~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/events/core.c

Version: ~ [ linux-6.2-rc3 ] ~ [ linux-6.1.5 ] ~ [ linux-6.0.19 ] ~ [ linux-5.19.17 ] ~ [ linux-5.18.19 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.87 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.162 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.228 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.269 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.302 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.337 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0
  2 /*
  3  * Performance events core code:
  4  *
  5  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
  6  *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
  7  *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
  8  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
  9  */
 10 
 11 #include <linux/fs.h>
 12 #include <linux/mm.h>
 13 #include <linux/cpu.h>
 14 #include <linux/smp.h>
 15 #include <linux/idr.h>
 16 #include <linux/file.h>
 17 #include <linux/poll.h>
 18 #include <linux/slab.h>
 19 #include <linux/hash.h>
 20 #include <linux/tick.h>
 21 #include <linux/sysfs.h>
 22 #include <linux/dcache.h>
 23 #include <linux/percpu.h>
 24 #include <linux/ptrace.h>
 25 #include <linux/reboot.h>
 26 #include <linux/vmstat.h>
 27 #include <linux/device.h>
 28 #include <linux/export.h>
 29 #include <linux/vmalloc.h>
 30 #include <linux/hardirq.h>
 31 #include <linux/hugetlb.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/cgroup.h>
 38 #include <linux/perf_event.h>
 39 #include <linux/trace_events.h>
 40 #include <linux/hw_breakpoint.h>
 41 #include <linux/mm_types.h>
 42 #include <linux/module.h>
 43 #include <linux/mman.h>
 44 #include <linux/compat.h>
 45 #include <linux/bpf.h>
 46 #include <linux/filter.h>
 47 #include <linux/namei.h>
 48 #include <linux/parser.h>
 49 #include <linux/sched/clock.h>
 50 #include <linux/sched/mm.h>
 51 #include <linux/proc_ns.h>
 52 #include <linux/mount.h>
 53 #include <linux/min_heap.h>
 54 #include <linux/highmem.h>
 55 #include <linux/pgtable.h>
 56 #include <linux/buildid.h>
 57 
 58 #include "internal.h"
 59 
 60 #include <asm/irq_regs.h>
 61 
 62 typedef int (*remote_function_f)(void *);
 63 
 64 struct remote_function_call {
 65         struct task_struct      *p;
 66         remote_function_f       func;
 67         void                    *info;
 68         int                     ret;
 69 };
 70 
 71 static void remote_function(void *data)
 72 {
 73         struct remote_function_call *tfc = data;
 74         struct task_struct *p = tfc->p;
 75 
 76         if (p) {
 77                 /* -EAGAIN */
 78                 if (task_cpu(p) != smp_processor_id())
 79                         return;
 80 
 81                 /*
 82                  * Now that we're on right CPU with IRQs disabled, we can test
 83                  * if we hit the right task without races.
 84                  */
 85 
 86                 tfc->ret = -ESRCH; /* No such (running) process */
 87                 if (p != current)
 88                         return;
 89         }
 90 
 91         tfc->ret = tfc->func(tfc->info);
 92 }
 93 
 94 /**
 95  * task_function_call - call a function on the cpu on which a task runs
 96  * @p:          the task to evaluate
 97  * @func:       the function to be called
 98  * @info:       the function call argument
 99  *
100  * Calls the function @func when the task is currently running. This might
101  * be on the current CPU, which just calls the function directly.  This will
102  * retry due to any failures in smp_call_function_single(), such as if the
103  * task_cpu() goes offline concurrently.
104  *
105  * returns @func return value or -ESRCH or -ENXIO when the process isn't running
106  */
107 static int
108 task_function_call(struct task_struct *p, remote_function_f func, void *info)
109 {
110         struct remote_function_call data = {
111                 .p      = p,
112                 .func   = func,
113                 .info   = info,
114                 .ret    = -EAGAIN,
115         };
116         int ret;
117 
118         for (;;) {
119                 ret = smp_call_function_single(task_cpu(p), remote_function,
120                                                &data, 1);
121                 if (!ret)
122                         ret = data.ret;
123 
124                 if (ret != -EAGAIN)
125                         break;
126 
127                 cond_resched();
128         }
129 
130         return ret;
131 }
132 
133 /**
134  * cpu_function_call - call a function on the cpu
135  * @cpu:        target cpu to queue this function
136  * @func:       the function to be called
137  * @info:       the function call argument
138  *
139  * Calls the function @func on the remote cpu.
140  *
141  * returns: @func return value or -ENXIO when the cpu is offline
142  */
143 static int cpu_function_call(int cpu, remote_function_f func, void *info)
144 {
145         struct remote_function_call data = {
146                 .p      = NULL,
147                 .func   = func,
148                 .info   = info,
149                 .ret    = -ENXIO, /* No such CPU */
150         };
151 
152         smp_call_function_single(cpu, remote_function, &data, 1);
153 
154         return data.ret;
155 }
156 
157 static inline struct perf_cpu_context *
158 __get_cpu_context(struct perf_event_context *ctx)
159 {
160         return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
161 }
162 
163 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
164                           struct perf_event_context *ctx)
165 {
166         raw_spin_lock(&cpuctx->ctx.lock);
167         if (ctx)
168                 raw_spin_lock(&ctx->lock);
169 }
170 
171 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
172                             struct perf_event_context *ctx)
173 {
174         if (ctx)
175                 raw_spin_unlock(&ctx->lock);
176         raw_spin_unlock(&cpuctx->ctx.lock);
177 }
178 
179 #define TASK_TOMBSTONE ((void *)-1L)
180 
181 static bool is_kernel_event(struct perf_event *event)
182 {
183         return READ_ONCE(event->owner) == TASK_TOMBSTONE;
184 }
185 
186 /*
187  * On task ctx scheduling...
188  *
189  * When !ctx->nr_events a task context will not be scheduled. This means
190  * we can disable the scheduler hooks (for performance) without leaving
191  * pending task ctx state.
192  *
193  * This however results in two special cases:
194  *
195  *  - removing the last event from a task ctx; this is relatively straight
196  *    forward and is done in __perf_remove_from_context.
197  *
198  *  - adding the first event to a task ctx; this is tricky because we cannot
199  *    rely on ctx->is_active and therefore cannot use event_function_call().
200  *    See perf_install_in_context().
201  *
202  * If ctx->nr_events, then ctx->is_active and cpuctx->task_ctx are set.
203  */
204 
205 typedef void (*event_f)(struct perf_event *, struct perf_cpu_context *,
206                         struct perf_event_context *, void *);
207 
208 struct event_function_struct {
209         struct perf_event *event;
210         event_f func;
211         void *data;
212 };
213 
214 static int event_function(void *info)
215 {
216         struct event_function_struct *efs = info;
217         struct perf_event *event = efs->event;
218         struct perf_event_context *ctx = event->ctx;
219         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
220         struct perf_event_context *task_ctx = cpuctx->task_ctx;
221         int ret = 0;
222 
223         lockdep_assert_irqs_disabled();
224 
225         perf_ctx_lock(cpuctx, task_ctx);
226         /*
227          * Since we do the IPI call without holding ctx->lock things can have
228          * changed, double check we hit the task we set out to hit.
229          */
230         if (ctx->task) {
231                 if (ctx->task != current) {
232                         ret = -ESRCH;
233                         goto unlock;
234                 }
235 
236                 /*
237                  * We only use event_function_call() on established contexts,
238                  * and event_function() is only ever called when active (or
239                  * rather, we'll have bailed in task_function_call() or the
240                  * above ctx->task != current test), therefore we must have
241                  * ctx->is_active here.
242                  */
243                 WARN_ON_ONCE(!ctx->is_active);
244                 /*
245                  * And since we have ctx->is_active, cpuctx->task_ctx must
246                  * match.
247                  */
248                 WARN_ON_ONCE(task_ctx != ctx);
249         } else {
250                 WARN_ON_ONCE(&cpuctx->ctx != ctx);
251         }
252 
253         efs->func(event, cpuctx, ctx, efs->data);
254 unlock:
255         perf_ctx_unlock(cpuctx, task_ctx);
256 
257         return ret;
258 }
259 
260 static void event_function_call(struct perf_event *event, event_f func, void *data)
261 {
262         struct perf_event_context *ctx = event->ctx;
263         struct task_struct *task = READ_ONCE(ctx->task); /* verified in event_function */
264         struct event_function_struct efs = {
265                 .event = event,
266                 .func = func,
267                 .data = data,
268         };
269 
270         if (!event->parent) {
271                 /*
272                  * If this is a !child event, we must hold ctx::mutex to
273                  * stabilize the event->ctx relation. See
274                  * perf_event_ctx_lock().
275                  */
276                 lockdep_assert_held(&ctx->mutex);
277         }
278 
279         if (!task) {
280                 cpu_function_call(event->cpu, event_function, &efs);
281                 return;
282         }
283 
284         if (task == TASK_TOMBSTONE)
285                 return;
286 
287 again:
288         if (!task_function_call(task, event_function, &efs))
289                 return;
290 
291         raw_spin_lock_irq(&ctx->lock);
292         /*
293          * Reload the task pointer, it might have been changed by
294          * a concurrent perf_event_context_sched_out().
295          */
296         task = ctx->task;
297         if (task == TASK_TOMBSTONE) {
298                 raw_spin_unlock_irq(&ctx->lock);
299                 return;
300         }
301         if (ctx->is_active) {
302                 raw_spin_unlock_irq(&ctx->lock);
303                 goto again;
304         }
305         func(event, NULL, ctx, data);
306         raw_spin_unlock_irq(&ctx->lock);
307 }
308 
309 /*
310  * Similar to event_function_call() + event_function(), but hard assumes IRQs
311  * are already disabled and we're on the right CPU.
312  */
313 static void event_function_local(struct perf_event *event, event_f func, void *data)
314 {
315         struct perf_event_context *ctx = event->ctx;
316         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
317         struct task_struct *task = READ_ONCE(ctx->task);
318         struct perf_event_context *task_ctx = NULL;
319 
320         lockdep_assert_irqs_disabled();
321 
322         if (task) {
323                 if (task == TASK_TOMBSTONE)
324                         return;
325 
326                 task_ctx = ctx;
327         }
328 
329         perf_ctx_lock(cpuctx, task_ctx);
330 
331         task = ctx->task;
332         if (task == TASK_TOMBSTONE)
333                 goto unlock;
334 
335         if (task) {
336                 /*
337                  * We must be either inactive or active and the right task,
338                  * otherwise we're screwed, since we cannot IPI to somewhere
339                  * else.
340                  */
341                 if (ctx->is_active) {
342                         if (WARN_ON_ONCE(task != current))
343                                 goto unlock;
344 
345                         if (WARN_ON_ONCE(cpuctx->task_ctx != ctx))
346                                 goto unlock;
347                 }
348         } else {
349                 WARN_ON_ONCE(&cpuctx->ctx != ctx);
350         }
351 
352         func(event, cpuctx, ctx, data);
353 unlock:
354         perf_ctx_unlock(cpuctx, task_ctx);
355 }
356 
357 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
358                        PERF_FLAG_FD_OUTPUT  |\
359                        PERF_FLAG_PID_CGROUP |\
360                        PERF_FLAG_FD_CLOEXEC)
361 
362 /*
363  * branch priv levels that need permission checks
364  */
365 #define PERF_SAMPLE_BRANCH_PERM_PLM \
366         (PERF_SAMPLE_BRANCH_KERNEL |\
367          PERF_SAMPLE_BRANCH_HV)
368 
369 enum event_type_t {
370         EVENT_FLEXIBLE = 0x1,
371         EVENT_PINNED = 0x2,
372         EVENT_TIME = 0x4,
373         /* see ctx_resched() for details */
374         EVENT_CPU = 0x8,
375         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
376 };
377 
378 /*
379  * perf_sched_events : >0 events exist
380  * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
381  */
382 
383 static void perf_sched_delayed(struct work_struct *work);
384 DEFINE_STATIC_KEY_FALSE(perf_sched_events);
385 static DECLARE_DELAYED_WORK(perf_sched_work, perf_sched_delayed);
386 static DEFINE_MUTEX(perf_sched_mutex);
387 static atomic_t perf_sched_count;
388 
389 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
390 static DEFINE_PER_CPU(int, perf_sched_cb_usages);
391 static DEFINE_PER_CPU(struct pmu_event_list, pmu_sb_events);
392 
393 static atomic_t nr_mmap_events __read_mostly;
394 static atomic_t nr_comm_events __read_mostly;
395 static atomic_t nr_namespaces_events __read_mostly;
396 static atomic_t nr_task_events __read_mostly;
397 static atomic_t nr_freq_events __read_mostly;
398 static atomic_t nr_switch_events __read_mostly;
399 static atomic_t nr_ksymbol_events __read_mostly;
400 static atomic_t nr_bpf_events __read_mostly;
401 static atomic_t nr_cgroup_events __read_mostly;
402 static atomic_t nr_text_poke_events __read_mostly;
403 static atomic_t nr_build_id_events __read_mostly;
404 
405 static LIST_HEAD(pmus);
406 static DEFINE_MUTEX(pmus_lock);
407 static struct srcu_struct pmus_srcu;
408 static cpumask_var_t perf_online_mask;
409 static struct kmem_cache *perf_event_cache;
410 
411 /*
412  * perf event paranoia level:
413  *  -1 - not paranoid at all
414  *   0 - disallow raw tracepoint access for unpriv
415  *   1 - disallow cpu events for unpriv
416  *   2 - disallow kernel profiling for unpriv
417  */
418 int sysctl_perf_event_paranoid __read_mostly = 2;
419 
420 /* Minimum for 512 kiB + 1 user control page */
421 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
422 
423 /*
424  * max perf event sample rate
425  */
426 #define DEFAULT_MAX_SAMPLE_RATE         100000
427 #define DEFAULT_SAMPLE_PERIOD_NS        (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
428 #define DEFAULT_CPU_TIME_MAX_PERCENT    25
429 
430 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
431 
432 static int max_samples_per_tick __read_mostly   = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
433 static int perf_sample_period_ns __read_mostly  = DEFAULT_SAMPLE_PERIOD_NS;
434 
435 static int perf_sample_allowed_ns __read_mostly =
436         DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
437 
438 static void update_perf_cpu_limits(void)
439 {
440         u64 tmp = perf_sample_period_ns;
441 
442         tmp *= sysctl_perf_cpu_time_max_percent;
443         tmp = div_u64(tmp, 100);
444         if (!tmp)
445                 tmp = 1;
446 
447         WRITE_ONCE(perf_sample_allowed_ns, tmp);
448 }
449 
450 static bool perf_rotate_context(struct perf_cpu_context *cpuctx);
451 
452 int perf_proc_update_handler(struct ctl_table *table, int write,
453                 void *buffer, size_t *lenp, loff_t *ppos)
454 {
455         int ret;
456         int perf_cpu = sysctl_perf_cpu_time_max_percent;
457         /*
458          * If throttling is disabled don't allow the write:
459          */
460         if (write && (perf_cpu == 100 || perf_cpu == 0))
461                 return -EINVAL;
462 
463         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
464         if (ret || !write)
465                 return ret;
466 
467         max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
468         perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
469         update_perf_cpu_limits();
470 
471         return 0;
472 }
473 
474 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
475 
476 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
477                 void *buffer, size_t *lenp, loff_t *ppos)
478 {
479         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
480 
481         if (ret || !write)
482                 return ret;
483 
484         if (sysctl_perf_cpu_time_max_percent == 100 ||
485             sysctl_perf_cpu_time_max_percent == 0) {
486                 printk(KERN_WARNING
487                        "perf: Dynamic interrupt throttling disabled, can hang your system!\n");
488                 WRITE_ONCE(perf_sample_allowed_ns, 0);
489         } else {
490                 update_perf_cpu_limits();
491         }
492 
493         return 0;
494 }
495 
496 /*
497  * perf samples are done in some very critical code paths (NMIs).
498  * If they take too much CPU time, the system can lock up and not
499  * get any real work done.  This will drop the sample rate when
500  * we detect that events are taking too long.
501  */
502 #define NR_ACCUMULATED_SAMPLES 128
503 static DEFINE_PER_CPU(u64, running_sample_length);
504 
505 static u64 __report_avg;
506 static u64 __report_allowed;
507 
508 static void perf_duration_warn(struct irq_work *w)
509 {
510         printk_ratelimited(KERN_INFO
511                 "perf: interrupt took too long (%lld > %lld), lowering "
512                 "kernel.perf_event_max_sample_rate to %d\n",
513                 __report_avg, __report_allowed,
514                 sysctl_perf_event_sample_rate);
515 }
516 
517 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
518 
519 void perf_sample_event_took(u64 sample_len_ns)
520 {
521         u64 max_len = READ_ONCE(perf_sample_allowed_ns);
522         u64 running_len;
523         u64 avg_len;
524         u32 max;
525 
526         if (max_len == 0)
527                 return;
528 
529         /* Decay the counter by 1 average sample. */
530         running_len = __this_cpu_read(running_sample_length);
531         running_len -= running_len/NR_ACCUMULATED_SAMPLES;
532         running_len += sample_len_ns;
533         __this_cpu_write(running_sample_length, running_len);
534 
535         /*
536          * Note: this will be biased artifically low until we have
537          * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
538          * from having to maintain a count.
539          */
540         avg_len = running_len/NR_ACCUMULATED_SAMPLES;
541         if (avg_len <= max_len)
542                 return;
543 
544         __report_avg = avg_len;
545         __report_allowed = max_len;
546 
547         /*
548          * Compute a throttle threshold 25% below the current duration.
549          */
550         avg_len += avg_len / 4;
551         max = (TICK_NSEC / 100) * sysctl_perf_cpu_time_max_percent;
552         if (avg_len < max)
553                 max /= (u32)avg_len;
554         else
555                 max = 1;
556 
557         WRITE_ONCE(perf_sample_allowed_ns, avg_len);
558         WRITE_ONCE(max_samples_per_tick, max);
559 
560         sysctl_perf_event_sample_rate = max * HZ;
561         perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
562 
563         if (!irq_work_queue(&perf_duration_work)) {
564                 early_printk("perf: interrupt took too long (%lld > %lld), lowering "
565                              "kernel.perf_event_max_sample_rate to %d\n",
566                              __report_avg, __report_allowed,
567                              sysctl_perf_event_sample_rate);
568         }
569 }
570 
571 static atomic64_t perf_event_id;
572 
573 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
574                               enum event_type_t event_type);
575 
576 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
577                              enum event_type_t event_type,
578                              struct task_struct *task);
579 
580 static void update_context_time(struct perf_event_context *ctx);
581 static u64 perf_event_time(struct perf_event *event);
582 
583 void __weak perf_event_print_debug(void)        { }
584 
585 static inline u64 perf_clock(void)
586 {
587         return local_clock();
588 }
589 
590 static inline u64 perf_event_clock(struct perf_event *event)
591 {
592         return event->clock();
593 }
594 
595 /*
596  * State based event timekeeping...
597  *
598  * The basic idea is to use event->state to determine which (if any) time
599  * fields to increment with the current delta. This means we only need to
600  * update timestamps when we change state or when they are explicitly requested
601  * (read).
602  *
603  * Event groups make things a little more complicated, but not terribly so. The
604  * rules for a group are that if the group leader is OFF the entire group is
605  * OFF, irrespecive of what the group member states are. This results in
606  * __perf_effective_state().
607  *
608  * A futher ramification is that when a group leader flips between OFF and
609  * !OFF, we need to update all group member times.
610  *
611  *
612  * NOTE: perf_event_time() is based on the (cgroup) context time, and thus we
613  * need to make sure the relevant context time is updated before we try and
614  * update our timestamps.
615  */
616 
617 static __always_inline enum perf_event_state
618 __perf_effective_state(struct perf_event *event)
619 {
620         struct perf_event *leader = event->group_leader;
621 
622         if (leader->state <= PERF_EVENT_STATE_OFF)
623                 return leader->state;
624 
625         return event->state;
626 }
627 
628 static __always_inline void
629 __perf_update_times(struct perf_event *event, u64 now, u64 *enabled, u64 *running)
630 {
631         enum perf_event_state state = __perf_effective_state(event);
632         u64 delta = now - event->tstamp;
633 
634         *enabled = event->total_time_enabled;
635         if (state >= PERF_EVENT_STATE_INACTIVE)
636                 *enabled += delta;
637 
638         *running = event->total_time_running;
639         if (state >= PERF_EVENT_STATE_ACTIVE)
640                 *running += delta;
641 }
642 
643 static void perf_event_update_time(struct perf_event *event)
644 {
645         u64 now = perf_event_time(event);
646 
647         __perf_update_times(event, now, &event->total_time_enabled,
648                                         &event->total_time_running);
649         event->tstamp = now;
650 }
651 
652 static void perf_event_update_sibling_time(struct perf_event *leader)
653 {
654         struct perf_event *sibling;
655 
656         for_each_sibling_event(sibling, leader)
657                 perf_event_update_time(sibling);
658 }
659 
660 static void
661 perf_event_set_state(struct perf_event *event, enum perf_event_state state)
662 {
663         if (event->state == state)
664                 return;
665 
666         perf_event_update_time(event);
667         /*
668          * If a group leader gets enabled/disabled all its siblings
669          * are affected too.
670          */
671         if ((event->state < 0) ^ (state < 0))
672                 perf_event_update_sibling_time(event);
673 
674         WRITE_ONCE(event->state, state);
675 }
676 
677 /*
678  * UP store-release, load-acquire
679  */
680 
681 #define __store_release(ptr, val)                                       \
682 do {                                                                    \
683         barrier();                                                      \
684         WRITE_ONCE(*(ptr), (val));                                      \
685 } while (0)
686 
687 #define __load_acquire(ptr)                                             \
688 ({                                                                      \
689         __unqual_scalar_typeof(*(ptr)) ___p = READ_ONCE(*(ptr));        \
690         barrier();                                                      \
691         ___p;                                                           \
692 })
693 
694 #ifdef CONFIG_CGROUP_PERF
695 
696 static inline bool
697 perf_cgroup_match(struct perf_event *event)
698 {
699         struct perf_event_context *ctx = event->ctx;
700         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
701 
702         /* @event doesn't care about cgroup */
703         if (!event->cgrp)
704                 return true;
705 
706         /* wants specific cgroup scope but @cpuctx isn't associated with any */
707         if (!cpuctx->cgrp)
708                 return false;
709 
710         /*
711          * Cgroup scoping is recursive.  An event enabled for a cgroup is
712          * also enabled for all its descendant cgroups.  If @cpuctx's
713          * cgroup is a descendant of @event's (the test covers identity
714          * case), it's a match.
