * stop the event and update event->count
2850 event->pmu->stop(event, PERF_EF_UPDATE);
2852 now = local64_read(&event->count);
2853 delta = now - hwc->freq_count_stamp;
2854 hwc->freq_count_stamp = now;
2858 * reload only if value has changed
2859 * we have stopped the event so tell that
2860 * to perf_adjust_period() to avoid stopping it
2864 perf_adjust_period(event, period, delta, false);
2866 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
2868 perf_pmu_enable(event->pmu);
2871 perf_pmu_enable(ctx->pmu);
2872 raw_spin_unlock(&ctx->lock);
2876 * Round-robin a context's events:
2878 static void rotate_ctx(struct perf_event_context *ctx)
2881 * Rotate the first entry last of non-pinned groups. Rotation might be
2882 * disabled by the inheritance code.
2884 if (!ctx->rotate_disable)
2885 list_rotate_left(&ctx->flexible_groups);
2889 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
2890 * because they're strictly cpu affine and rotate_start is called with IRQs
2891 * disabled, while rotate_context is called from IRQ context.
2893 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
2895 struct perf_event_context *ctx = NULL;
2896 int rotate = 0, remove = 1;
2898 if (cpuctx->ctx.nr_events) {
2900 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
2904 ctx = cpuctx->task_ctx;
2905 if (ctx && ctx->nr_events) {
2907 if (ctx->nr_events != ctx->nr_active)
2914 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2915 perf_pmu_disable(cpuctx->ctx.pmu);
2917 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2919 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
2921 rotate_ctx(&cpuctx->ctx);
2925 perf_event_sched_in(cpuctx, ctx, current);
2927 perf_pmu_enable(cpuctx->ctx.pmu);
2928 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2931 list_del_init(&cpuctx->rotation_list);
2936 #ifdef CONFIG_NO_HZ_FULL
2937 bool perf_event_can_stop_tick(void)
2939 if (atomic_read(&nr_freq_events) ||
2940 __this_cpu_read(perf_throttled_count))
2947 void perf_event_task_tick(void)
2949 struct list_head *head = &__get_cpu_var(rotation_list);
2950 struct perf_cpu_context *cpuctx, *tmp;
2951 struct perf_event_context *ctx;
2954 WARN_ON(!irqs_disabled());
2956 __this_cpu_inc(perf_throttled_seq);
2957 throttled = __this_cpu_xchg(perf_throttled_count, 0);
2959 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
2961 perf_adjust_freq_unthr_context(ctx, throttled);
2963 ctx = cpuctx->task_ctx;
2965 perf_adjust_freq_unthr_context(ctx, throttled);
2969 static int event_enable_on_exec(struct perf_event *event,
2970 struct perf_event_context *ctx)
2972 if (!event->attr.enable_on_exec)
2975 event->attr.enable_on_exec = 0;
2976 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2979 __perf_event_mark_enabled(event);
2985 * Enable all of a task's events that have been marked enable-on-exec.
2986 * This expects task == current.
2988 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
2990 struct perf_event *event;
2991 unsigned long flags;
2995 local_irq_save(flags);
2996 if (!ctx || !ctx->nr_events)
3000 * We must ctxsw out cgroup events to avoid conflict
3001 * when invoking perf_task_event_sched_in() later on
3002 * in this function. Otherwise we end up trying to
3003 * ctxswin cgroup events which are already scheduled
3006 perf_cgroup_sched_out(current, NULL);
3008 raw_spin_lock(&ctx->lock);
3009 task_ctx_sched_out(ctx);
3011 list_for_each_entry(event, &ctx->event_list, event_entry) {
3012 ret = event_enable_on_exec(event, ctx);
3018 * Unclone this context if we enabled any event.
3023 raw_spin_unlock(&ctx->lock);
3026 * Also calls ctxswin for cgroup events, if any:
3028 perf_event_context_sched_in(ctx, ctx->task);
3030 local_irq_restore(flags);
3033 void perf_event_exec(void)
3035 struct perf_event_context *ctx;
3039 for_each_task_context_nr(ctxn) {
3040 ctx = current->perf_event_ctxp[ctxn];
3044 perf_event_enable_on_exec(ctx);
3050 * Cross CPU call to read the hardware event
3052 static void __perf_event_read(void *info)
3054 struct perf_event *event = info;
3055 struct perf_event_context *ctx = event->ctx;
3056 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3059 * If this is a task context, we need to check whether it is
3060 * the current task context of this cpu. If not it has been
3061 * scheduled out before the smp call arrived. In that case
3062 * event->count would have been updated to a recent sample
3063 * when the event was scheduled out.
3065 if (ctx->task && cpuctx->task_ctx != ctx)
3068 raw_spin_lock(&ctx->lock);
3069 if (ctx->is_active) {
3070 update_context_time(ctx);
3071 update_cgrp_time_from_event(event);
3073 update_event_times(event);
3074 if (event->state == PERF_EVENT_STATE_ACTIVE)
3075 event->pmu->read(event);
3076 raw_spin_unlock(&ctx->lock);
3079 static inline u64 perf_event_count(struct perf_event *event)
3081 return local64_read(&event->count) + atomic64_read(&event->child_count);
3084 static u64 perf_event_read(struct perf_event *event)
3087 * If event is enabled and currently active on a CPU, update the
3088 * value in the event structure:
3090 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3091 smp_call_function_single(event->oncpu,
3092 __perf_event_read, event, 1);
3093 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3094 struct perf_event_context *ctx = event->ctx;
3095 unsigned long flags;
3097 raw_spin_lock_irqsave(&ctx->lock, flags);
3099 * may read while context is not active
3100 * (e.g., thread is blocked), in that case
3101 * we cannot update context time
3103 if (ctx->is_active) {
3104 update_context_time(ctx);
3105 update_cgrp_time_from_event(event);
3107 update_event_times(event);
3108 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3111 return perf_event_count(event);
3115 * Initialize the perf_event context in a task_struct:
3117 static void __perf_event_init_context(struct perf_event_context *ctx)
3119 raw_spin_lock_init(&ctx->lock);
3120 mutex_init(&ctx->mutex);
3121 INIT_LIST_HEAD(&ctx->pinned_groups);
3122 INIT_LIST_HEAD(&ctx->flexible_groups);
3123 INIT_LIST_HEAD(&ctx->event_list);
3124 atomic_set(&ctx->refcount, 1);
3125 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3128 static struct perf_event_context *
3129 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3131 struct perf_event_context *ctx;
3133 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3137 __perf_event_init_context(ctx);
3140 get_task_struct(task);
3147 static struct task_struct *
3148 find_lively_task_by_vpid(pid_t vpid)
3150 struct task_struct *task;
3157 task = find_task_by_vpid(vpid);
3159 get_task_struct(task);
3163 return ERR_PTR(-ESRCH);
3165 /* Reuse ptrace permission checks for now. */
3167 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3172 put_task_struct(task);
3173 return ERR_PTR(err);
3178 * Returns a matching context with refcount and pincount.
3180 static struct perf_event_context *
3181 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3183 struct perf_event_context *ctx;
3184 struct perf_cpu_context *cpuctx;
3185 unsigned long flags;
3189 /* Must be root to operate on a CPU event: */
3190 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3191 return ERR_PTR(-EACCES);
3194 * We could be clever and allow to attach a event to an
3195 * offline CPU and activate it when the CPU comes up, but
3198 if (!cpu_online(cpu))
3199 return ERR_PTR(-ENODEV);
3201 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3210 ctxn = pmu->task_ctx_nr;
3215 ctx = perf_lock_task_context(task, ctxn, &flags);
3219 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3221 ctx = alloc_perf_context(pmu, task);
3227 mutex_lock(&task->perf_event_mutex);
3229 * If it has already passed perf_event_exit_task().
3230 * we must see PF_EXITING, it takes this mutex too.
3232 if (task->flags & PF_EXITING)
3234 else if (task->perf_event_ctxp[ctxn])
3239 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3241 mutex_unlock(&task->perf_event_mutex);
3243 if (unlikely(err)) {
3255 return ERR_PTR(err);
3258 static void perf_event_free_filter(struct perf_event *event);
3260 static void free_event_rcu(struct rcu_head *head)
3262 struct perf_event *event;
3264 event = container_of(head, struct perf_event, rcu_head);
3266 put_pid_ns(event->ns);
3267 perf_event_free_filter(event);
3271 static void ring_buffer_put(struct ring_buffer *rb);
3272 static void ring_buffer_attach(struct perf_event *event,
3273 struct ring_buffer *rb);
3275 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3280 if (has_branch_stack(event)) {
3281 if (!(event->attach_state & PERF_ATTACH_TASK))
3282 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3284 if (is_cgroup_event(event))
3285 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3288 static void unaccount_event(struct perf_event *event)
3293 if (event->attach_state & PERF_ATTACH_TASK)
3294 static_key_slow_dec_deferred(&perf_sched_events);
3295 if (event->attr.mmap || event->attr.mmap_data)
3296 atomic_dec(&nr_mmap_events);
3297 if (event->attr.comm)
3298 atomic_dec(&nr_comm_events);
3299 if (event->attr.task)
3300 atomic_dec(&nr_task_events);
3301 if (event->attr.freq)
3302 atomic_dec(&nr_freq_events);
3303 if (is_cgroup_event(event))
3304 static_key_slow_dec_deferred(&perf_sched_events);
3305 if (has_branch_stack(event))
3306 static_key_slow_dec_deferred(&perf_sched_events);
3308 unaccount_event_cpu(event, event->cpu);
3311 static void __free_event(struct perf_event *event)
3313 if (!event->parent) {
3314 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3315 put_callchain_buffers();
3319 event->destroy(event);
3322 put_ctx(event->ctx);
3325 module_put(event->pmu->module);
3327 call_rcu(&event->rcu_head, free_event_rcu);
3330 static void _free_event(struct perf_event *event)
3332 irq_work_sync(&event->pending);
3334 unaccount_event(event);
3338 * Can happen when we close an event with re-directed output.
3340 * Since we have a 0 refcount, perf_mmap_close() will skip
3341 * over us; possibly making our ring_buffer_put() the last.
3343 mutex_lock(&event->mmap_mutex);
3344 ring_buffer_attach(event, NULL);
3345 mutex_unlock(&event->mmap_mutex);
3348 if (is_cgroup_event(event))
3349 perf_detach_cgroup(event);
3351 __free_event(event);
3355 * Used to free events which have a known refcount of 1, such as in error paths
3356 * where the event isn't exposed yet and inherited events.
3358 static void free_event(struct perf_event *event)
3360 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3361 "unexpected event refcount: %ld; ptr=%p\n",
3362 atomic_long_read(&event->refcount), event)) {
3363 /* leak to avoid use-after-free */
3371 * Remove user event from the owner task.
3373 static void perf_remove_from_owner(struct perf_event *event)
3375 struct task_struct *owner;
3378 owner = ACCESS_ONCE(event->owner);
3380 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3381 * !owner it means the list deletion is complete and we can indeed
3382 * free this event, otherwise we need to serialize on
3383 * owner->perf_event_mutex.
3385 smp_read_barrier_depends();
3388 * Since delayed_put_task_struct() also drops the last
3389 * task reference we can safely take a new reference
3390 * while holding the rcu_read_lock().
3392 get_task_struct(owner);
3397 mutex_lock(&owner->perf_event_mutex);
3399 * We have to re-check the event->owner field, if it is cleared
3400 * we raced with perf_event_exit_task(), acquiring the mutex
3401 * ensured they're done, and we can proceed with freeing the
3405 list_del_init(&event->owner_entry);
3406 mutex_unlock(&owner->perf_event_mutex);
3407 put_task_struct(owner);
3412 * Called when the last reference to the file is gone.
3414 static void put_event(struct perf_event *event)
3416 struct perf_event_context *ctx = event->ctx;
3418 if (!atomic_long_dec_and_test(&event->refcount))
3421 if (!is_kernel_event(event))
3422 perf_remove_from_owner(event);
3424 WARN_ON_ONCE(ctx->parent_ctx);
3426 * There are two ways this annotation is useful:
3428 * 1) there is a lock recursion from perf_event_exit_task
3429 * see the comment there.
3431 * 2) there is a lock-inversion with mmap_sem through
3432 * perf_event_read_group(), which takes faults while
3433 * holding ctx->mutex, however this is called after
3434 * the last filedesc died, so there is no possibility
3435 * to trigger the AB-BA case.
3437 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3438 perf_remove_from_context(event, true);
3439 mutex_unlock(&ctx->mutex);
3444 int perf_event_release_kernel(struct perf_event *event)
3449 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3451 static int perf_release(struct inode *inode, struct file *file)
3453 put_event(file->private_data);
3458 * Remove all orphanes events from the context.
3460 static void orphans_remove_work(struct work_struct *work)
3462 struct perf_event_context *ctx;
3463 struct perf_event *event, *tmp;
3465 ctx = container_of(work, struct perf_event_context,
3466 orphans_remove.work);
3468 mutex_lock(&ctx->mutex);
3469 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3470 struct perf_event *parent_event = event->parent;
3472 if (!is_orphaned_child(event))
3475 perf_remove_from_context(event, true);
3477 mutex_lock(&parent_event->child_mutex);
3478 list_del_init(&event->child_list);
3479 mutex_unlock(&parent_event->child_mutex);
3482 put_event(parent_event);
3485 raw_spin_lock_irq(&ctx->lock);
3486 ctx->orphans_remove_sched = false;
3487 raw_spin_unlock_irq(&ctx->lock);
3488 mutex_unlock(&ctx->mutex);
3493 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3495 struct perf_event *child;
3501 mutex_lock(&event->child_mutex);
3502 total += perf_event_read(event);
3503 *enabled += event->total_time_enabled +
3504 atomic64_read(&event->child_total_time_enabled);
3505 *running += event->total_time_running +
3506 atomic64_read(&event->child_total_time_running);
3508 list_for_each_entry(child, &event->child_list, child_list) {
3509 total += perf_event_read(child);
3510 *enabled += child->total_time_enabled;
3511 *running += child->total_time_running;
3513 mutex_unlock(&event->child_mutex);
3517 EXPORT_SYMBOL_GPL(perf_event_read_value);
3519 static int perf_event_read_group(struct perf_event *event,
3520 u64 read_format, char __user *buf)
3522 struct perf_event *leader = event->group_leader, *sub;
3523 int n = 0, size = 0, ret = -EFAULT;
3524 struct perf_event_context *ctx = leader->ctx;
3526 u64 count, enabled, running;
3528 mutex_lock(&ctx->mutex);
3529 count = perf_event_read_value(leader, &enabled, &running);
3531 values[n++] = 1 + leader->nr_siblings;
3532 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3533 values[n++] = enabled;
3534 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3535 values[n++] = running;
3536 values[n++] = count;
3537 if (read_format & PERF_FORMAT_ID)
3538 values[n++] = primary_event_id(leader);
3540 size = n * sizeof(u64);
3542 if (copy_to_user(buf, values, size))
3547 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3550 values[n++] = perf_event_read_value(sub, &enabled, &running);
3551 if (read_format & PERF_FORMAT_ID)
3552 values[n++] = primary_event_id(sub);
3554 size = n * sizeof(u64);
3556 if (copy_to_user(buf + ret, values, size)) {
3564 mutex_unlock(&ctx->mutex);
3569 static int perf_event_read_one(struct perf_event *event,
3570 u64 read_format, char __user *buf)
3572 u64 enabled, running;
3576 values[n++] = perf_event_read_value(event, &enabled, &running);
3577 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3578 values[n++] = enabled;
3579 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3580 values[n++] = running;
3581 if (read_format & PERF_FORMAT_ID)
3582 values[n++] = primary_event_id(event);
3584 if (copy_to_user(buf, values, n * sizeof(u64)))
3587 return n * sizeof(u64);
3591 * Read the performance event - simple non blocking version for now
3594 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3596 u64 read_format = event->attr.read_format;
3600 * Return end-of-file for a read on a event that is in
3601 * error state (i.e. because it was pinned but it couldn't be
3602 * scheduled on to the CPU at some point).
