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

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

Version: ~ [ linux-5.11-rc3 ] ~ [ linux-5.10.7 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.89 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.167 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.215 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.251 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.251 ] ~ [ 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 = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
440 
441         if (ret || !write)
442                 return ret;
443 
444         /*
445          * If throttling is disabled don't allow the write:
446          */
447         if (sysctl_perf_cpu_time_max_percent == 100 ||
448             sysctl_perf_cpu_time_max_percent == 0)
449                 return -EINVAL;
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 a 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 a 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 (event->pending_disable) {
2011                 event->pending_disable = 0;
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 a 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         event->pending_disable = 1;
2200         irq_work_queue(&event->pending);
2201 }
2202 
2203 static void perf_set_shadow_time(struct perf_event *event,
2204                                  struct perf_event_context *ctx)
2205 {
2206         /*
2207          * use the correct time source for the time snapshot
2208          *
2209          * We could get by without this by leveraging the
2210          * fact that to get to this function, the caller
2211          * has most likely already called update_context_time()
2212          * and update_cgrp_time_xx() and thus both timestamp
2213          * are identical (or very close). Given that tstamp is,
2214          * already adjusted for cgroup, we could say that:
2215          *    tstamp - ctx->timestamp
2216          * is equivalent to
2217          *    tstamp - cgrp->timestamp.
2218          *
2219          * Then, in perf_output_read(), the calculation would
2220          * work with no changes because:
2221          * - event is guaranteed scheduled in
2222          * - no scheduled out in between
2223          * - thus the timestamp would be the same
2224          *
2225          * But this is a bit hairy.
2226          *
2227          * So instead, we have an explicit cgroup call to remain
2228          * within the time time source all along. We believe it
2229          * is cleaner and simpler to understand.
2230          */
2231         if (is_cgroup_event(event))
2232                 perf_cgroup_set_shadow_time(event, event->tstamp);
2233         else
2234                 event->shadow_ctx_time = event->tstamp - ctx->timestamp;
2235 }
2236 
2237 #define MAX_INTERRUPTS (~0ULL)
2238 
2239 static void perf_log_throttle(struct perf_event *event, int enable);
2240 static void perf_log_itrace_start(struct perf_event *event);
2241 
2242 static int
2243 event_sched_in(struct perf_event *event,
2244                  struct perf_cpu_context *cpuctx,
2245                  struct perf_event_context *ctx)
2246 {
2247         int ret = 0;
2248 
2249         lockdep_assert_held(&ctx->lock);
2250 
2251         if (event->state <= PERF_EVENT_STATE_OFF)
2252                 return 0;
2253 
2254         WRITE_ONCE(event->oncpu, smp_processor_id());
2255         /*
2256          * Order event::oncpu write to happen before the ACTIVE state is
2257          * visible. This allows perf_event_{stop,read}() to observe the correct
2258          * ->oncpu if it sees ACTIVE.
2259          */
2260         smp_wmb();
2261         perf_event_set_state(event, PERF_EVENT_STATE_ACTIVE);
2262 
2263         /*
2264          * Unthrottle events, since we scheduled we might have missed several
2265          * ticks already, also for a heavily scheduling task there is little
2266          * guarantee it'll get a tick in a timely manner.
2267          */
2268         if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
2269                 perf_log_throttle(event, 1);
2270                 event->hw.interrupts = 0;
2271         }
2272 
2273         perf_pmu_disable(event->pmu);
2274 
2275         perf_set_shadow_time(event, ctx);
2276 
2277         perf_log_itrace_start(event);
2278 
2279         if (event->pmu->add(event, PERF_EF_START)) {
2280                 perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2281                 event->oncpu = -1;
2282                 ret = -EAGAIN;
2283                 goto out;
2284         }
2285 
2286         if (!is_software_event(event))
2287                 cpuctx->active_oncpu++;
2288         if (!ctx->nr_active++)
2289                 perf_event_ctx_activate(ctx);
2290         if (event->attr.freq && event->attr.sample_freq)
2291                 ctx->nr_freq++;
2292 
2293         if (event->attr.exclusive)
2294                 cpuctx->exclusive = 1;
2295 
2296 out:
2297         perf_pmu_enable(event->pmu);
2298 
2299         return ret;
2300 }
2301 
2302 static int
2303 group_sched_in(struct perf_event *group_event,
2304                struct perf_cpu_context *cpuctx,
2305                struct perf_event_context *ctx)
2306 {
2307         struct perf_event *event, *partial_group = NULL;
2308         struct pmu *pmu = ctx->pmu;
2309 
2310         if (group_event->state == PERF_EVENT_STATE_OFF)
2311                 return 0;
2312 
2313         pmu->start_txn(pmu, PERF_PMU_TXN_ADD);
2314 
2315         if (event_sched_in(group_event, cpuctx, ctx)) {
2316                 pmu->cancel_txn(pmu);
2317                 perf_mux_hrtimer_restart(cpuctx);
2318                 return -EAGAIN;
2319         }
2320 
2321         /*
2322          * Schedule in siblings as one group (if any):
2323          */
2324         for_each_sibling_event(event, group_event) {
2325                 if (event_sched_in(event, cpuctx, ctx)) {
2326                         partial_group = event;
2327                         goto group_error;
2328                 }
2329         }
2330 
2331         if (!pmu->commit_txn(pmu))
2332                 return 0;
2333 
2334 group_error:
2335         /*
2336          * Groups can be scheduled in as one unit only, so undo any
2337          * partial group before returning:
2338          * The events up to the failed event are scheduled out normally.
2339          */
2340         for_each_sibling_event(event, group_event) {
2341                 if (event == partial_group)
2342                         break;
2343 
2344                 event_sched_out(event, cpuctx, ctx);
2345         }
2346         event_sched_out(group_event, cpuctx, ctx);
2347 
2348         pmu->cancel_txn(pmu);
2349 
2350         perf_mux_hrtimer_restart(cpuctx);
2351 
2352         return -EAGAIN;
2353 }
2354 
2355 /*
2356  * Work out whether we can put this event group on the CPU now.
2357  */
2358 static int group_can_go_on(struct perf_event *event,
2359                            struct perf_cpu_context *cpuctx,
2360                            int can_add_hw)
2361 {
2362         /*
2363          * Groups consisting entirely of software events can always go on.
2364          */
2365         if (event->group_caps & PERF_EV_CAP_SOFTWARE)
2366                 return 1;
2367         /*
2368          * If an exclusive group is already on, no other hardware
2369          * events can go on.
2370          */
2371         if (cpuctx->exclusive)
2372                 return 0;
2373         /*
2374          * If this group is exclusive and there are already
2375          * events on the CPU, it can't go on.
2376          */
2377         if (event->attr.exclusive && cpuctx->active_oncpu)
2378                 return 0;
2379         /*
2380          * Otherwise, try to add it if all previous groups were able
2381          * to go on.
2382          */
2383         return can_add_hw;
2384 }
2385 
2386 static void add_event_to_ctx(struct perf_event *event,
2387                                struct perf_event_context *ctx)
2388 {
2389         list_add_event(event, ctx);
2390         perf_group_attach(event);
2391 }
2392 
2393 static void ctx_sched_out(struct perf_event_context *ctx,
2394                           struct perf_cpu_context *cpuctx,
2395                           enum event_type_t event_type);
2396 static void
2397 ctx_sched_in(struct perf_event_context *ctx,
2398              struct perf_cpu_context *cpuctx,
2399              enum event_type_t event_type,
2400              struct task_struct *task);
2401 
2402 static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
2403                                struct perf_event_context *ctx,
2404                                enum event_type_t event_type)
2405 {
2406         if (!cpuctx->task_ctx)
2407                 return;
2408 
2409         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2410                 return;
2411 
2412         ctx_sched_out(ctx, cpuctx, event_type);
2413 }
2414 
2415 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2416                                 struct perf_event_context *ctx,
2417                                 struct task_struct *task)
2418 {
2419         cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2420         if (ctx)
2421                 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2422         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2423         if (ctx)
2424                 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2425 }
2426 
2427 /*
2428  * We want to maintain the following priority of scheduling:
2429  *  - CPU pinned (EVENT_CPU | EVENT_PINNED)
2430  *  - task pinned (EVENT_PINNED)
2431  *  - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
2432  *  - task flexible (EVENT_FLEXIBLE).
2433  *
2434  * In order to avoid unscheduling and scheduling back in everything every
2435  * time an event is added, only do it for the groups of equal priority and
2436  * below.
2437  *
2438  * This can be called after a batch operation on task events, in which case
2439  * event_type is a bit mask of the types of events involved. For CPU events,
2440  * event_type is only either EVENT_PINNED or EVENT_FLEXIBLE.
2441  */
2442 static void ctx_resched(struct perf_cpu_context *cpuctx,
2443                         struct perf_event_context *task_ctx,
2444                         enum event_type_t event_type)
2445 {
2446         enum event_type_t ctx_event_type;
2447         bool cpu_event = !!(event_type & EVENT_CPU);
2448 
2449         /*
2450          * If pinned groups are involved, flexible groups also need to be
2451          * scheduled out.
2452          */
2453         if (event_type & EVENT_PINNED)
2454                 event_type |= EVENT_FLEXIBLE;
2455 
2456         ctx_event_type = event_type & EVENT_ALL;
2457 
2458         perf_pmu_disable(cpuctx->ctx.pmu);
2459         if (task_ctx)
2460                 task_ctx_sched_out(cpuctx, task_ctx, event_type);
2461 
2462         /*
2463          * Decide which cpu ctx groups to schedule out based on the types
2464          * of events that caused rescheduling:
2465          *  - EVENT_CPU: schedule out corresponding groups;
2466          *  - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
2467          *  - otherwise, do nothing more.
2468          */
2469         if (cpu_event)
2470                 cpu_ctx_sched_out(cpuctx, ctx_event_type);
2471         else if (ctx_event_type & EVENT_PINNED)
2472                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2473 
2474         perf_event_sched_in(cpuctx, task_ctx, current);
2475         perf_pmu_enable(cpuctx->ctx.pmu);
2476 }
2477 
2478 /*
2479  * Cross CPU call to install and enable a performance event
2480  *
2481  * Very similar to remote_function() + event_function() but cannot assume that
2482  * things like ctx->is_active and cpuctx->task_ctx are set.
2483  */
2484 static int  __perf_install_in_context(void *info)
2485 {
2486         struct perf_event *event = info;
2487         struct perf_event_context *ctx = event->ctx;
2488         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2489         struct perf_event_context *task_ctx = cpuctx->task_ctx;
2490         bool reprogram = true;
2491         int ret = 0;
2492 
2493         raw_spin_lock(&cpuctx->ctx.lock);
2494         if (ctx->task) {
2495                 raw_spin_lock(&ctx->lock);
2496                 task_ctx = ctx;
2497 
2498                 reprogram = (ctx->task == current);
2499 
2500                 /*
2501                  * If the task is running, it must be running on this CPU,
2502                  * otherwise we cannot reprogram things.
2503                  *
2504                  * If its not running, we don't care, ctx->lock will
2505                  * serialize against it becoming runnable.
2506                  */
2507                 if (task_curr(ctx->task) && !reprogram) {
2508                         ret = -ESRCH;
2509                         goto unlock;
2510                 }
2511 
2512                 WARN_ON_ONCE(reprogram && cpuctx->task_ctx && cpuctx->task_ctx != ctx);
2513         } else if (task_ctx) {
2514                 raw_spin_lock(&task_ctx->lock);
2515         }
2516 
2517 #ifdef CONFIG_CGROUP_PERF
2518         if (is_cgroup_event(event)) {
2519                 /*
2520                  * If the current cgroup doesn't match the event's
2521                  * cgroup, we should not try to schedule it.
2522                  */
2523                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
2524                 reprogram = cgroup_is_descendant(cgrp->css.cgroup,
2525                                         event->cgrp->css.cgroup);
2526         }
2527 #endif
2528 
2529         if (reprogram) {
2530                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2531                 add_event_to_ctx(event, ctx);
2532                 ctx_resched(cpuctx, task_ctx, get_event_type(event));
2533         } else {
2534                 add_event_to_ctx(event, ctx);
2535         }
2536 
2537 unlock:
2538         perf_ctx_unlock(cpuctx, task_ctx);
2539 
2540         return ret;
2541 }
2542 
2543 /*
2544  * Attach a performance event to a context.
2545  *
2546  * Very similar to event_function_call, see comment there.
2547  */
2548 static void
2549 perf_install_in_context(struct perf_event_context *ctx,
2550                         struct perf_event *event,
2551                         int cpu)
2552 {
2553         struct task_struct *task = READ_ONCE(ctx->task);
2554 
2555         lockdep_assert_held(&ctx->mutex);
2556 
2557         if (event->cpu != -1)
2558                 event->cpu = cpu;
2559 
2560         /*
2561          * Ensures that if we can observe event->ctx, both the event and ctx
2562          * will be 'complete'. See perf_iterate_sb_cpu().
2563          */
2564         smp_store_release(&event->ctx, ctx);
2565 
2566         if (!task) {
2567                 cpu_function_call(cpu, __perf_install_in_context, event);
2568                 return;
2569         }
2570 
2571         /*
2572          * Should not happen, we validate the ctx is still alive before calling.
2573          */
2574         if (WARN_ON_ONCE(task == TASK_TOMBSTONE))
2575                 return;
2576 
2577         /*
2578          * Installing events is tricky because we cannot rely on ctx->is_active
2579          * to be set in case this is the nr_events 0 -> 1 transition.
2580          *
2581          * Instead we use task_curr(), which tells us if the task is running.
2582          * However, since we use task_curr() outside of rq::lock, we can race
2583          * against the actual state. This means the result can be wrong.
2584          *
2585          * If we get a false positive, we retry, this is harmless.
2586          *
2587          * If we get a false negative, things are complicated. If we are after
2588          * perf_event_context_sched_in() ctx::lock will serialize us, and the
2589          * value must be correct. If we're before, it doesn't matter since
2590          * perf_event_context_sched_in() will program the counter.
2591          *
2592          * However, this hinges on the remote context switch having observed
2593          * our task->perf_event_ctxp[] store, such that it will in fact take
2594          * ctx::lock in perf_event_context_sched_in().
2595          *
2596          * We do this by task_function_call(), if the IPI fails to hit the task
2597          * we know any future context switch of task must see the
2598          * perf_event_ctpx[] store.
2599          */
2600 
2601         /*
2602          * This smp_mb() orders the task->perf_event_ctxp[] store with the
2603          * task_cpu() load, such that if the IPI then does not find the task
2604          * running, a future context switch of that task must observe the
2605          * store.
2606          */
2607         smp_mb();
2608 again:
2609         if (!task_function_call(task, __perf_install_in_context, event))
2610                 return;
2611 
2612         raw_spin_lock_irq(&ctx->lock);
2613         task = ctx->task;
2614         if (WARN_ON_ONCE(task == TASK_TOMBSTONE)) {
2615                 /*
2616                  * Cannot happen because we already checked above (which also
2617                  * cannot happen), and we hold ctx->mutex, which serializes us
2618                  * against perf_event_exit_task_context().
2619                  */
2620                 raw_spin_unlock_irq(&ctx->lock);
2621                 return;
2622         }
2623         /*
2624          * If the task is not running, ctx->lock will avoid it becoming so,
2625          * thus we can safely install the event.
