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 = __this_cpu_read(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 = __this_cpu_read(running_sample_length);
275 local_samples_len -= local_samples_len/NR_ACCUMULATED_SAMPLES;
276 local_samples_len += sample_len_ns;
277 __this_cpu_write(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_detach_cgroup(struct perf_event *event)
406 css_put(&event->cgrp->css);
410 static inline int is_cgroup_event(struct perf_event *event)
412 return event->cgrp != NULL;
415 static inline u64 perf_cgroup_event_time(struct perf_event *event)
417 struct perf_cgroup_info *t;
419 t = per_cpu_ptr(event->cgrp->info, event->cpu);
423 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
425 struct perf_cgroup_info *info;
430 info = this_cpu_ptr(cgrp->info);
432 info->time += now - info->timestamp;
433 info->timestamp = now;
436 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
438 struct perf_cgroup *cgrp_out = cpuctx->cgrp;
440 __update_cgrp_time(cgrp_out);
443 static inline void update_cgrp_time_from_event(struct perf_event *event)
445 struct perf_cgroup *cgrp;
448 * ensure we access cgroup data only when needed and
449 * when we know the cgroup is pinned (css_get)
451 if (!is_cgroup_event(event))
454 cgrp = perf_cgroup_from_task(current);
456 * Do not update time when cgroup is not active
458 if (cgrp == event->cgrp)
459 __update_cgrp_time(event->cgrp);
463 perf_cgroup_set_timestamp(struct task_struct *task,
464 struct perf_event_context *ctx)
466 struct perf_cgroup *cgrp;
467 struct perf_cgroup_info *info;
470 * ctx->lock held by caller
471 * ensure we do not access cgroup data
472 * unless we have the cgroup pinned (css_get)
474 if (!task || !ctx->nr_cgroups)
477 cgrp = perf_cgroup_from_task(task);
478 info = this_cpu_ptr(cgrp->info);
479 info->timestamp = ctx->timestamp;
482 #define PERF_CGROUP_SWOUT 0x1 /* cgroup switch out every event */
483 #define PERF_CGROUP_SWIN 0x2 /* cgroup switch in events based on task */
486 * reschedule events based on the cgroup constraint of task.
488 * mode SWOUT : schedule out everything
489 * mode SWIN : schedule in based on cgroup for next
491 void perf_cgroup_switch(struct task_struct *task, int mode)
493 struct perf_cpu_context *cpuctx;
498 * disable interrupts to avoid geting nr_cgroup
499 * changes via __perf_event_disable(). Also
502 local_irq_save(flags);
505 * we reschedule only in the presence of cgroup
506 * constrained events.
510 list_for_each_entry_rcu(pmu, &pmus, entry) {
511 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
512 if (cpuctx->unique_pmu != pmu)
513 continue; /* ensure we process each cpuctx once */
516 * perf_cgroup_events says at least one
517 * context on this CPU has cgroup events.
519 * ctx->nr_cgroups reports the number of cgroup
520 * events for a context.
522 if (cpuctx->ctx.nr_cgroups > 0) {
523 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
524 perf_pmu_disable(cpuctx->ctx.pmu);
526 if (mode & PERF_CGROUP_SWOUT) {
527 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
529 * must not be done before ctxswout due
530 * to event_filter_match() in event_sched_out()
535 if (mode & PERF_CGROUP_SWIN) {
536 WARN_ON_ONCE(cpuctx->cgrp);
538 * set cgrp before ctxsw in to allow
539 * event_filter_match() to not have to pass
542 cpuctx->cgrp = perf_cgroup_from_task(task);
543 cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
545 perf_pmu_enable(cpuctx->ctx.pmu);
546 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
552 local_irq_restore(flags);
555 static inline void perf_cgroup_sched_out(struct task_struct *task,
556 struct task_struct *next)
558 struct perf_cgroup *cgrp1;
559 struct perf_cgroup *cgrp2 = NULL;
562 * we come here when we know perf_cgroup_events > 0
564 cgrp1 = perf_cgroup_from_task(task);
567 * next is NULL when called from perf_event_enable_on_exec()
568 * that will systematically cause a cgroup_switch()
571 cgrp2 = perf_cgroup_from_task(next);
574 * only schedule out current cgroup events if we know
575 * that we are switching to a different cgroup. Otherwise,
576 * do no touch the cgroup events.
579 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
582 static inline void perf_cgroup_sched_in(struct task_struct *prev,
583 struct task_struct *task)
585 struct perf_cgroup *cgrp1;
586 struct perf_cgroup *cgrp2 = NULL;
589 * we come here when we know perf_cgroup_events > 0
591 cgrp1 = perf_cgroup_from_task(task);
593 /* prev can never be NULL */
594 cgrp2 = perf_cgroup_from_task(prev);
597 * only need to schedule in cgroup events if we are changing
598 * cgroup during ctxsw. Cgroup events were not scheduled
599 * out of ctxsw out if that was not the case.
602 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
605 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
606 struct perf_event_attr *attr,
607 struct perf_event *group_leader)
609 struct perf_cgroup *cgrp;
610 struct cgroup_subsys_state *css;
611 struct fd f = fdget(fd);
617 css = css_tryget_online_from_dir(f.file->f_path.dentry,
618 &perf_event_cgrp_subsys);
624 cgrp = container_of(css, struct perf_cgroup, css);
628 * all events in a group must monitor
629 * the same cgroup because a task belongs
630 * to only one perf cgroup at a time
632 if (group_leader && group_leader->cgrp != cgrp) {
633 perf_detach_cgroup(event);
642 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
644 struct perf_cgroup_info *t;
645 t = per_cpu_ptr(event->cgrp->info, event->cpu);
646 event->shadow_ctx_time = now - t->timestamp;
650 perf_cgroup_defer_enabled(struct perf_event *event)
653 * when the current task's perf cgroup does not match
654 * the event's, we need to remember to call the
655 * perf_mark_enable() function the first time a task with
656 * a matching perf cgroup is scheduled in.
658 if (is_cgroup_event(event) && !perf_cgroup_match(event))
659 event->cgrp_defer_enabled = 1;
663 perf_cgroup_mark_enabled(struct perf_event *event,
664 struct perf_event_context *ctx)
666 struct perf_event *sub;
667 u64 tstamp = perf_event_time(event);
669 if (!event->cgrp_defer_enabled)
672 event->cgrp_defer_enabled = 0;
674 event->tstamp_enabled = tstamp - event->total_time_enabled;
675 list_for_each_entry(sub, &event->sibling_list, group_entry) {
676 if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
677 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
678 sub->cgrp_defer_enabled = 0;
682 #else /* !CONFIG_CGROUP_PERF */
685 perf_cgroup_match(struct perf_event *event)
690 static inline void perf_detach_cgroup(struct perf_event *event)
693 static inline int is_cgroup_event(struct perf_event *event)
698 static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
703 static inline void update_cgrp_time_from_event(struct perf_event *event)
707 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
711 static inline void perf_cgroup_sched_out(struct task_struct *task,
712 struct task_struct *next)
716 static inline void perf_cgroup_sched_in(struct task_struct *prev,
717 struct task_struct *task)
721 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
722 struct perf_event_attr *attr,
723 struct perf_event *group_leader)
729 perf_cgroup_set_timestamp(struct task_struct *task,
730 struct perf_event_context *ctx)
735 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
740 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
744 static inline u64 perf_cgroup_event_time(struct perf_event *event)
750 perf_cgroup_defer_enabled(struct perf_event *event)
755 perf_cgroup_mark_enabled(struct perf_event *event,
756 struct perf_event_context *ctx)
762 * set default to be dependent on timer tick just
765 #define PERF_CPU_HRTIMER (1000 / HZ)
767 * function must be called with interrupts disbled
769 static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
771 struct perf_cpu_context *cpuctx;
772 enum hrtimer_restart ret = HRTIMER_NORESTART;
775 WARN_ON(!irqs_disabled());
777 cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
779 rotations = perf_rotate_context(cpuctx);
782 * arm timer if needed
785 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
786 ret = HRTIMER_RESTART;
792 /* CPU is going down */
793 void perf_cpu_hrtimer_cancel(int cpu)
795 struct perf_cpu_context *cpuctx;
799 if (WARN_ON(cpu != smp_processor_id()))
802 local_irq_save(flags);
806 list_for_each_entry_rcu(pmu, &pmus, entry) {
807 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
809 if (pmu->task_ctx_nr == perf_sw_context)
812 hrtimer_cancel(&cpuctx->hrtimer);
817 local_irq_restore(flags);
820 static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
822 struct hrtimer *hr = &cpuctx->hrtimer;
823 struct pmu *pmu = cpuctx->ctx.pmu;
826 /* no multiplexing needed for SW PMU */
827 if (pmu->task_ctx_nr == perf_sw_context)
831 * check default is sane, if not set then force to
832 * default interval (1/tick)
834 timer = pmu->hrtimer_interval_ms;
836 timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
838 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
840 hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
841 hr->function = perf_cpu_hrtimer_handler;
844 static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
846 struct hrtimer *hr = &cpuctx->hrtimer;
847 struct pmu *pmu = cpuctx->ctx.pmu;
850 if (pmu->task_ctx_nr == perf_sw_context)
853 if (hrtimer_active(hr))
856 if (!hrtimer_callback_running(hr))
857 __hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
858 0, HRTIMER_MODE_REL_PINNED, 0);
861 void perf_pmu_disable(struct pmu *pmu)
863 int *count = this_cpu_ptr(pmu->pmu_disable_count);
865 pmu->pmu_disable(pmu);
868 void perf_pmu_enable(struct pmu *pmu)
870 int *count = this_cpu_ptr(pmu->pmu_disable_count);
872 pmu->pmu_enable(pmu);
875 static DEFINE_PER_CPU(struct list_head, rotation_list);
878 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
879 * because they're strictly cpu affine and rotate_start is called with IRQs
880 * disabled, while rotate_context is called from IRQ context.
882 static void perf_pmu_rotate_start(struct pmu *pmu)
884 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
885 struct list_head *head = this_cpu_ptr(&rotation_list);
887 WARN_ON(!irqs_disabled());
889 if (list_empty(&cpuctx->rotation_list))
890 list_add(&cpuctx->rotation_list, head);
893 static void get_ctx(struct perf_event_context *ctx)
895 WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
898 static void put_ctx(struct perf_event_context *ctx)
900 if (atomic_dec_and_test(&ctx->refcount)) {
902 put_ctx(ctx->parent_ctx);
904 put_task_struct(ctx->task);
905 kfree_rcu(ctx, rcu_head);
910 * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
911 * perf_pmu_migrate_context() we need some magic.
913 * Those places that change perf_event::ctx will hold both
914 * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
916 * Lock ordering is by mutex address. There is one other site where
917 * perf_event_context::mutex nests and that is put_event(). But remember that
918 * that is a parent<->child context relation, and migration does not affect
919 * children, therefore these two orderings should not interact.
921 * The change in perf_event::ctx does not affect children (as claimed above)
922 * because the sys_perf_event_open() case will install a new event and break
923 * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
924 * concerned with cpuctx and that doesn't have children.
926 * The places that change perf_event::ctx will issue:
928 * perf_remove_from_context();
930 * perf_install_in_context();
932 * to affect the change. The remove_from_context() + synchronize_rcu() should
933 * quiesce the event, after which we can install it in the new location. This
934 * means that only external vectors (perf_fops, prctl) can perturb the event
935 * while in transit. Therefore all such accessors should also acquire
936 * perf_event_context::mutex to serialize against this.
938 * However; because event->ctx can change while we're waiting to acquire
939 * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
943 * task_struct::perf_event_mutex
944 * perf_event_context::mutex
945 * perf_event_context::lock
946 * perf_event::child_mutex;
947 * perf_event::mmap_mutex
950 static struct perf_event_context *perf_event_ctx_lock(struct perf_event *event)
952 struct perf_event_context *ctx;
956 ctx = ACCESS_ONCE(event->ctx);
957 if (!atomic_inc_not_zero(&ctx->refcount)) {
963 mutex_lock(&ctx->mutex);
964 if (event->ctx != ctx) {
965 mutex_unlock(&ctx->mutex);
973 static void perf_event_ctx_unlock(struct perf_event *event,
974 struct perf_event_context *ctx)
976 mutex_unlock(&ctx->mutex);
981 * This must be done under the ctx->lock, such as to serialize against
982 * context_equiv(), therefore we cannot call put_ctx() since that might end up
983 * calling scheduler related locks and ctx->lock nests inside those.
985 static __must_check struct perf_event_context *
986 unclone_ctx(struct perf_event_context *ctx)
988 struct perf_event_context *parent_ctx = ctx->parent_ctx;
990 lockdep_assert_held(&ctx->lock);
993 ctx->parent_ctx = NULL;
999 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1002 * only top level events have the pid namespace they were created in
1005 event = event->parent;
1007 return task_tgid_nr_ns(p, event->ns);
1010 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1013 * only top level events have the pid namespace they were created in
1016 event = event->parent;
1018 return task_pid_nr_ns(p, event->ns);
1022 * If we inherit events we want to return the parent event id
1025 static u64 primary_event_id(struct perf_event *event)
1030 id = event->parent->id;
1036 * Get the perf_event_context for a task and lock it.
1037 * This has to cope with with the fact that until it is locked,
1038 * the context could get moved to another task.
1040 static struct perf_event_context *
1041 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1043 struct perf_event_context *ctx;
1047 * One of the few rules of preemptible RCU is that one cannot do
1048 * rcu_read_unlock() while holding a scheduler (or nested) lock when
1049 * part of the read side critical section was preemptible -- see
1050 * rcu_read_unlock_special().
1052 * Since ctx->lock nests under rq->lock we must ensure the entire read
1053 * side critical section is non-preemptible.
1057 ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1060 * If this context is a clone of another, it might
1061 * get swapped for another underneath us by
1062 * perf_event_task_sched_out, though the
1063 * rcu_read_lock() protects us from any context
1064 * getting freed. Lock the context and check if it
1065 * got swapped before we could get the lock, and retry
1066 * if so. If we locked the right context, then it
1067 * can't get swapped on us any more.
1069 raw_spin_lock_irqsave(&ctx->lock, *flags);
1070 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1071 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1077 if (!atomic_inc_not_zero(&ctx->refcount)) {
1078 raw_spin_unlock_irqrestore(&ctx->lock, *flags);
1088 * Get the context for a task and increment its pin_count so it
1089 * can't get swapped to another task. This also increments its
1090 * reference count so that the context can't get freed.
1092 static struct perf_event_context *
1093 perf_pin_task_context(struct task_struct *task, int ctxn)
1095 struct perf_event_context *ctx;
1096 unsigned long flags;
1098 ctx = perf_lock_task_context(task, ctxn, &flags);
1101 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1106 static void perf_unpin_context(struct perf_event_context *ctx)
1108 unsigned long flags;
1110 raw_spin_lock_irqsave(&ctx->lock, flags);
1112 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1116 * Update the record of the current time in a context.
1118 static void update_context_time(struct perf_event_context *ctx)
1120 u64 now = perf_clock();
1122 ctx->time += now - ctx->timestamp;
1123 ctx->timestamp = now;
1126 static u64 perf_event_time(struct perf_event *event)
1128 struct perf_event_context *ctx = event->ctx;
1130 if (is_cgroup_event(event))
1131 return perf_cgroup_event_time(event);
1133 return ctx ? ctx->time : 0;
1137 * Update the total_time_enabled and total_time_running fields for a event.
1138 * The caller of this function needs to hold the ctx->lock.
1140 static void update_event_times(struct perf_event *event)
1142 struct perf_event_context *ctx = event->ctx;
1145 if (event->state < PERF_EVENT_STATE_INACTIVE ||
1146 event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
1149 * in cgroup mode, time_enabled represents
1150 * the time the event was enabled AND active
1151 * tasks were in the monitored cgroup. This is
1152 * independent of the activity of the context as
1153 * there may be a mix of cgroup and non-cgroup events.
1155 * That is why we treat cgroup events differently
1158 if (is_cgroup_event(event))
1159 run_end = perf_cgroup_event_time(event);
1160 else if (ctx->is_active)
1161 run_end = ctx->time;
1163 run_end = event->tstamp_stopped;
1165 event->total_time_enabled = run_end - event->tstamp_enabled;
1167 if (event->state == PERF_EVENT_STATE_INACTIVE)
1168 run_end = event->tstamp_stopped;
1170 run_end = perf_event_time(event);
1172 event->total_time_running = run_end - event->tstamp_running;
1177 * Update total_time_enabled and total_time_running for all events in a group.
1179 static void update_group_times(struct perf_event *leader)
1181 struct perf_event *event;
1183 update_event_times(leader);
1184 list_for_each_entry(event, &leader->sibling_list, group_entry)
1185 update_event_times(event);
1188 static struct list_head *
1189 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1191 if (event->attr.pinned)
1192 return &ctx->pinned_groups;
1194 return &ctx->flexible_groups;
1198 * Add a event from the lists for its context.
1199 * Must be called with ctx->mutex and ctx->lock held.
1202 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1204 WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1205 event->attach_state |= PERF_ATTACH_CONTEXT;
1208 * If we're a stand alone event or group leader, we go to the context
1209 * list, group events are kept attached to the group so that
1210 * perf_group_detach can, at all times, locate all siblings.
1212 if (event->group_leader == event) {
1213 struct list_head *list;
1215 if (is_software_event(event))
1216 event->group_flags |= PERF_GROUP_SOFTWARE;
1218 list = ctx_group_list(event, ctx);
1219 list_add_tail(&event->group_entry, list);
1222 if (is_cgroup_event(event))
1225 if (has_branch_stack(event))
1226 ctx->nr_branch_stack++;
1228 list_add_rcu(&event->event_entry, &ctx->event_list);
1229 if (!ctx->nr_events)
1230 perf_pmu_rotate_start(ctx->pmu);
1232 if (event->attr.inherit_stat)
1239 * Initialize event state based on the perf_event_attr::disabled.
1241 static inline void perf_event__state_init(struct perf_event *event)
1243 event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1244 PERF_EVENT_STATE_INACTIVE;
1248 * Called at perf_event creation and when events are attached/detached from a
1251 static void perf_event__read_size(struct perf_event *event)
1253 int entry = sizeof(u64); /* value */
1257 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1258 size += sizeof(u64);
1260 if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1261 size += sizeof(u64);
1263 if (event->attr.read_format & PERF_FORMAT_ID)
1264 entry += sizeof(u64);
1266 if (event->attr.read_format & PERF_FORMAT_GROUP) {
1267 nr += event->group_leader->nr_siblings;
1268 size += sizeof(u64);
1272 event->read_size = size;
1275 static void perf_event__header_size(struct perf_event *event)
1277 struct perf_sample_data *data;
1278 u64 sample_type = event->attr.sample_type;
1281 perf_event__read_size(event);
1283 if (sample_type & PERF_SAMPLE_IP)
1284 size += sizeof(data->ip);
1286 if (sample_type & PERF_SAMPLE_ADDR)
1287 size += sizeof(data->addr);
1289 if (sample_type & PERF_SAMPLE_PERIOD)
1290 size += sizeof(data->period);
1292 if (sample_type & PERF_SAMPLE_WEIGHT)
1293 size += sizeof(data->weight);
1295 if (sample_type & PERF_SAMPLE_READ)
1296 size += event->read_size;
1298 if (sample_type & PERF_SAMPLE_DATA_SRC)
1299 size += sizeof(data->data_src.val);
1301 if (sample_type & PERF_SAMPLE_TRANSACTION)
1302 size += sizeof(data->txn);
1304 event->header_size = size;
1307 static void perf_event__id_header_size(struct perf_event *event)
1309 struct perf_sample_data *data;
1310 u64 sample_type = event->attr.sample_type;
1313 if (sample_type & PERF_SAMPLE_TID)
1314 size += sizeof(data->tid_entry);
1316 if (sample_type & PERF_SAMPLE_TIME)
1317 size += sizeof(data->time);
1319 if (sample_type & PERF_SAMPLE_IDENTIFIER)
1320 size += sizeof(data->id);
1322 if (sample_type & PERF_SAMPLE_ID)
1323 size += sizeof(data->id);
1325 if (sample_type & PERF_SAMPLE_STREAM_ID)
1326 size += sizeof(data->stream_id);
1328 if (sample_type & PERF_SAMPLE_CPU)
1329 size += sizeof(data->cpu_entry);
1331 event->id_header_size = size;
1334 static void perf_group_attach(struct perf_event *event)
1336 struct perf_event *group_leader = event->group_leader, *pos;
1339 * We can have double attach due to group movement in perf_event_open.
1341 if (event->attach_state & PERF_ATTACH_GROUP)
1344 event->attach_state |= PERF_ATTACH_GROUP;
1346 if (group_leader == event)
1349 WARN_ON_ONCE(group_leader->ctx != event->ctx);
1351 if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
1352 !is_software_event(event))
1353 group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
1355 list_add_tail(&event->group_entry, &group_leader->sibling_list);
1356 group_leader->nr_siblings++;
1358 perf_event__header_size(group_leader);
1360 list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1361 perf_event__header_size(pos);
1365 * Remove a event from the lists for its context.
1366 * Must be called with ctx->mutex and ctx->lock held.
1369 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1371 struct perf_cpu_context *cpuctx;
1373 WARN_ON_ONCE(event->ctx != ctx);
1374 lockdep_assert_held(&ctx->lock);
1377 * We can have double detach due to exit/hot-unplug + close.
1379 if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1382 event->attach_state &= ~PERF_ATTACH_CONTEXT;
1384 if (is_cgroup_event(event)) {
1386 cpuctx = __get_cpu_context(ctx);
1388 * if there are no more cgroup events
1389 * then cler cgrp to avoid stale pointer
1390 * in update_cgrp_time_from_cpuctx()
1392 if (!ctx->nr_cgroups)
1393 cpuctx->cgrp = NULL;
1396 if (has_branch_stack(event))
1397 ctx->nr_branch_stack--;
1400 if (event->attr.inherit_stat)
1403 list_del_rcu(&event->event_entry);
1405 if (event->group_leader == event)
1406 list_del_init(&event->group_entry);
1408 update_group_times(event);
1411 * If event was in error state, then keep it
1412 * that way, otherwise bogus counts will be
1413 * returned on read(). The only way to get out
1414 * of error state is by explicit re-enabling
1417 if (event->state > PERF_EVENT_STATE_OFF)
1418 event->state = PERF_EVENT_STATE_OFF;
1423 static void perf_group_detach(struct perf_event *event)
1425 struct perf_event *sibling, *tmp;
1426 struct list_head *list = NULL;
1429 * We can have double detach due to exit/hot-unplug + close.
1431 if (!(event->attach_state & PERF_ATTACH_GROUP))
1434 event->attach_state &= ~PERF_ATTACH_GROUP;
1437 * If this is a sibling, remove it from its group.
1439 if (event->group_leader != event) {
1440 list_del_init(&event->group_entry);
1441 event->group_leader->nr_siblings--;
1445 if (!list_empty(&event->group_entry))
1446 list = &event->group_entry;
1449 * If this was a group event with sibling events then
1450 * upgrade the siblings to singleton events by adding them
1451 * to whatever list we are on.
