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

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

Version: ~ [ linux-5.11 ] ~ [ linux-5.10.17 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.99 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.176 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.221 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.257 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.257 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Performance events core code:
  3  *
  4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
  5  *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
  6  *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
  7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
  8  *
  9  * For licensing details see kernel-base/COPYING
 10  */
 11 
 12 #include <linux/fs.h>
 13 #include <linux/mm.h>
 14 #include <linux/cpu.h>
 15 #include <linux/smp.h>
 16 #include <linux/idr.h>
 17 #include <linux/file.h>
 18 #include <linux/poll.h>
 19 #include <linux/slab.h>
 20 #include <linux/hash.h>
 21 #include <linux/tick.h>
 22 #include <linux/sysfs.h>
 23 #include <linux/dcache.h>
 24 #include <linux/percpu.h>
 25 #include <linux/ptrace.h>
 26 #include <linux/reboot.h>
 27 #include <linux/vmstat.h>
 28 #include <linux/device.h>
 29 #include <linux/export.h>
 30 #include <linux/vmalloc.h>
 31 #include <linux/hardirq.h>
 32 #include <linux/rculist.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/syscalls.h>
 35 #include <linux/anon_inodes.h>
 36 #include <linux/kernel_stat.h>
 37 #include <linux/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 
 54 #include "internal.h"
 55 
 56 #include <asm/irq_regs.h>
 57 
 58 typedef int (*remote_function_f)(void *);
 59 
 60 struct remote_function_call {
 61         struct task_struct      *p;
 62         remote_function_f       func;
 63         void                    *info;
 64         int                     ret;
 65 };
 66 
 67 static void remote_function(void *data)
 68 {
 69         struct remote_function_call *tfc = data;
 70         struct task_struct *p = tfc->p;
 71 
 72         if (p) {
 73                 /* -EAGAIN */
 74                 if (task_cpu(p) != smp_processor_id())
 75                         return;
 76 
 77                 /*
 78                  * Now that we're on right CPU with IRQs disabled, we can test
 79                  * if we hit the right task without races.
 80                  */
 81 
 82                 tfc->ret = -ESRCH; /* No such (running) process */
 83                 if (p != current)
 84                         return;
 85         }
 86 
 87         tfc->ret = tfc->func(tfc->info);
 88 }
 89 
 90 /**
 91  * task_function_call - call a function on the cpu on which a task runs
 92  * @p:          the task to evaluate
 93  * @func:       the function to be called
 94  * @info:       the function call argument
 95  *
 96  * Calls the function @func when the task is currently running. This might
 97  * be on the current CPU, which just calls the function directly
 98  *
 99  * returns: @func return value, or
100  *          -ESRCH  - when the process isn't running
101  *          -EAGAIN - when the process moved away
102  */
103 static int
104 task_function_call(struct task_struct *p, remote_function_f func, void *info)
105 {
106         struct remote_function_call data = {
107                 .p      = p,
108                 .func   = func,
109                 .info   = info,
110                 .ret    = -EAGAIN,
111         };
112         int ret;
113 
114         do {
115                 ret = smp_call_function_single(task_cpu(p), remote_function, &data, 1);
116                 if (!ret)
117                         ret = data.ret;
118         } while (ret == -EAGAIN);
119 
120         return ret;
121 }
122 
123 /**
124  * cpu_function_call - call a function on the cpu
125  * @func:       the function to be called
126  * @info:       the function call argument
127  *
128  * Calls the function @func on the remote cpu.
129  *
130  * returns: @func return value or -ENXIO when the cpu is offline
131  */
132 static int cpu_function_call(int cpu, remote_function_f func, void *info)
133 {
134         struct remote_function_call data = {
135                 .p      = NULL,
136                 .func   = func,
137                 .info   = info,
138                 .ret    = -ENXIO, /* No such CPU */
139         };
140 
141         smp_call_function_single(cpu, remote_function, &data, 1);
142 
143         return data.ret;
144 }
145 
146 static inline struct perf_cpu_context *
147 __get_cpu_context(struct perf_event_context *ctx)
148 {
149         return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
150 }
151 
152 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
153                           struct perf_event_context *ctx)
154 {
155         raw_spin_lock(&cpuctx->ctx.lock);
156         if (ctx)
157                 raw_spin_lock(&ctx->lock);
158 }
159 
160 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
161                             struct perf_event_context *ctx)
162 {
163         if (ctx)
164                 raw_spin_unlock(&ctx->lock);
165         raw_spin_unlock(&cpuctx->ctx.lock);
166 }
167 
168 #define TASK_TOMBSTONE ((void *)-1L)
169 
170 static bool is_kernel_event(struct perf_event *event)
171 {
172         return READ_ONCE(event->owner) == TASK_TOMBSTONE;
173 }
174 
175 /*
176  * On task ctx scheduling...
177  *
178  * When !ctx->nr_events a task context will not be scheduled. This means
179  * we can disable the scheduler hooks (for performance) without leaving
180  * pending task ctx state.
181  *
182  * This however results in two special cases:
183  *
184  *  - removing the last event from a task ctx; this is relatively straight
185  *    forward and is done in __perf_remove_from_context.
186  *
187  *  - adding the first event to a task ctx; this is tricky because we cannot
188  *    rely on ctx->is_active and therefore cannot use event_function_call().
189  *    See perf_install_in_context().
190  *
191  * If ctx->nr_events, then ctx->is_active and cpuctx->task_ctx are set.
192  */
193 
194 typedef void (*event_f)(struct perf_event *, struct perf_cpu_context *,
195                         struct perf_event_context *, void *);
196 
197 struct event_function_struct {
198         struct perf_event *event;
199         event_f func;
200         void *data;
201 };
202 
203 static int event_function(void *info)
204 {
205         struct event_function_struct *efs = info;
206         struct perf_event *event = efs->event;
207         struct perf_event_context *ctx = event->ctx;
208         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
209         struct perf_event_context *task_ctx = cpuctx->task_ctx;
210         int ret = 0;
211 
212         lockdep_assert_irqs_disabled();
213 
214         perf_ctx_lock(cpuctx, task_ctx);
215         /*
216          * Since we do the IPI call without holding ctx->lock things can have
217          * changed, double check we hit the task we set out to hit.
218          */
219         if (ctx->task) {
220                 if (ctx->task != current) {
221                         ret = -ESRCH;
222                         goto unlock;
223                 }
224 
225                 /*
226                  * We only use event_function_call() on established contexts,
227                  * and event_function() is only ever called when active (or
228                  * rather, we'll have bailed in task_function_call() or the
229                  * above ctx->task != current test), therefore we must have
230                  * ctx->is_active here.
231                  */
232                 WARN_ON_ONCE(!ctx->is_active);
233                 /*
234                  * And since we have ctx->is_active, cpuctx->task_ctx must
235                  * match.
236                  */
237                 WARN_ON_ONCE(task_ctx != ctx);
238         } else {
239                 WARN_ON_ONCE(&cpuctx->ctx != ctx);
240         }
241 
242         efs->func(event, cpuctx, ctx, efs->data);
243 unlock:
244         perf_ctx_unlock(cpuctx, task_ctx);
245 
246         return ret;
247 }
248 
249 static void event_function_call(struct perf_event *event, event_f func, void *data)
250 {
251         struct perf_event_context *ctx = event->ctx;
252         struct task_struct *task = READ_ONCE(ctx->task); /* verified in event_function */
253         struct event_function_struct efs = {
254                 .event = event,
255                 .func = func,
256                 .data = data,
257         };
258 
259         if (!event->parent) {
260                 /*
261                  * If this is a !child event, we must hold ctx::mutex to
262                  * stabilize the the event->ctx relation. See
263                  * perf_event_ctx_lock().
264                  */
265                 lockdep_assert_held(&ctx->mutex);
266         }
267 
268         if (!task) {
269                 cpu_function_call(event->cpu, event_function, &efs);
270                 return;
271         }
272 
273         if (task == TASK_TOMBSTONE)
274                 return;
275 
276 again:
277         if (!task_function_call(task, event_function, &efs))
278                 return;
279 
280         raw_spin_lock_irq(&ctx->lock);
281         /*
282          * Reload the task pointer, it might have been changed by
283          * a concurrent perf_event_context_sched_out().
284          */
285         task = ctx->task;
286         if (task == TASK_TOMBSTONE) {
287                 raw_spin_unlock_irq(&ctx->lock);
288                 return;
289         }
290         if (ctx->is_active) {
291                 raw_spin_unlock_irq(&ctx->lock);
292                 goto again;
293         }
294         func(event, NULL, ctx, data);
295         raw_spin_unlock_irq(&ctx->lock);
296 }
297 
298 /*
299  * Similar to event_function_call() + event_function(), but hard assumes IRQs
300  * are already disabled and we're on the right CPU.
301  */
302 static void event_function_local(struct perf_event *event, event_f func, void *data)
303 {
304         struct perf_event_context *ctx = event->ctx;
305         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
306         struct task_struct *task = READ_ONCE(ctx->task);
307         struct perf_event_context *task_ctx = NULL;
308 
309         lockdep_assert_irqs_disabled();
310 
311         if (task) {
312                 if (task == TASK_TOMBSTONE)
313                         return;
314 
315                 task_ctx = ctx;
316         }
317 
318         perf_ctx_lock(cpuctx, task_ctx);
319 
320         task = ctx->task;
321         if (task == TASK_TOMBSTONE)
322                 goto unlock;
323 
324         if (task) {
325                 /*
326                  * We must be either inactive or active and the right task,
327                  * otherwise we're screwed, since we cannot IPI to somewhere
328                  * else.
329                  */
330                 if (ctx->is_active) {
331                         if (WARN_ON_ONCE(task != current))
332                                 goto unlock;
333 
334                         if (WARN_ON_ONCE(cpuctx->task_ctx != ctx))
335                                 goto unlock;
336                 }
337         } else {
338                 WARN_ON_ONCE(&cpuctx->ctx != ctx);
339         }
340 
341         func(event, cpuctx, ctx, data);
342 unlock:
343         perf_ctx_unlock(cpuctx, task_ctx);
344 }
345 
346 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
347                        PERF_FLAG_FD_OUTPUT  |\
348                        PERF_FLAG_PID_CGROUP |\
349                        PERF_FLAG_FD_CLOEXEC)
350 
351 /*
352  * branch priv levels that need permission checks
353  */
354 #define PERF_SAMPLE_BRANCH_PERM_PLM \
355         (PERF_SAMPLE_BRANCH_KERNEL |\
356          PERF_SAMPLE_BRANCH_HV)
357 
358 enum event_type_t {
359         EVENT_FLEXIBLE = 0x1,
360         EVENT_PINNED = 0x2,
361         EVENT_TIME = 0x4,
362         /* see ctx_resched() for details */
363         EVENT_CPU = 0x8,
364         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
365 };
366 
367 /*
368  * perf_sched_events : >0 events exist
369  * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
370  */
371 
372 static void perf_sched_delayed(struct work_struct *work);
373 DEFINE_STATIC_KEY_FALSE(perf_sched_events);
374 static DECLARE_DELAYED_WORK(perf_sched_work, perf_sched_delayed);
375 static DEFINE_MUTEX(perf_sched_mutex);
376 static atomic_t perf_sched_count;
377 
378 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
379 static DEFINE_PER_CPU(int, perf_sched_cb_usages);
380 static DEFINE_PER_CPU(struct pmu_event_list, pmu_sb_events);
381 
382 static atomic_t nr_mmap_events __read_mostly;
383 static atomic_t nr_comm_events __read_mostly;
384 static atomic_t nr_namespaces_events __read_mostly;
385 static atomic_t nr_task_events __read_mostly;
386 static atomic_t nr_freq_events __read_mostly;
387 static atomic_t nr_switch_events __read_mostly;
388 
389 static LIST_HEAD(pmus);
390 static DEFINE_MUTEX(pmus_lock);
391 static struct srcu_struct pmus_srcu;
392 static cpumask_var_t perf_online_mask;
393 
394 /*
395  * perf event paranoia level:
396  *  -1 - not paranoid at all
397  *   0 - disallow raw tracepoint access for unpriv
398  *   1 - disallow cpu events for unpriv
399  *   2 - disallow kernel profiling for unpriv
400  */
401 int sysctl_perf_event_paranoid __read_mostly = 2;
402 
403 /* Minimum for 512 kiB + 1 user control page */
404 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
405 
406 /*
407  * max perf event sample rate
408  */
409 #define DEFAULT_MAX_SAMPLE_RATE         100000
410 #define DEFAULT_SAMPLE_PERIOD_NS        (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
411 #define DEFAULT_CPU_TIME_MAX_PERCENT    25
412 
413 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
414 
415 static int max_samples_per_tick __read_mostly   = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
416 static int perf_sample_period_ns __read_mostly  = DEFAULT_SAMPLE_PERIOD_NS;
417 
418 static int perf_sample_allowed_ns __read_mostly =
419         DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
420 
421 static void update_perf_cpu_limits(void)
422 {
423         u64 tmp = perf_sample_period_ns;
424 
425         tmp *= sysctl_perf_cpu_time_max_percent;
426         tmp = div_u64(tmp, 100);
427         if (!tmp)
428                 tmp = 1;
429 
430         WRITE_ONCE(perf_sample_allowed_ns, tmp);
431 }
432 
433 static bool perf_rotate_context(struct perf_cpu_context *cpuctx);
434 
435 int perf_proc_update_handler(struct ctl_table *table, int write,
436                 void __user *buffer, size_t *lenp,
437                 loff_t *ppos)
438 {
439         int ret;
440         int perf_cpu = sysctl_perf_cpu_time_max_percent;
441         /*
442          * If throttling is disabled don't allow the write:
443          */
444         if (write && (perf_cpu == 100 || perf_cpu == 0))
445                 return -EINVAL;
446 
447         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
448         if (ret || !write)
449                 return ret;
450 
451         max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
452         perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
453         update_perf_cpu_limits();
454 
455         return 0;
456 }
457 
458 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
459 
460 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
461                                 void __user *buffer, size_t *lenp,
462                                 loff_t *ppos)
463 {
464         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
465 
466         if (ret || !write)
467                 return ret;
468 
469         if (sysctl_perf_cpu_time_max_percent == 100 ||
470             sysctl_perf_cpu_time_max_percent == 0) {
471                 printk(KERN_WARNING
472                        "perf: Dynamic interrupt throttling disabled, can hang your system!\n");
473                 WRITE_ONCE(perf_sample_allowed_ns, 0);
474         } else {
475                 update_perf_cpu_limits();
476         }
477 
478         return 0;
479 }
480 
481 /*
482  * perf samples are done in some very critical code paths (NMIs).
483  * If they take too much CPU time, the system can lock up and not
484  * get any real work done.  This will drop the sample rate when
485  * we detect that events are taking too long.
486  */
487 #define NR_ACCUMULATED_SAMPLES 128
488 static DEFINE_PER_CPU(u64, running_sample_length);
489 
490 static u64 __report_avg;
491 static u64 __report_allowed;
492 
493 static void perf_duration_warn(struct irq_work *w)
494 {
495         printk_ratelimited(KERN_INFO
496                 "perf: interrupt took too long (%lld > %lld), lowering "
497                 "kernel.perf_event_max_sample_rate to %d\n",
498                 __report_avg, __report_allowed,
499                 sysctl_perf_event_sample_rate);
500 }
501 
502 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
503 
504 void perf_sample_event_took(u64 sample_len_ns)
505 {
506         u64 max_len = READ_ONCE(perf_sample_allowed_ns);
507         u64 running_len;
508         u64 avg_len;
509         u32 max;
510 
511         if (max_len == 0)
512                 return;
513 
514         /* Decay the counter by 1 average sample. */
515         running_len = __this_cpu_read(running_sample_length);
516         running_len -= running_len/NR_ACCUMULATED_SAMPLES;
517         running_len += sample_len_ns;
518         __this_cpu_write(running_sample_length, running_len);
519 
520         /*
521          * Note: this will be biased artifically low until we have
522          * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
523          * from having to maintain a count.
524          */
525         avg_len = running_len/NR_ACCUMULATED_SAMPLES;
526         if (avg_len <= max_len)
527                 return;
528 
529         __report_avg = avg_len;
530         __report_allowed = max_len;
531 
532         /*
533          * Compute a throttle threshold 25% below the current duration.
534          */
535         avg_len += avg_len / 4;
536         max = (TICK_NSEC / 100) * sysctl_perf_cpu_time_max_percent;
537         if (avg_len < max)
538                 max /= (u32)avg_len;
539         else
540                 max = 1;
541 
542         WRITE_ONCE(perf_sample_allowed_ns, avg_len);
543         WRITE_ONCE(max_samples_per_tick, max);
544 
545         sysctl_perf_event_sample_rate = max * HZ;
546         perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
547 
548         if (!irq_work_queue(&perf_duration_work)) {
549                 early_printk("perf: interrupt took too long (%lld > %lld), lowering "
550                              "kernel.perf_event_max_sample_rate to %d\n",
551                              __report_avg, __report_allowed,
552                              sysctl_perf_event_sample_rate);
553         }
554 }
555 
556 static atomic64_t perf_event_id;
557 
558 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
559                               enum event_type_t event_type);
560 
561 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
562                              enum event_type_t event_type,
563                              struct task_struct *task);
564 
565 static void update_context_time(struct perf_event_context *ctx);
566 static u64 perf_event_time(struct perf_event *event);
567 
568 void __weak perf_event_print_debug(void)        { }
569 
570 extern __weak const char *perf_pmu_name(void)
571 {
572         return "pmu";
573 }
574 
575 static inline u64 perf_clock(void)
576 {
577         return local_clock();
578 }
579 
580 static inline u64 perf_event_clock(struct perf_event *event)
581 {
582         return event->clock();
583 }
584 
585 /*
586  * State based event timekeeping...
587  *
588  * The basic idea is to use event->state to determine which (if any) time
589  * fields to increment with the current delta. This means we only need to
590  * update timestamps when we change state or when they are explicitly requested
591  * (read).
592  *
593  * Event groups make things a little more complicated, but not terribly so. The
594  * rules for a group are that if the group leader is OFF the entire group is
595  * OFF, irrespecive of what the group member states are. This results in
596  * __perf_effective_state().
597  *
598  * A futher ramification is that when a group leader flips between OFF and
599  * !OFF, we need to update all group member times.
600  *
601  *
602  * NOTE: perf_event_time() is based on the (cgroup) context time, and thus we
603  * need to make sure the relevant context time is updated before we try and
604  * update our timestamps.
605  */
606 
607 static __always_inline enum perf_event_state
608 __perf_effective_state(struct perf_event *event)
609 {
610         struct perf_event *leader = event->group_leader;
611 
612         if (leader->state <= PERF_EVENT_STATE_OFF)
613                 return leader->state;
614 
615         return event->state;
616 }
617 
618 static __always_inline void
619 __perf_update_times(struct perf_event *event, u64 now, u64 *enabled, u64 *running)
620 {
621         enum perf_event_state state = __perf_effective_state(event);
622         u64 delta = now - event->tstamp;
623 
624         *enabled = event->total_time_enabled;
625         if (state >= PERF_EVENT_STATE_INACTIVE)
626                 *enabled += delta;
627 
628         *running = event->total_time_running;
629         if (state >= PERF_EVENT_STATE_ACTIVE)
630                 *running += delta;
631 }
632 
633 static void perf_event_update_time(struct perf_event *event)
634 {
635         u64 now = perf_event_time(event);
636 
637         __perf_update_times(event, now, &event->total_time_enabled,
638                                         &event->total_time_running);
639         event->tstamp = now;
640 }
641 
642 static void perf_event_update_sibling_time(struct perf_event *leader)
643 {
644         struct perf_event *sibling;
645 
646         for_each_sibling_event(sibling, leader)
647                 perf_event_update_time(sibling);
648 }
649 
650 static void
651 perf_event_set_state(struct perf_event *event, enum perf_event_state state)
652 {
653         if (event->state == state)
654                 return;
655 
656         perf_event_update_time(event);
657         /*
658          * If a group leader gets enabled/disabled all its siblings
659          * are affected too.
660          */
661         if ((event->state < 0) ^ (state < 0))
662                 perf_event_update_sibling_time(event);
663 
664         WRITE_ONCE(event->state, state);
665 }
666 
667 #ifdef CONFIG_CGROUP_PERF
668 
669 static inline bool
670 perf_cgroup_match(struct perf_event *event)
671 {
672         struct perf_event_context *ctx = event->ctx;
673         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
674 
675         /* @event doesn't care about cgroup */
676         if (!event->cgrp)
677                 return true;
678 
679         /* wants specific cgroup scope but @cpuctx isn't associated with any */
680         if (!cpuctx->cgrp)
681                 return false;
682 
683         /*
684          * Cgroup scoping is recursive.  An event enabled for a cgroup is
685          * also enabled for all its descendant cgroups.  If @cpuctx's
686          * cgroup is a descendant of @event's (the test covers identity
687          * case), it's a match.
688          */
689         return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
690                                     event->cgrp->css.cgroup);
691 }
692 
693 static inline void perf_detach_cgroup(struct perf_event *event)
694 {
695         css_put(&event->cgrp->css);
696         event->cgrp = NULL;
697 }
698 
699 static inline int is_cgroup_event(struct perf_event *event)
700 {
701         return event->cgrp != NULL;
702 }
703 
704 static inline u64 perf_cgroup_event_time(struct perf_event *event)
705 {
706         struct perf_cgroup_info *t;
707 
708         t = per_cpu_ptr(event->cgrp->info, event->cpu);
709         return t->time;
710 }
711 
712 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
713 {
714         struct perf_cgroup_info *info;
715         u64 now;
716 
717         now = perf_clock();
718 
719         info = this_cpu_ptr(cgrp->info);
720 
721         info->time += now - info->timestamp;
722         info->timestamp = now;
723 }
724 
725 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
726 {
727         struct perf_cgroup *cgrp = cpuctx->cgrp;
728         struct cgroup_subsys_state *css;
729 
730         if (cgrp) {
731                 for (css = &cgrp->css; css; css = css->parent) {
732                         cgrp = container_of(css, struct perf_cgroup, css);
733                         __update_cgrp_time(cgrp);
734                 }
735         }
736 }
737 
738 static inline void update_cgrp_time_from_event(struct perf_event *event)
739 {
740         struct perf_cgroup *cgrp;
741 
742         /*
743          * ensure we access cgroup data only when needed and
744          * when we know the cgroup is pinned (css_get)
745          */
746         if (!is_cgroup_event(event))
747                 return;
748 
749         cgrp = perf_cgroup_from_task(current, event->ctx);
750         /*
751          * Do not update time when cgroup is not active
752          */
753         if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
754                 __update_cgrp_time(event->cgrp);
755 }
756 
757 static inline void
758 perf_cgroup_set_timestamp(struct task_struct *task,
759                           struct perf_event_context *ctx)
760 {
761         struct perf_cgroup *cgrp;
762         struct perf_cgroup_info *info;
763         struct cgroup_subsys_state *css;
764 
765         /*
766          * ctx->lock held by caller
767          * ensure we do not access cgroup data
768          * unless we have the cgroup pinned (css_get)
769          */
770         if (!task || !ctx->nr_cgroups)
771                 return;
772 
773         cgrp = perf_cgroup_from_task(task, ctx);
774 
775         for (css = &cgrp->css; css; css = css->parent) {
776                 cgrp = container_of(css, struct perf_cgroup, css);
777                 info = this_cpu_ptr(cgrp->info);
778                 info->timestamp = ctx->timestamp;
779         }
780 }
781 
782 static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
783 
784 #define PERF_CGROUP_SWOUT       0x1 /* cgroup switch out every event */
785 #define PERF_CGROUP_SWIN        0x2 /* cgroup switch in events based on task */
786 
787 /*
788  * reschedule events based on the cgroup constraint of task.
789  *
790  * mode SWOUT : schedule out everything
791  * mode SWIN : schedule in based on cgroup for next
792  */
793 static void perf_cgroup_switch(struct task_struct *task, int mode)
794 {
795         struct perf_cpu_context *cpuctx;
796         struct list_head *list;
797         unsigned long flags;
798 
799         /*
800          * Disable interrupts and preemption to avoid this CPU's
801          * cgrp_cpuctx_entry to change under us.
