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

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

Version: ~ [ linux-5.9 ] ~ [ linux-5.8.14 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.70 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.150 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.200 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.238 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.238 ] ~ [ 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 int 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         list_for_each_entry(sibling, &leader->sibling_list, group_entry)
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_out = cpuctx->cgrp;
728         if (cgrp_out)
729                 __update_cgrp_time(cgrp_out);
730 }
731 
732 static inline void update_cgrp_time_from_event(struct perf_event *event)
733 {
734         struct perf_cgroup *cgrp;
735 
736         /*
737          * ensure we access cgroup data only when needed and
738          * when we know the cgroup is pinned (css_get)
739          */
740         if (!is_cgroup_event(event))
741                 return;
742 
743         cgrp = perf_cgroup_from_task(current, event->ctx);
744         /*
745          * Do not update time when cgroup is not active
746          */
747        if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
748                 __update_cgrp_time(event->cgrp);
749 }
750 
751 static inline void
752 perf_cgroup_set_timestamp(struct task_struct *task,
753                           struct perf_event_context *ctx)
754 {
755         struct perf_cgroup *cgrp;
756         struct perf_cgroup_info *info;
757 
758         /*
759          * ctx->lock held by caller
760          * ensure we do not access cgroup data
761          * unless we have the cgroup pinned (css_get)
762          */
763         if (!task || !ctx->nr_cgroups)
764                 return;
765 
766         cgrp = perf_cgroup_from_task(task, ctx);
767         info = this_cpu_ptr(cgrp->info);
768         info->timestamp = ctx->timestamp;
769 }
770 
771 static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
772 
773 #define PERF_CGROUP_SWOUT       0x1 /* cgroup switch out every event */
774 #define PERF_CGROUP_SWIN        0x2 /* cgroup switch in events based on task */
775 
776 /*
777  * reschedule events based on the cgroup constraint of task.
778  *
779  * mode SWOUT : schedule out everything
780  * mode SWIN : schedule in based on cgroup for next
781  */
782 static void perf_cgroup_switch(struct task_struct *task, int mode)
783 {
784         struct perf_cpu_context *cpuctx;
785         struct list_head *list;
786         unsigned long flags;
787 
788         /*
789          * Disable interrupts and preemption to avoid this CPU's
790          * cgrp_cpuctx_entry to change under us.
791          */
792         local_irq_save(flags);
793 
794         list = this_cpu_ptr(&cgrp_cpuctx_list);
795         list_for_each_entry(cpuctx, list, cgrp_cpuctx_entry) {
796                 WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
797 
798                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
799                 perf_pmu_disable(cpuctx->ctx.pmu);
800 
801                 if (mode & PERF_CGROUP_SWOUT) {
802                         cpu_ctx_sched_out(cpuctx, EVENT_ALL);
803                         /*
804                          * must not be done before ctxswout due
805                          * to event_filter_match() in event_sched_out()
806                          */
807                         cpuctx->cgrp = NULL;
808                 }
809 
810                 if (mode & PERF_CGROUP_SWIN) {
811                         WARN_ON_ONCE(cpuctx->cgrp);
812                         /*
813                          * set cgrp before ctxsw in to allow
814                          * event_filter_match() to not have to pass
815                          * task around
816                          * we pass the cpuctx->ctx to perf_cgroup_from_task()
817                          * because cgorup events are only per-cpu
818                          */
819                         cpuctx->cgrp = perf_cgroup_from_task(task,
820                                                              &cpuctx->ctx);
821                         cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
822                 }
823                 perf_pmu_enable(cpuctx->ctx.pmu);
824                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
825         }
826 
827         local_irq_restore(flags);
828 }
829 
830 static inline void perf_cgroup_sched_out(struct task_struct *task,
831                                          struct task_struct *next)
832 {
833         struct perf_cgroup *cgrp1;
834         struct perf_cgroup *cgrp2 = NULL;
835 
836         rcu_read_lock();
837         /*
838          * we come here when we know perf_cgroup_events > 0
839          * we do not need to pass the ctx here because we know
840          * we are holding the rcu lock
841          */
842         cgrp1 = perf_cgroup_from_task(task, NULL);
843         cgrp2 = perf_cgroup_from_task(next, NULL);
844 
845         /*
846          * only schedule out current cgroup events if we know
847          * that we are switching to a different cgroup. Otherwise,
848          * do no touch the cgroup events.
849          */
850         if (cgrp1 != cgrp2)
851                 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
852 
853         rcu_read_unlock();
854 }
855 
856 static inline void perf_cgroup_sched_in(struct task_struct *prev,
857                                         struct task_struct *task)
858 {
859         struct perf_cgroup *cgrp1;
860         struct perf_cgroup *cgrp2 = NULL;
861 
862         rcu_read_lock();
863         /*
864          * we come here when we know perf_cgroup_events > 0
865          * we do not need to pass the ctx here because we know
866          * we are holding the rcu lock
867          */
868         cgrp1 = perf_cgroup_from_task(task, NULL);
869         cgrp2 = perf_cgroup_from_task(prev, NULL);
870 
871         /*
872          * only need to schedule in cgroup events if we are changing
873          * cgroup during ctxsw. Cgroup events were not scheduled
874          * out of ctxsw out if that was not the case.
875          */
876         if (cgrp1 != cgrp2)
877                 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
878 
879         rcu_read_unlock();
880 }
881 
882 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
883                                       struct perf_event_attr *attr,
884                                       struct perf_event *group_leader)
885 {
886         struct perf_cgroup *cgrp;
887         struct cgroup_subsys_state *css;
888         struct fd f = fdget(fd);
889         int ret = 0;
890 
891         if (!f.file)
892                 return -EBADF;
893 
894         css = css_tryget_online_from_dir(f.file->f_path.dentry,
895                                          &perf_event_cgrp_subsys);
896         if (IS_ERR(css)) {
897                 ret = PTR_ERR(css);
898                 goto out;
899         }
900 
901         cgrp = container_of(css, struct perf_cgroup, css);
902         event->cgrp = cgrp;
903 
904         /*
905          * all events in a group must monitor
906          * the same cgroup because a task belongs
907          * to only one perf cgroup at a time
908          */
909         if (group_leader && group_leader->cgrp != cgrp) {
910                 perf_detach_cgroup(event);
911                 ret = -EINVAL;
912         }
913 out:
914         fdput(f);
915         return ret;
916 }
917 
918 static inline void
919 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
920 {
921         struct perf_cgroup_info *t;
922         t = per_cpu_ptr(event->cgrp->info, event->cpu);
923         event->shadow_ctx_time = now - t->timestamp;
924 }
925 
926 /*
927  * Update cpuctx->cgrp so that it is set when first cgroup event is added and
928  * cleared when last cgroup event is removed.
929  */
930 static inline void
931 list_update_cgroup_event(struct perf_event *event,
932                          struct perf_event_context *ctx, bool add)
933 {
934         struct perf_cpu_context *cpuctx;
935         struct list_head *cpuctx_entry;
936 
937         if (!is_cgroup_event(event))
938                 return;
939 
940         if (add && ctx->nr_cgroups++)
941                 return;
942         else if (!add && --ctx->nr_cgroups)
943                 return;
944         /*
945          * Because cgroup events are always per-cpu events,
946          * this will always be called from the right CPU.
947          */
948         cpuctx = __get_cpu_context(ctx);
949         cpuctx_entry = &cpuctx->cgrp_cpuctx_entry;
950         /* cpuctx->cgrp is NULL unless a cgroup event is active in this CPU .*/
951         if (add) {
952                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
953 
954                 list_add(cpuctx_entry, this_cpu_ptr(&cgrp_cpuctx_list));
955                 if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
956                         cpuctx->cgrp = cgrp;
957         } else {
958                 list_del(cpuctx_entry);
959                 cpuctx->cgrp = NULL;
960         }
961 }
962 
963 #else /* !CONFIG_CGROUP_PERF */
964 
965 static inline bool
966 perf_cgroup_match(struct perf_event *event)
967 {
968         return true;
969 }
970 
971 static inline void perf_detach_cgroup(struct perf_event *event)
972 {}
973 
974 static inline int is_cgroup_event(struct perf_event *event)
975 {
976         return 0;
977 }
978 
979 static inline void update_cgrp_time_from_event(struct perf_event *event)
980 {
981 }
982 
983 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
984 {
985 }
986 
987 static inline void perf_cgroup_sched_out(struct task_struct *task,
988                                          struct task_struct *next)
989 {
990 }
991 
992 static inline void perf_cgroup_sched_in(struct task_struct *prev,
993                                         struct task_struct *task)
994 {
995 }
996 
997 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
998                                       struct perf_event_attr *attr,
999                                       struct perf_event *group_leader)
1000 {
1001         return -EINVAL;
1002 }
1003 
1004 static inline void
1005 perf_cgroup_set_timestamp(struct task_struct *task,
1006                           struct perf_event_context *ctx)
1007 {
1008 }
1009 
1010 void
1011 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
1012 {
1013 }
1014 
1015 static inline void
1016 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
1017 {
1018 }
1019 
1020 static inline u64 perf_cgroup_event_time(struct perf_event *event)
1021 {
1022         return 0;
1023 }
1024 
1025 static inline void
1026 list_update_cgroup_event(struct perf_event *event,
1027                          struct perf_event_context *ctx, bool add)
1028 {
1029 }
1030 
1031 #endif
1032 
1033 /*
1034  * set default to be dependent on timer tick just
1035  * like original code
1036  */
1037 #define PERF_CPU_HRTIMER (1000 / HZ)
1038 /*
1039  * function must be called with interrupts disabled
1040  */
1041 static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr)
1042 {
1043         struct perf_cpu_context *cpuctx;
1044         int rotations = 0;
1045 
1046         lockdep_assert_irqs_disabled();
1047 
1048         cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
1049         rotations = perf_rotate_context(cpuctx);
1050 
1051         raw_spin_lock(&cpuctx->hrtimer_lock);
1052         if (rotations)
1053                 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
1054         else
1055                 cpuctx->hrtimer_active = 0;
1056         raw_spin_unlock(&cpuctx->hrtimer_lock);
1057 
1058         return rotations ? HRTIMER_RESTART : HRTIMER_NORESTART;
1059 }
1060 
1061 static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
1062 {
1063         struct hrtimer *timer = &cpuctx->hrtimer;
1064         struct pmu *pmu = cpuctx->ctx.pmu;
1065         u64 interval;
1066 
1067         /* no multiplexing needed for SW PMU */
1068         if (pmu->task_ctx_nr == perf_sw_context)
1069                 return;
1070 
1071         /*
1072          * check default is sane, if not set then force to
1073          * default interval (1/tick)
1074          */
1075         interval = pmu->hrtimer_interval_ms;
1076         if (interval < 1)
1077                 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
1078 
1079         cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval);
1080 
1081         raw_spin_lock_init(&cpuctx->hrtimer_lock);
1082         hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1083         timer->function = perf_mux_hrtimer_handler;
1084 }
1085 
1086 static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
1087 {
1088         struct hrtimer *timer = &cpuctx->hrtimer;
1089         struct pmu *pmu = cpuctx->ctx.pmu;
1090         unsigned long flags;
1091 
1092         /* not for SW PMU */
1093         if (pmu->task_ctx_nr == perf_sw_context)
1094                 return 0;
1095 
1096         raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags);
1097         if (!cpuctx->hrtimer_active) {
1098                 cpuctx->hrtimer_active = 1;
1099                 hrtimer_forward_now(timer, cpuctx->hrtimer_interval);
1100                 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
1101         }
1102         raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags);
1103 
1104         return 0;
1105 }
1106 
1107 void perf_pmu_disable(struct pmu *pmu)
1108 {
1109         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1110         if (!(*count)++)
1111                 pmu->pmu_disable(pmu);
1112 }
1113 
1114 void perf_pmu_enable(struct pmu *pmu)
1115 {
1116         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1117         if (!--(*count))
1118                 pmu->pmu_enable(pmu);
1119 }
1120 
1121 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
1122 
1123 /*
1124  * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
1125  * perf_event_task_tick() are fully serialized because they're strictly cpu
1126  * affine and perf_event_ctx{activate,deactivate} are called with IRQs
1127  * disabled, while perf_event_task_tick is called from IRQ context.
1128  */
1129 static void perf_event_ctx_activate(struct perf_event_context *ctx)
1130 {
1131         struct list_head *head = this_cpu_ptr(&active_ctx_list);
1132 
1133         lockdep_assert_irqs_disabled();
1134 
1135         WARN_ON(!list_empty(&ctx->active_ctx_list));
1136 
1137         list_add(&ctx->active_ctx_list, head);
1138 }
1139 
1140 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
1141 {
1142         lockdep_assert_irqs_disabled();
1143 
1144         WARN_ON(list_empty(&ctx->active_ctx_list));
1145 
1146         list_del_init(&ctx->active_ctx_list);
1147 }
1148 
1149 static void get_ctx(struct perf_event_context *ctx)
1150 {
1151         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
1152 }
1153 
1154 static void free_ctx(struct rcu_head *head)
1155 {
1156         struct perf_event_context *ctx;
1157 
1158         ctx = container_of(head, struct perf_event_context, rcu_head);
1159         kfree(ctx->task_ctx_data);
1160         kfree(ctx);
1161 }
1162 
1163 static void put_ctx(struct perf_event_context *ctx)
1164 {
1165         if (atomic_dec_and_test(&ctx->refcount)) {
1166                 if (ctx->parent_ctx)
1167                         put_ctx(ctx->parent_ctx);
1168                 if (ctx->task && ctx->task != TASK_TOMBSTONE)
1169                         put_task_struct(ctx->task);
1170                 call_rcu(&ctx->rcu_head, free_ctx);
1171         }
1172 }
1173 
1174 /*
1175  * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
1176  * perf_pmu_migrate_context() we need some magic.
1177  *
1178  * Those places that change perf_event::ctx will hold both
1179  * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
1180  *
1181  * Lock ordering is by mutex address. There are two other sites where
1182  * perf_event_context::mutex nests and those are:
1183  *
1184  *  - perf_event_exit_task_context()    [ child , 0 ]
1185  *      perf_event_exit_event()
1186  *        put_event()                   [ parent, 1 ]
1187  *
1188  *  - perf_event_init_context()         [ parent, 0 ]
1189  *      inherit_task_group()
1190  *        inherit_group()
1191  *          inherit_event()
1192  *            perf_event_alloc()
1193  *              perf_init_event()
1194  *                perf_try_init_event() [ child , 1 ]
1195  *
1196  * While it appears there is an obvious deadlock here -- the parent and child
1197  * nesting levels are inverted between the two. This is in fact safe because
1198  * life-time rules separate them. That is an exiting task cannot fork, and a
1199  * spawning task cannot (yet) exit.
1200  *
1201  * But remember that that these are parent<->child context relations, and
1202  * migration does not affect children, therefore these two orderings should not
1203  * interact.
1204  *
1205  * The change in perf_event::ctx does not affect children (as claimed above)
1206  * because the sys_perf_event_open() case will install a new event and break
1207  * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
1208  * concerned with cpuctx and that doesn't have children.
1209  *
1210  * The places that change perf_event::ctx will issue:
1211  *
1212  *   perf_remove_from_context();
1213  *   synchronize_rcu();
1214  *   perf_install_in_context();
1215  *
1216  * to affect the change. The remove_from_context() + synchronize_rcu() should
1217  * quiesce the event, after which we can install it in the new location. This
1218  * means that only external vectors (perf_fops, prctl) can perturb the event
1219  * while in transit. Therefore all such accessors should also acquire
1220  * perf_event_context::mutex to serialize against this.
1221  *
1222  * However; because event->ctx can change while we're waiting to acquire
1223  * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
1224  * function.
1225  *
1226  * Lock order:
1227  *    cred_guard_mutex
1228  *      task_struct::perf_event_mutex
1229  *        perf_event_context::mutex
1230  *          perf_event::child_mutex;
1231  *            perf_event_context::lock
1232  *          perf_event::mmap_mutex
1233  *          mmap_sem
1234  *
1235  *    cpu_hotplug_lock
1236  *      pmus_lock
1237  *        cpuctx->mutex / perf_event_context::mutex
1238  */
1239 static struct perf_event_context *
1240 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
1241 {
1242         struct perf_event_context *ctx;
1243 
1244 again:
1245         rcu_read_lock();
1246         ctx = READ_ONCE(event->ctx);
1247         if (!atomic_inc_not_zero(&ctx->refcount)) {
1248                 rcu_read_unlock();
1249                 goto again;
1250         }
1251         rcu_read_unlock();
1252 
1253         mutex_lock_nested(&ctx->mutex, nesting);
1254         if (event->ctx != ctx) {
1255                 mutex_unlock(&ctx->mutex);
1256                 put_ctx(ctx);
1257                 goto again;
1258         }
1259 
1260         return ctx;
1261 }
1262 
1263 static inline struct perf_event_context *
1264 perf_event_ctx_lock(struct perf_event *event)
1265 {
1266         return perf_event_ctx_lock_nested(event, 0);
1267 }
1268 
1269 static void perf_event_ctx_unlock(struct perf_event *event,
1270                                   struct perf_event_context *ctx)
1271 {
1272         mutex_unlock(&ctx->mutex);
1273         put_ctx(ctx);
1274 }
1275 
1276 /*
1277  * This must be done under the ctx->lock, such as to serialize against
1278  * context_equiv(), therefore we cannot call put_ctx() since that might end up
1279  * calling scheduler related locks and ctx->lock nests inside those.
1280  */
1281 static __must_check struct perf_event_context *
1282 unclone_ctx(struct perf_event_context *ctx)
1283 {
1284         struct perf_event_context *parent_ctx = ctx->parent_ctx;
1285 
1286         lockdep_assert_held(&ctx->lock);
1287 
1288         if (parent_ctx)
1289                 ctx->parent_ctx = NULL;
1290         ctx->generation++;
1291 
1292         return parent_ctx;
1293 }
1294 
1295 static u32 perf_event_pid_type(struct perf_event *event, struct task_struct *p,
1296                                 enum pid_type type)
1297 {
1298         u32 nr;
1299         /*
1300          * only top level events have the pid namespace they were created in
1301          */
1302         if (event->parent)
1303                 event = event->parent;
1304 
1305         nr = __task_pid_nr_ns(p, type, event->ns);
1306         /* avoid -1 if it is idle thread or runs in another ns */
1307         if (!nr && !pid_alive(p))
1308                 nr = -1;
1309         return nr;
1310 }
1311 
1312 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1313 {
1314         return perf_event_pid_type(event, p, __PIDTYPE_TGID);
1315 }
1316 
1317 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1318 {
1319         return perf_event_pid_type(event, p, PIDTYPE_PID);
1320 }
1321 
1322 /*
1323  * If we inherit events we want to return the parent event id
1324  * to userspace.
1325  */
1326 static u64 primary_event_id(struct perf_event *event)
1327 {
1328         u64 id = event->id;
1329 
1330         if (event->parent)
1331                 id = event->parent->id;
1332 
1333         return id;
1334 }
1335 
1336 /*
1337  * Get the perf_event_context for a task and lock it.
1338  *
1339  * This has to cope with with the fact that until it is locked,
1340  * the context could get moved to another task.
1341  */
1342 static struct perf_event_context *
1343 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1344 {
1345         struct perf_event_context *ctx;
1346 
1347 retry:
1348         /*
1349          * One of the few rules of preemptible RCU is that one cannot do
1350          * rcu_read_unlock() while holding a scheduler (or nested) lock when
1351          * part of the read side critical section was irqs-enabled -- see
1352          * rcu_read_unlock_special().
1353          *
1354          * Since ctx->lock nests under rq->lock we must ensure the entire read
1355          * side critical section has interrupts disabled.
1356          */
1357         local_irq_save(*flags);
1358         rcu_read_lock();
1359         ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1360         if (ctx) {
1361                 /*
1362                  * If this context is a clone of another, it might
1363                  * get swapped for another underneath us by
1364                  * perf_event_task_sched_out, though the
1365                  * rcu_read_lock() protects us from any context
1366                  * getting freed.  Lock the context and check if it
1367                  * got swapped before we could get the lock, and retry
1368                  * if so.  If we locked the right context, then it
1369                  * can't get swapped on us any more.
