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

TOMOYO Linux Cross Reference
Linux/kernel/perf_event.c

Version: ~ [ linux-5.10-rc6 ] ~ [ linux-5.9.12 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.81 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.161 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.210 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.247 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.247 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Performance events core code:
  3  *
  4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
  5  *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
  6  *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  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/file.h>
 17 #include <linux/poll.h>
 18 #include <linux/sysfs.h>
 19 #include <linux/dcache.h>
 20 #include <linux/percpu.h>
 21 #include <linux/ptrace.h>
 22 #include <linux/vmstat.h>
 23 #include <linux/vmalloc.h>
 24 #include <linux/hardirq.h>
 25 #include <linux/rculist.h>
 26 #include <linux/uaccess.h>
 27 #include <linux/syscalls.h>
 28 #include <linux/anon_inodes.h>
 29 #include <linux/kernel_stat.h>
 30 #include <linux/perf_event.h>
 31 
 32 #include <asm/irq_regs.h>
 33 
 34 /*
 35  * Each CPU has a list of per CPU events:
 36  */
 37 DEFINE_PER_CPU(struct perf_cpu_context, perf_cpu_context);
 38 
 39 int perf_max_events __read_mostly = 1;
 40 static int perf_reserved_percpu __read_mostly;
 41 static int perf_overcommit __read_mostly = 1;
 42 
 43 static atomic_t nr_events __read_mostly;
 44 static atomic_t nr_mmap_events __read_mostly;
 45 static atomic_t nr_comm_events __read_mostly;
 46 static atomic_t nr_task_events __read_mostly;
 47 
 48 /*
 49  * perf event paranoia level:
 50  *  -1 - not paranoid at all
 51  *   0 - disallow raw tracepoint access for unpriv
 52  *   1 - disallow cpu events for unpriv
 53  *   2 - disallow kernel profiling for unpriv
 54  */
 55 int sysctl_perf_event_paranoid __read_mostly = 1;
 56 
 57 static inline bool perf_paranoid_tracepoint_raw(void)
 58 {
 59         return sysctl_perf_event_paranoid > -1;
 60 }
 61 
 62 static inline bool perf_paranoid_cpu(void)
 63 {
 64         return sysctl_perf_event_paranoid > 0;
 65 }
 66 
 67 static inline bool perf_paranoid_kernel(void)
 68 {
 69         return sysctl_perf_event_paranoid > 1;
 70 }
 71 
 72 /* Minimum for 128 pages + 1 for the user control page */
 73 int sysctl_perf_event_mlock __read_mostly = 516; /* 'free' kb per user */
 74 
 75 /*
 76  * max perf event sample rate
 77  */
 78 int sysctl_perf_event_sample_rate __read_mostly = 100000;
 79 
 80 static atomic64_t perf_event_id;
 81 
 82 /*
 83  * Lock for (sysadmin-configurable) event reservations:
 84  */
 85 static DEFINE_SPINLOCK(perf_resource_lock);
 86 
 87 /*
 88  * Architecture provided APIs - weak aliases:
 89  */
 90 extern __weak const struct pmu *hw_perf_event_init(struct perf_event *event)
 91 {
 92         return NULL;
 93 }
 94 
 95 void __weak hw_perf_disable(void)               { barrier(); }
 96 void __weak hw_perf_enable(void)                { barrier(); }
 97 
 98 void __weak hw_perf_event_setup(int cpu)        { barrier(); }
 99 void __weak hw_perf_event_setup_online(int cpu) { barrier(); }
100 
101 int __weak
102 hw_perf_group_sched_in(struct perf_event *group_leader,
103                struct perf_cpu_context *cpuctx,
104                struct perf_event_context *ctx, int cpu)
105 {
106         return 0;
107 }
108 
109 void __weak perf_event_print_debug(void)        { }
110 
111 static DEFINE_PER_CPU(int, perf_disable_count);
112 
113 void __perf_disable(void)
114 {
115         __get_cpu_var(perf_disable_count)++;
116 }
117 
118 bool __perf_enable(void)
119 {
120         return !--__get_cpu_var(perf_disable_count);
121 }
122 
123 void perf_disable(void)
124 {
125         __perf_disable();
126         hw_perf_disable();
127 }
128 
129 void perf_enable(void)
130 {
131         if (__perf_enable())
132                 hw_perf_enable();
133 }
134 
135 static void get_ctx(struct perf_event_context *ctx)
136 {
137         WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
138 }
139 
140 static void free_ctx(struct rcu_head *head)
141 {
142         struct perf_event_context *ctx;
143 
144         ctx = container_of(head, struct perf_event_context, rcu_head);
145         kfree(ctx);
146 }
147 
148 static void put_ctx(struct perf_event_context *ctx)
149 {
150         if (atomic_dec_and_test(&ctx->refcount)) {
151                 if (ctx->parent_ctx)
152                         put_ctx(ctx->parent_ctx);
153                 if (ctx->task)
154                         put_task_struct(ctx->task);
155                 call_rcu(&ctx->rcu_head, free_ctx);
156         }
157 }
158 
159 static void unclone_ctx(struct perf_event_context *ctx)
160 {
161         if (ctx->parent_ctx) {
162                 put_ctx(ctx->parent_ctx);
163                 ctx->parent_ctx = NULL;
164         }
165 }
166 
167 /*
168  * If we inherit events we want to return the parent event id
169  * to userspace.
170  */
171 static u64 primary_event_id(struct perf_event *event)
172 {
173         u64 id = event->id;
174 
175         if (event->parent)
176                 id = event->parent->id;
177 
178         return id;
179 }
180 
181 /*
182  * Get the perf_event_context for a task and lock it.
183  * This has to cope with with the fact that until it is locked,
184  * the context could get moved to another task.
185  */
186 static struct perf_event_context *
187 perf_lock_task_context(struct task_struct *task, unsigned long *flags)
188 {
189         struct perf_event_context *ctx;
190 
191         rcu_read_lock();
192  retry:
193         ctx = rcu_dereference(task->perf_event_ctxp);
194         if (ctx) {
195                 /*
196                  * If this context is a clone of another, it might
197                  * get swapped for another underneath us by
198                  * perf_event_task_sched_out, though the
199                  * rcu_read_lock() protects us from any context
200                  * getting freed.  Lock the context and check if it
201                  * got swapped before we could get the lock, and retry
202                  * if so.  If we locked the right context, then it
203                  * can't get swapped on us any more.
204                  */
205                 spin_lock_irqsave(&ctx->lock, *flags);
206                 if (ctx != rcu_dereference(task->perf_event_ctxp)) {
207                         spin_unlock_irqrestore(&ctx->lock, *flags);
208                         goto retry;
209                 }
210 
211                 if (!atomic_inc_not_zero(&ctx->refcount)) {
212                         spin_unlock_irqrestore(&ctx->lock, *flags);
213                         ctx = NULL;
214                 }
215         }
216         rcu_read_unlock();
217         return ctx;
218 }
219 
220 /*
221  * Get the context for a task and increment its pin_count so it
222  * can't get swapped to another task.  This also increments its
223  * reference count so that the context can't get freed.
224  */
225 static struct perf_event_context *perf_pin_task_context(struct task_struct *task)
226 {
227         struct perf_event_context *ctx;
228         unsigned long flags;
229 
230         ctx = perf_lock_task_context(task, &flags);
231         if (ctx) {
232                 ++ctx->pin_count;
233                 spin_unlock_irqrestore(&ctx->lock, flags);
234         }
235         return ctx;
236 }
237 
238 static void perf_unpin_context(struct perf_event_context *ctx)
239 {
240         unsigned long flags;
241 
242         spin_lock_irqsave(&ctx->lock, flags);
243         --ctx->pin_count;
244         spin_unlock_irqrestore(&ctx->lock, flags);
245         put_ctx(ctx);
246 }
247 
248 /*
249  * Add a event from the lists for its context.
250  * Must be called with ctx->mutex and ctx->lock held.
251  */
252 static void
253 list_add_event(struct perf_event *event, struct perf_event_context *ctx)
254 {
255         struct perf_event *group_leader = event->group_leader;
256 
257         /*
258          * Depending on whether it is a standalone or sibling event,
259          * add it straight to the context's event list, or to the group
260          * leader's sibling list:
261          */
262         if (group_leader == event)
263                 list_add_tail(&event->group_entry, &ctx->group_list);
264         else {
265                 list_add_tail(&event->group_entry, &group_leader->sibling_list);
266                 group_leader->nr_siblings++;
267         }
268 
269         list_add_rcu(&event->event_entry, &ctx->event_list);
270         ctx->nr_events++;
271         if (event->attr.inherit_stat)
272                 ctx->nr_stat++;
273 }
274 
275 /*
276  * Remove a event from the lists for its context.
277  * Must be called with ctx->mutex and ctx->lock held.
278  */
279 static void
280 list_del_event(struct perf_event *event, struct perf_event_context *ctx)
281 {
282         struct perf_event *sibling, *tmp;
283 
284         if (list_empty(&event->group_entry))
285                 return;
286         ctx->nr_events--;
287         if (event->attr.inherit_stat)
288                 ctx->nr_stat--;
289 
290         list_del_init(&event->group_entry);
291         list_del_rcu(&event->event_entry);
292 
293         if (event->group_leader != event)
294                 event->group_leader->nr_siblings--;
295 
296         /*
297          * If this was a group event with sibling events then
298          * upgrade the siblings to singleton events by adding them
299          * to the context list directly:
300          */
301         list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
302 
303                 list_move_tail(&sibling->group_entry, &ctx->group_list);
304                 sibling->group_leader = sibling;
305         }
306 }
307 
308 static void
309 event_sched_out(struct perf_event *event,
310                   struct perf_cpu_context *cpuctx,
311                   struct perf_event_context *ctx)
312 {
313         if (event->state != PERF_EVENT_STATE_ACTIVE)
314                 return;
315 
316         event->state = PERF_EVENT_STATE_INACTIVE;
317         if (event->pending_disable) {
318                 event->pending_disable = 0;
319                 event->state = PERF_EVENT_STATE_OFF;
320         }
321         event->tstamp_stopped = ctx->time;
322         event->pmu->disable(event);
323         event->oncpu = -1;
324 
325         if (!is_software_event(event))
326                 cpuctx->active_oncpu--;
327         ctx->nr_active--;
328         if (event->attr.exclusive || !cpuctx->active_oncpu)
329                 cpuctx->exclusive = 0;
330 }
331 
332 static void
333 group_sched_out(struct perf_event *group_event,
334                 struct perf_cpu_context *cpuctx,
335                 struct perf_event_context *ctx)
336 {
337         struct perf_event *event;
338 
339         if (group_event->state != PERF_EVENT_STATE_ACTIVE)
340                 return;
341 
342         event_sched_out(group_event, cpuctx, ctx);
343 
344         /*
345          * Schedule out siblings (if any):
346          */
347         list_for_each_entry(event, &group_event->sibling_list, group_entry)
348                 event_sched_out(event, cpuctx, ctx);
349 
350         if (group_event->attr.exclusive)
351                 cpuctx->exclusive = 0;
352 }
353 
354 /*
355  * Cross CPU call to remove a performance event
356  *
357  * We disable the event on the hardware level first. After that we
358  * remove it from the context list.
359  */
360 static void __perf_event_remove_from_context(void *info)
361 {
362         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
363         struct perf_event *event = info;
364         struct perf_event_context *ctx = event->ctx;
365 
366         /*
367          * If this is a task context, we need to check whether it is
368          * the current task context of this cpu. If not it has been
369          * scheduled out before the smp call arrived.
370          */
371         if (ctx->task && cpuctx->task_ctx != ctx)
372                 return;
373 
374         spin_lock(&ctx->lock);
375         /*
376          * Protect the list operation against NMI by disabling the
377          * events on a global level.
378          */
379         perf_disable();
380 
381         event_sched_out(event, cpuctx, ctx);
382 
383         list_del_event(event, ctx);
384 
385         if (!ctx->task) {
386                 /*
387                  * Allow more per task events with respect to the
388                  * reservation:
389                  */
390                 cpuctx->max_pertask =
391                         min(perf_max_events - ctx->nr_events,
392                             perf_max_events - perf_reserved_percpu);
393         }
394 
395         perf_enable();
396         spin_unlock(&ctx->lock);
397 }
398 
399 
400 /*
401  * Remove the event from a task's (or a CPU's) list of events.
402  *
403  * Must be called with ctx->mutex held.
404  *
405  * CPU events are removed with a smp call. For task events we only
406  * call when the task is on a CPU.
407  *
408  * If event->ctx is a cloned context, callers must make sure that
409  * every task struct that event->ctx->task could possibly point to
410  * remains valid.  This is OK when called from perf_release since
411  * that only calls us on the top-level context, which can't be a clone.
412  * When called from perf_event_exit_task, it's OK because the
413  * context has been detached from its task.
414  */
415 static void perf_event_remove_from_context(struct perf_event *event)
416 {
417         struct perf_event_context *ctx = event->ctx;
418         struct task_struct *task = ctx->task;
419 
420         if (!task) {
421                 /*
422                  * Per cpu events are removed via an smp call and
423                  * the removal is always sucessful.
424                  */
425                 smp_call_function_single(event->cpu,
426                                          __perf_event_remove_from_context,
427                                          event, 1);
428                 return;
429         }
430 
431 retry:
432         task_oncpu_function_call(task, __perf_event_remove_from_context,
433                                  event);
434 
435         spin_lock_irq(&ctx->lock);
436         /*
437          * If the context is active we need to retry the smp call.
438          */
439         if (ctx->nr_active && !list_empty(&event->group_entry)) {
440                 spin_unlock_irq(&ctx->lock);
441                 goto retry;
442         }
443 
444         /*
445          * The lock prevents that this context is scheduled in so we
446          * can remove the event safely, if the call above did not
447          * succeed.
448          */
449         if (!list_empty(&event->group_entry)) {
450                 list_del_event(event, ctx);
451         }
452         spin_unlock_irq(&ctx->lock);
453 }
454 
455 static inline u64 perf_clock(void)
456 {
457         return cpu_clock(smp_processor_id());
458 }
459 
460 /*
461  * Update the record of the current time in a context.
462  */
463 static void update_context_time(struct perf_event_context *ctx)
464 {
465         u64 now = perf_clock();
466 
467         ctx->time += now - ctx->timestamp;
468         ctx->timestamp = now;
469 }
470 
471 /*
472  * Update the total_time_enabled and total_time_running fields for a event.