2626          */
2627         if (task_curr(task)) {
2628                 raw_spin_unlock_irq(&ctx->lock);
2629                 goto again;
2630         }
2631         add_event_to_ctx(event, ctx);
2632         raw_spin_unlock_irq(&ctx->lock);
2633 }
2634 
2635 /*
2636  * Cross CPU call to enable a performance event
2637  */
2638 static void __perf_event_enable(struct perf_event *event,
2639                                 struct perf_cpu_context *cpuctx,
2640                                 struct perf_event_context *ctx,
2641                                 void *info)
2642 {
2643         struct perf_event *leader = event->group_leader;
2644         struct perf_event_context *task_ctx;
2645 
2646         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2647             event->state <= PERF_EVENT_STATE_ERROR)
2648                 return;
2649 
2650         if (ctx->is_active)
2651                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2652 
2653         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2654 
2655         if (!ctx->is_active)
2656                 return;
2657 
2658         if (!event_filter_match(event)) {
2659                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2660                 return;
2661         }
2662 
2663         /*
2664          * If the event is in a group and isn't the group leader,
2665          * then don't put it on unless the group is on.
2666          */
2667         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) {
2668                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2669                 return;
2670         }
2671 
2672         task_ctx = cpuctx->task_ctx;
2673         if (ctx->task)
2674                 WARN_ON_ONCE(task_ctx != ctx);
2675 
2676         ctx_resched(cpuctx, task_ctx, get_event_type(event));
2677 }
2678 
2679 /*
2680  * Enable a event.
2681  *
2682  * If event->ctx is a cloned context, callers must make sure that
2683  * every task struct that event->ctx->task could possibly point to
2684  * remains valid.  This condition is satisfied when called through
2685  * perf_event_for_each_child or perf_event_for_each as described
2686  * for perf_event_disable.
2687  */
2688 static void _perf_event_enable(struct perf_event *event)
2689 {
2690         struct perf_event_context *ctx = event->ctx;
2691 
2692         raw_spin_lock_irq(&ctx->lock);
2693         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2694             event->state <  PERF_EVENT_STATE_ERROR) {
2695                 raw_spin_unlock_irq(&ctx->lock);
2696                 return;
2697         }
2698 
2699         /*
2700          * If the event is in error state, clear that first.
2701          *
2702          * That way, if we see the event in error state below, we know that it
2703          * has gone back into error state, as distinct from the task having
2704          * been scheduled away before the cross-call arrived.
2705          */
2706         if (event->state == PERF_EVENT_STATE_ERROR)
2707                 event->state = PERF_EVENT_STATE_OFF;
2708         raw_spin_unlock_irq(&ctx->lock);
2709 
2710         event_function_call(event, __perf_event_enable, NULL);
2711 }
2712 
2713 /*
2714  * See perf_event_disable();
2715  */
2716 void perf_event_enable(struct perf_event *event)
2717 {
2718         struct perf_event_context *ctx;
2719 
2720         ctx = perf_event_ctx_lock(event);
2721         _perf_event_enable(event);
2722         perf_event_ctx_unlock(event, ctx);
2723 }
2724 EXPORT_SYMBOL_GPL(perf_event_enable);
2725 
2726 struct stop_event_data {
2727         struct perf_event       *event;
2728         unsigned int            restart;
2729 };
2730 
2731 static int __perf_event_stop(void *info)
2732 {
2733         struct stop_event_data *sd = info;
2734         struct perf_event *event = sd->event;
2735 
2736         /* if it's already INACTIVE, do nothing */
2737         if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2738                 return 0;
2739 
2740         /* matches smp_wmb() in event_sched_in() */
2741         smp_rmb();
2742 
2743         /*
2744          * There is a window with interrupts enabled before we get here,
2745          * so we need to check again lest we try to stop another CPU's event.
2746          */
2747         if (READ_ONCE(event->oncpu) != smp_processor_id())
2748                 return -EAGAIN;
2749 
2750         event->pmu->stop(event, PERF_EF_UPDATE);
2751 
2752         /*
2753          * May race with the actual stop (through perf_pmu_output_stop()),
2754          * but it is only used for events with AUX ring buffer, and such
2755          * events will refuse to restart because of rb::aux_mmap_count==0,
2756          * see comments in perf_aux_output_begin().
2757          *
2758          * Since this is happening on a event-local CPU, no trace is lost
2759          * while restarting.
2760          */
2761         if (sd->restart)
2762                 event->pmu->start(event, 0);
2763 
2764         return 0;
2765 }
2766 
2767 static int perf_event_stop(struct perf_event *event, int restart)
2768 {
2769         struct stop_event_data sd = {
2770                 .event          = event,
2771                 .restart        = restart,
2772         };
2773         int ret = 0;
2774 
2775         do {
2776                 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2777                         return 0;
2778 
2779                 /* matches smp_wmb() in event_sched_in() */
2780                 smp_rmb();
2781 
2782                 /*
2783                  * We only want to restart ACTIVE events, so if the event goes
2784                  * inactive here (event->oncpu==-1), there's nothing more to do;
2785                  * fall through with ret==-ENXIO.
2786                  */
2787                 ret = cpu_function_call(READ_ONCE(event->oncpu),
2788                                         __perf_event_stop, &sd);
2789         } while (ret == -EAGAIN);
2790 
2791         return ret;
2792 }
2793 
2794 /*
2795  * In order to contain the amount of racy and tricky in the address filter
2796  * configuration management, it is a two part process:
2797  *
2798  * (p1) when userspace mappings change as a result of (1) or (2) or (3) below,
2799  *      we update the addresses of corresponding vmas in
2800  *      event::addr_filters_offs array and bump the event::addr_filters_gen;
2801  * (p2) when an event is scheduled in (pmu::add), it calls
2802  *      perf_event_addr_filters_sync() which calls pmu::addr_filters_sync()
2803  *      if the generation has changed since the previous call.
2804  *
2805  * If (p1) happens while the event is active, we restart it to force (p2).
2806  *
2807  * (1) perf_addr_filters_apply(): adjusting filters' offsets based on
2808  *     pre-existing mappings, called once when new filters arrive via SET_FILTER
2809  *     ioctl;
2810  * (2) perf_addr_filters_adjust(): adjusting filters' offsets based on newly
2811  *     registered mapping, called for every new mmap(), with mm::mmap_sem down
2812  *     for reading;
2813  * (3) perf_event_addr_filters_exec(): clearing filters' offsets in the process
2814  *     of exec.
2815  */
2816 void perf_event_addr_filters_sync(struct perf_event *event)
2817 {
2818         struct perf_addr_filters_head *ifh = perf_event_addr_filters(event);
2819 
2820         if (!has_addr_filter(event))
2821                 return;
2822 
2823         raw_spin_lock(&ifh->lock);
2824         if (event->addr_filters_gen != event->hw.addr_filters_gen) {
2825                 event->pmu->addr_filters_sync(event);
2826                 event->hw.addr_filters_gen = event->addr_filters_gen;
2827         }
2828         raw_spin_unlock(&ifh->lock);
2829 }
2830 EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
2831 
2832 static int _perf_event_refresh(struct perf_event *event, int refresh)
2833 {
2834         /*
2835          * not supported on inherited events
2836          */
2837         if (event->attr.inherit || !is_sampling_event(event))
2838                 return -EINVAL;
2839 
2840         atomic_add(refresh, &event->event_limit);
2841         _perf_event_enable(event);
2842 
2843         return 0;
2844 }
2845 
2846 /*
2847  * See perf_event_disable()
2848  */
2849 int perf_event_refresh(struct perf_event *event, int refresh)
2850 {
2851         struct perf_event_context *ctx;
2852         int ret;
2853 
2854         ctx = perf_event_ctx_lock(event);
2855         ret = _perf_event_refresh(event, refresh);
2856         perf_event_ctx_unlock(event, ctx);
2857 
2858         return ret;
2859 }
2860 EXPORT_SYMBOL_GPL(perf_event_refresh);
2861 
2862 static int perf_event_modify_breakpoint(struct perf_event *bp,
2863                                          struct perf_event_attr *attr)
2864 {
2865         int err;
2866 
2867         _perf_event_disable(bp);
2868 
2869         err = modify_user_hw_breakpoint_check(bp, attr, true);
2870         if (err) {
2871                 if (!bp->attr.disabled)
2872                         _perf_event_enable(bp);
2873 
2874                 return err;
2875         }
2876 
2877         if (!attr->disabled)
2878                 _perf_event_enable(bp);
2879         return 0;
2880 }
2881 
2882 static int perf_event_modify_attr(struct perf_event *event,
2883                                   struct perf_event_attr *attr)
2884 {
2885         if (event->attr.type != attr->type)
2886                 return -EINVAL;
2887 
2888         switch (event->attr.type) {
2889         case PERF_TYPE_BREAKPOINT:
2890                 return perf_event_modify_breakpoint(event, attr);
2891         default:
2892                 /* Place holder for future additions. */
2893                 return -EOPNOTSUPP;
2894         }
2895 }
2896 
2897 static void ctx_sched_out(struct perf_event_context *ctx,
2898                           struct perf_cpu_context *cpuctx,
2899                           enum event_type_t event_type)
2900 {
2901         struct perf_event *event, *tmp;
2902         int is_active = ctx->is_active;
2903 
2904         lockdep_assert_held(&ctx->lock);
2905 
2906         if (likely(!ctx->nr_events)) {
2907                 /*
2908                  * See __perf_remove_from_context().
2909                  */
2910                 WARN_ON_ONCE(ctx->is_active);
2911                 if (ctx->task)
2912                         WARN_ON_ONCE(cpuctx->task_ctx);
2913                 return;
2914         }
2915 
2916         ctx->is_active &= ~event_type;
2917         if (!(ctx->is_active & EVENT_ALL))
2918                 ctx->is_active = 0;
2919 
2920         if (ctx->task) {
2921                 WARN_ON_ONCE(cpuctx->task_ctx != ctx);
2922                 if (!ctx->is_active)
2923                         cpuctx->task_ctx = NULL;
2924         }
2925 
2926         /*
2927          * Always update time if it was set; not only when it changes.
2928          * Otherwise we can 'forget' to update time for any but the last
2929          * context we sched out. For example:
2930          *
2931          *   ctx_sched_out(.event_type = EVENT_FLEXIBLE)
2932          *   ctx_sched_out(.event_type = EVENT_PINNED)
2933          *
2934          * would only update time for the pinned events.
2935          */
2936         if (is_active & EVENT_TIME) {
2937                 /* update (and stop) ctx time */
2938                 update_context_time(ctx);
2939                 update_cgrp_time_from_cpuctx(cpuctx);
2940         }
2941 
2942         is_active ^= ctx->is_active; /* changed bits */
2943 
2944         if (!ctx->nr_active || !(is_active & EVENT_ALL))
2945                 return;
2946 
2947         perf_pmu_disable(ctx->pmu);
2948         if (is_active & EVENT_PINNED) {
2949                 list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list)
2950                         group_sched_out(event, cpuctx, ctx);
2951         }
2952 
2953         if (is_active & EVENT_FLEXIBLE) {
2954                 list_for_each_entry_safe(event, tmp, &ctx->flexible_active, active_list)
2955                         group_sched_out(event, cpuctx, ctx);
2956         }
2957         perf_pmu_enable(ctx->pmu);
2958 }
2959 
2960 /*
2961  * Test whether two contexts are equivalent, i.e. whether they have both been
2962  * cloned from the same version of the same context.
2963  *
2964  * Equivalence is measured using a generation number in the context that is
2965  * incremented on each modification to it; see unclone_ctx(), list_add_event()
2966  * and list_del_event().
2967  */
2968 static int context_equiv(struct perf_event_context *ctx1,
2969                          struct perf_event_context *ctx2)
2970 {
2971         lockdep_assert_held(&ctx1->lock);
2972         lockdep_assert_held(&ctx2->lock);
2973 
2974         /* Pinning disables the swap optimization */
2975         if (ctx1->pin_count || ctx2->pin_count)
2976                 return 0;
2977 
2978         /* If ctx1 is the parent of ctx2 */
2979         if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2980                 return 1;
2981 
2982         /* If ctx2 is the parent of ctx1 */
2983         if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2984                 return 1;
2985 
2986         /*
2987          * If ctx1 and ctx2 have the same parent; we flatten the parent
2988          * hierarchy, see perf_event_init_context().
2989          */
2990         if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2991                         ctx1->parent_gen == ctx2->parent_gen)
2992                 return 1;
2993 
2994         /* Unmatched */
2995         return 0;
2996 }
2997 
2998 static void __perf_event_sync_stat(struct perf_event *event,
2999                                      struct perf_event *next_event)
3000 {
3001         u64 value;
3002 
3003         if (!event->attr.inherit_stat)
3004                 return;
3005 
3006         /*
3007          * Update the event value, we cannot use perf_event_read()
3008          * because we're in the middle of a context switch and have IRQs
3009          * disabled, which upsets smp_call_function_single(), however
3010          * we know the event must be on the current CPU, therefore we
3011          * don't need to use it.
3012          */
3013         if (event->state == PERF_EVENT_STATE_ACTIVE)
3014                 event->pmu->read(event);
3015 
3016         perf_event_update_time(event);
3017 
3018         /*
3019          * In order to keep per-task stats reliable we need to flip the event
3020          * values when we flip the contexts.
3021          */
3022         value = local64_read(&next_event->count);
3023         value = local64_xchg(&event->count, value);
3024         local64_set(&next_event->count, value);
3025 
3026         swap(event->total_time_enabled, next_event->total_time_enabled);
3027         swap(event->total_time_running, next_event->total_time_running);
3028 
3029         /*
3030          * Since we swizzled the values, update the user visible data too.
3031          */
3032         perf_event_update_userpage(event);
3033         perf_event_update_userpage(next_event);
3034 }
3035 
3036 static void perf_event_sync_stat(struct perf_event_context *ctx,
3037                                    struct perf_event_context *next_ctx)
3038 {
3039         struct perf_event *event, *next_event;
3040 
3041         if (!ctx->nr_stat)
3042                 return;
3043 
3044         update_context_time(ctx);
3045 
3046         event = list_first_entry(&ctx->event_list,
3047                                    struct perf_event, event_entry);
3048 
3049         next_event = list_first_entry(&next_ctx->event_list,
3050                                         struct perf_event, event_entry);
3051 
3052         while (&event->event_entry != &ctx->event_list &&
3053                &next_event->event_entry != &next_ctx->event_list) {
3054 
3055                 __perf_event_sync_stat(event, next_event);
3056 
3057                 event = list_next_entry(event, event_entry);
3058                 next_event = list_next_entry(next_event, event_entry);
3059         }
3060 }
3061 
3062 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
3063                                          struct task_struct *next)
3064 {
3065         struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
3066         struct perf_event_context *next_ctx;
3067         struct perf_event_context *parent, *next_parent;
3068         struct perf_cpu_context *cpuctx;
3069         int do_switch = 1;
3070 
3071         if (likely(!ctx))
3072                 return;
3073 
3074         cpuctx = __get_cpu_context(ctx);
3075         if (!cpuctx->task_ctx)
3076                 return;
3077 
3078         rcu_read_lock();
3079         next_ctx = next->perf_event_ctxp[ctxn];
3080         if (!next_ctx)
3081                 goto unlock;
3082 
3083         parent = rcu_dereference(ctx->parent_ctx);
3084         next_parent = rcu_dereference(next_ctx->parent_ctx);
3085 
3086         /* If neither context have a parent context; they cannot be clones. */
3087         if (!parent && !next_parent)
3088                 goto unlock;
3089 
3090         if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
3091                 /*
3092                  * Looks like the two contexts are clones, so we might be
3093                  * able to optimize the context switch.  We lock both
3094                  * contexts and check that they are clones under the
3095                  * lock (including re-checking that neither has been
3096                  * uncloned in the meantime).  It doesn't matter which
3097                  * order we take the locks because no other cpu could
3098                  * be trying to lock both of these tasks.