1453 list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1455 list_move_tail(&sibling->group_entry, list);
1456 sibling->group_leader = sibling;
1458 /* Inherit group flags from the previous leader */
1459 sibling->group_flags = event->group_flags;
1461 WARN_ON_ONCE(sibling->ctx != event->ctx);
1465 perf_event__header_size(event->group_leader);
1467 list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1468 perf_event__header_size(tmp);
1472 * User event without the task.
1474 static bool is_orphaned_event(struct perf_event *event)
1476 return event && !is_kernel_event(event) && !event->owner;
1480 * Event has a parent but parent's task finished and it's
1481 * alive only because of children holding refference.
1483 static bool is_orphaned_child(struct perf_event *event)
1485 return is_orphaned_event(event->parent);
1488 static void orphans_remove_work(struct work_struct *work);
1490 static void schedule_orphans_remove(struct perf_event_context *ctx)
1492 if (!ctx->task || ctx->orphans_remove_sched || !perf_wq)
1495 if (queue_delayed_work(perf_wq, &ctx->orphans_remove, 1)) {
1497 ctx->orphans_remove_sched = true;
1501 static int __init perf_workqueue_init(void)
1503 perf_wq = create_singlethread_workqueue("perf");
1504 WARN(!perf_wq, "failed to create perf workqueue\n");
1505 return perf_wq ? 0 : -1;
1508 core_initcall(perf_workqueue_init);
1511 event_filter_match(struct perf_event *event)
1513 return (event->cpu == -1 || event->cpu == smp_processor_id())
1514 && perf_cgroup_match(event);
1518 event_sched_out(struct perf_event *event,
1519 struct perf_cpu_context *cpuctx,
1520 struct perf_event_context *ctx)
1522 u64 tstamp = perf_event_time(event);
1525 WARN_ON_ONCE(event->ctx != ctx);
1526 lockdep_assert_held(&ctx->lock);
1529 * An event which could not be activated because of
1530 * filter mismatch still needs to have its timings
1531 * maintained, otherwise bogus information is return
1532 * via read() for time_enabled, time_running:
1534 if (event->state == PERF_EVENT_STATE_INACTIVE
1535 && !event_filter_match(event)) {
1536 delta = tstamp - event->tstamp_stopped;
1537 event->tstamp_running += delta;
1538 event->tstamp_stopped = tstamp;
1541 if (event->state != PERF_EVENT_STATE_ACTIVE)
1544 perf_pmu_disable(event->pmu);
1546 event->state = PERF_EVENT_STATE_INACTIVE;
1547 if (event->pending_disable) {
1548 event->pending_disable = 0;
1549 event->state = PERF_EVENT_STATE_OFF;
1551 event->tstamp_stopped = tstamp;
1552 event->pmu->del(event, 0);
1555 if (!is_software_event(event))
1556 cpuctx->active_oncpu--;
1558 if (event->attr.freq && event->attr.sample_freq)
1560 if (event->attr.exclusive || !cpuctx->active_oncpu)
1561 cpuctx->exclusive = 0;
1563 if (is_orphaned_child(event))
1564 schedule_orphans_remove(ctx);
1566 perf_pmu_enable(event->pmu);
1570 group_sched_out(struct perf_event *group_event,
1571 struct perf_cpu_context *cpuctx,
1572 struct perf_event_context *ctx)
1574 struct perf_event *event;
1575 int state = group_event->state;
1577 event_sched_out(group_event, cpuctx, ctx);
1580 * Schedule out siblings (if any):
1582 list_for_each_entry(event, &group_event->sibling_list, group_entry)
1583 event_sched_out(event, cpuctx, ctx);
1585 if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
1586 cpuctx->exclusive = 0;
1589 struct remove_event {
1590 struct perf_event *event;
1595 * Cross CPU call to remove a performance event
1597 * We disable the event on the hardware level first. After that we
1598 * remove it from the context list.
1600 static int __perf_remove_from_context(void *info)
1602 struct remove_event *re = info;
1603 struct perf_event *event = re->event;
1604 struct perf_event_context *ctx = event->ctx;
1605 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1607 raw_spin_lock(&ctx->lock);
1608 event_sched_out(event, cpuctx, ctx);
1609 if (re->detach_group)
1610 perf_group_detach(event);
1611 list_del_event(event, ctx);
1612 if (!ctx->nr_events && cpuctx->task_ctx == ctx) {
1614 cpuctx->task_ctx = NULL;
1616 raw_spin_unlock(&ctx->lock);
1623 * Remove the event from a task's (or a CPU's) list of events.
1625 * CPU events are removed with a smp call. For task events we only
1626 * call when the task is on a CPU.
1628 * If event->ctx is a cloned context, callers must make sure that
1629 * every task struct that event->ctx->task could possibly point to
1630 * remains valid. This is OK when called from perf_release since
1631 * that only calls us on the top-level context, which can't be a clone.
1632 * When called from perf_event_exit_task, it's OK because the
1633 * context has been detached from its task.
1635 static void perf_remove_from_context(struct perf_event *event, bool detach_group)
1637 struct perf_event_context *ctx = event->ctx;
1638 struct task_struct *task = ctx->task;
1639 struct remove_event re = {
1641 .detach_group = detach_group,
1644 lockdep_assert_held(&ctx->mutex);
1648 * Per cpu events are removed via an smp call. The removal can
1649 * fail if the CPU is currently offline, but in that case we
1650 * already called __perf_remove_from_context from
1651 * perf_event_exit_cpu.
1653 cpu_function_call(event->cpu, __perf_remove_from_context, &re);
1658 if (!task_function_call(task, __perf_remove_from_context, &re))
1661 raw_spin_lock_irq(&ctx->lock);
1663 * If we failed to find a running task, but find the context active now
1664 * that we've acquired the ctx->lock, retry.
1666 if (ctx->is_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 the task isn't running, its safe to remove the event, us
1678 * holding the ctx->lock ensures the task won't get scheduled in.
1681 perf_group_detach(event);
1682 list_del_event(event, ctx);
1683 raw_spin_unlock_irq(&ctx->lock);
1687 * Cross CPU call to disable a performance event
1689 int __perf_event_disable(void *info)
1691 struct perf_event *event = info;
1692 struct perf_event_context *ctx = event->ctx;
1693 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
1696 * If this is a per-task event, need to check whether this
1697 * event's task is the current task on this cpu.
1699 * Can trigger due to concurrent perf_event_context_sched_out()
1700 * flipping contexts around.
1702 if (ctx->task && cpuctx->task_ctx != ctx)
1705 raw_spin_lock(&ctx->lock);
1708 * If the event is on, turn it off.
1709 * If it is in error state, leave it in error state.
1711 if (event->state >= PERF_EVENT_STATE_INACTIVE) {
1712 update_context_time(ctx);
1713 update_cgrp_time_from_event(event);
1714 update_group_times(event);
1715 if (event == event->group_leader)
1716 group_sched_out(event, cpuctx, ctx);
1718 event_sched_out(event, cpuctx, ctx);
1719 event->state = PERF_EVENT_STATE_OFF;
1722 raw_spin_unlock(&ctx->lock);
1730 * If event->ctx is a cloned context, callers must make sure that
1731 * every task struct that event->ctx->task could possibly point to
1732 * remains valid. This condition is satisifed when called through
1733 * perf_event_for_each_child or perf_event_for_each because they
1734 * hold the top-level event's child_mutex, so any descendant that
1735 * goes to exit will block in sync_child_event.
1736 * When called from perf_pending_event it's OK because event->ctx
1737 * is the current context on this CPU and preemption is disabled,
1738 * hence we can't get into perf_event_task_sched_out for this context.
1740 static void _perf_event_disable(struct perf_event *event)
1742 struct perf_event_context *ctx = event->ctx;
1743 struct task_struct *task = ctx->task;
1747 * Disable the event on the cpu that it's on
1749 cpu_function_call(event->cpu, __perf_event_disable, event);
1754 if (!task_function_call(task, __perf_event_disable, event))
1757 raw_spin_lock_irq(&ctx->lock);
1759 * If the event is still active, we need to retry the cross-call.
1761 if (event->state == PERF_EVENT_STATE_ACTIVE) {
1762 raw_spin_unlock_irq(&ctx->lock);
1764 * Reload the task pointer, it might have been changed by
1765 * a concurrent perf_event_context_sched_out().
1772 * Since we have the lock this context can't be scheduled
1773 * in, so we can change the state safely.
1775 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1776 update_group_times(event);
1777 event->state = PERF_EVENT_STATE_OFF;
1779 raw_spin_unlock_irq(&ctx->lock);
1783 * Strictly speaking kernel users cannot create groups and therefore this
1784 * interface does not need the perf_event_ctx_lock() magic.
1786 void perf_event_disable(struct perf_event *event)
1788 struct perf_event_context *ctx;
1790 ctx = perf_event_ctx_lock(event);
1791 _perf_event_disable(event);
1792 perf_event_ctx_unlock(event, ctx);
1794 EXPORT_SYMBOL_GPL(perf_event_disable);
1796 static void perf_set_shadow_time(struct perf_event *event,
1797 struct perf_event_context *ctx,
1801 * use the correct time source for the time snapshot
1803 * We could get by without this by leveraging the
1804 * fact that to get to this function, the caller
1805 * has most likely already called update_context_time()
1806 * and update_cgrp_time_xx() and thus both timestamp
1807 * are identical (or very close). Given that tstamp is,
1808 * already adjusted for cgroup, we could say that:
1809 * tstamp - ctx->timestamp
1811 * tstamp - cgrp->timestamp.
1813 * Then, in perf_output_read(), the calculation would
1814 * work with no changes because:
1815 * - event is guaranteed scheduled in
1816 * - no scheduled out in between
1817 * - thus the timestamp would be the same
1819 * But this is a bit hairy.
1821 * So instead, we have an explicit cgroup call to remain
1822 * within the time time source all along. We believe it
1823 * is cleaner and simpler to understand.
1825 if (is_cgroup_event(event))
1826 perf_cgroup_set_shadow_time(event, tstamp);
1828 event->shadow_ctx_time = tstamp - ctx->timestamp;
1831 #define MAX_INTERRUPTS (~0ULL)
1833 static void perf_log_throttle(struct perf_event *event, int enable);
1836 event_sched_in(struct perf_event *event,
1837 struct perf_cpu_context *cpuctx,
1838 struct perf_event_context *ctx)
1840 u64 tstamp = perf_event_time(event);
1843 lockdep_assert_held(&ctx->lock);
1845 if (event->state <= PERF_EVENT_STATE_OFF)
1848 event->state = PERF_EVENT_STATE_ACTIVE;
1849 event->oncpu = smp_processor_id();
1852 * Unthrottle events, since we scheduled we might have missed several
1853 * ticks already, also for a heavily scheduling task there is little
1854 * guarantee it'll get a tick in a timely manner.
1856 if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
1857 perf_log_throttle(event, 1);
1858 event->hw.interrupts = 0;
1862 * The new state must be visible before we turn it on in the hardware:
1866 perf_pmu_disable(event->pmu);
1868 if (event->pmu->add(event, PERF_EF_START)) {
1869 event->state = PERF_EVENT_STATE_INACTIVE;
1875 event->tstamp_running += tstamp - event->tstamp_stopped;
1877 perf_set_shadow_time(event, ctx, tstamp);
1879 if (!is_software_event(event))
1880 cpuctx->active_oncpu++;
1882 if (event->attr.freq && event->attr.sample_freq)
1885 if (event->attr.exclusive)
1886 cpuctx->exclusive = 1;
1888 if (is_orphaned_child(event))
1889 schedule_orphans_remove(ctx);
1892 perf_pmu_enable(event->pmu);
1898 group_sched_in(struct perf_event *group_event,
1899 struct perf_cpu_context *cpuctx,
1900 struct perf_event_context *ctx)
1902 struct perf_event *event, *partial_group = NULL;
1903 struct pmu *pmu = ctx->pmu;
1904 u64 now = ctx->time;
1905 bool simulate = false;
1907 if (group_event->state == PERF_EVENT_STATE_OFF)
1910 pmu->start_txn(pmu);
1912 if (event_sched_in(group_event, cpuctx, ctx)) {
1913 pmu->cancel_txn(pmu);
1914 perf_cpu_hrtimer_restart(cpuctx);
1919 * Schedule in siblings as one group (if any):
1921 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1922 if (event_sched_in(event, cpuctx, ctx)) {
1923 partial_group = event;
1928 if (!pmu->commit_txn(pmu))
1933 * Groups can be scheduled in as one unit only, so undo any
1934 * partial group before returning:
1935 * The events up to the failed event are scheduled out normally,
1936 * tstamp_stopped will be updated.
1938 * The failed events and the remaining siblings need to have
1939 * their timings updated as if they had gone thru event_sched_in()
1940 * and event_sched_out(). This is required to get consistent timings
1941 * across the group. This also takes care of the case where the group
1942 * could never be scheduled by ensuring tstamp_stopped is set to mark
1943 * the time the event was actually stopped, such that time delta
1944 * calculation in update_event_times() is correct.
1946 list_for_each_entry(event, &group_event->sibling_list, group_entry) {
1947 if (event == partial_group)
1951 event->tstamp_running += now - event->tstamp_stopped;
1952 event->tstamp_stopped = now;
1954 event_sched_out(event, cpuctx, ctx);
1957 event_sched_out(group_event, cpuctx, ctx);
1959 pmu->cancel_txn(pmu);
1961 perf_cpu_hrtimer_restart(cpuctx);
1967 * Work out whether we can put this event group on the CPU now.
1969 static int group_can_go_on(struct perf_event *event,
1970 struct perf_cpu_context *cpuctx,
1974 * Groups consisting entirely of software events can always go on.
1976 if (event->group_flags & PERF_GROUP_SOFTWARE)
1979 * If an exclusive group is already on, no other hardware
1982 if (cpuctx->exclusive)
1985 * If this group is exclusive and there are already
1986 * events on the CPU, it can't go on.
1988 if (event->attr.exclusive && cpuctx->active_oncpu)
1991 * Otherwise, try to add it if all previous groups were able
1997 static void add_event_to_ctx(struct perf_event *event,
1998 struct perf_event_context *ctx)
2000 u64 tstamp = perf_event_time(event);
2002 list_add_event(event, ctx);
2003 perf_group_attach(event);
2004 event->tstamp_enabled = tstamp;
2005 event->tstamp_running = tstamp;
2006 event->tstamp_stopped = tstamp;
2009 static void task_ctx_sched_out(struct perf_event_context *ctx);
2011 ctx_sched_in(struct perf_event_context *ctx,
2012 struct perf_cpu_context *cpuctx,
2013 enum event_type_t event_type,
2014 struct task_struct *task);
2016 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2017 struct perf_event_context *ctx,
2018 struct task_struct *task)
2020 cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2022 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2023 cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2025 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2029 * Cross CPU call to install and enable a performance event
2031 * Must be called with ctx->mutex held
2033 static int __perf_install_in_context(void *info)
2035 struct perf_event *event = info;
2036 struct perf_event_context *ctx = event->ctx;
2037 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2038 struct perf_event_context *task_ctx = cpuctx->task_ctx;
2039 struct task_struct *task = current;
2041 perf_ctx_lock(cpuctx, task_ctx);
2042 perf_pmu_disable(cpuctx->ctx.pmu);
2045 * If there was an active task_ctx schedule it out.
2048 task_ctx_sched_out(task_ctx);
2051 * If the context we're installing events in is not the
2052 * active task_ctx, flip them.
2054 if (ctx->task && task_ctx != ctx) {
2056 raw_spin_unlock(&task_ctx->lock);
2057 raw_spin_lock(&ctx->lock);
2062 cpuctx->task_ctx = task_ctx;
2063 task = task_ctx->task;
2066 cpu_ctx_sched_out(cpuctx, EVENT_ALL);
2068 update_context_time(ctx);
2070 * update cgrp time only if current cgrp
2071 * matches event->cgrp. Must be done before
2072 * calling add_event_to_ctx()
2074 update_cgrp_time_from_event(event);
2076 add_event_to_ctx(event, ctx);
2079 * Schedule everything back in
2081 perf_event_sched_in(cpuctx, task_ctx, task);
2083 perf_pmu_enable(cpuctx->ctx.pmu);
2084 perf_ctx_unlock(cpuctx, task_ctx);
2090 * Attach a performance event to a context
2092 * First we add the event to the list with the hardware enable bit
2093 * in event->hw_config cleared.
2095 * If the event is attached to a task which is on a CPU we use a smp
2096 * call to enable it in the task context. The task might have been
2097 * scheduled away, but we check this in the smp call again.
2100 perf_install_in_context(struct perf_event_context *ctx,
2101 struct perf_event *event,
2104 struct task_struct *task = ctx->task;
2106 lockdep_assert_held(&ctx->mutex);
2109 if (event->cpu != -1)
2114 * Per cpu events are installed via an smp call and
2115 * the install is always successful.
2117 cpu_function_call(cpu, __perf_install_in_context, event);
2122 if (!task_function_call(task, __perf_install_in_context, event))
2125 raw_spin_lock_irq(&ctx->lock);
2127 * If we failed to find a running task, but find the context active now
2128 * that we've acquired the ctx->lock, retry.
2130 if (ctx->is_active) {
2131 raw_spin_unlock_irq(&ctx->lock);
2133 * Reload the task pointer, it might have been changed by
2134 * a concurrent perf_event_context_sched_out().
2141 * Since the task isn't running, its safe to add the event, us holding
2142 * the ctx->lock ensures the task won't get scheduled in.
2144 add_event_to_ctx(event, ctx);
2145 raw_spin_unlock_irq(&ctx->lock);
2149 * Put a event into inactive state and update time fields.
2150 * Enabling the leader of a group effectively enables all
2151 * the group members that aren't explicitly disabled, so we
2152 * have to update their ->tstamp_enabled also.
2153 * Note: this works for group members as well as group leaders
2154 * since the non-leader members' sibling_lists will be empty.
2156 static void __perf_event_mark_enabled(struct perf_event *event)
2158 struct perf_event *sub;
2159 u64 tstamp = perf_event_time(event);
2161 event->state = PERF_EVENT_STATE_INACTIVE;
2162 event->tstamp_enabled = tstamp - event->total_time_enabled;
2163 list_for_each_entry(sub, &event->sibling_list, group_entry) {
2164 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
2165 sub->tstamp_enabled = tstamp - sub->total_time_enabled;
2170 * Cross CPU call to enable a performance event
2172 static int __perf_event_enable(void *info)
2174 struct perf_event *event = info;
2175 struct perf_event_context *ctx = event->ctx;
2176 struct perf_event *leader = event->group_leader;
2177 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2181 * There's a time window between 'ctx->is_active' check
2182 * in perf_event_enable function and this place having:
2184 * - ctx->lock unlocked
2186 * where the task could be killed and 'ctx' deactivated
2187 * by perf_event_exit_task.
2189 if (!ctx->is_active)
2192 raw_spin_lock(&ctx->lock);
2193 update_context_time(ctx);
2195 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2199 * set current task's cgroup time reference point
2201 perf_cgroup_set_timestamp(current, ctx);
2203 __perf_event_mark_enabled(event);
2205 if (!event_filter_match(event)) {
2206 if (is_cgroup_event(event))
2207 perf_cgroup_defer_enabled(event);
2212 * If the event is in a group and isn't the group leader,
2213 * then don't put it on unless the group is on.
2215 if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
2218 if (!group_can_go_on(event, cpuctx, 1)) {
2221 if (event == leader)
2222 err = group_sched_in(event, cpuctx, ctx);
2224 err = event_sched_in(event, cpuctx, ctx);
2229 * If this event can't go on and it's part of a
2230 * group, then the whole group has to come off.
2232 if (leader != event) {
2233 group_sched_out(leader, cpuctx, ctx);
2234 perf_cpu_hrtimer_restart(cpuctx);
2236 if (leader->attr.pinned) {
2237 update_group_times(leader);
2238 leader->state = PERF_EVENT_STATE_ERROR;
2243 raw_spin_unlock(&ctx->lock);
2251 * If event->ctx is a cloned context, callers must make sure that
2252 * every task struct that event->ctx->task could possibly point to
2253 * remains valid. This condition is satisfied when called through
2254 * perf_event_for_each_child or perf_event_for_each as described
2255 * for perf_event_disable.
2257 static void _perf_event_enable(struct perf_event *event)
2259 struct perf_event_context *ctx = event->ctx;
2260 struct task_struct *task = ctx->task;
2264 * Enable the event on the cpu that it's on
2266 cpu_function_call(event->cpu, __perf_event_enable, event);
2270 raw_spin_lock_irq(&ctx->lock);
2271 if (event->state >= PERF_EVENT_STATE_INACTIVE)
2275 * If the event is in error state, clear that first.
2276 * That way, if we see the event in error state below, we
2277 * know that it has gone back into error state, as distinct
2278 * from the task having been scheduled away before the
2279 * cross-call arrived.
2281 if (event->state == PERF_EVENT_STATE_ERROR)
2282 event->state = PERF_EVENT_STATE_OFF;
2285 if (!ctx->is_active) {
2286 __perf_event_mark_enabled(event);
2290 raw_spin_unlock_irq(&ctx->lock);
2292 if (!task_function_call(task, __perf_event_enable, event))
2295 raw_spin_lock_irq(&ctx->lock);
2298 * If the context is active and the event is still off,
2299 * we need to retry the cross-call.
2301 if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
2303 * task could have been flipped by a concurrent
2304 * perf_event_context_sched_out()
2311 raw_spin_unlock_irq(&ctx->lock);
2315 * See perf_event_disable();
2317 void perf_event_enable(struct perf_event *event)
2319 struct perf_event_context *ctx;
2321 ctx = perf_event_ctx_lock(event);
2322 _perf_event_enable(event);
2323 perf_event_ctx_unlock(event, ctx);
2325 EXPORT_SYMBOL_GPL(perf_event_enable);
2327 static int _perf_event_refresh(struct perf_event *event, int refresh)
2330 * not supported on inherited events
2332 if (event->attr.inherit || !is_sampling_event(event))
2335 atomic_add(refresh, &event->event_limit);
2336 _perf_event_enable(event);
2342 * See perf_event_disable()
2344 int perf_event_refresh(struct perf_event *event, int refresh)
2346 struct perf_event_context *ctx;
2349 ctx = perf_event_ctx_lock(event);
2350 ret = _perf_event_refresh(event, refresh);
2351 perf_event_ctx_unlock(event, ctx);
2355 EXPORT_SYMBOL_GPL(perf_event_refresh);
2357 static void ctx_sched_out(struct perf_event_context *ctx,
2358 struct perf_cpu_context *cpuctx,
2359 enum event_type_t event_type)
2361 struct perf_event *event;
2362 int is_active = ctx->is_active;
2364 ctx->is_active &= ~event_type;
2365 if (likely(!ctx->nr_events))
2368 update_context_time(ctx);
2369 update_cgrp_time_from_cpuctx(cpuctx);
2370 if (!ctx->nr_active)
2373 perf_pmu_disable(ctx->pmu);
2374 if ((is_active & EVENT_PINNED) && (event_type & EVENT_PINNED)) {
2375 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2376 group_sched_out(event, cpuctx, ctx);
2379 if ((is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE)) {
2380 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2381 group_sched_out(event, cpuctx, ctx);
2383 perf_pmu_enable(ctx->pmu);
2387 * Test whether two contexts are equivalent, i.e. whether they have both been
2388 * cloned from the same version of the same context.
2390 * Equivalence is measured using a generation number in the context that is
2391 * incremented on each modification to it; see unclone_ctx(), list_add_event()
2392 * and list_del_event().