802          */
803         local_irq_save(flags);
804 
805         list = this_cpu_ptr(&cgrp_cpuctx_list);
806         list_for_each_entry(cpuctx, list, cgrp_cpuctx_entry) {
807                 WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
808 
809                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
810                 perf_pmu_disable(cpuctx->ctx.pmu);
811 
812                 if (mode & PERF_CGROUP_SWOUT) {
813                         cpu_ctx_sched_out(cpuctx, EVENT_ALL);
814                         /*
815                          * must not be done before ctxswout due
816                          * to event_filter_match() in event_sched_out()
817                          */
818                         cpuctx->cgrp = NULL;
819                 }
820 
821                 if (mode & PERF_CGROUP_SWIN) {
822                         WARN_ON_ONCE(cpuctx->cgrp);
823                         /*
824                          * set cgrp before ctxsw in to allow
825                          * event_filter_match() to not have to pass
826                          * task around
827                          * we pass the cpuctx->ctx to perf_cgroup_from_task()
828                          * because cgorup events are only per-cpu
829                          */
830                         cpuctx->cgrp = perf_cgroup_from_task(task,
831                                                              &cpuctx->ctx);
832                         cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
833                 }
834                 perf_pmu_enable(cpuctx->ctx.pmu);
835                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
836         }
837 
838         local_irq_restore(flags);
839 }
840 
841 static inline void perf_cgroup_sched_out(struct task_struct *task,
842                                          struct task_struct *next)
843 {
844         struct perf_cgroup *cgrp1;
845         struct perf_cgroup *cgrp2 = NULL;
846 
847         rcu_read_lock();
848         /*
849          * we come here when we know perf_cgroup_events > 0
850          * we do not need to pass the ctx here because we know
851          * we are holding the rcu lock
852          */
853         cgrp1 = perf_cgroup_from_task(task, NULL);
854         cgrp2 = perf_cgroup_from_task(next, NULL);
855 
856         /*
857          * only schedule out current cgroup events if we know
858          * that we are switching to a different cgroup. Otherwise,
859          * do no touch the cgroup events.
860          */
861         if (cgrp1 != cgrp2)
862                 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
863 
864         rcu_read_unlock();
865 }
866 
867 static inline void perf_cgroup_sched_in(struct task_struct *prev,
868                                         struct task_struct *task)
869 {
870         struct perf_cgroup *cgrp1;
871         struct perf_cgroup *cgrp2 = NULL;
872 
873         rcu_read_lock();
874         /*
875          * we come here when we know perf_cgroup_events > 0
876          * we do not need to pass the ctx here because we know
877          * we are holding the rcu lock
878          */
879         cgrp1 = perf_cgroup_from_task(task, NULL);
880         cgrp2 = perf_cgroup_from_task(prev, NULL);
881 
882         /*
883          * only need to schedule in cgroup events if we are changing
884          * cgroup during ctxsw. Cgroup events were not scheduled
885          * out of ctxsw out if that was not the case.
886          */
887         if (cgrp1 != cgrp2)
888                 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
889 
890         rcu_read_unlock();
891 }
892 
893 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
894                                       struct perf_event_attr *attr,
895                                       struct perf_event *group_leader)
896 {
897         struct perf_cgroup *cgrp;
898         struct cgroup_subsys_state *css;
899         struct fd f = fdget(fd);
900         int ret = 0;
901 
902         if (!f.file)
903                 return -EBADF;
904 
905         css = css_tryget_online_from_dir(f.file->f_path.dentry,
906                                          &perf_event_cgrp_subsys);
907         if (IS_ERR(css)) {
908                 ret = PTR_ERR(css);
909                 goto out;
910         }
911 
912         cgrp = container_of(css, struct perf_cgroup, css);
913         event->cgrp = cgrp;
914 
915         /*
916          * all events in a group must monitor
917          * the same cgroup because a task belongs
918          * to only one perf cgroup at a time
919          */
920         if (group_leader && group_leader->cgrp != cgrp) {
921                 perf_detach_cgroup(event);
922                 ret = -EINVAL;
923         }
924 out:
925         fdput(f);
926         return ret;
927 }
928 
929 static inline void
930 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
931 {
932         struct perf_cgroup_info *t;
933         t = per_cpu_ptr(event->cgrp->info, event->cpu);
934         event->shadow_ctx_time = now - t->timestamp;
935 }
936 
937 /*
938  * Update cpuctx->cgrp so that it is set when first cgroup event is added and
939  * cleared when last cgroup event is removed.
940  */
941 static inline void
942 list_update_cgroup_event(struct perf_event *event,
943                          struct perf_event_context *ctx, bool add)
944 {
945         struct perf_cpu_context *cpuctx;
946         struct list_head *cpuctx_entry;
947 
948         if (!is_cgroup_event(event))
949                 return;
950 
951         /*
952          * Because cgroup events are always per-cpu events,
953          * this will always be called from the right CPU.
954          */
955         cpuctx = __get_cpu_context(ctx);
956 
957         /*
958          * Since setting cpuctx->cgrp is conditional on the current @cgrp
959          * matching the event's cgroup, we must do this for every new event,
960          * because if the first would mismatch, the second would not try again
961          * and we would leave cpuctx->cgrp unset.
962          */
963         if (add && !cpuctx->cgrp) {
964                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
965 
966                 if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
967                         cpuctx->cgrp = cgrp;
968         }
969 
970         if (add && ctx->nr_cgroups++)
971                 return;
972         else if (!add && --ctx->nr_cgroups)
973                 return;
974 
975         /* no cgroup running */
976         if (!add)
977                 cpuctx->cgrp = NULL;
978 
979         cpuctx_entry = &cpuctx->cgrp_cpuctx_entry;
980         if (add)
981                 list_add(cpuctx_entry, this_cpu_ptr(&cgrp_cpuctx_list));
982         else
983                 list_del(cpuctx_entry);
984 }
985 
986 #else /* !CONFIG_CGROUP_PERF */
987 
988 static inline bool
989 perf_cgroup_match(struct perf_event *event)
990 {
991         return true;
992 }
993 
994 static inline void perf_detach_cgroup(struct perf_event *event)
995 {}
996 
997 static inline int is_cgroup_event(struct perf_event *event)
998 {
999         return 0;
1000 }
1001 
1002 static inline void update_cgrp_time_from_event(struct perf_event *event)
1003 {
1004 }
1005 
1006 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
1007 {
1008 }
1009 
1010 static inline void perf_cgroup_sched_out(struct task_struct *task,
1011                                          struct task_struct *next)
1012 {
1013 }
1014 
1015 static inline void perf_cgroup_sched_in(struct task_struct *prev,
1016                                         struct task_struct *task)
1017 {
1018 }
1019 
1020 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
1021                                       struct perf_event_attr *attr,
1022                                       struct perf_event *group_leader)
1023 {
1024         return -EINVAL;
1025 }
1026 
1027 static inline void
1028 perf_cgroup_set_timestamp(struct task_struct *task,
1029                           struct perf_event_context *ctx)
1030 {
1031 }
1032 
1033 void
1034 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
1035 {
1036 }
1037 
1038 static inline void
1039 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
1040 {
1041 }
1042 
1043 static inline u64 perf_cgroup_event_time(struct perf_event *event)
1044 {
1045         return 0;
1046 }
1047 
1048 static inline void
1049 list_update_cgroup_event(struct perf_event *event,
1050                          struct perf_event_context *ctx, bool add)
1051 {
1052 }
1053 
1054 #endif
1055 
1056 /*
1057  * set default to be dependent on timer tick just
1058  * like original code
1059  */
1060 #define PERF_CPU_HRTIMER (1000 / HZ)
1061 /*
1062  * function must be called with interrupts disabled
1063  */
1064 static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr)
1065 {
1066         struct perf_cpu_context *cpuctx;
1067         bool rotations;
1068 
1069         lockdep_assert_irqs_disabled();
1070 
1071         cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
1072         rotations = perf_rotate_context(cpuctx);
1073 
1074         raw_spin_lock(&cpuctx->hrtimer_lock);
1075         if (rotations)
1076                 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
1077         else
1078                 cpuctx->hrtimer_active = 0;
1079         raw_spin_unlock(&cpuctx->hrtimer_lock);
1080 
1081         return rotations ? HRTIMER_RESTART : HRTIMER_NORESTART;
1082 }
1083 
1084 static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
1085 {
1086         struct hrtimer *timer = &cpuctx->hrtimer;
1087         struct pmu *pmu = cpuctx->ctx.pmu;
1088         u64 interval;
1089 
1090         /* no multiplexing needed for SW PMU */
1091         if (pmu->task_ctx_nr == perf_sw_context)
1092                 return;
1093 
1094         /*
1095          * check default is sane, if not set then force to
1096          * default interval (1/tick)
1097          */
1098         interval = pmu->hrtimer_interval_ms;
1099         if (interval < 1)
1100                 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
1101 
1102         cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval);
1103 
1104         raw_spin_lock_init(&cpuctx->hrtimer_lock);
1105         hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1106         timer->function = perf_mux_hrtimer_handler;
1107 }
1108 
1109 static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
1110 {
1111         struct hrtimer *timer = &cpuctx->hrtimer;
1112         struct pmu *pmu = cpuctx->ctx.pmu;
1113         unsigned long flags;
1114 
1115         /* not for SW PMU */
1116         if (pmu->task_ctx_nr == perf_sw_context)
1117                 return 0;
1118 
1119         raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags);
1120         if (!cpuctx->hrtimer_active) {
1121                 cpuctx->hrtimer_active = 1;
1122                 hrtimer_forward_now(timer, cpuctx->hrtimer_interval);
1123                 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
1124         }
1125         raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags);
1126 
1127         return 0;
1128 }
1129 
1130 void perf_pmu_disable(struct pmu *pmu)
1131 {
1132         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1133         if (!(*count)++)
1134                 pmu->pmu_disable(pmu);
1135 }
1136 
1137 void perf_pmu_enable(struct pmu *pmu)
1138 {
1139         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1140         if (!--(*count))
1141                 pmu->pmu_enable(pmu);
1142 }
1143 
1144 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
1145 
1146 /*
1147  * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
1148  * perf_event_task_tick() are fully serialized because they're strictly cpu
1149  * affine and perf_event_ctx{activate,deactivate} are called with IRQs
1150  * disabled, while perf_event_task_tick is called from IRQ context.
1151  */
1152 static void perf_event_ctx_activate(struct perf_event_context *ctx)
1153 {
1154         struct list_head *head = this_cpu_ptr(&active_ctx_list);
1155 
1156         lockdep_assert_irqs_disabled();
1157 
1158         WARN_ON(!list_empty(&ctx->active_ctx_list));
1159 
1160         list_add(&ctx->active_ctx_list, head);
1161 }
1162 
1163 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
1164 {
1165         lockdep_assert_irqs_disabled();
1166 
1167         WARN_ON(list_empty(&ctx->active_ctx_list));
1168 
1169         list_del_init(&ctx->active_ctx_list);
1170 }
1171 
1172 static void get_ctx(struct perf_event_context *ctx)
1173 {
1174         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
1175 }
1176 
1177 static void free_ctx(struct rcu_head *head)
1178 {
1179         struct perf_event_context *ctx;
1180 
1181         ctx = container_of(head, struct perf_event_context, rcu_head);
1182         kfree(ctx->task_ctx_data);
1183         kfree(ctx);
1184 }
1185 
1186 static void put_ctx(struct perf_event_context *ctx)
1187 {
1188         if (atomic_dec_and_test(&ctx->refcount)) {
1189                 if (ctx->parent_ctx)
1190                         put_ctx(ctx->parent_ctx);
1191                 if (ctx->task && ctx->task != TASK_TOMBSTONE)
1192                         put_task_struct(ctx->task);
1193                 call_rcu(&ctx->rcu_head, free_ctx);
1194         }
1195 }
1196 
1197 /*
1198  * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
1199  * perf_pmu_migrate_context() we need some magic.
1200  *
1201  * Those places that change perf_event::ctx will hold both
1202  * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
1203  *
1204  * Lock ordering is by mutex address. There are two other sites where
1205  * perf_event_context::mutex nests and those are:
1206  *
1207  *  - perf_event_exit_task_context()    [ child , 0 ]
1208  *      perf_event_exit_event()
1209  *        put_event()                   [ parent, 1 ]
1210  *
1211  *  - perf_event_init_context()         [ parent, 0 ]
1212  *      inherit_task_group()
1213  *        inherit_group()
1214  *          inherit_event()
1215  *            perf_event_alloc()
1216  *              perf_init_event()
1217  *                perf_try_init_event() [ child , 1 ]
1218  *
1219  * While it appears there is an obvious deadlock here -- the parent and child
1220  * nesting levels are inverted between the two. This is in fact safe because
1221  * life-time rules separate them. That is an exiting task cannot fork, and a
1222  * spawning task cannot (yet) exit.
1223  *
1224  * But remember that that these are parent<->child context relations, and
1225  * migration does not affect children, therefore these two orderings should not
1226  * interact.
1227  *
1228  * The change in perf_event::ctx does not affect children (as claimed above)
1229  * because the sys_perf_event_open() case will install a new event and break
1230  * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
1231  * concerned with cpuctx and that doesn't have children.
1232  *
1233  * The places that change perf_event::ctx will issue:
1234  *
1235  *   perf_remove_from_context();
1236  *   synchronize_rcu();
1237  *   perf_install_in_context();
1238  *
1239  * to affect the change. The remove_from_context() + synchronize_rcu() should
1240  * quiesce the event, after which we can install it in the new location. This
1241  * means that only external vectors (perf_fops, prctl) can perturb the event
1242  * while in transit. Therefore all such accessors should also acquire
1243  * perf_event_context::mutex to serialize against this.
1244  *
1245  * However; because event->ctx can change while we're waiting to acquire
1246  * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
1247  * function.
1248  *
1249  * Lock order:
1250  *    cred_guard_mutex
1251  *      task_struct::perf_event_mutex
1252  *        perf_event_context::mutex
1253  *          perf_event::child_mutex;
1254  *            perf_event_context::lock
1255  *          perf_event::mmap_mutex
1256  *          mmap_sem
1257  *
1258  *    cpu_hotplug_lock
1259  *      pmus_lock
1260  *        cpuctx->mutex / perf_event_context::mutex
1261  */
1262 static struct perf_event_context *
1263 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
1264 {
1265         struct perf_event_context *ctx;
1266 
1267 again:
1268         rcu_read_lock();
1269         ctx = READ_ONCE(event->ctx);
1270         if (!atomic_inc_not_zero(&ctx->refcount)) {
1271                 rcu_read_unlock();
1272                 goto again;
1273         }
1274         rcu_read_unlock();
1275 
1276         mutex_lock_nested(&ctx->mutex, nesting);
1277         if (event->ctx != ctx) {
1278                 mutex_unlock(&ctx->mutex);
1279                 put_ctx(ctx);
1280                 goto again;
1281         }
1282 
1283         return ctx;
1284 }
1285 
1286 static inline struct perf_event_context *
1287 perf_event_ctx_lock(struct perf_event *event)
1288 {
1289         return perf_event_ctx_lock_nested(event, 0);
1290 }
1291 
1292 static void perf_event_ctx_unlock(struct perf_event *event,
1293                                   struct perf_event_context *ctx)
1294 {
1295         mutex_unlock(&ctx->mutex);
1296         put_ctx(ctx);
1297 }
1298 
1299 /*
1300  * This must be done under the ctx->lock, such as to serialize against
1301  * context_equiv(), therefore we cannot call put_ctx() since that might end up
1302  * calling scheduler related locks and ctx->lock nests inside those.
1303  */
1304 static __must_check struct perf_event_context *
1305 unclone_ctx(struct perf_event_context *ctx)
1306 {
1307         struct perf_event_context *parent_ctx = ctx->parent_ctx;
1308 
1309         lockdep_assert_held(&ctx->lock);
1310 
1311         if (parent_ctx)
1312                 ctx->parent_ctx = NULL;
1313         ctx->generation++;
1314 
1315         return parent_ctx;
1316 }
1317 
1318 static u32 perf_event_pid_type(struct perf_event *event, struct task_struct *p,
1319                                 enum pid_type type)
1320 {
1321         u32 nr;
1322         /*
1323          * only top level events have the pid namespace they were created in
1324          */
1325         if (event->parent)
1326                 event = event->parent;
1327 
1328         nr = __task_pid_nr_ns(p, type, event->ns);
1329         /* avoid -1 if it is idle thread or runs in another ns */
1330         if (!nr && !pid_alive(p))
1331                 nr = -1;
1332         return nr;
1333 }
1334 
1335 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1336 {
1337         return perf_event_pid_type(event, p, PIDTYPE_TGID);
1338 }
1339 
1340 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1341 {
1342         return perf_event_pid_type(event, p, PIDTYPE_PID);
1343 }
1344 
1345 /*
1346  * If we inherit events we want to return the parent event id
1347  * to userspace.
1348  */
1349 static u64 primary_event_id(struct perf_event *event)
1350 {
1351         u64 id = event->id;
1352 
1353         if (event->parent)
1354                 id = event->parent->id;
1355 
1356         return id;
1357 }
1358 
1359 /*
1360  * Get the perf_event_context for a task and lock it.
1361  *
1362  * This has to cope with with the fact that until it is locked,
1363  * the context could get moved to another task.
1364  */
1365 static struct perf_event_context *
1366 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1367 {
1368         struct perf_event_context *ctx;
1369 
1370 retry:
1371         /*
1372          * One of the few rules of preemptible RCU is that one cannot do
1373          * rcu_read_unlock() while holding a scheduler (or nested) lock when
1374          * part of the read side critical section was irqs-enabled -- see
1375          * rcu_read_unlock_special().
1376          *
1377          * Since ctx->lock nests under rq->lock we must ensure the entire read
1378          * side critical section has interrupts disabled.
1379          */
1380         local_irq_save(*flags);
1381         rcu_read_lock();
1382         ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1383         if (ctx) {
1384                 /*
1385                  * If this context is a clone of another, it might
1386                  * get swapped for another underneath us by
1387                  * perf_event_task_sched_out, though the
1388                  * rcu_read_lock() protects us from any context
1389                  * getting freed.  Lock the context and check if it
1390                  * got swapped before we could get the lock, and retry
1391                  * if so.  If we locked the right context, then it
1392                  * can't get swapped on us any more.
1393                  */
1394                 raw_spin_lock(&ctx->lock);
1395                 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1396                         raw_spin_unlock(&ctx->lock);
1397                         rcu_read_unlock();
1398                         local_irq_restore(*flags);
1399                         goto retry;
1400                 }
1401 
1402                 if (ctx->task == TASK_TOMBSTONE ||
1403                     !atomic_inc_not_zero(&ctx->refcount)) {
1404                         raw_spin_unlock(&ctx->lock);
1405                         ctx = NULL;
1406                 } else {
1407                         WARN_ON_ONCE(ctx->task != task);
1408                 }
1409         }
1410         rcu_read_unlock();
1411         if (!ctx)
1412                 local_irq_restore(*flags);
1413         return ctx;
1414 }
1415 
1416 /*
1417  * Get the context for a task and increment its pin_count so it
1418  * can't get swapped to another task.  This also increments its
1419  * reference count so that the context can't get freed.
1420  */
1421 static struct perf_event_context *
1422 perf_pin_task_context(struct task_struct *task, int ctxn)
1423 {
1424         struct perf_event_context *ctx;
1425         unsigned long flags;
1426 
1427         ctx = perf_lock_task_context(task, ctxn, &flags);
1428         if (ctx) {
1429                 ++ctx->pin_count;
1430                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1431         }
1432         return ctx;
1433 }
1434 
1435 static void perf_unpin_context(struct perf_event_context *ctx)
1436 {
1437         unsigned long flags;
1438 
1439         raw_spin_lock_irqsave(&ctx->lock, flags);
1440         --ctx->pin_count;
1441         raw_spin_unlock_irqrestore(&ctx->lock, flags);
1442 }
1443 
1444 /*
1445  * Update the record of the current time in a context.
1446  */
1447 static void update_context_time(struct perf_event_context *ctx)
1448 {
1449         u64 now = perf_clock();
1450 
1451         ctx->time += now - ctx->timestamp;
1452         ctx->timestamp = now;
1453 }
1454 
1455 static u64 perf_event_time(struct perf_event *event)
1456 {
1457         struct perf_event_context *ctx = event->ctx;
1458 
1459         if (is_cgroup_event(event))
1460                 return perf_cgroup_event_time(event);
1461 
1462         return ctx ? ctx->time : 0;
1463 }
1464 
1465 static enum event_type_t get_event_type(struct perf_event *event)
1466 {
1467         struct perf_event_context *ctx = event->ctx;
1468         enum event_type_t event_type;
1469 
1470         lockdep_assert_held(&ctx->lock);
1471 
1472         /*
1473          * It's 'group type', really, because if our group leader is
1474          * pinned, so are we.
1475          */
1476         if (event->group_leader != event)
1477                 event = event->group_leader;
1478 
1479         event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE;
1480         if (!ctx->task)
1481                 event_type |= EVENT_CPU;
1482 
1483         return event_type;
1484 }
1485 
1486 /*
1487  * Helper function to initialize event group nodes.
1488  */
1489 static void init_event_group(struct perf_event *event)
1490 {
1491         RB_CLEAR_NODE(&event->group_node);
1492         event->group_index = 0;
1493 }
1494 
1495 /*
1496  * Extract pinned or flexible groups from the context
1497  * based on event attrs bits.
1498  */
1499 static struct perf_event_groups *
1500 get_event_groups(struct perf_event *event, struct perf_event_context *ctx)
1501 {
1502         if (event->attr.pinned)
1503                 return &ctx->pinned_groups;
1504         else
1505                 return &ctx->flexible_groups;
1506 }
1507 
1508 /*
1509  * Helper function to initializes perf_event_group trees.
1510  */
1511 static void perf_event_groups_init(struct perf_event_groups *groups)
1512 {
1513         groups->tree = RB_ROOT;
1514         groups->index = 0;
1515 }
1516 
1517 /*
1518  * Compare function for event groups;
1519  *
1520  * Implements complex key that first sorts by CPU and then by virtual index
1521  * which provides ordering when rotating groups for the same CPU.
1522  */
1523 static bool
1524 perf_event_groups_less(struct perf_event *left, struct perf_event *right)
1525 {
1526         if (left->cpu < right->cpu)
1527                 return true;
1528         if (left->cpu > right->cpu)
1529                 return false;
1530 
1531         if (left->group_index < right->group_index)
1532                 return true;
1533         if (left->group_index > right->group_index)
1534                 return false;
1535 
1536         return false;
1537 }
1538 
1539 /*
1540  * Insert @event into @groups' tree; using {@event->cpu, ++@groups->index} for
1541  * key (see perf_event_groups_less). This places it last inside the CPU
1542  * subtree.
1543  */
1544 static void
1545 perf_event_groups_insert(struct perf_event_groups *groups,
1546                          struct perf_event *event)
1547 {
1548         struct perf_event *node_event;
1549         struct rb_node *parent;
1550         struct rb_node **node;
1551 
1552         event->group_index = ++groups->index;
1553 
1554         node = &groups->tree.rb_node;
1555         parent = *node;
1556 
1557         while (*node) {
1558                 parent = *node;
1559                 node_event = container_of(*node, struct perf_event, group_node);
1560 
1561                 if (perf_event_groups_less(event, node_event))
1562                         node = &parent->rb_left;
1563                 else
1564                         node = &parent->rb_right;
1565         }
1566 
1567         rb_link_node(&event->group_node, parent, node);
1568         rb_insert_color(&event->group_node, &groups->tree);
1569 }
1570 
1571 /*
1572  * Helper function to insert event into the pinned or flexible groups.
1573  */
1574 static void
1575 add_event_to_groups(struct perf_event *event, struct perf_event_context *ctx)
1576 {
1577         struct perf_event_groups *groups;
1578 
1579         groups = get_event_groups(event, ctx);
1580         perf_event_groups_insert(groups, event);
1581 }
1582 
1583 /*
1584  * Delete a group from a tree.