1370                  */
1371                 raw_spin_lock(&ctx->lock);
1372                 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1373                         raw_spin_unlock(&ctx->lock);
1374                         rcu_read_unlock();
1375                         local_irq_restore(*flags);
1376                         goto retry;
1377                 }
1378 
1379                 if (ctx->task == TASK_TOMBSTONE ||
1380                     !atomic_inc_not_zero(&ctx->refcount)) {
1381                         raw_spin_unlock(&ctx->lock);
1382                         ctx = NULL;
1383                 } else {
1384                         WARN_ON_ONCE(ctx->task != task);
1385                 }
1386         }
1387         rcu_read_unlock();
1388         if (!ctx)
1389                 local_irq_restore(*flags);
1390         return ctx;
1391 }
1392 
1393 /*
1394  * Get the context for a task and increment its pin_count so it
1395  * can't get swapped to another task.  This also increments its
1396  * reference count so that the context can't get freed.
1397  */
1398 static struct perf_event_context *
1399 perf_pin_task_context(struct task_struct *task, int ctxn)
1400 {
1401         struct perf_event_context *ctx;
1402         unsigned long flags;
1403 
1404         ctx = perf_lock_task_context(task, ctxn, &flags);
1405         if (ctx) {
1406                 ++ctx->pin_count;
1407                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1408         }
1409         return ctx;
1410 }
1411 
1412 static void perf_unpin_context(struct perf_event_context *ctx)
1413 {
1414         unsigned long flags;
1415 
1416         raw_spin_lock_irqsave(&ctx->lock, flags);
1417         --ctx->pin_count;
1418         raw_spin_unlock_irqrestore(&ctx->lock, flags);
1419 }
1420 
1421 /*
1422  * Update the record of the current time in a context.
1423  */
1424 static void update_context_time(struct perf_event_context *ctx)
1425 {
1426         u64 now = perf_clock();
1427 
1428         ctx->time += now - ctx->timestamp;
1429         ctx->timestamp = now;
1430 }
1431 
1432 static u64 perf_event_time(struct perf_event *event)
1433 {
1434         struct perf_event_context *ctx = event->ctx;
1435 
1436         if (is_cgroup_event(event))
1437                 return perf_cgroup_event_time(event);
1438 
1439         return ctx ? ctx->time : 0;
1440 }
1441 
1442 static enum event_type_t get_event_type(struct perf_event *event)
1443 {
1444         struct perf_event_context *ctx = event->ctx;
1445         enum event_type_t event_type;
1446 
1447         lockdep_assert_held(&ctx->lock);
1448 
1449         /*
1450          * It's 'group type', really, because if our group leader is
1451          * pinned, so are we.
1452          */
1453         if (event->group_leader != event)
1454                 event = event->group_leader;
1455 
1456         event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE;
1457         if (!ctx->task)
1458                 event_type |= EVENT_CPU;
1459 
1460         return event_type;
1461 }
1462 
1463 static struct list_head *
1464 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
1465 {
1466         if (event->attr.pinned)
1467                 return &ctx->pinned_groups;
1468         else
1469                 return &ctx->flexible_groups;
1470 }
1471 
1472 /*
1473  * Add a event from the lists for its context.
1474  * Must be called with ctx->mutex and ctx->lock held.
1475  */
1476 static void
1477 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1478 {
1479         lockdep_assert_held(&ctx->lock);
1480 
1481         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1482         event->attach_state |= PERF_ATTACH_CONTEXT;
1483 
1484         event->tstamp = perf_event_time(event);
1485 
1486         /*
1487          * If we're a stand alone event or group leader, we go to the context
1488          * list, group events are kept attached to the group so that
1489          * perf_group_detach can, at all times, locate all siblings.
1490          */
1491         if (event->group_leader == event) {
1492                 struct list_head *list;
1493 
1494                 event->group_caps = event->event_caps;
1495 
1496                 list = ctx_group_list(event, ctx);
1497                 list_add_tail(&event->group_entry, list);
1498         }
1499 
1500         list_update_cgroup_event(event, ctx, true);
1501 
1502         list_add_rcu(&event->event_entry, &ctx->event_list);
1503         ctx->nr_events++;
1504         if (event->attr.inherit_stat)
1505                 ctx->nr_stat++;
1506 
1507         ctx->generation++;
1508 }
1509 
1510 /*
1511  * Initialize event state based on the perf_event_attr::disabled.
1512  */
1513 static inline void perf_event__state_init(struct perf_event *event)
1514 {
1515         event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1516                                               PERF_EVENT_STATE_INACTIVE;
1517 }
1518 
1519 static void __perf_event_read_size(struct perf_event *event, int nr_siblings)
1520 {
1521         int entry = sizeof(u64); /* value */
1522         int size = 0;
1523         int nr = 1;
1524 
1525         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1526                 size += sizeof(u64);
1527 
1528         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1529                 size += sizeof(u64);
1530 
1531         if (event->attr.read_format & PERF_FORMAT_ID)
1532                 entry += sizeof(u64);
1533 
1534         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1535                 nr += nr_siblings;
1536                 size += sizeof(u64);
1537         }
1538 
1539         size += entry * nr;
1540         event->read_size = size;
1541 }
1542 
1543 static void __perf_event_header_size(struct perf_event *event, u64 sample_type)
1544 {
1545         struct perf_sample_data *data;
1546         u16 size = 0;
1547 
1548         if (sample_type & PERF_SAMPLE_IP)
1549                 size += sizeof(data->ip);
1550 
1551         if (sample_type & PERF_SAMPLE_ADDR)
1552                 size += sizeof(data->addr);
1553 
1554         if (sample_type & PERF_SAMPLE_PERIOD)
1555                 size += sizeof(data->period);
1556 
1557         if (sample_type & PERF_SAMPLE_WEIGHT)
1558                 size += sizeof(data->weight);
1559 
1560         if (sample_type & PERF_SAMPLE_READ)
1561                 size += event->read_size;
1562 
1563         if (sample_type & PERF_SAMPLE_DATA_SRC)
1564                 size += sizeof(data->data_src.val);
1565 
1566         if (sample_type & PERF_SAMPLE_TRANSACTION)
1567                 size += sizeof(data->txn);
1568 
1569         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
1570                 size += sizeof(data->phys_addr);
1571 
1572         event->header_size = size;
1573 }
1574 
1575 /*
1576  * Called at perf_event creation and when events are attached/detached from a
1577  * group.
1578  */
1579 static void perf_event__header_size(struct perf_event *event)
1580 {
1581         __perf_event_read_size(event,
1582                                event->group_leader->nr_siblings);
1583         __perf_event_header_size(event, event->attr.sample_type);
1584 }
1585 
1586 static void perf_event__id_header_size(struct perf_event *event)
1587 {
1588         struct perf_sample_data *data;
1589         u64 sample_type = event->attr.sample_type;
1590         u16 size = 0;
1591 
1592         if (sample_type & PERF_SAMPLE_TID)
1593                 size += sizeof(data->tid_entry);
1594 
1595         if (sample_type & PERF_SAMPLE_TIME)
1596                 size += sizeof(data->time);
1597 
1598         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1599                 size += sizeof(data->id);
1600 
1601         if (sample_type & PERF_SAMPLE_ID)
1602                 size += sizeof(data->id);
1603 
1604         if (sample_type & PERF_SAMPLE_STREAM_ID)
1605                 size += sizeof(data->stream_id);
1606 
1607         if (sample_type & PERF_SAMPLE_CPU)
1608                 size += sizeof(data->cpu_entry);
1609 
1610         event->id_header_size = size;
1611 }
1612 
1613 static bool perf_event_validate_size(struct perf_event *event)
1614 {
1615         /*
1616          * The values computed here will be over-written when we actually
1617          * attach the event.
1618          */
1619         __perf_event_read_size(event, event->group_leader->nr_siblings + 1);
1620         __perf_event_header_size(event, event->attr.sample_type & ~PERF_SAMPLE_READ);
1621         perf_event__id_header_size(event);
1622 
1623         /*
1624          * Sum the lot; should not exceed the 64k limit we have on records.
1625          * Conservative limit to allow for callchains and other variable fields.
1626          */
1627         if (event->read_size + event->header_size +
1628             event->id_header_size + sizeof(struct perf_event_header) >= 16*1024)
1629                 return false;
1630 
1631         return true;
1632 }
1633 
1634 static void perf_group_attach(struct perf_event *event)
1635 {
1636         struct perf_event *group_leader = event->group_leader, *pos;
1637 
1638         lockdep_assert_held(&event->ctx->lock);
1639 
1640         /*
1641          * We can have double attach due to group movement in perf_event_open.
1642          */
1643         if (event->attach_state & PERF_ATTACH_GROUP)
1644                 return;
1645 
1646         event->attach_state |= PERF_ATTACH_GROUP;
1647 
1648         if (group_leader == event)
1649                 return;
1650 
1651         WARN_ON_ONCE(group_leader->ctx != event->ctx);
1652 
1653         group_leader->group_caps &= event->event_caps;
1654 
1655         list_add_tail(&event->group_entry, &group_leader->sibling_list);
1656         group_leader->nr_siblings++;
1657 
1658         perf_event__header_size(group_leader);
1659 
1660         list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
1661                 perf_event__header_size(pos);
1662 }
1663 
1664 /*
1665  * Remove a event from the lists for its context.
1666  * Must be called with ctx->mutex and ctx->lock held.
1667  */
1668 static void
1669 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1670 {
1671         WARN_ON_ONCE(event->ctx != ctx);
1672         lockdep_assert_held(&ctx->lock);
1673 
1674         /*
1675          * We can have double detach due to exit/hot-unplug + close.
1676          */
1677         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1678                 return;
1679 
1680         event->attach_state &= ~PERF_ATTACH_CONTEXT;
1681 
1682         list_update_cgroup_event(event, ctx, false);
1683 
1684         ctx->nr_events--;
1685         if (event->attr.inherit_stat)
1686                 ctx->nr_stat--;
1687 
1688         list_del_rcu(&event->event_entry);
1689 
1690         if (event->group_leader == event)
1691                 list_del_init(&event->group_entry);
1692 
1693         /*
1694          * If event was in error state, then keep it
1695          * that way, otherwise bogus counts will be
1696          * returned on read(). The only way to get out
1697          * of error state is by explicit re-enabling
1698          * of the event
1699          */
1700         if (event->state > PERF_EVENT_STATE_OFF)
1701                 perf_event_set_state(event, PERF_EVENT_STATE_OFF);
1702 
1703         ctx->generation++;
1704 }
1705 
1706 static void perf_group_detach(struct perf_event *event)
1707 {
1708         struct perf_event *sibling, *tmp;
1709         struct list_head *list = NULL;
1710 
1711         lockdep_assert_held(&event->ctx->lock);
1712 
1713         /*
1714          * We can have double detach due to exit/hot-unplug + close.
1715          */
1716         if (!(event->attach_state & PERF_ATTACH_GROUP))
1717                 return;
1718 
1719         event->attach_state &= ~PERF_ATTACH_GROUP;
1720 
1721         /*
1722          * If this is a sibling, remove it from its group.
1723          */
1724         if (event->group_leader != event) {
1725                 list_del_init(&event->group_entry);
1726                 event->group_leader->nr_siblings--;
1727                 goto out;
1728         }
1729 
1730         if (!list_empty(&event->group_entry))
1731                 list = &event->group_entry;
1732 
1733         /*
1734          * If this was a group event with sibling events then
1735          * upgrade the siblings to singleton events by adding them
1736          * to whatever list we are on.
1737          */
1738         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
1739                 if (list)
1740                         list_move_tail(&sibling->group_entry, list);
1741                 sibling->group_leader = sibling;
1742 
1743                 /* Inherit group flags from the previous leader */
1744                 sibling->group_caps = event->group_caps;
1745 
1746                 WARN_ON_ONCE(sibling->ctx != event->ctx);
1747         }
1748 
1749 out:
1750         perf_event__header_size(event->group_leader);
1751 
1752         list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
1753                 perf_event__header_size(tmp);
1754 }
1755 
1756 static bool is_orphaned_event(struct perf_event *event)
1757 {
1758         return event->state == PERF_EVENT_STATE_DEAD;
1759 }
1760 
1761 static inline int __pmu_filter_match(struct perf_event *event)
1762 {
1763         struct pmu *pmu = event->pmu;
1764         return pmu->filter_match ? pmu->filter_match(event) : 1;
1765 }
1766 
1767 /*
1768  * Check whether we should attempt to schedule an event group based on
1769  * PMU-specific filtering. An event group can consist of HW and SW events,
1770  * potentially with a SW leader, so we must check all the filters, to
1771  * determine whether a group is schedulable:
1772  */
1773 static inline int pmu_filter_match(struct perf_event *event)
1774 {
1775         struct perf_event *child;
1776 
1777         if (!__pmu_filter_match(event))
1778                 return 0;
1779 
1780         list_for_each_entry(child, &event->sibling_list, group_entry) {
1781                 if (!__pmu_filter_match(child))
1782                         return 0;
1783         }
1784 
1785         return 1;
1786 }
1787 
1788 static inline int
1789 event_filter_match(struct perf_event *event)
1790 {
1791         return (event->cpu == -1 || event->cpu == smp_processor_id()) &&
1792                perf_cgroup_match(event) && pmu_filter_match(event);
1793 }
1794 
1795 static void
1796 event_sched_out(struct perf_event *event,
1797                   struct perf_cpu_context *cpuctx,
1798                   struct perf_event_context *ctx)
1799 {
1800         enum perf_event_state state = PERF_EVENT_STATE_INACTIVE;
1801 
1802         WARN_ON_ONCE(event->ctx != ctx);
1803         lockdep_assert_held(&ctx->lock);
1804 
1805         if (event->state != PERF_EVENT_STATE_ACTIVE)
1806                 return;
1807 
1808         perf_pmu_disable(event->pmu);
1809 
1810         event->pmu->del(event, 0);
1811         event->oncpu = -1;
1812 
1813         if (event->pending_disable) {
1814                 event->pending_disable = 0;
1815                 state = PERF_EVENT_STATE_OFF;
1816         }
1817         perf_event_set_state(event, state);
1818 
1819         if (!is_software_event(event))
1820                 cpuctx->active_oncpu--;
1821         if (!--ctx->nr_active)
1822                 perf_event_ctx_deactivate(ctx);
1823         if (event->attr.freq && event->attr.sample_freq)
1824                 ctx->nr_freq--;
1825         if (event->attr.exclusive || !cpuctx->active_oncpu)
1826                 cpuctx->exclusive = 0;
1827 
1828         perf_pmu_enable(event->pmu);
1829 }
1830 
1831 static void
1832 group_sched_out(struct perf_event *group_event,
1833                 struct perf_cpu_context *cpuctx,
1834                 struct perf_event_context *ctx)
1835 {
1836         struct perf_event *event;
1837 
1838         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
1839                 return;
1840 
1841         perf_pmu_disable(ctx->pmu);
1842 
1843         event_sched_out(group_event, cpuctx, ctx);
1844 
1845         /*
1846          * Schedule out siblings (if any):
1847          */
1848         list_for_each_entry(event, &group_event->sibling_list, group_entry)
1849                 event_sched_out(event, cpuctx, ctx);
1850 
1851         perf_pmu_enable(ctx->pmu);
1852 
1853         if (group_event->attr.exclusive)
1854                 cpuctx->exclusive = 0;
1855 }
1856 
1857 #define DETACH_GROUP    0x01UL
1858 
1859 /*
1860  * Cross CPU call to remove a performance event
1861  *
1862  * We disable the event on the hardware level first. After that we
1863  * remove it from the context list.
1864  */
1865 static void
1866 __perf_remove_from_context(struct perf_event *event,
1867                            struct perf_cpu_context *cpuctx,
1868                            struct perf_event_context *ctx,
1869                            void *info)
1870 {
1871         unsigned long flags = (unsigned long)info;
1872 
1873         if (ctx->is_active & EVENT_TIME) {
1874                 update_context_time(ctx);
1875                 update_cgrp_time_from_cpuctx(cpuctx);
1876         }
1877 
1878         event_sched_out(event, cpuctx, ctx);
1879         if (flags & DETACH_GROUP)
1880                 perf_group_detach(event);
1881         list_del_event(event, ctx);
1882 
1883         if (!ctx->nr_events && ctx->is_active) {
1884                 ctx->is_active = 0;
1885                 if (ctx->task) {
1886                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
1887                         cpuctx->task_ctx = NULL;
1888                 }
1889         }
1890 }
1891 
1892 /*
1893  * Remove the event from a task's (or a CPU's) list of events.
1894  *
1895  * If event->ctx is a cloned context, callers must make sure that
1896  * every task struct that event->ctx->task could possibly point to
1897  * remains valid.  This is OK when called from perf_release since
1898  * that only calls us on the top-level context, which can't be a clone.
1899  * When called from perf_event_exit_task, it's OK because the
1900  * context has been detached from its task.
1901  */
1902 static void perf_remove_from_context(struct perf_event *event, unsigned long flags)
1903 {
1904         struct perf_event_context *ctx = event->ctx;
1905 
1906         lockdep_assert_held(&ctx->mutex);
1907 
1908         event_function_call(event, __perf_remove_from_context, (void *)flags);
1909 
1910         /*
1911          * The above event_function_call() can NO-OP when it hits
1912          * TASK_TOMBSTONE. In that case we must already have been detached
1913          * from the context (by perf_event_exit_event()) but the grouping
1914          * might still be in-tact.
1915          */
1916         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1917         if ((flags & DETACH_GROUP) &&
1918             (event->attach_state & PERF_ATTACH_GROUP)) {
1919                 /*
1920                  * Since in that case we cannot possibly be scheduled, simply
1921                  * detach now.
1922                  */
1923                 raw_spin_lock_irq(&ctx->lock);
1924                 perf_group_detach(event);
1925                 raw_spin_unlock_irq(&ctx->lock);
1926         }
1927 }
1928 
1929 /*
1930  * Cross CPU call to disable a performance event
1931  */
1932 static void __perf_event_disable(struct perf_event *event,
1933                                  struct perf_cpu_context *cpuctx,
1934                                  struct perf_event_context *ctx,
1935                                  void *info)
1936 {
1937         if (event->state < PERF_EVENT_STATE_INACTIVE)
1938                 return;
1939 
1940         if (ctx->is_active & EVENT_TIME) {
1941                 update_context_time(ctx);
1942                 update_cgrp_time_from_event(event);
1943         }
1944 
1945         if (event == event->group_leader)
1946                 group_sched_out(event, cpuctx, ctx);
1947         else
1948                 event_sched_out(event, cpuctx, ctx);
1949 
1950         perf_event_set_state(event, PERF_EVENT_STATE_OFF);
1951 }
1952 
1953 /*
1954  * Disable a event.
1955  *
1956  * If event->ctx is a cloned context, callers must make sure that
1957  * every task struct that event->ctx->task could possibly point to
1958  * remains valid.  This condition is satisifed when called through
1959  * perf_event_for_each_child or perf_event_for_each because they
1960  * hold the top-level event's child_mutex, so any descendant that
1961  * goes to exit will block in perf_event_exit_event().
1962  *
1963  * When called from perf_pending_event it's OK because event->ctx
1964  * is the current context on this CPU and preemption is disabled,
1965  * hence we can't get into perf_event_task_sched_out for this context.
1966  */
1967 static void _perf_event_disable(struct perf_event *event)
1968 {
1969         struct perf_event_context *ctx = event->ctx;
1970 
1971         raw_spin_lock_irq(&ctx->lock);
1972         if (event->state <= PERF_EVENT_STATE_OFF) {
1973                 raw_spin_unlock_irq(&ctx->lock);
1974                 return;
1975         }
1976         raw_spin_unlock_irq(&ctx->lock);
1977 
1978         event_function_call(event, __perf_event_disable, NULL);
1979 }
1980 
1981 void perf_event_disable_local(struct perf_event *event)
1982 {
1983         event_function_local(event, __perf_event_disable, NULL);
1984 }
1985 
1986 /*
1987  * Strictly speaking kernel users cannot create groups and therefore this
1988  * interface does not need the perf_event_ctx_lock() magic.
1989  */
1990 void perf_event_disable(struct perf_event *event)
1991 {
1992         struct perf_event_context *ctx;
1993 
1994         ctx = perf_event_ctx_lock(event);
1995         _perf_event_disable(event);
1996         perf_event_ctx_unlock(event, ctx);
1997 }
1998 EXPORT_SYMBOL_GPL(perf_event_disable);
1999 
2000 void perf_event_disable_inatomic(struct perf_event *event)
2001 {
2002         event->pending_disable = 1;
2003         irq_work_queue(&event->pending);
2004 }
2005 
2006 static void perf_set_shadow_time(struct perf_event *event,
2007                                  struct perf_event_context *ctx)
2008 {
2009         /*
2010          * use the correct time source for the time snapshot
2011          *
2012          * We could get by without this by leveraging the
2013          * fact that to get to this function, the caller
2014          * has most likely already called update_context_time()
2015          * and update_cgrp_time_xx() and thus both timestamp
2016          * are identical (or very close). Given that tstamp is,
2017          * already adjusted for cgroup, we could say that:
2018          *    tstamp - ctx->timestamp
2019          * is equivalent to
2020          *    tstamp - cgrp->timestamp.