3604 if (event->state == PERF_EVENT_STATE_ERROR)
3607 if (count < event->read_size)
3610 WARN_ON_ONCE(event->ctx->parent_ctx);
3611 if (read_format & PERF_FORMAT_GROUP)
3612 ret = perf_event_read_group(event, read_format, buf);
3614 ret = perf_event_read_one(event, read_format, buf);
3620 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3622 struct perf_event *event = file->private_data;
3624 return perf_read_hw(event, buf, count);
3627 static unsigned int perf_poll(struct file *file, poll_table *wait)
3629 struct perf_event *event = file->private_data;
3630 struct ring_buffer *rb;
3631 unsigned int events = POLLHUP;
3633 poll_wait(file, &event->waitq, wait);
3635 if (event->state == PERF_EVENT_STATE_EXIT)
3639 * Pin the event->rb by taking event->mmap_mutex; otherwise
3640 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3642 mutex_lock(&event->mmap_mutex);
3645 events = atomic_xchg(&rb->poll, 0);
3646 mutex_unlock(&event->mmap_mutex);
3650 static void perf_event_reset(struct perf_event *event)
3652 (void)perf_event_read(event);
3653 local64_set(&event->count, 0);
3654 perf_event_update_userpage(event);
3658 * Holding the top-level event's child_mutex means that any
3659 * descendant process that has inherited this event will block
3660 * in sync_child_event if it goes to exit, thus satisfying the
3661 * task existence requirements of perf_event_enable/disable.
3663 static void perf_event_for_each_child(struct perf_event *event,
3664 void (*func)(struct perf_event *))
3666 struct perf_event *child;
3668 WARN_ON_ONCE(event->ctx->parent_ctx);
3669 mutex_lock(&event->child_mutex);
3671 list_for_each_entry(child, &event->child_list, child_list)
3673 mutex_unlock(&event->child_mutex);
3676 static void perf_event_for_each(struct perf_event *event,
3677 void (*func)(struct perf_event *))
3679 struct perf_event_context *ctx = event->ctx;
3680 struct perf_event *sibling;
3682 WARN_ON_ONCE(ctx->parent_ctx);
3683 mutex_lock(&ctx->mutex);
3684 event = event->group_leader;
3686 perf_event_for_each_child(event, func);
3687 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3688 perf_event_for_each_child(sibling, func);
3689 mutex_unlock(&ctx->mutex);
3692 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3694 struct perf_event_context *ctx = event->ctx;
3695 int ret = 0, active;
3698 if (!is_sampling_event(event))
3701 if (copy_from_user(&value, arg, sizeof(value)))
3707 raw_spin_lock_irq(&ctx->lock);
3708 if (event->attr.freq) {
3709 if (value > sysctl_perf_event_sample_rate) {
3714 event->attr.sample_freq = value;
3716 event->attr.sample_period = value;
3717 event->hw.sample_period = value;
3720 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3722 perf_pmu_disable(ctx->pmu);
3723 event->pmu->stop(event, PERF_EF_UPDATE);
3726 local64_set(&event->hw.period_left, 0);
3729 event->pmu->start(event, PERF_EF_RELOAD);
3730 perf_pmu_enable(ctx->pmu);
3734 raw_spin_unlock_irq(&ctx->lock);
3739 static const struct file_operations perf_fops;
3741 static inline int perf_fget_light(int fd, struct fd *p)
3743 struct fd f = fdget(fd);
3747 if (f.file->f_op != &perf_fops) {
3755 static int perf_event_set_output(struct perf_event *event,
3756 struct perf_event *output_event);
3757 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3759 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
3761 struct perf_event *event = file->private_data;
3762 void (*func)(struct perf_event *);
3766 case PERF_EVENT_IOC_ENABLE:
3767 func = perf_event_enable;
3769 case PERF_EVENT_IOC_DISABLE:
3770 func = perf_event_disable;
3772 case PERF_EVENT_IOC_RESET:
3773 func = perf_event_reset;
3776 case PERF_EVENT_IOC_REFRESH:
3777 return perf_event_refresh(event, arg);
3779 case PERF_EVENT_IOC_PERIOD:
3780 return perf_event_period(event, (u64 __user *)arg);
3782 case PERF_EVENT_IOC_ID:
3784 u64 id = primary_event_id(event);
3786 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3791 case PERF_EVENT_IOC_SET_OUTPUT:
3795 struct perf_event *output_event;
3797 ret = perf_fget_light(arg, &output);
3800 output_event = output.file->private_data;
3801 ret = perf_event_set_output(event, output_event);
3804 ret = perf_event_set_output(event, NULL);
3809 case PERF_EVENT_IOC_SET_FILTER:
3810 return perf_event_set_filter(event, (void __user *)arg);
3816 if (flags & PERF_IOC_FLAG_GROUP)
3817 perf_event_for_each(event, func);
3819 perf_event_for_each_child(event, func);
3824 #ifdef CONFIG_COMPAT
3825 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
3828 switch (_IOC_NR(cmd)) {
3829 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
3830 case _IOC_NR(PERF_EVENT_IOC_ID):
3831 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
3832 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
3833 cmd &= ~IOCSIZE_MASK;
3834 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
3838 return perf_ioctl(file, cmd, arg);
3841 # define perf_compat_ioctl NULL
3844 int perf_event_task_enable(void)
3846 struct perf_event *event;
3848 mutex_lock(¤t->perf_event_mutex);
3849 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3850 perf_event_for_each_child(event, perf_event_enable);
3851 mutex_unlock(¤t->perf_event_mutex);
3856 int perf_event_task_disable(void)
3858 struct perf_event *event;
3860 mutex_lock(¤t->perf_event_mutex);
3861 list_for_each_entry(event, ¤t->perf_event_list, owner_entry)
3862 perf_event_for_each_child(event, perf_event_disable);
3863 mutex_unlock(¤t->perf_event_mutex);
3868 static int perf_event_index(struct perf_event *event)
3870 if (event->hw.state & PERF_HES_STOPPED)
3873 if (event->state != PERF_EVENT_STATE_ACTIVE)
3876 return event->pmu->event_idx(event);
3879 static void calc_timer_values(struct perf_event *event,
3886 *now = perf_clock();
3887 ctx_time = event->shadow_ctx_time + *now;
3888 *enabled = ctx_time - event->tstamp_enabled;
3889 *running = ctx_time - event->tstamp_running;
3892 static void perf_event_init_userpage(struct perf_event *event)
3894 struct perf_event_mmap_page *userpg;
3895 struct ring_buffer *rb;
3898 rb = rcu_dereference(event->rb);
3902 userpg = rb->user_page;
3904 /* Allow new userspace to detect that bit 0 is deprecated */
3905 userpg->cap_bit0_is_deprecated = 1;
3906 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
3912 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
3917 * Callers need to ensure there can be no nesting of this function, otherwise
3918 * the seqlock logic goes bad. We can not serialize this because the arch
3919 * code calls this from NMI context.
3921 void perf_event_update_userpage(struct perf_event *event)
3923 struct perf_event_mmap_page *userpg;
3924 struct ring_buffer *rb;
3925 u64 enabled, running, now;
3928 rb = rcu_dereference(event->rb);
3933 * compute total_time_enabled, total_time_running
3934 * based on snapshot values taken when the event
3935 * was last scheduled in.
3937 * we cannot simply called update_context_time()
3938 * because of locking issue as we can be called in
3941 calc_timer_values(event, &now, &enabled, &running);
3943 userpg = rb->user_page;
3945 * Disable preemption so as to not let the corresponding user-space
3946 * spin too long if we get preempted.
3951 userpg->index = perf_event_index(event);
3952 userpg->offset = perf_event_count(event);
3954 userpg->offset -= local64_read(&event->hw.prev_count);
3956 userpg->time_enabled = enabled +
3957 atomic64_read(&event->child_total_time_enabled);
3959 userpg->time_running = running +
3960 atomic64_read(&event->child_total_time_running);
3962 arch_perf_update_userpage(userpg, now);
3971 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
3973 struct perf_event *event = vma->vm_file->private_data;
3974 struct ring_buffer *rb;
3975 int ret = VM_FAULT_SIGBUS;
3977 if (vmf->flags & FAULT_FLAG_MKWRITE) {
3978 if (vmf->pgoff == 0)
3984 rb = rcu_dereference(event->rb);
3988 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
3991 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
3995 get_page(vmf->page);
3996 vmf->page->mapping = vma->vm_file->f_mapping;
3997 vmf->page->index = vmf->pgoff;
4006 static void ring_buffer_attach(struct perf_event *event,
4007 struct ring_buffer *rb)
4009 struct ring_buffer *old_rb = NULL;
4010 unsigned long flags;
4014 * Should be impossible, we set this when removing
4015 * event->rb_entry and wait/clear when adding event->rb_entry.
4017 WARN_ON_ONCE(event->rcu_pending);
4020 event->rcu_batches = get_state_synchronize_rcu();
4021 event->rcu_pending = 1;
4023 spin_lock_irqsave(&old_rb->event_lock, flags);
4024 list_del_rcu(&event->rb_entry);
4025 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4028 if (event->rcu_pending && rb) {
4029 cond_synchronize_rcu(event->rcu_batches);
4030 event->rcu_pending = 0;
4034 spin_lock_irqsave(&rb->event_lock, flags);
4035 list_add_rcu(&event->rb_entry, &rb->event_list);
4036 spin_unlock_irqrestore(&rb->event_lock, flags);
4039 rcu_assign_pointer(event->rb, rb);
4042 ring_buffer_put(old_rb);
4044 * Since we detached before setting the new rb, so that we
4045 * could attach the new rb, we could have missed a wakeup.
4048 wake_up_all(&event->waitq);
4052 static void ring_buffer_wakeup(struct perf_event *event)
4054 struct ring_buffer *rb;
4057 rb = rcu_dereference(event->rb);
4059 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4060 wake_up_all(&event->waitq);
4065 static void rb_free_rcu(struct rcu_head *rcu_head)
4067 struct ring_buffer *rb;
4069 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4073 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
4075 struct ring_buffer *rb;
4078 rb = rcu_dereference(event->rb);
4080 if (!atomic_inc_not_zero(&rb->refcount))
4088 static void ring_buffer_put(struct ring_buffer *rb)
4090 if (!atomic_dec_and_test(&rb->refcount))
4093 WARN_ON_ONCE(!list_empty(&rb->event_list));
4095 call_rcu(&rb->rcu_head, rb_free_rcu);
4098 static void perf_mmap_open(struct vm_area_struct *vma)
4100 struct perf_event *event = vma->vm_file->private_data;
4102 atomic_inc(&event->mmap_count);
4103 atomic_inc(&event->rb->mmap_count);
4107 * A buffer can be mmap()ed multiple times; either directly through the same
4108 * event, or through other events by use of perf_event_set_output().
4110 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4111 * the buffer here, where we still have a VM context. This means we need
4112 * to detach all events redirecting to us.
4114 static void perf_mmap_close(struct vm_area_struct *vma)
4116 struct perf_event *event = vma->vm_file->private_data;
4118 struct ring_buffer *rb = ring_buffer_get(event);
4119 struct user_struct *mmap_user = rb->mmap_user;
4120 int mmap_locked = rb->mmap_locked;
4121 unsigned long size = perf_data_size(rb);
4123 atomic_dec(&rb->mmap_count);
4125 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4128 ring_buffer_attach(event, NULL);
4129 mutex_unlock(&event->mmap_mutex);
4131 /* If there's still other mmap()s of this buffer, we're done. */
4132 if (atomic_read(&rb->mmap_count))
4136 * No other mmap()s, detach from all other events that might redirect
4137 * into the now unreachable buffer. Somewhat complicated by the
4138 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4142 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4143 if (!atomic_long_inc_not_zero(&event->refcount)) {
4145 * This event is en-route to free_event() which will
4146 * detach it and remove it from the list.
4152 mutex_lock(&event->mmap_mutex);
4154 * Check we didn't race with perf_event_set_output() which can
4155 * swizzle the rb from under us while we were waiting to
4156 * acquire mmap_mutex.
4158 * If we find a different rb; ignore this event, a next
4159 * iteration will no longer find it on the list. We have to
4160 * still restart the iteration to make sure we're not now
4161 * iterating the wrong list.
4163 if (event->rb == rb)
4164 ring_buffer_attach(event, NULL);
4166 mutex_unlock(&event->mmap_mutex);
4170 * Restart the iteration; either we're on the wrong list or
4171 * destroyed its integrity by doing a deletion.
4178 * It could be there's still a few 0-ref events on the list; they'll
4179 * get cleaned up by free_event() -- they'll also still have their
4180 * ref on the rb and will free it whenever they are done with it.
4182 * Aside from that, this buffer is 'fully' detached and unmapped,
4183 * undo the VM accounting.