2394 static int context_equiv(struct perf_event_context *ctx1,
2395 struct perf_event_context *ctx2)
2397 lockdep_assert_held(&ctx1->lock);
2398 lockdep_assert_held(&ctx2->lock);
2400 /* Pinning disables the swap optimization */
2401 if (ctx1->pin_count || ctx2->pin_count)
2404 /* If ctx1 is the parent of ctx2 */
2405 if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2408 /* If ctx2 is the parent of ctx1 */
2409 if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2413 * If ctx1 and ctx2 have the same parent; we flatten the parent
2414 * hierarchy, see perf_event_init_context().
2416 if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2417 ctx1->parent_gen == ctx2->parent_gen)
2424 static void __perf_event_sync_stat(struct perf_event *event,
2425 struct perf_event *next_event)
2429 if (!event->attr.inherit_stat)
2433 * Update the event value, we cannot use perf_event_read()
2434 * because we're in the middle of a context switch and have IRQs
2435 * disabled, which upsets smp_call_function_single(), however
2436 * we know the event must be on the current CPU, therefore we
2437 * don't need to use it.
2439 switch (event->state) {
2440 case PERF_EVENT_STATE_ACTIVE:
2441 event->pmu->read(event);
2444 case PERF_EVENT_STATE_INACTIVE:
2445 update_event_times(event);
2453 * In order to keep per-task stats reliable we need to flip the event
2454 * values when we flip the contexts.
2456 value = local64_read(&next_event->count);
2457 value = local64_xchg(&event->count, value);
2458 local64_set(&next_event->count, value);
2460 swap(event->total_time_enabled, next_event->total_time_enabled);
2461 swap(event->total_time_running, next_event->total_time_running);
2464 * Since we swizzled the values, update the user visible data too.
2466 perf_event_update_userpage(event);
2467 perf_event_update_userpage(next_event);
2470 static void perf_event_sync_stat(struct perf_event_context *ctx,
2471 struct perf_event_context *next_ctx)
2473 struct perf_event *event, *next_event;
2478 update_context_time(ctx);
2480 event = list_first_entry(&ctx->event_list,
2481 struct perf_event, event_entry);
2483 next_event = list_first_entry(&next_ctx->event_list,
2484 struct perf_event, event_entry);
2486 while (&event->event_entry != &ctx->event_list &&
2487 &next_event->event_entry != &next_ctx->event_list) {
2489 __perf_event_sync_stat(event, next_event);
2491 event = list_next_entry(event, event_entry);
2492 next_event = list_next_entry(next_event, event_entry);
2496 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2497 struct task_struct *next)
2499 struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2500 struct perf_event_context *next_ctx;
2501 struct perf_event_context *parent, *next_parent;
2502 struct perf_cpu_context *cpuctx;
2508 cpuctx = __get_cpu_context(ctx);
2509 if (!cpuctx->task_ctx)
2513 next_ctx = next->perf_event_ctxp[ctxn];
2517 parent = rcu_dereference(ctx->parent_ctx);
2518 next_parent = rcu_dereference(next_ctx->parent_ctx);
2520 /* If neither context have a parent context; they cannot be clones. */
2521 if (!parent && !next_parent)
2524 if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2526 * Looks like the two contexts are clones, so we might be
2527 * able to optimize the context switch. We lock both
2528 * contexts and check that they are clones under the
2529 * lock (including re-checking that neither has been
2530 * uncloned in the meantime). It doesn't matter which
2531 * order we take the locks because no other cpu could
2532 * be trying to lock both of these tasks.
2534 raw_spin_lock(&ctx->lock);
2535 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2536 if (context_equiv(ctx, next_ctx)) {
2538 * XXX do we need a memory barrier of sorts
2539 * wrt to rcu_dereference() of perf_event_ctxp
2541 task->perf_event_ctxp[ctxn] = next_ctx;
2542 next->perf_event_ctxp[ctxn] = ctx;
2544 next_ctx->task = task;
2547 perf_event_sync_stat(ctx, next_ctx);
2549 raw_spin_unlock(&next_ctx->lock);
2550 raw_spin_unlock(&ctx->lock);
2556 raw_spin_lock(&ctx->lock);
2557 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2558 cpuctx->task_ctx = NULL;
2559 raw_spin_unlock(&ctx->lock);
2563 #define for_each_task_context_nr(ctxn) \
2564 for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2567 * Called from scheduler to remove the events of the current task,
2568 * with interrupts disabled.
2570 * We stop each event and update the event value in event->count.
2572 * This does not protect us against NMI, but disable()
2573 * sets the disabled bit in the control field of event _before_
2574 * accessing the event control register. If a NMI hits, then it will
2575 * not restart the event.
2577 void __perf_event_task_sched_out(struct task_struct *task,
2578 struct task_struct *next)
2582 for_each_task_context_nr(ctxn)
2583 perf_event_context_sched_out(task, ctxn, next);
2586 * if cgroup events exist on this CPU, then we need
2587 * to check if we have to switch out PMU state.
2588 * cgroup event are system-wide mode only
2590 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2591 perf_cgroup_sched_out(task, next);
2594 static void task_ctx_sched_out(struct perf_event_context *ctx)
2596 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2598 if (!cpuctx->task_ctx)
2601 if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2604 ctx_sched_out(ctx, cpuctx, EVENT_ALL);
2605 cpuctx->task_ctx = NULL;
2609 * Called with IRQs disabled
2611 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2612 enum event_type_t event_type)
2614 ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2618 ctx_pinned_sched_in(struct perf_event_context *ctx,
2619 struct perf_cpu_context *cpuctx)
2621 struct perf_event *event;
2623 list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
2624 if (event->state <= PERF_EVENT_STATE_OFF)
2626 if (!event_filter_match(event))
2629 /* may need to reset tstamp_enabled */
2630 if (is_cgroup_event(event))
2631 perf_cgroup_mark_enabled(event, ctx);
2633 if (group_can_go_on(event, cpuctx, 1))
2634 group_sched_in(event, cpuctx, ctx);
2637 * If this pinned group hasn't been scheduled,
2638 * put it in error state.
2640 if (event->state == PERF_EVENT_STATE_INACTIVE) {
2641 update_group_times(event);
2642 event->state = PERF_EVENT_STATE_ERROR;
2648 ctx_flexible_sched_in(struct perf_event_context *ctx,
2649 struct perf_cpu_context *cpuctx)
2651 struct perf_event *event;
2654 list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
2655 /* Ignore events in OFF or ERROR state */
2656 if (event->state <= PERF_EVENT_STATE_OFF)
2659 * Listen to the 'cpu' scheduling filter constraint
2662 if (!event_filter_match(event))
2665 /* may need to reset tstamp_enabled */
2666 if (is_cgroup_event(event))
2667 perf_cgroup_mark_enabled(event, ctx);
2669 if (group_can_go_on(event, cpuctx, can_add_hw)) {
2670 if (group_sched_in(event, cpuctx, ctx))
2677 ctx_sched_in(struct perf_event_context *ctx,
2678 struct perf_cpu_context *cpuctx,
2679 enum event_type_t event_type,
2680 struct task_struct *task)
2683 int is_active = ctx->is_active;
2685 ctx->is_active |= event_type;
2686 if (likely(!ctx->nr_events))
2690 ctx->timestamp = now;
2691 perf_cgroup_set_timestamp(task, ctx);
2693 * First go through the list and put on any pinned groups
2694 * in order to give them the best chance of going on.
2696 if (!(is_active & EVENT_PINNED) && (event_type & EVENT_PINNED))
2697 ctx_pinned_sched_in(ctx, cpuctx);
2699 /* Then walk through the lower prio flexible groups */
2700 if (!(is_active & EVENT_FLEXIBLE) && (event_type & EVENT_FLEXIBLE))
2701 ctx_flexible_sched_in(ctx, cpuctx);
2704 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
2705 enum event_type_t event_type,
2706 struct task_struct *task)
2708 struct perf_event_context *ctx = &cpuctx->ctx;
2710 ctx_sched_in(ctx, cpuctx, event_type, task);
2713 static void perf_event_context_sched_in(struct perf_event_context *ctx,
2714 struct task_struct *task)
2716 struct perf_cpu_context *cpuctx;
2718 cpuctx = __get_cpu_context(ctx);
2719 if (cpuctx->task_ctx == ctx)
2722 perf_ctx_lock(cpuctx, ctx);
2723 perf_pmu_disable(ctx->pmu);
2725 * We want to keep the following priority order:
2726 * cpu pinned (that don't need to move), task pinned,
2727 * cpu flexible, task flexible.
2729 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2732 cpuctx->task_ctx = ctx;
2734 perf_event_sched_in(cpuctx, cpuctx->task_ctx, task);
2736 perf_pmu_enable(ctx->pmu);
2737 perf_ctx_unlock(cpuctx, ctx);
2740 * Since these rotations are per-cpu, we need to ensure the
2741 * cpu-context we got scheduled on is actually rotating.
2743 perf_pmu_rotate_start(ctx->pmu);
2747 * When sampling the branck stack in system-wide, it may be necessary
2748 * to flush the stack on context switch. This happens when the branch
2749 * stack does not tag its entries with the pid of the current task.
2750 * Otherwise it becomes impossible to associate a branch entry with a
2751 * task. This ambiguity is more likely to appear when the branch stack
2752 * supports priv level filtering and the user sets it to monitor only
2753 * at the user level (which could be a useful measurement in system-wide
2754 * mode). In that case, the risk is high of having a branch stack with
2755 * branch from multiple tasks. Flushing may mean dropping the existing
2756 * entries or stashing them somewhere in the PMU specific code layer.
2758 * This function provides the context switch callback to the lower code
2759 * layer. It is invoked ONLY when there is at least one system-wide context
2760 * with at least one active event using taken branch sampling.
2762 static void perf_branch_stack_sched_in(struct task_struct *prev,
2763 struct task_struct *task)
2765 struct perf_cpu_context *cpuctx;
2767 unsigned long flags;
2769 /* no need to flush branch stack if not changing task */
2773 local_irq_save(flags);
2777 list_for_each_entry_rcu(pmu, &pmus, entry) {
2778 cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2781 * check if the context has at least one
2782 * event using PERF_SAMPLE_BRANCH_STACK
2784 if (cpuctx->ctx.nr_branch_stack > 0
2785 && pmu->flush_branch_stack) {
2787 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2789 perf_pmu_disable(pmu);
2791 pmu->flush_branch_stack();
2793 perf_pmu_enable(pmu);
2795 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2801 local_irq_restore(flags);
2805 * Called from scheduler to add the events of the current task
2806 * with interrupts disabled.
2808 * We restore the event value and then enable it.
2810 * This does not protect us against NMI, but enable()
2811 * sets the enabled bit in the control field of event _before_
2812 * accessing the event control register. If a NMI hits, then it will
2813 * keep the event running.
2815 void __perf_event_task_sched_in(struct task_struct *prev,
2816 struct task_struct *task)
2818 struct perf_event_context *ctx;
2821 for_each_task_context_nr(ctxn) {
2822 ctx = task->perf_event_ctxp[ctxn];
2826 perf_event_context_sched_in(ctx, task);
2829 * if cgroup events exist on this CPU, then we need
2830 * to check if we have to switch in PMU state.
2831 * cgroup event are system-wide mode only
2833 if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2834 perf_cgroup_sched_in(prev, task);
2836 /* check for system-wide branch_stack events */
2837 if (atomic_read(this_cpu_ptr(&perf_branch_stack_events)))
2838 perf_branch_stack_sched_in(prev, task);
2841 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
2843 u64 frequency = event->attr.sample_freq;
2844 u64 sec = NSEC_PER_SEC;
2845 u64 divisor, dividend;
2847 int count_fls, nsec_fls, frequency_fls, sec_fls;
2849 count_fls = fls64(count);
2850 nsec_fls = fls64(nsec);
2851 frequency_fls = fls64(frequency);
2855 * We got @count in @nsec, with a target of sample_freq HZ
2856 * the target period becomes:
2859 * period = -------------------
2860 * @nsec * sample_freq
2865 * Reduce accuracy by one bit such that @a and @b converge
2866 * to a similar magnitude.
2868 #define REDUCE_FLS(a, b) \
2870 if (a##_fls > b##_fls) { \
2880 * Reduce accuracy until either term fits in a u64, then proceed with
2881 * the other, so that finally we can do a u64/u64 division.
2883 while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
2884 REDUCE_FLS(nsec, frequency);
2885 REDUCE_FLS(sec, count);
2888 if (count_fls + sec_fls > 64) {
2889 divisor = nsec * frequency;
2891 while (count_fls + sec_fls > 64) {
2892 REDUCE_FLS(count, sec);
2896 dividend = count * sec;
2898 dividend = count * sec;
2900 while (nsec_fls + frequency_fls > 64) {
2901 REDUCE_FLS(nsec, frequency);
2905 divisor = nsec * frequency;
2911 return div64_u64(dividend, divisor);
2914 static DEFINE_PER_CPU(int, perf_throttled_count);
2915 static DEFINE_PER_CPU(u64, perf_throttled_seq);
2917 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
2919 struct hw_perf_event *hwc = &event->hw;
2920 s64 period, sample_period;
2923 period = perf_calculate_period(event, nsec, count);
2925 delta = (s64)(period - hwc->sample_period);
2926 delta = (delta + 7) / 8; /* low pass filter */
2928 sample_period = hwc->sample_period + delta;
2933 hwc->sample_period = sample_period;
2935 if (local64_read(&hwc->period_left) > 8*sample_period) {
2937 event->pmu->stop(event, PERF_EF_UPDATE);
2939 local64_set(&hwc->period_left, 0);
2942 event->pmu->start(event, PERF_EF_RELOAD);
2947 * combine freq adjustment with unthrottling to avoid two passes over the
2948 * events. At the same time, make sure, having freq events does not change
2949 * the rate of unthrottling as that would introduce bias.
2951 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
2954 struct perf_event *event;
2955 struct hw_perf_event *hwc;
2956 u64 now, period = TICK_NSEC;
2960 * only need to iterate over all events iff:
2961 * - context have events in frequency mode (needs freq adjust)
2962 * - there are events to unthrottle on this cpu
2964 if (!(ctx->nr_freq || needs_unthr))
2967 raw_spin_lock(&ctx->lock);
2968 perf_pmu_disable(ctx->pmu);
2970 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
2971 if (event->state != PERF_EVENT_STATE_ACTIVE)
2974 if (!event_filter_match(event))
2977 perf_pmu_disable(event->pmu);
2981 if (hwc->interrupts == MAX_INTERRUPTS) {
2982 hwc->interrupts = 0;
2983 perf_log_throttle(event, 1);
2984 event->pmu->start(event, 0);
2987 if (!event->attr.freq || !event->attr.sample_freq)
2991 * stop the event and update event->count
2993 event->pmu->stop(event, PERF_EF_UPDATE);
2995 now = local64_read(&event->count);
2996 delta = now - hwc->freq_count_stamp;
2997 hwc->freq_count_stamp = now;
3001 * reload only if value has changed
3002 * we have stopped the event so tell that
3003 * to perf_adjust_period() to avoid stopping it
3007 perf_adjust_period(event, period, delta, false);
3009 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3011 perf_pmu_enable(event->pmu);
3014 perf_pmu_enable(ctx->pmu);
3015 raw_spin_unlock(&ctx->lock);
3019 * Round-robin a context's events:
3021 static void rotate_ctx(struct perf_event_context *ctx)
3024 * Rotate the first entry last of non-pinned groups. Rotation might be
3025 * disabled by the inheritance code.
3027 if (!ctx->rotate_disable)
3028 list_rotate_left(&ctx->flexible_groups);
3032 * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
3033 * because they're strictly cpu affine and rotate_start is called with IRQs
3034 * disabled, while rotate_context is called from IRQ context.
3036 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
3038 struct perf_event_context *ctx = NULL;
3039 int rotate = 0, remove = 1;
3041 if (cpuctx->ctx.nr_events) {
3043 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3047 ctx = cpuctx->task_ctx;
3048 if (ctx && ctx->nr_events) {
3050 if (ctx->nr_events != ctx->nr_active)
3057 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3058 perf_pmu_disable(cpuctx->ctx.pmu);
3060 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3062 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3064 rotate_ctx(&cpuctx->ctx);
3068 perf_event_sched_in(cpuctx, ctx, current);
3070 perf_pmu_enable(cpuctx->ctx.pmu);
3071 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3074 list_del_init(&cpuctx->rotation_list);
3079 #ifdef CONFIG_NO_HZ_FULL
3080 bool perf_event_can_stop_tick(void)
3082 if (atomic_read(&nr_freq_events) ||
3083 __this_cpu_read(perf_throttled_count))
3090 void perf_event_task_tick(void)
3092 struct list_head *head = this_cpu_ptr(&rotation_list);
3093 struct perf_cpu_context *cpuctx, *tmp;
3094 struct perf_event_context *ctx;
3097 WARN_ON(!irqs_disabled());
3099 __this_cpu_inc(perf_throttled_seq);
3100 throttled = __this_cpu_xchg(perf_throttled_count, 0);
3102 list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
3104 perf_adjust_freq_unthr_context(ctx, throttled);
3106 ctx = cpuctx->task_ctx;
3108 perf_adjust_freq_unthr_context(ctx, throttled);
3112 static int event_enable_on_exec(struct perf_event *event,
3113 struct perf_event_context *ctx)
3115 if (!event->attr.enable_on_exec)
3118 event->attr.enable_on_exec = 0;
3119 if (event->state >= PERF_EVENT_STATE_INACTIVE)
3122 __perf_event_mark_enabled(event);
3128 * Enable all of a task's events that have been marked enable-on-exec.
3129 * This expects task == current.
3131 static void perf_event_enable_on_exec(struct perf_event_context *ctx)
3133 struct perf_event_context *clone_ctx = NULL;
3134 struct perf_event *event;
3135 unsigned long flags;
3139 local_irq_save(flags);
3140 if (!ctx || !ctx->nr_events)
3144 * We must ctxsw out cgroup events to avoid conflict
3145 * when invoking perf_task_event_sched_in() later on
3146 * in this function. Otherwise we end up trying to
3147 * ctxswin cgroup events which are already scheduled
3150 perf_cgroup_sched_out(current, NULL);
3152 raw_spin_lock(&ctx->lock);
3153 task_ctx_sched_out(ctx);
3155 list_for_each_entry(event, &ctx->event_list, event_entry) {
3156 ret = event_enable_on_exec(event, ctx);
3162 * Unclone this context if we enabled any event.
3165 clone_ctx = unclone_ctx(ctx);
3167 raw_spin_unlock(&ctx->lock);
3170 * Also calls ctxswin for cgroup events, if any:
3172 perf_event_context_sched_in(ctx, ctx->task);
3174 local_irq_restore(flags);
3180 void perf_event_exec(void)
3182 struct perf_event_context *ctx;
3186 for_each_task_context_nr(ctxn) {
3187 ctx = current->perf_event_ctxp[ctxn];
3191 perf_event_enable_on_exec(ctx);
3197 * Cross CPU call to read the hardware event
3199 static void __perf_event_read(void *info)
3201 struct perf_event *event = info;
3202 struct perf_event_context *ctx = event->ctx;
3203 struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3206 * If this is a task context, we need to check whether it is
3207 * the current task context of this cpu. If not it has been
3208 * scheduled out before the smp call arrived. In that case
3209 * event->count would have been updated to a recent sample
3210 * when the event was scheduled out.
3212 if (ctx->task && cpuctx->task_ctx != ctx)
3215 raw_spin_lock(&ctx->lock);
3216 if (ctx->is_active) {
3217 update_context_time(ctx);
3218 update_cgrp_time_from_event(event);
3220 update_event_times(event);
3221 if (event->state == PERF_EVENT_STATE_ACTIVE)
3222 event->pmu->read(event);
3223 raw_spin_unlock(&ctx->lock);
3226 static inline u64 perf_event_count(struct perf_event *event)
3228 return local64_read(&event->count) + atomic64_read(&event->child_count);
3231 static u64 perf_event_read(struct perf_event *event)
3234 * If event is enabled and currently active on a CPU, update the
3235 * value in the event structure:
3237 if (event->state == PERF_EVENT_STATE_ACTIVE) {
3238 smp_call_function_single(event->oncpu,
3239 __perf_event_read, event, 1);
3240 } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
3241 struct perf_event_context *ctx = event->ctx;
3242 unsigned long flags;
3244 raw_spin_lock_irqsave(&ctx->lock, flags);
3246 * may read while context is not active
3247 * (e.g., thread is blocked), in that case
3248 * we cannot update context time
3250 if (ctx->is_active) {
3251 update_context_time(ctx);
3252 update_cgrp_time_from_event(event);
3254 update_event_times(event);
3255 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3258 return perf_event_count(event);
3262 * Initialize the perf_event context in a task_struct:
3264 static void __perf_event_init_context(struct perf_event_context *ctx)
3266 raw_spin_lock_init(&ctx->lock);
3267 mutex_init(&ctx->mutex);
3268 INIT_LIST_HEAD(&ctx->pinned_groups);
3269 INIT_LIST_HEAD(&ctx->flexible_groups);
3270 INIT_LIST_HEAD(&ctx->event_list);
3271 atomic_set(&ctx->refcount, 1);
3272 INIT_DELAYED_WORK(&ctx->orphans_remove, orphans_remove_work);
3275 static struct perf_event_context *
3276 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3278 struct perf_event_context *ctx;
3280 ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3284 __perf_event_init_context(ctx);
3287 get_task_struct(task);
3294 static struct task_struct *
3295 find_lively_task_by_vpid(pid_t vpid)
3297 struct task_struct *task;
3304 task = find_task_by_vpid(vpid);
3306 get_task_struct(task);
3310 return ERR_PTR(-ESRCH);
3312 /* Reuse ptrace permission checks for now. */
3314 if (!ptrace_may_access(task, PTRACE_MODE_READ))
3319 put_task_struct(task);
3320 return ERR_PTR(err);
3325 * Returns a matching context with refcount and pincount.
3327 static struct perf_event_context *
3328 find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
3330 struct perf_event_context *ctx, *clone_ctx = NULL;
3331 struct perf_cpu_context *cpuctx;
3332 unsigned long flags;
3336 /* Must be root to operate on a CPU event: */
3337 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3338 return ERR_PTR(-EACCES);
3341 * We could be clever and allow to attach a event to an
3342 * offline CPU and activate it when the CPU comes up, but
3345 if (!cpu_online(cpu))
3346 return ERR_PTR(-ENODEV);
3348 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3357 ctxn = pmu->task_ctx_nr;
3362 ctx = perf_lock_task_context(task, ctxn, &flags);
3364 clone_ctx = unclone_ctx(ctx);
3366 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3371 ctx = alloc_perf_context(pmu, task);
3377 mutex_lock(&task->perf_event_mutex);
3379 * If it has already passed perf_event_exit_task().
3380 * we must see PF_EXITING, it takes this mutex too.