3099                  */
3100                 raw_spin_lock(&ctx->lock);
3101                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
3102                 if (context_equiv(ctx, next_ctx)) {
3103                         WRITE_ONCE(ctx->task, next);
3104                         WRITE_ONCE(next_ctx->task, task);
3105 
3106                         swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
3107 
3108                         /*
3109                          * RCU_INIT_POINTER here is safe because we've not
3110                          * modified the ctx and the above modification of
3111                          * ctx->task and ctx->task_ctx_data are immaterial
3112                          * since those values are always verified under
3113                          * ctx->lock which we're now holding.
3114                          */
3115                         RCU_INIT_POINTER(task->perf_event_ctxp[ctxn], next_ctx);
3116                         RCU_INIT_POINTER(next->perf_event_ctxp[ctxn], ctx);
3117 
3118                         do_switch = 0;
3119 
3120                         perf_event_sync_stat(ctx, next_ctx);
3121                 }
3122                 raw_spin_unlock(&next_ctx->lock);
3123                 raw_spin_unlock(&ctx->lock);
3124         }
3125 unlock:
3126         rcu_read_unlock();
3127 
3128         if (do_switch) {
3129                 raw_spin_lock(&ctx->lock);
3130                 task_ctx_sched_out(cpuctx, ctx, EVENT_ALL);
3131                 raw_spin_unlock(&ctx->lock);
3132         }
3133 }
3134 
3135 static DEFINE_PER_CPU(struct list_head, sched_cb_list);
3136 
3137 void perf_sched_cb_dec(struct pmu *pmu)
3138 {
3139         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3140 
3141         this_cpu_dec(perf_sched_cb_usages);
3142 
3143         if (!--cpuctx->sched_cb_usage)
3144                 list_del(&cpuctx->sched_cb_entry);
3145 }
3146 
3147 
3148 void perf_sched_cb_inc(struct pmu *pmu)
3149 {
3150         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3151 
3152         if (!cpuctx->sched_cb_usage++)
3153                 list_add(&cpuctx->sched_cb_entry, this_cpu_ptr(&sched_cb_list));
3154 
3155         this_cpu_inc(perf_sched_cb_usages);
3156 }
3157 
3158 /*
3159  * This function provides the context switch callback to the lower code
3160  * layer. It is invoked ONLY when the context switch callback is enabled.
3161  *
3162  * This callback is relevant even to per-cpu events; for example multi event
3163  * PEBS requires this to provide PID/TID information. This requires we flush
3164  * all queued PEBS records before we context switch to a new task.
3165  */
3166 static void perf_pmu_sched_task(struct task_struct *prev,
3167                                 struct task_struct *next,
3168                                 bool sched_in)
3169 {
3170         struct perf_cpu_context *cpuctx;
3171         struct pmu *pmu;
3172 
3173         if (prev == next)
3174                 return;
3175 
3176         list_for_each_entry(cpuctx, this_cpu_ptr(&sched_cb_list), sched_cb_entry) {
3177                 pmu = cpuctx->ctx.pmu; /* software PMUs will not have sched_task */
3178 
3179                 if (WARN_ON_ONCE(!pmu->sched_task))
3180                         continue;
3181 
3182                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3183                 perf_pmu_disable(pmu);
3184 
3185                 pmu->sched_task(cpuctx->task_ctx, sched_in);
3186 
3187                 perf_pmu_enable(pmu);
3188                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3189         }
3190 }
3191 
3192 static void perf_event_switch(struct task_struct *task,
3193                               struct task_struct *next_prev, bool sched_in);
3194 
3195 #define for_each_task_context_nr(ctxn)                                  \
3196         for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
3197 
3198 /*
3199  * Called from scheduler to remove the events of the current task,
3200  * with interrupts disabled.
3201  *
3202  * We stop each event and update the event value in event->count.
3203  *
3204  * This does not protect us against NMI, but disable()
3205  * sets the disabled bit in the control field of event _before_
3206  * accessing the event control register. If a NMI hits, then it will
3207  * not restart the event.
3208  */
3209 void __perf_event_task_sched_out(struct task_struct *task,
3210                                  struct task_struct *next)
3211 {
3212         int ctxn;
3213 
3214         if (__this_cpu_read(perf_sched_cb_usages))
3215                 perf_pmu_sched_task(task, next, false);
3216 
3217         if (atomic_read(&nr_switch_events))
3218                 perf_event_switch(task, next, false);
3219 
3220         for_each_task_context_nr(ctxn)
3221                 perf_event_context_sched_out(task, ctxn, next);
3222 
3223         /*
3224          * if cgroup events exist on this CPU, then we need
3225          * to check if we have to switch out PMU state.
3226          * cgroup event are system-wide mode only
3227          */
3228         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3229                 perf_cgroup_sched_out(task, next);
3230 }
3231 
3232 /*
3233  * Called with IRQs disabled
3234  */
3235 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
3236                               enum event_type_t event_type)
3237 {
3238         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
3239 }
3240 
3241 static int visit_groups_merge(struct perf_event_groups *groups, int cpu,
3242                               int (*func)(struct perf_event *, void *), void *data)
3243 {
3244         struct perf_event **evt, *evt1, *evt2;
3245         int ret;
3246 
3247         evt1 = perf_event_groups_first(groups, -1);
3248         evt2 = perf_event_groups_first(groups, cpu);
3249 
3250         while (evt1 || evt2) {
3251                 if (evt1 && evt2) {
3252                         if (evt1->group_index < evt2->group_index)
3253                                 evt = &evt1;
3254                         else
3255                                 evt = &evt2;
3256                 } else if (evt1) {
3257                         evt = &evt1;
3258                 } else {
3259                         evt = &evt2;
3260                 }
3261 
3262                 ret = func(*evt, data);
3263                 if (ret)
3264                         return ret;
3265 
3266                 *evt = perf_event_groups_next(*evt);
3267         }
3268 
3269         return 0;
3270 }
3271 
3272 struct sched_in_data {
3273         struct perf_event_context *ctx;
3274         struct perf_cpu_context *cpuctx;
3275         int can_add_hw;
3276 };
3277 
3278 static int pinned_sched_in(struct perf_event *event, void *data)
3279 {
3280         struct sched_in_data *sid = data;
3281 
3282         if (event->state <= PERF_EVENT_STATE_OFF)
3283                 return 0;
3284 
3285         if (!event_filter_match(event))
3286                 return 0;
3287 
3288         if (group_can_go_on(event, sid->cpuctx, sid->can_add_hw)) {
3289                 if (!group_sched_in(event, sid->cpuctx, sid->ctx))
3290                         list_add_tail(&event->active_list, &sid->ctx->pinned_active);
3291         }
3292 
3293         /*
3294          * If this pinned group hasn't been scheduled,
3295          * put it in error state.
3296          */
3297         if (event->state == PERF_EVENT_STATE_INACTIVE)
3298                 perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
3299 
3300         return 0;
3301 }
3302 
3303 static int flexible_sched_in(struct perf_event *event, void *data)
3304 {
3305         struct sched_in_data *sid = data;
3306 
3307         if (event->state <= PERF_EVENT_STATE_OFF)
3308                 return 0;
3309 
3310         if (!event_filter_match(event))
3311                 return 0;
3312 
3313         if (group_can_go_on(event, sid->cpuctx, sid->can_add_hw)) {
3314                 if (!group_sched_in(event, sid->cpuctx, sid->ctx))
3315                         list_add_tail(&event->active_list, &sid->ctx->flexible_active);
3316                 else
3317                         sid->can_add_hw = 0;
3318         }
3319 
3320         return 0;
3321 }
3322 
3323 static void
3324 ctx_pinned_sched_in(struct perf_event_context *ctx,
3325                     struct perf_cpu_context *cpuctx)
3326 {
3327         struct sched_in_data sid = {
3328                 .ctx = ctx,
3329                 .cpuctx = cpuctx,
3330                 .can_add_hw = 1,
3331         };
3332 
3333         visit_groups_merge(&ctx->pinned_groups,
3334                            smp_processor_id(),
3335                            pinned_sched_in, &sid);
3336 }
3337 
3338 static void
3339 ctx_flexible_sched_in(struct perf_event_context *ctx,
3340                       struct perf_cpu_context *cpuctx)
3341 {
3342         struct sched_in_data sid = {
3343                 .ctx = ctx,
3344                 .cpuctx = cpuctx,
3345                 .can_add_hw = 1,
3346         };
3347 
3348         visit_groups_merge(&ctx->flexible_groups,
3349                            smp_processor_id(),
3350                            flexible_sched_in, &sid);
3351 }
3352 
3353 static void
3354 ctx_sched_in(struct perf_event_context *ctx,
3355              struct perf_cpu_context *cpuctx,
3356              enum event_type_t event_type,
3357              struct task_struct *task)
3358 {
3359         int is_active = ctx->is_active;
3360         u64 now;
3361 
3362         lockdep_assert_held(&ctx->lock);
3363 
3364         if (likely(!ctx->nr_events))
3365                 return;
3366 
3367         ctx->is_active |= (event_type | EVENT_TIME);
3368         if (ctx->task) {
3369                 if (!is_active)
3370                         cpuctx->task_ctx = ctx;
3371                 else
3372                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
3373         }
3374 
3375         is_active ^= ctx->is_active; /* changed bits */
3376 
3377         if (is_active & EVENT_TIME) {
3378                 /* start ctx time */
3379                 now = perf_clock();
3380                 ctx->timestamp = now;
3381                 perf_cgroup_set_timestamp(task, ctx);
3382         }
3383 
3384         /*
3385          * First go through the list and put on any pinned groups
3386          * in order to give them the best chance of going on.
3387          */
3388         if (is_active & EVENT_PINNED)
3389                 ctx_pinned_sched_in(ctx, cpuctx);
3390 
3391         /* Then walk through the lower prio flexible groups */
3392         if (is_active & EVENT_FLEXIBLE)
3393                 ctx_flexible_sched_in(ctx, cpuctx);
3394 }
3395 
3396 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
3397                              enum event_type_t event_type,
3398                              struct task_struct *task)
3399 {
3400         struct perf_event_context *ctx = &cpuctx->ctx;
3401 
3402         ctx_sched_in(ctx, cpuctx, event_type, task);
3403 }
3404 
3405 static void perf_event_context_sched_in(struct perf_event_context *ctx,
3406                                         struct task_struct *task)
3407 {
3408         struct perf_cpu_context *cpuctx;
3409 
3410         cpuctx = __get_cpu_context(ctx);
3411         if (cpuctx->task_ctx == ctx)
3412                 return;
3413 
3414         perf_ctx_lock(cpuctx, ctx);
3415         /*
3416          * We must check ctx->nr_events while holding ctx->lock, such
3417          * that we serialize against perf_install_in_context().
3418          */
3419         if (!ctx->nr_events)
3420                 goto unlock;
3421 
3422         perf_pmu_disable(ctx->pmu);
3423         /*
3424          * We want to keep the following priority order:
3425          * cpu pinned (that don't need to move), task pinned,
3426          * cpu flexible, task flexible.
3427          *
3428          * However, if task's ctx is not carrying any pinned
3429          * events, no need to flip the cpuctx's events around.
3430          */
3431         if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree))
3432                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3433         perf_event_sched_in(cpuctx, ctx, task);
3434         perf_pmu_enable(ctx->pmu);
3435 
3436 unlock:
3437         perf_ctx_unlock(cpuctx, ctx);
3438 }
3439 
3440 /*
3441  * Called from scheduler to add the events of the current task
3442  * with interrupts disabled.
3443  *
3444  * We restore the event value and then enable it.
3445  *
3446  * This does not protect us against NMI, but enable()
3447  * sets the enabled bit in the control field of event _before_
3448  * accessing the event control register. If a NMI hits, then it will
3449  * keep the event running.
3450  */
3451 void __perf_event_task_sched_in(struct task_struct *prev,
3452                                 struct task_struct *task)
3453 {
3454         struct perf_event_context *ctx;
3455         int ctxn;
3456 
3457         /*
3458          * If cgroup events exist on this CPU, then we need to check if we have
3459          * to switch in PMU state; cgroup event are system-wide mode only.
3460          *
3461          * Since cgroup events are CPU events, we must schedule these in before
3462          * we schedule in the task events.
3463          */
3464         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3465                 perf_cgroup_sched_in(prev, task);
3466 
3467         for_each_task_context_nr(ctxn) {
3468                 ctx = task->perf_event_ctxp[ctxn];
3469                 if (likely(!ctx))
3470                         continue;
3471 
3472                 perf_event_context_sched_in(ctx, task);
3473         }
3474 
3475         if (atomic_read(&nr_switch_events))
3476                 perf_event_switch(task, prev, true);
3477 
3478         if (__this_cpu_read(perf_sched_cb_usages))
3479                 perf_pmu_sched_task(prev, task, true);
3480 }
3481 
3482 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
3483 {
3484         u64 frequency = event->attr.sample_freq;
3485         u64 sec = NSEC_PER_SEC;
3486         u64 divisor, dividend;
3487 
3488         int count_fls, nsec_fls, frequency_fls, sec_fls;
3489 
3490         count_fls = fls64(count);
3491         nsec_fls = fls64(nsec);
3492         frequency_fls = fls64(frequency);
3493         sec_fls = 30;
3494 
3495         /*
3496          * We got @count in @nsec, with a target of sample_freq HZ
3497          * the target period becomes:
3498          *
3499          *             @count * 10^9
3500          * period = -------------------
3501          *          @nsec * sample_freq
3502          *
3503          */
3504 
3505         /*
3506          * Reduce accuracy by one bit such that @a and @b converge
3507          * to a similar magnitude.
3508          */
3509 #define REDUCE_FLS(a, b)                \
3510 do {                                    \
3511         if (a##_fls > b##_fls) {        \
3512                 a >>= 1;                \
3513                 a##_fls--;              \
3514         } else {                        \
3515                 b >>= 1;                \
3516                 b##_fls--;              \
3517         }                               \
3518 } while (0)
3519 
3520         /*
3521          * Reduce accuracy until either term fits in a u64, then proceed with
3522          * the other, so that finally we can do a u64/u64 division.