4186 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4187 vma->vm_mm->pinned_vm -= mmap_locked;
4188 free_uid(mmap_user);
4191 ring_buffer_put(rb); /* could be last */
4194 static const struct vm_operations_struct perf_mmap_vmops = {
4195 .open = perf_mmap_open,
4196 .close = perf_mmap_close,
4197 .fault = perf_mmap_fault,
4198 .page_mkwrite = perf_mmap_fault,
4201 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4203 struct perf_event *event = file->private_data;
4204 unsigned long user_locked, user_lock_limit;
4205 struct user_struct *user = current_user();
4206 unsigned long locked, lock_limit;
4207 struct ring_buffer *rb;
4208 unsigned long vma_size;
4209 unsigned long nr_pages;
4210 long user_extra, extra;
4211 int ret = 0, flags = 0;
4214 * Don't allow mmap() of inherited per-task counters. This would
4215 * create a performance issue due to all children writing to the
4218 if (event->cpu == -1 && event->attr.inherit)
4221 if (!(vma->vm_flags & VM_SHARED))
4224 vma_size = vma->vm_end - vma->vm_start;
4225 nr_pages = (vma_size / PAGE_SIZE) - 1;
4228 * If we have rb pages ensure they're a power-of-two number, so we
4229 * can do bitmasks instead of modulo.
4231 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4234 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4237 if (vma->vm_pgoff != 0)
4240 WARN_ON_ONCE(event->ctx->parent_ctx);
4242 mutex_lock(&event->mmap_mutex);
4244 if (event->rb->nr_pages != nr_pages) {
4249 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4251 * Raced against perf_mmap_close() through
4252 * perf_event_set_output(). Try again, hope for better
4255 mutex_unlock(&event->mmap_mutex);
4262 user_extra = nr_pages + 1;
4263 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4266 * Increase the limit linearly with more CPUs:
4268 user_lock_limit *= num_online_cpus();
4270 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4273 if (user_locked > user_lock_limit)
4274 extra = user_locked - user_lock_limit;
4276 lock_limit = rlimit(RLIMIT_MEMLOCK);
4277 lock_limit >>= PAGE_SHIFT;
4278 locked = vma->vm_mm->pinned_vm + extra;
4280 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4281 !capable(CAP_IPC_LOCK)) {
4288 if (vma->vm_flags & VM_WRITE)
4289 flags |= RING_BUFFER_WRITABLE;
4291 rb = rb_alloc(nr_pages,
4292 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4300 atomic_set(&rb->mmap_count, 1);
4301 rb->mmap_locked = extra;
4302 rb->mmap_user = get_current_user();
4304 atomic_long_add(user_extra, &user->locked_vm);
4305 vma->vm_mm->pinned_vm += extra;
4307 ring_buffer_attach(event, rb);
4309 perf_event_init_userpage(event);
4310 perf_event_update_userpage(event);
4314 atomic_inc(&event->mmap_count);
4315 mutex_unlock(&event->mmap_mutex);
4318 * Since pinned accounting is per vm we cannot allow fork() to copy our
4321 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4322 vma->vm_ops = &perf_mmap_vmops;
4327 static int perf_fasync(int fd, struct file *filp, int on)
4329 struct inode *inode = file_inode(filp);
4330 struct perf_event *event = filp->private_data;
4333 mutex_lock(&inode->i_mutex);
4334 retval = fasync_helper(fd, filp, on, &event->fasync);
4335 mutex_unlock(&inode->i_mutex);
4343 static const struct file_operations perf_fops = {
4344 .llseek = no_llseek,
4345 .release = perf_release,
4348 .unlocked_ioctl = perf_ioctl,
4349 .compat_ioctl = perf_compat_ioctl,
4351 .fasync = perf_fasync,
4357 * If there's data, ensure we set the poll() state and publish everything
4358 * to user-space before waking everybody up.
4361 void perf_event_wakeup(struct perf_event *event)
4363 ring_buffer_wakeup(event);
4365 if (event->pending_kill) {
4366 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4367 event->pending_kill = 0;
4371 static void perf_pending_event(struct irq_work *entry)
4373 struct perf_event *event = container_of(entry,
4374 struct perf_event, pending);
4376 if (event->pending_disable) {
4377 event->pending_disable = 0;
4378 __perf_event_disable(event);
4381 if (event->pending_wakeup) {
4382 event->pending_wakeup = 0;
4383 perf_event_wakeup(event);
4388 * We assume there is only KVM supporting the callbacks.
4389 * Later on, we might change it to a list if there is
4390 * another virtualization implementation supporting the callbacks.
4392 struct perf_guest_info_callbacks *perf_guest_cbs;
4394 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4396 perf_guest_cbs = cbs;
4399 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4401 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4403 perf_guest_cbs = NULL;
4406 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4409 perf_output_sample_regs(struct perf_output_handle *handle,
4410 struct pt_regs *regs, u64 mask)
4414 for_each_set_bit(bit, (const unsigned long *) &mask,
4415 sizeof(mask) * BITS_PER_BYTE) {
4418 val = perf_reg_value(regs, bit);
4419 perf_output_put(handle, val);
4423 static void perf_sample_regs_user(struct perf_regs_user *regs_user,
4424 struct pt_regs *regs)
4426 if (!user_mode(regs)) {
4428 regs = task_pt_regs(current);
4434 regs_user->regs = regs;
4435 regs_user->abi = perf_reg_abi(current);
4440 * Get remaining task size from user stack pointer.
4442 * It'd be better to take stack vma map and limit this more
4443 * precisly, but there's no way to get it safely under interrupt,
4444 * so using TASK_SIZE as limit.
4446 static u64 perf_ustack_task_size(struct pt_regs *regs)
4448 unsigned long addr = perf_user_stack_pointer(regs);
4450 if (!addr || addr >= TASK_SIZE)
4453 return TASK_SIZE - addr;
4457 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4458 struct pt_regs *regs)
4462 /* No regs, no stack pointer, no dump. */
4467 * Check if we fit in with the requested stack size into the:
4469 * If we don't, we limit the size to the TASK_SIZE.
4471 * - remaining sample size
4472 * If we don't, we customize the stack size to
4473 * fit in to the remaining sample size.
4476 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4477 stack_size = min(stack_size, (u16) task_size);
4479 /* Current header size plus static size and dynamic size. */
4480 header_size += 2 * sizeof(u64);
4482 /* Do we fit in with the current stack dump size? */
4483 if ((u16) (header_size + stack_size) < header_size) {
4485 * If we overflow the maximum size for the sample,
4486 * we customize the stack dump size to fit in.
4488 stack_size = USHRT_MAX - header_size - sizeof(u64);
4489 stack_size = round_up(stack_size, sizeof(u64));
4496 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4497 struct pt_regs *regs)
4499 /* Case of a kernel thread, nothing to dump */
4502 perf_output_put(handle, size);
4511 * - the size requested by user or the best one we can fit
4512 * in to the sample max size
4514 * - user stack dump data
4516 * - the actual dumped size
4520 perf_output_put(handle, dump_size);
4523 sp = perf_user_stack_pointer(regs);
4524 rem = __output_copy_user(handle, (void *) sp, dump_size);
4525 dyn_size = dump_size - rem;
4527 perf_output_skip(handle, rem);
4530 perf_output_put(handle, dyn_size);
4534 static void __perf_event_header__init_id(struct perf_event_header *header,
4535 struct perf_sample_data *data,
4536 struct perf_event *event)
4538 u64 sample_type = event->attr.sample_type;
4540 data->type = sample_type;
4541 header->size += event->id_header_size;
4543 if (sample_type & PERF_SAMPLE_TID) {
4544 /* namespace issues */
4545 data->tid_entry.pid = perf_event_pid(event, current);
4546 data->tid_entry.tid = perf_event_tid(event, current);
4549 if (sample_type & PERF_SAMPLE_TIME)
4550 data->time = perf_clock();
4552 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4553 data->id = primary_event_id(event);
4555 if (sample_type & PERF_SAMPLE_STREAM_ID)
4556 data->stream_id = event->id;
4558 if (sample_type & PERF_SAMPLE_CPU) {
4559 data->cpu_entry.cpu = raw_smp_processor_id();
4560 data->cpu_entry.reserved = 0;
4564 void perf_event_header__init_id(struct perf_event_header *header,
4565 struct perf_sample_data *data,
4566 struct perf_event *event)
4568 if (event->attr.sample_id_all)
4569 __perf_event_header__init_id(header, data, event);
4572 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4573 struct perf_sample_data *data)
4575 u64 sample_type = data->type;
4577 if (sample_type & PERF_SAMPLE_TID)
4578 perf_output_put(handle, data->tid_entry);
4580 if (sample_type & PERF_SAMPLE_TIME)
4581 perf_output_put(handle, data->time);
4583 if (sample_type & PERF_SAMPLE_ID)
4584 perf_output_put(handle, data->id);
4586 if (sample_type & PERF_SAMPLE_STREAM_ID)
4587 perf_output_put(handle, data->stream_id);
4589 if (sample_type & PERF_SAMPLE_CPU)
4590 perf_output_put(handle, data->cpu_entry);
4592 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4593 perf_output_put(handle, data->id);
4596 void perf_event__output_id_sample(struct perf_event *event,
4597 struct perf_output_handle *handle,
4598 struct perf_sample_data *sample)
4600 if (event->attr.sample_id_all)
4601 __perf_event__output_id_sample(handle, sample);
4604 static void perf_output_read_one(struct perf_output_handle *handle,
4605 struct perf_event *event,
4606 u64 enabled, u64 running)
4608 u64 read_format = event->attr.read_format;
4612 values[n++] = perf_event_count(event);
4613 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4614 values[n++] = enabled +
4615 atomic64_read(&event->child_total_time_enabled);
4617 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4618 values[n++] = running +
4619 atomic64_read(&event->child_total_time_running);
4621 if (read_format & PERF_FORMAT_ID)
4622 values[n++] = primary_event_id(event);
4624 __output_copy(handle, values, n * sizeof(u64));
4628 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4630 static void perf_output_read_group(struct perf_output_handle *handle,
4631 struct perf_event *event,
4632 u64 enabled, u64 running)
4634 struct perf_event *leader = event->group_leader, *sub;
4635 u64 read_format = event->attr.read_format;
4639 values[n++] = 1 + leader->nr_siblings;
4641 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4642 values[n++] = enabled;
4644 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4645 values[n++] = running;
4647 if (leader != event)
4648 leader->pmu->read(leader);
4650 values[n++] = perf_event_count(leader);
4651 if (read_format & PERF_FORMAT_ID)
4652 values[n++] = primary_event_id(leader);
4654 __output_copy(handle, values, n * sizeof(u64));
4656 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4659 if ((sub != event) &&
4660 (sub->state == PERF_EVENT_STATE_ACTIVE))
4661 sub->pmu->read(sub);
4663 values[n++] = perf_event_count(sub);
4664 if (read_format & PERF_FORMAT_ID)
4665 values[n++] = primary_event_id(sub);
4667 __output_copy(handle, values, n * sizeof(u64));
4671 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4672 PERF_FORMAT_TOTAL_TIME_RUNNING)
4674 static void perf_output_read(struct perf_output_handle *handle,
4675 struct perf_event *event)
4677 u64 enabled = 0, running = 0, now;
4678 u64 read_format = event->attr.read_format;
4681 * compute total_time_enabled, total_time_running
4682 * based on snapshot values taken when the event
4683 * was last scheduled in.
4685 * we cannot simply called update_context_time()
4686 * because of locking issue as we are called in
4689 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4690 calc_timer_values(event, &now, &enabled, &running);
4692 if (event->attr.read_format & PERF_FORMAT_GROUP)
4693 perf_output_read_group(handle, event, enabled, running);
4695 perf_output_read_one(handle, event, enabled, running);
4698 void perf_output_sample(struct perf_output_handle *handle,
4699 struct perf_event_header *header,
4700 struct perf_sample_data *data,
4701 struct perf_event *event)
4703 u64 sample_type = data->type;
4705 perf_output_put(handle, *header);
4707 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4708 perf_output_put(handle, data->id);
4710 if (sample_type & PERF_SAMPLE_IP)
4711 perf_output_put(handle, data->ip);
4713 if (sample_type & PERF_SAMPLE_TID)
4714 perf_output_put(handle, data->tid_entry);
4716 if (sample_type & PERF_SAMPLE_TIME)
4717 perf_output_put(handle, data->time);
4719 if (sample_type & PERF_SAMPLE_ADDR)
4720 perf_output_put(handle, data->addr);
4722 if (sample_type & PERF_SAMPLE_ID)
4723 perf_output_put(handle, data->id);
4725 if (sample_type & PERF_SAMPLE_STREAM_ID)
4726 perf_output_put(handle, data->stream_id);
4728 if (sample_type & PERF_SAMPLE_CPU)
4729 perf_output_put(handle, data->cpu_entry);
4731 if (sample_type & PERF_SAMPLE_PERIOD)
4732 perf_output_put(handle, data->period);
4734 if (sample_type & PERF_SAMPLE_READ)
4735 perf_output_read(handle, event);
4737 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4738 if (data->callchain) {
4741 if (data->callchain)
4742 size += data->callchain->nr;
4744 size *= sizeof(u64);
4746 __output_copy(handle, data->callchain, size);
4749 perf_output_put(handle, nr);
4753 if (sample_type & PERF_SAMPLE_RAW) {
4755 perf_output_put(handle, data->raw->size);
4756 __output_copy(handle, data->raw->data,
4763 .size = sizeof(u32),
4766 perf_output_put(handle, raw);
4770 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4771 if (data->br_stack) {
4774 size = data->br_stack->nr
4775 * sizeof(struct perf_branch_entry);
4777 perf_output_put(handle, data->br_stack->nr);
4778 perf_output_copy(handle, data->br_stack->entries, size);
4781 * we always store at least the value of nr
4784 perf_output_put(handle, nr);
4788 if (sample_type & PERF_SAMPLE_REGS_USER) {
4789 u64 abi = data->regs_user.abi;
4792 * If there are no regs to dump, notice it through
4793 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4795 perf_output_put(handle, abi);
4798 u64 mask = event->attr.sample_regs_user;
4799 perf_output_sample_regs(handle,
4800 data->regs_user.regs,
4805 if (sample_type & PERF_SAMPLE_STACK_USER) {
4806 perf_output_sample_ustack(handle,
4807 data->stack_user_size,
4808 data->regs_user.regs);
4811 if (sample_type & PERF_SAMPLE_WEIGHT)
4812 perf_output_put(handle, data->weight);
4814 if (sample_type & PERF_SAMPLE_DATA_SRC)
4815 perf_output_put(handle, data->data_src.val);
4817 if (sample_type & PERF_SAMPLE_TRANSACTION)
4818 perf_output_put(handle, data->txn);
4820 if (!event->attr.watermark) {
4821 int wakeup_events = event->attr.wakeup_events;
4823 if (wakeup_events) {
4824 struct ring_buffer *rb = handle->rb;
4825 int events = local_inc_return(&rb->events);
4827 if (events >= wakeup_events) {
4828 local_sub(wakeup_events, &rb->events);
4829 local_inc(&rb->wakeup);
4835 void perf_prepare_sample(struct perf_event_header *header,
4836 struct perf_sample_data *data,
4837 struct perf_event *event,
4838 struct pt_regs *regs)
4840 u64 sample_type = event->attr.sample_type;
4842 header->type = PERF_RECORD_SAMPLE;
4843 header->size = sizeof(*header) + event->header_size;
4846 header->misc |= perf_misc_flags(regs);
4848 __perf_event_header__init_id(header, data, event);
4850 if (sample_type & PERF_SAMPLE_IP)
4851 data->ip = perf_instruction_pointer(regs);
4853 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4856 data->callchain = perf_callchain(event, regs);
4858 if (data->callchain)
4859 size += data->callchain->nr;
4861 header->size += size * sizeof(u64);
4864 if (sample_type & PERF_SAMPLE_RAW) {
4865 int size = sizeof(u32);
4868 size += data->raw->size;
4870 size += sizeof(u32);
4872 WARN_ON_ONCE(size & (sizeof(u64)-1));
4873 header->size += size;
4876 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4877 int size = sizeof(u64); /* nr */
4878 if (data->br_stack) {
4879 size += data->br_stack->nr
4880 * sizeof(struct perf_branch_entry);
4882 header->size += size;
4885 if (sample_type & PERF_SAMPLE_REGS_USER) {
4886 /* regs dump ABI info */
4887 int size = sizeof(u64);
4889 perf_sample_regs_user(&data->regs_user, regs);
4891 if (data->regs_user.regs) {
4892 u64 mask = event->attr.sample_regs_user;
4893 size += hweight64(mask) * sizeof(u64);
4896 header->size += size;
4899 if (sample_type & PERF_SAMPLE_STACK_USER) {
4901 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
4902 * processed as the last one or have additional check added
4903 * in case new sample type is added, because we could eat
4904 * up the rest of the sample size.