3382 if (task->flags & PF_EXITING)
3384 else if (task->perf_event_ctxp[ctxn])
3389 rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3391 mutex_unlock(&task->perf_event_mutex);
3393 if (unlikely(err)) {
3405 return ERR_PTR(err);
3408 static void perf_event_free_filter(struct perf_event *event);
3410 static void free_event_rcu(struct rcu_head *head)
3412 struct perf_event *event;
3414 event = container_of(head, struct perf_event, rcu_head);
3416 put_pid_ns(event->ns);
3417 perf_event_free_filter(event);
3421 static void ring_buffer_put(struct ring_buffer *rb);
3422 static void ring_buffer_attach(struct perf_event *event,
3423 struct ring_buffer *rb);
3425 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3430 if (has_branch_stack(event)) {
3431 if (!(event->attach_state & PERF_ATTACH_TASK))
3432 atomic_dec(&per_cpu(perf_branch_stack_events, cpu));
3434 if (is_cgroup_event(event))
3435 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3438 static void unaccount_event(struct perf_event *event)
3443 if (event->attach_state & PERF_ATTACH_TASK)
3444 static_key_slow_dec_deferred(&perf_sched_events);
3445 if (event->attr.mmap || event->attr.mmap_data)
3446 atomic_dec(&nr_mmap_events);
3447 if (event->attr.comm)
3448 atomic_dec(&nr_comm_events);
3449 if (event->attr.task)
3450 atomic_dec(&nr_task_events);
3451 if (event->attr.freq)
3452 atomic_dec(&nr_freq_events);
3453 if (is_cgroup_event(event))
3454 static_key_slow_dec_deferred(&perf_sched_events);
3455 if (has_branch_stack(event))
3456 static_key_slow_dec_deferred(&perf_sched_events);
3458 unaccount_event_cpu(event, event->cpu);
3461 static void __free_event(struct perf_event *event)
3463 if (!event->parent) {
3464 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
3465 put_callchain_buffers();
3469 event->destroy(event);
3472 put_ctx(event->ctx);
3475 module_put(event->pmu->module);
3477 call_rcu(&event->rcu_head, free_event_rcu);
3480 static void _free_event(struct perf_event *event)
3482 irq_work_sync(&event->pending);
3484 unaccount_event(event);
3488 * Can happen when we close an event with re-directed output.
3490 * Since we have a 0 refcount, perf_mmap_close() will skip
3491 * over us; possibly making our ring_buffer_put() the last.
3493 mutex_lock(&event->mmap_mutex);
3494 ring_buffer_attach(event, NULL);
3495 mutex_unlock(&event->mmap_mutex);
3498 if (is_cgroup_event(event))
3499 perf_detach_cgroup(event);
3501 __free_event(event);
3505 * Used to free events which have a known refcount of 1, such as in error paths
3506 * where the event isn't exposed yet and inherited events.
3508 static void free_event(struct perf_event *event)
3510 if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
3511 "unexpected event refcount: %ld; ptr=%p\n",
3512 atomic_long_read(&event->refcount), event)) {
3513 /* leak to avoid use-after-free */
3521 * Remove user event from the owner task.
3523 static void perf_remove_from_owner(struct perf_event *event)
3525 struct task_struct *owner;
3528 owner = ACCESS_ONCE(event->owner);
3530 * Matches the smp_wmb() in perf_event_exit_task(). If we observe
3531 * !owner it means the list deletion is complete and we can indeed
3532 * free this event, otherwise we need to serialize on
3533 * owner->perf_event_mutex.
3535 smp_read_barrier_depends();
3538 * Since delayed_put_task_struct() also drops the last
3539 * task reference we can safely take a new reference
3540 * while holding the rcu_read_lock().
3542 get_task_struct(owner);
3548 * If we're here through perf_event_exit_task() we're already
3549 * holding ctx->mutex which would be an inversion wrt. the
3550 * normal lock order.
3552 * However we can safely take this lock because its the child
3555 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
3558 * We have to re-check the event->owner field, if it is cleared
3559 * we raced with perf_event_exit_task(), acquiring the mutex
3560 * ensured they're done, and we can proceed with freeing the
3564 list_del_init(&event->owner_entry);
3565 mutex_unlock(&owner->perf_event_mutex);
3566 put_task_struct(owner);
3571 * Called when the last reference to the file is gone.
3573 static void put_event(struct perf_event *event)
3575 struct perf_event_context *ctx = event->ctx;
3577 if (!atomic_long_dec_and_test(&event->refcount))
3580 if (!is_kernel_event(event))
3581 perf_remove_from_owner(event);
3583 WARN_ON_ONCE(ctx->parent_ctx);
3585 * There are two ways this annotation is useful:
3587 * 1) there is a lock recursion from perf_event_exit_task
3588 * see the comment there.
3590 * 2) there is a lock-inversion with mmap_sem through
3591 * perf_event_read_group(), which takes faults while
3592 * holding ctx->mutex, however this is called after
3593 * the last filedesc died, so there is no possibility
3594 * to trigger the AB-BA case.
3596 mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
3597 perf_remove_from_context(event, true);
3598 mutex_unlock(&ctx->mutex);
3603 int perf_event_release_kernel(struct perf_event *event)
3608 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
3610 static int perf_release(struct inode *inode, struct file *file)
3612 put_event(file->private_data);
3617 * Remove all orphanes events from the context.
3619 static void orphans_remove_work(struct work_struct *work)
3621 struct perf_event_context *ctx;
3622 struct perf_event *event, *tmp;
3624 ctx = container_of(work, struct perf_event_context,
3625 orphans_remove.work);
3627 mutex_lock(&ctx->mutex);
3628 list_for_each_entry_safe(event, tmp, &ctx->event_list, event_entry) {
3629 struct perf_event *parent_event = event->parent;
3631 if (!is_orphaned_child(event))
3634 perf_remove_from_context(event, true);
3636 mutex_lock(&parent_event->child_mutex);
3637 list_del_init(&event->child_list);
3638 mutex_unlock(&parent_event->child_mutex);
3641 put_event(parent_event);
3644 raw_spin_lock_irq(&ctx->lock);
3645 ctx->orphans_remove_sched = false;
3646 raw_spin_unlock_irq(&ctx->lock);
3647 mutex_unlock(&ctx->mutex);
3652 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
3654 struct perf_event *child;
3660 mutex_lock(&event->child_mutex);
3661 total += perf_event_read(event);
3662 *enabled += event->total_time_enabled +
3663 atomic64_read(&event->child_total_time_enabled);
3664 *running += event->total_time_running +
3665 atomic64_read(&event->child_total_time_running);
3667 list_for_each_entry(child, &event->child_list, child_list) {
3668 total += perf_event_read(child);
3669 *enabled += child->total_time_enabled;
3670 *running += child->total_time_running;
3672 mutex_unlock(&event->child_mutex);
3676 EXPORT_SYMBOL_GPL(perf_event_read_value);
3678 static int perf_event_read_group(struct perf_event *event,
3679 u64 read_format, char __user *buf)
3681 struct perf_event *leader = event->group_leader, *sub;
3682 struct perf_event_context *ctx = leader->ctx;
3683 int n = 0, size = 0, ret;
3684 u64 count, enabled, running;
3687 lockdep_assert_held(&ctx->mutex);
3689 count = perf_event_read_value(leader, &enabled, &running);
3691 values[n++] = 1 + leader->nr_siblings;
3692 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3693 values[n++] = enabled;
3694 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3695 values[n++] = running;
3696 values[n++] = count;
3697 if (read_format & PERF_FORMAT_ID)
3698 values[n++] = primary_event_id(leader);
3700 size = n * sizeof(u64);
3702 if (copy_to_user(buf, values, size))
3707 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
3710 values[n++] = perf_event_read_value(sub, &enabled, &running);
3711 if (read_format & PERF_FORMAT_ID)
3712 values[n++] = primary_event_id(sub);
3714 size = n * sizeof(u64);
3716 if (copy_to_user(buf + ret, values, size)) {
3726 static int perf_event_read_one(struct perf_event *event,
3727 u64 read_format, char __user *buf)
3729 u64 enabled, running;
3733 values[n++] = perf_event_read_value(event, &enabled, &running);
3734 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
3735 values[n++] = enabled;
3736 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
3737 values[n++] = running;
3738 if (read_format & PERF_FORMAT_ID)
3739 values[n++] = primary_event_id(event);
3741 if (copy_to_user(buf, values, n * sizeof(u64)))
3744 return n * sizeof(u64);
3747 static bool is_event_hup(struct perf_event *event)
3751 if (event->state != PERF_EVENT_STATE_EXIT)
3754 mutex_lock(&event->child_mutex);
3755 no_children = list_empty(&event->child_list);
3756 mutex_unlock(&event->child_mutex);
3761 * Read the performance event - simple non blocking version for now
3764 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
3766 u64 read_format = event->attr.read_format;
3770 * Return end-of-file for a read on a event that is in
3771 * error state (i.e. because it was pinned but it couldn't be
3772 * scheduled on to the CPU at some point).
3774 if (event->state == PERF_EVENT_STATE_ERROR)
3777 if (count < event->read_size)
3780 WARN_ON_ONCE(event->ctx->parent_ctx);
3781 if (read_format & PERF_FORMAT_GROUP)
3782 ret = perf_event_read_group(event, read_format, buf);
3784 ret = perf_event_read_one(event, read_format, buf);
3790 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
3792 struct perf_event *event = file->private_data;
3793 struct perf_event_context *ctx;
3796 ctx = perf_event_ctx_lock(event);
3797 ret = perf_read_hw(event, buf, count);
3798 perf_event_ctx_unlock(event, ctx);
3803 static unsigned int perf_poll(struct file *file, poll_table *wait)
3805 struct perf_event *event = file->private_data;
3806 struct ring_buffer *rb;
3807 unsigned int events = POLLHUP;
3809 poll_wait(file, &event->waitq, wait);
3811 if (is_event_hup(event))
3815 * Pin the event->rb by taking event->mmap_mutex; otherwise
3816 * perf_event_set_output() can swizzle our rb and make us miss wakeups.
3818 mutex_lock(&event->mmap_mutex);
3821 events = atomic_xchg(&rb->poll, 0);
3822 mutex_unlock(&event->mmap_mutex);
3826 static void _perf_event_reset(struct perf_event *event)
3828 (void)perf_event_read(event);
3829 local64_set(&event->count, 0);
3830 perf_event_update_userpage(event);
3834 * Holding the top-level event's child_mutex means that any
3835 * descendant process that has inherited this event will block
3836 * in sync_child_event if it goes to exit, thus satisfying the
3837 * task existence requirements of perf_event_enable/disable.
3839 static void perf_event_for_each_child(struct perf_event *event,
3840 void (*func)(struct perf_event *))
3842 struct perf_event *child;
3844 WARN_ON_ONCE(event->ctx->parent_ctx);
3846 mutex_lock(&event->child_mutex);
3848 list_for_each_entry(child, &event->child_list, child_list)
3850 mutex_unlock(&event->child_mutex);
3853 static void perf_event_for_each(struct perf_event *event,
3854 void (*func)(struct perf_event *))
3856 struct perf_event_context *ctx = event->ctx;
3857 struct perf_event *sibling;
3859 lockdep_assert_held(&ctx->mutex);
3861 event = event->group_leader;
3863 perf_event_for_each_child(event, func);
3864 list_for_each_entry(sibling, &event->sibling_list, group_entry)
3865 perf_event_for_each_child(sibling, func);
3868 static int perf_event_period(struct perf_event *event, u64 __user *arg)
3870 struct perf_event_context *ctx = event->ctx;
3871 int ret = 0, active;
3874 if (!is_sampling_event(event))
3877 if (copy_from_user(&value, arg, sizeof(value)))
3883 raw_spin_lock_irq(&ctx->lock);
3884 if (event->attr.freq) {
3885 if (value > sysctl_perf_event_sample_rate) {
3890 event->attr.sample_freq = value;
3892 event->attr.sample_period = value;
3893 event->hw.sample_period = value;
3896 active = (event->state == PERF_EVENT_STATE_ACTIVE);
3898 perf_pmu_disable(ctx->pmu);
3899 event->pmu->stop(event, PERF_EF_UPDATE);
3902 local64_set(&event->hw.period_left, 0);
3905 event->pmu->start(event, PERF_EF_RELOAD);
3906 perf_pmu_enable(ctx->pmu);
3910 raw_spin_unlock_irq(&ctx->lock);
3915 static const struct file_operations perf_fops;
3917 static inline int perf_fget_light(int fd, struct fd *p)
3919 struct fd f = fdget(fd);
3923 if (f.file->f_op != &perf_fops) {
3931 static int perf_event_set_output(struct perf_event *event,
3932 struct perf_event *output_event);
3933 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
3935 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
3937 void (*func)(struct perf_event *);
3941 case PERF_EVENT_IOC_ENABLE:
3942 func = _perf_event_enable;
3944 case PERF_EVENT_IOC_DISABLE:
3945 func = _perf_event_disable;
3947 case PERF_EVENT_IOC_RESET:
3948 func = _perf_event_reset;
3951 case PERF_EVENT_IOC_REFRESH:
3952 return _perf_event_refresh(event, arg);
3954 case PERF_EVENT_IOC_PERIOD:
3955 return perf_event_period(event, (u64 __user *)arg);
3957 case PERF_EVENT_IOC_ID:
3959 u64 id = primary_event_id(event);
3961 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
3966 case PERF_EVENT_IOC_SET_OUTPUT:
3970 struct perf_event *output_event;
3972 ret = perf_fget_light(arg, &output);
3975 output_event = output.file->private_data;
3976 ret = perf_event_set_output(event, output_event);
3979 ret = perf_event_set_output(event, NULL);
3984 case PERF_EVENT_IOC_SET_FILTER:
3985 return perf_event_set_filter(event, (void __user *)arg);
3991 if (flags & PERF_IOC_FLAG_GROUP)
3992 perf_event_for_each(event, func);
3994 perf_event_for_each_child(event, func);
3999 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
4001 struct perf_event *event = file->private_data;
4002 struct perf_event_context *ctx;
4005 ctx = perf_event_ctx_lock(event);
4006 ret = _perf_ioctl(event, cmd, arg);
4007 perf_event_ctx_unlock(event, ctx);
4012 #ifdef CONFIG_COMPAT
4013 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4016 switch (_IOC_NR(cmd)) {
4017 case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4018 case _IOC_NR(PERF_EVENT_IOC_ID):
4019 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4020 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4021 cmd &= ~IOCSIZE_MASK;
4022 cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4026 return perf_ioctl(file, cmd, arg);
4029 # define perf_compat_ioctl NULL
4032 int perf_event_task_enable(void)
4034 struct perf_event_context *ctx;
4035 struct perf_event *event;
4037 mutex_lock(¤t->perf_event_mutex);
4038 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4039 ctx = perf_event_ctx_lock(event);
4040 perf_event_for_each_child(event, _perf_event_enable);
4041 perf_event_ctx_unlock(event, ctx);
4043 mutex_unlock(¤t->perf_event_mutex);
4048 int perf_event_task_disable(void)
4050 struct perf_event_context *ctx;
4051 struct perf_event *event;
4053 mutex_lock(¤t->perf_event_mutex);
4054 list_for_each_entry(event, ¤t->perf_event_list, owner_entry) {
4055 ctx = perf_event_ctx_lock(event);
4056 perf_event_for_each_child(event, _perf_event_disable);
4057 perf_event_ctx_unlock(event, ctx);
4059 mutex_unlock(¤t->perf_event_mutex);
4064 static int perf_event_index(struct perf_event *event)
4066 if (event->hw.state & PERF_HES_STOPPED)
4069 if (event->state != PERF_EVENT_STATE_ACTIVE)
4072 return event->pmu->event_idx(event);
4075 static void calc_timer_values(struct perf_event *event,
4082 *now = perf_clock();
4083 ctx_time = event->shadow_ctx_time + *now;
4084 *enabled = ctx_time - event->tstamp_enabled;
4085 *running = ctx_time - event->tstamp_running;
4088 static void perf_event_init_userpage(struct perf_event *event)
4090 struct perf_event_mmap_page *userpg;
4091 struct ring_buffer *rb;
4094 rb = rcu_dereference(event->rb);
4098 userpg = rb->user_page;
4100 /* Allow new userspace to detect that bit 0 is deprecated */
4101 userpg->cap_bit0_is_deprecated = 1;
4102 userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4108 void __weak arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
4113 * Callers need to ensure there can be no nesting of this function, otherwise
4114 * the seqlock logic goes bad. We can not serialize this because the arch
4115 * code calls this from NMI context.
4117 void perf_event_update_userpage(struct perf_event *event)
4119 struct perf_event_mmap_page *userpg;
4120 struct ring_buffer *rb;
4121 u64 enabled, running, now;
4124 rb = rcu_dereference(event->rb);
4129 * compute total_time_enabled, total_time_running
4130 * based on snapshot values taken when the event
4131 * was last scheduled in.
4133 * we cannot simply called update_context_time()
4134 * because of locking issue as we can be called in
4137 calc_timer_values(event, &now, &enabled, &running);
4139 userpg = rb->user_page;
4141 * Disable preemption so as to not let the corresponding user-space
4142 * spin too long if we get preempted.
4147 userpg->index = perf_event_index(event);
4148 userpg->offset = perf_event_count(event);
4150 userpg->offset -= local64_read(&event->hw.prev_count);
4152 userpg->time_enabled = enabled +
4153 atomic64_read(&event->child_total_time_enabled);
4155 userpg->time_running = running +
4156 atomic64_read(&event->child_total_time_running);
4158 arch_perf_update_userpage(userpg, now);
4167 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
4169 struct perf_event *event = vma->vm_file->private_data;
4170 struct ring_buffer *rb;
4171 int ret = VM_FAULT_SIGBUS;
4173 if (vmf->flags & FAULT_FLAG_MKWRITE) {
4174 if (vmf->pgoff == 0)
4180 rb = rcu_dereference(event->rb);
4184 if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4187 vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4191 get_page(vmf->page);
4192 vmf->page->mapping = vma->vm_file->f_mapping;
4193 vmf->page->index = vmf->pgoff;
4202 static void ring_buffer_attach(struct perf_event *event,
4203 struct ring_buffer *rb)
4205 struct ring_buffer *old_rb = NULL;
4206 unsigned long flags;
4210 * Should be impossible, we set this when removing
4211 * event->rb_entry and wait/clear when adding event->rb_entry.
4213 WARN_ON_ONCE(event->rcu_pending);
4216 event->rcu_batches = get_state_synchronize_rcu();
4217 event->rcu_pending = 1;
4219 spin_lock_irqsave(&old_rb->event_lock, flags);
4220 list_del_rcu(&event->rb_entry);
4221 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4224 if (event->rcu_pending && rb) {
4225 cond_synchronize_rcu(event->rcu_batches);
4226 event->rcu_pending = 0;
4230 spin_lock_irqsave(&rb->event_lock, flags);
4231 list_add_rcu(&event->rb_entry, &rb->event_list);
4232 spin_unlock_irqrestore(&rb->event_lock, flags);
4235 rcu_assign_pointer(event->rb, rb);
4238 ring_buffer_put(old_rb);
4240 * Since we detached before setting the new rb, so that we
4241 * could attach the new rb, we could have missed a wakeup.
4244 wake_up_all(&event->waitq);
4248 static void ring_buffer_wakeup(struct perf_event *event)
4250 struct ring_buffer *rb;
4253 rb = rcu_dereference(event->rb);
4255 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
4256 wake_up_all(&event->waitq);
4261 static void rb_free_rcu(struct rcu_head *rcu_head)
4263 struct ring_buffer *rb;
4265 rb = container_of(rcu_head, struct ring_buffer, rcu_head);
4269 static struct ring_buffer *ring_buffer_get(struct perf_event *event)
4271 struct ring_buffer *rb;
4274 rb = rcu_dereference(event->rb);
4276 if (!atomic_inc_not_zero(&rb->refcount))
4284 static void ring_buffer_put(struct ring_buffer *rb)
4286 if (!atomic_dec_and_test(&rb->refcount))
4289 WARN_ON_ONCE(!list_empty(&rb->event_list));
4291 call_rcu(&rb->rcu_head, rb_free_rcu);
4294 static void perf_mmap_open(struct vm_area_struct *vma)
4296 struct perf_event *event = vma->vm_file->private_data;
4298 atomic_inc(&event->mmap_count);
4299 atomic_inc(&event->rb->mmap_count);
4303 * A buffer can be mmap()ed multiple times; either directly through the same
4304 * event, or through other events by use of perf_event_set_output().
4306 * In order to undo the VM accounting done by perf_mmap() we need to destroy
4307 * the buffer here, where we still have a VM context. This means we need
4308 * to detach all events redirecting to us.
4310 static void perf_mmap_close(struct vm_area_struct *vma)
4312 struct perf_event *event = vma->vm_file->private_data;
4314 struct ring_buffer *rb = ring_buffer_get(event);
4315 struct user_struct *mmap_user = rb->mmap_user;
4316 int mmap_locked = rb->mmap_locked;
4317 unsigned long size = perf_data_size(rb);
4319 atomic_dec(&rb->mmap_count);
4321 if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
4324 ring_buffer_attach(event, NULL);
4325 mutex_unlock(&event->mmap_mutex);
4327 /* If there's still other mmap()s of this buffer, we're done. */
4328 if (atomic_read(&rb->mmap_count))
4332 * No other mmap()s, detach from all other events that might redirect
4333 * into the now unreachable buffer. Somewhat complicated by the
4334 * fact that rb::event_lock otherwise nests inside mmap_mutex.
4338 list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
4339 if (!atomic_long_inc_not_zero(&event->refcount)) {
4341 * This event is en-route to free_event() which will
4342 * detach it and remove it from the list.
4348 mutex_lock(&event->mmap_mutex);
4350 * Check we didn't race with perf_event_set_output() which can
4351 * swizzle the rb from under us while we were waiting to
4352 * acquire mmap_mutex.
4354 * If we find a different rb; ignore this event, a next
4355 * iteration will no longer find it on the list. We have to
4356 * still restart the iteration to make sure we're not now
4357 * iterating the wrong list.
4359 if (event->rb == rb)
4360 ring_buffer_attach(event, NULL);
4362 mutex_unlock(&event->mmap_mutex);
4366 * Restart the iteration; either we're on the wrong list or
4367 * destroyed its integrity by doing a deletion.
4374 * It could be there's still a few 0-ref events on the list; they'll
4375 * get cleaned up by free_event() -- they'll also still have their
4376 * ref on the rb and will free it whenever they are done with it.
4378 * Aside from that, this buffer is 'fully' detached and unmapped,
4379 * undo the VM accounting.
4382 atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
4383 vma->vm_mm->pinned_vm -= mmap_locked;
4384 free_uid(mmap_user);
4387 ring_buffer_put(rb); /* could be last */
4390 static const struct vm_operations_struct perf_mmap_vmops = {
4391 .open = perf_mmap_open,
4392 .close = perf_mmap_close,
4393 .fault = perf_mmap_fault,
4394 .page_mkwrite = perf_mmap_fault,
4397 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
4399 struct perf_event *event = file->private_data;
4400 unsigned long user_locked, user_lock_limit;
4401 struct user_struct *user = current_user();
4402 unsigned long locked, lock_limit;
4403 struct ring_buffer *rb;
4404 unsigned long vma_size;
4405 unsigned long nr_pages;
4406 long user_extra, extra;
4407 int ret = 0, flags = 0;
4410 * Don't allow mmap() of inherited per-task counters. This would
4411 * create a performance issue due to all children writing to the
4414 if (event->cpu == -1 && event->attr.inherit)
4417 if (!(vma->vm_flags & VM_SHARED))
4420 vma_size = vma->vm_end - vma->vm_start;
4421 nr_pages = (vma_size / PAGE_SIZE) - 1;
4424 * If we have rb pages ensure they're a power-of-two number, so we
4425 * can do bitmasks instead of modulo.