715          */
716         return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
717                                     event->cgrp->css.cgroup);
718 }
719 
720 static inline void perf_detach_cgroup(struct perf_event *event)
721 {
722         css_put(&event->cgrp->css);
723         event->cgrp = NULL;
724 }
725 
726 static inline int is_cgroup_event(struct perf_event *event)
727 {
728         return event->cgrp != NULL;
729 }
730 
731 static inline u64 perf_cgroup_event_time(struct perf_event *event)
732 {
733         struct perf_cgroup_info *t;
734 
735         t = per_cpu_ptr(event->cgrp->info, event->cpu);
736         return t->time;
737 }
738 
739 static inline u64 perf_cgroup_event_time_now(struct perf_event *event, u64 now)
740 {
741         struct perf_cgroup_info *t;
742 
743         t = per_cpu_ptr(event->cgrp->info, event->cpu);
744         if (!__load_acquire(&t->active))
745                 return t->time;
746         now += READ_ONCE(t->timeoffset);
747         return now;
748 }
749 
750 static inline void __update_cgrp_time(struct perf_cgroup_info *info, u64 now, bool adv)
751 {
752         if (adv)
753                 info->time += now - info->timestamp;
754         info->timestamp = now;
755         /*
756          * see update_context_time()
757          */
758         WRITE_ONCE(info->timeoffset, info->time - info->timestamp);
759 }
760 
761 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx, bool final)
762 {
763         struct perf_cgroup *cgrp = cpuctx->cgrp;
764         struct cgroup_subsys_state *css;
765         struct perf_cgroup_info *info;
766 
767         if (cgrp) {
768                 u64 now = perf_clock();
769 
770                 for (css = &cgrp->css; css; css = css->parent) {
771                         cgrp = container_of(css, struct perf_cgroup, css);
772                         info = this_cpu_ptr(cgrp->info);
773 
774                         __update_cgrp_time(info, now, true);
775                         if (final)
776                                 __store_release(&info->active, 0);
777                 }
778         }
779 }
780 
781 static inline void update_cgrp_time_from_event(struct perf_event *event)
782 {
783         struct perf_cgroup_info *info;
784         struct perf_cgroup *cgrp;
785 
786         /*
787          * ensure we access cgroup data only when needed and
788          * when we know the cgroup is pinned (css_get)
789          */
790         if (!is_cgroup_event(event))
791                 return;
792 
793         cgrp = perf_cgroup_from_task(current, event->ctx);
794         /*
795          * Do not update time when cgroup is not active
796          */
797         if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) {
798                 info = this_cpu_ptr(event->cgrp->info);
799                 __update_cgrp_time(info, perf_clock(), true);
800         }
801 }
802 
803 static inline void
804 perf_cgroup_set_timestamp(struct task_struct *task,
805                           struct perf_event_context *ctx)
806 {
807         struct perf_cgroup *cgrp;
808         struct perf_cgroup_info *info;
809         struct cgroup_subsys_state *css;
810 
811         /*
812          * ctx->lock held by caller
813          * ensure we do not access cgroup data
814          * unless we have the cgroup pinned (css_get)
815          */
816         if (!task || !ctx->nr_cgroups)
817                 return;
818 
819         cgrp = perf_cgroup_from_task(task, ctx);
820 
821         for (css = &cgrp->css; css; css = css->parent) {
822                 cgrp = container_of(css, struct perf_cgroup, css);
823                 info = this_cpu_ptr(cgrp->info);
824                 __update_cgrp_time(info, ctx->timestamp, false);
825                 __store_release(&info->active, 1);
826         }
827 }
828 
829 static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
830 
831 #define PERF_CGROUP_SWOUT       0x1 /* cgroup switch out every event */
832 #define PERF_CGROUP_SWIN        0x2 /* cgroup switch in events based on task */
833 
834 /*
835  * reschedule events based on the cgroup constraint of task.
836  *
837  * mode SWOUT : schedule out everything
838  * mode SWIN : schedule in based on cgroup for next
839  */
840 static void perf_cgroup_switch(struct task_struct *task, int mode)
841 {
842         struct perf_cpu_context *cpuctx, *tmp;
843         struct list_head *list;
844         unsigned long flags;
845 
846         /*
847          * Disable interrupts and preemption to avoid this CPU's
848          * cgrp_cpuctx_entry to change under us.
849          */
850         local_irq_save(flags);
851 
852         list = this_cpu_ptr(&cgrp_cpuctx_list);
853         list_for_each_entry_safe(cpuctx, tmp, list, cgrp_cpuctx_entry) {
854                 WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
855 
856                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
857                 perf_pmu_disable(cpuctx->ctx.pmu);
858 
859                 if (mode & PERF_CGROUP_SWOUT) {
860                         cpu_ctx_sched_out(cpuctx, EVENT_ALL);
861                         /*
862                          * must not be done before ctxswout due
863                          * to event_filter_match() in event_sched_out()
864                          */
865                         cpuctx->cgrp = NULL;
866                 }
867 
868                 if (mode & PERF_CGROUP_SWIN) {
869                         WARN_ON_ONCE(cpuctx->cgrp);
870                         /*
871                          * set cgrp before ctxsw in to allow
872                          * event_filter_match() to not have to pass
873                          * task around
874                          * we pass the cpuctx->ctx to perf_cgroup_from_task()
875                          * because cgorup events are only per-cpu
876                          */
877                         cpuctx->cgrp = perf_cgroup_from_task(task,
878                                                              &cpuctx->ctx);
879                         cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
880                 }
881                 perf_pmu_enable(cpuctx->ctx.pmu);
882                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
883         }
884 
885         local_irq_restore(flags);
886 }
887 
888 static inline void perf_cgroup_sched_out(struct task_struct *task,
889                                          struct task_struct *next)
890 {
891         struct perf_cgroup *cgrp1;
892         struct perf_cgroup *cgrp2 = NULL;
893 
894         rcu_read_lock();
895         /*
896          * we come here when we know perf_cgroup_events > 0
897          * we do not need to pass the ctx here because we know
898          * we are holding the rcu lock
899          */
900         cgrp1 = perf_cgroup_from_task(task, NULL);
901         cgrp2 = perf_cgroup_from_task(next, NULL);
902 
903         /*
904          * only schedule out current cgroup events if we know
905          * that we are switching to a different cgroup. Otherwise,
906          * do no touch the cgroup events.
907          */
908         if (cgrp1 != cgrp2)
909                 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
910 
911         rcu_read_unlock();
912 }
913 
914 static inline void perf_cgroup_sched_in(struct task_struct *prev,
915                                         struct task_struct *task)
916 {
917         struct perf_cgroup *cgrp1;
918         struct perf_cgroup *cgrp2 = NULL;
919 
920         rcu_read_lock();
921         /*
922          * we come here when we know perf_cgroup_events > 0
923          * we do not need to pass the ctx here because we know
924          * we are holding the rcu lock
925          */
926         cgrp1 = perf_cgroup_from_task(task, NULL);
927         cgrp2 = perf_cgroup_from_task(prev, NULL);
928 
929         /*
930          * only need to schedule in cgroup events if we are changing
931          * cgroup during ctxsw. Cgroup events were not scheduled
932          * out of ctxsw out if that was not the case.
933          */
934         if (cgrp1 != cgrp2)
935                 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
936 
937         rcu_read_unlock();
938 }
939 
940 static int perf_cgroup_ensure_storage(struct perf_event *event,
941                                 struct cgroup_subsys_state *css)
942 {
943         struct perf_cpu_context *cpuctx;
944         struct perf_event **storage;
945         int cpu, heap_size, ret = 0;
946 
947         /*
948          * Allow storage to have sufficent space for an iterator for each
949          * possibly nested cgroup plus an iterator for events with no cgroup.
950          */
951         for (heap_size = 1; css; css = css->parent)
952                 heap_size++;
953 
954         for_each_possible_cpu(cpu) {
955                 cpuctx = per_cpu_ptr(event->pmu->pmu_cpu_context, cpu);
956                 if (heap_size <= cpuctx->heap_size)
957                         continue;
958 
959                 storage = kmalloc_node(heap_size * sizeof(struct perf_event *),
960                                        GFP_KERNEL, cpu_to_node(cpu));
961                 if (!storage) {
962                         ret = -ENOMEM;
963                         break;
964                 }
965 
966                 raw_spin_lock_irq(&cpuctx->ctx.lock);
967                 if (cpuctx->heap_size < heap_size) {
968                         swap(cpuctx->heap, storage);
969                         if (storage == cpuctx->heap_default)
970                                 storage = NULL;
971                         cpuctx->heap_size = heap_size;
972                 }
973                 raw_spin_unlock_irq(&cpuctx->ctx.lock);
974 
975                 kfree(storage);
976         }
977 
978         return ret;
979 }
980 
981 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
982                                       struct perf_event_attr *attr,
983                                       struct perf_event *group_leader)
984 {
985         struct perf_cgroup *cgrp;
986         struct cgroup_subsys_state *css;
987         struct fd f = fdget(fd);
988         int ret = 0;
989 
990         if (!f.file)
991                 return -EBADF;
992 
993         css = css_tryget_online_from_dir(f.file->f_path.dentry,
994                                          &perf_event_cgrp_subsys);
995         if (IS_ERR(css)) {
996                 ret = PTR_ERR(css);
997                 goto out;
998         }
999 
1000         ret = perf_cgroup_ensure_storage(event, css);
1001         if (ret)
1002                 goto out;
1003 
1004         cgrp = container_of(css, struct perf_cgroup, css);
1005         event->cgrp = cgrp;
1006 
1007         /*
1008          * all events in a group must monitor
1009          * the same cgroup because a task belongs
1010          * to only one perf cgroup at a time
1011          */
1012         if (group_leader && group_leader->cgrp != cgrp) {
1013                 perf_detach_cgroup(event);
1014                 ret = -EINVAL;
1015         }
1016 out:
1017         fdput(f);
1018         return ret;
1019 }
1020 
1021 static inline void
1022 perf_cgroup_event_enable(struct perf_event *event, struct perf_event_context *ctx)
1023 {
1024         struct perf_cpu_context *cpuctx;
1025 
1026         if (!is_cgroup_event(event))
1027                 return;
1028 
1029         /*
1030          * Because cgroup events are always per-cpu events,
1031          * @ctx == &cpuctx->ctx.
1032          */
1033         cpuctx = container_of(ctx, struct perf_cpu_context, ctx);
1034 
1035         /*
1036          * Since setting cpuctx->cgrp is conditional on the current @cgrp
1037          * matching the event's cgroup, we must do this for every new event,
1038          * because if the first would mismatch, the second would not try again
1039          * and we would leave cpuctx->cgrp unset.
1040          */
1041         if (ctx->is_active && !cpuctx->cgrp) {
1042                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
1043 
1044                 if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
1045                         cpuctx->cgrp = cgrp;
1046         }
1047 
1048         if (ctx->nr_cgroups++)
1049                 return;
1050 
1051         list_add(&cpuctx->cgrp_cpuctx_entry,
1052                         per_cpu_ptr(&cgrp_cpuctx_list, event->cpu));
1053 }
1054 
1055 static inline void
1056 perf_cgroup_event_disable(struct perf_event *event, struct perf_event_context *ctx)
1057 {
1058         struct perf_cpu_context *cpuctx;
1059 
1060         if (!is_cgroup_event(event))
1061                 return;
1062 
1063         /*
1064          * Because cgroup events are always per-cpu events,
1065          * @ctx == &cpuctx->ctx.
1066          */
1067         cpuctx = container_of(ctx, struct perf_cpu_context, ctx);
1068 
1069         if (--ctx->nr_cgroups)
1070                 return;
1071 
1072         if (ctx->is_active && cpuctx->cgrp)
1073                 cpuctx->cgrp = NULL;
1074 
1075         list_del(&cpuctx->cgrp_cpuctx_entry);
1076 }
1077 
1078 #else /* !CONFIG_CGROUP_PERF */
1079 
1080 static inline bool
1081 perf_cgroup_match(struct perf_event *event)
1082 {
1083         return true;
1084 }
1085 
1086 static inline void perf_detach_cgroup(struct perf_event *event)
1087 {}
1088 
1089 static inline int is_cgroup_event(struct perf_event *event)
1090 {
1091         return 0;
1092 }
1093 
1094 static inline void update_cgrp_time_from_event(struct perf_event *event)
1095 {
1096 }
1097 
1098 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx,
1099                                                 bool final)
1100 {
1101 }
1102 
1103 static inline void perf_cgroup_sched_out(struct task_struct *task,
1104                                          struct task_struct *next)
1105 {
1106 }
1107 
1108 static inline void perf_cgroup_sched_in(struct task_struct *prev,
1109                                         struct task_struct *task)
1110 {
1111 }
1112 
1113 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
1114                                       struct perf_event_attr *attr,
1115                                       struct perf_event *group_leader)
1116 {
1117         return -EINVAL;
1118 }
1119 
1120 static inline void
1121 perf_cgroup_set_timestamp(struct task_struct *task,
1122                           struct perf_event_context *ctx)
1123 {
1124 }
1125 
1126 static inline void
1127 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
1128 {
1129 }
1130 
1131 static inline u64 perf_cgroup_event_time(struct perf_event *event)
1132 {
1133         return 0;
1134 }
1135 
1136 static inline u64 perf_cgroup_event_time_now(struct perf_event *event, u64 now)
1137 {
1138         return 0;
1139 }
1140 
1141 static inline void
1142 perf_cgroup_event_enable(struct perf_event *event, struct perf_event_context *ctx)
1143 {
1144 }
1145 
1146 static inline void
1147 perf_cgroup_event_disable(struct perf_event *event, struct perf_event_context *ctx)
1148 {
1149 }
1150 #endif
1151 
1152 /*
1153  * set default to be dependent on timer tick just
1154  * like original code
1155  */
1156 #define PERF_CPU_HRTIMER (1000 / HZ)
1157 /*
1158  * function must be called with interrupts disabled
1159  */
1160 static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr)
1161 {
1162         struct perf_cpu_context *cpuctx;
1163         bool rotations;
1164 
1165         lockdep_assert_irqs_disabled();
1166 
1167         cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
1168         rotations = perf_rotate_context(cpuctx);
1169 
1170         raw_spin_lock(&cpuctx->hrtimer_lock);
1171         if (rotations)
1172                 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
1173         else
1174                 cpuctx->hrtimer_active = 0;
1175         raw_spin_unlock(&cpuctx->hrtimer_lock);
1176 
1177         return rotations ? HRTIMER_RESTART : HRTIMER_NORESTART;
1178 }
1179 
1180 static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
1181 {
1182         struct hrtimer *timer = &cpuctx->hrtimer;
1183         struct pmu *pmu = cpuctx->ctx.pmu;
1184         u64 interval;
1185 
1186         /* no multiplexing needed for SW PMU */
1187         if (pmu->task_ctx_nr == perf_sw_context)
1188                 return;
1189 
1190         /*
1191          * check default is sane, if not set then force to
1192          * default interval (1/tick)
1193          */
1194         interval = pmu->hrtimer_interval_ms;
1195         if (interval < 1)
1196                 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
1197 
1198         cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval);
1199 
1200         raw_spin_lock_init(&cpuctx->hrtimer_lock);
1201         hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED_HARD);
1202         timer->function = perf_mux_hrtimer_handler;
1203 }
1204 
1205 static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
1206 {
1207         struct hrtimer *timer = &cpuctx->hrtimer;
1208         struct pmu *pmu = cpuctx->ctx.pmu;
1209         unsigned long flags;
1210 
1211         /* not for SW PMU */
1212         if (pmu->task_ctx_nr == perf_sw_context)
1213                 return 0;
1214 
1215         raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags);
1216         if (!cpuctx->hrtimer_active) {
1217                 cpuctx->hrtimer_active = 1;
1218                 hrtimer_forward_now(timer, cpuctx->hrtimer_interval);
1219                 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED_HARD);
1220         }
1221         raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags);
1222 
1223         return 0;
1224 }
1225 
1226 void perf_pmu_disable(struct pmu *pmu)
1227 {
1228         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1229         if (!(*count)++)
1230                 pmu->pmu_disable(pmu);
1231 }
1232 
1233 void perf_pmu_enable(struct pmu *pmu)
1234 {
1235         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1236         if (!--(*count))
1237                 pmu->pmu_enable(pmu);
1238 }
1239 
1240 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
1241 
1242 /*
1243  * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
1244  * perf_event_task_tick() are fully serialized because they're strictly cpu
1245  * affine and perf_event_ctx{activate,deactivate} are called with IRQs
1246  * disabled, while perf_event_task_tick is called from IRQ context.
1247  */
1248 static void perf_event_ctx_activate(struct perf_event_context *ctx)
1249 {
1250         struct list_head *head = this_cpu_ptr(&active_ctx_list);
1251 
1252         lockdep_assert_irqs_disabled();
1253 
1254         WARN_ON(!list_empty(&ctx->active_ctx_list));
1255 
1256         list_add(&ctx->active_ctx_list, head);
1257 }
1258 
1259 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
1260 {
1261         lockdep_assert_irqs_disabled();
1262 
1263         WARN_ON(list_empty(&ctx->active_ctx_list));
1264 
1265         list_del_init(&ctx->active_ctx_list);
1266 }
1267 
1268 static void get_ctx(struct perf_event_context *ctx)
1269 {
1270         refcount_inc(&ctx->refcount);
1271 }
1272 
1273 static void *alloc_task_ctx_data(struct pmu *pmu)
1274 {
1275         if (pmu->task_ctx_cache)
1276                 return kmem_cache_zalloc(pmu->task_ctx_cache, GFP_KERNEL);
1277 
1278         return NULL;
1279 }
1280 
1281 static void free_task_ctx_data(struct pmu *pmu, void *task_ctx_data)
1282 {
1283         if (pmu->task_ctx_cache && task_ctx_data)
1284                 kmem_cache_free(pmu->task_ctx_cache, task_ctx_data);
1285 }
1286 
1287 static void free_ctx(struct rcu_head *head)
1288 {
1289         struct perf_event_context *ctx;
1290 
1291         ctx = container_of(head, struct perf_event_context, rcu_head);
1292         free_task_ctx_data(ctx->pmu, ctx->task_ctx_data);
1293         kfree(ctx);
1294 }
1295 
1296 static void put_ctx(struct perf_event_context *ctx)
1297 {
1298         if (refcount_dec_and_test(&ctx->refcount)) {
1299                 if (ctx->parent_ctx)
1300                         put_ctx(ctx->parent_ctx);
1301                 if (ctx->task && ctx->task != TASK_TOMBSTONE)
1302                         put_task_struct(ctx->task);
1303                 call_rcu(&ctx->rcu_head, free_ctx);
1304         }
1305 }
1306 
1307 /*
1308  * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
1309  * perf_pmu_migrate_context() we need some magic.
1310  *
1311  * Those places that change perf_event::ctx will hold both
1312  * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
1313  *
1314  * Lock ordering is by mutex address. There are two other sites where
1315  * perf_event_context::mutex nests and those are:
1316  *
1317  *  - perf_event_exit_task_context()    [ child , 0 ]
1318  *      perf_event_exit_event()
1319  *        put_event()                   [ parent, 1 ]
1320  *
1321  *  - perf_event_init_context()         [ parent, 0 ]
1322  *      inherit_task_group()
1323  *        inherit_group()
1324  *          inherit_event()
1325  *            perf_event_alloc()
1326  *              perf_init_event()
1327  *                perf_try_init_event() [ child , 1 ]
1328  *
1329  * While it appears there is an obvious deadlock here -- the parent and child
1330  * nesting levels are inverted between the two. This is in fact safe because
1331  * life-time rules separate them. That is an exiting task cannot fork, and a
1332  * spawning task cannot (yet) exit.
1333  *
1334  * But remember that these are parent<->child context relations, and
1335  * migration does not affect children, therefore these two orderings should not
1336  * interact.
1337  *
1338  * The change in perf_event::ctx does not affect children (as claimed above)
1339  * because the sys_perf_event_open() case will install a new event and break
1340  * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
1341  * concerned with cpuctx and that doesn't have children.
1342  *
1343  * The places that change perf_event::ctx will issue:
1344  *
1345  *   perf_remove_from_context();
1346  *   synchronize_rcu();
1347  *   perf_install_in_context();
1348  *
1349  * to affect the change. The remove_from_context() + synchronize_rcu() should
1350  * quiesce the event, after which we can install it in the new location. This
1351  * means that only external vectors (perf_fops, prctl) can perturb the event
1352  * while in transit. Therefore all such accessors should also acquire
1353  * perf_event_context::mutex to serialize against this.
1354  *
1355  * However; because event->ctx can change while we're waiting to acquire
1356  * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
1357  * function.
1358  *
1359  * Lock order:
1360  *    exec_update_lock
1361  *      task_struct::perf_event_mutex
1362  *        perf_event_context::mutex
1363  *          perf_event::child_mutex;
1364  *            perf_event_context::lock
1365  *          perf_event::mmap_mutex
1366  *          mmap_lock
1367  *            perf_addr_filters_head::lock
1368  *
1369  *    cpu_hotplug_lock
1370  *      pmus_lock
1371  *        cpuctx->mutex / perf_event_context::mutex
1372  */
1373 static struct perf_event_context *
1374 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
1375 {
1376         struct perf_event_context *ctx;
1377 
1378 again:
1379         rcu_read_lock();
1380         ctx = READ_ONCE(event->ctx);
1381         if (!refcount_inc_not_zero(&ctx->refcount)) {
1382                 rcu_read_unlock();
1383                 goto again;
1384         }
1385         rcu_read_unlock();
1386 
1387         mutex_lock_nested(&ctx->mutex, nesting);
1388         if (event->ctx != ctx) {
1389                 mutex_unlock(&ctx->mutex);
1390                 put_ctx(ctx);
1391                 goto again;
1392         }
1393 
1394         return ctx;
1395 }
1396 
1397 static inline struct perf_event_context *
1398 perf_event_ctx_lock(struct perf_event *event)
1399 {
1400         return perf_event_ctx_lock_nested(event, 0);
1401 }
1402 
1403 static void perf_event_ctx_unlock(struct perf_event *event,
1404                                   struct perf_event_context *ctx)
1405 {
1406         mutex_unlock(&ctx->mutex);
1407         put_ctx(ctx);
1408 }
1409 
1410 /*
1411  * This must be done under the ctx->lock, such as to serialize against
1412  * context_equiv(), therefore we cannot call put_ctx() since that might end up
1413  * calling scheduler related locks and ctx->lock nests inside those.
1414  */
1415 static __must_check struct perf_event_context *
1416 unclone_ctx(struct perf_event_context *ctx)
1417 {
1418         struct perf_event_context *parent_ctx = ctx->parent_ctx;
1419 
1420         lockdep_assert_held(&ctx->lock);
1421 
1422         if (parent_ctx)
1423                 ctx->parent_ctx = NULL;
1424         ctx->generation++;
1425 
1426         return parent_ctx;
1427 }
1428 
1429 static u32 perf_event_pid_type(struct perf_event *event, struct task_struct *p,
1430                                 enum pid_type type)
1431 {
1432         u32 nr;
1433         /*
1434          * only top level events have the pid namespace they were created in
1435          */
1436         if (event->parent)
1437                 event = event->parent;
1438 
1439         nr = __task_pid_nr_ns(p, type, event->ns);
1440         /* avoid -1 if it is idle thread or runs in another ns */
1441         if (!nr && !pid_alive(p))
1442                 nr = -1;
1443         return nr;
1444 }
1445 
1446 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1447 {
1448         return perf_event_pid_type(event, p, PIDTYPE_TGID);
1449 }
1450 
1451 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1452 {
1453         return perf_event_pid_type(event, p, PIDTYPE_PID);
1454 }
1455 
1456 /*
1457  * If we inherit events we want to return the parent event id
1458  * to userspace.
1459  */
1460 static u64 primary_event_id(struct perf_event *event)
1461 {
1462         u64 id = event->id;
1463 
1464         if (event->parent)
1465                 id = event->parent->id;
1466 
1467         return id;
1468 }
1469 
1470 /*
1471  * Get the perf_event_context for a task and lock it.
1472  *
1473  * This has to cope with the fact that until it is locked,
1474  * the context could get moved to another task.
1475  */
1476 static struct perf_event_context *
1477 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1478 {
1479         struct perf_event_context *ctx;
1480 
1481 retry:
1482         /*
1483          * One of the few rules of preemptible RCU is that one cannot do
1484          * rcu_read_unlock() while holding a scheduler (or nested) lock when
1485          * part of the read side critical section was irqs-enabled -- see
1486          * rcu_read_unlock_special().
1487          *
1488          * Since ctx->lock nests under rq->lock we must ensure the entire read
1489          * side critical section has interrupts disabled.
1490          */
1491         local_irq_save(*flags);
1492         rcu_read_lock();
1493         ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1494         if (ctx) {
1495                 /*
1496                  * If this context is a clone of another, it might
1497                  * get swapped for another underneath us by
1498                  * perf_event_task_sched_out, though the
1499                  * rcu_read_lock() protects us from any context
1500                  * getting freed.  Lock the context and check if it
1501                  * got swapped before we could get the lock, and retry
1502                  * if so.  If we locked the right context, then it
1503                  * can't get swapped on us any more.
