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

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

Version: ~ [ linux-4.19-rc7 ] ~ [ linux-4.18.16 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.78 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.135 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.162 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.124 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.59 ] ~ [ 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-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.31.14 ] ~ [ linux-2.6.30.10 ] ~ [ linux-2.6.29.6 ] ~ [ linux-2.6.28.10 ] ~ [ linux-2.6.27.62 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Performance events core code:
  3  *
  4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
  5  *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
  6  *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
  7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
  8  *
  9  * For licensing details see kernel-base/COPYING
 10  */
 11 
 12 #include <linux/fs.h>
 13 #include <linux/mm.h>
 14 #include <linux/cpu.h>
 15 #include <linux/smp.h>
 16 #include <linux/idr.h>
 17 #include <linux/file.h>
 18 #include <linux/poll.h>
 19 #include <linux/slab.h>
 20 #include <linux/hash.h>
 21 #include <linux/tick.h>
 22 #include <linux/sysfs.h>
 23 #include <linux/dcache.h>
 24 #include <linux/percpu.h>
 25 #include <linux/ptrace.h>
 26 #include <linux/reboot.h>
 27 #include <linux/vmstat.h>
 28 #include <linux/device.h>
 29 #include <linux/export.h>
 30 #include <linux/vmalloc.h>
 31 #include <linux/hardirq.h>
 32 #include <linux/rculist.h>
 33 #include <linux/uaccess.h>
 34 #include <linux/syscalls.h>
 35 #include <linux/anon_inodes.h>
 36 #include <linux/kernel_stat.h>
 37 #include <linux/cgroup.h>
 38 #include <linux/perf_event.h>
 39 #include <linux/trace_events.h>
 40 #include <linux/hw_breakpoint.h>
 41 #include <linux/mm_types.h>
 42 #include <linux/module.h>
 43 #include <linux/mman.h>
 44 #include <linux/compat.h>
 45 #include <linux/bpf.h>
 46 #include <linux/filter.h>
 47 #include <linux/namei.h>
 48 #include <linux/parser.h>
 49 #include <linux/sched/clock.h>
 50 #include <linux/sched/mm.h>
 51 #include <linux/proc_ns.h>
 52 #include <linux/mount.h>
 53 
 54 #include "internal.h"
 55 
 56 #include <asm/irq_regs.h>
 57 
 58 typedef int (*remote_function_f)(void *);
 59 
 60 struct remote_function_call {
 61         struct task_struct      *p;
 62         remote_function_f       func;
 63         void                    *info;
 64         int                     ret;
 65 };
 66 
 67 static void remote_function(void *data)
 68 {
 69         struct remote_function_call *tfc = data;
 70         struct task_struct *p = tfc->p;
 71 
 72         if (p) {
 73                 /* -EAGAIN */
 74                 if (task_cpu(p) != smp_processor_id())
 75                         return;
 76 
 77                 /*
 78                  * Now that we're on right CPU with IRQs disabled, we can test
 79                  * if we hit the right task without races.
 80                  */
 81 
 82                 tfc->ret = -ESRCH; /* No such (running) process */
 83                 if (p != current)
 84                         return;
 85         }
 86 
 87         tfc->ret = tfc->func(tfc->info);
 88 }
 89 
 90 /**
 91  * task_function_call - call a function on the cpu on which a task runs
 92  * @p:          the task to evaluate
 93  * @func:       the function to be called
 94  * @info:       the function call argument
 95  *
 96  * Calls the function @func when the task is currently running. This might
 97  * be on the current CPU, which just calls the function directly
 98  *
 99  * returns: @func return value, or
100  *          -ESRCH  - when the process isn't running
101  *          -EAGAIN - when the process moved away
102  */
103 static int
104 task_function_call(struct task_struct *p, remote_function_f func, void *info)
105 {
106         struct remote_function_call data = {
107                 .p      = p,
108                 .func   = func,
109                 .info   = info,
110                 .ret    = -EAGAIN,
111         };
112         int ret;
113 
114         do {
115                 ret = smp_call_function_single(task_cpu(p), remote_function, &data, 1);
116                 if (!ret)
117                         ret = data.ret;
118         } while (ret == -EAGAIN);
119 
120         return ret;
121 }
122 
123 /**
124  * cpu_function_call - call a function on the cpu
125  * @func:       the function to be called
126  * @info:       the function call argument
127  *
128  * Calls the function @func on the remote cpu.
129  *
130  * returns: @func return value or -ENXIO when the cpu is offline
131  */
132 static int cpu_function_call(int cpu, remote_function_f func, void *info)
133 {
134         struct remote_function_call data = {
135                 .p      = NULL,
136                 .func   = func,
137                 .info   = info,
138                 .ret    = -ENXIO, /* No such CPU */
139         };
140 
141         smp_call_function_single(cpu, remote_function, &data, 1);
142 
143         return data.ret;
144 }
145 
146 static inline struct perf_cpu_context *
147 __get_cpu_context(struct perf_event_context *ctx)
148 {
149         return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
150 }
151 
152 static void perf_ctx_lock(struct perf_cpu_context *cpuctx,
153                           struct perf_event_context *ctx)
154 {
155         raw_spin_lock(&cpuctx->ctx.lock);
156         if (ctx)
157                 raw_spin_lock(&ctx->lock);
158 }
159 
160 static void perf_ctx_unlock(struct perf_cpu_context *cpuctx,
161                             struct perf_event_context *ctx)
162 {
163         if (ctx)
164                 raw_spin_unlock(&ctx->lock);
165         raw_spin_unlock(&cpuctx->ctx.lock);
166 }
167 
168 #define TASK_TOMBSTONE ((void *)-1L)
169 
170 static bool is_kernel_event(struct perf_event *event)
171 {
172         return READ_ONCE(event->owner) == TASK_TOMBSTONE;
173 }
174 
175 /*
176  * On task ctx scheduling...
177  *
178  * When !ctx->nr_events a task context will not be scheduled. This means
179  * we can disable the scheduler hooks (for performance) without leaving
180  * pending task ctx state.
181  *
182  * This however results in two special cases:
183  *
184  *  - removing the last event from a task ctx; this is relatively straight
185  *    forward and is done in __perf_remove_from_context.
186  *
187  *  - adding the first event to a task ctx; this is tricky because we cannot
188  *    rely on ctx->is_active and therefore cannot use event_function_call().
189  *    See perf_install_in_context().
190  *
191  * If ctx->nr_events, then ctx->is_active and cpuctx->task_ctx are set.
192  */
193 
194 typedef void (*event_f)(struct perf_event *, struct perf_cpu_context *,
195                         struct perf_event_context *, void *);
196 
197 struct event_function_struct {
198         struct perf_event *event;
199         event_f func;
200         void *data;
201 };
202 
203 static int event_function(void *info)
204 {
205         struct event_function_struct *efs = info;
206         struct perf_event *event = efs->event;
207         struct perf_event_context *ctx = event->ctx;
208         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
209         struct perf_event_context *task_ctx = cpuctx->task_ctx;
210         int ret = 0;
211 
212         lockdep_assert_irqs_disabled();
213 
214         perf_ctx_lock(cpuctx, task_ctx);
215         /*
216          * Since we do the IPI call without holding ctx->lock things can have
217          * changed, double check we hit the task we set out to hit.
218          */
219         if (ctx->task) {
220                 if (ctx->task != current) {
221                         ret = -ESRCH;
222                         goto unlock;
223                 }
224 
225                 /*
226                  * We only use event_function_call() on established contexts,
227                  * and event_function() is only ever called when active (or
228                  * rather, we'll have bailed in task_function_call() or the
229                  * above ctx->task != current test), therefore we must have
230                  * ctx->is_active here.
231                  */
232                 WARN_ON_ONCE(!ctx->is_active);
233                 /*
234                  * And since we have ctx->is_active, cpuctx->task_ctx must
235                  * match.
236                  */
237                 WARN_ON_ONCE(task_ctx != ctx);
238         } else {
239                 WARN_ON_ONCE(&cpuctx->ctx != ctx);
240         }
241 
242         efs->func(event, cpuctx, ctx, efs->data);
243 unlock:
244         perf_ctx_unlock(cpuctx, task_ctx);
245 
246         return ret;
247 }
248 
249 static void event_function_call(struct perf_event *event, event_f func, void *data)
250 {
251         struct perf_event_context *ctx = event->ctx;
252         struct task_struct *task = READ_ONCE(ctx->task); /* verified in event_function */
253         struct event_function_struct efs = {
254                 .event = event,
255                 .func = func,
256                 .data = data,
257         };
258 
259         if (!event->parent) {
260                 /*
261                  * If this is a !child event, we must hold ctx::mutex to
262                  * stabilize the the event->ctx relation. See
263                  * perf_event_ctx_lock().
264                  */
265                 lockdep_assert_held(&ctx->mutex);
266         }
267 
268         if (!task) {
269                 cpu_function_call(event->cpu, event_function, &efs);
270                 return;
271         }
272 
273         if (task == TASK_TOMBSTONE)
274                 return;
275 
276 again:
277         if (!task_function_call(task, event_function, &efs))
278                 return;
279 
280         raw_spin_lock_irq(&ctx->lock);
281         /*
282          * Reload the task pointer, it might have been changed by
283          * a concurrent perf_event_context_sched_out().
284          */
285         task = ctx->task;
286         if (task == TASK_TOMBSTONE) {
287                 raw_spin_unlock_irq(&ctx->lock);
288                 return;
289         }
290         if (ctx->is_active) {
291                 raw_spin_unlock_irq(&ctx->lock);
292                 goto again;
293         }
294         func(event, NULL, ctx, data);
295         raw_spin_unlock_irq(&ctx->lock);
296 }
297 
298 /*
299  * Similar to event_function_call() + event_function(), but hard assumes IRQs
300  * are already disabled and we're on the right CPU.
301  */
302 static void event_function_local(struct perf_event *event, event_f func, void *data)
303 {
304         struct perf_event_context *ctx = event->ctx;
305         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
306         struct task_struct *task = READ_ONCE(ctx->task);
307         struct perf_event_context *task_ctx = NULL;
308 
309         lockdep_assert_irqs_disabled();
310 
311         if (task) {
312                 if (task == TASK_TOMBSTONE)
313                         return;
314 
315                 task_ctx = ctx;
316         }
317 
318         perf_ctx_lock(cpuctx, task_ctx);
319 
320         task = ctx->task;
321         if (task == TASK_TOMBSTONE)
322                 goto unlock;
323 
324         if (task) {
325                 /*
326                  * We must be either inactive or active and the right task,
327                  * otherwise we're screwed, since we cannot IPI to somewhere
328                  * else.
329                  */
330                 if (ctx->is_active) {
331                         if (WARN_ON_ONCE(task != current))
332                                 goto unlock;
333 
334                         if (WARN_ON_ONCE(cpuctx->task_ctx != ctx))
335                                 goto unlock;
336                 }
337         } else {
338                 WARN_ON_ONCE(&cpuctx->ctx != ctx);
339         }
340 
341         func(event, cpuctx, ctx, data);
342 unlock:
343         perf_ctx_unlock(cpuctx, task_ctx);
344 }
345 
346 #define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
347                        PERF_FLAG_FD_OUTPUT  |\
348                        PERF_FLAG_PID_CGROUP |\
349                        PERF_FLAG_FD_CLOEXEC)
350 
351 /*
352  * branch priv levels that need permission checks
353  */
354 #define PERF_SAMPLE_BRANCH_PERM_PLM \
355         (PERF_SAMPLE_BRANCH_KERNEL |\
356          PERF_SAMPLE_BRANCH_HV)
357 
358 enum event_type_t {
359         EVENT_FLEXIBLE = 0x1,
360         EVENT_PINNED = 0x2,
361         EVENT_TIME = 0x4,
362         /* see ctx_resched() for details */
363         EVENT_CPU = 0x8,
364         EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
365 };
366 
367 /*
368  * perf_sched_events : >0 events exist
369  * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
370  */
371 
372 static void perf_sched_delayed(struct work_struct *work);
373 DEFINE_STATIC_KEY_FALSE(perf_sched_events);
374 static DECLARE_DELAYED_WORK(perf_sched_work, perf_sched_delayed);
375 static DEFINE_MUTEX(perf_sched_mutex);
376 static atomic_t perf_sched_count;
377 
378 static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
379 static DEFINE_PER_CPU(int, perf_sched_cb_usages);
380 static DEFINE_PER_CPU(struct pmu_event_list, pmu_sb_events);
381 
382 static atomic_t nr_mmap_events __read_mostly;
383 static atomic_t nr_comm_events __read_mostly;
384 static atomic_t nr_namespaces_events __read_mostly;
385 static atomic_t nr_task_events __read_mostly;
386 static atomic_t nr_freq_events __read_mostly;
387 static atomic_t nr_switch_events __read_mostly;
388 
389 static LIST_HEAD(pmus);
390 static DEFINE_MUTEX(pmus_lock);
391 static struct srcu_struct pmus_srcu;
392 static cpumask_var_t perf_online_mask;
393 
394 /*
395  * perf event paranoia level:
396  *  -1 - not paranoid at all
397  *   0 - disallow raw tracepoint access for unpriv
398  *   1 - disallow cpu events for unpriv
399  *   2 - disallow kernel profiling for unpriv
400  */
401 int sysctl_perf_event_paranoid __read_mostly = 2;
402 
403 /* Minimum for 512 kiB + 1 user control page */
404 int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
405 
406 /*
407  * max perf event sample rate
408  */
409 #define DEFAULT_MAX_SAMPLE_RATE         100000
410 #define DEFAULT_SAMPLE_PERIOD_NS        (NSEC_PER_SEC / DEFAULT_MAX_SAMPLE_RATE)
411 #define DEFAULT_CPU_TIME_MAX_PERCENT    25
412 
413 int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
414 
415 static int max_samples_per_tick __read_mostly   = DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
416 static int perf_sample_period_ns __read_mostly  = DEFAULT_SAMPLE_PERIOD_NS;
417 
418 static int perf_sample_allowed_ns __read_mostly =
419         DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
420 
421 static void update_perf_cpu_limits(void)
422 {
423         u64 tmp = perf_sample_period_ns;
424 
425         tmp *= sysctl_perf_cpu_time_max_percent;
426         tmp = div_u64(tmp, 100);
427         if (!tmp)
428                 tmp = 1;
429 
430         WRITE_ONCE(perf_sample_allowed_ns, tmp);
431 }
432 
433 static bool perf_rotate_context(struct perf_cpu_context *cpuctx);
434 
435 int perf_proc_update_handler(struct ctl_table *table, int write,
436                 void __user *buffer, size_t *lenp,
437                 loff_t *ppos)
438 {
439         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
440 
441         if (ret || !write)
442                 return ret;
443 
444         /*
445          * If throttling is disabled don't allow the write:
446          */
447         if (sysctl_perf_cpu_time_max_percent == 100 ||
448             sysctl_perf_cpu_time_max_percent == 0)
449                 return -EINVAL;
450 
451         max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
452         perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
453         update_perf_cpu_limits();
454 
455         return 0;
456 }
457 
458 int sysctl_perf_cpu_time_max_percent __read_mostly = DEFAULT_CPU_TIME_MAX_PERCENT;
459 
460 int perf_cpu_time_max_percent_handler(struct ctl_table *table, int write,
461                                 void __user *buffer, size_t *lenp,
462                                 loff_t *ppos)
463 {
464         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
465 
466         if (ret || !write)
467                 return ret;
468 
469         if (sysctl_perf_cpu_time_max_percent == 100 ||
470             sysctl_perf_cpu_time_max_percent == 0) {
471                 printk(KERN_WARNING
472                        "perf: Dynamic interrupt throttling disabled, can hang your system!\n");
473                 WRITE_ONCE(perf_sample_allowed_ns, 0);
474         } else {
475                 update_perf_cpu_limits();
476         }
477 
478         return 0;
479 }
480 
481 /*
482  * perf samples are done in some very critical code paths (NMIs).
483  * If they take too much CPU time, the system can lock up and not
484  * get any real work done.  This will drop the sample rate when
485  * we detect that events are taking too long.
486  */
487 #define NR_ACCUMULATED_SAMPLES 128
488 static DEFINE_PER_CPU(u64, running_sample_length);
489 
490 static u64 __report_avg;
491 static u64 __report_allowed;
492 
493 static void perf_duration_warn(struct irq_work *w)
494 {
495         printk_ratelimited(KERN_INFO
496                 "perf: interrupt took too long (%lld > %lld), lowering "
497                 "kernel.perf_event_max_sample_rate to %d\n",
498                 __report_avg, __report_allowed,
499                 sysctl_perf_event_sample_rate);
500 }
501 
502 static DEFINE_IRQ_WORK(perf_duration_work, perf_duration_warn);
503 
504 void perf_sample_event_took(u64 sample_len_ns)
505 {
506         u64 max_len = READ_ONCE(perf_sample_allowed_ns);
507         u64 running_len;
508         u64 avg_len;
509         u32 max;
510 
511         if (max_len == 0)
512                 return;
513 
514         /* Decay the counter by 1 average sample. */
515         running_len = __this_cpu_read(running_sample_length);
516         running_len -= running_len/NR_ACCUMULATED_SAMPLES;
517         running_len += sample_len_ns;
518         __this_cpu_write(running_sample_length, running_len);
519 
520         /*
521          * Note: this will be biased artifically low until we have
522          * seen NR_ACCUMULATED_SAMPLES. Doing it this way keeps us
523          * from having to maintain a count.
524          */
525         avg_len = running_len/NR_ACCUMULATED_SAMPLES;
526         if (avg_len <= max_len)
527                 return;
528 
529         __report_avg = avg_len;
530         __report_allowed = max_len;
531 
532         /*
533          * Compute a throttle threshold 25% below the current duration.
534          */
535         avg_len += avg_len / 4;
536         max = (TICK_NSEC / 100) * sysctl_perf_cpu_time_max_percent;
537         if (avg_len < max)
538                 max /= (u32)avg_len;
539         else
540                 max = 1;
541 
542         WRITE_ONCE(perf_sample_allowed_ns, avg_len);
543         WRITE_ONCE(max_samples_per_tick, max);
544 
545         sysctl_perf_event_sample_rate = max * HZ;
546         perf_sample_period_ns = NSEC_PER_SEC / sysctl_perf_event_sample_rate;
547 
548         if (!irq_work_queue(&perf_duration_work)) {
549                 early_printk("perf: interrupt took too long (%lld > %lld), lowering "
550                              "kernel.perf_event_max_sample_rate to %d\n",
551                              __report_avg, __report_allowed,
552                              sysctl_perf_event_sample_rate);
553         }
554 }
555 
556 static atomic64_t perf_event_id;
557 
558 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
559                               enum event_type_t event_type);
560 
561 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
562                              enum event_type_t event_type,
563                              struct task_struct *task);
564 
565 static void update_context_time(struct perf_event_context *ctx);
566 static u64 perf_event_time(struct perf_event *event);
567 
568 void __weak perf_event_print_debug(void)        { }
569 
570 extern __weak const char *perf_pmu_name(void)
571 {
572         return "pmu";
573 }
574 
575 static inline u64 perf_clock(void)
576 {
577         return local_clock();
578 }
579 
580 static inline u64 perf_event_clock(struct perf_event *event)
581 {
582         return event->clock();
583 }
584 
585 /*
586  * State based event timekeeping...
587  *
588  * The basic idea is to use event->state to determine which (if any) time
589  * fields to increment with the current delta. This means we only need to
590  * update timestamps when we change state or when they are explicitly requested
591  * (read).
592  *
593  * Event groups make things a little more complicated, but not terribly so. The
594  * rules for a group are that if the group leader is OFF the entire group is
595  * OFF, irrespecive of what the group member states are. This results in
596  * __perf_effective_state().
597  *
598  * A futher ramification is that when a group leader flips between OFF and
599  * !OFF, we need to update all group member times.
600  *
601  *
602  * NOTE: perf_event_time() is based on the (cgroup) context time, and thus we
603  * need to make sure the relevant context time is updated before we try and
604  * update our timestamps.
605  */
606 
607 static __always_inline enum perf_event_state
608 __perf_effective_state(struct perf_event *event)
609 {
610         struct perf_event *leader = event->group_leader;
611 
612         if (leader->state <= PERF_EVENT_STATE_OFF)
613                 return leader->state;
614 
615         return event->state;
616 }
617 
618 static __always_inline void
619 __perf_update_times(struct perf_event *event, u64 now, u64 *enabled, u64 *running)
620 {
621         enum perf_event_state state = __perf_effective_state(event);
622         u64 delta = now - event->tstamp;
623 
624         *enabled = event->total_time_enabled;
625         if (state >= PERF_EVENT_STATE_INACTIVE)
626                 *enabled += delta;
627 
628         *running = event->total_time_running;
629         if (state >= PERF_EVENT_STATE_ACTIVE)
630                 *running += delta;
631 }
632 
633 static void perf_event_update_time(struct perf_event *event)
634 {
635         u64 now = perf_event_time(event);
636 
637         __perf_update_times(event, now, &event->total_time_enabled,
638                                         &event->total_time_running);
639         event->tstamp = now;
640 }
641 
642 static void perf_event_update_sibling_time(struct perf_event *leader)
643 {
644         struct perf_event *sibling;
645 
646         for_each_sibling_event(sibling, leader)
647                 perf_event_update_time(sibling);
648 }
649 
650 static void
651 perf_event_set_state(struct perf_event *event, enum perf_event_state state)
652 {
653         if (event->state == state)
654                 return;
655 
656         perf_event_update_time(event);
657         /*
658          * If a group leader gets enabled/disabled all its siblings
659          * are affected too.
660          */
661         if ((event->state < 0) ^ (state < 0))
662                 perf_event_update_sibling_time(event);
663 
664         WRITE_ONCE(event->state, state);
665 }
666 
667 #ifdef CONFIG_CGROUP_PERF
668 
669 static inline bool
670 perf_cgroup_match(struct perf_event *event)
671 {
672         struct perf_event_context *ctx = event->ctx;
673         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
674 
675         /* @event doesn't care about cgroup */
676         if (!event->cgrp)
677                 return true;
678 
679         /* wants specific cgroup scope but @cpuctx isn't associated with any */
680         if (!cpuctx->cgrp)
681                 return false;
682 
683         /*
684          * Cgroup scoping is recursive.  An event enabled for a cgroup is
685          * also enabled for all its descendant cgroups.  If @cpuctx's
686          * cgroup is a descendant of @event's (the test covers identity
687          * case), it's a match.
688          */
689         return cgroup_is_descendant(cpuctx->cgrp->css.cgroup,
690                                     event->cgrp->css.cgroup);
691 }
692 
693 static inline void perf_detach_cgroup(struct perf_event *event)
694 {
695         css_put(&event->cgrp->css);
696         event->cgrp = NULL;
697 }
698 
699 static inline int is_cgroup_event(struct perf_event *event)
700 {
701         return event->cgrp != NULL;
702 }
703 
704 static inline u64 perf_cgroup_event_time(struct perf_event *event)
705 {
706         struct perf_cgroup_info *t;
707 
708         t = per_cpu_ptr(event->cgrp->info, event->cpu);
709         return t->time;
710 }
711 
712 static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
713 {
714         struct perf_cgroup_info *info;
715         u64 now;
716 
717         now = perf_clock();
718 
719         info = this_cpu_ptr(cgrp->info);
720 
721         info->time += now - info->timestamp;
722         info->timestamp = now;
723 }
724 
725 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
726 {
727         struct perf_cgroup *cgrp = cpuctx->cgrp;
728         struct cgroup_subsys_state *css;
729 
730         if (cgrp) {
731                 for (css = &cgrp->css; css; css = css->parent) {
732                         cgrp = container_of(css, struct perf_cgroup, css);
733                         __update_cgrp_time(cgrp);
734                 }
735         }
736 }
737 
738 static inline void update_cgrp_time_from_event(struct perf_event *event)
739 {
740         struct perf_cgroup *cgrp;
741 
742         /*
743          * ensure we access cgroup data only when needed and
744          * when we know the cgroup is pinned (css_get)
745          */
746         if (!is_cgroup_event(event))
747                 return;
748 
749         cgrp = perf_cgroup_from_task(current, event->ctx);
750         /*
751          * Do not update time when cgroup is not active
752          */
753        if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
754                 __update_cgrp_time(event->cgrp);
755 }
756 
757 static inline void
758 perf_cgroup_set_timestamp(struct task_struct *task,
759                           struct perf_event_context *ctx)
760 {
761         struct perf_cgroup *cgrp;
762         struct perf_cgroup_info *info;
763         struct cgroup_subsys_state *css;
764 
765         /*
766          * ctx->lock held by caller
767          * ensure we do not access cgroup data
768          * unless we have the cgroup pinned (css_get)
769          */
770         if (!task || !ctx->nr_cgroups)
771                 return;
772 
773         cgrp = perf_cgroup_from_task(task, ctx);
774 
775         for (css = &cgrp->css; css; css = css->parent) {
776                 cgrp = container_of(css, struct perf_cgroup, css);
777                 info = this_cpu_ptr(cgrp->info);
778                 info->timestamp = ctx->timestamp;
779         }
780 }
781 
782 static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
783 
784 #define PERF_CGROUP_SWOUT       0x1 /* cgroup switch out every event */
785 #define PERF_CGROUP_SWIN        0x2 /* cgroup switch in events based on task */
786 
787 /*
788  * reschedule events based on the cgroup constraint of task.
789  *
790  * mode SWOUT : schedule out everything
791  * mode SWIN : schedule in based on cgroup for next
792  */
793 static void perf_cgroup_switch(struct task_struct *task, int mode)
794 {
795         struct perf_cpu_context *cpuctx;
796         struct list_head *list;
797         unsigned long flags;
798 
799         /*
800          * Disable interrupts and preemption to avoid this CPU's
801          * cgrp_cpuctx_entry to change under us.
802          */
803         local_irq_save(flags);
804 
805         list = this_cpu_ptr(&cgrp_cpuctx_list);
806         list_for_each_entry(cpuctx, list, cgrp_cpuctx_entry) {
807                 WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
808 
809                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
810                 perf_pmu_disable(cpuctx->ctx.pmu);
811 
812                 if (mode & PERF_CGROUP_SWOUT) {
813                         cpu_ctx_sched_out(cpuctx, EVENT_ALL);
814                         /*
815                          * must not be done before ctxswout due
816                          * to event_filter_match() in event_sched_out()
817                          */
818                         cpuctx->cgrp = NULL;
819                 }
820 
821                 if (mode & PERF_CGROUP_SWIN) {
822                         WARN_ON_ONCE(cpuctx->cgrp);
823                         /*
824                          * set cgrp before ctxsw in to allow
825                          * event_filter_match() to not have to pass
826                          * task around
827                          * we pass the cpuctx->ctx to perf_cgroup_from_task()
828                          * because cgorup events are only per-cpu
829                          */
830                         cpuctx->cgrp = perf_cgroup_from_task(task,
831                                                              &cpuctx->ctx);
832                         cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
833                 }
834                 perf_pmu_enable(cpuctx->ctx.pmu);
835                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
836         }
837 
838         local_irq_restore(flags);
839 }
840 
841 static inline void perf_cgroup_sched_out(struct task_struct *task,
842                                          struct task_struct *next)
843 {
844         struct perf_cgroup *cgrp1;
845         struct perf_cgroup *cgrp2 = NULL;
846 
847         rcu_read_lock();
848         /*
849          * we come here when we know perf_cgroup_events > 0
850          * we do not need to pass the ctx here because we know
851          * we are holding the rcu lock
852          */
853         cgrp1 = perf_cgroup_from_task(task, NULL);
854         cgrp2 = perf_cgroup_from_task(next, NULL);
855 
856         /*
857          * only schedule out current cgroup events if we know
858          * that we are switching to a different cgroup. Otherwise,
859          * do no touch the cgroup events.
860          */
861         if (cgrp1 != cgrp2)
862                 perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
863 
864         rcu_read_unlock();
865 }
866 
867 static inline void perf_cgroup_sched_in(struct task_struct *prev,
868                                         struct task_struct *task)
869 {
870         struct perf_cgroup *cgrp1;
871         struct perf_cgroup *cgrp2 = NULL;
872 
873         rcu_read_lock();
874         /*
875          * we come here when we know perf_cgroup_events > 0
876          * we do not need to pass the ctx here because we know
877          * we are holding the rcu lock
878          */
879         cgrp1 = perf_cgroup_from_task(task, NULL);
880         cgrp2 = perf_cgroup_from_task(prev, NULL);
881 
882         /*
883          * only need to schedule in cgroup events if we are changing
884          * cgroup during ctxsw. Cgroup events were not scheduled
885          * out of ctxsw out if that was not the case.
886          */
887         if (cgrp1 != cgrp2)
888                 perf_cgroup_switch(task, PERF_CGROUP_SWIN);
889 
890         rcu_read_unlock();
891 }
892 
893 static inline int perf_cgroup_connect(int fd, struct perf_event *event,
894                                       struct perf_event_attr *attr,
895                                       struct perf_event *group_leader)
896 {
897         struct perf_cgroup *cgrp;
898         struct cgroup_subsys_state *css;
899         struct fd f = fdget(fd);
900         int ret = 0;
901 
902         if (!f.file)
903                 return -EBADF;
904 
905         css = css_tryget_online_from_dir(f.file->f_path.dentry,
906                                          &perf_event_cgrp_subsys);
907         if (IS_ERR(css)) {
908                 ret = PTR_ERR(css);
909                 goto out;
910         }
911 
912         cgrp = container_of(css, struct perf_cgroup, css);
913         event->cgrp = cgrp;
914 
915         /*
916          * all events in a group must monitor
917          * the same cgroup because a task belongs
918          * to only one perf cgroup at a time
919          */
920         if (group_leader && group_leader->cgrp != cgrp) {
921                 perf_detach_cgroup(event);
922                 ret = -EINVAL;
923         }
924 out:
925         fdput(f);
926         return ret;
927 }
928 
929 static inline void
930 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
931 {
932         struct perf_cgroup_info *t;
933         t = per_cpu_ptr(event->cgrp->info, event->cpu);
934         event->shadow_ctx_time = now - t->timestamp;
935 }
936 
937 /*
938  * Update cpuctx->cgrp so that it is set when first cgroup event is added and
939  * cleared when last cgroup event is removed.
940  */
941 static inline void
942 list_update_cgroup_event(struct perf_event *event,
943                          struct perf_event_context *ctx, bool add)
944 {
945         struct perf_cpu_context *cpuctx;
946         struct list_head *cpuctx_entry;
947 
948         if (!is_cgroup_event(event))
949                 return;
950 
951         /*
952          * Because cgroup events are always per-cpu events,
953          * this will always be called from the right CPU.
954          */
955         cpuctx = __get_cpu_context(ctx);
956 
957         /*
958          * Since setting cpuctx->cgrp is conditional on the current @cgrp
959          * matching the event's cgroup, we must do this for every new event,
960          * because if the first would mismatch, the second would not try again
961          * and we would leave cpuctx->cgrp unset.