2021          *
2022          * Then, in perf_output_read(), the calculation would
2023          * work with no changes because:
2024          * - event is guaranteed scheduled in
2025          * - no scheduled out in between
2026          * - thus the timestamp would be the same
2027          *
2028          * But this is a bit hairy.
2029          *
2030          * So instead, we have an explicit cgroup call to remain
2031          * within the time time source all along. We believe it
2032          * is cleaner and simpler to understand.
2033          */
2034         if (is_cgroup_event(event))
2035                 perf_cgroup_set_shadow_time(event, event->tstamp);
2036         else
2037                 event->shadow_ctx_time = event->tstamp - ctx->timestamp;
2038 }
2039 
2040 #define MAX_INTERRUPTS (~0ULL)
2041 
2042 static void perf_log_throttle(struct perf_event *event, int enable);
2043 static void perf_log_itrace_start(struct perf_event *event);
2044 
2045 static int
2046 event_sched_in(struct perf_event *event,
2047                  struct perf_cpu_context *cpuctx,
2048                  struct perf_event_context *ctx)
2049 {
2050         int ret = 0;
2051 
2052         lockdep_assert_held(&ctx->lock);
2053 
2054         if (event->state <= PERF_EVENT_STATE_OFF)
2055                 return 0;
2056 
2057         WRITE_ONCE(event->oncpu, smp_processor_id());
2058         /*
2059          * Order event::oncpu write to happen before the ACTIVE state is
2060          * visible. This allows perf_event_{stop,read}() to observe the correct
2061          * ->oncpu if it sees ACTIVE.
2062          */
2063         smp_wmb();
2064         perf_event_set_state(event, PERF_EVENT_STATE_ACTIVE);
2065 
2066         /*
2067          * Unthrottle events, since we scheduled we might have missed several
2068          * ticks already, also for a heavily scheduling task there is little
2069          * guarantee it'll get a tick in a timely manner.
2070          */
2071         if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
2072                 perf_log_throttle(event, 1);
2073                 event->hw.interrupts = 0;
2074         }
2075 
2076         perf_pmu_disable(event->pmu);
2077 
2078         perf_set_shadow_time(event, ctx);
2079 
2080         perf_log_itrace_start(event);
2081 
2082         if (event->pmu->add(event, PERF_EF_START)) {
2083                 perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2084                 event->oncpu = -1;
2085                 ret = -EAGAIN;
2086                 goto out;
2087         }
2088 
2089         if (!is_software_event(event))
2090                 cpuctx->active_oncpu++;
2091         if (!ctx->nr_active++)
2092                 perf_event_ctx_activate(ctx);
2093         if (event->attr.freq && event->attr.sample_freq)
2094                 ctx->nr_freq++;
2095 
2096         if (event->attr.exclusive)
2097                 cpuctx->exclusive = 1;
2098 
2099 out:
2100         perf_pmu_enable(event->pmu);
2101 
2102         return ret;
2103 }
2104 
2105 static int
2106 group_sched_in(struct perf_event *group_event,
2107                struct perf_cpu_context *cpuctx,
2108                struct perf_event_context *ctx)
2109 {
2110         struct perf_event *event, *partial_group = NULL;
2111         struct pmu *pmu = ctx->pmu;
2112 
2113         if (group_event->state == PERF_EVENT_STATE_OFF)
2114                 return 0;
2115 
2116         pmu->start_txn(pmu, PERF_PMU_TXN_ADD);
2117 
2118         if (event_sched_in(group_event, cpuctx, ctx)) {
2119                 pmu->cancel_txn(pmu);
2120                 perf_mux_hrtimer_restart(cpuctx);
2121                 return -EAGAIN;
2122         }
2123 
2124         /*
2125          * Schedule in siblings as one group (if any):
2126          */
2127         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
2128                 if (event_sched_in(event, cpuctx, ctx)) {
2129                         partial_group = event;
2130                         goto group_error;
2131                 }
2132         }
2133 
2134         if (!pmu->commit_txn(pmu))
2135                 return 0;
2136 
2137 group_error:
2138         /*
2139          * Groups can be scheduled in as one unit only, so undo any
2140          * partial group before returning:
2141          * The events up to the failed event are scheduled out normally.
2142          */
2143         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
2144                 if (event == partial_group)
2145                         break;
2146 
2147                 event_sched_out(event, cpuctx, ctx);
2148         }
2149         event_sched_out(group_event, cpuctx, ctx);
2150 
2151         pmu->cancel_txn(pmu);
2152 
2153         perf_mux_hrtimer_restart(cpuctx);
2154 
2155         return -EAGAIN;
2156 }
2157 
2158 /*
2159  * Work out whether we can put this event group on the CPU now.
2160  */
2161 static int group_can_go_on(struct perf_event *event,
2162                            struct perf_cpu_context *cpuctx,
2163                            int can_add_hw)
2164 {
2165         /*
2166          * Groups consisting entirely of software events can always go on.
2167          */
2168         if (event->group_caps & PERF_EV_CAP_SOFTWARE)
2169                 return 1;
2170         /*
2171          * If an exclusive group is already on, no other hardware
2172          * events can go on.
2173          */
2174         if (cpuctx->exclusive)
2175                 return 0;
2176         /*
2177          * If this group is exclusive and there are already
2178          * events on the CPU, it can't go on.
2179          */
2180         if (event->attr.exclusive && cpuctx->active_oncpu)
2181                 return 0;
2182         /*
2183          * Otherwise, try to add it if all previous groups were able
2184          * to go on.
2185          */
2186         return can_add_hw;
2187 }
2188 
2189 static void add_event_to_ctx(struct perf_event *event,
2190                                struct perf_event_context *ctx)
2191 {
2192         list_add_event(event, ctx);
2193         perf_group_attach(event);
2194 }
2195 
2196 static void ctx_sched_out(struct perf_event_context *ctx,
2197                           struct perf_cpu_context *cpuctx,
2198                           enum event_type_t event_type);
2199 static void
2200 ctx_sched_in(struct perf_event_context *ctx,
2201              struct perf_cpu_context *cpuctx,
2202              enum event_type_t event_type,
2203              struct task_struct *task);
2204 
2205 static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
2206                                struct perf_event_context *ctx,
2207                                enum event_type_t event_type)
2208 {
2209         if (!cpuctx->task_ctx)
2210                 return;
2211 
2212         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2213                 return;
2214 
2215         ctx_sched_out(ctx, cpuctx, event_type);
2216 }
2217 
2218 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2219                                 struct perf_event_context *ctx,
2220                                 struct task_struct *task)
2221 {
2222         cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2223         if (ctx)
2224                 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2225         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2226         if (ctx)
2227                 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2228 }
2229 
2230 /*
2231  * We want to maintain the following priority of scheduling:
2232  *  - CPU pinned (EVENT_CPU | EVENT_PINNED)
2233  *  - task pinned (EVENT_PINNED)
2234  *  - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
2235  *  - task flexible (EVENT_FLEXIBLE).
2236  *
2237  * In order to avoid unscheduling and scheduling back in everything every
2238  * time an event is added, only do it for the groups of equal priority and
2239  * below.
2240  *
2241  * This can be called after a batch operation on task events, in which case
2242  * event_type is a bit mask of the types of events involved. For CPU events,
2243  * event_type is only either EVENT_PINNED or EVENT_FLEXIBLE.
2244  */
2245 static void ctx_resched(struct perf_cpu_context *cpuctx,
2246                         struct perf_event_context *task_ctx,
2247                         enum event_type_t event_type)
2248 {
2249         enum event_type_t ctx_event_type;
2250         bool cpu_event = !!(event_type & EVENT_CPU);
2251 
2252         /*
2253          * If pinned groups are involved, flexible groups also need to be
2254          * scheduled out.
2255          */
2256         if (event_type & EVENT_PINNED)
2257                 event_type |= EVENT_FLEXIBLE;
2258 
2259         ctx_event_type = event_type & EVENT_ALL;
2260 
2261         perf_pmu_disable(cpuctx->ctx.pmu);
2262         if (task_ctx)
2263                 task_ctx_sched_out(cpuctx, task_ctx, event_type);
2264 
2265         /*
2266          * Decide which cpu ctx groups to schedule out based on the types
2267          * of events that caused rescheduling:
2268          *  - EVENT_CPU: schedule out corresponding groups;
2269          *  - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
2270          *  - otherwise, do nothing more.
2271          */
2272         if (cpu_event)
2273                 cpu_ctx_sched_out(cpuctx, ctx_event_type);
2274         else if (ctx_event_type & EVENT_PINNED)
2275                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2276 
2277         perf_event_sched_in(cpuctx, task_ctx, current);
2278         perf_pmu_enable(cpuctx->ctx.pmu);
2279 }
2280 
2281 /*
2282  * Cross CPU call to install and enable a performance event
2283  *
2284  * Very similar to remote_function() + event_function() but cannot assume that
2285  * things like ctx->is_active and cpuctx->task_ctx are set.
2286  */
2287 static int  __perf_install_in_context(void *info)
2288 {
2289         struct perf_event *event = info;
2290         struct perf_event_context *ctx = event->ctx;
2291         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2292         struct perf_event_context *task_ctx = cpuctx->task_ctx;
2293         bool reprogram = true;
2294         int ret = 0;
2295 
2296         raw_spin_lock(&cpuctx->ctx.lock);
2297         if (ctx->task) {
2298                 raw_spin_lock(&ctx->lock);
2299                 task_ctx = ctx;
2300 
2301                 reprogram = (ctx->task == current);
2302 
2303                 /*
2304                  * If the task is running, it must be running on this CPU,
2305                  * otherwise we cannot reprogram things.
2306                  *
2307                  * If its not running, we don't care, ctx->lock will
2308                  * serialize against it becoming runnable.
2309                  */
2310                 if (task_curr(ctx->task) && !reprogram) {
2311                         ret = -ESRCH;
2312                         goto unlock;
2313                 }
2314 
2315                 WARN_ON_ONCE(reprogram && cpuctx->task_ctx && cpuctx->task_ctx != ctx);
2316         } else if (task_ctx) {
2317                 raw_spin_lock(&task_ctx->lock);
2318         }
2319 
2320         if (reprogram) {
2321                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2322                 add_event_to_ctx(event, ctx);
2323                 ctx_resched(cpuctx, task_ctx, get_event_type(event));
2324         } else {
2325                 add_event_to_ctx(event, ctx);
2326         }
2327 
2328 unlock:
2329         perf_ctx_unlock(cpuctx, task_ctx);
2330 
2331         return ret;
2332 }
2333 
2334 /*
2335  * Attach a performance event to a context.
2336  *
2337  * Very similar to event_function_call, see comment there.
2338  */
2339 static void
2340 perf_install_in_context(struct perf_event_context *ctx,
2341                         struct perf_event *event,
2342                         int cpu)
2343 {
2344         struct task_struct *task = READ_ONCE(ctx->task);
2345 
2346         lockdep_assert_held(&ctx->mutex);
2347 
2348         if (event->cpu != -1)
2349                 event->cpu = cpu;
2350 
2351         /*
2352          * Ensures that if we can observe event->ctx, both the event and ctx
2353          * will be 'complete'. See perf_iterate_sb_cpu().
2354          */
2355         smp_store_release(&event->ctx, ctx);
2356 
2357         if (!task) {
2358                 cpu_function_call(cpu, __perf_install_in_context, event);
2359                 return;
2360         }
2361 
2362         /*
2363          * Should not happen, we validate the ctx is still alive before calling.
2364          */
2365         if (WARN_ON_ONCE(task == TASK_TOMBSTONE))
2366                 return;
2367 
2368         /*
2369          * Installing events is tricky because we cannot rely on ctx->is_active
2370          * to be set in case this is the nr_events 0 -> 1 transition.
2371          *
2372          * Instead we use task_curr(), which tells us if the task is running.
2373          * However, since we use task_curr() outside of rq::lock, we can race
2374          * against the actual state. This means the result can be wrong.
2375          *
2376          * If we get a false positive, we retry, this is harmless.
2377          *
2378          * If we get a false negative, things are complicated. If we are after
2379          * perf_event_context_sched_in() ctx::lock will serialize us, and the
2380          * value must be correct. If we're before, it doesn't matter since
2381          * perf_event_context_sched_in() will program the counter.
2382          *
2383          * However, this hinges on the remote context switch having observed
2384          * our task->perf_event_ctxp[] store, such that it will in fact take
2385          * ctx::lock in perf_event_context_sched_in().
2386          *
2387          * We do this by task_function_call(), if the IPI fails to hit the task
2388          * we know any future context switch of task must see the
2389          * perf_event_ctpx[] store.
2390          */
2391 
2392         /*
2393          * This smp_mb() orders the task->perf_event_ctxp[] store with the
2394          * task_cpu() load, such that if the IPI then does not find the task
2395          * running, a future context switch of that task must observe the
2396          * store.
2397          */
2398         smp_mb();
2399 again:
2400         if (!task_function_call(task, __perf_install_in_context, event))
2401                 return;
2402 
2403         raw_spin_lock_irq(&ctx->lock);
2404         task = ctx->task;
2405         if (WARN_ON_ONCE(task == TASK_TOMBSTONE)) {
2406                 /*
2407                  * Cannot happen because we already checked above (which also
2408                  * cannot happen), and we hold ctx->mutex, which serializes us
2409                  * against perf_event_exit_task_context().
2410                  */
2411                 raw_spin_unlock_irq(&ctx->lock);
2412                 return;
2413         }
2414         /*
2415          * If the task is not running, ctx->lock will avoid it becoming so,
2416          * thus we can safely install the event.
2417          */
2418         if (task_curr(task)) {
2419                 raw_spin_unlock_irq(&ctx->lock);
2420                 goto again;
2421         }
2422         add_event_to_ctx(event, ctx);
2423         raw_spin_unlock_irq(&ctx->lock);
2424 }
2425 
2426 /*
2427  * Cross CPU call to enable a performance event
2428  */
2429 static void __perf_event_enable(struct perf_event *event,
2430                                 struct perf_cpu_context *cpuctx,
2431                                 struct perf_event_context *ctx,
2432                                 void *info)
2433 {
2434         struct perf_event *leader = event->group_leader;
2435         struct perf_event_context *task_ctx;
2436 
2437         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2438             event->state <= PERF_EVENT_STATE_ERROR)
2439                 return;
2440 
2441         if (ctx->is_active)
2442                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2443 
2444         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2445 
2446         if (!ctx->is_active)
2447                 return;
2448 
2449         if (!event_filter_match(event)) {
2450                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2451                 return;
2452         }
2453 
2454         /*
2455          * If the event is in a group and isn't the group leader,
2456          * then don't put it on unless the group is on.
2457          */
2458         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) {
2459                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2460                 return;
2461         }
2462 
2463         task_ctx = cpuctx->task_ctx;
2464         if (ctx->task)
2465                 WARN_ON_ONCE(task_ctx != ctx);
2466 
2467         ctx_resched(cpuctx, task_ctx, get_event_type(event));
2468 }
2469 
2470 /*
2471  * Enable a event.
2472  *
2473  * If event->ctx is a cloned context, callers must make sure that
2474  * every task struct that event->ctx->task could possibly point to
2475  * remains valid.  This condition is satisfied when called through
2476  * perf_event_for_each_child or perf_event_for_each as described
2477  * for perf_event_disable.
2478  */
2479 static void _perf_event_enable(struct perf_event *event)
2480 {
2481         struct perf_event_context *ctx = event->ctx;
2482 
2483         raw_spin_lock_irq(&ctx->lock);
2484         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2485             event->state <  PERF_EVENT_STATE_ERROR) {
2486                 raw_spin_unlock_irq(&ctx->lock);
2487                 return;
2488         }
2489 
2490         /*
2491          * If the event is in error state, clear that first.
2492          *
2493          * That way, if we see the event in error state below, we know that it
2494          * has gone back into error state, as distinct from the task having
2495          * been scheduled away before the cross-call arrived.
2496          */
2497         if (event->state == PERF_EVENT_STATE_ERROR)
2498                 event->state = PERF_EVENT_STATE_OFF;
2499         raw_spin_unlock_irq(&ctx->lock);
2500 
2501         event_function_call(event, __perf_event_enable, NULL);
2502 }
2503 
2504 /*
2505  * See perf_event_disable();
2506  */
2507 void perf_event_enable(struct perf_event *event)
2508 {
2509         struct perf_event_context *ctx;
2510 
2511         ctx = perf_event_ctx_lock(event);
2512         _perf_event_enable(event);
2513         perf_event_ctx_unlock(event, ctx);
2514 }
2515 EXPORT_SYMBOL_GPL(perf_event_enable);
2516 
2517 struct stop_event_data {
2518         struct perf_event       *event;
2519         unsigned int            restart;
2520 };
2521 
2522 static int __perf_event_stop(void *info)
2523 {
2524         struct stop_event_data *sd = info;
2525         struct perf_event *event = sd->event;
2526 
2527         /* if it's already INACTIVE, do nothing */
2528         if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2529                 return 0;
2530 
2531         /* matches smp_wmb() in event_sched_in() */
2532         smp_rmb();
2533 
2534         /*
2535          * There is a window with interrupts enabled before we get here,
2536          * so we need to check again lest we try to stop another CPU's event.
2537          */
2538         if (READ_ONCE(event->oncpu) != smp_processor_id())
2539                 return -EAGAIN;
2540 
2541         event->pmu->stop(event, PERF_EF_UPDATE);
2542 
2543         /*
2544          * May race with the actual stop (through perf_pmu_output_stop()),
2545          * but it is only used for events with AUX ring buffer, and such
2546          * events will refuse to restart because of rb::aux_mmap_count==0,
2547          * see comments in perf_aux_output_begin().
2548          *
2549          * Since this is happening on a event-local CPU, no trace is lost
2550          * while restarting.
2551          */
2552         if (sd->restart)
2553                 event->pmu->start(event, 0);
2554 
2555         return 0;
2556 }
2557 
2558 static int perf_event_stop(struct perf_event *event, int restart)
2559 {
2560         struct stop_event_data sd = {
2561                 .event          = event,
2562                 .restart        = restart,
2563         };
2564         int ret = 0;
2565 
2566         do {
2567                 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2568                         return 0;
2569 
2570                 /* matches smp_wmb() in event_sched_in() */
2571                 smp_rmb();
2572 
2573                 /*
2574                  * We only want to restart ACTIVE events, so if the event goes
2575                  * inactive here (event->oncpu==-1), there's nothing more to do;
2576                  * fall through with ret==-ENXIO.
2577                  */
2578                 ret = cpu_function_call(READ_ONCE(event->oncpu),
2579                                         __perf_event_stop, &sd);
2580         } while (ret == -EAGAIN);
2581 
2582         return ret;
2583 }
2584 
2585 /*
2586  * In order to contain the amount of racy and tricky in the address filter
2587  * configuration management, it is a two part process:
2588  *
2589  * (p1) when userspace mappings change as a result of (1) or (2) or (3) below,
2590  *      we update the addresses of corresponding vmas in
2591  *      event::addr_filters_offs array and bump the event::addr_filters_gen;
2592  * (p2) when an event is scheduled in (pmu::add), it calls
2593  *      perf_event_addr_filters_sync() which calls pmu::addr_filters_sync()
2594  *      if the generation has changed since the previous call.
2595  *
2596  * If (p1) happens while the event is active, we restart it to force (p2).
2597  *
2598  * (1) perf_addr_filters_apply(): adjusting filters' offsets based on
2599  *     pre-existing mappings, called once when new filters arrive via SET_FILTER
2600  *     ioctl;
2601  * (2) perf_addr_filters_adjust(): adjusting filters' offsets based on newly
2602  *     registered mapping, called for every new mmap(), with mm::mmap_sem down
2603  *     for reading;
2604  * (3) perf_event_addr_filters_exec(): clearing filters' offsets in the process
2605  *     of exec.