4427 if (nr_pages != 0 && !is_power_of_2(nr_pages))
4430 if (vma_size != PAGE_SIZE * (1 + nr_pages))
4433 if (vma->vm_pgoff != 0)
4436 WARN_ON_ONCE(event->ctx->parent_ctx);
4438 mutex_lock(&event->mmap_mutex);
4440 if (event->rb->nr_pages != nr_pages) {
4445 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
4447 * Raced against perf_mmap_close() through
4448 * perf_event_set_output(). Try again, hope for better
4451 mutex_unlock(&event->mmap_mutex);
4458 user_extra = nr_pages + 1;
4459 user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
4462 * Increase the limit linearly with more CPUs:
4464 user_lock_limit *= num_online_cpus();
4466 user_locked = atomic_long_read(&user->locked_vm) + user_extra;
4469 if (user_locked > user_lock_limit)
4470 extra = user_locked - user_lock_limit;
4472 lock_limit = rlimit(RLIMIT_MEMLOCK);
4473 lock_limit >>= PAGE_SHIFT;
4474 locked = vma->vm_mm->pinned_vm + extra;
4476 if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
4477 !capable(CAP_IPC_LOCK)) {
4484 if (vma->vm_flags & VM_WRITE)
4485 flags |= RING_BUFFER_WRITABLE;
4487 rb = rb_alloc(nr_pages,
4488 event->attr.watermark ? event->attr.wakeup_watermark : 0,
4496 atomic_set(&rb->mmap_count, 1);
4497 rb->mmap_locked = extra;
4498 rb->mmap_user = get_current_user();
4500 atomic_long_add(user_extra, &user->locked_vm);
4501 vma->vm_mm->pinned_vm += extra;
4503 ring_buffer_attach(event, rb);
4505 perf_event_init_userpage(event);
4506 perf_event_update_userpage(event);
4510 atomic_inc(&event->mmap_count);
4511 mutex_unlock(&event->mmap_mutex);
4514 * Since pinned accounting is per vm we cannot allow fork() to copy our
4517 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
4518 vma->vm_ops = &perf_mmap_vmops;
4523 static int perf_fasync(int fd, struct file *filp, int on)
4525 struct inode *inode = file_inode(filp);
4526 struct perf_event *event = filp->private_data;
4529 mutex_lock(&inode->i_mutex);
4530 retval = fasync_helper(fd, filp, on, &event->fasync);
4531 mutex_unlock(&inode->i_mutex);
4539 static const struct file_operations perf_fops = {
4540 .llseek = no_llseek,
4541 .release = perf_release,
4544 .unlocked_ioctl = perf_ioctl,
4545 .compat_ioctl = perf_compat_ioctl,
4547 .fasync = perf_fasync,
4553 * If there's data, ensure we set the poll() state and publish everything
4554 * to user-space before waking everybody up.
4557 void perf_event_wakeup(struct perf_event *event)
4559 ring_buffer_wakeup(event);
4561 if (event->pending_kill) {
4562 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
4563 event->pending_kill = 0;
4567 static void perf_pending_event(struct irq_work *entry)
4569 struct perf_event *event = container_of(entry,
4570 struct perf_event, pending);
4572 if (event->pending_disable) {
4573 event->pending_disable = 0;
4574 __perf_event_disable(event);
4577 if (event->pending_wakeup) {
4578 event->pending_wakeup = 0;
4579 perf_event_wakeup(event);
4584 * We assume there is only KVM supporting the callbacks.
4585 * Later on, we might change it to a list if there is
4586 * another virtualization implementation supporting the callbacks.
4588 struct perf_guest_info_callbacks *perf_guest_cbs;
4590 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4592 perf_guest_cbs = cbs;
4595 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
4597 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
4599 perf_guest_cbs = NULL;
4602 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
4605 perf_output_sample_regs(struct perf_output_handle *handle,
4606 struct pt_regs *regs, u64 mask)
4610 for_each_set_bit(bit, (const unsigned long *) &mask,
4611 sizeof(mask) * BITS_PER_BYTE) {
4614 val = perf_reg_value(regs, bit);
4615 perf_output_put(handle, val);
4619 static void perf_sample_regs_user(struct perf_regs *regs_user,
4620 struct pt_regs *regs,
4621 struct pt_regs *regs_user_copy)
4623 if (user_mode(regs)) {
4624 regs_user->abi = perf_reg_abi(current);
4625 regs_user->regs = regs;
4626 } else if (current->mm) {
4627 perf_get_regs_user(regs_user, regs, regs_user_copy);
4629 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
4630 regs_user->regs = NULL;
4634 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
4635 struct pt_regs *regs)
4637 regs_intr->regs = regs;
4638 regs_intr->abi = perf_reg_abi(current);
4643 * Get remaining task size from user stack pointer.
4645 * It'd be better to take stack vma map and limit this more
4646 * precisly, but there's no way to get it safely under interrupt,
4647 * so using TASK_SIZE as limit.
4649 static u64 perf_ustack_task_size(struct pt_regs *regs)
4651 unsigned long addr = perf_user_stack_pointer(regs);
4653 if (!addr || addr >= TASK_SIZE)
4656 return TASK_SIZE - addr;
4660 perf_sample_ustack_size(u16 stack_size, u16 header_size,
4661 struct pt_regs *regs)
4665 /* No regs, no stack pointer, no dump. */
4670 * Check if we fit in with the requested stack size into the:
4672 * If we don't, we limit the size to the TASK_SIZE.
4674 * - remaining sample size
4675 * If we don't, we customize the stack size to
4676 * fit in to the remaining sample size.
4679 task_size = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
4680 stack_size = min(stack_size, (u16) task_size);
4682 /* Current header size plus static size and dynamic size. */
4683 header_size += 2 * sizeof(u64);
4685 /* Do we fit in with the current stack dump size? */
4686 if ((u16) (header_size + stack_size) < header_size) {
4688 * If we overflow the maximum size for the sample,
4689 * we customize the stack dump size to fit in.
4691 stack_size = USHRT_MAX - header_size - sizeof(u64);
4692 stack_size = round_up(stack_size, sizeof(u64));
4699 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
4700 struct pt_regs *regs)
4702 /* Case of a kernel thread, nothing to dump */
4705 perf_output_put(handle, size);
4714 * - the size requested by user or the best one we can fit
4715 * in to the sample max size
4717 * - user stack dump data
4719 * - the actual dumped size
4723 perf_output_put(handle, dump_size);
4726 sp = perf_user_stack_pointer(regs);
4727 rem = __output_copy_user(handle, (void *) sp, dump_size);
4728 dyn_size = dump_size - rem;
4730 perf_output_skip(handle, rem);
4733 perf_output_put(handle, dyn_size);
4737 static void __perf_event_header__init_id(struct perf_event_header *header,
4738 struct perf_sample_data *data,
4739 struct perf_event *event)
4741 u64 sample_type = event->attr.sample_type;
4743 data->type = sample_type;
4744 header->size += event->id_header_size;
4746 if (sample_type & PERF_SAMPLE_TID) {
4747 /* namespace issues */
4748 data->tid_entry.pid = perf_event_pid(event, current);
4749 data->tid_entry.tid = perf_event_tid(event, current);
4752 if (sample_type & PERF_SAMPLE_TIME)
4753 data->time = perf_clock();
4755 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
4756 data->id = primary_event_id(event);
4758 if (sample_type & PERF_SAMPLE_STREAM_ID)
4759 data->stream_id = event->id;
4761 if (sample_type & PERF_SAMPLE_CPU) {
4762 data->cpu_entry.cpu = raw_smp_processor_id();
4763 data->cpu_entry.reserved = 0;
4767 void perf_event_header__init_id(struct perf_event_header *header,
4768 struct perf_sample_data *data,
4769 struct perf_event *event)
4771 if (event->attr.sample_id_all)
4772 __perf_event_header__init_id(header, data, event);
4775 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
4776 struct perf_sample_data *data)
4778 u64 sample_type = data->type;
4780 if (sample_type & PERF_SAMPLE_TID)
4781 perf_output_put(handle, data->tid_entry);
4783 if (sample_type & PERF_SAMPLE_TIME)
4784 perf_output_put(handle, data->time);
4786 if (sample_type & PERF_SAMPLE_ID)
4787 perf_output_put(handle, data->id);
4789 if (sample_type & PERF_SAMPLE_STREAM_ID)
4790 perf_output_put(handle, data->stream_id);
4792 if (sample_type & PERF_SAMPLE_CPU)
4793 perf_output_put(handle, data->cpu_entry);
4795 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4796 perf_output_put(handle, data->id);
4799 void perf_event__output_id_sample(struct perf_event *event,
4800 struct perf_output_handle *handle,
4801 struct perf_sample_data *sample)
4803 if (event->attr.sample_id_all)
4804 __perf_event__output_id_sample(handle, sample);
4807 static void perf_output_read_one(struct perf_output_handle *handle,
4808 struct perf_event *event,
4809 u64 enabled, u64 running)
4811 u64 read_format = event->attr.read_format;
4815 values[n++] = perf_event_count(event);
4816 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4817 values[n++] = enabled +
4818 atomic64_read(&event->child_total_time_enabled);
4820 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4821 values[n++] = running +
4822 atomic64_read(&event->child_total_time_running);
4824 if (read_format & PERF_FORMAT_ID)
4825 values[n++] = primary_event_id(event);
4827 __output_copy(handle, values, n * sizeof(u64));
4831 * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
4833 static void perf_output_read_group(struct perf_output_handle *handle,
4834 struct perf_event *event,
4835 u64 enabled, u64 running)
4837 struct perf_event *leader = event->group_leader, *sub;
4838 u64 read_format = event->attr.read_format;
4842 values[n++] = 1 + leader->nr_siblings;
4844 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4845 values[n++] = enabled;
4847 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4848 values[n++] = running;
4850 if (leader != event)
4851 leader->pmu->read(leader);
4853 values[n++] = perf_event_count(leader);
4854 if (read_format & PERF_FORMAT_ID)
4855 values[n++] = primary_event_id(leader);
4857 __output_copy(handle, values, n * sizeof(u64));
4859 list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4862 if ((sub != event) &&
4863 (sub->state == PERF_EVENT_STATE_ACTIVE))
4864 sub->pmu->read(sub);
4866 values[n++] = perf_event_count(sub);
4867 if (read_format & PERF_FORMAT_ID)
4868 values[n++] = primary_event_id(sub);
4870 __output_copy(handle, values, n * sizeof(u64));
4874 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
4875 PERF_FORMAT_TOTAL_TIME_RUNNING)
4877 static void perf_output_read(struct perf_output_handle *handle,
4878 struct perf_event *event)
4880 u64 enabled = 0, running = 0, now;
4881 u64 read_format = event->attr.read_format;
4884 * compute total_time_enabled, total_time_running
4885 * based on snapshot values taken when the event
4886 * was last scheduled in.
4888 * we cannot simply called update_context_time()
4889 * because of locking issue as we are called in
4892 if (read_format & PERF_FORMAT_TOTAL_TIMES)
4893 calc_timer_values(event, &now, &enabled, &running);
4895 if (event->attr.read_format & PERF_FORMAT_GROUP)
4896 perf_output_read_group(handle, event, enabled, running);
4898 perf_output_read_one(handle, event, enabled, running);
4901 void perf_output_sample(struct perf_output_handle *handle,
4902 struct perf_event_header *header,
4903 struct perf_sample_data *data,
4904 struct perf_event *event)
4906 u64 sample_type = data->type;
4908 perf_output_put(handle, *header);
4910 if (sample_type & PERF_SAMPLE_IDENTIFIER)
4911 perf_output_put(handle, data->id);
4913 if (sample_type & PERF_SAMPLE_IP)
4914 perf_output_put(handle, data->ip);
4916 if (sample_type & PERF_SAMPLE_TID)
4917 perf_output_put(handle, data->tid_entry);
4919 if (sample_type & PERF_SAMPLE_TIME)
4920 perf_output_put(handle, data->time);
4922 if (sample_type & PERF_SAMPLE_ADDR)
4923 perf_output_put(handle, data->addr);
4925 if (sample_type & PERF_SAMPLE_ID)
4926 perf_output_put(handle, data->id);
4928 if (sample_type & PERF_SAMPLE_STREAM_ID)
4929 perf_output_put(handle, data->stream_id);
4931 if (sample_type & PERF_SAMPLE_CPU)
4932 perf_output_put(handle, data->cpu_entry);
4934 if (sample_type & PERF_SAMPLE_PERIOD)
4935 perf_output_put(handle, data->period);
4937 if (sample_type & PERF_SAMPLE_READ)
4938 perf_output_read(handle, event);
4940 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
4941 if (data->callchain) {
4944 if (data->callchain)
4945 size += data->callchain->nr;
4947 size *= sizeof(u64);
4949 __output_copy(handle, data->callchain, size);
4952 perf_output_put(handle, nr);
4956 if (sample_type & PERF_SAMPLE_RAW) {
4958 perf_output_put(handle, data->raw->size);
4959 __output_copy(handle, data->raw->data,
4966 .size = sizeof(u32),
4969 perf_output_put(handle, raw);
4973 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
4974 if (data->br_stack) {
4977 size = data->br_stack->nr
4978 * sizeof(struct perf_branch_entry);
4980 perf_output_put(handle, data->br_stack->nr);
4981 perf_output_copy(handle, data->br_stack->entries, size);
4984 * we always store at least the value of nr
4987 perf_output_put(handle, nr);
4991 if (sample_type & PERF_SAMPLE_REGS_USER) {
4992 u64 abi = data->regs_user.abi;
4995 * If there are no regs to dump, notice it through
4996 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
4998 perf_output_put(handle, abi);
5001 u64 mask = event->attr.sample_regs_user;
5002 perf_output_sample_regs(handle,
5003 data->regs_user.regs,
5008 if (sample_type & PERF_SAMPLE_STACK_USER) {
5009 perf_output_sample_ustack(handle,
5010 data->stack_user_size,
5011 data->regs_user.regs);
5014 if (sample_type & PERF_SAMPLE_WEIGHT)
5015 perf_output_put(handle, data->weight);
5017 if (sample_type & PERF_SAMPLE_DATA_SRC)
5018 perf_output_put(handle, data->data_src.val);
5020 if (sample_type & PERF_SAMPLE_TRANSACTION)
5021 perf_output_put(handle, data->txn);
5023 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5024 u64 abi = data->regs_intr.abi;
5026 * If there are no regs to dump, notice it through
5027 * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
5029 perf_output_put(handle, abi);
5032 u64 mask = event->attr.sample_regs_intr;
5034 perf_output_sample_regs(handle,
5035 data->regs_intr.regs,
5040 if (!event->attr.watermark) {
5041 int wakeup_events = event->attr.wakeup_events;
5043 if (wakeup_events) {
5044 struct ring_buffer *rb = handle->rb;
5045 int events = local_inc_return(&rb->events);
5047 if (events >= wakeup_events) {
5048 local_sub(wakeup_events, &rb->events);
5049 local_inc(&rb->wakeup);
5055 void perf_prepare_sample(struct perf_event_header *header,
5056 struct perf_sample_data *data,
5057 struct perf_event *event,
5058 struct pt_regs *regs)
5060 u64 sample_type = event->attr.sample_type;
5062 header->type = PERF_RECORD_SAMPLE;
5063 header->size = sizeof(*header) + event->header_size;
5066 header->misc |= perf_misc_flags(regs);
5068 __perf_event_header__init_id(header, data, event);
5070 if (sample_type & PERF_SAMPLE_IP)
5071 data->ip = perf_instruction_pointer(regs);
5073 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
5076 data->callchain = perf_callchain(event, regs);
5078 if (data->callchain)
5079 size += data->callchain->nr;
5081 header->size += size * sizeof(u64);
5084 if (sample_type & PERF_SAMPLE_RAW) {
5085 int size = sizeof(u32);
5088 size += data->raw->size;
5090 size += sizeof(u32);
5092 WARN_ON_ONCE(size & (sizeof(u64)-1));
5093 header->size += size;
5096 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
5097 int size = sizeof(u64); /* nr */
5098 if (data->br_stack) {
5099 size += data->br_stack->nr
5100 * sizeof(struct perf_branch_entry);
5102 header->size += size;
5105 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
5106 perf_sample_regs_user(&data->regs_user, regs,
5107 &data->regs_user_copy);
5109 if (sample_type & PERF_SAMPLE_REGS_USER) {
5110 /* regs dump ABI info */
5111 int size = sizeof(u64);
5113 if (data->regs_user.regs) {
5114 u64 mask = event->attr.sample_regs_user;
5115 size += hweight64(mask) * sizeof(u64);
5118 header->size += size;
5121 if (sample_type & PERF_SAMPLE_STACK_USER) {
5123 * Either we need PERF_SAMPLE_STACK_USER bit to be allways
5124 * processed as the last one or have additional check added
5125 * in case new sample type is added, because we could eat
5126 * up the rest of the sample size.
5128 u16 stack_size = event->attr.sample_stack_user;
5129 u16 size = sizeof(u64);
5131 stack_size = perf_sample_ustack_size(stack_size, header->size,
5132 data->regs_user.regs);
5135 * If there is something to dump, add space for the dump
5136 * itself and for the field that tells the dynamic size,
5137 * which is how many have been actually dumped.
5140 size += sizeof(u64) + stack_size;
5142 data->stack_user_size = stack_size;
5143 header->size += size;
5146 if (sample_type & PERF_SAMPLE_REGS_INTR) {
5147 /* regs dump ABI info */
5148 int size = sizeof(u64);
5150 perf_sample_regs_intr(&data->regs_intr, regs);
5152 if (data->regs_intr.regs) {
5153 u64 mask = event->attr.sample_regs_intr;
5155 size += hweight64(mask) * sizeof(u64);
5158 header->size += size;
5162 static void perf_event_output(struct perf_event *event,
5163 struct perf_sample_data *data,
5164 struct pt_regs *regs)
5166 struct perf_output_handle handle;
5167 struct perf_event_header header;
5169 /* protect the callchain buffers */
5172 perf_prepare_sample(&header, data, event, regs);
5174 if (perf_output_begin(&handle, event, header.size))
5177 perf_output_sample(&handle, &header, data, event);
5179 perf_output_end(&handle);
5189 struct perf_read_event {
5190 struct perf_event_header header;
5197 perf_event_read_event(struct perf_event *event,
5198 struct task_struct *task)
5200 struct perf_output_handle handle;
5201 struct perf_sample_data sample;
5202 struct perf_read_event read_event = {
5204 .type = PERF_RECORD_READ,
5206 .size = sizeof(read_event) + event->read_size,
5208 .pid = perf_event_pid(event, task),
5209 .tid = perf_event_tid(event, task),
5213 perf_event_header__init_id(&read_event.header, &sample, event);
5214 ret = perf_output_begin(&handle, event, read_event.header.size);
5218 perf_output_put(&handle, read_event);
5219 perf_output_read(&handle, event);
5220 perf_event__output_id_sample(event, &handle, &sample);
5222 perf_output_end(&handle);
5225 typedef void (perf_event_aux_output_cb)(struct perf_event *event, void *data);
5228 perf_event_aux_ctx(struct perf_event_context *ctx,
5229 perf_event_aux_output_cb output,
5232 struct perf_event *event;
5234 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
5235 if (event->state < PERF_EVENT_STATE_INACTIVE)
5237 if (!event_filter_match(event))
5239 output(event, data);
5244 perf_event_aux(perf_event_aux_output_cb output, void *data,
5245 struct perf_event_context *task_ctx)
5247 struct perf_cpu_context *cpuctx;
5248 struct perf_event_context *ctx;
5253 list_for_each_entry_rcu(pmu, &pmus, entry) {
5254 cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
5255 if (cpuctx->unique_pmu != pmu)
5257 perf_event_aux_ctx(&cpuctx->ctx, output, data);
5260 ctxn = pmu->task_ctx_nr;
5263 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
5265 perf_event_aux_ctx(ctx, output, data);
5267 put_cpu_ptr(pmu->pmu_cpu_context);
5272 perf_event_aux_ctx(task_ctx, output, data);
5279 * task tracking -- fork/exit
5281 * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
5284 struct perf_task_event {
5285 struct task_struct *task;
5286 struct perf_event_context *task_ctx;
5289 struct perf_event_header header;
5299 static int perf_event_task_match(struct perf_event *event)
5301 return event->attr.comm || event->attr.mmap ||
5302 event->attr.mmap2 || event->attr.mmap_data ||
5306 static void perf_event_task_output(struct perf_event *event,
5309 struct perf_task_event *task_event = data;
5310 struct perf_output_handle handle;
5311 struct perf_sample_data sample;
5312 struct task_struct *task = task_event->task;
5313 int ret, size = task_event->event_id.header.size;
5315 if (!perf_event_task_match(event))
5318 perf_event_header__init_id(&task_event->event_id.header, &sample, event);
5320 ret = perf_output_begin(&handle, event,
5321 task_event->event_id.header.size);
5325 task_event->event_id.pid = perf_event_pid(event, task);
5326 task_event->event_id.ppid = perf_event_pid(event, current);
5328 task_event->event_id.tid = perf_event_tid(event, task);
5329 task_event->event_id.ptid = perf_event_tid(event, current);
5331 perf_output_put(&handle, task_event->event_id);
5333 perf_event__output_id_sample(event, &handle, &sample);
5335 perf_output_end(&handle);
5337 task_event->event_id.header.size = size;
5340 static void perf_event_task(struct task_struct *task,
5341 struct perf_event_context *task_ctx,
5344 struct perf_task_event task_event;
5346 if (!atomic_read(&nr_comm_events) &&
5347 !atomic_read(&nr_mmap_events) &&
5348 !atomic_read(&nr_task_events))
5351 task_event = (struct perf_task_event){
5353 .task_ctx = task_ctx,
5356 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
5358 .size = sizeof(task_event.event_id),
5364 .time = perf_clock(),
5368 perf_event_aux(perf_event_task_output,
5373 void perf_event_fork(struct task_struct *task)
5375 perf_event_task(task, NULL, 1);
5382 struct perf_comm_event {
5383 struct task_struct *task;
5388 struct perf_event_header header;
5395 static int perf_event_comm_match(struct perf_event *event)
5397 return event->attr.comm;
5400 static void perf_event_comm_output(struct perf_event *event,
5403 struct perf_comm_event *comm_event = data;
5404 struct perf_output_handle handle;
5405 struct perf_sample_data sample;
5406 int size = comm_event->event_id.header.size;
5409 if (!perf_event_comm_match(event))
5412 perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
5413 ret = perf_output_begin(&handle, event,
5414 comm_event->event_id.header.size);
5419 comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
5420 comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
5422 perf_output_put(&handle, comm_event->event_id);
5423 __output_copy(&handle, comm_event->comm,
5424 comm_event->comm_size);
5426 perf_event__output_id_sample(event, &handle, &sample);
5428 perf_output_end(&handle);
5430 comm_event->event_id.header.size = size;
5433 static void perf_event_comm_event(struct perf_comm_event *comm_event)
5435 char comm[TASK_COMM_LEN];
5438 memset(comm, 0, sizeof(comm));
5439 strlcpy(comm, comm_event->task->comm, sizeof(comm));
5440 size = ALIGN(strlen(comm)+1, sizeof(u64));
5442 comm_event->comm = comm;
5443 comm_event->comm_size = size;
5445 comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
5447 perf_event_aux(perf_event_comm_output,
5452 void perf_event_comm(struct task_struct *task, bool exec)
5454 struct perf_comm_event comm_event;
5456 if (!atomic_read(&nr_comm_events))
5459 comm_event = (struct perf_comm_event){
5465 .type = PERF_RECORD_COMM,
5466 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
5474 perf_event_comm_event(&comm_event);
5481 struct perf_mmap_event {
5482 struct vm_area_struct *vma;
5484 const char *file_name;
5492 struct perf_event_header header;
5502 static int perf_event_mmap_match(struct perf_event *event,
5505 struct perf_mmap_event *mmap_event = data;
5506 struct vm_area_struct *vma = mmap_event->vma;
5507 int executable = vma->vm_flags & VM_EXEC;
5509 return (!executable && event->attr.mmap_data) ||
5510 (executable && (event->attr.mmap || event->attr.mmap2));
5513 static void perf_event_mmap_output(struct perf_event *event,
5516 struct perf_mmap_event *mmap_event = data;
5517 struct perf_output_handle handle;
5518 struct perf_sample_data sample;
5519 int size = mmap_event->event_id.header.size;
5522 if (!perf_event_mmap_match(event, data))
5525 if (event->attr.mmap2) {
5526 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
5527 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
5528 mmap_event->event_id.header.size += sizeof(mmap_event->min);
5529 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
5530 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
5531 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
5532 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
5535 perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
5536 ret = perf_output_begin(&handle, event,
5537 mmap_event->event_id.header.size);
5541 mmap_event->event_id.pid = perf_event_pid(event, current);
5542 mmap_event->event_id.tid = perf_event_tid(event, current);
5544 perf_output_put(&handle, mmap_event->event_id);
5546 if (event->attr.mmap2) {
5547 perf_output_put(&handle, mmap_event->maj);
5548 perf_output_put(&handle, mmap_event->min);
5549 perf_output_put(&handle, mmap_event->ino);
5550 perf_output_put(&handle, mmap_event->ino_generation);
5551 perf_output_put(&handle, mmap_event->prot);
5552 perf_output_put(&handle, mmap_event->flags);
5555 __output_copy(&handle, mmap_event->file_name,
5556 mmap_event->file_size);
5558 perf_event__output_id_sample(event, &handle, &sample);
5560 perf_output_end(&handle);
5562 mmap_event->event_id.header.size = size;
5565 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
5567 struct vm_area_struct *vma = mmap_event->vma;
5568 struct file *file = vma->vm_file;
5569 int maj = 0, min = 0;
5570 u64 ino = 0, gen = 0;
5571 u32 prot = 0, flags = 0;
5578 struct inode *inode;
5581 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5587 * d_path() works from the end of the rb backwards, so we
5588 * need to add enough zero bytes after the string to handle
5589 * the 64bit alignment we do later.