1585  */
1586 static void
1587 perf_event_groups_delete(struct perf_event_groups *groups,
1588                          struct perf_event *event)
1589 {
1590         WARN_ON_ONCE(RB_EMPTY_NODE(&event->group_node) ||
1591                      RB_EMPTY_ROOT(&groups->tree));
1592 
1593         rb_erase(&event->group_node, &groups->tree);
1594         init_event_group(event);
1595 }
1596 
1597 /*
1598  * Helper function to delete event from its groups.
1599  */
1600 static void
1601 del_event_from_groups(struct perf_event *event, struct perf_event_context *ctx)
1602 {
1603         struct perf_event_groups *groups;
1604 
1605         groups = get_event_groups(event, ctx);
1606         perf_event_groups_delete(groups, event);
1607 }
1608 
1609 /*
1610  * Get the leftmost event in the @cpu subtree.
1611  */
1612 static struct perf_event *
1613 perf_event_groups_first(struct perf_event_groups *groups, int cpu)
1614 {
1615         struct perf_event *node_event = NULL, *match = NULL;
1616         struct rb_node *node = groups->tree.rb_node;
1617 
1618         while (node) {
1619                 node_event = container_of(node, struct perf_event, group_node);
1620 
1621                 if (cpu < node_event->cpu) {
1622                         node = node->rb_left;
1623                 } else if (cpu > node_event->cpu) {
1624                         node = node->rb_right;
1625                 } else {
1626                         match = node_event;
1627                         node = node->rb_left;
1628                 }
1629         }
1630 
1631         return match;
1632 }
1633 
1634 /*
1635  * Like rb_entry_next_safe() for the @cpu subtree.
1636  */
1637 static struct perf_event *
1638 perf_event_groups_next(struct perf_event *event)
1639 {
1640         struct perf_event *next;
1641 
1642         next = rb_entry_safe(rb_next(&event->group_node), typeof(*event), group_node);
1643         if (next && next->cpu == event->cpu)
1644                 return next;
1645 
1646         return NULL;
1647 }
1648 
1649 /*
1650  * Iterate through the whole groups tree.
1651  */
1652 #define perf_event_groups_for_each(event, groups)                       \
1653         for (event = rb_entry_safe(rb_first(&((groups)->tree)),         \
1654                                 typeof(*event), group_node); event;     \
1655                 event = rb_entry_safe(rb_next(&event->group_node),      \
1656                                 typeof(*event), group_node))
1657 
1658 /*
1659  * Add an event from the lists for its context.
1660  * Must be called with ctx->mutex and ctx->lock held.
1661  */
1662 static void
1663 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1664 {
1665         lockdep_assert_held(&ctx->lock);
1666 
1667         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1668         event->attach_state |= PERF_ATTACH_CONTEXT;
1669 
1670         event->tstamp = perf_event_time(event);
1671 
1672         /*
1673          * If we're a stand alone event or group leader, we go to the context
1674          * list, group events are kept attached to the group so that
1675          * perf_group_detach can, at all times, locate all siblings.
1676          */
1677         if (event->group_leader == event) {
1678                 event->group_caps = event->event_caps;
1679                 add_event_to_groups(event, ctx);
1680         }
1681 
1682         list_update_cgroup_event(event, ctx, true);
1683 
1684         list_add_rcu(&event->event_entry, &ctx->event_list);
1685         ctx->nr_events++;
1686         if (event->attr.inherit_stat)
1687                 ctx->nr_stat++;
1688 
1689         ctx->generation++;
1690 }
1691 
1692 /*
1693  * Initialize event state based on the perf_event_attr::disabled.
1694  */
1695 static inline void perf_event__state_init(struct perf_event *event)
1696 {
1697         event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1698                                               PERF_EVENT_STATE_INACTIVE;
1699 }
1700 
1701 static void __perf_event_read_size(struct perf_event *event, int nr_siblings)
1702 {
1703         int entry = sizeof(u64); /* value */
1704         int size = 0;
1705         int nr = 1;
1706 
1707         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1708                 size += sizeof(u64);
1709 
1710         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1711                 size += sizeof(u64);
1712 
1713         if (event->attr.read_format & PERF_FORMAT_ID)
1714                 entry += sizeof(u64);
1715 
1716         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1717                 nr += nr_siblings;
1718                 size += sizeof(u64);
1719         }
1720 
1721         size += entry * nr;
1722         event->read_size = size;
1723 }
1724 
1725 static void __perf_event_header_size(struct perf_event *event, u64 sample_type)
1726 {
1727         struct perf_sample_data *data;
1728         u16 size = 0;
1729 
1730         if (sample_type & PERF_SAMPLE_IP)
1731                 size += sizeof(data->ip);
1732 
1733         if (sample_type & PERF_SAMPLE_ADDR)
1734                 size += sizeof(data->addr);
1735 
1736         if (sample_type & PERF_SAMPLE_PERIOD)
1737                 size += sizeof(data->period);
1738 
1739         if (sample_type & PERF_SAMPLE_WEIGHT)
1740                 size += sizeof(data->weight);
1741 
1742         if (sample_type & PERF_SAMPLE_READ)
1743                 size += event->read_size;
1744 
1745         if (sample_type & PERF_SAMPLE_DATA_SRC)
1746                 size += sizeof(data->data_src.val);
1747 
1748         if (sample_type & PERF_SAMPLE_TRANSACTION)
1749                 size += sizeof(data->txn);
1750 
1751         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
1752                 size += sizeof(data->phys_addr);
1753 
1754         event->header_size = size;
1755 }
1756 
1757 /*
1758  * Called at perf_event creation and when events are attached/detached from a
1759  * group.
1760  */
1761 static void perf_event__header_size(struct perf_event *event)
1762 {
1763         __perf_event_read_size(event,
1764                                event->group_leader->nr_siblings);
1765         __perf_event_header_size(event, event->attr.sample_type);
1766 }
1767 
1768 static void perf_event__id_header_size(struct perf_event *event)
1769 {
1770         struct perf_sample_data *data;
1771         u64 sample_type = event->attr.sample_type;
1772         u16 size = 0;
1773 
1774         if (sample_type & PERF_SAMPLE_TID)
1775                 size += sizeof(data->tid_entry);
1776 
1777         if (sample_type & PERF_SAMPLE_TIME)
1778                 size += sizeof(data->time);
1779 
1780         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1781                 size += sizeof(data->id);
1782 
1783         if (sample_type & PERF_SAMPLE_ID)
1784                 size += sizeof(data->id);
1785 
1786         if (sample_type & PERF_SAMPLE_STREAM_ID)
1787                 size += sizeof(data->stream_id);
1788 
1789         if (sample_type & PERF_SAMPLE_CPU)
1790                 size += sizeof(data->cpu_entry);
1791 
1792         event->id_header_size = size;
1793 }
1794 
1795 static bool perf_event_validate_size(struct perf_event *event)
1796 {
1797         /*
1798          * The values computed here will be over-written when we actually
1799          * attach the event.
1800          */
1801         __perf_event_read_size(event, event->group_leader->nr_siblings + 1);
1802         __perf_event_header_size(event, event->attr.sample_type & ~PERF_SAMPLE_READ);
1803         perf_event__id_header_size(event);
1804 
1805         /*
1806          * Sum the lot; should not exceed the 64k limit we have on records.
1807          * Conservative limit to allow for callchains and other variable fields.
1808          */
1809         if (event->read_size + event->header_size +
1810             event->id_header_size + sizeof(struct perf_event_header) >= 16*1024)
1811                 return false;
1812 
1813         return true;
1814 }
1815 
1816 static void perf_group_attach(struct perf_event *event)
1817 {
1818         struct perf_event *group_leader = event->group_leader, *pos;
1819 
1820         lockdep_assert_held(&event->ctx->lock);
1821 
1822         /*
1823          * We can have double attach due to group movement in perf_event_open.
1824          */
1825         if (event->attach_state & PERF_ATTACH_GROUP)
1826                 return;
1827 
1828         event->attach_state |= PERF_ATTACH_GROUP;
1829 
1830         if (group_leader == event)
1831                 return;
1832 
1833         WARN_ON_ONCE(group_leader->ctx != event->ctx);
1834 
1835         group_leader->group_caps &= event->event_caps;
1836 
1837         list_add_tail(&event->sibling_list, &group_leader->sibling_list);
1838         group_leader->nr_siblings++;
1839 
1840         perf_event__header_size(group_leader);
1841 
1842         for_each_sibling_event(pos, group_leader)
1843                 perf_event__header_size(pos);
1844 }
1845 
1846 /*
1847  * Remove an event from the lists for its context.
1848  * Must be called with ctx->mutex and ctx->lock held.
1849  */
1850 static void
1851 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1852 {
1853         WARN_ON_ONCE(event->ctx != ctx);
1854         lockdep_assert_held(&ctx->lock);
1855 
1856         /*
1857          * We can have double detach due to exit/hot-unplug + close.
1858          */
1859         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1860                 return;
1861 
1862         event->attach_state &= ~PERF_ATTACH_CONTEXT;
1863 
1864         list_update_cgroup_event(event, ctx, false);
1865 
1866         ctx->nr_events--;
1867         if (event->attr.inherit_stat)
1868                 ctx->nr_stat--;
1869 
1870         list_del_rcu(&event->event_entry);
1871 
1872         if (event->group_leader == event)
1873                 del_event_from_groups(event, ctx);
1874 
1875         /*
1876          * If event was in error state, then keep it
1877          * that way, otherwise bogus counts will be
1878          * returned on read(). The only way to get out
1879          * of error state is by explicit re-enabling
1880          * of the event
1881          */
1882         if (event->state > PERF_EVENT_STATE_OFF)
1883                 perf_event_set_state(event, PERF_EVENT_STATE_OFF);
1884 
1885         ctx->generation++;
1886 }
1887 
1888 static void perf_group_detach(struct perf_event *event)
1889 {
1890         struct perf_event *sibling, *tmp;
1891         struct perf_event_context *ctx = event->ctx;
1892 
1893         lockdep_assert_held(&ctx->lock);
1894 
1895         /*
1896          * We can have double detach due to exit/hot-unplug + close.
1897          */
1898         if (!(event->attach_state & PERF_ATTACH_GROUP))
1899                 return;
1900 
1901         event->attach_state &= ~PERF_ATTACH_GROUP;
1902 
1903         /*
1904          * If this is a sibling, remove it from its group.
1905          */
1906         if (event->group_leader != event) {
1907                 list_del_init(&event->sibling_list);
1908                 event->group_leader->nr_siblings--;
1909                 goto out;
1910         }
1911 
1912         /*
1913          * If this was a group event with sibling events then
1914          * upgrade the siblings to singleton events by adding them
1915          * to whatever list we are on.
1916          */
1917         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, sibling_list) {
1918 
1919                 sibling->group_leader = sibling;
1920                 list_del_init(&sibling->sibling_list);
1921 
1922                 /* Inherit group flags from the previous leader */
1923                 sibling->group_caps = event->group_caps;
1924 
1925                 if (!RB_EMPTY_NODE(&event->group_node)) {
1926                         add_event_to_groups(sibling, event->ctx);
1927 
1928                         if (sibling->state == PERF_EVENT_STATE_ACTIVE) {
1929                                 struct list_head *list = sibling->attr.pinned ?
1930                                         &ctx->pinned_active : &ctx->flexible_active;
1931 
1932                                 list_add_tail(&sibling->active_list, list);
1933                         }
1934                 }
1935 
1936                 WARN_ON_ONCE(sibling->ctx != event->ctx);
1937         }
1938 
1939 out:
1940         perf_event__header_size(event->group_leader);
1941 
1942         for_each_sibling_event(tmp, event->group_leader)
1943                 perf_event__header_size(tmp);
1944 }
1945 
1946 static bool is_orphaned_event(struct perf_event *event)
1947 {
1948         return event->state == PERF_EVENT_STATE_DEAD;
1949 }
1950 
1951 static inline int __pmu_filter_match(struct perf_event *event)
1952 {
1953         struct pmu *pmu = event->pmu;
1954         return pmu->filter_match ? pmu->filter_match(event) : 1;
1955 }
1956 
1957 /*
1958  * Check whether we should attempt to schedule an event group based on
1959  * PMU-specific filtering. An event group can consist of HW and SW events,
1960  * potentially with a SW leader, so we must check all the filters, to
1961  * determine whether a group is schedulable:
1962  */
1963 static inline int pmu_filter_match(struct perf_event *event)
1964 {
1965         struct perf_event *sibling;
1966 
1967         if (!__pmu_filter_match(event))
1968                 return 0;
1969 
1970         for_each_sibling_event(sibling, event) {
1971                 if (!__pmu_filter_match(sibling))
1972                         return 0;
1973         }
1974 
1975         return 1;
1976 }
1977 
1978 static inline int
1979 event_filter_match(struct perf_event *event)
1980 {
1981         return (event->cpu == -1 || event->cpu == smp_processor_id()) &&
1982                perf_cgroup_match(event) && pmu_filter_match(event);
1983 }
1984 
1985 static void
1986 event_sched_out(struct perf_event *event,
1987                   struct perf_cpu_context *cpuctx,
1988                   struct perf_event_context *ctx)
1989 {
1990         enum perf_event_state state = PERF_EVENT_STATE_INACTIVE;
1991 
1992         WARN_ON_ONCE(event->ctx != ctx);
1993         lockdep_assert_held(&ctx->lock);
1994 
1995         if (event->state != PERF_EVENT_STATE_ACTIVE)
1996                 return;
1997 
1998         /*
1999          * Asymmetry; we only schedule events _IN_ through ctx_sched_in(), but
2000          * we can schedule events _OUT_ individually through things like
2001          * __perf_remove_from_context().
2002          */
2003         list_del_init(&event->active_list);
2004 
2005         perf_pmu_disable(event->pmu);
2006 
2007         event->pmu->del(event, 0);
2008         event->oncpu = -1;
2009 
2010         if (READ_ONCE(event->pending_disable) >= 0) {
2011                 WRITE_ONCE(event->pending_disable, -1);
2012                 state = PERF_EVENT_STATE_OFF;
2013         }
2014         perf_event_set_state(event, state);
2015 
2016         if (!is_software_event(event))
2017                 cpuctx->active_oncpu--;
2018         if (!--ctx->nr_active)
2019                 perf_event_ctx_deactivate(ctx);
2020         if (event->attr.freq && event->attr.sample_freq)
2021                 ctx->nr_freq--;
2022         if (event->attr.exclusive || !cpuctx->active_oncpu)
2023                 cpuctx->exclusive = 0;
2024 
2025         perf_pmu_enable(event->pmu);
2026 }
2027 
2028 static void
2029 group_sched_out(struct perf_event *group_event,
2030                 struct perf_cpu_context *cpuctx,
2031                 struct perf_event_context *ctx)
2032 {
2033         struct perf_event *event;
2034 
2035         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
2036                 return;
2037 
2038         perf_pmu_disable(ctx->pmu);
2039 
2040         event_sched_out(group_event, cpuctx, ctx);
2041 
2042         /*
2043          * Schedule out siblings (if any):
2044          */
2045         for_each_sibling_event(event, group_event)
2046                 event_sched_out(event, cpuctx, ctx);
2047 
2048         perf_pmu_enable(ctx->pmu);
2049 
2050         if (group_event->attr.exclusive)
2051                 cpuctx->exclusive = 0;
2052 }
2053 
2054 #define DETACH_GROUP    0x01UL
2055 
2056 /*
2057  * Cross CPU call to remove a performance event
2058  *
2059  * We disable the event on the hardware level first. After that we
2060  * remove it from the context list.
2061  */
2062 static void
2063 __perf_remove_from_context(struct perf_event *event,
2064                            struct perf_cpu_context *cpuctx,
2065                            struct perf_event_context *ctx,
2066                            void *info)
2067 {
2068         unsigned long flags = (unsigned long)info;
2069 
2070         if (ctx->is_active & EVENT_TIME) {
2071                 update_context_time(ctx);
2072                 update_cgrp_time_from_cpuctx(cpuctx);
2073         }
2074 
2075         event_sched_out(event, cpuctx, ctx);
2076         if (flags & DETACH_GROUP)
2077                 perf_group_detach(event);
2078         list_del_event(event, ctx);
2079 
2080         if (!ctx->nr_events && ctx->is_active) {
2081                 ctx->is_active = 0;
2082                 if (ctx->task) {
2083                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
2084                         cpuctx->task_ctx = NULL;
2085                 }
2086         }
2087 }
2088 
2089 /*
2090  * Remove the event from a task's (or a CPU's) list of events.
2091  *
2092  * If event->ctx is a cloned context, callers must make sure that
2093  * every task struct that event->ctx->task could possibly point to
2094  * remains valid.  This is OK when called from perf_release since
2095  * that only calls us on the top-level context, which can't be a clone.
2096  * When called from perf_event_exit_task, it's OK because the
2097  * context has been detached from its task.
2098  */
2099 static void perf_remove_from_context(struct perf_event *event, unsigned long flags)
2100 {
2101         struct perf_event_context *ctx = event->ctx;
2102 
2103         lockdep_assert_held(&ctx->mutex);
2104 
2105         event_function_call(event, __perf_remove_from_context, (void *)flags);
2106 
2107         /*
2108          * The above event_function_call() can NO-OP when it hits
2109          * TASK_TOMBSTONE. In that case we must already have been detached
2110          * from the context (by perf_event_exit_event()) but the grouping
2111          * might still be in-tact.
2112          */
2113         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
2114         if ((flags & DETACH_GROUP) &&
2115             (event->attach_state & PERF_ATTACH_GROUP)) {
2116                 /*
2117                  * Since in that case we cannot possibly be scheduled, simply
2118                  * detach now.
2119                  */
2120                 raw_spin_lock_irq(&ctx->lock);
2121                 perf_group_detach(event);
2122                 raw_spin_unlock_irq(&ctx->lock);
2123         }
2124 }
2125 
2126 /*
2127  * Cross CPU call to disable a performance event
2128  */
2129 static void __perf_event_disable(struct perf_event *event,
2130                                  struct perf_cpu_context *cpuctx,
2131                                  struct perf_event_context *ctx,
2132                                  void *info)
2133 {
2134         if (event->state < PERF_EVENT_STATE_INACTIVE)
2135                 return;
2136 
2137         if (ctx->is_active & EVENT_TIME) {
2138                 update_context_time(ctx);
2139                 update_cgrp_time_from_event(event);
2140         }
2141 
2142         if (event == event->group_leader)
2143                 group_sched_out(event, cpuctx, ctx);
2144         else
2145                 event_sched_out(event, cpuctx, ctx);
2146 
2147         perf_event_set_state(event, PERF_EVENT_STATE_OFF);
2148 }
2149 
2150 /*
2151  * Disable an event.
2152  *
2153  * If event->ctx is a cloned context, callers must make sure that
2154  * every task struct that event->ctx->task could possibly point to
2155  * remains valid.  This condition is satisifed when called through
2156  * perf_event_for_each_child or perf_event_for_each because they
2157  * hold the top-level event's child_mutex, so any descendant that
2158  * goes to exit will block in perf_event_exit_event().
2159  *
2160  * When called from perf_pending_event it's OK because event->ctx
2161  * is the current context on this CPU and preemption is disabled,
2162  * hence we can't get into perf_event_task_sched_out for this context.
2163  */
2164 static void _perf_event_disable(struct perf_event *event)
2165 {
2166         struct perf_event_context *ctx = event->ctx;
2167 
2168         raw_spin_lock_irq(&ctx->lock);
2169         if (event->state <= PERF_EVENT_STATE_OFF) {
2170                 raw_spin_unlock_irq(&ctx->lock);
2171                 return;
2172         }
2173         raw_spin_unlock_irq(&ctx->lock);
2174 
2175         event_function_call(event, __perf_event_disable, NULL);
2176 }
2177 
2178 void perf_event_disable_local(struct perf_event *event)
2179 {
2180         event_function_local(event, __perf_event_disable, NULL);
2181 }
2182 
2183 /*
2184  * Strictly speaking kernel users cannot create groups and therefore this
2185  * interface does not need the perf_event_ctx_lock() magic.
2186  */
2187 void perf_event_disable(struct perf_event *event)
2188 {
2189         struct perf_event_context *ctx;
2190 
2191         ctx = perf_event_ctx_lock(event);
2192         _perf_event_disable(event);
2193         perf_event_ctx_unlock(event, ctx);
2194 }
2195 EXPORT_SYMBOL_GPL(perf_event_disable);
2196 
2197 void perf_event_disable_inatomic(struct perf_event *event)
2198 {
2199         WRITE_ONCE(event->pending_disable, smp_processor_id());
2200         /* can fail, see perf_pending_event_disable() */
2201         irq_work_queue(&event->pending);
2202 }
2203 
2204 static void perf_set_shadow_time(struct perf_event *event,
2205                                  struct perf_event_context *ctx)
2206 {
2207         /*
2208          * use the correct time source for the time snapshot
2209          *
2210          * We could get by without this by leveraging the
2211          * fact that to get to this function, the caller
2212          * has most likely already called update_context_time()
2213          * and update_cgrp_time_xx() and thus both timestamp
2214          * are identical (or very close). Given that tstamp is,
2215          * already adjusted for cgroup, we could say that:
2216          *    tstamp - ctx->timestamp
2217          * is equivalent to
2218          *    tstamp - cgrp->timestamp.
2219          *
2220          * Then, in perf_output_read(), the calculation would
2221          * work with no changes because:
2222          * - event is guaranteed scheduled in
2223          * - no scheduled out in between
2224          * - thus the timestamp would be the same
2225          *
2226          * But this is a bit hairy.
2227          *
2228          * So instead, we have an explicit cgroup call to remain
2229          * within the time time source all along. We believe it
2230          * is cleaner and simpler to understand.
2231          */
2232         if (is_cgroup_event(event))
2233                 perf_cgroup_set_shadow_time(event, event->tstamp);
2234         else
2235                 event->shadow_ctx_time = event->tstamp - ctx->timestamp;
2236 }
2237 
2238 #define MAX_INTERRUPTS (~0ULL)
2239 
2240 static void perf_log_throttle(struct perf_event *event, int enable);
2241 static void perf_log_itrace_start(struct perf_event *event);
2242 
2243 static int
2244 event_sched_in(struct perf_event *event,
2245                  struct perf_cpu_context *cpuctx,
2246                  struct perf_event_context *ctx)
2247 {
2248         int ret = 0;
2249 
2250         lockdep_assert_held(&ctx->lock);
2251 
2252         if (event->state <= PERF_EVENT_STATE_OFF)
2253                 return 0;
2254 
2255         WRITE_ONCE(event->oncpu, smp_processor_id());
2256         /*
2257          * Order event::oncpu write to happen before the ACTIVE state is
2258          * visible. This allows perf_event_{stop,read}() to observe the correct
2259          * ->oncpu if it sees ACTIVE.
2260          */
2261         smp_wmb();
2262         perf_event_set_state(event, PERF_EVENT_STATE_ACTIVE);
2263 
2264         /*
2265          * Unthrottle events, since we scheduled we might have missed several
2266          * ticks already, also for a heavily scheduling task there is little
2267          * guarantee it'll get a tick in a timely manner.
2268          */
2269         if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
2270                 perf_log_throttle(event, 1);
2271                 event->hw.interrupts = 0;
2272         }
2273 
2274         perf_pmu_disable(event->pmu);
2275 
2276         perf_set_shadow_time(event, ctx);
2277 
2278         perf_log_itrace_start(event);
2279 
2280         if (event->pmu->add(event, PERF_EF_START)) {
2281                 perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2282                 event->oncpu = -1;
2283                 ret = -EAGAIN;
2284                 goto out;
2285         }
2286 
2287         if (!is_software_event(event))
2288                 cpuctx->active_oncpu++;
2289         if (!ctx->nr_active++)
2290                 perf_event_ctx_activate(ctx);
2291         if (event->attr.freq && event->attr.sample_freq)
2292                 ctx->nr_freq++;
2293 
2294         if (event->attr.exclusive)
2295                 cpuctx->exclusive = 1;
2296 
2297 out:
2298         perf_pmu_enable(event->pmu);
2299 
2300         return ret;
2301 }
2302 
2303 static int
2304 group_sched_in(struct perf_event *group_event,
2305                struct perf_cpu_context *cpuctx,
2306                struct perf_event_context *ctx)
2307 {
2308         struct perf_event *event, *partial_group = NULL;
2309         struct pmu *pmu = ctx->pmu;
2310 
2311         if (group_event->state == PERF_EVENT_STATE_OFF)
2312                 return 0;
2313 
2314         pmu->start_txn(pmu, PERF_PMU_TXN_ADD);
2315 
2316         if (event_sched_in(group_event, cpuctx, ctx)) {
2317                 pmu->cancel_txn(pmu);
2318                 perf_mux_hrtimer_restart(cpuctx);
2319                 return -EAGAIN;
2320         }
2321 
2322         /*
2323          * Schedule in siblings as one group (if any):
2324          */
2325         for_each_sibling_event(event, group_event) {
2326                 if (event_sched_in(event, cpuctx, ctx)) {
2327                         partial_group = event;
2328                         goto group_error;
2329                 }
2330         }
2331 
2332         if (!pmu->commit_txn(pmu))
2333                 return 0;
2334 
2335 group_error:
2336         /*
2337          * Groups can be scheduled in as one unit only, so undo any
2338          * partial group before returning:
2339          * The events up to the failed event are scheduled out normally.