473  */
474 static void update_event_times(struct perf_event *event)
475 {
476         struct perf_event_context *ctx = event->ctx;
477         u64 run_end;
478 
479         if (event->state < PERF_EVENT_STATE_INACTIVE ||
480             event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
481                 return;
482 
483         event->total_time_enabled = ctx->time - event->tstamp_enabled;
484 
485         if (event->state == PERF_EVENT_STATE_INACTIVE)
486                 run_end = event->tstamp_stopped;
487         else
488                 run_end = ctx->time;
489 
490         event->total_time_running = run_end - event->tstamp_running;
491 }
492 
493 /*
494  * Update total_time_enabled and total_time_running for all events in a group.
495  */
496 static void update_group_times(struct perf_event *leader)
497 {
498         struct perf_event *event;
499 
500         update_event_times(leader);
501         list_for_each_entry(event, &leader->sibling_list, group_entry)
502                 update_event_times(event);
503 }
504 
505 /*
506  * Cross CPU call to disable a performance event
507  */
508 static void __perf_event_disable(void *info)
509 {
510         struct perf_event *event = info;
511         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
512         struct perf_event_context *ctx = event->ctx;
513 
514         /*
515          * If this is a per-task event, need to check whether this
516          * event's task is the current task on this cpu.
517          */
518         if (ctx->task && cpuctx->task_ctx != ctx)
519                 return;
520 
521         spin_lock(&ctx->lock);
522 
523         /*
524          * If the event is on, turn it off.
525          * If it is in error state, leave it in error state.
526          */
527         if (event->state >= PERF_EVENT_STATE_INACTIVE) {
528                 update_context_time(ctx);
529                 update_group_times(event);
530                 if (event == event->group_leader)
531                         group_sched_out(event, cpuctx, ctx);
532                 else
533                         event_sched_out(event, cpuctx, ctx);
534                 event->state = PERF_EVENT_STATE_OFF;
535         }
536 
537         spin_unlock(&ctx->lock);
538 }
539 
540 /*
541  * Disable a event.
542  *
543  * If event->ctx is a cloned context, callers must make sure that
544  * every task struct that event->ctx->task could possibly point to
545  * remains valid.  This condition is satisifed when called through
546  * perf_event_for_each_child or perf_event_for_each because they
547  * hold the top-level event's child_mutex, so any descendant that
548  * goes to exit will block in sync_child_event.
549  * When called from perf_pending_event it's OK because event->ctx
550  * is the current context on this CPU and preemption is disabled,
551  * hence we can't get into perf_event_task_sched_out for this context.
552  */
553 static void perf_event_disable(struct perf_event *event)
554 {
555         struct perf_event_context *ctx = event->ctx;
556         struct task_struct *task = ctx->task;
557 
558         if (!task) {
559                 /*
560                  * Disable the event on the cpu that it's on
561                  */
562                 smp_call_function_single(event->cpu, __perf_event_disable,
563                                          event, 1);
564                 return;
565         }
566 
567  retry:
568         task_oncpu_function_call(task, __perf_event_disable, event);
569 
570         spin_lock_irq(&ctx->lock);
571         /*
572          * If the event is still active, we need to retry the cross-call.
573          */
574         if (event->state == PERF_EVENT_STATE_ACTIVE) {
575                 spin_unlock_irq(&ctx->lock);
576                 goto retry;
577         }
578 
579         /*
580          * Since we have the lock this context can't be scheduled
581          * in, so we can change the state safely.
582          */
583         if (event->state == PERF_EVENT_STATE_INACTIVE) {
584                 update_group_times(event);
585                 event->state = PERF_EVENT_STATE_OFF;
586         }
587 
588         spin_unlock_irq(&ctx->lock);
589 }
590 
591 static int
592 event_sched_in(struct perf_event *event,
593                  struct perf_cpu_context *cpuctx,
594                  struct perf_event_context *ctx,
595                  int cpu)
596 {
597         if (event->state <= PERF_EVENT_STATE_OFF)
598                 return 0;
599 
600         event->state = PERF_EVENT_STATE_ACTIVE;
601         event->oncpu = cpu;     /* TODO: put 'cpu' into cpuctx->cpu */
602         /*
603          * The new state must be visible before we turn it on in the hardware:
604          */
605         smp_wmb();
606 
607         if (event->pmu->enable(event)) {
608                 event->state = PERF_EVENT_STATE_INACTIVE;
609                 event->oncpu = -1;
610                 return -EAGAIN;
611         }
612 
613         event->tstamp_running += ctx->time - event->tstamp_stopped;
614 
615         if (!is_software_event(event))
616                 cpuctx->active_oncpu++;
617         ctx->nr_active++;
618 
619         if (event->attr.exclusive)
620                 cpuctx->exclusive = 1;
621 
622         return 0;
623 }
624 
625 static int
626 group_sched_in(struct perf_event *group_event,
627                struct perf_cpu_context *cpuctx,
628                struct perf_event_context *ctx,
629                int cpu)
630 {
631         struct perf_event *event, *partial_group;
632         int ret;
633 
634         if (group_event->state == PERF_EVENT_STATE_OFF)
635                 return 0;
636 
637         ret = hw_perf_group_sched_in(group_event, cpuctx, ctx, cpu);
638         if (ret)
639                 return ret < 0 ? ret : 0;
640 
641         if (event_sched_in(group_event, cpuctx, ctx, cpu))
642                 return -EAGAIN;
643 
644         /*
645          * Schedule in siblings as one group (if any):
646          */
647         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
648                 if (event_sched_in(event, cpuctx, ctx, cpu)) {
649                         partial_group = event;
650                         goto group_error;
651                 }
652         }
653 
654         return 0;
655 
656 group_error:
657         /*
658          * Groups can be scheduled in as one unit only, so undo any
659          * partial group before returning:
660          */
661         list_for_each_entry(event, &group_event->sibling_list, group_entry) {
662                 if (event == partial_group)
663                         break;
664                 event_sched_out(event, cpuctx, ctx);
665         }
666         event_sched_out(group_event, cpuctx, ctx);
667 
668         return -EAGAIN;
669 }
670 
671 /*
672  * Return 1 for a group consisting entirely of software events,
673  * 0 if the group contains any hardware events.
674  */
675 static int is_software_only_group(struct perf_event *leader)
676 {
677         struct perf_event *event;
678 
679         if (!is_software_event(leader))
680                 return 0;
681 
682         list_for_each_entry(event, &leader->sibling_list, group_entry)
683                 if (!is_software_event(event))
684                         return 0;
685 
686         return 1;
687 }
688 
689 /*
690  * Work out whether we can put this event group on the CPU now.
691  */
692 static int group_can_go_on(struct perf_event *event,
693                            struct perf_cpu_context *cpuctx,
694                            int can_add_hw)
695 {
696         /*
697          * Groups consisting entirely of software events can always go on.
698          */
699         if (is_software_only_group(event))
700                 return 1;
701         /*
702          * If an exclusive group is already on, no other hardware
703          * events can go on.
704          */
705         if (cpuctx->exclusive)
706                 return 0;
707         /*
708          * If this group is exclusive and there are already
709          * events on the CPU, it can't go on.
710          */
711         if (event->attr.exclusive && cpuctx->active_oncpu)
712                 return 0;
713         /*
714          * Otherwise, try to add it if all previous groups were able
715          * to go on.
716          */
717         return can_add_hw;
718 }
719 
720 static void add_event_to_ctx(struct perf_event *event,
721                                struct perf_event_context *ctx)
722 {
723         list_add_event(event, ctx);
724         event->tstamp_enabled = ctx->time;
725         event->tstamp_running = ctx->time;
726         event->tstamp_stopped = ctx->time;
727 }
728 
729 /*
730  * Cross CPU call to install and enable a performance event
731  *
732  * Must be called with ctx->mutex held
733  */
734 static void __perf_install_in_context(void *info)
735 {
736         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
737         struct perf_event *event = info;
738         struct perf_event_context *ctx = event->ctx;
739         struct perf_event *leader = event->group_leader;
740         int cpu = smp_processor_id();
741         int err;
742 
743         /*
744          * If this is a task context, we need to check whether it is
745          * the current task context of this cpu. If not it has been
746          * scheduled out before the smp call arrived.
747          * Or possibly this is the right context but it isn't
748          * on this cpu because it had no events.
749          */
750         if (ctx->task && cpuctx->task_ctx != ctx) {
751                 if (cpuctx->task_ctx || ctx->task != current)
752                         return;
753                 cpuctx->task_ctx = ctx;
754         }
755 
756         spin_lock(&ctx->lock);
757         ctx->is_active = 1;
758         update_context_time(ctx);
759 
760         /*
761          * Protect the list operation against NMI by disabling the
762          * events on a global level. NOP for non NMI based events.
763          */
764         perf_disable();
765 
766         add_event_to_ctx(event, ctx);
767 
768         /*
769          * Don't put the event on if it is disabled or if
770          * it is in a group and the group isn't on.
771          */
772         if (event->state != PERF_EVENT_STATE_INACTIVE ||
773             (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
774                 goto unlock;
775 
776         /*
777          * An exclusive event can't go on if there are already active
778          * hardware events, and no hardware event can go on if there
779          * is already an exclusive event on.
780          */
781         if (!group_can_go_on(event, cpuctx, 1))
782                 err = -EEXIST;
783         else
784                 err = event_sched_in(event, cpuctx, ctx, cpu);
785 
786         if (err) {
787                 /*
788                  * This event couldn't go on.  If it is in a group
789                  * then we have to pull the whole group off.
790                  * If the event group is pinned then put it in error state.
791                  */
792                 if (leader != event)
793                         group_sched_out(leader, cpuctx, ctx);
794                 if (leader->attr.pinned) {
795                         update_group_times(leader);
796                         leader->state = PERF_EVENT_STATE_ERROR;
797                 }
798         }
799 
800         if (!err && !ctx->task && cpuctx->max_pertask)
801                 cpuctx->max_pertask--;
802 
803  unlock:
804         perf_enable();
805 
806         spin_unlock(&ctx->lock);
807 }
808 
809 /*
810  * Attach a performance event to a context
811  *
812  * First we add the event to the list with the hardware enable bit
813  * in event->hw_config cleared.
814  *
815  * If the event is attached to a task which is on a CPU we use a smp
816  * call to enable it in the task context. The task might have been
817  * scheduled away, but we check this in the smp call again.
818  *
819  * Must be called with ctx->mutex held.
820  */
821 static void
822 perf_install_in_context(struct perf_event_context *ctx,
823                         struct perf_event *event,
824                         int cpu)
825 {
826         struct task_struct *task = ctx->task;
827 
828         if (!task) {
829                 /*
830                  * Per cpu events are installed via an smp call and
831                  * the install is always sucessful.
832                  */
833                 smp_call_function_single(cpu, __perf_install_in_context,
834                                          event, 1);
835                 return;
836         }
837 
838 retry:
839         task_oncpu_function_call(task, __perf_install_in_context,
840                                  event);
841 
842         spin_lock_irq(&ctx->lock);
843         /*
844          * we need to retry the smp call.
845          */
846         if (ctx->is_active && list_empty(&event->group_entry)) {
847                 spin_unlock_irq(&ctx->lock);
848                 goto retry;
849         }
850 
851         /*
852          * The lock prevents that this context is scheduled in so we
853          * can add the event safely, if it the call above did not
854          * succeed.
855          */
856         if (list_empty(&event->group_entry))
857                 add_event_to_ctx(event, ctx);
858         spin_unlock_irq(&ctx->lock);
859 }
860 
861 /*
862  * Put a event into inactive state and update time fields.
863  * Enabling the leader of a group effectively enables all
864  * the group members that aren't explicitly disabled, so we
865  * have to update their ->tstamp_enabled also.
866  * Note: this works for group members as well as group leaders
867  * since the non-leader members' sibling_lists will be empty.
868  */
869 static void __perf_event_mark_enabled(struct perf_event *event,
870                                         struct perf_event_context *ctx)
871 {
872         struct perf_event *sub;
873 
874         event->state = PERF_EVENT_STATE_INACTIVE;
875         event->tstamp_enabled = ctx->time - event->total_time_enabled;
876         list_for_each_entry(sub, &event->sibling_list, group_entry)
877                 if (sub->state >= PERF_EVENT_STATE_INACTIVE)
878                         sub->tstamp_enabled =
879                                 ctx->time - sub->total_time_enabled;
880 }
881 
882 /*
883  * Cross CPU call to enable a performance event
884  */
885 static void __perf_event_enable(void *info)
886 {
887         struct perf_event *event = info;
888         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
889         struct perf_event_context *ctx = event->ctx;
890         struct perf_event *leader = event->group_leader;
891         int err;
892 
893         /*
894          * If this is a per-task event, need to check whether this
895          * event's task is the current task on this cpu.
896          */
897         if (ctx->task && cpuctx->task_ctx != ctx) {
898                 if (cpuctx->task_ctx || ctx->task != current)
899                         return;
900                 cpuctx->task_ctx = ctx;
901         }
902 
903         spin_lock(&ctx->lock);
904         ctx->is_active = 1;
905         update_context_time(ctx);
906 
907         if (event->state >= PERF_EVENT_STATE_INACTIVE)
908                 goto unlock;
909         __perf_event_mark_enabled(event, ctx);
910 
911         /*
912          * If the event is in a group and isn't the group leader,
913          * then don't put it on unless the group is on.
914          */
915         if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
916                 goto unlock;
917 
918         if (!group_can_go_on(event, cpuctx, 1)) {
919                 err = -EEXIST;
920         } else {
921                 perf_disable();
922                 if (event == leader)
923                         err = group_sched_in(event, cpuctx, ctx,
924                                              smp_processor_id());
925                 else
926                         err = event_sched_in(event, cpuctx, ctx,
927                                                smp_processor_id());
928                 perf_enable();
929         }
930 
931         if (err) {
932                 /*
933                  * If this event can't go on and it's part of a
934                  * group, then the whole group has to come off.
935                  */
936                 if (leader != event)
937                         group_sched_out(leader, cpuctx, ctx);
938                 if (leader->attr.pinned) {
939                         update_group_times(leader);
940                         leader->state = PERF_EVENT_STATE_ERROR;
941                 }
942         }
943 
944  unlock:
945         spin_unlock(&ctx->lock);
946 }
947 
948 /*
949  * Enable a event.
950  *
951  * If event->ctx is a cloned context, callers must make sure that
952  * every task struct that event->ctx->task could possibly point to
953  * remains valid.  This condition is satisfied when called through
954  * perf_event_for_each_child or perf_event_for_each as described
955  * for perf_event_disable.
956  */
957 static void perf_event_enable(struct perf_event *event)
958 {
959         struct perf_event_context *ctx = event->ctx;
960         struct task_struct *task = ctx->task;
961 
962         if (!task) {
963                 /*
964                  * Enable the event on the cpu that it's on
965                  */
966                 smp_call_function_single(event->cpu, __perf_event_enable,
967                                          event, 1);
968                 return;
969         }
970 
971         spin_lock_irq(&ctx->lock);
972         if (event->state >= PERF_EVENT_STATE_INACTIVE)
973                 goto out;
974 
975         /*
976          * If the event is in error state, clear that first.