962          */
963         if (add && !cpuctx->cgrp) {
964                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
965 
966                 if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
967                         cpuctx->cgrp = cgrp;
968         }
969 
970         if (add && ctx->nr_cgroups++)
971                 return;
972         else if (!add && --ctx->nr_cgroups)
973                 return;
974 
975         /* no cgroup running */
976         if (!add)
977                 cpuctx->cgrp = NULL;
978 
979         cpuctx_entry = &cpuctx->cgrp_cpuctx_entry;
980         if (add)
981                 list_add(cpuctx_entry, this_cpu_ptr(&cgrp_cpuctx_list));
982         else
983                 list_del(cpuctx_entry);
984 }
985 
986 #else /* !CONFIG_CGROUP_PERF */
987 
988 static inline bool
989 perf_cgroup_match(struct perf_event *event)
990 {
991         return true;
992 }
993 
994 static inline void perf_detach_cgroup(struct perf_event *event)
995 {}
996 
997 static inline int is_cgroup_event(struct perf_event *event)
998 {
999         return 0;
1000 }
1001 
1002 static inline void update_cgrp_time_from_event(struct perf_event *event)
1003 {
1004 }
1005 
1006 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
1007 {
1008 }
1009 
1010 static inline void perf_cgroup_sched_out(struct task_struct *task,
1011                                          struct task_struct *next)
1012 {
1013 }
1014 
1015 static inline void perf_cgroup_sched_in(struct task_struct *prev,
1016                                         struct task_struct *task)
1017 {
1018 }
1019 
1020 static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
1021                                       struct perf_event_attr *attr,
1022                                       struct perf_event *group_leader)
1023 {
1024         return -EINVAL;
1025 }
1026 
1027 static inline void
1028 perf_cgroup_set_timestamp(struct task_struct *task,
1029                           struct perf_event_context *ctx)
1030 {
1031 }
1032 
1033 void
1034 perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
1035 {
1036 }
1037 
1038 static inline void
1039 perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
1040 {
1041 }
1042 
1043 static inline u64 perf_cgroup_event_time(struct perf_event *event)
1044 {
1045         return 0;
1046 }
1047 
1048 static inline void
1049 list_update_cgroup_event(struct perf_event *event,
1050                          struct perf_event_context *ctx, bool add)
1051 {
1052 }
1053 
1054 #endif
1055 
1056 /*
1057  * set default to be dependent on timer tick just
1058  * like original code
1059  */
1060 #define PERF_CPU_HRTIMER (1000 / HZ)
1061 /*
1062  * function must be called with interrupts disabled
1063  */
1064 static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr)
1065 {
1066         struct perf_cpu_context *cpuctx;
1067         bool rotations;
1068 
1069         lockdep_assert_irqs_disabled();
1070 
1071         cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
1072         rotations = perf_rotate_context(cpuctx);
1073 
1074         raw_spin_lock(&cpuctx->hrtimer_lock);
1075         if (rotations)
1076                 hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
1077         else
1078                 cpuctx->hrtimer_active = 0;
1079         raw_spin_unlock(&cpuctx->hrtimer_lock);
1080 
1081         return rotations ? HRTIMER_RESTART : HRTIMER_NORESTART;
1082 }
1083 
1084 static void __perf_mux_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
1085 {
1086         struct hrtimer *timer = &cpuctx->hrtimer;
1087         struct pmu *pmu = cpuctx->ctx.pmu;
1088         u64 interval;
1089 
1090         /* no multiplexing needed for SW PMU */
1091         if (pmu->task_ctx_nr == perf_sw_context)
1092                 return;
1093 
1094         /*
1095          * check default is sane, if not set then force to
1096          * default interval (1/tick)
1097          */
1098         interval = pmu->hrtimer_interval_ms;
1099         if (interval < 1)
1100                 interval = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
1101 
1102         cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * interval);
1103 
1104         raw_spin_lock_init(&cpuctx->hrtimer_lock);
1105         hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
1106         timer->function = perf_mux_hrtimer_handler;
1107 }
1108 
1109 static int perf_mux_hrtimer_restart(struct perf_cpu_context *cpuctx)
1110 {
1111         struct hrtimer *timer = &cpuctx->hrtimer;
1112         struct pmu *pmu = cpuctx->ctx.pmu;
1113         unsigned long flags;
1114 
1115         /* not for SW PMU */
1116         if (pmu->task_ctx_nr == perf_sw_context)
1117                 return 0;
1118 
1119         raw_spin_lock_irqsave(&cpuctx->hrtimer_lock, flags);
1120         if (!cpuctx->hrtimer_active) {
1121                 cpuctx->hrtimer_active = 1;
1122                 hrtimer_forward_now(timer, cpuctx->hrtimer_interval);
1123                 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
1124         }
1125         raw_spin_unlock_irqrestore(&cpuctx->hrtimer_lock, flags);
1126 
1127         return 0;
1128 }
1129 
1130 void perf_pmu_disable(struct pmu *pmu)
1131 {
1132         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1133         if (!(*count)++)
1134                 pmu->pmu_disable(pmu);
1135 }
1136 
1137 void perf_pmu_enable(struct pmu *pmu)
1138 {
1139         int *count = this_cpu_ptr(pmu->pmu_disable_count);
1140         if (!--(*count))
1141                 pmu->pmu_enable(pmu);
1142 }
1143 
1144 static DEFINE_PER_CPU(struct list_head, active_ctx_list);
1145 
1146 /*
1147  * perf_event_ctx_activate(), perf_event_ctx_deactivate(), and
1148  * perf_event_task_tick() are fully serialized because they're strictly cpu
1149  * affine and perf_event_ctx{activate,deactivate} are called with IRQs
1150  * disabled, while perf_event_task_tick is called from IRQ context.
1151  */
1152 static void perf_event_ctx_activate(struct perf_event_context *ctx)
1153 {
1154         struct list_head *head = this_cpu_ptr(&active_ctx_list);
1155 
1156         lockdep_assert_irqs_disabled();
1157 
1158         WARN_ON(!list_empty(&ctx->active_ctx_list));
1159 
1160         list_add(&ctx->active_ctx_list, head);
1161 }
1162 
1163 static void perf_event_ctx_deactivate(struct perf_event_context *ctx)
1164 {
1165         lockdep_assert_irqs_disabled();
1166 
1167         WARN_ON(list_empty(&ctx->active_ctx_list));
1168 
1169         list_del_init(&ctx->active_ctx_list);
1170 }
1171 
1172 static void get_ctx(struct perf_event_context *ctx)
1173 {
1174         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
1175 }
1176 
1177 static void free_ctx(struct rcu_head *head)
1178 {
1179         struct perf_event_context *ctx;
1180 
1181         ctx = container_of(head, struct perf_event_context, rcu_head);
1182         kfree(ctx->task_ctx_data);
1183         kfree(ctx);
1184 }
1185 
1186 static void put_ctx(struct perf_event_context *ctx)
1187 {
1188         if (atomic_dec_and_test(&ctx->refcount)) {
1189                 if (ctx->parent_ctx)
1190                         put_ctx(ctx->parent_ctx);
1191                 if (ctx->task && ctx->task != TASK_TOMBSTONE)
1192                         put_task_struct(ctx->task);
1193                 call_rcu(&ctx->rcu_head, free_ctx);
1194         }
1195 }
1196 
1197 /*
1198  * Because of perf_event::ctx migration in sys_perf_event_open::move_group and
1199  * perf_pmu_migrate_context() we need some magic.
1200  *
1201  * Those places that change perf_event::ctx will hold both
1202  * perf_event_ctx::mutex of the 'old' and 'new' ctx value.
1203  *
1204  * Lock ordering is by mutex address. There are two other sites where
1205  * perf_event_context::mutex nests and those are:
1206  *
1207  *  - perf_event_exit_task_context()    [ child , 0 ]
1208  *      perf_event_exit_event()
1209  *        put_event()                   [ parent, 1 ]
1210  *
1211  *  - perf_event_init_context()         [ parent, 0 ]
1212  *      inherit_task_group()
1213  *        inherit_group()
1214  *          inherit_event()
1215  *            perf_event_alloc()
1216  *              perf_init_event()
1217  *                perf_try_init_event() [ child , 1 ]
1218  *
1219  * While it appears there is an obvious deadlock here -- the parent and child
1220  * nesting levels are inverted between the two. This is in fact safe because
1221  * life-time rules separate them. That is an exiting task cannot fork, and a
1222  * spawning task cannot (yet) exit.
1223  *
1224  * But remember that that these are parent<->child context relations, and
1225  * migration does not affect children, therefore these two orderings should not
1226  * interact.
1227  *
1228  * The change in perf_event::ctx does not affect children (as claimed above)
1229  * because the sys_perf_event_open() case will install a new event and break
1230  * the ctx parent<->child relation, and perf_pmu_migrate_context() is only
1231  * concerned with cpuctx and that doesn't have children.
1232  *
1233  * The places that change perf_event::ctx will issue:
1234  *
1235  *   perf_remove_from_context();
1236  *   synchronize_rcu();
1237  *   perf_install_in_context();
1238  *
1239  * to affect the change. The remove_from_context() + synchronize_rcu() should
1240  * quiesce the event, after which we can install it in the new location. This
1241  * means that only external vectors (perf_fops, prctl) can perturb the event
1242  * while in transit. Therefore all such accessors should also acquire
1243  * perf_event_context::mutex to serialize against this.
1244  *
1245  * However; because event->ctx can change while we're waiting to acquire
1246  * ctx->mutex we must be careful and use the below perf_event_ctx_lock()
1247  * function.
1248  *
1249  * Lock order:
1250  *    cred_guard_mutex
1251  *      task_struct::perf_event_mutex
1252  *        perf_event_context::mutex
1253  *          perf_event::child_mutex;
1254  *            perf_event_context::lock
1255  *          perf_event::mmap_mutex
1256  *          mmap_sem
1257  *
1258  *    cpu_hotplug_lock
1259  *      pmus_lock
1260  *        cpuctx->mutex / perf_event_context::mutex
1261  */
1262 static struct perf_event_context *
1263 perf_event_ctx_lock_nested(struct perf_event *event, int nesting)
1264 {
1265         struct perf_event_context *ctx;
1266 
1267 again:
1268         rcu_read_lock();
1269         ctx = READ_ONCE(event->ctx);
1270         if (!atomic_inc_not_zero(&ctx->refcount)) {
1271                 rcu_read_unlock();
1272                 goto again;
1273         }
1274         rcu_read_unlock();
1275 
1276         mutex_lock_nested(&ctx->mutex, nesting);
1277         if (event->ctx != ctx) {
1278                 mutex_unlock(&ctx->mutex);
1279                 put_ctx(ctx);
1280                 goto again;
1281         }
1282 
1283         return ctx;
1284 }
1285 
1286 static inline struct perf_event_context *
1287 perf_event_ctx_lock(struct perf_event *event)
1288 {
1289         return perf_event_ctx_lock_nested(event, 0);
1290 }
1291 
1292 static void perf_event_ctx_unlock(struct perf_event *event,
1293                                   struct perf_event_context *ctx)
1294 {
1295         mutex_unlock(&ctx->mutex);
1296         put_ctx(ctx);
1297 }
1298 
1299 /*
1300  * This must be done under the ctx->lock, such as to serialize against
1301  * context_equiv(), therefore we cannot call put_ctx() since that might end up
1302  * calling scheduler related locks and ctx->lock nests inside those.
1303  */
1304 static __must_check struct perf_event_context *
1305 unclone_ctx(struct perf_event_context *ctx)
1306 {
1307         struct perf_event_context *parent_ctx = ctx->parent_ctx;
1308 
1309         lockdep_assert_held(&ctx->lock);
1310 
1311         if (parent_ctx)
1312                 ctx->parent_ctx = NULL;
1313         ctx->generation++;
1314 
1315         return parent_ctx;
1316 }
1317 
1318 static u32 perf_event_pid_type(struct perf_event *event, struct task_struct *p,
1319                                 enum pid_type type)
1320 {
1321         u32 nr;
1322         /*
1323          * only top level events have the pid namespace they were created in
1324          */
1325         if (event->parent)
1326                 event = event->parent;
1327 
1328         nr = __task_pid_nr_ns(p, type, event->ns);
1329         /* avoid -1 if it is idle thread or runs in another ns */
1330         if (!nr && !pid_alive(p))
1331                 nr = -1;
1332         return nr;
1333 }
1334 
1335 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
1336 {
1337         return perf_event_pid_type(event, p, PIDTYPE_TGID);
1338 }
1339 
1340 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
1341 {
1342         return perf_event_pid_type(event, p, PIDTYPE_PID);
1343 }
1344 
1345 /*
1346  * If we inherit events we want to return the parent event id
1347  * to userspace.
1348  */
1349 static u64 primary_event_id(struct perf_event *event)
1350 {
1351         u64 id = event->id;
1352 
1353         if (event->parent)
1354                 id = event->parent->id;
1355 
1356         return id;
1357 }
1358 
1359 /*
1360  * Get the perf_event_context for a task and lock it.
1361  *
1362  * This has to cope with with the fact that until it is locked,
1363  * the context could get moved to another task.
1364  */
1365 static struct perf_event_context *
1366 perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
1367 {
1368         struct perf_event_context *ctx;
1369 
1370 retry:
1371         /*
1372          * One of the few rules of preemptible RCU is that one cannot do
1373          * rcu_read_unlock() while holding a scheduler (or nested) lock when
1374          * part of the read side critical section was irqs-enabled -- see
1375          * rcu_read_unlock_special().
1376          *
1377          * Since ctx->lock nests under rq->lock we must ensure the entire read
1378          * side critical section has interrupts disabled.
1379          */
1380         local_irq_save(*flags);
1381         rcu_read_lock();
1382         ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
1383         if (ctx) {
1384                 /*
1385                  * If this context is a clone of another, it might
1386                  * get swapped for another underneath us by
1387                  * perf_event_task_sched_out, though the
1388                  * rcu_read_lock() protects us from any context
1389                  * getting freed.  Lock the context and check if it
1390                  * got swapped before we could get the lock, and retry
1391                  * if so.  If we locked the right context, then it
1392                  * can't get swapped on us any more.
1393                  */
1394                 raw_spin_lock(&ctx->lock);
1395                 if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
1396                         raw_spin_unlock(&ctx->lock);
1397                         rcu_read_unlock();
1398                         local_irq_restore(*flags);
1399                         goto retry;
1400                 }
1401 
1402                 if (ctx->task == TASK_TOMBSTONE ||
1403                     !atomic_inc_not_zero(&ctx->refcount)) {
1404                         raw_spin_unlock(&ctx->lock);
1405                         ctx = NULL;
1406                 } else {
1407                         WARN_ON_ONCE(ctx->task != task);
1408                 }
1409         }
1410         rcu_read_unlock();
1411         if (!ctx)
1412                 local_irq_restore(*flags);
1413         return ctx;
1414 }
1415 
1416 /*
1417  * Get the context for a task and increment its pin_count so it
1418  * can't get swapped to another task.  This also increments its
1419  * reference count so that the context can't get freed.
1420  */
1421 static struct perf_event_context *
1422 perf_pin_task_context(struct task_struct *task, int ctxn)
1423 {
1424         struct perf_event_context *ctx;
1425         unsigned long flags;
1426 
1427         ctx = perf_lock_task_context(task, ctxn, &flags);
1428         if (ctx) {
1429                 ++ctx->pin_count;
1430                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
1431         }
1432         return ctx;
1433 }
1434 
1435 static void perf_unpin_context(struct perf_event_context *ctx)
1436 {
1437         unsigned long flags;
1438 
1439         raw_spin_lock_irqsave(&ctx->lock, flags);
1440         --ctx->pin_count;
1441         raw_spin_unlock_irqrestore(&ctx->lock, flags);
1442 }
1443 
1444 /*
1445  * Update the record of the current time in a context.
1446  */
1447 static void update_context_time(struct perf_event_context *ctx)
1448 {
1449         u64 now = perf_clock();
1450 
1451         ctx->time += now - ctx->timestamp;
1452         ctx->timestamp = now;
1453 }
1454 
1455 static u64 perf_event_time(struct perf_event *event)
1456 {
1457         struct perf_event_context *ctx = event->ctx;
1458 
1459         if (is_cgroup_event(event))
1460                 return perf_cgroup_event_time(event);
1461 
1462         return ctx ? ctx->time : 0;
1463 }
1464 
1465 static enum event_type_t get_event_type(struct perf_event *event)
1466 {
1467         struct perf_event_context *ctx = event->ctx;
1468         enum event_type_t event_type;
1469 
1470         lockdep_assert_held(&ctx->lock);
1471 
1472         /*
1473          * It's 'group type', really, because if our group leader is
1474          * pinned, so are we.
1475          */
1476         if (event->group_leader != event)
1477                 event = event->group_leader;
1478 
1479         event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE;
1480         if (!ctx->task)
1481                 event_type |= EVENT_CPU;
1482 
1483         return event_type;
1484 }
1485 
1486 /*
1487  * Helper function to initialize event group nodes.
1488  */
1489 static void init_event_group(struct perf_event *event)
1490 {
1491         RB_CLEAR_NODE(&event->group_node);
1492         event->group_index = 0;
1493 }
1494 
1495 /*
1496  * Extract pinned or flexible groups from the context
1497  * based on event attrs bits.
1498  */
1499 static struct perf_event_groups *
1500 get_event_groups(struct perf_event *event, struct perf_event_context *ctx)
1501 {
1502         if (event->attr.pinned)
1503                 return &ctx->pinned_groups;
1504         else
1505                 return &ctx->flexible_groups;
1506 }
1507 
1508 /*
1509  * Helper function to initializes perf_event_group trees.
1510  */
1511 static void perf_event_groups_init(struct perf_event_groups *groups)
1512 {
1513         groups->tree = RB_ROOT;
1514         groups->index = 0;
1515 }
1516 
1517 /*
1518  * Compare function for event groups;
1519  *
1520  * Implements complex key that first sorts by CPU and then by virtual index
1521  * which provides ordering when rotating groups for the same CPU.
1522  */
1523 static bool
1524 perf_event_groups_less(struct perf_event *left, struct perf_event *right)
1525 {
1526         if (left->cpu < right->cpu)
1527                 return true;
1528         if (left->cpu > right->cpu)
1529                 return false;
1530 
1531         if (left->group_index < right->group_index)
1532                 return true;
1533         if (left->group_index > right->group_index)
1534                 return false;
1535 
1536         return false;
1537 }
1538 
1539 /*
1540  * Insert @event into @groups' tree; using {@event->cpu, ++@groups->index} for
1541  * key (see perf_event_groups_less). This places it last inside the CPU
1542  * subtree.
1543  */
1544 static void
1545 perf_event_groups_insert(struct perf_event_groups *groups,
1546                          struct perf_event *event)
1547 {
1548         struct perf_event *node_event;
1549         struct rb_node *parent;
1550         struct rb_node **node;
1551 
1552         event->group_index = ++groups->index;
1553 
1554         node = &groups->tree.rb_node;
1555         parent = *node;
1556 
1557         while (*node) {
1558                 parent = *node;
1559                 node_event = container_of(*node, struct perf_event, group_node);
1560 
1561                 if (perf_event_groups_less(event, node_event))
1562                         node = &parent->rb_left;
1563                 else
1564                         node = &parent->rb_right;
1565         }
1566 
1567         rb_link_node(&event->group_node, parent, node);
1568         rb_insert_color(&event->group_node, &groups->tree);
1569 }
1570 
1571 /*
1572  * Helper function to insert event into the pinned or flexible groups.
1573  */
1574 static void
1575 add_event_to_groups(struct perf_event *event, struct perf_event_context *ctx)
1576 {
1577         struct perf_event_groups *groups;
1578 
1579         groups = get_event_groups(event, ctx);
1580         perf_event_groups_insert(groups, event);
1581 }
1582 
1583 /*
1584  * Delete a group from a tree.
1585  */
1586 static void
1587 perf_event_groups_delete(struct perf_event_groups *groups,
1588                          struct perf_event *event)
1589 {
1590         WARN_ON_ONCE(RB_EMPTY_NODE(&event->group_node) ||
1591                      RB_EMPTY_ROOT(&groups->tree));
1592 
1593         rb_erase(&event->group_node, &groups->tree);
1594         init_event_group(event);
1595 }
1596 
1597 /*
1598  * Helper function to delete event from its groups.
1599  */
1600 static void
1601 del_event_from_groups(struct perf_event *event, struct perf_event_context *ctx)
1602 {
1603         struct perf_event_groups *groups;
1604 
1605         groups = get_event_groups(event, ctx);
1606         perf_event_groups_delete(groups, event);
1607 }
1608 
1609 /*
1610  * Get the leftmost event in the @cpu subtree.
1611  */
1612 static struct perf_event *
1613 perf_event_groups_first(struct perf_event_groups *groups, int cpu)
1614 {
1615         struct perf_event *node_event = NULL, *match = NULL;
1616         struct rb_node *node = groups->tree.rb_node;
1617 
1618         while (node) {
1619                 node_event = container_of(node, struct perf_event, group_node);
1620 
1621                 if (cpu < node_event->cpu) {
1622                         node = node->rb_left;
1623                 } else if (cpu > node_event->cpu) {
1624                         node = node->rb_right;
1625                 } else {
1626                         match = node_event;
1627                         node = node->rb_left;
1628                 }
1629         }
1630 
1631         return match;
1632 }
1633 
1634 /*
1635  * Like rb_entry_next_safe() for the @cpu subtree.
1636  */
1637 static struct perf_event *
1638 perf_event_groups_next(struct perf_event *event)
1639 {
1640         struct perf_event *next;
1641 
1642         next = rb_entry_safe(rb_next(&event->group_node), typeof(*event), group_node);
1643         if (next && next->cpu == event->cpu)
1644                 return next;
1645 
1646         return NULL;
1647 }
1648 
1649 /*
1650  * Iterate through the whole groups tree.
1651  */
1652 #define perf_event_groups_for_each(event, groups)                       \
1653         for (event = rb_entry_safe(rb_first(&((groups)->tree)),         \
1654                                 typeof(*event), group_node); event;     \
1655                 event = rb_entry_safe(rb_next(&event->group_node),      \
1656                                 typeof(*event), group_node))
1657 
1658 /*
1659  * Add an event from the lists for its context.
1660  * Must be called with ctx->mutex and ctx->lock held.
1661  */
1662 static void
1663 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
1664 {
1665         lockdep_assert_held(&ctx->lock);
1666 
1667         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
1668         event->attach_state |= PERF_ATTACH_CONTEXT;
1669 
1670         event->tstamp = perf_event_time(event);
1671 
1672         /*
1673          * If we're a stand alone event or group leader, we go to the context
1674          * list, group events are kept attached to the group so that
1675          * perf_group_detach can, at all times, locate all siblings.
1676          */
1677         if (event->group_leader == event) {
1678                 event->group_caps = event->event_caps;
1679                 add_event_to_groups(event, ctx);
1680         }
1681 
1682         list_update_cgroup_event(event, ctx, true);
1683 
1684         list_add_rcu(&event->event_entry, &ctx->event_list);
1685         ctx->nr_events++;
1686         if (event->attr.inherit_stat)
1687                 ctx->nr_stat++;
1688 
1689         ctx->generation++;
1690 }
1691 
1692 /*
1693  * Initialize event state based on the perf_event_attr::disabled.
1694  */
1695 static inline void perf_event__state_init(struct perf_event *event)
1696 {
1697         event->state = event->attr.disabled ? PERF_EVENT_STATE_OFF :
1698                                               PERF_EVENT_STATE_INACTIVE;
1699 }
1700 
1701 static void __perf_event_read_size(struct perf_event *event, int nr_siblings)
1702 {
1703         int entry = sizeof(u64); /* value */
1704         int size = 0;
1705         int nr = 1;
1706 
1707         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1708                 size += sizeof(u64);
1709 
1710         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1711                 size += sizeof(u64);
1712 
1713         if (event->attr.read_format & PERF_FORMAT_ID)
1714                 entry += sizeof(u64);
1715 
1716         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1717                 nr += nr_siblings;
1718                 size += sizeof(u64);
1719         }
1720 
1721         size += entry * nr;
1722         event->read_size = size;
1723 }
1724 
1725 static void __perf_event_header_size(struct perf_event *event, u64 sample_type)
1726 {
1727         struct perf_sample_data *data;
1728         u16 size = 0;
1729 
1730         if (sample_type & PERF_SAMPLE_IP)
1731                 size += sizeof(data->ip);
1732 
1733         if (sample_type & PERF_SAMPLE_ADDR)
1734                 size += sizeof(data->addr);
1735 
1736         if (sample_type & PERF_SAMPLE_PERIOD)
1737                 size += sizeof(data->period);
1738 
1739         if (sample_type & PERF_SAMPLE_WEIGHT)
1740                 size += sizeof(data->weight);
1741 
1742         if (sample_type & PERF_SAMPLE_READ)
1743                 size += event->read_size;
1744 
1745         if (sample_type & PERF_SAMPLE_DATA_SRC)
1746                 size += sizeof(data->data_src.val);
1747 
1748         if (sample_type & PERF_SAMPLE_TRANSACTION)
1749                 size += sizeof(data->txn);
1750 
1751         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
1752                 size += sizeof(data->phys_addr);
1753 
1754         event->header_size = size;
1755 }
1756 
1757 /*
1758  * Called at perf_event creation and when events are attached/detached from a
1759  * group.
1760  */
1761 static void perf_event__header_size(struct perf_event *event)
1762 {
1763         __perf_event_read_size(event,
1764                                event->group_leader->nr_siblings);
1765         __perf_event_header_size(event, event->attr.sample_type);
1766 }
1767 
1768 static void perf_event__id_header_size(struct perf_event *event)
1769 {
1770         struct perf_sample_data *data;
1771         u64 sample_type = event->attr.sample_type;
1772         u16 size = 0;
1773 
1774         if (sample_type & PERF_SAMPLE_TID)
1775                 size += sizeof(data->tid_entry);
1776 
1777         if (sample_type & PERF_SAMPLE_TIME)
1778                 size += sizeof(data->time);
1779 
1780         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1781                 size += sizeof(data->id);
1782 
1783         if (sample_type & PERF_SAMPLE_ID)
1784                 size += sizeof(data->id);
1785 
1786         if (sample_type & PERF_SAMPLE_STREAM_ID)
1787                 size += sizeof(data->stream_id);
1788 
1789         if (sample_type & PERF_SAMPLE_CPU)
1790                 size += sizeof(data->cpu_entry);
1791 
1792         event->id_header_size = size;
1793 }
1794 
1795 static bool perf_event_validate_size(struct perf_event *event)
1796 {
1797         /*
1798          * The values computed here will be over-written when we actually
1799          * attach the event.
1800          */
1801         __perf_event_read_size(event, event->group_leader->nr_siblings + 1);
1802         __perf_event_header_size(event, event->attr.sample_type & ~PERF_SAMPLE_READ);
1803         perf_event__id_header_size(event);
1804 
1805         /*
1806          * Sum the lot; should not exceed the 64k limit we have on records.
1807          * Conservative limit to allow for callchains and other variable fields.
1808          */
1809         if (event->read_size + event->header_size +
1810             event->id_header_size + sizeof(struct perf_event_header) >= 16*1024)
1811                 return false;
1812 
1813         return true;
1814 }
1815 
1816 static void perf_group_attach(struct perf_event *event)
1817 {
1818         struct perf_event *group_leader = event->group_leader, *pos;
1819 
1820         lockdep_assert_held(&event->ctx->lock);
1821 
1822         /*
1823          * We can have double attach due to group movement in perf_event_open.
1824          */
1825         if (event->attach_state & PERF_ATTACH_GROUP)
1826                 return;
1827 
1828         event->attach_state |= PERF_ATTACH_GROUP;
1829 
1830         if (group_leader == event)
1831                 return;
1832 
1833         WARN_ON_ONCE(group_leader->ctx != event->ctx);
1834 
1835         group_leader->group_caps &= event->event_caps;
1836 
1837         list_add_tail(&event->sibling_list, &group_leader->sibling_list);
1838         group_leader->nr_siblings++;
1839 
1840         perf_event__header_size(group_leader);
1841 
1842         for_each_sibling_event(pos, group_leader)
1843                 perf_event__header_size(pos);
1844 }
1845 
1846 /*
1847  * Remove an event from the lists for its context.
1848  * Must be called with ctx->mutex and ctx->lock held.
1849  */
1850 static void
1851 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
1852 {
1853         WARN_ON_ONCE(event->ctx != ctx);
1854         lockdep_assert_held(&ctx->lock);
1855 
1856         /*
1857          * We can have double detach due to exit/hot-unplug + close.
1858          */
1859         if (!(event->attach_state & PERF_ATTACH_CONTEXT))
1860                 return;
1861 
1862         event->attach_state &= ~PERF_ATTACH_CONTEXT;
1863 
1864         list_update_cgroup_event(event, ctx, false);
1865 
1866         ctx->nr_events--;
1867         if (event->attr.inherit_stat)
1868                 ctx->nr_stat--;
1869 
1870         list_del_rcu(&event->event_entry);
1871 
1872         if (event->group_leader == event)
1873                 del_event_from_groups(event, ctx);
1874 
1875         /*
1876          * If event was in error state, then keep it
1877          * that way, otherwise bogus counts will be
1878          * returned on read(). The only way to get out
1879          * of error state is by explicit re-enabling
1880          * of the event
1881          */
1882         if (event->state > PERF_EVENT_STATE_OFF)
1883                 perf_event_set_state(event, PERF_EVENT_STATE_OFF);
1884 
1885         ctx->generation++;
1886 }
1887 
1888 static void perf_group_detach(struct perf_event *event)
1889 {
1890         struct perf_event *sibling, *tmp;
1891         struct perf_event_context *ctx = event->ctx;
1892 
1893         lockdep_assert_held(&ctx->lock);
1894 
1895         /*
1896          * We can have double detach due to exit/hot-unplug + close.
1897          */
1898         if (!(event->attach_state & PERF_ATTACH_GROUP))
1899                 return;
1900 
1901         event->attach_state &= ~PERF_ATTACH_GROUP;
1902 
1903         /*
1904          * If this is a sibling, remove it from its group.
1905          */
1906         if (event->group_leader != event) {
1907                 list_del_init(&event->sibling_list);
1908                 event->group_leader->nr_siblings--;
1909                 goto out;
1910         }
1911 
1912         /*
1913          * If this was a group event with sibling events then
1914          * upgrade the siblings to singleton events by adding them
1915          * to whatever list we are on.
1916          */
1917         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, sibling_list) {
1918 
1919                 sibling->group_leader = sibling;
1920                 list_del_init(&sibling->sibling_list);
1921 
1922                 /* Inherit group flags from the previous leader */
1923                 sibling->group_caps = event->group_caps;
1924 
1925                 if (!RB_EMPTY_NODE(&event->group_node)) {
1926                         add_event_to_groups(sibling, event->ctx);
1927 
1928                         if (sibling->state == PERF_EVENT_STATE_ACTIVE) {
1929                                 struct list_head *list = sibling->attr.pinned ?
1930                                         &ctx->pinned_active : &ctx->flexible_active;
1931 
1932                                 list_add_tail(&sibling->active_list, list);
1933                         }
1934                 }
1935 
1936                 WARN_ON_ONCE(sibling->ctx != event->ctx);
1937         }
1938 
1939 out:
1940         perf_event__header_size(event->group_leader);
1941 
1942         for_each_sibling_event(tmp, event->group_leader)
1943                 perf_event__header_size(tmp);
1944 }
1945 
1946 static bool is_orphaned_event(struct perf_event *event)
1947 {
1948         return event->state == PERF_EVENT_STATE_DEAD;
1949 }
1950 
1951 static inline int __pmu_filter_match(struct perf_event *event)
1952 {
1953         struct pmu *pmu = event->pmu;
1954         return pmu->filter_match ? pmu->filter_match(event) : 1;
1955 }
1956 
1957 /*
1958  * Check whether we should attempt to schedule an event group based on
1959  * PMU-specific filtering. An event group can consist of HW and SW events,
1960  * potentially with a SW leader, so we must check all the filters, to
1961  * determine whether a group is schedulable:
1962  */
1963 static inline int pmu_filter_match(struct perf_event *event)
1964 {
1965         struct perf_event *sibling;
1966 
1967         if (!__pmu_filter_match(event))
1968                 return 0;
1969 
1970         for_each_sibling_event(sibling, event) {
1971                 if (!__pmu_filter_match(sibling))
1972                         return 0;
1973         }
1974 
1975         return 1;
1976 }
1977 
1978 static inline int
1979 event_filter_match(struct perf_event *event)
1980 {
1981         return (event->cpu == -1 || event->cpu == smp_processor_id()) &&
1982                perf_cgroup_match(event) && pmu_filter_match(event);
1983 }
1984 
1985 static void
1986 event_sched_out(struct perf_event *event,
1987                   struct perf_cpu_context *cpuctx,
1988                   struct perf_event_context *ctx)
1989 {
1990         enum perf_event_state state = PERF_EVENT_STATE_INACTIVE;
1991 
1992         WARN_ON_ONCE(event->ctx != ctx);
1993         lockdep_assert_held(&ctx->lock);
1994 
1995         if (event->state != PERF_EVENT_STATE_ACTIVE)
1996                 return;
1997 
1998         /*
1999          * Asymmetry; we only schedule events _IN_ through ctx_sched_in(), but
2000          * we can schedule events _OUT_ individually through things like
2001          * __perf_remove_from_context().
2002          */
2003         list_del_init(&event->active_list);
2004 
2005         perf_pmu_disable(event->pmu);
2006 
2007         event->pmu->del(event, 0);
2008         event->oncpu = -1;
2009 
2010         if (event->pending_disable) {
2011                 event->pending_disable = 0;
2012                 state = PERF_EVENT_STATE_OFF;
2013         }
2014         perf_event_set_state(event, state);
2015 
2016         if (!is_software_event(event))
2017                 cpuctx->active_oncpu--;
2018         if (!--ctx->nr_active)
2019                 perf_event_ctx_deactivate(ctx);
2020         if (event->attr.freq && event->attr.sample_freq)
2021                 ctx->nr_freq--;
2022         if (event->attr.exclusive || !cpuctx->active_oncpu)
2023                 cpuctx->exclusive = 0;
2024 
2025         perf_pmu_enable(event->pmu);
2026 }
2027 
2028 static void
2029 group_sched_out(struct perf_event *group_event,
2030                 struct perf_cpu_context *cpuctx,
2031                 struct perf_event_context *ctx)
2032 {
2033         struct perf_event *event;
2034 
2035         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
2036                 return;
2037 
2038         perf_pmu_disable(ctx->pmu);
2039 
2040         event_sched_out(group_event, cpuctx, ctx);
2041 
2042         /*
2043          * Schedule out siblings (if any):
2044          */
2045         for_each_sibling_event(event, group_event)
2046                 event_sched_out(event, cpuctx, ctx);
2047 
2048         perf_pmu_enable(ctx->pmu);
2049 
2050         if (group_event->attr.exclusive)
2051                 cpuctx->exclusive = 0;
2052 }
2053 
2054 #define DETACH_GROUP    0x01UL
2055 
2056 /*
2057  * Cross CPU call to remove a performance event
2058  *
2059  * We disable the event on the hardware level first. After that we
2060  * remove it from the context list.