4906 struct perf_regs_user *uregs = &data->regs_user;
4907 u16 stack_size = event->attr.sample_stack_user;
4908 u16 size = sizeof(u64);
4911 perf_sample_regs_user(uregs, regs);
4913 stack_size = perf_sample_ustack_size(stack_size, header->size,
4917 * If there is something to dump, add space for the dump
4918 * itself and for the field that tells the dynamic size,
4919 * which is how many have been actually dumped.
4922 size += sizeof(u64) + stack_size;
4924 data->stack_user_size = stack_size;
4925 header->size += size;
4929 static void perf_event_output(struct perf_event *event,
4930 struct perf_sample_data *data,
4931 struct pt_regs *regs)
4933 struct perf_output_handle handle;
4934 struct perf_event_header header;
4936 /* protect the callchain buffers */
4939 perf_prepare_sample(&header, data, event, regs);
4941 if (perf_output_begin(&handle, event, header.size))
4944 perf_output_sample(&handle, &header, data, event);
4946 perf_output_end(&handle);
4956 struct perf_read_event {
4957 struct perf_event_header header;
4964 perf_event_read_event(struct perf_event *event,
4965 struct task_struct *task)
4967 struct perf_output_handle handle;
4968 struct perf_sample_data sample;
4969 struct perf_read_event read_event = {
4971 .type = PERF_RECORD_READ,
4973 .size = sizeof(read_event) + event->read_size,
4975 .pid = perf_event_pid(event, task),
4976 .tid = perf_event_tid(event, task),
4980 perf_event_header__init_id(&read_event.header, &sample, event);
4981 ret = perf_output_begin(&handle, event, read_event.header.size);
4985 perf_output_put(&handle, read_event);
4986 perf_output_read(&handle, event);
4987 perf_event__output_id_sample(event, &handle, &sample);
4989 perf_output_end(&handle);
4992 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
4995 perf_event_aux_ctx(struct perf_event_context *ctx,
4996 perf_event_aux_output_cb output,
4999 struct perf_event *event;
5001 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5002 if (event->state < PERF_EVENT_STATE_INACTIVE)
5004 if (!event_filter_match(event))
5006 output(event, data);
5011 perf_event_aux(perf_event_aux_output_cb output, void *data,
5012 struct perf_event_context *task_ctx)
5014 struct perf_cpu_context *cpuctx;
5015 struct perf_event_context *ctx;
5020 list_for_each_entry_rcu(pmu, &pmus, entry) {
5021 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5022 if (cpuctx->unique_pmu != pmu)
5024 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5027 ctxn = pmu->task_ctx_nr;
5030 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5032 perf_event_aux_ctx(ctx, output, data);
5034 put_cpu_ptr(pmu->pmu_cpu_context);
5039 perf_event_aux_ctx(task_ctx, output, data);
5046 * task tracking -- fork/exit
5048 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5051 struct perf_task_event {
5052 struct task_struct *task;
5053 struct perf_event_context *task_ctx;
5056 struct perf_event_header header;
5066 static int perf_event_task_match(struct perf_event *event)
5068 return event->attr.comm || event->attr.mmap ||
5069 event->attr.mmap2 || event->attr.mmap_data ||
5073 static void perf_event_task_output(struct perf_event *event,
5076 struct perf_task_event *task_event = data;
5077 struct perf_output_handle handle;
5078 struct perf_sample_data sample;
5079 struct task_struct *task = task_event->task;
5080 int ret, size = task_event->event_id.header.size;
5082 if (!perf_event_task_match(event))
5085 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5087 ret = perf_output_begin(&handle, event,
5088 task_event->event_id.header.size);
5092 task_event->event_id.pid = perf_event_pid(event, task);
5093 task_event->event_id.ppid = perf_event_pid(event, current);
5095 task_event->event_id.tid = perf_event_tid(event, task);
5096 task_event->event_id.ptid = perf_event_tid(event, current);
5098 perf_output_put(&handle, task_event->event_id);
5100 perf_event__output_id_sample(event, &handle, &sample);
5102 perf_output_end(&handle);
5104 task_event->event_id.header.size = size;
5107 static void perf_event_task(struct task_struct *task,
5108 struct perf_event_context *task_ctx,
5111 struct perf_task_event task_event;
5113 if (!atomic_read(&nr_comm_events) &&
5114 !atomic_read(&nr_mmap_events) &&
5115 !atomic_read(&nr_task_events))
5118 task_event = (struct perf_task_event){
5120 .task_ctx = task_ctx,
5123 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5125 .size = sizeof(task_event.event_id),
5131 .time = perf_clock(),
5135 perf_event_aux(perf_event_task_output,
5140 void perf_event_fork(struct task_struct *task)
5142 perf_event_task(task, NULL, 1);
5149 struct perf_comm_event {
5150 struct task_struct *task;
5155 struct perf_event_header header;
5162 static int perf_event_comm_match(struct perf_event *event)
5164 return event->attr.comm;
5167 static void perf_event_comm_output(struct perf_event *event,
5170 struct perf_comm_event *comm_event = data;
5171 struct perf_output_handle handle;
5172 struct perf_sample_data sample;
5173 int size = comm_event->event_id.header.size;
5176 if (!perf_event_comm_match(event))
5179 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5180 ret = perf_output_begin(&handle, event,
5181 comm_event->event_id.header.size);
5186 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5187 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5189 perf_output_put(&handle, comm_event->event_id);
5190 __output_copy(&handle, comm_event->comm,
5191 comm_event->comm_size);
5193 perf_event__output_id_sample(event, &handle, &sample);
5195 perf_output_end(&handle);
5197 comm_event->event_id.header.size = size;
5200 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5202 char comm[TASK_COMM_LEN];
5205 memset(comm, 0, sizeof(comm));
5206 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5207 size = ALIGN(strlen(comm)+1, sizeof(u64));
5209 comm_event->comm = comm;
5210 comm_event->comm_size = size;
5212 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5214 perf_event_aux(perf_event_comm_output,
5219 void perf_event_comm(struct task_struct *task, bool exec)
5221 struct perf_comm_event comm_event;
5223 if (!atomic_read(&nr_comm_events))
5226 comm_event = (struct perf_comm_event){
5232 .type = PERF_RECORD_COMM,
5233 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5241 perf_event_comm_event(&comm_event);
5248 struct perf_mmap_event {
5249 struct vm_area_struct *vma;
5251 const char *file_name;
5259 struct perf_event_header header;
5269 static int perf_event_mmap_match(struct perf_event *event,
5272 struct perf_mmap_event *mmap_event = data;
5273 struct vm_area_struct *vma = mmap_event->vma;
5274 int executable = vma->vm_flags & VM_EXEC;
5276 return (!executable && event->attr.mmap_data) ||
5277 (executable && (event->attr.mmap || event->attr.mmap2));
5280 static void perf_event_mmap_output(struct perf_event *event,
5283 struct perf_mmap_event *mmap_event = data;
5284 struct perf_output_handle handle;
5285 struct perf_sample_data sample;
5286 int size = mmap_event->event_id.header.size;
5289 if (!perf_event_mmap_match(event, data))
5292 if (event->attr.mmap2) {
5293 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5294 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5295 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5296 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5297 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5298 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5299 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5302 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5303 ret = perf_output_begin(&handle, event,
5304 mmap_event->event_id.header.size);
5308 mmap_event->event_id.pid = perf_event_pid(event, current);
5309 mmap_event->event_id.tid = perf_event_tid(event, current);
5311 perf_output_put(&handle, mmap_event->event_id);
5313 if (event->attr.mmap2) {
5314 perf_output_put(&handle, mmap_event->maj);
5315 perf_output_put(&handle, mmap_event->min);
5316 perf_output_put(&handle, mmap_event->ino);
5317 perf_output_put(&handle, mmap_event->ino_generation);
5318 perf_output_put(&handle, mmap_event->prot);
5319 perf_output_put(&handle, mmap_event->flags);
5322 __output_copy(&handle, mmap_event->file_name,
5323 mmap_event->file_size);
5325 perf_event__output_id_sample(event, &handle, &sample);
5327 perf_output_end(&handle);
5329 mmap_event->event_id.header.size = size;
5332 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5334 struct vm_area_struct *vma = mmap_event->vma;
5335 struct file *file = vma->vm_file;
5336 int maj = 0, min = 0;
5337 u64 ino = 0, gen = 0;
5338 u32 prot = 0, flags = 0;
5345 struct inode *inode;
5348 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5354 * d_path() works from the end of the rb backwards, so we
5355 * need to add enough zero bytes after the string to handle
5356 * the 64bit alignment we do later.
5358 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5363 inode = file_inode(vma->vm_file);
5364 dev = inode->i_sb->s_dev;
5366 gen = inode->i_generation;
5370 if (vma->vm_flags & VM_READ)
5372 if (vma->vm_flags & VM_WRITE)
5374 if (vma->vm_flags & VM_EXEC)
5377 if (vma->vm_flags & VM_MAYSHARE)
5380 flags = MAP_PRIVATE;
5382 if (vma->vm_flags & VM_DENYWRITE)
5383 flags |= MAP_DENYWRITE;
5384 if (vma->vm_flags & VM_MAYEXEC)
5385 flags |= MAP_EXECUTABLE;
5386 if (vma->vm_flags & VM_LOCKED)
5387 flags |= MAP_LOCKED;
5388 if (vma->vm_flags & VM_HUGETLB)
5389 flags |= MAP_HUGETLB;
5393 if (vma->vm_ops && vma->vm_ops->name) {
5394 name = (char *) vma->vm_ops->name(vma);
5399 name = (char *)arch_vma_name(vma);
5403 if (vma->vm_start <= vma->vm_mm->start_brk &&
5404 vma->vm_end >= vma->vm_mm->brk) {
5408 if (vma->vm_start <= vma->vm_mm->start_stack &&
5409 vma->vm_end >= vma->vm_mm->start_stack) {
5419 strlcpy(tmp, name, sizeof(tmp));
5423 * Since our buffer works in 8 byte units we need to align our string
5424 * size to a multiple of 8. However, we must guarantee the tail end is
5425 * zero'd out to avoid leaking random bits to userspace.
5427 size = strlen(name)+1;
5428 while (!IS_ALIGNED(size, sizeof(u64)))
5429 name[size++] = '\0';
5431 mmap_event->file_name = name;
5432 mmap_event->file_size = size;
5433 mmap_event->maj = maj;
5434 mmap_event->min = min;
5435 mmap_event->ino = ino;
5436 mmap_event->ino_generation = gen;
5437 mmap_event->prot = prot;
5438 mmap_event->flags = flags;
5440 if (!(vma->vm_flags & VM_EXEC))
5441 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5443 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5445 perf_event_aux(perf_event_mmap_output,
5452 void perf_event_mmap(struct vm_area_struct *vma)
5454 struct perf_mmap_event mmap_event;
5456 if (!atomic_read(&nr_mmap_events))
5459 mmap_event = (struct perf_mmap_event){
5465 .type = PERF_RECORD_MMAP,
5466 .misc = PERF_RECORD_MISC_USER,
5471 .start = vma->vm_start,
5472 .len = vma->vm_end - vma->vm_start,
5473 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5475 /* .maj (attr_mmap2 only) */
5476 /* .min (attr_mmap2 only) */
5477 /* .ino (attr_mmap2 only) */
5478 /* .ino_generation (attr_mmap2 only) */
5479 /* .prot (attr_mmap2 only) */
5480 /* .flags (attr_mmap2 only) */
5483 perf_event_mmap_event(&mmap_event);
5487 * IRQ throttle logging
5490 static void perf_log_throttle(struct perf_event *event, int enable)
5492 struct perf_output_handle handle;
5493 struct perf_sample_data sample;
5497 struct perf_event_header header;
5501 } throttle_event = {
5503 .type = PERF_RECORD_THROTTLE,
5505 .size = sizeof(throttle_event),
5507 .time = perf_clock(),
5508 .id = primary_event_id(event),
5509 .stream_id = event->id,
5513 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5515 perf_event_header__init_id(&throttle_event.header, &sample, event);
5517 ret = perf_output_begin(&handle, event,
5518 throttle_event.header.size);
5522 perf_output_put(&handle, throttle_event);
5523 perf_event__output_id_sample(event, &handle, &sample);
5524 perf_output_end(&handle);
5528 * Generic event overflow handling, sampling.