977          * That way, if we see the event in error state below, we
978          * know that it has gone back into error state, as distinct
979          * from the task having been scheduled away before the
980          * cross-call arrived.
981          */
982         if (event->state == PERF_EVENT_STATE_ERROR)
983                 event->state = PERF_EVENT_STATE_OFF;
984 
985  retry:
986         spin_unlock_irq(&ctx->lock);
987         task_oncpu_function_call(task, __perf_event_enable, event);
988 
989         spin_lock_irq(&ctx->lock);
990 
991         /*
992          * If the context is active and the event is still off,
993          * we need to retry the cross-call.
994          */
995         if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
996                 goto retry;
997 
998         /*
999          * Since we have the lock this context can't be scheduled
1000          * in, so we can change the state safely.
1001          */
1002         if (event->state == PERF_EVENT_STATE_OFF)
1003                 __perf_event_mark_enabled(event, ctx);
1004 
1005  out:
1006         spin_unlock_irq(&ctx->lock);
1007 }
1008 
1009 static int perf_event_refresh(struct perf_event *event, int refresh)
1010 {
1011         /*
1012          * not supported on inherited events
1013          */
1014         if (event->attr.inherit)
1015                 return -EINVAL;
1016 
1017         atomic_add(refresh, &event->event_limit);
1018         perf_event_enable(event);
1019 
1020         return 0;
1021 }
1022 
1023 void __perf_event_sched_out(struct perf_event_context *ctx,
1024                               struct perf_cpu_context *cpuctx)
1025 {
1026         struct perf_event *event;
1027 
1028         spin_lock(&ctx->lock);
1029         ctx->is_active = 0;
1030         if (likely(!ctx->nr_events))
1031                 goto out;
1032         update_context_time(ctx);
1033 
1034         perf_disable();
1035         if (ctx->nr_active)
1036                 list_for_each_entry(event, &ctx->group_list, group_entry)
1037                         group_sched_out(event, cpuctx, ctx);
1038 
1039         perf_enable();
1040  out:
1041         spin_unlock(&ctx->lock);
1042 }
1043 
1044 /*
1045  * Test whether two contexts are equivalent, i.e. whether they
1046  * have both been cloned from the same version of the same context
1047  * and they both have the same number of enabled events.
1048  * If the number of enabled events is the same, then the set
1049  * of enabled events should be the same, because these are both
1050  * inherited contexts, therefore we can't access individual events
1051  * in them directly with an fd; we can only enable/disable all
1052  * events via prctl, or enable/disable all events in a family
1053  * via ioctl, which will have the same effect on both contexts.
1054  */
1055 static int context_equiv(struct perf_event_context *ctx1,
1056                          struct perf_event_context *ctx2)
1057 {
1058         return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
1059                 && ctx1->parent_gen == ctx2->parent_gen
1060                 && !ctx1->pin_count && !ctx2->pin_count;
1061 }
1062 
1063 static void __perf_event_read(void *event);
1064 
1065 static void __perf_event_sync_stat(struct perf_event *event,
1066                                      struct perf_event *next_event)
1067 {
1068         u64 value;
1069 
1070         if (!event->attr.inherit_stat)
1071                 return;
1072 
1073         /*
1074          * Update the event value, we cannot use perf_event_read()
1075          * because we're in the middle of a context switch and have IRQs
1076          * disabled, which upsets smp_call_function_single(), however
1077          * we know the event must be on the current CPU, therefore we
1078          * don't need to use it.
1079          */
1080         switch (event->state) {
1081         case PERF_EVENT_STATE_ACTIVE:
1082                 __perf_event_read(event);
1083                 break;
1084 
1085         case PERF_EVENT_STATE_INACTIVE:
1086                 update_event_times(event);
1087                 break;
1088 
1089         default:
1090                 break;
1091         }
1092 
1093         /*
1094          * In order to keep per-task stats reliable we need to flip the event
1095          * values when we flip the contexts.
1096          */
1097         value = atomic64_read(&next_event->count);
1098         value = atomic64_xchg(&event->count, value);
1099         atomic64_set(&next_event->count, value);
1100 
1101         swap(event->total_time_enabled, next_event->total_time_enabled);
1102         swap(event->total_time_running, next_event->total_time_running);
1103 
1104         /*
1105          * Since we swizzled the values, update the user visible data too.
1106          */
1107         perf_event_update_userpage(event);
1108         perf_event_update_userpage(next_event);
1109 }
1110 
1111 #define list_next_entry(pos, member) \
1112         list_entry(pos->member.next, typeof(*pos), member)
1113 
1114 static void perf_event_sync_stat(struct perf_event_context *ctx,
1115                                    struct perf_event_context *next_ctx)
1116 {
1117         struct perf_event *event, *next_event;
1118 
1119         if (!ctx->nr_stat)
1120                 return;
1121 
1122         event = list_first_entry(&ctx->event_list,
1123                                    struct perf_event, event_entry);
1124 
1125         next_event = list_first_entry(&next_ctx->event_list,
1126                                         struct perf_event, event_entry);
1127 
1128         while (&event->event_entry != &ctx->event_list &&
1129                &next_event->event_entry != &next_ctx->event_list) {
1130 
1131                 __perf_event_sync_stat(event, next_event);
1132 
1133                 event = list_next_entry(event, event_entry);
1134                 next_event = list_next_entry(next_event, event_entry);
1135         }
1136 }
1137 
1138 /*
1139  * Called from scheduler to remove the events of the current task,
1140  * with interrupts disabled.
1141  *
1142  * We stop each event and update the event value in event->count.
1143  *
1144  * This does not protect us against NMI, but disable()
1145  * sets the disabled bit in the control field of event _before_
1146  * accessing the event control register. If a NMI hits, then it will
1147  * not restart the event.
1148  */
1149 void perf_event_task_sched_out(struct task_struct *task,
1150                                  struct task_struct *next, int cpu)
1151 {
1152         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1153         struct perf_event_context *ctx = task->perf_event_ctxp;
1154         struct perf_event_context *next_ctx;
1155         struct perf_event_context *parent;
1156         struct pt_regs *regs;
1157         int do_switch = 1;
1158 
1159         regs = task_pt_regs(task);
1160         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, regs, 0);
1161 
1162         if (likely(!ctx || !cpuctx->task_ctx))
1163                 return;
1164 
1165         update_context_time(ctx);
1166 
1167         rcu_read_lock();
1168         parent = rcu_dereference(ctx->parent_ctx);
1169         next_ctx = next->perf_event_ctxp;
1170         if (parent && next_ctx &&
1171             rcu_dereference(next_ctx->parent_ctx) == parent) {
1172                 /*
1173                  * Looks like the two contexts are clones, so we might be
1174                  * able to optimize the context switch.  We lock both
1175                  * contexts and check that they are clones under the
1176                  * lock (including re-checking that neither has been
1177                  * uncloned in the meantime).  It doesn't matter which
1178                  * order we take the locks because no other cpu could
1179                  * be trying to lock both of these tasks.
1180                  */
1181                 spin_lock(&ctx->lock);
1182                 spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
1183                 if (context_equiv(ctx, next_ctx)) {
1184                         /*
1185                          * XXX do we need a memory barrier of sorts
1186                          * wrt to rcu_dereference() of perf_event_ctxp
1187                          */
1188                         task->perf_event_ctxp = next_ctx;
1189                         next->perf_event_ctxp = ctx;
1190                         ctx->task = next;
1191                         next_ctx->task = task;
1192                         do_switch = 0;
1193 
1194                         perf_event_sync_stat(ctx, next_ctx);
1195                 }
1196                 spin_unlock(&next_ctx->lock);
1197                 spin_unlock(&ctx->lock);
1198         }
1199         rcu_read_unlock();
1200 
1201         if (do_switch) {
1202                 __perf_event_sched_out(ctx, cpuctx);
1203                 cpuctx->task_ctx = NULL;
1204         }
1205 }
1206 
1207 /*
1208  * Called with IRQs disabled
1209  */
1210 static void __perf_event_task_sched_out(struct perf_event_context *ctx)
1211 {
1212         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1213 
1214         if (!cpuctx->task_ctx)
1215                 return;
1216 
1217         if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
1218                 return;
1219 
1220         __perf_event_sched_out(ctx, cpuctx);
1221         cpuctx->task_ctx = NULL;
1222 }
1223 
1224 /*
1225  * Called with IRQs disabled
1226  */
1227 static void perf_event_cpu_sched_out(struct perf_cpu_context *cpuctx)
1228 {
1229         __perf_event_sched_out(&cpuctx->ctx, cpuctx);
1230 }
1231 
1232 static void
1233 __perf_event_sched_in(struct perf_event_context *ctx,
1234                         struct perf_cpu_context *cpuctx, int cpu)
1235 {
1236         struct perf_event *event;
1237         int can_add_hw = 1;
1238 
1239         spin_lock(&ctx->lock);
1240         ctx->is_active = 1;
1241         if (likely(!ctx->nr_events))
1242                 goto out;
1243 
1244         ctx->timestamp = perf_clock();
1245 
1246         perf_disable();
1247 
1248         /*
1249          * First go through the list and put on any pinned groups
1250          * in order to give them the best chance of going on.
1251          */
1252         list_for_each_entry(event, &ctx->group_list, group_entry) {
1253                 if (event->state <= PERF_EVENT_STATE_OFF ||
1254                     !event->attr.pinned)
1255                         continue;
1256                 if (event->cpu != -1 && event->cpu != cpu)
1257                         continue;
1258 
1259                 if (group_can_go_on(event, cpuctx, 1))
1260                         group_sched_in(event, cpuctx, ctx, cpu);
1261 
1262                 /*
1263                  * If this pinned group hasn't been scheduled,
1264                  * put it in error state.
1265                  */
1266                 if (event->state == PERF_EVENT_STATE_INACTIVE) {
1267                         update_group_times(event);
1268                         event->state = PERF_EVENT_STATE_ERROR;
1269                 }
1270         }
1271 
1272         list_for_each_entry(event, &ctx->group_list, group_entry) {
1273                 /*
1274                  * Ignore events in OFF or ERROR state, and
1275                  * ignore pinned events since we did them already.
1276                  */
1277                 if (event->state <= PERF_EVENT_STATE_OFF ||
1278                     event->attr.pinned)
1279                         continue;
1280 
1281                 /*
1282                  * Listen to the 'cpu' scheduling filter constraint
1283                  * of events:
1284                  */
1285                 if (event->cpu != -1 && event->cpu != cpu)
1286                         continue;
1287 
1288                 if (group_can_go_on(event, cpuctx, can_add_hw))
1289                         if (group_sched_in(event, cpuctx, ctx, cpu))
1290                                 can_add_hw = 0;
1291         }
1292         perf_enable();
1293  out:
1294         spin_unlock(&ctx->lock);
1295 }
1296 
1297 /*
1298  * Called from scheduler to add the events of the current task
1299  * with interrupts disabled.
1300  *
1301  * We restore the event value and then enable it.
1302  *
1303  * This does not protect us against NMI, but enable()
1304  * sets the enabled bit in the control field of event _before_
1305  * accessing the event control register. If a NMI hits, then it will
1306  * keep the event running.
1307  */
1308 void perf_event_task_sched_in(struct task_struct *task, int cpu)
1309 {
1310         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
1311         struct perf_event_context *ctx = task->perf_event_ctxp;
1312 
1313         if (likely(!ctx))
1314                 return;
1315         if (cpuctx->task_ctx == ctx)
1316                 return;
1317         __perf_event_sched_in(ctx, cpuctx, cpu);
1318         cpuctx->task_ctx = ctx;
1319 }
1320 
1321 static void perf_event_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
1322 {
1323         struct perf_event_context *ctx = &cpuctx->ctx;
1324 
1325         __perf_event_sched_in(ctx, cpuctx, cpu);
1326 }
1327 
1328 #define MAX_INTERRUPTS (~0ULL)
1329 
1330 static void perf_log_throttle(struct perf_event *event, int enable);
1331 
1332 static void perf_adjust_period(struct perf_event *event, u64 events)
1333 {
1334         struct hw_perf_event *hwc = &event->hw;
1335         u64 period, sample_period;
1336         s64 delta;
1337 
1338         events *= hwc->sample_period;
1339         period = div64_u64(events, event->attr.sample_freq);
1340 
1341         delta = (s64)(period - hwc->sample_period);
1342         delta = (delta + 7) / 8; /* low pass filter */
1343 
1344         sample_period = hwc->sample_period + delta;
1345 
1346         if (!sample_period)
1347                 sample_period = 1;
1348 
1349         hwc->sample_period = sample_period;
1350 }
1351 
1352 static void perf_ctx_adjust_freq(struct perf_event_context *ctx)
1353 {
1354         struct perf_event *event;
1355         struct hw_perf_event *hwc;
1356         u64 interrupts, freq;
1357 
1358         spin_lock(&ctx->lock);
1359         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
1360                 if (event->state != PERF_EVENT_STATE_ACTIVE)
1361                         continue;
1362 
1363                 if (event->cpu != -1 && event->cpu != smp_processor_id())
1364                         continue;
1365 
1366                 hwc = &event->hw;
1367 
1368                 interrupts = hwc->interrupts;
1369                 hwc->interrupts = 0;
1370 
1371                 /*
1372                  * unthrottle events on the tick
1373                  */
1374                 if (interrupts == MAX_INTERRUPTS) {
1375                         perf_log_throttle(event, 1);
1376                         event->pmu->unthrottle(event);
1377                         interrupts = 2*sysctl_perf_event_sample_rate/HZ;
1378                 }
1379 
1380                 if (!event->attr.freq || !event->attr.sample_freq)
1381                         continue;
1382 
1383                 /*
1384                  * if the specified freq < HZ then we need to skip ticks
1385                  */
1386                 if (event->attr.sample_freq < HZ) {
1387                         freq = event->attr.sample_freq;
1388 
1389                         hwc->freq_count += freq;
1390                         hwc->freq_interrupts += interrupts;
1391 
1392                         if (hwc->freq_count < HZ)
1393                                 continue;
1394 
1395                         interrupts = hwc->freq_interrupts;
1396                         hwc->freq_interrupts = 0;
1397                         hwc->freq_count -= HZ;
1398                 } else
1399                         freq = HZ;
1400 
1401                 perf_adjust_period(event, freq * interrupts);
1402 
1403                 /*
1404                  * In order to avoid being stalled by an (accidental) huge
1405                  * sample period, force reset the sample period if we didn't
1406                  * get any events in this freq period.