5591 name = d_path(&file->f_path, buf, PATH_MAX - sizeof(u64));
5596 inode = file_inode(vma->vm_file);
5597 dev = inode->i_sb->s_dev;
5599 gen = inode->i_generation;
5603 if (vma->vm_flags & VM_READ)
5605 if (vma->vm_flags & VM_WRITE)
5607 if (vma->vm_flags & VM_EXEC)
5610 if (vma->vm_flags & VM_MAYSHARE)
5613 flags = MAP_PRIVATE;
5615 if (vma->vm_flags & VM_DENYWRITE)
5616 flags |= MAP_DENYWRITE;
5617 if (vma->vm_flags & VM_MAYEXEC)
5618 flags |= MAP_EXECUTABLE;
5619 if (vma->vm_flags & VM_LOCKED)
5620 flags |= MAP_LOCKED;
5621 if (vma->vm_flags & VM_HUGETLB)
5622 flags |= MAP_HUGETLB;
5626 if (vma->vm_ops && vma->vm_ops->name) {
5627 name = (char *) vma->vm_ops->name(vma);
5632 name = (char *)arch_vma_name(vma);
5636 if (vma->vm_start <= vma->vm_mm->start_brk &&
5637 vma->vm_end >= vma->vm_mm->brk) {
5641 if (vma->vm_start <= vma->vm_mm->start_stack &&
5642 vma->vm_end >= vma->vm_mm->start_stack) {
5652 strlcpy(tmp, name, sizeof(tmp));
5656 * Since our buffer works in 8 byte units we need to align our string
5657 * size to a multiple of 8. However, we must guarantee the tail end is
5658 * zero'd out to avoid leaking random bits to userspace.
5660 size = strlen(name)+1;
5661 while (!IS_ALIGNED(size, sizeof(u64)))
5662 name[size++] = '\0';
5664 mmap_event->file_name = name;
5665 mmap_event->file_size = size;
5666 mmap_event->maj = maj;
5667 mmap_event->min = min;
5668 mmap_event->ino = ino;
5669 mmap_event->ino_generation = gen;
5670 mmap_event->prot = prot;
5671 mmap_event->flags = flags;
5673 if (!(vma->vm_flags & VM_EXEC))
5674 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
5676 mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
5678 perf_event_aux(perf_event_mmap_output,
5685 void perf_event_mmap(struct vm_area_struct *vma)
5687 struct perf_mmap_event mmap_event;
5689 if (!atomic_read(&nr_mmap_events))
5692 mmap_event = (struct perf_mmap_event){
5698 .type = PERF_RECORD_MMAP,
5699 .misc = PERF_RECORD_MISC_USER,
5704 .start = vma->vm_start,
5705 .len = vma->vm_end - vma->vm_start,
5706 .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
5708 /* .maj (attr_mmap2 only) */
5709 /* .min (attr_mmap2 only) */
5710 /* .ino (attr_mmap2 only) */
5711 /* .ino_generation (attr_mmap2 only) */
5712 /* .prot (attr_mmap2 only) */
5713 /* .flags (attr_mmap2 only) */
5716 perf_event_mmap_event(&mmap_event);
5720 * IRQ throttle logging
5723 static void perf_log_throttle(struct perf_event *event, int enable)
5725 struct perf_output_handle handle;
5726 struct perf_sample_data sample;
5730 struct perf_event_header header;
5734 } throttle_event = {
5736 .type = PERF_RECORD_THROTTLE,
5738 .size = sizeof(throttle_event),
5740 .time = perf_clock(),
5741 .id = primary_event_id(event),
5742 .stream_id = event->id,
5746 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
5748 perf_event_header__init_id(&throttle_event.header, &sample, event);
5750 ret = perf_output_begin(&handle, event,
5751 throttle_event.header.size);
5755 perf_output_put(&handle, throttle_event);
5756 perf_event__output_id_sample(event, &handle, &sample);
5757 perf_output_end(&handle);
5761 * Generic event overflow handling, sampling.
5764 static int __perf_event_overflow(struct perf_event *event,
5765 int throttle, struct perf_sample_data *data,
5766 struct pt_regs *regs)
5768 int events = atomic_read(&event->event_limit);
5769 struct hw_perf_event *hwc = &event->hw;
5774 * Non-sampling counters might still use the PMI to fold short
5775 * hardware counters, ignore those.
5777 if (unlikely(!is_sampling_event(event)))
5780 seq = __this_cpu_read(perf_throttled_seq);
5781 if (seq != hwc->interrupts_seq) {
5782 hwc->interrupts_seq = seq;
5783 hwc->interrupts = 1;
5786 if (unlikely(throttle
5787 && hwc->interrupts >= max_samples_per_tick)) {
5788 __this_cpu_inc(perf_throttled_count);
5789 hwc->interrupts = MAX_INTERRUPTS;
5790 perf_log_throttle(event, 0);
5791 tick_nohz_full_kick();
5796 if (event->attr.freq) {
5797 u64 now = perf_clock();
5798 s64 delta = now - hwc->freq_time_stamp;
5800 hwc->freq_time_stamp = now;
5802 if (delta > 0 && delta < 2*TICK_NSEC)
5803 perf_adjust_period(event, delta, hwc->last_period, true);
5807 * XXX event_limit might not quite work as expected on inherited
5811 event->pending_kill = POLL_IN;
5812 if (events && atomic_dec_and_test(&event->event_limit)) {
5814 event->pending_kill = POLL_HUP;
5815 event->pending_disable = 1;
5816 irq_work_queue(&event->pending);
5819 if (event->overflow_handler)
5820 event->overflow_handler(event, data, regs);
5822 perf_event_output(event, data, regs);
5824 if (event->fasync && event->pending_kill) {
5825 event->pending_wakeup = 1;
5826 irq_work_queue(&event->pending);
5832 int perf_event_overflow(struct perf_event *event,
5833 struct perf_sample_data *data,
5834 struct pt_regs *regs)
5836 return __perf_event_overflow(event, 1, data, regs);
5840 * Generic software event infrastructure
5843 struct swevent_htable {
5844 struct swevent_hlist *swevent_hlist;
5845 struct mutex hlist_mutex;
5848 /* Recursion avoidance in each contexts */
5849 int recursion[PERF_NR_CONTEXTS];
5851 /* Keeps track of cpu being initialized/exited */
5855 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
5858 * We directly increment event->count and keep a second value in
5859 * event->hw.period_left to count intervals. This period event
5860 * is kept in the range [-sample_period, 0] so that we can use the
5864 u64 perf_swevent_set_period(struct perf_event *event)
5866 struct hw_perf_event *hwc = &event->hw;
5867 u64 period = hwc->last_period;
5871 hwc->last_period = hwc->sample_period;
5874 old = val = local64_read(&hwc->period_left);
5878 nr = div64_u64(period + val, period);
5879 offset = nr * period;
5881 if (local64_cmpxchg(&hwc->period_left, old, val) != old)
5887 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
5888 struct perf_sample_data *data,
5889 struct pt_regs *regs)
5891 struct hw_perf_event *hwc = &event->hw;
5895 overflow = perf_swevent_set_period(event);
5897 if (hwc->interrupts == MAX_INTERRUPTS)
5900 for (; overflow; overflow--) {
5901 if (__perf_event_overflow(event, throttle,
5904 * We inhibit the overflow from happening when
5905 * hwc->interrupts == MAX_INTERRUPTS.
5913 static void perf_swevent_event(struct perf_event *event, u64 nr,
5914 struct perf_sample_data *data,
5915 struct pt_regs *regs)
5917 struct hw_perf_event *hwc = &event->hw;
5919 local64_add(nr, &event->count);
5924 if (!is_sampling_event(event))
5927 if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
5929 return perf_swevent_overflow(event, 1, data, regs);
5931 data->period = event->hw.last_period;
5933 if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
5934 return perf_swevent_overflow(event, 1, data, regs);
5936 if (local64_add_negative(nr, &hwc->period_left))
5939 perf_swevent_overflow(event, 0, data, regs);
5942 static int perf_exclude_event(struct perf_event *event,
5943 struct pt_regs *regs)
5945 if (event->hw.state & PERF_HES_STOPPED)
5949 if (event->attr.exclude_user && user_mode(regs))
5952 if (event->attr.exclude_kernel && !user_mode(regs))
5959 static int perf_swevent_match(struct perf_event *event,
5960 enum perf_type_id type,
5962 struct perf_sample_data *data,
5963 struct pt_regs *regs)
5965 if (event->attr.type != type)
5968 if (event->attr.config != event_id)
5971 if (perf_exclude_event(event, regs))
5977 static inline u64 swevent_hash(u64 type, u32 event_id)
5979 u64 val = event_id | (type << 32);
5981 return hash_64(val, SWEVENT_HLIST_BITS);
5984 static inline struct hlist_head *
5985 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
5987 u64 hash = swevent_hash(type, event_id);
5989 return &hlist->heads[hash];
5992 /* For the read side: events when they trigger */
5993 static inline struct hlist_head *
5994 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
5996 struct swevent_hlist *hlist;
5998 hlist = rcu_dereference(swhash->swevent_hlist);
6002 return __find_swevent_head(hlist, type, event_id);
6005 /* For the event head insertion and removal in the hlist */
6006 static inline struct hlist_head *
6007 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
6009 struct swevent_hlist *hlist;
6010 u32 event_id = event->attr.config;
6011 u64 type = event->attr.type;
6014 * Event scheduling is always serialized against hlist allocation
6015 * and release. Which makes the protected version suitable here.
6016 * The context lock guarantees that.
6018 hlist = rcu_dereference_protected(swhash->swevent_hlist,
6019 lockdep_is_held(&event->ctx->lock));
6023 return __find_swevent_head(hlist, type, event_id);
6026 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
6028 struct perf_sample_data *data,
6029 struct pt_regs *regs)
6031 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6032 struct perf_event *event;
6033 struct hlist_head *head;
6036 head = find_swevent_head_rcu(swhash, type, event_id);
6040 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6041 if (perf_swevent_match(event, type, event_id, data, regs))
6042 perf_swevent_event(event, nr, data, regs);
6048 DEFINE_PER_CPU(struct pt_regs, __perf_regs[4]);
6050 int perf_swevent_get_recursion_context(void)
6052 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6054 return get_recursion_context(swhash->recursion);
6056 EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
6058 inline void perf_swevent_put_recursion_context(int rctx)
6060 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6062 put_recursion_context(swhash->recursion, rctx);
6065 void ___perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6067 struct perf_sample_data data;
6069 if (WARN_ON_ONCE(!regs))
6072 perf_sample_data_init(&data, addr, 0);
6073 do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, &data, regs);
6076 void __perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
6080 preempt_disable_notrace();
6081 rctx = perf_swevent_get_recursion_context();
6082 if (unlikely(rctx < 0))
6085 ___perf_sw_event(event_id, nr, regs, addr);
6087 perf_swevent_put_recursion_context(rctx);
6089 preempt_enable_notrace();
6092 static void perf_swevent_read(struct perf_event *event)
6096 static int perf_swevent_add(struct perf_event *event, int flags)
6098 struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
6099 struct hw_perf_event *hwc = &event->hw;
6100 struct hlist_head *head;
6102 if (is_sampling_event(event)) {
6103 hwc->last_period = hwc->sample_period;
6104 perf_swevent_set_period(event);
6107 hwc->state = !(flags & PERF_EF_START);
6109 head = find_swevent_head(swhash, event);
6112 * We can race with cpu hotplug code. Do not
6113 * WARN if the cpu just got unplugged.
6115 WARN_ON_ONCE(swhash->online);
6119 hlist_add_head_rcu(&event->hlist_entry, head);
6124 static void perf_swevent_del(struct perf_event *event, int flags)
6126 hlist_del_rcu(&event->hlist_entry);
6129 static void perf_swevent_start(struct perf_event *event, int flags)
6131 event->hw.state = 0;
6134 static void perf_swevent_stop(struct perf_event *event, int flags)
6136 event->hw.state = PERF_HES_STOPPED;
6139 /* Deref the hlist from the update side */
6140 static inline struct swevent_hlist *
6141 swevent_hlist_deref(struct swevent_htable *swhash)
6143 return rcu_dereference_protected(swhash->swevent_hlist,
6144 lockdep_is_held(&swhash->hlist_mutex));
6147 static void swevent_hlist_release(struct swevent_htable *swhash)
6149 struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
6154 RCU_INIT_POINTER(swhash->swevent_hlist, NULL);
6155 kfree_rcu(hlist, rcu_head);
6158 static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
6160 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6162 mutex_lock(&swhash->hlist_mutex);
6164 if (!--swhash->hlist_refcount)
6165 swevent_hlist_release(swhash);
6167 mutex_unlock(&swhash->hlist_mutex);
6170 static void swevent_hlist_put(struct perf_event *event)
6174 for_each_possible_cpu(cpu)
6175 swevent_hlist_put_cpu(event, cpu);
6178 static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
6180 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
6183 mutex_lock(&swhash->hlist_mutex);
6185 if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
6186 struct swevent_hlist *hlist;
6188 hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
6193 rcu_assign_pointer(swhash->swevent_hlist, hlist);
6195 swhash->hlist_refcount++;
6197 mutex_unlock(&swhash->hlist_mutex);
6202 static int swevent_hlist_get(struct perf_event *event)
6205 int cpu, failed_cpu;
6208 for_each_possible_cpu(cpu) {
6209 err = swevent_hlist_get_cpu(event, cpu);
6219 for_each_possible_cpu(cpu) {
6220 if (cpu == failed_cpu)
6222 swevent_hlist_put_cpu(event, cpu);
6229 struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
6231 static void sw_perf_event_destroy(struct perf_event *event)
6233 u64 event_id = event->attr.config;
6235 WARN_ON(event->parent);
6237 static_key_slow_dec(&perf_swevent_enabled[event_id]);
6238 swevent_hlist_put(event);
6241 static int perf_swevent_init(struct perf_event *event)
6243 u64 event_id = event->attr.config;
6245 if (event->attr.type != PERF_TYPE_SOFTWARE)
6249 * no branch sampling for software events
6251 if (has_branch_stack(event))
6255 case PERF_COUNT_SW_CPU_CLOCK:
6256 case PERF_COUNT_SW_TASK_CLOCK:
6263 if (event_id >= PERF_COUNT_SW_MAX)
6266 if (!event->parent) {
6269 err = swevent_hlist_get(event);
6273 static_key_slow_inc(&perf_swevent_enabled[event_id]);
6274 event->destroy = sw_perf_event_destroy;
6280 static struct pmu perf_swevent = {
6281 .task_ctx_nr = perf_sw_context,
6283 .event_init = perf_swevent_init,
6284 .add = perf_swevent_add,
6285 .del = perf_swevent_del,
6286 .start = perf_swevent_start,
6287 .stop = perf_swevent_stop,
6288 .read = perf_swevent_read,
6291 #ifdef CONFIG_EVENT_TRACING
6293 static int perf_tp_filter_match(struct perf_event *event,
6294 struct perf_sample_data *data)
6296 void *record = data->raw->data;
6298 if (likely(!event->filter) || filter_match_preds(event->filter, record))
6303 static int perf_tp_event_match(struct perf_event *event,
6304 struct perf_sample_data *data,
6305 struct pt_regs *regs)
6307 if (event->hw.state & PERF_HES_STOPPED)
6310 * All tracepoints are from kernel-space.
6312 if (event->attr.exclude_kernel)
6315 if (!perf_tp_filter_match(event, data))
6321 void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
6322 struct pt_regs *regs, struct hlist_head *head, int rctx,
6323 struct task_struct *task)
6325 struct perf_sample_data data;
6326 struct perf_event *event;
6328 struct perf_raw_record raw = {
6333 perf_sample_data_init(&data, addr, 0);
6336 hlist_for_each_entry_rcu(event, head, hlist_entry) {
6337 if (perf_tp_event_match(event, &data, regs))
6338 perf_swevent_event(event, count, &data, regs);
6342 * If we got specified a target task, also iterate its context and
6343 * deliver this event there too.
6345 if (task && task != current) {
6346 struct perf_event_context *ctx;
6347 struct trace_entry *entry = record;
6350 ctx = rcu_dereference(task->perf_event_ctxp[perf_sw_context]);
6354 list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6355 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6357 if (event->attr.config != entry->type)
6359 if (perf_tp_event_match(event, &data, regs))
6360 perf_swevent_event(event, count, &data, regs);
6366 perf_swevent_put_recursion_context(rctx);
6368 EXPORT_SYMBOL_GPL(perf_tp_event);
6370 static void tp_perf_event_destroy(struct perf_event *event)
6372 perf_trace_destroy(event);
6375 static int perf_tp_event_init(struct perf_event *event)
6379 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6383 * no branch sampling for tracepoint events
6385 if (has_branch_stack(event))
6388 err = perf_trace_init(event);
6392 event->destroy = tp_perf_event_destroy;
6397 static struct pmu perf_tracepoint = {
6398 .task_ctx_nr = perf_sw_context,
6400 .event_init = perf_tp_event_init,
6401 .add = perf_trace_add,
6402 .del = perf_trace_del,
6403 .start = perf_swevent_start,
6404 .stop = perf_swevent_stop,
6405 .read = perf_swevent_read,
6408 static inline void perf_tp_register(void)
6410 perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
6413 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6418 if (event->attr.type != PERF_TYPE_TRACEPOINT)
6421 filter_str = strndup_user(arg, PAGE_SIZE);
6422 if (IS_ERR(filter_str))
6423 return PTR_ERR(filter_str);
6425 ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
6431 static void perf_event_free_filter(struct perf_event *event)
6433 ftrace_profile_free_filter(event);
6438 static inline void perf_tp_register(void)
6442 static int perf_event_set_filter(struct perf_event *event, void __user *arg)
6447 static void perf_event_free_filter(struct perf_event *event)
6451 #endif /* CONFIG_EVENT_TRACING */
6453 #ifdef CONFIG_HAVE_HW_BREAKPOINT
6454 void perf_bp_event(struct perf_event *bp, void *data)
6456 struct perf_sample_data sample;
6457 struct pt_regs *regs = data;
6459 perf_sample_data_init(&sample, bp->attr.bp_addr, 0);
6461 if (!bp->hw.state && !perf_exclude_event(bp, regs))
6462 perf_swevent_event(bp, 1, &sample, regs);
6467 * hrtimer based swevent callback
6470 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
6472 enum hrtimer_restart ret = HRTIMER_RESTART;
6473 struct perf_sample_data data;
6474 struct pt_regs *regs;
6475 struct perf_event *event;
6478 event = container_of(hrtimer, struct perf_event, hw.hrtimer);
6480 if (event->state != PERF_EVENT_STATE_ACTIVE)
6481 return HRTIMER_NORESTART;
6483 event->pmu->read(event);
6485 perf_sample_data_init(&data, 0, event->hw.last_period);
6486 regs = get_irq_regs();
6488 if (regs && !perf_exclude_event(event, regs)) {
6489 if (!(event->attr.exclude_idle && is_idle_task(current)))
6490 if (__perf_event_overflow(event, 1, &data, regs))
6491 ret = HRTIMER_NORESTART;
6494 period = max_t(u64, 10000, event->hw.sample_period);
6495 hrtimer_forward_now(hrtimer, ns_to_ktime(period));
6500 static void perf_swevent_start_hrtimer(struct perf_event *event)
6502 struct hw_perf_event *hwc = &event->hw;
6505 if (!is_sampling_event(event))
6508 period = local64_read(&hwc->period_left);
6513 local64_set(&hwc->period_left, 0);
6515 period = max_t(u64, 10000, hwc->sample_period);
6517 __hrtimer_start_range_ns(&hwc->hrtimer,
6518 ns_to_ktime(period), 0,
6519 HRTIMER_MODE_REL_PINNED, 0);
6522 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
6524 struct hw_perf_event *hwc = &event->hw;
6526 if (is_sampling_event(event)) {
6527 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
6528 local64_set(&hwc->period_left, ktime_to_ns(remaining));
6530 hrtimer_cancel(&hwc->hrtimer);
6534 static void perf_swevent_init_hrtimer(struct perf_event *event)
6536 struct hw_perf_event *hwc = &event->hw;
6538 if (!is_sampling_event(event))
6541 hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6542 hwc->hrtimer.function = perf_swevent_hrtimer;
6545 * Since hrtimers have a fixed rate, we can do a static freq->period
6546 * mapping and avoid the whole period adjust feedback stuff.