2340          */
2341         for_each_sibling_event(event, group_event) {
2342                 if (event == partial_group)
2343                         break;
2344 
2345                 event_sched_out(event, cpuctx, ctx);
2346         }
2347         event_sched_out(group_event, cpuctx, ctx);
2348 
2349         pmu->cancel_txn(pmu);
2350 
2351         perf_mux_hrtimer_restart(cpuctx);
2352 
2353         return -EAGAIN;
2354 }
2355 
2356 /*
2357  * Work out whether we can put this event group on the CPU now.
2358  */
2359 static int group_can_go_on(struct perf_event *event,
2360                            struct perf_cpu_context *cpuctx,
2361                            int can_add_hw)
2362 {
2363         /*
2364          * Groups consisting entirely of software events can always go on.
2365          */
2366         if (event->group_caps & PERF_EV_CAP_SOFTWARE)
2367                 return 1;
2368         /*
2369          * If an exclusive group is already on, no other hardware
2370          * events can go on.
2371          */
2372         if (cpuctx->exclusive)
2373                 return 0;
2374         /*
2375          * If this group is exclusive and there are already
2376          * events on the CPU, it can't go on.
2377          */
2378         if (event->attr.exclusive && cpuctx->active_oncpu)
2379                 return 0;
2380         /*
2381          * Otherwise, try to add it if all previous groups were able
2382          * to go on.
2383          */
2384         return can_add_hw;
2385 }
2386 
2387 static void add_event_to_ctx(struct perf_event *event,
2388                                struct perf_event_context *ctx)
2389 {
2390         list_add_event(event, ctx);
2391         perf_group_attach(event);
2392 }
2393 
2394 static void ctx_sched_out(struct perf_event_context *ctx,
2395                           struct perf_cpu_context *cpuctx,
2396                           enum event_type_t event_type);
2397 static void
2398 ctx_sched_in(struct perf_event_context *ctx,
2399              struct perf_cpu_context *cpuctx,
2400              enum event_type_t event_type,
2401              struct task_struct *task);
2402 
2403 static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
2404                                struct perf_event_context *ctx,
2405                                enum event_type_t event_type)
2406 {
2407         if (!cpuctx->task_ctx)
2408                 return;
2409 
2410         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2411                 return;
2412 
2413         ctx_sched_out(ctx, cpuctx, event_type);
2414 }
2415 
2416 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2417                                 struct perf_event_context *ctx,
2418                                 struct task_struct *task)
2419 {
2420         cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2421         if (ctx)
2422                 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2423         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2424         if (ctx)
2425                 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2426 }
2427 
2428 /*
2429  * We want to maintain the following priority of scheduling:
2430  *  - CPU pinned (EVENT_CPU | EVENT_PINNED)
2431  *  - task pinned (EVENT_PINNED)
2432  *  - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
2433  *  - task flexible (EVENT_FLEXIBLE).
2434  *
2435  * In order to avoid unscheduling and scheduling back in everything every
2436  * time an event is added, only do it for the groups of equal priority and
2437  * below.
2438  *
2439  * This can be called after a batch operation on task events, in which case
2440  * event_type is a bit mask of the types of events involved. For CPU events,
2441  * event_type is only either EVENT_PINNED or EVENT_FLEXIBLE.
2442  */
2443 static void ctx_resched(struct perf_cpu_context *cpuctx,
2444                         struct perf_event_context *task_ctx,
2445                         enum event_type_t event_type)
2446 {
2447         enum event_type_t ctx_event_type;
2448         bool cpu_event = !!(event_type & EVENT_CPU);
2449 
2450         /*
2451          * If pinned groups are involved, flexible groups also need to be
2452          * scheduled out.
2453          */
2454         if (event_type & EVENT_PINNED)
2455                 event_type |= EVENT_FLEXIBLE;
2456 
2457         ctx_event_type = event_type & EVENT_ALL;
2458 
2459         perf_pmu_disable(cpuctx->ctx.pmu);
2460         if (task_ctx)
2461                 task_ctx_sched_out(cpuctx, task_ctx, event_type);
2462 
2463         /*
2464          * Decide which cpu ctx groups to schedule out based on the types
2465          * of events that caused rescheduling:
2466          *  - EVENT_CPU: schedule out corresponding groups;
2467          *  - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
2468          *  - otherwise, do nothing more.
2469          */
2470         if (cpu_event)
2471                 cpu_ctx_sched_out(cpuctx, ctx_event_type);
2472         else if (ctx_event_type & EVENT_PINNED)
2473                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2474 
2475         perf_event_sched_in(cpuctx, task_ctx, current);
2476         perf_pmu_enable(cpuctx->ctx.pmu);
2477 }
2478 
2479 /*
2480  * Cross CPU call to install and enable a performance event
2481  *
2482  * Very similar to remote_function() + event_function() but cannot assume that
2483  * things like ctx->is_active and cpuctx->task_ctx are set.
2484  */
2485 static int  __perf_install_in_context(void *info)
2486 {
2487         struct perf_event *event = info;
2488         struct perf_event_context *ctx = event->ctx;
2489         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2490         struct perf_event_context *task_ctx = cpuctx->task_ctx;
2491         bool reprogram = true;
2492         int ret = 0;
2493 
2494         raw_spin_lock(&cpuctx->ctx.lock);
2495         if (ctx->task) {
2496                 raw_spin_lock(&ctx->lock);
2497                 task_ctx = ctx;
2498 
2499                 reprogram = (ctx->task == current);
2500 
2501                 /*
2502                  * If the task is running, it must be running on this CPU,
2503                  * otherwise we cannot reprogram things.
2504                  *
2505                  * If its not running, we don't care, ctx->lock will
2506                  * serialize against it becoming runnable.
2507                  */
2508                 if (task_curr(ctx->task) && !reprogram) {
2509                         ret = -ESRCH;
2510                         goto unlock;
2511                 }
2512 
2513                 WARN_ON_ONCE(reprogram && cpuctx->task_ctx && cpuctx->task_ctx != ctx);
2514         } else if (task_ctx) {
2515                 raw_spin_lock(&task_ctx->lock);
2516         }
2517 
2518 #ifdef CONFIG_CGROUP_PERF
2519         if (is_cgroup_event(event)) {
2520                 /*
2521                  * If the current cgroup doesn't match the event's
2522                  * cgroup, we should not try to schedule it.
2523                  */
2524                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
2525                 reprogram = cgroup_is_descendant(cgrp->css.cgroup,
2526                                         event->cgrp->css.cgroup);
2527         }
2528 #endif
2529 
2530         if (reprogram) {
2531                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2532                 add_event_to_ctx(event, ctx);
2533                 ctx_resched(cpuctx, task_ctx, get_event_type(event));
2534         } else {
2535                 add_event_to_ctx(event, ctx);
2536         }
2537 
2538 unlock:
2539         perf_ctx_unlock(cpuctx, task_ctx);
2540 
2541         return ret;
2542 }
2543 
2544 /*
2545  * Attach a performance event to a context.
2546  *
2547  * Very similar to event_function_call, see comment there.
2548  */
2549 static void
2550 perf_install_in_context(struct perf_event_context *ctx,
2551                         struct perf_event *event,
2552                         int cpu)
2553 {
2554         struct task_struct *task = READ_ONCE(ctx->task);
2555 
2556         lockdep_assert_held(&ctx->mutex);
2557 
2558         if (event->cpu != -1)
2559                 event->cpu = cpu;
2560 
2561         /*
2562          * Ensures that if we can observe event->ctx, both the event and ctx
2563          * will be 'complete'. See perf_iterate_sb_cpu().
2564          */
2565         smp_store_release(&event->ctx, ctx);
2566 
2567         if (!task) {
2568                 cpu_function_call(cpu, __perf_install_in_context, event);
2569                 return;
2570         }
2571 
2572         /*
2573          * Should not happen, we validate the ctx is still alive before calling.
2574          */
2575         if (WARN_ON_ONCE(task == TASK_TOMBSTONE))
2576                 return;
2577 
2578         /*
2579          * Installing events is tricky because we cannot rely on ctx->is_active
2580          * to be set in case this is the nr_events 0 -> 1 transition.
2581          *
2582          * Instead we use task_curr(), which tells us if the task is running.
2583          * However, since we use task_curr() outside of rq::lock, we can race
2584          * against the actual state. This means the result can be wrong.
2585          *
2586          * If we get a false positive, we retry, this is harmless.
2587          *
2588          * If we get a false negative, things are complicated. If we are after
2589          * perf_event_context_sched_in() ctx::lock will serialize us, and the
2590          * value must be correct. If we're before, it doesn't matter since
2591          * perf_event_context_sched_in() will program the counter.
2592          *
2593          * However, this hinges on the remote context switch having observed
2594          * our task->perf_event_ctxp[] store, such that it will in fact take
2595          * ctx::lock in perf_event_context_sched_in().
2596          *
2597          * We do this by task_function_call(), if the IPI fails to hit the task
2598          * we know any future context switch of task must see the
2599          * perf_event_ctpx[] store.
2600          */
2601 
2602         /*
2603          * This smp_mb() orders the task->perf_event_ctxp[] store with the
2604          * task_cpu() load, such that if the IPI then does not find the task
2605          * running, a future context switch of that task must observe the
2606          * store.
2607          */
2608         smp_mb();
2609 again:
2610         if (!task_function_call(task, __perf_install_in_context, event))
2611                 return;
2612 
2613         raw_spin_lock_irq(&ctx->lock);
2614         task = ctx->task;
2615         if (WARN_ON_ONCE(task == TASK_TOMBSTONE)) {
2616                 /*
2617                  * Cannot happen because we already checked above (which also
2618                  * cannot happen), and we hold ctx->mutex, which serializes us
2619                  * against perf_event_exit_task_context().
2620                  */
2621                 raw_spin_unlock_irq(&ctx->lock);
2622                 return;
2623         }
2624         /*
2625          * If the task is not running, ctx->lock will avoid it becoming so,
2626          * thus we can safely install the event.
2627          */
2628         if (task_curr(task)) {
2629                 raw_spin_unlock_irq(&ctx->lock);
2630                 goto again;
2631         }
2632         add_event_to_ctx(event, ctx);
2633         raw_spin_unlock_irq(&ctx->lock);
2634 }
2635 
2636 /*
2637  * Cross CPU call to enable a performance event
2638  */
2639 static void __perf_event_enable(struct perf_event *event,
2640                                 struct perf_cpu_context *cpuctx,
2641                                 struct perf_event_context *ctx,
2642                                 void *info)
2643 {
2644         struct perf_event *leader = event->group_leader;
2645         struct perf_event_context *task_ctx;
2646 
2647         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2648             event->state <= PERF_EVENT_STATE_ERROR)
2649                 return;
2650 
2651         if (ctx->is_active)
2652                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2653 
2654         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2655 
2656         if (!ctx->is_active)
2657                 return;
2658 
2659         if (!event_filter_match(event)) {
2660                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2661                 return;
2662         }
2663 
2664         /*
2665          * If the event is in a group and isn't the group leader,
2666          * then don't put it on unless the group is on.
2667          */
2668         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) {
2669                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2670                 return;
2671         }
2672 
2673         task_ctx = cpuctx->task_ctx;
2674         if (ctx->task)
2675                 WARN_ON_ONCE(task_ctx != ctx);
2676 
2677         ctx_resched(cpuctx, task_ctx, get_event_type(event));
2678 }
2679 
2680 /*
2681  * Enable an event.
2682  *
2683  * If event->ctx is a cloned context, callers must make sure that
2684  * every task struct that event->ctx->task could possibly point to
2685  * remains valid.  This condition is satisfied when called through
2686  * perf_event_for_each_child or perf_event_for_each as described
2687  * for perf_event_disable.
2688  */
2689 static void _perf_event_enable(struct perf_event *event)
2690 {
2691         struct perf_event_context *ctx = event->ctx;
2692 
2693         raw_spin_lock_irq(&ctx->lock);
2694         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2695             event->state <  PERF_EVENT_STATE_ERROR) {
2696                 raw_spin_unlock_irq(&ctx->lock);
2697                 return;
2698         }
2699 
2700         /*
2701          * If the event is in error state, clear that first.
2702          *
2703          * That way, if we see the event in error state below, we know that it
2704          * has gone back into error state, as distinct from the task having
2705          * been scheduled away before the cross-call arrived.
2706          */
2707         if (event->state == PERF_EVENT_STATE_ERROR)
2708                 event->state = PERF_EVENT_STATE_OFF;
2709         raw_spin_unlock_irq(&ctx->lock);
2710 
2711         event_function_call(event, __perf_event_enable, NULL);
2712 }
2713 
2714 /*
2715  * See perf_event_disable();
2716  */
2717 void perf_event_enable(struct perf_event *event)
2718 {
2719         struct perf_event_context *ctx;
2720 
2721         ctx = perf_event_ctx_lock(event);
2722         _perf_event_enable(event);
2723         perf_event_ctx_unlock(event, ctx);
2724 }
2725 EXPORT_SYMBOL_GPL(perf_event_enable);
2726 
2727 struct stop_event_data {
2728         struct perf_event       *event;
2729         unsigned int            restart;
2730 };
2731 
2732 static int __perf_event_stop(void *info)
2733 {
2734         struct stop_event_data *sd = info;
2735         struct perf_event *event = sd->event;
2736 
2737         /* if it's already INACTIVE, do nothing */
2738         if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2739                 return 0;
2740 
2741         /* matches smp_wmb() in event_sched_in() */
2742         smp_rmb();
2743 
2744         /*
2745          * There is a window with interrupts enabled before we get here,
2746          * so we need to check again lest we try to stop another CPU's event.
2747          */
2748         if (READ_ONCE(event->oncpu) != smp_processor_id())
2749                 return -EAGAIN;
2750 
2751         event->pmu->stop(event, PERF_EF_UPDATE);
2752 
2753         /*
2754          * May race with the actual stop (through perf_pmu_output_stop()),
2755          * but it is only used for events with AUX ring buffer, and such
2756          * events will refuse to restart because of rb::aux_mmap_count==0,
2757          * see comments in perf_aux_output_begin().
2758          *
2759          * Since this is happening on an event-local CPU, no trace is lost
2760          * while restarting.
2761          */
2762         if (sd->restart)
2763                 event->pmu->start(event, 0);
2764 
2765         return 0;
2766 }
2767 
2768 static int perf_event_stop(struct perf_event *event, int restart)
2769 {
2770         struct stop_event_data sd = {
2771                 .event          = event,
2772                 .restart        = restart,
2773         };
2774         int ret = 0;
2775 
2776         do {
2777                 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2778                         return 0;
2779 
2780                 /* matches smp_wmb() in event_sched_in() */
2781                 smp_rmb();
2782 
2783                 /*
2784                  * We only want to restart ACTIVE events, so if the event goes
2785                  * inactive here (event->oncpu==-1), there's nothing more to do;
2786                  * fall through with ret==-ENXIO.
2787                  */
2788                 ret = cpu_function_call(READ_ONCE(event->oncpu),
2789                                         __perf_event_stop, &sd);
2790         } while (ret == -EAGAIN);
2791 
2792         return ret;
2793 }
2794 
2795 /*
2796  * In order to contain the amount of racy and tricky in the address filter
2797  * configuration management, it is a two part process:
2798  *
2799  * (p1) when userspace mappings change as a result of (1) or (2) or (3) below,
2800  *      we update the addresses of corresponding vmas in
2801  *      event::addr_filters_offs array and bump the event::addr_filters_gen;
2802  * (p2) when an event is scheduled in (pmu::add), it calls
2803  *      perf_event_addr_filters_sync() which calls pmu::addr_filters_sync()
2804  *      if the generation has changed since the previous call.
2805  *
2806  * If (p1) happens while the event is active, we restart it to force (p2).
2807  *
2808  * (1) perf_addr_filters_apply(): adjusting filters' offsets based on
2809  *     pre-existing mappings, called once when new filters arrive via SET_FILTER
2810  *     ioctl;
2811  * (2) perf_addr_filters_adjust(): adjusting filters' offsets based on newly
2812  *     registered mapping, called for every new mmap(), with mm::mmap_sem down
2813  *     for reading;
2814  * (3) perf_event_addr_filters_exec(): clearing filters' offsets in the process
2815  *     of exec.
2816  */
2817 void perf_event_addr_filters_sync(struct perf_event *event)
2818 {
2819         struct perf_addr_filters_head *ifh = perf_event_addr_filters(event);
2820 
2821         if (!has_addr_filter(event))
2822                 return;
2823 
2824         raw_spin_lock(&ifh->lock);
2825         if (event->addr_filters_gen != event->hw.addr_filters_gen) {
2826                 event->pmu->addr_filters_sync(event);
2827                 event->hw.addr_filters_gen = event->addr_filters_gen;
2828         }
2829         raw_spin_unlock(&ifh->lock);
2830 }
2831 EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
2832 
2833 static int _perf_event_refresh(struct perf_event *event, int refresh)
2834 {
2835         /*
2836          * not supported on inherited events
2837          */
2838         if (event->attr.inherit || !is_sampling_event(event))
2839                 return -EINVAL;
2840 
2841         atomic_add(refresh, &event->event_limit);
2842         _perf_event_enable(event);
2843 
2844         return 0;
2845 }
2846 
2847 /*
2848  * See perf_event_disable()
2849  */
2850 int perf_event_refresh(struct perf_event *event, int refresh)
2851 {
2852         struct perf_event_context *ctx;
2853         int ret;
2854 
2855         ctx = perf_event_ctx_lock(event);
2856         ret = _perf_event_refresh(event, refresh);
2857         perf_event_ctx_unlock(event, ctx);
2858 
2859         return ret;
2860 }
2861 EXPORT_SYMBOL_GPL(perf_event_refresh);
2862 
2863 static int perf_event_modify_breakpoint(struct perf_event *bp,
2864                                          struct perf_event_attr *attr)
2865 {
2866         int err;
2867 
2868         _perf_event_disable(bp);
2869 
2870         err = modify_user_hw_breakpoint_check(bp, attr, true);
2871 
2872         if (!bp->attr.disabled)
2873                 _perf_event_enable(bp);
2874 
2875         return err;
2876 }
2877 
2878 static int perf_event_modify_attr(struct perf_event *event,
2879                                   struct perf_event_attr *attr)
2880 {
2881         if (event->attr.type != attr->type)
2882                 return -EINVAL;
2883 
2884         switch (event->attr.type) {
2885         case PERF_TYPE_BREAKPOINT:
2886                 return perf_event_modify_breakpoint(event, attr);
2887         default:
2888                 /* Place holder for future additions. */
2889                 return -EOPNOTSUPP;
2890         }
2891 }
2892 
2893 static void ctx_sched_out(struct perf_event_context *ctx,
2894                           struct perf_cpu_context *cpuctx,
2895                           enum event_type_t event_type)
2896 {
2897         struct perf_event *event, *tmp;
2898         int is_active = ctx->is_active;
2899 
2900         lockdep_assert_held(&ctx->lock);
2901 
2902         if (likely(!ctx->nr_events)) {
2903                 /*
2904                  * See __perf_remove_from_context().
2905                  */
2906                 WARN_ON_ONCE(ctx->is_active);
2907                 if (ctx->task)
2908                         WARN_ON_ONCE(cpuctx->task_ctx);
2909                 return;
2910         }
2911 
2912         ctx->is_active &= ~event_type;
2913         if (!(ctx->is_active & EVENT_ALL))
2914                 ctx->is_active = 0;
2915 
2916         if (ctx->task) {
2917                 WARN_ON_ONCE(cpuctx->task_ctx != ctx);
2918                 if (!ctx->is_active)
2919                         cpuctx->task_ctx = NULL;
2920         }
2921 
2922         /*
2923          * Always update time if it was set; not only when it changes.
2924          * Otherwise we can 'forget' to update time for any but the last
2925          * context we sched out. For example:
2926          *
2927          *   ctx_sched_out(.event_type = EVENT_FLEXIBLE)
2928          *   ctx_sched_out(.event_type = EVENT_PINNED)
2929          *
2930          * would only update time for the pinned events.
2931          */
2932         if (is_active & EVENT_TIME) {
2933                 /* update (and stop) ctx time */
2934                 update_context_time(ctx);
2935                 update_cgrp_time_from_cpuctx(cpuctx);
2936         }
2937 
2938         is_active ^= ctx->is_active; /* changed bits */
2939 
2940         if (!ctx->nr_active || !(is_active & EVENT_ALL))
2941                 return;
2942 
2943         perf_pmu_disable(ctx->pmu);
2944         if (is_active & EVENT_PINNED) {
2945                 list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list)
2946                         group_sched_out(event, cpuctx, ctx);
2947         }
2948 
2949         if (is_active & EVENT_FLEXIBLE) {
2950                 list_for_each_entry_safe(event, tmp, &ctx->flexible_active, active_list)
2951                         group_sched_out(event, cpuctx, ctx);
2952         }
2953         perf_pmu_enable(ctx->pmu);
2954 }
2955 
2956 /*
2957  * Test whether two contexts are equivalent, i.e. whether they have both been
2958  * cloned from the same version of the same context.
2959  *
2960  * Equivalence is measured using a generation number in the context that is
2961  * incremented on each modification to it; see unclone_ctx(), list_add_event()
2962  * and list_del_event().
2963  */
2964 static int context_equiv(struct perf_event_context *ctx1,
2965                          struct perf_event_context *ctx2)
2966 {
2967         lockdep_assert_held(&ctx1->lock);
2968         lockdep_assert_held(&ctx2->lock);
2969 
2970         /* Pinning disables the swap optimization */
2971         if (ctx1->pin_count || ctx2->pin_count)
2972                 return 0;
2973 
2974         /* If ctx1 is the parent of ctx2 */
2975         if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2976                 return 1;
2977 
2978         /* If ctx2 is the parent of ctx1 */
2979         if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2980                 return 1;
2981 
2982         /*
2983          * If ctx1 and ctx2 have the same parent; we flatten the parent
2984          * hierarchy, see perf_event_init_context().
2985          */
2986         if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2987                         ctx1->parent_gen == ctx2->parent_gen)
2988                 return 1;
2989 
2990         /* Unmatched */
2991         return 0;
2992 }
2993 
2994 static void __perf_event_sync_stat(struct perf_event *event,
2995                                      struct perf_event *next_event)
2996 {
2997         u64 value;
2998 
2999         if (!event->attr.inherit_stat)
3000                 return;
3001 
3002         /*
3003          * Update the event value, we cannot use perf_event_read()
3004          * because we're in the middle of a context switch and have IRQs
3005          * disabled, which upsets smp_call_function_single(), however
3006          * we know the event must be on the current CPU, therefore we
3007          * don't need to use it.
3008          */
3009         if (event->state == PERF_EVENT_STATE_ACTIVE)
3010                 event->pmu->read(event);
3011 
3012         perf_event_update_time(event);
3013 
3014         /*
3015          * In order to keep per-task stats reliable we need to flip the event
3016          * values when we flip the contexts.
3017          */
3018         value = local64_read(&next_event->count);
3019         value = local64_xchg(&event->count, value);
3020         local64_set(&next_event->count, value);
3021 
3022         swap(event->total_time_enabled, next_event->total_time_enabled);
3023         swap(event->total_time_running, next_event->total_time_running);
3024 
3025         /*
3026          * Since we swizzled the values, update the user visible data too.