5531 static int __perf_event_overflow(struct perf_event *event,
5532 int throttle, struct perf_sample_data *data,
5533 struct pt_regs *regs)
5535 int events = atomic_read(&event->event_limit);
5536 struct hw_perf_event *hwc = &event->hw;
5541 * Non-sampling counters might still use the PMI to fold short
5542 * hardware counters, ignore those.
5544 if (unlikely(!is_sampling_event(event)))
5547 seq = __this_cpu_read(perf_throttled_seq);
5548 if (seq != hwc->interrupts_seq) {
5549 hwc->interrupts_seq = seq;
5550 hwc->interrupts = 1;
5553 if (unlikely(throttle
5554 && hwc->interrupts >= max_samples_per_tick)) {
5555 __this_cpu_inc(perf_throttled_count);
5556 hwc->interrupts = MAX_INTERRUPTS;
5557 perf_log_throttle(event, 0);
5558 tick_nohz_full_kick();
5563 if (event->attr.freq) {
5564 u64 now = perf_clock();
5565 s64 delta = now - hwc->freq_time_stamp;
5567 hwc->freq_time_stamp = now;
5569 if (delta > 0 && delta < 2*TICK_NSEC)
5570 perf_adjust_period(event, delta, hwc->last_period, true);
5574 * XXX event_limit might not quite work as expected on inherited
5578 event->pending_kill = POLL_IN;
5579 if (events && atomic_dec_and_test(&event->event_limit)) {
5581 event->pending_kill = POLL_HUP;
5582 event->pending_disable = 1;
5583 irq_work_queue(&event->pending);
5586 if (event->overflow_handler)
5587 event->overflow_handler(event, data, regs);
5589 perf_event_output(event, data, regs);
5591 if (event->fasync && event->pending_kill) {
5592 event->pending_wakeup = 1;
5593 irq_work_queue(&event->pending);
5599 int perf_event_overflow(struct perf_event *event,
5600 struct perf_sample_data *data,
5601 struct pt_regs *regs)
5603 return __perf_event_overflow(event, 1, data, regs);
5607 * Generic software event infrastructure
5610 struct swevent_htable {
5611 struct swevent_hlist *swevent_hlist;
5612 struct mutex hlist_mutex;
5615 /* Recursion avoidance in each contexts */
5616 int recursion[PERF_NR_CONTEXTS];
5618 /* Keeps track of cpu being initialized/exited */
5622 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5625 * We directly increment event->count and keep a second value in
5626 * event->hw.period_left to count intervals. This period event
5627 * is kept in the range [-sample_period, 0] so that we can use the
5631 u64 perf_swevent_set_period(struct perf_event *event)
5633 struct hw_perf_event *hwc = &event->hw;
5634 u64 period = hwc->last_period;
5638 hwc->last_period = hwc->sample_period;
5641 old = val = local64_read(&hwc->period_left);
5645 nr = div64_u64(period + val, period);
5646 offset = nr * period;
5648 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5654 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5655 struct perf_sample_data *data,
5656 struct pt_regs *regs)
5658 struct hw_perf_event *hwc = &event->hw;
5662 overflow = perf_swevent_set_period(event);
5664 if (hwc->interrupts == MAX_INTERRUPTS)
5667 for (; overflow; overflow--) {
5668 if (__perf_event_overflow(event, throttle,
5671 * We inhibit the overflow from happening when
5672 * hwc->interrupts == MAX_INTERRUPTS.
5680 static void perf_swevent_event(struct perf_event *event, u64 nr,
5681 struct perf_sample_data *data,
5682 struct pt_regs *regs)
5684 struct hw_perf_event *hwc = &event->hw;
5686 local64_add(nr, &event->count);
5691 if (!is_sampling_event(event))
5694 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5696 return perf_swevent_overflow(event, 1, data, regs);
5698 data->period = event->hw.last_period;
5700 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5701 return perf_swevent_overflow(event, 1, data, regs);
5703 if (local64_add_negative(nr, &hwc->period_left))
5706 perf_swevent_overflow(event, 0, data, regs);
5709 static int perf_exclude_event(struct perf_event *event,
5710 struct pt_regs *regs)
5712 if (event->hw.state & PERF_HES_STOPPED)
5716 if (event->attr.exclude_user && user_mode(regs))
5719 if (event->attr.exclude_kernel && !user_mode(regs))
5726 static int perf_swevent_match(struct perf_event *event,
5727 enum perf_type_id type,
5729 struct perf_sample_data *data,
5730 struct pt_regs *regs)
5732 if (event->attr.type != type)
5735 if (event->attr.config != event_id)
5738 if (perf_exclude_event(event, regs))
5744 static inline u64 swevent_hash(u64 type, u32 event_id)
5746 u64 val = event_id | (type << 32);
5748 return hash_64(val, SWEVENT_HLIST_BITS);
5751 static inline struct hlist_head *
5752 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5754 u64 hash = swevent_hash(type, event_id);
5756 return &hlist->heads[hash];
5759 /* For the read side: events when they trigger */
5760 static inline struct hlist_head *
5761 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5763 struct swevent_hlist *hlist;
5765 hlist = rcu_dereference(swhash->swevent_hlist);
5769 return __find_swevent_head(hlist, type, event_id);
5772 /* For the event head insertion and removal in the hlist */
5773 static inline struct hlist_head *
5774 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
5776 struct swevent_hlist *hlist;
5777 u32 event_id = event->attr.config;
5778 u64 type = event->attr.type;
5781 * Event scheduling is always serialized against hlist allocation
5782 * and release. Which makes the protected version suitable here.
5783 * The context lock guarantees that.
5785 hlist = rcu_dereference_protected(swhash->swevent_hlist,
5786 lockdep_is_held(&event->ctx->lock));
5790 return __find_swevent_head(hlist, type, event_id);
5793 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
5795 struct perf_sample_data *data,
5796 struct pt_regs *regs)
5798 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5799 struct perf_event *event;
5800 struct hlist_head *head;
5803 head = find_swevent_head_rcu(swhash, type, event_id);
5807 hlist_for_each_entry_rcu(event, head, hlist_entry) {
5808 if (perf_swevent_match(event, type, event_id, data, regs))
5809 perf_swevent_event(event, nr, data, regs);
5815 int perf_swevent_get_recursion_context(void)
5817 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5819 return get_recursion_context(swhash->recursion);
5821 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
5823 inline void perf_swevent_put_recursion_context(int rctx)
5825 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5827 put_recursion_context(swhash->recursion, rctx);
5830 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
5832 struct perf_sample_data data;
5835 preempt_disable_notrace();
5836 rctx = perf_swevent_get_recursion_context();
5840 perf_sample_data_init(&data, addr, 0);
5842 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
5844 perf_swevent_put_recursion_context(rctx);
5845 preempt_enable_notrace();
5848 static void perf_swevent_read(struct perf_event *event)
5852 static int perf_swevent_add(struct perf_event *event, int flags)
5854 struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
5855 struct hw_perf_event *hwc = &event->hw;
5856 struct hlist_head *head;
5858 if (is_sampling_event(event)) {
5859 hwc->last_period = hwc->sample_period;
5860 perf_swevent_set_period(event);
5863 hwc->state = !(flags & PERF_EF_START);
5865 head = find_swevent_head(swhash, event);
5868 * We can race with cpu hotplug code. Do not
5869 * WARN if the cpu just got unplugged.
5871 WARN_ON_ONCE(swhash->online);
5875 hlist_add_head_rcu(&event->hlist_entry, head);
5880 static void perf_swevent_del(struct perf_event *event, int flags)
5882 hlist_del_rcu(&event->hlist_entry);
5885 static void perf_swevent_start(struct perf_event *event, int flags)
5887 event->hw.state = 0;
5890 static void perf_swevent_stop(struct perf_event *event, int flags)
5892 event->hw.state = PERF_HES_STOPPED;
5895 /* Deref the hlist from the update side */
5896 static inline struct swevent_hlist *
5897 swevent_hlist_deref(struct swevent_htable *swhash)
5899 return rcu_dereference_protected(swhash->swevent_hlist,
5900 lockdep_is_held(&swhash->hlist_mutex));
5903 static void swevent_hlist_release(struct swevent_htable *swhash)
5905 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
5910 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
5911 kfree_rcu(hlist, rcu_head);
5914 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
5916 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5918 mutex_lock(&swhash->hlist_mutex);
5920 if (!--swhash->hlist_refcount)
5921 swevent_hlist_release(swhash);
5923 mutex_unlock(&swhash->hlist_mutex);
5926 static void swevent_hlist_put(struct perf_event *event)
5930 for_each_possible_cpu(cpu)
5931 swevent_hlist_put_cpu(event, cpu);
5934 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
5936 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
5939 mutex_lock(&swhash->hlist_mutex);
5941 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
5942 struct swevent_hlist *hlist;
5944 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
5949 rcu_assign_pointer(swhash->swevent_hlist, hlist);
5951 swhash->hlist_refcount++;
5953 mutex_unlock(&swhash->hlist_mutex);
5958 static int swevent_hlist_get(struct perf_event *event)
5961 int cpu, failed_cpu;
5964 for_each_possible_cpu(cpu) {
5965 err = swevent_hlist_get_cpu(event, cpu);
5975 for_each_possible_cpu(cpu) {
5976 if (cpu == failed_cpu)
5978 swevent_hlist_put_cpu(event, cpu);
5985 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
5987 static void sw_perf_event_destroy(struct perf_event *event)
5989 u64 event_id = event->attr.config;
5991 WARN_ON(event->parent);
5993 static_key_slow_dec(&perf_swevent_enabled[event_id]);
5994 swevent_hlist_put(event);
5997 static int perf_swevent_init(struct perf_event *event)
5999 u64 event_id = event->attr.config;
6001 if (event->attr.type != PERF_TYPE_SOFTWARE)
6005 * no branch sampling for software events
6007 if (has_branch_stack(event))
6011 case PERF_COUNT_SW_CPU_CLOCK:
6012 case PERF_COUNT_SW_TASK_CLOCK:
6019 if (event_id >= PERF_COUNT_SW_MAX)
6022 if (!event->parent) {
6025 err = swevent_hlist_get(event);
6029 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6030 event->destroy = sw_perf_event_destroy;
6036 static int perf_swevent_event_idx(struct perf_event *event)
6041 static struct pmu perf_swevent = {
6042 .task_ctx_nr = perf_sw_context,
6044 .event_init = perf_swevent_init,
6045 .add = perf_swevent_add,
6046 .del = perf_swevent_del,
6047 .start = perf_swevent_start,
6048 .stop = perf_swevent_stop,
6049 .read = perf_swevent_read,
6051 .event_idx = perf_swevent_event_idx,
6054 #ifdef CONFIG_EVENT_TRACING
6056 static int perf_tp_filter_match(struct perf_event *event,
6057 struct perf_sample_data *data)
6059 void *record = data->raw->data;
6061 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6066 static int perf_tp_event_match(struct perf_event *event,
6067 struct perf_sample_data *data,
6068 struct pt_regs *regs)
6070 if (event->hw.state & PERF_HES_STOPPED)
6073 * All tracepoints are from kernel-space.
6075 if (event->attr.exclude_kernel)
6078 if (!perf_tp_filter_match(event, data))
6084 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6085 struct pt_regs *regs, struct hlist_head *head, int rctx,
6086 struct task_struct *task)
6088 struct perf_sample_data data;
6089 struct perf_event *event;
6091 struct perf_raw_record raw = {
6096 perf_sample_data_init(&data, addr, 0);
6099 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6100 if (perf_tp_event_match(event, &data, regs))
6101 perf_swevent_event(event, count, &data, regs);
6105 * If we got specified a target task, also iterate its context and
6106 * deliver this event there too.
6108 if (task && task != current) {
6109 struct perf_event_context *ctx;
6110 struct trace_entry *entry = record;
6113 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6117 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6118 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6120 if (event->attr.config != entry->type)
6122 if (perf_tp_event_match(event, &data, regs))
6123 perf_swevent_event(event, count, &data, regs);
6129 perf_swevent_put_recursion_context(rctx);
6131 EXPORT_SYMBOL_GPL(perf_tp_event);
6133 static void tp_perf_event_destroy(struct perf_event *event)
6135 perf_trace_destroy(event);
6138 static int perf_tp_event_init(struct perf_event *event)
6142 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6146 * no branch sampling for tracepoint events
6148 if (has_branch_stack(event))
6151 err = perf_trace_init(event);
6155 event->destroy = tp_perf_event_destroy;
6160 static struct pmu perf_tracepoint = {
6161 .task_ctx_nr = perf_sw_context,
6163 .event_init = perf_tp_event_init,
6164 .add = perf_trace_add,
6165 .del = perf_trace_del,
6166 .start = perf_swevent_start,
6167 .stop = perf_swevent_stop,
6168 .read = perf_swevent_read,
6170 .event_idx = perf_swevent_event_idx,
6173 static inline void perf_tp_register(void)
6175 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6178 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6183 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6186 filter_str = strndup_user(arg, PAGE_SIZE);
6187 if (IS_ERR(filter_str))
6188 return PTR_ERR(filter_str);
6190 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6196 static void perf_event_free_filter(struct perf_event *event)
6198 ftrace_profile_free_filter(event);
6203 static inline void perf_tp_register(void)
6207 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6212 static void perf_event_free_filter(struct perf_event *event)
6216 #endif /* CONFIG_EVENT_TRACING */
6218 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6219 void perf_bp_event(struct perf_event *bp, void *data)
6221 struct perf_sample_data sample;
6222 struct pt_regs *regs = data;
6224 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6226 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6227 perf_swevent_event(bp, 1, &sample, regs);
6232 * hrtimer based swevent callback
6235 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6237 enum hrtimer_restart ret = HRTIMER_RESTART;
6238 struct perf_sample_data data;
6239 struct pt_regs *regs;
6240 struct perf_event *event;
6243 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6245 if (event->state != PERF_EVENT_STATE_ACTIVE)
6246 return HRTIMER_NORESTART;
6248 event->pmu->read(event);
6250 perf_sample_data_init(&data, 0, event->hw.last_period);
6251 regs = get_irq_regs();
6253 if (regs && !perf_exclude_event(event, regs)) {
6254 if (!(event->attr.exclude_idle && is_idle_task(current)))
6255 if (__perf_event_overflow(event, 1, &data, regs))
6256 ret = HRTIMER_NORESTART;
6259 period = max_t(u64, 10000, event->hw.sample_period);
6260 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6265 static void perf_swevent_start_hrtimer(struct perf_event *event)
6267 struct hw_perf_event *hwc = &event->hw;
6270 if (!is_sampling_event(event))
6273 period = local64_read(&hwc->period_left);
6278 local64_set(&hwc->period_left, 0);
6280 period = max_t(u64, 10000, hwc->sample_period);
6282 __hrtimer_start_range_ns(&hwc->hrtimer,
6283 ns_to_ktime(period), 0,
6284 HRTIMER_MODE_REL_PINNED, 0);
6287 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6289 struct hw_perf_event *hwc = &event->hw;
6291 if (is_sampling_event(event)) {
6292 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6293 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6295 hrtimer_cancel(&hwc->hrtimer);
6299 static void perf_swevent_init_hrtimer(struct perf_event *event)
6301 struct hw_perf_event *hwc = &event->hw;
6303 if (!is_sampling_event(event))
6306 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6307 hwc->hrtimer.function = perf_swevent_hrtimer;
6310 * Since hrtimers have a fixed rate, we can do a static freq->period
6311 * mapping and avoid the whole period adjust feedback stuff.