3523          */
3524         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
3525                 REDUCE_FLS(nsec, frequency);
3526                 REDUCE_FLS(sec, count);
3527         }
3528 
3529         if (count_fls + sec_fls > 64) {
3530                 divisor = nsec * frequency;
3531 
3532                 while (count_fls + sec_fls > 64) {
3533                         REDUCE_FLS(count, sec);
3534                         divisor >>= 1;
3535                 }
3536 
3537                 dividend = count * sec;
3538         } else {
3539                 dividend = count * sec;
3540 
3541                 while (nsec_fls + frequency_fls > 64) {
3542                         REDUCE_FLS(nsec, frequency);
3543                         dividend >>= 1;
3544                 }
3545 
3546                 divisor = nsec * frequency;
3547         }
3548 
3549         if (!divisor)
3550                 return dividend;
3551 
3552         return div64_u64(dividend, divisor);
3553 }
3554 
3555 static DEFINE_PER_CPU(int, perf_throttled_count);
3556 static DEFINE_PER_CPU(u64, perf_throttled_seq);
3557 
3558 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
3559 {
3560         struct hw_perf_event *hwc = &event->hw;
3561         s64 period, sample_period;
3562         s64 delta;
3563 
3564         period = perf_calculate_period(event, nsec, count);
3565 
3566         delta = (s64)(period - hwc->sample_period);
3567         delta = (delta + 7) / 8; /* low pass filter */
3568 
3569         sample_period = hwc->sample_period + delta;
3570 
3571         if (!sample_period)
3572                 sample_period = 1;
3573 
3574         hwc->sample_period = sample_period;
3575 
3576         if (local64_read(&hwc->period_left) > 8*sample_period) {
3577                 if (disable)
3578                         event->pmu->stop(event, PERF_EF_UPDATE);
3579 
3580                 local64_set(&hwc->period_left, 0);
3581 
3582                 if (disable)
3583                         event->pmu->start(event, PERF_EF_RELOAD);
3584         }
3585 }
3586 
3587 /*
3588  * combine freq adjustment with unthrottling to avoid two passes over the
3589  * events. At the same time, make sure, having freq events does not change
3590  * the rate of unthrottling as that would introduce bias.
3591  */
3592 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
3593                                            int needs_unthr)
3594 {
3595         struct perf_event *event;
3596         struct hw_perf_event *hwc;
3597         u64 now, period = TICK_NSEC;
3598         s64 delta;
3599 
3600         /*
3601          * only need to iterate over all events iff:
3602          * - context have events in frequency mode (needs freq adjust)
3603          * - there are events to unthrottle on this cpu
3604          */
3605         if (!(ctx->nr_freq || needs_unthr))
3606                 return;
3607 
3608         raw_spin_lock(&ctx->lock);
3609         perf_pmu_disable(ctx->pmu);
3610 
3611         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3612                 if (event->state != PERF_EVENT_STATE_ACTIVE)
3613                         continue;
3614 
3615                 if (!event_filter_match(event))
3616                         continue;
3617 
3618                 perf_pmu_disable(event->pmu);
3619 
3620                 hwc = &event->hw;
3621 
3622                 if (hwc->interrupts == MAX_INTERRUPTS) {
3623                         hwc->interrupts = 0;
3624                         perf_log_throttle(event, 1);
3625                         event->pmu->start(event, 0);
3626                 }
3627 
3628                 if (!event->attr.freq || !event->attr.sample_freq)
3629                         goto next;
3630 
3631                 /*
3632                  * stop the event and update event->count
3633                  */
3634                 event->pmu->stop(event, PERF_EF_UPDATE);
3635 
3636                 now = local64_read(&event->count);
3637                 delta = now - hwc->freq_count_stamp;
3638                 hwc->freq_count_stamp = now;
3639 
3640                 /*
3641                  * restart the event
3642                  * reload only if value has changed
3643                  * we have stopped the event so tell that
3644                  * to perf_adjust_period() to avoid stopping it
3645                  * twice.
3646                  */
3647                 if (delta > 0)
3648                         perf_adjust_period(event, period, delta, false);
3649 
3650                 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3651         next:
3652                 perf_pmu_enable(event->pmu);
3653         }
3654 
3655         perf_pmu_enable(ctx->pmu);
3656         raw_spin_unlock(&ctx->lock);
3657 }
3658 
3659 /*
3660  * Move @event to the tail of the @ctx's elegible events.
3661  */
3662 static void rotate_ctx(struct perf_event_context *ctx, struct perf_event *event)
3663 {
3664         /*
3665          * Rotate the first entry last of non-pinned groups. Rotation might be
3666          * disabled by the inheritance code.
3667          */
3668         if (ctx->rotate_disable)
3669                 return;
3670 
3671         perf_event_groups_delete(&ctx->flexible_groups, event);
3672         perf_event_groups_insert(&ctx->flexible_groups, event);
3673 }
3674 
3675 static inline struct perf_event *
3676 ctx_first_active(struct perf_event_context *ctx)
3677 {
3678         return list_first_entry_or_null(&ctx->flexible_active,
3679                                         struct perf_event, active_list);
3680 }
3681 
3682 static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
3683 {
3684         struct perf_event *cpu_event = NULL, *task_event = NULL;
3685         bool cpu_rotate = false, task_rotate = false;
3686         struct perf_event_context *ctx = NULL;
3687 
3688         /*
3689          * Since we run this from IRQ context, nobody can install new
3690          * events, thus the event count values are stable.
3691          */
3692 
3693         if (cpuctx->ctx.nr_events) {
3694                 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3695                         cpu_rotate = true;
3696         }
3697 
3698         ctx = cpuctx->task_ctx;
3699         if (ctx && ctx->nr_events) {
3700                 if (ctx->nr_events != ctx->nr_active)
3701                         task_rotate = true;
3702         }
3703 
3704         if (!(cpu_rotate || task_rotate))
3705                 return false;
3706 
3707         perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3708         perf_pmu_disable(cpuctx->ctx.pmu);
3709 
3710         if (task_rotate)
3711                 task_event = ctx_first_active(ctx);
3712         if (cpu_rotate)
3713                 cpu_event = ctx_first_active(&cpuctx->ctx);
3714 
3715         /*
3716          * As per the order given at ctx_resched() first 'pop' task flexible
3717          * and then, if needed CPU flexible.
3718          */
3719         if (task_event || (ctx && cpu_event))
3720                 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3721         if (cpu_event)
3722                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3723 
3724         if (task_event)
3725                 rotate_ctx(ctx, task_event);
3726         if (cpu_event)
3727                 rotate_ctx(&cpuctx->ctx, cpu_event);
3728 
3729         perf_event_sched_in(cpuctx, ctx, current);
3730 
3731         perf_pmu_enable(cpuctx->ctx.pmu);
3732         perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3733 
3734         return true;
3735 }
3736 
3737 void perf_event_task_tick(void)
3738 {
3739         struct list_head *head = this_cpu_ptr(&active_ctx_list);
3740         struct perf_event_context *ctx, *tmp;
3741         int throttled;
3742 
3743         lockdep_assert_irqs_disabled();
3744 
3745         __this_cpu_inc(perf_throttled_seq);
3746         throttled = __this_cpu_xchg(perf_throttled_count, 0);
3747         tick_dep_clear_cpu(smp_processor_id(), TICK_DEP_BIT_PERF_EVENTS);
3748 
3749         list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3750                 perf_adjust_freq_unthr_context(ctx, throttled);
3751 }
3752 
3753 static int event_enable_on_exec(struct perf_event *event,
3754                                 struct perf_event_context *ctx)
3755 {
3756         if (!event->attr.enable_on_exec)
3757                 return 0;
3758 
3759         event->attr.enable_on_exec = 0;
3760         if (event->state >= PERF_EVENT_STATE_INACTIVE)
3761                 return 0;
3762 
3763         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
3764 
3765         return 1;
3766 }
3767 
3768 /*
3769  * Enable all of a task's events that have been marked enable-on-exec.
3770  * This expects task == current.
3771  */
3772 static void perf_event_enable_on_exec(int ctxn)
3773 {
3774         struct perf_event_context *ctx, *clone_ctx = NULL;
3775         enum event_type_t event_type = 0;
3776         struct perf_cpu_context *cpuctx;
3777         struct perf_event *event;
3778         unsigned long flags;
3779         int enabled = 0;
3780 
3781         local_irq_save(flags);
3782         ctx = current->perf_event_ctxp[ctxn];
3783         if (!ctx || !ctx->nr_events)
3784                 goto out;
3785 
3786         cpuctx = __get_cpu_context(ctx);
3787         perf_ctx_lock(cpuctx, ctx);
3788         ctx_sched_out(ctx, cpuctx, EVENT_TIME);
3789         list_for_each_entry(event, &ctx->event_list, event_entry) {
3790                 enabled |= event_enable_on_exec(event, ctx);
3791                 event_type |= get_event_type(event);
3792         }
3793 
3794         /*
3795          * Unclone and reschedule this context if we enabled any event.
3796          */
3797         if (enabled) {
3798                 clone_ctx = unclone_ctx(ctx);
3799                 ctx_resched(cpuctx, ctx, event_type);
3800         } else {
3801                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
3802         }
3803         perf_ctx_unlock(cpuctx, ctx);
3804 
3805 out:
3806         local_irq_restore(flags);
3807 
3808         if (clone_ctx)
3809                 put_ctx(clone_ctx);
3810 }
3811 
3812 struct perf_read_data {
3813         struct perf_event *event;
3814         bool group;
3815         int ret;
3816 };
3817 
3818 static int __perf_event_read_cpu(struct perf_event *event, int event_cpu)
3819 {
3820         u16 local_pkg, event_pkg;
3821 
3822         if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) {
3823                 int local_cpu = smp_processor_id();
3824 
3825                 event_pkg = topology_physical_package_id(event_cpu);
3826                 local_pkg = topology_physical_package_id(local_cpu);
3827 
3828                 if (event_pkg == local_pkg)
3829                         return local_cpu;
3830         }
3831 
3832         return event_cpu;
3833 }
3834 
3835 /*
3836  * Cross CPU call to read the hardware event
3837  */
3838 static void __perf_event_read(void *info)
3839 {
3840         struct perf_read_data *data = info;
3841         struct perf_event *sub, *event = data->event;
3842         struct perf_event_context *ctx = event->ctx;
3843         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3844         struct pmu *pmu = event->pmu;
3845 
3846         /*
3847          * If this is a task context, we need to check whether it is
3848          * the current task context of this cpu.  If not it has been
3849          * scheduled out before the smp call arrived.  In that case
3850          * event->count would have been updated to a recent sample
3851          * when the event was scheduled out.
3852          */
3853         if (ctx->task && cpuctx->task_ctx != ctx)
3854                 return;
3855 
3856         raw_spin_lock(&ctx->lock);
3857         if (ctx->is_active & EVENT_TIME) {
3858                 update_context_time(ctx);
3859                 update_cgrp_time_from_event(event);
3860         }
3861 
3862         perf_event_update_time(event);
3863         if (data->group)
3864                 perf_event_update_sibling_time(event);
3865 
3866         if (event->state != PERF_EVENT_STATE_ACTIVE)
3867                 goto unlock;
3868 
3869         if (!data->group) {
3870                 pmu->read(event);
3871                 data->ret = 0;
3872                 goto unlock;
3873         }
3874 
3875         pmu->start_txn(pmu, PERF_PMU_TXN_READ);
3876 
3877         pmu->read(event);
3878 
3879         for_each_sibling_event(sub, event) {
3880                 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
3881                         /*
3882                          * Use sibling's PMU rather than @event's since
3883                          * sibling could be on different (eg: software) PMU.
3884                          */
3885                         sub->pmu->read(sub);
3886                 }
3887         }
3888 
3889         data->ret = pmu->commit_txn(pmu);
3890 
3891 unlock:
3892         raw_spin_unlock(&ctx->lock);
3893 }
3894 
3895 static inline u64 perf_event_count(struct perf_event *event)
3896 {
3897         return local64_read(&event->count) + atomic64_read(&event->child_count);
3898 }
3899 
3900 /*
3901  * NMI-safe method to read a local event, that is an event that
3902  * is:
3903  *   - either for the current task, or for this CPU
3904  *   - does not have inherit set, for inherited task events
3905  *     will not be local and we cannot read them atomically
3906  *   - must not have a pmu::count method
3907  */
3908 int perf_event_read_local(struct perf_event *event, u64 *value,
3909                           u64 *enabled, u64 *running)
3910 {
3911         unsigned long flags;
3912         int ret = 0;
3913 
3914         /*
3915          * Disabling interrupts avoids all counter scheduling (context
3916          * switches, timer based rotation and IPIs).
3917          */
3918         local_irq_save(flags);
3919 
3920         /*
3921          * It must not be an event with inherit set, we cannot read
3922          * all child counters from atomic context.
3923          */
3924         if (event->attr.inherit) {
3925                 ret = -EOPNOTSUPP;
3926                 goto out;
3927         }
3928 
3929         /* If this is a per-task event, it must be for current */
3930         if ((event->attach_state & PERF_ATTACH_TASK) &&
3931             event->hw.target != current) {
3932                 ret = -EINVAL;
3933                 goto out;
3934         }
3935 
3936         /* If this is a per-CPU event, it must be for this CPU */
3937         if (!(event->attach_state & PERF_ATTACH_TASK) &&
3938             event->cpu != smp_processor_id()) {
3939                 ret = -EINVAL;
3940                 goto out;
3941         }
3942 
3943         /*
3944          * If the event is currently on this CPU, its either a per-task event,
3945          * or local to this CPU. Furthermore it means its ACTIVE (otherwise
3946          * oncpu == -1).
3947          */
3948         if (event->oncpu == smp_processor_id())
3949                 event->pmu->read(event);
3950 
3951         *value = local64_read(&event->count);
3952         if (enabled || running) {
3953                 u64 now = event->shadow_ctx_time + perf_clock();
3954                 u64 __enabled, __running;
3955 
3956                 __perf_update_times(event, now, &__enabled, &__running);
3957                 if (enabled)
3958                         *enabled = __enabled;
3959                 if (running)
3960                         *running = __running;
3961         }
3962 out:
3963         local_irq_restore(flags);
3964 
3965         return ret;
3966 }
3967 
3968 static int perf_event_read(struct perf_event *event, bool group)
3969 {
3970         enum perf_event_state state = READ_ONCE(event->state);
3971         int event_cpu, ret = 0;
3972 
3973         /*
3974          * If event is enabled and currently active on a CPU, update the
3975          * value in the event structure:
3976          */
3977 again:
3978         if (state == PERF_EVENT_STATE_ACTIVE) {
3979                 struct perf_read_data data;
3980 
3981                 /*
3982                  * Orders the ->state and ->oncpu loads such that if we see
3983                  * ACTIVE we must also see the right ->oncpu.
3984                  *
3985                  * Matches the smp_wmb() from event_sched_in().
3986                  */
3987                 smp_rmb();
3988 
3989                 event_cpu = READ_ONCE(event->oncpu);
3990                 if ((unsigned)event_cpu >= nr_cpu_ids)
3991                         return 0;
3992 
3993                 data = (struct perf_read_data){
3994                         .event = event,
3995                         .group = group,
3996                         .ret = 0,
3997                 };
3998 
3999                 preempt_disable();
4000                 event_cpu = __perf_event_read_cpu(event, event_cpu);
4001 
4002                 /*
4003                  * Purposely ignore the smp_call_function_single() return
4004                  * value.
4005                  *
4006                  * If event_cpu isn't a valid CPU it means the event got
4007                  * scheduled out and that will have updated the event count.
4008                  *
4009                  * Therefore, either way, we'll have an up-to-date event count
4010                  * after this.