3027          */
3028         perf_event_update_userpage(event);
3029         perf_event_update_userpage(next_event);
3030 }
3031 
3032 static void perf_event_sync_stat(struct perf_event_context *ctx,
3033                                    struct perf_event_context *next_ctx)
3034 {
3035         struct perf_event *event, *next_event;
3036 
3037         if (!ctx->nr_stat)
3038                 return;
3039 
3040         update_context_time(ctx);
3041 
3042         event = list_first_entry(&ctx->event_list,
3043                                    struct perf_event, event_entry);
3044 
3045         next_event = list_first_entry(&next_ctx->event_list,
3046                                         struct perf_event, event_entry);
3047 
3048         while (&event->event_entry != &ctx->event_list &&
3049                &next_event->event_entry != &next_ctx->event_list) {
3050 
3051                 __perf_event_sync_stat(event, next_event);
3052 
3053                 event = list_next_entry(event, event_entry);
3054                 next_event = list_next_entry(next_event, event_entry);
3055         }
3056 }
3057 
3058 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
3059                                          struct task_struct *next)
3060 {
3061         struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
3062         struct perf_event_context *next_ctx;
3063         struct perf_event_context *parent, *next_parent;
3064         struct perf_cpu_context *cpuctx;
3065         int do_switch = 1;
3066 
3067         if (likely(!ctx))
3068                 return;
3069 
3070         cpuctx = __get_cpu_context(ctx);
3071         if (!cpuctx->task_ctx)
3072                 return;
3073 
3074         rcu_read_lock();
3075         next_ctx = next->perf_event_ctxp[ctxn];
3076         if (!next_ctx)
3077                 goto unlock;
3078 
3079         parent = rcu_dereference(ctx->parent_ctx);
3080         next_parent = rcu_dereference(next_ctx->parent_ctx);
3081 
3082         /* If neither context have a parent context; they cannot be clones. */
3083         if (!parent && !next_parent)
3084                 goto unlock;
3085 
3086         if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
3087                 /*
3088                  * Looks like the two contexts are clones, so we might be
3089                  * able to optimize the context switch.  We lock both
3090                  * contexts and check that they are clones under the
3091                  * lock (including re-checking that neither has been
3092                  * uncloned in the meantime).  It doesn't matter which
3093                  * order we take the locks because no other cpu could
3094                  * be trying to lock both of these tasks.
3095                  */
3096                 raw_spin_lock(&ctx->lock);
3097                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
3098                 if (context_equiv(ctx, next_ctx)) {
3099                         WRITE_ONCE(ctx->task, next);
3100                         WRITE_ONCE(next_ctx->task, task);
3101 
3102                         swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
3103 
3104                         /*
3105                          * RCU_INIT_POINTER here is safe because we've not
3106                          * modified the ctx and the above modification of
3107                          * ctx->task and ctx->task_ctx_data are immaterial
3108                          * since those values are always verified under
3109                          * ctx->lock which we're now holding.
3110                          */
3111                         RCU_INIT_POINTER(task->perf_event_ctxp[ctxn], next_ctx);
3112                         RCU_INIT_POINTER(next->perf_event_ctxp[ctxn], ctx);
3113 
3114                         do_switch = 0;
3115 
3116                         perf_event_sync_stat(ctx, next_ctx);
3117                 }
3118                 raw_spin_unlock(&next_ctx->lock);
3119                 raw_spin_unlock(&ctx->lock);
3120         }
3121 unlock:
3122         rcu_read_unlock();
3123 
3124         if (do_switch) {
3125                 raw_spin_lock(&ctx->lock);
3126                 task_ctx_sched_out(cpuctx, ctx, EVENT_ALL);
3127                 raw_spin_unlock(&ctx->lock);
3128         }
3129 }
3130 
3131 static DEFINE_PER_CPU(struct list_head, sched_cb_list);
3132 
3133 void perf_sched_cb_dec(struct pmu *pmu)
3134 {
3135         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3136 
3137         this_cpu_dec(perf_sched_cb_usages);
3138 
3139         if (!--cpuctx->sched_cb_usage)
3140                 list_del(&cpuctx->sched_cb_entry);
3141 }
3142 
3143 
3144 void perf_sched_cb_inc(struct pmu *pmu)
3145 {
3146         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3147 
3148         if (!cpuctx->sched_cb_usage++)
3149                 list_add(&cpuctx->sched_cb_entry, this_cpu_ptr(&sched_cb_list));
3150 
3151         this_cpu_inc(perf_sched_cb_usages);
3152 }
3153 
3154 /*
3155  * This function provides the context switch callback to the lower code
3156  * layer. It is invoked ONLY when the context switch callback is enabled.
3157  *
3158  * This callback is relevant even to per-cpu events; for example multi event
3159  * PEBS requires this to provide PID/TID information. This requires we flush
3160  * all queued PEBS records before we context switch to a new task.
3161  */
3162 static void perf_pmu_sched_task(struct task_struct *prev,
3163                                 struct task_struct *next,
3164                                 bool sched_in)
3165 {
3166         struct perf_cpu_context *cpuctx;
3167         struct pmu *pmu;
3168 
3169         if (prev == next)
3170                 return;
3171 
3172         list_for_each_entry(cpuctx, this_cpu_ptr(&sched_cb_list), sched_cb_entry) {
3173                 pmu = cpuctx->ctx.pmu; /* software PMUs will not have sched_task */
3174 
3175                 if (WARN_ON_ONCE(!pmu->sched_task))
3176                         continue;
3177 
3178                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3179                 perf_pmu_disable(pmu);
3180 
3181                 pmu->sched_task(cpuctx->task_ctx, sched_in);
3182 
3183                 perf_pmu_enable(pmu);
3184                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3185         }
3186 }
3187 
3188 static void perf_event_switch(struct task_struct *task,
3189                               struct task_struct *next_prev, bool sched_in);
3190 
3191 #define for_each_task_context_nr(ctxn)                                  \
3192         for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
3193 
3194 /*
3195  * Called from scheduler to remove the events of the current task,
3196  * with interrupts disabled.
3197  *
3198  * We stop each event and update the event value in event->count.
3199  *
3200  * This does not protect us against NMI, but disable()
3201  * sets the disabled bit in the control field of event _before_
3202  * accessing the event control register. If a NMI hits, then it will
3203  * not restart the event.
3204  */
3205 void __perf_event_task_sched_out(struct task_struct *task,
3206                                  struct task_struct *next)
3207 {
3208         int ctxn;
3209 
3210         if (__this_cpu_read(perf_sched_cb_usages))
3211                 perf_pmu_sched_task(task, next, false);
3212 
3213         if (atomic_read(&nr_switch_events))
3214                 perf_event_switch(task, next, false);
3215 
3216         for_each_task_context_nr(ctxn)
3217                 perf_event_context_sched_out(task, ctxn, next);
3218 
3219         /*
3220          * if cgroup events exist on this CPU, then we need
3221          * to check if we have to switch out PMU state.
3222          * cgroup event are system-wide mode only
3223          */
3224         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3225                 perf_cgroup_sched_out(task, next);
3226 }
3227 
3228 /*
3229  * Called with IRQs disabled
3230  */
3231 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
3232                               enum event_type_t event_type)
3233 {
3234         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
3235 }
3236 
3237 static int visit_groups_merge(struct perf_event_groups *groups, int cpu,
3238                               int (*func)(struct perf_event *, void *), void *data)
3239 {
3240         struct perf_event **evt, *evt1, *evt2;
3241         int ret;
3242 
3243         evt1 = perf_event_groups_first(groups, -1);
3244         evt2 = perf_event_groups_first(groups, cpu);
3245 
3246         while (evt1 || evt2) {
3247                 if (evt1 && evt2) {
3248                         if (evt1->group_index < evt2->group_index)
3249                                 evt = &evt1;
3250                         else
3251                                 evt = &evt2;
3252                 } else if (evt1) {
3253                         evt = &evt1;
3254                 } else {
3255                         evt = &evt2;
3256                 }
3257 
3258                 ret = func(*evt, data);
3259                 if (ret)
3260                         return ret;
3261 
3262                 *evt = perf_event_groups_next(*evt);
3263         }
3264 
3265         return 0;
3266 }
3267 
3268 struct sched_in_data {
3269         struct perf_event_context *ctx;
3270         struct perf_cpu_context *cpuctx;
3271         int can_add_hw;
3272 };
3273 
3274 static int pinned_sched_in(struct perf_event *event, void *data)
3275 {
3276         struct sched_in_data *sid = data;
3277 
3278         if (event->state <= PERF_EVENT_STATE_OFF)
3279                 return 0;
3280 
3281         if (!event_filter_match(event))
3282                 return 0;
3283 
3284         if (group_can_go_on(event, sid->cpuctx, sid->can_add_hw)) {
3285                 if (!group_sched_in(event, sid->cpuctx, sid->ctx))
3286                         list_add_tail(&event->active_list, &sid->ctx->pinned_active);
3287         }
3288 
3289         /*
3290          * If this pinned group hasn't been scheduled,
3291          * put it in error state.
3292          */
3293         if (event->state == PERF_EVENT_STATE_INACTIVE)
3294                 perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
3295 
3296         return 0;
3297 }
3298 
3299 static int flexible_sched_in(struct perf_event *event, void *data)
3300 {
3301         struct sched_in_data *sid = data;
3302 
3303         if (event->state <= PERF_EVENT_STATE_OFF)
3304                 return 0;
3305 
3306         if (!event_filter_match(event))
3307                 return 0;
3308 
3309         if (group_can_go_on(event, sid->cpuctx, sid->can_add_hw)) {
3310                 if (!group_sched_in(event, sid->cpuctx, sid->ctx))
3311                         list_add_tail(&event->active_list, &sid->ctx->flexible_active);
3312                 else
3313                         sid->can_add_hw = 0;
3314         }
3315 
3316         return 0;
3317 }
3318 
3319 static void
3320 ctx_pinned_sched_in(struct perf_event_context *ctx,
3321                     struct perf_cpu_context *cpuctx)
3322 {
3323         struct sched_in_data sid = {
3324                 .ctx = ctx,
3325                 .cpuctx = cpuctx,
3326                 .can_add_hw = 1,
3327         };
3328 
3329         visit_groups_merge(&ctx->pinned_groups,
3330                            smp_processor_id(),
3331                            pinned_sched_in, &sid);
3332 }
3333 
3334 static void
3335 ctx_flexible_sched_in(struct perf_event_context *ctx,
3336                       struct perf_cpu_context *cpuctx)
3337 {
3338         struct sched_in_data sid = {
3339                 .ctx = ctx,
3340                 .cpuctx = cpuctx,
3341                 .can_add_hw = 1,
3342         };
3343 
3344         visit_groups_merge(&ctx->flexible_groups,
3345                            smp_processor_id(),
3346                            flexible_sched_in, &sid);
3347 }
3348 
3349 static void
3350 ctx_sched_in(struct perf_event_context *ctx,
3351              struct perf_cpu_context *cpuctx,
3352              enum event_type_t event_type,
3353              struct task_struct *task)
3354 {
3355         int is_active = ctx->is_active;
3356         u64 now;
3357 
3358         lockdep_assert_held(&ctx->lock);
3359 
3360         if (likely(!ctx->nr_events))
3361                 return;
3362 
3363         ctx->is_active |= (event_type | EVENT_TIME);
3364         if (ctx->task) {
3365                 if (!is_active)
3366                         cpuctx->task_ctx = ctx;
3367                 else
3368                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
3369         }
3370 
3371         is_active ^= ctx->is_active; /* changed bits */
3372 
3373         if (is_active & EVENT_TIME) {
3374                 /* start ctx time */
3375                 now = perf_clock();
3376                 ctx->timestamp = now;
3377                 perf_cgroup_set_timestamp(task, ctx);
3378         }
3379 
3380         /*
3381          * First go through the list and put on any pinned groups
3382          * in order to give them the best chance of going on.
3383          */
3384         if (is_active & EVENT_PINNED)
3385                 ctx_pinned_sched_in(ctx, cpuctx);
3386 
3387         /* Then walk through the lower prio flexible groups */
3388         if (is_active & EVENT_FLEXIBLE)
3389                 ctx_flexible_sched_in(ctx, cpuctx);
3390 }
3391 
3392 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
3393                              enum event_type_t event_type,
3394                              struct task_struct *task)
3395 {
3396         struct perf_event_context *ctx = &cpuctx->ctx;
3397 
3398         ctx_sched_in(ctx, cpuctx, event_type, task);
3399 }
3400 
3401 static void perf_event_context_sched_in(struct perf_event_context *ctx,
3402                                         struct task_struct *task)
3403 {
3404         struct perf_cpu_context *cpuctx;
3405 
3406         cpuctx = __get_cpu_context(ctx);
3407         if (cpuctx->task_ctx == ctx)
3408                 return;
3409 
3410         perf_ctx_lock(cpuctx, ctx);
3411         /*
3412          * We must check ctx->nr_events while holding ctx->lock, such
3413          * that we serialize against perf_install_in_context().
3414          */
3415         if (!ctx->nr_events)
3416                 goto unlock;
3417 
3418         perf_pmu_disable(ctx->pmu);
3419         /*
3420          * We want to keep the following priority order:
3421          * cpu pinned (that don't need to move), task pinned,
3422          * cpu flexible, task flexible.
3423          *
3424          * However, if task's ctx is not carrying any pinned
3425          * events, no need to flip the cpuctx's events around.
3426          */
3427         if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree))
3428                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3429         perf_event_sched_in(cpuctx, ctx, task);
3430         perf_pmu_enable(ctx->pmu);
3431 
3432 unlock:
3433         perf_ctx_unlock(cpuctx, ctx);
3434 }
3435 
3436 /*
3437  * Called from scheduler to add the events of the current task
3438  * with interrupts disabled.
3439  *
3440  * We restore the event value and then enable it.
3441  *
3442  * This does not protect us against NMI, but enable()
3443  * sets the enabled bit in the control field of event _before_
3444  * accessing the event control register. If a NMI hits, then it will
3445  * keep the event running.
3446  */
3447 void __perf_event_task_sched_in(struct task_struct *prev,
3448                                 struct task_struct *task)
3449 {
3450         struct perf_event_context *ctx;
3451         int ctxn;
3452 
3453         /*
3454          * If cgroup events exist on this CPU, then we need to check if we have
3455          * to switch in PMU state; cgroup event are system-wide mode only.
3456          *
3457          * Since cgroup events are CPU events, we must schedule these in before
3458          * we schedule in the task events.
3459          */
3460         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3461                 perf_cgroup_sched_in(prev, task);
3462 
3463         for_each_task_context_nr(ctxn) {
3464                 ctx = task->perf_event_ctxp[ctxn];
3465                 if (likely(!ctx))
3466                         continue;
3467 
3468                 perf_event_context_sched_in(ctx, task);
3469         }
3470 
3471         if (atomic_read(&nr_switch_events))
3472                 perf_event_switch(task, prev, true);
3473 
3474         if (__this_cpu_read(perf_sched_cb_usages))
3475                 perf_pmu_sched_task(prev, task, true);
3476 }
3477 
3478 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
3479 {
3480         u64 frequency = event->attr.sample_freq;
3481         u64 sec = NSEC_PER_SEC;
3482         u64 divisor, dividend;
3483 
3484         int count_fls, nsec_fls, frequency_fls, sec_fls;
3485 
3486         count_fls = fls64(count);
3487         nsec_fls = fls64(nsec);
3488         frequency_fls = fls64(frequency);
3489         sec_fls = 30;
3490 
3491         /*
3492          * We got @count in @nsec, with a target of sample_freq HZ
3493          * the target period becomes:
3494          *
3495          *             @count * 10^9
3496          * period = -------------------
3497          *          @nsec * sample_freq
3498          *
3499          */
3500 
3501         /*
3502          * Reduce accuracy by one bit such that @a and @b converge
3503          * to a similar magnitude.
3504          */
3505 #define REDUCE_FLS(a, b)                \
3506 do {                                    \
3507         if (a##_fls > b##_fls) {        \
3508                 a >>= 1;                \
3509                 a##_fls--;              \
3510         } else {                        \
3511                 b >>= 1;                \
3512                 b##_fls--;              \
3513         }                               \
3514 } while (0)
3515 
3516         /*
3517          * Reduce accuracy until either term fits in a u64, then proceed with
3518          * the other, so that finally we can do a u64/u64 division.
3519          */
3520         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
3521                 REDUCE_FLS(nsec, frequency);
3522                 REDUCE_FLS(sec, count);
3523         }
3524 
3525         if (count_fls + sec_fls > 64) {
3526                 divisor = nsec * frequency;
3527 
3528                 while (count_fls + sec_fls > 64) {
3529                         REDUCE_FLS(count, sec);
3530                         divisor >>= 1;
3531                 }
3532 
3533                 dividend = count * sec;
3534         } else {
3535                 dividend = count * sec;
3536 
3537                 while (nsec_fls + frequency_fls > 64) {
3538                         REDUCE_FLS(nsec, frequency);
3539                         dividend >>= 1;
3540                 }
3541 
3542                 divisor = nsec * frequency;
3543         }
3544 
3545         if (!divisor)
3546                 return dividend;
3547 
3548         return div64_u64(dividend, divisor);
3549 }
3550 
3551 static DEFINE_PER_CPU(int, perf_throttled_count);
3552 static DEFINE_PER_CPU(u64, perf_throttled_seq);
3553 
3554 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
3555 {
3556         struct hw_perf_event *hwc = &event->hw;
3557         s64 period, sample_period;
3558         s64 delta;
3559 
3560         period = perf_calculate_period(event, nsec, count);
3561 
3562         delta = (s64)(period - hwc->sample_period);
3563         delta = (delta + 7) / 8; /* low pass filter */
3564 
3565         sample_period = hwc->sample_period + delta;
3566 
3567         if (!sample_period)
3568                 sample_period = 1;
3569 
3570         hwc->sample_period = sample_period;
3571 
3572         if (local64_read(&hwc->period_left) > 8*sample_period) {
3573                 if (disable)
3574                         event->pmu->stop(event, PERF_EF_UPDATE);
3575 
3576                 local64_set(&hwc->period_left, 0);
3577 
3578                 if (disable)
3579                         event->pmu->start(event, PERF_EF_RELOAD);
3580         }
3581 }
3582 
3583 /*
3584  * combine freq adjustment with unthrottling to avoid two passes over the
3585  * events. At the same time, make sure, having freq events does not change
3586  * the rate of unthrottling as that would introduce bias.
3587  */
3588 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
3589                                            int needs_unthr)
3590 {
3591         struct perf_event *event;
3592         struct hw_perf_event *hwc;
3593         u64 now, period = TICK_NSEC;
3594         s64 delta;
3595 
3596         /*
3597          * only need to iterate over all events iff:
3598          * - context have events in frequency mode (needs freq adjust)
3599          * - there are events to unthrottle on this cpu
3600          */
3601         if (!(ctx->nr_freq || needs_unthr))
3602                 return;
3603 
3604         raw_spin_lock(&ctx->lock);
3605         perf_pmu_disable(ctx->pmu);
3606 
3607         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3608                 if (event->state != PERF_EVENT_STATE_ACTIVE)
3609                         continue;
3610 
3611                 if (!event_filter_match(event))
3612                         continue;
3613 
3614                 perf_pmu_disable(event->pmu);
3615 
3616                 hwc = &event->hw;
3617 
3618                 if (hwc->interrupts == MAX_INTERRUPTS) {
3619                         hwc->interrupts = 0;
3620                         perf_log_throttle(event, 1);
3621                         event->pmu->start(event, 0);
3622                 }
3623 
3624                 if (!event->attr.freq || !event->attr.sample_freq)
3625                         goto next;
3626 
3627                 /*
3628                  * stop the event and update event->count
3629                  */
3630                 event->pmu->stop(event, PERF_EF_UPDATE);
3631 
3632                 now = local64_read(&event->count);
3633                 delta = now - hwc->freq_count_stamp;
3634                 hwc->freq_count_stamp = now;
3635 
3636                 /*
3637                  * restart the event
3638                  * reload only if value has changed
3639                  * we have stopped the event so tell that
3640                  * to perf_adjust_period() to avoid stopping it
3641                  * twice.
3642                  */
3643                 if (delta > 0)
3644                         perf_adjust_period(event, period, delta, false);
3645 
3646                 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3647         next:
3648                 perf_pmu_enable(event->pmu);
3649         }
3650 
3651         perf_pmu_enable(ctx->pmu);
3652         raw_spin_unlock(&ctx->lock);
3653 }
3654 
3655 /*
3656  * Move @event to the tail of the @ctx's elegible events.
3657  */
3658 static void rotate_ctx(struct perf_event_context *ctx, struct perf_event *event)
3659 {
3660         /*
3661          * Rotate the first entry last of non-pinned groups. Rotation might be
3662          * disabled by the inheritance code.
3663          */
3664         if (ctx->rotate_disable)
3665                 return;
3666 
3667         perf_event_groups_delete(&ctx->flexible_groups, event);
3668         perf_event_groups_insert(&ctx->flexible_groups, event);
3669 }
3670 
3671 static inline struct perf_event *
3672 ctx_first_active(struct perf_event_context *ctx)
3673 {
3674         return list_first_entry_or_null(&ctx->flexible_active,
3675                                         struct perf_event, active_list);
3676 }
3677 
3678 static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
3679 {
3680         struct perf_event *cpu_event = NULL, *task_event = NULL;
3681         bool cpu_rotate = false, task_rotate = false;
3682         struct perf_event_context *ctx = NULL;
3683 
3684         /*
3685          * Since we run this from IRQ context, nobody can install new
3686          * events, thus the event count values are stable.
3687          */
3688 
3689         if (cpuctx->ctx.nr_events) {
3690                 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3691                         cpu_rotate = true;
3692         }
3693 
3694         ctx = cpuctx->task_ctx;
3695         if (ctx && ctx->nr_events) {
3696                 if (ctx->nr_events != ctx->nr_active)
3697                         task_rotate = true;
3698         }
3699 
3700         if (!(cpu_rotate || task_rotate))
3701                 return false;
3702 
3703         perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3704         perf_pmu_disable(cpuctx->ctx.pmu);
3705 
3706         if (task_rotate)
3707                 task_event = ctx_first_active(ctx);
3708         if (cpu_rotate)
3709                 cpu_event = ctx_first_active(&cpuctx->ctx);
3710 
3711         /*
3712          * As per the order given at ctx_resched() first 'pop' task flexible
3713          * and then, if needed CPU flexible.
3714          */
3715         if (task_event || (ctx && cpu_event))
3716                 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3717         if (cpu_event)
3718                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3719 
3720         if (task_event)
3721                 rotate_ctx(ctx, task_event);
3722         if (cpu_event)
3723                 rotate_ctx(&cpuctx->ctx, cpu_event);
3724 
3725         perf_event_sched_in(cpuctx, ctx, current);
3726 
3727         perf_pmu_enable(cpuctx->ctx.pmu);
3728         perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3729 
3730         return true;
3731 }
3732 
3733 void perf_event_task_tick(void)
3734 {
3735         struct list_head *head = this_cpu_ptr(&active_ctx_list);
3736         struct perf_event_context *ctx, *tmp;
3737         int throttled;
3738 
3739         lockdep_assert_irqs_disabled();
3740 
3741         __this_cpu_inc(perf_throttled_seq);
3742         throttled = __this_cpu_xchg(perf_throttled_count, 0);
3743         tick_dep_clear_cpu(smp_processor_id(), TICK_DEP_BIT_PERF_EVENTS);
3744 
3745         list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3746                 perf_adjust_freq_unthr_context(ctx, throttled);
3747 }
3748 
3749 static int event_enable_on_exec(struct perf_event *event,
3750                                 struct perf_event_context *ctx)
3751 {
3752         if (!event->attr.enable_on_exec)
3753                 return 0;
3754 
3755         event->attr.enable_on_exec = 0;
3756         if (event->state >= PERF_EVENT_STATE_INACTIVE)
3757                 return 0;
3758 
3759         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
3760 
3761         return 1;
3762 }
3763 
3764 /*
3765  * Enable all of a task's events that have been marked enable-on-exec.
3766  * This expects task == current.