1504                  */
1505                 raw_spin_lock(&ctx->lock);
1506                 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1507                         raw_spin_unlock(&ctx->lock);
1508                         rcu_read_unlock();
1509                         local_irq_restore(*flags);
1510                         goto retry;
1511                 }
1512 
1513                 if (ctx->task == TASK_TOMBSTONE ||
1514                     !refcount_inc_not_zero(&ctx->refcount)) {
1515                         raw_spin_unlock(&ctx->lock);
1516                         ctx = NULL;
1517                 } else {
1518                         WARN_ON_ONCE(ctx->task != task);
1519                 }
1520         }
1521         rcu_read_unlock();
1522         if (!ctx)
1523                 local_irq_restore(*flags);
1524         return ctx;
1525 }
1526 
1527 /*
1528  * Get the context for a task and increment its pin_count so it
1529  * can't get swapped to another task.  This also increments its
1530  * reference count so that the context can't get freed.
1531  */
1532 static struct perf_event_context *
1533 perf_pin_task_context(struct task_struct *task, int ctxn)
1534 {
1535         struct perf_event_context *ctx;
1536         unsigned long flags;
1537 
1538         ctx = perf_lock_task_context(task, ctxn, &flags);
1539         if (ctx) {
1540                 ++ctx->pin_count;
1541                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1542         }
1543         return ctx;
1544 }
1545 
1546 static void perf_unpin_context(struct perf_event_context *ctx)
1547 {
1548         unsigned long flags;
1549 
1550         raw_spin_lock_irqsave(&ctx->lock, flags);
1551         --ctx->pin_count;
1552         raw_spin_unlock_irqrestore(&ctx->lock, flags);
1553 }
1554 
1555 /*
1556  * Update the record of the current time in a context.
1557  */
1558 static void __update_context_time(struct perf_event_context *ctx, bool adv)
1559 {
1560         u64 now = perf_clock();
1561 
1562         if (adv)
1563                 ctx->time += now - ctx->timestamp;
1564         ctx->timestamp = now;
1565 
1566         /*
1567          * The above: time' = time + (now - timestamp), can be re-arranged
1568          * into: time` = now + (time - timestamp), which gives a single value
1569          * offset to compute future time without locks on.
1570          *
1571          * See perf_event_time_now(), which can be used from NMI context where
1572          * it's (obviously) not possible to acquire ctx->lock in order to read
1573          * both the above values in a consistent manner.
1574          */
1575         WRITE_ONCE(ctx->timeoffset, ctx->time - ctx->timestamp);
1576 }
1577 
1578 static void update_context_time(struct perf_event_context *ctx)
1579 {
1580         __update_context_time(ctx, true);
1581 }
1582 
1583 static u64 perf_event_time(struct perf_event *event)
1584 {
1585         struct perf_event_context *ctx = event->ctx;
1586 
1587         if (unlikely(!ctx))
1588                 return 0;
1589 
1590         if (is_cgroup_event(event))
1591                 return perf_cgroup_event_time(event);
1592 
1593         return ctx->time;
1594 }
1595 
1596 static u64 perf_event_time_now(struct perf_event *event, u64 now)
1597 {
1598         struct perf_event_context *ctx = event->ctx;
1599 
1600         if (unlikely(!ctx))
1601                 return 0;
1602 
1603         if (is_cgroup_event(event))
1604                 return perf_cgroup_event_time_now(event, now);
1605 
1606         if (!(__load_acquire(&ctx->is_active) & EVENT_TIME))
1607                 return ctx->time;
1608 
1609         now += READ_ONCE(ctx->timeoffset);
1610         return now;
1611 }
1612 
1613 static enum event_type_t get_event_type(struct perf_event *event)
1614 {
1615         struct perf_event_context *ctx = event->ctx;
1616         enum event_type_t event_type;
1617 
1618         lockdep_assert_held(&ctx->lock);
1619 
1620         /*
1621          * It's 'group type', really, because if our group leader is
1622          * pinned, so are we.
1623          */
1624         if (event->group_leader != event)
1625                 event = event->group_leader;
1626 
1627         event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE;
1628         if (!ctx->task)
1629                 event_type |= EVENT_CPU;
1630 
1631         return event_type;
1632 }
1633 
1634 /*
1635  * Helper function to initialize event group nodes.
1636  */
1637 static void init_event_group(struct perf_event *event)
1638 {
1639         RB_CLEAR_NODE(&event->group_node);
1640         event->group_index = 0;
1641 }
1642 
1643 /*
1644  * Extract pinned or flexible groups from the context
1645  * based on event attrs bits.
1646  */
1647 static struct perf_event_groups *
1648 get_event_groups(struct perf_event *event, struct perf_event_context *ctx)
1649 {
1650         if (event->attr.pinned)
1651                 return &ctx->pinned_groups;
1652         else
1653                 return &ctx->flexible_groups;
1654 }
1655 
1656 /*
1657  * Helper function to initializes perf_event_group trees.
1658  */
1659 static void perf_event_groups_init(struct perf_event_groups *groups)
1660 {
1661         groups->tree = RB_ROOT;
1662         groups->index = 0;
1663 }
1664 
1665 static inline struct cgroup *event_cgroup(const struct perf_event *event)
1666 {
1667         struct cgroup *cgroup = NULL;
1668 
1669 #ifdef CONFIG_CGROUP_PERF
1670         if (event->cgrp)
1671                 cgroup = event->cgrp->css.cgroup;
1672 #endif
1673 
1674         return cgroup;
1675 }
1676 
1677 /*
1678  * Compare function for event groups;
1679  *
1680  * Implements complex key that first sorts by CPU and then by virtual index
1681  * which provides ordering when rotating groups for the same CPU.
1682  */
1683 static __always_inline int
1684 perf_event_groups_cmp(const int left_cpu, const struct cgroup *left_cgroup,
1685                       const u64 left_group_index, const struct perf_event *right)
1686 {
1687         if (left_cpu < right->cpu)
1688                 return -1;
1689         if (left_cpu > right->cpu)
1690                 return 1;
1691 
1692 #ifdef CONFIG_CGROUP_PERF
1693         {
1694                 const struct cgroup *right_cgroup = event_cgroup(right);
1695 
1696                 if (left_cgroup != right_cgroup) {
1697                         if (!left_cgroup) {
1698                                 /*
1699                                  * Left has no cgroup but right does, no
1700                                  * cgroups come first.
1701                                  */
1702                                 return -1;
1703                         }
1704                         if (!right_cgroup) {
1705                                 /*
1706                                  * Right has no cgroup but left does, no
1707                                  * cgroups come first.
1708                                  */
1709                                 return 1;
1710                         }
1711                         /* Two dissimilar cgroups, order by id. */
1712                         if (cgroup_id(left_cgroup) < cgroup_id(right_cgroup))
1713                                 return -1;
1714 
1715                         return 1;
1716                 }
1717         }
1718 #endif
1719 
1720         if (left_group_index < right->group_index)
1721                 return -1;
1722         if (left_group_index > right->group_index)
1723                 return 1;
1724 
1725         return 0;
1726 }
1727 
1728 #define __node_2_pe(node) \
1729         rb_entry((node), struct perf_event, group_node)
1730 
1731 static inline bool __group_less(struct rb_node *a, const struct rb_node *b)
1732 {
1733         struct perf_event *e = __node_2_pe(a);
1734         return perf_event_groups_cmp(e->cpu, event_cgroup(e), e->group_index,
1735                                      __node_2_pe(b)) < 0;
1736 }
1737 
1738 struct __group_key {
1739         int cpu;
1740         struct cgroup *cgroup;
1741 };
1742 
1743 static inline int __group_cmp(const void *key, const struct rb_node *node)
1744 {
1745         const struct __group_key *a = key;
1746         const struct perf_event *b = __node_2_pe(node);
1747 
1748         /* partial/subtree match: @cpu, @cgroup; ignore: @group_index */
1749         return perf_event_groups_cmp(a->cpu, a->cgroup, b->group_index, b);
1750 }
1751 
1752 /*
1753  * Insert @event into @groups' tree; using {@event->cpu, ++@groups->index} for
1754  * key (see perf_event_groups_less). This places it last inside the CPU
1755  * subtree.
1756  */
1757 static void
1758 perf_event_groups_insert(struct perf_event_groups *groups,
1759                          struct perf_event *event)
1760 {
1761         event->group_index = ++groups->index;
1762 
1763         rb_add(&event->group_node, &groups->tree, __group_less);
1764 }
1765 
1766 /*
1767  * Helper function to insert event into the pinned or flexible groups.
1768  */
1769 static void
1770 add_event_to_groups(struct perf_event *event, struct perf_event_context *ctx)
1771 {
1772         struct perf_event_groups *groups;
1773 
1774         groups = get_event_groups(event, ctx);
1775         perf_event_groups_insert(groups, event);
1776 }
1777 
1778 /*
1779  * Delete a group from a tree.
1780  */
1781 static void
1782 perf_event_groups_delete(struct perf_event_groups *groups,
1783                          struct perf_event *event)
1784 {
1785         WARN_ON_ONCE(RB_EMPTY_NODE(&event->group_node) ||
1786                      RB_EMPTY_ROOT(&groups->tree));
1787 
1788         rb_erase(&event->group_node, &groups->tree);
1789         init_event_group(event);
1790 }
1791 
1792 /*
1793  * Helper function to delete event from its groups.
1794  */
1795 static void
1796 del_event_from_groups(struct perf_event *event, struct perf_event_context *ctx)
1797 {
1798         struct perf_event_groups *groups;
1799 
1800         groups = get_event_groups(event, ctx);
1801         perf_event_groups_delete(groups, event);
1802 }
1803 
1804 /*
1805  * Get the leftmost event in the cpu/cgroup subtree.
1806  */
1807 static struct perf_event *
1808 perf_event_groups_first(struct perf_event_groups *groups, int cpu,
1809                         struct cgroup *cgrp)
1810 {
1811         struct __group_key key = {
1812                 .cpu = cpu,
1813                 .cgroup = cgrp,
1814         };
1815         struct rb_node *node;
1816 
1817         node = rb_find_first(&key, &groups->tree, __group_cmp);
1818         if (node)
1819                 return __node_2_pe(node);
1820 
1821         return NULL;
1822 }
1823 
1824 /*
1825  * Like rb_entry_next_safe() for the @cpu subtree.
1826  */
1827 static struct perf_event *
1828 perf_event_groups_next(struct perf_event *event)
1829 {
1830         struct __group_key key = {
1831                 .cpu = event->cpu,
1832                 .cgroup = event_cgroup(event),
1833         };
1834         struct rb_node *next;
1835 
1836         next = rb_next_match(&key, &event->group_node, __group_cmp);
1837         if (next)
1838                 return __node_2_pe(next);
1839 
1840         return NULL;
1841 }
1842 
1843 /*
1844  * Iterate through the whole groups tree.
1845  */
1846 #define perf_event_groups_for_each(event, groups)                       \
1847         for (event = rb_entry_safe(rb_first(&((groups)->tree)),         \
1848                                 typeof(*event), group_node); event;     \
1849                 event = rb_entry_safe(rb_next(&event->group_node),      \
1850                                 typeof(*event), group_node))
1851 
1852 /*
1853  * Add an event from the lists for its context.
1854  * Must be called with ctx->mutex and ctx->lock held.
1855  */
1856 static void
1857 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1858 {
1859         lockdep_assert_held(&ctx->lock);
1860 
1861         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1862         event->attach_state |= PERF_ATTACH_CONTEXT;
1863 
1864         event->tstamp = perf_event_time(event);
1865 
1866         /*
1867          * If we're a stand alone event or group leader, we go to the context
1868          * list, group events are kept attached to the group so that
1869          * perf_group_detach can, at all times, locate all siblings.
1870          */
1871         if (event->group_leader == event) {
1872                 event->group_caps = event->event_caps;
1873                 add_event_to_groups(event, ctx);
1874         }
1875 
1876         list_add_rcu(&event->event_entry, &ctx->event_list);
1877         ctx->nr_events++;
1878         if (event->attr.inherit_stat)
1879                 ctx->nr_stat++;
1880 
1881         if (event->state > PERF_EVENT_STATE_OFF)
1882                 perf_cgroup_event_enable(event, ctx);
1883 
1884         ctx->generation++;
1885 }
1886 
1887 /*
1888  * Initialize event state based on the perf_event_attr::disabled.
1889  */
1890 static inline void perf_event__state_init(struct perf_event *event)
1891 {
1892         event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1893                                               PERF_EVENT_STATE_INACTIVE;
1894 }
1895 
1896 static void __perf_event_read_size(struct perf_event *event, int nr_siblings)
1897 {
1898         int entry = sizeof(u64); /* value */
1899         int size = 0;
1900         int nr = 1;
1901 
1902         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1903                 size += sizeof(u64);
1904 
1905         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1906                 size += sizeof(u64);
1907 
1908         if (event->attr.read_format & PERF_FORMAT_ID)
1909                 entry += sizeof(u64);
1910 
1911         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1912                 nr += nr_siblings;
1913                 size += sizeof(u64);
1914         }
1915 
1916         size += entry * nr;
1917         event->read_size = size;
1918 }
1919 
1920 static void __perf_event_header_size(struct perf_event *event, u64 sample_type)
1921 {
1922         struct perf_sample_data *data;
1923         u16 size = 0;
1924 
1925         if (sample_type & PERF_SAMPLE_IP)
1926                 size += sizeof(data->ip);
1927 
1928         if (sample_type & PERF_SAMPLE_ADDR)
1929                 size += sizeof(data->addr);
1930 
1931         if (sample_type & PERF_SAMPLE_PERIOD)
1932                 size += sizeof(data->period);
1933 
1934         if (sample_type & PERF_SAMPLE_WEIGHT_TYPE)
1935                 size += sizeof(data->weight.full);
1936 
1937         if (sample_type & PERF_SAMPLE_READ)
1938                 size += event->read_size;
1939 
1940         if (sample_type & PERF_SAMPLE_DATA_SRC)
1941                 size += sizeof(data->data_src.val);
1942 
1943         if (sample_type & PERF_SAMPLE_TRANSACTION)
1944                 size += sizeof(data->txn);
1945 
1946         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
1947                 size += sizeof(data->phys_addr);
1948 
1949         if (sample_type & PERF_SAMPLE_CGROUP)
1950                 size += sizeof(data->cgroup);
1951 
1952         if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)
1953                 size += sizeof(data->data_page_size);
1954 
1955         if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)
1956                 size += sizeof(data->code_page_size);
1957 
1958         event->header_size = size;
1959 }
1960 
1961 /*
1962  * Called at perf_event creation and when events are attached/detached from a
1963  * group.
1964  */
1965 static void perf_event__header_size(struct perf_event *event)
1966 {
1967         __perf_event_read_size(event,
1968                                event->group_leader->nr_siblings);
1969         __perf_event_header_size(event, event->attr.sample_type);
1970 }
1971 
1972 static void perf_event__id_header_size(struct perf_event *event)
1973 {
1974         struct perf_sample_data *data;
1975         u64 sample_type = event->attr.sample_type;
1976         u16 size = 0;
1977 
1978         if (sample_type & PERF_SAMPLE_TID)
1979                 size += sizeof(data->tid_entry);
1980 
1981         if (sample_type & PERF_SAMPLE_TIME)
1982                 size += sizeof(data->time);
1983 
1984         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1985                 size += sizeof(data->id);
1986 
1987         if (sample_type & PERF_SAMPLE_ID)
1988                 size += sizeof(data->id);
1989 
1990         if (sample_type & PERF_SAMPLE_STREAM_ID)
1991                 size += sizeof(data->stream_id);
1992 
1993         if (sample_type & PERF_SAMPLE_CPU)
1994                 size += sizeof(data->cpu_entry);
1995 
1996         event->id_header_size = size;
1997 }
1998 
1999 static bool perf_event_validate_size(struct perf_event *event)
2000 {
2001         /*
2002          * The values computed here will be over-written when we actually
2003          * attach the event.
2004          */
2005         __perf_event_read_size(event, event->group_leader->nr_siblings + 1);
2006         __perf_event_header_size(event, event->attr.sample_type & ~PERF_SAMPLE_READ);
2007         perf_event__id_header_size(event);
2008 
2009         /*
2010          * Sum the lot; should not exceed the 64k limit we have on records.
2011          * Conservative limit to allow for callchains and other variable fields.
2012          */
2013         if (event->read_size + event->header_size +
2014             event->id_header_size + sizeof(struct perf_event_header) >= 16*1024)
2015                 return false;
2016 
2017         return true;
2018 }
2019 
2020 static void perf_group_attach(struct perf_event *event)
2021 {
2022         struct perf_event *group_leader = event->group_leader, *pos;
2023 
2024         lockdep_assert_held(&event->ctx->lock);
2025 
2026         /*
2027          * We can have double attach due to group movement in perf_event_open.
2028          */
2029         if (event->attach_state & PERF_ATTACH_GROUP)
2030                 return;
2031 
2032         event->attach_state |= PERF_ATTACH_GROUP;
2033 
2034         if (group_leader == event)
2035                 return;
2036 
2037         WARN_ON_ONCE(group_leader->ctx != event->ctx);
2038 
2039         group_leader->group_caps &= event->event_caps;
2040 
2041         list_add_tail(&event->sibling_list, &group_leader->sibling_list);
2042         group_leader->nr_siblings++;
2043 
2044         perf_event__header_size(group_leader);
2045 
2046         for_each_sibling_event(pos, group_leader)
2047                 perf_event__header_size(pos);
2048 }
2049 
2050 /*
2051  * Remove an event from the lists for its context.
2052  * Must be called with ctx->mutex and ctx->lock held.
2053  */
2054 static void
2055 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
2056 {
2057         WARN_ON_ONCE(event->ctx != ctx);
2058         lockdep_assert_held(&ctx->lock);
2059 
2060         /*
2061          * We can have double detach due to exit/hot-unplug + close.
2062          */
2063         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
2064                 return;
2065 
2066         event->attach_state &= ~PERF_ATTACH_CONTEXT;
2067 
2068         ctx->nr_events--;
2069         if (event->attr.inherit_stat)
2070                 ctx->nr_stat--;
2071 
2072         list_del_rcu(&event->event_entry);
2073 
2074         if (event->group_leader == event)
2075                 del_event_from_groups(event, ctx);
2076 
2077         /*
2078          * If event was in error state, then keep it
2079          * that way, otherwise bogus counts will be
2080          * returned on read(). The only way to get out
2081          * of error state is by explicit re-enabling
2082          * of the event
2083          */
2084         if (event->state > PERF_EVENT_STATE_OFF) {
2085                 perf_cgroup_event_disable(event, ctx);
2086                 perf_event_set_state(event, PERF_EVENT_STATE_OFF);
2087         }
2088 
2089         ctx->generation++;
2090 }
2091 
2092 static int
2093 perf_aux_output_match(struct perf_event *event, struct perf_event *aux_event)
2094 {
2095         if (!has_aux(aux_event))
2096                 return 0;
2097 
2098         if (!event->pmu->aux_output_match)
2099                 return 0;
2100 
2101         return event->pmu->aux_output_match(aux_event);
2102 }
2103 
2104 static void put_event(struct perf_event *event);
2105 static void event_sched_out(struct perf_event *event,
2106                             struct perf_cpu_context *cpuctx,
2107                             struct perf_event_context *ctx);
2108 
2109 static void perf_put_aux_event(struct perf_event *event)
2110 {
2111         struct perf_event_context *ctx = event->ctx;
2112         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2113         struct perf_event *iter;
2114 
2115         /*
2116          * If event uses aux_event tear down the link
2117          */
2118         if (event->aux_event) {
2119                 iter = event->aux_event;
2120                 event->aux_event = NULL;
2121                 put_event(iter);
2122                 return;
2123         }
2124 
2125         /*
2126          * If the event is an aux_event, tear down all links to
2127          * it from other events.
2128          */
2129         for_each_sibling_event(iter, event->group_leader) {
2130                 if (iter->aux_event != event)
2131                         continue;
2132 
2133                 iter->aux_event = NULL;
2134                 put_event(event);
2135 
2136                 /*
2137                  * If it's ACTIVE, schedule it out and put it into ERROR
2138                  * state so that we don't try to schedule it again. Note
2139                  * that perf_event_enable() will clear the ERROR status.
2140                  */
2141                 event_sched_out(iter, cpuctx, ctx);
2142                 perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
2143         }
2144 }
2145 
2146 static bool perf_need_aux_event(struct perf_event *event)
2147 {
2148         return !!event->attr.aux_output || !!event->attr.aux_sample_size;
2149 }
2150 
2151 static int perf_get_aux_event(struct perf_event *event,
2152                               struct perf_event *group_leader)
2153 {
2154         /*
2155          * Our group leader must be an aux event if we want to be
2156          * an aux_output. This way, the aux event will precede its
2157          * aux_output events in the group, and therefore will always
2158          * schedule first.
2159          */
2160         if (!group_leader)
2161                 return 0;
2162 
2163         /*
2164          * aux_output and aux_sample_size are mutually exclusive.
2165          */
2166         if (event->attr.aux_output && event->attr.aux_sample_size)
2167                 return 0;
2168 
2169         if (event->attr.aux_output &&
2170             !perf_aux_output_match(event, group_leader))
2171                 return 0;
2172 
2173         if (event->attr.aux_sample_size && !group_leader->pmu->snapshot_aux)
2174                 return 0;
2175 
2176         if (!atomic_long_inc_not_zero(&group_leader->refcount))
2177                 return 0;
2178 
2179         /*
2180          * Link aux_outputs to their aux event; this is undone in
2181          * perf_group_detach() by perf_put_aux_event(). When the
2182          * group in torn down, the aux_output events loose their
2183          * link to the aux_event and can't schedule any more.
2184          */
2185         event->aux_event = group_leader;
2186 
2187         return 1;
2188 }
2189 
2190 static inline struct list_head *get_event_list(struct perf_event *event)
2191 {
2192         struct perf_event_context *ctx = event->ctx;
2193         return event->attr.pinned ? &ctx->pinned_active : &ctx->flexible_active;
2194 }
2195 
2196 /*
2197  * Events that have PERF_EV_CAP_SIBLING require being part of a group and
2198  * cannot exist on their own, schedule them out and move them into the ERROR
2199  * state. Also see _perf_event_enable(), it will not be able to recover
2200  * this ERROR state.
2201  */
2202 static inline void perf_remove_sibling_event(struct perf_event *event)
2203 {
2204         struct perf_event_context *ctx = event->ctx;
2205         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2206 
2207         event_sched_out(event, cpuctx, ctx);
2208         perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
2209 }
2210 
2211 static void perf_group_detach(struct perf_event *event)
2212 {
2213         struct perf_event *leader = event->group_leader;
2214         struct perf_event *sibling, *tmp;
2215         struct perf_event_context *ctx = event->ctx;
2216 
2217         lockdep_assert_held(&ctx->lock);
2218 
2219         /*
2220          * We can have double detach due to exit/hot-unplug + close.
2221          */
2222         if (!(event->attach_state & PERF_ATTACH_GROUP))
2223                 return;
2224 
2225         event->attach_state &= ~PERF_ATTACH_GROUP;
2226 
2227         perf_put_aux_event(event);
2228 
2229         /*
2230          * If this is a sibling, remove it from its group.
2231          */
2232         if (leader != event) {
2233                 list_del_init(&event->sibling_list);
2234                 event->group_leader->nr_siblings--;
2235                 goto out;
2236         }
2237 
2238         /*
2239          * If this was a group event with sibling events then
2240          * upgrade the siblings to singleton events by adding them
2241          * to whatever list we are on.