2606  */
2607 void perf_event_addr_filters_sync(struct perf_event *event)
2608 {
2609         struct perf_addr_filters_head *ifh = perf_event_addr_filters(event);
2610 
2611         if (!has_addr_filter(event))
2612                 return;
2613 
2614         raw_spin_lock(&ifh->lock);
2615         if (event->addr_filters_gen != event->hw.addr_filters_gen) {
2616                 event->pmu->addr_filters_sync(event);
2617                 event->hw.addr_filters_gen = event->addr_filters_gen;
2618         }
2619         raw_spin_unlock(&ifh->lock);
2620 }
2621 EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
2622 
2623 static int _perf_event_refresh(struct perf_event *event, int refresh)
2624 {
2625         /*
2626          * not supported on inherited events
2627          */
2628         if (event->attr.inherit || !is_sampling_event(event))
2629                 return -EINVAL;
2630 
2631         atomic_add(refresh, &event->event_limit);
2632         _perf_event_enable(event);
2633 
2634         return 0;
2635 }
2636 
2637 /*
2638  * See perf_event_disable()
2639  */
2640 int perf_event_refresh(struct perf_event *event, int refresh)
2641 {
2642         struct perf_event_context *ctx;
2643         int ret;
2644 
2645         ctx = perf_event_ctx_lock(event);
2646         ret = _perf_event_refresh(event, refresh);
2647         perf_event_ctx_unlock(event, ctx);
2648 
2649         return ret;
2650 }
2651 EXPORT_SYMBOL_GPL(perf_event_refresh);
2652 
2653 static void ctx_sched_out(struct perf_event_context *ctx,
2654                           struct perf_cpu_context *cpuctx,
2655                           enum event_type_t event_type)
2656 {
2657         int is_active = ctx->is_active;
2658         struct perf_event *event;
2659 
2660         lockdep_assert_held(&ctx->lock);
2661 
2662         if (likely(!ctx->nr_events)) {
2663                 /*
2664                  * See __perf_remove_from_context().
2665                  */
2666                 WARN_ON_ONCE(ctx->is_active);
2667                 if (ctx->task)
2668                         WARN_ON_ONCE(cpuctx->task_ctx);
2669                 return;
2670         }
2671 
2672         ctx->is_active &= ~event_type;
2673         if (!(ctx->is_active & EVENT_ALL))
2674                 ctx->is_active = 0;
2675 
2676         if (ctx->task) {
2677                 WARN_ON_ONCE(cpuctx->task_ctx != ctx);
2678                 if (!ctx->is_active)
2679                         cpuctx->task_ctx = NULL;
2680         }
2681 
2682         /*
2683          * Always update time if it was set; not only when it changes.
2684          * Otherwise we can 'forget' to update time for any but the last
2685          * context we sched out. For example:
2686          *
2687          *   ctx_sched_out(.event_type = EVENT_FLEXIBLE)
2688          *   ctx_sched_out(.event_type = EVENT_PINNED)
2689          *
2690          * would only update time for the pinned events.
2691          */
2692         if (is_active & EVENT_TIME) {
2693                 /* update (and stop) ctx time */
2694                 update_context_time(ctx);
2695                 update_cgrp_time_from_cpuctx(cpuctx);
2696         }
2697 
2698         is_active ^= ctx->is_active; /* changed bits */
2699 
2700         if (!ctx->nr_active || !(is_active & EVENT_ALL))
2701                 return;
2702 
2703         perf_pmu_disable(ctx->pmu);
2704         if (is_active & EVENT_PINNED) {
2705                 list_for_each_entry(event, &ctx->pinned_groups, group_entry)
2706                         group_sched_out(event, cpuctx, ctx);
2707         }
2708 
2709         if (is_active & EVENT_FLEXIBLE) {
2710                 list_for_each_entry(event, &ctx->flexible_groups, group_entry)
2711                         group_sched_out(event, cpuctx, ctx);
2712         }
2713         perf_pmu_enable(ctx->pmu);
2714 }
2715 
2716 /*
2717  * Test whether two contexts are equivalent, i.e. whether they have both been
2718  * cloned from the same version of the same context.
2719  *
2720  * Equivalence is measured using a generation number in the context that is
2721  * incremented on each modification to it; see unclone_ctx(), list_add_event()
2722  * and list_del_event().
2723  */
2724 static int context_equiv(struct perf_event_context *ctx1,
2725                          struct perf_event_context *ctx2)
2726 {
2727         lockdep_assert_held(&ctx1->lock);
2728         lockdep_assert_held(&ctx2->lock);
2729 
2730         /* Pinning disables the swap optimization */
2731         if (ctx1->pin_count || ctx2->pin_count)
2732                 return 0;
2733 
2734         /* If ctx1 is the parent of ctx2 */
2735         if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2736                 return 1;
2737 
2738         /* If ctx2 is the parent of ctx1 */
2739         if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2740                 return 1;
2741 
2742         /*
2743          * If ctx1 and ctx2 have the same parent; we flatten the parent
2744          * hierarchy, see perf_event_init_context().
2745          */
2746         if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2747                         ctx1->parent_gen == ctx2->parent_gen)
2748                 return 1;
2749 
2750         /* Unmatched */
2751         return 0;
2752 }
2753 
2754 static void __perf_event_sync_stat(struct perf_event *event,
2755                                      struct perf_event *next_event)
2756 {
2757         u64 value;
2758 
2759         if (!event->attr.inherit_stat)
2760                 return;
2761 
2762         /*
2763          * Update the event value, we cannot use perf_event_read()
2764          * because we're in the middle of a context switch and have IRQs
2765          * disabled, which upsets smp_call_function_single(), however
2766          * we know the event must be on the current CPU, therefore we
2767          * don't need to use it.
2768          */
2769         if (event->state == PERF_EVENT_STATE_ACTIVE)
2770                 event->pmu->read(event);
2771 
2772         perf_event_update_time(event);
2773 
2774         /*
2775          * In order to keep per-task stats reliable we need to flip the event
2776          * values when we flip the contexts.
2777          */
2778         value = local64_read(&next_event->count);
2779         value = local64_xchg(&event->count, value);
2780         local64_set(&next_event->count, value);
2781 
2782         swap(event->total_time_enabled, next_event->total_time_enabled);
2783         swap(event->total_time_running, next_event->total_time_running);
2784 
2785         /*
2786          * Since we swizzled the values, update the user visible data too.
2787          */
2788         perf_event_update_userpage(event);
2789         perf_event_update_userpage(next_event);
2790 }
2791 
2792 static void perf_event_sync_stat(struct perf_event_context *ctx,
2793                                    struct perf_event_context *next_ctx)
2794 {
2795         struct perf_event *event, *next_event;
2796 
2797         if (!ctx->nr_stat)
2798                 return;
2799 
2800         update_context_time(ctx);
2801 
2802         event = list_first_entry(&ctx->event_list,
2803                                    struct perf_event, event_entry);
2804 
2805         next_event = list_first_entry(&next_ctx->event_list,
2806                                         struct perf_event, event_entry);
2807 
2808         while (&event->event_entry != &ctx->event_list &&
2809                &next_event->event_entry != &next_ctx->event_list) {
2810 
2811                 __perf_event_sync_stat(event, next_event);
2812 
2813                 event = list_next_entry(event, event_entry);
2814                 next_event = list_next_entry(next_event, event_entry);
2815         }
2816 }
2817 
2818 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
2819                                          struct task_struct *next)
2820 {
2821         struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
2822         struct perf_event_context *next_ctx;
2823         struct perf_event_context *parent, *next_parent;
2824         struct perf_cpu_context *cpuctx;
2825         int do_switch = 1;
2826 
2827         if (likely(!ctx))
2828                 return;
2829 
2830         cpuctx = __get_cpu_context(ctx);
2831         if (!cpuctx->task_ctx)
2832                 return;
2833 
2834         rcu_read_lock();
2835         next_ctx = next->perf_event_ctxp[ctxn];
2836         if (!next_ctx)
2837                 goto unlock;
2838 
2839         parent = rcu_dereference(ctx->parent_ctx);
2840         next_parent = rcu_dereference(next_ctx->parent_ctx);
2841 
2842         /* If neither context have a parent context; they cannot be clones. */
2843         if (!parent && !next_parent)
2844                 goto unlock;
2845 
2846         if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
2847                 /*
2848                  * Looks like the two contexts are clones, so we might be
2849                  * able to optimize the context switch.  We lock both
2850                  * contexts and check that they are clones under the
2851                  * lock (including re-checking that neither has been
2852                  * uncloned in the meantime).  It doesn't matter which
2853                  * order we take the locks because no other cpu could
2854                  * be trying to lock both of these tasks.
2855                  */
2856                 raw_spin_lock(&ctx->lock);
2857                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
2858                 if (context_equiv(ctx, next_ctx)) {
2859                         WRITE_ONCE(ctx->task, next);
2860                         WRITE_ONCE(next_ctx->task, task);
2861 
2862                         swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
2863 
2864                         /*
2865                          * RCU_INIT_POINTER here is safe because we've not
2866                          * modified the ctx and the above modification of
2867                          * ctx->task and ctx->task_ctx_data are immaterial
2868                          * since those values are always verified under
2869                          * ctx->lock which we're now holding.
2870                          */
2871                         RCU_INIT_POINTER(task->perf_event_ctxp[ctxn], next_ctx);
2872                         RCU_INIT_POINTER(next->perf_event_ctxp[ctxn], ctx);
2873 
2874                         do_switch = 0;
2875 
2876                         perf_event_sync_stat(ctx, next_ctx);
2877                 }
2878                 raw_spin_unlock(&next_ctx->lock);
2879                 raw_spin_unlock(&ctx->lock);
2880         }
2881 unlock:
2882         rcu_read_unlock();
2883 
2884         if (do_switch) {
2885                 raw_spin_lock(&ctx->lock);
2886                 task_ctx_sched_out(cpuctx, ctx, EVENT_ALL);
2887                 raw_spin_unlock(&ctx->lock);
2888         }
2889 }
2890 
2891 static DEFINE_PER_CPU(struct list_head, sched_cb_list);
2892 
2893 void perf_sched_cb_dec(struct pmu *pmu)
2894 {
2895         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2896 
2897         this_cpu_dec(perf_sched_cb_usages);
2898 
2899         if (!--cpuctx->sched_cb_usage)
2900                 list_del(&cpuctx->sched_cb_entry);
2901 }
2902 
2903 
2904 void perf_sched_cb_inc(struct pmu *pmu)
2905 {
2906         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
2907 
2908         if (!cpuctx->sched_cb_usage++)
2909                 list_add(&cpuctx->sched_cb_entry, this_cpu_ptr(&sched_cb_list));
2910 
2911         this_cpu_inc(perf_sched_cb_usages);
2912 }
2913 
2914 /*
2915  * This function provides the context switch callback to the lower code
2916  * layer. It is invoked ONLY when the context switch callback is enabled.
2917  *
2918  * This callback is relevant even to per-cpu events; for example multi event
2919  * PEBS requires this to provide PID/TID information. This requires we flush
2920  * all queued PEBS records before we context switch to a new task.
2921  */
2922 static void perf_pmu_sched_task(struct task_struct *prev,
2923                                 struct task_struct *next,
2924                                 bool sched_in)
2925 {
2926         struct perf_cpu_context *cpuctx;
2927         struct pmu *pmu;
2928 
2929         if (prev == next)
2930                 return;
2931 
2932         list_for_each_entry(cpuctx, this_cpu_ptr(&sched_cb_list), sched_cb_entry) {
2933                 pmu = cpuctx->ctx.pmu; /* software PMUs will not have sched_task */
2934 
2935                 if (WARN_ON_ONCE(!pmu->sched_task))
2936                         continue;
2937 
2938                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
2939                 perf_pmu_disable(pmu);
2940 
2941                 pmu->sched_task(cpuctx->task_ctx, sched_in);
2942 
2943                 perf_pmu_enable(pmu);
2944                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
2945         }
2946 }
2947 
2948 static void perf_event_switch(struct task_struct *task,
2949                               struct task_struct *next_prev, bool sched_in);
2950 
2951 #define for_each_task_context_nr(ctxn)                                  \
2952         for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
2953 
2954 /*
2955  * Called from scheduler to remove the events of the current task,
2956  * with interrupts disabled.
2957  *
2958  * We stop each event and update the event value in event->count.
2959  *
2960  * This does not protect us against NMI, but disable()
2961  * sets the disabled bit in the control field of event _before_
2962  * accessing the event control register. If a NMI hits, then it will
2963  * not restart the event.
2964  */
2965 void __perf_event_task_sched_out(struct task_struct *task,
2966                                  struct task_struct *next)
2967 {
2968         int ctxn;
2969 
2970         if (__this_cpu_read(perf_sched_cb_usages))
2971                 perf_pmu_sched_task(task, next, false);
2972 
2973         if (atomic_read(&nr_switch_events))
2974                 perf_event_switch(task, next, false);
2975 
2976         for_each_task_context_nr(ctxn)
2977                 perf_event_context_sched_out(task, ctxn, next);
2978 
2979         /*
2980          * if cgroup events exist on this CPU, then we need
2981          * to check if we have to switch out PMU state.
2982          * cgroup event are system-wide mode only
2983          */
2984         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
2985                 perf_cgroup_sched_out(task, next);
2986 }
2987 
2988 /*
2989  * Called with IRQs disabled
2990  */
2991 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
2992                               enum event_type_t event_type)
2993 {
2994         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
2995 }
2996 
2997 static void
2998 ctx_pinned_sched_in(struct perf_event_context *ctx,
2999                     struct perf_cpu_context *cpuctx)
3000 {
3001         struct perf_event *event;
3002 
3003         list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
3004                 if (event->state <= PERF_EVENT_STATE_OFF)
3005                         continue;
3006                 if (!event_filter_match(event))
3007                         continue;
3008 
3009                 if (group_can_go_on(event, cpuctx, 1))
3010                         group_sched_in(event, cpuctx, ctx);
3011 
3012                 /*
3013                  * If this pinned group hasn't been scheduled,
3014                  * put it in error state.
3015                  */
3016                 if (event->state == PERF_EVENT_STATE_INACTIVE)
3017                         perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
3018         }
3019 }
3020 
3021 static void
3022 ctx_flexible_sched_in(struct perf_event_context *ctx,
3023                       struct perf_cpu_context *cpuctx)
3024 {
3025         struct perf_event *event;
3026         int can_add_hw = 1;
3027 
3028         list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
3029                 /* Ignore events in OFF or ERROR state */
3030                 if (event->state <= PERF_EVENT_STATE_OFF)
3031                         continue;
3032                 /*
3033                  * Listen to the 'cpu' scheduling filter constraint
3034                  * of events:
3035                  */
3036                 if (!event_filter_match(event))
3037                         continue;
3038 
3039                 if (group_can_go_on(event, cpuctx, can_add_hw)) {
3040                         if (group_sched_in(event, cpuctx, ctx))
3041                                 can_add_hw = 0;
3042                 }
3043         }
3044 }
3045 
3046 static void
3047 ctx_sched_in(struct perf_event_context *ctx,
3048              struct perf_cpu_context *cpuctx,
3049              enum event_type_t event_type,
3050              struct task_struct *task)
3051 {
3052         int is_active = ctx->is_active;
3053         u64 now;
3054 
3055         lockdep_assert_held(&ctx->lock);
3056 
3057         if (likely(!ctx->nr_events))
3058                 return;
3059 
3060         ctx->is_active |= (event_type | EVENT_TIME);
3061         if (ctx->task) {
3062                 if (!is_active)
3063                         cpuctx->task_ctx = ctx;
3064                 else
3065                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
3066         }
3067 
3068         is_active ^= ctx->is_active; /* changed bits */
3069 
3070         if (is_active & EVENT_TIME) {
3071                 /* start ctx time */
3072                 now = perf_clock();
3073                 ctx->timestamp = now;
3074                 perf_cgroup_set_timestamp(task, ctx);
3075         }
3076 
3077         /*
3078          * First go through the list and put on any pinned groups
3079          * in order to give them the best chance of going on.
3080          */
3081         if (is_active & EVENT_PINNED)
3082                 ctx_pinned_sched_in(ctx, cpuctx);
3083 
3084         /* Then walk through the lower prio flexible groups */
3085         if (is_active & EVENT_FLEXIBLE)
3086                 ctx_flexible_sched_in(ctx, cpuctx);
3087 }
3088 
3089 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
3090                              enum event_type_t event_type,
3091                              struct task_struct *task)
3092 {
3093         struct perf_event_context *ctx = &cpuctx->ctx;
3094 
3095         ctx_sched_in(ctx, cpuctx, event_type, task);
3096 }
3097 
3098 static void perf_event_context_sched_in(struct perf_event_context *ctx,
3099                                         struct task_struct *task)
3100 {
3101         struct perf_cpu_context *cpuctx;
3102 
3103         cpuctx = __get_cpu_context(ctx);
3104         if (cpuctx->task_ctx == ctx)
3105                 return;
3106 
3107         perf_ctx_lock(cpuctx, ctx);
3108         /*
3109          * We must check ctx->nr_events while holding ctx->lock, such
3110          * that we serialize against perf_install_in_context().
3111          */
3112         if (!ctx->nr_events)
3113                 goto unlock;
3114 
3115         perf_pmu_disable(ctx->pmu);
3116         /*
3117          * We want to keep the following priority order:
3118          * cpu pinned (that don't need to move), task pinned,
3119          * cpu flexible, task flexible.
3120          *
3121          * However, if task's ctx is not carrying any pinned
3122          * events, no need to flip the cpuctx's events around.
3123          */
3124         if (!list_empty(&ctx->pinned_groups))
3125                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3126         perf_event_sched_in(cpuctx, ctx, task);
3127         perf_pmu_enable(ctx->pmu);
3128 
3129 unlock:
3130         perf_ctx_unlock(cpuctx, ctx);
3131 }
3132 
3133 /*
3134  * Called from scheduler to add the events of the current task
3135  * with interrupts disabled.
3136  *
3137  * We restore the event value and then enable it.
3138  *
3139  * This does not protect us against NMI, but enable()
3140  * sets the enabled bit in the control field of event _before_
3141  * accessing the event control register. If a NMI hits, then it will
3142  * keep the event running.
3143  */
3144 void __perf_event_task_sched_in(struct task_struct *prev,
3145                                 struct task_struct *task)
3146 {
3147         struct perf_event_context *ctx;
3148         int ctxn;
3149 
3150         /*
3151          * If cgroup events exist on this CPU, then we need to check if we have
3152          * to switch in PMU state; cgroup event are system-wide mode only.
3153          *
3154          * Since cgroup events are CPU events, we must schedule these in before
3155          * we schedule in the task events.
3156          */
3157         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3158                 perf_cgroup_sched_in(prev, task);
3159 
3160         for_each_task_context_nr(ctxn) {
3161                 ctx = task->perf_event_ctxp[ctxn];
3162                 if (likely(!ctx))
3163                         continue;
3164 
3165                 perf_event_context_sched_in(ctx, task);
3166         }
3167 
3168         if (atomic_read(&nr_switch_events))
3169                 perf_event_switch(task, prev, true);
3170 
3171         if (__this_cpu_read(perf_sched_cb_usages))
3172                 perf_pmu_sched_task(prev, task, true);
3173 }
3174 
3175 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
3176 {
3177         u64 frequency = event->attr.sample_freq;
3178         u64 sec = NSEC_PER_SEC;
3179         u64 divisor, dividend;
3180 
3181         int count_fls, nsec_fls, frequency_fls, sec_fls;
3182 
3183         count_fls = fls64(count);
3184         nsec_fls = fls64(nsec);
3185         frequency_fls = fls64(frequency);
3186         sec_fls = 30;
3187 
3188         /*
3189          * We got @count in @nsec, with a target of sample_freq HZ
3190          * the target period becomes:
3191          *
3192          *             @count * 10^9
3193          * period = -------------------
3194          *          @nsec * sample_freq
3195          *
3196          */
3197 
3198         /*
3199          * Reduce accuracy by one bit such that @a and @b converge
3200          * to a similar magnitude.
3201          */
3202 #define REDUCE_FLS(a, b)                \
3203 do {                                    \
3204         if (a##_fls > b##_fls) {        \
3205                 a >>= 1;                \
3206                 a##_fls--;              \
3207         } else {                        \
3208                 b >>= 1;                \
3209                 b##_fls--;              \
3210         }                               \
3211 } while (0)
3212 
3213         /*
3214          * Reduce accuracy until either term fits in a u64, then proceed with
3215          * the other, so that finally we can do a u64/u64 division.