6313 if (event->attr.freq) {
6314 long freq = event->attr.sample_freq;
6316 event->attr.sample_period = NSEC_PER_SEC / freq;
6317 hwc->sample_period = event->attr.sample_period;
6318 local64_set(&hwc->period_left, hwc->sample_period);
6319 hwc->last_period = hwc->sample_period;
6320 event->attr.freq = 0;
6325 * Software event: cpu wall time clock
6328 static void cpu_clock_event_update(struct perf_event *event)
6333 now = local_clock();
6334 prev = local64_xchg(&event->hw.prev_count, now);
6335 local64_add(now - prev, &event->count);
6338 static void cpu_clock_event_start(struct perf_event *event, int flags)
6340 local64_set(&event->hw.prev_count, local_clock());
6341 perf_swevent_start_hrtimer(event);
6344 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6346 perf_swevent_cancel_hrtimer(event);
6347 cpu_clock_event_update(event);
6350 static int cpu_clock_event_add(struct perf_event *event, int flags)
6352 if (flags & PERF_EF_START)
6353 cpu_clock_event_start(event, flags);
6358 static void cpu_clock_event_del(struct perf_event *event, int flags)
6360 cpu_clock_event_stop(event, flags);
6363 static void cpu_clock_event_read(struct perf_event *event)
6365 cpu_clock_event_update(event);
6368 static int cpu_clock_event_init(struct perf_event *event)
6370 if (event->attr.type != PERF_TYPE_SOFTWARE)
6373 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6377 * no branch sampling for software events
6379 if (has_branch_stack(event))
6382 perf_swevent_init_hrtimer(event);
6387 static struct pmu perf_cpu_clock = {
6388 .task_ctx_nr = perf_sw_context,
6390 .event_init = cpu_clock_event_init,
6391 .add = cpu_clock_event_add,
6392 .del = cpu_clock_event_del,
6393 .start = cpu_clock_event_start,
6394 .stop = cpu_clock_event_stop,
6395 .read = cpu_clock_event_read,
6397 .event_idx = perf_swevent_event_idx,
6401 * Software event: task time clock
6404 static void task_clock_event_update(struct perf_event *event, u64 now)
6409 prev = local64_xchg(&event->hw.prev_count, now);
6411 local64_add(delta, &event->count);
6414 static void task_clock_event_start(struct perf_event *event, int flags)
6416 local64_set(&event->hw.prev_count, event->ctx->time);
6417 perf_swevent_start_hrtimer(event);
6420 static void task_clock_event_stop(struct perf_event *event, int flags)
6422 perf_swevent_cancel_hrtimer(event);
6423 task_clock_event_update(event, event->ctx->time);
6426 static int task_clock_event_add(struct perf_event *event, int flags)
6428 if (flags & PERF_EF_START)
6429 task_clock_event_start(event, flags);
6434 static void task_clock_event_del(struct perf_event *event, int flags)
6436 task_clock_event_stop(event, PERF_EF_UPDATE);
6439 static void task_clock_event_read(struct perf_event *event)
6441 u64 now = perf_clock();
6442 u64 delta = now - event->ctx->timestamp;
6443 u64 time = event->ctx->time + delta;
6445 task_clock_event_update(event, time);
6448 static int task_clock_event_init(struct perf_event *event)
6450 if (event->attr.type != PERF_TYPE_SOFTWARE)
6453 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6457 * no branch sampling for software events
6459 if (has_branch_stack(event))
6462 perf_swevent_init_hrtimer(event);
6467 static struct pmu perf_task_clock = {
6468 .task_ctx_nr = perf_sw_context,
6470 .event_init = task_clock_event_init,
6471 .add = task_clock_event_add,
6472 .del = task_clock_event_del,
6473 .start = task_clock_event_start,
6474 .stop = task_clock_event_stop,
6475 .read = task_clock_event_read,
6477 .event_idx = perf_swevent_event_idx,
6480 static void perf_pmu_nop_void(struct pmu *pmu)
6484 static int perf_pmu_nop_int(struct pmu *pmu)
6489 static void perf_pmu_start_txn(struct pmu *pmu)
6491 perf_pmu_disable(pmu);
6494 static int perf_pmu_commit_txn(struct pmu *pmu)
6496 perf_pmu_enable(pmu);
6500 static void perf_pmu_cancel_txn(struct pmu *pmu)
6502 perf_pmu_enable(pmu);
6505 static int perf_event_idx_default(struct perf_event *event)
6507 return event->hw.idx + 1;
6511 * Ensures all contexts with the same task_ctx_nr have the same
6512 * pmu_cpu_context too.
6514 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6521 list_for_each_entry(pmu, &pmus, entry) {
6522 if (pmu->task_ctx_nr == ctxn)
6523 return pmu->pmu_cpu_context;
6529 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6533 for_each_possible_cpu(cpu) {
6534 struct perf_cpu_context *cpuctx;
6536 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6538 if (cpuctx->unique_pmu == old_pmu)
6539 cpuctx->unique_pmu = pmu;
6543 static void free_pmu_context(struct pmu *pmu)
6547 mutex_lock(&pmus_lock);
6549 * Like a real lame refcount.
6551 list_for_each_entry(i, &pmus, entry) {
6552 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6553 update_pmu_context(i, pmu);
6558 free_percpu(pmu->pmu_cpu_context);
6560 mutex_unlock(&pmus_lock);
6562 static struct idr pmu_idr;
6565 type_show(struct device *dev, struct device_attribute *attr, char *page)
6567 struct pmu *pmu = dev_get_drvdata(dev);
6569 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6571 static DEVICE_ATTR_RO(type);
6574 perf_event_mux_interval_ms_show(struct device *dev,
6575 struct device_attribute *attr,
6578 struct pmu *pmu = dev_get_drvdata(dev);
6580 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6584 perf_event_mux_interval_ms_store(struct device *dev,
6585 struct device_attribute *attr,
6586 const char *buf, size_t count)
6588 struct pmu *pmu = dev_get_drvdata(dev);
6589 int timer, cpu, ret;
6591 ret = kstrtoint(buf, 0, &timer);
6598 /* same value, noting to do */
6599 if (timer == pmu->hrtimer_interval_ms)
6602 pmu->hrtimer_interval_ms = timer;
6604 /* update all cpuctx for this PMU */
6605 for_each_possible_cpu(cpu) {
6606 struct perf_cpu_context *cpuctx;
6607 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6608 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6610 if (hrtimer_active(&cpuctx->hrtimer))
6611 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6616 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6618 static struct attribute *pmu_dev_attrs[] = {
6619 &dev_attr_type.attr,
6620 &dev_attr_perf_event_mux_interval_ms.attr,
6623 ATTRIBUTE_GROUPS(pmu_dev);
6625 static int pmu_bus_running;
6626 static struct bus_type pmu_bus = {
6627 .name = "event_source",
6628 .dev_groups = pmu_dev_groups,
6631 static void pmu_dev_release(struct device *dev)
6636 static int pmu_dev_alloc(struct pmu *pmu)
6640 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6644 pmu->dev->groups = pmu->attr_groups;
6645 device_initialize(pmu->dev);
6646 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6650 dev_set_drvdata(pmu->dev, pmu);
6651 pmu->dev->bus = &pmu_bus;
6652 pmu->dev->release = pmu_dev_release;
6653 ret = device_add(pmu->dev);
6661 put_device(pmu->dev);
6665 static struct lock_class_key cpuctx_mutex;
6666 static struct lock_class_key cpuctx_lock;
6668 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6672 mutex_lock(&pmus_lock);
6674 pmu->pmu_disable_count = alloc_percpu(int);
6675 if (!pmu->pmu_disable_count)
6684 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6692 if (pmu_bus_running) {
6693 ret = pmu_dev_alloc(pmu);
6699 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6700 if (pmu->pmu_cpu_context)
6701 goto got_cpu_context;
6704 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6705 if (!pmu->pmu_cpu_context)
6708 for_each_possible_cpu(cpu) {
6709 struct perf_cpu_context *cpuctx;
6711 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6712 __perf_event_init_context(&cpuctx->ctx);
6713 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6714 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6715 cpuctx->ctx.type = cpu_context;
6716 cpuctx->ctx.pmu = pmu;
6718 __perf_cpu_hrtimer_init(cpuctx, cpu);
6720 INIT_LIST_HEAD(&cpuctx->rotation_list);
6721 cpuctx->unique_pmu = pmu;
6725 if (!pmu->start_txn) {
6726 if (pmu->pmu_enable) {
6728 * If we have pmu_enable/pmu_disable calls, install
6729 * transaction stubs that use that to try and batch
6730 * hardware accesses.
6732 pmu->start_txn = perf_pmu_start_txn;
6733 pmu->commit_txn = perf_pmu_commit_txn;
6734 pmu->cancel_txn = perf_pmu_cancel_txn;
6736 pmu->start_txn = perf_pmu_nop_void;
6737 pmu->commit_txn = perf_pmu_nop_int;
6738 pmu->cancel_txn = perf_pmu_nop_void;
6742 if (!pmu->pmu_enable) {
6743 pmu->pmu_enable = perf_pmu_nop_void;
6744 pmu->pmu_disable = perf_pmu_nop_void;
6747 if (!pmu->event_idx)
6748 pmu->event_idx = perf_event_idx_default;
6750 list_add_rcu(&pmu->entry, &pmus);
6753 mutex_unlock(&pmus_lock);
6758 device_del(pmu->dev);
6759 put_device(pmu->dev);
6762 if (pmu->type >= PERF_TYPE_MAX)
6763 idr_remove(&pmu_idr, pmu->type);
6766 free_percpu(pmu->pmu_disable_count);
6769 EXPORT_SYMBOL_GPL(perf_pmu_register);
6771 void perf_pmu_unregister(struct pmu *pmu)
6773 mutex_lock(&pmus_lock);
6774 list_del_rcu(&pmu->entry);
6775 mutex_unlock(&pmus_lock);
6778 * We dereference the pmu list under both SRCU and regular RCU, so
6779 * synchronize against both of those.
6781 synchronize_srcu(&pmus_srcu);
6784 free_percpu(pmu->pmu_disable_count);
6785 if (pmu->type >= PERF_TYPE_MAX)
6786 idr_remove(&pmu_idr, pmu->type);
6787 device_del(pmu->dev);
6788 put_device(pmu->dev);
6789 free_pmu_context(pmu);
6791 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
6793 struct pmu *perf_init_event(struct perf_event *event)
6795 struct pmu *pmu = NULL;
6799 idx = srcu_read_lock(&pmus_srcu);
6802 pmu = idr_find(&pmu_idr, event->attr.type);
6805 if (!try_module_get(pmu->module)) {
6806 pmu = ERR_PTR(-ENODEV);
6810 ret = pmu->event_init(event);
6816 list_for_each_entry_rcu(pmu, &pmus, entry) {
6817 if (!try_module_get(pmu->module)) {
6818 pmu = ERR_PTR(-ENODEV);
6822 ret = pmu->event_init(event);
6826 if (ret != -ENOENT) {
6831 pmu = ERR_PTR(-ENOENT);
6833 srcu_read_unlock(&pmus_srcu, idx);
6838 static void account_event_cpu(struct perf_event *event, int cpu)
6843 if (has_branch_stack(event)) {
6844 if (!(event->attach_state & PERF_ATTACH_TASK))
6845 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
6847 if (is_cgroup_event(event))
6848 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
6851 static void account_event(struct perf_event *event)
6856 if (event->attach_state & PERF_ATTACH_TASK)
6857 static_key_slow_inc(&perf_sched_events.key);
6858 if (event->attr.mmap || event->attr.mmap_data)
6859 atomic_inc(&nr_mmap_events);
6860 if (event->attr.comm)
6861 atomic_inc(&nr_comm_events);
6862 if (event->attr.task)
6863 atomic_inc(&nr_task_events);
6864 if (event->attr.freq) {
6865 if (atomic_inc_return(&nr_freq_events) == 1)
6866 tick_nohz_full_kick_all();
6868 if (has_branch_stack(event))
6869 static_key_slow_inc(&perf_sched_events.key);
6870 if (is_cgroup_event(event))
6871 static_key_slow_inc(&perf_sched_events.key);
6873 account_event_cpu(event, event->cpu);
6877 * Allocate and initialize a event structure
6879 static struct perf_event *
6880 perf_event_alloc(struct perf_event_attr *attr, int cpu,
6881 struct task_struct *task,
6882 struct perf_event *group_leader,
6883 struct perf_event *parent_event,
6884 perf_overflow_handler_t overflow_handler,
6888 struct perf_event *event;
6889 struct hw_perf_event *hwc;
6892 if ((unsigned)cpu >= nr_cpu_ids) {
6893 if (!task || cpu != -1)
6894 return ERR_PTR(-EINVAL);
6897 event = kzalloc(sizeof(*event), GFP_KERNEL);
6899 return ERR_PTR(-ENOMEM);
6902 * Single events are their own group leaders, with an
6903 * empty sibling list:
6906 group_leader = event;
6908 mutex_init(&event->child_mutex);
6909 INIT_LIST_HEAD(&event->child_list);
6911 INIT_LIST_HEAD(&event->group_entry);
6912 INIT_LIST_HEAD(&event->event_entry);
6913 INIT_LIST_HEAD(&event->sibling_list);
6914 INIT_LIST_HEAD(&event->rb_entry);
6915 INIT_LIST_HEAD(&event->active_entry);
6916 INIT_HLIST_NODE(&event->hlist_entry);
6919 init_waitqueue_head(&event->waitq);
6920 init_irq_work(&event->pending, perf_pending_event);
6922 mutex_init(&event->mmap_mutex);
6924 atomic_long_set(&event->refcount, 1);
6926 event->attr = *attr;
6927 event->group_leader = group_leader;
6931 event->parent = parent_event;
6933 event->ns = get_pid_ns(task_active_pid_ns(current));
6934 event->id = atomic64_inc_return(&perf_event_id);
6936 event->state = PERF_EVENT_STATE_INACTIVE;
6939 event->attach_state = PERF_ATTACH_TASK;
6941 if (attr->type == PERF_TYPE_TRACEPOINT)
6942 event->hw.tp_target = task;
6943 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6945 * hw_breakpoint is a bit difficult here..