1407                  */
1408                 if (!interrupts) {
1409                         perf_disable();
1410                         event->pmu->disable(event);
1411                         atomic64_set(&hwc->period_left, 0);
1412                         event->pmu->enable(event);
1413                         perf_enable();
1414                 }
1415         }
1416         spin_unlock(&ctx->lock);
1417 }
1418 
1419 /*
1420  * Round-robin a context's events:
1421  */
1422 static void rotate_ctx(struct perf_event_context *ctx)
1423 {
1424         struct perf_event *event;
1425 
1426         if (!ctx->nr_events)
1427                 return;
1428 
1429         spin_lock(&ctx->lock);
1430         /*
1431          * Rotate the first entry last (works just fine for group events too):
1432          */
1433         perf_disable();
1434         list_for_each_entry(event, &ctx->group_list, group_entry) {
1435                 list_move_tail(&event->group_entry, &ctx->group_list);
1436                 break;
1437         }
1438         perf_enable();
1439 
1440         spin_unlock(&ctx->lock);
1441 }
1442 
1443 void perf_event_task_tick(struct task_struct *curr, int cpu)
1444 {
1445         struct perf_cpu_context *cpuctx;
1446         struct perf_event_context *ctx;
1447 
1448         if (!atomic_read(&nr_events))
1449                 return;
1450 
1451         cpuctx = &per_cpu(perf_cpu_context, cpu);
1452         ctx = curr->perf_event_ctxp;
1453 
1454         perf_ctx_adjust_freq(&cpuctx->ctx);
1455         if (ctx)
1456                 perf_ctx_adjust_freq(ctx);
1457 
1458         perf_event_cpu_sched_out(cpuctx);
1459         if (ctx)
1460                 __perf_event_task_sched_out(ctx);
1461 
1462         rotate_ctx(&cpuctx->ctx);
1463         if (ctx)
1464                 rotate_ctx(ctx);
1465 
1466         perf_event_cpu_sched_in(cpuctx, cpu);
1467         if (ctx)
1468                 perf_event_task_sched_in(curr, cpu);
1469 }
1470 
1471 /*
1472  * Enable all of a task's events that have been marked enable-on-exec.
1473  * This expects task == current.
1474  */
1475 static void perf_event_enable_on_exec(struct task_struct *task)
1476 {
1477         struct perf_event_context *ctx;
1478         struct perf_event *event;
1479         unsigned long flags;
1480         int enabled = 0;
1481 
1482         local_irq_save(flags);
1483         ctx = task->perf_event_ctxp;
1484         if (!ctx || !ctx->nr_events)
1485                 goto out;
1486 
1487         __perf_event_task_sched_out(ctx);
1488 
1489         spin_lock(&ctx->lock);
1490 
1491         list_for_each_entry(event, &ctx->group_list, group_entry) {
1492                 if (!event->attr.enable_on_exec)
1493                         continue;
1494                 event->attr.enable_on_exec = 0;
1495                 if (event->state >= PERF_EVENT_STATE_INACTIVE)
1496                         continue;
1497                 __perf_event_mark_enabled(event, ctx);
1498                 enabled = 1;
1499         }
1500 
1501         /*
1502          * Unclone this context if we enabled any event.
1503          */
1504         if (enabled)
1505                 unclone_ctx(ctx);
1506 
1507         spin_unlock(&ctx->lock);
1508 
1509         perf_event_task_sched_in(task, smp_processor_id());
1510  out:
1511         local_irq_restore(flags);
1512 }
1513 
1514 /*
1515  * Cross CPU call to read the hardware event
1516  */
1517 static void __perf_event_read(void *info)
1518 {
1519         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
1520         struct perf_event *event = info;
1521         struct perf_event_context *ctx = event->ctx;
1522         unsigned long flags;
1523 
1524         /*
1525          * If this is a task context, we need to check whether it is
1526          * the current task context of this cpu.  If not it has been
1527          * scheduled out before the smp call arrived.  In that case
1528          * event->count would have been updated to a recent sample
1529          * when the event was scheduled out.
1530          */
1531         if (ctx->task && cpuctx->task_ctx != ctx)
1532                 return;
1533 
1534         local_irq_save(flags);
1535         if (ctx->is_active)
1536                 update_context_time(ctx);
1537         event->pmu->read(event);
1538         update_event_times(event);
1539         local_irq_restore(flags);
1540 }
1541 
1542 static u64 perf_event_read(struct perf_event *event)
1543 {
1544         /*
1545          * If event is enabled and currently active on a CPU, update the
1546          * value in the event structure:
1547          */
1548         if (event->state == PERF_EVENT_STATE_ACTIVE) {
1549                 smp_call_function_single(event->oncpu,
1550                                          __perf_event_read, event, 1);
1551         } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
1552                 update_event_times(event);
1553         }
1554 
1555         return atomic64_read(&event->count);
1556 }
1557 
1558 /*
1559  * Initialize the perf_event context in a task_struct:
1560  */
1561 static void
1562 __perf_event_init_context(struct perf_event_context *ctx,
1563                             struct task_struct *task)
1564 {
1565         memset(ctx, 0, sizeof(*ctx));
1566         spin_lock_init(&ctx->lock);
1567         mutex_init(&ctx->mutex);
1568         INIT_LIST_HEAD(&ctx->group_list);
1569         INIT_LIST_HEAD(&ctx->event_list);
1570         atomic_set(&ctx->refcount, 1);
1571         ctx->task = task;
1572 }
1573 
1574 static struct perf_event_context *find_get_context(pid_t pid, int cpu)
1575 {
1576         struct perf_event_context *ctx;
1577         struct perf_cpu_context *cpuctx;
1578         struct task_struct *task;
1579         unsigned long flags;
1580         int err;
1581 
1582         /*
1583          * If cpu is not a wildcard then this is a percpu event:
1584          */
1585         if (cpu != -1) {
1586                 /* Must be root to operate on a CPU event: */
1587                 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1588                         return ERR_PTR(-EACCES);
1589 
1590                 if (cpu < 0 || cpu >= nr_cpumask_bits)
1591                         return ERR_PTR(-EINVAL);
1592 
1593                 /*
1594                  * We could be clever and allow to attach a event to an
1595                  * offline CPU and activate it when the CPU comes up, but
1596                  * that's for later.
1597                  */
1598                 if (!cpu_isset(cpu, cpu_online_map))
1599                         return ERR_PTR(-ENODEV);
1600 
1601                 cpuctx = &per_cpu(perf_cpu_context, cpu);
1602                 ctx = &cpuctx->ctx;
1603                 get_ctx(ctx);
1604 
1605                 return ctx;
1606         }
1607 
1608         rcu_read_lock();
1609         if (!pid)
1610                 task = current;
1611         else
1612                 task = find_task_by_vpid(pid);
1613         if (task)
1614                 get_task_struct(task);
1615         rcu_read_unlock();
1616 
1617         if (!task)
1618                 return ERR_PTR(-ESRCH);
1619 
1620         /*
1621          * Can't attach events to a dying task.
1622          */
1623         err = -ESRCH;
1624         if (task->flags & PF_EXITING)
1625                 goto errout;
1626 
1627         /* Reuse ptrace permission checks for now. */
1628         err = -EACCES;
1629         if (!ptrace_may_access(task, PTRACE_MODE_READ))
1630                 goto errout;
1631 
1632  retry:
1633         ctx = perf_lock_task_context(task, &flags);
1634         if (ctx) {
1635                 unclone_ctx(ctx);
1636                 spin_unlock_irqrestore(&ctx->lock, flags);
1637         }
1638 
1639         if (!ctx) {
1640                 ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
1641                 err = -ENOMEM;
1642                 if (!ctx)
1643                         goto errout;
1644                 __perf_event_init_context(ctx, task);
1645                 get_ctx(ctx);
1646                 if (cmpxchg(&task->perf_event_ctxp, NULL, ctx)) {
1647                         /*
1648                          * We raced with some other task; use
1649                          * the context they set.
1650                          */
1651                         kfree(ctx);
1652                         goto retry;
1653                 }
1654                 get_task_struct(task);
1655         }
1656 
1657         put_task_struct(task);
1658         return ctx;
1659 
1660  errout:
1661         put_task_struct(task);
1662         return ERR_PTR(err);
1663 }
1664 
1665 static void free_event_rcu(struct rcu_head *head)
1666 {
1667         struct perf_event *event;
1668 
1669         event = container_of(head, struct perf_event, rcu_head);
1670         if (event->ns)
1671                 put_pid_ns(event->ns);
1672         kfree(event);
1673 }
1674 
1675 static void perf_pending_sync(struct perf_event *event);
1676 
1677 static void free_event(struct perf_event *event)
1678 {
1679         perf_pending_sync(event);
1680 
1681         if (!event->parent) {
1682                 atomic_dec(&nr_events);
1683                 if (event->attr.mmap)
1684                         atomic_dec(&nr_mmap_events);
1685                 if (event->attr.comm)
1686                         atomic_dec(&nr_comm_events);
1687                 if (event->attr.task)
1688                         atomic_dec(&nr_task_events);
1689         }
1690 
1691         if (event->output) {
1692                 fput(event->output->filp);
1693                 event->output = NULL;
1694         }
1695 
1696         if (event->destroy)
1697                 event->destroy(event);
1698 
1699         put_ctx(event->ctx);
1700         call_rcu(&event->rcu_head, free_event_rcu);
1701 }
1702 
1703 /*
1704  * Called when the last reference to the file is gone.
1705  */
1706 static int perf_release(struct inode *inode, struct file *file)
1707 {
1708         struct perf_event *event = file->private_data;
1709         struct perf_event_context *ctx = event->ctx;
1710 
1711         file->private_data = NULL;
1712 
1713         WARN_ON_ONCE(ctx->parent_ctx);
1714         mutex_lock(&ctx->mutex);
1715         perf_event_remove_from_context(event);
1716         mutex_unlock(&ctx->mutex);
1717 
1718         mutex_lock(&event->owner->perf_event_mutex);
1719         list_del_init(&event->owner_entry);
1720         mutex_unlock(&event->owner->perf_event_mutex);
1721         put_task_struct(event->owner);
1722 
1723         free_event(event);
1724 
1725         return 0;
1726 }
1727 
1728 static int perf_event_read_size(struct perf_event *event)
1729 {
1730         int entry = sizeof(u64); /* value */
1731         int size = 0;
1732         int nr = 1;
1733 
1734         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1735                 size += sizeof(u64);
1736 
1737         if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1738                 size += sizeof(u64);
1739 
1740         if (event->attr.read_format & PERF_FORMAT_ID)
1741                 entry += sizeof(u64);
1742 
1743         if (event->attr.read_format & PERF_FORMAT_GROUP) {
1744                 nr += event->group_leader->nr_siblings;
1745                 size += sizeof(u64);
1746         }
1747 
1748         size += entry * nr;
1749 
1750         return size;
1751 }
1752 
1753 static u64 perf_event_read_value(struct perf_event *event)
1754 {
1755         struct perf_event *child;
1756         u64 total = 0;
1757 
1758         total += perf_event_read(event);
1759         list_for_each_entry(child, &event->child_list, child_list)
1760                 total += perf_event_read(child);
1761 
1762         return total;
1763 }
1764 
1765 static int perf_event_read_entry(struct perf_event *event,
1766                                    u64 read_format, char __user *buf)
1767 {
1768         int n = 0, count = 0;
1769         u64 values[2];
1770 
1771         values[n++] = perf_event_read_value(event);
1772         if (read_format & PERF_FORMAT_ID)
1773                 values[n++] = primary_event_id(event);
1774 
1775         count = n * sizeof(u64);
1776 
1777         if (copy_to_user(buf, values, count))
1778                 return -EFAULT;
1779 
1780         return count;
1781 }
1782 
1783 static int perf_event_read_group(struct perf_event *event,
1784                                    u64 read_format, char __user *buf)
1785 {
1786         struct perf_event *leader = event->group_leader, *sub;
1787         int n = 0, size = 0, err = -EFAULT;
1788         u64 values[3];
1789 
1790         values[n++] = 1 + leader->nr_siblings;
1791         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1792                 values[n++] = leader->total_time_enabled +
1793                         atomic64_read(&leader->child_total_time_enabled);
1794         }
1795         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1796                 values[n++] = leader->total_time_running +
1797                         atomic64_read(&leader->child_total_time_running);
1798         }
1799 
1800         size = n * sizeof(u64);
1801 
1802         if (copy_to_user(buf, values, size))
1803                 return -EFAULT;
1804 
1805         err = perf_event_read_entry(leader, read_format, buf + size);
1806         if (err < 0)
1807                 return err;
1808 
1809         size += err;
1810 
1811         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
1812                 err = perf_event_read_entry(sub, read_format,
1813                                 buf + size);
1814                 if (err < 0)
1815                         return err;
1816 
1817                 size += err;
1818         }
1819 
1820         return size;
1821 }
1822 
1823 static int perf_event_read_one(struct perf_event *event,
1824                                  u64 read_format, char __user *buf)
1825 {
1826         u64 values[4];
1827         int n = 0;
1828 
1829         values[n++] = perf_event_read_value(event);
1830         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1831                 values[n++] = event->total_time_enabled +
1832                         atomic64_read(&event->child_total_time_enabled);
1833         }
1834         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1835                 values[n++] = event->total_time_running +
1836                         atomic64_read(&event->child_total_time_running);
1837         }
1838         if (read_format & PERF_FORMAT_ID)
1839                 values[n++] = primary_event_id(event);
1840 
1841         if (copy_to_user(buf, values, n * sizeof(u64)))
1842                 return -EFAULT;
1843 
1844         return n * sizeof(u64);
1845 }
1846 
1847 /*
1848  * Read the performance event - simple non blocking version for now
1849  */
1850 static ssize_t
1851 perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
1852 {
1853         u64 read_format = event->attr.read_format;
1854         int ret;
1855 
1856         /*
1857          * Return end-of-file for a read on a event that is in
1858          * error state (i.e. because it was pinned but it couldn't be
1859          * scheduled on to the CPU at some point).