2242          */
2243         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, sibling_list) {
2244 
2245                 if (sibling->event_caps & PERF_EV_CAP_SIBLING)
2246                         perf_remove_sibling_event(sibling);
2247 
2248                 sibling->group_leader = sibling;
2249                 list_del_init(&sibling->sibling_list);
2250 
2251                 /* Inherit group flags from the previous leader */
2252                 sibling->group_caps = event->group_caps;
2253 
2254                 if (!RB_EMPTY_NODE(&event->group_node)) {
2255                         add_event_to_groups(sibling, event->ctx);
2256 
2257                         if (sibling->state == PERF_EVENT_STATE_ACTIVE)
2258                                 list_add_tail(&sibling->active_list, get_event_list(sibling));
2259                 }
2260 
2261                 WARN_ON_ONCE(sibling->ctx != event->ctx);
2262         }
2263 
2264 out:
2265         for_each_sibling_event(tmp, leader)
2266                 perf_event__header_size(tmp);
2267 
2268         perf_event__header_size(leader);
2269 }
2270 
2271 static void sync_child_event(struct perf_event *child_event);
2272 
2273 static void perf_child_detach(struct perf_event *event)
2274 {
2275         struct perf_event *parent_event = event->parent;
2276 
2277         if (!(event->attach_state & PERF_ATTACH_CHILD))
2278                 return;
2279 
2280         event->attach_state &= ~PERF_ATTACH_CHILD;
2281 
2282         if (WARN_ON_ONCE(!parent_event))
2283                 return;
2284 
2285         lockdep_assert_held(&parent_event->child_mutex);
2286 
2287         sync_child_event(event);
2288         list_del_init(&event->child_list);
2289 }
2290 
2291 static bool is_orphaned_event(struct perf_event *event)
2292 {
2293         return event->state == PERF_EVENT_STATE_DEAD;
2294 }
2295 
2296 static inline int __pmu_filter_match(struct perf_event *event)
2297 {
2298         struct pmu *pmu = event->pmu;
2299         return pmu->filter_match ? pmu->filter_match(event) : 1;
2300 }
2301 
2302 /*
2303  * Check whether we should attempt to schedule an event group based on
2304  * PMU-specific filtering. An event group can consist of HW and SW events,
2305  * potentially with a SW leader, so we must check all the filters, to
2306  * determine whether a group is schedulable:
2307  */
2308 static inline int pmu_filter_match(struct perf_event *event)
2309 {
2310         struct perf_event *sibling;
2311 
2312         if (!__pmu_filter_match(event))
2313                 return 0;
2314 
2315         for_each_sibling_event(sibling, event) {
2316                 if (!__pmu_filter_match(sibling))
2317                         return 0;
2318         }
2319 
2320         return 1;
2321 }
2322 
2323 static inline int
2324 event_filter_match(struct perf_event *event)
2325 {
2326         return (event->cpu == -1 || event->cpu == smp_processor_id()) &&
2327                perf_cgroup_match(event) && pmu_filter_match(event);
2328 }
2329 
2330 static void
2331 event_sched_out(struct perf_event *event,
2332                   struct perf_cpu_context *cpuctx,
2333                   struct perf_event_context *ctx)
2334 {
2335         enum perf_event_state state = PERF_EVENT_STATE_INACTIVE;
2336 
2337         WARN_ON_ONCE(event->ctx != ctx);
2338         lockdep_assert_held(&ctx->lock);
2339 
2340         if (event->state != PERF_EVENT_STATE_ACTIVE)
2341                 return;
2342 
2343         /*
2344          * Asymmetry; we only schedule events _IN_ through ctx_sched_in(), but
2345          * we can schedule events _OUT_ individually through things like
2346          * __perf_remove_from_context().
2347          */
2348         list_del_init(&event->active_list);
2349 
2350         perf_pmu_disable(event->pmu);
2351 
2352         event->pmu->del(event, 0);
2353         event->oncpu = -1;
2354 
2355         if (READ_ONCE(event->pending_disable) >= 0) {
2356                 WRITE_ONCE(event->pending_disable, -1);
2357                 perf_cgroup_event_disable(event, ctx);
2358                 state = PERF_EVENT_STATE_OFF;
2359         }
2360         perf_event_set_state(event, state);
2361 
2362         if (!is_software_event(event))
2363                 cpuctx->active_oncpu--;
2364         if (!--ctx->nr_active)
2365                 perf_event_ctx_deactivate(ctx);
2366         if (event->attr.freq && event->attr.sample_freq)
2367                 ctx->nr_freq--;
2368         if (event->attr.exclusive || !cpuctx->active_oncpu)
2369                 cpuctx->exclusive = 0;
2370 
2371         perf_pmu_enable(event->pmu);
2372 }
2373 
2374 static void
2375 group_sched_out(struct perf_event *group_event,
2376                 struct perf_cpu_context *cpuctx,
2377                 struct perf_event_context *ctx)
2378 {
2379         struct perf_event *event;
2380 
2381         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
2382                 return;
2383 
2384         perf_pmu_disable(ctx->pmu);
2385 
2386         event_sched_out(group_event, cpuctx, ctx);
2387 
2388         /*
2389          * Schedule out siblings (if any):
2390          */
2391         for_each_sibling_event(event, group_event)
2392                 event_sched_out(event, cpuctx, ctx);
2393 
2394         perf_pmu_enable(ctx->pmu);
2395 }
2396 
2397 #define DETACH_GROUP    0x01UL
2398 #define DETACH_CHILD    0x02UL
2399 
2400 /*
2401  * Cross CPU call to remove a performance event
2402  *
2403  * We disable the event on the hardware level first. After that we
2404  * remove it from the context list.
2405  */
2406 static void
2407 __perf_remove_from_context(struct perf_event *event,
2408                            struct perf_cpu_context *cpuctx,
2409                            struct perf_event_context *ctx,
2410                            void *info)
2411 {
2412         unsigned long flags = (unsigned long)info;
2413 
2414         if (ctx->is_active & EVENT_TIME) {
2415                 update_context_time(ctx);
2416                 update_cgrp_time_from_cpuctx(cpuctx, false);
2417         }
2418 
2419         event_sched_out(event, cpuctx, ctx);
2420         if (flags & DETACH_GROUP)
2421                 perf_group_detach(event);
2422         if (flags & DETACH_CHILD)
2423                 perf_child_detach(event);
2424         list_del_event(event, ctx);
2425 
2426         if (!ctx->nr_events && ctx->is_active) {
2427                 if (ctx == &cpuctx->ctx)
2428                         update_cgrp_time_from_cpuctx(cpuctx, true);
2429 
2430                 ctx->is_active = 0;
2431                 ctx->rotate_necessary = 0;
2432                 if (ctx->task) {
2433                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
2434                         cpuctx->task_ctx = NULL;
2435                 }
2436         }
2437 }
2438 
2439 /*
2440  * Remove the event from a task's (or a CPU's) list of events.
2441  *
2442  * If event->ctx is a cloned context, callers must make sure that
2443  * every task struct that event->ctx->task could possibly point to
2444  * remains valid.  This is OK when called from perf_release since
2445  * that only calls us on the top-level context, which can't be a clone.
2446  * When called from perf_event_exit_task, it's OK because the
2447  * context has been detached from its task.
2448  */
2449 static void perf_remove_from_context(struct perf_event *event, unsigned long flags)
2450 {
2451         struct perf_event_context *ctx = event->ctx;
2452 
2453         lockdep_assert_held(&ctx->mutex);
2454 
2455         /*
2456          * Because of perf_event_exit_task(), perf_remove_from_context() ought
2457          * to work in the face of TASK_TOMBSTONE, unlike every other
2458          * event_function_call() user.
2459          */
2460         raw_spin_lock_irq(&ctx->lock);
2461         /*
2462          * Cgroup events are per-cpu events, and must IPI because of
2463          * cgrp_cpuctx_list.
2464          */
2465         if (!ctx->is_active && !is_cgroup_event(event)) {
2466                 __perf_remove_from_context(event, __get_cpu_context(ctx),
2467                                            ctx, (void *)flags);
2468                 raw_spin_unlock_irq(&ctx->lock);
2469                 return;
2470         }
2471         raw_spin_unlock_irq(&ctx->lock);
2472 
2473         event_function_call(event, __perf_remove_from_context, (void *)flags);
2474 }
2475 
2476 /*
2477  * Cross CPU call to disable a performance event
2478  */
2479 static void __perf_event_disable(struct perf_event *event,
2480                                  struct perf_cpu_context *cpuctx,
2481                                  struct perf_event_context *ctx,
2482                                  void *info)
2483 {
2484         if (event->state < PERF_EVENT_STATE_INACTIVE)
2485                 return;
2486 
2487         if (ctx->is_active & EVENT_TIME) {
2488                 update_context_time(ctx);
2489                 update_cgrp_time_from_event(event);
2490         }
2491 
2492         if (event == event->group_leader)
2493                 group_sched_out(event, cpuctx, ctx);
2494         else
2495                 event_sched_out(event, cpuctx, ctx);
2496 
2497         perf_event_set_state(event, PERF_EVENT_STATE_OFF);
2498         perf_cgroup_event_disable(event, ctx);
2499 }
2500 
2501 /*
2502  * Disable an event.
2503  *
2504  * If event->ctx is a cloned context, callers must make sure that
2505  * every task struct that event->ctx->task could possibly point to
2506  * remains valid.  This condition is satisfied when called through
2507  * perf_event_for_each_child or perf_event_for_each because they
2508  * hold the top-level event's child_mutex, so any descendant that
2509  * goes to exit will block in perf_event_exit_event().
2510  *
2511  * When called from perf_pending_event it's OK because event->ctx
2512  * is the current context on this CPU and preemption is disabled,
2513  * hence we can't get into perf_event_task_sched_out for this context.
2514  */
2515 static void _perf_event_disable(struct perf_event *event)
2516 {
2517         struct perf_event_context *ctx = event->ctx;
2518 
2519         raw_spin_lock_irq(&ctx->lock);
2520         if (event->state <= PERF_EVENT_STATE_OFF) {
2521                 raw_spin_unlock_irq(&ctx->lock);
2522                 return;
2523         }
2524         raw_spin_unlock_irq(&ctx->lock);
2525 
2526         event_function_call(event, __perf_event_disable, NULL);
2527 }
2528 
2529 void perf_event_disable_local(struct perf_event *event)
2530 {
2531         event_function_local(event, __perf_event_disable, NULL);
2532 }
2533 
2534 /*
2535  * Strictly speaking kernel users cannot create groups and therefore this
2536  * interface does not need the perf_event_ctx_lock() magic.
2537  */
2538 void perf_event_disable(struct perf_event *event)
2539 {
2540         struct perf_event_context *ctx;
2541 
2542         ctx = perf_event_ctx_lock(event);
2543         _perf_event_disable(event);
2544         perf_event_ctx_unlock(event, ctx);
2545 }
2546 EXPORT_SYMBOL_GPL(perf_event_disable);
2547 
2548 void perf_event_disable_inatomic(struct perf_event *event)
2549 {
2550         WRITE_ONCE(event->pending_disable, smp_processor_id());
2551         /* can fail, see perf_pending_event_disable() */
2552         irq_work_queue(&event->pending);
2553 }
2554 
2555 #define MAX_INTERRUPTS (~0ULL)
2556 
2557 static void perf_log_throttle(struct perf_event *event, int enable);
2558 static void perf_log_itrace_start(struct perf_event *event);
2559 
2560 static int
2561 event_sched_in(struct perf_event *event,
2562                  struct perf_cpu_context *cpuctx,
2563                  struct perf_event_context *ctx)
2564 {
2565         int ret = 0;
2566 
2567         WARN_ON_ONCE(event->ctx != ctx);
2568 
2569         lockdep_assert_held(&ctx->lock);
2570 
2571         if (event->state <= PERF_EVENT_STATE_OFF)
2572                 return 0;
2573 
2574         WRITE_ONCE(event->oncpu, smp_processor_id());
2575         /*
2576          * Order event::oncpu write to happen before the ACTIVE state is
2577          * visible. This allows perf_event_{stop,read}() to observe the correct
2578          * ->oncpu if it sees ACTIVE.
2579          */
2580         smp_wmb();
2581         perf_event_set_state(event, PERF_EVENT_STATE_ACTIVE);
2582 
2583         /*
2584          * Unthrottle events, since we scheduled we might have missed several
2585          * ticks already, also for a heavily scheduling task there is little
2586          * guarantee it'll get a tick in a timely manner.
2587          */
2588         if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
2589                 perf_log_throttle(event, 1);
2590                 event->hw.interrupts = 0;
2591         }
2592 
2593         perf_pmu_disable(event->pmu);
2594 
2595         perf_log_itrace_start(event);
2596 
2597         if (event->pmu->add(event, PERF_EF_START)) {
2598                 perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2599                 event->oncpu = -1;
2600                 ret = -EAGAIN;
2601                 goto out;
2602         }
2603 
2604         if (!is_software_event(event))
2605                 cpuctx->active_oncpu++;
2606         if (!ctx->nr_active++)
2607                 perf_event_ctx_activate(ctx);
2608         if (event->attr.freq && event->attr.sample_freq)
2609                 ctx->nr_freq++;
2610 
2611         if (event->attr.exclusive)
2612                 cpuctx->exclusive = 1;
2613 
2614 out:
2615         perf_pmu_enable(event->pmu);
2616 
2617         return ret;
2618 }
2619 
2620 static int
2621 group_sched_in(struct perf_event *group_event,
2622                struct perf_cpu_context *cpuctx,
2623                struct perf_event_context *ctx)
2624 {
2625         struct perf_event *event, *partial_group = NULL;
2626         struct pmu *pmu = ctx->pmu;
2627 
2628         if (group_event->state == PERF_EVENT_STATE_OFF)
2629                 return 0;
2630 
2631         pmu->start_txn(pmu, PERF_PMU_TXN_ADD);
2632 
2633         if (event_sched_in(group_event, cpuctx, ctx))
2634                 goto error;
2635 
2636         /*
2637          * Schedule in siblings as one group (if any):
2638          */
2639         for_each_sibling_event(event, group_event) {
2640                 if (event_sched_in(event, cpuctx, ctx)) {
2641                         partial_group = event;
2642                         goto group_error;
2643                 }
2644         }
2645 
2646         if (!pmu->commit_txn(pmu))
2647                 return 0;
2648 
2649 group_error:
2650         /*
2651          * Groups can be scheduled in as one unit only, so undo any
2652          * partial group before returning:
2653          * The events up to the failed event are scheduled out normally.
2654          */
2655         for_each_sibling_event(event, group_event) {
2656                 if (event == partial_group)
2657                         break;
2658 
2659                 event_sched_out(event, cpuctx, ctx);
2660         }
2661         event_sched_out(group_event, cpuctx, ctx);
2662 
2663 error:
2664         pmu->cancel_txn(pmu);
2665         return -EAGAIN;
2666 }
2667 
2668 /*
2669  * Work out whether we can put this event group on the CPU now.
2670  */
2671 static int group_can_go_on(struct perf_event *event,
2672                            struct perf_cpu_context *cpuctx,
2673                            int can_add_hw)
2674 {
2675         /*
2676          * Groups consisting entirely of software events can always go on.
2677          */
2678         if (event->group_caps & PERF_EV_CAP_SOFTWARE)
2679                 return 1;
2680         /*
2681          * If an exclusive group is already on, no other hardware
2682          * events can go on.
2683          */
2684         if (cpuctx->exclusive)
2685                 return 0;
2686         /*
2687          * If this group is exclusive and there are already
2688          * events on the CPU, it can't go on.
2689          */
2690         if (event->attr.exclusive && !list_empty(get_event_list(event)))
2691                 return 0;
2692         /*
2693          * Otherwise, try to add it if all previous groups were able
2694          * to go on.
2695          */
2696         return can_add_hw;
2697 }
2698 
2699 static void add_event_to_ctx(struct perf_event *event,
2700                                struct perf_event_context *ctx)
2701 {
2702         list_add_event(event, ctx);
2703         perf_group_attach(event);
2704 }
2705 
2706 static void ctx_sched_out(struct perf_event_context *ctx,
2707                           struct perf_cpu_context *cpuctx,
2708                           enum event_type_t event_type);
2709 static void
2710 ctx_sched_in(struct perf_event_context *ctx,
2711              struct perf_cpu_context *cpuctx,
2712              enum event_type_t event_type,
2713              struct task_struct *task);
2714 
2715 static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
2716                                struct perf_event_context *ctx,
2717                                enum event_type_t event_type)
2718 {
2719         if (!cpuctx->task_ctx)
2720                 return;
2721 
2722         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2723                 return;
2724 
2725         ctx_sched_out(ctx, cpuctx, event_type);
2726 }
2727 
2728 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2729                                 struct perf_event_context *ctx,
2730                                 struct task_struct *task)
2731 {
2732         cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2733         if (ctx)
2734                 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2735         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2736         if (ctx)
2737                 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2738 }
2739 
2740 /*
2741  * We want to maintain the following priority of scheduling:
2742  *  - CPU pinned (EVENT_CPU | EVENT_PINNED)
2743  *  - task pinned (EVENT_PINNED)
2744  *  - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
2745  *  - task flexible (EVENT_FLEXIBLE).
2746  *
2747  * In order to avoid unscheduling and scheduling back in everything every
2748  * time an event is added, only do it for the groups of equal priority and
2749  * below.
2750  *
2751  * This can be called after a batch operation on task events, in which case
2752  * event_type is a bit mask of the types of events involved. For CPU events,
2753  * event_type is only either EVENT_PINNED or EVENT_FLEXIBLE.
2754  */
2755 static void ctx_resched(struct perf_cpu_context *cpuctx,
2756                         struct perf_event_context *task_ctx,
2757                         enum event_type_t event_type)
2758 {
2759         enum event_type_t ctx_event_type;
2760         bool cpu_event = !!(event_type & EVENT_CPU);
2761 
2762         /*
2763          * If pinned groups are involved, flexible groups also need to be
2764          * scheduled out.
2765          */
2766         if (event_type & EVENT_PINNED)
2767                 event_type |= EVENT_FLEXIBLE;
2768 
2769         ctx_event_type = event_type & EVENT_ALL;
2770 
2771         perf_pmu_disable(cpuctx->ctx.pmu);
2772         if (task_ctx)
2773                 task_ctx_sched_out(cpuctx, task_ctx, event_type);
2774 
2775         /*
2776          * Decide which cpu ctx groups to schedule out based on the types
2777          * of events that caused rescheduling:
2778          *  - EVENT_CPU: schedule out corresponding groups;
2779          *  - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
2780          *  - otherwise, do nothing more.
2781          */
2782         if (cpu_event)
2783                 cpu_ctx_sched_out(cpuctx, ctx_event_type);
2784         else if (ctx_event_type & EVENT_PINNED)
2785                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2786 
2787         perf_event_sched_in(cpuctx, task_ctx, current);
2788         perf_pmu_enable(cpuctx->ctx.pmu);
2789 }
2790 
2791 void perf_pmu_resched(struct pmu *pmu)
2792 {
2793         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2794         struct perf_event_context *task_ctx = cpuctx->task_ctx;
2795 
2796         perf_ctx_lock(cpuctx, task_ctx);
2797         ctx_resched(cpuctx, task_ctx, EVENT_ALL|EVENT_CPU);
2798         perf_ctx_unlock(cpuctx, task_ctx);
2799 }
2800 
2801 /*
2802  * Cross CPU call to install and enable a performance event
2803  *
2804  * Very similar to remote_function() + event_function() but cannot assume that
2805  * things like ctx->is_active and cpuctx->task_ctx are set.
2806  */
2807 static int  __perf_install_in_context(void *info)
2808 {
2809         struct perf_event *event = info;
2810         struct perf_event_context *ctx = event->ctx;
2811         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2812         struct perf_event_context *task_ctx = cpuctx->task_ctx;
2813         bool reprogram = true;
2814         int ret = 0;
2815 
2816         raw_spin_lock(&cpuctx->ctx.lock);
2817         if (ctx->task) {
2818                 raw_spin_lock(&ctx->lock);
2819                 task_ctx = ctx;
2820 
2821                 reprogram = (ctx->task == current);
2822 
2823                 /*
2824                  * If the task is running, it must be running on this CPU,
2825                  * otherwise we cannot reprogram things.
2826                  *
2827                  * If its not running, we don't care, ctx->lock will
2828                  * serialize against it becoming runnable.
2829                  */
2830                 if (task_curr(ctx->task) && !reprogram) {
2831                         ret = -ESRCH;
2832                         goto unlock;
2833                 }
2834 
2835                 WARN_ON_ONCE(reprogram && cpuctx->task_ctx && cpuctx->task_ctx != ctx);
2836         } else if (task_ctx) {
2837                 raw_spin_lock(&task_ctx->lock);
2838         }
2839 
2840 #ifdef CONFIG_CGROUP_PERF
2841         if (event->state > PERF_EVENT_STATE_OFF && is_cgroup_event(event)) {
2842                 /*
2843                  * If the current cgroup doesn't match the event's
2844                  * cgroup, we should not try to schedule it.
2845                  */
2846                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
2847                 reprogram = cgroup_is_descendant(cgrp->css.cgroup,
2848                                         event->cgrp->css.cgroup);
2849         }
2850 #endif
2851 
2852         if (reprogram) {
2853                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2854                 add_event_to_ctx(event, ctx);
2855                 ctx_resched(cpuctx, task_ctx, get_event_type(event));
2856         } else {
2857                 add_event_to_ctx(event, ctx);
2858         }
2859 
2860 unlock:
2861         perf_ctx_unlock(cpuctx, task_ctx);
2862 
2863         return ret;
2864 }
2865 
2866 static bool exclusive_event_installable(struct perf_event *event,
2867                                         struct perf_event_context *ctx);
2868 
2869 /*
2870  * Attach a performance event to a context.
2871  *
2872  * Very similar to event_function_call, see comment there.
2873  */
2874 static void
2875 perf_install_in_context(struct perf_event_context *ctx,
2876                         struct perf_event *event,
2877                         int cpu)
2878 {
2879         struct task_struct *task = READ_ONCE(ctx->task);
2880 
2881         lockdep_assert_held(&ctx->mutex);
2882 
2883         WARN_ON_ONCE(!exclusive_event_installable(event, ctx));
2884 
2885         if (event->cpu != -1)
2886                 event->cpu = cpu;
2887 
2888         /*
2889          * Ensures that if we can observe event->ctx, both the event and ctx
2890          * will be 'complete'. See perf_iterate_sb_cpu().
2891          */
2892         smp_store_release(&event->ctx, ctx);
2893 
2894         /*
2895          * perf_event_attr::disabled events will not run and can be initialized
2896          * without IPI. Except when this is the first event for the context, in
2897          * that case we need the magic of the IPI to set ctx->is_active.
2898          * Similarly, cgroup events for the context also needs the IPI to
2899          * manipulate the cgrp_cpuctx_list.
2900          *
2901          * The IOC_ENABLE that is sure to follow the creation of a disabled
2902          * event will issue the IPI and reprogram the hardware.
2903          */
2904         if (__perf_effective_state(event) == PERF_EVENT_STATE_OFF &&
2905             ctx->nr_events && !is_cgroup_event(event)) {
2906                 raw_spin_lock_irq(&ctx->lock);
2907                 if (ctx->task == TASK_TOMBSTONE) {
2908                         raw_spin_unlock_irq(&ctx->lock);
2909                         return;
2910                 }
2911                 add_event_to_ctx(event, ctx);
2912                 raw_spin_unlock_irq(&ctx->lock);
2913                 return;
2914         }
2915 
2916         if (!task) {
2917                 cpu_function_call(cpu, __perf_install_in_context, event);
2918                 return;
2919         }
2920 
2921         /*
2922          * Should not happen, we validate the ctx is still alive before calling.
2923          */
2924         if (WARN_ON_ONCE(task == TASK_TOMBSTONE))
2925                 return;
2926 
2927         /*
2928          * Installing events is tricky because we cannot rely on ctx->is_active
2929          * to be set in case this is the nr_events 0 -> 1 transition.
2930          *
2931          * Instead we use task_curr(), which tells us if the task is running.
2932          * However, since we use task_curr() outside of rq::lock, we can race
2933          * against the actual state. This means the result can be wrong.
2934          *
2935          * If we get a false positive, we retry, this is harmless.
2936          *
2937          * If we get a false negative, things are complicated. If we are after
2938          * perf_event_context_sched_in() ctx::lock will serialize us, and the
2939          * value must be correct. If we're before, it doesn't matter since
2940          * perf_event_context_sched_in() will program the counter.
2941          *
2942          * However, this hinges on the remote context switch having observed
2943          * our task->perf_event_ctxp[] store, such that it will in fact take
2944          * ctx::lock in perf_event_context_sched_in().
2945          *
2946          * We do this by task_function_call(), if the IPI fails to hit the task
2947          * we know any future context switch of task must see the
2948          * perf_event_ctpx[] store.