2061  */
2062 static void
2063 __perf_remove_from_context(struct perf_event *event,
2064                            struct perf_cpu_context *cpuctx,
2065                            struct perf_event_context *ctx,
2066                            void *info)
2067 {
2068         unsigned long flags = (unsigned long)info;
2069 
2070         if (ctx->is_active & EVENT_TIME) {
2071                 update_context_time(ctx);
2072                 update_cgrp_time_from_cpuctx(cpuctx);
2073         }
2074 
2075         event_sched_out(event, cpuctx, ctx);
2076         if (flags & DETACH_GROUP)
2077                 perf_group_detach(event);
2078         list_del_event(event, ctx);
2079 
2080         if (!ctx->nr_events && ctx->is_active) {
2081                 ctx->is_active = 0;
2082                 if (ctx->task) {
2083                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
2084                         cpuctx->task_ctx = NULL;
2085                 }
2086         }
2087 }
2088 
2089 /*
2090  * Remove the event from a task's (or a CPU's) list of events.
2091  *
2092  * If event->ctx is a cloned context, callers must make sure that
2093  * every task struct that event->ctx->task could possibly point to
2094  * remains valid.  This is OK when called from perf_release since
2095  * that only calls us on the top-level context, which can't be a clone.
2096  * When called from perf_event_exit_task, it's OK because the
2097  * context has been detached from its task.
2098  */
2099 static void perf_remove_from_context(struct perf_event *event, unsigned long flags)
2100 {
2101         struct perf_event_context *ctx = event->ctx;
2102 
2103         lockdep_assert_held(&ctx->mutex);
2104 
2105         event_function_call(event, __perf_remove_from_context, (void *)flags);
2106 
2107         /*
2108          * The above event_function_call() can NO-OP when it hits
2109          * TASK_TOMBSTONE. In that case we must already have been detached
2110          * from the context (by perf_event_exit_event()) but the grouping
2111          * might still be in-tact.
2112          */
2113         WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
2114         if ((flags & DETACH_GROUP) &&
2115             (event->attach_state & PERF_ATTACH_GROUP)) {
2116                 /*
2117                  * Since in that case we cannot possibly be scheduled, simply
2118                  * detach now.
2119                  */
2120                 raw_spin_lock_irq(&ctx->lock);
2121                 perf_group_detach(event);
2122                 raw_spin_unlock_irq(&ctx->lock);
2123         }
2124 }
2125 
2126 /*
2127  * Cross CPU call to disable a performance event
2128  */
2129 static void __perf_event_disable(struct perf_event *event,
2130                                  struct perf_cpu_context *cpuctx,
2131                                  struct perf_event_context *ctx,
2132                                  void *info)
2133 {
2134         if (event->state < PERF_EVENT_STATE_INACTIVE)
2135                 return;
2136 
2137         if (ctx->is_active & EVENT_TIME) {
2138                 update_context_time(ctx);
2139                 update_cgrp_time_from_event(event);
2140         }
2141 
2142         if (event == event->group_leader)
2143                 group_sched_out(event, cpuctx, ctx);
2144         else
2145                 event_sched_out(event, cpuctx, ctx);
2146 
2147         perf_event_set_state(event, PERF_EVENT_STATE_OFF);
2148 }
2149 
2150 /*
2151  * Disable an event.
2152  *
2153  * If event->ctx is a cloned context, callers must make sure that
2154  * every task struct that event->ctx->task could possibly point to
2155  * remains valid.  This condition is satisifed when called through
2156  * perf_event_for_each_child or perf_event_for_each because they
2157  * hold the top-level event's child_mutex, so any descendant that
2158  * goes to exit will block in perf_event_exit_event().
2159  *
2160  * When called from perf_pending_event it's OK because event->ctx
2161  * is the current context on this CPU and preemption is disabled,
2162  * hence we can't get into perf_event_task_sched_out for this context.
2163  */
2164 static void _perf_event_disable(struct perf_event *event)
2165 {
2166         struct perf_event_context *ctx = event->ctx;
2167 
2168         raw_spin_lock_irq(&ctx->lock);
2169         if (event->state <= PERF_EVENT_STATE_OFF) {
2170                 raw_spin_unlock_irq(&ctx->lock);
2171                 return;
2172         }
2173         raw_spin_unlock_irq(&ctx->lock);
2174 
2175         event_function_call(event, __perf_event_disable, NULL);
2176 }
2177 
2178 void perf_event_disable_local(struct perf_event *event)
2179 {
2180         event_function_local(event, __perf_event_disable, NULL);
2181 }
2182 
2183 /*
2184  * Strictly speaking kernel users cannot create groups and therefore this
2185  * interface does not need the perf_event_ctx_lock() magic.
2186  */
2187 void perf_event_disable(struct perf_event *event)
2188 {
2189         struct perf_event_context *ctx;
2190 
2191         ctx = perf_event_ctx_lock(event);
2192         _perf_event_disable(event);
2193         perf_event_ctx_unlock(event, ctx);
2194 }
2195 EXPORT_SYMBOL_GPL(perf_event_disable);
2196 
2197 void perf_event_disable_inatomic(struct perf_event *event)
2198 {
2199         event->pending_disable = 1;
2200         irq_work_queue(&event->pending);
2201 }
2202 
2203 static void perf_set_shadow_time(struct perf_event *event,
2204                                  struct perf_event_context *ctx)
2205 {
2206         /*
2207          * use the correct time source for the time snapshot
2208          *
2209          * We could get by without this by leveraging the
2210          * fact that to get to this function, the caller
2211          * has most likely already called update_context_time()
2212          * and update_cgrp_time_xx() and thus both timestamp
2213          * are identical (or very close). Given that tstamp is,
2214          * already adjusted for cgroup, we could say that:
2215          *    tstamp - ctx->timestamp
2216          * is equivalent to
2217          *    tstamp - cgrp->timestamp.
2218          *
2219          * Then, in perf_output_read(), the calculation would
2220          * work with no changes because:
2221          * - event is guaranteed scheduled in
2222          * - no scheduled out in between
2223          * - thus the timestamp would be the same
2224          *
2225          * But this is a bit hairy.
2226          *
2227          * So instead, we have an explicit cgroup call to remain
2228          * within the time time source all along. We believe it
2229          * is cleaner and simpler to understand.
2230          */
2231         if (is_cgroup_event(event))
2232                 perf_cgroup_set_shadow_time(event, event->tstamp);
2233         else
2234                 event->shadow_ctx_time = event->tstamp - ctx->timestamp;
2235 }
2236 
2237 #define MAX_INTERRUPTS (~0ULL)
2238 
2239 static void perf_log_throttle(struct perf_event *event, int enable);
2240 static void perf_log_itrace_start(struct perf_event *event);
2241 
2242 static int
2243 event_sched_in(struct perf_event *event,
2244                  struct perf_cpu_context *cpuctx,
2245                  struct perf_event_context *ctx)
2246 {
2247         int ret = 0;
2248 
2249         lockdep_assert_held(&ctx->lock);
2250 
2251         if (event->state <= PERF_EVENT_STATE_OFF)
2252                 return 0;
2253 
2254         WRITE_ONCE(event->oncpu, smp_processor_id());
2255         /*
2256          * Order event::oncpu write to happen before the ACTIVE state is
2257          * visible. This allows perf_event_{stop,read}() to observe the correct
2258          * ->oncpu if it sees ACTIVE.
2259          */
2260         smp_wmb();
2261         perf_event_set_state(event, PERF_EVENT_STATE_ACTIVE);
2262 
2263         /*
2264          * Unthrottle events, since we scheduled we might have missed several
2265          * ticks already, also for a heavily scheduling task there is little
2266          * guarantee it'll get a tick in a timely manner.
2267          */
2268         if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
2269                 perf_log_throttle(event, 1);
2270                 event->hw.interrupts = 0;
2271         }
2272 
2273         perf_pmu_disable(event->pmu);
2274 
2275         perf_set_shadow_time(event, ctx);
2276 
2277         perf_log_itrace_start(event);
2278 
2279         if (event->pmu->add(event, PERF_EF_START)) {
2280                 perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2281                 event->oncpu = -1;
2282                 ret = -EAGAIN;
2283                 goto out;
2284         }
2285 
2286         if (!is_software_event(event))
2287                 cpuctx->active_oncpu++;
2288         if (!ctx->nr_active++)
2289                 perf_event_ctx_activate(ctx);
2290         if (event->attr.freq && event->attr.sample_freq)
2291                 ctx->nr_freq++;
2292 
2293         if (event->attr.exclusive)
2294                 cpuctx->exclusive = 1;
2295 
2296 out:
2297         perf_pmu_enable(event->pmu);
2298 
2299         return ret;
2300 }
2301 
2302 static int
2303 group_sched_in(struct perf_event *group_event,
2304                struct perf_cpu_context *cpuctx,
2305                struct perf_event_context *ctx)
2306 {
2307         struct perf_event *event, *partial_group = NULL;
2308         struct pmu *pmu = ctx->pmu;
2309 
2310         if (group_event->state == PERF_EVENT_STATE_OFF)
2311                 return 0;
2312 
2313         pmu->start_txn(pmu, PERF_PMU_TXN_ADD);
2314 
2315         if (event_sched_in(group_event, cpuctx, ctx)) {
2316                 pmu->cancel_txn(pmu);
2317                 perf_mux_hrtimer_restart(cpuctx);
2318                 return -EAGAIN;
2319         }
2320 
2321         /*
2322          * Schedule in siblings as one group (if any):
2323          */
2324         for_each_sibling_event(event, group_event) {
2325                 if (event_sched_in(event, cpuctx, ctx)) {
2326                         partial_group = event;
2327                         goto group_error;
2328                 }
2329         }
2330 
2331         if (!pmu->commit_txn(pmu))
2332                 return 0;
2333 
2334 group_error:
2335         /*
2336          * Groups can be scheduled in as one unit only, so undo any
2337          * partial group before returning:
2338          * The events up to the failed event are scheduled out normally.
2339          */
2340         for_each_sibling_event(event, group_event) {
2341                 if (event == partial_group)
2342                         break;
2343 
2344                 event_sched_out(event, cpuctx, ctx);
2345         }
2346         event_sched_out(group_event, cpuctx, ctx);
2347 
2348         pmu->cancel_txn(pmu);
2349 
2350         perf_mux_hrtimer_restart(cpuctx);
2351 
2352         return -EAGAIN;
2353 }
2354 
2355 /*
2356  * Work out whether we can put this event group on the CPU now.
2357  */
2358 static int group_can_go_on(struct perf_event *event,
2359                            struct perf_cpu_context *cpuctx,
2360                            int can_add_hw)
2361 {
2362         /*
2363          * Groups consisting entirely of software events can always go on.
2364          */
2365         if (event->group_caps & PERF_EV_CAP_SOFTWARE)
2366                 return 1;
2367         /*
2368          * If an exclusive group is already on, no other hardware
2369          * events can go on.
2370          */
2371         if (cpuctx->exclusive)
2372                 return 0;
2373         /*
2374          * If this group is exclusive and there are already
2375          * events on the CPU, it can't go on.
2376          */
2377         if (event->attr.exclusive && cpuctx->active_oncpu)
2378                 return 0;
2379         /*
2380          * Otherwise, try to add it if all previous groups were able
2381          * to go on.
2382          */
2383         return can_add_hw;
2384 }
2385 
2386 static void add_event_to_ctx(struct perf_event *event,
2387                                struct perf_event_context *ctx)
2388 {
2389         list_add_event(event, ctx);
2390         perf_group_attach(event);
2391 }
2392 
2393 static void ctx_sched_out(struct perf_event_context *ctx,
2394                           struct perf_cpu_context *cpuctx,
2395                           enum event_type_t event_type);
2396 static void
2397 ctx_sched_in(struct perf_event_context *ctx,
2398              struct perf_cpu_context *cpuctx,
2399              enum event_type_t event_type,
2400              struct task_struct *task);
2401 
2402 static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
2403                                struct perf_event_context *ctx,
2404                                enum event_type_t event_type)
2405 {
2406         if (!cpuctx->task_ctx)
2407                 return;
2408 
2409         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
2410                 return;
2411 
2412         ctx_sched_out(ctx, cpuctx, event_type);
2413 }
2414 
2415 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
2416                                 struct perf_event_context *ctx,
2417                                 struct task_struct *task)
2418 {
2419         cpu_ctx_sched_in(cpuctx, EVENT_PINNED, task);
2420         if (ctx)
2421                 ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
2422         cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
2423         if (ctx)
2424                 ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
2425 }
2426 
2427 /*
2428  * We want to maintain the following priority of scheduling:
2429  *  - CPU pinned (EVENT_CPU | EVENT_PINNED)
2430  *  - task pinned (EVENT_PINNED)
2431  *  - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
2432  *  - task flexible (EVENT_FLEXIBLE).
2433  *
2434  * In order to avoid unscheduling and scheduling back in everything every
2435  * time an event is added, only do it for the groups of equal priority and
2436  * below.
2437  *
2438  * This can be called after a batch operation on task events, in which case
2439  * event_type is a bit mask of the types of events involved. For CPU events,
2440  * event_type is only either EVENT_PINNED or EVENT_FLEXIBLE.
2441  */
2442 static void ctx_resched(struct perf_cpu_context *cpuctx,
2443                         struct perf_event_context *task_ctx,
2444                         enum event_type_t event_type)
2445 {
2446         enum event_type_t ctx_event_type;
2447         bool cpu_event = !!(event_type & EVENT_CPU);
2448 
2449         /*
2450          * If pinned groups are involved, flexible groups also need to be
2451          * scheduled out.
2452          */
2453         if (event_type & EVENT_PINNED)
2454                 event_type |= EVENT_FLEXIBLE;
2455 
2456         ctx_event_type = event_type & EVENT_ALL;
2457 
2458         perf_pmu_disable(cpuctx->ctx.pmu);
2459         if (task_ctx)
2460                 task_ctx_sched_out(cpuctx, task_ctx, event_type);
2461 
2462         /*
2463          * Decide which cpu ctx groups to schedule out based on the types
2464          * of events that caused rescheduling:
2465          *  - EVENT_CPU: schedule out corresponding groups;
2466          *  - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
2467          *  - otherwise, do nothing more.
2468          */
2469         if (cpu_event)
2470                 cpu_ctx_sched_out(cpuctx, ctx_event_type);
2471         else if (ctx_event_type & EVENT_PINNED)
2472                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
2473 
2474         perf_event_sched_in(cpuctx, task_ctx, current);
2475         perf_pmu_enable(cpuctx->ctx.pmu);
2476 }
2477 
2478 /*
2479  * Cross CPU call to install and enable a performance event
2480  *
2481  * Very similar to remote_function() + event_function() but cannot assume that
2482  * things like ctx->is_active and cpuctx->task_ctx are set.
2483  */
2484 static int  __perf_install_in_context(void *info)
2485 {
2486         struct perf_event *event = info;
2487         struct perf_event_context *ctx = event->ctx;
2488         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
2489         struct perf_event_context *task_ctx = cpuctx->task_ctx;
2490         bool reprogram = true;
2491         int ret = 0;
2492 
2493         raw_spin_lock(&cpuctx->ctx.lock);
2494         if (ctx->task) {
2495                 raw_spin_lock(&ctx->lock);
2496                 task_ctx = ctx;
2497 
2498                 reprogram = (ctx->task == current);
2499 
2500                 /*
2501                  * If the task is running, it must be running on this CPU,
2502                  * otherwise we cannot reprogram things.
2503                  *
2504                  * If its not running, we don't care, ctx->lock will
2505                  * serialize against it becoming runnable.
2506                  */
2507                 if (task_curr(ctx->task) && !reprogram) {
2508                         ret = -ESRCH;
2509                         goto unlock;
2510                 }
2511 
2512                 WARN_ON_ONCE(reprogram && cpuctx->task_ctx && cpuctx->task_ctx != ctx);
2513         } else if (task_ctx) {
2514                 raw_spin_lock(&task_ctx->lock);
2515         }
2516 
2517 #ifdef CONFIG_CGROUP_PERF
2518         if (is_cgroup_event(event)) {
2519                 /*
2520                  * If the current cgroup doesn't match the event's
2521                  * cgroup, we should not try to schedule it.
2522                  */
2523                 struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
2524                 reprogram = cgroup_is_descendant(cgrp->css.cgroup,
2525                                         event->cgrp->css.cgroup);
2526         }
2527 #endif
2528 
2529         if (reprogram) {
2530                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2531                 add_event_to_ctx(event, ctx);
2532                 ctx_resched(cpuctx, task_ctx, get_event_type(event));
2533         } else {
2534                 add_event_to_ctx(event, ctx);
2535         }
2536 
2537 unlock:
2538         perf_ctx_unlock(cpuctx, task_ctx);
2539 
2540         return ret;
2541 }
2542 
2543 /*
2544  * Attach a performance event to a context.
2545  *
2546  * Very similar to event_function_call, see comment there.
2547  */
2548 static void
2549 perf_install_in_context(struct perf_event_context *ctx,
2550                         struct perf_event *event,
2551                         int cpu)
2552 {
2553         struct task_struct *task = READ_ONCE(ctx->task);
2554 
2555         lockdep_assert_held(&ctx->mutex);
2556 
2557         if (event->cpu != -1)
2558                 event->cpu = cpu;
2559 
2560         /*
2561          * Ensures that if we can observe event->ctx, both the event and ctx
2562          * will be 'complete'. See perf_iterate_sb_cpu().
2563          */
2564         smp_store_release(&event->ctx, ctx);
2565 
2566         if (!task) {
2567                 cpu_function_call(cpu, __perf_install_in_context, event);
2568                 return;
2569         }
2570 
2571         /*
2572          * Should not happen, we validate the ctx is still alive before calling.
2573          */
2574         if (WARN_ON_ONCE(task == TASK_TOMBSTONE))
2575                 return;
2576 
2577         /*
2578          * Installing events is tricky because we cannot rely on ctx->is_active
2579          * to be set in case this is the nr_events 0 -> 1 transition.
2580          *
2581          * Instead we use task_curr(), which tells us if the task is running.
2582          * However, since we use task_curr() outside of rq::lock, we can race
2583          * against the actual state. This means the result can be wrong.
2584          *
2585          * If we get a false positive, we retry, this is harmless.
2586          *
2587          * If we get a false negative, things are complicated. If we are after
2588          * perf_event_context_sched_in() ctx::lock will serialize us, and the
2589          * value must be correct. If we're before, it doesn't matter since
2590          * perf_event_context_sched_in() will program the counter.
2591          *
2592          * However, this hinges on the remote context switch having observed
2593          * our task->perf_event_ctxp[] store, such that it will in fact take
2594          * ctx::lock in perf_event_context_sched_in().
2595          *
2596          * We do this by task_function_call(), if the IPI fails to hit the task
2597          * we know any future context switch of task must see the
2598          * perf_event_ctpx[] store.
2599          */
2600 
2601         /*
2602          * This smp_mb() orders the task->perf_event_ctxp[] store with the
2603          * task_cpu() load, such that if the IPI then does not find the task
2604          * running, a future context switch of that task must observe the
2605          * store.
2606          */
2607         smp_mb();
2608 again:
2609         if (!task_function_call(task, __perf_install_in_context, event))
2610                 return;
2611 
2612         raw_spin_lock_irq(&ctx->lock);
2613         task = ctx->task;
2614         if (WARN_ON_ONCE(task == TASK_TOMBSTONE)) {
2615                 /*
2616                  * Cannot happen because we already checked above (which also
2617                  * cannot happen), and we hold ctx->mutex, which serializes us
2618                  * against perf_event_exit_task_context().
2619                  */
2620                 raw_spin_unlock_irq(&ctx->lock);
2621                 return;
2622         }
2623         /*
2624          * If the task is not running, ctx->lock will avoid it becoming so,
2625          * thus we can safely install the event.
2626          */
2627         if (task_curr(task)) {
2628                 raw_spin_unlock_irq(&ctx->lock);
2629                 goto again;
2630         }
2631         add_event_to_ctx(event, ctx);
2632         raw_spin_unlock_irq(&ctx->lock);
2633 }
2634 
2635 /*
2636  * Cross CPU call to enable a performance event
2637  */
2638 static void __perf_event_enable(struct perf_event *event,
2639                                 struct perf_cpu_context *cpuctx,
2640                                 struct perf_event_context *ctx,
2641                                 void *info)
2642 {
2643         struct perf_event *leader = event->group_leader;
2644         struct perf_event_context *task_ctx;
2645 
2646         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2647             event->state <= PERF_EVENT_STATE_ERROR)
2648                 return;
2649 
2650         if (ctx->is_active)
2651                 ctx_sched_out(ctx, cpuctx, EVENT_TIME);
2652 
2653         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
2654 
2655         if (!ctx->is_active)
2656                 return;
2657 
2658         if (!event_filter_match(event)) {
2659                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2660                 return;
2661         }
2662 
2663         /*
2664          * If the event is in a group and isn't the group leader,
2665          * then don't put it on unless the group is on.
2666          */
2667         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE) {
2668                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
2669                 return;
2670         }
2671 
2672         task_ctx = cpuctx->task_ctx;
2673         if (ctx->task)
2674                 WARN_ON_ONCE(task_ctx != ctx);
2675 
2676         ctx_resched(cpuctx, task_ctx, get_event_type(event));
2677 }
2678 
2679 /*
2680  * Enable an event.
2681  *
2682  * If event->ctx is a cloned context, callers must make sure that
2683  * every task struct that event->ctx->task could possibly point to
2684  * remains valid.  This condition is satisfied when called through
2685  * perf_event_for_each_child or perf_event_for_each as described
2686  * for perf_event_disable.
2687  */
2688 static void _perf_event_enable(struct perf_event *event)
2689 {
2690         struct perf_event_context *ctx = event->ctx;
2691 
2692         raw_spin_lock_irq(&ctx->lock);
2693         if (event->state >= PERF_EVENT_STATE_INACTIVE ||
2694             event->state <  PERF_EVENT_STATE_ERROR) {
2695                 raw_spin_unlock_irq(&ctx->lock);
2696                 return;
2697         }
2698 
2699         /*
2700          * If the event is in error state, clear that first.
2701          *
2702          * That way, if we see the event in error state below, we know that it
2703          * has gone back into error state, as distinct from the task having
2704          * been scheduled away before the cross-call arrived.
2705          */
2706         if (event->state == PERF_EVENT_STATE_ERROR)
2707                 event->state = PERF_EVENT_STATE_OFF;
2708         raw_spin_unlock_irq(&ctx->lock);
2709 
2710         event_function_call(event, __perf_event_enable, NULL);
2711 }
2712 
2713 /*
2714  * See perf_event_disable();
2715  */
2716 void perf_event_enable(struct perf_event *event)
2717 {
2718         struct perf_event_context *ctx;
2719 
2720         ctx = perf_event_ctx_lock(event);
2721         _perf_event_enable(event);
2722         perf_event_ctx_unlock(event, ctx);
2723 }
2724 EXPORT_SYMBOL_GPL(perf_event_enable);
2725 
2726 struct stop_event_data {
2727         struct perf_event       *event;
2728         unsigned int            restart;
2729 };
2730 
2731 static int __perf_event_stop(void *info)
2732 {
2733         struct stop_event_data *sd = info;
2734         struct perf_event *event = sd->event;
2735 
2736         /* if it's already INACTIVE, do nothing */
2737         if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2738                 return 0;
2739 
2740         /* matches smp_wmb() in event_sched_in() */
2741         smp_rmb();
2742 
2743         /*
2744          * There is a window with interrupts enabled before we get here,
2745          * so we need to check again lest we try to stop another CPU's event.
2746          */
2747         if (READ_ONCE(event->oncpu) != smp_processor_id())
2748                 return -EAGAIN;
2749 
2750         event->pmu->stop(event, PERF_EF_UPDATE);
2751 
2752         /*
2753          * May race with the actual stop (through perf_pmu_output_stop()),
2754          * but it is only used for events with AUX ring buffer, and such
2755          * events will refuse to restart because of rb::aux_mmap_count==0,
2756          * see comments in perf_aux_output_begin().
2757          *
2758          * Since this is happening on an event-local CPU, no trace is lost
2759          * while restarting.
2760          */
2761         if (sd->restart)
2762                 event->pmu->start(event, 0);
2763 
2764         return 0;
2765 }
2766 
2767 static int perf_event_stop(struct perf_event *event, int restart)
2768 {
2769         struct stop_event_data sd = {
2770                 .event          = event,
2771                 .restart        = restart,
2772         };
2773         int ret = 0;
2774 
2775         do {
2776                 if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
2777                         return 0;
2778 
2779                 /* matches smp_wmb() in event_sched_in() */
2780                 smp_rmb();
2781 
2782                 /*
2783                  * We only want to restart ACTIVE events, so if the event goes
2784                  * inactive here (event->oncpu==-1), there's nothing more to do;
2785                  * fall through with ret==-ENXIO.
2786                  */
2787                 ret = cpu_function_call(READ_ONCE(event->oncpu),
2788                                         __perf_event_stop, &sd);
2789         } while (ret == -EAGAIN);
2790 
2791         return ret;
2792 }
2793 
2794 /*
2795  * In order to contain the amount of racy and tricky in the address filter
2796  * configuration management, it is a two part process:
2797  *
2798  * (p1) when userspace mappings change as a result of (1) or (2) or (3) below,
2799  *      we update the addresses of corresponding vmas in
2800  *      event::addr_filters_offs array and bump the event::addr_filters_gen;
2801  * (p2) when an event is scheduled in (pmu::add), it calls
2802  *      perf_event_addr_filters_sync() which calls pmu::addr_filters_sync()
2803  *      if the generation has changed since the previous call.
2804  *
2805  * If (p1) happens while the event is active, we restart it to force (p2).
2806  *
2807  * (1) perf_addr_filters_apply(): adjusting filters' offsets based on
2808  *     pre-existing mappings, called once when new filters arrive via SET_FILTER
2809  *     ioctl;
2810  * (2) perf_addr_filters_adjust(): adjusting filters' offsets based on newly
2811  *     registered mapping, called for every new mmap(), with mm::mmap_sem down
2812  *     for reading;
2813  * (3) perf_event_addr_filters_exec(): clearing filters' offsets in the process
2814  *     of exec.
2815  */
2816 void perf_event_addr_filters_sync(struct perf_event *event)
2817 {
2818         struct perf_addr_filters_head *ifh = perf_event_addr_filters(event);
2819 
2820         if (!has_addr_filter(event))
2821                 return;
2822 
2823         raw_spin_lock(&ifh->lock);
2824         if (event->addr_filters_gen != event->hw.addr_filters_gen) {
2825                 event->pmu->addr_filters_sync(event);
2826                 event->hw.addr_filters_gen = event->addr_filters_gen;
2827         }
2828         raw_spin_unlock(&ifh->lock);
2829 }
2830 EXPORT_SYMBOL_GPL(perf_event_addr_filters_sync);
2831 
2832 static int _perf_event_refresh(struct perf_event *event, int refresh)
2833 {
2834         /*
2835          * not supported on inherited events
2836          */
2837         if (event->attr.inherit || !is_sampling_event(event))
2838                 return -EINVAL;
2839 
2840         atomic_add(refresh, &event->event_limit);
2841         _perf_event_enable(event);
2842 
2843         return 0;
2844 }
2845 
2846 /*
2847  * See perf_event_disable()
2848  */
2849 int perf_event_refresh(struct perf_event *event, int refresh)
2850 {
2851         struct perf_event_context *ctx;
2852         int ret;
2853 
2854         ctx = perf_event_ctx_lock(event);
2855         ret = _perf_event_refresh(event, refresh);
2856         perf_event_ctx_unlock(event, ctx);
2857 
2858         return ret;
2859 }
2860 EXPORT_SYMBOL_GPL(perf_event_refresh);
2861 
2862 static int perf_event_modify_breakpoint(struct perf_event *bp,
2863                                          struct perf_event_attr *attr)
2864 {
2865         int err;
2866 
2867         _perf_event_disable(bp);
2868 
2869         err = modify_user_hw_breakpoint_check(bp, attr, true);
2870 
2871         if (!bp->attr.disabled)
2872                 _perf_event_enable(bp);
2873 
2874         return err;
2875 }
2876 
2877 static int perf_event_modify_attr(struct perf_event *event,
2878                                   struct perf_event_attr *attr)
2879 {
2880         if (event->attr.type != attr->type)
2881                 return -EINVAL;
2882 
2883         switch (event->attr.type) {
2884         case PERF_TYPE_BREAKPOINT:
2885                 return perf_event_modify_breakpoint(event, attr);
2886         default:
2887                 /* Place holder for future additions. */
2888                 return -EOPNOTSUPP;
2889         }
2890 }
2891 
2892 static void ctx_sched_out(struct perf_event_context *ctx,
2893                           struct perf_cpu_context *cpuctx,
2894                           enum event_type_t event_type)
2895 {
2896         struct perf_event *event, *tmp;
2897         int is_active = ctx->is_active;
2898 
2899         lockdep_assert_held(&ctx->lock);
2900 
2901         if (likely(!ctx->nr_events)) {
2902                 /*
2903                  * See __perf_remove_from_context().
2904                  */
2905                 WARN_ON_ONCE(ctx->is_active);
2906                 if (ctx->task)
2907                         WARN_ON_ONCE(cpuctx->task_ctx);
2908                 return;
2909         }
2910 
2911         ctx->is_active &= ~event_type;
2912         if (!(ctx->is_active & EVENT_ALL))
2913                 ctx->is_active = 0;
2914 
2915         if (ctx->task) {
2916                 WARN_ON_ONCE(cpuctx->task_ctx != ctx);
2917                 if (!ctx->is_active)
2918                         cpuctx->task_ctx = NULL;
2919         }
2920 
2921         /*
2922          * Always update time if it was set; not only when it changes.
2923          * Otherwise we can 'forget' to update time for any but the last
2924          * context we sched out. For example:
2925          *
2926          *   ctx_sched_out(.event_type = EVENT_FLEXIBLE)
2927          *   ctx_sched_out(.event_type = EVENT_PINNED)
2928          *
2929          * would only update time for the pinned events.
2930          */
2931         if (is_active & EVENT_TIME) {
2932                 /* update (and stop) ctx time */
2933                 update_context_time(ctx);
2934                 update_cgrp_time_from_cpuctx(cpuctx);
2935         }
2936 
2937         is_active ^= ctx->is_active; /* changed bits */
2938 
2939         if (!ctx->nr_active || !(is_active & EVENT_ALL))
2940                 return;
2941 
2942         perf_pmu_disable(ctx->pmu);
2943         if (is_active & EVENT_PINNED) {
2944                 list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list)
2945                         group_sched_out(event, cpuctx, ctx);
2946         }
2947 
2948         if (is_active & EVENT_FLEXIBLE) {
2949                 list_for_each_entry_safe(event, tmp, &ctx->flexible_active, active_list)
2950                         group_sched_out(event, cpuctx, ctx);
2951         }
2952         perf_pmu_enable(ctx->pmu);
2953 }
2954 
2955 /*
2956  * Test whether two contexts are equivalent, i.e. whether they have both been
2957  * cloned from the same version of the same context.
2958  *
2959  * Equivalence is measured using a generation number in the context that is
2960  * incremented on each modification to it; see unclone_ctx(), list_add_event()
2961  * and list_del_event().
2962  */
2963 static int context_equiv(struct perf_event_context *ctx1,
2964                          struct perf_event_context *ctx2)
2965 {
2966         lockdep_assert_held(&ctx1->lock);
2967         lockdep_assert_held(&ctx2->lock);
2968 
2969         /* Pinning disables the swap optimization */
2970         if (ctx1->pin_count || ctx2->pin_count)
2971                 return 0;
2972 
2973         /* If ctx1 is the parent of ctx2 */
2974         if (ctx1 == ctx2->parent_ctx && ctx1->generation == ctx2->parent_gen)
2975                 return 1;
2976 
2977         /* If ctx2 is the parent of ctx1 */
2978         if (ctx1->parent_ctx == ctx2 && ctx1->parent_gen == ctx2->generation)
2979                 return 1;
2980 
2981         /*
2982          * If ctx1 and ctx2 have the same parent; we flatten the parent
2983          * hierarchy, see perf_event_init_context().
2984          */
2985         if (ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx &&
2986                         ctx1->parent_gen == ctx2->parent_gen)
2987                 return 1;
2988 
2989         /* Unmatched */
2990         return 0;
2991 }
2992 
2993 static void __perf_event_sync_stat(struct perf_event *event,
2994                                      struct perf_event *next_event)
2995 {
2996         u64 value;
2997 
2998         if (!event->attr.inherit_stat)
2999                 return;
3000 
3001         /*
3002          * Update the event value, we cannot use perf_event_read()
3003          * because we're in the middle of a context switch and have IRQs
3004          * disabled, which upsets smp_call_function_single(), however
3005          * we know the event must be on the current CPU, therefore we
3006          * don't need to use it.
3007          */
3008         if (event->state == PERF_EVENT_STATE_ACTIVE)
3009                 event->pmu->read(event);
3010 
3011         perf_event_update_time(event);
3012 
3013         /*
3014          * In order to keep per-task stats reliable we need to flip the event
3015          * values when we flip the contexts.