6947 else if (attr->type == PERF_TYPE_BREAKPOINT)
6948 event->hw.bp_target = task;
6952 if (!overflow_handler && parent_event) {
6953 overflow_handler = parent_event->overflow_handler;
6954 context = parent_event->overflow_handler_context;
6957 event->overflow_handler = overflow_handler;
6958 event->overflow_handler_context = context;
6960 perf_event__state_init(event);
6965 hwc->sample_period = attr->sample_period;
6966 if (attr->freq && attr->sample_freq)
6967 hwc->sample_period = 1;
6968 hwc->last_period = hwc->sample_period;
6970 local64_set(&hwc->period_left, hwc->sample_period);
6973 * we currently do not support PERF_FORMAT_GROUP on inherited events
6975 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
6978 pmu = perf_init_event(event);
6981 else if (IS_ERR(pmu)) {
6986 if (!event->parent) {
6987 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
6988 err = get_callchain_buffers();
6998 event->destroy(event);
6999 module_put(pmu->module);
7002 put_pid_ns(event->ns);
7005 return ERR_PTR(err);
7008 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7009 struct perf_event_attr *attr)
7014 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7018 * zero the full structure, so that a short copy will be nice.
7020 memset(attr, 0, sizeof(*attr));
7022 ret = get_user(size, &uattr->size);
7026 if (size > PAGE_SIZE) /* silly large */
7029 if (!size) /* abi compat */
7030 size = PERF_ATTR_SIZE_VER0;
7032 if (size < PERF_ATTR_SIZE_VER0)
7036 * If we're handed a bigger struct than we know of,
7037 * ensure all the unknown bits are 0 - i.e. new
7038 * user-space does not rely on any kernel feature
7039 * extensions we dont know about yet.
7041 if (size > sizeof(*attr)) {
7042 unsigned char __user *addr;
7043 unsigned char __user *end;
7046 addr = (void __user *)uattr + sizeof(*attr);
7047 end = (void __user *)uattr + size;
7049 for (; addr < end; addr++) {
7050 ret = get_user(val, addr);
7056 size = sizeof(*attr);
7059 ret = copy_from_user(attr, uattr, size);
7063 if (attr->__reserved_1)
7066 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7069 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7072 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7073 u64 mask = attr->branch_sample_type;
7075 /* only using defined bits */
7076 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7079 /* at least one branch bit must be set */
7080 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7083 /* propagate priv level, when not set for branch */
7084 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7086 /* exclude_kernel checked on syscall entry */
7087 if (!attr->exclude_kernel)
7088 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7090 if (!attr->exclude_user)
7091 mask |= PERF_SAMPLE_BRANCH_USER;
7093 if (!attr->exclude_hv)
7094 mask |= PERF_SAMPLE_BRANCH_HV;
7096 * adjust user setting (for HW filter setup)
7098 attr->branch_sample_type = mask;
7100 /* privileged levels capture (kernel, hv): check permissions */
7101 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7102 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7106 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7107 ret = perf_reg_validate(attr->sample_regs_user);
7112 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7113 if (!arch_perf_have_user_stack_dump())
7117 * We have __u32 type for the size, but so far
7118 * we can only use __u16 as maximum due to the
7119 * __u16 sample size limit.
7121 if (attr->sample_stack_user >= USHRT_MAX)
7123 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7131 put_user(sizeof(*attr), &uattr->size);
7137 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7139 struct ring_buffer *rb = NULL;
7145 /* don't allow circular references */
7146 if (event == output_event)
7150 * Don't allow cross-cpu buffers
7152 if (output_event->cpu != event->cpu)
7156 * If its not a per-cpu rb, it must be the same task.
7158 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7162 mutex_lock(&event->mmap_mutex);
7163 /* Can't redirect output if we've got an active mmap() */
7164 if (atomic_read(&event->mmap_count))
7168 /* get the rb we want to redirect to */
7169 rb = ring_buffer_get(output_event);
7174 ring_buffer_attach(event, rb);
7178 mutex_unlock(&event->mmap_mutex);
7185 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7187 * @attr_uptr: event_id type attributes for monitoring/sampling
7190 * @group_fd: group leader event fd
7192 SYSCALL_DEFINE5(perf_event_open,
7193 struct perf_event_attr __user *, attr_uptr,
7194 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7196 struct perf_event *group_leader = NULL, *output_event = NULL;
7197 struct perf_event *event, *sibling;
7198 struct perf_event_attr attr;
7199 struct perf_event_context *ctx;
7200 struct file *event_file = NULL;
7201 struct fd group = {NULL, 0};
7202 struct task_struct *task = NULL;
7207 int f_flags = O_RDWR;
7209 /* for future expandability... */
7210 if (flags & ~PERF_FLAG_ALL)
7213 err = perf_copy_attr(attr_uptr, &attr);
7217 if (!attr.exclude_kernel) {
7218 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7223 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7226 if (attr.sample_period & (1ULL << 63))
7231 * In cgroup mode, the pid argument is used to pass the fd
7232 * opened to the cgroup directory in cgroupfs. The cpu argument
7233 * designates the cpu on which to monitor threads from that
7236 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7239 if (flags & PERF_FLAG_FD_CLOEXEC)
7240 f_flags |= O_CLOEXEC;
7242 event_fd = get_unused_fd_flags(f_flags);
7246 if (group_fd != -1) {
7247 err = perf_fget_light(group_fd, &group);
7250 group_leader = group.file->private_data;
7251 if (flags & PERF_FLAG_FD_OUTPUT)
7252 output_event = group_leader;
7253 if (flags & PERF_FLAG_FD_NO_GROUP)
7254 group_leader = NULL;
7257 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7258 task = find_lively_task_by_vpid(pid);
7260 err = PTR_ERR(task);
7265 if (task && group_leader &&
7266 group_leader->attr.inherit != attr.inherit) {
7273 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7275 if (IS_ERR(event)) {
7276 err = PTR_ERR(event);
7280 if (flags & PERF_FLAG_PID_CGROUP) {
7281 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7283 __free_event(event);
7288 if (is_sampling_event(event)) {
7289 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7295 account_event(event);
7298 * Special case software events and allow them to be part of
7299 * any hardware group.
7304 (is_software_event(event) != is_software_event(group_leader))) {
7305 if (is_software_event(event)) {
7307 * If event and group_leader are not both a software
7308 * event, and event is, then group leader is not.
7310 * Allow the addition of software events to !software
7311 * groups, this is safe because software events never
7314 pmu = group_leader->pmu;
7315 } else if (is_software_event(group_leader) &&
7316 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7318 * In case the group is a pure software group, and we
7319 * try to add a hardware event, move the whole group to
7320 * the hardware context.
7327 * Get the target context (task or percpu):
7329 ctx = find_get_context(pmu, task, event->cpu);
7336 put_task_struct(task);
7341 * Look up the group leader (we will attach this event to it):
7347 * Do not allow a recursive hierarchy (this new sibling
7348 * becoming part of another group-sibling):
7350 if (group_leader->group_leader != group_leader)
7353 * Do not allow to attach to a group in a different
7354 * task or CPU context:
7357 if (group_leader->ctx->type != ctx->type)
7360 if (group_leader->ctx != ctx)
7365 * Only a group leader can be exclusive or pinned
7367 if (attr.exclusive || attr.pinned)
7372 err = perf_event_set_output(event, output_event);
7377 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7379 if (IS_ERR(event_file)) {
7380 err = PTR_ERR(event_file);
7385 struct perf_event_context *gctx = group_leader->ctx;
7387 mutex_lock(&gctx->mutex);
7388 perf_remove_from_context(group_leader, false);
7391 * Removing from the context ends up with disabled
7392 * event. What we want here is event in the initial
7393 * startup state, ready to be add into new context.
7395 perf_event__state_init(group_leader);
7396 list_for_each_entry(sibling, &group_leader->sibling_list,
7398 perf_remove_from_context(sibling, false);
7399 perf_event__state_init(sibling);
7402 mutex_unlock(&gctx->mutex);
7406 WARN_ON_ONCE(ctx->parent_ctx);
7407 mutex_lock(&ctx->mutex);
7411 perf_install_in_context(ctx, group_leader, event->cpu);
7413 list_for_each_entry(sibling, &group_leader->sibling_list,
7415 perf_install_in_context(ctx, sibling, event->cpu);
7420 perf_install_in_context(ctx, event, event->cpu);
7421 perf_unpin_context(ctx);
7422 mutex_unlock(&ctx->mutex);
7426 event->owner = current;
7428 mutex_lock(¤t->perf_event_mutex);
7429 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7430 mutex_unlock(¤t->perf_event_mutex);
7433 * Precalculate sample_data sizes
7435 perf_event__header_size(event);
7436 perf_event__id_header_size(event);
7439 * Drop the reference on the group_event after placing the
7440 * new event on the sibling_list. This ensures destruction
7441 * of the group leader will find the pointer to itself in
7442 * perf_group_detach().
7445 fd_install(event_fd, event_file);
7449 perf_unpin_context(ctx);
7457 put_task_struct(task);
7461 put_unused_fd(event_fd);
7466 * perf_event_create_kernel_counter
7468 * @attr: attributes of the counter to create
7469 * @cpu: cpu in which the counter is bound
7470 * @task: task to profile (NULL for percpu)
7473 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7474 struct task_struct *task,
7475 perf_overflow_handler_t overflow_handler,
7478 struct perf_event_context *ctx;
7479 struct perf_event *event;
7483 * Get the target context (task or percpu):
7486 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7487 overflow_handler, context);
7488 if (IS_ERR(event)) {
7489 err = PTR_ERR(event);
7493 /* Mark owner so we could distinguish it from user events. */
7494 event->owner = EVENT_OWNER_KERNEL;
7496 account_event(event);
7498 ctx = find_get_context(event->pmu, task, cpu);
7504 WARN_ON_ONCE(ctx->parent_ctx);
7505 mutex_lock(&ctx->mutex);
7506 perf_install_in_context(ctx, event, cpu);
7507 perf_unpin_context(ctx);
7508 mutex_unlock(&ctx->mutex);
7515 return ERR_PTR(err);
7517 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7519 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7521 struct perf_event_context *src_ctx;
7522 struct perf_event_context *dst_ctx;
7523 struct perf_event *event, *tmp;
7526 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7527 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7529 mutex_lock(&src_ctx->mutex);
7530 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7532 perf_remove_from_context(event, false);
7533 unaccount_event_cpu(event, src_cpu);
7535 list_add(&event->migrate_entry, &events);
7537 mutex_unlock(&src_ctx->mutex);
7541 mutex_lock(&dst_ctx->mutex);
7542 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7543 list_del(&event->migrate_entry);
7544 if (event->state >= PERF_EVENT_STATE_OFF)
7545 event->state = PERF_EVENT_STATE_INACTIVE;
7546 account_event_cpu(event, dst_cpu);
7547 perf_install_in_context(dst_ctx, event, dst_cpu);
7550 mutex_unlock(&dst_ctx->mutex);
7552 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7554 static void sync_child_event(struct perf_event *child_event,
7555 struct task_struct *child)
7557 struct perf_event *parent_event = child_event->parent;
7560 if (child_event->attr.inherit_stat)
7561 perf_event_read_event(child_event, child);
7563 child_val = perf_event_count(child_event);
7566 * Add back the child's count to the parent's count:
7568 atomic64_add(child_val, &parent_event->child_count);
7569 atomic64_add(child_event->total_time_enabled,
7570 &parent_event->child_total_time_enabled);
7571 atomic64_add(child_event->total_time_running,
7572 &parent_event->child_total_time_running);
7575 * Remove this event from the parent's list
7577 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7578 mutex_lock(&parent_event->child_mutex);
7579 list_del_init(&child_event->child_list);
7580 mutex_unlock(&parent_event->child_mutex);
7583 * Release the parent event, if this was the last
7586 put_event(parent_event);
7590 __perf_event_exit_task(struct perf_event *child_event,
7591 struct perf_event_context *child_ctx,
7592 struct task_struct *child)
7595 * Do not destroy the 'original' grouping; because of the context
7596 * switch optimization the original events could've ended up in a
7597 * random child task.
7599 * If we were to destroy the original group, all group related
7600 * operations would cease to function properly after this random
7603 * Do destroy all inherited groups, we don't care about those
7604 * and being thorough is better.
7606 perf_remove_from_context(child_event, !!child_event->parent);
7609 * It can happen that the parent exits first, and has events
7610 * that are still around due to the child reference. These
7611 * events need to be zapped.
7613 if (child_event->parent) {
7614 sync_child_event(child_event, child);
7615 free_event(child_event);
7617 child_event->state = PERF_EVENT_STATE_EXIT;
7618 perf_event_wakeup(child_event);
7622 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7624 struct perf_event *child_event, *next;
7625 struct perf_event_context *child_ctx, *parent_ctx;
7626 unsigned long flags;
7628 if (likely(!child->perf_event_ctxp[ctxn])) {
7629 perf_event_task(child, NULL, 0);
7633 local_irq_save(flags);
7635 * We can't reschedule here because interrupts are disabled,
7636 * and either child is current or it is a task that can't be
7637 * scheduled, so we are now safe from rescheduling changing
7640 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7643 * Take the context lock here so that if find_get_context is
7644 * reading child->perf_event_ctxp, we wait until it has
7645 * incremented the context's refcount before we do put_ctx below.
7647 raw_spin_lock(&child_ctx->lock);
7648 task_ctx_sched_out(child_ctx);
7649 child->perf_event_ctxp[ctxn] = NULL;
7652 * In order to avoid freeing: child_ctx->parent_ctx->task
7653 * under perf_event_context::lock, grab another reference.