2949          */
2950 
2951         /*
2952          * This smp_mb() orders the task->perf_event_ctxp[] store with the
2953          * task_cpu() load, such that if the IPI then does not find the task
2954          * running, a future context switch of that task must observe the
2955          * store.
2956          */
2957         smp_mb();
2958 again:
2959         if (!task_function_call(task, __perf_install_in_context, event))
2960                 return;
2961 
2962         raw_spin_lock_irq(&ctx->lock);
2963         task = ctx->task;
2964         if (WARN_ON_ONCE(task == TASK_TOMBSTONE)) {
2965                 /*
2966                  * Cannot happen because we already checked above (which also
2967                  * cannot happen), and we hold ctx->mutex, which serializes us
2968                  * against perf_event_exit_task_context().
2969                  */
2970                 raw_spin_unlock_irq(&ctx->lock);
2971                 return;
2972         }
2973         /*
2974          * If the task is not running, ctx->lock will avoid it becoming so,
2975          * thus we can safely install the event.
2976          */
2977         if (task_curr(task)) {
2978                 raw_spin_unlock_irq(&ctx->lock);
2979                 goto again;
2980         }
2981         add_event_to_ctx(event, ctx);
2982         raw_spin_unlock_irq(&ctx->lock);
2983 }
2984 
2985 /*
2986  * Cross CPU call to enable a performance event
2987  */
2988 static void __perf_event_enable(struct perf_event *event,
2989                                 struct perf_cpu_context *cpuctx,
2990                                 struct perf_event_context *ctx,
2991                                 void *info)
2992 {
2993         struct perf_event *leader = event->group_leader;
2994         struct perf_event_context *task_ctx;
2995 
2996         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2997             event->state <= PERF_EVENT_STATE_ERROR)
2998                 return;
2999 
3000         if (ctx->is_active)
3001                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
3002 
3003         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
3004         perf_cgroup_event_enable(event, ctx);
3005 
3006         if (!ctx->is_active)
3007                 return;
3008 
3009         if (!event_filter_match(event)) {
3010                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
3011                 return;
3012         }
3013 
3014         /*
3015          * If the event is in a group and isn't the group leader,
3016          * then don't put it on unless the group is on.
3017          */
3018         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) {
3019                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
3020                 return;
3021         }
3022 
3023         task_ctx = cpuctx->task_ctx;
3024         if (ctx->task)
3025                 WARN_ON_ONCE(task_ctx != ctx);
3026 
3027         ctx_resched(cpuctx, task_ctx, get_event_type(event));
3028 }
3029 
3030 /*
3031  * Enable an event.
3032  *
3033  * If event->ctx is a cloned context, callers must make sure that
3034  * every task struct that event->ctx->task could possibly point to
3035  * remains valid.  This condition is satisfied when called through
3036  * perf_event_for_each_child or perf_event_for_each as described
3037  * for perf_event_disable.
3038  */
3039 static void _perf_event_enable(struct perf_event *event)
3040 {
3041         struct perf_event_context *ctx = event->ctx;
3042 
3043         raw_spin_lock_irq(&ctx->lock);
3044         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
3045             event->state <  PERF_EVENT_STATE_ERROR) {
3046 out:
3047                 raw_spin_unlock_irq(&ctx->lock);
3048                 return;
3049         }
3050 
3051         /*
3052          * If the event is in error state, clear that first.
3053          *
3054          * That way, if we see the event in error state below, we know that it
3055          * has gone back into error state, as distinct from the task having
3056          * been scheduled away before the cross-call arrived.
3057          */
3058         if (event->state == PERF_EVENT_STATE_ERROR) {
3059                 /*
3060                  * Detached SIBLING events cannot leave ERROR state.
3061                  */
3062                 if (event->event_caps & PERF_EV_CAP_SIBLING &&
3063                     event->group_leader == event)
3064                         goto out;
3065 
3066                 event->state = PERF_EVENT_STATE_OFF;
3067         }
3068         raw_spin_unlock_irq(&ctx->lock);
3069 
3070         event_function_call(event, __perf_event_enable, NULL);
3071 }
3072 
3073 /*
3074  * See perf_event_disable();
3075  */
3076 void perf_event_enable(struct perf_event *event)
3077 {
3078         struct perf_event_context *ctx;
3079 
3080         ctx = perf_event_ctx_lock(event);
3081         _perf_event_enable(event);
3082         perf_event_ctx_unlock(event, ctx);
3083 }
3084 EXPORT_SYMBOL_GPL(perf_event_enable);
3085 
3086 struct stop_event_data {
3087         struct perf_event       *event;
3088         unsigned int            restart;
3089 };
3090 
3091 static int __perf_event_stop(void *info)
3092 {
3093         struct stop_event_data *sd = info;
3094         struct perf_event *event = sd->event;
3095 
3096         /* if it's already INACTIVE, do nothing */
3097         if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
3098                 return 0;
3099 
3100         /* matches smp_wmb() in event_sched_in() */
3101         smp_rmb();
3102 
3103         /*
3104          * There is a window with interrupts enabled before we get here,
3105          * so we need to check again lest we try to stop another CPU's event.
3106          */
3107         if (READ_ONCE(event->oncpu) != smp_processor_id())
3108                 return -EAGAIN;
3109 
3110         event->pmu->stop(event, PERF_EF_UPDATE);
3111 
3112         /*
3113          * May race with the actual stop (through perf_pmu_output_stop()),
3114          * but it is only used for events with AUX ring buffer, and such
3115          * events will refuse to restart because of rb::aux_mmap_count==0,
3116          * see comments in perf_aux_output_begin().
3117          *
3118          * Since this is happening on an event-local CPU, no trace is lost
3119          * while restarting.
3120          */
3121         if (sd->restart)
3122                 event->pmu->start(event, 0);
3123 
3124         return 0;
3125 }
3126 
3127 static int perf_event_stop(struct perf_event *event, int restart)
3128 {
3129         struct stop_event_data sd = {
3130                 .event          = event,
3131                 .restart        = restart,
3132         };
3133         int ret = 0;
3134 
3135         do {
3136                 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
3137                         return 0;
3138 
3139                 /* matches smp_wmb() in event_sched_in() */
3140                 smp_rmb();
3141 
3142                 /*
3143                  * We only want to restart ACTIVE events, so if the event goes
3144                  * inactive here (event->oncpu==-1), there's nothing more to do;
3145                  * fall through with ret==-ENXIO.
3146                  */
3147                 ret = cpu_function_call(READ_ONCE(event->oncpu),
3148                                         __perf_event_stop, &sd);
3149         } while (ret == -EAGAIN);
3150 
3151         return ret;
3152 }
3153 
3154 /*
3155  * In order to contain the amount of racy and tricky in the address filter
3156  * configuration management, it is a two part process:
3157  *
3158  * (p1) when userspace mappings change as a result of (1) or (2) or (3) below,
3159  *      we update the addresses of corresponding vmas in
3160  *      event::addr_filter_ranges array and bump the event::addr_filters_gen;
3161  * (p2) when an event is scheduled in (pmu::add), it calls
3162  *      perf_event_addr_filters_sync() which calls pmu::addr_filters_sync()
3163  *      if the generation has changed since the previous call.
3164  *
3165  * If (p1) happens while the event is active, we restart it to force (p2).
3166  *
3167  * (1) perf_addr_filters_apply(): adjusting filters' offsets based on
3168  *     pre-existing mappings, called once when new filters arrive via SET_FILTER
3169  *     ioctl;
3170  * (2) perf_addr_filters_adjust(): adjusting filters' offsets based on newly
3171  *     registered mapping, called for every new mmap(), with mm::mmap_lock down
3172  *     for reading;
3173  * (3) perf_event_addr_filters_exec(): clearing filters' offsets in the process
3174  *     of exec.
3175  */
3176 void perf_event_addr_filters_sync(struct perf_event *event)
3177 {
3178         struct perf_addr_filters_head *ifh = perf_event_addr_filters(event);
3179 
3180         if (!has_addr_filter(event))
3181                 return;
3182 
3183         raw_spin_lock(&ifh->lock);
3184         if (event->addr_filters_gen != event->hw.addr_filters_gen) {
3185                 event->pmu->addr_filters_sync(event);
3186                 event->hw.addr_filters_gen = event->addr_filters_gen;
3187         }
3188         raw_spin_unlock(&ifh->lock);
3189 }
3190 EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
3191 
3192 static int _perf_event_refresh(struct perf_event *event, int refresh)
3193 {
3194         /*
3195          * not supported on inherited events
3196          */
3197         if (event->attr.inherit || !is_sampling_event(event))
3198                 return -EINVAL;
3199 
3200         atomic_add(refresh, &event->event_limit);
3201         _perf_event_enable(event);
3202 
3203         return 0;
3204 }
3205 
3206 /*
3207  * See perf_event_disable()
3208  */
3209 int perf_event_refresh(struct perf_event *event, int refresh)
3210 {
3211         struct perf_event_context *ctx;
3212         int ret;
3213 
3214         ctx = perf_event_ctx_lock(event);
3215         ret = _perf_event_refresh(event, refresh);
3216         perf_event_ctx_unlock(event, ctx);
3217 
3218         return ret;
3219 }
3220 EXPORT_SYMBOL_GPL(perf_event_refresh);
3221 
3222 static int perf_event_modify_breakpoint(struct perf_event *bp,
3223                                          struct perf_event_attr *attr)
3224 {
3225         int err;
3226 
3227         _perf_event_disable(bp);
3228 
3229         err = modify_user_hw_breakpoint_check(bp, attr, true);
3230 
3231         if (!bp->attr.disabled)
3232                 _perf_event_enable(bp);
3233 
3234         return err;
3235 }
3236 
3237 /*
3238  * Copy event-type-independent attributes that may be modified.
3239  */
3240 static void perf_event_modify_copy_attr(struct perf_event_attr *to,
3241                                         const struct perf_event_attr *from)
3242 {
3243         to->sig_data = from->sig_data;
3244 }
3245 
3246 static int perf_event_modify_attr(struct perf_event *event,
3247                                   struct perf_event_attr *attr)
3248 {
3249         int (*func)(struct perf_event *, struct perf_event_attr *);
3250         struct perf_event *child;
3251         int err;
3252 
3253         if (event->attr.type != attr->type)
3254                 return -EINVAL;
3255 
3256         switch (event->attr.type) {
3257         case PERF_TYPE_BREAKPOINT:
3258                 func = perf_event_modify_breakpoint;
3259                 break;
3260         default:
3261                 /* Place holder for future additions. */
3262                 return -EOPNOTSUPP;
3263         }
3264 
3265         WARN_ON_ONCE(event->ctx->parent_ctx);
3266 
3267         mutex_lock(&event->child_mutex);
3268         /*
3269          * Event-type-independent attributes must be copied before event-type
3270          * modification, which will validate that final attributes match the
3271          * source attributes after all relevant attributes have been copied.
3272          */
3273         perf_event_modify_copy_attr(&event->attr, attr);
3274         err = func(event, attr);
3275         if (err)
3276                 goto out;
3277         list_for_each_entry(child, &event->child_list, child_list) {
3278                 perf_event_modify_copy_attr(&child->attr, attr);
3279                 err = func(child, attr);
3280                 if (err)
3281                         goto out;
3282         }
3283 out:
3284         mutex_unlock(&event->child_mutex);
3285         return err;
3286 }
3287 
3288 static void ctx_sched_out(struct perf_event_context *ctx,
3289                           struct perf_cpu_context *cpuctx,
3290                           enum event_type_t event_type)
3291 {
3292         struct perf_event *event, *tmp;
3293         int is_active = ctx->is_active;
3294 
3295         lockdep_assert_held(&ctx->lock);
3296 
3297         if (likely(!ctx->nr_events)) {
3298                 /*
3299                  * See __perf_remove_from_context().
3300                  */
3301                 WARN_ON_ONCE(ctx->is_active);
3302                 if (ctx->task)
3303                         WARN_ON_ONCE(cpuctx->task_ctx);
3304                 return;
3305         }
3306 
3307         /*
3308          * Always update time if it was set; not only when it changes.
3309          * Otherwise we can 'forget' to update time for any but the last
3310          * context we sched out. For example:
3311          *
3312          *   ctx_sched_out(.event_type = EVENT_FLEXIBLE)
3313          *   ctx_sched_out(.event_type = EVENT_PINNED)
3314          *
3315          * would only update time for the pinned events.
3316          */
3317         if (is_active & EVENT_TIME) {
3318                 /* update (and stop) ctx time */
3319                 update_context_time(ctx);
3320                 update_cgrp_time_from_cpuctx(cpuctx, ctx == &cpuctx->ctx);
3321                 /*
3322                  * CPU-release for the below ->is_active store,
3323                  * see __load_acquire() in perf_event_time_now()
3324                  */
3325                 barrier();
3326         }
3327 
3328         ctx->is_active &= ~event_type;
3329         if (!(ctx->is_active & EVENT_ALL))
3330                 ctx->is_active = 0;
3331 
3332         if (ctx->task) {
3333                 WARN_ON_ONCE(cpuctx->task_ctx != ctx);
3334                 if (!ctx->is_active)
3335                         cpuctx->task_ctx = NULL;
3336         }
3337 
3338         is_active ^= ctx->is_active; /* changed bits */
3339 
3340         if (!ctx->nr_active || !(is_active & EVENT_ALL))
3341                 return;
3342 
3343         perf_pmu_disable(ctx->pmu);
3344         if (is_active & EVENT_PINNED) {
3345                 list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list)
3346                         group_sched_out(event, cpuctx, ctx);
3347         }
3348 
3349         if (is_active & EVENT_FLEXIBLE) {
3350                 list_for_each_entry_safe(event, tmp, &ctx->flexible_active, active_list)
3351                         group_sched_out(event, cpuctx, ctx);
3352 
3353                 /*
3354                  * Since we cleared EVENT_FLEXIBLE, also clear
3355                  * rotate_necessary, is will be reset by
3356                  * ctx_flexible_sched_in() when needed.
3357                  */
3358                 ctx->rotate_necessary = 0;
3359         }
3360         perf_pmu_enable(ctx->pmu);
3361 }
3362 
3363 /*
3364  * Test whether two contexts are equivalent, i.e. whether they have both been
3365  * cloned from the same version of the same context.
3366  *
3367  * Equivalence is measured using a generation number in the context that is
3368  * incremented on each modification to it; see unclone_ctx(), list_add_event()
3369  * and list_del_event().
3370  */
3371 static int context_equiv(struct perf_event_context *ctx1,
3372                          struct perf_event_context *ctx2)
3373 {
3374         lockdep_assert_held(&ctx1->lock);
3375         lockdep_assert_held(&ctx2->lock);
3376 
3377         /* Pinning disables the swap optimization */
3378         if (ctx1->pin_count || ctx2->pin_count)
3379                 return 0;
3380 
3381         /* If ctx1 is the parent of ctx2 */
3382         if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
3383                 return 1;
3384 
3385         /* If ctx2 is the parent of ctx1 */
3386         if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
3387                 return 1;
3388 
3389         /*
3390          * If ctx1 and ctx2 have the same parent; we flatten the parent
3391          * hierarchy, see perf_event_init_context().
3392          */
3393         if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
3394                         ctx1->parent_gen == ctx2->parent_gen)
3395                 return 1;
3396 
3397         /* Unmatched */
3398         return 0;
3399 }
3400 
3401 static void __perf_event_sync_stat(struct perf_event *event,
3402                                      struct perf_event *next_event)
3403 {
3404         u64 value;
3405 
3406         if (!event->attr.inherit_stat)
3407                 return;
3408 
3409         /*
3410          * Update the event value, we cannot use perf_event_read()
3411          * because we're in the middle of a context switch and have IRQs
3412          * disabled, which upsets smp_call_function_single(), however
3413          * we know the event must be on the current CPU, therefore we
3414          * don't need to use it.
3415          */
3416         if (event->state == PERF_EVENT_STATE_ACTIVE)
3417                 event->pmu->read(event);
3418 
3419         perf_event_update_time(event);
3420 
3421         /*
3422          * In order to keep per-task stats reliable we need to flip the event
3423          * values when we flip the contexts.
3424          */
3425         value = local64_read(&next_event->count);
3426         value = local64_xchg(&event->count, value);
3427         local64_set(&next_event->count, value);
3428 
3429         swap(event->total_time_enabled, next_event->total_time_enabled);
3430         swap(event->total_time_running, next_event->total_time_running);
3431 
3432         /*
3433          * Since we swizzled the values, update the user visible data too.
3434          */
3435         perf_event_update_userpage(event);
3436         perf_event_update_userpage(next_event);
3437 }
3438 
3439 static void perf_event_sync_stat(struct perf_event_context *ctx,
3440                                    struct perf_event_context *next_ctx)
3441 {
3442         struct perf_event *event, *next_event;
3443 
3444         if (!ctx->nr_stat)
3445                 return;
3446 
3447         update_context_time(ctx);
3448 
3449         event = list_first_entry(&ctx->event_list,
3450                                    struct perf_event, event_entry);
3451 
3452         next_event = list_first_entry(&next_ctx->event_list,
3453                                         struct perf_event, event_entry);
3454 
3455         while (&event->event_entry != &ctx->event_list &&
3456                &next_event->event_entry != &next_ctx->event_list) {
3457 
3458                 __perf_event_sync_stat(event, next_event);
3459 
3460                 event = list_next_entry(event, event_entry);
3461                 next_event = list_next_entry(next_event, event_entry);
3462         }
3463 }
3464 
3465 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
3466                                          struct task_struct *next)
3467 {
3468         struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
3469         struct perf_event_context *next_ctx;
3470         struct perf_event_context *parent, *next_parent;
3471         struct perf_cpu_context *cpuctx;
3472         int do_switch = 1;
3473         struct pmu *pmu;
3474 
3475         if (likely(!ctx))
3476                 return;
3477 
3478         pmu = ctx->pmu;
3479         cpuctx = __get_cpu_context(ctx);
3480         if (!cpuctx->task_ctx)
3481                 return;
3482 
3483         rcu_read_lock();
3484         next_ctx = next->perf_event_ctxp[ctxn];
3485         if (!next_ctx)
3486                 goto unlock;
3487 
3488         parent = rcu_dereference(ctx->parent_ctx);
3489         next_parent = rcu_dereference(next_ctx->parent_ctx);
3490 
3491         /* If neither context have a parent context; they cannot be clones. */
3492         if (!parent && !next_parent)
3493                 goto unlock;
3494 
3495         if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
3496                 /*
3497                  * Looks like the two contexts are clones, so we might be
3498                  * able to optimize the context switch.  We lock both
3499                  * contexts and check that they are clones under the
3500                  * lock (including re-checking that neither has been
3501                  * uncloned in the meantime).  It doesn't matter which
3502                  * order we take the locks because no other cpu could
3503                  * be trying to lock both of these tasks.
3504                  */
3505                 raw_spin_lock(&ctx->lock);
3506                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
3507                 if (context_equiv(ctx, next_ctx)) {
3508 
3509                         WRITE_ONCE(ctx->task, next);
3510                         WRITE_ONCE(next_ctx->task, task);
3511 
3512                         perf_pmu_disable(pmu);
3513 
3514                         if (cpuctx->sched_cb_usage && pmu->sched_task)
3515                                 pmu->sched_task(ctx, false);
3516 
3517                         /*
3518                          * PMU specific parts of task perf context can require
3519                          * additional synchronization. As an example of such
3520                          * synchronization see implementation details of Intel
3521                          * LBR call stack data profiling;
3522                          */
3523                         if (pmu->swap_task_ctx)
3524                                 pmu->swap_task_ctx(ctx, next_ctx);
3525                         else
3526                                 swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
3527 
3528                         perf_pmu_enable(pmu);
3529 
3530                         /*
3531                          * RCU_INIT_POINTER here is safe because we've not
3532                          * modified the ctx and the above modification of
3533                          * ctx->task and ctx->task_ctx_data are immaterial
3534                          * since those values are always verified under
3535                          * ctx->lock which we're now holding.
3536                          */
3537                         RCU_INIT_POINTER(task->perf_event_ctxp[ctxn], next_ctx);
3538                         RCU_INIT_POINTER(next->perf_event_ctxp[ctxn], ctx);
3539 
3540                         do_switch = 0;
3541 
3542                         perf_event_sync_stat(ctx, next_ctx);
3543                 }
3544                 raw_spin_unlock(&next_ctx->lock);
3545                 raw_spin_unlock(&ctx->lock);
3546         }
3547 unlock:
3548         rcu_read_unlock();
3549 
3550         if (do_switch) {
3551                 raw_spin_lock(&ctx->lock);
3552                 perf_pmu_disable(pmu);
3553 
3554                 if (cpuctx->sched_cb_usage && pmu->sched_task)
3555                         pmu->sched_task(ctx, false);
3556                 task_ctx_sched_out(cpuctx, ctx, EVENT_ALL);
3557 
3558                 perf_pmu_enable(pmu);
3559                 raw_spin_unlock(&ctx->lock);
3560         }
3561 }
3562 
3563 static DEFINE_PER_CPU(struct list_head, sched_cb_list);
3564 
3565 void perf_sched_cb_dec(struct pmu *pmu)
3566 {
3567         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3568 
3569         this_cpu_dec(perf_sched_cb_usages);
3570 
3571         if (!--cpuctx->sched_cb_usage)
3572                 list_del(&cpuctx->sched_cb_entry);
3573 }
3574 
3575 
3576 void perf_sched_cb_inc(struct pmu *pmu)
3577 {
3578         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3579 
3580         if (!cpuctx->sched_cb_usage++)
3581                 list_add(&cpuctx->sched_cb_entry, this_cpu_ptr(&sched_cb_list));
3582 
3583         this_cpu_inc(perf_sched_cb_usages);
3584 }
3585 
3586 /*
3587  * This function provides the context switch callback to the lower code
3588  * layer. It is invoked ONLY when the context switch callback is enabled.
3589  *
3590  * This callback is relevant even to per-cpu events; for example multi event
3591  * PEBS requires this to provide PID/TID information. This requires we flush
3592  * all queued PEBS records before we context switch to a new task.
3593  */
3594 static void __perf_pmu_sched_task(struct perf_cpu_context *cpuctx, bool sched_in)
3595 {
3596         struct pmu *pmu;
3597 
3598         pmu = cpuctx->ctx.pmu; /* software PMUs will not have sched_task */
3599 
3600         if (WARN_ON_ONCE(!pmu->sched_task))
3601                 return;
3602 
3603         perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3604         perf_pmu_disable(pmu);
3605 
3606         pmu->sched_task(cpuctx->task_ctx, sched_in);
3607 
3608         perf_pmu_enable(pmu);
3609         perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3610 }
3611 
3612 static void perf_pmu_sched_task(struct task_struct *prev,
3613                                 struct task_struct *next,
3614                                 bool sched_in)
3615 {
3616         struct perf_cpu_context *cpuctx;
3617 
3618         if (prev == next)
3619                 return;
3620 
3621         list_for_each_entry(cpuctx, this_cpu_ptr(&sched_cb_list), sched_cb_entry) {
3622                 /* will be handled in perf_event_context_sched_in/out */
3623                 if (cpuctx->task_ctx)
3624                         continue;
3625 
3626                 __perf_pmu_sched_task(cpuctx, sched_in);
3627         }
3628 }
3629 
3630 static void perf_event_switch(struct task_struct *task,
3631                               struct task_struct *next_prev, bool sched_in);
3632 
3633 #define for_each_task_context_nr(ctxn)                                  \
3634         for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
3635 
3636 /*
3637  * Called from scheduler to remove the events of the current task,
3638  * with interrupts disabled.
3639  *
3640  * We stop each event and update the event value in event->count.
3641  *
3642  * This does not protect us against NMI, but disable()
3643  * sets the disabled bit in the control field of event _before_
3644  * accessing the event control register. If a NMI hits, then it will
3645  * not restart the event.
3646  */
3647 void __perf_event_task_sched_out(struct task_struct *task,
3648                                  struct task_struct *next)
3649 {
3650         int ctxn;
3651 
3652         if (__this_cpu_read(perf_sched_cb_usages))
3653                 perf_pmu_sched_task(task, next, false);
3654 
3655         if (atomic_read(&nr_switch_events))
3656                 perf_event_switch(task, next, false);
3657 
3658         for_each_task_context_nr(ctxn)
3659                 perf_event_context_sched_out(task, ctxn, next);
3660 
3661         /*
3662          * if cgroup events exist on this CPU, then we need
3663          * to check if we have to switch out PMU state.