3216          */
3217         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
3218                 REDUCE_FLS(nsec, frequency);
3219                 REDUCE_FLS(sec, count);
3220         }
3221 
3222         if (count_fls + sec_fls > 64) {
3223                 divisor = nsec * frequency;
3224 
3225                 while (count_fls + sec_fls > 64) {
3226                         REDUCE_FLS(count, sec);
3227                         divisor >>= 1;
3228                 }
3229 
3230                 dividend = count * sec;
3231         } else {
3232                 dividend = count * sec;
3233 
3234                 while (nsec_fls + frequency_fls > 64) {
3235                         REDUCE_FLS(nsec, frequency);
3236                         dividend >>= 1;
3237                 }
3238 
3239                 divisor = nsec * frequency;
3240         }
3241 
3242         if (!divisor)
3243                 return dividend;
3244 
3245         return div64_u64(dividend, divisor);
3246 }
3247 
3248 static DEFINE_PER_CPU(int, perf_throttled_count);
3249 static DEFINE_PER_CPU(u64, perf_throttled_seq);
3250 
3251 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
3252 {
3253         struct hw_perf_event *hwc = &event->hw;
3254         s64 period, sample_period;
3255         s64 delta;
3256 
3257         period = perf_calculate_period(event, nsec, count);
3258 
3259         delta = (s64)(period - hwc->sample_period);
3260         delta = (delta + 7) / 8; /* low pass filter */
3261 
3262         sample_period = hwc->sample_period + delta;
3263 
3264         if (!sample_period)
3265                 sample_period = 1;
3266 
3267         hwc->sample_period = sample_period;
3268 
3269         if (local64_read(&hwc->period_left) > 8*sample_period) {
3270                 if (disable)
3271                         event->pmu->stop(event, PERF_EF_UPDATE);
3272 
3273                 local64_set(&hwc->period_left, 0);
3274 
3275                 if (disable)
3276                         event->pmu->start(event, PERF_EF_RELOAD);
3277         }
3278 }
3279 
3280 /*
3281  * combine freq adjustment with unthrottling to avoid two passes over the
3282  * events. At the same time, make sure, having freq events does not change
3283  * the rate of unthrottling as that would introduce bias.
3284  */
3285 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
3286                                            int needs_unthr)
3287 {
3288         struct perf_event *event;
3289         struct hw_perf_event *hwc;
3290         u64 now, period = TICK_NSEC;
3291         s64 delta;
3292 
3293         /*
3294          * only need to iterate over all events iff:
3295          * - context have events in frequency mode (needs freq adjust)
3296          * - there are events to unthrottle on this cpu
3297          */
3298         if (!(ctx->nr_freq || needs_unthr))
3299                 return;
3300 
3301         raw_spin_lock(&ctx->lock);
3302         perf_pmu_disable(ctx->pmu);
3303 
3304         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3305                 if (event->state != PERF_EVENT_STATE_ACTIVE)
3306                         continue;
3307 
3308                 if (!event_filter_match(event))
3309                         continue;
3310 
3311                 perf_pmu_disable(event->pmu);
3312 
3313                 hwc = &event->hw;
3314 
3315                 if (hwc->interrupts == MAX_INTERRUPTS) {
3316                         hwc->interrupts = 0;
3317                         perf_log_throttle(event, 1);
3318                         event->pmu->start(event, 0);
3319                 }
3320 
3321                 if (!event->attr.freq || !event->attr.sample_freq)
3322                         goto next;
3323 
3324                 /*
3325                  * stop the event and update event->count
3326                  */
3327                 event->pmu->stop(event, PERF_EF_UPDATE);
3328 
3329                 now = local64_read(&event->count);
3330                 delta = now - hwc->freq_count_stamp;
3331                 hwc->freq_count_stamp = now;
3332 
3333                 /*
3334                  * restart the event
3335                  * reload only if value has changed
3336                  * we have stopped the event so tell that
3337                  * to perf_adjust_period() to avoid stopping it
3338                  * twice.
3339                  */
3340                 if (delta > 0)
3341                         perf_adjust_period(event, period, delta, false);
3342 
3343                 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3344         next:
3345                 perf_pmu_enable(event->pmu);
3346         }
3347 
3348         perf_pmu_enable(ctx->pmu);
3349         raw_spin_unlock(&ctx->lock);
3350 }
3351 
3352 /*
3353  * Round-robin a context's events:
3354  */
3355 static void rotate_ctx(struct perf_event_context *ctx)
3356 {
3357         /*
3358          * Rotate the first entry last of non-pinned groups. Rotation might be
3359          * disabled by the inheritance code.
3360          */
3361         if (!ctx->rotate_disable)
3362                 list_rotate_left(&ctx->flexible_groups);
3363 }
3364 
3365 static int perf_rotate_context(struct perf_cpu_context *cpuctx)
3366 {
3367         struct perf_event_context *ctx = NULL;
3368         int rotate = 0;
3369 
3370         if (cpuctx->ctx.nr_events) {
3371                 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3372                         rotate = 1;
3373         }
3374 
3375         ctx = cpuctx->task_ctx;
3376         if (ctx && ctx->nr_events) {
3377                 if (ctx->nr_events != ctx->nr_active)
3378                         rotate = 1;
3379         }
3380 
3381         if (!rotate)
3382                 goto done;
3383 
3384         perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3385         perf_pmu_disable(cpuctx->ctx.pmu);
3386 
3387         cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3388         if (ctx)
3389                 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3390 
3391         rotate_ctx(&cpuctx->ctx);
3392         if (ctx)
3393                 rotate_ctx(ctx);
3394 
3395         perf_event_sched_in(cpuctx, ctx, current);
3396 
3397         perf_pmu_enable(cpuctx->ctx.pmu);
3398         perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3399 done:
3400 
3401         return rotate;
3402 }
3403 
3404 void perf_event_task_tick(void)
3405 {
3406         struct list_head *head = this_cpu_ptr(&active_ctx_list);
3407         struct perf_event_context *ctx, *tmp;
3408         int throttled;
3409 
3410         lockdep_assert_irqs_disabled();
3411 
3412         __this_cpu_inc(perf_throttled_seq);
3413         throttled = __this_cpu_xchg(perf_throttled_count, 0);
3414         tick_dep_clear_cpu(smp_processor_id(), TICK_DEP_BIT_PERF_EVENTS);
3415 
3416         list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3417                 perf_adjust_freq_unthr_context(ctx, throttled);
3418 }
3419 
3420 static int event_enable_on_exec(struct perf_event *event,
3421                                 struct perf_event_context *ctx)
3422 {
3423         if (!event->attr.enable_on_exec)
3424                 return 0;
3425 
3426         event->attr.enable_on_exec = 0;
3427         if (event->state >= PERF_EVENT_STATE_INACTIVE)
3428                 return 0;
3429 
3430         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
3431 
3432         return 1;
3433 }
3434 
3435 /*
3436  * Enable all of a task's events that have been marked enable-on-exec.
3437  * This expects task == current.
3438  */
3439 static void perf_event_enable_on_exec(int ctxn)
3440 {
3441         struct perf_event_context *ctx, *clone_ctx = NULL;
3442         enum event_type_t event_type = 0;
3443         struct perf_cpu_context *cpuctx;
3444         struct perf_event *event;
3445         unsigned long flags;
3446         int enabled = 0;
3447 
3448         local_irq_save(flags);
3449         ctx = current->perf_event_ctxp[ctxn];
3450         if (!ctx || !ctx->nr_events)
3451                 goto out;
3452 
3453         cpuctx = __get_cpu_context(ctx);
3454         perf_ctx_lock(cpuctx, ctx);
3455         ctx_sched_out(ctx, cpuctx, EVENT_TIME);
3456         list_for_each_entry(event, &ctx->event_list, event_entry) {
3457                 enabled |= event_enable_on_exec(event, ctx);
3458                 event_type |= get_event_type(event);
3459         }
3460 
3461         /*
3462          * Unclone and reschedule this context if we enabled any event.
3463          */
3464         if (enabled) {
3465                 clone_ctx = unclone_ctx(ctx);
3466                 ctx_resched(cpuctx, ctx, event_type);
3467         } else {
3468                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
3469         }
3470         perf_ctx_unlock(cpuctx, ctx);
3471 
3472 out:
3473         local_irq_restore(flags);
3474 
3475         if (clone_ctx)
3476                 put_ctx(clone_ctx);
3477 }
3478 
3479 struct perf_read_data {
3480         struct perf_event *event;
3481         bool group;
3482         int ret;
3483 };
3484 
3485 static int __perf_event_read_cpu(struct perf_event *event, int event_cpu)
3486 {
3487         u16 local_pkg, event_pkg;
3488 
3489         if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) {
3490                 int local_cpu = smp_processor_id();
3491 
3492                 event_pkg = topology_physical_package_id(event_cpu);
3493                 local_pkg = topology_physical_package_id(local_cpu);
3494 
3495                 if (event_pkg == local_pkg)
3496                         return local_cpu;
3497         }
3498 
3499         return event_cpu;
3500 }
3501 
3502 /*
3503  * Cross CPU call to read the hardware event
3504  */
3505 static void __perf_event_read(void *info)
3506 {
3507         struct perf_read_data *data = info;
3508         struct perf_event *sub, *event = data->event;
3509         struct perf_event_context *ctx = event->ctx;
3510         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3511         struct pmu *pmu = event->pmu;
3512 
3513         /*
3514          * If this is a task context, we need to check whether it is
3515          * the current task context of this cpu.  If not it has been
3516          * scheduled out before the smp call arrived.  In that case
3517          * event->count would have been updated to a recent sample
3518          * when the event was scheduled out.
3519          */
3520         if (ctx->task && cpuctx->task_ctx != ctx)
3521                 return;
3522 
3523         raw_spin_lock(&ctx->lock);
3524         if (ctx->is_active & EVENT_TIME) {
3525                 update_context_time(ctx);
3526                 update_cgrp_time_from_event(event);
3527         }
3528 
3529         perf_event_update_time(event);
3530         if (data->group)
3531                 perf_event_update_sibling_time(event);
3532 
3533         if (event->state != PERF_EVENT_STATE_ACTIVE)
3534                 goto unlock;
3535 
3536         if (!data->group) {
3537                 pmu->read(event);
3538                 data->ret = 0;
3539                 goto unlock;
3540         }
3541 
3542         pmu->start_txn(pmu, PERF_PMU_TXN_READ);
3543 
3544         pmu->read(event);
3545 
3546         list_for_each_entry(sub, &event->sibling_list, group_entry) {
3547                 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
3548                         /*
3549                          * Use sibling's PMU rather than @event's since
3550                          * sibling could be on different (eg: software) PMU.
3551                          */
3552                         sub->pmu->read(sub);
3553                 }
3554         }
3555 
3556         data->ret = pmu->commit_txn(pmu);
3557 
3558 unlock:
3559         raw_spin_unlock(&ctx->lock);
3560 }
3561 
3562 static inline u64 perf_event_count(struct perf_event *event)
3563 {
3564         return local64_read(&event->count) + atomic64_read(&event->child_count);
3565 }
3566 
3567 /*
3568  * NMI-safe method to read a local event, that is an event that
3569  * is:
3570  *   - either for the current task, or for this CPU
3571  *   - does not have inherit set, for inherited task events
3572  *     will not be local and we cannot read them atomically
3573  *   - must not have a pmu::count method
3574  */
3575 int perf_event_read_local(struct perf_event *event, u64 *value,
3576                           u64 *enabled, u64 *running)
3577 {
3578         unsigned long flags;
3579         int ret = 0;
3580 
3581         /*
3582          * Disabling interrupts avoids all counter scheduling (context
3583          * switches, timer based rotation and IPIs).
3584          */
3585         local_irq_save(flags);
3586 
3587         /*
3588          * It must not be an event with inherit set, we cannot read
3589          * all child counters from atomic context.
3590          */
3591         if (event->attr.inherit) {
3592                 ret = -EOPNOTSUPP;
3593                 goto out;
3594         }
3595 
3596         /* If this is a per-task event, it must be for current */
3597         if ((event->attach_state & PERF_ATTACH_TASK) &&
3598             event->hw.target != current) {
3599                 ret = -EINVAL;
3600                 goto out;
3601         }
3602 
3603         /* If this is a per-CPU event, it must be for this CPU */
3604         if (!(event->attach_state & PERF_ATTACH_TASK) &&
3605             event->cpu != smp_processor_id()) {
3606                 ret = -EINVAL;
3607                 goto out;
3608         }
3609 
3610         /*
3611          * If the event is currently on this CPU, its either a per-task event,
3612          * or local to this CPU. Furthermore it means its ACTIVE (otherwise
3613          * oncpu == -1).
3614          */
3615         if (event->oncpu == smp_processor_id())
3616                 event->pmu->read(event);
3617 
3618         *value = local64_read(&event->count);
3619         if (enabled || running) {
3620                 u64 now = event->shadow_ctx_time + perf_clock();
3621                 u64 __enabled, __running;
3622 
3623                 __perf_update_times(event, now, &__enabled, &__running);
3624                 if (enabled)
3625                         *enabled = __enabled;
3626                 if (running)
3627                         *running = __running;
3628         }
3629 out:
3630         local_irq_restore(flags);
3631 
3632         return ret;
3633 }
3634 
3635 static int perf_event_read(struct perf_event *event, bool group)
3636 {
3637         enum perf_event_state state = READ_ONCE(event->state);
3638         int event_cpu, ret = 0;
3639 
3640         /*
3641          * If event is enabled and currently active on a CPU, update the
3642          * value in the event structure:
3643          */
3644 again:
3645         if (state == PERF_EVENT_STATE_ACTIVE) {
3646                 struct perf_read_data data;
3647 
3648                 /*
3649                  * Orders the ->state and ->oncpu loads such that if we see
3650                  * ACTIVE we must also see the right ->oncpu.
3651                  *
3652                  * Matches the smp_wmb() from event_sched_in().
3653                  */
3654                 smp_rmb();
3655 
3656                 event_cpu = READ_ONCE(event->oncpu);
3657                 if ((unsigned)event_cpu >= nr_cpu_ids)
3658                         return 0;
3659 
3660                 data = (struct perf_read_data){
3661                         .event = event,
3662                         .group = group,
3663                         .ret = 0,
3664                 };
3665 
3666                 preempt_disable();
3667                 event_cpu = __perf_event_read_cpu(event, event_cpu);
3668 
3669                 /*
3670                  * Purposely ignore the smp_call_function_single() return
3671                  * value.
3672                  *
3673                  * If event_cpu isn't a valid CPU it means the event got
3674                  * scheduled out and that will have updated the event count.
3675                  *
3676                  * Therefore, either way, we'll have an up-to-date event count
3677                  * after this.
3678                  */
3679                 (void)smp_call_function_single(event_cpu, __perf_event_read, &data, 1);
3680                 preempt_enable();
3681                 ret = data.ret;
3682 
3683         } else if (state == PERF_EVENT_STATE_INACTIVE) {
3684                 struct perf_event_context *ctx = event->ctx;
3685                 unsigned long flags;
3686 
3687                 raw_spin_lock_irqsave(&ctx->lock, flags);
3688                 state = event->state;
3689                 if (state != PERF_EVENT_STATE_INACTIVE) {
3690                         raw_spin_unlock_irqrestore(&ctx->lock, flags);
3691                         goto again;
3692                 }
3693 
3694                 /*
3695                  * May read while context is not active (e.g., thread is
3696                  * blocked), in that case we cannot update context time
3697                  */
3698                 if (ctx->is_active & EVENT_TIME) {
3699                         update_context_time(ctx);
3700                         update_cgrp_time_from_event(event);
3701                 }
3702 
3703                 perf_event_update_time(event);
3704                 if (group)
3705                         perf_event_update_sibling_time(event);
3706                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3707         }
3708 
3709         return ret;
3710 }
3711 
3712 /*
3713  * Initialize the perf_event context in a task_struct:
3714  */
3715 static void __perf_event_init_context(struct perf_event_context *ctx)
3716 {
3717         raw_spin_lock_init(&ctx->lock);
3718         mutex_init(&ctx->mutex);
3719         INIT_LIST_HEAD(&ctx->active_ctx_list);
3720         INIT_LIST_HEAD(&ctx->pinned_groups);
3721         INIT_LIST_HEAD(&ctx->flexible_groups);
3722         INIT_LIST_HEAD(&ctx->event_list);
3723         atomic_set(&ctx->refcount, 1);
3724 }
3725 
3726 static struct perf_event_context *
3727 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
3728 {
3729         struct perf_event_context *ctx;
3730 
3731         ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
3732         if (!ctx)
3733                 return NULL;
3734 
3735         __perf_event_init_context(ctx);
3736         if (task) {
3737                 ctx->task = task;
3738                 get_task_struct(task);
3739         }
3740         ctx->pmu = pmu;
3741 
3742         return ctx;
3743 }
3744 
3745 static struct task_struct *
3746 find_lively_task_by_vpid(pid_t vpid)
3747 {
3748         struct task_struct *task;
3749 
3750         rcu_read_lock();
3751         if (!vpid)
3752                 task = current;
3753         else
3754                 task = find_task_by_vpid(vpid);
3755         if (task)
3756                 get_task_struct(task);
3757         rcu_read_unlock();
3758 
3759         if (!task)
3760                 return ERR_PTR(-ESRCH);
3761 
3762         return task;
3763 }
3764 
3765 /*
3766  * Returns a matching context with refcount and pincount.
3767  */
3768 static struct perf_event_context *
3769 find_get_context(struct pmu *pmu, struct task_struct *task,
3770                 struct perf_event *event)
3771 {
3772         struct perf_event_context *ctx, *clone_ctx = NULL;
3773         struct perf_cpu_context *cpuctx;
3774         void *task_ctx_data = NULL;
3775         unsigned long flags;
3776         int ctxn, err;
3777         int cpu = event->cpu;
3778 
3779         if (!task) {
3780                 /* Must be root to operate on a CPU event: */
3781                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
3782                         return ERR_PTR(-EACCES);
3783 
3784                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
3785                 ctx = &cpuctx->ctx;
3786                 get_ctx(ctx);
3787                 ++ctx->pin_count;
3788 
3789                 return ctx;
3790         }
3791 
3792         err = -EINVAL;
3793         ctxn = pmu->task_ctx_nr;
3794         if (ctxn < 0)
3795                 goto errout;
3796 
3797         if (event->attach_state & PERF_ATTACH_TASK_DATA) {
3798                 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
3799                 if (!task_ctx_data) {
3800                         err = -ENOMEM;
3801                         goto errout;
3802                 }
3803         }
3804 
3805 retry:
3806         ctx = perf_lock_task_context(task, ctxn, &flags);
3807         if (ctx) {
3808                 clone_ctx = unclone_ctx(ctx);
3809                 ++ctx->pin_count;
3810 
3811                 if (task_ctx_data && !ctx->task_ctx_data) {
3812                         ctx->task_ctx_data = task_ctx_data;
3813                         task_ctx_data = NULL;
3814                 }
3815                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
3816 
3817                 if (clone_ctx)
3818                         put_ctx(clone_ctx);
3819         } else {
3820                 ctx = alloc_perf_context(pmu, task);
3821                 err = -ENOMEM;
3822                 if (!ctx)
3823                         goto errout;
3824 
3825                 if (task_ctx_data) {
3826                         ctx->task_ctx_data = task_ctx_data;
3827                         task_ctx_data = NULL;
3828                 }
3829 
3830                 err = 0;
3831                 mutex_lock(&task->perf_event_mutex);
3832                 /*
3833                  * If it has already passed perf_event_exit_task().
3834                  * we must see PF_EXITING, it takes this mutex too.