4011                  */
4012                 (void)smp_call_function_single(event_cpu, __perf_event_read, &data, 1);
4013                 preempt_enable();
4014                 ret = data.ret;
4015 
4016         } else if (state == PERF_EVENT_STATE_INACTIVE) {
4017                 struct perf_event_context *ctx = event->ctx;
4018                 unsigned long flags;
4019 
4020                 raw_spin_lock_irqsave(&ctx->lock, flags);
4021                 state = event->state;
4022                 if (state != PERF_EVENT_STATE_INACTIVE) {
4023                         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4024                         goto again;
4025                 }
4026 
4027                 /*
4028                  * May read while context is not active (e.g., thread is
4029                  * blocked), in that case we cannot update context time
4030                  */
4031                 if (ctx->is_active & EVENT_TIME) {
4032                         update_context_time(ctx);
4033                         update_cgrp_time_from_event(event);
4034                 }
4035 
4036                 perf_event_update_time(event);
4037                 if (group)
4038                         perf_event_update_sibling_time(event);
4039                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4040         }
4041 
4042         return ret;
4043 }
4044 
4045 /*
4046  * Initialize the perf_event context in a task_struct:
4047  */
4048 static void __perf_event_init_context(struct perf_event_context *ctx)
4049 {
4050         raw_spin_lock_init(&ctx->lock);
4051         mutex_init(&ctx->mutex);
4052         INIT_LIST_HEAD(&ctx->active_ctx_list);
4053         perf_event_groups_init(&ctx->pinned_groups);
4054         perf_event_groups_init(&ctx->flexible_groups);
4055         INIT_LIST_HEAD(&ctx->event_list);
4056         INIT_LIST_HEAD(&ctx->pinned_active);
4057         INIT_LIST_HEAD(&ctx->flexible_active);
4058         atomic_set(&ctx->refcount, 1);
4059 }
4060 
4061 static struct perf_event_context *
4062 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
4063 {
4064         struct perf_event_context *ctx;
4065 
4066         ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4067         if (!ctx)
4068                 return NULL;
4069 
4070         __perf_event_init_context(ctx);
4071         if (task) {
4072                 ctx->task = task;
4073                 get_task_struct(task);
4074         }
4075         ctx->pmu = pmu;
4076 
4077         return ctx;
4078 }
4079 
4080 static struct task_struct *
4081 find_lively_task_by_vpid(pid_t vpid)
4082 {
4083         struct task_struct *task;
4084 
4085         rcu_read_lock();
4086         if (!vpid)
4087                 task = current;
4088         else
4089                 task = find_task_by_vpid(vpid);
4090         if (task)
4091                 get_task_struct(task);
4092         rcu_read_unlock();
4093 
4094         if (!task)
4095                 return ERR_PTR(-ESRCH);
4096 
4097         return task;
4098 }
4099 
4100 /*
4101  * Returns a matching context with refcount and pincount.
4102  */
4103 static struct perf_event_context *
4104 find_get_context(struct pmu *pmu, struct task_struct *task,
4105                 struct perf_event *event)
4106 {
4107         struct perf_event_context *ctx, *clone_ctx = NULL;
4108         struct perf_cpu_context *cpuctx;
4109         void *task_ctx_data = NULL;
4110         unsigned long flags;
4111         int ctxn, err;
4112         int cpu = event->cpu;
4113 
4114         if (!task) {
4115                 /* Must be root to operate on a CPU event: */
4116                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
4117                         return ERR_PTR(-EACCES);
4118 
4119                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
4120                 ctx = &cpuctx->ctx;
4121                 get_ctx(ctx);
4122                 ++ctx->pin_count;
4123 
4124                 return ctx;
4125         }
4126 
4127         err = -EINVAL;
4128         ctxn = pmu->task_ctx_nr;
4129         if (ctxn < 0)
4130                 goto errout;
4131 
4132         if (event->attach_state & PERF_ATTACH_TASK_DATA) {
4133                 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
4134                 if (!task_ctx_data) {
4135                         err = -ENOMEM;
4136                         goto errout;
4137                 }
4138         }
4139 
4140 retry:
4141         ctx = perf_lock_task_context(task, ctxn, &flags);
4142         if (ctx) {
4143                 clone_ctx = unclone_ctx(ctx);
4144                 ++ctx->pin_count;
4145 
4146                 if (task_ctx_data && !ctx->task_ctx_data) {
4147                         ctx->task_ctx_data = task_ctx_data;
4148                         task_ctx_data = NULL;
4149                 }
4150                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4151 
4152                 if (clone_ctx)
4153                         put_ctx(clone_ctx);
4154         } else {
4155                 ctx = alloc_perf_context(pmu, task);
4156                 err = -ENOMEM;
4157                 if (!ctx)
4158                         goto errout;
4159 
4160                 if (task_ctx_data) {
4161                         ctx->task_ctx_data = task_ctx_data;
4162                         task_ctx_data = NULL;
4163                 }
4164 
4165                 err = 0;
4166                 mutex_lock(&task->perf_event_mutex);
4167                 /*
4168                  * If it has already passed perf_event_exit_task().
4169                  * we must see PF_EXITING, it takes this mutex too.
4170                  */
4171                 if (task->flags & PF_EXITING)
4172                         err = -ESRCH;
4173                 else if (task->perf_event_ctxp[ctxn])
4174                         err = -EAGAIN;
4175                 else {
4176                         get_ctx(ctx);
4177                         ++ctx->pin_count;
4178                         rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
4179                 }
4180                 mutex_unlock(&task->perf_event_mutex);
4181 
4182                 if (unlikely(err)) {
4183                         put_ctx(ctx);
4184 
4185                         if (err == -EAGAIN)
4186                                 goto retry;
4187                         goto errout;
4188                 }
4189         }
4190 
4191         kfree(task_ctx_data);
4192         return ctx;
4193 
4194 errout:
4195         kfree(task_ctx_data);
4196         return ERR_PTR(err);
4197 }
4198 
4199 static void perf_event_free_filter(struct perf_event *event);
4200 static void perf_event_free_bpf_prog(struct perf_event *event);
4201 
4202 static void free_event_rcu(struct rcu_head *head)
4203 {
4204         struct perf_event *event;
4205 
4206         event = container_of(head, struct perf_event, rcu_head);
4207         if (event->ns)
4208                 put_pid_ns(event->ns);
4209         perf_event_free_filter(event);
4210         kfree(event);
4211 }
4212 
4213 static void ring_buffer_attach(struct perf_event *event,
4214                                struct ring_buffer *rb);
4215 
4216 static void detach_sb_event(struct perf_event *event)
4217 {
4218         struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu);
4219 
4220         raw_spin_lock(&pel->lock);
4221         list_del_rcu(&event->sb_list);
4222         raw_spin_unlock(&pel->lock);
4223 }
4224 
4225 static bool is_sb_event(struct perf_event *event)
4226 {
4227         struct perf_event_attr *attr = &event->attr;
4228 
4229         if (event->parent)
4230                 return false;
4231 
4232         if (event->attach_state & PERF_ATTACH_TASK)
4233                 return false;
4234 
4235         if (attr->mmap || attr->mmap_data || attr->mmap2 ||
4236             attr->comm || attr->comm_exec ||
4237             attr->task ||
4238             attr->context_switch)
4239                 return true;
4240         return false;
4241 }
4242 
4243 static void unaccount_pmu_sb_event(struct perf_event *event)
4244 {
4245         if (is_sb_event(event))
4246                 detach_sb_event(event);
4247 }
4248 
4249 static void unaccount_event_cpu(struct perf_event *event, int cpu)
4250 {
4251         if (event->parent)
4252                 return;
4253 
4254         if (is_cgroup_event(event))
4255                 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
4256 }
4257 
4258 #ifdef CONFIG_NO_HZ_FULL
4259 static DEFINE_SPINLOCK(nr_freq_lock);
4260 #endif
4261 
4262 static void unaccount_freq_event_nohz(void)
4263 {
4264 #ifdef CONFIG_NO_HZ_FULL
4265         spin_lock(&nr_freq_lock);
4266         if (atomic_dec_and_test(&nr_freq_events))
4267                 tick_nohz_dep_clear(TICK_DEP_BIT_PERF_EVENTS);
4268         spin_unlock(&nr_freq_lock);
4269 #endif
4270 }
4271 
4272 static void unaccount_freq_event(void)
4273 {
4274         if (tick_nohz_full_enabled())
4275                 unaccount_freq_event_nohz();
4276         else
4277                 atomic_dec(&nr_freq_events);
4278 }
4279 
4280 static void unaccount_event(struct perf_event *event)
4281 {
4282         bool dec = false;
4283 
4284         if (event->parent)
4285                 return;
4286 
4287         if (event->attach_state & PERF_ATTACH_TASK)
4288                 dec = true;
4289         if (event->attr.mmap || event->attr.mmap_data)
4290                 atomic_dec(&nr_mmap_events);
4291         if (event->attr.comm)
4292                 atomic_dec(&nr_comm_events);
4293         if (event->attr.namespaces)
4294                 atomic_dec(&nr_namespaces_events);
4295         if (event->attr.task)
4296                 atomic_dec(&nr_task_events);
4297         if (event->attr.freq)
4298                 unaccount_freq_event();
4299         if (event->attr.context_switch) {
4300                 dec = true;
4301                 atomic_dec(&nr_switch_events);
4302         }
4303         if (is_cgroup_event(event))
4304                 dec = true;
4305         if (has_branch_stack(event))
4306                 dec = true;
4307 
4308         if (dec) {
4309                 if (!atomic_add_unless(&perf_sched_count, -1, 1))
4310                         schedule_delayed_work(&perf_sched_work, HZ);
4311         }
4312 
4313         unaccount_event_cpu(event, event->cpu);
4314 
4315         unaccount_pmu_sb_event(event);
4316 }
4317 
4318 static void perf_sched_delayed(struct work_struct *work)
4319 {
4320         mutex_lock(&perf_sched_mutex);
4321         if (atomic_dec_and_test(&perf_sched_count))
4322                 static_branch_disable(&perf_sched_events);
4323         mutex_unlock(&perf_sched_mutex);
4324 }
4325 
4326 /*
4327  * The following implement mutual exclusion of events on "exclusive" pmus
4328  * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
4329  * at a time, so we disallow creating events that might conflict, namely:
4330  *
4331  *  1) cpu-wide events in the presence of per-task events,
4332  *  2) per-task events in the presence of cpu-wide events,
4333  *  3) two matching events on the same context.
4334  *
4335  * The former two cases are handled in the allocation path (perf_event_alloc(),
4336  * _free_event()), the latter -- before the first perf_install_in_context().
4337  */
4338 static int exclusive_event_init(struct perf_event *event)
4339 {
4340         struct pmu *pmu = event->pmu;
4341 
4342         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4343                 return 0;
4344 
4345         /*
4346          * Prevent co-existence of per-task and cpu-wide events on the
4347          * same exclusive pmu.
4348          *
4349          * Negative pmu::exclusive_cnt means there are cpu-wide
4350          * events on this "exclusive" pmu, positive means there are
4351          * per-task events.
4352          *
4353          * Since this is called in perf_event_alloc() path, event::ctx
4354          * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
4355          * to mean "per-task event", because unlike other attach states it
4356          * never gets cleared.
4357          */
4358         if (event->attach_state & PERF_ATTACH_TASK) {
4359                 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
4360                         return -EBUSY;
4361         } else {
4362                 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
4363                         return -EBUSY;
4364         }
4365 
4366         return 0;
4367 }
4368 
4369 static void exclusive_event_destroy(struct perf_event *event)
4370 {
4371         struct pmu *pmu = event->pmu;
4372 
4373         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4374                 return;
4375 
4376         /* see comment in exclusive_event_init() */
4377         if (event->attach_state & PERF_ATTACH_TASK)
4378                 atomic_dec(&pmu->exclusive_cnt);
4379         else
4380                 atomic_inc(&pmu->exclusive_cnt);
4381 }
4382 
4383 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
4384 {
4385         if ((e1->pmu == e2->pmu) &&
4386             (e1->cpu == e2->cpu ||
4387              e1->cpu == -1 ||
4388              e2->cpu == -1))
4389                 return true;
4390         return false;
4391 }
4392 
4393 /* Called under the same ctx::mutex as perf_install_in_context() */
4394 static bool exclusive_event_installable(struct perf_event *event,
4395                                         struct perf_event_context *ctx)
4396 {
4397         struct perf_event *iter_event;
4398         struct pmu *pmu = event->pmu;
4399 
4400         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4401                 return true;
4402 
4403         list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
4404                 if (exclusive_event_match(iter_event, event))
4405                         return false;
4406         }
4407 
4408         return true;
4409 }
4410 
4411 static void perf_addr_filters_splice(struct perf_event *event,
4412                                        struct list_head *head);
4413 
4414 static void _free_event(struct perf_event *event)
4415 {
4416         irq_work_sync(&event->pending);
4417 
4418         unaccount_event(event);
4419 
4420         if (event->rb) {
4421                 /*
4422                  * Can happen when we close an event with re-directed output.
4423                  *
4424                  * Since we have a 0 refcount, perf_mmap_close() will skip
4425                  * over us; possibly making our ring_buffer_put() the last.
4426                  */
4427                 mutex_lock(&event->mmap_mutex);
4428                 ring_buffer_attach(event, NULL);
4429                 mutex_unlock(&event->mmap_mutex);
4430         }
4431 
4432         if (is_cgroup_event(event))
4433                 perf_detach_cgroup(event);
4434 
4435         if (!event->parent) {
4436                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
4437                         put_callchain_buffers();
4438         }
4439 
4440         perf_event_free_bpf_prog(event);
4441         perf_addr_filters_splice(event, NULL);
4442         kfree(event->addr_filters_offs);
4443 
4444         if (event->destroy)
4445                 event->destroy(event);
4446 
4447         if (event->ctx)
4448                 put_ctx(event->ctx);
4449 
4450         if (event->hw.target)
4451                 put_task_struct(event->hw.target);
4452 
4453         exclusive_event_destroy(event);
4454         module_put(event->pmu->module);
4455 
4456         call_rcu(&event->rcu_head, free_event_rcu);
4457 }
4458 
4459 /*
4460  * Used to free events which have a known refcount of 1, such as in error paths
4461  * where the event isn't exposed yet and inherited events.
4462  */
4463 static void free_event(struct perf_event *event)
4464 {
4465         if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
4466                                 "unexpected event refcount: %ld; ptr=%p\n",
4467                                 atomic_long_read(&event->refcount), event)) {
4468                 /* leak to avoid use-after-free */
4469                 return;
4470         }
4471 
4472         _free_event(event);
4473 }
4474 
4475 /*
4476  * Remove user event from the owner task.
4477  */
4478 static void perf_remove_from_owner(struct perf_event *event)
4479 {
4480         struct task_struct *owner;
4481 
4482         rcu_read_lock();
4483         /*
4484          * Matches the smp_store_release() in perf_event_exit_task(). If we
4485          * observe !owner it means the list deletion is complete and we can
4486          * indeed free this event, otherwise we need to serialize on
4487          * owner->perf_event_mutex.
4488          */
4489         owner = READ_ONCE(event->owner);
4490         if (owner) {
4491                 /*
4492                  * Since delayed_put_task_struct() also drops the last
4493                  * task reference we can safely take a new reference
4494                  * while holding the rcu_read_lock().
4495                  */
4496                 get_task_struct(owner);
4497         }
4498         rcu_read_unlock();
4499 
4500         if (owner) {
4501                 /*
4502                  * If we're here through perf_event_exit_task() we're already
4503                  * holding ctx->mutex which would be an inversion wrt. the
4504                  * normal lock order.
4505                  *
4506                  * However we can safely take this lock because its the child
4507                  * ctx->mutex.
4508                  */
4509                 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
4510 
4511                 /*
4512                  * We have to re-check the event->owner field, if it is cleared
4513                  * we raced with perf_event_exit_task(), acquiring the mutex
4514                  * ensured they're done, and we can proceed with freeing the
4515                  * event.