3664          * cgroup event are system-wide mode only
3665          */
3666         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3667                 perf_cgroup_sched_out(task, next);
3668 }
3669 
3670 /*
3671  * Called with IRQs disabled
3672  */
3673 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
3674                               enum event_type_t event_type)
3675 {
3676         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
3677 }
3678 
3679 static bool perf_less_group_idx(const void *l, const void *r)
3680 {
3681         const struct perf_event *le = *(const struct perf_event **)l;
3682         const struct perf_event *re = *(const struct perf_event **)r;
3683 
3684         return le->group_index < re->group_index;
3685 }
3686 
3687 static void swap_ptr(void *l, void *r)
3688 {
3689         void **lp = l, **rp = r;
3690 
3691         swap(*lp, *rp);
3692 }
3693 
3694 static const struct min_heap_callbacks perf_min_heap = {
3695         .elem_size = sizeof(struct perf_event *),
3696         .less = perf_less_group_idx,
3697         .swp = swap_ptr,
3698 };
3699 
3700 static void __heap_add(struct min_heap *heap, struct perf_event *event)
3701 {
3702         struct perf_event **itrs = heap->data;
3703 
3704         if (event) {
3705                 itrs[heap->nr] = event;
3706                 heap->nr++;
3707         }
3708 }
3709 
3710 static noinline int visit_groups_merge(struct perf_cpu_context *cpuctx,
3711                                 struct perf_event_groups *groups, int cpu,
3712                                 int (*func)(struct perf_event *, void *),
3713                                 void *data)
3714 {
3715 #ifdef CONFIG_CGROUP_PERF
3716         struct cgroup_subsys_state *css = NULL;
3717 #endif
3718         /* Space for per CPU and/or any CPU event iterators. */
3719         struct perf_event *itrs[2];
3720         struct min_heap event_heap;
3721         struct perf_event **evt;
3722         int ret;
3723 
3724         if (cpuctx) {
3725                 event_heap = (struct min_heap){
3726                         .data = cpuctx->heap,
3727                         .nr = 0,
3728                         .size = cpuctx->heap_size,
3729                 };
3730 
3731                 lockdep_assert_held(&cpuctx->ctx.lock);
3732 
3733 #ifdef CONFIG_CGROUP_PERF
3734                 if (cpuctx->cgrp)
3735                         css = &cpuctx->cgrp->css;
3736 #endif
3737         } else {
3738                 event_heap = (struct min_heap){
3739                         .data = itrs,
3740                         .nr = 0,
3741                         .size = ARRAY_SIZE(itrs),
3742                 };
3743                 /* Events not within a CPU context may be on any CPU. */
3744                 __heap_add(&event_heap, perf_event_groups_first(groups, -1, NULL));
3745         }
3746         evt = event_heap.data;
3747 
3748         __heap_add(&event_heap, perf_event_groups_first(groups, cpu, NULL));
3749 
3750 #ifdef CONFIG_CGROUP_PERF
3751         for (; css; css = css->parent)
3752                 __heap_add(&event_heap, perf_event_groups_first(groups, cpu, css->cgroup));
3753 #endif
3754 
3755         min_heapify_all(&event_heap, &perf_min_heap);
3756 
3757         while (event_heap.nr) {
3758                 ret = func(*evt, data);
3759                 if (ret)
3760                         return ret;
3761 
3762                 *evt = perf_event_groups_next(*evt);
3763                 if (*evt)
3764                         min_heapify(&event_heap, 0, &perf_min_heap);
3765                 else
3766                         min_heap_pop(&event_heap, &perf_min_heap);
3767         }
3768 
3769         return 0;
3770 }
3771 
3772 /*
3773  * Because the userpage is strictly per-event (there is no concept of context,
3774  * so there cannot be a context indirection), every userpage must be updated
3775  * when context time starts :-(
3776  *
3777  * IOW, we must not miss EVENT_TIME edges.
3778  */
3779 static inline bool event_update_userpage(struct perf_event *event)
3780 {
3781         if (likely(!atomic_read(&event->mmap_count)))
3782                 return false;
3783 
3784         perf_event_update_time(event);
3785         perf_event_update_userpage(event);
3786 
3787         return true;
3788 }
3789 
3790 static inline void group_update_userpage(struct perf_event *group_event)
3791 {
3792         struct perf_event *event;
3793 
3794         if (!event_update_userpage(group_event))
3795                 return;
3796 
3797         for_each_sibling_event(event, group_event)
3798                 event_update_userpage(event);
3799 }
3800 
3801 static int merge_sched_in(struct perf_event *event, void *data)
3802 {
3803         struct perf_event_context *ctx = event->ctx;
3804         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3805         int *can_add_hw = data;
3806 
3807         if (event->state <= PERF_EVENT_STATE_OFF)
3808                 return 0;
3809 
3810         if (!event_filter_match(event))
3811                 return 0;
3812 
3813         if (group_can_go_on(event, cpuctx, *can_add_hw)) {
3814                 if (!group_sched_in(event, cpuctx, ctx))
3815                         list_add_tail(&event->active_list, get_event_list(event));
3816         }
3817 
3818         if (event->state == PERF_EVENT_STATE_INACTIVE) {
3819                 *can_add_hw = 0;
3820                 if (event->attr.pinned) {
3821                         perf_cgroup_event_disable(event, ctx);
3822                         perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
3823                 } else {
3824                         ctx->rotate_necessary = 1;
3825                         perf_mux_hrtimer_restart(cpuctx);
3826                         group_update_userpage(event);
3827                 }
3828         }
3829 
3830         return 0;
3831 }
3832 
3833 static void
3834 ctx_pinned_sched_in(struct perf_event_context *ctx,
3835                     struct perf_cpu_context *cpuctx)
3836 {
3837         int can_add_hw = 1;
3838 
3839         if (ctx != &cpuctx->ctx)
3840                 cpuctx = NULL;
3841 
3842         visit_groups_merge(cpuctx, &ctx->pinned_groups,
3843                            smp_processor_id(),
3844                            merge_sched_in, &can_add_hw);
3845 }
3846 
3847 static void
3848 ctx_flexible_sched_in(struct perf_event_context *ctx,
3849                       struct perf_cpu_context *cpuctx)
3850 {
3851         int can_add_hw = 1;
3852 
3853         if (ctx != &cpuctx->ctx)
3854                 cpuctx = NULL;
3855 
3856         visit_groups_merge(cpuctx, &ctx->flexible_groups,
3857                            smp_processor_id(),
3858                            merge_sched_in, &can_add_hw);
3859 }
3860 
3861 static void
3862 ctx_sched_in(struct perf_event_context *ctx,
3863              struct perf_cpu_context *cpuctx,
3864              enum event_type_t event_type,
3865              struct task_struct *task)
3866 {
3867         int is_active = ctx->is_active;
3868 
3869         lockdep_assert_held(&ctx->lock);
3870 
3871         if (likely(!ctx->nr_events))
3872                 return;
3873 
3874         if (is_active ^ EVENT_TIME) {
3875                 /* start ctx time */
3876                 __update_context_time(ctx, false);
3877                 perf_cgroup_set_timestamp(task, ctx);
3878                 /*
3879                  * CPU-release for the below ->is_active store,
3880                  * see __load_acquire() in perf_event_time_now()
3881                  */
3882                 barrier();
3883         }
3884 
3885         ctx->is_active |= (event_type | EVENT_TIME);
3886         if (ctx->task) {
3887                 if (!is_active)
3888                         cpuctx->task_ctx = ctx;
3889                 else
3890                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
3891         }
3892 
3893         is_active ^= ctx->is_active; /* changed bits */
3894 
3895         /*
3896          * First go through the list and put on any pinned groups
3897          * in order to give them the best chance of going on.
3898          */
3899         if (is_active & EVENT_PINNED)
3900                 ctx_pinned_sched_in(ctx, cpuctx);
3901 
3902         /* Then walk through the lower prio flexible groups */
3903         if (is_active & EVENT_FLEXIBLE)
3904                 ctx_flexible_sched_in(ctx, cpuctx);
3905 }
3906 
3907 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
3908                              enum event_type_t event_type,
3909                              struct task_struct *task)
3910 {
3911         struct perf_event_context *ctx = &cpuctx->ctx;
3912 
3913         ctx_sched_in(ctx, cpuctx, event_type, task);
3914 }
3915 
3916 static void perf_event_context_sched_in(struct perf_event_context *ctx,
3917                                         struct task_struct *task)
3918 {
3919         struct perf_cpu_context *cpuctx;
3920         struct pmu *pmu;
3921 
3922         cpuctx = __get_cpu_context(ctx);
3923 
3924         /*
3925          * HACK: for HETEROGENEOUS the task context might have switched to a
3926          * different PMU, force (re)set the context,
3927          */
3928         pmu = ctx->pmu = cpuctx->ctx.pmu;
3929 
3930         if (cpuctx->task_ctx == ctx) {
3931                 if (cpuctx->sched_cb_usage)
3932                         __perf_pmu_sched_task(cpuctx, true);
3933                 return;
3934         }
3935 
3936         perf_ctx_lock(cpuctx, ctx);
3937         /*
3938          * We must check ctx->nr_events while holding ctx->lock, such
3939          * that we serialize against perf_install_in_context().
3940          */
3941         if (!ctx->nr_events)
3942                 goto unlock;
3943 
3944         perf_pmu_disable(pmu);
3945         /*
3946          * We want to keep the following priority order:
3947          * cpu pinned (that don't need to move), task pinned,
3948          * cpu flexible, task flexible.
3949          *
3950          * However, if task's ctx is not carrying any pinned
3951          * events, no need to flip the cpuctx's events around.
3952          */
3953         if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree))
3954                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3955         perf_event_sched_in(cpuctx, ctx, task);
3956 
3957         if (cpuctx->sched_cb_usage && pmu->sched_task)
3958                 pmu->sched_task(cpuctx->task_ctx, true);
3959 
3960         perf_pmu_enable(pmu);
3961 
3962 unlock:
3963         perf_ctx_unlock(cpuctx, ctx);
3964 }
3965 
3966 /*
3967  * Called from scheduler to add the events of the current task
3968  * with interrupts disabled.
3969  *
3970  * We restore the event value and then enable it.
3971  *
3972  * This does not protect us against NMI, but enable()
3973  * sets the enabled bit in the control field of event _before_
3974  * accessing the event control register. If a NMI hits, then it will
3975  * keep the event running.
3976  */
3977 void __perf_event_task_sched_in(struct task_struct *prev,
3978                                 struct task_struct *task)
3979 {
3980         struct perf_event_context *ctx;
3981         int ctxn;
3982 
3983         /*
3984          * If cgroup events exist on this CPU, then we need to check if we have
3985          * to switch in PMU state; cgroup event are system-wide mode only.
3986          *
3987          * Since cgroup events are CPU events, we must schedule these in before
3988          * we schedule in the task events.
3989          */
3990         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3991                 perf_cgroup_sched_in(prev, task);
3992 
3993         for_each_task_context_nr(ctxn) {
3994                 ctx = task->perf_event_ctxp[ctxn];
3995                 if (likely(!ctx))
3996                         continue;
3997 
3998                 perf_event_context_sched_in(ctx, task);
3999         }
4000 
4001         if (atomic_read(&nr_switch_events))
4002                 perf_event_switch(task, prev, true);
4003 
4004         if (__this_cpu_read(perf_sched_cb_usages))
4005                 perf_pmu_sched_task(prev, task, true);
4006 }
4007 
4008 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
4009 {
4010         u64 frequency = event->attr.sample_freq;
4011         u64 sec = NSEC_PER_SEC;
4012         u64 divisor, dividend;
4013 
4014         int count_fls, nsec_fls, frequency_fls, sec_fls;
4015 
4016         count_fls = fls64(count);
4017         nsec_fls = fls64(nsec);
4018         frequency_fls = fls64(frequency);
4019         sec_fls = 30;
4020 
4021         /*
4022          * We got @count in @nsec, with a target of sample_freq HZ
4023          * the target period becomes:
4024          *
4025          *             @count * 10^9
4026          * period = -------------------
4027          *          @nsec * sample_freq
4028          *
4029          */
4030 
4031         /*
4032          * Reduce accuracy by one bit such that @a and @b converge
4033          * to a similar magnitude.
4034          */
4035 #define REDUCE_FLS(a, b)                \
4036 do {                                    \
4037         if (a##_fls > b##_fls) {        \
4038                 a >>= 1;                \
4039                 a##_fls--;              \
4040         } else {                        \
4041                 b >>= 1;                \
4042                 b##_fls--;              \
4043         }                               \
4044 } while (0)
4045 
4046         /*
4047          * Reduce accuracy until either term fits in a u64, then proceed with
4048          * the other, so that finally we can do a u64/u64 division.
4049          */
4050         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
4051                 REDUCE_FLS(nsec, frequency);
4052                 REDUCE_FLS(sec, count);
4053         }
4054 
4055         if (count_fls + sec_fls > 64) {
4056                 divisor = nsec * frequency;
4057 
4058                 while (count_fls + sec_fls > 64) {
4059                         REDUCE_FLS(count, sec);
4060                         divisor >>= 1;
4061                 }
4062 
4063                 dividend = count * sec;
4064         } else {
4065                 dividend = count * sec;
4066 
4067                 while (nsec_fls + frequency_fls > 64) {
4068                         REDUCE_FLS(nsec, frequency);
4069                         dividend >>= 1;
4070                 }
4071 
4072                 divisor = nsec * frequency;
4073         }
4074 
4075         if (!divisor)
4076                 return dividend;
4077 
4078         return div64_u64(dividend, divisor);
4079 }
4080 
4081 static DEFINE_PER_CPU(int, perf_throttled_count);
4082 static DEFINE_PER_CPU(u64, perf_throttled_seq);
4083 
4084 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
4085 {
4086         struct hw_perf_event *hwc = &event->hw;
4087         s64 period, sample_period;
4088         s64 delta;
4089 
4090         period = perf_calculate_period(event, nsec, count);
4091 
4092         delta = (s64)(period - hwc->sample_period);
4093         delta = (delta + 7) / 8; /* low pass filter */
4094 
4095         sample_period = hwc->sample_period + delta;
4096 
4097         if (!sample_period)
4098                 sample_period = 1;
4099 
4100         hwc->sample_period = sample_period;
4101 
4102         if (local64_read(&hwc->period_left) > 8*sample_period) {
4103                 if (disable)
4104                         event->pmu->stop(event, PERF_EF_UPDATE);
4105 
4106                 local64_set(&hwc->period_left, 0);
4107 
4108                 if (disable)
4109                         event->pmu->start(event, PERF_EF_RELOAD);
4110         }
4111 }
4112 
4113 /*
4114  * combine freq adjustment with unthrottling to avoid two passes over the
4115  * events. At the same time, make sure, having freq events does not change
4116  * the rate of unthrottling as that would introduce bias.
4117  */
4118 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
4119                                            int needs_unthr)
4120 {
4121         struct perf_event *event;
4122         struct hw_perf_event *hwc;
4123         u64 now, period = TICK_NSEC;
4124         s64 delta;
4125 
4126         /*
4127          * only need to iterate over all events iff:
4128          * - context have events in frequency mode (needs freq adjust)
4129          * - there are events to unthrottle on this cpu
4130          */
4131         if (!(ctx->nr_freq || needs_unthr))
4132                 return;
4133 
4134         raw_spin_lock(&ctx->lock);
4135         perf_pmu_disable(ctx->pmu);
4136 
4137         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
4138                 if (event->state != PERF_EVENT_STATE_ACTIVE)
4139                         continue;
4140 
4141                 if (!event_filter_match(event))
4142                         continue;
4143 
4144                 perf_pmu_disable(event->pmu);
4145 
4146                 hwc = &event->hw;
4147 
4148                 if (hwc->interrupts == MAX_INTERRUPTS) {
4149                         hwc->interrupts = 0;
4150                         perf_log_throttle(event, 1);
4151                         event->pmu->start(event, 0);
4152                 }
4153 
4154                 if (!event->attr.freq || !event->attr.sample_freq)
4155                         goto next;
4156 
4157                 /*
4158                  * stop the event and update event->count
4159                  */
4160                 event->pmu->stop(event, PERF_EF_UPDATE);
4161 
4162                 now = local64_read(&event->count);
4163                 delta = now - hwc->freq_count_stamp;
4164                 hwc->freq_count_stamp = now;
4165 
4166                 /*
4167                  * restart the event
4168                  * reload only if value has changed
4169                  * we have stopped the event so tell that
4170                  * to perf_adjust_period() to avoid stopping it
4171                  * twice.
4172                  */
4173                 if (delta > 0)
4174                         perf_adjust_period(event, period, delta, false);
4175 
4176                 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
4177         next:
4178                 perf_pmu_enable(event->pmu);
4179         }
4180 
4181         perf_pmu_enable(ctx->pmu);
4182         raw_spin_unlock(&ctx->lock);
4183 }
4184 
4185 /*
4186  * Move @event to the tail of the @ctx's elegible events.
4187  */
4188 static void rotate_ctx(struct perf_event_context *ctx, struct perf_event *event)
4189 {
4190         /*
4191          * Rotate the first entry last of non-pinned groups. Rotation might be
4192          * disabled by the inheritance code.
4193          */
4194         if (ctx->rotate_disable)
4195                 return;
4196 
4197         perf_event_groups_delete(&ctx->flexible_groups, event);
4198         perf_event_groups_insert(&ctx->flexible_groups, event);
4199 }
4200 
4201 /* pick an event from the flexible_groups to rotate */
4202 static inline struct perf_event *
4203 ctx_event_to_rotate(struct perf_event_context *ctx)
4204 {
4205         struct perf_event *event;
4206 
4207         /* pick the first active flexible event */
4208         event = list_first_entry_or_null(&ctx->flexible_active,
4209                                          struct perf_event, active_list);
4210 
4211         /* if no active flexible event, pick the first event */
4212         if (!event) {
4213                 event = rb_entry_safe(rb_first(&ctx->flexible_groups.tree),
4214                                       typeof(*event), group_node);
4215         }
4216 
4217         /*
4218          * Unconditionally clear rotate_necessary; if ctx_flexible_sched_in()
4219          * finds there are unschedulable events, it will set it again.
4220          */
4221         ctx->rotate_necessary = 0;
4222 
4223         return event;
4224 }
4225 
4226 static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
4227 {
4228         struct perf_event *cpu_event = NULL, *task_event = NULL;
4229         struct perf_event_context *task_ctx = NULL;
4230         int cpu_rotate, task_rotate;
4231 
4232         /*
4233          * Since we run this from IRQ context, nobody can install new
4234          * events, thus the event count values are stable.
4235          */
4236 
4237         cpu_rotate = cpuctx->ctx.rotate_necessary;
4238         task_ctx = cpuctx->task_ctx;
4239         task_rotate = task_ctx ? task_ctx->rotate_necessary : 0;
4240 
4241         if (!(cpu_rotate || task_rotate))
4242                 return false;
4243 
4244         perf_ctx_lock(cpuctx, cpuctx->task_ctx);
4245         perf_pmu_disable(cpuctx->ctx.pmu);
4246 
4247         if (task_rotate)
4248                 task_event = ctx_event_to_rotate(task_ctx);
4249         if (cpu_rotate)
4250                 cpu_event = ctx_event_to_rotate(&cpuctx->ctx);
4251 
4252         /*
4253          * As per the order given at ctx_resched() first 'pop' task flexible
4254          * and then, if needed CPU flexible.
4255          */
4256         if (task_event || (task_ctx && cpu_event))
4257                 ctx_sched_out(task_ctx, cpuctx, EVENT_FLEXIBLE);
4258         if (cpu_event)
4259                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
4260 
4261         if (task_event)
4262                 rotate_ctx(task_ctx, task_event);
4263         if (cpu_event)
4264                 rotate_ctx(&cpuctx->ctx, cpu_event);
4265 
4266         perf_event_sched_in(cpuctx, task_ctx, current);
4267 
4268         perf_pmu_enable(cpuctx->ctx.pmu);
4269         perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
4270 
4271         return true;
4272 }
4273 
4274 void perf_event_task_tick(void)
4275 {
4276         struct list_head *head = this_cpu_ptr(&active_ctx_list);
4277         struct perf_event_context *ctx, *tmp;
4278         int throttled;
4279 
4280         lockdep_assert_irqs_disabled();
4281 
4282         __this_cpu_inc(perf_throttled_seq);
4283         throttled = __this_cpu_xchg(perf_throttled_count, 0);
4284         tick_dep_clear_cpu(smp_processor_id(), TICK_DEP_BIT_PERF_EVENTS);
4285 
4286         list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
4287                 perf_adjust_freq_unthr_context(ctx, throttled);
4288 }
4289 
4290 static int event_enable_on_exec(struct perf_event *event,
4291                                 struct perf_event_context *ctx)
4292 {
4293         if (!event->attr.enable_on_exec)
4294                 return 0;
4295 
4296         event->attr.enable_on_exec = 0;
4297         if (event->state >= PERF_EVENT_STATE_INACTIVE)
4298                 return 0;
4299 
4300         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
4301 
4302         return 1;
4303 }
4304 
4305 /*
4306  * Enable all of a task's events that have been marked enable-on-exec.
4307  * This expects task == current.
4308  */
4309 static void perf_event_enable_on_exec(int ctxn)
4310 {
4311         struct perf_event_context *ctx, *clone_ctx = NULL;
4312         enum event_type_t event_type = 0;
4313         struct perf_cpu_context *cpuctx;
4314         struct perf_event *event;
4315         unsigned long flags;
4316         int enabled = 0;
4317 
4318         local_irq_save(flags);
4319         ctx = current->perf_event_ctxp[ctxn];
4320         if (!ctx || !ctx->nr_events)
4321                 goto out;
4322 
4323         cpuctx = __get_cpu_context(ctx);
4324         perf_ctx_lock(cpuctx, ctx);
4325         ctx_sched_out(ctx, cpuctx, EVENT_TIME);
4326         list_for_each_entry(event, &ctx->event_list, event_entry) {
4327                 enabled |= event_enable_on_exec(event, ctx);
4328                 event_type |= get_event_type(event);
4329         }
4330 
4331         /*
4332          * Unclone and reschedule this context if we enabled any event.
4333          */
4334         if (enabled) {
4335                 clone_ctx = unclone_ctx(ctx);
4336                 ctx_resched(cpuctx, ctx, event_type);
4337         } else {
4338                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
4339         }
4340         perf_ctx_unlock(cpuctx, ctx);
4341 
4342 out:
4343         local_irq_restore(flags);
4344 
4345         if (clone_ctx)
4346                 put_ctx(clone_ctx);
4347 }
4348 
4349 static void perf_remove_from_owner(struct perf_event *event);
4350 static void perf_event_exit_event(struct perf_event *event,
4351                                   struct perf_event_context *ctx);
4352 
4353 /*
4354  * Removes all events from the current task that have been marked
4355  * remove-on-exec, and feeds their values back to parent events.