1860          */
1861         if (event->state == PERF_EVENT_STATE_ERROR)
1862                 return 0;
1863 
1864         if (count < perf_event_read_size(event))
1865                 return -ENOSPC;
1866 
1867         WARN_ON_ONCE(event->ctx->parent_ctx);
1868         mutex_lock(&event->child_mutex);
1869         if (read_format & PERF_FORMAT_GROUP)
1870                 ret = perf_event_read_group(event, read_format, buf);
1871         else
1872                 ret = perf_event_read_one(event, read_format, buf);
1873         mutex_unlock(&event->child_mutex);
1874 
1875         return ret;
1876 }
1877 
1878 static ssize_t
1879 perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
1880 {
1881         struct perf_event *event = file->private_data;
1882 
1883         return perf_read_hw(event, buf, count);
1884 }
1885 
1886 static unsigned int perf_poll(struct file *file, poll_table *wait)
1887 {
1888         struct perf_event *event = file->private_data;
1889         struct perf_mmap_data *data;
1890         unsigned int events = POLL_HUP;
1891 
1892         rcu_read_lock();
1893         data = rcu_dereference(event->data);
1894         if (data)
1895                 events = atomic_xchg(&data->poll, 0);
1896         rcu_read_unlock();
1897 
1898         poll_wait(file, &event->waitq, wait);
1899 
1900         return events;
1901 }
1902 
1903 static void perf_event_reset(struct perf_event *event)
1904 {
1905         (void)perf_event_read(event);
1906         atomic64_set(&event->count, 0);
1907         perf_event_update_userpage(event);
1908 }
1909 
1910 /*
1911  * Holding the top-level event's child_mutex means that any
1912  * descendant process that has inherited this event will block
1913  * in sync_child_event if it goes to exit, thus satisfying the
1914  * task existence requirements of perf_event_enable/disable.
1915  */
1916 static void perf_event_for_each_child(struct perf_event *event,
1917                                         void (*func)(struct perf_event *))
1918 {
1919         struct perf_event *child;
1920 
1921         WARN_ON_ONCE(event->ctx->parent_ctx);
1922         mutex_lock(&event->child_mutex);
1923         func(event);
1924         list_for_each_entry(child, &event->child_list, child_list)
1925                 func(child);
1926         mutex_unlock(&event->child_mutex);
1927 }
1928 
1929 static void perf_event_for_each(struct perf_event *event,
1930                                   void (*func)(struct perf_event *))
1931 {
1932         struct perf_event_context *ctx = event->ctx;
1933         struct perf_event *sibling;
1934 
1935         WARN_ON_ONCE(ctx->parent_ctx);
1936         mutex_lock(&ctx->mutex);
1937         event = event->group_leader;
1938 
1939         perf_event_for_each_child(event, func);
1940         func(event);
1941         list_for_each_entry(sibling, &event->sibling_list, group_entry)
1942                 perf_event_for_each_child(event, func);
1943         mutex_unlock(&ctx->mutex);
1944 }
1945 
1946 static int perf_event_period(struct perf_event *event, u64 __user *arg)
1947 {
1948         struct perf_event_context *ctx = event->ctx;
1949         unsigned long size;
1950         int ret = 0;
1951         u64 value;
1952 
1953         if (!event->attr.sample_period)
1954                 return -EINVAL;
1955 
1956         size = copy_from_user(&value, arg, sizeof(value));
1957         if (size != sizeof(value))
1958                 return -EFAULT;
1959 
1960         if (!value)
1961                 return -EINVAL;
1962 
1963         spin_lock_irq(&ctx->lock);
1964         if (event->attr.freq) {
1965                 if (value > sysctl_perf_event_sample_rate) {
1966                         ret = -EINVAL;
1967                         goto unlock;
1968                 }
1969 
1970                 event->attr.sample_freq = value;
1971         } else {
1972                 event->attr.sample_period = value;
1973                 event->hw.sample_period = value;
1974         }
1975 unlock:
1976         spin_unlock_irq(&ctx->lock);
1977 
1978         return ret;
1979 }
1980 
1981 int perf_event_set_output(struct perf_event *event, int output_fd);
1982 
1983 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1984 {
1985         struct perf_event *event = file->private_data;
1986         void (*func)(struct perf_event *);
1987         u32 flags = arg;
1988 
1989         switch (cmd) {
1990         case PERF_EVENT_IOC_ENABLE:
1991                 func = perf_event_enable;
1992                 break;
1993         case PERF_EVENT_IOC_DISABLE:
1994                 func = perf_event_disable;
1995                 break;
1996         case PERF_EVENT_IOC_RESET:
1997                 func = perf_event_reset;
1998                 break;
1999 
2000         case PERF_EVENT_IOC_REFRESH:
2001                 return perf_event_refresh(event, arg);
2002 
2003         case PERF_EVENT_IOC_PERIOD:
2004                 return perf_event_period(event, (u64 __user *)arg);
2005 
2006         case PERF_EVENT_IOC_SET_OUTPUT:
2007                 return perf_event_set_output(event, arg);
2008 
2009         default:
2010                 return -ENOTTY;
2011         }
2012 
2013         if (flags & PERF_IOC_FLAG_GROUP)
2014                 perf_event_for_each(event, func);
2015         else
2016                 perf_event_for_each_child(event, func);
2017 
2018         return 0;
2019 }
2020 
2021 int perf_event_task_enable(void)
2022 {
2023         struct perf_event *event;
2024 
2025         mutex_lock(&current->perf_event_mutex);
2026         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2027                 perf_event_for_each_child(event, perf_event_enable);
2028         mutex_unlock(&current->perf_event_mutex);
2029 
2030         return 0;
2031 }
2032 
2033 int perf_event_task_disable(void)
2034 {
2035         struct perf_event *event;
2036 
2037         mutex_lock(&current->perf_event_mutex);
2038         list_for_each_entry(event, &current->perf_event_list, owner_entry)
2039                 perf_event_for_each_child(event, perf_event_disable);
2040         mutex_unlock(&current->perf_event_mutex);
2041 
2042         return 0;
2043 }
2044 
2045 #ifndef PERF_EVENT_INDEX_OFFSET
2046 # define PERF_EVENT_INDEX_OFFSET 0
2047 #endif
2048 
2049 static int perf_event_index(struct perf_event *event)
2050 {
2051         if (event->state != PERF_EVENT_STATE_ACTIVE)
2052                 return 0;
2053 
2054         return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
2055 }
2056 
2057 /*
2058  * Callers need to ensure there can be no nesting of this function, otherwise
2059  * the seqlock logic goes bad. We can not serialize this because the arch
2060  * code calls this from NMI context.
2061  */
2062 void perf_event_update_userpage(struct perf_event *event)
2063 {
2064         struct perf_event_mmap_page *userpg;
2065         struct perf_mmap_data *data;
2066 
2067         rcu_read_lock();
2068         data = rcu_dereference(event->data);
2069         if (!data)
2070                 goto unlock;
2071 
2072         userpg = data->user_page;
2073 
2074         /*
2075          * Disable preemption so as to not let the corresponding user-space
2076          * spin too long if we get preempted.
2077          */
2078         preempt_disable();
2079         ++userpg->lock;
2080         barrier();
2081         userpg->index = perf_event_index(event);
2082         userpg->offset = atomic64_read(&event->count);
2083         if (event->state == PERF_EVENT_STATE_ACTIVE)
2084                 userpg->offset -= atomic64_read(&event->hw.prev_count);
2085 
2086         userpg->time_enabled = event->total_time_enabled +
2087                         atomic64_read(&event->child_total_time_enabled);
2088 
2089         userpg->time_running = event->total_time_running +
2090                         atomic64_read(&event->child_total_time_running);
2091 
2092         barrier();
2093         ++userpg->lock;
2094         preempt_enable();
2095 unlock:
2096         rcu_read_unlock();
2097 }
2098 
2099 static unsigned long perf_data_size(struct perf_mmap_data *data)
2100 {
2101         return data->nr_pages << (PAGE_SHIFT + data->data_order);
2102 }
2103 
2104 #ifndef CONFIG_PERF_USE_VMALLOC
2105 
2106 /*
2107  * Back perf_mmap() with regular GFP_KERNEL-0 pages.
2108  */
2109 
2110 static struct page *
2111 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2112 {
2113         if (pgoff > data->nr_pages)
2114                 return NULL;
2115 
2116         if (pgoff == 0)
2117                 return virt_to_page(data->user_page);
2118 
2119         return virt_to_page(data->data_pages[pgoff - 1]);
2120 }
2121 
2122 static struct perf_mmap_data *
2123 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2124 {
2125         struct perf_mmap_data *data;
2126         unsigned long size;
2127         int i;
2128 
2129         WARN_ON(atomic_read(&event->mmap_count));
2130 
2131         size = sizeof(struct perf_mmap_data);
2132         size += nr_pages * sizeof(void *);
2133 
2134         data = kzalloc(size, GFP_KERNEL);
2135         if (!data)
2136                 goto fail;
2137 
2138         data->user_page = (void *)get_zeroed_page(GFP_KERNEL);
2139         if (!data->user_page)
2140                 goto fail_user_page;
2141 
2142         for (i = 0; i < nr_pages; i++) {
2143                 data->data_pages[i] = (void *)get_zeroed_page(GFP_KERNEL);
2144                 if (!data->data_pages[i])
2145                         goto fail_data_pages;
2146         }
2147 
2148         data->data_order = 0;
2149         data->nr_pages = nr_pages;
2150 
2151         return data;
2152 
2153 fail_data_pages:
2154         for (i--; i >= 0; i--)
2155                 free_page((unsigned long)data->data_pages[i]);
2156 
2157         free_page((unsigned long)data->user_page);
2158 
2159 fail_user_page:
2160         kfree(data);
2161 
2162 fail:
2163         return NULL;
2164 }
2165 
2166 static void perf_mmap_free_page(unsigned long addr)
2167 {
2168         struct page *page = virt_to_page((void *)addr);
2169 
2170         page->mapping = NULL;
2171         __free_page(page);
2172 }
2173 
2174 static void perf_mmap_data_free(struct perf_mmap_data *data)
2175 {
2176         int i;
2177 
2178         perf_mmap_free_page((unsigned long)data->user_page);
2179         for (i = 0; i < data->nr_pages; i++)
2180                 perf_mmap_free_page((unsigned long)data->data_pages[i]);
2181         kfree(data);
2182 }
2183 
2184 #else
2185 
2186 /*
2187  * Back perf_mmap() with vmalloc memory.
2188  *
2189  * Required for architectures that have d-cache aliasing issues.
2190  */
2191 
2192 static struct page *
2193 perf_mmap_to_page(struct perf_mmap_data *data, unsigned long pgoff)
2194 {
2195         if (pgoff > (1UL << data->data_order))
2196                 return NULL;
2197 
2198         return vmalloc_to_page((void *)data->user_page + pgoff * PAGE_SIZE);
2199 }
2200 
2201 static void perf_mmap_unmark_page(void *addr)
2202 {
2203         struct page *page = vmalloc_to_page(addr);
2204 
2205         page->mapping = NULL;
2206 }
2207 
2208 static void perf_mmap_data_free_work(struct work_struct *work)
2209 {
2210         struct perf_mmap_data *data;
2211         void *base;
2212         int i, nr;
2213 
2214         data = container_of(work, struct perf_mmap_data, work);
2215         nr = 1 << data->data_order;
2216 
2217         base = data->user_page;
2218         for (i = 0; i < nr + 1; i++)
2219                 perf_mmap_unmark_page(base + (i * PAGE_SIZE));
2220 
2221         vfree(base);
2222         kfree(data);
2223 }
2224 
2225 static void perf_mmap_data_free(struct perf_mmap_data *data)
2226 {
2227         schedule_work(&data->work);
2228 }
2229 
2230 static struct perf_mmap_data *
2231 perf_mmap_data_alloc(struct perf_event *event, int nr_pages)
2232 {
2233         struct perf_mmap_data *data;
2234         unsigned long size;
2235         void *all_buf;
2236 
2237         WARN_ON(atomic_read(&event->mmap_count));
2238 
2239         size = sizeof(struct perf_mmap_data);
2240         size += sizeof(void *);
2241 
2242         data = kzalloc(size, GFP_KERNEL);
2243         if (!data)
2244                 goto fail;
2245 
2246         INIT_WORK(&data->work, perf_mmap_data_free_work);
2247 
2248         all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
2249         if (!all_buf)
2250                 goto fail_all_buf;
2251 
2252         data->user_page = all_buf;
2253         data->data_pages[0] = all_buf + PAGE_SIZE;
2254         data->data_order = ilog2(nr_pages);
2255         data->nr_pages = 1;
2256 
2257         return data;
2258 
2259 fail_all_buf:
2260         kfree(data);
2261 
2262 fail:
2263         return NULL;
2264 }
2265 
2266 #endif
2267 
2268 static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2269 {
2270         struct perf_event *event = vma->vm_file->private_data;
2271         struct perf_mmap_data *data;
2272         int ret = VM_FAULT_SIGBUS;
2273 
2274         if (vmf->flags & FAULT_FLAG_MKWRITE) {
2275                 if (vmf->pgoff == 0)
2276                         ret = 0;
2277                 return ret;
2278         }
2279 
2280         rcu_read_lock();
2281         data = rcu_dereference(event->data);
2282         if (!data)
2283                 goto unlock;
2284 
2285         if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
2286                 goto unlock;
2287 
2288         vmf->page = perf_mmap_to_page(data, vmf->pgoff);
2289         if (!vmf->page)
2290                 goto unlock;
2291 
2292         get_page(vmf->page);
2293         vmf->page->mapping = vma->vm_file->f_mapping;
2294         vmf->page->index   = vmf->pgoff;
2295 
2296         ret = 0;
2297 unlock:
2298         rcu_read_unlock();
2299 
2300         return ret;
2301 }
2302 
2303 static void
2304 perf_mmap_data_init(struct perf_event *event, struct perf_mmap_data *data)
2305 {
2306         long max_size = perf_data_size(data);
2307 
2308         atomic_set(&data->lock, -1);
2309 
2310         if (event->attr.watermark) {
2311                 data->watermark = min_t(long, max_size,
2312                                         event->attr.wakeup_watermark);
2313         }
2314 
2315         if (!data->watermark)
2316                 data->watermark = max_t(long, PAGE_SIZE, max_size / 2);
2317 
2318 
2319         rcu_assign_pointer(event->data, data);
2320 }
2321 
2322 static void perf_mmap_data_free_rcu(struct rcu_head *rcu_head)
2323 {
2324         struct perf_mmap_data *data;
2325 
2326         data = container_of(rcu_head, struct perf_mmap_data, rcu_head);
2327         perf_mmap_data_free(data);
2328 }
2329 
2330 static void perf_mmap_data_release(struct perf_event *event)
2331 {
2332         struct perf_mmap_data *data = event->data;
2333 
2334         WARN_ON(atomic_read(&event->mmap_count));
2335 
2336         rcu_assign_pointer(event->data, NULL);
2337         call_rcu(&data->rcu_head, perf_mmap_data_free_rcu);
2338 }
2339 
2340 static void perf_mmap_open(struct vm_area_struct *vma)
2341 {
2342         struct perf_event *event = vma->vm_file->private_data;
2343 
2344         atomic_inc(&event->mmap_count);
2345 }
2346 
2347 static void perf_mmap_close(struct vm_area_struct *vma)
2348 {
2349         struct perf_event *event = vma->vm_file->private_data;
2350 
2351         WARN_ON_ONCE(event->ctx->parent_ctx);
2352         if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
2353                 unsigned long size = perf_data_size(event->data);
2354                 struct user_struct *user = current_user();
2355 
2356                 atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
2357                 vma->vm_mm->locked_vm -= event->data->nr_locked;
2358                 perf_mmap_data_release(event);
2359                 mutex_unlock(&event->mmap_mutex);
2360         }
2361 }
2362 
2363 static const struct vm_operations_struct perf_mmap_vmops = {
2364         .open           = perf_mmap_open,
2365         .close          = perf_mmap_close,
2366         .fault          = perf_mmap_fault,
2367         .page_mkwrite   = perf_mmap_fault,
2368 };
2369 
2370 static int perf_mmap(struct file *file, struct vm_area_struct *vma)
2371 {
2372         struct perf_event *event = file->private_data;
2373         unsigned long user_locked, user_lock_limit;
2374         struct user_struct *user = current_user();
2375         unsigned long locked, lock_limit;
2376         struct perf_mmap_data *data;
2377         unsigned long vma_size;
2378         unsigned long nr_pages;
2379         long user_extra, extra;
2380         int ret = 0;
2381 
2382         if (!(vma->vm_flags & VM_SHARED))
2383                 return -EINVAL;
2384 
2385         vma_size = vma->vm_end - vma->vm_start;
2386         nr_pages = (vma_size / PAGE_SIZE) - 1;
2387 
2388         /*
2389          * If we have data pages ensure they're a power-of-two number, so we
2390          * can do bitmasks instead of modulo.