4516                  */
4517                 if (event->owner) {
4518                         list_del_init(&event->owner_entry);
4519                         smp_store_release(&event->owner, NULL);
4520                 }
4521                 mutex_unlock(&owner->perf_event_mutex);
4522                 put_task_struct(owner);
4523         }
4524 }
4525 
4526 static void put_event(struct perf_event *event)
4527 {
4528         if (!atomic_long_dec_and_test(&event->refcount))
4529                 return;
4530 
4531         _free_event(event);
4532 }
4533 
4534 /*
4535  * Kill an event dead; while event:refcount will preserve the event
4536  * object, it will not preserve its functionality. Once the last 'user'
4537  * gives up the object, we'll destroy the thing.
4538  */
4539 int perf_event_release_kernel(struct perf_event *event)
4540 {
4541         struct perf_event_context *ctx = event->ctx;
4542         struct perf_event *child, *tmp;
4543         LIST_HEAD(free_list);
4544 
4545         /*
4546          * If we got here through err_file: fput(event_file); we will not have
4547          * attached to a context yet.
4548          */
4549         if (!ctx) {
4550                 WARN_ON_ONCE(event->attach_state &
4551                                 (PERF_ATTACH_CONTEXT|PERF_ATTACH_GROUP));
4552                 goto no_ctx;
4553         }
4554 
4555         if (!is_kernel_event(event))
4556                 perf_remove_from_owner(event);
4557 
4558         ctx = perf_event_ctx_lock(event);
4559         WARN_ON_ONCE(ctx->parent_ctx);
4560         perf_remove_from_context(event, DETACH_GROUP);
4561 
4562         raw_spin_lock_irq(&ctx->lock);
4563         /*
4564          * Mark this event as STATE_DEAD, there is no external reference to it
4565          * anymore.
4566          *
4567          * Anybody acquiring event->child_mutex after the below loop _must_
4568          * also see this, most importantly inherit_event() which will avoid
4569          * placing more children on the list.
4570          *
4571          * Thus this guarantees that we will in fact observe and kill _ALL_
4572          * child events.
4573          */
4574         event->state = PERF_EVENT_STATE_DEAD;
4575         raw_spin_unlock_irq(&ctx->lock);
4576 
4577         perf_event_ctx_unlock(event, ctx);
4578 
4579 again:
4580         mutex_lock(&event->child_mutex);
4581         list_for_each_entry(child, &event->child_list, child_list) {
4582 
4583                 /*
4584                  * Cannot change, child events are not migrated, see the
4585                  * comment with perf_event_ctx_lock_nested().
4586                  */
4587                 ctx = READ_ONCE(child->ctx);
4588                 /*
4589                  * Since child_mutex nests inside ctx::mutex, we must jump
4590                  * through hoops. We start by grabbing a reference on the ctx.
4591                  *
4592                  * Since the event cannot get freed while we hold the
4593                  * child_mutex, the context must also exist and have a !0
4594                  * reference count.
4595                  */
4596                 get_ctx(ctx);
4597 
4598                 /*
4599                  * Now that we have a ctx ref, we can drop child_mutex, and
4600                  * acquire ctx::mutex without fear of it going away. Then we
4601                  * can re-acquire child_mutex.
4602                  */
4603                 mutex_unlock(&event->child_mutex);
4604                 mutex_lock(&ctx->mutex);
4605                 mutex_lock(&event->child_mutex);
4606 
4607                 /*
4608                  * Now that we hold ctx::mutex and child_mutex, revalidate our
4609                  * state, if child is still the first entry, it didn't get freed
4610                  * and we can continue doing so.
4611                  */
4612                 tmp = list_first_entry_or_null(&event->child_list,
4613                                                struct perf_event, child_list);
4614                 if (tmp == child) {
4615                         perf_remove_from_context(child, DETACH_GROUP);
4616                         list_move(&child->child_list, &free_list);
4617                         /*
4618                          * This matches the refcount bump in inherit_event();
4619                          * this can't be the last reference.
4620                          */
4621                         put_event(event);
4622                 }
4623 
4624                 mutex_unlock(&event->child_mutex);
4625                 mutex_unlock(&ctx->mutex);
4626                 put_ctx(ctx);
4627                 goto again;
4628         }
4629         mutex_unlock(&event->child_mutex);
4630 
4631         list_for_each_entry_safe(child, tmp, &free_list, child_list) {
4632                 list_del(&child->child_list);
4633                 free_event(child);
4634         }
4635 
4636 no_ctx:
4637         put_event(event); /* Must be the 'last' reference */
4638         return 0;
4639 }
4640 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
4641 
4642 /*
4643  * Called when the last reference to the file is gone.
4644  */
4645 static int perf_release(struct inode *inode, struct file *file)
4646 {
4647         perf_event_release_kernel(file->private_data);
4648         return 0;
4649 }
4650 
4651 static u64 __perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4652 {
4653         struct perf_event *child;
4654         u64 total = 0;
4655 
4656         *enabled = 0;
4657         *running = 0;
4658 
4659         mutex_lock(&event->child_mutex);
4660 
4661         (void)perf_event_read(event, false);
4662         total += perf_event_count(event);
4663 
4664         *enabled += event->total_time_enabled +
4665                         atomic64_read(&event->child_total_time_enabled);
4666         *running += event->total_time_running +
4667                         atomic64_read(&event->child_total_time_running);
4668 
4669         list_for_each_entry(child, &event->child_list, child_list) {
4670                 (void)perf_event_read(child, false);
4671                 total += perf_event_count(child);
4672                 *enabled += child->total_time_enabled;
4673                 *running += child->total_time_running;
4674         }
4675         mutex_unlock(&event->child_mutex);
4676 
4677         return total;
4678 }
4679 
4680 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4681 {
4682         struct perf_event_context *ctx;
4683         u64 count;
4684 
4685         ctx = perf_event_ctx_lock(event);
4686         count = __perf_event_read_value(event, enabled, running);
4687         perf_event_ctx_unlock(event, ctx);
4688 
4689         return count;
4690 }
4691 EXPORT_SYMBOL_GPL(perf_event_read_value);
4692 
4693 static int __perf_read_group_add(struct perf_event *leader,
4694                                         u64 read_format, u64 *values)
4695 {
4696         struct perf_event_context *ctx = leader->ctx;
4697         struct perf_event *sub;
4698         unsigned long flags;
4699         int n = 1; /* skip @nr */
4700         int ret;
4701 
4702         ret = perf_event_read(leader, true);
4703         if (ret)
4704                 return ret;
4705 
4706         raw_spin_lock_irqsave(&ctx->lock, flags);
4707 
4708         /*
4709          * Since we co-schedule groups, {enabled,running} times of siblings
4710          * will be identical to those of the leader, so we only publish one
4711          * set.
4712          */
4713         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4714                 values[n++] += leader->total_time_enabled +
4715                         atomic64_read(&leader->child_total_time_enabled);
4716         }
4717 
4718         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4719                 values[n++] += leader->total_time_running +
4720                         atomic64_read(&leader->child_total_time_running);
4721         }
4722 
4723         /*
4724          * Write {count,id} tuples for every sibling.
4725          */
4726         values[n++] += perf_event_count(leader);
4727         if (read_format & PERF_FORMAT_ID)
4728                 values[n++] = primary_event_id(leader);
4729 
4730         for_each_sibling_event(sub, leader) {
4731                 values[n++] += perf_event_count(sub);
4732                 if (read_format & PERF_FORMAT_ID)
4733                         values[n++] = primary_event_id(sub);
4734         }
4735 
4736         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4737         return 0;
4738 }
4739 
4740 static int perf_read_group(struct perf_event *event,
4741                                    u64 read_format, char __user *buf)
4742 {
4743         struct perf_event *leader = event->group_leader, *child;
4744         struct perf_event_context *ctx = leader->ctx;
4745         int ret;
4746         u64 *values;
4747 
4748         lockdep_assert_held(&ctx->mutex);
4749 
4750         values = kzalloc(event->read_size, GFP_KERNEL);
4751         if (!values)
4752                 return -ENOMEM;
4753 
4754         values[0] = 1 + leader->nr_siblings;
4755 
4756         /*
4757          * By locking the child_mutex of the leader we effectively
4758          * lock the child list of all siblings.. XXX explain how.
4759          */
4760         mutex_lock(&leader->child_mutex);
4761 
4762         ret = __perf_read_group_add(leader, read_format, values);
4763         if (ret)
4764                 goto unlock;
4765 
4766         list_for_each_entry(child, &leader->child_list, child_list) {
4767                 ret = __perf_read_group_add(child, read_format, values);
4768                 if (ret)
4769                         goto unlock;
4770         }
4771 
4772         mutex_unlock(&leader->child_mutex);
4773 
4774         ret = event->read_size;
4775         if (copy_to_user(buf, values, event->read_size))
4776                 ret = -EFAULT;
4777         goto out;
4778 
4779 unlock:
4780         mutex_unlock(&leader->child_mutex);
4781 out:
4782         kfree(values);
4783         return ret;
4784 }
4785 
4786 static int perf_read_one(struct perf_event *event,
4787                                  u64 read_format, char __user *buf)
4788 {
4789         u64 enabled, running;
4790         u64 values[4];
4791         int n = 0;
4792 
4793         values[n++] = __perf_event_read_value(event, &enabled, &running);
4794         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4795                 values[n++] = enabled;
4796         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4797                 values[n++] = running;
4798         if (read_format & PERF_FORMAT_ID)
4799                 values[n++] = primary_event_id(event);
4800 
4801         if (copy_to_user(buf, values, n * sizeof(u64)))
4802                 return -EFAULT;
4803 
4804         return n * sizeof(u64);
4805 }
4806 
4807 static bool is_event_hup(struct perf_event *event)
4808 {
4809         bool no_children;
4810 
4811         if (event->state > PERF_EVENT_STATE_EXIT)
4812                 return false;
4813 
4814         mutex_lock(&event->child_mutex);
4815         no_children = list_empty(&event->child_list);
4816         mutex_unlock(&event->child_mutex);
4817         return no_children;
4818 }
4819 
4820 /*
4821  * Read the performance event - simple non blocking version for now
4822  */
4823 static ssize_t
4824 __perf_read(struct perf_event *event, char __user *buf, size_t count)
4825 {
4826         u64 read_format = event->attr.read_format;
4827         int ret;
4828 
4829         /*
4830          * Return end-of-file for a read on a event that is in
4831          * error state (i.e. because it was pinned but it couldn't be
4832          * scheduled on to the CPU at some point).
4833          */
4834         if (event->state == PERF_EVENT_STATE_ERROR)
4835                 return 0;
4836 
4837         if (count < event->read_size)
4838                 return -ENOSPC;
4839 
4840         WARN_ON_ONCE(event->ctx->parent_ctx);
4841         if (read_format & PERF_FORMAT_GROUP)
4842                 ret = perf_read_group(event, read_format, buf);
4843         else
4844                 ret = perf_read_one(event, read_format, buf);
4845 
4846         return ret;
4847 }
4848 
4849 static ssize_t
4850 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
4851 {
4852         struct perf_event *event = file->private_data;
4853         struct perf_event_context *ctx;
4854         int ret;
4855 
4856         ctx = perf_event_ctx_lock(event);
4857         ret = __perf_read(event, buf, count);
4858         perf_event_ctx_unlock(event, ctx);
4859 
4860         return ret;
4861 }
4862 
4863 static __poll_t perf_poll(struct file *file, poll_table *wait)
4864 {
4865         struct perf_event *event = file->private_data;
4866         struct ring_buffer *rb;
4867         __poll_t events = EPOLLHUP;
4868 
4869         poll_wait(file, &event->waitq, wait);
4870 
4871         if (is_event_hup(event))
4872                 return events;
4873 
4874         /*
4875          * Pin the event->rb by taking event->mmap_mutex; otherwise
4876          * perf_event_set_output() can swizzle our rb and make us miss wakeups.
4877          */
4878         mutex_lock(&event->mmap_mutex);
4879         rb = event->rb;
4880         if (rb)
4881                 events = atomic_xchg(&rb->poll, 0);
4882         mutex_unlock(&event->mmap_mutex);
4883         return events;
4884 }
4885 
4886 static void _perf_event_reset(struct perf_event *event)
4887 {
4888         (void)perf_event_read(event, false);
4889         local64_set(&event->count, 0);
4890         perf_event_update_userpage(event);
4891 }
4892 
4893 /*
4894  * Holding the top-level event's child_mutex means that any
4895  * descendant process that has inherited this event will block
4896  * in perf_event_exit_event() if it goes to exit, thus satisfying the
4897  * task existence requirements of perf_event_enable/disable.
4898  */
4899 static void perf_event_for_each_child(struct perf_event *event,
4900                                         void (*func)(struct perf_event *))
4901 {
4902         struct perf_event *child;
4903 
4904         WARN_ON_ONCE(event->ctx->parent_ctx);
4905 
4906         mutex_lock(&event->child_mutex);
4907         func(event);
4908         list_for_each_entry(child, &event->child_list, child_list)
4909                 func(child);
4910         mutex_unlock(&event->child_mutex);
4911 }
4912 
4913 static void perf_event_for_each(struct perf_event *event,
4914                                   void (*func)(struct perf_event *))
4915 {
4916         struct perf_event_context *ctx = event->ctx;
4917         struct perf_event *sibling;
4918 
4919         lockdep_assert_held(&ctx->mutex);
4920 
4921         event = event->group_leader;
4922 
4923         perf_event_for_each_child(event, func);
4924         for_each_sibling_event(sibling, event)
4925                 perf_event_for_each_child(sibling, func);
4926 }
4927 
4928 static void __perf_event_period(struct perf_event *event,
4929                                 struct perf_cpu_context *cpuctx,
4930                                 struct perf_event_context *ctx,
4931                                 void *info)
4932 {
4933         u64 value = *((u64 *)info);
4934         bool active;
4935 
4936         if (event->attr.freq) {
4937                 event->attr.sample_freq = value;
4938         } else {
4939                 event->attr.sample_period = value;
4940                 event->hw.sample_period = value;
4941         }
4942 
4943         active = (event->state == PERF_EVENT_STATE_ACTIVE);
4944         if (active) {
4945                 perf_pmu_disable(ctx->pmu);
4946                 /*
4947                  * We could be throttled; unthrottle now to avoid the tick
4948                  * trying to unthrottle while we already re-started the event.