4356  */
4357 static void perf_event_remove_on_exec(int ctxn)
4358 {
4359         struct perf_event_context *ctx, *clone_ctx = NULL;
4360         struct perf_event *event, *next;
4361         LIST_HEAD(free_list);
4362         unsigned long flags;
4363         bool modified = false;
4364 
4365         ctx = perf_pin_task_context(current, ctxn);
4366         if (!ctx)
4367                 return;
4368 
4369         mutex_lock(&ctx->mutex);
4370 
4371         if (WARN_ON_ONCE(ctx->task != current))
4372                 goto unlock;
4373 
4374         list_for_each_entry_safe(event, next, &ctx->event_list, event_entry) {
4375                 if (!event->attr.remove_on_exec)
4376                         continue;
4377 
4378                 if (!is_kernel_event(event))
4379                         perf_remove_from_owner(event);
4380 
4381                 modified = true;
4382 
4383                 perf_event_exit_event(event, ctx);
4384         }
4385 
4386         raw_spin_lock_irqsave(&ctx->lock, flags);
4387         if (modified)
4388                 clone_ctx = unclone_ctx(ctx);
4389         --ctx->pin_count;
4390         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4391 
4392 unlock:
4393         mutex_unlock(&ctx->mutex);
4394 
4395         put_ctx(ctx);
4396         if (clone_ctx)
4397                 put_ctx(clone_ctx);
4398 }
4399 
4400 struct perf_read_data {
4401         struct perf_event *event;
4402         bool group;
4403         int ret;
4404 };
4405 
4406 static int __perf_event_read_cpu(struct perf_event *event, int event_cpu)
4407 {
4408         u16 local_pkg, event_pkg;
4409 
4410         if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) {
4411                 int local_cpu = smp_processor_id();
4412 
4413                 event_pkg = topology_physical_package_id(event_cpu);
4414                 local_pkg = topology_physical_package_id(local_cpu);
4415 
4416                 if (event_pkg == local_pkg)
4417                         return local_cpu;
4418         }
4419 
4420         return event_cpu;
4421 }
4422 
4423 /*
4424  * Cross CPU call to read the hardware event
4425  */
4426 static void __perf_event_read(void *info)
4427 {
4428         struct perf_read_data *data = info;
4429         struct perf_event *sub, *event = data->event;
4430         struct perf_event_context *ctx = event->ctx;
4431         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
4432         struct pmu *pmu = event->pmu;
4433 
4434         /*
4435          * If this is a task context, we need to check whether it is
4436          * the current task context of this cpu.  If not it has been
4437          * scheduled out before the smp call arrived.  In that case
4438          * event->count would have been updated to a recent sample
4439          * when the event was scheduled out.
4440          */
4441         if (ctx->task && cpuctx->task_ctx != ctx)
4442                 return;
4443 
4444         raw_spin_lock(&ctx->lock);
4445         if (ctx->is_active & EVENT_TIME) {
4446                 update_context_time(ctx);
4447                 update_cgrp_time_from_event(event);
4448         }
4449 
4450         perf_event_update_time(event);
4451         if (data->group)
4452                 perf_event_update_sibling_time(event);
4453 
4454         if (event->state != PERF_EVENT_STATE_ACTIVE)
4455                 goto unlock;
4456 
4457         if (!data->group) {
4458                 pmu->read(event);
4459                 data->ret = 0;
4460                 goto unlock;
4461         }
4462 
4463         pmu->start_txn(pmu, PERF_PMU_TXN_READ);
4464 
4465         pmu->read(event);
4466 
4467         for_each_sibling_event(sub, event) {
4468                 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
4469                         /*
4470                          * Use sibling's PMU rather than @event's since
4471                          * sibling could be on different (eg: software) PMU.
4472                          */
4473                         sub->pmu->read(sub);
4474                 }
4475         }
4476 
4477         data->ret = pmu->commit_txn(pmu);
4478 
4479 unlock:
4480         raw_spin_unlock(&ctx->lock);
4481 }
4482 
4483 static inline u64 perf_event_count(struct perf_event *event)
4484 {
4485         return local64_read(&event->count) + atomic64_read(&event->child_count);
4486 }
4487 
4488 static void calc_timer_values(struct perf_event *event,
4489                                 u64 *now,
4490                                 u64 *enabled,
4491                                 u64 *running)
4492 {
4493         u64 ctx_time;
4494 
4495         *now = perf_clock();
4496         ctx_time = perf_event_time_now(event, *now);
4497         __perf_update_times(event, ctx_time, enabled, running);
4498 }
4499 
4500 /*
4501  * NMI-safe method to read a local event, that is an event that
4502  * is:
4503  *   - either for the current task, or for this CPU
4504  *   - does not have inherit set, for inherited task events
4505  *     will not be local and we cannot read them atomically
4506  *   - must not have a pmu::count method
4507  */
4508 int perf_event_read_local(struct perf_event *event, u64 *value,
4509                           u64 *enabled, u64 *running)
4510 {
4511         unsigned long flags;
4512         int ret = 0;
4513 
4514         /*
4515          * Disabling interrupts avoids all counter scheduling (context
4516          * switches, timer based rotation and IPIs).
4517          */
4518         local_irq_save(flags);
4519 
4520         /*
4521          * It must not be an event with inherit set, we cannot read
4522          * all child counters from atomic context.
4523          */
4524         if (event->attr.inherit) {
4525                 ret = -EOPNOTSUPP;
4526                 goto out;
4527         }
4528 
4529         /* If this is a per-task event, it must be for current */
4530         if ((event->attach_state & PERF_ATTACH_TASK) &&
4531             event->hw.target != current) {
4532                 ret = -EINVAL;
4533                 goto out;
4534         }
4535 
4536         /* If this is a per-CPU event, it must be for this CPU */
4537         if (!(event->attach_state & PERF_ATTACH_TASK) &&
4538             event->cpu != smp_processor_id()) {
4539                 ret = -EINVAL;
4540                 goto out;
4541         }
4542 
4543         /* If this is a pinned event it must be running on this CPU */
4544         if (event->attr.pinned && event->oncpu != smp_processor_id()) {
4545                 ret = -EBUSY;
4546                 goto out;
4547         }
4548 
4549         /*
4550          * If the event is currently on this CPU, its either a per-task event,
4551          * or local to this CPU. Furthermore it means its ACTIVE (otherwise
4552          * oncpu == -1).
4553          */
4554         if (event->oncpu == smp_processor_id())
4555                 event->pmu->read(event);
4556 
4557         *value = local64_read(&event->count);
4558         if (enabled || running) {
4559                 u64 __enabled, __running, __now;;
4560 
4561                 calc_timer_values(event, &__now, &__enabled, &__running);
4562                 if (enabled)
4563                         *enabled = __enabled;
4564                 if (running)
4565                         *running = __running;
4566         }
4567 out:
4568         local_irq_restore(flags);
4569 
4570         return ret;
4571 }
4572 
4573 static int perf_event_read(struct perf_event *event, bool group)
4574 {
4575         enum perf_event_state state = READ_ONCE(event->state);
4576         int event_cpu, ret = 0;
4577 
4578         /*
4579          * If event is enabled and currently active on a CPU, update the
4580          * value in the event structure:
4581          */
4582 again:
4583         if (state == PERF_EVENT_STATE_ACTIVE) {
4584                 struct perf_read_data data;
4585 
4586                 /*
4587                  * Orders the ->state and ->oncpu loads such that if we see
4588                  * ACTIVE we must also see the right ->oncpu.
4589                  *
4590                  * Matches the smp_wmb() from event_sched_in().
4591                  */
4592                 smp_rmb();
4593 
4594                 event_cpu = READ_ONCE(event->oncpu);
4595                 if ((unsigned)event_cpu >= nr_cpu_ids)
4596                         return 0;
4597 
4598                 data = (struct perf_read_data){
4599                         .event = event,
4600                         .group = group,
4601                         .ret = 0,
4602                 };
4603 
4604                 preempt_disable();
4605                 event_cpu = __perf_event_read_cpu(event, event_cpu);
4606 
4607                 /*
4608                  * Purposely ignore the smp_call_function_single() return
4609                  * value.
4610                  *
4611                  * If event_cpu isn't a valid CPU it means the event got
4612                  * scheduled out and that will have updated the event count.
4613                  *
4614                  * Therefore, either way, we'll have an up-to-date event count
4615                  * after this.
4616                  */
4617                 (void)smp_call_function_single(event_cpu, __perf_event_read, &data, 1);
4618                 preempt_enable();
4619                 ret = data.ret;
4620 
4621         } else if (state == PERF_EVENT_STATE_INACTIVE) {
4622                 struct perf_event_context *ctx = event->ctx;
4623                 unsigned long flags;
4624 
4625                 raw_spin_lock_irqsave(&ctx->lock, flags);
4626                 state = event->state;
4627                 if (state != PERF_EVENT_STATE_INACTIVE) {
4628                         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4629                         goto again;
4630                 }
4631 
4632                 /*
4633                  * May read while context is not active (e.g., thread is
4634                  * blocked), in that case we cannot update context time
4635                  */
4636                 if (ctx->is_active & EVENT_TIME) {
4637                         update_context_time(ctx);
4638                         update_cgrp_time_from_event(event);
4639                 }
4640 
4641                 perf_event_update_time(event);
4642                 if (group)
4643                         perf_event_update_sibling_time(event);
4644                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4645         }
4646 
4647         return ret;
4648 }
4649 
4650 /*
4651  * Initialize the perf_event context in a task_struct:
4652  */
4653 static void __perf_event_init_context(struct perf_event_context *ctx)
4654 {
4655         raw_spin_lock_init(&ctx->lock);
4656         mutex_init(&ctx->mutex);
4657         INIT_LIST_HEAD(&ctx->active_ctx_list);
4658         perf_event_groups_init(&ctx->pinned_groups);
4659         perf_event_groups_init(&ctx->flexible_groups);
4660         INIT_LIST_HEAD(&ctx->event_list);
4661         INIT_LIST_HEAD(&ctx->pinned_active);
4662         INIT_LIST_HEAD(&ctx->flexible_active);
4663         refcount_set(&ctx->refcount, 1);
4664 }
4665 
4666 static struct perf_event_context *
4667 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
4668 {
4669         struct perf_event_context *ctx;
4670 
4671         ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4672         if (!ctx)
4673                 return NULL;
4674 
4675         __perf_event_init_context(ctx);
4676         if (task)
4677                 ctx->task = get_task_struct(task);
4678         ctx->pmu = pmu;
4679 
4680         return ctx;
4681 }
4682 
4683 static struct task_struct *
4684 find_lively_task_by_vpid(pid_t vpid)
4685 {
4686         struct task_struct *task;
4687 
4688         rcu_read_lock();
4689         if (!vpid)
4690                 task = current;
4691         else
4692                 task = find_task_by_vpid(vpid);
4693         if (task)
4694                 get_task_struct(task);
4695         rcu_read_unlock();
4696 
4697         if (!task)
4698                 return ERR_PTR(-ESRCH);
4699 
4700         return task;
4701 }
4702 
4703 /*
4704  * Returns a matching context with refcount and pincount.
4705  */
4706 static struct perf_event_context *
4707 find_get_context(struct pmu *pmu, struct task_struct *task,
4708                 struct perf_event *event)
4709 {
4710         struct perf_event_context *ctx, *clone_ctx = NULL;
4711         struct perf_cpu_context *cpuctx;
4712         void *task_ctx_data = NULL;
4713         unsigned long flags;
4714         int ctxn, err;
4715         int cpu = event->cpu;
4716 
4717         if (!task) {
4718                 /* Must be root to operate on a CPU event: */
4719                 err = perf_allow_cpu(&event->attr);
4720                 if (err)
4721                         return ERR_PTR(err);
4722 
4723                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
4724                 ctx = &cpuctx->ctx;
4725                 get_ctx(ctx);
4726                 raw_spin_lock_irqsave(&ctx->lock, flags);
4727                 ++ctx->pin_count;
4728                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4729 
4730                 return ctx;
4731         }
4732 
4733         err = -EINVAL;
4734         ctxn = pmu->task_ctx_nr;
4735         if (ctxn < 0)
4736                 goto errout;
4737 
4738         if (event->attach_state & PERF_ATTACH_TASK_DATA) {
4739                 task_ctx_data = alloc_task_ctx_data(pmu);
4740                 if (!task_ctx_data) {
4741                         err = -ENOMEM;
4742                         goto errout;
4743                 }
4744         }
4745 
4746 retry:
4747         ctx = perf_lock_task_context(task, ctxn, &flags);
4748         if (ctx) {
4749                 clone_ctx = unclone_ctx(ctx);
4750                 ++ctx->pin_count;
4751 
4752                 if (task_ctx_data && !ctx->task_ctx_data) {
4753                         ctx->task_ctx_data = task_ctx_data;
4754                         task_ctx_data = NULL;
4755                 }
4756                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4757 
4758                 if (clone_ctx)
4759                         put_ctx(clone_ctx);
4760         } else {
4761                 ctx = alloc_perf_context(pmu, task);
4762                 err = -ENOMEM;
4763                 if (!ctx)
4764                         goto errout;
4765 
4766                 if (task_ctx_data) {
4767                         ctx->task_ctx_data = task_ctx_data;
4768                         task_ctx_data = NULL;
4769                 }
4770 
4771                 err = 0;
4772                 mutex_lock(&task->perf_event_mutex);
4773                 /*
4774                  * If it has already passed perf_event_exit_task().
4775                  * we must see PF_EXITING, it takes this mutex too.
4776                  */
4777                 if (task->flags & PF_EXITING)
4778                         err = -ESRCH;
4779                 else if (task->perf_event_ctxp[ctxn])
4780                         err = -EAGAIN;
4781                 else {
4782                         get_ctx(ctx);
4783                         ++ctx->pin_count;
4784                         rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
4785                 }
4786                 mutex_unlock(&task->perf_event_mutex);
4787 
4788                 if (unlikely(err)) {
4789                         put_ctx(ctx);
4790 
4791                         if (err == -EAGAIN)
4792                                 goto retry;
4793                         goto errout;
4794                 }
4795         }
4796 
4797         free_task_ctx_data(pmu, task_ctx_data);
4798         return ctx;
4799 
4800 errout:
4801         free_task_ctx_data(pmu, task_ctx_data);
4802         return ERR_PTR(err);
4803 }
4804 
4805 static void perf_event_free_filter(struct perf_event *event);
4806 
4807 static void free_event_rcu(struct rcu_head *head)
4808 {
4809         struct perf_event *event;
4810 
4811         event = container_of(head, struct perf_event, rcu_head);
4812         if (event->ns)
4813                 put_pid_ns(event->ns);
4814         perf_event_free_filter(event);
4815         kmem_cache_free(perf_event_cache, event);
4816 }
4817 
4818 static void ring_buffer_attach(struct perf_event *event,
4819                                struct perf_buffer *rb);
4820 
4821 static void detach_sb_event(struct perf_event *event)
4822 {
4823         struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu);
4824 
4825         raw_spin_lock(&pel->lock);
4826         list_del_rcu(&event->sb_list);
4827         raw_spin_unlock(&pel->lock);
4828 }
4829 
4830 static bool is_sb_event(struct perf_event *event)
4831 {
4832         struct perf_event_attr *attr = &event->attr;
4833 
4834         if (event->parent)
4835                 return false;
4836 
4837         if (event->attach_state & PERF_ATTACH_TASK)
4838                 return false;
4839 
4840         if (attr->mmap || attr->mmap_data || attr->mmap2 ||
4841             attr->comm || attr->comm_exec ||
4842             attr->task || attr->ksymbol ||
4843             attr->context_switch || attr->text_poke ||
4844             attr->bpf_event)
4845                 return true;
4846         return false;
4847 }
4848 
4849 static void unaccount_pmu_sb_event(struct perf_event *event)
4850 {
4851         if (is_sb_event(event))
4852                 detach_sb_event(event);
4853 }
4854 
4855 static void unaccount_event_cpu(struct perf_event *event, int cpu)
4856 {
4857         if (event->parent)
4858                 return;
4859 
4860         if (is_cgroup_event(event))
4861                 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
4862 }
4863 
4864 #ifdef CONFIG_NO_HZ_FULL
4865 static DEFINE_SPINLOCK(nr_freq_lock);
4866 #endif
4867 
4868 static void unaccount_freq_event_nohz(void)
4869 {
4870 #ifdef CONFIG_NO_HZ_FULL
4871         spin_lock(&nr_freq_lock);
4872         if (atomic_dec_and_test(&nr_freq_events))
4873                 tick_nohz_dep_clear(TICK_DEP_BIT_PERF_EVENTS);
4874         spin_unlock(&nr_freq_lock);
4875 #endif
4876 }
4877 
4878 static void unaccount_freq_event(void)
4879 {
4880         if (tick_nohz_full_enabled())
4881                 unaccount_freq_event_nohz();
4882         else
4883                 atomic_dec(&nr_freq_events);
4884 }
4885 
4886 static void unaccount_event(struct perf_event *event)
4887 {
4888         bool dec = false;
4889 
4890         if (event->parent)
4891                 return;
4892 
4893         if (event->attach_state & (PERF_ATTACH_TASK | PERF_ATTACH_SCHED_CB))
4894                 dec = true;
4895         if (event->attr.mmap || event->attr.mmap_data)
4896                 atomic_dec(&nr_mmap_events);
4897         if (event->attr.build_id)
4898                 atomic_dec(&nr_build_id_events);
4899         if (event->attr.comm)
4900                 atomic_dec(&nr_comm_events);
4901         if (event->attr.namespaces)
4902                 atomic_dec(&nr_namespaces_events);
4903         if (event->attr.cgroup)
4904                 atomic_dec(&nr_cgroup_events);
4905         if (event->attr.task)
4906                 atomic_dec(&nr_task_events);
4907         if (event->attr.freq)
4908                 unaccount_freq_event();
4909         if (event->attr.context_switch) {
4910                 dec = true;
4911                 atomic_dec(&nr_switch_events);
4912         }
4913         if (is_cgroup_event(event))
4914                 dec = true;
4915         if (has_branch_stack(event))
4916                 dec = true;
4917         if (event->attr.ksymbol)
4918                 atomic_dec(&nr_ksymbol_events);
4919         if (event->attr.bpf_event)
4920                 atomic_dec(&nr_bpf_events);
4921         if (event->attr.text_poke)
4922                 atomic_dec(&nr_text_poke_events);
4923 
4924         if (dec) {
4925                 if (!atomic_add_unless(&perf_sched_count, -1, 1))
4926                         schedule_delayed_work(&perf_sched_work, HZ);
4927         }
4928 
4929         unaccount_event_cpu(event, event->cpu);
4930 
4931         unaccount_pmu_sb_event(event);
4932 }
4933 
4934 static void perf_sched_delayed(struct work_struct *work)
4935 {
4936         mutex_lock(&perf_sched_mutex);
4937         if (atomic_dec_and_test(&perf_sched_count))
4938                 static_branch_disable(&perf_sched_events);
4939         mutex_unlock(&perf_sched_mutex);
4940 }
4941 
4942 /*
4943  * The following implement mutual exclusion of events on "exclusive" pmus
4944  * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
4945  * at a time, so we disallow creating events that might conflict, namely:
4946  *
4947  *  1) cpu-wide events in the presence of per-task events,
4948  *  2) per-task events in the presence of cpu-wide events,
4949  *  3) two matching events on the same context.
4950  *
4951  * The former two cases are handled in the allocation path (perf_event_alloc(),
4952  * _free_event()), the latter -- before the first perf_install_in_context().
4953  */
4954 static int exclusive_event_init(struct perf_event *event)
4955 {
4956         struct pmu *pmu = event->pmu;
4957 
4958         if (!is_exclusive_pmu(pmu))
4959                 return 0;
4960 
4961         /*
4962          * Prevent co-existence of per-task and cpu-wide events on the
4963          * same exclusive pmu.
4964          *
4965          * Negative pmu::exclusive_cnt means there are cpu-wide
4966          * events on this "exclusive" pmu, positive means there are
4967          * per-task events.
4968          *
4969          * Since this is called in perf_event_alloc() path, event::ctx
4970          * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
4971          * to mean "per-task event", because unlike other attach states it
4972          * never gets cleared.
4973          */
4974         if (event->attach_state & PERF_ATTACH_TASK) {
4975                 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
4976                         return -EBUSY;
4977         } else {
4978                 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
4979                         return -EBUSY;
4980         }
4981 
4982         return 0;
4983 }
4984 
4985 static void exclusive_event_destroy(struct perf_event *event)
4986 {
4987         struct pmu *pmu = event->pmu;
4988 
4989         if (!is_exclusive_pmu(pmu))
4990                 return;
4991 
4992         /* see comment in exclusive_event_init() */
4993         if (event->attach_state & PERF_ATTACH_TASK)
4994                 atomic_dec(&pmu->exclusive_cnt);
4995         else
4996                 atomic_inc(&pmu->exclusive_cnt);
4997 }
4998 
4999 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
5000 {
5001         if ((e1->pmu == e2->pmu) &&
5002             (e1->cpu == e2->cpu ||
5003              e1->cpu == -1 ||
5004              e2->cpu == -1))
5005                 return true;
5006         return false;
5007 }
5008 
5009 static bool exclusive_event_installable(struct perf_event *event,
5010                                         struct perf_event_context *ctx)
5011 {
5012         struct perf_event *iter_event;
5013         struct pmu *pmu = event->pmu;
5014 
5015         lockdep_assert_held(&ctx->mutex);
5016 
5017         if (!is_exclusive_pmu(pmu))
5018                 return true;
5019 
5020         list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
5021                 if (exclusive_event_match(iter_event, event))
5022                         return false;
5023         }
5024 
5025         return true;
5026 }
5027 
5028 static void perf_addr_filters_splice(struct perf_event *event,
5029                                        struct list_head *head);
5030 
5031 static void _free_event(struct perf_event *event)
5032 {
5033         irq_work_sync(&event->pending);
5034 
5035         unaccount_event(event);
5036 
5037         security_perf_event_free(event);
5038 
5039         if (event->rb) {
5040                 /*
5041                  * Can happen when we close an event with re-directed output.
5042                  *
5043                  * Since we have a 0 refcount, perf_mmap_close() will skip
5044                  * over us; possibly making our ring_buffer_put() the last.
5045                  */
5046                 mutex_lock(&event->mmap_mutex);
5047                 ring_buffer_attach(event, NULL);
5048                 mutex_unlock(&event->mmap_mutex);
5049         }
5050 
5051         if (is_cgroup_event(event))
5052                 perf_detach_cgroup(event);
5053 
5054         if (!event->parent) {
5055                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
5056                         put_callchain_buffers();
5057         }
5058 
5059         perf_event_free_bpf_prog(event);
5060         perf_addr_filters_splice(event, NULL);
5061         kfree(event->addr_filter_ranges);
5062 
5063         if (event->destroy)
5064                 event->destroy(event);
5065 
5066         /*
5067          * Must be after ->destroy(), due to uprobe_perf_close() using
5068          * hw.target.
5069          */
5070         if (event->hw.target)
5071                 put_task_struct(event->hw.target);
5072 
5073         /*
5074          * perf_event_free_task() relies on put_ctx() being 'last', in particular
5075          * all task references must be cleaned up.
5076          */
5077         if (event->ctx)
5078                 put_ctx(event->ctx);
5079 
5080         exclusive_event_destroy(event);
5081         module_put(event->pmu->module);
5082 
5083         call_rcu(&event->rcu_head, free_event_rcu);
5084 }
5085 
5086 /*
5087  * Used to free events which have a known refcount of 1, such as in error paths
5088  * where the event isn't exposed yet and inherited events.
5089  */
5090 static void free_event(struct perf_event *event)
5091 {
5092         if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
5093                                 "unexpected event refcount: %ld; ptr=%p\n",
5094                                 atomic_long_read(&event->refcount), event)) {
5095                 /* leak to avoid use-after-free */
5096                 return;
5097         }
5098 
5099         _free_event(event);
5100 }
5101 
5102 /*
5103  * Remove user event from the owner task.
5104  */
5105 static void perf_remove_from_owner(struct perf_event *event)
5106 {
5107         struct task_struct *owner;
5108 
5109         rcu_read_lock();
5110         /*
5111          * Matches the smp_store_release() in perf_event_exit_task(). If we
5112          * observe !owner it means the list deletion is complete and we can
5113          * indeed free this event, otherwise we need to serialize on
5114          * owner->perf_event_mutex.
5115          */
5116         owner = READ_ONCE(event->owner);
5117         if (owner) {
5118                 /*
5119                  * Since delayed_put_task_struct() also drops the last
5120                  * task reference we can safely take a new reference
5121                  * while holding the rcu_read_lock().
5122                  */
5123                 get_task_struct(owner);
5124         }
5125         rcu_read_unlock();
5126 
5127         if (owner) {
5128                 /*
5129                  * If we're here through perf_event_exit_task() we're already
5130                  * holding ctx->mutex which would be an inversion wrt. the
5131                  * normal lock order.
5132                  *
5133                  * However we can safely take this lock because its the child
5134                  * ctx->mutex.