3835                  */
3836                 if (task->flags & PF_EXITING)
3837                         err = -ESRCH;
3838                 else if (task->perf_event_ctxp[ctxn])
3839                         err = -EAGAIN;
3840                 else {
3841                         get_ctx(ctx);
3842                         ++ctx->pin_count;
3843                         rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
3844                 }
3845                 mutex_unlock(&task->perf_event_mutex);
3846 
3847                 if (unlikely(err)) {
3848                         put_ctx(ctx);
3849 
3850                         if (err == -EAGAIN)
3851                                 goto retry;
3852                         goto errout;
3853                 }
3854         }
3855 
3856         kfree(task_ctx_data);
3857         return ctx;
3858 
3859 errout:
3860         kfree(task_ctx_data);
3861         return ERR_PTR(err);
3862 }
3863 
3864 static void perf_event_free_filter(struct perf_event *event);
3865 static void perf_event_free_bpf_prog(struct perf_event *event);
3866 
3867 static void free_event_rcu(struct rcu_head *head)
3868 {
3869         struct perf_event *event;
3870 
3871         event = container_of(head, struct perf_event, rcu_head);
3872         if (event->ns)
3873                 put_pid_ns(event->ns);
3874         perf_event_free_filter(event);
3875         kfree(event);
3876 }
3877 
3878 static void ring_buffer_attach(struct perf_event *event,
3879                                struct ring_buffer *rb);
3880 
3881 static void detach_sb_event(struct perf_event *event)
3882 {
3883         struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu);
3884 
3885         raw_spin_lock(&pel->lock);
3886         list_del_rcu(&event->sb_list);
3887         raw_spin_unlock(&pel->lock);
3888 }
3889 
3890 static bool is_sb_event(struct perf_event *event)
3891 {
3892         struct perf_event_attr *attr = &event->attr;
3893 
3894         if (event->parent)
3895                 return false;
3896 
3897         if (event->attach_state & PERF_ATTACH_TASK)
3898                 return false;
3899 
3900         if (attr->mmap || attr->mmap_data || attr->mmap2 ||
3901             attr->comm || attr->comm_exec ||
3902             attr->task ||
3903             attr->context_switch)
3904                 return true;
3905         return false;
3906 }
3907 
3908 static void unaccount_pmu_sb_event(struct perf_event *event)
3909 {
3910         if (is_sb_event(event))
3911                 detach_sb_event(event);
3912 }
3913 
3914 static void unaccount_event_cpu(struct perf_event *event, int cpu)
3915 {
3916         if (event->parent)
3917                 return;
3918 
3919         if (is_cgroup_event(event))
3920                 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
3921 }
3922 
3923 #ifdef CONFIG_NO_HZ_FULL
3924 static DEFINE_SPINLOCK(nr_freq_lock);
3925 #endif
3926 
3927 static void unaccount_freq_event_nohz(void)
3928 {
3929 #ifdef CONFIG_NO_HZ_FULL
3930         spin_lock(&nr_freq_lock);
3931         if (atomic_dec_and_test(&nr_freq_events))
3932                 tick_nohz_dep_clear(TICK_DEP_BIT_PERF_EVENTS);
3933         spin_unlock(&nr_freq_lock);
3934 #endif
3935 }
3936 
3937 static void unaccount_freq_event(void)
3938 {
3939         if (tick_nohz_full_enabled())
3940                 unaccount_freq_event_nohz();
3941         else
3942                 atomic_dec(&nr_freq_events);
3943 }
3944 
3945 static void unaccount_event(struct perf_event *event)
3946 {
3947         bool dec = false;
3948 
3949         if (event->parent)
3950                 return;
3951 
3952         if (event->attach_state & PERF_ATTACH_TASK)
3953                 dec = true;
3954         if (event->attr.mmap || event->attr.mmap_data)
3955                 atomic_dec(&nr_mmap_events);
3956         if (event->attr.comm)
3957                 atomic_dec(&nr_comm_events);
3958         if (event->attr.namespaces)
3959                 atomic_dec(&nr_namespaces_events);
3960         if (event->attr.task)
3961                 atomic_dec(&nr_task_events);
3962         if (event->attr.freq)
3963                 unaccount_freq_event();
3964         if (event->attr.context_switch) {
3965                 dec = true;
3966                 atomic_dec(&nr_switch_events);
3967         }
3968         if (is_cgroup_event(event))
3969                 dec = true;
3970         if (has_branch_stack(event))
3971                 dec = true;
3972 
3973         if (dec) {
3974                 if (!atomic_add_unless(&perf_sched_count, -1, 1))
3975                         schedule_delayed_work(&perf_sched_work, HZ);
3976         }
3977 
3978         unaccount_event_cpu(event, event->cpu);
3979 
3980         unaccount_pmu_sb_event(event);
3981 }
3982 
3983 static void perf_sched_delayed(struct work_struct *work)
3984 {
3985         mutex_lock(&perf_sched_mutex);
3986         if (atomic_dec_and_test(&perf_sched_count))
3987                 static_branch_disable(&perf_sched_events);
3988         mutex_unlock(&perf_sched_mutex);
3989 }
3990 
3991 /*
3992  * The following implement mutual exclusion of events on "exclusive" pmus
3993  * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
3994  * at a time, so we disallow creating events that might conflict, namely:
3995  *
3996  *  1) cpu-wide events in the presence of per-task events,
3997  *  2) per-task events in the presence of cpu-wide events,
3998  *  3) two matching events on the same context.
3999  *
4000  * The former two cases are handled in the allocation path (perf_event_alloc(),
4001  * _free_event()), the latter -- before the first perf_install_in_context().
4002  */
4003 static int exclusive_event_init(struct perf_event *event)
4004 {
4005         struct pmu *pmu = event->pmu;
4006 
4007         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4008                 return 0;
4009 
4010         /*
4011          * Prevent co-existence of per-task and cpu-wide events on the
4012          * same exclusive pmu.
4013          *
4014          * Negative pmu::exclusive_cnt means there are cpu-wide
4015          * events on this "exclusive" pmu, positive means there are
4016          * per-task events.
4017          *
4018          * Since this is called in perf_event_alloc() path, event::ctx
4019          * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
4020          * to mean "per-task event", because unlike other attach states it
4021          * never gets cleared.
4022          */
4023         if (event->attach_state & PERF_ATTACH_TASK) {
4024                 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
4025                         return -EBUSY;
4026         } else {
4027                 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
4028                         return -EBUSY;
4029         }
4030 
4031         return 0;
4032 }
4033 
4034 static void exclusive_event_destroy(struct perf_event *event)
4035 {
4036         struct pmu *pmu = event->pmu;
4037 
4038         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4039                 return;
4040 
4041         /* see comment in exclusive_event_init() */
4042         if (event->attach_state & PERF_ATTACH_TASK)
4043                 atomic_dec(&pmu->exclusive_cnt);
4044         else
4045                 atomic_inc(&pmu->exclusive_cnt);
4046 }
4047 
4048 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
4049 {
4050         if ((e1->pmu == e2->pmu) &&
4051             (e1->cpu == e2->cpu ||
4052              e1->cpu == -1 ||
4053              e2->cpu == -1))
4054                 return true;
4055         return false;
4056 }
4057 
4058 /* Called under the same ctx::mutex as perf_install_in_context() */
4059 static bool exclusive_event_installable(struct perf_event *event,
4060                                         struct perf_event_context *ctx)
4061 {
4062         struct perf_event *iter_event;
4063         struct pmu *pmu = event->pmu;
4064 
4065         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4066                 return true;
4067 
4068         list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
4069                 if (exclusive_event_match(iter_event, event))
4070                         return false;
4071         }
4072 
4073         return true;
4074 }
4075 
4076 static void perf_addr_filters_splice(struct perf_event *event,
4077                                        struct list_head *head);
4078 
4079 static void _free_event(struct perf_event *event)
4080 {
4081         irq_work_sync(&event->pending);
4082 
4083         unaccount_event(event);
4084 
4085         if (event->rb) {
4086                 /*
4087                  * Can happen when we close an event with re-directed output.
4088                  *
4089                  * Since we have a 0 refcount, perf_mmap_close() will skip
4090                  * over us; possibly making our ring_buffer_put() the last.
4091                  */
4092                 mutex_lock(&event->mmap_mutex);
4093                 ring_buffer_attach(event, NULL);
4094                 mutex_unlock(&event->mmap_mutex);
4095         }
4096 
4097         if (is_cgroup_event(event))
4098                 perf_detach_cgroup(event);
4099 
4100         if (!event->parent) {
4101                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
4102                         put_callchain_buffers();
4103         }
4104 
4105         perf_event_free_bpf_prog(event);
4106         perf_addr_filters_splice(event, NULL);
4107         kfree(event->addr_filters_offs);
4108 
4109         if (event->destroy)
4110                 event->destroy(event);
4111 
4112         if (event->ctx)
4113                 put_ctx(event->ctx);
4114 
4115         if (event->hw.target)
4116                 put_task_struct(event->hw.target);
4117 
4118         exclusive_event_destroy(event);
4119         module_put(event->pmu->module);
4120 
4121         call_rcu(&event->rcu_head, free_event_rcu);
4122 }
4123 
4124 /*
4125  * Used to free events which have a known refcount of 1, such as in error paths
4126  * where the event isn't exposed yet and inherited events.
4127  */
4128 static void free_event(struct perf_event *event)
4129 {
4130         if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
4131                                 "unexpected event refcount: %ld; ptr=%p\n",
4132                                 atomic_long_read(&event->refcount), event)) {
4133                 /* leak to avoid use-after-free */
4134                 return;
4135         }
4136 
4137         _free_event(event);
4138 }
4139 
4140 /*
4141  * Remove user event from the owner task.
4142  */
4143 static void perf_remove_from_owner(struct perf_event *event)
4144 {
4145         struct task_struct *owner;
4146 
4147         rcu_read_lock();
4148         /*
4149          * Matches the smp_store_release() in perf_event_exit_task(). If we
4150          * observe !owner it means the list deletion is complete and we can
4151          * indeed free this event, otherwise we need to serialize on
4152          * owner->perf_event_mutex.
4153          */
4154         owner = READ_ONCE(event->owner);
4155         if (owner) {
4156                 /*
4157                  * Since delayed_put_task_struct() also drops the last
4158                  * task reference we can safely take a new reference
4159                  * while holding the rcu_read_lock().
4160                  */
4161                 get_task_struct(owner);
4162         }
4163         rcu_read_unlock();
4164 
4165         if (owner) {
4166                 /*
4167                  * If we're here through perf_event_exit_task() we're already
4168                  * holding ctx->mutex which would be an inversion wrt. the
4169                  * normal lock order.
4170                  *
4171                  * However we can safely take this lock because its the child
4172                  * ctx->mutex.
4173                  */
4174                 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
4175 
4176                 /*
4177                  * We have to re-check the event->owner field, if it is cleared
4178                  * we raced with perf_event_exit_task(), acquiring the mutex
4179                  * ensured they're done, and we can proceed with freeing the
4180                  * event.
4181                  */
4182                 if (event->owner) {
4183                         list_del_init(&event->owner_entry);
4184                         smp_store_release(&event->owner, NULL);
4185                 }
4186                 mutex_unlock(&owner->perf_event_mutex);
4187                 put_task_struct(owner);
4188         }
4189 }
4190 
4191 static void put_event(struct perf_event *event)
4192 {
4193         if (!atomic_long_dec_and_test(&event->refcount))
4194                 return;
4195 
4196         _free_event(event);
4197 }
4198 
4199 /*
4200  * Kill an event dead; while event:refcount will preserve the event
4201  * object, it will not preserve its functionality. Once the last 'user'
4202  * gives up the object, we'll destroy the thing.
4203  */
4204 int perf_event_release_kernel(struct perf_event *event)
4205 {
4206         struct perf_event_context *ctx = event->ctx;
4207         struct perf_event *child, *tmp;
4208         LIST_HEAD(free_list);
4209 
4210         /*
4211          * If we got here through err_file: fput(event_file); we will not have
4212          * attached to a context yet.
4213          */
4214         if (!ctx) {
4215                 WARN_ON_ONCE(event->attach_state &
4216                                 (PERF_ATTACH_CONTEXT|PERF_ATTACH_GROUP));
4217                 goto no_ctx;
4218         }
4219 
4220         if (!is_kernel_event(event))
4221                 perf_remove_from_owner(event);
4222 
4223         ctx = perf_event_ctx_lock(event);
4224         WARN_ON_ONCE(ctx->parent_ctx);
4225         perf_remove_from_context(event, DETACH_GROUP);
4226 
4227         raw_spin_lock_irq(&ctx->lock);
4228         /*
4229          * Mark this event as STATE_DEAD, there is no external reference to it
4230          * anymore.
4231          *
4232          * Anybody acquiring event->child_mutex after the below loop _must_
4233          * also see this, most importantly inherit_event() which will avoid
4234          * placing more children on the list.
4235          *
4236          * Thus this guarantees that we will in fact observe and kill _ALL_
4237          * child events.
4238          */
4239         event->state = PERF_EVENT_STATE_DEAD;
4240         raw_spin_unlock_irq(&ctx->lock);
4241 
4242         perf_event_ctx_unlock(event, ctx);
4243 
4244 again:
4245         mutex_lock(&event->child_mutex);
4246         list_for_each_entry(child, &event->child_list, child_list) {
4247 
4248                 /*
4249                  * Cannot change, child events are not migrated, see the
4250                  * comment with perf_event_ctx_lock_nested().
4251                  */
4252                 ctx = READ_ONCE(child->ctx);
4253                 /*
4254                  * Since child_mutex nests inside ctx::mutex, we must jump
4255                  * through hoops. We start by grabbing a reference on the ctx.
4256                  *
4257                  * Since the event cannot get freed while we hold the
4258                  * child_mutex, the context must also exist and have a !0
4259                  * reference count.
4260                  */
4261                 get_ctx(ctx);
4262 
4263                 /*
4264                  * Now that we have a ctx ref, we can drop child_mutex, and
4265                  * acquire ctx::mutex without fear of it going away. Then we
4266                  * can re-acquire child_mutex.
4267                  */
4268                 mutex_unlock(&event->child_mutex);
4269                 mutex_lock(&ctx->mutex);
4270                 mutex_lock(&event->child_mutex);
4271 
4272                 /*
4273                  * Now that we hold ctx::mutex and child_mutex, revalidate our
4274                  * state, if child is still the first entry, it didn't get freed
4275                  * and we can continue doing so.
4276                  */
4277                 tmp = list_first_entry_or_null(&event->child_list,
4278                                                struct perf_event, child_list);
4279                 if (tmp == child) {
4280                         perf_remove_from_context(child, DETACH_GROUP);
4281                         list_move(&child->child_list, &free_list);
4282                         /*
4283                          * This matches the refcount bump in inherit_event();
4284                          * this can't be the last reference.
4285                          */
4286                         put_event(event);
4287                 }
4288 
4289                 mutex_unlock(&event->child_mutex);
4290                 mutex_unlock(&ctx->mutex);
4291                 put_ctx(ctx);
4292                 goto again;
4293         }
4294         mutex_unlock(&event->child_mutex);
4295 
4296         list_for_each_entry_safe(child, tmp, &free_list, child_list) {
4297                 list_del(&child->child_list);
4298                 free_event(child);
4299         }
4300 
4301 no_ctx:
4302         put_event(event); /* Must be the 'last' reference */
4303         return 0;
4304 }
4305 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
4306 
4307 /*
4308  * Called when the last reference to the file is gone.
4309  */
4310 static int perf_release(struct inode *inode, struct file *file)
4311 {
4312         perf_event_release_kernel(file->private_data);
4313         return 0;
4314 }
4315 
4316 static u64 __perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4317 {
4318         struct perf_event *child;
4319         u64 total = 0;
4320 
4321         *enabled = 0;
4322         *running = 0;
4323 
4324         mutex_lock(&event->child_mutex);
4325 
4326         (void)perf_event_read(event, false);
4327         total += perf_event_count(event);
4328 
4329         *enabled += event->total_time_enabled +
4330                         atomic64_read(&event->child_total_time_enabled);
4331         *running += event->total_time_running +
4332                         atomic64_read(&event->child_total_time_running);
4333 
4334         list_for_each_entry(child, &event->child_list, child_list) {
4335                 (void)perf_event_read(child, false);
4336                 total += perf_event_count(child);
4337                 *enabled += child->total_time_enabled;
4338                 *running += child->total_time_running;
4339         }
4340         mutex_unlock(&event->child_mutex);
4341 
4342         return total;
4343 }
4344 
4345 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4346 {
4347         struct perf_event_context *ctx;
4348         u64 count;
4349 
4350         ctx = perf_event_ctx_lock(event);
4351         count = __perf_event_read_value(event, enabled, running);
4352         perf_event_ctx_unlock(event, ctx);
4353 
4354         return count;
4355 }
4356 EXPORT_SYMBOL_GPL(perf_event_read_value);
4357 
4358 static int __perf_read_group_add(struct perf_event *leader,
4359                                         u64 read_format, u64 *values)
4360 {
4361         struct perf_event_context *ctx = leader->ctx;
4362         struct perf_event *sub;
4363         unsigned long flags;
4364         int n = 1; /* skip @nr */
4365         int ret;
4366 
4367         ret = perf_event_read(leader, true);
4368         if (ret)
4369                 return ret;
4370 
4371         raw_spin_lock_irqsave(&ctx->lock, flags);
4372 
4373         /*
4374          * Since we co-schedule groups, {enabled,running} times of siblings
4375          * will be identical to those of the leader, so we only publish one
4376          * set.
4377          */
4378         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4379                 values[n++] += leader->total_time_enabled +
4380                         atomic64_read(&leader->child_total_time_enabled);
4381         }
4382 
4383         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4384                 values[n++] += leader->total_time_running +
4385                         atomic64_read(&leader->child_total_time_running);
4386         }
4387 
4388         /*
4389          * Write {count,id} tuples for every sibling.
4390          */
4391         values[n++] += perf_event_count(leader);
4392         if (read_format & PERF_FORMAT_ID)
4393                 values[n++] = primary_event_id(leader);
4394 
4395         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
4396                 values[n++] += perf_event_count(sub);
4397                 if (read_format & PERF_FORMAT_ID)
4398                         values[n++] = primary_event_id(sub);
4399         }
4400 
4401         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4402         return 0;
4403 }
4404 
4405 static int perf_read_group(struct perf_event *event,
4406                                    u64 read_format, char __user *buf)
4407 {
4408         struct perf_event *leader = event->group_leader, *child;
4409         struct perf_event_context *ctx = leader->ctx;
4410         int ret;
4411         u64 *values;
4412 
4413         lockdep_assert_held(&ctx->mutex);
4414 
4415         values = kzalloc(event->read_size, GFP_KERNEL);
4416         if (!values)
4417                 return -ENOMEM;
4418 
4419         values[0] = 1 + leader->nr_siblings;
4420 
4421         /*
4422          * By locking the child_mutex of the leader we effectively
4423          * lock the child list of all siblings.. XXX explain how.
4424          */
4425         mutex_lock(&leader->child_mutex);
4426 
4427         ret = __perf_read_group_add(leader, read_format, values);
4428         if (ret)
4429                 goto unlock;
4430 
4431         list_for_each_entry(child, &leader->child_list, child_list) {
4432                 ret = __perf_read_group_add(child, read_format, values);
4433                 if (ret)
4434                         goto unlock;
4435         }
4436 
4437         mutex_unlock(&leader->child_mutex);
4438 
4439         ret = event->read_size;
4440         if (copy_to_user(buf, values, event->read_size))
4441                 ret = -EFAULT;
4442         goto out;
4443 
4444 unlock:
4445         mutex_unlock(&leader->child_mutex);
4446 out:
4447         kfree(values);
4448         return ret;
4449 }
4450 
4451 static int perf_read_one(struct perf_event *event,
4452                                  u64 read_format, char __user *buf)
4453 {
4454         u64 enabled, running;
4455         u64 values[4];
4456         int n = 0;
4457 
4458         values[n++] = __perf_event_read_value(event, &enabled, &running);
4459         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4460                 values[n++] = enabled;
4461         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4462                 values[n++] = running;
4463         if (read_format & PERF_FORMAT_ID)
4464                 values[n++] = primary_event_id(event);
4465 
4466         if (copy_to_user(buf, values, n * sizeof(u64)))
4467                 return -EFAULT;
4468 
4469         return n * sizeof(u64);
4470 }
4471 
4472 static bool is_event_hup(struct perf_event *event)
4473 {
4474         bool no_children;
4475 
4476         if (event->state > PERF_EVENT_STATE_EXIT)
4477                 return false;
4478 
4479         mutex_lock(&event->child_mutex);
4480         no_children = list_empty(&event->child_list);
4481         mutex_unlock(&event->child_mutex);
4482         return no_children;
4483 }
4484 
4485 /*
4486  * Read the performance event - simple non blocking version for now
4487  */
4488 static ssize_t
4489 __perf_read(struct perf_event *event, char __user *buf, size_t count)
4490 {
4491         u64 read_format = event->attr.read_format;
4492         int ret;
4493 
4494         /*
4495          * Return end-of-file for a read on a event that is in
4496          * error state (i.e. because it was pinned but it couldn't be
4497          * scheduled on to the CPU at some point).
4498          */
4499         if (event->state == PERF_EVENT_STATE_ERROR)
4500                 return 0;
4501 
4502         if (count < event->read_size)
4503                 return -ENOSPC;
4504 
4505         WARN_ON_ONCE(event->ctx->parent_ctx);
4506         if (read_format & PERF_FORMAT_GROUP)
4507                 ret = perf_read_group(event, read_format, buf);
4508         else
4509                 ret = perf_read_one(event, read_format, buf);
4510 
4511         return ret;
4512 }
4513 
4514 static ssize_t
4515 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
4516 {
4517         struct perf_event *event = file->private_data;
4518         struct perf_event_context *ctx;
4519         int ret;
4520 
4521         ctx = perf_event_ctx_lock(event);
4522         ret = __perf_read(event, buf, count);
4523         perf_event_ctx_unlock(event, ctx);
4524 
4525         return ret;
4526 }
4527 
4528 static unsigned int perf_poll(struct file *file, poll_table *wait)
4529 {
4530         struct perf_event *event = file->private_data;
4531         struct ring_buffer *rb;
4532         unsigned int events = POLLHUP;
4533 
4534         poll_wait(file, &event->waitq, wait);
4535 
4536         if (is_event_hup(event))
4537                 return events;
4538 
4539         /*
4540          * Pin the event->rb by taking event->mmap_mutex; otherwise
4541          * perf_event_set_output() can swizzle our rb and make us miss wakeups.