4949                  */
4950                 if (event->hw.interrupts == MAX_INTERRUPTS) {
4951                         event->hw.interrupts = 0;
4952                         perf_log_throttle(event, 1);
4953                 }
4954                 event->pmu->stop(event, PERF_EF_UPDATE);
4955         }
4956 
4957         local64_set(&event->hw.period_left, 0);
4958 
4959         if (active) {
4960                 event->pmu->start(event, PERF_EF_RELOAD);
4961                 perf_pmu_enable(ctx->pmu);
4962         }
4963 }
4964 
4965 static int perf_event_period(struct perf_event *event, u64 __user *arg)
4966 {
4967         u64 value;
4968 
4969         if (!is_sampling_event(event))
4970                 return -EINVAL;
4971 
4972         if (copy_from_user(&value, arg, sizeof(value)))
4973                 return -EFAULT;
4974 
4975         if (!value)
4976                 return -EINVAL;
4977 
4978         if (event->attr.freq && value > sysctl_perf_event_sample_rate)
4979                 return -EINVAL;
4980 
4981         event_function_call(event, __perf_event_period, &value);
4982 
4983         return 0;
4984 }
4985 
4986 static const struct file_operations perf_fops;
4987 
4988 static inline int perf_fget_light(int fd, struct fd *p)
4989 {
4990         struct fd f = fdget(fd);
4991         if (!f.file)
4992                 return -EBADF;
4993 
4994         if (f.file->f_op != &perf_fops) {
4995                 fdput(f);
4996                 return -EBADF;
4997         }
4998         *p = f;
4999         return 0;
5000 }
5001 
5002 static int perf_event_set_output(struct perf_event *event,
5003                                  struct perf_event *output_event);
5004 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
5005 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
5006 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5007                           struct perf_event_attr *attr);
5008 
5009 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
5010 {
5011         void (*func)(struct perf_event *);
5012         u32 flags = arg;
5013 
5014         switch (cmd) {
5015         case PERF_EVENT_IOC_ENABLE:
5016                 func = _perf_event_enable;
5017                 break;
5018         case PERF_EVENT_IOC_DISABLE:
5019                 func = _perf_event_disable;
5020                 break;
5021         case PERF_EVENT_IOC_RESET:
5022                 func = _perf_event_reset;
5023                 break;
5024 
5025         case PERF_EVENT_IOC_REFRESH:
5026                 return _perf_event_refresh(event, arg);
5027 
5028         case PERF_EVENT_IOC_PERIOD:
5029                 return perf_event_period(event, (u64 __user *)arg);
5030 
5031         case PERF_EVENT_IOC_ID:
5032         {
5033                 u64 id = primary_event_id(event);
5034 
5035                 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
5036                         return -EFAULT;
5037                 return 0;
5038         }
5039 
5040         case PERF_EVENT_IOC_SET_OUTPUT:
5041         {
5042                 int ret;
5043                 if (arg != -1) {
5044                         struct perf_event *output_event;
5045                         struct fd output;
5046                         ret = perf_fget_light(arg, &output);
5047                         if (ret)
5048                                 return ret;
5049                         output_event = output.file->private_data;
5050                         ret = perf_event_set_output(event, output_event);
5051                         fdput(output);
5052                 } else {
5053                         ret = perf_event_set_output(event, NULL);
5054                 }
5055                 return ret;
5056         }
5057 
5058         case PERF_EVENT_IOC_SET_FILTER:
5059                 return perf_event_set_filter(event, (void __user *)arg);
5060 
5061         case PERF_EVENT_IOC_SET_BPF:
5062                 return perf_event_set_bpf_prog(event, arg);
5063 
5064         case PERF_EVENT_IOC_PAUSE_OUTPUT: {
5065                 struct ring_buffer *rb;
5066 
5067                 rcu_read_lock();
5068                 rb = rcu_dereference(event->rb);
5069                 if (!rb || !rb->nr_pages) {
5070                         rcu_read_unlock();
5071                         return -EINVAL;
5072                 }
5073                 rb_toggle_paused(rb, !!arg);
5074                 rcu_read_unlock();
5075                 return 0;
5076         }
5077 
5078         case PERF_EVENT_IOC_QUERY_BPF:
5079                 return perf_event_query_prog_array(event, (void __user *)arg);
5080 
5081         case PERF_EVENT_IOC_MODIFY_ATTRIBUTES: {
5082                 struct perf_event_attr new_attr;
5083                 int err = perf_copy_attr((struct perf_event_attr __user *)arg,
5084                                          &new_attr);
5085 
5086                 if (err)
5087                         return err;
5088 
5089                 return perf_event_modify_attr(event,  &new_attr);
5090         }
5091         default:
5092                 return -ENOTTY;
5093         }
5094 
5095         if (flags & PERF_IOC_FLAG_GROUP)
5096                 perf_event_for_each(event, func);
5097         else
5098                 perf_event_for_each_child(event, func);
5099 
5100         return 0;
5101 }
5102 
5103 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
5104 {
5105         struct perf_event *event = file->private_data;
5106         struct perf_event_context *ctx;
5107         long ret;
5108 
5109         ctx = perf_event_ctx_lock(event);
5110         ret = _perf_ioctl(event, cmd, arg);
5111         perf_event_ctx_unlock(event, ctx);
5112 
5113         return ret;
5114 }
5115 
5116 #ifdef CONFIG_COMPAT
5117 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
5118                                 unsigned long arg)
5119 {
5120         switch (_IOC_NR(cmd)) {
5121         case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
5122         case _IOC_NR(PERF_EVENT_IOC_ID):
5123                 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
5124                 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
5125                         cmd &= ~IOCSIZE_MASK;
5126                         cmd |= sizeof(void *) << IOCSIZE_SHIFT;
5127                 }
5128                 break;
5129         }
5130         return perf_ioctl(file, cmd, arg);
5131 }
5132 #else
5133 # define perf_compat_ioctl NULL
5134 #endif
5135 
5136 int perf_event_task_enable(void)
5137 {
5138         struct perf_event_context *ctx;
5139         struct perf_event *event;
5140 
5141         mutex_lock(&current->perf_event_mutex);
5142         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5143                 ctx = perf_event_ctx_lock(event);
5144                 perf_event_for_each_child(event, _perf_event_enable);
5145                 perf_event_ctx_unlock(event, ctx);
5146         }
5147         mutex_unlock(&current->perf_event_mutex);
5148 
5149         return 0;
5150 }
5151 
5152 int perf_event_task_disable(void)
5153 {
5154         struct perf_event_context *ctx;
5155         struct perf_event *event;
5156 
5157         mutex_lock(&current->perf_event_mutex);
5158         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5159                 ctx = perf_event_ctx_lock(event);
5160                 perf_event_for_each_child(event, _perf_event_disable);
5161                 perf_event_ctx_unlock(event, ctx);
5162         }
5163         mutex_unlock(&current->perf_event_mutex);
5164 
5165         return 0;
5166 }
5167 
5168 static int perf_event_index(struct perf_event *event)
5169 {
5170         if (event->hw.state & PERF_HES_STOPPED)
5171                 return 0;
5172 
5173         if (event->state != PERF_EVENT_STATE_ACTIVE)
5174                 return 0;
5175 
5176         return event->pmu->event_idx(event);
5177 }
5178 
5179 static void calc_timer_values(struct perf_event *event,
5180                                 u64 *now,
5181                                 u64 *enabled,
5182                                 u64 *running)
5183 {
5184         u64 ctx_time;
5185 
5186         *now = perf_clock();
5187         ctx_time = event->shadow_ctx_time + *now;
5188         __perf_update_times(event, ctx_time, enabled, running);
5189 }
5190 
5191 static void perf_event_init_userpage(struct perf_event *event)
5192 {
5193         struct perf_event_mmap_page *userpg;
5194         struct ring_buffer *rb;
5195 
5196         rcu_read_lock();
5197         rb = rcu_dereference(event->rb);
5198         if (!rb)
5199                 goto unlock;
5200 
5201         userpg = rb->user_page;
5202 
5203         /* Allow new userspace to detect that bit 0 is deprecated */
5204         userpg->cap_bit0_is_deprecated = 1;
5205         userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
5206         userpg->data_offset = PAGE_SIZE;
5207         userpg->data_size = perf_data_size(rb);
5208 
5209 unlock:
5210         rcu_read_unlock();
5211 }
5212 
5213 void __weak arch_perf_update_userpage(
5214         struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
5215 {
5216 }
5217 
5218 /*
5219  * Callers need to ensure there can be no nesting of this function, otherwise
5220  * the seqlock logic goes bad. We can not serialize this because the arch
5221  * code calls this from NMI context.
5222  */
5223 void perf_event_update_userpage(struct perf_event *event)
5224 {
5225         struct perf_event_mmap_page *userpg;
5226         struct ring_buffer *rb;
5227         u64 enabled, running, now;
5228 
5229         rcu_read_lock();
5230         rb = rcu_dereference(event->rb);
5231         if (!rb)
5232                 goto unlock;
5233 
5234         /*
5235          * compute total_time_enabled, total_time_running
5236          * based on snapshot values taken when the event
5237          * was last scheduled in.
5238          *
5239          * we cannot simply called update_context_time()
5240          * because of locking issue as we can be called in
5241          * NMI context
5242          */
5243         calc_timer_values(event, &now, &enabled, &running);
5244 
5245         userpg = rb->user_page;
5246         /*
5247          * Disable preemption so as to not let the corresponding user-space
5248          * spin too long if we get preempted.
5249          */
5250         preempt_disable();
5251         ++userpg->lock;
5252         barrier();
5253         userpg->index = perf_event_index(event);
5254         userpg->offset = perf_event_count(event);
5255         if (userpg->index)
5256                 userpg->offset -= local64_read(&event->hw.prev_count);
5257 
5258         userpg->time_enabled = enabled +
5259                         atomic64_read(&event->child_total_time_enabled);
5260 
5261         userpg->time_running = running +
5262                         atomic64_read(&event->child_total_time_running);
5263 
5264         arch_perf_update_userpage(event, userpg, now);
5265 
5266         barrier();
5267         ++userpg->lock;
5268         preempt_enable();
5269 unlock:
5270         rcu_read_unlock();
5271 }
5272 EXPORT_SYMBOL_GPL(perf_event_update_userpage);
5273 
5274 static int perf_mmap_fault(struct vm_fault *vmf)
5275 {
5276         struct perf_event *event = vmf->vma->vm_file->private_data;
5277         struct ring_buffer *rb;
5278         int ret = VM_FAULT_SIGBUS;
5279 
5280         if (vmf->flags & FAULT_FLAG_MKWRITE) {
5281                 if (vmf->pgoff == 0)
5282                         ret = 0;
5283                 return ret;
5284         }
5285 
5286         rcu_read_lock();
5287         rb = rcu_dereference(event->rb);
5288         if (!rb)
5289                 goto unlock;
5290 
5291         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
5292                 goto unlock;
5293 
5294         vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
5295         if (!vmf->page)
5296                 goto unlock;
5297 
5298         get_page(vmf->page);
5299         vmf->page->mapping = vmf->vma->vm_file->f_mapping;
5300         vmf->page->index   = vmf->pgoff;
5301 
5302         ret = 0;
5303 unlock:
5304         rcu_read_unlock();
5305 
5306         return ret;
5307 }
5308 
5309 static void ring_buffer_attach(struct perf_event *event,
5310                                struct ring_buffer *rb)
5311 {
5312         struct ring_buffer *old_rb = NULL;
5313         unsigned long flags;
5314 
5315         if (event->rb) {
5316                 /*
5317                  * Should be impossible, we set this when removing
5318                  * event->rb_entry and wait/clear when adding event->rb_entry.
5319                  */
5320                 WARN_ON_ONCE(event->rcu_pending);
5321 
5322                 old_rb = event->rb;
5323                 spin_lock_irqsave(&old_rb->event_lock, flags);
5324                 list_del_rcu(&event->rb_entry);
5325                 spin_unlock_irqrestore(&old_rb->event_lock, flags);
5326 
5327                 event->rcu_batches = get_state_synchronize_rcu();
5328                 event->rcu_pending = 1;
5329         }
5330 
5331         if (rb) {
5332                 if (event->rcu_pending) {
5333                         cond_synchronize_rcu(event->rcu_batches);
5334                         event->rcu_pending = 0;
5335                 }
5336 
5337                 spin_lock_irqsave(&rb->event_lock, flags);
5338                 list_add_rcu(&event->rb_entry, &rb->event_list);
5339                 spin_unlock_irqrestore(&rb->event_lock, flags);
5340         }
5341 
5342         /*
5343          * Avoid racing with perf_mmap_close(AUX): stop the event
5344          * before swizzling the event::rb pointer; if it's getting
5345          * unmapped, its aux_mmap_count will be 0 and it won't
5346          * restart. See the comment in __perf_pmu_output_stop().
5347          *
5348          * Data will inevitably be lost when set_output is done in
5349          * mid-air, but then again, whoever does it like this is
5350          * not in for the data anyway.
5351          */
5352         if (has_aux(event))
5353                 perf_event_stop(event, 0);
5354 
5355         rcu_assign_pointer(event->rb, rb);
5356 
5357         if (old_rb) {
5358                 ring_buffer_put(old_rb);
5359                 /*
5360                  * Since we detached before setting the new rb, so that we
5361                  * could attach the new rb, we could have missed a wakeup.
5362                  * Provide it now.
5363                  */
5364                 wake_up_all(&event->waitq);
5365         }
5366 }
5367 
5368 static void ring_buffer_wakeup(struct perf_event *event)
5369 {
5370         struct ring_buffer *rb;
5371 
5372         rcu_read_lock();
5373         rb = rcu_dereference(event->rb);
5374         if (rb) {
5375                 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
5376                         wake_up_all(&event->waitq);
5377         }
5378         rcu_read_unlock();
5379 }
5380 
5381 struct ring_buffer *ring_buffer_get(struct perf_event *event)
5382 {
5383         struct ring_buffer *rb;
5384 
5385         rcu_read_lock();
5386         rb = rcu_dereference(event->rb);
5387         if (rb) {
5388                 if (!atomic_inc_not_zero(&rb->refcount))
5389                         rb = NULL;
5390         }
5391         rcu_read_unlock();
5392 
5393         return rb;
5394 }
5395 
5396 void ring_buffer_put(struct ring_buffer *rb)
5397 {
5398         if (!atomic_dec_and_test(&rb->refcount))
5399                 return;
5400 
5401         WARN_ON_ONCE(!list_empty(&rb->event_list));
5402 
5403         call_rcu(&rb->rcu_head, rb_free_rcu);
5404 }
5405 
5406 static void perf_mmap_open(struct vm_area_struct *vma)
5407 {
5408         struct perf_event *event = vma->vm_file->private_data;
5409 
5410         atomic_inc(&event->mmap_count);
5411         atomic_inc(&event->rb->mmap_count);
5412 
5413         if (vma->vm_pgoff)
5414                 atomic_inc(&event->rb->aux_mmap_count);
5415 
5416         if (event->pmu->event_mapped)
5417                 event->pmu->event_mapped(event, vma->vm_mm);
5418 }
5419 
5420 static void perf_pmu_output_stop(struct perf_event *event);
5421 
5422 /*
5423  * A buffer can be mmap()ed multiple times; either directly through the same
5424  * event, or through other events by use of perf_event_set_output().
5425  *
5426  * In order to undo the VM accounting done by perf_mmap() we need to destroy
5427  * the buffer here, where we still have a VM context. This means we need
5428  * to detach all events redirecting to us.
5429  */
5430 static void perf_mmap_close(struct vm_area_struct *vma)
5431 {
5432         struct perf_event *event = vma->vm_file->private_data;
5433 
5434         struct ring_buffer *rb = ring_buffer_get(event);
5435         struct user_struct *mmap_user = rb->mmap_user;
5436         int mmap_locked = rb->mmap_locked;
5437         unsigned long size = perf_data_size(rb);
5438 
5439         if (event->pmu->event_unmapped)
5440                 event->pmu->event_unmapped(event, vma->vm_mm);
5441 
5442         /*
5443          * rb->aux_mmap_count will always drop before rb->mmap_count and
5444          * event->mmap_count, so it is ok to use event->mmap_mutex to
5445          * serialize with perf_mmap here.
5446          */
5447         if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
5448             atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
5449                 /*
5450                  * Stop all AUX events that are writing to this buffer,
5451                  * so that we can free its AUX pages and corresponding PMU
5452                  * data. Note that after rb::aux_mmap_count dropped to zero,
5453                  * they won't start any more (see perf_aux_output_begin()).