3016          */
3017         value = local64_read(&next_event->count);
3018         value = local64_xchg(&event->count, value);
3019         local64_set(&next_event->count, value);
3020 
3021         swap(event->total_time_enabled, next_event->total_time_enabled);
3022         swap(event->total_time_running, next_event->total_time_running);
3023 
3024         /*
3025          * Since we swizzled the values, update the user visible data too.
3026          */
3027         perf_event_update_userpage(event);
3028         perf_event_update_userpage(next_event);
3029 }
3030 
3031 static void perf_event_sync_stat(struct perf_event_context *ctx,
3032                                    struct perf_event_context *next_ctx)
3033 {
3034         struct perf_event *event, *next_event;
3035 
3036         if (!ctx->nr_stat)
3037                 return;
3038 
3039         update_context_time(ctx);
3040 
3041         event = list_first_entry(&ctx->event_list,
3042                                    struct perf_event, event_entry);
3043 
3044         next_event = list_first_entry(&next_ctx->event_list,
3045                                         struct perf_event, event_entry);
3046 
3047         while (&event->event_entry != &ctx->event_list &&
3048                &next_event->event_entry != &next_ctx->event_list) {
3049 
3050                 __perf_event_sync_stat(event, next_event);
3051 
3052                 event = list_next_entry(event, event_entry);
3053                 next_event = list_next_entry(next_event, event_entry);
3054         }
3055 }
3056 
3057 static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
3058                                          struct task_struct *next)
3059 {
3060         struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
3061         struct perf_event_context *next_ctx;
3062         struct perf_event_context *parent, *next_parent;
3063         struct perf_cpu_context *cpuctx;
3064         int do_switch = 1;
3065 
3066         if (likely(!ctx))
3067                 return;
3068 
3069         cpuctx = __get_cpu_context(ctx);
3070         if (!cpuctx->task_ctx)
3071                 return;
3072 
3073         rcu_read_lock();
3074         next_ctx = next->perf_event_ctxp[ctxn];
3075         if (!next_ctx)
3076                 goto unlock;
3077 
3078         parent = rcu_dereference(ctx->parent_ctx);
3079         next_parent = rcu_dereference(next_ctx->parent_ctx);
3080 
3081         /* If neither context have a parent context; they cannot be clones. */
3082         if (!parent && !next_parent)
3083                 goto unlock;
3084 
3085         if (next_parent == ctx || next_ctx == parent || next_parent == parent) {
3086                 /*
3087                  * Looks like the two contexts are clones, so we might be
3088                  * able to optimize the context switch.  We lock both
3089                  * contexts and check that they are clones under the
3090                  * lock (including re-checking that neither has been
3091                  * uncloned in the meantime).  It doesn't matter which
3092                  * order we take the locks because no other cpu could
3093                  * be trying to lock both of these tasks.
3094                  */
3095                 raw_spin_lock(&ctx->lock);
3096                 raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
3097                 if (context_equiv(ctx, next_ctx)) {
3098                         WRITE_ONCE(ctx->task, next);
3099                         WRITE_ONCE(next_ctx->task, task);
3100 
3101                         swap(ctx->task_ctx_data, next_ctx->task_ctx_data);
3102 
3103                         /*
3104                          * RCU_INIT_POINTER here is safe because we've not
3105                          * modified the ctx and the above modification of
3106                          * ctx->task and ctx->task_ctx_data are immaterial
3107                          * since those values are always verified under
3108                          * ctx->lock which we're now holding.
3109                          */
3110                         RCU_INIT_POINTER(task->perf_event_ctxp[ctxn], next_ctx);
3111                         RCU_INIT_POINTER(next->perf_event_ctxp[ctxn], ctx);
3112 
3113                         do_switch = 0;
3114 
3115                         perf_event_sync_stat(ctx, next_ctx);
3116                 }
3117                 raw_spin_unlock(&next_ctx->lock);
3118                 raw_spin_unlock(&ctx->lock);
3119         }
3120 unlock:
3121         rcu_read_unlock();
3122 
3123         if (do_switch) {
3124                 raw_spin_lock(&ctx->lock);
3125                 task_ctx_sched_out(cpuctx, ctx, EVENT_ALL);
3126                 raw_spin_unlock(&ctx->lock);
3127         }
3128 }
3129 
3130 static DEFINE_PER_CPU(struct list_head, sched_cb_list);
3131 
3132 void perf_sched_cb_dec(struct pmu *pmu)
3133 {
3134         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3135 
3136         this_cpu_dec(perf_sched_cb_usages);
3137 
3138         if (!--cpuctx->sched_cb_usage)
3139                 list_del(&cpuctx->sched_cb_entry);
3140 }
3141 
3142 
3143 void perf_sched_cb_inc(struct pmu *pmu)
3144 {
3145         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
3146 
3147         if (!cpuctx->sched_cb_usage++)
3148                 list_add(&cpuctx->sched_cb_entry, this_cpu_ptr(&sched_cb_list));
3149 
3150         this_cpu_inc(perf_sched_cb_usages);
3151 }
3152 
3153 /*
3154  * This function provides the context switch callback to the lower code
3155  * layer. It is invoked ONLY when the context switch callback is enabled.
3156  *
3157  * This callback is relevant even to per-cpu events; for example multi event
3158  * PEBS requires this to provide PID/TID information. This requires we flush
3159  * all queued PEBS records before we context switch to a new task.
3160  */
3161 static void perf_pmu_sched_task(struct task_struct *prev,
3162                                 struct task_struct *next,
3163                                 bool sched_in)
3164 {
3165         struct perf_cpu_context *cpuctx;
3166         struct pmu *pmu;
3167 
3168         if (prev == next)
3169                 return;
3170 
3171         list_for_each_entry(cpuctx, this_cpu_ptr(&sched_cb_list), sched_cb_entry) {
3172                 pmu = cpuctx->ctx.pmu; /* software PMUs will not have sched_task */
3173 
3174                 if (WARN_ON_ONCE(!pmu->sched_task))
3175                         continue;
3176 
3177                 perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3178                 perf_pmu_disable(pmu);
3179 
3180                 pmu->sched_task(cpuctx->task_ctx, sched_in);
3181 
3182                 perf_pmu_enable(pmu);
3183                 perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3184         }
3185 }
3186 
3187 static void perf_event_switch(struct task_struct *task,
3188                               struct task_struct *next_prev, bool sched_in);
3189 
3190 #define for_each_task_context_nr(ctxn)                                  \
3191         for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
3192 
3193 /*
3194  * Called from scheduler to remove the events of the current task,
3195  * with interrupts disabled.
3196  *
3197  * We stop each event and update the event value in event->count.
3198  *
3199  * This does not protect us against NMI, but disable()
3200  * sets the disabled bit in the control field of event _before_
3201  * accessing the event control register. If a NMI hits, then it will
3202  * not restart the event.
3203  */
3204 void __perf_event_task_sched_out(struct task_struct *task,
3205                                  struct task_struct *next)
3206 {
3207         int ctxn;
3208 
3209         if (__this_cpu_read(perf_sched_cb_usages))
3210                 perf_pmu_sched_task(task, next, false);
3211 
3212         if (atomic_read(&nr_switch_events))
3213                 perf_event_switch(task, next, false);
3214 
3215         for_each_task_context_nr(ctxn)
3216                 perf_event_context_sched_out(task, ctxn, next);
3217 
3218         /*
3219          * if cgroup events exist on this CPU, then we need
3220          * to check if we have to switch out PMU state.
3221          * cgroup event are system-wide mode only
3222          */
3223         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3224                 perf_cgroup_sched_out(task, next);
3225 }
3226 
3227 /*
3228  * Called with IRQs disabled
3229  */
3230 static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
3231                               enum event_type_t event_type)
3232 {
3233         ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
3234 }
3235 
3236 static int visit_groups_merge(struct perf_event_groups *groups, int cpu,
3237                               int (*func)(struct perf_event *, void *), void *data)
3238 {
3239         struct perf_event **evt, *evt1, *evt2;
3240         int ret;
3241 
3242         evt1 = perf_event_groups_first(groups, -1);
3243         evt2 = perf_event_groups_first(groups, cpu);
3244 
3245         while (evt1 || evt2) {
3246                 if (evt1 && evt2) {
3247                         if (evt1->group_index < evt2->group_index)
3248                                 evt = &evt1;
3249                         else
3250                                 evt = &evt2;
3251                 } else if (evt1) {
3252                         evt = &evt1;
3253                 } else {
3254                         evt = &evt2;
3255                 }
3256 
3257                 ret = func(*evt, data);
3258                 if (ret)
3259                         return ret;
3260 
3261                 *evt = perf_event_groups_next(*evt);
3262         }
3263 
3264         return 0;
3265 }
3266 
3267 struct sched_in_data {
3268         struct perf_event_context *ctx;
3269         struct perf_cpu_context *cpuctx;
3270         int can_add_hw;
3271 };
3272 
3273 static int pinned_sched_in(struct perf_event *event, void *data)
3274 {
3275         struct sched_in_data *sid = data;
3276 
3277         if (event->state <= PERF_EVENT_STATE_OFF)
3278                 return 0;
3279 
3280         if (!event_filter_match(event))
3281                 return 0;
3282 
3283         if (group_can_go_on(event, sid->cpuctx, sid->can_add_hw)) {
3284                 if (!group_sched_in(event, sid->cpuctx, sid->ctx))
3285                         list_add_tail(&event->active_list, &sid->ctx->pinned_active);
3286         }
3287 
3288         /*
3289          * If this pinned group hasn't been scheduled,
3290          * put it in error state.
3291          */
3292         if (event->state == PERF_EVENT_STATE_INACTIVE)
3293                 perf_event_set_state(event, PERF_EVENT_STATE_ERROR);
3294 
3295         return 0;
3296 }
3297 
3298 static int flexible_sched_in(struct perf_event *event, void *data)
3299 {
3300         struct sched_in_data *sid = data;
3301 
3302         if (event->state <= PERF_EVENT_STATE_OFF)
3303                 return 0;
3304 
3305         if (!event_filter_match(event))
3306                 return 0;
3307 
3308         if (group_can_go_on(event, sid->cpuctx, sid->can_add_hw)) {
3309                 if (!group_sched_in(event, sid->cpuctx, sid->ctx))
3310                         list_add_tail(&event->active_list, &sid->ctx->flexible_active);
3311                 else
3312                         sid->can_add_hw = 0;
3313         }
3314 
3315         return 0;
3316 }
3317 
3318 static void
3319 ctx_pinned_sched_in(struct perf_event_context *ctx,
3320                     struct perf_cpu_context *cpuctx)
3321 {
3322         struct sched_in_data sid = {
3323                 .ctx = ctx,
3324                 .cpuctx = cpuctx,
3325                 .can_add_hw = 1,
3326         };
3327 
3328         visit_groups_merge(&ctx->pinned_groups,
3329                            smp_processor_id(),
3330                            pinned_sched_in, &sid);
3331 }
3332 
3333 static void
3334 ctx_flexible_sched_in(struct perf_event_context *ctx,
3335                       struct perf_cpu_context *cpuctx)
3336 {
3337         struct sched_in_data sid = {
3338                 .ctx = ctx,
3339                 .cpuctx = cpuctx,
3340                 .can_add_hw = 1,
3341         };
3342 
3343         visit_groups_merge(&ctx->flexible_groups,
3344                            smp_processor_id(),
3345                            flexible_sched_in, &sid);
3346 }
3347 
3348 static void
3349 ctx_sched_in(struct perf_event_context *ctx,
3350              struct perf_cpu_context *cpuctx,
3351              enum event_type_t event_type,
3352              struct task_struct *task)
3353 {
3354         int is_active = ctx->is_active;
3355         u64 now;
3356 
3357         lockdep_assert_held(&ctx->lock);
3358 
3359         if (likely(!ctx->nr_events))
3360                 return;
3361 
3362         ctx->is_active |= (event_type | EVENT_TIME);
3363         if (ctx->task) {
3364                 if (!is_active)
3365                         cpuctx->task_ctx = ctx;
3366                 else
3367                         WARN_ON_ONCE(cpuctx->task_ctx != ctx);
3368         }
3369 
3370         is_active ^= ctx->is_active; /* changed bits */
3371 
3372         if (is_active & EVENT_TIME) {
3373                 /* start ctx time */
3374                 now = perf_clock();
3375                 ctx->timestamp = now;
3376                 perf_cgroup_set_timestamp(task, ctx);
3377         }
3378 
3379         /*
3380          * First go through the list and put on any pinned groups
3381          * in order to give them the best chance of going on.
3382          */
3383         if (is_active & EVENT_PINNED)
3384                 ctx_pinned_sched_in(ctx, cpuctx);
3385 
3386         /* Then walk through the lower prio flexible groups */
3387         if (is_active & EVENT_FLEXIBLE)
3388                 ctx_flexible_sched_in(ctx, cpuctx);
3389 }
3390 
3391 static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
3392                              enum event_type_t event_type,
3393                              struct task_struct *task)
3394 {
3395         struct perf_event_context *ctx = &cpuctx->ctx;
3396 
3397         ctx_sched_in(ctx, cpuctx, event_type, task);
3398 }
3399 
3400 static void perf_event_context_sched_in(struct perf_event_context *ctx,
3401                                         struct task_struct *task)
3402 {
3403         struct perf_cpu_context *cpuctx;
3404 
3405         cpuctx = __get_cpu_context(ctx);
3406         if (cpuctx->task_ctx == ctx)
3407                 return;
3408 
3409         perf_ctx_lock(cpuctx, ctx);
3410         /*
3411          * We must check ctx->nr_events while holding ctx->lock, such
3412          * that we serialize against perf_install_in_context().
3413          */
3414         if (!ctx->nr_events)
3415                 goto unlock;
3416 
3417         perf_pmu_disable(ctx->pmu);
3418         /*
3419          * We want to keep the following priority order:
3420          * cpu pinned (that don't need to move), task pinned,
3421          * cpu flexible, task flexible.
3422          *
3423          * However, if task's ctx is not carrying any pinned
3424          * events, no need to flip the cpuctx's events around.
3425          */
3426         if (!RB_EMPTY_ROOT(&ctx->pinned_groups.tree))
3427                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3428         perf_event_sched_in(cpuctx, ctx, task);
3429         perf_pmu_enable(ctx->pmu);
3430 
3431 unlock:
3432         perf_ctx_unlock(cpuctx, ctx);
3433 }
3434 
3435 /*
3436  * Called from scheduler to add the events of the current task
3437  * with interrupts disabled.
3438  *
3439  * We restore the event value and then enable it.
3440  *
3441  * This does not protect us against NMI, but enable()
3442  * sets the enabled bit in the control field of event _before_
3443  * accessing the event control register. If a NMI hits, then it will
3444  * keep the event running.
3445  */
3446 void __perf_event_task_sched_in(struct task_struct *prev,
3447                                 struct task_struct *task)
3448 {
3449         struct perf_event_context *ctx;
3450         int ctxn;
3451 
3452         /*
3453          * If cgroup events exist on this CPU, then we need to check if we have
3454          * to switch in PMU state; cgroup event are system-wide mode only.
3455          *
3456          * Since cgroup events are CPU events, we must schedule these in before
3457          * we schedule in the task events.
3458          */
3459         if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
3460                 perf_cgroup_sched_in(prev, task);
3461 
3462         for_each_task_context_nr(ctxn) {
3463                 ctx = task->perf_event_ctxp[ctxn];
3464                 if (likely(!ctx))
3465                         continue;
3466 
3467                 perf_event_context_sched_in(ctx, task);
3468         }
3469 
3470         if (atomic_read(&nr_switch_events))
3471                 perf_event_switch(task, prev, true);
3472 
3473         if (__this_cpu_read(perf_sched_cb_usages))
3474                 perf_pmu_sched_task(prev, task, true);
3475 }
3476 
3477 static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
3478 {
3479         u64 frequency = event->attr.sample_freq;
3480         u64 sec = NSEC_PER_SEC;
3481         u64 divisor, dividend;
3482 
3483         int count_fls, nsec_fls, frequency_fls, sec_fls;
3484 
3485         count_fls = fls64(count);
3486         nsec_fls = fls64(nsec);
3487         frequency_fls = fls64(frequency);
3488         sec_fls = 30;
3489 
3490         /*
3491          * We got @count in @nsec, with a target of sample_freq HZ
3492          * the target period becomes:
3493          *
3494          *             @count * 10^9
3495          * period = -------------------
3496          *          @nsec * sample_freq
3497          *
3498          */
3499 
3500         /*
3501          * Reduce accuracy by one bit such that @a and @b converge
3502          * to a similar magnitude.
3503          */
3504 #define REDUCE_FLS(a, b)                \
3505 do {                                    \
3506         if (a##_fls > b##_fls) {        \
3507                 a >>= 1;                \
3508                 a##_fls--;              \
3509         } else {                        \
3510                 b >>= 1;                \
3511                 b##_fls--;              \
3512         }                               \
3513 } while (0)
3514 
3515         /*
3516          * Reduce accuracy until either term fits in a u64, then proceed with
3517          * the other, so that finally we can do a u64/u64 division.
3518          */
3519         while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
3520                 REDUCE_FLS(nsec, frequency);
3521                 REDUCE_FLS(sec, count);
3522         }
3523 
3524         if (count_fls + sec_fls > 64) {
3525                 divisor = nsec * frequency;
3526 
3527                 while (count_fls + sec_fls > 64) {
3528                         REDUCE_FLS(count, sec);
3529                         divisor >>= 1;
3530                 }
3531 
3532                 dividend = count * sec;
3533         } else {
3534                 dividend = count * sec;
3535 
3536                 while (nsec_fls + frequency_fls > 64) {
3537                         REDUCE_FLS(nsec, frequency);
3538                         dividend >>= 1;
3539                 }
3540 
3541                 divisor = nsec * frequency;
3542         }
3543 
3544         if (!divisor)
3545                 return dividend;
3546 
3547         return div64_u64(dividend, divisor);
3548 }
3549 
3550 static DEFINE_PER_CPU(int, perf_throttled_count);
3551 static DEFINE_PER_CPU(u64, perf_throttled_seq);
3552 
3553 static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bool disable)
3554 {
3555         struct hw_perf_event *hwc = &event->hw;
3556         s64 period, sample_period;
3557         s64 delta;
3558 
3559         period = perf_calculate_period(event, nsec, count);
3560 
3561         delta = (s64)(period - hwc->sample_period);
3562         delta = (delta + 7) / 8; /* low pass filter */
3563 
3564         sample_period = hwc->sample_period + delta;
3565 
3566         if (!sample_period)
3567                 sample_period = 1;
3568 
3569         hwc->sample_period = sample_period;
3570 
3571         if (local64_read(&hwc->period_left) > 8*sample_period) {
3572                 if (disable)
3573                         event->pmu->stop(event, PERF_EF_UPDATE);
3574 
3575                 local64_set(&hwc->period_left, 0);
3576 
3577                 if (disable)
3578                         event->pmu->start(event, PERF_EF_RELOAD);
3579         }
3580 }
3581 
3582 /*
3583  * combine freq adjustment with unthrottling to avoid two passes over the
3584  * events. At the same time, make sure, having freq events does not change
3585  * the rate of unthrottling as that would introduce bias.
3586  */
3587 static void perf_adjust_freq_unthr_context(struct perf_event_context *ctx,
3588                                            int needs_unthr)
3589 {
3590         struct perf_event *event;
3591         struct hw_perf_event *hwc;
3592         u64 now, period = TICK_NSEC;
3593         s64 delta;
3594 
3595         /*
3596          * only need to iterate over all events iff:
3597          * - context have events in frequency mode (needs freq adjust)
3598          * - there are events to unthrottle on this cpu
3599          */
3600         if (!(ctx->nr_freq || needs_unthr))
3601                 return;
3602 
3603         raw_spin_lock(&ctx->lock);
3604         perf_pmu_disable(ctx->pmu);
3605 
3606         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3607                 if (event->state != PERF_EVENT_STATE_ACTIVE)
3608                         continue;
3609 
3610                 if (!event_filter_match(event))
3611                         continue;
3612 
3613                 perf_pmu_disable(event->pmu);
3614 
3615                 hwc = &event->hw;
3616 
3617                 if (hwc->interrupts == MAX_INTERRUPTS) {
3618                         hwc->interrupts = 0;
3619                         perf_log_throttle(event, 1);
3620                         event->pmu->start(event, 0);
3621                 }
3622 
3623                 if (!event->attr.freq || !event->attr.sample_freq)
3624                         goto next;
3625 
3626                 /*
3627                  * stop the event and update event->count
3628                  */
3629                 event->pmu->stop(event, PERF_EF_UPDATE);
3630 
3631                 now = local64_read(&event->count);
3632                 delta = now - hwc->freq_count_stamp;
3633                 hwc->freq_count_stamp = now;
3634 
3635                 /*
3636                  * restart the event
3637                  * reload only if value has changed
3638                  * we have stopped the event so tell that
3639                  * to perf_adjust_period() to avoid stopping it
3640                  * twice.
3641                  */
3642                 if (delta > 0)
3643                         perf_adjust_period(event, period, delta, false);
3644 
3645                 event->pmu->start(event, delta > 0 ? PERF_EF_RELOAD : 0);
3646         next:
3647                 perf_pmu_enable(event->pmu);
3648         }
3649 
3650         perf_pmu_enable(ctx->pmu);
3651         raw_spin_unlock(&ctx->lock);
3652 }
3653 
3654 /*
3655  * Move @event to the tail of the @ctx's elegible events.
3656  */
3657 static void rotate_ctx(struct perf_event_context *ctx, struct perf_event *event)
3658 {
3659         /*
3660          * Rotate the first entry last of non-pinned groups. Rotation might be
3661          * disabled by the inheritance code.
3662          */
3663         if (ctx->rotate_disable)
3664                 return;
3665 
3666         perf_event_groups_delete(&ctx->flexible_groups, event);
3667         perf_event_groups_insert(&ctx->flexible_groups, event);
3668 }
3669 
3670 static inline struct perf_event *
3671 ctx_first_active(struct perf_event_context *ctx)
3672 {
3673         return list_first_entry_or_null(&ctx->flexible_active,
3674                                         struct perf_event, active_list);
3675 }
3676 
3677 static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
3678 {
3679         struct perf_event *cpu_event = NULL, *task_event = NULL;
3680         bool cpu_rotate = false, task_rotate = false;
3681         struct perf_event_context *ctx = NULL;
3682 
3683         /*
3684          * Since we run this from IRQ context, nobody can install new
3685          * events, thus the event count values are stable.
3686          */
3687 
3688         if (cpuctx->ctx.nr_events) {
3689                 if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
3690                         cpu_rotate = true;
3691         }
3692 
3693         ctx = cpuctx->task_ctx;
3694         if (ctx && ctx->nr_events) {
3695                 if (ctx->nr_events != ctx->nr_active)
3696                         task_rotate = true;
3697         }
3698 
3699         if (!(cpu_rotate || task_rotate))
3700                 return false;
3701 
3702         perf_ctx_lock(cpuctx, cpuctx->task_ctx);
3703         perf_pmu_disable(cpuctx->ctx.pmu);
3704 
3705         if (task_rotate)
3706                 task_event = ctx_first_active(ctx);
3707         if (cpu_rotate)
3708                 cpu_event = ctx_first_active(&cpuctx->ctx);
3709 
3710         /*
3711          * As per the order given at ctx_resched() first 'pop' task flexible
3712          * and then, if needed CPU flexible.
3713          */
3714         if (task_event || (ctx && cpu_event))
3715                 ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
3716         if (cpu_event)
3717                 cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
3718 
3719         if (task_event)
3720                 rotate_ctx(ctx, task_event);
3721         if (cpu_event)
3722                 rotate_ctx(&cpuctx->ctx, cpu_event);
3723 
3724         perf_event_sched_in(cpuctx, ctx, current);
3725 
3726         perf_pmu_enable(cpuctx->ctx.pmu);
3727         perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
3728 
3729         return true;
3730 }
3731 
3732 void perf_event_task_tick(void)
3733 {
3734         struct list_head *head = this_cpu_ptr(&active_ctx_list);
3735         struct perf_event_context *ctx, *tmp;
3736         int throttled;
3737 
3738         lockdep_assert_irqs_disabled();
3739 
3740         __this_cpu_inc(perf_throttled_seq);
3741         throttled = __this_cpu_xchg(perf_throttled_count, 0);
3742         tick_dep_clear_cpu(smp_processor_id(), TICK_DEP_BIT_PERF_EVENTS);
3743 
3744         list_for_each_entry_safe(ctx, tmp, head, active_ctx_list)
3745                 perf_adjust_freq_unthr_context(ctx, throttled);
3746 }
3747 
3748 static int event_enable_on_exec(struct perf_event *event,
3749                                 struct perf_event_context *ctx)
3750 {
3751         if (!event->attr.enable_on_exec)
3752                 return 0;
3753 
3754         event->attr.enable_on_exec = 0;
3755         if (event->state >= PERF_EVENT_STATE_INACTIVE)
3756                 return 0;
3757 
3758         perf_event_set_state(event, PERF_EVENT_STATE_INACTIVE);
3759 
3760         return 1;
3761 }
3762 
3763 /*
3764  * Enable all of a task's events that have been marked enable-on-exec.
3765  * This expects task == current.
3766  */
3767 static void perf_event_enable_on_exec(int ctxn)
3768 {
3769         struct perf_event_context *ctx, *clone_ctx = NULL;
3770         enum event_type_t event_type = 0;
3771         struct perf_cpu_context *cpuctx;
3772         struct perf_event *event;
3773         unsigned long flags;
3774         int enabled = 0;
3775 
3776         local_irq_save(flags);
3777         ctx = current->perf_event_ctxp[ctxn];
3778         if (!ctx || !ctx->nr_events)
3779                 goto out;
3780 
3781         cpuctx = __get_cpu_context(ctx);
3782         perf_ctx_lock(cpuctx, ctx);
3783         ctx_sched_out(ctx, cpuctx, EVENT_TIME);
3784         list_for_each_entry(event, &ctx->event_list, event_entry) {
3785                 enabled |= event_enable_on_exec(event, ctx);
3786                 event_type |= get_event_type(event);
3787         }
3788 
3789         /*
3790          * Unclone and reschedule this context if we enabled any event.
3791          */
3792         if (enabled) {
3793                 clone_ctx = unclone_ctx(ctx);
3794                 ctx_resched(cpuctx, ctx, event_type);
3795         } else {
3796                 ctx_sched_in(ctx, cpuctx, EVENT_TIME, current);
3797         }
3798         perf_ctx_unlock(cpuctx, ctx);
3799 
3800 out:
3801         local_irq_restore(flags);
3802 
3803         if (clone_ctx)
3804                 put_ctx(clone_ctx);
3805 }
3806 
3807 struct perf_read_data {
3808         struct perf_event *event;
3809         bool group;
3810         int ret;
3811 };
3812 
3813 static int __perf_event_read_cpu(struct perf_event *event, int event_cpu)
3814 {
3815         u16 local_pkg, event_pkg;
3816 
3817         if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) {
3818                 int local_cpu = smp_processor_id();
3819 
3820                 event_pkg = topology_physical_package_id(event_cpu);
3821                 local_pkg = topology_physical_package_id(local_cpu);
3822 
3823                 if (event_pkg == local_pkg)
3824                         return local_cpu;
3825         }
3826 
3827         return event_cpu;
3828 }
3829 
3830 /*
3831  * Cross CPU call to read the hardware event
3832  */
3833 static void __perf_event_read(void *info)
3834 {
3835         struct perf_read_data *data = info;
3836         struct perf_event *sub, *event = data->event;
3837         struct perf_event_context *ctx = event->ctx;
3838         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
3839         struct pmu *pmu = event->pmu;
3840 
3841         /*
3842          * If this is a task context, we need to check whether it is
3843          * the current task context of this cpu.  If not it has been
3844          * scheduled out before the smp call arrived.  In that case
3845          * event->count would have been updated to a recent sample
3846          * when the event was scheduled out.
3847          */
3848         if (ctx->task && cpuctx->task_ctx != ctx)
3849                 return;
3850 
3851         raw_spin_lock(&ctx->lock);
3852         if (ctx->is_active & EVENT_TIME) {
3853                 update_context_time(ctx);
3854                 update_cgrp_time_from_event(event);
3855         }
3856 
3857         perf_event_update_time(event);
3858         if (data->group)
3859                 perf_event_update_sibling_time(event);
3860 
3861         if (event->state != PERF_EVENT_STATE_ACTIVE)
3862                 goto unlock;
3863 
3864         if (!data->group) {
3865                 pmu->read(event);
3866                 data->ret = 0;
3867                 goto unlock;
3868         }
3869 
3870         pmu->start_txn(pmu, PERF_PMU_TXN_READ);
3871 
3872         pmu->read(event);
3873 
3874         for_each_sibling_event(sub, event) {
3875                 if (sub->state == PERF_EVENT_STATE_ACTIVE) {
3876                         /*
3877                          * Use sibling's PMU rather than @event's since
3878                          * sibling could be on different (eg: software) PMU.
3879                          */
3880                         sub->pmu->read(sub);
3881                 }
3882         }
3883 
3884         data->ret = pmu->commit_txn(pmu);
3885 
3886 unlock:
3887         raw_spin_unlock(&ctx->lock);
3888 }
3889 
3890 static inline u64 perf_event_count(struct perf_event *event)
3891 {
3892         return local64_read(&event->count) + atomic64_read(&event->child_count);
3893 }
3894 
3895 /*
3896  * NMI-safe method to read a local event, that is an event that
3897  * is:
3898  *   - either for the current task, or for this CPU
3899  *   - does not have inherit set, for inherited task events
3900  *     will not be local and we cannot read them atomically
3901  *   - must not have a pmu::count method
3902  */
3903 int perf_event_read_local(struct perf_event *event, u64 *value,
3904                           u64 *enabled, u64 *running)
3905 {
3906         unsigned long flags;
3907         int ret = 0;
3908 
3909         /*
3910          * Disabling interrupts avoids all counter scheduling (context
3911          * switches, timer based rotation and IPIs).
3912          */
3913         local_irq_save(flags);
3914 
3915         /*
3916          * It must not be an event with inherit set, we cannot read
3917          * all child counters from atomic context.
3918          */
3919         if (event->attr.inherit) {
3920                 ret = -EOPNOTSUPP;
3921                 goto out;
3922         }
3923 
3924         /* If this is a per-task event, it must be for current */
3925         if ((event->attach_state & PERF_ATTACH_TASK) &&
3926             event->hw.target != current) {
3927                 ret = -EINVAL;
3928                 goto out;
3929         }
3930 
3931         /* If this is a per-CPU event, it must be for this CPU */
3932         if (!(event->attach_state & PERF_ATTACH_TASK) &&
3933             event->cpu != smp_processor_id()) {
3934                 ret = -EINVAL;
3935                 goto out;
3936         }
3937 
3938         /* If this is a pinned event it must be running on this CPU */
3939         if (event->attr.pinned && event->oncpu != smp_processor_id()) {
3940                 ret = -EBUSY;
3941                 goto out;
3942         }
3943 
3944         /*
3945          * If the event is currently on this CPU, its either a per-task event,
3946          * or local to this CPU. Furthermore it means its ACTIVE (otherwise
3947          * oncpu == -1).
3948          */
3949         if (event->oncpu == smp_processor_id())
3950                 event->pmu->read(event);
3951 
3952         *value = local64_read(&event->count);
3953         if (enabled || running) {
3954                 u64 now = event->shadow_ctx_time + perf_clock();
3955                 u64 __enabled, __running;
3956 
3957                 __perf_update_times(event, now, &__enabled, &__running);
3958                 if (enabled)
3959                         *enabled = __enabled;
3960                 if (running)
3961                         *running = __running;
3962         }
3963 out:
3964         local_irq_restore(flags);
3965 
3966         return ret;
3967 }
3968 
3969 static int perf_event_read(struct perf_event *event, bool group)
3970 {
3971         enum perf_event_state state = READ_ONCE(event->state);
3972         int event_cpu, ret = 0;
3973 
3974         /*
3975          * If event is enabled and currently active on a CPU, update the
3976          * value in the event structure:
3977          */
3978 again:
3979         if (state == PERF_EVENT_STATE_ACTIVE) {
3980                 struct perf_read_data data;
3981 
3982                 /*
3983                  * Orders the ->state and ->oncpu loads such that if we see
3984                  * ACTIVE we must also see the right ->oncpu.
3985                  *
3986                  * Matches the smp_wmb() from event_sched_in().
3987                  */
3988                 smp_rmb();
3989 
3990                 event_cpu = READ_ONCE(event->oncpu);
3991                 if ((unsigned)event_cpu >= nr_cpu_ids)
3992                         return 0;
3993 
3994                 data = (struct perf_read_data){
3995                         .event = event,
3996                         .group = group,
3997                         .ret = 0,
3998                 };
3999 
4000                 preempt_disable();
4001                 event_cpu = __perf_event_read_cpu(event, event_cpu);
4002 
4003                 /*
4004                  * Purposely ignore the smp_call_function_single() return
4005                  * value.
4006                  *
4007                  * If event_cpu isn't a valid CPU it means the event got
4008                  * scheduled out and that will have updated the event count.