4542          */
4543         mutex_lock(&event->mmap_mutex);
4544         rb = event->rb;
4545         if (rb)
4546                 events = atomic_xchg(&rb->poll, 0);
4547         mutex_unlock(&event->mmap_mutex);
4548         return events;
4549 }
4550 
4551 static void _perf_event_reset(struct perf_event *event)
4552 {
4553         (void)perf_event_read(event, false);
4554         local64_set(&event->count, 0);
4555         perf_event_update_userpage(event);
4556 }
4557 
4558 /*
4559  * Holding the top-level event's child_mutex means that any
4560  * descendant process that has inherited this event will block
4561  * in perf_event_exit_event() if it goes to exit, thus satisfying the
4562  * task existence requirements of perf_event_enable/disable.
4563  */
4564 static void perf_event_for_each_child(struct perf_event *event,
4565                                         void (*func)(struct perf_event *))
4566 {
4567         struct perf_event *child;
4568 
4569         WARN_ON_ONCE(event->ctx->parent_ctx);
4570 
4571         mutex_lock(&event->child_mutex);
4572         func(event);
4573         list_for_each_entry(child, &event->child_list, child_list)
4574                 func(child);
4575         mutex_unlock(&event->child_mutex);
4576 }
4577 
4578 static void perf_event_for_each(struct perf_event *event,
4579                                   void (*func)(struct perf_event *))
4580 {
4581         struct perf_event_context *ctx = event->ctx;
4582         struct perf_event *sibling;
4583 
4584         lockdep_assert_held(&ctx->mutex);
4585 
4586         event = event->group_leader;
4587 
4588         perf_event_for_each_child(event, func);
4589         list_for_each_entry(sibling, &event->sibling_list, group_entry)
4590                 perf_event_for_each_child(sibling, func);
4591 }
4592 
4593 static void __perf_event_period(struct perf_event *event,
4594                                 struct perf_cpu_context *cpuctx,
4595                                 struct perf_event_context *ctx,
4596                                 void *info)
4597 {
4598         u64 value = *((u64 *)info);
4599         bool active;
4600 
4601         if (event->attr.freq) {
4602                 event->attr.sample_freq = value;
4603         } else {
4604                 event->attr.sample_period = value;
4605                 event->hw.sample_period = value;
4606         }
4607 
4608         active = (event->state == PERF_EVENT_STATE_ACTIVE);
4609         if (active) {
4610                 perf_pmu_disable(ctx->pmu);
4611                 /*
4612                  * We could be throttled; unthrottle now to avoid the tick
4613                  * trying to unthrottle while we already re-started the event.
4614                  */
4615                 if (event->hw.interrupts == MAX_INTERRUPTS) {
4616                         event->hw.interrupts = 0;
4617                         perf_log_throttle(event, 1);
4618                 }
4619                 event->pmu->stop(event, PERF_EF_UPDATE);
4620         }
4621 
4622         local64_set(&event->hw.period_left, 0);
4623 
4624         if (active) {
4625                 event->pmu->start(event, PERF_EF_RELOAD);
4626                 perf_pmu_enable(ctx->pmu);
4627         }
4628 }
4629 
4630 static int perf_event_period(struct perf_event *event, u64 __user *arg)
4631 {
4632         u64 value;
4633 
4634         if (!is_sampling_event(event))
4635                 return -EINVAL;
4636 
4637         if (copy_from_user(&value, arg, sizeof(value)))
4638                 return -EFAULT;
4639 
4640         if (!value)
4641                 return -EINVAL;
4642 
4643         if (event->attr.freq && value > sysctl_perf_event_sample_rate)
4644                 return -EINVAL;
4645 
4646         event_function_call(event, __perf_event_period, &value);
4647 
4648         return 0;
4649 }
4650 
4651 static const struct file_operations perf_fops;
4652 
4653 static inline int perf_fget_light(int fd, struct fd *p)
4654 {
4655         struct fd f = fdget(fd);
4656         if (!f.file)
4657                 return -EBADF;
4658 
4659         if (f.file->f_op != &perf_fops) {
4660                 fdput(f);
4661                 return -EBADF;
4662         }
4663         *p = f;
4664         return 0;
4665 }
4666 
4667 static int perf_event_set_output(struct perf_event *event,
4668                                  struct perf_event *output_event);
4669 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
4670 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
4671 
4672 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
4673 {
4674         void (*func)(struct perf_event *);
4675         u32 flags = arg;
4676 
4677         switch (cmd) {
4678         case PERF_EVENT_IOC_ENABLE:
4679                 func = _perf_event_enable;
4680                 break;
4681         case PERF_EVENT_IOC_DISABLE:
4682                 func = _perf_event_disable;
4683                 break;
4684         case PERF_EVENT_IOC_RESET:
4685                 func = _perf_event_reset;
4686                 break;
4687 
4688         case PERF_EVENT_IOC_REFRESH:
4689                 return _perf_event_refresh(event, arg);
4690 
4691         case PERF_EVENT_IOC_PERIOD:
4692                 return perf_event_period(event, (u64 __user *)arg);
4693 
4694         case PERF_EVENT_IOC_ID:
4695         {
4696                 u64 id = primary_event_id(event);
4697 
4698                 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
4699                         return -EFAULT;
4700                 return 0;
4701         }
4702 
4703         case PERF_EVENT_IOC_SET_OUTPUT:
4704         {
4705                 int ret;
4706                 if (arg != -1) {
4707                         struct perf_event *output_event;
4708                         struct fd output;
4709                         ret = perf_fget_light(arg, &output);
4710                         if (ret)
4711                                 return ret;
4712                         output_event = output.file->private_data;
4713                         ret = perf_event_set_output(event, output_event);
4714                         fdput(output);
4715                 } else {
4716                         ret = perf_event_set_output(event, NULL);
4717                 }
4718                 return ret;
4719         }
4720 
4721         case PERF_EVENT_IOC_SET_FILTER:
4722                 return perf_event_set_filter(event, (void __user *)arg);
4723 
4724         case PERF_EVENT_IOC_SET_BPF:
4725                 return perf_event_set_bpf_prog(event, arg);
4726 
4727         case PERF_EVENT_IOC_PAUSE_OUTPUT: {
4728                 struct ring_buffer *rb;
4729 
4730                 rcu_read_lock();
4731                 rb = rcu_dereference(event->rb);
4732                 if (!rb || !rb->nr_pages) {
4733                         rcu_read_unlock();
4734                         return -EINVAL;
4735                 }
4736                 rb_toggle_paused(rb, !!arg);
4737                 rcu_read_unlock();
4738                 return 0;
4739         }
4740         default:
4741                 return -ENOTTY;
4742         }
4743 
4744         if (flags & PERF_IOC_FLAG_GROUP)
4745                 perf_event_for_each(event, func);
4746         else
4747                 perf_event_for_each_child(event, func);
4748 
4749         return 0;
4750 }
4751 
4752 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
4753 {
4754         struct perf_event *event = file->private_data;
4755         struct perf_event_context *ctx;
4756         long ret;
4757 
4758         ctx = perf_event_ctx_lock(event);
4759         ret = _perf_ioctl(event, cmd, arg);
4760         perf_event_ctx_unlock(event, ctx);
4761 
4762         return ret;
4763 }
4764 
4765 #ifdef CONFIG_COMPAT
4766 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
4767                                 unsigned long arg)
4768 {
4769         switch (_IOC_NR(cmd)) {
4770         case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
4771         case _IOC_NR(PERF_EVENT_IOC_ID):
4772                 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
4773                 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
4774                         cmd &= ~IOCSIZE_MASK;
4775                         cmd |= sizeof(void *) << IOCSIZE_SHIFT;
4776                 }
4777                 break;
4778         }
4779         return perf_ioctl(file, cmd, arg);
4780 }
4781 #else
4782 # define perf_compat_ioctl NULL
4783 #endif
4784 
4785 int perf_event_task_enable(void)
4786 {
4787         struct perf_event_context *ctx;
4788         struct perf_event *event;
4789 
4790         mutex_lock(&current->perf_event_mutex);
4791         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
4792                 ctx = perf_event_ctx_lock(event);
4793                 perf_event_for_each_child(event, _perf_event_enable);
4794                 perf_event_ctx_unlock(event, ctx);
4795         }
4796         mutex_unlock(&current->perf_event_mutex);
4797 
4798         return 0;
4799 }
4800 
4801 int perf_event_task_disable(void)
4802 {
4803         struct perf_event_context *ctx;
4804         struct perf_event *event;
4805 
4806         mutex_lock(&current->perf_event_mutex);
4807         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
4808                 ctx = perf_event_ctx_lock(event);
4809                 perf_event_for_each_child(event, _perf_event_disable);
4810                 perf_event_ctx_unlock(event, ctx);
4811         }
4812         mutex_unlock(&current->perf_event_mutex);
4813 
4814         return 0;
4815 }
4816 
4817 static int perf_event_index(struct perf_event *event)
4818 {
4819         if (event->hw.state & PERF_HES_STOPPED)
4820                 return 0;
4821 
4822         if (event->state != PERF_EVENT_STATE_ACTIVE)
4823                 return 0;
4824 
4825         return event->pmu->event_idx(event);
4826 }
4827 
4828 static void calc_timer_values(struct perf_event *event,
4829                                 u64 *now,
4830                                 u64 *enabled,
4831                                 u64 *running)
4832 {
4833         u64 ctx_time;
4834 
4835         *now = perf_clock();
4836         ctx_time = event->shadow_ctx_time + *now;
4837         __perf_update_times(event, ctx_time, enabled, running);
4838 }
4839 
4840 static void perf_event_init_userpage(struct perf_event *event)
4841 {
4842         struct perf_event_mmap_page *userpg;
4843         struct ring_buffer *rb;
4844 
4845         rcu_read_lock();
4846         rb = rcu_dereference(event->rb);
4847         if (!rb)
4848                 goto unlock;
4849 
4850         userpg = rb->user_page;
4851 
4852         /* Allow new userspace to detect that bit 0 is deprecated */
4853         userpg->cap_bit0_is_deprecated = 1;
4854         userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
4855         userpg->data_offset = PAGE_SIZE;
4856         userpg->data_size = perf_data_size(rb);
4857 
4858 unlock:
4859         rcu_read_unlock();
4860 }
4861 
4862 void __weak arch_perf_update_userpage(
4863         struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
4864 {
4865 }
4866 
4867 /*
4868  * Callers need to ensure there can be no nesting of this function, otherwise
4869  * the seqlock logic goes bad. We can not serialize this because the arch
4870  * code calls this from NMI context.
4871  */
4872 void perf_event_update_userpage(struct perf_event *event)
4873 {
4874         struct perf_event_mmap_page *userpg;
4875         struct ring_buffer *rb;
4876         u64 enabled, running, now;
4877 
4878         rcu_read_lock();
4879         rb = rcu_dereference(event->rb);
4880         if (!rb)
4881                 goto unlock;
4882 
4883         /*
4884          * compute total_time_enabled, total_time_running
4885          * based on snapshot values taken when the event
4886          * was last scheduled in.
4887          *
4888          * we cannot simply called update_context_time()
4889          * because of locking issue as we can be called in
4890          * NMI context
4891          */
4892         calc_timer_values(event, &now, &enabled, &running);
4893 
4894         userpg = rb->user_page;
4895         /*
4896          * Disable preemption so as to not let the corresponding user-space
4897          * spin too long if we get preempted.
4898          */
4899         preempt_disable();
4900         ++userpg->lock;
4901         barrier();
4902         userpg->index = perf_event_index(event);
4903         userpg->offset = perf_event_count(event);
4904         if (userpg->index)
4905                 userpg->offset -= local64_read(&event->hw.prev_count);
4906 
4907         userpg->time_enabled = enabled +
4908                         atomic64_read(&event->child_total_time_enabled);
4909 
4910         userpg->time_running = running +
4911                         atomic64_read(&event->child_total_time_running);
4912 
4913         arch_perf_update_userpage(event, userpg, now);
4914 
4915         barrier();
4916         ++userpg->lock;
4917         preempt_enable();
4918 unlock:
4919         rcu_read_unlock();
4920 }
4921 
4922 static int perf_mmap_fault(struct vm_fault *vmf)
4923 {
4924         struct perf_event *event = vmf->vma->vm_file->private_data;
4925         struct ring_buffer *rb;
4926         int ret = VM_FAULT_SIGBUS;
4927 
4928         if (vmf->flags & FAULT_FLAG_MKWRITE) {
4929                 if (vmf->pgoff == 0)
4930                         ret = 0;
4931                 return ret;
4932         }
4933 
4934         rcu_read_lock();
4935         rb = rcu_dereference(event->rb);
4936         if (!rb)
4937                 goto unlock;
4938 
4939         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
4940                 goto unlock;
4941 
4942         vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
4943         if (!vmf->page)
4944                 goto unlock;
4945 
4946         get_page(vmf->page);
4947         vmf->page->mapping = vmf->vma->vm_file->f_mapping;
4948         vmf->page->index   = vmf->pgoff;
4949 
4950         ret = 0;
4951 unlock:
4952         rcu_read_unlock();
4953 
4954         return ret;
4955 }
4956 
4957 static void ring_buffer_attach(struct perf_event *event,
4958                                struct ring_buffer *rb)
4959 {
4960         struct ring_buffer *old_rb = NULL;
4961         unsigned long flags;
4962 
4963         if (event->rb) {
4964                 /*
4965                  * Should be impossible, we set this when removing
4966                  * event->rb_entry and wait/clear when adding event->rb_entry.
4967                  */
4968                 WARN_ON_ONCE(event->rcu_pending);
4969 
4970                 old_rb = event->rb;
4971                 spin_lock_irqsave(&old_rb->event_lock, flags);
4972                 list_del_rcu(&event->rb_entry);
4973                 spin_unlock_irqrestore(&old_rb->event_lock, flags);
4974 
4975                 event->rcu_batches = get_state_synchronize_rcu();
4976                 event->rcu_pending = 1;
4977         }
4978 
4979         if (rb) {
4980                 if (event->rcu_pending) {
4981                         cond_synchronize_rcu(event->rcu_batches);
4982                         event->rcu_pending = 0;
4983                 }
4984 
4985                 spin_lock_irqsave(&rb->event_lock, flags);
4986                 list_add_rcu(&event->rb_entry, &rb->event_list);
4987                 spin_unlock_irqrestore(&rb->event_lock, flags);
4988         }
4989 
4990         /*
4991          * Avoid racing with perf_mmap_close(AUX): stop the event
4992          * before swizzling the event::rb pointer; if it's getting
4993          * unmapped, its aux_mmap_count will be 0 and it won't
4994          * restart. See the comment in __perf_pmu_output_stop().
4995          *
4996          * Data will inevitably be lost when set_output is done in
4997          * mid-air, but then again, whoever does it like this is
4998          * not in for the data anyway.
4999          */
5000         if (has_aux(event))
5001                 perf_event_stop(event, 0);
5002 
5003         rcu_assign_pointer(event->rb, rb);
5004 
5005         if (old_rb) {
5006                 ring_buffer_put(old_rb);
5007                 /*
5008                  * Since we detached before setting the new rb, so that we
5009                  * could attach the new rb, we could have missed a wakeup.
5010                  * Provide it now.
5011                  */
5012                 wake_up_all(&event->waitq);
5013         }
5014 }
5015 
5016 static void ring_buffer_wakeup(struct perf_event *event)
5017 {
5018         struct ring_buffer *rb;
5019 
5020         rcu_read_lock();
5021         rb = rcu_dereference(event->rb);
5022         if (rb) {
5023                 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
5024                         wake_up_all(&event->waitq);
5025         }
5026         rcu_read_unlock();
5027 }
5028 
5029 struct ring_buffer *ring_buffer_get(struct perf_event *event)
5030 {
5031         struct ring_buffer *rb;
5032 
5033         rcu_read_lock();
5034         rb = rcu_dereference(event->rb);
5035         if (rb) {
5036                 if (!atomic_inc_not_zero(&rb->refcount))
5037                         rb = NULL;
5038         }
5039         rcu_read_unlock();
5040 
5041         return rb;
5042 }
5043 
5044 void ring_buffer_put(struct ring_buffer *rb)
5045 {
5046         if (!atomic_dec_and_test(&rb->refcount))
5047                 return;
5048 
5049         WARN_ON_ONCE(!list_empty(&rb->event_list));
5050 
5051         call_rcu(&rb->rcu_head, rb_free_rcu);
5052 }
5053 
5054 static void perf_mmap_open(struct vm_area_struct *vma)
5055 {
5056         struct perf_event *event = vma->vm_file->private_data;
5057 
5058         atomic_inc(&event->mmap_count);
5059         atomic_inc(&event->rb->mmap_count);
5060 
5061         if (vma->vm_pgoff)
5062                 atomic_inc(&event->rb->aux_mmap_count);
5063 
5064         if (event->pmu->event_mapped)
5065                 event->pmu->event_mapped(event, vma->vm_mm);
5066 }
5067 
5068 static void perf_pmu_output_stop(struct perf_event *event);
5069 
5070 /*
5071  * A buffer can be mmap()ed multiple times; either directly through the same
5072  * event, or through other events by use of perf_event_set_output().
5073  *
5074  * In order to undo the VM accounting done by perf_mmap() we need to destroy
5075  * the buffer here, where we still have a VM context. This means we need
5076  * to detach all events redirecting to us.
5077  */
5078 static void perf_mmap_close(struct vm_area_struct *vma)
5079 {
5080         struct perf_event *event = vma->vm_file->private_data;
5081 
5082         struct ring_buffer *rb = ring_buffer_get(event);
5083         struct user_struct *mmap_user = rb->mmap_user;
5084         int mmap_locked = rb->mmap_locked;
5085         unsigned long size = perf_data_size(rb);
5086 
5087         if (event->pmu->event_unmapped)
5088                 event->pmu->event_unmapped(event, vma->vm_mm);
5089 
5090         /*
5091          * rb->aux_mmap_count will always drop before rb->mmap_count and
5092          * event->mmap_count, so it is ok to use event->mmap_mutex to
5093          * serialize with perf_mmap here.
5094          */
5095         if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
5096             atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
5097                 /*
5098                  * Stop all AUX events that are writing to this buffer,
5099                  * so that we can free its AUX pages and corresponding PMU
5100                  * data. Note that after rb::aux_mmap_count dropped to zero,
5101                  * they won't start any more (see perf_aux_output_begin()).
5102                  */
5103                 perf_pmu_output_stop(event);
5104 
5105                 /* now it's safe to free the pages */
5106                 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
5107                 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
5108 
5109                 /* this has to be the last one */
5110                 rb_free_aux(rb);
5111                 WARN_ON_ONCE(atomic_read(&rb->aux_refcount));
5112 
5113                 mutex_unlock(&event->mmap_mutex);
5114         }
5115 
5116         atomic_dec(&rb->mmap_count);
5117 
5118         if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
5119                 goto out_put;
5120 
5121         ring_buffer_attach(event, NULL);
5122         mutex_unlock(&event->mmap_mutex);
5123 
5124         /* If there's still other mmap()s of this buffer, we're done. */
5125         if (atomic_read(&rb->mmap_count))
5126                 goto out_put;
5127 
5128         /*
5129          * No other mmap()s, detach from all other events that might redirect
5130          * into the now unreachable buffer. Somewhat complicated by the
5131          * fact that rb::event_lock otherwise nests inside mmap_mutex.
5132          */
5133 again:
5134         rcu_read_lock();
5135         list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
5136                 if (!atomic_long_inc_not_zero(&event->refcount)) {
5137                         /*
5138                          * This event is en-route to free_event() which will
5139                          * detach it and remove it from the list.
5140                          */
5141                         continue;
5142                 }
5143                 rcu_read_unlock();
5144 
5145                 mutex_lock(&event->mmap_mutex);
5146                 /*
5147                  * Check we didn't race with perf_event_set_output() which can
5148                  * swizzle the rb from under us while we were waiting to
5149                  * acquire mmap_mutex.
5150                  *
5151                  * If we find a different rb; ignore this event, a next
5152                  * iteration will no longer find it on the list. We have to
5153                  * still restart the iteration to make sure we're not now
5154                  * iterating the wrong list.
5155                  */
5156                 if (event->rb == rb)
5157                         ring_buffer_attach(event, NULL);
5158 
5159                 mutex_unlock(&event->mmap_mutex);
5160                 put_event(event);
5161 
5162                 /*
5163                  * Restart the iteration; either we're on the wrong list or
5164                  * destroyed its integrity by doing a deletion.