5135                  */
5136                 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
5137 
5138                 /*
5139                  * We have to re-check the event->owner field, if it is cleared
5140                  * we raced with perf_event_exit_task(), acquiring the mutex
5141                  * ensured they're done, and we can proceed with freeing the
5142                  * event.
5143                  */
5144                 if (event->owner) {
5145                         list_del_init(&event->owner_entry);
5146                         smp_store_release(&event->owner, NULL);
5147                 }
5148                 mutex_unlock(&owner->perf_event_mutex);
5149                 put_task_struct(owner);
5150         }
5151 }
5152 
5153 static void put_event(struct perf_event *event)
5154 {
5155         if (!atomic_long_dec_and_test(&event->refcount))
5156                 return;
5157 
5158         _free_event(event);
5159 }
5160 
5161 /*
5162  * Kill an event dead; while event:refcount will preserve the event
5163  * object, it will not preserve its functionality. Once the last 'user'
5164  * gives up the object, we'll destroy the thing.
5165  */
5166 int perf_event_release_kernel(struct perf_event *event)
5167 {
5168         struct perf_event_context *ctx = event->ctx;
5169         struct perf_event *child, *tmp;
5170         LIST_HEAD(free_list);
5171 
5172         /*
5173          * If we got here through err_file: fput(event_file); we will not have
5174          * attached to a context yet.
5175          */
5176         if (!ctx) {
5177                 WARN_ON_ONCE(event->attach_state &
5178                                 (PERF_ATTACH_CONTEXT|PERF_ATTACH_GROUP));
5179                 goto no_ctx;
5180         }
5181 
5182         if (!is_kernel_event(event))
5183                 perf_remove_from_owner(event);
5184 
5185         ctx = perf_event_ctx_lock(event);
5186         WARN_ON_ONCE(ctx->parent_ctx);
5187         perf_remove_from_context(event, DETACH_GROUP);
5188 
5189         raw_spin_lock_irq(&ctx->lock);
5190         /*
5191          * Mark this event as STATE_DEAD, there is no external reference to it
5192          * anymore.
5193          *
5194          * Anybody acquiring event->child_mutex after the below loop _must_
5195          * also see this, most importantly inherit_event() which will avoid
5196          * placing more children on the list.
5197          *
5198          * Thus this guarantees that we will in fact observe and kill _ALL_
5199          * child events.
5200          */
5201         event->state = PERF_EVENT_STATE_DEAD;
5202         raw_spin_unlock_irq(&ctx->lock);
5203 
5204         perf_event_ctx_unlock(event, ctx);
5205 
5206 again:
5207         mutex_lock(&event->child_mutex);
5208         list_for_each_entry(child, &event->child_list, child_list) {
5209 
5210                 /*
5211                  * Cannot change, child events are not migrated, see the
5212                  * comment with perf_event_ctx_lock_nested().
5213                  */
5214                 ctx = READ_ONCE(child->ctx);
5215                 /*
5216                  * Since child_mutex nests inside ctx::mutex, we must jump
5217                  * through hoops. We start by grabbing a reference on the ctx.
5218                  *
5219                  * Since the event cannot get freed while we hold the
5220                  * child_mutex, the context must also exist and have a !0
5221                  * reference count.
5222                  */
5223                 get_ctx(ctx);
5224 
5225                 /*
5226                  * Now that we have a ctx ref, we can drop child_mutex, and
5227                  * acquire ctx::mutex without fear of it going away. Then we
5228                  * can re-acquire child_mutex.
5229                  */
5230                 mutex_unlock(&event->child_mutex);
5231                 mutex_lock(&ctx->mutex);
5232                 mutex_lock(&event->child_mutex);
5233 
5234                 /*
5235                  * Now that we hold ctx::mutex and child_mutex, revalidate our
5236                  * state, if child is still the first entry, it didn't get freed
5237                  * and we can continue doing so.
5238                  */
5239                 tmp = list_first_entry_or_null(&event->child_list,
5240                                                struct perf_event, child_list);
5241                 if (tmp == child) {
5242                         perf_remove_from_context(child, DETACH_GROUP);
5243                         list_move(&child->child_list, &free_list);
5244                         /*
5245                          * This matches the refcount bump in inherit_event();
5246                          * this can't be the last reference.
5247                          */
5248                         put_event(event);
5249                 }
5250 
5251                 mutex_unlock(&event->child_mutex);
5252                 mutex_unlock(&ctx->mutex);
5253                 put_ctx(ctx);
5254                 goto again;
5255         }
5256         mutex_unlock(&event->child_mutex);
5257 
5258         list_for_each_entry_safe(child, tmp, &free_list, child_list) {
5259                 void *var = &child->ctx->refcount;
5260 
5261                 list_del(&child->child_list);
5262                 free_event(child);
5263 
5264                 /*
5265                  * Wake any perf_event_free_task() waiting for this event to be
5266                  * freed.
5267                  */
5268                 smp_mb(); /* pairs with wait_var_event() */
5269                 wake_up_var(var);
5270         }
5271 
5272 no_ctx:
5273         put_event(event); /* Must be the 'last' reference */
5274         return 0;
5275 }
5276 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
5277 
5278 /*
5279  * Called when the last reference to the file is gone.
5280  */
5281 static int perf_release(struct inode *inode, struct file *file)
5282 {
5283         perf_event_release_kernel(file->private_data);
5284         return 0;
5285 }
5286 
5287 static u64 __perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
5288 {
5289         struct perf_event *child;
5290         u64 total = 0;
5291 
5292         *enabled = 0;
5293         *running = 0;
5294 
5295         mutex_lock(&event->child_mutex);
5296 
5297         (void)perf_event_read(event, false);
5298         total += perf_event_count(event);
5299 
5300         *enabled += event->total_time_enabled +
5301                         atomic64_read(&event->child_total_time_enabled);
5302         *running += event->total_time_running +
5303                         atomic64_read(&event->child_total_time_running);
5304 
5305         list_for_each_entry(child, &event->child_list, child_list) {
5306                 (void)perf_event_read(child, false);
5307                 total += perf_event_count(child);
5308                 *enabled += child->total_time_enabled;
5309                 *running += child->total_time_running;
5310         }
5311         mutex_unlock(&event->child_mutex);
5312 
5313         return total;
5314 }
5315 
5316 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
5317 {
5318         struct perf_event_context *ctx;
5319         u64 count;
5320 
5321         ctx = perf_event_ctx_lock(event);
5322         count = __perf_event_read_value(event, enabled, running);
5323         perf_event_ctx_unlock(event, ctx);
5324 
5325         return count;
5326 }
5327 EXPORT_SYMBOL_GPL(perf_event_read_value);
5328 
5329 static int __perf_read_group_add(struct perf_event *leader,
5330                                         u64 read_format, u64 *values)
5331 {
5332         struct perf_event_context *ctx = leader->ctx;
5333         struct perf_event *sub;
5334         unsigned long flags;
5335         int n = 1; /* skip @nr */
5336         int ret;
5337 
5338         ret = perf_event_read(leader, true);
5339         if (ret)
5340                 return ret;
5341 
5342         raw_spin_lock_irqsave(&ctx->lock, flags);
5343 
5344         /*
5345          * Since we co-schedule groups, {enabled,running} times of siblings
5346          * will be identical to those of the leader, so we only publish one
5347          * set.
5348          */
5349         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
5350                 values[n++] += leader->total_time_enabled +
5351                         atomic64_read(&leader->child_total_time_enabled);
5352         }
5353 
5354         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
5355                 values[n++] += leader->total_time_running +
5356                         atomic64_read(&leader->child_total_time_running);
5357         }
5358 
5359         /*
5360          * Write {count,id} tuples for every sibling.
5361          */
5362         values[n++] += perf_event_count(leader);
5363         if (read_format & PERF_FORMAT_ID)
5364                 values[n++] = primary_event_id(leader);
5365 
5366         for_each_sibling_event(sub, leader) {
5367                 values[n++] += perf_event_count(sub);
5368                 if (read_format & PERF_FORMAT_ID)
5369                         values[n++] = primary_event_id(sub);
5370         }
5371 
5372         raw_spin_unlock_irqrestore(&ctx->lock, flags);
5373         return 0;
5374 }
5375 
5376 static int perf_read_group(struct perf_event *event,
5377                                    u64 read_format, char __user *buf)
5378 {
5379         struct perf_event *leader = event->group_leader, *child;
5380         struct perf_event_context *ctx = leader->ctx;
5381         int ret;
5382         u64 *values;
5383 
5384         lockdep_assert_held(&ctx->mutex);
5385 
5386         values = kzalloc(event->read_size, GFP_KERNEL);
5387         if (!values)
5388                 return -ENOMEM;
5389 
5390         values[0] = 1 + leader->nr_siblings;
5391 
5392         /*
5393          * By locking the child_mutex of the leader we effectively
5394          * lock the child list of all siblings.. XXX explain how.
5395          */
5396         mutex_lock(&leader->child_mutex);
5397 
5398         ret = __perf_read_group_add(leader, read_format, values);
5399         if (ret)
5400                 goto unlock;
5401 
5402         list_for_each_entry(child, &leader->child_list, child_list) {
5403                 ret = __perf_read_group_add(child, read_format, values);
5404                 if (ret)
5405                         goto unlock;
5406         }
5407 
5408         mutex_unlock(&leader->child_mutex);
5409 
5410         ret = event->read_size;
5411         if (copy_to_user(buf, values, event->read_size))
5412                 ret = -EFAULT;
5413         goto out;
5414 
5415 unlock:
5416         mutex_unlock(&leader->child_mutex);
5417 out:
5418         kfree(values);
5419         return ret;
5420 }
5421 
5422 static int perf_read_one(struct perf_event *event,
5423                                  u64 read_format, char __user *buf)
5424 {
5425         u64 enabled, running;
5426         u64 values[4];
5427         int n = 0;
5428 
5429         values[n++] = __perf_event_read_value(event, &enabled, &running);
5430         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
5431                 values[n++] = enabled;
5432         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
5433                 values[n++] = running;
5434         if (read_format & PERF_FORMAT_ID)
5435                 values[n++] = primary_event_id(event);
5436 
5437         if (copy_to_user(buf, values, n * sizeof(u64)))
5438                 return -EFAULT;
5439 
5440         return n * sizeof(u64);
5441 }
5442 
5443 static bool is_event_hup(struct perf_event *event)
5444 {
5445         bool no_children;
5446 
5447         if (event->state > PERF_EVENT_STATE_EXIT)
5448                 return false;
5449 
5450         mutex_lock(&event->child_mutex);
5451         no_children = list_empty(&event->child_list);
5452         mutex_unlock(&event->child_mutex);
5453         return no_children;
5454 }
5455 
5456 /*
5457  * Read the performance event - simple non blocking version for now
5458  */
5459 static ssize_t
5460 __perf_read(struct perf_event *event, char __user *buf, size_t count)
5461 {
5462         u64 read_format = event->attr.read_format;
5463         int ret;
5464 
5465         /*
5466          * Return end-of-file for a read on an event that is in
5467          * error state (i.e. because it was pinned but it couldn't be
5468          * scheduled on to the CPU at some point).
5469          */
5470         if (event->state == PERF_EVENT_STATE_ERROR)
5471                 return 0;
5472 
5473         if (count < event->read_size)
5474                 return -ENOSPC;
5475 
5476         WARN_ON_ONCE(event->ctx->parent_ctx);
5477         if (read_format & PERF_FORMAT_GROUP)
5478                 ret = perf_read_group(event, read_format, buf);
5479         else
5480                 ret = perf_read_one(event, read_format, buf);
5481 
5482         return ret;
5483 }
5484 
5485 static ssize_t
5486 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
5487 {
5488         struct perf_event *event = file->private_data;
5489         struct perf_event_context *ctx;
5490         int ret;
5491 
5492         ret = security_perf_event_read(event);
5493         if (ret)
5494                 return ret;
5495 
5496         ctx = perf_event_ctx_lock(event);
5497         ret = __perf_read(event, buf, count);
5498         perf_event_ctx_unlock(event, ctx);
5499 
5500         return ret;
5501 }
5502 
5503 static __poll_t perf_poll(struct file *file, poll_table *wait)
5504 {
5505         struct perf_event *event = file->private_data;
5506         struct perf_buffer *rb;
5507         __poll_t events = EPOLLHUP;
5508 
5509         poll_wait(file, &event->waitq, wait);
5510 
5511         if (is_event_hup(event))
5512                 return events;
5513 
5514         /*
5515          * Pin the event->rb by taking event->mmap_mutex; otherwise
5516          * perf_event_set_output() can swizzle our rb and make us miss wakeups.
5517          */
5518         mutex_lock(&event->mmap_mutex);
5519         rb = event->rb;
5520         if (rb)
5521                 events = atomic_xchg(&rb->poll, 0);
5522         mutex_unlock(&event->mmap_mutex);
5523         return events;
5524 }
5525 
5526 static void _perf_event_reset(struct perf_event *event)
5527 {
5528         (void)perf_event_read(event, false);
5529         local64_set(&event->count, 0);
5530         perf_event_update_userpage(event);
5531 }
5532 
5533 /* Assume it's not an event with inherit set. */
5534 u64 perf_event_pause(struct perf_event *event, bool reset)
5535 {
5536         struct perf_event_context *ctx;
5537         u64 count;
5538 
5539         ctx = perf_event_ctx_lock(event);
5540         WARN_ON_ONCE(event->attr.inherit);
5541         _perf_event_disable(event);
5542         count = local64_read(&event->count);
5543         if (reset)
5544                 local64_set(&event->count, 0);
5545         perf_event_ctx_unlock(event, ctx);
5546 
5547         return count;
5548 }
5549 EXPORT_SYMBOL_GPL(perf_event_pause);
5550 
5551 /*
5552  * Holding the top-level event's child_mutex means that any
5553  * descendant process that has inherited this event will block
5554  * in perf_event_exit_event() if it goes to exit, thus satisfying the
5555  * task existence requirements of perf_event_enable/disable.
5556  */
5557 static void perf_event_for_each_child(struct perf_event *event,
5558                                         void (*func)(struct perf_event *))
5559 {
5560         struct perf_event *child;
5561 
5562         WARN_ON_ONCE(event->ctx->parent_ctx);
5563 
5564         mutex_lock(&event->child_mutex);
5565         func(event);
5566         list_for_each_entry(child, &event->child_list, child_list)
5567                 func(child);
5568         mutex_unlock(&event->child_mutex);
5569 }
5570 
5571 static void perf_event_for_each(struct perf_event *event,
5572                                   void (*func)(struct perf_event *))
5573 {
5574         struct perf_event_context *ctx = event->ctx;
5575         struct perf_event *sibling;
5576 
5577         lockdep_assert_held(&ctx->mutex);
5578 
5579         event = event->group_leader;
5580 
5581         perf_event_for_each_child(event, func);
5582         for_each_sibling_event(sibling, event)
5583                 perf_event_for_each_child(sibling, func);
5584 }
5585 
5586 static void __perf_event_period(struct perf_event *event,
5587                                 struct perf_cpu_context *cpuctx,
5588                                 struct perf_event_context *ctx,
5589                                 void *info)
5590 {
5591         u64 value = *((u64 *)info);
5592         bool active;
5593 
5594         if (event->attr.freq) {
5595                 event->attr.sample_freq = value;
5596         } else {
5597                 event->attr.sample_period = value;
5598                 event->hw.sample_period = value;
5599         }
5600 
5601         active = (event->state == PERF_EVENT_STATE_ACTIVE);
5602         if (active) {
5603                 perf_pmu_disable(ctx->pmu);
5604                 /*
5605                  * We could be throttled; unthrottle now to avoid the tick
5606                  * trying to unthrottle while we already re-started the event.
5607                  */
5608                 if (event->hw.interrupts == MAX_INTERRUPTS) {
5609                         event->hw.interrupts = 0;
5610                         perf_log_throttle(event, 1);
5611                 }
5612                 event->pmu->stop(event, PERF_EF_UPDATE);
5613         }
5614 
5615         local64_set(&event->hw.period_left, 0);
5616 
5617         if (active) {
5618                 event->pmu->start(event, PERF_EF_RELOAD);
5619                 perf_pmu_enable(ctx->pmu);
5620         }
5621 }
5622 
5623 static int perf_event_check_period(struct perf_event *event, u64 value)
5624 {
5625         return event->pmu->check_period(event, value);
5626 }
5627 
5628 static int _perf_event_period(struct perf_event *event, u64 value)
5629 {
5630         if (!is_sampling_event(event))
5631                 return -EINVAL;
5632 
5633         if (!value)
5634                 return -EINVAL;
5635 
5636         if (event->attr.freq && value > sysctl_perf_event_sample_rate)
5637                 return -EINVAL;
5638 
5639         if (perf_event_check_period(event, value))
5640                 return -EINVAL;
5641 
5642         if (!event->attr.freq && (value & (1ULL << 63)))
5643                 return -EINVAL;
5644 
5645         event_function_call(event, __perf_event_period, &value);
5646 
5647         return 0;
5648 }
5649 
5650 int perf_event_period(struct perf_event *event, u64 value)
5651 {
5652         struct perf_event_context *ctx;
5653         int ret;
5654 
5655         ctx = perf_event_ctx_lock(event);
5656         ret = _perf_event_period(event, value);
5657         perf_event_ctx_unlock(event, ctx);
5658 
5659         return ret;
5660 }
5661 EXPORT_SYMBOL_GPL(perf_event_period);
5662 
5663 static const struct file_operations perf_fops;
5664 
5665 static inline int perf_fget_light(int fd, struct fd *p)
5666 {
5667         struct fd f = fdget(fd);
5668         if (!f.file)
5669                 return -EBADF;
5670 
5671         if (f.file->f_op != &perf_fops) {
5672                 fdput(f);
5673                 return -EBADF;
5674         }
5675         *p = f;
5676         return 0;
5677 }
5678 
5679 static int perf_event_set_output(struct perf_event *event,
5680                                  struct perf_event *output_event);
5681 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
5682 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5683                           struct perf_event_attr *attr);
5684 
5685 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
5686 {
5687         void (*func)(struct perf_event *);
5688         u32 flags = arg;
5689 
5690         switch (cmd) {
5691         case PERF_EVENT_IOC_ENABLE:
5692                 func = _perf_event_enable;
5693                 break;
5694         case PERF_EVENT_IOC_DISABLE:
5695                 func = _perf_event_disable;
5696                 break;
5697         case PERF_EVENT_IOC_RESET:
5698                 func = _perf_event_reset;
5699                 break;
5700 
5701         case PERF_EVENT_IOC_REFRESH:
5702                 return _perf_event_refresh(event, arg);
5703 
5704         case PERF_EVENT_IOC_PERIOD:
5705         {
5706                 u64 value;
5707 
5708                 if (copy_from_user(&value, (u64 __user *)arg, sizeof(value)))
5709                         return -EFAULT;
5710 
5711                 return _perf_event_period(event, value);
5712         }
5713         case PERF_EVENT_IOC_ID:
5714         {
5715                 u64 id = primary_event_id(event);
5716 
5717                 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
5718                         return -EFAULT;
5719                 return 0;
5720         }
5721 
5722         case PERF_EVENT_IOC_SET_OUTPUT:
5723         {
5724                 int ret;
5725                 if (arg != -1) {
5726                         struct perf_event *output_event;
5727                         struct fd output;
5728                         ret = perf_fget_light(arg, &output);
5729                         if (ret)
5730                                 return ret;
5731                         output_event = output.file->private_data;
5732                         ret = perf_event_set_output(event, output_event);
5733                         fdput(output);
5734                 } else {
5735                         ret = perf_event_set_output(event, NULL);
5736                 }
5737                 return ret;
5738         }
5739 
5740         case PERF_EVENT_IOC_SET_FILTER:
5741                 return perf_event_set_filter(event, (void __user *)arg);
5742 
5743         case PERF_EVENT_IOC_SET_BPF:
5744         {
5745                 struct bpf_prog *prog;
5746                 int err;
5747 
5748                 prog = bpf_prog_get(arg);
5749                 if (IS_ERR(prog))
5750                         return PTR_ERR(prog);
5751 
5752                 err = perf_event_set_bpf_prog(event, prog, 0);
5753                 if (err) {
5754                         bpf_prog_put(prog);
5755                         return err;
5756                 }
5757 
5758                 return 0;
5759         }
5760 
5761         case PERF_EVENT_IOC_PAUSE_OUTPUT: {
5762                 struct perf_buffer *rb;
5763 
5764                 rcu_read_lock();
5765                 rb = rcu_dereference(event->rb);
5766                 if (!rb || !rb->nr_pages) {
5767                         rcu_read_unlock();
5768                         return -EINVAL;
5769                 }
5770                 rb_toggle_paused(rb, !!arg);
5771                 rcu_read_unlock();
5772                 return 0;
5773         }
5774 
5775         case PERF_EVENT_IOC_QUERY_BPF:
5776                 return perf_event_query_prog_array(event, (void __user *)arg);
5777 
5778         case PERF_EVENT_IOC_MODIFY_ATTRIBUTES: {
5779                 struct perf_event_attr new_attr;
5780                 int err = perf_copy_attr((struct perf_event_attr __user *)arg,
5781                                          &new_attr);
5782 
5783                 if (err)
5784                         return err;
5785 
5786                 return perf_event_modify_attr(event,  &new_attr);
5787         }
5788         default:
5789                 return -ENOTTY;
5790         }
5791 
5792         if (flags & PERF_IOC_FLAG_GROUP)
5793                 perf_event_for_each(event, func);
5794         else
5795                 perf_event_for_each_child(event, func);
5796 
5797         return 0;
5798 }
5799 
5800 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
5801 {
5802         struct perf_event *event = file->private_data;
5803         struct perf_event_context *ctx;
5804         long ret;
5805 
5806         /* Treat ioctl like writes as it is likely a mutating operation. */
5807         ret = security_perf_event_write(event);
5808         if (ret)
5809                 return ret;
5810 
5811         ctx = perf_event_ctx_lock(event);
5812         ret = _perf_ioctl(event, cmd, arg);
5813         perf_event_ctx_unlock(event, ctx);
5814 
5815         return ret;
5816 }
5817 
5818 #ifdef CONFIG_COMPAT
5819 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
5820                                 unsigned long arg)
5821 {
5822         switch (_IOC_NR(cmd)) {
5823         case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
5824         case _IOC_NR(PERF_EVENT_IOC_ID):
5825         case _IOC_NR(PERF_EVENT_IOC_QUERY_BPF):
5826         case _IOC_NR(PERF_EVENT_IOC_MODIFY_ATTRIBUTES):
5827                 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
5828                 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
5829                         cmd &= ~IOCSIZE_MASK;
5830                         cmd |= sizeof(void *) << IOCSIZE_SHIFT;
5831                 }
5832                 break;
5833         }
5834         return perf_ioctl(file, cmd, arg);
5835 }
5836 #else
5837 # define perf_compat_ioctl NULL
5838 #endif
5839 
5840 int perf_event_task_enable(void)
5841 {
5842         struct perf_event_context *ctx;
5843         struct perf_event *event;
5844 
5845         mutex_lock(&current->perf_event_mutex);
5846         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5847                 ctx = perf_event_ctx_lock(event);
5848                 perf_event_for_each_child(event, _perf_event_enable);
5849                 perf_event_ctx_unlock(event, ctx);
5850         }
5851         mutex_unlock(&current->perf_event_mutex);
5852 
5853         return 0;
5854 }
5855 
5856 int perf_event_task_disable(void)
5857 {
5858         struct perf_event_context *ctx;
5859         struct perf_event *event;
5860 
5861         mutex_lock(&current->perf_event_mutex);
5862         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5863                 ctx = perf_event_ctx_lock(event);
5864                 perf_event_for_each_child(event, _perf_event_disable);
5865                 perf_event_ctx_unlock(event, ctx);
5866         }
5867         mutex_unlock(&current->perf_event_mutex);
5868 
5869         return 0;
5870 }
5871 
5872 static int perf_event_index(struct perf_event *event)
5873 {
5874         if (event->hw.state & PERF_HES_STOPPED)
5875                 return 0;