3767  */
3768 static void perf_event_enable_on_exec(int ctxn)
3769 {
3770         struct perf_event_context *ctx, *clone_ctx = NULL;
3771         enum event_type_t event_type = 0;
3772         struct perf_cpu_context *cpuctx;
3773         struct perf_event *event;
3774         unsigned long flags;
3775         int enabled = 0;
3776 
3777         local_irq_save(flags);
3778         ctx = current->perf_event_ctxp[ctxn];
3779         if (!ctx || !ctx->nr_events)
3780                 goto out;
3781 
3782         cpuctx = __get_cpu_context(ctx);
3783         perf_ctx_lock(cpuctx, ctx);
3784         ctx_sched_out(ctx, cpuctx, EVENT_TIME);
3785         list_for_each_entry(event, &ctx->event_list, event_entry) {
3786                 enabled |= event_enable_on_exec(event, ctx);
3787                 event_type |= get_event_type(event);
3788         }
3789 
3790         /*
3791          * Unclone and reschedule this context if we enabled any event.
3792          */
3793         if (enabled) {
3794                 clone_ctx = unclone_ctx(ctx);
3795                 ctx_resched(cpuctx, ctx, event_type);
3796         } else {
3797                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
3798         }
3799         perf_ctx_unlock(cpuctx, ctx);
3800 
3801 out:
3802         local_irq_restore(flags);
3803 
3804         if (clone_ctx)
3805                 put_ctx(clone_ctx);
3806 }
3807 
3808 struct perf_read_data {
3809         struct perf_event *event;
3810         bool group;
3811         int ret;
3812 };
3813 
3814 static int __perf_event_read_cpu(struct perf_event *event, int event_cpu)
3815 {
3816         u16 local_pkg, event_pkg;
3817 
3818         if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) {
3819                 int local_cpu = smp_processor_id();
3820 
3821                 event_pkg = topology_physical_package_id(event_cpu);
3822                 local_pkg = topology_physical_package_id(local_cpu);
3823 
3824                 if (event_pkg == local_pkg)
3825                         return local_cpu;
3826         }
3827 
3828         return event_cpu;
3829 }
3830 
3831 /*
3832  * Cross CPU call to read the hardware event
3833  */
3834 static void __perf_event_read(void *info)
3835 {
3836         struct perf_read_data *data = info;
3837         struct perf_event *sub, *event = data->event;
3838         struct perf_event_context *ctx = event->ctx;
3839         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3840         struct pmu *pmu = event->pmu;
3841 
3842         /*
3843          * If this is a task context, we need to check whether it is
3844          * the current task context of this cpu.  If not it has been
3845          * scheduled out before the smp call arrived.  In that case
3846          * event->count would have been updated to a recent sample
3847          * when the event was scheduled out.
3848          */
3849         if (ctx->task && cpuctx->task_ctx != ctx)
3850                 return;
3851 
3852         raw_spin_lock(&ctx->lock);
3853         if (ctx->is_active & EVENT_TIME) {
3854                 update_context_time(ctx);
3855                 update_cgrp_time_from_event(event);
3856         }
3857 
3858         perf_event_update_time(event);
3859         if (data->group)
3860                 perf_event_update_sibling_time(event);
3861 
3862         if (event->state != PERF_EVENT_STATE_ACTIVE)
3863                 goto unlock;
3864 
3865         if (!data->group) {
3866                 pmu->read(event);
3867                 data->ret = 0;
3868                 goto unlock;
3869         }
3870 
3871         pmu->start_txn(pmu, PERF_PMU_TXN_READ);
3872 
3873         pmu->read(event);
3874 
3875         for_each_sibling_event(sub, event) {
3876                 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
3877                         /*
3878                          * Use sibling's PMU rather than @event's since
3879                          * sibling could be on different (eg: software) PMU.
3880                          */
3881                         sub->pmu->read(sub);
3882                 }
3883         }
3884 
3885         data->ret = pmu->commit_txn(pmu);
3886 
3887 unlock:
3888         raw_spin_unlock(&ctx->lock);
3889 }
3890 
3891 static inline u64 perf_event_count(struct perf_event *event)
3892 {
3893         return local64_read(&event->count) + atomic64_read(&event->child_count);
3894 }
3895 
3896 /*
3897  * NMI-safe method to read a local event, that is an event that
3898  * is:
3899  *   - either for the current task, or for this CPU
3900  *   - does not have inherit set, for inherited task events
3901  *     will not be local and we cannot read them atomically
3902  *   - must not have a pmu::count method
3903  */
3904 int perf_event_read_local(struct perf_event *event, u64 *value,
3905                           u64 *enabled, u64 *running)
3906 {
3907         unsigned long flags;
3908         int ret = 0;
3909 
3910         /*
3911          * Disabling interrupts avoids all counter scheduling (context
3912          * switches, timer based rotation and IPIs).
3913          */
3914         local_irq_save(flags);
3915 
3916         /*
3917          * It must not be an event with inherit set, we cannot read
3918          * all child counters from atomic context.
3919          */
3920         if (event->attr.inherit) {
3921                 ret = -EOPNOTSUPP;
3922                 goto out;
3923         }
3924 
3925         /* If this is a per-task event, it must be for current */
3926         if ((event->attach_state & PERF_ATTACH_TASK) &&
3927             event->hw.target != current) {
3928                 ret = -EINVAL;
3929                 goto out;
3930         }
3931 
3932         /* If this is a per-CPU event, it must be for this CPU */
3933         if (!(event->attach_state & PERF_ATTACH_TASK) &&
3934             event->cpu != smp_processor_id()) {
3935                 ret = -EINVAL;
3936                 goto out;
3937         }
3938 
3939         /* If this is a pinned event it must be running on this CPU */
3940         if (event->attr.pinned && event->oncpu != smp_processor_id()) {
3941                 ret = -EBUSY;
3942                 goto out;
3943         }
3944 
3945         /*
3946          * If the event is currently on this CPU, its either a per-task event,
3947          * or local to this CPU. Furthermore it means its ACTIVE (otherwise
3948          * oncpu == -1).
3949          */
3950         if (event->oncpu == smp_processor_id())
3951                 event->pmu->read(event);
3952 
3953         *value = local64_read(&event->count);
3954         if (enabled || running) {
3955                 u64 now = event->shadow_ctx_time + perf_clock();
3956                 u64 __enabled, __running;
3957 
3958                 __perf_update_times(event, now, &__enabled, &__running);
3959                 if (enabled)
3960                         *enabled = __enabled;
3961                 if (running)
3962                         *running = __running;
3963         }
3964 out:
3965         local_irq_restore(flags);
3966 
3967         return ret;
3968 }
3969 
3970 static int perf_event_read(struct perf_event *event, bool group)
3971 {
3972         enum perf_event_state state = READ_ONCE(event->state);
3973         int event_cpu, ret = 0;
3974 
3975         /*
3976          * If event is enabled and currently active on a CPU, update the
3977          * value in the event structure:
3978          */
3979 again:
3980         if (state == PERF_EVENT_STATE_ACTIVE) {
3981                 struct perf_read_data data;
3982 
3983                 /*
3984                  * Orders the ->state and ->oncpu loads such that if we see
3985                  * ACTIVE we must also see the right ->oncpu.
3986                  *
3987                  * Matches the smp_wmb() from event_sched_in().
3988                  */
3989                 smp_rmb();
3990 
3991                 event_cpu = READ_ONCE(event->oncpu);
3992                 if ((unsigned)event_cpu >= nr_cpu_ids)
3993                         return 0;
3994 
3995                 data = (struct perf_read_data){
3996                         .event = event,
3997                         .group = group,
3998                         .ret = 0,
3999                 };
4000 
4001                 preempt_disable();
4002                 event_cpu = __perf_event_read_cpu(event, event_cpu);
4003 
4004                 /*
4005                  * Purposely ignore the smp_call_function_single() return
4006                  * value.
4007                  *
4008                  * If event_cpu isn't a valid CPU it means the event got
4009                  * scheduled out and that will have updated the event count.
4010                  *
4011                  * Therefore, either way, we'll have an up-to-date event count
4012                  * after this.
4013                  */
4014                 (void)smp_call_function_single(event_cpu, __perf_event_read, &data, 1);
4015                 preempt_enable();
4016                 ret = data.ret;
4017 
4018         } else if (state == PERF_EVENT_STATE_INACTIVE) {
4019                 struct perf_event_context *ctx = event->ctx;
4020                 unsigned long flags;
4021 
4022                 raw_spin_lock_irqsave(&ctx->lock, flags);
4023                 state = event->state;
4024                 if (state != PERF_EVENT_STATE_INACTIVE) {
4025                         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4026                         goto again;
4027                 }
4028 
4029                 /*
4030                  * May read while context is not active (e.g., thread is
4031                  * blocked), in that case we cannot update context time
4032                  */
4033                 if (ctx->is_active & EVENT_TIME) {
4034                         update_context_time(ctx);
4035                         update_cgrp_time_from_event(event);
4036                 }
4037 
4038                 perf_event_update_time(event);
4039                 if (group)
4040                         perf_event_update_sibling_time(event);
4041                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4042         }
4043 
4044         return ret;
4045 }
4046 
4047 /*
4048  * Initialize the perf_event context in a task_struct:
4049  */
4050 static void __perf_event_init_context(struct perf_event_context *ctx)
4051 {
4052         raw_spin_lock_init(&ctx->lock);
4053         mutex_init(&ctx->mutex);
4054         INIT_LIST_HEAD(&ctx->active_ctx_list);
4055         perf_event_groups_init(&ctx->pinned_groups);
4056         perf_event_groups_init(&ctx->flexible_groups);
4057         INIT_LIST_HEAD(&ctx->event_list);
4058         INIT_LIST_HEAD(&ctx->pinned_active);
4059         INIT_LIST_HEAD(&ctx->flexible_active);
4060         atomic_set(&ctx->refcount, 1);
4061 }
4062 
4063 static struct perf_event_context *
4064 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
4065 {
4066         struct perf_event_context *ctx;
4067 
4068         ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4069         if (!ctx)
4070                 return NULL;
4071 
4072         __perf_event_init_context(ctx);
4073         if (task) {
4074                 ctx->task = task;
4075                 get_task_struct(task);
4076         }
4077         ctx->pmu = pmu;
4078 
4079         return ctx;
4080 }
4081 
4082 static struct task_struct *
4083 find_lively_task_by_vpid(pid_t vpid)
4084 {
4085         struct task_struct *task;
4086 
4087         rcu_read_lock();
4088         if (!vpid)
4089                 task = current;
4090         else
4091                 task = find_task_by_vpid(vpid);
4092         if (task)
4093                 get_task_struct(task);
4094         rcu_read_unlock();
4095 
4096         if (!task)
4097                 return ERR_PTR(-ESRCH);
4098 
4099         return task;
4100 }
4101 
4102 /*
4103  * Returns a matching context with refcount and pincount.
4104  */
4105 static struct perf_event_context *
4106 find_get_context(struct pmu *pmu, struct task_struct *task,
4107                 struct perf_event *event)
4108 {
4109         struct perf_event_context *ctx, *clone_ctx = NULL;
4110         struct perf_cpu_context *cpuctx;
4111         void *task_ctx_data = NULL;
4112         unsigned long flags;
4113         int ctxn, err;
4114         int cpu = event->cpu;
4115 
4116         if (!task) {
4117                 /* Must be root to operate on a CPU event: */
4118                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
4119                         return ERR_PTR(-EACCES);
4120 
4121                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
4122                 ctx = &cpuctx->ctx;
4123                 get_ctx(ctx);
4124                 ++ctx->pin_count;
4125 
4126                 return ctx;
4127         }
4128 
4129         err = -EINVAL;
4130         ctxn = pmu->task_ctx_nr;
4131         if (ctxn < 0)
4132                 goto errout;
4133 
4134         if (event->attach_state & PERF_ATTACH_TASK_DATA) {
4135                 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
4136                 if (!task_ctx_data) {
4137                         err = -ENOMEM;
4138                         goto errout;
4139                 }
4140         }
4141 
4142 retry:
4143         ctx = perf_lock_task_context(task, ctxn, &flags);
4144         if (ctx) {
4145                 clone_ctx = unclone_ctx(ctx);
4146                 ++ctx->pin_count;
4147 
4148                 if (task_ctx_data && !ctx->task_ctx_data) {
4149                         ctx->task_ctx_data = task_ctx_data;
4150                         task_ctx_data = NULL;
4151                 }
4152                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4153 
4154                 if (clone_ctx)
4155                         put_ctx(clone_ctx);
4156         } else {
4157                 ctx = alloc_perf_context(pmu, task);
4158                 err = -ENOMEM;
4159                 if (!ctx)
4160                         goto errout;
4161 
4162                 if (task_ctx_data) {
4163                         ctx->task_ctx_data = task_ctx_data;
4164                         task_ctx_data = NULL;
4165                 }
4166 
4167                 err = 0;
4168                 mutex_lock(&task->perf_event_mutex);
4169                 /*
4170                  * If it has already passed perf_event_exit_task().
4171                  * we must see PF_EXITING, it takes this mutex too.
4172                  */
4173                 if (task->flags & PF_EXITING)
4174                         err = -ESRCH;
4175                 else if (task->perf_event_ctxp[ctxn])
4176                         err = -EAGAIN;
4177                 else {
4178                         get_ctx(ctx);
4179                         ++ctx->pin_count;
4180                         rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
4181                 }
4182                 mutex_unlock(&task->perf_event_mutex);
4183 
4184                 if (unlikely(err)) {
4185                         put_ctx(ctx);
4186 
4187                         if (err == -EAGAIN)
4188                                 goto retry;
4189                         goto errout;
4190                 }
4191         }
4192 
4193         kfree(task_ctx_data);
4194         return ctx;
4195 
4196 errout:
4197         kfree(task_ctx_data);
4198         return ERR_PTR(err);
4199 }
4200 
4201 static void perf_event_free_filter(struct perf_event *event);
4202 static void perf_event_free_bpf_prog(struct perf_event *event);
4203 
4204 static void free_event_rcu(struct rcu_head *head)
4205 {
4206         struct perf_event *event;
4207 
4208         event = container_of(head, struct perf_event, rcu_head);
4209         if (event->ns)
4210                 put_pid_ns(event->ns);
4211         perf_event_free_filter(event);
4212         kfree(event);
4213 }
4214 
4215 static void ring_buffer_attach(struct perf_event *event,
4216                                struct ring_buffer *rb);
4217 
4218 static void detach_sb_event(struct perf_event *event)
4219 {
4220         struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu);
4221 
4222         raw_spin_lock(&pel->lock);
4223         list_del_rcu(&event->sb_list);
4224         raw_spin_unlock(&pel->lock);
4225 }
4226 
4227 static bool is_sb_event(struct perf_event *event)
4228 {
4229         struct perf_event_attr *attr = &event->attr;
4230 
4231         if (event->parent)
4232                 return false;
4233 
4234         if (event->attach_state & PERF_ATTACH_TASK)
4235                 return false;
4236 
4237         if (attr->mmap || attr->mmap_data || attr->mmap2 ||
4238             attr->comm || attr->comm_exec ||
4239             attr->task ||
4240             attr->context_switch)
4241                 return true;
4242         return false;
4243 }
4244 
4245 static void unaccount_pmu_sb_event(struct perf_event *event)
4246 {
4247         if (is_sb_event(event))
4248                 detach_sb_event(event);
4249 }
4250 
4251 static void unaccount_event_cpu(struct perf_event *event, int cpu)
4252 {
4253         if (event->parent)
4254                 return;
4255 
4256         if (is_cgroup_event(event))
4257                 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
4258 }
4259 
4260 #ifdef CONFIG_NO_HZ_FULL
4261 static DEFINE_SPINLOCK(nr_freq_lock);
4262 #endif
4263 
4264 static void unaccount_freq_event_nohz(void)
4265 {
4266 #ifdef CONFIG_NO_HZ_FULL
4267         spin_lock(&nr_freq_lock);
4268         if (atomic_dec_and_test(&nr_freq_events))
4269                 tick_nohz_dep_clear(TICK_DEP_BIT_PERF_EVENTS);
4270         spin_unlock(&nr_freq_lock);
4271 #endif
4272 }
4273 
4274 static void unaccount_freq_event(void)
4275 {
4276         if (tick_nohz_full_enabled())
4277                 unaccount_freq_event_nohz();
4278         else
4279                 atomic_dec(&nr_freq_events);
4280 }
4281 
4282 static void unaccount_event(struct perf_event *event)
4283 {
4284         bool dec = false;
4285 
4286         if (event->parent)
4287                 return;
4288 
4289         if (event->attach_state & PERF_ATTACH_TASK)
4290                 dec = true;
4291         if (event->attr.mmap || event->attr.mmap_data)
4292                 atomic_dec(&nr_mmap_events);
4293         if (event->attr.comm)
4294                 atomic_dec(&nr_comm_events);
4295         if (event->attr.namespaces)
4296                 atomic_dec(&nr_namespaces_events);
4297         if (event->attr.task)
4298                 atomic_dec(&nr_task_events);
4299         if (event->attr.freq)
4300                 unaccount_freq_event();
4301         if (event->attr.context_switch) {
4302                 dec = true;
4303                 atomic_dec(&nr_switch_events);
4304         }
4305         if (is_cgroup_event(event))
4306                 dec = true;
4307         if (has_branch_stack(event))
4308                 dec = true;
4309 
4310         if (dec) {
4311                 if (!atomic_add_unless(&perf_sched_count, -1, 1))
4312                         schedule_delayed_work(&perf_sched_work, HZ);
4313         }
4314 
4315         unaccount_event_cpu(event, event->cpu);
4316 
4317         unaccount_pmu_sb_event(event);
4318 }
4319 
4320 static void perf_sched_delayed(struct work_struct *work)
4321 {
4322         mutex_lock(&perf_sched_mutex);
4323         if (atomic_dec_and_test(&perf_sched_count))
4324                 static_branch_disable(&perf_sched_events);
4325         mutex_unlock(&perf_sched_mutex);
4326 }
4327 
4328 /*
4329  * The following implement mutual exclusion of events on "exclusive" pmus
4330  * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
4331  * at a time, so we disallow creating events that might conflict, namely:
4332  *
4333  *  1) cpu-wide events in the presence of per-task events,
4334  *  2) per-task events in the presence of cpu-wide events,
4335  *  3) two matching events on the same context.
4336  *
4337  * The former two cases are handled in the allocation path (perf_event_alloc(),
4338  * _free_event()), the latter -- before the first perf_install_in_context().
4339  */
4340 static int exclusive_event_init(struct perf_event *event)
4341 {
4342         struct pmu *pmu = event->pmu;
4343 
4344         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4345                 return 0;
4346 
4347         /*
4348          * Prevent co-existence of per-task and cpu-wide events on the
4349          * same exclusive pmu.
4350          *
4351          * Negative pmu::exclusive_cnt means there are cpu-wide
4352          * events on this "exclusive" pmu, positive means there are
4353          * per-task events.
4354          *
4355          * Since this is called in perf_event_alloc() path, event::ctx
4356          * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
4357          * to mean "per-task event", because unlike other attach states it
4358          * never gets cleared.
4359          */
4360         if (event->attach_state & PERF_ATTACH_TASK) {
4361                 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
4362                         return -EBUSY;
4363         } else {
4364                 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
4365                         return -EBUSY;
4366         }
4367 
4368         return 0;
4369 }
4370 
4371 static void exclusive_event_destroy(struct perf_event *event)
4372 {
4373         struct pmu *pmu = event->pmu;
4374 
4375         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4376                 return;
4377 
4378         /* see comment in exclusive_event_init() */
4379         if (event->attach_state & PERF_ATTACH_TASK)
4380                 atomic_dec(&pmu->exclusive_cnt);
4381         else
4382                 atomic_inc(&pmu->exclusive_cnt);
4383 }
4384 
4385 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
4386 {
4387         if ((e1->pmu == e2->pmu) &&
4388             (e1->cpu == e2->cpu ||
4389              e1->cpu == -1 ||
4390              e2->cpu == -1))
4391                 return true;
4392         return false;
4393 }
4394 
4395 /* Called under the same ctx::mutex as perf_install_in_context() */
4396 static bool exclusive_event_installable(struct perf_event *event,
4397                                         struct perf_event_context *ctx)
4398 {
4399         struct perf_event *iter_event;
4400         struct pmu *pmu = event->pmu;
4401 
4402         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4403                 return true;
4404 
4405         list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
4406                 if (exclusive_event_match(iter_event, event))
4407                         return false;
4408         }
4409 
4410         return true;
4411 }
4412 
4413 static void perf_addr_filters_splice(struct perf_event *event,
4414                                        struct list_head *head);
4415 
4416 static void _free_event(struct perf_event *event)
4417 {
4418         irq_work_sync(&event->pending);
4419 
4420         unaccount_event(event);
4421 
4422         if (event->rb) {
4423                 /*
4424                  * Can happen when we close an event with re-directed output.
4425                  *
4426                  * Since we have a 0 refcount, perf_mmap_close() will skip
4427                  * over us; possibly making our ring_buffer_put() the last.
4428                  */
4429                 mutex_lock(&event->mmap_mutex);
4430                 ring_buffer_attach(event, NULL);
4431                 mutex_unlock(&event->mmap_mutex);
4432         }
4433 
4434         if (is_cgroup_event(event))
4435                 perf_detach_cgroup(event);
4436 
4437         if (!event->parent) {
4438                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
4439                         put_callchain_buffers();
4440         }
4441 
4442         perf_event_free_bpf_prog(event);
4443         perf_addr_filters_splice(event, NULL);
4444         kfree(event->addr_filters_offs);
4445 
4446         if (event->destroy)
4447                 event->destroy(event);
4448 
4449         if (event->ctx)
4450                 put_ctx(event->ctx);
4451 
4452         if (event->hw.target)
4453                 put_task_struct(event->hw.target);
4454 
4455         exclusive_event_destroy(event);
4456         module_put(event->pmu->module);
4457 
4458         call_rcu(&event->rcu_head, free_event_rcu);
4459 }
4460 
4461 /*
4462  * Used to free events which have a known refcount of 1, such as in error paths
4463  * where the event isn't exposed yet and inherited events.
4464  */
4465 static void free_event(struct perf_event *event)
4466 {
4467         if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
4468                                 "unexpected event refcount: %ld; ptr=%p\n",
4469                                 atomic_long_read(&event->refcount), event)) {
4470                 /* leak to avoid use-after-free */
4471                 return;
4472         }
4473 
4474         _free_event(event);
4475 }
4476 
4477 /*
4478  * Remove user event from the owner task.
4479  */
4480 static void perf_remove_from_owner(struct perf_event *event)
4481 {
4482         struct task_struct *owner;
4483 
4484         rcu_read_lock();
4485         /*
4486          * Matches the smp_store_release() in perf_event_exit_task(). If we
4487          * observe !owner it means the list deletion is complete and we can
4488          * indeed free this event, otherwise we need to serialize on
4489          * owner->perf_event_mutex.
4490          */
4491         owner = READ_ONCE(event->owner);
4492         if (owner) {
4493                 /*
4494                  * Since delayed_put_task_struct() also drops the last
4495                  * task reference we can safely take a new reference
4496                  * while holding the rcu_read_lock().
4497                  */
4498                 get_task_struct(owner);
4499         }
4500         rcu_read_unlock();
4501 
4502         if (owner) {
4503                 /*
4504                  * If we're here through perf_event_exit_task() we're already
4505                  * holding ctx->mutex which would be an inversion wrt. the
4506                  * normal lock order.
4507                  *
4508                  * However we can safely take this lock because its the child
4509                  * ctx->mutex.
4510                  */
4511                 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
4512 
4513                 /*
4514                  * We have to re-check the event->owner field, if it is cleared
4515                  * we raced with perf_event_exit_task(), acquiring the mutex
4516                  * ensured they're done, and we can proceed with freeing the
4517                  * event.
4518                  */
4519                 if (event->owner) {
4520                         list_del_init(&event->owner_entry);
4521                         smp_store_release(&event->owner, NULL);
4522                 }
4523                 mutex_unlock(&owner->perf_event_mutex);
4524                 put_task_struct(owner);
4525         }
4526 }
4527 
4528 static void put_event(struct perf_event *event)
4529 {
4530         if (!atomic_long_dec_and_test(&event->refcount))
4531                 return;
4532 
4533         _free_event(event);
4534 }
4535 
4536 /*
4537  * Kill an event dead; while event:refcount will preserve the event
4538  * object, it will not preserve its functionality. Once the last 'user'
4539  * gives up the object, we'll destroy the thing.