5454                  */
5455                 perf_pmu_output_stop(event);
5456 
5457                 /* now it's safe to free the pages */
5458                 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
5459                 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
5460 
5461                 /* this has to be the last one */
5462                 rb_free_aux(rb);
5463                 WARN_ON_ONCE(atomic_read(&rb->aux_refcount));
5464 
5465                 mutex_unlock(&event->mmap_mutex);
5466         }
5467 
5468         atomic_dec(&rb->mmap_count);
5469 
5470         if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
5471                 goto out_put;
5472 
5473         ring_buffer_attach(event, NULL);
5474         mutex_unlock(&event->mmap_mutex);
5475 
5476         /* If there's still other mmap()s of this buffer, we're done. */
5477         if (atomic_read(&rb->mmap_count))
5478                 goto out_put;
5479 
5480         /*
5481          * No other mmap()s, detach from all other events that might redirect
5482          * into the now unreachable buffer. Somewhat complicated by the
5483          * fact that rb::event_lock otherwise nests inside mmap_mutex.
5484          */
5485 again:
5486         rcu_read_lock();
5487         list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
5488                 if (!atomic_long_inc_not_zero(&event->refcount)) {
5489                         /*
5490                          * This event is en-route to free_event() which will
5491                          * detach it and remove it from the list.
5492                          */
5493                         continue;
5494                 }
5495                 rcu_read_unlock();
5496 
5497                 mutex_lock(&event->mmap_mutex);
5498                 /*
5499                  * Check we didn't race with perf_event_set_output() which can
5500                  * swizzle the rb from under us while we were waiting to
5501                  * acquire mmap_mutex.
5502                  *
5503                  * If we find a different rb; ignore this event, a next
5504                  * iteration will no longer find it on the list. We have to
5505                  * still restart the iteration to make sure we're not now
5506                  * iterating the wrong list.
5507                  */
5508                 if (event->rb == rb)
5509                         ring_buffer_attach(event, NULL);
5510 
5511                 mutex_unlock(&event->mmap_mutex);
5512                 put_event(event);
5513 
5514                 /*
5515                  * Restart the iteration; either we're on the wrong list or
5516                  * destroyed its integrity by doing a deletion.
5517                  */
5518                 goto again;
5519         }
5520         rcu_read_unlock();
5521 
5522         /*
5523          * It could be there's still a few 0-ref events on the list; they'll
5524          * get cleaned up by free_event() -- they'll also still have their
5525          * ref on the rb and will free it whenever they are done with it.
5526          *
5527          * Aside from that, this buffer is 'fully' detached and unmapped,
5528          * undo the VM accounting.
5529          */
5530 
5531         atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
5532         vma->vm_mm->pinned_vm -= mmap_locked;
5533         free_uid(mmap_user);
5534 
5535 out_put:
5536         ring_buffer_put(rb); /* could be last */
5537 }
5538 
5539 static const struct vm_operations_struct perf_mmap_vmops = {
5540         .open           = perf_mmap_open,
5541         .close          = perf_mmap_close, /* non mergable */
5542         .fault          = perf_mmap_fault,
5543         .page_mkwrite   = perf_mmap_fault,
5544 };
5545 
5546 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
5547 {
5548         struct perf_event *event = file->private_data;
5549         unsigned long user_locked, user_lock_limit;
5550         struct user_struct *user = current_user();
5551         unsigned long locked, lock_limit;
5552         struct ring_buffer *rb = NULL;
5553         unsigned long vma_size;
5554         unsigned long nr_pages;
5555         long user_extra = 0, extra = 0;
5556         int ret = 0, flags = 0;
5557 
5558         /*
5559          * Don't allow mmap() of inherited per-task counters. This would
5560          * create a performance issue due to all children writing to the
5561          * same rb.
5562          */
5563         if (event->cpu == -1 && event->attr.inherit)
5564                 return -EINVAL;
5565 
5566         if (!(vma->vm_flags & VM_SHARED))
5567                 return -EINVAL;
5568 
5569         vma_size = vma->vm_end - vma->vm_start;
5570 
5571         if (vma->vm_pgoff == 0) {
5572                 nr_pages = (vma_size / PAGE_SIZE) - 1;
5573         } else {
5574                 /*
5575                  * AUX area mapping: if rb->aux_nr_pages != 0, it's already
5576                  * mapped, all subsequent mappings should have the same size
5577                  * and offset. Must be above the normal perf buffer.
5578                  */
5579                 u64 aux_offset, aux_size;
5580 
5581                 if (!event->rb)
5582                         return -EINVAL;
5583 
5584                 nr_pages = vma_size / PAGE_SIZE;
5585 
5586                 mutex_lock(&event->mmap_mutex);
5587                 ret = -EINVAL;
5588 
5589                 rb = event->rb;
5590                 if (!rb)
5591                         goto aux_unlock;
5592 
5593                 aux_offset = READ_ONCE(rb->user_page->aux_offset);
5594                 aux_size = READ_ONCE(rb->user_page->aux_size);
5595 
5596                 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
5597                         goto aux_unlock;
5598 
5599                 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
5600                         goto aux_unlock;
5601 
5602                 /* already mapped with a different offset */
5603                 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
5604                         goto aux_unlock;
5605 
5606                 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
5607                         goto aux_unlock;
5608 
5609                 /* already mapped with a different size */
5610                 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
5611                         goto aux_unlock;
5612 
5613                 if (!is_power_of_2(nr_pages))
5614                         goto aux_unlock;
5615 
5616                 if (!atomic_inc_not_zero(&rb->mmap_count))
5617                         goto aux_unlock;
5618 
5619                 if (rb_has_aux(rb)) {
5620                         atomic_inc(&rb->aux_mmap_count);
5621                         ret = 0;
5622                         goto unlock;
5623                 }
5624 
5625                 atomic_set(&rb->aux_mmap_count, 1);
5626                 user_extra = nr_pages;
5627 
5628                 goto accounting;
5629         }
5630 
5631         /*
5632          * If we have rb pages ensure they're a power-of-two number, so we
5633          * can do bitmasks instead of modulo.
5634          */
5635         if (nr_pages != 0 && !is_power_of_2(nr_pages))
5636                 return -EINVAL;
5637 
5638         if (vma_size != PAGE_SIZE * (1 + nr_pages))
5639                 return -EINVAL;
5640 
5641         WARN_ON_ONCE(event->ctx->parent_ctx);
5642 again:
5643         mutex_lock(&event->mmap_mutex);
5644         if (event->rb) {
5645                 if (event->rb->nr_pages != nr_pages) {
5646                         ret = -EINVAL;
5647                         goto unlock;
5648                 }
5649 
5650                 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
5651                         /*
5652                          * Raced against perf_mmap_close() through
5653                          * perf_event_set_output(). Try again, hope for better
5654                          * luck.
5655                          */
5656                         mutex_unlock(&event->mmap_mutex);
5657                         goto again;
5658                 }
5659 
5660                 goto unlock;
5661         }
5662 
5663         user_extra = nr_pages + 1;
5664 
5665 accounting:
5666         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
5667 
5668         /*
5669          * Increase the limit linearly with more CPUs:
5670          */
5671         user_lock_limit *= num_online_cpus();
5672 
5673         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
5674 
5675         if (user_locked > user_lock_limit)
5676                 extra = user_locked - user_lock_limit;
5677 
5678         lock_limit = rlimit(RLIMIT_MEMLOCK);
5679         lock_limit >>= PAGE_SHIFT;
5680         locked = vma->vm_mm->pinned_vm + extra;
5681 
5682         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
5683                 !capable(CAP_IPC_LOCK)) {
5684                 ret = -EPERM;
5685                 goto unlock;
5686         }
5687 
5688         WARN_ON(!rb && event->rb);
5689 
5690         if (vma->vm_flags & VM_WRITE)
5691                 flags |= RING_BUFFER_WRITABLE;
5692 
5693         if (!rb) {
5694                 rb = rb_alloc(nr_pages,
5695                               event->attr.watermark ? event->attr.wakeup_watermark : 0,
5696                               event->cpu, flags);
5697 
5698                 if (!rb) {
5699                         ret = -ENOMEM;
5700                         goto unlock;
5701                 }
5702 
5703                 atomic_set(&rb->mmap_count, 1);
5704                 rb->mmap_user = get_current_user();
5705                 rb->mmap_locked = extra;
5706 
5707                 ring_buffer_attach(event, rb);
5708 
5709                 perf_event_init_userpage(event);
5710                 perf_event_update_userpage(event);
5711         } else {
5712                 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
5713                                    event->attr.aux_watermark, flags);
5714                 if (!ret)
5715                         rb->aux_mmap_locked = extra;
5716         }
5717 
5718 unlock:
5719         if (!ret) {
5720                 atomic_long_add(user_extra, &user->locked_vm);
5721                 vma->vm_mm->pinned_vm += extra;
5722 
5723                 atomic_inc(&event->mmap_count);
5724         } else if (rb) {
5725                 atomic_dec(&rb->mmap_count);
5726         }
5727 aux_unlock:
5728         mutex_unlock(&event->mmap_mutex);
5729 
5730         /*
5731          * Since pinned accounting is per vm we cannot allow fork() to copy our
5732          * vma.
5733          */
5734         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
5735         vma->vm_ops = &perf_mmap_vmops;
5736 
5737         if (event->pmu->event_mapped)
5738                 event->pmu->event_mapped(event, vma->vm_mm);
5739 
5740         return ret;
5741 }
5742 
5743 static int perf_fasync(int fd, struct file *filp, int on)
5744 {
5745         struct inode *inode = file_inode(filp);
5746         struct perf_event *event = filp->private_data;
5747         int retval;
5748 
5749         inode_lock(inode);
5750         retval = fasync_helper(fd, filp, on, &event->fasync);
5751         inode_unlock(inode);
5752 
5753         if (retval < 0)
5754                 return retval;
5755 
5756         return 0;
5757 }
5758 
5759 static const struct file_operations perf_fops = {
5760         .llseek                 = no_llseek,
5761         .release                = perf_release,
5762         .read                   = perf_read,
5763         .poll                   = perf_poll,
5764         .unlocked_ioctl         = perf_ioctl,
5765         .compat_ioctl           = perf_compat_ioctl,
5766         .mmap                   = perf_mmap,
5767         .fasync                 = perf_fasync,
5768 };
5769 
5770 /*
5771  * Perf event wakeup
5772  *
5773  * If there's data, ensure we set the poll() state and publish everything
5774  * to user-space before waking everybody up.
5775  */
5776 
5777 static inline struct fasync_struct **perf_event_fasync(struct perf_event *event)
5778 {
5779         /* only the parent has fasync state */
5780         if (event->parent)
5781                 event = event->parent;
5782         return &event->fasync;
5783 }
5784 
5785 void perf_event_wakeup(struct perf_event *event)
5786 {
5787         ring_buffer_wakeup(event);
5788 
5789         if (event->pending_kill) {
5790                 kill_fasync(perf_event_fasync(event), SIGIO, event->pending_kill);
5791                 event->pending_kill = 0;
5792         }
5793 }
5794 
5795 static void perf_pending_event(struct irq_work *entry)
5796 {
5797         struct perf_event *event = container_of(entry,
5798                         struct perf_event, pending);
5799         int rctx;
5800 
5801         rctx = perf_swevent_get_recursion_context();
5802         /*
5803          * If we 'fail' here, that's OK, it means recursion is already disabled
5804          * and we won't recurse 'further'.
5805          */
5806 
5807         if (event->pending_disable) {
5808                 event->pending_disable = 0;
5809                 perf_event_disable_local(event);
5810         }
5811 
5812         if (event->pending_wakeup) {
5813                 event->pending_wakeup = 0;
5814                 perf_event_wakeup(event);
5815         }
5816 
5817         if (rctx >= 0)
5818                 perf_swevent_put_recursion_context(rctx);
5819 }
5820 
5821 /*
5822  * We assume there is only KVM supporting the callbacks.
5823  * Later on, we might change it to a list if there is
5824  * another virtualization implementation supporting the callbacks.
5825  */
5826 struct perf_guest_info_callbacks *perf_guest_cbs;
5827 
5828 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5829 {
5830         perf_guest_cbs = cbs;
5831         return 0;
5832 }
5833 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
5834 
5835 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5836 {
5837         perf_guest_cbs = NULL;
5838         return 0;
5839 }
5840 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
5841 
5842 static void
5843 perf_output_sample_regs(struct perf_output_handle *handle,
5844                         struct pt_regs *regs, u64 mask)
5845 {
5846         int bit;
5847         DECLARE_BITMAP(_mask, 64);
5848 
5849         bitmap_from_u64(_mask, mask);
5850         for_each_set_bit(bit, _mask, sizeof(mask) * BITS_PER_BYTE) {
5851                 u64 val;
5852 
5853                 val = perf_reg_value(regs, bit);
5854                 perf_output_put(handle, val);
5855         }
5856 }
5857 
5858 static void perf_sample_regs_user(struct perf_regs *regs_user,
5859                                   struct pt_regs *regs,
5860                                   struct pt_regs *regs_user_copy)
5861 {
5862         if (user_mode(regs)) {
5863                 regs_user->abi = perf_reg_abi(current);
5864                 regs_user->regs = regs;
5865         } else if (current->mm) {
5866                 perf_get_regs_user(regs_user, regs, regs_user_copy);
5867         } else {
5868                 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
5869                 regs_user->regs = NULL;
5870         }
5871 }
5872 
5873 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
5874                                   struct pt_regs *regs)
5875 {
5876         regs_intr->regs = regs;
5877         regs_intr->abi  = perf_reg_abi(current);
5878 }
5879 
5880 
5881 /*
5882  * Get remaining task size from user stack pointer.
5883  *
5884  * It'd be better to take stack vma map and limit this more
5885  * precisly, but there's no way to get it safely under interrupt,
5886  * so using TASK_SIZE as limit.
5887  */
5888 static u64 perf_ustack_task_size(struct pt_regs *regs)
5889 {
5890         unsigned long addr = perf_user_stack_pointer(regs);
5891 
5892         if (!addr || addr >= TASK_SIZE)
5893                 return 0;
5894 
5895         return TASK_SIZE - addr;
5896 }
5897 
5898 static u16
5899 perf_sample_ustack_size(u16 stack_size, u16 header_size,
5900                         struct pt_regs *regs)
5901 {
5902         u64 task_size;
5903 
5904         /* No regs, no stack pointer, no dump. */
5905         if (!regs)
5906                 return 0;
5907 
5908         /*
5909          * Check if we fit in with the requested stack size into the:
5910          * - TASK_SIZE
5911          *   If we don't, we limit the size to the TASK_SIZE.
5912          *
5913          * - remaining sample size
5914          *   If we don't, we customize the stack size to
5915          *   fit in to the remaining sample size.
5916          */
5917 
5918         task_size  = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
5919         stack_size = min(stack_size, (u16) task_size);
5920 
5921         /* Current header size plus static size and dynamic size. */
5922         header_size += 2 * sizeof(u64);
5923 
5924         /* Do we fit in with the current stack dump size? */
5925         if ((u16) (header_size + stack_size) < header_size) {
5926                 /*
5927                  * If we overflow the maximum size for the sample,
5928                  * we customize the stack dump size to fit in.
5929                  */
5930                 stack_size = USHRT_MAX - header_size - sizeof(u64);
5931                 stack_size = round_up(stack_size, sizeof(u64));
5932         }
5933 
5934         return stack_size;
5935 }
5936 
5937 static void
5938 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
5939                           struct pt_regs *regs)
5940 {
5941         /* Case of a kernel thread, nothing