7655 parent_ctx = child_ctx->parent_ctx;
7657 get_ctx(parent_ctx);
7660 * If this context is a clone; unclone it so it can't get
7661 * swapped to another process while we're removing all
7662 * the events from it.
7664 unclone_ctx(child_ctx);
7665 update_context_time(child_ctx);
7666 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7669 * Now that we no longer hold perf_event_context::lock, drop
7670 * our extra child_ctx->parent_ctx reference.
7673 put_ctx(parent_ctx);
7676 * Report the task dead after unscheduling the events so that we
7677 * won't get any samples after PERF_RECORD_EXIT. We can however still
7678 * get a few PERF_RECORD_READ events.
7680 perf_event_task(child, child_ctx, 0);
7683 * We can recurse on the same lock type through:
7685 * __perf_event_exit_task()
7686 * sync_child_event()
7688 * mutex_lock(&ctx->mutex)
7690 * But since its the parent context it won't be the same instance.
7692 mutex_lock(&child_ctx->mutex);
7694 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7695 __perf_event_exit_task(child_event, child_ctx, child);
7697 mutex_unlock(&child_ctx->mutex);
7703 * When a child task exits, feed back event values to parent events.
7705 void perf_event_exit_task(struct task_struct *child)
7707 struct perf_event *event, *tmp;
7710 mutex_lock(&child->perf_event_mutex);
7711 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7713 list_del_init(&event->owner_entry);
7716 * Ensure the list deletion is visible before we clear
7717 * the owner, closes a race against perf_release() where
7718 * we need to serialize on the owner->perf_event_mutex.
7721 event->owner = NULL;
7723 mutex_unlock(&child->perf_event_mutex);
7725 for_each_task_context_nr(ctxn)
7726 perf_event_exit_task_context(child, ctxn);
7729 static void perf_free_event(struct perf_event *event,
7730 struct perf_event_context *ctx)
7732 struct perf_event *parent = event->parent;
7734 if (WARN_ON_ONCE(!parent))
7737 mutex_lock(&parent->child_mutex);
7738 list_del_init(&event->child_list);
7739 mutex_unlock(&parent->child_mutex);
7743 perf_group_detach(event);
7744 list_del_event(event, ctx);
7749 * free an unexposed, unused context as created by inheritance by
7750 * perf_event_init_task below, used by fork() in case of fail.
7752 void perf_event_free_task(struct task_struct *task)
7754 struct perf_event_context *ctx;
7755 struct perf_event *event, *tmp;
7758 for_each_task_context_nr(ctxn) {
7759 ctx = task->perf_event_ctxp[ctxn];
7763 mutex_lock(&ctx->mutex);
7765 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
7767 perf_free_event(event, ctx);
7769 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
7771 perf_free_event(event, ctx);
7773 if (!list_empty(&ctx->pinned_groups) ||
7774 !list_empty(&ctx->flexible_groups))
7777 mutex_unlock(&ctx->mutex);
7783 void perf_event_delayed_put(struct task_struct *task)
7787 for_each_task_context_nr(ctxn)
7788 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
7792 * inherit a event from parent task to child task:
7794 static struct perf_event *
7795 inherit_event(struct perf_event *parent_event,
7796 struct task_struct *parent,
7797 struct perf_event_context *parent_ctx,
7798 struct task_struct *child,
7799 struct perf_event *group_leader,
7800 struct perf_event_context *child_ctx)
7802 struct perf_event *child_event;
7803 unsigned long flags;
7806 * Instead of creating recursive hierarchies of events,
7807 * we link inherited events back to the original parent,
7808 * which has a filp for sure, which we use as the reference
7811 if (parent_event->parent)
7812 parent_event = parent_event->parent;
7814 child_event = perf_event_alloc(&parent_event->attr,
7817 group_leader, parent_event,
7819 if (IS_ERR(child_event))
7822 if (is_orphaned_event(parent_event) ||
7823 !atomic_long_inc_not_zero(&parent_event->refcount)) {
7824 free_event(child_event);
7831 * Make the child state follow the state of the parent event,
7832 * not its attr.disabled bit. We hold the parent's mutex,
7833 * so we won't race with perf_event_{en, dis}able_family.
7835 if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
7836 child_event->state = PERF_EVENT_STATE_INACTIVE;
7838 child_event->state = PERF_EVENT_STATE_OFF;
7840 if (parent_event->attr.freq) {
7841 u64 sample_period = parent_event->hw.sample_period;
7842 struct hw_perf_event *hwc = &child_event->hw;
7844 hwc->sample_period = sample_period;
7845 hwc->last_period = sample_period;
7847 local64_set(&hwc->period_left, sample_period);
7850 child_event->ctx = child_ctx;
7851 child_event->overflow_handler = parent_event->overflow_handler;
7852 child_event->overflow_handler_context
7853 = parent_event->overflow_handler_context;
7856 * Precalculate sample_data sizes
7858 perf_event__header_size(child_event);
7859 perf_event__id_header_size(child_event);
7862 * Link it up in the child's context:
7864 raw_spin_lock_irqsave(&child_ctx->lock, flags);
7865 add_event_to_ctx(child_event, child_ctx);
7866 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7869 * Link this into the parent event's child list
7871 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7872 mutex_lock(&parent_event->child_mutex);
7873 list_add_tail(&child_event->child_list, &parent_event->child_list);
7874 mutex_unlock(&parent_event->child_mutex);
7879 static int inherit_group(struct perf_event *parent_event,
7880 struct task_struct *parent,
7881 struct perf_event_context *parent_ctx,
7882 struct task_struct *child,
7883 struct perf_event_context *child_ctx)
7885 struct perf_event *leader;
7886 struct perf_event *sub;
7887 struct perf_event *child_ctr;
7889 leader = inherit_event(parent_event, parent, parent_ctx,
7890 child, NULL, child_ctx);
7892 return PTR_ERR(leader);
7893 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
7894 child_ctr = inherit_event(sub, parent, parent_ctx,
7895 child, leader, child_ctx);
7896 if (IS_ERR(child_ctr))
7897 return PTR_ERR(child_ctr);
7903 inherit_task_group(struct perf_event *event, struct task_struct *parent,
7904 struct perf_event_context *parent_ctx,
7905 struct task_struct *child, int ctxn,
7909 struct perf_event_context *child_ctx;
7911 if (!event->attr.inherit) {
7916 child_ctx = child->perf_event_ctxp[ctxn];
7919 * This is executed from the parent task context, so
7920 * inherit events that have been marked for cloning.
7921 * First allocate and initialize a context for the
7925 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
7929 child->perf_event_ctxp[ctxn] = child_ctx;
7932 ret = inherit_group(event, parent, parent_ctx,
7942 * Initialize the perf_event context in task_struct
7944 static int perf_event_init_context(struct task_struct *child, int ctxn)
7946 struct perf_event_context *child_ctx, *parent_ctx;
7947 struct perf_event_context *cloned_ctx;
7948 struct perf_event *event;
7949 struct task_struct *parent = current;
7950 int inherited_all = 1;
7951 unsigned long flags;
7954 if (likely(!parent->perf_event_ctxp[ctxn]))
7958 * If the parent's context is a clone, pin it so it won't get
7961 parent_ctx = perf_pin_task_context(parent, ctxn);
7966 * No need to check if parent_ctx != NULL here; since we saw
7967 * it non-NULL earlier, the only reason for it to become NULL
7968 * is if we exit, and since we're currently in the middle of
7969 * a fork we can't be exiting at the same time.
7973 * Lock the parent list. No need to lock the child - not PID
7974 * hashed yet and not running, so nobody can access it.
7976 mutex_lock(&parent_ctx->mutex);
7979 * We dont have to disable NMIs - we are only looking at
7980 * the list, not manipulating it:
7982 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
7983 ret = inherit_task_group(event, parent, parent_ctx,
7984 child, ctxn, &inherited_all);
7990 * We can't hold ctx->lock when iterating the ->flexible_group list due
7991 * to allocations, but we need to prevent rotation because
7992 * rotate_ctx() will change the list from interrupt context.
7994 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
7995 parent_ctx->rotate_disable = 1;
7996 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
7998 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
7999 ret = inherit_task_group(event, parent, parent_ctx,
8000 child, ctxn, &inherited_all);
8005 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8006 parent_ctx->rotate_disable = 0;
8008 child_ctx = child->perf_event_ctxp[ctxn];
8010 if (child_ctx && inherited_all) {
8012 * Mark the child context as a clone of the parent
8013 * context, or of whatever the parent is a clone of.
8015 * Note that if the parent is a clone, the holding of
8016 * parent_ctx->lock avoids it from being uncloned.
8018 cloned_ctx = parent_ctx->parent_ctx;
8020 child_ctx->parent_ctx = cloned_ctx;
8021 child_ctx->parent_gen = parent_ctx->parent_gen;
8023 child_ctx->parent_ctx = parent_ctx;
8024 child_ctx->parent_gen = parent_ctx->generation;
8026 get_ctx(child_ctx->parent_ctx);
8029 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8030 mutex_unlock(&parent_ctx->mutex);
8032 perf_unpin_context(parent_ctx);
8033 put_ctx(parent_ctx);
8039 * Initialize the perf_event context in task_struct
8041 int perf_event_init_task(struct task_struct *child)
8045 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8046 mutex_init(&child->perf_event_mutex);
8047 INIT_LIST_HEAD(&child->perf_event_list);
8049 for_each_task_context_nr(ctxn) {
8050 ret = perf_event_init_context(child, ctxn);
8058 static void __init perf_event_init_all_cpus(void)
8060 struct swevent_htable *swhash;
8063 for_each_possible_cpu(cpu) {
8064 swhash = &per_cpu(swevent_htable, cpu);
8065 mutex_init(&swhash->hlist_mutex);
8066 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
8070 static void perf_event_init_cpu(int cpu)
8072 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8074 mutex_lock(&swhash->hlist_mutex);
8075 swhash->online = true;
8076 if (swhash->hlist_refcount > 0) {
8077 struct swevent_hlist *hlist;
8079 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8081 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8083 mutex_unlock(&swhash->hlist_mutex);
8086 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8087 static void perf_pmu_rotate_stop(struct pmu *pmu)
8089 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
8091 WARN_ON(!irqs_disabled());
8093 list_del_init(&cpuctx->rotation_list);
8096 static void __perf_event_exit_context(void *__info)
8098 struct remove_event re = { .detach_group = false };
8099 struct perf_event_context *ctx = __info;
8101 perf_pmu_rotate_stop(ctx->pmu);
8104 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8105 __perf_remove_from_context(&re);
8109 static void perf_event_exit_cpu_context(int cpu)
8111 struct perf_event_context *ctx;
8115 idx = srcu_read_lock(&pmus_srcu);
8116 list_for_each_entry_rcu(pmu, &pmus, entry) {
8117 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8119 mutex_lock(&ctx->mutex);
8120 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8121 mutex_unlock(&ctx->mutex);
8123 srcu_read_unlock(&pmus_srcu, idx);
8126 static void perf_event_exit_cpu(int cpu)
8128 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8130 perf_event_exit_cpu_context(cpu);
8132 mutex_lock(&swhash->hlist_mutex);
8133 swhash->online = false;
8134 swevent_hlist_release(swhash);
8135 mutex_unlock(&swhash->hlist_mutex);
8138 static inline void perf_event_exit_cpu(int cpu) { }
8142 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8146 for_each_online_cpu(cpu)
8147 perf_event_exit_cpu(cpu);
8153 * Run the perf reboot notifier at the very last possible moment so that
8154 * the generic watchdog code runs as long as possible.
8156 static struct notifier_block perf_reboot_notifier = {
8157 .notifier_call = perf_reboot,
8158 .priority = INT_MIN,
8162 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8164 unsigned int cpu = (long)hcpu;
8166 switch (action & ~CPU_TASKS_FROZEN) {
8168 case CPU_UP_PREPARE:
8169 case CPU_DOWN_FAILED:
8170 perf_event_init_cpu(cpu);
8173 case CPU_UP_CANCELED:
8174 case CPU_DOWN_PREPARE:
8175 perf_event_exit_cpu(cpu);
8184 void __init perf_event_init(void)
8190 perf_event_init_all_cpus();
8191 init_srcu_struct(&pmus_srcu);
8192 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8193 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8194 perf_pmu_register(&perf_task_clock, NULL, -1);
8196 perf_cpu_notifier(perf_cpu_notify);
8197 register_reboot_notifier(&perf_reboot_notifier);
8199 ret = init_hw_breakpoint();
8200 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8202 /* do not patch jump label more than once per second */
8203 jump_label_rate_limit(&perf_sched_events, HZ);
8206 * Build time assertion that we keep the data_head at the intended
8207 * location. IOW, validation we got the __reserved[] size right.
8209 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8213 static int __init perf_event_sysfs_init(void)
8218 mutex_lock(&pmus_lock);
8220 ret = bus_register(&pmu_bus);
8224 list_for_each_entry(pmu, &pmus, entry) {
8225 if (!pmu->name || pmu->type < 0)
8228 ret = pmu_dev_alloc(pmu);
8229 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8231 pmu_bus_running = 1;
8235 mutex_unlock(&pmus_lock);
8239 device_initcall(perf_event_sysfs_init);
8241 #ifdef CONFIG_CGROUP_PERF
8242 static struct cgroup_subsys_state *
8243 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8245 struct perf_cgroup *jc;
8247 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8249 return ERR_PTR(-ENOMEM);
8251 jc->info = alloc_percpu(struct perf_cgroup_info);
8254 return ERR_PTR(-ENOMEM);
8260 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8262 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8264 free_percpu(jc->info);
8268 static int __perf_cgroup_move(void *info)
8270 struct task_struct *task = info;
8271 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8275 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8276 struct cgroup_taskset *tset)
8278 struct task_struct *task;
8280 cgroup_taskset_for_each(task, tset)
8281 task_function_call(task, __perf_cgroup_move, task);
8284 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8285 struct cgroup_subsys_state *old_css,
8286 struct task_struct *task)
8289 * cgroup_exit() is called in the copy_process() failure path.
8290 * Ignore this case since the task hasn't ran yet, this avoids
8291 * trying to poke a half freed task state from generic code.
8293 if (!(task->flags & PF_EXITING))
8296 task_function_call(task, __perf_cgroup_move, task);
8299 struct cgroup_subsys perf_event_cgrp_subsys = {
8300 .css_alloc = perf_cgroup_css_alloc,
8301 .css_free = perf_cgroup_css_free,
8302 .exit = perf_cgroup_exit,
8303 .attach = perf_cgroup_attach,
8305 #endif /* CONFIG_CGROUP_PERF */