4540  */
4541 int perf_event_release_kernel(struct perf_event *event)
4542 {
4543         struct perf_event_context *ctx = event->ctx;
4544         struct perf_event *child, *tmp;
4545         LIST_HEAD(free_list);
4546 
4547         /*
4548          * If we got here through err_file: fput(event_file); we will not have
4549          * attached to a context yet.
4550          */
4551         if (!ctx) {
4552                 WARN_ON_ONCE(event->attach_state &
4553                                 (PERF_ATTACH_CONTEXT|PERF_ATTACH_GROUP));
4554                 goto no_ctx;
4555         }
4556 
4557         if (!is_kernel_event(event))
4558                 perf_remove_from_owner(event);
4559 
4560         ctx = perf_event_ctx_lock(event);
4561         WARN_ON_ONCE(ctx->parent_ctx);
4562         perf_remove_from_context(event, DETACH_GROUP);
4563 
4564         raw_spin_lock_irq(&ctx->lock);
4565         /*
4566          * Mark this event as STATE_DEAD, there is no external reference to it
4567          * anymore.
4568          *
4569          * Anybody acquiring event->child_mutex after the below loop _must_
4570          * also see this, most importantly inherit_event() which will avoid
4571          * placing more children on the list.
4572          *
4573          * Thus this guarantees that we will in fact observe and kill _ALL_
4574          * child events.
4575          */
4576         event->state = PERF_EVENT_STATE_DEAD;
4577         raw_spin_unlock_irq(&ctx->lock);
4578 
4579         perf_event_ctx_unlock(event, ctx);
4580 
4581 again:
4582         mutex_lock(&event->child_mutex);
4583         list_for_each_entry(child, &event->child_list, child_list) {
4584 
4585                 /*
4586                  * Cannot change, child events are not migrated, see the
4587                  * comment with perf_event_ctx_lock_nested().
4588                  */
4589                 ctx = READ_ONCE(child->ctx);
4590                 /*
4591                  * Since child_mutex nests inside ctx::mutex, we must jump
4592                  * through hoops. We start by grabbing a reference on the ctx.
4593                  *
4594                  * Since the event cannot get freed while we hold the
4595                  * child_mutex, the context must also exist and have a !0
4596                  * reference count.
4597                  */
4598                 get_ctx(ctx);
4599 
4600                 /*
4601                  * Now that we have a ctx ref, we can drop child_mutex, and
4602                  * acquire ctx::mutex without fear of it going away. Then we
4603                  * can re-acquire child_mutex.
4604                  */
4605                 mutex_unlock(&event->child_mutex);
4606                 mutex_lock(&ctx->mutex);
4607                 mutex_lock(&event->child_mutex);
4608 
4609                 /*
4610                  * Now that we hold ctx::mutex and child_mutex, revalidate our
4611                  * state, if child is still the first entry, it didn't get freed
4612                  * and we can continue doing so.
4613                  */
4614                 tmp = list_first_entry_or_null(&event->child_list,
4615                                                struct perf_event, child_list);
4616                 if (tmp == child) {
4617                         perf_remove_from_context(child, DETACH_GROUP);
4618                         list_move(&child->child_list, &free_list);
4619                         /*
4620                          * This matches the refcount bump in inherit_event();
4621                          * this can't be the last reference.
4622                          */
4623                         put_event(event);
4624                 }
4625 
4626                 mutex_unlock(&event->child_mutex);
4627                 mutex_unlock(&ctx->mutex);
4628                 put_ctx(ctx);
4629                 goto again;
4630         }
4631         mutex_unlock(&event->child_mutex);
4632 
4633         list_for_each_entry_safe(child, tmp, &free_list, child_list) {
4634                 list_del(&child->child_list);
4635                 free_event(child);
4636         }
4637 
4638 no_ctx:
4639         put_event(event); /* Must be the 'last' reference */
4640         return 0;
4641 }
4642 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
4643 
4644 /*
4645  * Called when the last reference to the file is gone.
4646  */
4647 static int perf_release(struct inode *inode, struct file *file)
4648 {
4649         perf_event_release_kernel(file->private_data);
4650         return 0;
4651 }
4652 
4653 static u64 __perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4654 {
4655         struct perf_event *child;
4656         u64 total = 0;
4657 
4658         *enabled = 0;
4659         *running = 0;
4660 
4661         mutex_lock(&event->child_mutex);
4662 
4663         (void)perf_event_read(event, false);
4664         total += perf_event_count(event);
4665 
4666         *enabled += event->total_time_enabled +
4667                         atomic64_read(&event->child_total_time_enabled);
4668         *running += event->total_time_running +
4669                         atomic64_read(&event->child_total_time_running);
4670 
4671         list_for_each_entry(child, &event->child_list, child_list) {
4672                 (void)perf_event_read(child, false);
4673                 total += perf_event_count(child);
4674                 *enabled += child->total_time_enabled;
4675                 *running += child->total_time_running;
4676         }
4677         mutex_unlock(&event->child_mutex);
4678 
4679         return total;
4680 }
4681 
4682 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4683 {
4684         struct perf_event_context *ctx;
4685         u64 count;
4686 
4687         ctx = perf_event_ctx_lock(event);
4688         count = __perf_event_read_value(event, enabled, running);
4689         perf_event_ctx_unlock(event, ctx);
4690 
4691         return count;
4692 }
4693 EXPORT_SYMBOL_GPL(perf_event_read_value);
4694 
4695 static int __perf_read_group_add(struct perf_event *leader,
4696                                         u64 read_format, u64 *values)
4697 {
4698         struct perf_event_context *ctx = leader->ctx;
4699         struct perf_event *sub;
4700         unsigned long flags;
4701         int n = 1; /* skip @nr */
4702         int ret;
4703 
4704         ret = perf_event_read(leader, true);
4705         if (ret)
4706                 return ret;
4707 
4708         raw_spin_lock_irqsave(&ctx->lock, flags);
4709 
4710         /*
4711          * Since we co-schedule groups, {enabled,running} times of siblings
4712          * will be identical to those of the leader, so we only publish one
4713          * set.
4714          */
4715         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4716                 values[n++] += leader->total_time_enabled +
4717                         atomic64_read(&leader->child_total_time_enabled);
4718         }
4719 
4720         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4721                 values[n++] += leader->total_time_running +
4722                         atomic64_read(&leader->child_total_time_running);
4723         }
4724 
4725         /*
4726          * Write {count,id} tuples for every sibling.
4727          */
4728         values[n++] += perf_event_count(leader);
4729         if (read_format & PERF_FORMAT_ID)
4730                 values[n++] = primary_event_id(leader);
4731 
4732         for_each_sibling_event(sub, leader) {
4733                 values[n++] += perf_event_count(sub);
4734                 if (read_format & PERF_FORMAT_ID)
4735                         values[n++] = primary_event_id(sub);
4736         }
4737 
4738         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4739         return 0;
4740 }
4741 
4742 static int perf_read_group(struct perf_event *event,
4743                                    u64 read_format, char __user *buf)
4744 {
4745         struct perf_event *leader = event->group_leader, *child;
4746         struct perf_event_context *ctx = leader->ctx;
4747         int ret;
4748         u64 *values;
4749 
4750         lockdep_assert_held(&ctx->mutex);
4751 
4752         values = kzalloc(event->read_size, GFP_KERNEL);
4753         if (!values)
4754                 return -ENOMEM;
4755 
4756         values[0] = 1 + leader->nr_siblings;
4757 
4758         /*
4759          * By locking the child_mutex of the leader we effectively
4760          * lock the child list of all siblings.. XXX explain how.
4761          */
4762         mutex_lock(&leader->child_mutex);
4763 
4764         ret = __perf_read_group_add(leader, read_format, values);
4765         if (ret)
4766                 goto unlock;
4767 
4768         list_for_each_entry(child, &leader->child_list, child_list) {
4769                 ret = __perf_read_group_add(child, read_format, values);
4770                 if (ret)
4771                         goto unlock;
4772         }
4773 
4774         mutex_unlock(&leader->child_mutex);
4775 
4776         ret = event->read_size;
4777         if (copy_to_user(buf, values, event->read_size))
4778                 ret = -EFAULT;
4779         goto out;
4780 
4781 unlock:
4782         mutex_unlock(&leader->child_mutex);
4783 out:
4784         kfree(values);
4785         return ret;
4786 }
4787 
4788 static int perf_read_one(struct perf_event *event,
4789                                  u64 read_format, char __user *buf)
4790 {
4791         u64 enabled, running;
4792         u64 values[4];
4793         int n = 0;
4794 
4795         values[n++] = __perf_event_read_value(event, &enabled, &running);
4796         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4797                 values[n++] = enabled;
4798         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4799                 values[n++] = running;
4800         if (read_format & PERF_FORMAT_ID)
4801                 values[n++] = primary_event_id(event);
4802 
4803         if (copy_to_user(buf, values, n * sizeof(u64)))
4804                 return -EFAULT;
4805 
4806         return n * sizeof(u64);
4807 }
4808 
4809 static bool is_event_hup(struct perf_event *event)
4810 {
4811         bool no_children;
4812 
4813         if (event->state > PERF_EVENT_STATE_EXIT)
4814                 return false;
4815 
4816         mutex_lock(&event->child_mutex);
4817         no_children = list_empty(&event->child_list);
4818         mutex_unlock(&event->child_mutex);
4819         return no_children;
4820 }
4821 
4822 /*
4823  * Read the performance event - simple non blocking version for now
4824  */
4825 static ssize_t
4826 __perf_read(struct perf_event *event, char __user *buf, size_t count)
4827 {
4828         u64 read_format = event->attr.read_format;
4829         int ret;
4830 
4831         /*
4832          * Return end-of-file for a read on an event that is in
4833          * error state (i.e. because it was pinned but it couldn't be
4834          * scheduled on to the CPU at some point).
4835          */
4836         if (event->state == PERF_EVENT_STATE_ERROR)
4837                 return 0;
4838 
4839         if (count < event->read_size)
4840                 return -ENOSPC;
4841 
4842         WARN_ON_ONCE(event->ctx->parent_ctx);
4843         if (read_format & PERF_FORMAT_GROUP)
4844                 ret = perf_read_group(event, read_format, buf);
4845         else
4846                 ret = perf_read_one(event, read_format, buf);
4847 
4848         return ret;
4849 }
4850 
4851 static ssize_t
4852 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
4853 {
4854         struct perf_event *event = file->private_data;
4855         struct perf_event_context *ctx;
4856         int ret;
4857 
4858         ctx = perf_event_ctx_lock(event);
4859         ret = __perf_read(event, buf, count);
4860         perf_event_ctx_unlock(event, ctx);
4861 
4862         return ret;
4863 }
4864 
4865 static __poll_t perf_poll(struct file *file, poll_table *wait)
4866 {
4867         struct perf_event *event = file->private_data;
4868         struct ring_buffer *rb;
4869         __poll_t events = EPOLLHUP;
4870 
4871         poll_wait(file, &event->waitq, wait);
4872 
4873         if (is_event_hup(event))
4874                 return events;
4875 
4876         /*
4877          * Pin the event->rb by taking event->mmap_mutex; otherwise
4878          * perf_event_set_output() can swizzle our rb and make us miss wakeups.
4879          */
4880         mutex_lock(&event->mmap_mutex);
4881         rb = event->rb;
4882         if (rb)
4883                 events = atomic_xchg(&rb->poll, 0);
4884         mutex_unlock(&event->mmap_mutex);
4885         return events;
4886 }
4887 
4888 static void _perf_event_reset(struct perf_event *event)
4889 {
4890         (void)perf_event_read(event, false);
4891         local64_set(&event->count, 0);
4892         perf_event_update_userpage(event);
4893 }
4894 
4895 /*
4896  * Holding the top-level event's child_mutex means that any
4897  * descendant process that has inherited this event will block
4898  * in perf_event_exit_event() if it goes to exit, thus satisfying the
4899  * task existence requirements of perf_event_enable/disable.
4900  */
4901 static void perf_event_for_each_child(struct perf_event *event,
4902                                         void (*func)(struct perf_event *))
4903 {
4904         struct perf_event *child;
4905 
4906         WARN_ON_ONCE(event->ctx->parent_ctx);
4907 
4908         mutex_lock(&event->child_mutex);
4909         func(event);
4910         list_for_each_entry(child, &event->child_list, child_list)
4911                 func(child);
4912         mutex_unlock(&event->child_mutex);
4913 }
4914 
4915 static void perf_event_for_each(struct perf_event *event,
4916                                   void (*func)(struct perf_event *))
4917 {
4918         struct perf_event_context *ctx = event->ctx;
4919         struct perf_event *sibling;
4920 
4921         lockdep_assert_held(&ctx->mutex);
4922 
4923         event = event->group_leader;
4924 
4925         perf_event_for_each_child(event, func);
4926         for_each_sibling_event(sibling, event)
4927                 perf_event_for_each_child(sibling, func);
4928 }
4929 
4930 static void __perf_event_period(struct perf_event *event,
4931                                 struct perf_cpu_context *cpuctx,
4932                                 struct perf_event_context *ctx,
4933                                 void *info)
4934 {
4935         u64 value = *((u64 *)info);
4936         bool active;
4937 
4938         if (event->attr.freq) {
4939                 event->attr.sample_freq = value;
4940         } else {
4941                 event->attr.sample_period = value;
4942                 event->hw.sample_period = value;
4943         }
4944 
4945         active = (event->state == PERF_EVENT_STATE_ACTIVE);
4946         if (active) {
4947                 perf_pmu_disable(ctx->pmu);
4948                 /*
4949                  * We could be throttled; unthrottle now to avoid the tick
4950                  * trying to unthrottle while we already re-started the event.
4951                  */
4952                 if (event->hw.interrupts == MAX_INTERRUPTS) {
4953                         event->hw.interrupts = 0;
4954                         perf_log_throttle(event, 1);
4955                 }
4956                 event->pmu->stop(event, PERF_EF_UPDATE);
4957         }
4958 
4959         local64_set(&event->hw.period_left, 0);
4960 
4961         if (active) {
4962                 event->pmu->start(event, PERF_EF_RELOAD);
4963                 perf_pmu_enable(ctx->pmu);
4964         }
4965 }
4966 
4967 static int perf_event_check_period(struct perf_event *event, u64 value)
4968 {
4969         return event->pmu->check_period(event, value);
4970 }
4971 
4972 static int perf_event_period(struct perf_event *event, u64 __user *arg)
4973 {
4974         u64 value;
4975 
4976         if (!is_sampling_event(event))
4977                 return -EINVAL;
4978 
4979         if (copy_from_user(&value, arg, sizeof(value)))
4980                 return -EFAULT;
4981 
4982         if (!value)
4983                 return -EINVAL;
4984 
4985         if (event->attr.freq && value > sysctl_perf_event_sample_rate)
4986                 return -EINVAL;
4987 
4988         if (perf_event_check_period(event, value))
4989                 return -EINVAL;
4990 
4991         event_function_call(event, __perf_event_period, &value);
4992 
4993         return 0;
4994 }
4995 
4996 static const struct file_operations perf_fops;
4997 
4998 static inline int perf_fget_light(int fd, struct fd *p)
4999 {
5000         struct fd f = fdget(fd);
5001         if (!f.file)
5002                 return -EBADF;
5003 
5004         if (f.file->f_op != &perf_fops) {
5005                 fdput(f);
5006                 return -EBADF;
5007         }
5008         *p = f;
5009         return 0;
5010 }
5011 
5012 static int perf_event_set_output(struct perf_event *event,
5013                                  struct perf_event *output_event);
5014 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
5015 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
5016 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5017                           struct perf_event_attr *attr);
5018 
5019 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
5020 {
5021         void (*func)(struct perf_event *);
5022         u32 flags = arg;
5023 
5024         switch (cmd) {
5025         case PERF_EVENT_IOC_ENABLE:
5026                 func = _perf_event_enable;
5027                 break;
5028         case PERF_EVENT_IOC_DISABLE:
5029                 func = _perf_event_disable;
5030                 break;
5031         case PERF_EVENT_IOC_RESET:
5032                 func = _perf_event_reset;
5033                 break;
5034 
5035         case PERF_EVENT_IOC_REFRESH:
5036                 return _perf_event_refresh(event, arg);
5037 
5038         case PERF_EVENT_IOC_PERIOD:
5039                 return perf_event_period(event, (u64 __user *)arg);
5040 
5041         case PERF_EVENT_IOC_ID:
5042         {
5043                 u64 id = primary_event_id(event);
5044 
5045                 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
5046                         return -EFAULT;
5047                 return 0;
5048         }
5049 
5050         case PERF_EVENT_IOC_SET_OUTPUT:
5051         {
5052                 int ret;
5053                 if (arg != -1) {
5054                         struct perf_event *output_event;
5055                         struct fd output;
5056                         ret = perf_fget_light(arg, &output);
5057                         if (ret)
5058                                 return ret;
5059                         output_event = output.file->private_data;
5060                         ret = perf_event_set_output(event, output_event);
5061                         fdput(output);
5062                 } else {
5063                         ret = perf_event_set_output(event, NULL);
5064                 }
5065                 return ret;
5066         }
5067 
5068         case PERF_EVENT_IOC_SET_FILTER:
5069                 return perf_event_set_filter(event, (void __user *)arg);
5070 
5071         case PERF_EVENT_IOC_SET_BPF:
5072                 return perf_event_set_bpf_prog(event, arg);
5073 
5074         case PERF_EVENT_IOC_PAUSE_OUTPUT: {
5075                 struct ring_buffer *rb;
5076 
5077                 rcu_read_lock();
5078                 rb = rcu_dereference(event->rb);
5079                 if (!rb || !rb->nr_pages) {
5080                         rcu_read_unlock();
5081                         return -EINVAL;
5082                 }
5083                 rb_toggle_paused(rb, !!arg);
5084                 rcu_read_unlock();
5085                 return 0;
5086         }
5087 
5088         case PERF_EVENT_IOC_QUERY_BPF:
5089                 return perf_event_query_prog_array(event, (void __user *)arg);
5090 
5091         case PERF_EVENT_IOC_MODIFY_ATTRIBUTES: {
5092                 struct perf_event_attr new_attr;
5093                 int err = perf_copy_attr((struct perf_event_attr __user *)arg,
5094                                          &new_attr);
5095 
5096                 if (err)
5097                         return err;
5098 
5099                 return perf_event_modify_attr(event,  &new_attr);
5100         }
5101         default:
5102                 return -ENOTTY;
5103         }
5104 
5105         if (flags & PERF_IOC_FLAG_GROUP)
5106                 perf_event_for_each(event, func);
5107         else
5108                 perf_event_for_each_child(event, func);
5109 
5110         return 0;
5111 }
5112 
5113 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
5114 {
5115         struct perf_event *event = file->private_data;
5116         struct perf_event_context *ctx;
5117         long ret;
5118 
5119         ctx = perf_event_ctx_lock(event);
5120         ret = _perf_ioctl(event, cmd, arg);
5121         perf_event_ctx_unlock(event, ctx);
5122 
5123         return ret;
5124 }
5125 
5126 #ifdef CONFIG_COMPAT
5127 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
5128                                 unsigned long arg)
5129 {
5130         switch (_IOC_NR(cmd)) {
5131         case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
5132         case _IOC_NR(PERF_EVENT_IOC_ID):
5133         case _IOC_NR(PERF_EVENT_IOC_QUERY_BPF):
5134         case _IOC_NR(PERF_EVENT_IOC_MODIFY_ATTRIBUTES):
5135                 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
5136                 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
5137                         cmd &= ~IOCSIZE_MASK;
5138                         cmd |= sizeof(void *) << IOCSIZE_SHIFT;
5139                 }
5140                 break;
5141         }
5142         return perf_ioctl(file, cmd, arg);
5143 }
5144 #else
5145 # define perf_compat_ioctl NULL
5146 #endif
5147 
5148 int perf_event_task_enable(void)
5149 {
5150         struct perf_event_context *ctx;
5151         struct perf_event *event;
5152 
5153         mutex_lock(&current->perf_event_mutex);
5154         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5155                 ctx = perf_event_ctx_lock(event);
5156                 perf_event_for_each_child(event, _perf_event_enable);
5157                 perf_event_ctx_unlock(event, ctx);
5158         }
5159         mutex_unlock(&current->perf_event_mutex);
5160 
5161         return 0;
5162 }
5163 
5164 int perf_event_task_disable(void)
5165 {
5166         struct perf_event_context *ctx;
5167         struct perf_event *event;
5168 
5169         mutex_lock(&current->perf_event_mutex);
5170         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5171                 ctx = perf_event_ctx_lock(event);
5172                 perf_event_for_each_child(event, _perf_event_disable);
5173                 perf_event_ctx_unlock(event, ctx);
5174         }
5175         mutex_unlock(&current->perf_event_mutex);
5176 
5177         return 0;
5178 }
5179 
5180 static int perf_event_index(struct perf_event *event)
5181 {
5182         if (event->hw.state & PERF_HES_STOPPED)
5183                 return 0;
5184 
5185         if (event->state != PERF_EVENT_STATE_ACTIVE)
5186                 return 0;
5187 
5188         return event->pmu->event_idx(event);
5189 }
5190 
5191 static void calc_timer_values(struct perf_event *event,
5192                                 u64 *now,
5193                                 u64 *enabled,
5194                                 u64 *running)
5195 {
5196         u64 ctx_time;
5197 
5198         *now = perf_clock();
5199         ctx_time = event->shadow_ctx_time + *now;
5200         __perf_update_times(event, ctx_time, enabled, running);
5201 }
5202 
5203 static void perf_event_init_userpage(struct perf_event *event)
5204 {
5205         struct perf_event_mmap_page *userpg;
5206         struct ring_buffer *rb;
5207 
5208         rcu_read_lock();
5209         rb = rcu_dereference(event->rb);
5210         if (!rb)
5211                 goto unlock;
5212 
5213         userpg = rb->user_page;
5214 
5215         /* Allow new userspace to detect that bit 0 is deprecated */
5216         userpg->cap_bit0_is_deprecated = 1;
5217         userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
5218         userpg->data_offset = PAGE_SIZE;
5219         userpg->data_size = perf_data_size(rb);
5220 
5221 unlock:
5222         rcu_read_unlock();
5223 }
5224 
5225 void __weak arch_perf_update_userpage(
5226         struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
5227 {
5228 }
5229 
5230 /*
5231  * Callers need to ensure there can be no nesting of this function, otherwise
5232  * the seqlock logic goes bad. We can not serialize this because the arch
5233  * code calls this from NMI context.
5234  */
5235 void perf_event_update_userpage(struct perf_event *event)
5236 {
5237         struct perf_event_mmap_page *userpg;
5238         struct ring_buffer *rb;
5239         u64 enabled, running, now;
5240 
5241         rcu_read_lock();
5242         rb = rcu_dereference(event->rb);
5243         if (!rb)
5244                 goto unlock;
5245 
5246         /*
5247          * compute total_time_enabled, total_time_running
5248          * based on snapshot values taken when the event
5249          * was last scheduled in.
5250          *
5251          * we cannot simply called update_context_time()
5252          * because of locking issue as we can be called in
5253          * NMI context
5254          */
5255         calc_timer_values(event, &now, &enabled, &running);
5256 
5257         userpg = rb->user_page;
5258         /*
5259          * Disable preemption to guarantee consistent time stamps are stored to
5260          * the user page.