6548 if (event->attr.freq) {
6549 long freq = event->attr.sample_freq;
6551 event->attr.sample_period = NSEC_PER_SEC / freq;
6552 hwc->sample_period = event->attr.sample_period;
6553 local64_set(&hwc->period_left, hwc->sample_period);
6554 hwc->last_period = hwc->sample_period;
6555 event->attr.freq = 0;
6560 * Software event: cpu wall time clock
6563 static void cpu_clock_event_update(struct perf_event *event)
6568 now = local_clock();
6569 prev = local64_xchg(&event->hw.prev_count, now);
6570 local64_add(now - prev, &event->count);
6573 static void cpu_clock_event_start(struct perf_event *event, int flags)
6575 local64_set(&event->hw.prev_count, local_clock());
6576 perf_swevent_start_hrtimer(event);
6579 static void cpu_clock_event_stop(struct perf_event *event, int flags)
6581 perf_swevent_cancel_hrtimer(event);
6582 cpu_clock_event_update(event);
6585 static int cpu_clock_event_add(struct perf_event *event, int flags)
6587 if (flags & PERF_EF_START)
6588 cpu_clock_event_start(event, flags);
6593 static void cpu_clock_event_del(struct perf_event *event, int flags)
6595 cpu_clock_event_stop(event, flags);
6598 static void cpu_clock_event_read(struct perf_event *event)
6600 cpu_clock_event_update(event);
6603 static int cpu_clock_event_init(struct perf_event *event)
6605 if (event->attr.type != PERF_TYPE_SOFTWARE)
6608 if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
6612 * no branch sampling for software events
6614 if (has_branch_stack(event))
6617 perf_swevent_init_hrtimer(event);
6622 static struct pmu perf_cpu_clock = {
6623 .task_ctx_nr = perf_sw_context,
6625 .event_init = cpu_clock_event_init,
6626 .add = cpu_clock_event_add,
6627 .del = cpu_clock_event_del,
6628 .start = cpu_clock_event_start,
6629 .stop = cpu_clock_event_stop,
6630 .read = cpu_clock_event_read,
6634 * Software event: task time clock
6637 static void task_clock_event_update(struct perf_event *event, u64 now)
6642 prev = local64_xchg(&event->hw.prev_count, now);
6644 local64_add(delta, &event->count);
6647 static void task_clock_event_start(struct perf_event *event, int flags)
6649 local64_set(&event->hw.prev_count, event->ctx->time);
6650 perf_swevent_start_hrtimer(event);
6653 static void task_clock_event_stop(struct perf_event *event, int flags)
6655 perf_swevent_cancel_hrtimer(event);
6656 task_clock_event_update(event, event->ctx->time);
6659 static int task_clock_event_add(struct perf_event *event, int flags)
6661 if (flags & PERF_EF_START)
6662 task_clock_event_start(event, flags);
6667 static void task_clock_event_del(struct perf_event *event, int flags)
6669 task_clock_event_stop(event, PERF_EF_UPDATE);
6672 static void task_clock_event_read(struct perf_event *event)
6674 u64 now = perf_clock();
6675 u64 delta = now - event->ctx->timestamp;
6676 u64 time = event->ctx->time + delta;
6678 task_clock_event_update(event, time);
6681 static int task_clock_event_init(struct perf_event *event)
6683 if (event->attr.type != PERF_TYPE_SOFTWARE)
6686 if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
6690 * no branch sampling for software events
6692 if (has_branch_stack(event))
6695 perf_swevent_init_hrtimer(event);
6700 static struct pmu perf_task_clock = {
6701 .task_ctx_nr = perf_sw_context,
6703 .event_init = task_clock_event_init,
6704 .add = task_clock_event_add,
6705 .del = task_clock_event_del,
6706 .start = task_clock_event_start,
6707 .stop = task_clock_event_stop,
6708 .read = task_clock_event_read,
6711 static void perf_pmu_nop_void(struct pmu *pmu)
6715 static int perf_pmu_nop_int(struct pmu *pmu)
6720 static void perf_pmu_start_txn(struct pmu *pmu)
6722 perf_pmu_disable(pmu);
6725 static int perf_pmu_commit_txn(struct pmu *pmu)
6727 perf_pmu_enable(pmu);
6731 static void perf_pmu_cancel_txn(struct pmu *pmu)
6733 perf_pmu_enable(pmu);
6736 static int perf_event_idx_default(struct perf_event *event)
6742 * Ensures all contexts with the same task_ctx_nr have the same
6743 * pmu_cpu_context too.
6745 static struct perf_cpu_context __percpu *find_pmu_context(int ctxn)
6752 list_for_each_entry(pmu, &pmus, entry) {
6753 if (pmu->task_ctx_nr == ctxn)
6754 return pmu->pmu_cpu_context;
6760 static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
6764 for_each_possible_cpu(cpu) {
6765 struct perf_cpu_context *cpuctx;
6767 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6769 if (cpuctx->unique_pmu == old_pmu)
6770 cpuctx->unique_pmu = pmu;
6774 static void free_pmu_context(struct pmu *pmu)
6778 mutex_lock(&pmus_lock);
6780 * Like a real lame refcount.
6782 list_for_each_entry(i, &pmus, entry) {
6783 if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
6784 update_pmu_context(i, pmu);
6789 free_percpu(pmu->pmu_cpu_context);
6791 mutex_unlock(&pmus_lock);
6793 static struct idr pmu_idr;
6796 type_show(struct device *dev, struct device_attribute *attr, char *page)
6798 struct pmu *pmu = dev_get_drvdata(dev);
6800 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
6802 static DEVICE_ATTR_RO(type);
6805 perf_event_mux_interval_ms_show(struct device *dev,
6806 struct device_attribute *attr,
6809 struct pmu *pmu = dev_get_drvdata(dev);
6811 return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
6815 perf_event_mux_interval_ms_store(struct device *dev,
6816 struct device_attribute *attr,
6817 const char *buf, size_t count)
6819 struct pmu *pmu = dev_get_drvdata(dev);
6820 int timer, cpu, ret;
6822 ret = kstrtoint(buf, 0, &timer);
6829 /* same value, noting to do */
6830 if (timer == pmu->hrtimer_interval_ms)
6833 pmu->hrtimer_interval_ms = timer;
6835 /* update all cpuctx for this PMU */
6836 for_each_possible_cpu(cpu) {
6837 struct perf_cpu_context *cpuctx;
6838 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6839 cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
6841 if (hrtimer_active(&cpuctx->hrtimer))
6842 hrtimer_forward_now(&cpuctx->hrtimer, cpuctx->hrtimer_interval);
6847 static DEVICE_ATTR_RW(perf_event_mux_interval_ms);
6849 static struct attribute *pmu_dev_attrs[] = {
6850 &dev_attr_type.attr,
6851 &dev_attr_perf_event_mux_interval_ms.attr,
6854 ATTRIBUTE_GROUPS(pmu_dev);
6856 static int pmu_bus_running;
6857 static struct bus_type pmu_bus = {
6858 .name = "event_source",
6859 .dev_groups = pmu_dev_groups,
6862 static void pmu_dev_release(struct device *dev)
6867 static int pmu_dev_alloc(struct pmu *pmu)
6871 pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
6875 pmu->dev->groups = pmu->attr_groups;
6876 device_initialize(pmu->dev);
6877 ret = dev_set_name(pmu->dev, "%s", pmu->name);
6881 dev_set_drvdata(pmu->dev, pmu);
6882 pmu->dev->bus = &pmu_bus;
6883 pmu->dev->release = pmu_dev_release;
6884 ret = device_add(pmu->dev);
6892 put_device(pmu->dev);
6896 static struct lock_class_key cpuctx_mutex;
6897 static struct lock_class_key cpuctx_lock;
6899 int perf_pmu_register(struct pmu *pmu, const char *name, int type)
6903 mutex_lock(&pmus_lock);
6905 pmu->pmu_disable_count = alloc_percpu(int);
6906 if (!pmu->pmu_disable_count)
6915 type = idr_alloc(&pmu_idr, pmu, PERF_TYPE_MAX, 0, GFP_KERNEL);
6923 if (pmu_bus_running) {
6924 ret = pmu_dev_alloc(pmu);
6930 pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
6931 if (pmu->pmu_cpu_context)
6932 goto got_cpu_context;
6935 pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
6936 if (!pmu->pmu_cpu_context)
6939 for_each_possible_cpu(cpu) {
6940 struct perf_cpu_context *cpuctx;
6942 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
6943 __perf_event_init_context(&cpuctx->ctx);
6944 lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
6945 lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
6946 cpuctx->ctx.pmu = pmu;
6948 __perf_cpu_hrtimer_init(cpuctx, cpu);
6950 INIT_LIST_HEAD(&cpuctx->rotation_list);
6951 cpuctx->unique_pmu = pmu;
6955 if (!pmu->start_txn) {
6956 if (pmu->pmu_enable) {
6958 * If we have pmu_enable/pmu_disable calls, install
6959 * transaction stubs that use that to try and batch
6960 * hardware accesses.
6962 pmu->start_txn = perf_pmu_start_txn;
6963 pmu->commit_txn = perf_pmu_commit_txn;
6964 pmu->cancel_txn = perf_pmu_cancel_txn;
6966 pmu->start_txn = perf_pmu_nop_void;
6967 pmu->commit_txn = perf_pmu_nop_int;
6968 pmu->cancel_txn = perf_pmu_nop_void;
6972 if (!pmu->pmu_enable) {
6973 pmu->pmu_enable = perf_pmu_nop_void;
6974 pmu->pmu_disable = perf_pmu_nop_void;
6977 if (!pmu->event_idx)
6978 pmu->event_idx = perf_event_idx_default;
6980 list_add_rcu(&pmu->entry, &pmus);
6983 mutex_unlock(&pmus_lock);
6988 device_del(pmu->dev);
6989 put_device(pmu->dev);
6992 if (pmu->type >= PERF_TYPE_MAX)
6993 idr_remove(&pmu_idr, pmu->type);
6996 free_percpu(pmu->pmu_disable_count);
6999 EXPORT_SYMBOL_GPL(perf_pmu_register);
7001 void perf_pmu_unregister(struct pmu *pmu)
7003 mutex_lock(&pmus_lock);
7004 list_del_rcu(&pmu->entry);
7005 mutex_unlock(&pmus_lock);
7008 * We dereference the pmu list under both SRCU and regular RCU, so
7009 * synchronize against both of those.
7011 synchronize_srcu(&pmus_srcu);
7014 free_percpu(pmu->pmu_disable_count);
7015 if (pmu->type >= PERF_TYPE_MAX)
7016 idr_remove(&pmu_idr, pmu->type);
7017 device_del(pmu->dev);
7018 put_device(pmu->dev);
7019 free_pmu_context(pmu);
7021 EXPORT_SYMBOL_GPL(perf_pmu_unregister);
7023 struct pmu *perf_init_event(struct perf_event *event)
7025 struct pmu *pmu = NULL;
7029 idx = srcu_read_lock(&pmus_srcu);
7032 pmu = idr_find(&pmu_idr, event->attr.type);
7035 if (!try_module_get(pmu->module)) {
7036 pmu = ERR_PTR(-ENODEV);
7040 ret = pmu->event_init(event);
7046 list_for_each_entry_rcu(pmu, &pmus, entry) {
7047 if (!try_module_get(pmu->module)) {
7048 pmu = ERR_PTR(-ENODEV);
7052 ret = pmu->event_init(event);
7056 if (ret != -ENOENT) {
7061 pmu = ERR_PTR(-ENOENT);
7063 srcu_read_unlock(&pmus_srcu, idx);
7068 static void account_event_cpu(struct perf_event *event, int cpu)
7073 if (has_branch_stack(event)) {
7074 if (!(event->attach_state & PERF_ATTACH_TASK))
7075 atomic_inc(&per_cpu(perf_branch_stack_events, cpu));
7077 if (is_cgroup_event(event))
7078 atomic_inc(&per_cpu(perf_cgroup_events, cpu));
7081 static void account_event(struct perf_event *event)
7086 if (event->attach_state & PERF_ATTACH_TASK)
7087 static_key_slow_inc(&perf_sched_events.key);
7088 if (event->attr.mmap || event->attr.mmap_data)
7089 atomic_inc(&nr_mmap_events);
7090 if (event->attr.comm)
7091 atomic_inc(&nr_comm_events);
7092 if (event->attr.task)
7093 atomic_inc(&nr_task_events);
7094 if (event->attr.freq) {
7095 if (atomic_inc_return(&nr_freq_events) == 1)
7096 tick_nohz_full_kick_all();
7098 if (has_branch_stack(event))
7099 static_key_slow_inc(&perf_sched_events.key);
7100 if (is_cgroup_event(event))
7101 static_key_slow_inc(&perf_sched_events.key);
7103 account_event_cpu(event, event->cpu);
7107 * Allocate and initialize a event structure
7109 static struct perf_event *
7110 perf_event_alloc(struct perf_event_attr *attr, int cpu,
7111 struct task_struct *task,
7112 struct perf_event *group_leader,
7113 struct perf_event *parent_event,
7114 perf_overflow_handler_t overflow_handler,
7118 struct perf_event *event;
7119 struct hw_perf_event *hwc;
7122 if ((unsigned)cpu >= nr_cpu_ids) {
7123 if (!task || cpu != -1)
7124 return ERR_PTR(-EINVAL);
7127 event = kzalloc(sizeof(*event), GFP_KERNEL);
7129 return ERR_PTR(-ENOMEM);
7132 * Single events are their own group leaders, with an
7133 * empty sibling list:
7136 group_leader = event;
7138 mutex_init(&event->child_mutex);
7139 INIT_LIST_HEAD(&event->child_list);
7141 INIT_LIST_HEAD(&event->group_entry);
7142 INIT_LIST_HEAD(&event->event_entry);
7143 INIT_LIST_HEAD(&event->sibling_list);
7144 INIT_LIST_HEAD(&event->rb_entry);
7145 INIT_LIST_HEAD(&event->active_entry);
7146 INIT_HLIST_NODE(&event->hlist_entry);
7149 init_waitqueue_head(&event->waitq);
7150 init_irq_work(&event->pending, perf_pending_event);
7152 mutex_init(&event->mmap_mutex);
7154 atomic_long_set(&event->refcount, 1);
7156 event->attr = *attr;
7157 event->group_leader = group_leader;
7161 event->parent = parent_event;
7163 event->ns = get_pid_ns(task_active_pid_ns(current));
7164 event->id = atomic64_inc_return(&perf_event_id);
7166 event->state = PERF_EVENT_STATE_INACTIVE;
7169 event->attach_state = PERF_ATTACH_TASK;
7171 if (attr->type == PERF_TYPE_TRACEPOINT)
7172 event->hw.tp_target = task;
7173 #ifdef CONFIG_HAVE_HW_BREAKPOINT
7175 * hw_breakpoint is a bit difficult here..
7177 else if (attr->type == PERF_TYPE_BREAKPOINT)
7178 event->hw.bp_target = task;
7182 if (!overflow_handler && parent_event) {
7183 overflow_handler = parent_event->overflow_handler;
7184 context = parent_event->overflow_handler_context;
7187 event->overflow_handler = overflow_handler;
7188 event->overflow_handler_context = context;
7190 perf_event__state_init(event);
7195 hwc->sample_period = attr->sample_period;
7196 if (attr->freq && attr->sample_freq)
7197 hwc->sample_period = 1;
7198 hwc->last_period = hwc->sample_period;
7200 local64_set(&hwc->period_left, hwc->sample_period);
7203 * we currently do not support PERF_FORMAT_GROUP on inherited events
7205 if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
7208 pmu = perf_init_event(event);
7211 else if (IS_ERR(pmu)) {
7216 if (!event->parent) {
7217 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
7218 err = get_callchain_buffers();
7228 event->destroy(event);
7229 module_put(pmu->module);
7232 put_pid_ns(event->ns);
7235 return ERR_PTR(err);
7238 static int perf_copy_attr(struct perf_event_attr __user *uattr,
7239 struct perf_event_attr *attr)
7244 if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
7248 * zero the full structure, so that a short copy will be nice.
7250 memset(attr, 0, sizeof(*attr));
7252 ret = get_user(size, &uattr->size);
7256 if (size > PAGE_SIZE) /* silly large */
7259 if (!size) /* abi compat */
7260 size = PERF_ATTR_SIZE_VER0;
7262 if (size < PERF_ATTR_SIZE_VER0)
7266 * If we're handed a bigger struct than we know of,
7267 * ensure all the unknown bits are 0 - i.e. new
7268 * user-space does not rely on any kernel feature
7269 * extensions we dont know about yet.
7271 if (size > sizeof(*attr)) {
7272 unsigned char __user *addr;
7273 unsigned char __user *end;
7276 addr = (void __user *)uattr + sizeof(*attr);
7277 end = (void __user *)uattr + size;
7279 for (; addr < end; addr++) {
7280 ret = get_user(val, addr);
7286 size = sizeof(*attr);
7289 ret = copy_from_user(attr, uattr, size);
7293 if (attr->__reserved_1)
7296 if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
7299 if (attr->read_format & ~(PERF_FORMAT_MAX-1))
7302 if (attr->sample_type & PERF_SAMPLE_BRANCH_STACK) {
7303 u64 mask = attr->branch_sample_type;
7305 /* only using defined bits */
7306 if (mask & ~(PERF_SAMPLE_BRANCH_MAX-1))
7309 /* at least one branch bit must be set */
7310 if (!(mask & ~PERF_SAMPLE_BRANCH_PLM_ALL))
7313 /* propagate priv level, when not set for branch */
7314 if (!(mask & PERF_SAMPLE_BRANCH_PLM_ALL)) {
7316 /* exclude_kernel checked on syscall entry */
7317 if (!attr->exclude_kernel)
7318 mask |= PERF_SAMPLE_BRANCH_KERNEL;
7320 if (!attr->exclude_user)
7321 mask |= PERF_SAMPLE_BRANCH_USER;
7323 if (!attr->exclude_hv)
7324 mask |= PERF_SAMPLE_BRANCH_HV;
7326 * adjust user setting (for HW filter setup)
7328 attr->branch_sample_type = mask;
7330 /* privileged levels capture (kernel, hv): check permissions */
7331 if ((mask & PERF_SAMPLE_BRANCH_PERM_PLM)
7332 && perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7336 if (attr->sample_type & PERF_SAMPLE_REGS_USER) {
7337 ret = perf_reg_validate(attr->sample_regs_user);
7342 if (attr->sample_type & PERF_SAMPLE_STACK_USER) {
7343 if (!arch_perf_have_user_stack_dump())
7347 * We have __u32 type for the size, but so far
7348 * we can only use __u16 as maximum due to the
7349 * __u16 sample size limit.
7351 if (attr->sample_stack_user >= USHRT_MAX)
7353 else if (!IS_ALIGNED(attr->sample_stack_user, sizeof(u64)))
7357 if (attr->sample_type & PERF_SAMPLE_REGS_INTR)
7358 ret = perf_reg_validate(attr->sample_regs_intr);
7363 put_user(sizeof(*attr), &uattr->size);
7369 perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
7371 struct ring_buffer *rb = NULL;
7377 /* don't allow circular references */
7378 if (event == output_event)
7382 * Don't allow cross-cpu buffers
7384 if (output_event->cpu != event->cpu)
7388 * If its not a per-cpu rb, it must be the same task.
7390 if (output_event->cpu == -1 && output_event->ctx != event->ctx)
7394 mutex_lock(&event->mmap_mutex);
7395 /* Can't redirect output if we've got an active mmap() */
7396 if (atomic_read(&event->mmap_count))
7400 /* get the rb we want to redirect to */
7401 rb = ring_buffer_get(output_event);
7406 ring_buffer_attach(event, rb);
7410 mutex_unlock(&event->mmap_mutex);
7416 static void mutex_lock_double(struct mutex *a, struct mutex *b)
7422 mutex_lock_nested(b, SINGLE_DEPTH_NESTING);
7426 * sys_perf_event_open - open a performance event, associate it to a task/cpu
7428 * @attr_uptr: event_id type attributes for monitoring/sampling
7431 * @group_fd: group leader event fd
7433 SYSCALL_DEFINE5(perf_event_open,
7434 struct perf_event_attr __user *, attr_uptr,
7435 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
7437 struct perf_event *group_leader = NULL, *output_event = NULL;
7438 struct perf_event *event, *sibling;
7439 struct perf_event_attr attr;
7440 struct perf_event_context *ctx, *uninitialized_var(gctx);
7441 struct file *event_file = NULL;
7442 struct fd group = {NULL, 0};
7443 struct task_struct *task = NULL;
7448 int f_flags = O_RDWR;
7450 /* for future expandability... */
7451 if (flags & ~PERF_FLAG_ALL)
7454 err = perf_copy_attr(attr_uptr, &attr);
7458 if (!attr.exclude_kernel) {
7459 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
7464 if (attr.sample_freq > sysctl_perf_event_sample_rate)
7467 if (attr.sample_period & (1ULL << 63))
7472 * In cgroup mode, the pid argument is used to pass the fd
7473 * opened to the cgroup directory in cgroupfs. The cpu argument
7474 * designates the cpu on which to monitor threads from that
7477 if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
7480 if (flags & PERF_FLAG_FD_CLOEXEC)
7481 f_flags |= O_CLOEXEC;
7483 event_fd = get_unused_fd_flags(f_flags);
7487 if (group_fd != -1) {
7488 err = perf_fget_light(group_fd, &group);
7491 group_leader = group.file->private_data;
7492 if (flags & PERF_FLAG_FD_OUTPUT)
7493 output_event = group_leader;
7494 if (flags & PERF_FLAG_FD_NO_GROUP)
7495 group_leader = NULL;
7498 if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
7499 task = find_lively_task_by_vpid(pid);
7501 err = PTR_ERR(task);
7506 if (task && group_leader &&
7507 group_leader->attr.inherit != attr.inherit) {
7514 event = perf_event_alloc(&attr, cpu, task, group_leader, NULL,
7516 if (IS_ERR(event)) {
7517 err = PTR_ERR(event);
7521 if (flags & PERF_FLAG_PID_CGROUP) {
7522 err = perf_cgroup_connect(pid, event, &attr, group_leader);
7524 __free_event(event);
7529 if (is_sampling_event(event)) {
7530 if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
7536 account_event(event);
7539 * Special case software events and allow them to be part of
7540 * any hardware group.
7545 (is_software_event(event) != is_software_event(group_leader))) {
7546 if (is_software_event(event)) {
7548 * If event and group_leader are not both a software
7549 * event, and event is, then group leader is not.
7551 * Allow the addition of software events to !software
7552 * groups, this is safe because software events never
7555 pmu = group_leader->pmu;
7556 } else if (is_software_event(group_leader) &&
7557 (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
7559 * In case the group is a pure software group, and we
7560 * try to add a hardware event, move the whole group to
7561 * the hardware context.
7568 * Get the target context (task or percpu):
7570 ctx = find_get_context(pmu, task, event->cpu);
7577 put_task_struct(task);
7582 * Look up the group leader (we will attach this event to it):
7588 * Do not allow a recursive hierarchy (this new sibling
7589 * becoming part of another group-sibling):
7591 if (group_leader->group_leader != group_leader)
7594 * Do not allow to attach to a group in a different
7595 * task or CPU context:
7599 * Make sure we're both on the same task, or both
7602 if (group_leader->ctx->task != ctx->task)
7606 * Make sure we're both events for the same CPU;
7607 * grouping events for different CPUs is broken; since
7608 * you can never concurrently schedule them anyhow.