4009                  *
4010                  * Therefore, either way, we'll have an up-to-date event count
4011                  * after this.
4012                  */
4013                 (void)smp_call_function_single(event_cpu, __perf_event_read, &data, 1);
4014                 preempt_enable();
4015                 ret = data.ret;
4016 
4017         } else if (state == PERF_EVENT_STATE_INACTIVE) {
4018                 struct perf_event_context *ctx = event->ctx;
4019                 unsigned long flags;
4020 
4021                 raw_spin_lock_irqsave(&ctx->lock, flags);
4022                 state = event->state;
4023                 if (state != PERF_EVENT_STATE_INACTIVE) {
4024                         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4025                         goto again;
4026                 }
4027 
4028                 /*
4029                  * May read while context is not active (e.g., thread is
4030                  * blocked), in that case we cannot update context time
4031                  */
4032                 if (ctx->is_active & EVENT_TIME) {
4033                         update_context_time(ctx);
4034                         update_cgrp_time_from_event(event);
4035                 }
4036 
4037                 perf_event_update_time(event);
4038                 if (group)
4039                         perf_event_update_sibling_time(event);
4040                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4041         }
4042 
4043         return ret;
4044 }
4045 
4046 /*
4047  * Initialize the perf_event context in a task_struct:
4048  */
4049 static void __perf_event_init_context(struct perf_event_context *ctx)
4050 {
4051         raw_spin_lock_init(&ctx->lock);
4052         mutex_init(&ctx->mutex);
4053         INIT_LIST_HEAD(&ctx->active_ctx_list);
4054         perf_event_groups_init(&ctx->pinned_groups);
4055         perf_event_groups_init(&ctx->flexible_groups);
4056         INIT_LIST_HEAD(&ctx->event_list);
4057         INIT_LIST_HEAD(&ctx->pinned_active);
4058         INIT_LIST_HEAD(&ctx->flexible_active);
4059         atomic_set(&ctx->refcount, 1);
4060 }
4061 
4062 static struct perf_event_context *
4063 alloc_perf_context(struct pmu *pmu, struct task_struct *task)
4064 {
4065         struct perf_event_context *ctx;
4066 
4067         ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4068         if (!ctx)
4069                 return NULL;
4070 
4071         __perf_event_init_context(ctx);
4072         if (task) {
4073                 ctx->task = task;
4074                 get_task_struct(task);
4075         }
4076         ctx->pmu = pmu;
4077 
4078         return ctx;
4079 }
4080 
4081 static struct task_struct *
4082 find_lively_task_by_vpid(pid_t vpid)
4083 {
4084         struct task_struct *task;
4085 
4086         rcu_read_lock();
4087         if (!vpid)
4088                 task = current;
4089         else
4090                 task = find_task_by_vpid(vpid);
4091         if (task)
4092                 get_task_struct(task);
4093         rcu_read_unlock();
4094 
4095         if (!task)
4096                 return ERR_PTR(-ESRCH);
4097 
4098         return task;
4099 }
4100 
4101 /*
4102  * Returns a matching context with refcount and pincount.
4103  */
4104 static struct perf_event_context *
4105 find_get_context(struct pmu *pmu, struct task_struct *task,
4106                 struct perf_event *event)
4107 {
4108         struct perf_event_context *ctx, *clone_ctx = NULL;
4109         struct perf_cpu_context *cpuctx;
4110         void *task_ctx_data = NULL;
4111         unsigned long flags;
4112         int ctxn, err;
4113         int cpu = event->cpu;
4114 
4115         if (!task) {
4116                 /* Must be root to operate on a CPU event: */
4117                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
4118                         return ERR_PTR(-EACCES);
4119 
4120                 cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
4121                 ctx = &cpuctx->ctx;
4122                 get_ctx(ctx);
4123                 ++ctx->pin_count;
4124 
4125                 return ctx;
4126         }
4127 
4128         err = -EINVAL;
4129         ctxn = pmu->task_ctx_nr;
4130         if (ctxn < 0)
4131                 goto errout;
4132 
4133         if (event->attach_state & PERF_ATTACH_TASK_DATA) {
4134                 task_ctx_data = kzalloc(pmu->task_ctx_size, GFP_KERNEL);
4135                 if (!task_ctx_data) {
4136                         err = -ENOMEM;
4137                         goto errout;
4138                 }
4139         }
4140 
4141 retry:
4142         ctx = perf_lock_task_context(task, ctxn, &flags);
4143         if (ctx) {
4144                 clone_ctx = unclone_ctx(ctx);
4145                 ++ctx->pin_count;
4146 
4147                 if (task_ctx_data && !ctx->task_ctx_data) {
4148                         ctx->task_ctx_data = task_ctx_data;
4149                         task_ctx_data = NULL;
4150                 }
4151                 raw_spin_unlock_irqrestore(&ctx->lock, flags);
4152 
4153                 if (clone_ctx)
4154                         put_ctx(clone_ctx);
4155         } else {
4156                 ctx = alloc_perf_context(pmu, task);
4157                 err = -ENOMEM;
4158                 if (!ctx)
4159                         goto errout;
4160 
4161                 if (task_ctx_data) {
4162                         ctx->task_ctx_data = task_ctx_data;
4163                         task_ctx_data = NULL;
4164                 }
4165 
4166                 err = 0;
4167                 mutex_lock(&task->perf_event_mutex);
4168                 /*
4169                  * If it has already passed perf_event_exit_task().
4170                  * we must see PF_EXITING, it takes this mutex too.
4171                  */
4172                 if (task->flags & PF_EXITING)
4173                         err = -ESRCH;
4174                 else if (task->perf_event_ctxp[ctxn])
4175                         err = -EAGAIN;
4176                 else {
4177                         get_ctx(ctx);
4178                         ++ctx->pin_count;
4179                         rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
4180                 }
4181                 mutex_unlock(&task->perf_event_mutex);
4182 
4183                 if (unlikely(err)) {
4184                         put_ctx(ctx);
4185 
4186                         if (err == -EAGAIN)
4187                                 goto retry;
4188                         goto errout;
4189                 }
4190         }
4191 
4192         kfree(task_ctx_data);
4193         return ctx;
4194 
4195 errout:
4196         kfree(task_ctx_data);
4197         return ERR_PTR(err);
4198 }
4199 
4200 static void perf_event_free_filter(struct perf_event *event);
4201 static void perf_event_free_bpf_prog(struct perf_event *event);
4202 
4203 static void free_event_rcu(struct rcu_head *head)
4204 {
4205         struct perf_event *event;
4206 
4207         event = container_of(head, struct perf_event, rcu_head);
4208         if (event->ns)
4209                 put_pid_ns(event->ns);
4210         perf_event_free_filter(event);
4211         kfree(event);
4212 }
4213 
4214 static void ring_buffer_attach(struct perf_event *event,
4215                                struct ring_buffer *rb);
4216 
4217 static void detach_sb_event(struct perf_event *event)
4218 {
4219         struct pmu_event_list *pel = per_cpu_ptr(&pmu_sb_events, event->cpu);
4220 
4221         raw_spin_lock(&pel->lock);
4222         list_del_rcu(&event->sb_list);
4223         raw_spin_unlock(&pel->lock);
4224 }
4225 
4226 static bool is_sb_event(struct perf_event *event)
4227 {
4228         struct perf_event_attr *attr = &event->attr;
4229 
4230         if (event->parent)
4231                 return false;
4232 
4233         if (event->attach_state & PERF_ATTACH_TASK)
4234                 return false;
4235 
4236         if (attr->mmap || attr->mmap_data || attr->mmap2 ||
4237             attr->comm || attr->comm_exec ||
4238             attr->task ||
4239             attr->context_switch)
4240                 return true;
4241         return false;
4242 }
4243 
4244 static void unaccount_pmu_sb_event(struct perf_event *event)
4245 {
4246         if (is_sb_event(event))
4247                 detach_sb_event(event);
4248 }
4249 
4250 static void unaccount_event_cpu(struct perf_event *event, int cpu)
4251 {
4252         if (event->parent)
4253                 return;
4254 
4255         if (is_cgroup_event(event))
4256                 atomic_dec(&per_cpu(perf_cgroup_events, cpu));
4257 }
4258 
4259 #ifdef CONFIG_NO_HZ_FULL
4260 static DEFINE_SPINLOCK(nr_freq_lock);
4261 #endif
4262 
4263 static void unaccount_freq_event_nohz(void)
4264 {
4265 #ifdef CONFIG_NO_HZ_FULL
4266         spin_lock(&nr_freq_lock);
4267         if (atomic_dec_and_test(&nr_freq_events))
4268                 tick_nohz_dep_clear(TICK_DEP_BIT_PERF_EVENTS);
4269         spin_unlock(&nr_freq_lock);
4270 #endif
4271 }
4272 
4273 static void unaccount_freq_event(void)
4274 {
4275         if (tick_nohz_full_enabled())
4276                 unaccount_freq_event_nohz();
4277         else
4278                 atomic_dec(&nr_freq_events);
4279 }
4280 
4281 static void unaccount_event(struct perf_event *event)
4282 {
4283         bool dec = false;
4284 
4285         if (event->parent)
4286                 return;
4287 
4288         if (event->attach_state & PERF_ATTACH_TASK)
4289                 dec = true;
4290         if (event->attr.mmap || event->attr.mmap_data)
4291                 atomic_dec(&nr_mmap_events);
4292         if (event->attr.comm)
4293                 atomic_dec(&nr_comm_events);
4294         if (event->attr.namespaces)
4295                 atomic_dec(&nr_namespaces_events);
4296         if (event->attr.task)
4297                 atomic_dec(&nr_task_events);
4298         if (event->attr.freq)
4299                 unaccount_freq_event();
4300         if (event->attr.context_switch) {
4301                 dec = true;
4302                 atomic_dec(&nr_switch_events);
4303         }
4304         if (is_cgroup_event(event))
4305                 dec = true;
4306         if (has_branch_stack(event))
4307                 dec = true;
4308 
4309         if (dec) {
4310                 if (!atomic_add_unless(&perf_sched_count, -1, 1))
4311                         schedule_delayed_work(&perf_sched_work, HZ);
4312         }
4313 
4314         unaccount_event_cpu(event, event->cpu);
4315 
4316         unaccount_pmu_sb_event(event);
4317 }
4318 
4319 static void perf_sched_delayed(struct work_struct *work)
4320 {
4321         mutex_lock(&perf_sched_mutex);
4322         if (atomic_dec_and_test(&perf_sched_count))
4323                 static_branch_disable(&perf_sched_events);
4324         mutex_unlock(&perf_sched_mutex);
4325 }
4326 
4327 /*
4328  * The following implement mutual exclusion of events on "exclusive" pmus
4329  * (PERF_PMU_CAP_EXCLUSIVE). Such pmus can only have one event scheduled
4330  * at a time, so we disallow creating events that might conflict, namely:
4331  *
4332  *  1) cpu-wide events in the presence of per-task events,
4333  *  2) per-task events in the presence of cpu-wide events,
4334  *  3) two matching events on the same context.
4335  *
4336  * The former two cases are handled in the allocation path (perf_event_alloc(),
4337  * _free_event()), the latter -- before the first perf_install_in_context().
4338  */
4339 static int exclusive_event_init(struct perf_event *event)
4340 {
4341         struct pmu *pmu = event->pmu;
4342 
4343         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4344                 return 0;
4345 
4346         /*
4347          * Prevent co-existence of per-task and cpu-wide events on the
4348          * same exclusive pmu.
4349          *
4350          * Negative pmu::exclusive_cnt means there are cpu-wide
4351          * events on this "exclusive" pmu, positive means there are
4352          * per-task events.
4353          *
4354          * Since this is called in perf_event_alloc() path, event::ctx
4355          * doesn't exist yet; it is, however, safe to use PERF_ATTACH_TASK
4356          * to mean "per-task event", because unlike other attach states it
4357          * never gets cleared.
4358          */
4359         if (event->attach_state & PERF_ATTACH_TASK) {
4360                 if (!atomic_inc_unless_negative(&pmu->exclusive_cnt))
4361                         return -EBUSY;
4362         } else {
4363                 if (!atomic_dec_unless_positive(&pmu->exclusive_cnt))
4364                         return -EBUSY;
4365         }
4366 
4367         return 0;
4368 }
4369 
4370 static void exclusive_event_destroy(struct perf_event *event)
4371 {
4372         struct pmu *pmu = event->pmu;
4373 
4374         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4375                 return;
4376 
4377         /* see comment in exclusive_event_init() */
4378         if (event->attach_state & PERF_ATTACH_TASK)
4379                 atomic_dec(&pmu->exclusive_cnt);
4380         else
4381                 atomic_inc(&pmu->exclusive_cnt);
4382 }
4383 
4384 static bool exclusive_event_match(struct perf_event *e1, struct perf_event *e2)
4385 {
4386         if ((e1->pmu == e2->pmu) &&
4387             (e1->cpu == e2->cpu ||
4388              e1->cpu == -1 ||
4389              e2->cpu == -1))
4390                 return true;
4391         return false;
4392 }
4393 
4394 /* Called under the same ctx::mutex as perf_install_in_context() */
4395 static bool exclusive_event_installable(struct perf_event *event,
4396                                         struct perf_event_context *ctx)
4397 {
4398         struct perf_event *iter_event;
4399         struct pmu *pmu = event->pmu;
4400 
4401         if (!(pmu->capabilities & PERF_PMU_CAP_EXCLUSIVE))
4402                 return true;
4403 
4404         list_for_each_entry(iter_event, &ctx->event_list, event_entry) {
4405                 if (exclusive_event_match(iter_event, event))
4406                         return false;
4407         }
4408 
4409         return true;
4410 }
4411 
4412 static void perf_addr_filters_splice(struct perf_event *event,
4413                                        struct list_head *head);
4414 
4415 static void _free_event(struct perf_event *event)
4416 {
4417         irq_work_sync(&event->pending);
4418 
4419         unaccount_event(event);
4420 
4421         if (event->rb) {
4422                 /*
4423                  * Can happen when we close an event with re-directed output.
4424                  *
4425                  * Since we have a 0 refcount, perf_mmap_close() will skip
4426                  * over us; possibly making our ring_buffer_put() the last.
4427                  */
4428                 mutex_lock(&event->mmap_mutex);
4429                 ring_buffer_attach(event, NULL);
4430                 mutex_unlock(&event->mmap_mutex);
4431         }
4432 
4433         if (is_cgroup_event(event))
4434                 perf_detach_cgroup(event);
4435 
4436         if (!event->parent) {
4437                 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
4438                         put_callchain_buffers();
4439         }
4440 
4441         perf_event_free_bpf_prog(event);
4442         perf_addr_filters_splice(event, NULL);
4443         kfree(event->addr_filters_offs);
4444 
4445         if (event->destroy)
4446                 event->destroy(event);
4447 
4448         if (event->ctx)
4449                 put_ctx(event->ctx);
4450 
4451         if (event->hw.target)
4452                 put_task_struct(event->hw.target);
4453 
4454         exclusive_event_destroy(event);
4455         module_put(event->pmu->module);
4456 
4457         call_rcu(&event->rcu_head, free_event_rcu);
4458 }
4459 
4460 /*
4461  * Used to free events which have a known refcount of 1, such as in error paths
4462  * where the event isn't exposed yet and inherited events.
4463  */
4464 static void free_event(struct perf_event *event)
4465 {
4466         if (WARN(atomic_long_cmpxchg(&event->refcount, 1, 0) != 1,
4467                                 "unexpected event refcount: %ld; ptr=%p\n",
4468                                 atomic_long_read(&event->refcount), event)) {
4469                 /* leak to avoid use-after-free */
4470                 return;
4471         }
4472 
4473         _free_event(event);
4474 }
4475 
4476 /*
4477  * Remove user event from the owner task.
4478  */
4479 static void perf_remove_from_owner(struct perf_event *event)
4480 {
4481         struct task_struct *owner;
4482 
4483         rcu_read_lock();
4484         /*
4485          * Matches the smp_store_release() in perf_event_exit_task(). If we
4486          * observe !owner it means the list deletion is complete and we can
4487          * indeed free this event, otherwise we need to serialize on
4488          * owner->perf_event_mutex.
4489          */
4490         owner = READ_ONCE(event->owner);
4491         if (owner) {
4492                 /*
4493                  * Since delayed_put_task_struct() also drops the last
4494                  * task reference we can safely take a new reference
4495                  * while holding the rcu_read_lock().
4496                  */
4497                 get_task_struct(owner);
4498         }
4499         rcu_read_unlock();
4500 
4501         if (owner) {
4502                 /*
4503                  * If we're here through perf_event_exit_task() we're already
4504                  * holding ctx->mutex which would be an inversion wrt. the
4505                  * normal lock order.
4506                  *
4507                  * However we can safely take this lock because its the child
4508                  * ctx->mutex.
4509                  */
4510                 mutex_lock_nested(&owner->perf_event_mutex, SINGLE_DEPTH_NESTING);
4511 
4512                 /*
4513                  * We have to re-check the event->owner field, if it is cleared
4514                  * we raced with perf_event_exit_task(), acquiring the mutex
4515                  * ensured they're done, and we can proceed with freeing the
4516                  * event.
4517                  */
4518                 if (event->owner) {
4519                         list_del_init(&event->owner_entry);
4520                         smp_store_release(&event->owner, NULL);
4521                 }
4522                 mutex_unlock(&owner->perf_event_mutex);
4523                 put_task_struct(owner);
4524         }
4525 }
4526 
4527 static void put_event(struct perf_event *event)
4528 {
4529         if (!atomic_long_dec_and_test(&event->refcount))
4530                 return;
4531 
4532         _free_event(event);
4533 }
4534 
4535 /*
4536  * Kill an event dead; while event:refcount will preserve the event
4537  * object, it will not preserve its functionality. Once the last 'user'
4538  * gives up the object, we'll destroy the thing.
4539  */
4540 int perf_event_release_kernel(struct perf_event *event)
4541 {
4542         struct perf_event_context *ctx = event->ctx;
4543         struct perf_event *child, *tmp;
4544         LIST_HEAD(free_list);
4545 
4546         /*
4547          * If we got here through err_file: fput(event_file); we will not have
4548          * attached to a context yet.
4549          */
4550         if (!ctx) {
4551                 WARN_ON_ONCE(event->attach_state &
4552                                 (PERF_ATTACH_CONTEXT|PERF_ATTACH_GROUP));
4553                 goto no_ctx;
4554         }
4555 
4556         if (!is_kernel_event(event))
4557                 perf_remove_from_owner(event);
4558 
4559         ctx = perf_event_ctx_lock(event);
4560         WARN_ON_ONCE(ctx->parent_ctx);
4561         perf_remove_from_context(event, DETACH_GROUP);
4562 
4563         raw_spin_lock_irq(&ctx->lock);
4564         /*
4565          * Mark this event as STATE_DEAD, there is no external reference to it
4566          * anymore.
4567          *
4568          * Anybody acquiring event->child_mutex after the below loop _must_
4569          * also see this, most importantly inherit_event() which will avoid
4570          * placing more children on the list.
4571          *
4572          * Thus this guarantees that we will in fact observe and kill _ALL_
4573          * child events.
4574          */
4575         event->state = PERF_EVENT_STATE_DEAD;
4576         raw_spin_unlock_irq(&ctx->lock);
4577 
4578         perf_event_ctx_unlock(event, ctx);
4579 
4580 again:
4581         mutex_lock(&event->child_mutex);
4582         list_for_each_entry(child, &event->child_list, child_list) {
4583 
4584                 /*
4585                  * Cannot change, child events are not migrated, see the
4586                  * comment with perf_event_ctx_lock_nested().
4587                  */
4588                 ctx = READ_ONCE(child->ctx);
4589                 /*
4590                  * Since child_mutex nests inside ctx::mutex, we must jump
4591                  * through hoops. We start by grabbing a reference on the ctx.
4592                  *
4593                  * Since the event cannot get freed while we hold the
4594                  * child_mutex, the context must also exist and have a !0
4595                  * reference count.
4596                  */
4597                 get_ctx(ctx);
4598 
4599                 /*
4600                  * Now that we have a ctx ref, we can drop child_mutex, and
4601                  * acquire ctx::mutex without fear of it going away. Then we
4602                  * can re-acquire child_mutex.
4603                  */
4604                 mutex_unlock(&event->child_mutex);
4605                 mutex_lock(&ctx->mutex);
4606                 mutex_lock(&event->child_mutex);
4607 
4608                 /*
4609                  * Now that we hold ctx::mutex and child_mutex, revalidate our
4610                  * state, if child is still the first entry, it didn't get freed
4611                  * and we can continue doing so.
4612                  */
4613                 tmp = list_first_entry_or_null(&event->child_list,
4614                                                struct perf_event, child_list);
4615                 if (tmp == child) {
4616                         perf_remove_from_context(child, DETACH_GROUP);
4617                         list_move(&child->child_list, &free_list);
4618                         /*
4619                          * This matches the refcount bump in inherit_event();
4620                          * this can't be the last reference.
4621                          */
4622                         put_event(event);
4623                 }
4624 
4625                 mutex_unlock(&event->child_mutex);
4626                 mutex_unlock(&ctx->mutex);
4627                 put_ctx(ctx);
4628                 goto again;
4629         }
4630         mutex_unlock(&event->child_mutex);
4631 
4632         list_for_each_entry_safe(child, tmp, &free_list, child_list) {
4633                 list_del(&child->child_list);
4634                 free_event(child);
4635         }
4636 
4637 no_ctx:
4638         put_event(event); /* Must be the 'last' reference */
4639         return 0;
4640 }
4641 EXPORT_SYMBOL_GPL(perf_event_release_kernel);
4642 
4643 /*
4644  * Called when the last reference to the file is gone.
4645  */
4646 static int perf_release(struct inode *inode, struct file *file)
4647 {
4648         perf_event_release_kernel(file->private_data);
4649         return 0;
4650 }
4651 
4652 static u64 __perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4653 {
4654         struct perf_event *child;
4655         u64 total = 0;
4656 
4657         *enabled = 0;
4658         *running = 0;
4659 
4660         mutex_lock(&event->child_mutex);
4661 
4662         (void)perf_event_read(event, false);
4663         total += perf_event_count(event);
4664 
4665         *enabled += event->total_time_enabled +
4666                         atomic64_read(&event->child_total_time_enabled);
4667         *running += event->total_time_running +
4668                         atomic64_read(&event->child_total_time_running);
4669 
4670         list_for_each_entry(child, &event->child_list, child_list) {
4671                 (void)perf_event_read(child, false);
4672                 total += perf_event_count(child);
4673                 *enabled += child->total_time_enabled;
4674                 *running += child->total_time_running;
4675         }
4676         mutex_unlock(&event->child_mutex);
4677 
4678         return total;
4679 }
4680 
4681 u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
4682 {
4683         struct perf_event_context *ctx;
4684         u64 count;
4685 
4686         ctx = perf_event_ctx_lock(event);
4687         count = __perf_event_read_value(event, enabled, running);
4688         perf_event_ctx_unlock(event, ctx);
4689 
4690         return count;
4691 }
4692 EXPORT_SYMBOL_GPL(perf_event_read_value);
4693 
4694 static int __perf_read_group_add(struct perf_event *leader,
4695                                         u64 read_format, u64 *values)
4696 {
4697         struct perf_event_context *ctx = leader->ctx;
4698         struct perf_event *sub;
4699         unsigned long flags;
4700         int n = 1; /* skip @nr */
4701         int ret;
4702 
4703         ret = perf_event_read(leader, true);
4704         if (ret)
4705                 return ret;
4706 
4707         raw_spin_lock_irqsave(&ctx->lock, flags);
4708 
4709         /*
4710          * Since we co-schedule groups, {enabled,running} times of siblings
4711          * will be identical to those of the leader, so we only publish one
4712          * set.
4713          */
4714         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
4715                 values[n++] += leader->total_time_enabled +
4716                         atomic64_read(&leader->child_total_time_enabled);
4717         }
4718 
4719         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
4720                 values[n++] += leader->total_time_running +
4721                         atomic64_read(&leader->child_total_time_running);
4722         }
4723 
4724         /*
4725          * Write {count,id} tuples for every sibling.
4726          */
4727         values[n++] += perf_event_count(leader);
4728         if (read_format & PERF_FORMAT_ID)
4729                 values[n++] = primary_event_id(leader);
4730 
4731         for_each_sibling_event(sub, leader) {
4732                 values[n++] += perf_event_count(sub);
4733                 if (read_format & PERF_FORMAT_ID)
4734                         values[n++] = primary_event_id(sub);
4735         }
4736 
4737         raw_spin_unlock_irqrestore(&ctx->lock, flags);
4738         return 0;
4739 }
4740 
4741 static int perf_read_group(struct perf_event *event,
4742                                    u64 read_format, char __user *buf)
4743 {
4744         struct perf_event *leader = event->group_leader, *child;
4745         struct perf_event_context *ctx = leader->ctx;
4746         int ret;
4747         u64 *values;
4748 
4749         lockdep_assert_held(&ctx->mutex);
4750 
4751         values = kzalloc(event->read_size, GFP_KERNEL);
4752         if (!values)
4753                 return -ENOMEM;
4754 
4755         values[0] = 1 + leader->nr_siblings;
4756 
4757         /*
4758          * By locking the child_mutex of the leader we effectively
4759          * lock the child list of all siblings.. XXX explain how.
4760          */
4761         mutex_lock(&leader->child_mutex);
4762 
4763         ret = __perf_read_group_add(leader, read_format, values);
4764         if (ret)
4765                 goto unlock;
4766 
4767         list_for_each_entry(child, &leader->child_list, child_list) {
4768                 ret = __perf_read_group_add(child, read_format, values);
4769                 if (ret)
4770                         goto unlock;
4771         }
4772 
4773         mutex_unlock(&leader->child_mutex);
4774 
4775         ret = event->read_size;
4776         if (copy_to_user(buf, values, event->read_size))
4777                 ret = -EFAULT;
4778         goto out;
4779 
4780 unlock:
4781         mutex_unlock(&leader->child_mutex);
4782 out:
4783         kfree(values);
4784         return ret;
4785 }
4786 
4787 static int perf_read_one(struct perf_event *event,
4788                                  u64 read_format, char __user *buf)
4789 {
4790         u64 enabled, running;
4791         u64 values[4];
4792         int n = 0;
4793 
4794         values[n++] = __perf_event_read_value(event, &enabled, &running);
4795         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
4796                 values[n++] = enabled;
4797         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
4798                 values[n++] = running;
4799         if (read_format & PERF_FORMAT_ID)
4800                 values[n++] = primary_event_id(event);
4801 
4802         if (copy_to_user(buf, values, n * sizeof(u64)))
4803                 return -EFAULT;
4804 
4805         return n * sizeof(u64);
4806 }
4807 
4808 static bool is_event_hup(struct perf_event *event)
4809 {
4810         bool no_children;
4811 
4812         if (event->state > PERF_EVENT_STATE_EXIT)
4813                 return false;
4814 
4815         mutex_lock(&event->child_mutex);
4816         no_children = list_empty(&event->child_list);
4817         mutex_unlock(&event->child_mutex);
4818         return no_children;
4819 }
4820 
4821 /*
4822  * Read the performance event - simple non blocking version for now
4823  */
4824 static ssize_t
4825 __perf_read(struct perf_event *event, char __user *buf, size_t count)
4826 {
4827         u64 read_format = event->attr.read_format;
4828         int ret;
4829 
4830         /*
4831          * Return end-of-file for a read on an event that is in
4832          * error state (i.e. because it was pinned but it couldn't be
4833          * scheduled on to the CPU at some point).
4834          */
4835         if (event->state == PERF_EVENT_STATE_ERROR)
4836                 return 0;
4837 
4838         if (count < event->read_size)
4839                 return -ENOSPC;
4840 
4841         WARN_ON_ONCE(event->ctx->parent_ctx);
4842         if (read_format & PERF_FORMAT_GROUP)
4843                 ret = perf_read_group(event, read_format, buf);
4844         else
4845                 ret = perf_read_one(event, read_format, buf);
4846 
4847         return ret;
4848 }
4849 
4850 static ssize_t
4851 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
4852 {
4853         struct perf_event *event = file->private_data;
4854         struct perf_event_context *ctx;
4855         int ret;
4856 
4857         ctx = perf_event_ctx_lock(event);
4858         ret = __perf_read(event, buf, count);
4859         perf_event_ctx_unlock(event, ctx);
4860 
4861         return ret;
4862 }
4863 
4864 static __poll_t perf_poll(struct file *file, poll_table *wait)
4865 {
4866         struct perf_event *event = file->private_data;
4867         struct ring_buffer *rb;
4868         __poll_t events = EPOLLHUP;
4869 
4870         poll_wait(file, &event->waitq, wait);
4871 
4872         if (is_event_hup(event))
4873                 return events;
4874 
4875         /*
4876          * Pin the event->rb by taking event->mmap_mutex; otherwise
4877          * perf_event_set_output() can swizzle our rb and make us miss wakeups.
4878          */
4879         mutex_lock(&event->mmap_mutex);
4880         rb = event->rb;
4881         if (rb)
4882                 events = atomic_xchg(&rb->poll, 0);
4883         mutex_unlock(&event->mmap_mutex);
4884         return events;
4885 }
4886 
4887 static void _perf_event_reset(struct perf_event *event)
4888 {
4889         (void)perf_event_read(event, false);
4890         local64_set(&event->count, 0);
4891         perf_event_update_userpage(event);
4892 }
4893 
4894 /*
4895  * Holding the top-level event's child_mutex means that any
4896  * descendant process that has inherited this event will block
4897  * in perf_event_exit_event() if it goes to exit, thus satisfying the
4898  * task existence requirements of perf_event_enable/disable.
4899  */
4900 static void perf_event_for_each_child(struct perf_event *event,
4901                                         void (*func)(struct perf_event *))
4902 {
4903         struct perf_event *child;
4904 
4905         WARN_ON_ONCE(event->ctx->parent_ctx);
4906 
4907         mutex_lock(&event->child_mutex);
4908         func(event);
4909         list_for_each_entry(child, &event->child_list, child_list)
4910                 func(child);
4911         mutex_unlock(&event->child_mutex);
4912 }
4913 
4914 static void perf_event_for_each(struct perf_event *event,
4915                                   void (*func)(struct perf_event *))
4916 {
4917         struct perf_event_context *ctx = event->ctx;
4918         struct perf_event *sibling;
4919 
4920         lockdep_assert_held(&ctx->mutex);
4921 
4922         event = event->group_leader;
4923 
4924         perf_event_for_each_child(event, func);
4925         for_each_sibling_event(sibling, event)
4926                 perf_event_for_each_child(sibling, func);
4927 }
4928 
4929 static void __perf_event_period(struct perf_event *event,
4930                                 struct perf_cpu_context *cpuctx,
4931                                 struct perf_event_context *ctx,
4932                                 void *info)
4933 {
4934         u64 value = *((u64 *)info);
4935         bool active;
4936 
4937         if (event->attr.freq) {
4938                 event->attr.sample_freq = value;
4939         } else {
4940                 event->attr.sample_period = value;
4941                 event->hw.sample_period = value;
4942         }
4943 
4944         active = (event->state == PERF_EVENT_STATE_ACTIVE);
4945         if (active) {
4946                 perf_pmu_disable(ctx->pmu);
4947                 /*
4948                  * We could be throttled; unthrottle now to avoid the tick
4949                  * trying to unthrottle while we already re-started the event.