5165                  */
5166                 goto again;
5167         }
5168         rcu_read_unlock();
5169 
5170         /*
5171          * It could be there's still a few 0-ref events on the list; they'll
5172          * get cleaned up by free_event() -- they'll also still have their
5173          * ref on the rb and will free it whenever they are done with it.
5174          *
5175          * Aside from that, this buffer is 'fully' detached and unmapped,
5176          * undo the VM accounting.
5177          */
5178 
5179         atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
5180         vma->vm_mm->pinned_vm -= mmap_locked;
5181         free_uid(mmap_user);
5182 
5183 out_put:
5184         ring_buffer_put(rb); /* could be last */
5185 }
5186 
5187 static const struct vm_operations_struct perf_mmap_vmops = {
5188         .open           = perf_mmap_open,
5189         .close          = perf_mmap_close, /* non mergable */
5190         .fault          = perf_mmap_fault,
5191         .page_mkwrite   = perf_mmap_fault,
5192 };
5193 
5194 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
5195 {
5196         struct perf_event *event = file->private_data;
5197         unsigned long user_locked, user_lock_limit;
5198         struct user_struct *user = current_user();
5199         unsigned long locked, lock_limit;
5200         struct ring_buffer *rb = NULL;
5201         unsigned long vma_size;
5202         unsigned long nr_pages;
5203         long user_extra = 0, extra = 0;
5204         int ret = 0, flags = 0;
5205 
5206         /*
5207          * Don't allow mmap() of inherited per-task counters. This would
5208          * create a performance issue due to all children writing to the
5209          * same rb.
5210          */
5211         if (event->cpu == -1 && event->attr.inherit)
5212                 return -EINVAL;
5213 
5214         if (!(vma->vm_flags & VM_SHARED))
5215                 return -EINVAL;
5216 
5217         vma_size = vma->vm_end - vma->vm_start;
5218 
5219         if (vma->vm_pgoff == 0) {
5220                 nr_pages = (vma_size / PAGE_SIZE) - 1;
5221         } else {
5222                 /*
5223                  * AUX area mapping: if rb->aux_nr_pages != 0, it's already
5224                  * mapped, all subsequent mappings should have the same size
5225                  * and offset. Must be above the normal perf buffer.
5226                  */
5227                 u64 aux_offset, aux_size;
5228 
5229                 if (!event->rb)
5230                         return -EINVAL;
5231 
5232                 nr_pages = vma_size / PAGE_SIZE;
5233 
5234                 mutex_lock(&event->mmap_mutex);
5235                 ret = -EINVAL;
5236 
5237                 rb = event->rb;
5238                 if (!rb)
5239                         goto aux_unlock;
5240 
5241                 aux_offset = READ_ONCE(rb->user_page->aux_offset);
5242                 aux_size = READ_ONCE(rb->user_page->aux_size);
5243 
5244                 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
5245                         goto aux_unlock;
5246 
5247                 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
5248                         goto aux_unlock;
5249 
5250                 /* already mapped with a different offset */
5251                 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
5252                         goto aux_unlock;
5253 
5254                 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
5255                         goto aux_unlock;
5256 
5257                 /* already mapped with a different size */
5258                 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
5259                         goto aux_unlock;
5260 
5261                 if (!is_power_of_2(nr_pages))
5262                         goto aux_unlock;
5263 
5264                 if (!atomic_inc_not_zero(&rb->mmap_count))
5265                         goto aux_unlock;
5266 
5267                 if (rb_has_aux(rb)) {
5268                         atomic_inc(&rb->aux_mmap_count);
5269                         ret = 0;
5270                         goto unlock;
5271                 }
5272 
5273                 atomic_set(&rb->aux_mmap_count, 1);
5274                 user_extra = nr_pages;
5275 
5276                 goto accounting;
5277         }
5278 
5279         /*
5280          * If we have rb pages ensure they're a power-of-two number, so we
5281          * can do bitmasks instead of modulo.
5282          */
5283         if (nr_pages != 0 && !is_power_of_2(nr_pages))
5284                 return -EINVAL;
5285 
5286         if (vma_size != PAGE_SIZE * (1 + nr_pages))
5287                 return -EINVAL;
5288 
5289         WARN_ON_ONCE(event->ctx->parent_ctx);
5290 again:
5291         mutex_lock(&event->mmap_mutex);
5292         if (event->rb) {
5293                 if (event->rb->nr_pages != nr_pages) {
5294                         ret = -EINVAL;
5295                         goto unlock;
5296                 }
5297 
5298                 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
5299                         /*
5300                          * Raced against perf_mmap_close() through
5301                          * perf_event_set_output(). Try again, hope for better
5302                          * luck.
5303                          */
5304                         mutex_unlock(&event->mmap_mutex);
5305                         goto again;
5306                 }
5307 
5308                 goto unlock;
5309         }
5310 
5311         user_extra = nr_pages + 1;
5312 
5313 accounting:
5314         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
5315 
5316         /*
5317          * Increase the limit linearly with more CPUs:
5318          */
5319         user_lock_limit *= num_online_cpus();
5320 
5321         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
5322 
5323         if (user_locked > user_lock_limit)
5324                 extra = user_locked - user_lock_limit;
5325 
5326         lock_limit = rlimit(RLIMIT_MEMLOCK);
5327         lock_limit >>= PAGE_SHIFT;
5328         locked = vma->vm_mm->pinned_vm + extra;
5329 
5330         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
5331                 !capable(CAP_IPC_LOCK)) {
5332                 ret = -EPERM;
5333                 goto unlock;
5334         }
5335 
5336         WARN_ON(!rb && event->rb);
5337 
5338         if (vma->vm_flags & VM_WRITE)
5339                 flags |= RING_BUFFER_WRITABLE;
5340 
5341         if (!rb) {
5342                 rb = rb_alloc(nr_pages,
5343                               event->attr.watermark ? event->attr.wakeup_watermark : 0,
5344                               event->cpu, flags);
5345 
5346                 if (!rb) {
5347                         ret = -ENOMEM;
5348                         goto unlock;
5349                 }
5350 
5351                 atomic_set(&rb->mmap_count, 1);
5352                 rb->mmap_user = get_current_user();
5353                 rb->mmap_locked = extra;
5354 
5355                 ring_buffer_attach(event, rb);
5356 
5357                 perf_event_init_userpage(event);
5358                 perf_event_update_userpage(event);
5359         } else {
5360                 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
5361                                    event->attr.aux_watermark, flags);
5362                 if (!ret)
5363                         rb->aux_mmap_locked = extra;
5364         }
5365 
5366 unlock:
5367         if (!ret) {
5368                 atomic_long_add(user_extra, &user->locked_vm);
5369                 vma->vm_mm->pinned_vm += extra;
5370 
5371                 atomic_inc(&event->mmap_count);
5372         } else if (rb) {
5373                 atomic_dec(&rb->mmap_count);
5374         }
5375 aux_unlock:
5376         mutex_unlock(&event->mmap_mutex);
5377 
5378         /*
5379          * Since pinned accounting is per vm we cannot allow fork() to copy our
5380          * vma.
5381          */
5382         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
5383         vma->vm_ops = &perf_mmap_vmops;
5384 
5385         if (event->pmu->event_mapped)
5386                 event->pmu->event_mapped(event, vma->vm_mm);
5387 
5388         return ret;
5389 }
5390 
5391 static int perf_fasync(int fd, struct file *filp, int on)
5392 {
5393         struct inode *inode = file_inode(filp);
5394         struct perf_event *event = filp->private_data;
5395         int retval;
5396 
5397         inode_lock(inode);
5398         retval = fasync_helper(fd, filp, on, &event->fasync);
5399         inode_unlock(inode);
5400 
5401         if (retval < 0)
5402                 return retval;
5403 
5404         return 0;
5405 }
5406 
5407 static const struct file_operations perf_fops = {
5408         .llseek                 = no_llseek,
5409         .release                = perf_release,
5410         .read                   = perf_read,
5411         .poll                   = perf_poll,
5412         .unlocked_ioctl         = perf_ioctl,
5413         .compat_ioctl           = perf_compat_ioctl,
5414         .mmap                   = perf_mmap,
5415         .fasync                 = perf_fasync,
5416 };
5417 
5418 /*
5419  * Perf event wakeup
5420  *
5421  * If there's data, ensure we set the poll() state and publish everything
5422  * to user-space before waking everybody up.
5423  */
5424 
5425 static inline struct fasync_struct **perf_event_fasync(struct perf_event *event)
5426 {
5427         /* only the parent has fasync state */
5428         if (event->parent)
5429                 event = event->parent;
5430         return &event->fasync;
5431 }
5432 
5433 void perf_event_wakeup(struct perf_event *event)
5434 {
5435         ring_buffer_wakeup(event);
5436 
5437         if (event->pending_kill) {
5438                 kill_fasync(perf_event_fasync(event), SIGIO, event->pending_kill);
5439                 event->pending_kill = 0;
5440         }
5441 }
5442 
5443 static void perf_pending_event(struct irq_work *entry)
5444 {
5445         struct perf_event *event = container_of(entry,
5446                         struct perf_event, pending);
5447         int rctx;
5448 
5449         rctx = perf_swevent_get_recursion_context();
5450         /*
5451          * If we 'fail' here, that's OK, it means recursion is already disabled
5452          * and we won't recurse 'further'.
5453          */
5454 
5455         if (event->pending_disable) {
5456                 event->pending_disable = 0;
5457                 perf_event_disable_local(event);
5458         }
5459 
5460         if (event->pending_wakeup) {
5461                 event->pending_wakeup = 0;
5462                 perf_event_wakeup(event);
5463         }
5464 
5465         if (rctx >= 0)
5466                 perf_swevent_put_recursion_context(rctx);
5467 }
5468 
5469 /*
5470  * We assume there is only KVM supporting the callbacks.
5471  * Later on, we might change it to a list if there is
5472  * another virtualization implementation supporting the callbacks.
5473  */
5474 struct perf_guest_info_callbacks *perf_guest_cbs;
5475 
5476 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5477 {
5478         perf_guest_cbs = cbs;
5479         return 0;
5480 }
5481 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
5482 
5483 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5484 {
5485         perf_guest_cbs = NULL;
5486         return 0;
5487 }
5488 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
5489 
5490 static void
5491 perf_output_sample_regs(struct perf_output_handle *handle,
5492                         struct pt_regs *regs, u64 mask)
5493 {
5494         int bit;
5495         DECLARE_BITMAP(_mask, 64);
5496 
5497         bitmap_from_u64(_mask, mask);
5498         for_each_set_bit(bit, _mask, sizeof(mask) * BITS_PER_BYTE) {
5499                 u64 val;
5500 
5501                 val = perf_reg_value(regs, bit);
5502                 perf_output_put(handle, val);
5503         }
5504 }
5505 
5506 static void perf_sample_regs_user(struct perf_regs *regs_user,
5507                                   struct pt_regs *regs,
5508                                   struct pt_regs *regs_user_copy)
5509 {
5510         if (user_mode(regs)) {
5511                 regs_user->abi = perf_reg_abi(current);
5512                 regs_user->regs = regs;
5513         } else if (current->mm) {
5514                 perf_get_regs_user(regs_user, regs, regs_user_copy);
5515         } else {
5516                 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
5517                 regs_user->regs = NULL;
5518         }
5519 }
5520 
5521 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
5522                                   struct pt_regs *regs)
5523 {
5524         regs_intr->regs = regs;
5525         regs_intr->abi  = perf_reg_abi(current);
5526 }
5527 
5528 
5529 /*
5530  * Get remaining task size from user stack pointer.
5531  *
5532  * It'd be better to take stack vma map and limit this more
5533  * precisly, but there's no way to get it safely under interrupt,
5534  * so using TASK_SIZE as limit.
5535  */
5536 static u64 perf_ustack_task_size(struct pt_regs *regs)
5537 {
5538         unsigned long addr = perf_user_stack_pointer(regs);
5539 
5540         if (!addr || addr >= TASK_SIZE)
5541                 return 0;
5542 
5543         return TASK_SIZE - addr;
5544 }
5545 
5546 static u16
5547 perf_sample_ustack_size(u16 stack_size, u16 header_size,
5548                         struct pt_regs *regs)
5549 {
5550         u64 task_size;
5551 
5552         /* No regs, no stack pointer, no dump. */
5553         if (!regs)
5554                 return 0;
5555 
5556         /*
5557          * Check if we fit in with the requested stack size into the:
5558          * - TASK_SIZE
5559          *   If we don't, we limit the size to the TASK_SIZE.
5560          *
5561          * - remaining sample size
5562          *   If we don't, we customize the stack size to
5563          *   fit in to the remaining sample size.
5564          */
5565 
5566         task_size  = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
5567         stack_size = min(stack_size, (u16) task_size);
5568 
5569         /* Current header size plus static size and dynamic size. */
5570         header_size += 2 * sizeof(u64);
5571 
5572         /* Do we fit in with the current stack dump size? */
5573         if ((u16) (header_size + stack_size) < header_size) {
5574                 /*
5575                  * If we overflow the maximum size for the sample,
5576                  * we customize the stack dump size to fit in.
5577                  */
5578                 stack_size = USHRT_MAX - header_size - sizeof(u64);
5579                 stack_size = round_up(stack_size, sizeof(u64));
5580         }
5581 
5582         return stack_size;
5583 }
5584 
5585 static void
5586 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
5587                           struct pt_regs *regs)
5588 {
5589         /* Case of a kernel thread, nothing to dump */
5590         if (!regs) {
5591                 u64 size = 0;
5592                 perf_output_put(handle, size);
5593         } else {
5594                 unsigned long sp;
5595                 unsigned int rem;
5596                 u64 dyn_size;
5597 
5598                 /*
5599                  * We dump:
5600                  * static size
5601                  *   - the size requested by user or the best one we can fit
5602                  *     in to the sample max size
5603                  * data
5604                  *   - user stack dump data
5605                  * dynamic size
5606                  *   - the actual dumped size
5607                  */
5608 
5609                 /* Static size. */
5610                 perf_output_put(handle, dump_size);
5611 
5612                 /* Data. */
5613                 sp = perf_user_stack_pointer(regs);
5614                 rem = __output_copy_user(handle, (void *) sp, dump_size);
5615                 dyn_size = dump_size - rem;
5616 
5617                 perf_output_skip(handle, rem);
5618 
5619                 /* Dynamic size. */
5620                 perf_output_put(handle, dyn_size);
5621         }
5622 }
5623 
5624 static void __perf_event_header__init_id(struct perf_event_header *header,
5625                                          struct perf_sample_data *data,
5626                                          struct perf_event *event)
5627 {
5628         u64 sample_type = event->attr.sample_type;
5629 
5630         data->type = sample_type;
5631         header->size += event->id_header_size;
5632 
5633         if (sample_type & PERF_SAMPLE_TID) {
5634                 /* namespace issues */
5635                 data->tid_entry.pid = perf_event_pid(event, current);
5636                 data->tid_entry.tid = perf_event_tid(event, current);
5637         }
5638 
5639         if (sample_type & PERF_SAMPLE_TIME)
5640                 data->time = perf_event_clock(event);
5641 
5642         if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
5643                 data->id = primary_event_id(event);
5644 
5645         if (sample_type & PERF_SAMPLE_STREAM_ID)
5646                 data->stream_id = event->id;
5647 
5648         if (sample_type & PERF_SAMPLE_CPU) {
5649                 data->cpu_entry.cpu      = raw_smp_processor_id();
5650                 data->cpu_entry.reserved = 0;
5651         }
5652 }
5653 
5654 void perf_event_header__init_id(struct perf_event_header *header,
5655                                 struct perf_sample_data *data,
5656                                 struct perf_event *event)
5657 {
5658         if (event->attr.sample_id_all)
5659                 __perf_event_header__init_id(header, data, event);
5660 }
5661 
5662 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
5663                                            struct perf_sample_data *data)
5664 {
5665         u64 sample_type = data->type;
5666 
5667         if (sample_type & PERF_SAMPLE_TID)
5668                 perf_output_put(handle, data->tid_entry);
5669 
5670         if (sample_type & PERF_SAMPLE_TIME)
5671                 perf_output_put(handle, data->time);
5672 
5673         if (sample_type & PERF_SAMPLE_ID)
5674                 perf_output_put(handle, data->id);
5675 
5676         if (sample_type & PERF_SAMPLE_STREAM_ID)
5677                 perf_output_put(handle, data->stream_id);
5678 
5679         if (sample_type & PERF_SAMPLE_CPU)
5680                 perf_output_put(handle, data->cpu_entry);
5681 
5682         if (sample_type & PERF_SAMPLE_IDENTIFIER)
5683                 perf_output_put(handle, data->id);
5684 }
5685 
5686 void perf_event__output_id_sample(struct perf_event *event,
5687                                   struct perf_output_handle *handle,
5688                                   struct perf_sample_data *sample)
5689 {
5690         if (event->attr.sample_id_all)
5691                 __perf_event__output_id_sample(handle, sample);
5692 }
5693 
5694 static void perf_output_read_one(struct perf_output_handle *handle,
5695                                  struct perf_event *event,
5696                                  u64 enabled, u64 running)
5697 {
5698         u64 read_format = event->attr.read_format;
5699         u64 values[4];
5700         int n = 0;
5701 
5702         values[n++] = perf_event_count(event);
5703         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
5704                 values[n++] = enabled +
5705                         atomic64_read(&event->child_total_time_enabled);
5706         }
5707         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
5708                 values[n++] = running +
5709                         atomic64_read(&event->child_total_time_running);
5710         }
5711         if (read_format & PERF_FORMAT_ID)
5712                 values[n++] = primary_event_id(event);
5713 
5714         __output_copy(handle, values, n * sizeof(u64));
5715 }
5716 
5717 static void perf_output_read_group(struct perf_output_handle *handle,
5718                             struct perf_event *event,
5719                             u64 enabled, u64 running)
5720 {
5721         struct perf_event *leader = event->group_leader, *sub;
5722         u64 read_format = event->attr.read_format;
5723         u64 values[5];
5724         int n = 0;
5725 
5726         values[n++] = 1 + leader->nr_siblings;
5727 
5728         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
5729                 values[n++] = enabled;
5730 
5731         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
5732                 values[n++] = running;
5733 
5734         if (leader != event)
5735                 leader->pmu->read(leader);
5736 
5737         values[n++] = perf_event_count(leader);
5738         if (read_format & PERF_FORMAT_ID)
5739                 values[n++] = primary_event_id(leader);
5740 
5741         __output_copy(handle, values, n * sizeof(u64));
5742 
5743         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
5744                 n = 0;
5745 
5746                 if ((sub != event) &&
5747                     (sub->state == PERF_EVENT_STATE_ACTIVE))
5748                         sub->pmu->read(sub);
5749 
5750                 values[n++] = perf_event_count(sub);
5751                 if (read_format & PERF_FORMAT_ID)
5752                         values[n++] = primary_event_id(sub);
5753 
5754                 __output_copy(handle, values, n * sizeof(u64));
5755         }
5756 }
5757 
5758 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
5759                                  PERF_FORMAT_TOTAL_TIME_RUNNING)
5760 
5761 /*
5762  * XXX PERF_SAMPLE_READ vs inherited events seems difficult.
5763  *
5764  * The problem is that its both hard and excessively expensive to iterate the
5765  * child list, not to mention that its impossible to IPI the children running
5766  * on another CPU, from interrupt/NMI context.
5767  */
5768 static void perf_output_read(struct perf_output_handle *handle,
5769                              struct perf_event *event)
5770 {
5771         u64 enabled = 0, running = 0, now;
5772         u64 read_format = event->attr.read_format;
5773 
5774         /*
5775          * compute total_time_enabled, total_time_running
5776          * based on snapshot values taken when the event
5777          * was last scheduled in.
5778          *
5779          * we cannot simply called update_context_time()
5780          * because of locking issue as we are called in
5781          * NMI context
5782          */
5783         if (read_format & PERF_FORMAT_TOTAL_TIMES)
5784                 calc_timer_values(event, &now, &enabled, &running);
5785 
5786         if (event->attr.read_format & PERF_FORMAT_GROUP)
5787                 perf_output_read_group(handle, event, enabled, running);
5788         else
5789                 perf_output_read_one(handle, event, enabled, running);
5790 }
5791 
5792 void perf_output_sample(struct perf_output_handle *handle,
5793                         struct perf_event_header *header,
5794                         struct perf_sample_data *data,
5795                         struct perf_event *event)
5796 {
5797         u64 sample_type = data->type;
5798 
5799         perf_output_put(handle, *header);
5800 
5801         if (sample_type & PERF_SAMPLE_IDENTIFIER)
5802                 perf_output_put(handle, data->id);
5803