2391          */
2392         if (nr_pages != 0 && !is_power_of_2(nr_pages))
2393                 return -EINVAL;
2394 
2395         if (vma_size != PAGE_SIZE * (1 + nr_pages))
2396                 return -EINVAL;
2397 
2398         if (vma->vm_pgoff != 0)
2399                 return -EINVAL;
2400 
2401         WARN_ON_ONCE(event->ctx->parent_ctx);
2402         mutex_lock(&event->mmap_mutex);
2403         if (event->output) {
2404                 ret = -EINVAL;
2405                 goto unlock;
2406         }
2407 
2408         if (atomic_inc_not_zero(&event->mmap_count)) {
2409                 if (nr_pages != event->data->nr_pages)
2410                         ret = -EINVAL;
2411                 goto unlock;
2412         }
2413 
2414         user_extra = nr_pages + 1;
2415         user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
2416 
2417         /*
2418          * Increase the limit linearly with more CPUs:
2419          */
2420         user_lock_limit *= num_online_cpus();
2421 
2422         user_locked = atomic_long_read(&user->locked_vm) + user_extra;
2423 
2424         extra = 0;
2425         if (user_locked > user_lock_limit)
2426                 extra = user_locked - user_lock_limit;
2427 
2428         lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2429         lock_limit >>= PAGE_SHIFT;
2430         locked = vma->vm_mm->locked_vm + extra;
2431 
2432         if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
2433                 !capable(CAP_IPC_LOCK)) {
2434                 ret = -EPERM;
2435                 goto unlock;
2436         }
2437 
2438         WARN_ON(event->data);
2439 
2440         data = perf_mmap_data_alloc(event, nr_pages);
2441         ret = -ENOMEM;
2442         if (!data)
2443                 goto unlock;
2444 
2445         ret = 0;
2446         perf_mmap_data_init(event, data);
2447 
2448         atomic_set(&event->mmap_count, 1);
2449         atomic_long_add(user_extra, &user->locked_vm);
2450         vma->vm_mm->locked_vm += extra;
2451         event->data->nr_locked = extra;
2452         if (vma->vm_flags & VM_WRITE)
2453                 event->data->writable = 1;
2454 
2455 unlock:
2456         mutex_unlock(&event->mmap_mutex);
2457 
2458         vma->vm_flags |= VM_RESERVED;
2459         vma->vm_ops = &perf_mmap_vmops;
2460 
2461         return ret;
2462 }
2463 
2464 static int perf_fasync(int fd, struct file *filp, int on)
2465 {
2466         struct inode *inode = filp->f_path.dentry->d_inode;
2467         struct perf_event *event = filp->private_data;
2468         int retval;
2469 
2470         mutex_lock(&inode->i_mutex);
2471         retval = fasync_helper(fd, filp, on, &event->fasync);
2472         mutex_unlock(&inode->i_mutex);
2473 
2474         if (retval < 0)
2475                 return retval;
2476 
2477         return 0;
2478 }
2479 
2480 static const struct file_operations perf_fops = {
2481         .release                = perf_release,
2482         .read                   = perf_read,
2483         .poll                   = perf_poll,
2484         .unlocked_ioctl         = perf_ioctl,
2485         .compat_ioctl           = perf_ioctl,
2486         .mmap                   = perf_mmap,
2487         .fasync                 = perf_fasync,
2488 };
2489 
2490 /*
2491  * Perf event wakeup
2492  *
2493  * If there's data, ensure we set the poll() state and publish everything
2494  * to user-space before waking everybody up.
2495  */
2496 
2497 void perf_event_wakeup(struct perf_event *event)
2498 {
2499         wake_up_all(&event->waitq);
2500 
2501         if (event->pending_kill) {
2502                 kill_fasync(&event->fasync, SIGIO, event->pending_kill);
2503                 event->pending_kill = 0;
2504         }
2505 }
2506 
2507 /*
2508  * Pending wakeups
2509  *
2510  * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
2511  *
2512  * The NMI bit means we cannot possibly take locks. Therefore, maintain a
2513  * single linked list and use cmpxchg() to add entries lockless.
2514  */
2515 
2516 static void perf_pending_event(struct perf_pending_entry *entry)
2517 {
2518         struct perf_event *event = container_of(entry,
2519                         struct perf_event, pending);
2520 
2521         if (event->pending_disable) {
2522                 event->pending_disable = 0;
2523                 __perf_event_disable(event);
2524         }
2525 
2526         if (event->pending_wakeup) {
2527                 event->pending_wakeup = 0;
2528                 perf_event_wakeup(event);
2529         }
2530 }
2531 
2532 #define PENDING_TAIL ((struct perf_pending_entry *)-1UL)
2533 
2534 static DEFINE_PER_CPU(struct perf_pending_entry *, perf_pending_head) = {
2535         PENDING_TAIL,
2536 };
2537 
2538 static void perf_pending_queue(struct perf_pending_entry *entry,
2539                                void (*func)(struct perf_pending_entry *))
2540 {
2541         struct perf_pending_entry **head;
2542 
2543         if (cmpxchg(&entry->next, NULL, PENDING_TAIL) != NULL)
2544                 return;
2545 
2546         entry->func = func;
2547 
2548         head = &get_cpu_var(perf_pending_head);
2549 
2550         do {
2551                 entry->next = *head;
2552         } while (cmpxchg(head, entry->next, entry) != entry->next);
2553 
2554         set_perf_event_pending();
2555 
2556         put_cpu_var(perf_pending_head);
2557 }
2558 
2559 static int __perf_pending_run(void)
2560 {
2561         struct perf_pending_entry *list;
2562         int nr = 0;
2563 
2564         list = xchg(&__get_cpu_var(perf_pending_head), PENDING_TAIL);
2565         while (list != PENDING_TAIL) {
2566                 void (*func)(struct perf_pending_entry *);
2567                 struct perf_pending_entry *entry = list;
2568 
2569                 list = list->next;
2570 
2571                 func = entry->func;
2572                 entry->next = NULL;
2573                 /*
2574                  * Ensure we observe the unqueue before we issue the wakeup,
2575                  * so that we won't be waiting forever.
2576                  * -- see perf_not_pending().
2577                  */
2578                 smp_wmb();
2579 
2580                 func(entry);
2581                 nr++;
2582         }
2583 
2584         return nr;
2585 }
2586 
2587 static inline int perf_not_pending(struct perf_event *event)
2588 {
2589         /*
2590          * If we flush on whatever cpu we run, there is a chance we don't
2591          * need to wait.
2592          */
2593         get_cpu();
2594         __perf_pending_run();
2595         put_cpu();
2596 
2597         /*
2598          * Ensure we see the proper queue state before going to sleep
2599          * so that we do not miss the wakeup. -- see perf_pending_handle()
2600          */
2601         smp_rmb();
2602         return event->pending.next == NULL;
2603 }
2604 
2605 static void perf_pending_sync(struct perf_event *event)
2606 {
2607         wait_event(event->waitq, perf_not_pending(event));
2608 }
2609 
2610 void perf_event_do_pending(void)
2611 {
2612         __perf_pending_run();
2613 }
2614 
2615 /*
2616  * Callchain support -- arch specific
2617  */
2618 
2619 __weak struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
2620 {
2621         return NULL;
2622 }
2623 
2624 /*
2625  * Output
2626  */
2627 static bool perf_output_space(struct perf_mmap_data *data, unsigned long tail,
2628                               unsigned long offset, unsigned long head)
2629 {
2630         unsigned long mask;
2631 
2632         if (!data->writable)
2633                 return true;
2634 
2635         mask = perf_data_size(data) - 1;
2636 
2637         offset = (offset - tail) & mask;
2638         head   = (head   - tail) & mask;
2639 
2640         if ((int)(head - offset) < 0)
2641                 return false;
2642 
2643         return true;
2644 }
2645 
2646 static void perf_output_wakeup(struct perf_output_handle *handle)
2647 {
2648         atomic_set(&handle->data->poll, POLL_IN);
2649 
2650         if (handle->nmi) {
2651                 handle->event->pending_wakeup = 1;
2652                 perf_pending_queue(&handle->event->pending,
2653                                    perf_pending_event);
2654         } else
2655                 perf_event_wakeup(handle->event);
2656 }
2657 
2658 /*
2659  * Curious locking construct.
2660  *
2661  * We need to ensure a later event_id doesn't publish a head when a former
2662  * event_id isn't done writing. However since we need to deal with NMIs we
2663  * cannot fully serialize things.
2664  *
2665  * What we do is serialize between CPUs so we only have to deal with NMI
2666  * nesting on a single CPU.
2667  *
2668  * We only publish the head (and generate a wakeup) when the outer-most
2669  * event_id completes.
2670  */
2671 static void perf_output_lock(struct perf_output_handle *handle)
2672 {
2673         struct perf_mmap_data *data = handle->data;
2674         int cpu;
2675 
2676         handle->locked = 0;
2677 
2678         local_irq_save(handle->flags);
2679         cpu = smp_processor_id();
2680 
2681         if (in_nmi() && atomic_read(&data->lock) == cpu)
2682                 return;
2683 
2684         while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2685                 cpu_relax();
2686 
2687         handle->locked = 1;
2688 }
2689 
2690 static void perf_output_unlock(struct perf_output_handle *handle)
2691 {
2692         struct perf_mmap_data *data = handle->data;
2693         unsigned long head;
2694         int cpu;
2695 
2696         data->done_head = data->head;
2697 
2698         if (!handle->locked)
2699                 goto out;
2700 
2701 again:
2702         /*
2703          * The xchg implies a full barrier that ensures all writes are done
2704          * before we publish the new head, matched by a rmb() in userspace when
2705          * reading this position.
2706          */
2707         while ((head = atomic_long_xchg(&data->done_head, 0)))
2708                 data->user_page->data_head = head;
2709 
2710         /*
2711          * NMI can happen here, which means we can miss a done_head update.
2712          */
2713 
2714         cpu = atomic_xchg(&data->lock, -1);
2715         WARN_ON_ONCE(cpu != smp_processor_id());
2716 
2717         /*
2718          * Therefore we have to validate we did not indeed do so.
2719          */
2720         if (unlikely(atomic_long_read(&data->done_head))) {
2721                 /*
2722                  * Since we had it locked, we can lock it again.
2723                  */
2724                 while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
2725                         cpu_relax();
2726 
2727                 goto again;
2728         }
2729 
2730         if (atomic_xchg(&data->wakeup, 0))
2731                 perf_output_wakeup(handle);
2732 out:
2733         local_irq_restore(handle->flags);
2734 }
2735 
2736 void perf_output_copy(struct perf_output_handle *handle,
2737                       const void *buf, unsigned int len)
2738 {
2739         unsigned int pages_mask;
2740         unsigned long offset;
2741         unsigned int size;
2742         void **pages;
2743 
2744         offset          = handle->offset;
2745         pages_mask      = handle->data->nr_pages - 1;
2746         pages           = handle->data->data_pages;
2747 
2748         do {
2749                 unsigned long page_offset;
2750                 unsigned long page_size;
2751                 int nr;
2752 
2753                 nr          = (offset >> PAGE_SHIFT) & pages_mask;
2754                 page_size   = 1UL << (handle->data->data_order + PAGE_SHIFT);
2755                 page_offset = offset & (page_size - 1);
2756                 size        = min_t(unsigned int, page_size - page_offset, len);
2757 
2758                 memcpy(pages[nr] + page_offset, buf, size);
2759 
2760                 len         -= size;
2761                 buf         += size;
2762                 offset      += size;
2763         } while (len);
2764 
2765         handle->offset = offset;
2766 
2767         /*
2768          * Check we didn't copy past our reservation window, taking the
2769          * possible unsigned int wrap into account.
2770          */
2771         WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0);
2772 }
2773 
2774 int perf_output_begin(struct perf_output_handle *handle,
2775                       struct perf_event *event, unsigned int size,
2776                       int nmi, int sample)
2777 {
2778         struct perf_event *output_event;
2779         struct perf_mmap_data *data;
2780         unsigned long tail, offset, head;
2781         int have_lost;
2782         struct {
2783                 struct perf_event_header header;
2784                 u64                      id;
2785                 u64                      lost;
2786         } lost_event;
2787 
2788         rcu_read_lock();
2789         /*
2790          * For inherited events we send all the output towards the parent.
2791          */
2792         if (event->parent)
2793                 event = event->parent;
2794 
2795         output_event = rcu_dereference(event->output);
2796         if (output_event)
2797                 event = output_event;
2798 
2799         data = rcu_dereference(event->data);
2800         if (!data)
2801                 goto out;
2802 
2803         handle->data    = data;
2804         handle->event   = event;
2805         handle->nmi     = nmi;
2806         handle->sample  = sample;
2807 
2808         if (!data->nr_pages)
2809                 goto fail;
2810 
2811         have_lost = atomic_read(&data->lost);
2812         if (have_lost)
2813                 size += sizeof(lost_event);
2814 
2815         perf_output_lock(handle);
2816 
2817         do {
2818                 /*
2819                  * Userspace could choose to issue a mb() before updating the
2820                  * tail pointer. So that all reads will be completed before the
2821                  * write is issued.