4950                  */
4951                 if (event->hw.interrupts == MAX_INTERRUPTS) {
4952                         event->hw.interrupts = 0;
4953                         perf_log_throttle(event, 1);
4954                 }
4955                 event->pmu->stop(event, PERF_EF_UPDATE);
4956         }
4957 
4958         local64_set(&event->hw.period_left, 0);
4959 
4960         if (active) {
4961                 event->pmu->start(event, PERF_EF_RELOAD);
4962                 perf_pmu_enable(ctx->pmu);
4963         }
4964 }
4965 
4966 static int perf_event_period(struct perf_event *event, u64 __user *arg)
4967 {
4968         u64 value;
4969 
4970         if (!is_sampling_event(event))
4971                 return -EINVAL;
4972 
4973         if (copy_from_user(&value, arg, sizeof(value)))
4974                 return -EFAULT;
4975 
4976         if (!value)
4977                 return -EINVAL;
4978 
4979         if (event->attr.freq && value > sysctl_perf_event_sample_rate)
4980                 return -EINVAL;
4981 
4982         event_function_call(event, __perf_event_period, &value);
4983 
4984         return 0;
4985 }
4986 
4987 static const struct file_operations perf_fops;
4988 
4989 static inline int perf_fget_light(int fd, struct fd *p)
4990 {
4991         struct fd f = fdget(fd);
4992         if (!f.file)
4993                 return -EBADF;
4994 
4995         if (f.file->f_op != &perf_fops) {
4996                 fdput(f);
4997                 return -EBADF;
4998         }
4999         *p = f;
5000         return 0;
5001 }
5002 
5003 static int perf_event_set_output(struct perf_event *event,
5004                                  struct perf_event *output_event);
5005 static int perf_event_set_filter(struct perf_event *event, void __user *arg);
5006 static int perf_event_set_bpf_prog(struct perf_event *event, u32 prog_fd);
5007 static int perf_copy_attr(struct perf_event_attr __user *uattr,
5008                           struct perf_event_attr *attr);
5009 
5010 static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned long arg)
5011 {
5012         void (*func)(struct perf_event *);
5013         u32 flags = arg;
5014 
5015         switch (cmd) {
5016         case PERF_EVENT_IOC_ENABLE:
5017                 func = _perf_event_enable;
5018                 break;
5019         case PERF_EVENT_IOC_DISABLE:
5020                 func = _perf_event_disable;
5021                 break;
5022         case PERF_EVENT_IOC_RESET:
5023                 func = _perf_event_reset;
5024                 break;
5025 
5026         case PERF_EVENT_IOC_REFRESH:
5027                 return _perf_event_refresh(event, arg);
5028 
5029         case PERF_EVENT_IOC_PERIOD:
5030                 return perf_event_period(event, (u64 __user *)arg);
5031 
5032         case PERF_EVENT_IOC_ID:
5033         {
5034                 u64 id = primary_event_id(event);
5035 
5036                 if (copy_to_user((void __user *)arg, &id, sizeof(id)))
5037                         return -EFAULT;
5038                 return 0;
5039         }
5040 
5041         case PERF_EVENT_IOC_SET_OUTPUT:
5042         {
5043                 int ret;
5044                 if (arg != -1) {
5045                         struct perf_event *output_event;
5046                         struct fd output;
5047                         ret = perf_fget_light(arg, &output);
5048                         if (ret)
5049                                 return ret;
5050                         output_event = output.file->private_data;
5051                         ret = perf_event_set_output(event, output_event);
5052                         fdput(output);
5053                 } else {
5054                         ret = perf_event_set_output(event, NULL);
5055                 }
5056                 return ret;
5057         }
5058 
5059         case PERF_EVENT_IOC_SET_FILTER:
5060                 return perf_event_set_filter(event, (void __user *)arg);
5061 
5062         case PERF_EVENT_IOC_SET_BPF:
5063                 return perf_event_set_bpf_prog(event, arg);
5064 
5065         case PERF_EVENT_IOC_PAUSE_OUTPUT: {
5066                 struct ring_buffer *rb;
5067 
5068                 rcu_read_lock();
5069                 rb = rcu_dereference(event->rb);
5070                 if (!rb || !rb->nr_pages) {
5071                         rcu_read_unlock();
5072                         return -EINVAL;
5073                 }
5074                 rb_toggle_paused(rb, !!arg);
5075                 rcu_read_unlock();
5076                 return 0;
5077         }
5078 
5079         case PERF_EVENT_IOC_QUERY_BPF:
5080                 return perf_event_query_prog_array(event, (void __user *)arg);
5081 
5082         case PERF_EVENT_IOC_MODIFY_ATTRIBUTES: {
5083                 struct perf_event_attr new_attr;
5084                 int err = perf_copy_attr((struct perf_event_attr __user *)arg,
5085                                          &new_attr);
5086 
5087                 if (err)
5088                         return err;
5089 
5090                 return perf_event_modify_attr(event,  &new_attr);
5091         }
5092         default:
5093                 return -ENOTTY;
5094         }
5095 
5096         if (flags & PERF_IOC_FLAG_GROUP)
5097                 perf_event_for_each(event, func);
5098         else
5099                 perf_event_for_each_child(event, func);
5100 
5101         return 0;
5102 }
5103 
5104 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
5105 {
5106         struct perf_event *event = file->private_data;
5107         struct perf_event_context *ctx;
5108         long ret;
5109 
5110         ctx = perf_event_ctx_lock(event);
5111         ret = _perf_ioctl(event, cmd, arg);
5112         perf_event_ctx_unlock(event, ctx);
5113 
5114         return ret;
5115 }
5116 
5117 #ifdef CONFIG_COMPAT
5118 static long perf_compat_ioctl(struct file *file, unsigned int cmd,
5119                                 unsigned long arg)
5120 {
5121         switch (_IOC_NR(cmd)) {
5122         case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
5123         case _IOC_NR(PERF_EVENT_IOC_ID):
5124         case _IOC_NR(PERF_EVENT_IOC_QUERY_BPF):
5125         case _IOC_NR(PERF_EVENT_IOC_MODIFY_ATTRIBUTES):
5126                 /* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
5127                 if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
5128                         cmd &= ~IOCSIZE_MASK;
5129                         cmd |= sizeof(void *) << IOCSIZE_SHIFT;
5130                 }
5131                 break;
5132         }
5133         return perf_ioctl(file, cmd, arg);
5134 }
5135 #else
5136 # define perf_compat_ioctl NULL
5137 #endif
5138 
5139 int perf_event_task_enable(void)
5140 {
5141         struct perf_event_context *ctx;
5142         struct perf_event *event;
5143 
5144         mutex_lock(&current->perf_event_mutex);
5145         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5146                 ctx = perf_event_ctx_lock(event);
5147                 perf_event_for_each_child(event, _perf_event_enable);
5148                 perf_event_ctx_unlock(event, ctx);
5149         }
5150         mutex_unlock(&current->perf_event_mutex);
5151 
5152         return 0;
5153 }
5154 
5155 int perf_event_task_disable(void)
5156 {
5157         struct perf_event_context *ctx;
5158         struct perf_event *event;
5159 
5160         mutex_lock(&current->perf_event_mutex);
5161         list_for_each_entry(event, &current->perf_event_list, owner_entry) {
5162                 ctx = perf_event_ctx_lock(event);
5163                 perf_event_for_each_child(event, _perf_event_disable);
5164                 perf_event_ctx_unlock(event, ctx);
5165         }
5166         mutex_unlock(&current->perf_event_mutex);
5167 
5168         return 0;
5169 }
5170 
5171 static int perf_event_index(struct perf_event *event)
5172 {
5173         if (event->hw.state & PERF_HES_STOPPED)
5174                 return 0;
5175 
5176         if (event->state != PERF_EVENT_STATE_ACTIVE)
5177                 return 0;
5178 
5179         return event->pmu->event_idx(event);
5180 }
5181 
5182 static void calc_timer_values(struct perf_event *event,
5183                                 u64 *now,
5184                                 u64 *enabled,
5185                                 u64 *running)
5186 {
5187         u64 ctx_time;
5188 
5189         *now = perf_clock();
5190         ctx_time = event->shadow_ctx_time + *now;
5191         __perf_update_times(event, ctx_time, enabled, running);
5192 }
5193 
5194 static void perf_event_init_userpage(struct perf_event *event)
5195 {
5196         struct perf_event_mmap_page *userpg;
5197         struct ring_buffer *rb;
5198 
5199         rcu_read_lock();
5200         rb = rcu_dereference(event->rb);
5201         if (!rb)
5202                 goto unlock;
5203 
5204         userpg = rb->user_page;
5205 
5206         /* Allow new userspace to detect that bit 0 is deprecated */
5207         userpg->cap_bit0_is_deprecated = 1;
5208         userpg->size = offsetof(struct perf_event_mmap_page, __reserved);
5209         userpg->data_offset = PAGE_SIZE;
5210         userpg->data_size = perf_data_size(rb);
5211 
5212 unlock:
5213         rcu_read_unlock();
5214 }
5215 
5216 void __weak arch_perf_update_userpage(
5217         struct perf_event *event, struct perf_event_mmap_page *userpg, u64 now)
5218 {
5219 }
5220 
5221 /*
5222  * Callers need to ensure there can be no nesting of this function, otherwise
5223  * the seqlock logic goes bad. We can not serialize this because the arch
5224  * code calls this from NMI context.
5225  */
5226 void perf_event_update_userpage(struct perf_event *event)
5227 {
5228         struct perf_event_mmap_page *userpg;
5229         struct ring_buffer *rb;
5230         u64 enabled, running, now;
5231 
5232         rcu_read_lock();
5233         rb = rcu_dereference(event->rb);
5234         if (!rb)
5235                 goto unlock;
5236 
5237         /*
5238          * compute total_time_enabled, total_time_running
5239          * based on snapshot values taken when the event
5240          * was last scheduled in.
5241          *
5242          * we cannot simply called update_context_time()
5243          * because of locking issue as we can be called in
5244          * NMI context
5245          */
5246         calc_timer_values(event, &now, &enabled, &running);
5247 
5248         userpg = rb->user_page;
5249         /*
5250          * Disable preemption to guarantee consistent time stamps are stored to
5251          * the user page.
5252          */
5253         preempt_disable();
5254         ++userpg->lock;
5255         barrier();
5256         userpg->index = perf_event_index(event);
5257         userpg->offset = perf_event_count(event);
5258         if (userpg->index)
5259                 userpg->offset -= local64_read(&event->hw.prev_count);
5260 
5261         userpg->time_enabled = enabled +
5262                         atomic64_read(&event->child_total_time_enabled);
5263 
5264         userpg->time_running = running +
5265                         atomic64_read(&event->child_total_time_running);
5266 
5267         arch_perf_update_userpage(event, userpg, now);
5268 
5269         barrier();
5270         ++userpg->lock;
5271         preempt_enable();
5272 unlock:
5273         rcu_read_unlock();
5274 }
5275 EXPORT_SYMBOL_GPL(perf_event_update_userpage);
5276 
5277 static vm_fault_t perf_mmap_fault(struct vm_fault *vmf)
5278 {
5279         struct perf_event *event = vmf->vma->vm_file->private_data;
5280         struct ring_buffer *rb;
5281         vm_fault_t ret = VM_FAULT_SIGBUS;
5282 
5283         if (vmf->flags & FAULT_FLAG_MKWRITE) {
5284                 if (vmf->pgoff == 0)
5285                         ret = 0;
5286                 return ret;
5287         }
5288 
5289         rcu_read_lock();
5290         rb = rcu_dereference(event->rb);
5291         if (!rb)
5292                 goto unlock;
5293 
5294         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
5295                 goto unlock;
5296 
5297         vmf->page = perf_mmap_to_page(rb, vmf->pgoff);
5298         if (!vmf->page)
5299                 goto unlock;
5300 
5301         get_page(vmf->page);
5302         vmf->page->mapping = vmf->vma->vm_file->f_mapping;
5303         vmf->page->index   = vmf->pgoff;
5304 
5305         ret = 0;
5306 unlock:
5307         rcu_read_unlock();
5308 
5309         return ret;
5310 }
5311 
5312 static void ring_buffer_attach(struct perf_event *event,
5313                                struct ring_buffer *rb)
5314 {
5315         struct ring_buffer *old_rb = NULL;
5316         unsigned long flags;
5317 
5318         if (event->rb) {
5319                 /*
5320                  * Should be impossible, we set this when removing
5321                  * event->rb_entry and wait/clear when adding event->rb_entry.
5322                  */
5323                 WARN_ON_ONCE(event->rcu_pending);
5324 
5325                 old_rb = event->rb;
5326                 spin_lock_irqsave(&old_rb->event_lock, flags);
5327                 list_del_rcu(&event->rb_entry);
5328                 spin_unlock_irqrestore(&old_rb->event_lock, flags);
5329 
5330                 event->rcu_batches = get_state_synchronize_rcu();
5331                 event->rcu_pending = 1;
5332         }
5333 
5334         if (rb) {
5335                 if (event->rcu_pending) {
5336                         cond_synchronize_rcu(event->rcu_batches);
5337                         event->rcu_pending = 0;
5338                 }
5339 
5340                 spin_lock_irqsave(&rb->event_lock, flags);
5341                 list_add_rcu(&event->rb_entry, &rb->event_list);
5342                 spin_unlock_irqrestore(&rb->event_lock, flags);
5343         }
5344 
5345         /*
5346          * Avoid racing with perf_mmap_close(AUX): stop the event
5347          * before swizzling the event::rb pointer; if it's getting
5348          * unmapped, its aux_mmap_count will be 0 and it won't
5349          * restart. See the comment in __perf_pmu_output_stop().
5350          *
5351          * Data will inevitably be lost when set_output is done in
5352          * mid-air, but then again, whoever does it like this is
5353          * not in for the data anyway.
5354          */
5355         if (has_aux(event))
5356                 perf_event_stop(event, 0);
5357 
5358         rcu_assign_pointer(event->rb, rb);
5359 
5360         if (old_rb) {
5361                 ring_buffer_put(old_rb);
5362                 /*
5363                  * Since we detached before setting the new rb, so that we
5364                  * could attach the new rb, we could have missed a wakeup.
5365                  * Provide it now.
5366                  */
5367                 wake_up_all(&event->waitq);
5368         }
5369 }
5370 
5371 static void ring_buffer_wakeup(struct perf_event *event)
5372 {
5373         struct ring_buffer *rb;
5374 
5375         rcu_read_lock();
5376         rb = rcu_dereference(event->rb);
5377         if (rb) {
5378                 list_for_each_entry_rcu(event, &rb->event_list, rb_entry)
5379                         wake_up_all(&event->waitq);
5380         }
5381         rcu_read_unlock();
5382 }
5383 
5384 struct ring_buffer *ring_buffer_get(struct perf_event *event)
5385 {
5386         struct ring_buffer *rb;
5387 
5388         rcu_read_lock();
5389         rb = rcu_dereference(event->rb);
5390         if (rb) {
5391                 if (!atomic_inc_not_zero(&rb->refcount))
5392                         rb = NULL;
5393         }
5394         rcu_read_unlock();
5395 
5396         return rb;
5397 }
5398 
5399 void ring_buffer_put(struct ring_buffer *rb)
5400 {
5401         if (!atomic_dec_and_test(&rb->refcount))
5402                 return;
5403 
5404         WARN_ON_ONCE(!list_empty(&rb->event_list));
5405 
5406         call_rcu(&rb->rcu_head, rb_free_rcu);
5407 }
5408 
5409 static void perf_mmap_open(struct vm_area_struct *vma)
5410 {
5411         struct perf_event *event = vma->vm_file->private_data;
5412 
5413         atomic_inc(&event->mmap_count);
5414         atomic_inc(&event->rb->mmap_count);
5415 
5416         if (vma->vm_pgoff)
5417                 atomic_inc(&event->rb->aux_mmap_count);
5418 
5419         if (event->pmu->event_mapped)
5420                 event->pmu->event_mapped(event, vma->vm_mm);
5421 }
5422 
5423 static void perf_pmu_output_stop(struct perf_event *event);
5424 
5425 /*
5426  * A buffer can be mmap()ed multiple times; either directly through the same
5427  * event, or through other events by use of perf_event_set_output().
5428  *
5429  * In order to undo the VM accounting done by perf_mmap() we need to destroy
5430  * the buffer here, where we still have a VM context. This means we need
5431  * to detach all events redirecting to us.
5432  */
5433 static void perf_mmap_close(struct vm_area_struct *vma)
5434 {
5435         struct perf_event *event = vma->vm_file->private_data;
5436 
5437         struct ring_buffer *rb = ring_buffer_get(event);
5438         struct user_struct *mmap_user = rb->mmap_user;
5439         int mmap_locked = rb->mmap_locked;
5440         unsigned long size = perf_data_size(rb);
5441 
5442         if (event->pmu->event_unmapped)
5443                 event->pmu->event_unmapped(event, vma->vm_mm);
5444 
5445         /*
5446          * rb->aux_mmap_count will always drop before rb->mmap_count and
5447          * event->mmap_count, so it is ok to use event->mmap_mutex to
5448          * serialize with perf_mmap here.
5449          */
5450         if (rb_has_aux(rb) && vma->vm_pgoff == rb->aux_pgoff &&
5451             atomic_dec_and_mutex_lock(&rb->aux_mmap_count, &event->mmap_mutex)) {
5452                 /*
5453                  * Stop all AUX events that are writing to this buffer,
5454                  * so that we can free its AUX pages and corresponding PMU
5455                  * data. Note that after rb::aux_mmap_count dropped to zero,
5456                  * they won't start any more (see perf_aux_output_begin()).
5457                  */
5458                 perf_pmu_output_stop(event);
5459 
5460                 /* now it's safe to free the pages */
5461                 atomic_long_sub(rb->aux_nr_pages, &mmap_user->locked_vm);
5462                 vma->vm_mm->pinned_vm -= rb->aux_mmap_locked;
5463 
5464                 /* this has to be the last one */
5465                 rb_free_aux(rb);
5466                 WARN_ON_ONCE(atomic_read(&rb->aux_refcount));
5467 
5468                 mutex_unlock(&event->mmap_mutex);
5469         }
5470 
5471         atomic_dec(&rb->mmap_count);
5472 
5473         if (!atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
5474                 goto out_put;
5475 
5476         ring_buffer_attach(event, NULL);
5477         mutex_unlock(&event->mmap_mutex);
5478 
5479         /* If there's still other mmap()s of this buffer, we're done. */
5480         if (atomic_read(&rb->mmap_count))
5481                 goto out_put;
5482 
5483         /*
5484          * No other mmap()s, detach from all other events that might redirect
5485          * into the now unreachable buffer. Somewhat complicated by the
5486          * fact that rb::event_lock otherwise nests inside mmap_mutex.
5487          */
5488 again:
5489         rcu_read_lock();
5490         list_for_each_entry_rcu(event, &rb->event_list, rb_entry) {
5491                 if (!atomic_long_inc_not_zero(&event->refcount)) {
5492                         /*
5493                          * This event is en-route to free_event() which will
5494                          * detach it and remove it from the list.
5495                          */
5496                         continue;
5497                 }
5498                 rcu_read_unlock();
5499 
5500                 mutex_lock(&event->mmap_mutex);
5501                 /*
5502                  * Check we didn't race with perf_event_set_output() which can
5503                  * swizzle the rb from under us while we were waiting to
5504                  * acquire mmap_mutex.
5505                  *
5506                  * If we find a different rb; ignore this event, a next
5507                  * iteration will no longer find it on the list. We have to
5508                  * still restart the iteration to make sure we're not now
5509                  * iterating the wrong list.
5510                  */
5511                 if (event->rb == rb)
5512                         ring_buffer_attach(event, NULL);
5513 
5514                 mutex_unlock(&event->mmap_mutex);
5515                 put_event(event);
5516 
5517                 /*
5518                  * Restart the iteration; either we're on the wrong list or
5519                  * destroyed its integrity by doing a deletion.
5520                  */
5521                 goto again;
5522         }
5523         rcu_read_unlock();
5524 
5525         /*
5526          * It could be there's still a few 0-ref events on the list; they'll
5527          * get cleaned up by free_event() -- they'll also still have their
5528          * ref on the rb and will free it whenever they are done with it.
5529          *
5530          * Aside from that, this buffer is 'fully' detached and unmapped,
5531          * undo the VM accounting.
5532          */
5533 
5534         atomic_long_sub((size >> PAGE_SHIFT) + 1, &mmap_user->locked_vm);
5535         vma->vm_mm->pinned_vm -= mmap_locked;
5536         free_uid(mmap_user);
5537 
5538 out_put:
5539         ring_buffer_put(rb); /* could be last */
5540 }
5541 
5542 static const struct vm_operations_struct perf_mmap_vmops = {
5543         .open           = perf_mmap_open,
5544         .close          = perf_mmap_close, /* non mergable */
5545         .fault          = perf_mmap_fault,
5546         .page_mkwrite   = perf_mmap_fault,
5547 };
5548 
5549 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
5550 {
5551         struct perf_event *event = file->private_data;
5552         unsigned long user_locked, user_lock_limit;
5553         struct user_struct *user = current_user();
5554         unsigned long locked, lock_limit;
5555         struct ring_buffer *rb = NULL;
5556         unsigned long vma_size;
5557         unsigned long nr_pages;
5558         long user_extra = 0, extra = 0;
5559         int ret = 0, flags = 0;
5560 
5561         /*
5562          * Don't allow mmap() of inherited per-task counters. This would
5563          * create a performance issue due to all children writing to the
5564          * same rb.
5565          */
5566         if (event->cpu == -1 && event->attr.inherit)
5567                 return -EINVAL;
5568 
5569         if (!(vma->vm_flags & VM_SHARED))
5570                 return -EINVAL;
5571 
5572         vma_size = vma->vm_end - vma->vm_start;
5573 
5574         if (vma->vm_pgoff == 0) {
5575                 nr_pages = (vma_size / PAGE_SIZE) - 1;
5576         } else {
5577                 /*
5578                  * AUX area mapping: if rb->aux_nr_pages != 0, it's already
5579                  * mapped, all subsequent mappings should have the same size
5580                  * and offset. Must be above the normal perf buffer.
5581                  */
5582                 u64 aux_offset, aux_size;
5583 
5584                 if (!event->rb)
5585                         return -EINVAL;
5586 
5587                 nr_pages = vma_size / PAGE_SIZE;
5588 
5589                 mutex_lock(&event->mmap_mutex);
5590                 ret = -EINVAL;
5591 
5592                 rb = event->rb;
5593                 if (!rb)
5594                         goto aux_unlock;
5595 
5596                 aux_offset = READ_ONCE(rb->user_page->aux_offset);
5597                 aux_size = READ_ONCE(rb->user_page->aux_size);
5598 
5599                 if (aux_offset < perf_data_size(rb) + PAGE_SIZE)
5600                         goto aux_unlock;
5601 
5602                 if (aux_offset != vma->vm_pgoff << PAGE_SHIFT)
5603                         goto aux_unlock;
5604 
5605                 /* already mapped with a different offset */
5606                 if (rb_has_aux(rb) && rb->aux_pgoff != vma->vm_pgoff)
5607                         goto aux_unlock;
5608 
5609                 if (aux_size != vma_size || aux_size != nr_pages * PAGE_SIZE)
5610                         goto aux_unlock;
5611 
5612                 /* already mapped with a different size */
5613                 if (rb_has_aux(rb) && rb->aux_nr_pages != nr_pages)
5614                         goto aux_unlock;
5615 
5616                 if (!is_power_of_2(nr_pages))
5617                         goto aux_unlock;
5618 
5619                 if (!atomic_inc_not_zero(&rb->mmap_count))
5620                         goto aux_unlock;
5621 
5622                 if (rb_has_aux(rb)) {
5623                         atomic_inc(&rb->aux_mmap_count);
5624                         ret = 0;
5625                         goto unlock;
5626                 }
5627 
5628                 atomic_set(&rb->aux_mmap_count, 1);
5629                 user_extra = nr_pages;
5630 
5631                 goto accounting;
5632         }
5633 
5634         /*
5635          * If we have rb pages ensure they're a power-of-two number, so we
5636          * can do bitmasks instead of modulo.
5637          */
5638         if (nr_pages != 0 && !is_power_of_2(nr_pages))
5639                 return -EINVAL;
5640 
5641         if (vma_size != PAGE_SIZE * (1 + nr_pages))
5642                 return -EINVAL;
5643 
5644         WARN_ON_ONCE(event->ctx->parent_ctx);
5645 again:
5646         mutex_lock(&event->mmap_mutex);
5647         if (event->rb) {
5648                 if (event->rb->nr_pages != nr_pages) {
5649                         ret = -EINVAL;
5650                         goto unlock;
5651                 }
5652 
5653                 if (!atomic_inc_not_zero(&event->rb->mmap_count)) {
5654                         /*
5655                          * Raced against perf_mmap_close() through
5656                          * perf_event_set_output(). Try again, hope for better
5657                          * luck.
5658                          */
5659                         mutex_unlock(&event->mmap_mutex);
5660                         goto again;
5661                 }
5662 
5663                 goto unlock;
5664         }
5665 
5666         user_extra = nr_pages + 1;
5667 
5668 accounting:
5669         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
5670 
5671         /*
5672          * Increase the limit linearly with more CPUs:
5673          */
5674         user_lock_limit *= num_online_cpus();
5675 
5676         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
5677 
5678         if (user_locked > user_lock_limit)
5679                 extra = user_locked - user_lock_limit;
5680 
5681         lock_limit = rlimit(RLIMIT_MEMLOCK);
5682         lock_limit >>= PAGE_SHIFT;
5683         locked = vma->vm_mm->pinned_vm + extra;
5684 
5685         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
5686                 !capable(CAP_IPC_LOCK)) {
5687                 ret = -EPERM;
5688                 goto unlock;
5689         }
5690 
5691         WARN_ON(!rb && event->rb);
5692 
5693         if (vma->vm_flags & VM_WRITE)
5694                 flags |= RING_BUFFER_WRITABLE;
5695 
5696         if (!rb) {
5697                 rb = rb_alloc(nr_pages,
5698                               event->attr.watermark ? event->attr.wakeup_watermark : 0,
5699                               event->cpu, flags);
5700 
5701                 if (!rb) {
5702                         ret = -ENOMEM;
5703                         goto unlock;
5704                 }
5705 
5706                 atomic_set(&rb->mmap_count, 1);
5707                 rb->mmap_user = get_current_user();
5708                 rb->mmap_locked = extra;
5709 
5710                 ring_buffer_attach(event, rb);
5711 
5712                 perf_event_init_userpage(event);
5713                 perf_event_update_userpage(event);
5714         } else {
5715                 ret = rb_alloc_aux(rb, event, vma->vm_pgoff, nr_pages,
5716                                    event->attr.aux_watermark, flags);
5717                 if (!ret)
5718                         rb->aux_mmap_locked = extra;
5719         }
5720 
5721 unlock:
5722         if (!ret) {
5723                 atomic_long_add(user_extra, &user->locked_vm);
5724                 vma->vm_mm->pinned_vm += extra;
5725 
5726                 atomic_inc(&event->mmap_count);
5727         } else if (rb) {
5728                 atomic_dec(&rb->mmap_count);
5729         }
5730 aux_unlock:
5731         mutex_unlock(&event->mmap_mutex);
5732 
5733         /*
5734          * Since pinned accounting is per vm we cannot allow fork() to copy our
5735          * vma.
5736          */
5737         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP;
5738         vma->vm_ops = &perf_mmap_vmops;
5739 
5740         if (event->pmu->event_mapped)
5741                 event->pmu->event_mapped(event, vma->vm_mm);
5742 
5743         return ret;
5744 }
5745 
5746 static int perf_fasync(int fd, struct file *filp, int on)
5747 {
5748         struct inode *inode = file_inode(filp);
5749         struct perf_event *event = filp->private_data;
5750         int retval;
5751 
5752         inode_lock(inode);
5753         retval = fasync_helper(fd, filp, on, &event->fasync);
5754         inode_unlock(inode);
5755 
5756         if (retval < 0)
5757                 return retval;
5758 
5759         return 0;
5760 }
5761 
5762 static const struct file_operations perf_fops = {
5763         .llseek                 = no_llseek,
5764         .release                = perf_release,
5765         .read                   = perf_read,
5766         .poll                   = perf_poll,
5767         .unlocked_ioctl         = perf_ioctl,
5768         .compat_ioctl           = perf_compat_ioctl,
5769         .mmap                   = perf_mmap,
5770         .fasync                 = perf_fasync,
5771 };
5772 
5773 /*
5774  * Perf event wakeup
5775  *
5776  * If there's data, ensure we set the poll() state and publish everything
5777  * to user-space before waking everybody up.
5778  */
5779 
5780 static inline struct fasync_struct **perf_event_fasync(struct perf_event *event)
5781 {
5782         /* only the parent has fasync state */
5783         if (event->parent)
5784                 event = event->parent;
5785         return &event->fasync;
5786 }
5787 
5788 void perf_event_wakeup(struct perf_event *event)
5789 {
5790         ring_buffer_wakeup(event);
5791 
5792         if (event->pending_kill) {
5793                 kill_fasync(perf_event_fasync(event), SIGIO, event->pending_kill);
5794                 event->pending_kill = 0;
5795         }
5796 }
5797 
5798 static void perf_pending_event(struct irq_work *entry)
5799 {
5800         struct perf_event *event = container_of(entry,
5801                         struct perf_event, pending);
5802         int rctx;
5803 
5804         rctx = perf_swevent_get_recursion_context();
5805         /*
5806          * If we 'fail' here, that's OK, it means recursion is already disabled
5807          * and we won't recurse 'further'.
5808          */
5809 
5810         if (event->pending_disable) {
5811                 event->pending_disable = 0;
5812                 perf_event_disable_local(event);
5813         }
5814 
5815         if (event->pending_wakeup) {
5816                 event->pending_wakeup = 0;
5817                 perf_event_wakeup(event);
5818         }
5819 
5820         if (rctx >= 0)
5821                 perf_swevent_put_recursion_context(rctx);
5822 }
5823 
5824 /*
5825  * We assume there is only KVM supporting the callbacks.
5826  * Later on, we might change it to a list if there is
5827  * another virtualization implementation supporting the callbacks.
5828  */
5829 struct perf_guest_info_callbacks *perf_guest_cbs;
5830 
5831 int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5832 {
5833         perf_guest_cbs = cbs;
5834         return 0;
5835 }
5836 EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
5837 
5838 int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
5839 {
5840         perf_guest_cbs = NULL;
5841         return 0;
5842 }
5843 EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
5844 
5845 static void
5846 perf_output_sample_regs(struct perf_output_handle *handle,
5847                         struct pt_regs *regs, u64 mask)
5848 {
5849         int bit;
5850         DECLARE_BITMAP(_mask, 64);
5851 
5852         bitmap_from_u64(_mask, mask);
5853         for_each_set_bit(bit, _mask, sizeof(mask) * BITS_PER_BYTE) {
5854                 u64 val;
5855 
5856                 val = perf_reg_value(regs, bit);
5857                 perf_output_put(handle, val);
5858         }
5859 }
5860 
5861 static void perf_sample_regs_user(struct perf_regs *regs_user,
5862                                   struct pt_regs *regs,
5863                                   struct pt_regs *regs_user_copy)
5864 {
5865         if (user_mode(regs)) {
5866                 regs_user->abi = perf_reg_abi(current);
5867                 regs_user->regs = regs;
5868         } else if (current->mm) {
5869                 perf_get_regs_user(regs_user, regs, regs_user_copy);
5870         } else {
5871                 regs_user->abi = PERF_SAMPLE_REGS_ABI_NONE;
5872                 regs_user->regs = NULL;
5873         }
5874 }
5875 
5876 static void perf_sample_regs_intr(struct perf_regs *regs_intr,
5877                                   struct pt_regs *regs)
5878 {
5879         regs_intr->regs = regs;
5880         regs_intr->abi  = perf_reg_abi(current);
5881 }
5882 
5883 
5884 /*
5885  * Get remaining task size from user stack pointer.
5886  *
5887  * It'd be better to take stack vma map and limit this more
5888  * precisly, but there's no way to get it safely under interrupt,
5889  * so using TASK_SIZE as limit.
5890  */
5891 static u64 perf_ustack_task_size(struct pt_regs *regs)
5892 {
5893         unsigned long addr = perf_user_stack_pointer(regs);
5894 
5895         if (!addr || addr >= TASK_SIZE)
5896                 return 0;
5897 
5898         return TASK_SIZE - addr;
5899 }
5900 
5901 static u16
5902 perf_sample_ustack_size(u16 stack_size, u16 header_size,
5903                         struct pt_regs *regs)
5904 {
5905         u64 task_size;
5906 
5907         /* No regs, no stack pointer, no dump. */
5908         if (!regs)
5909                 return 0;
5910 
5911         /*
5912          * Check if we fit in with the requested stack size into the:
5913          * - TASK_SIZE
5914          *   If we don't, we limit the size to the TASK_SIZE.
5915          *
5916          * - remaining sample size
5917          *   If we don't, we customize the stack size to
5918          *   fit in to the remaining sample size.
5919          */
5920 
5921         task_size  = min((u64) USHRT_MAX, perf_ustack_task_size(regs));
5922         stack_size = min(stack_size, (u16) task_size);
5923 
5924         /* Current header size plus static size and dynamic size. */
5925         header_size += 2 * sizeof(u64);
5926 
5927         /* Do we fit in with the current stack dump size? */
5928         if ((u16) (header_size + stack_size) < header_size) {
5929                 /*
5930                  * If we overflow the maximum size for the sample,
5931                  * we customize the stack dump size to fit in.