5261          */
5262         preempt_disable();
5263         ++userpg->lock;
5264         barrier();
5265         userpg->index = perf_event_index(event);
5266         userpg->offset = perf_event_count(event);
5267         if (userpg->index)
5268                 userpg->offset -= local64_read(&event->hw.prev_count);
5269 
5270         userpg->time_enabled = enabled +
5271                         atomic64_read(&event->child_total_time_enabled);
5272 
5273         userpg->time_running = running +
5274                         atomic64_read(&event->child_total_time_running);
5275 
5276         arch_perf_update_userpage(event, userpg, now);
5277 
5278         barrier();
5279         ++userpg->lock;
5280         preempt_enable();
5281 unlock:
5282         rcu_read_unlock();
5283 }
5284 EXPORT_SYMBOL_GPL(perf_event_update_userpage);
5285 
5286 static vm_fault_t perf_mmap_fault(struct vm_fault *vmf)
5287 {
5288         struct perf_event *event = vmf->vma->vm_file->private_data;
5289         struct ring_buffer *rb;
5290         vm_fault_t ret = VM_FAULT_SIGBUS;
5291 
5292         if (vmf->flags & FAULT_FLAG_MKWRITE) {
5293                 if (vmf->pgoff == 0)
5294                         ret = 0;
5295                 return ret;
5296         }
5297 
5298         rcu_read_lock();
5299         rb = rcu_dereference(event->rb);
5300         if (!rb)
5301                 goto unlock;
5302 
5303         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
5304                 goto unlock;
5305 
5306         vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
5307         if (!vmf->page)
5308                 goto unlock;
5309 
5310         get_page(vmf->page);
5311         vmf->page->mapping = vmf->vma->vm_file->f_mapping;
5312         vmf->page->index   = vmf->pgoff;
5313 
5314         ret = 0;
5315 unlock:
5316         rcu_read_unlock();
5317 
5318         return ret;
5319 }
5320 
5321 static void ring_buffer_attach(struct perf_event *event,
5322                                struct ring_buffer *rb)
5323 {
5324         struct ring_buffer *old_rb = NULL;
5325         unsigned long flags;
5326 
5327         if (event->rb) {
5328                 /*
5329                  * Should be impossible, we set this when removing
5330                  * event->rb_entry and wait/clear when adding event->rb_entry.
5331                  */
5332                 WARN_ON_ONCE(event->rcu_pending);
5333 
5334                 old_rb = event->rb;
5335                 spin_lock_irqsave(&old_rb->event_lock, flags);
5336                 list_del_rcu(&event->rb_entry);
5337                 spin_unlock_irqrestore(&old_rb->event_lock, flags);
5338 
5339                 event->rcu_batches = get_state_synchronize_rcu();
5340                 event->rcu_pending = 1;
5341         }
5342 
5343         if (rb) {
5344                 if (event->rcu_pending) {
5345                         cond_synchronize_rcu(event->rcu_batches);
5346                         event->rcu_pending = 0;
5347                 }
5348 
5349                 spin_lock_irqsave(&rb->event_lock, flags);
5350                 list_add_rcu(&event->rb_entry, &rb->event_list);
5351                 spin_unlock_irqrestore(&rb->event_lock, flags);
5352         }
5353 
5354         /*
5355          * Avoid racing with perf_mmap_close(AUX): stop the event
5356          * before swizzling the event::rb pointer; if it's getting
5357          * unmapped, its aux_mmap_count will be 0 and it won't
5358          * restart. See the comment in __perf_pmu_output_stop().
5359          *
5360          * Data will inevitably be lost when set_output is done in
5361          * mid-air, but then again, whoever does it like this is
5362          * not in for the data anyway.
5363          */
5364         if (has_aux(event))
5365                 perf_event_stop(event, 0);
5366 
5367         rcu_assign_pointer(event->rb, rb);
5368 
5369         if (old_rb) {
5370                 ring_buffer_put(old_rb);
5371                 /*
5372                  * Since we detached before setting the new rb, so that we
5373                  * could attach the new rb, we could have missed a wakeup.
5374                  * Provide it now.
5375                  */
5376                 wake_up_all(&event->waitq);
5377         }
5378 }
5379 
5380 static void ring_buffer_wakeup(struct perf_event *event)
5381 {
5382         struct ring_buffer *rb;
5383 
5384         rcu_read_lock();
5385         rb = rcu_dereference(event->rb);
5386         if (rb) {
5387                 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
5388                         wake_up_all(&event->waitq);
5389         }
5390         rcu_read_unlock();
5391 }
5392 
5393 struct ring_buffer *ring_buffer_get(struct perf_event *event)
5394 {
5395         struct ring_buffer *rb;
5396 
5397         rcu_read_lock();
5398         rb = rcu_dereference(event->rb);
5399         if (rb) {
5400                 if (!atomic_inc_not_zero(&rb->refcount))
5401                         rb = NULL;
5402         }
5403         rcu_read_unlock();
5404 
5405         return rb;
5406 }
5407 
5408 void ring_buffer_put(struct ring_buffer *rb)
5409 {
5410         if (!atomic_dec_and_test(&rb->refcount))
5411                 return;
5412 
5413         WARN_ON_ONCE(!list_empty(&rb->event_list));
5414 
5415         call_rcu(&rb->rcu_head, rb_free_rcu);
5416 }
5417 
5418 static void perf_mmap_open(struct vm_area_struct *vma)
5419 {
5420         struct perf_event *event = vma->vm_file->private_data;
5421 
5422         atomic_inc(&event->mmap_count);
5423         atomic_inc(&event->rb->mmap_count);
5424 
5425         if (vma->vm_pgoff)
5426                 atomic_inc(&event->rb->aux_mmap_count);
5427 
5428         if (event->pmu->event_mapped)
5429                 event->pmu->event_mapped(event, vma->vm_mm);
5430 }
5431 
5432 static void perf_pmu_output_stop(struct perf_event *event);
5433 
5434 /*
5435  * A buffer can be mmap()ed multiple times; either directly through the same
5436  * event, or through other events by use of perf_event_set_output().
5437  *
5438  * In order to undo the VM accounting done by perf_mmap() we need to destroy
5439  * the buffer here, where we still have a VM context. This means we need
5440  * to detach all events redirecting to us.
5441  */
5442 static void perf_mmap_close(struct vm_area_struct *vma)
5443 {
5444         struct perf_event *event = vma->vm_file->private_data;
5445 
5446         struct ring_buffer *rb = ring_buffer_get(event);
5447         struct user_struct *mmap_user = rb->mmap_user;
5448         int mmap_locked = rb->mmap_locked;
5449         unsigned long size = perf_data_size(rb);
5450 
5451         if (event->pmu->event_unmapped)
5452                 event->pmu->event_unmapped(event, vma->vm_mm);
5453 
5454         /*
5455          * rb->aux_mmap_count will always drop before rb->mmap_count and
5456          * event->mmap_count, so it is ok to use event->mmap_mutex to
5457          * serialize with perf_mmap here.
5458          */
5459         if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
5460             atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
5461                 /*
5462                  * Stop all AUX events that are writing to this buffer,
5463                  * so that we can free its AUX pages and corresponding PMU
5464                  * data. Note that after rb::aux_mmap_count dropped to zero,
5465                  * they won't start any more (see perf_aux_output_begin()).
5466                  */
5467                 perf_pmu_output_stop(event);
5468 
5469                 /* now it's safe to free the pages */
5470                 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
5471                 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
5472 
5473                 /* this has to be the last one */
5474                 rb_free_aux(rb);
5475                 WARN_ON_ONCE(atomic_read(&rb->aux_refcount));
5476 
5477                 mutex_unlock(&event->mmap_mutex);
5478         }
5479 
5480         atomic_dec(&rb->mmap_count);
5481 
5482         if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
5483                 goto out_put;
5484 
5485         ring_buffer_attach(event, NULL);
5486         mutex_unlock(&event->mmap_mutex);
5487 
5488         /* If there's still other mmap()s of this buffer, we're done. */
5489         if (atomic_read(&rb->mmap_count))
5490                 goto out_put;
5491 
5492         /*
5493          * No other mmap()s, detach from all other events that might redirect
5494          * into the now unreachable buffer. Somewhat complicated by the
5495          * fact that rb::event_lock otherwise nests inside mmap_mutex.
5496          */
5497 again:
5498         rcu_read_lock();
5499         list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
5500                 if (!atomic_long_inc_not_zero(&event->refcount)) {
5501                         /*
5502                          * This event is en-route to free_event() which will
5503                          * detach it and remove it from the list.
5504                          */
5505                         continue;
5506                 }
5507                 rcu_read_unlock();
5508 
5509                 mutex_lock(&event->mmap_mutex);
5510                 /*
5511                  * Check we didn't race with perf_event_set_output() which can
5512                  * swizzle the rb from under us while we were waiting to
5513                  * acquire mmap_mutex.
5514                  *
5515                  * If we find a different rb; ignore this event, a next
5516                  * iteration will no longer find it on the list. We have to
5517                  * still restart the iteration to make sure we're not now
5518                  * iterating the wrong list.
5519                  */
5520                 if (event->rb == rb)
5521                         ring_buffer_attach(event, NULL);
5522 
5523                 mutex_unlock(&event->mmap_mutex);
5524                 put_event(event);
5525 
5526                 /*
5527                  * Restart the iteration; either we're on the wrong list or
5528                  * destroyed its integrity by doing a deletion.
5529                  */
5530                 goto again;
5531         }
5532         rcu_read_unlock();
5533 
5534         /*
5535          * It could be there's still a few 0-ref events on the list; they'll
5536          * get cleaned up by free_event() -- they'll also still have their
5537          * ref on the rb and will free it whenever they are done with it.
5538          *
5539          * Aside from that, this buffer is 'fully' detached and unmapped,
5540          * undo the VM accounting.
5541          */
5542 
5543         atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
5544         vma->vm_mm->pinned_vm -= mmap_locked;
5545         free_uid(mmap_user);
5546 
5547 out_put:
5548         ring_buffer_put(rb); /* could be last */
5549 }
5550 
5551 static const struct vm_operations_struct perf_mmap_vmops = {
5552         .open           = perf_mmap_open,
5553         .close          = perf_mmap_close, /* non mergeable */
5554         .fault          = perf_mmap_fault,
5555         .page_mkwrite   = perf_mmap_fault,
5556 };
5557 
5558 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
5559 {
5560         struct perf_event *event = file->private_data;
5561         unsigned long user_locked, user_lock_limit;
5562         struct user_struct *user = current_user();
5563         unsigned long locked, lock_limit;
5564         struct ring_buffer *rb = NULL;
5565         unsigned long vma_size;
5566         unsigned long nr_pages;
5567         long user_extra = 0, extra = 0;
5568         int ret = 0, flags = 0;
5569 
5570         /*
5571          * Don't allow mmap() of inherited per-task counters. This would
5572          * create a performance issue due to all children writing to the
5573          * same rb.
5574          */
5575         if (event->cpu == -1 && event->attr.inherit)
5576                 return -EINVAL;
5577 
5578         if (!(vma->vm_flags & VM_SHARED))
5579                 return -EINVAL;
5580 
5581         vma_size = vma->vm_end - vma->vm_start;
5582 
5583         if (vma->vm_pgoff == 0) {
5584                 nr_pages = (vma_size / PAGE_SIZE) - 1;
5585         } else {
5586                 /*
5587                  * AUX area mapping: if rb->aux_nr_pages != 0, it's already
5588                  * mapped, all subsequent mappings should have the same size
5589                  * and offset. Must be above the normal perf buffer.
5590                  */
5591                 u64 aux_offset, aux_size;
5592 
5593                 if (!event->rb)
5594                         return -EINVAL;
5595 
5596                 nr_pages = vma_size / PAGE_SIZE;
5597 
5598                 mutex_lock(&event->mmap_mutex);
5599                 ret = -EINVAL;
5600 
5601                 rb = event->rb;
5602                 if (!rb)
5603                         goto aux_unlock;
5604 
5605                 aux_offset = READ_ONCE(rb->user_page->aux_offset);
5606                 aux_size = READ_ONCE(rb->user_page->aux_size);
5607 
5608                 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
5609                         goto aux_unlock;
5610 
5611                 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
5612                         goto aux_unlock;
5613 
5614                 /* already mapped with a different offset */
5615                 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
5616                         goto aux_unlock;
5617 
5618                 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
5619                         goto aux_unlock;
5620 
5621                 /* already mapped with a different size */
5622                 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
5623                         goto aux_unlock;
5624 
5625                 if (!is_power_of_2(nr_pages))
5626                         goto aux_unlock;
5627 
5628                 if (!atomic_inc_not_zero(&rb->mmap_count))
5629                         goto aux_unlock;
5630 
5631                 if (rb_has_aux(rb)) {
5632                         atomic_inc(&rb->aux_mmap_count);
5633                         ret = 0;
5634                         goto unlock;
5635                 }
5636 
5637                 atomic_set(&rb->aux_mmap_count, 1);
5638                 user_extra = nr_pages;
5639 
5640                 goto accounting;
5641         }
5642 
5643         /*
5644          * If we have rb pages ensure they're a power-of-two number, so we
5645          * can do bitmasks instead of modulo.
5646          */
5647         if (nr_pages != 0 && !is_power_of_2(nr_pages))
5648                 return -EINVAL;
5649 
5650         if (vma_size != PAGE_SIZE * (1 + nr_pages))
5651                 return -EINVAL;
5652 
5653         WARN_ON_ONCE(event->ctx->parent_ctx);
5654 again:
5655         mutex_lock(&event->mmap_mutex);
5656         if (event->rb) {
5657                 if (event->rb->nr_pages != nr_pages) {
5658                         ret = -EINVAL;
5659                         goto unlock;
5660                 }
5661 
5662                 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
5663                         /*
5664                          * Raced against perf_mmap_close() through
5665                          * perf_event_set_output(). Try again, hope for better
5666                          * luck.
5667                          */
5668                         mutex_unlock(&event->mmap_mutex);
5669                         goto again;
5670                 }
5671 
5672                 goto unlock;
5673         }
5674 
5675         user_extra = nr_pages + 1;
5676 
5677 accounting:
5678         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
5679 
5680         /*
5681          * Increase the limit linearly with more CPUs:
5682          */
5683         user_lock_limit *= num_online_cpus();
5684 
5685         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
5686 
5687         if (user_locked > user_lock_limit)
5688                 extra = user_locked - user_lock_limit;
5689 
5690         lock_limit = rlimit(RLIMIT_MEMLOCK);
5691         lock_limit >>= PAGE_SHIFT;
5692         locked = vma->vm_mm->pinned_vm + extra;
5693 
5694         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
5695                 !capable(CAP_IPC_LOCK)) {
5696                 ret = -EPERM;
5697                 goto unlock;
5698         }
5699 
5700         WARN_ON(!rb && event->rb);
5701 
5702         if (vma->vm_flags & VM_WRITE)
5703                 flags |= RING_BUFFER_WRITABLE;
5704 
5705         if (!rb) {
5706                 rb = rb_alloc(nr_pages,
5707                               event->attr.watermark ? event->attr.wakeup_watermark : 0,
5708                               event->cpu, flags);
5709 
5710                 if (!rb) {
5711                         ret = -ENOMEM;
5712                         goto unlock;
5713                 }
5714 
5715                 atomic_set(&rb->mmap_count, 1);
5716                 rb->mmap_user = get_current_user();
5717                 rb->mmap_locked = extra;
5718 
5719                 ring_buffer_attach(event, rb);
5720 
5721                 perf_event_init_userpage(event);
5722                 perf_event_update_userpage(event);
5723         } else {
5724                 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
5725                                    event->attr.aux_watermark, flags);
5726                 if (!ret)
5727                         rb->aux_mmap_locked = extra;
5728         }
5729 
5730 unlock:
5731         if (!ret) {
5732                 atomic_long_add(user_extra, &user->locked_vm);
5733                 vma->vm_mm->pinned_vm += extra;
5734 
5735                 atomic_inc(&event->mmap_count);
5736         } else if (rb) {
5737                 atomic_dec(&rb->mmap_count);
5738         }
5739 aux_unlock:
5740         mutex_unlock(&event->mmap_mutex);
5741 
5742         /*
5743          * Since pinned accounting is per vm we cannot allow fork() to copy our
5744          * vma.
5745          */
5746         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
5747         vma->vm_ops = &perf_mmap_vmops;
5748 
5749         if (event->pmu->event_mapped)
5750                 event->pmu->event_mapped(event, vma->vm_mm);
5751 
5752         return ret;
5753 }
5754 
5755 static int perf_fasync(int fd, struct file *filp, int on)
5756 {
5757         struct inode *inode = file_inode(filp);
5758         struct perf_event *event = filp->private_data;
5759         int retval;
5760 
5761         inode_lock(inode);
5762         retval = fasync_helper(fd, filp, on, &event->fasync);
5763         inode_unlock(inode);
5764 
5765         if (retval < 0)
5766                 return retval;
5767 
5768         return 0;
5769 }
5770 
5771 static const struct file_operations perf_fops = {
5772         .llseek                 = no_llseek,
5773         .release                = perf_release,
5774         .read                   = perf_read,
5775         .poll                   = perf_poll,
5776         .unlocked_ioctl         = perf_ioctl,
5777         .compat_ioctl           = perf_compat_ioctl,
5778         .mmap                   = perf_mmap,
5779         .fasync                 = perf_fasync,
5780 };
5781 
5782 /*
5783  * Perf event wakeup
5784  *
5785  * If there's data, ensure we set the poll() state and publish everything
5786  * to user-space before waking everybody up.
5787  */
5788 
5789 static inline struct fasync_struct **perf_event_fasync(struct perf_event *event)
5790 {
5791         /* only the parent has fasync state */
5792         if (event->parent)
5793                 event = event->parent;
5794         return &event->fasync;
5795 }
5796 
5797 void perf_event_wakeup(struct perf_event *event)
5798 {
5799         ring_buffer_wakeup(event);
5800 
5801         if (event->pending_kill) {
5802                 kill_fasync(perf_event_fasync(event), SIGIO, event->pending_kill);
5803                 event->pending_kill = 0;
5804         }
5805 }
5806 
5807 static void perf_pending_event_disable(struct perf_event *event)
5808 {
5809         int cpu = READ_ONCE(event->pending_disable);
5810 
5811         if (cpu < 0)
5812                 return;
5813 
5814         if (cpu == smp_processor_id()) {
5815                 WRITE_ONCE(event->pending_disable, -1);
5816                 perf_event_disable_local(event);
5817                 return;
5818         }
5819 
5820         /*
5821          *  CPU-A                       CPU-B
5822          *
5823          *  perf_event_disable_inatomic()
5824          *    @pending_disable = CPU-A;
5825          *    irq_work_queue();
5826          *
5827          *  sched-out
5828          *    @pending_disable = -1;
5829          *
5830          *                              sched-in
5831          *                              perf_event_disable_inatomic()
5832          *                                @pending_disable = CPU-B;
5833          *                                irq_work_queue(); // FAILS
5834          *
5835          *  irq_work_run()
5836          *    perf_pending_event()
5837          *
5838          * But the event runs on CPU-B and wants disabling there.
5839          */
5840         irq_work_queue_on(&event->pending, cpu);
5841 }
5842 
5843 static void perf_pending_event(struct irq_work *entry)
5844 {
5845         struct perf_event *event = container_of(entry, struct perf_event, pending);
5846         int rctx;
5847 
5848         rctx = perf_swevent_get_recursion_context();
5849         /*
5850          * If we 'fail' here, that's OK, it means recursion is already disabled
5851          * and we won't recurse 'further'.
5852          */
5853 
5854         perf_pending_event_disable(event);
5855 
5856         if (event->pending_wakeup) {
5857                 event->pending_wakeup = 0;
5858                 perf_event_wakeup(event);
5859         }
5860 
5861         if (rctx >= 0)
5862                 perf_swevent_put_recursion_context(rctx);
5863 }
5864 
5865 /*
5866  * We assume there is only KVM supporting the callbacks.
5867  * Later on, we might change it to a list if there is
5868  * another virtualization implementation supporting the callbacks.
5869  */
5870 struct perf_guest_info_callbacks *perf_guest_cbs;
5871 
5872 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5873 {
5874         perf_guest_cbs = cbs;
5875         return 0;
5876 }
5877 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
5878 
5879 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5880 {
5881         perf_guest_cbs = NULL;
5882         return 0;
5883 }
5884 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
5885 
5886 static void
5887 perf_output_sample_regs(struct perf_output_handle *handle,
5888                         struct pt_regs *regs, u64 mask)
5889 {
5890         int bit;
5891         DECLARE_BITMAP(_mask, 64);
5892 
5893         bitmap_from_u64(_mask, mask);
5894         for_each_set_bit(bit, _mask, sizeof(mask) * BITS_PER_BYTE) {
5895                 u64 val;
5896 
5897                 val = perf_reg_value(regs, bit);
5898                 perf_output_put(handle, val);
5899         }
5900 }
5901 
5902 static void perf_sample_regs_user(struct perf_regs *regs_user,
5903                                   struct pt_regs *regs,
5904                                   struct pt_regs *regs_user_copy)
5905 {
5906         if (user_mode(regs)) {
5907                 regs_user->abi = perf_reg_abi(current);
5908                 regs_user->regs = regs;
5909         } else if (current->mm) {
5910                 perf_get_regs_user(regs_user, regs, regs_user_copy);
5911         } else {
5912                 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
5913                 regs_user->regs = NULL;
5914         }
5915 }
5916 
5917 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
5918                                   struct pt_regs *regs)
5919 {
5920         regs_intr->regs = regs;
5921         regs_intr->abi  = perf_reg_abi(current);
5922 }
5923 
5924 
5925 /*
5926  * Get remaining task size from user stack pointer.
5927  *
5928  * It'd be better to take stack vma map and limit this more
5929  * precisly, but there's no way to get it safely under interrupt,
5930  * so using TASK_SIZE as limit.
5931  */
5932 static u64 perf_ustack_task_size(struct pt_regs *regs)
5933 {
5934         unsigned long addr = perf_user_stack_pointer(regs);
5935 
5936         if (!addr || addr >= TASK_SIZE)
5937                 return 0;
5938 
5939         return TASK_SIZE - addr;
5940 }
5941 
5942 static u16
5943 perf_sample_ustack_size(u16 stack_size, u16 header_size,
5944                         struct pt_regs *regs)
5945 {
5946         u64 task_size;
5947 
5948         /* No regs, no stack pointer, no dump. */
5949         if (!regs)
5950                 return 0;
5951 
5952         /*
5953          * Check if we fit in with the requested stack size into the:
5954          * - TASK_SIZE
5955          *   If we don't, we limit the size to the TASK_SIZE.
5956          *
5957          * - remaining sample size
5958          *   If we don't, we customize the stack size to
5959          *   fit in to the remaining sample size.
5960          */
5961 
5962         task_size  = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
5963         stack_size = min(stack_size, (u16) task_size);
5964 
5965         /* Current header size plus static size and dynamic size. */
5966         header_size += 2 * sizeof(u64);
5967 
5968         /* Do we fit in with the current stack dump size? */
5969         if ((u16) (header_size + stack_size) < header_size) {
5970                 /*
5971                  * If we overflow the maximum size for the sample,
5972                  * we customize the stack dump size to fit in.
5973                  */
5974                 stack_size = USHRT_MAX - header_size - sizeof(u64);
5975                 stack_size = round_up(stack_size, sizeof(u64));
5976         }
5977 
5978         return stack_size;
5979 }
5980 
5981 static void
5982 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
5983                           struct pt_regs *regs)
5984 {
5985         /* Case of a kernel thread, nothing to dump */
5986         if (!regs) {
5987                 u64 size = 0;
5988                 perf_output_put(handle, size);
5989         } else {
5990                 unsigned long sp;
5991                 unsigned int rem;
5992                 u64 dyn_size;
5993                 mm_segment_t fs;
5994 
5995                 /*
5996                  * We dump:
5997                  * static size
5998                  *   - the size requested by user or the best one we can fit
5999                  *     in to the sample max size
6000                  * data
6001                  *   - user stack dump data
6002                  * dynamic size
6003                  *   - the actual dumped size
6004                  */
6005 
6006                 /* Static size. */
6007                 perf_output_put(handle, dump_size);
6008 
6009                 /* Data. */
6010                 sp = perf_user_stack_pointer(regs);
6011                 fs = get_fs();
6012                 set_fs(USER_DS);
6013                 rem = __output_copy_user(handle, (void *) sp, dump_size);
6014                 set_fs(fs);
6015                 dyn_size = dump_size - rem;
6016 
6017                 perf_output_skip(handle, rem);
6018 
6019                 /* Dynamic size. */
6020                 perf_output_put(handle, dyn_size);
6021         }
6022 }
6023 
6024 static void __p