2822                  */
2823                 tail = ACCESS_ONCE(data->user_page->data_tail);
2824                 smp_rmb();
2825                 offset = head = atomic_long_read(&data->head);
2826                 head += size;
2827                 if (unlikely(!perf_output_space(data, tail, offset, head)))
2828                         goto fail;
2829         } while (atomic_long_cmpxchg(&data->head, offset, head) != offset);
2830 
2831         handle->offset  = offset;
2832         handle->head    = head;
2833 
2834         if (head - tail > data->watermark)
2835                 atomic_set(&data->wakeup, 1);
2836 
2837         if (have_lost) {
2838                 lost_event.header.type = PERF_RECORD_LOST;
2839                 lost_event.header.misc = 0;
2840                 lost_event.header.size = sizeof(lost_event);
2841                 lost_event.id          = event->id;
2842                 lost_event.lost        = atomic_xchg(&data->lost, 0);
2843 
2844                 perf_output_put(handle, lost_event);
2845         }
2846 
2847         return 0;
2848 
2849 fail:
2850         atomic_inc(&data->lost);
2851         perf_output_unlock(handle);
2852 out:
2853         rcu_read_unlock();
2854 
2855         return -ENOSPC;
2856 }
2857 
2858 void perf_output_end(struct perf_output_handle *handle)
2859 {
2860         struct perf_event *event = handle->event;
2861         struct perf_mmap_data *data = handle->data;
2862 
2863         int wakeup_events = event->attr.wakeup_events;
2864 
2865         if (handle->sample && wakeup_events) {
2866                 int events = atomic_inc_return(&data->events);
2867                 if (events >= wakeup_events) {
2868                         atomic_sub(wakeup_events, &data->events);
2869                         atomic_set(&data->wakeup, 1);
2870                 }
2871         }
2872 
2873         perf_output_unlock(handle);
2874         rcu_read_unlock();
2875 }
2876 
2877 static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
2878 {
2879         /*
2880          * only top level events have the pid namespace they were created in
2881          */
2882         if (event->parent)
2883                 event = event->parent;
2884 
2885         return task_tgid_nr_ns(p, event->ns);
2886 }
2887 
2888 static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
2889 {
2890         /*
2891          * only top level events have the pid namespace they were created in
2892          */
2893         if (event->parent)
2894                 event = event->parent;
2895 
2896         return task_pid_nr_ns(p, event->ns);
2897 }
2898 
2899 static void perf_output_read_one(struct perf_output_handle *handle,
2900                                  struct perf_event *event)
2901 {
2902         u64 read_format = event->attr.read_format;
2903         u64 values[4];
2904         int n = 0;
2905 
2906         values[n++] = atomic64_read(&event->count);
2907         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2908                 values[n++] = event->total_time_enabled +
2909                         atomic64_read(&event->child_total_time_enabled);
2910         }
2911         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2912                 values[n++] = event->total_time_running +
2913                         atomic64_read(&event->child_total_time_running);
2914         }
2915         if (read_format & PERF_FORMAT_ID)
2916                 values[n++] = primary_event_id(event);
2917 
2918         perf_output_copy(handle, values, n * sizeof(u64));
2919 }
2920 
2921 /*
2922  * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
2923  */
2924 static void perf_output_read_group(struct perf_output_handle *handle,
2925                             struct perf_event *event)
2926 {
2927         struct perf_event *leader = event->group_leader, *sub;
2928         u64 read_format = event->attr.read_format;
2929         u64 values[5];
2930         int n = 0;
2931 
2932         values[n++] = 1 + leader->nr_siblings;
2933 
2934         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2935                 values[n++] = leader->total_time_enabled;
2936 
2937         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2938                 values[n++] = leader->total_time_running;
2939 
2940         if (leader != event)
2941                 leader->pmu->read(leader);
2942 
2943         values[n++] = atomic64_read(&leader->count);
2944         if (read_format & PERF_FORMAT_ID)
2945                 values[n++] = primary_event_id(leader);
2946 
2947         perf_output_copy(handle, values, n * sizeof(u64));
2948 
2949         list_for_each_entry(sub, &leader->sibling_list, group_entry) {
2950                 n = 0;
2951 
2952                 if (sub != event)
2953                         sub->pmu->read(sub);
2954 
2955                 values[n++] = atomic64_read(&sub->count);
2956                 if (read_format & PERF_FORMAT_ID)
2957                         values[n++] = primary_event_id(sub);
2958 
2959                 perf_output_copy(handle, values, n * sizeof(u64));
2960         }
2961 }
2962 
2963 static void perf_output_read(struct perf_output_handle *handle,
2964                              struct perf_event *event)
2965 {
2966         if (event->attr.read_format & PERF_FORMAT_GROUP)
2967                 perf_output_read_group(handle, event);
2968         else
2969                 perf_output_read_one(handle, event);
2970 }
2971 
2972 void perf_output_sample(struct perf_output_handle *handle,
2973                         struct perf_event_header *header,
2974                         struct perf_sample_data *data,
2975                         struct perf_event *event)
2976 {
2977         u64 sample_type = data->type;
2978 
2979         perf_output_put(handle, *header);
2980 
2981         if (sample_type & PERF_SAMPLE_IP)
2982                 perf_output_put(handle, data->ip);
2983 
2984         if (sample_type & PERF_SAMPLE_TID)
2985                 perf_output_put(handle, data->tid_entry);
2986 
2987         if (sample_type & PERF_SAMPLE_TIME)
2988                 perf_output_put(handle, data->time);
2989 
2990         if (sample_type & PERF_SAMPLE_ADDR)
2991                 perf_output_put(handle, data->addr);
2992 
2993         if (sample_type & PERF_SAMPLE_ID)
2994                 perf_output_put(handle, data->id);
2995 
2996         if (sample_type & PERF_SAMPLE_STREAM_ID)
2997                 perf_output_put(handle, data->stream_id);
2998 
2999         if (sample_type & PERF_SAMPLE_CPU)
3000                 perf_output_put(handle, data->cpu_entry);
3001 
3002         if (sample_type & PERF_SAMPLE_PERIOD)
3003                 perf_output_put(handle, data->period);
3004 
3005         if (sample_type & PERF_SAMPLE_READ)
3006                 perf_output_read(handle, event);
3007 
3008         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3009                 if (data->callchain) {
3010                         int size = 1;
3011 
3012                         if (data->callchain)
3013                                 size += data->callchain->nr;
3014 
3015                         size *= sizeof(u64);
3016 
3017                         perf_output_copy(handle, data->callchain, size);
3018                 } else {
3019                         u64 nr = 0;
3020                         perf_output_put(handle, nr);
3021                 }
3022         }
3023 
3024         if (sample_type & PERF_SAMPLE_RAW) {
3025                 if (data->raw) {
3026                         perf_output_put(handle, data->raw->size);
3027                         perf_output_copy(handle, data->raw->data,
3028                                          data->raw->size);
3029                 } else {
3030                         struct {
3031                                 u32     size;
3032                                 u32     data;
3033                         } raw = {
3034                                 .size = sizeof(u32),
3035                                 .data = 0,
3036                         };
3037                         perf_output_put(handle, raw);
3038                 }
3039         }
3040 }
3041 
3042 void perf_prepare_sample(struct perf_event_header *header,
3043                          struct perf_sample_data *data,
3044                          struct perf_event *event,
3045                          struct pt_regs *regs)
3046 {
3047         u64 sample_type = event->attr.sample_type;
3048 
3049         data->type = sample_type;
3050 
3051         header->type = PERF_RECORD_SAMPLE;
3052         header->size = sizeof(*header);
3053 
3054         header->misc = 0;
3055         header->misc |= perf_misc_flags(regs);
3056 
3057         if (sample_type & PERF_SAMPLE_IP) {
3058                 data->ip = perf_instruction_pointer(regs);
3059 
3060                 header->size += sizeof(data->ip);
3061         }
3062 
3063         if (sample_type & PERF_SAMPLE_TID) {
3064                 /* namespace issues */
3065                 data->tid_entry.pid = perf_event_pid(event, current);
3066                 data->tid_entry.tid = perf_event_tid(event, current);
3067 
3068                 header->size += sizeof(data->tid_entry);
3069         }
3070 
3071         if (sample_type & PERF_SAMPLE_TIME) {
3072                 data->time = perf_clock();
3073 
3074                 header->size += sizeof(data->time);
3075         }
3076 
3077         if (sample_type & PERF_SAMPLE_ADDR)
3078                 header->size += sizeof(data->addr);
3079 
3080         if (sample_type & PERF_SAMPLE_ID) {
3081                 data->id = primary_event_id(event);
3082 
3083                 header->size += sizeof(data->id);
3084         }
3085 
3086         if (sample_type & PERF_SAMPLE_STREAM_ID) {
3087                 data->stream_id = event->id;
3088 
3089                 header->size += sizeof(data->stream_id);
3090         }
3091 
3092         if (sample_type & PERF_SAMPLE_CPU) {
3093                 data->cpu_entry.cpu             = raw_smp_processor_id();
3094                 data->cpu_entry.reserved        = 0;
3095 
3096                 header->size += sizeof(data->cpu_entry);
3097         }
3098 
3099         if (sample_type & PERF_SAMPLE_PERIOD)
3100                 header->size += sizeof(data->period);
3101 
3102         if (sample_type & PERF_SAMPLE_READ)
3103                 header->size += perf_event_read_size(event);
3104 
3105         if (sample_type & PERF_SAMPLE_CALLCHAIN) {
3106                 int size = 1;
3107 
3108                 data->callchain = perf_callchain(regs);
3109 
3110                 if (data->callchain)
3111                         size += data->callchain->nr;
3112 
3113                 header->size += size * sizeof(u64);
3114         }
3115 
3116         if (sample_type & PERF_SAMPLE_RAW) {
3117                 int size = sizeof(u32);
3118 
3119                 if (data->raw)
3120                         size += data->raw->size;
3121                 else
3122                         size += sizeof(u32);
3123 
3124                 WARN_ON_ONCE(size & (sizeof(u64)-1));
3125                 header->size += size;
3126         }
3127 }
3128 
3129 static void perf_event_output(struct perf_event *event, int nmi,
3130                                 struct perf_sample_data *data,
3131                                 struct pt_regs *regs)
3132 {
3133         struct perf_output_handle handle;
3134         struct perf_event_header header;
3135 
3136         perf_prepare_sample(&header, data, event, regs);
3137 
3138         if (perf_output_begin(&handle, event, header.size, nmi, 1))
3139                 return;
3140 
3141         perf_output_sample(&handle, &header, data, event);
3142 
3143         perf_output_end(&handle);
3144 }
3145 
3146 /*
3147  * read event_id
3148  */
3149 
3150 struct perf_read_event {
3151         struct perf_event_header        header;
3152 
3153         u32                             pid;
3154         u32                             tid;
3155 };
3156 
3157 static void
3158 perf_event_read_event(struct perf_event *event,
3159                         struct task_struct *task)
3160 {
3161         struct perf_output_handle handle;
3162         struct perf_read_event read_event = {
3163                 .header = {
3164                         .type = PERF_RECORD_READ,
3165                         .misc = 0,
3166                         .size = sizeof(read_event) + perf_event_read_size(event),
3167                 },
3168                 .pid = perf_event_pid(event, task),
3169                 .tid = perf_event_tid(event, task),
3170         };
3171         int ret;
3172 
3173         ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
3174         if (ret)
3175                 return;
3176 
3177         perf_output_put(&handle, read_event);
3178         perf_output_read(&handle, event);
3179 
3180         perf_output_end(&handle);
3181 }
3182 
3183 /*
3184  * task tracking -- fork/exit
3185  *
3186  * enabled by: attr.comm | attr.mmap | attr.task
3187  */
3188 
3189 struct perf_task_event {
3190         struct task_struct              *task;
3191         struct perf_event_context       *task_ctx;
3192 
3193         struct {
3194                 struct perf_event_header        header;
3195 
3196                 u32                             pid;
3197                 u32                             ppid;
3198                 u32                             tid;
3199                 u32                             ptid;
3200                 u64                             time;
3201         } event_id;
3202 };
3203 
3204 static void perf_event_task_output(struct perf_event *event,
3205                                      struct perf_task_event *task_event)
3206 {
3207         struct perf_output_handle handle;
3208         int size;
3209         struct task_struct *task = task_event->task;
3210         int ret;
3211 
3212         size  = task_event->event_id.header.size;
3213         ret = perf_output_begin(&handle, event, size, 0, 0);
3214 
3215         if (ret)
3216                 return;
3217 
3218         task_event->event_id.pid = perf_event_pid(event, task);
3219         task_event->event_id.ppid = perf_event_pid(event, current);
3220 
3221         task_event->event_id.tid = perf_event_tid(event, task);
3222         task_event->event_id.ptid = perf_event_tid(event, current);
3223 
3224         task_event->event_id.time = perf_clock();
3225 
3226         perf_output_put(&handle, task_event->event_id);
3227 
3228         perf_output_end(&handle);
3229 }
3230 
3231 static int perf_event_task_match(struct perf_event *event)
3232 {
3233         if (event->state != PERF_EVENT_STATE_ACTIVE)
3234                 return 0;
3235 
3236         if (event->cpu != -1 && event->cpu != smp_processor_id())
3237                 return 0;
3238 
3239         if (event->attr.comm || event->attr.mmap || event->attr.task)
3240                 return 1;
3241 
3242         return 0;
3243 }
3244 
3245 static void perf_event_task_ctx(struct perf_event_context *ctx,
3246                                   struct perf_task_event *task_event)
3247 {
3248         struct perf_event *event;
3249 
3250         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3251                 return;
3252 
3253         rcu_read_lock();
3254         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3255                 if (perf_event_task_match(event))
3256                         perf_event_task_output(event, task_event);
3257         }
3258         rcu_read_unlock();
3259 }
3260 
3261 static void perf_event_task_event(struct perf_task_event *task_event)
3262 {
3263         struct perf_cpu_context *cpuctx;
3264         struct perf_event_context *ctx = task_event->task_ctx;
3265 
3266         cpuctx = &get_cpu_var(perf_cpu_context);
3267         perf_event_task_ctx(&cpuctx->ctx, task_event);
3268 
3269         rcu_read_lock();
3270         if (!ctx)
3271                 ctx = rcu_dereference(task_event->task->perf_event_ctxp);
3272         if (ctx)
3273                 perf_event_task_ctx(ctx, task_event);
3274         put_cpu_var(perf_cpu_context);
3275         rcu_read_unlock();
3276 }
3277 
3278 static void perf_event_task(struct task_struct *task,
3279                               struct perf_event_context *task_ctx,
3280                               int new)
3281 {
3282         struct perf_task_event task_event;
3283 
3284         if (!atomic_read(&nr_comm_events) &&
3285             !atomic_read(&nr_mmap_events) &&
3286             !atomic_read(&nr_task_events))
3287                 return;
3288 
3289         task_event = (struct perf_task_event){
3290                 .task     = task,
3291                 .task_ctx = task_ctx,
3292                 .event_id    = {
3293                         .header = {
3294                                 .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
3295                                 .misc = 0,
3296                                 .size = sizeof(task_event.event_id),
3297                         },
3298                         /* .pid  */
3299                         /* .ppid */
3300                         /* .tid  */
3301                         /* .ptid */
3302                 },
3303         };
3304 
3305         perf_event_task_event(&task_event);
3306 }
3307 
3308 void perf_event_fork(struct task_struct *task)
3309 {
3310         perf_event_task(task, NULL, 1);
3311 }
3312 
3313 /*
3314  * comm tracking
3315  */
3316 
3317 struct perf_comm_event {
3318         struct task_struct      *task;
3319         char                    *comm;
3320         int                     comm_size;
3321 
3322         struct {
3323                 struct perf_event_header        header;
3324 
3325                 u32                             pid;
3326                 u32                             tid;
3327         } event_id;
3328 };
3329 
3330 static void perf_event_comm_output(struct perf_event *event,
3331                                      struct perf_comm_event *comm_event)
3332 {
3333         struct perf_output_handle handle;
3334         int size = comm_event->event_id.header.size;
3335         int ret = perf_output_begin(&handle, event, size, 0, 0);
3336 
3337         if (ret)
3338                 return;
3339 
3340         comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
3341         comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
3342 
3343         perf_output_put(&handle, comm_event->event_id);
3344         perf_output_copy(&handle, comm_event->comm,
3345                                    comm_event->comm_size);
3346         perf_output_end(&handle);
3347 }
3348 
3349 static int perf_event_comm_match(struct perf_event *event)
3350 {
3351         if (event->state != PERF_EVENT_STATE_ACTIVE)
3352                 return 0;
3353 
3354         if (event->cpu != -1 && event->cpu != smp_processor_id())
3355                 return 0;
3356 
3357         if (event->attr.comm)
3358                 return 1;
3359 
3360         return 0;
3361 }
3362 
3363 static void perf_event_comm_ctx(struct perf_event_context *ctx,
3364                                   struct perf_comm_event *comm_event)
3365 {
3366         struct perf_event *event;
3367 
3368         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3369                 return;
3370 
3371         rcu_read_lock();
3372         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3373                 if (perf_event_comm_match(event))
3374                         perf_event_comm_output(event, comm_event);
3375         }
3376         rcu_read_unlock();
3377 }
3378 
3379 static void perf_event_comm_event(struct perf_comm_event *comm_event)
3380 {
3381         struct perf_cpu_context *cpuctx;
3382         struct perf_event_context *ctx;
3383         unsigned int size;
3384         char comm[TASK_COMM_LEN];
3385 
3386         memset(comm, 0, sizeof(comm));
3387         strncpy(comm, comm_event->task->comm, sizeof(comm));
3388         size = ALIGN(strlen(comm)+1, sizeof(u64));
3389 
3390         comm_event->comm = comm;
3391         comm_event->comm_size = size;
3392 
3393         comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
3394 
3395         cpuctx = &get_cpu_var(perf_cpu_context);
3396         perf_event_comm_ctx(&cpuctx->ctx, comm_event);
3397 
3398         rcu_read_lock();
3399         /*
3400          * doesn't really matter which of the child contexts the
3401          * events ends up in.