5932                  */
5933                 stack_size = USHRT_MAX - header_size - sizeof(u64);
5934                 stack_size = round_up(stack_size, sizeof(u64));
5935         }
5936 
5937         return stack_size;
5938 }
5939 
5940 static void
5941 perf_output_sample_ustack(struct perf_output_handle *handle, u64 dump_size,
5942                           struct pt_regs *regs)
5943 {
5944         /* Case of a kernel thread, nothing to dump */
5945         if (!regs) {
5946                 u64 size = 0;
5947                 perf_output_put(handle, size);
5948         } else {
5949                 unsigned long sp;
5950                 unsigned int rem;
5951                 u64 dyn_size;
5952                 mm_segment_t fs;
5953 
5954                 /*
5955                  * We dump:
5956                  * static size
5957                  *   - the size requested by user or the best one we can fit
5958                  *     in to the sample max size
5959                  * data
5960                  *   - user stack dump data
5961                  * dynamic size
5962                  *   - the actual dumped size
5963                  */
5964 
5965                 /* Static size. */
5966                 perf_output_put(handle, dump_size);
5967 
5968                 /* Data. */
5969                 sp = perf_user_stack_pointer(regs);
5970                 fs = get_fs();
5971                 set_fs(USER_DS);
5972                 rem = __output_copy_user(handle, (void *) sp, dump_size);
5973                 set_fs(fs);
5974                 dyn_size = dump_size - rem;
5975 
5976                 perf_output_skip(handle, rem);
5977 
5978                 /* Dynamic size. */
5979                 perf_output_put(handle, dyn_size);
5980         }
5981 }
5982 
5983 static void __perf_event_header__init_id(struct perf_event_header *header,
5984                                          struct perf_sample_data *data,
5985                                          struct perf_event *event)
5986 {
5987         u64 sample_type = event->attr.sample_type;
5988 
5989         data->type = sample_type;
5990         header->size += event->id_header_size;
5991 
5992         if (sample_type & PERF_SAMPLE_TID) {
5993                 /* namespace issues */
5994                 data->tid_entry.pid = perf_event_pid(event, current);
5995                 data->tid_entry.tid = perf_event_tid(event, current);
5996         }
5997 
5998         if (sample_type & PERF_SAMPLE_TIME)
5999                 data->time = perf_event_clock(event);
6000 
6001         if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER))
6002                 data->id = primary_event_id(event);
6003 
6004         if (sample_type & PERF_SAMPLE_STREAM_ID)
6005                 data->stream_id = event->id;
6006 
6007         if (sample_type & PERF_SAMPLE_CPU) {
6008                 data->cpu_entry.cpu      = raw_smp_processor_id();
6009                 data->cpu_entry.reserved = 0;
6010         }
6011 }
6012 
6013 void perf_event_header__init_id(struct perf_event_header *header,
6014                                 struct perf_sample_data *data,
6015                                 struct perf_event *event)
6016 {
6017         if (event->attr.sample_id_all)
6018                 __perf_event_header__init_id(header, data, event);
6019 }
6020 
6021 static void __perf_event__output_id_sample(struct perf_output_handle *handle,
6022                                            struct perf_sample_data *data)
6023 {
6024         u64 sample_type = data->type;
6025 
6026         if (sample_type & PERF_SAMPLE_TID)
6027                 perf_output_put(handle, data->tid_entry);
6028 
6029         if (sample_type & PERF_SAMPLE_TIME)
6030                 perf_output_put(handle, data->time);
6031 
6032         if (sample_type & PERF_SAMPLE_ID)
6033                 perf_output_put(handle, data->id);
6034 
6035         if (sample_type & PERF_SAMPLE_STREAM_ID)
6036                 perf_output_put(handle, data->stream_id);
6037 
6038         if (sample_type & PERF_SAMPLE_CPU)
6039                 perf_output_put(handle, data->cpu_entry);
6040 
6041         if (sample_type & PERF_SAMPLE_IDENTIFIER)
6042                 perf_output_put(handle, data->id);
6043 }
6044 
6045 void perf_event__output_id_sample(struct perf_event *event,
6046                                   struct perf_output_handle *handle,
6047                                   struct perf_sample_data *sample)
6048 {
6049         if (event->attr.sample_id_all)
6050                 __perf_event__output_id_sample(handle, sample);
6051 }
6052 
6053 static void perf_output_read_one(struct perf_output_handle *handle,
6054                                  struct perf_event *event,
6055                                  u64 enabled, u64 running)
6056 {
6057         u64 read_format = event->attr.read_format;
6058         u64 values[4];
6059         int n = 0;
6060 
6061         values[n++] = perf_event_count(event);
6062         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
6063                 values[n++] = enabled +
6064                         atomic64_read(&event->child_total_time_enabled);
6065         }
6066         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
6067                 values[n++] = running +
6068                         atomic64_read(&event->child_total_time_running);
6069         }
6070         if (read_format & PERF_FORMAT_ID)
6071                 values[n++] = primary_event_id(event);
6072 
6073         __output_copy(handle, values, n * sizeof(u64));
6074 }
6075 
6076 static void perf_output_read_group(struct perf_output_handle *handle,
6077                             struct perf_event *event,
6078                             u64 enabled, u64 running)
6079 {
6080         struct perf_event *leader = event->group_leader, *sub;
6081         u64 read_format = event->attr.read_format;
6082         u64 values[5];
6083         int n = 0;
6084 
6085         values[n++] = 1 + leader->nr_siblings;
6086 
6087         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
6088                 values[n++] = enabled;
6089 
6090         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
6091                 values[n++] = running;
6092 
6093         if ((leader != event) &&
6094             (leader->state == PERF_EVENT_STATE_ACTIVE))
6095                 leader->pmu->read(leader);
6096 
6097         values[n++] = perf_event_count(leader);
6098         if (read_format & PERF_FORMAT_ID)
6099                 values[n++] = primary_event_id(leader);
6100 
6101         __output_copy(handle, values, n * sizeof(u64));
6102 
6103         for_each_sibling_event(sub, leader) {
6104                 n = 0;
6105 
6106                 if ((sub != event) &&
6107                     (sub->state == PERF_EVENT_STATE_ACTIVE))
6108                         sub->pmu->read(sub);
6109 
6110                 values[n++] = perf_event_count(sub);
6111                 if (read_format & PERF_FORMAT_ID)
6112                         values[n++] = primary_event_id(sub);
6113 
6114                 __output_copy(handle, values, n * sizeof(u64));
6115         }
6116 }
6117 
6118 #define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
6119                                  PERF_FORMAT_TOTAL_TIME_RUNNING)
6120 
6121 /*
6122  * XXX PERF_SAMPLE_READ vs inherited events seems difficult.
6123  *
6124  * The problem is that its both hard and excessively expensive to iterate the
6125  * child list, not to mention that its impossible to IPI the children running
6126  * on another CPU, from interrupt/NMI context.
6127  */
6128 static void perf_output_read(struct perf_output_handle *handle,
6129                              struct perf_event *event)
6130 {
6131         u64 enabled = 0, running = 0, now;
6132         u64 read_format = event->attr.read_format;
6133 
6134         /*
6135          * compute total_time_enabled, total_time_running
6136          * based on snapshot values taken when the event
6137          * was last scheduled in.
6138          *
6139          * we cannot simply called update_context_time()
6140          * because of locking issue as we are called in
6141          * NMI context
6142          */
6143         if (read_format & PERF_FORMAT_TOTAL_TIMES)
6144                 calc_timer_values(event, &now, &enabled, &running);
6145 
6146         if (event->attr.read_format & PERF_FORMAT_GROUP)
6147                 perf_output_read_group(handle, event, enabled, running);
6148         else
6149                 perf_output_read_one(handle, event, enabled, running);
6150 }
6151 
6152 void perf_output_sample(struct perf_output_handle *handle,
6153                         struct perf_event_header *header,
6154                         struct perf_sample_data *data,
6155                         struct perf_event *event)
6156 {
6157         u64 sample_type = data->type;
6158 
6159         perf_output_put(handle, *header);
6160 
6161         if (sample_type & PERF_SAMPLE_IDENTIFIER)
6162                 perf_output_put(handle, data->id);
6163 
6164         if (sample_type & PERF_SAMPLE_IP)
6165                 perf_output_put(handle, data->ip);
6166 
6167         if (sample_type & PERF_SAMPLE_TID)
6168                 perf_output_put(handle, data->tid_entry);
6169 
6170         if (sample_type & PERF_SAMPLE_TIME)
6171                 perf_output_put(handle, data->time);
6172 
6173         if (sample_type & PERF_SAMPLE_ADDR)
6174                 perf_output_put(handle, data->addr);
6175 
6176         if (sample_type & PERF_SAMPLE_ID)
6177                 perf_output_put(handle, data->id);
6178 
6179         if (sample_type & PERF_SAMPLE_STREAM_ID)
6180                 perf_output_put(handle, data->stream_id);
6181 
6182         if (sample_type & PERF_SAMPLE_CPU)
6183                 perf_output_put(handle, data->cpu_entry);
6184 
6185         if (sample_type & PERF_SAMPLE_PERIOD)
6186                 perf_output_put(handle, data->period);
6187 
6188         if (sample_type & PERF_SAMPLE_READ)
6189                 perf_output_read(handle, event);
6190 
6191         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
6192                 int size = 1;
6193 
6194                 size += data->callchain->nr;
6195                 size *= sizeof(u64);
6196                 __output_copy(handle, data->callchain, size);
6197         }
6198 
6199         if (sample_type & PERF_SAMPLE_RAW) {
6200                 struct perf_raw_record *raw = data->raw;
6201 
6202                 if (raw) {
6203                         struct perf_raw_frag *frag = &raw->frag;
6204 
6205                         perf_output_put(handle, raw->size);
6206                         do {
6207                                 if (frag->copy) {
6208                                         __output_custom(handle, frag->copy,
6209                                                         frag->data, frag->size);
6210                                 } else {
6211                                         __output_copy(handle, frag->data,
6212                                                       frag->size);
6213                                 }
6214                                 if (perf_raw_frag_last(frag))
6215                                         break;
6216                                 frag = frag->next;
6217                         } while (1);
6218                         if (frag->pad)
6219                                 __output_skip(handle, NULL, frag->pad);
6220                 } else {
6221                         struct {
6222                                 u32     size;
6223                                 u32     data;
6224                         } raw = {
6225                                 .size = sizeof(u32),
6226                                 .data = 0,
6227                         };
6228                         perf_output_put(handle, raw);
6229                 }
6230         }
6231 
6232         if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
6233                 if (data->br_stack) {
6234                         size_t size;
6235 
6236                         size = data->br_stack->nr
6237                              * sizeof(struct perf_branch_entry);
6238 
6239                         perf_output_put(handle, data->br_stack->nr);
6240                         perf_output_copy(handle, data->br_stack->entries, size);
6241                 } else {
6242                         /*
6243                          * we always store at least the value of nr
6244                          */
6245                         u64 nr = 0;
6246                         perf_output_put(handle, nr);
6247                 }
6248         }
6249 
6250         if (sample_type & PERF_SAMPLE_REGS_USER) {
6251                 u64 abi = data->regs_user.abi;
6252 
6253                 /*
6254                  * If there are no regs to dump, notice it through
6255                  * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
6256                  */
6257                 perf_output_put(handle, abi);
6258 
6259                 if (abi) {
6260                         u64 mask = event->attr.sample_regs_user;
6261                         perf_output_sample_regs(handle,
6262                                                 data->regs_user.regs,
6263                                                 mask);
6264                 }
6265         }
6266 
6267         if (sample_type & PERF_SAMPLE_STACK_USER) {
6268                 perf_output_sample_ustack(handle,
6269                                           data->stack_user_size,
6270                                           data->regs_user.regs);
6271         }
6272 
6273         if (sample_type & PERF_SAMPLE_WEIGHT)
6274                 perf_output_put(handle, data->weight);
6275 
6276         if (sample_type & PERF_SAMPLE_DATA_SRC)
6277                 perf_output_put(handle, data->data_src.val);
6278 
6279         if (sample_type & PERF_SAMPLE_TRANSACTION)
6280                 perf_output_put(handle, data->txn);
6281 
6282         if (sample_type & PERF_SAMPLE_REGS_INTR) {
6283                 u64 abi = data->regs_intr.abi;
6284                 /*
6285                  * If there are no regs to dump, notice it through
6286                  * first u64 being zero (PERF_SAMPLE_REGS_ABI_NONE).
6287                  */
6288                 perf_output_put(handle, abi);
6289 
6290                 if (abi) {
6291                         u64 mask = event->attr.sample_regs_intr;
6292 
6293                         perf_output_sample_regs(handle,
6294                                                 data->regs_intr.regs,
6295                                                 mask);
6296                 }
6297         }
6298 
6299         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
6300                 perf_output_put(handle, data->phys_addr);
6301 
6302         if (!event->attr.watermark) {
6303                 int wakeup_events = event->attr.wakeup_events;
6304 
6305                 if (wakeup_events) {
6306                         struct ring_buffer *rb = handle->rb;
6307                         int events = local_inc_return(&rb->events);
6308 
6309                         if (events >= wakeup_events) {
6310                                 local_sub(wakeup_events, &rb->events);
6311                                 local_inc(&rb->wakeup);
6312                         }
6313                 }
6314         }
6315 }
6316 
6317 static u64 perf_virt_to_phys(u64 virt)
6318 {
6319         u64 phys_addr = 0;
6320         struct page *p = NULL;
6321 
6322         if (!virt)
6323                 return 0;
6324 
6325         if (virt >= TASK_SIZE) {
6326                 /* If it's vmalloc()d memory, leave phys_addr as 0 */
6327                 if (virt_addr_valid((void *)(uintptr_t)virt) &&
6328                     !(virt >= VMALLOC_START && virt < VMALLOC_END))
6329                         phys_addr = (u64)virt_to_phys((void *)(uintptr_t)virt);
6330         } else {
6331                 /*
6332                  * Walking the pages tables for user address.
6333                  * Interrupts are disabled, so it prevents any tear down
6334                  * of the page tables.
6335                  * Try IRQ-safe __get_user_pages_fast first.
6336                  * If failed, leave phys_addr as 0.
6337                  */
6338                 if ((current->mm != NULL) &&
6339                     (__get_user_pages_fast(virt, 1, 0, &p) == 1))
6340                         phys_addr = page_to_phys(p) + virt % PAGE_SIZE;
6341 
6342                 if (p)
6343                         put_page(p);
6344         }
6345 
6346         return phys_addr;
6347 }
6348 
6349 static struct perf_callchain_entry __empty_callchain = { .nr = 0, };
6350 
6351 struct perf_callchain_entry *
6352 perf_callchain(struct perf_event *event, struct pt_regs *regs)
6353 {
6354         bool kernel = !event->attr.exclude_callchain_kernel;
6355         bool user   = !event->attr.exclude_callchain_user;
6356         /* Disallow cross-task user callchains. */
6357         bool crosstask = event->ctx->task && event->ctx->task != current;
6358         const u32 max_stack = event->attr.sample_max_stack;
6359         struct perf_callchain_entry *callchain;
6360 
6361         if (!kernel && !user)
6362                 return &__empty_callchain;
6363 
6364         callchain = get_perf_callchain(regs, 0, kernel, user,
6365                                        max_stack, crosstask, true);
6366         return callchain ?: &__empty_callchain;
6367 }
6368 
6369 void perf_prepare_sample(struct perf_event_header *header,
6370                          struct perf_sample_data *data,
6371                          struct perf_event *event,
6372                          struct pt_regs *regs)
6373 {
6374         u64 sample_type = event->attr.sample_type;
6375 
6376         header->type = PERF_RECORD_SAMPLE;
6377         header->size = sizeof(*header) + event->header_size;
6378 
6379         header->misc = 0;
6380         header->misc |= perf_misc_flags(regs);
6381 
6382         __perf_event_header__init_id(header, data, event);
6383 
6384         if (sample_type & PERF_SAMPLE_IP)
6385                 data->ip = perf_instruction_pointer(regs);
6386 
6387         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
6388                 int size = 1;
6389 
6390                 if (!(sample_type & __PERF_SAMPLE_CALLCHAIN_EARLY))
6391                         data->callchain = perf_callchain(event, regs);
6392 
6393                 size += data->callchain->nr;
6394 
6395                 header->size += size * sizeof(u64);
6396         }
6397 
6398         if (sample_type & PERF_SAMPLE_RAW) {
6399                 struct perf_raw_record *raw = data->raw;
6400                 int size;
6401 
6402                 if (raw) {
6403                         struct perf_raw_frag *frag = &raw->frag;
6404                         u32 sum = 0;
6405 
6406                         do {
6407                                 sum += frag->size;
6408                                 if (perf_raw_frag_last(frag))
6409                                         break;
6410                                 frag = frag->next;
6411                         } while (1);
6412 
6413                         size = round_up(sum + sizeof(u32), sizeof(u64));
6414                         raw->size = size - sizeof(u32);
6415                         frag->pad = raw->size - sum;
6416                 } else {
6417                         size = sizeof(u64);
6418                 }
6419 
6420                 header->size += size;
6421         }
6422 
6423         if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
6424                 int size = sizeof(u64); /* nr */
6425                 if (data->br_stack) {
6426                         size += data->br_stack->nr
6427                               * sizeof(struct perf_branch_entry);
6428                 }
6429                 header->size += size;
6430         }
6431 
6432         if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_STACK_USER))
6433                 perf_sample_regs_user(&data->regs_user, regs,
6434                                       &data->regs_user_copy);
6435 
6436         if (sample_type & PERF_SAMPLE_REGS_USER) {
6437                 /* regs dump ABI info */
6438                 int size = sizeof(u64);
6439 
6440                 if (data->regs_user.regs) {
6441                         u64 mask = event->attr.sample_regs_user;
6442                         size += hweight64(mask) * sizeof(u64);
6443                 }
6444 
6445                 header->size += size;
6446         }
6447 
6448         if (sample_type & PERF_SAMPLE_STACK_USER) {
6449                 /*
6450                  * Either we need PERF_SAMPLE_STACK_USER bit to be allways
6451                  * processed as the last one or have additional check added
6452                  * in case new sample type is added, because we could eat
6453                  * up the rest of the sample size.
6454                  */
6455                 u16 stack_size = event->attr.sample_stack_user;
6456                 u16 size = sizeof(u64);
6457 
6458                 stack_size = perf_sample_ustack_size(stack_size, header->size,
6459                                                      data->regs_user.regs);
6460 
6461                 /*
6462                  * If there is something to dump, add space for the dump
6463                  * itself and for the field that tells the dynamic size,
6464                  * which is how many have been actually dumped.
6465                  */
6466                 if (stack_size)
6467                         size += sizeof(u64) + stack_size;
6468 
6469                 data->stack_user_size = stack_size;
6470                 header->size += size;
6471         }
6472 
6473         if (sample_type & PERF_SAMPLE_REGS_INTR) {
6474                 /* regs dump ABI info */
6475                 int size = sizeof(u64);
6476 
6477                 perf_sample_regs_intr(&data->regs_intr, regs);
6478 
6479                 if (data->regs_intr.regs) {
6480                         u64 mask = event->attr.sample_regs_intr;
6481 
6482                         size += hweight64(mask) * sizeof(u64);
6483                 }
6484 
6485                 header->size += size;
6486         }
6487 
6488         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
6489                 data->phys_addr = perf_virt_to_phys(data->addr);
6490 }
6491 
6492 static __always_inline void
6493 __perf_event_output(struct perf_event *event,
6494                     struct perf_sample_data *data,
6495                     struct pt_regs *regs,
6496                     int (*output_begin)(struct perf_output_handle *,
6497                                         struct perf_event *,
6498                                         unsigned int))
6499 {
6500         struct perf_output_handle handle;
6501         struct perf_event_header header;
6502 
6503         /* protect the callchain buffers */
6504         rcu_read_lock();
6505 
6506         perf_prepare_sample(&header, data, event, regs);
6507 
6508         if (output_begin(&handle, event, header.size))
6509                 goto exit;
6510 
6511         perf_output_sample(&handle, &header, data, event);
6512 
6513         perf_output_end(&handle);
6514 
6515 exit:
6516         rcu_read_unlock();
6517 }
6518 
6519 void
6520 perf_event_output_forward(struct perf_event *event,
6521                          struct perf_sample_data *data,
6522                          struct pt_regs *regs)
6523 {
6524         __perf_event_output(event, data, regs, perf_output_begin_forward);
6525 }
6526 
6527 void
6528 perf_event_output_backward(struct perf_event *event,
6529                            struct perf_sample_data *data,
6530                            struct pt_regs *regs)
6531 {
6532         __perf_event_output(event, data, regs, perf_output_begin_backward);
6533 }
6534 
6535 void
6536 perf_event_output(struct perf_event *event,
6537                   struct perf_sample_data *data,
6538                   struct pt_regs *regs)
6539 {
6540         __perf_event_output(event, data, regs, perf_output_begin);
6541 }
6542 
6543 /*
6544  * read event_id
6545  */
6546 
6547 struct perf_read_event {
6548         struct perf_event_header        header;
6549 
6550         u32                             pid;
6551         u32                             tid;
6552 };
6553 
6554 static void
6555 perf_event_read_event(struct perf_event *event,
6556                         struct task_struct *task)
6557 {
6558         struct perf_output_handle handle;
6559         struct perf_sample_data sample;
6560         struct perf_read_event read_event = {
6561                 .header = {
6562                         .type = PERF_RECORD_READ,
6563                         .misc = 0,
6564                         .size = sizeof(read_event) + event->read_size,
6565                 },
6566                 .pid = perf_event_pid(event, task),
6567                 .tid = perf_event_tid(event, task),
6568         };
6569         int ret;
6570 
6571         perf_event_header__init_id(&read_event.header, &sample, event);
6572         ret = perf_output_begin(&handle, event, read_event.header.size);
6573         if (ret)
6574                 return;
6575 
6576         perf_output_put(&handle, read_event);
6577         perf_output_read(&handle, event);
6578         perf_event__output_id_sample(event, &handle, &sample);
6579 
6580         perf_output_end(&handle);
6581 }
6582 
6583 typedef void (perf_iterate_f)(struct perf_event *event, void *data);
6584 
6585 static void
6586 perf_iterate_ctx(struct perf_event_context *ctx,
6587                    perf_iterate_f output,
6588                    void *data, bool all)
6589 {
6590         struct perf_event *event;
6591 
6592         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
6593                 if (!all) {
6594                         if (event->state < PERF_EVENT_STATE_INACTIVE)
6595                                 continue;
6596                         if (!event_filter_match(event))
6597                                 continue;
6598                 }
6599 
6600                 output(event, data);
6601         }
6602 }
6603 
6604 static void perf_iterate_sb_cpu(perf_iterate_f output, void *data)
6605 {
6606         struct pmu_event_list *pel = this_cpu_ptr(&pmu_sb_events);
6607         struct perf_event *event;
6608 
6609         list_for_each_entry_rcu(event, &pel->list, sb_list) {
6610                 /*
6611                  * Skip events that are not fully formed yet; ensure that
6612                  * if we observe event->ctx, both event and ctx will be
6613                  * complete enough. See perf_install_in_context().
6614                  */
6615                 if (!smp_load_acquire(&event->ctx))
6616                         continue;
6617 
6618                 if (event->state < PERF_EVENT_STATE_INACTIVE)
6619                         continue;
6620                 if (!event_filter_match(event))
6621                         continue;
6622                 output(event, data);
6623         }
6624 }
6625 
6626 /*
6627  * Iterate all events that need to receive side-band events.
6628  *
6629  * For new callers; ensure that account_pmu_sb_event() includes
6630  * your event, otherwise it might not get delivered.
6631  */
6632 static void
6633 perf_iterate_sb(perf_iterate_f output, void *data,
6634                struct perf_event_context *task_ctx)
6635 {
6636         struct perf_event_context *ctx;
6637         int ctxn;
6638 
6639         rcu_read_lock();
6640         preempt_disable();
6641 
6642         /*
6643          * If we have task_ctx != NULL we only notify the task context itself.
6644          * The task_ctx is set only for EXIT events before releasing task
6645          * context.
6646          */
6647         if (task_ctx) {
6648                 perf_iterate_ctx(task_ctx, output, data, false);
6649                 goto done;
6650         }
6651 
6652         perf_iterate_sb_cpu(output, data);
6653 
6654         for_each_task_context_nr(ctxn) {
6655                 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
6656                 if (ctx)
6657                         perf_iterate_ctx(ctx, output, data, false);
6658         }
6659 done:
6660         preempt_enable();
6661         rcu_read_unlock();
6662 }
6663 
6664 /*
6665  * Clear all file-based filters at exec, they'll have to be
6666  * re-instated when/if these objects are mmapped again.
6667  */
6668 static void perf_event_addr_filters_exec(struct perf_event *event, void *data)
6669 {
6670         struct perf_addr_filters_head *ifh = perf_event_addr_filters(event);
6671         struct perf_addr_filter *filter;
6672         unsigned int restart = 0, count = 0;
6673         unsigned long flags;
6674 
6675         if (!has_addr_filter(event))
6676                 return;
6677 
6678         raw_spin_lock_irqsave(&ifh->lock, flags);
6679         list_for_each_entry(filter, &ifh->list, entry) {
6680                 if (filter->path.dentry) {
6681                         event->addr_filters_offs[count] = 0;
6682                         restart++;
6683                 }
6684 
6685                 count++;
6686         }
6687 
6688         if (restart)
6689                 event->addr_filters_gen++;
6690         raw_spin_unlock_irqrestore(&ifh->lock, flags);
6691 
6692         if (restart)
6693                 perf_event_stop(event, 1);
6694 }
6695 
6696 void perf_event_exec(void)
6697 {
6698         struct perf_event_context *ctx;
6699         int ctxn;
6700 
6701         rcu_read_lock();
6702         for_each_task_context_nr(ctxn) {
6703                 ctx = current->perf_event_ctxp[ctxn];
6704                 if (!ctx)
6705                         continue;
6706 
6707                 perf_event_enable_on_exec(ctxn);
6708 
6709                 perf_iterate_ctx(ctx, perf_event_addr_filters_exec, NULL,
6710                                    true);
6711         }
6712         rcu_read_unlock();
6713 }
6714 
6715 struct remote_output {
6716         struct ring_buffer      *rb;
6717         int                     err;
6718 };
6719 
6720 static void __perf_event_output_stop(struct perf_event *event, void *data)
6721 {
6722         struct perf_event *parent = event->parent;
6723         struct remote_output *ro = data;
6724         struct ring_buffer *rb = ro->rb;
6725         struct stop_event_data sd = {
6726                 .event  = event,
6727         };
6728 
6729         if (!has_aux(event))
6730                 return;
6731 
6732         if (!parent)
6733                 parent = event;
6734 
6735         /*
6736          * In case of inheritance, it will be the parent that links to the
6737          * ring-buffer, but it will be the child that's actually using it.
6738          *
6739          * We are using event::rb to determine if the event should be stopped,
6740          * however this may race with ring_buffer_attach() (through set_output),
6741          * which will make us skip the event that actually needs to be stopped.
6742          * So ring_buffer_attach() has to stop an aux event before re-assigning
6743          * its rb pointer.
6744          */
6745         if (rcu_dereference(parent->rb) == rb)
6746                 ro->err = __perf_event_stop(&sd);
6747 }
6748 
6749 static int __perf_pmu_output_stop(void *info)
6750 {
6751         struct perf_event *event = info;
6752         struct pmu *pmu = event->pmu;
6753         struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
6754         struct remote_output ro = {
6755                 .rb     = event->rb,
6756         };
6757 
6758         rcu_read_lock();
6759         perf_iterate_ctx(&cpuctx->ctx, __perf_event_output_stop, &ro, false);
6760         if (cpuctx->task_ctx)
6761                 perf_iterate_ctx(cpuctx->task_ctx, __perf_event_output_stop,
6762                                    &ro, false);
6763         rcu_read_unlock();
6764 
6765         return ro.err;
6766 }
6767 
6768 static void perf_pmu_output_stop(struct perf_event *event)
6769 {
6770         struct perf_event *iter;
6771         int err, cpu;
6772 
6773 restart:
6774         rcu_read_lock();
6775         list_for_each_entry_rcu(iter, &event->rb->event_list, rb_entry) {
6776                 /*
6777                  * For per-CPU events, we need to make sure that neither they
6778                  * nor their children are running; for cpu==-1 events it's
6779                  * sufficient to stop the event itself if it's active, since
6780                  * it can't have children.