7610 if (group_leader->cpu != event->cpu)
7613 if (group_leader->ctx != ctx)
7618 * Only a group leader can be exclusive or pinned
7620 if (attr.exclusive || attr.pinned)
7625 err = perf_event_set_output(event, output_event);
7630 event_file = anon_inode_getfile("[perf_event]", &perf_fops, event,
7632 if (IS_ERR(event_file)) {
7633 err = PTR_ERR(event_file);
7638 gctx = group_leader->ctx;
7641 * See perf_event_ctx_lock() for comments on the details
7642 * of swizzling perf_event::ctx.
7644 mutex_lock_double(&gctx->mutex, &ctx->mutex);
7646 perf_remove_from_context(group_leader, false);
7649 * Removing from the context ends up with disabled
7650 * event. What we want here is event in the initial
7651 * startup state, ready to be add into new context.
7653 perf_event__state_init(group_leader);
7654 list_for_each_entry(sibling, &group_leader->sibling_list,
7656 perf_remove_from_context(sibling, false);
7657 perf_event__state_init(sibling);
7661 mutex_lock(&ctx->mutex);
7664 WARN_ON_ONCE(ctx->parent_ctx);
7668 * Wait for everybody to stop referencing the events through
7669 * the old lists, before installing it on new lists.
7673 perf_install_in_context(ctx, group_leader, group_leader->cpu);
7675 list_for_each_entry(sibling, &group_leader->sibling_list,
7677 perf_install_in_context(ctx, sibling, sibling->cpu);
7682 perf_install_in_context(ctx, event, event->cpu);
7683 perf_unpin_context(ctx);
7686 mutex_unlock(&gctx->mutex);
7689 mutex_unlock(&ctx->mutex);
7693 event->owner = current;
7695 mutex_lock(¤t->perf_event_mutex);
7696 list_add_tail(&event->owner_entry, ¤t->perf_event_list);
7697 mutex_unlock(¤t->perf_event_mutex);
7700 * Precalculate sample_data sizes
7702 perf_event__header_size(event);
7703 perf_event__id_header_size(event);
7706 * Drop the reference on the group_event after placing the
7707 * new event on the sibling_list. This ensures destruction
7708 * of the group leader will find the pointer to itself in
7709 * perf_group_detach().
7712 fd_install(event_fd, event_file);
7716 perf_unpin_context(ctx);
7724 put_task_struct(task);
7728 put_unused_fd(event_fd);
7733 * perf_event_create_kernel_counter
7735 * @attr: attributes of the counter to create
7736 * @cpu: cpu in which the counter is bound
7737 * @task: task to profile (NULL for percpu)
7740 perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
7741 struct task_struct *task,
7742 perf_overflow_handler_t overflow_handler,
7745 struct perf_event_context *ctx;
7746 struct perf_event *event;
7750 * Get the target context (task or percpu):
7753 event = perf_event_alloc(attr, cpu, task, NULL, NULL,
7754 overflow_handler, context);
7755 if (IS_ERR(event)) {
7756 err = PTR_ERR(event);
7760 /* Mark owner so we could distinguish it from user events. */
7761 event->owner = EVENT_OWNER_KERNEL;
7763 account_event(event);
7765 ctx = find_get_context(event->pmu, task, cpu);
7771 WARN_ON_ONCE(ctx->parent_ctx);
7772 mutex_lock(&ctx->mutex);
7773 perf_install_in_context(ctx, event, cpu);
7774 perf_unpin_context(ctx);
7775 mutex_unlock(&ctx->mutex);
7782 return ERR_PTR(err);
7784 EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
7786 void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
7788 struct perf_event_context *src_ctx;
7789 struct perf_event_context *dst_ctx;
7790 struct perf_event *event, *tmp;
7793 src_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, src_cpu)->ctx;
7794 dst_ctx = &per_cpu_ptr(pmu->pmu_cpu_context, dst_cpu)->ctx;
7797 * See perf_event_ctx_lock() for comments on the details
7798 * of swizzling perf_event::ctx.
7800 mutex_lock_double(&src_ctx->mutex, &dst_ctx->mutex);
7801 list_for_each_entry_safe(event, tmp, &src_ctx->event_list,
7803 perf_remove_from_context(event, false);
7804 unaccount_event_cpu(event, src_cpu);
7806 list_add(&event->migrate_entry, &events);
7811 list_for_each_entry_safe(event, tmp, &events, migrate_entry) {
7812 list_del(&event->migrate_entry);
7813 if (event->state >= PERF_EVENT_STATE_OFF)
7814 event->state = PERF_EVENT_STATE_INACTIVE;
7815 account_event_cpu(event, dst_cpu);
7816 perf_install_in_context(dst_ctx, event, dst_cpu);
7819 mutex_unlock(&dst_ctx->mutex);
7820 mutex_unlock(&src_ctx->mutex);
7822 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
7824 static void sync_child_event(struct perf_event *child_event,
7825 struct task_struct *child)
7827 struct perf_event *parent_event = child_event->parent;
7830 if (child_event->attr.inherit_stat)
7831 perf_event_read_event(child_event, child);
7833 child_val = perf_event_count(child_event);
7836 * Add back the child's count to the parent's count:
7838 atomic64_add(child_val, &parent_event->child_count);
7839 atomic64_add(child_event->total_time_enabled,
7840 &parent_event->child_total_time_enabled);
7841 atomic64_add(child_event->total_time_running,
7842 &parent_event->child_total_time_running);
7845 * Remove this event from the parent's list
7847 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
7848 mutex_lock(&parent_event->child_mutex);
7849 list_del_init(&child_event->child_list);
7850 mutex_unlock(&parent_event->child_mutex);
7853 * Make sure user/parent get notified, that we just
7856 perf_event_wakeup(parent_event);
7859 * Release the parent event, if this was the last
7862 put_event(parent_event);
7866 __perf_event_exit_task(struct perf_event *child_event,
7867 struct perf_event_context *child_ctx,
7868 struct task_struct *child)
7871 * Do not destroy the 'original' grouping; because of the context
7872 * switch optimization the original events could've ended up in a
7873 * random child task.
7875 * If we were to destroy the original group, all group related
7876 * operations would cease to function properly after this random
7879 * Do destroy all inherited groups, we don't care about those
7880 * and being thorough is better.
7882 perf_remove_from_context(child_event, !!child_event->parent);
7885 * It can happen that the parent exits first, and has events
7886 * that are still around due to the child reference. These
7887 * events need to be zapped.
7889 if (child_event->parent) {
7890 sync_child_event(child_event, child);
7891 free_event(child_event);
7893 child_event->state = PERF_EVENT_STATE_EXIT;
7894 perf_event_wakeup(child_event);
7898 static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
7900 struct perf_event *child_event, *next;
7901 struct perf_event_context *child_ctx, *clone_ctx = NULL;
7902 unsigned long flags;
7904 if (likely(!child->perf_event_ctxp[ctxn])) {
7905 perf_event_task(child, NULL, 0);
7909 local_irq_save(flags);
7911 * We can't reschedule here because interrupts are disabled,
7912 * and either child is current or it is a task that can't be
7913 * scheduled, so we are now safe from rescheduling changing
7916 child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
7919 * Take the context lock here so that if find_get_context is
7920 * reading child->perf_event_ctxp, we wait until it has
7921 * incremented the context's refcount before we do put_ctx below.
7923 raw_spin_lock(&child_ctx->lock);
7924 task_ctx_sched_out(child_ctx);
7925 child->perf_event_ctxp[ctxn] = NULL;
7928 * If this context is a clone; unclone it so it can't get
7929 * swapped to another process while we're removing all
7930 * the events from it.
7932 clone_ctx = unclone_ctx(child_ctx);
7933 update_context_time(child_ctx);
7934 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
7940 * Report the task dead after unscheduling the events so that we
7941 * won't get any samples after PERF_RECORD_EXIT. We can however still
7942 * get a few PERF_RECORD_READ events.
7944 perf_event_task(child, child_ctx, 0);
7947 * We can recurse on the same lock type through:
7949 * __perf_event_exit_task()
7950 * sync_child_event()
7952 * mutex_lock(&ctx->mutex)
7954 * But since its the parent context it won't be the same instance.
7956 mutex_lock(&child_ctx->mutex);
7958 list_for_each_entry_safe(child_event, next, &child_ctx->event_list, event_entry)
7959 __perf_event_exit_task(child_event, child_ctx, child);
7961 mutex_unlock(&child_ctx->mutex);
7967 * When a child task exits, feed back event values to parent events.
7969 void perf_event_exit_task(struct task_struct *child)
7971 struct perf_event *event, *tmp;
7974 mutex_lock(&child->perf_event_mutex);
7975 list_for_each_entry_safe(event, tmp, &child->perf_event_list,
7977 list_del_init(&event->owner_entry);
7980 * Ensure the list deletion is visible before we clear
7981 * the owner, closes a race against perf_release() where
7982 * we need to serialize on the owner->perf_event_mutex.
7985 event->owner = NULL;
7987 mutex_unlock(&child->perf_event_mutex);
7989 for_each_task_context_nr(ctxn)
7990 perf_event_exit_task_context(child, ctxn);
7993 static void perf_free_event(struct perf_event *event,
7994 struct perf_event_context *ctx)
7996 struct perf_event *parent = event->parent;
7998 if (WARN_ON_ONCE(!parent))
8001 mutex_lock(&parent->child_mutex);
8002 list_del_init(&event->child_list);
8003 mutex_unlock(&parent->child_mutex);
8007 raw_spin_lock_irq(&ctx->lock);
8008 perf_group_detach(event);
8009 list_del_event(event, ctx);
8010 raw_spin_unlock_irq(&ctx->lock);
8015 * Free an unexposed, unused context as created by inheritance by
8016 * perf_event_init_task below, used by fork() in case of fail.
8018 * Not all locks are strictly required, but take them anyway to be nice and
8019 * help out with the lockdep assertions.
8021 void perf_event_free_task(struct task_struct *task)
8023 struct perf_event_context *ctx;
8024 struct perf_event *event, *tmp;
8027 for_each_task_context_nr(ctxn) {
8028 ctx = task->perf_event_ctxp[ctxn];
8032 mutex_lock(&ctx->mutex);
8034 list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
8036 perf_free_event(event, ctx);
8038 list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
8040 perf_free_event(event, ctx);
8042 if (!list_empty(&ctx->pinned_groups) ||
8043 !list_empty(&ctx->flexible_groups))
8046 mutex_unlock(&ctx->mutex);
8052 void perf_event_delayed_put(struct task_struct *task)
8056 for_each_task_context_nr(ctxn)
8057 WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
8061 * inherit a event from parent task to child task:
8063 static struct perf_event *
8064 inherit_event(struct perf_event *parent_event,
8065 struct task_struct *parent,
8066 struct perf_event_context *parent_ctx,
8067 struct task_struct *child,
8068 struct perf_event *group_leader,
8069 struct perf_event_context *child_ctx)
8071 enum perf_event_active_state parent_state = parent_event->state;
8072 struct perf_event *child_event;
8073 unsigned long flags;
8076 * Instead of creating recursive hierarchies of events,
8077 * we link inherited events back to the original parent,
8078 * which has a filp for sure, which we use as the reference
8081 if (parent_event->parent)
8082 parent_event = parent_event->parent;
8084 child_event = perf_event_alloc(&parent_event->attr,
8087 group_leader, parent_event,
8089 if (IS_ERR(child_event))
8092 if (is_orphaned_event(parent_event) ||
8093 !atomic_long_inc_not_zero(&parent_event->refcount)) {
8094 free_event(child_event);
8101 * Make the child state follow the state of the parent event,
8102 * not its attr.disabled bit. We hold the parent's mutex,
8103 * so we won't race with perf_event_{en, dis}able_family.
8105 if (parent_state >= PERF_EVENT_STATE_INACTIVE)
8106 child_event->state = PERF_EVENT_STATE_INACTIVE;
8108 child_event->state = PERF_EVENT_STATE_OFF;
8110 if (parent_event->attr.freq) {
8111 u64 sample_period = parent_event->hw.sample_period;
8112 struct hw_perf_event *hwc = &child_event->hw;
8114 hwc->sample_period = sample_period;
8115 hwc->last_period = sample_period;
8117 local64_set(&hwc->period_left, sample_period);
8120 child_event->ctx = child_ctx;
8121 child_event->overflow_handler = parent_event->overflow_handler;
8122 child_event->overflow_handler_context
8123 = parent_event->overflow_handler_context;
8126 * Precalculate sample_data sizes
8128 perf_event__header_size(child_event);
8129 perf_event__id_header_size(child_event);
8132 * Link it up in the child's context:
8134 raw_spin_lock_irqsave(&child_ctx->lock, flags);
8135 add_event_to_ctx(child_event, child_ctx);
8136 raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
8139 * Link this into the parent event's child list
8141 WARN_ON_ONCE(parent_event->ctx->parent_ctx);
8142 mutex_lock(&parent_event->child_mutex);
8143 list_add_tail(&child_event->child_list, &parent_event->child_list);
8144 mutex_unlock(&parent_event->child_mutex);
8149 static int inherit_group(struct perf_event *parent_event,
8150 struct task_struct *parent,
8151 struct perf_event_context *parent_ctx,
8152 struct task_struct *child,
8153 struct perf_event_context *child_ctx)
8155 struct perf_event *leader;
8156 struct perf_event *sub;
8157 struct perf_event *child_ctr;
8159 leader = inherit_event(parent_event, parent, parent_ctx,
8160 child, NULL, child_ctx);
8162 return PTR_ERR(leader);
8163 list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
8164 child_ctr = inherit_event(sub, parent, parent_ctx,
8165 child, leader, child_ctx);
8166 if (IS_ERR(child_ctr))
8167 return PTR_ERR(child_ctr);
8173 inherit_task_group(struct perf_event *event, struct task_struct *parent,
8174 struct perf_event_context *parent_ctx,
8175 struct task_struct *child, int ctxn,
8179 struct perf_event_context *child_ctx;
8181 if (!event->attr.inherit) {
8186 child_ctx = child->perf_event_ctxp[ctxn];
8189 * This is executed from the parent task context, so
8190 * inherit events that have been marked for cloning.
8191 * First allocate and initialize a context for the
8195 child_ctx = alloc_perf_context(parent_ctx->pmu, child);
8199 child->perf_event_ctxp[ctxn] = child_ctx;
8202 ret = inherit_group(event, parent, parent_ctx,
8212 * Initialize the perf_event context in task_struct
8214 static int perf_event_init_context(struct task_struct *child, int ctxn)
8216 struct perf_event_context *child_ctx, *parent_ctx;
8217 struct perf_event_context *cloned_ctx;
8218 struct perf_event *event;
8219 struct task_struct *parent = current;
8220 int inherited_all = 1;
8221 unsigned long flags;
8224 if (likely(!parent->perf_event_ctxp[ctxn]))
8228 * If the parent's context is a clone, pin it so it won't get
8231 parent_ctx = perf_pin_task_context(parent, ctxn);
8236 * No need to check if parent_ctx != NULL here; since we saw
8237 * it non-NULL earlier, the only reason for it to become NULL
8238 * is if we exit, and since we're currently in the middle of
8239 * a fork we can't be exiting at the same time.
8243 * Lock the parent list. No need to lock the child - not PID
8244 * hashed yet and not running, so nobody can access it.
8246 mutex_lock(&parent_ctx->mutex);
8249 * We dont have to disable NMIs - we are only looking at
8250 * the list, not manipulating it:
8252 list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
8253 ret = inherit_task_group(event, parent, parent_ctx,
8254 child, ctxn, &inherited_all);
8260 * We can't hold ctx->lock when iterating the ->flexible_group list due
8261 * to allocations, but we need to prevent rotation because
8262 * rotate_ctx() will change the list from interrupt context.
8264 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8265 parent_ctx->rotate_disable = 1;
8266 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8268 list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
8269 ret = inherit_task_group(event, parent, parent_ctx,
8270 child, ctxn, &inherited_all);
8275 raw_spin_lock_irqsave(&parent_ctx->lock, flags);
8276 parent_ctx->rotate_disable = 0;
8278 child_ctx = child->perf_event_ctxp[ctxn];
8280 if (child_ctx && inherited_all) {
8282 * Mark the child context as a clone of the parent
8283 * context, or of whatever the parent is a clone of.
8285 * Note that if the parent is a clone, the holding of
8286 * parent_ctx->lock avoids it from being uncloned.
8288 cloned_ctx = parent_ctx->parent_ctx;
8290 child_ctx->parent_ctx = cloned_ctx;
8291 child_ctx->parent_gen = parent_ctx->parent_gen;
8293 child_ctx->parent_ctx = parent_ctx;
8294 child_ctx->parent_gen = parent_ctx->generation;
8296 get_ctx(child_ctx->parent_ctx);
8299 raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
8300 mutex_unlock(&parent_ctx->mutex);
8302 perf_unpin_context(parent_ctx);
8303 put_ctx(parent_ctx);
8309 * Initialize the perf_event context in task_struct
8311 int perf_event_init_task(struct task_struct *child)
8315 memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
8316 mutex_init(&child->perf_event_mutex);
8317 INIT_LIST_HEAD(&child->perf_event_list);
8319 for_each_task_context_nr(ctxn) {
8320 ret = perf_event_init_context(child, ctxn);
8322 perf_event_free_task(child);
8330 static void __init perf_event_init_all_cpus(void)
8332 struct swevent_htable *swhash;
8335 for_each_possible_cpu(cpu) {
8336 swhash = &per_cpu(swevent_htable, cpu);
8337 mutex_init(&swhash->hlist_mutex);
8338 INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
8342 static void perf_event_init_cpu(int cpu)
8344 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8346 mutex_lock(&swhash->hlist_mutex);
8347 swhash->online = true;
8348 if (swhash->hlist_refcount > 0) {
8349 struct swevent_hlist *hlist;
8351 hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
8353 rcu_assign_pointer(swhash->swevent_hlist, hlist);
8355 mutex_unlock(&swhash->hlist_mutex);
8358 #if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
8359 static void perf_pmu_rotate_stop(struct pmu *pmu)
8361 struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
8363 WARN_ON(!irqs_disabled());
8365 list_del_init(&cpuctx->rotation_list);
8368 static void __perf_event_exit_context(void *__info)
8370 struct remove_event re = { .detach_group = true };
8371 struct perf_event_context *ctx = __info;
8373 perf_pmu_rotate_stop(ctx->pmu);
8376 list_for_each_entry_rcu(re.event, &ctx->event_list, event_entry)
8377 __perf_remove_from_context(&re);
8381 static void perf_event_exit_cpu_context(int cpu)
8383 struct perf_event_context *ctx;
8387 idx = srcu_read_lock(&pmus_srcu);
8388 list_for_each_entry_rcu(pmu, &pmus, entry) {
8389 ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
8391 mutex_lock(&ctx->mutex);
8392 smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
8393 mutex_unlock(&ctx->mutex);
8395 srcu_read_unlock(&pmus_srcu, idx);
8398 static void perf_event_exit_cpu(int cpu)
8400 struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
8402 perf_event_exit_cpu_context(cpu);
8404 mutex_lock(&swhash->hlist_mutex);
8405 swhash->online = false;
8406 swevent_hlist_release(swhash);
8407 mutex_unlock(&swhash->hlist_mutex);
8410 static inline void perf_event_exit_cpu(int cpu) { }
8414 perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
8418 for_each_online_cpu(cpu)
8419 perf_event_exit_cpu(cpu);
8425 * Run the perf reboot notifier at the very last possible moment so that
8426 * the generic watchdog code runs as long as possible.
8428 static struct notifier_block perf_reboot_notifier = {
8429 .notifier_call = perf_reboot,
8430 .priority = INT_MIN,
8434 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
8436 unsigned int cpu = (long)hcpu;
8438 switch (action & ~CPU_TASKS_FROZEN) {
8440 case CPU_UP_PREPARE:
8441 case CPU_DOWN_FAILED:
8442 perf_event_init_cpu(cpu);
8445 case CPU_UP_CANCELED:
8446 case CPU_DOWN_PREPARE:
8447 perf_event_exit_cpu(cpu);
8456 void __init perf_event_init(void)
8462 perf_event_init_all_cpus();
8463 init_srcu_struct(&pmus_srcu);
8464 perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
8465 perf_pmu_register(&perf_cpu_clock, NULL, -1);
8466 perf_pmu_register(&perf_task_clock, NULL, -1);
8468 perf_cpu_notifier(perf_cpu_notify);
8469 register_reboot_notifier(&perf_reboot_notifier);
8471 ret = init_hw_breakpoint();
8472 WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
8474 /* do not patch jump label more than once per second */
8475 jump_label_rate_limit(&perf_sched_events, HZ);
8478 * Build time assertion that we keep the data_head at the intended
8479 * location. IOW, validation we got the __reserved[] size right.
8481 BUILD_BUG_ON((offsetof(struct perf_event_mmap_page, data_head))
8485 static int __init perf_event_sysfs_init(void)
8490 mutex_lock(&pmus_lock);
8492 ret = bus_register(&pmu_bus);
8496 list_for_each_entry(pmu, &pmus, entry) {
8497 if (!pmu->name || pmu->type < 0)
8500 ret = pmu_dev_alloc(pmu);
8501 WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
8503 pmu_bus_running = 1;
8507 mutex_unlock(&pmus_lock);
8511 device_initcall(perf_event_sysfs_init);
8513 #ifdef CONFIG_CGROUP_PERF
8514 static struct cgroup_subsys_state *
8515 perf_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8517 struct perf_cgroup *jc;
8519 jc = kzalloc(sizeof(*jc), GFP_KERNEL);
8521 return ERR_PTR(-ENOMEM);
8523 jc->info = alloc_percpu(struct perf_cgroup_info);
8526 return ERR_PTR(-ENOMEM);
8532 static void perf_cgroup_css_free(struct cgroup_subsys_state *css)
8534 struct perf_cgroup *jc = container_of(css, struct perf_cgroup, css);
8536 free_percpu(jc->info);
8540 static int __perf_cgroup_move(void *info)
8542 struct task_struct *task = info;
8543 perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
8547 static void perf_cgroup_attach(struct cgroup_subsys_state *css,
8548 struct cgroup_taskset *tset)
8550 struct task_struct *task;
8552 cgroup_taskset_for_each(task, tset)
8553 task_function_call(task, __perf_cgroup_move, task);
8556 static void perf_cgroup_exit(struct cgroup_subsys_state *css,
8557 struct cgroup_subsys_state *old_css,
8558 struct task_struct *task)
8561 * cgroup_exit() is called in the copy_process() failure path.
8562 * Ignore this case since the task hasn't ran yet, this avoids
8563 * trying to poke a half freed task state from generic code.
8565 if (!(task->flags & PF_EXITING))
8568 task_function_call(task, __perf_cgroup_move, task);
8571 struct cgroup_subsys perf_event_cgrp_subsys = {
8572 .css_alloc = perf_cgroup_css_alloc,
8573 .css_free = perf_cgroup_css_free,
8574 .exit = perf_cgroup_exit,
8575 .attach = perf_cgroup_attach,
8577 #endif /* CONFIG_CGROUP_PERF */