6781                  */
6782                 cpu = iter->cpu;
6783                 if (cpu == -1)
6784                         cpu = READ_ONCE(iter->oncpu);
6785 
6786                 if (cpu == -1)
6787                         continue;
6788 
6789                 err = cpu_function_call(cpu, __perf_pmu_output_stop, event);
6790                 if (err == -EAGAIN) {
6791                         rcu_read_unlock();
6792                         goto restart;
6793                 }
6794         }
6795         rcu_read_unlock();
6796 }
6797 
6798 /*
6799  * task tracking -- fork/exit
6800  *
6801  * enabled by: attr.comm | attr.mmap | attr.mmap2 | attr.mmap_data | attr.task
6802  */
6803 
6804 struct perf_task_event {
6805         struct task_struct              *task;
6806         struct perf_event_context       *task_ctx;
6807 
6808         struct {
6809                 struct perf_event_header        header;
6810 
6811                 u32                             pid;
6812                 u32                             ppid;
6813                 u32                             tid;
6814                 u32                             ptid;
6815                 u64                             time;
6816         } event_id;
6817 };
6818 
6819 static int perf_event_task_match(struct perf_event *event)
6820 {
6821         return event->attr.comm  || event->attr.mmap ||
6822                event->attr.mmap2 || event->attr.mmap_data ||
6823                event->attr.task;
6824 }
6825 
6826 static void perf_event_task_output(struct perf_event *event,
6827                                    void *data)
6828 {
6829         struct perf_task_event *task_event = data;
6830         struct perf_output_handle handle;
6831         struct perf_sample_data sample;
6832         struct task_struct *task = task_event->task;
6833         int ret, size = task_event->event_id.header.size;
6834 
6835         if (!perf_event_task_match(event))
6836                 return;
6837 
6838         perf_event_header__init_id(&task_event->event_id.header, &sample, event);
6839 
6840         ret = perf_output_begin(&handle, event,
6841                                 task_event->event_id.header.size);
6842         if (ret)
6843                 goto out;
6844 
6845         task_event->event_id.pid = perf_event_pid(event, task);
6846         task_event->event_id.ppid = perf_event_pid(event, current);
6847 
6848         task_event->event_id.tid = perf_event_tid(event, task);
6849         task_event->event_id.ptid = perf_event_tid(event, current);
6850 
6851         task_event->event_id.time = perf_event_clock(event);
6852 
6853         perf_output_put(&handle, task_event->event_id);
6854 
6855         perf_event__output_id_sample(event, &handle, &sample);
6856 
6857         perf_output_end(&handle);
6858 out:
6859         task_event->event_id.header.size = size;
6860 }
6861 
6862 static void perf_event_task(struct task_struct *task,
6863                               struct perf_event_context *task_ctx,
6864                               int new)
6865 {
6866         struct perf_task_event task_event;
6867 
6868         if (!atomic_read(&nr_comm_events) &&
6869             !atomic_read(&nr_mmap_events) &&
6870             !atomic_read(&nr_task_events))
6871                 return;
6872 
6873         task_event = (struct perf_task_event){
6874                 .task     = task,
6875                 .task_ctx = task_ctx,
6876                 .event_id    = {
6877                         .header = {
6878                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
6879                                 .misc = 0,
6880                                 .size = sizeof(task_event.event_id),
6881                         },
6882                         /* .pid  */
6883                         /* .ppid */
6884                         /* .tid  */
6885                         /* .ptid */
6886                         /* .time */
6887                 },
6888         };
6889 
6890         perf_iterate_sb(perf_event_task_output,
6891                        &task_event,
6892                        task_ctx);
6893 }
6894 
6895 void perf_event_fork(struct task_struct *task)
6896 {
6897         perf_event_task(task, NULL, 1);
6898         perf_event_namespaces(task);
6899 }
6900 
6901 /*
6902  * comm tracking
6903  */
6904 
6905 struct perf_comm_event {
6906         struct task_struct      *task;
6907         char                    *comm;
6908         int                     comm_size;
6909 
6910         struct {
6911                 struct perf_event_header        header;
6912 
6913                 u32                             pid;
6914                 u32                             tid;
6915         } event_id;
6916 };
6917 
6918 static int perf_event_comm_match(struct perf_event *event)
6919 {
6920         return event->attr.comm;
6921 }
6922 
6923 static void perf_event_comm_output(struct perf_event *event,
6924                                    void *data)
6925 {
6926         struct perf_comm_event *comm_event = data;
6927         struct perf_output_handle handle;
6928         struct perf_sample_data sample;
6929         int size = comm_event->event_id.header.size;
6930         int ret;
6931 
6932         if (!perf_event_comm_match(event))
6933                 return;
6934 
6935         perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
6936         ret = perf_output_begin(&handle, event,
6937                                 comm_event->event_id.header.size);
6938 
6939         if (ret)
6940                 goto out;
6941 
6942         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
6943         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
6944 
6945         perf_output_put(&handle, comm_event->event_id);
6946         __output_copy(&handle, comm_event->comm,
6947                                    comm_event->comm_size);
6948 
6949         perf_event__output_id_sample(event, &handle, &sample);
6950 
6951         perf_output_end(&handle);
6952 out:
6953         comm_event->event_id.header.size = size;
6954 }
6955 
6956 static void perf_event_comm_event(struct perf_comm_event *comm_event)
6957 {
6958         char comm[TASK_COMM_LEN];
6959         unsigned int size;
6960 
6961         memset(comm, 0, sizeof(comm));
6962         strlcpy(comm, comm_event->task->comm, sizeof(comm));
6963         size = ALIGN(strlen(comm)+1, sizeof(u64));
6964 
6965         comm_event->comm = comm;
6966         comm_event->comm_size = size;
6967 
6968         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
6969 
6970         perf_iterate_sb(perf_event_comm_output,
6971                        comm_event,
6972                        NULL);
6973 }
6974 
6975 void perf_event_comm(struct task_struct *task, bool exec)
6976 {
6977         struct perf_comm_event comm_event;
6978 
6979         if (!atomic_read(&nr_comm_events))
6980                 return;
6981 
6982         comm_event = (struct perf_comm_event){
6983                 .task   = task,
6984                 /* .comm      */
6985                 /* .comm_size */
6986                 .event_id  = {
6987                         .header = {
6988                                 .type = PERF_RECORD_COMM,
6989                                 .misc = exec ? PERF_RECORD_MISC_COMM_EXEC : 0,
6990                                 /* .size */
6991                         },
6992                         /* .pid */
6993                         /* .tid */
6994                 },
6995         };
6996 
6997         perf_event_comm_event(&comm_event);
6998 }
6999 
7000 /*
7001  * namespaces tracking
7002  */
7003 
7004 struct perf_namespaces_event {
7005         struct task_struct              *task;
7006 
7007         struct {
7008                 struct perf_event_header        header;
7009 
7010                 u32                             pid;
7011                 u32                             tid;
7012                 u64                             nr_namespaces;
7013                 struct perf_ns_link_info        link_info[NR_NAMESPACES];
7014         } event_id;
7015 };
7016 
7017 static int perf_event_namespaces_match(struct perf_event *event)
7018 {
7019         return event->attr.namespaces;
7020 }
7021 
7022 static void perf_event_namespaces_output(struct perf_event *event,
7023                                          void *data)
7024 {
7025         struct perf_namespaces_event *namespaces_event = data;
7026         struct perf_output_handle handle;
7027         struct perf_sample_data sample;
7028         u16 header_size = namespaces_event->event_id.header.size;
7029         int ret;
7030 
7031         if (!perf_event_namespaces_match(event))
7032                 return;
7033 
7034         perf_event_header__init_id(&namespaces_event->event_id.header,
7035                                    &sample, event);
7036         ret = perf_output_begin(&handle, event,
7037                                 namespaces_event->event_id.header.size);
7038         if (ret)
7039                 goto out;
7040 
7041         namespaces_event->event_id.pid = perf_event_pid(event,
7042                                                         namespaces_event->task);
7043         namespaces_event->event_id.tid = perf_event_tid(event,
7044                                                         namespaces_event->task);
7045 
7046         perf_output_put(&handle, namespaces_event->event_id);
7047 
7048         perf_event__output_id_sample(event, &handle, &sample);
7049 
7050         perf_output_end(&handle);
7051 out:
7052         namespaces_event->event_id.header.size = header_size;
7053 }
7054 
7055 static void perf_fill_ns_link_info(struct perf_ns_link_info *ns_link_info,
7056                                    struct task_struct *task,
7057                                    const struct proc_ns_operations *ns_ops)
7058 {
7059         struct path ns_path;
7060         struct inode *ns_inode;
7061         void *error;
7062 
7063         error = ns_get_path(&ns_path, task, ns_ops);
7064         if (!error) {
7065                 ns_inode = ns_path.dentry->d_inode;
7066                 ns_link_info->dev = new_encode_dev(ns_inode->i_sb->s_dev);
7067                 ns_link_info->ino = ns_inode->i_ino;
7068                 path_put(&ns_path);
7069         }
7070 }
7071 
7072 void perf_event_namespaces(struct task_struct *task)
7073 {
7074         struct perf_namespaces_event namespaces_event;
7075         struct perf_ns_link_info *ns_link_info;
7076 
7077         if (!atomic_read(&nr_namespaces_events))
7078                 return;
7079 
7080         namespaces_event = (struct perf_namespaces_event){
7081                 .task   = task,
7082                 .event_id  = {
7083                         .header = {
7084                                 .type = PERF_RECORD_NAMESPACES,
7085                                 .misc = 0,
7086                                 .size = sizeof(namespaces_event.event_id),
7087                         },
7088                         /* .pid */
7089                         /* .tid */
7090                         .nr_namespaces = NR_NAMESPACES,
7091                         /* .link_info[NR_NAMESPACES] */
7092                 },
7093         };
7094 
7095         ns_link_info = namespaces_event.event_id.link_info;
7096 
7097         perf_fill_ns_link_info(&ns_link_info[MNT_NS_INDEX],
7098                                task, &mntns_operations);
7099 
7100 #ifdef CONFIG_USER_NS
7101         perf_fill_ns_link_info(&ns_link_info[USER_NS_INDEX],
7102                                task, &userns_operations);
7103 #endif
7104 #ifdef CONFIG_NET_NS
7105         perf_fill_ns_link_info(&ns_link_info[NET_NS_INDEX],
7106                                task, &netns_operations);
7107 #endif
7108 #ifdef CONFIG_UTS_NS
7109         perf_fill_ns_link_info(&ns_link_info[UTS_NS_INDEX],
7110                                task, &utsns_operations);
7111 #endif
7112 #ifdef CONFIG_IPC_NS
7113         perf_fill_ns_link_info(&ns_link_info[IPC_NS_INDEX],
7114                                task, &ipcns_operations);
7115 #endif
7116 #ifdef CONFIG_PID_NS
7117         perf_fill_ns_link_info(&ns_link_info[PID_NS_INDEX],
7118                                task, &pidns_operations);
7119 #endif
7120 #ifdef CONFIG_CGROUPS
7121         perf_fill_ns_link_info(&ns_link_info[CGROUP_NS_INDEX],
7122                                task, &cgroupns_operations);
7123 #endif
7124 
7125         perf_iterate_sb(perf_event_namespaces_output,
7126                         &namespaces_event,
7127                         NULL);
7128 }
7129 
7130 /*
7131  * mmap tracking
7132  */
7133 
7134 struct perf_mmap_event {
7135         struct vm_area_struct   *vma;
7136 
7137         const char              *file_name;
7138         int                     file_size;
7139         int                     maj, min;
7140         u64                     ino;
7141         u64                     ino_generation;
7142         u32                     prot, flags;
7143 
7144         struct {
7145                 struct perf_event_header        header;
7146 
7147                 u32                             pid;
7148                 u32                             tid;
7149                 u64                             start;
7150                 u64                             len;
7151                 u64                             pgoff;
7152         } event_id;
7153 };
7154 
7155 static int perf_event_mmap_match(struct perf_event *event,
7156                                  void *data)
7157 {
7158         struct perf_mmap_event *mmap_event = data;
7159         struct vm_area_struct *vma = mmap_event->vma;
7160         int executable = vma->vm_flags & VM_EXEC;
7161 
7162         return (!executable && event->attr.mmap_data) ||
7163                (executable && (event->attr.mmap || event->attr.mmap2));
7164 }
7165 
7166 static void perf_event_mmap_output(struct perf_event *event,
7167                                    void *data)
7168 {
7169         struct perf_mmap_event *mmap_event = data;
7170         struct perf_output_handle handle;
7171         struct perf_sample_data sample;
7172         int size = mmap_event->event_id.header.size;
7173         int ret;
7174 
7175         if (!perf_event_mmap_match(event, data))
7176                 return;
7177 
7178         if (event->attr.mmap2) {
7179                 mmap_event->event_id.header.type = PERF_RECORD_MMAP2;
7180                 mmap_event->event_id.header.size += sizeof(mmap_event->maj);
7181                 mmap_event->event_id.header.size += sizeof(mmap_event->min);
7182                 mmap_event->event_id.header.size += sizeof(mmap_event->ino);
7183                 mmap_event->event_id.header.size += sizeof(mmap_event->ino_generation);
7184                 mmap_event->event_id.header.size += sizeof(mmap_event->prot);
7185                 mmap_event->event_id.header.size += sizeof(mmap_event->flags);
7186         }
7187 
7188         perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
7189         ret = perf_output_begin(&handle, event,
7190                                 mmap_event->event_id.header.size);
7191         if (ret)
7192                 goto out;
7193 
7194         mmap_event->event_id.pid = perf_event_pid(event, current);
7195         mmap_event->event_id.tid = perf_event_tid(event, current);
7196 
7197         perf_output_put(&handle, mmap_event->event_id);
7198 
7199         if (event->attr.mmap2) {
7200                 perf_output_put(&handle, mmap_event->maj);
7201                 perf_output_put(&handle, mmap_event->min);
7202                 perf_output_put(&handle, mmap_event->ino);
7203                 perf_output_put(&handle, mmap_event->ino_generation);
7204                 perf_output_put(&handle, mmap_event->prot);
7205                 perf_output_put(&handle, mmap_event->flags);
7206         }
7207 
7208         __output_copy(&handle, mmap_event->file_name,
7209                                    mmap_event->file_size);
7210 
7211         perf_event__output_id_sample(event, &handle, &sample);
7212 
7213         perf_output_end(&handle);
7214 out:
7215         mmap_event->event_id.header.size = size;
7216 }
7217 
7218 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
7219 {
7220         struct vm_area_struct *vma = mmap_event->vma;
7221         struct file *file = vma->vm_file;
7222         int maj = 0, min = 0;
7223         u64 ino = 0, gen = 0;
7224         u32 prot = 0, flags = 0;
7225         unsigned int size;
7226         char tmp[16];
7227         char *buf = NULL;
7228         char *name;
7229 
7230         if (vma->vm_flags & VM_READ)
7231                 prot |= PROT_READ;
7232         if (vma->vm_flags & VM_WRITE)
7233                 prot |= PROT_WRITE;
7234         if (vma->vm_flags & VM_EXEC)
7235                 prot |= PROT_EXEC;
7236 
7237         if (vma->vm_flags & VM_MAYSHARE)
7238                 flags = MAP_SHARED;
7239         else
7240                 flags = MAP_PRIVATE;
7241 
7242         if (vma->vm_flags & VM_DENYWRITE)
7243                 flags |= MAP_DENYWRITE;
7244         if (vma->vm_flags & VM_MAYEXEC)
7245                 flags |= MAP_EXECUTABLE;
7246         if (vma->vm_flags & VM_LOCKED)
7247                 flags |= MAP_LOCKED;
7248         if (vma->vm_flags & VM_HUGETLB)
7249                 flags |= MAP_HUGETLB;
7250 
7251         if (file) {
7252                 struct inode *inode;
7253                 dev_t dev;
7254 
7255                 buf = kmalloc(PATH_MAX, GFP_KERNEL);
7256                 if (!buf) {
7257                         name = "//enomem";
7258                         goto cpy_name;
7259                 }
7260                 /*
7261                  * d_path() works from the end of the rb backwards, so we
7262                  * need to add enough zero bytes after the string to handle
7263                  * the 64bit alignment we do later.
7264                  */
7265                 name = file_path(file, buf, PATH_MAX - sizeof(u64));
7266                 if (IS_ERR(name)) {
7267                         name = "//toolong";
7268                         goto cpy_name;
7269                 }
7270                 inode = file_inode(vma->vm_file);
7271                 dev = inode->i_sb->s_dev;
7272                 ino = inode->i_ino;
7273                 gen = inode->i_generation;
7274                 maj = MAJOR(dev);
7275                 min = MINOR(dev);
7276 
7277                 goto got_name;
7278         } else {
7279                 if (vma->vm_ops && vma->vm_ops->name) {
7280                         name = (char *) vma->vm_ops->name(vma);
7281                         if (name)
7282                                 goto cpy_name;
7283                 }
7284 
7285                 name = (char *)arch_vma_name(vma);
7286                 if (name)
7287                         goto cpy_name;
7288 
7289                 if (vma->vm_start <= vma->vm_mm->start_brk &&
7290                                 vma->vm_end >= vma->vm_mm->brk) {
7291                         name = "[heap]";
7292                         goto cpy_name;
7293                 }
7294                 if (vma->vm_start <= vma->vm_mm->start_stack &&
7295                                 vma->vm_end >= vma->vm_mm->start_stack) {
7296                         name = "[stack]";
7297                         goto cpy_name;
7298                 }
7299 
7300                 name = "//anon";
7301                 goto cpy_name;
7302         }
7303 
7304 cpy_name:
7305         strlcpy(tmp, name, sizeof(tmp));
7306         name = tmp;
7307 got_name:
7308         /*
7309          * Since our buffer works in 8 byte units we need to align our string
7310          * size to a multiple of 8. However, we must guarantee the tail end is
7311          * zero'd out to avoid leaking random bits to userspace.
7312          */
7313         size = strlen(name)+1;
7314         while (!IS_ALIGNED(size, sizeof(u64)))
7315                 name[size++] = '\0';
7316 
7317         mmap_event->file_name = name;
7318         mmap_event->file_size = size;
7319         mmap_event->maj = maj;
7320         mmap_event->min = min;
7321         mmap_event->ino = ino;
7322         mmap_event->ino_generation = gen;
7323         mmap_event->prot = prot;
7324         mmap_event->flags = flags;
7325 
7326         if (!(vma->vm_flags & VM_EXEC))
7327                 mmap_event->event_id.header.misc |= PERF_RECORD_MISC_MMAP_DATA;
7328 
7329         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
7330 
7331         perf_iterate_sb(perf_event_mmap_output,
7332                        mmap_event,
7333                        NULL);
7334 
7335         kfree(buf);
7336 }
7337 
7338 /*
7339  * Check whether inode and address range match filter criteria.
7340  */
7341 static bool perf_addr_filter_match(struct perf_addr_filter *filter,
7342                                      struct file *file, unsigned long offset,
7343                                      unsigned long size)
7344 {
7345         /* d_inode(NULL) won't be equal to any mapped user-space file */
7346         if (!filter->path.dentry)
7347                 return false;
7348 
7349         if (d_inode(filter->path.dentry) != file_inode(file))
7350                 return false;
7351 
7352         if (filter->offset > offset + size)
7353                 return false;
7354 
7355         if (filter->offset + filter->size < offset)
7356                 return false;
7357 
7358         return true;
7359 }
7360 
7361 static void __perf_addr_filters_adjust(struct perf_event *event, void *data)
7362 {
7363         struct perf_addr_filters_head *ifh = perf_event_addr_filters(event);
7364         struct vm_area_struct *vma = data;
7365         unsigned long off = vma->vm_pgoff << PAGE_SHIFT, flags;
7366         struct file *file = vma->vm_file;
7367         struct perf_addr_filter *filter;
7368         unsigned int restart = 0, count = 0;
7369 
7370         if (!has_addr_filter(event))
7371                 return;
7372 
7373         if (!file)
7374                 return;
7375 
7376         raw_spin_lock_irqsave(&ifh->lock, flags);
7377         list_for_each_entry(filter, &ifh->list, entry) {
7378                 if (perf_addr_filter_match(filter, file, off,
7379                                              vma->vm_end - vma->vm_start)) {
7380                         event->addr_filters_offs[count] = vma->vm_start;
7381                         restart++;
7382                 }
7383 
7384                 count++;
7385         }
7386 
7387         if (restart)
7388                 event->addr_filters_gen++;
7389         raw_spin_unlock_irqrestore(&ifh->lock, flags);
7390 
7391         if (restart)
7392                 perf_event_stop(event, 1);
7393 }
7394 
7395 /*
7396  * Adjust all task's events' filters to the new vma
7397  */
7398 static void perf_addr_filters_adjust(struct vm_area_struct *vma)
7399 {
7400         struct perf_event_context *ctx;
7401         int ctxn;
7402 
7403         /*
7404          * Data tracing isn't supported yet and as such there is no need
7405          * to keep track of anything that isn't related to executable code:
7406          */
7407         if (!(vma->vm_flags & VM_EXEC))
7408                 return;
7409 
7410         rcu_read_lock();
7411         for_each_task_context_nr(ctxn) {
7412                 ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
7413                 if (!ctx)
7414                         continue;
7415 
7416                 perf_iterate_ctx(ctx, __perf_addr_filters_adjust, vma, true);
7417         }
7418         rcu_read_unlock();
7419 }
7420 
7421 void perf_event_mmap(struct vm_area_struct *vma)
7422 {
7423         struct perf_mmap_event mmap_event;
7424 
7425         if (!atomic_read(&nr_mmap_events))
7426                 return;
7427 
7428         mmap_event = (struct perf_mmap_event){
7429                 .vma    = vma,
7430                 /* .file_name */
7431                 /* .file_size */
7432                 .event_id  = {
7433                         .header = {
7434                                 .type = PERF_RECORD_MMAP,
7435                                 .misc = PERF_RECORD_MISC_USER,
7436                                 /* .size */
7437                         },
7438                         /* .pid */
7439                         /* .tid */
7440                         .start  = vma->vm_start,
7441                         .len    = vma->vm_end - vma->vm_start,
7442                         .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
7443                 },
7444                 /* .maj (attr_mmap2 only) */
7445                 /* .min (attr_mmap2 only) */
7446                 /* .ino (attr_mmap2 only) */
7447                 /* .ino_generation (attr_mmap2 only) */
7448                 /* .prot (attr_mmap2 only) */
7449                 /* .flags (attr_mmap2 only) */
7450         };
7451 
7452         perf_addr_filters_adjust(vma);
7453         perf_event_mmap_event(&mmap_event);
7454 }
7455 
7456 void perf_event_aux_event(struct perf_event *event, unsigned long head,
7457                           unsigned long size, u64 flags)
7458 {
7459         struct perf_output_handle handle;
7460         struct perf_sample_data sample;
7461         struct perf_aux_event {
7462                 struct perf_event_header        header;
7463                 u64                             offset;
7464                 u64                             size;
7465                 u64                             flags;
7466         } rec = {
7467                 .header = {
7468                         .type = PERF_RECORD_AUX,
7469                         .misc = 0,
7470                         .size = sizeof(rec),
7471                 },
7472                 .offset         = head,
7473                 .size           = size,
7474                 .flags          = flags,
7475         };
7476         int ret;
7477 
7478         perf_event_header__init_id(&rec.header, &sample, event);
7479         ret = perf_output_begin(&handle, event, rec.header.size);
7480 
7481         if (ret)
7482                 return;
7483 
7484         perf_output_put(&handle, rec);
7485         perf_event__output_id_sample(event, &handle, &sample);
7486 
7487         perf_output_end(&handle);
7488 }
7489 
7490 /*
7491  * Lost/dropped samples logging
7492  */
7493 void perf_log_lost_samples(struct perf_event *event, u64 lost)
7494 {
7495         struct perf_output_handle handle;
7496         struct perf_sample_data sample;
7497         int ret;
7498 
7499         struct {
7500                 struct perf_event_header        header;
7501                 u64                             lost;
7502         } lost_samples_event = {
7503                 .header = {
7504                         .type = PERF_RECORD_LOST_SAMPLES,
7505                         .misc = 0,
7506                         .size = sizeof(lost_samples_event),
7507                 },
7508                 .lost           = lost,
7509         };
7510 
7511         perf_event_header__init_id(&lost_samples_event.header, &sample, event);
7512 
7513         ret = perf_output_begin(&handle, event,
7514                                 lost_samples_event.header.size);
7515         if (ret)
7516                 return;
7517 
7518         perf_output_put(&handle, lost_samples_event);
7519         perf_event__output_id_sample(event, &handle, &sample);
7520         perf_output_end(&handle);
7521 }
7522 
7523 /*
7524  * context_switch tracking
7525  */
7526 
7527 struct perf_switch_event {
7528         struct task_struct      *task;
7529         struct task_struct      *next_prev;
7530 
7531         struct {
7532                 struct perf_event_header        header;
7533                 u32                             next_prev_pid;
7534                 u32                             next_prev_tid;
7535         } event_id;
7536 };
7537 
7538 static int perf_event_switch_match(struct perf_event *event)
7539 {
7540         return event->attr.context_switch;
7541 }
7542 
7543 static void perf_event_switch_output(struct perf_event *event, void *data)
7544 {
7545         struct perf_switch_event *se = data;
7546         struct perf_output_handle handle;
7547         struct perf_sample_data sample;
7548         int ret;
7549 
7550         if (!perf_event_switch_match(event))
7551                 return;
7552 
7553         /* Only CPU-wide events are allowed to see next/prev pid/tid */
7554         if (event->ctx->task) {
7555                 se->event_id.header.type = PERF_RECORD_SWITCH;
7556                 se->event_id.header.size = sizeof(se->event_id.header);
7557         } else {
7558                 se->event_id.header.type = PERF_RECORD_SWITCH_CPU_WIDE;
7559                 se->event_id.header.size = sizeof(se->event_id);
7560                 se->event_id.next_prev_pid =
7561                                         perf_event_pid(event, se->next_prev);
7562                 se->event_id.next_prev_tid =
7563                                         perf_event_tid(event, se->next_prev);
7564         }
7565 
7566         perf_event_header__init_id(&se->event_id.header, &sample, event);
7567 
7568         ret = perf_output_begin(&handle, event, se->event_id.header.size);
7569         if (ret)
7570                 return;
7571 
7572         if (event->ctx->task)
7573                 perf_output_put(&handle, se->event_id.header);
7574         else
7575                 perf_output_put(&handle, se->event_id);
7576 
7577         perf_event__output_id_sample(event, &handle, &sample);
7578 
7579         perf_output_end(&handle);
7580 }
7581 
7582 static void perf_event_switch(struct task_struct *task,
7583                               struct task_struct *next_prev, bool sched_in)
7584 {
7585         struct perf_switch_event switch_event;
7586 
7587         /* N.B. caller checks nr_switch_events != 0 */
7588 
7589         switch_event = (struct perf_switch_event){
7590                 .task           = task,
7591                 .next_prev      = next_prev,
7592                 .event_id       = {
7593                         .header = {
7594                                 /* .type */
7595                                 .misc = sched_in ? 0 : PERF_RECORD_MISC_SWITCH_OUT,
7596                                 /* .size */
7597                         },
7598                         /* .next_prev_pid */
7599                         /* .next_prev_tid */
7600                 },
7601         };
7602 
7603         if (!sched_in && task->state == TASK_RUNNING)
7604                 switch_event.event_id.header.misc |=
7605                                 PERF_RECORD_MISC_SWITCH_OUT_PREEMPT;
7606 
7607         perf_iterate_sb(perf_event_switch_output,
7608                        &switch_event,
7609                        NULL);
7610 }
7611 
7612 /*
7613  * IRQ throttle logging
7614  */
7615 
7616 static void perf_log_throttle(struct perf_event *event, int enable)
7617 {
7618         struct perf_output_handle handle;
7619         struct perf_sample_data sample;
7620         int ret;
7621 
7622         struct {
7623                 struct perf_event_header        header;
7624                 u64                             time;
7625                 u64                             id;
7626                 u64                             stream_id;
7627         } throttle_event = {
7628                 .header = {
7629                         .type = PERF_RECORD_THROTTLE,
7630                         .misc = 0,
7631                         .size = sizeof(throttle_event),
7632                 },
7633                 .time           = perf_event_clock(event),
7634                 .id             = primary_event_id(event),
7635                 .stream_id      = event->id,
7636         };
7637 
7638         if (enable)
7639                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
7640 
7641         perf_event_header__init_id(&throttle_event.header, &sample, event);
7642 
7643         ret = perf_output_begin(&handle, event,
7644                                 throttle_event.header.size);
7645         if (ret)
7646                 return;
7647 
7648         perf_output_put(&handle, throttle_event);
7649         perf_event__output_id_sample(event, &handle, &sample);
7650         perf_output_end(&handle);
7651 }
7652 
7653 void perf_event_itrace_started(struct perf_event *event)
7654 {
7655         event->attach_state |= PERF_ATTACH_ITRACE;
7656 }
7657 
7658 static void perf_log_itrace_start(struct perf_event *event)
7659 {
7660         struct perf_output_handle handle;
7661         struct perf_sample_data sample;
7662         struct perf_aux_event {
7663                 struct perf_event_header        header;
7664                 u32                             pid;
7665                 u32                             tid;
7666         } rec;
7667         int ret;
7668 
7669         if (event->parent)
7670                 event = event->parent;
7671 
7672         if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
7673             event->attach_state & PERF_ATTACH_ITRACE)
7674                 return;
7675 
7676         rec.header.type = PERF_RECORD_ITRACE_START;
7677         rec.header.misc = 0;
7678         rec.header.size = sizeof(rec);
7679         rec.pid = perf_event_pid(event, current);
7680         rec.tid = perf_event_tid(event, current);
7681 
7682         perf_event_header__init_id(&rec.header, &sample, event);
7683         ret = perf_output_begin(&handle, event, rec.header.size);
7684 
7685         if (ret)
7686                 return;
7687 
7688         perf_output_put(&handle, rec);
7689         perf_event__output_id_sample(event, &handle, &sample);
7690 
7691         perf_output_end(&handle);
7692 }
7693 
7694 static int
7695 __perf_event_account_interrupt(struct perf_event *event, int throttle)
7696 {
7697         struct hw_perf_event *hwc = &event->hw;
7698         int ret = 0;
7699         u64 seq;
7700 
7701         seq = __this_cpu_read(perf_throttled_seq);
7702         if (seq != hwc->interrupts_seq) {
7703                 hwc->interrupts_seq = seq;
7704                 hwc->interrupts = 1;
7705         } else {
7706                 hwc->interrupts++;
7707                 if (unlikely(throttle
7708                              && hwc->interrupts >= max_samples_per_tick)) {
7709                         __this_cpu_inc(perf_throttled_count);
7710                         tick_dep_set_cpu(smp_processor_id(), TICK_DEP_BIT_PERF_EVENTS);
7711                         hwc->interrupts = MAX_INTERRUPTS;
7712                         perf_log_throttle(event, 0);
7713                         ret = 1;
7714                 }
7715         }
7716 
7717         if (event->attr.freq) {
7718                 u64 now = perf_clock();
7719                 s64 delta = now - hwc->freq_time_stamp;
7720 
7721                 hwc->freq_time_stamp = now;
7722 
7723                 if (delta > 0 && delta < 2*TICK_NSEC)
7724                         perf_adjust_period(event, delta, hwc->last_period, true);
7725         }
7726 
7727         return ret;
7728 }
7729 
7730 int perf_event_account_interrupt(struct perf_event *event)
7731 {
7732         return __perf_event_account_interrupt(event, 1);
7733 }
7734 
7735 /*
7736  * Generic event overflow handling, sampling.
7737  */
7738 
7739 static int __perf_event_overflow(struct perf_event *event,
7740                                    int throttle, struct perf_sample_data *data,
7741                                    struct pt_regs *regs)
7742 {
7743         int events = atomic_read(&event->event_limit);
7744         int ret = 0;
7745 
7746         /*
7747          * Non-sampling counters might still use the PMI to fold short
7748          * hardware counters, ignore those.
7749          */
7750         if (unlikely(!is_sampling_event(event)))
7751                 return 0;
7752 
7753         ret = __perf_event_account_interrupt(event, throttle);
7754 
7755         /*
7756          * XXX event_limit might not quite work as expected on inherited
7757          * events
7758          */
7759 
7760         event->pending_kill = POLL_IN;
7761         if (events && atomic_dec_and_test(&event->event_limit)) {
7762                 ret = 1;
7763                 event->pending_kill = POLL_HUP;
7764 
7765                 perf_event_disable_inatomic(event);
7766         }
7767 
7768         READ_ONCE(event->overflow_handler)(event, data, regs);
7769 
7770         if (*perf_event_fasync(event) && event->pending_kill) {
7771                 event->pending_wakeup = 1;
7772                 irq_work_queue(&event->pending);
7773         }
7774 
7775         return ret;
7776 }
7777 
7778 int perf_event_overflow(struct perf_event *event,
7779                           struct perf_sample_data *data,
7780                           struct pt_regs *regs)
7781 {
7782         return __perf_event_overflow(event, 1, data, regs);
7783 }
7784 
7785 /*
7786  * Generic software event infrastructure
7787  */
7788 
7789 struct swevent_htable {
7790         struct swevent_hlist            *swevent_hlist;
7791         struct mutex                    hlist_mutex;
7792         int                             hlist_refcount;
7793 
7794         /* Recursion avoidance in each contexts */
7795         int                             recursion[PERF_NR_CONTEXTS];
7796 };
7797 
7798 static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
7799 
7800 /*
7801  * We directly increment event->count and keep a second value in
7802  * event->hw.period_left to count intervals. This period event
7803  * is kept in the range [-sample_period, 0] so that we can use the
7804  * sign as trigger.
7805  */
7806 
7807 u64 perf_swevent_set_period(struct perf_event *event)
7808 {
7809         struct hw_perf_event *hwc = &event->hw;
7810         u64 period = hwc->last_period;
7811         u64 nr, offset;
7812         s64 old, val;
7813 
7814         hwc->last_period = hwc->sample_period;
7815 
7816 again:
7817         old = val = local64_read(&hwc->period_left);
7818         if (val < 0)
7819                 return 0;
7820 
7821         nr = div64_u64(period + val, period);
7822         offset = nr * period;
7823         val -= offset;
7824         if (local64_cmpxchg(&hwc->period_left, old, val) != old)
7825                 goto again;
7826 
7827         return nr;
7828 }
7829 
7830 static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
7831                                     struct perf_sample_data *data,
7832                                     struct pt_regs *regs)
7833 {
7834         struct hw_perf_event *hwc = &event->hw;
7835         int throttle = 0;
7836 
7837         if (!overflow)
7838                 overflow = perf_swevent_set_period(event);
7839 
7840         if (hwc->interrupts == MAX_INTERRUPTS)
7841                 return;
7842 
7843         for (; overflow; overflow--) {
7844                 if (__perf_event_overflow(event, throttle,
7845                                             data, regs)) {
7846                         /*
7847                          * We inhibit the overflow from happening when
7848                          * hwc->interrupts == MAX_INTERRUPTS.
7849                          */
7850                         break;
7851                 }
7852                 throttle = 1;
7853         }
7854 }
7855 
7856 static void perf_swevent_event(struct perf_event *event, u64 nr,
7857                                struct perf_sample_data *data,
7858                                struct pt_regs *regs)
7859 {
7860         struct hw_perf_event *hwc = &event->hw;
7861 
7862         local64_add(nr, &event->count);
7863 
7864         if (!regs)
7865                 return;
7866 
7867         if (!is_sampling_event(event))
7868                 return;
7869 
7870         if ((event->attr.sample_type & PERF_SAMPLE_PERIOD) && !event->attr.freq) {
7871                 data->period = nr;
7872                 return perf_swevent_overflow(event, 1, data, regs);
7873         } else
7874                 data->period = event->hw.last_period;
7875 
7876         if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
7877                 return perf_swevent_overflow(event, 1, data, regs);
7878 
7879         if (local64_add_negative(nr, &hwc->period_left))
7880                 return;
7881 
7882         perf_swevent_overflow(event, 0, data, regs);
7883 }
7884 
7885 static int perf_exclude_event(struct perf_event *event,
7886                               struct pt_regs *regs)
7887 {
7888         if (event->hw.state & PERF_HES_STOPPED)
7889                 return 1;
7890 
7891         if (regs) {
7892                 if (event->attr.exclude_user && user_mode(regs))
7893                         return 1;
7894 
7895                 if (event->attr.exclude_kernel && !user_mode(regs))
7896                         return 1;
7897         }
7898 
7899         return 0;
7900 }
7901 
7902 static int perf_swevent_match(struct perf_event *event,
7903                                 enum perf_type_id type,
7904                                 u32 event_id,
7905                                 struct perf_sample_data *data,
7906                                 struct pt_regs *regs)
7907 {
7908         if (event->attr.type != type)
7909                 return 0;
7910 
7911         if (event->attr.config != event_id)
7912                 return 0;
7913 
7914         if (perf_exclude_event(event, regs))
7915                 return 0;
7916 
7917         return 1;
7918 }
7919 
7920 static inline u64 swevent_hash(u64 type, u32 event_id)
7921 {
7922         u64 val = event_id | (type << 32);
7923 
7924         return hash_64(val, SWEVENT_HLIST_BITS);
7925 }
7926 
7927 static inline struct hlist_head *
7928 __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
7929 {
7930         u64 hash = swevent_hash(type, event_id);
7931 
7932         return &hlist->heads[hash];
7933 }
7934 
7935 /* For the read side: events when they trigger */
7936 static inline struct hlist_head *
7937 find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
7938 {
7939         struct swevent_hlist *hlist;
7940 
7941         hlist = rcu_dereference(swhash->swevent_hlist);
7942         if (!hlist)
7943                 return NULL;
7944 
7945         return __find_swevent_head(hlist, type, event_id);
7946 }
7947 
7948 /* For the event head insertion and removal in the hlist */
7949 static inline struct hlist_head *
7950 find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
7951 {
7952         struct swevent_hlist *hlist;
7953         u32 event_id = event->attr.config;
7954         u64 type = event->attr.type;
7955 
7956         /*
7957          * Event scheduling is always serialized against hlist allocation
7958          * and release. Which makes the protected version suitable here.
7959          * The context lock guarantees that.
7960          */
7961         hlist = rcu_dereference_protected(swhash->swevent_hlist,
7962                                           lockdep_is_held(&event->ctx->lock));
7963         if (!hlist)
7964                 return NULL;
7965 
7966         return __find_swevent_head(hlist, type, event_id);
7967 }
7968 
7969 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
7970                                     u64 nr,
7971                                     struct perf_sample_data *data,
7972                                     struct pt_regs *regs)
7973 {
7974         struct swevent_htable *swhash = this_cpu_ptr(&swevent_htable);
7975         struct perf_event *event;
7976         struct hlist_head *head;
7977 
7978         rcu_read_lock();
7979         head = find_swevent_head_rcu(swhash, type, event_id);
7980         if (!head)