3402          */
3403         ctx = rcu_dereference(current->perf_event_ctxp);
3404         if (ctx)
3405                 perf_event_comm_ctx(ctx, comm_event);
3406         put_cpu_var(perf_cpu_context);
3407         rcu_read_unlock();
3408 }
3409 
3410 void perf_event_comm(struct task_struct *task)
3411 {
3412         struct perf_comm_event comm_event;
3413 
3414         if (task->perf_event_ctxp)
3415                 perf_event_enable_on_exec(task);
3416 
3417         if (!atomic_read(&nr_comm_events))
3418                 return;
3419 
3420         comm_event = (struct perf_comm_event){
3421                 .task   = task,
3422                 /* .comm      */
3423                 /* .comm_size */
3424                 .event_id  = {
3425                         .header = {
3426                                 .type = PERF_RECORD_COMM,
3427                                 .misc = 0,
3428                                 /* .size */
3429                         },
3430                         /* .pid */
3431                         /* .tid */
3432                 },
3433         };
3434 
3435         perf_event_comm_event(&comm_event);
3436 }
3437 
3438 /*
3439  * mmap tracking
3440  */
3441 
3442 struct perf_mmap_event {
3443         struct vm_area_struct   *vma;
3444 
3445         const char              *file_name;
3446         int                     file_size;
3447 
3448         struct {
3449                 struct perf_event_header        header;
3450 
3451                 u32                             pid;
3452                 u32                             tid;
3453                 u64                             start;
3454                 u64                             len;
3455                 u64                             pgoff;
3456         } event_id;
3457 };
3458 
3459 static void perf_event_mmap_output(struct perf_event *event,
3460                                      struct perf_mmap_event *mmap_event)
3461 {
3462         struct perf_output_handle handle;
3463         int size = mmap_event->event_id.header.size;
3464         int ret = perf_output_begin(&handle, event, size, 0, 0);
3465 
3466         if (ret)
3467                 return;
3468 
3469         mmap_event->event_id.pid = perf_event_pid(event, current);
3470         mmap_event->event_id.tid = perf_event_tid(event, current);
3471 
3472         perf_output_put(&handle, mmap_event->event_id);
3473         perf_output_copy(&handle, mmap_event->file_name,
3474                                    mmap_event->file_size);
3475         perf_output_end(&handle);
3476 }
3477 
3478 static int perf_event_mmap_match(struct perf_event *event,
3479                                    struct perf_mmap_event *mmap_event)
3480 {
3481         if (event->state != PERF_EVENT_STATE_ACTIVE)
3482                 return 0;
3483 
3484         if (event->cpu != -1 && event->cpu != smp_processor_id())
3485                 return 0;
3486 
3487         if (event->attr.mmap)
3488                 return 1;
3489 
3490         return 0;
3491 }
3492 
3493 static void perf_event_mmap_ctx(struct perf_event_context *ctx,
3494                                   struct perf_mmap_event *mmap_event)
3495 {
3496         struct perf_event *event;
3497 
3498         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3499                 return;
3500 
3501         rcu_read_lock();
3502         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3503                 if (perf_event_mmap_match(event, mmap_event))
3504                         perf_event_mmap_output(event, mmap_event);
3505         }
3506         rcu_read_unlock();
3507 }
3508 
3509 static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
3510 {
3511         struct perf_cpu_context *cpuctx;
3512         struct perf_event_context *ctx;
3513         struct vm_area_struct *vma = mmap_event->vma;
3514         struct file *file = vma->vm_file;
3515         unsigned int size;
3516         char tmp[16];
3517         char *buf = NULL;
3518         const char *name;
3519 
3520         memset(tmp, 0, sizeof(tmp));
3521 
3522         if (file) {
3523                 /*
3524                  * d_path works from the end of the buffer backwards, so we
3525                  * need to add enough zero bytes after the string to handle
3526                  * the 64bit alignment we do later.
3527                  */
3528                 buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
3529                 if (!buf) {
3530                         name = strncpy(tmp, "//enomem", sizeof(tmp));
3531                         goto got_name;
3532                 }
3533                 name = d_path(&file->f_path, buf, PATH_MAX);
3534                 if (IS_ERR(name)) {
3535                         name = strncpy(tmp, "//toolong", sizeof(tmp));
3536                         goto got_name;
3537                 }
3538         } else {
3539                 if (arch_vma_name(mmap_event->vma)) {
3540                         name = strncpy(tmp, arch_vma_name(mmap_event->vma),
3541                                        sizeof(tmp));
3542                         goto got_name;
3543                 }
3544 
3545                 if (!vma->vm_mm) {
3546                         name = strncpy(tmp, "[vdso]", sizeof(tmp));
3547                         goto got_name;
3548                 }
3549 
3550                 name = strncpy(tmp, "//anon", sizeof(tmp));
3551                 goto got_name;
3552         }
3553 
3554 got_name:
3555         size = ALIGN(strlen(name)+1, sizeof(u64));
3556 
3557         mmap_event->file_name = name;
3558         mmap_event->file_size = size;
3559 
3560         mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
3561 
3562         cpuctx = &get_cpu_var(perf_cpu_context);
3563         perf_event_mmap_ctx(&cpuctx->ctx, mmap_event);
3564 
3565         rcu_read_lock();
3566         /*
3567          * doesn't really matter which of the child contexts the
3568          * events ends up in.
3569          */
3570         ctx = rcu_dereference(current->perf_event_ctxp);
3571         if (ctx)
3572                 perf_event_mmap_ctx(ctx, mmap_event);
3573         put_cpu_var(perf_cpu_context);
3574         rcu_read_unlock();
3575 
3576         kfree(buf);
3577 }
3578 
3579 void __perf_event_mmap(struct vm_area_struct *vma)
3580 {
3581         struct perf_mmap_event mmap_event;
3582 
3583         if (!atomic_read(&nr_mmap_events))
3584                 return;
3585 
3586         mmap_event = (struct perf_mmap_event){
3587                 .vma    = vma,
3588                 /* .file_name */
3589                 /* .file_size */
3590                 .event_id  = {
3591                         .header = {
3592                                 .type = PERF_RECORD_MMAP,
3593                                 .misc = 0,
3594                                 /* .size */
3595                         },
3596                         /* .pid */
3597                         /* .tid */
3598                         .start  = vma->vm_start,
3599                         .len    = vma->vm_end - vma->vm_start,
3600                         .pgoff  = vma->vm_pgoff,
3601                 },
3602         };
3603 
3604         perf_event_mmap_event(&mmap_event);
3605 }
3606 
3607 /*
3608  * IRQ throttle logging
3609  */
3610 
3611 static void perf_log_throttle(struct perf_event *event, int enable)
3612 {
3613         struct perf_output_handle handle;
3614         int ret;
3615 
3616         struct {
3617                 struct perf_event_header        header;
3618                 u64                             time;
3619                 u64                             id;
3620                 u64                             stream_id;
3621         } throttle_event = {
3622                 .header = {
3623                         .type = PERF_RECORD_THROTTLE,
3624                         .misc = 0,
3625                         .size = sizeof(throttle_event),
3626                 },
3627                 .time           = perf_clock(),
3628                 .id             = primary_event_id(event),
3629                 .stream_id      = event->id,
3630         };
3631 
3632         if (enable)
3633                 throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
3634 
3635         ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
3636         if (ret)
3637                 return;
3638 
3639         perf_output_put(&handle, throttle_event);
3640         perf_output_end(&handle);
3641 }
3642 
3643 /*
3644  * Generic event overflow handling, sampling.
3645  */
3646 
3647 static int __perf_event_overflow(struct perf_event *event, int nmi,
3648                                    int throttle, struct perf_sample_data *data,
3649                                    struct pt_regs *regs)
3650 {
3651         int events = atomic_read(&event->event_limit);
3652         struct hw_perf_event *hwc = &event->hw;
3653         int ret = 0;
3654 
3655         throttle = (throttle && event->pmu->unthrottle != NULL);
3656 
3657         if (!throttle) {
3658                 hwc->interrupts++;
3659         } else {
3660                 if (hwc->interrupts != MAX_INTERRUPTS) {
3661                         hwc->interrupts++;
3662                         if (HZ * hwc->interrupts >
3663                                         (u64)sysctl_perf_event_sample_rate) {
3664                                 hwc->interrupts = MAX_INTERRUPTS;
3665                                 perf_log_throttle(event, 0);
3666                                 ret = 1;
3667                         }
3668                 } else {
3669                         /*
3670                          * Keep re-disabling events even though on the previous
3671                          * pass we disabled it - just in case we raced with a
3672                          * sched-in and the event got enabled again:
3673                          */
3674                         ret = 1;
3675                 }
3676         }
3677 
3678         if (event->attr.freq) {
3679                 u64 now = perf_clock();
3680                 s64 delta = now - hwc->freq_stamp;
3681 
3682                 hwc->freq_stamp = now;
3683 
3684                 if (delta > 0 && delta < TICK_NSEC)
3685                         perf_adjust_period(event, NSEC_PER_SEC / (int)delta);
3686         }
3687 
3688         /*
3689          * XXX event_limit might not quite work as expected on inherited
3690          * events
3691          */
3692 
3693         event->pending_kill = POLL_IN;
3694         if (events && atomic_dec_and_test(&event->event_limit)) {
3695                 ret = 1;
3696                 event->pending_kill = POLL_HUP;
3697                 event->pending_disable = 1;
3698                 perf_pending_queue(&event->pending, perf_pending_event);
3699         }
3700 
3701         perf_event_output(event, nmi, data, regs);
3702         return ret;
3703 }
3704 
3705 int perf_event_overflow(struct perf_event *event, int nmi,
3706                           struct perf_sample_data *data,
3707                           struct pt_regs *regs)
3708 {
3709         return __perf_event_overflow(event, nmi, 1, data, regs);
3710 }
3711 
3712 /*
3713  * Generic software event infrastructure
3714  */
3715 
3716 /*
3717  * We directly increment event->count and keep a second value in
3718  * event->hw.period_left to count intervals. This period event
3719  * is kept in the range [-sample_period, 0] so that we can use the
3720  * sign as trigger.
3721  */
3722 
3723 static u64 perf_swevent_set_period(struct perf_event *event)
3724 {
3725         struct hw_perf_event *hwc = &event->hw;
3726         u64 period = hwc->last_period;
3727         u64 nr, offset;
3728         s64 old, val;
3729 
3730         hwc->last_period = hwc->sample_period;
3731 
3732 again:
3733         old = val = atomic64_read(&hwc->period_left);
3734         if (val < 0)
3735                 return 0;
3736 
3737         nr = div64_u64(period + val, period);
3738         offset = nr * period;
3739         val -= offset;
3740         if (atomic64_cmpxchg(&hwc->period_left, old, val) != old)
3741                 goto again;
3742 
3743         return nr;
3744 }
3745 
3746 static void perf_swevent_overflow(struct perf_event *event,
3747                                     int nmi, struct perf_sample_data *data,
3748                                     struct pt_regs *regs)
3749 {
3750         struct hw_perf_event *hwc = &event->hw;
3751         int throttle = 0;
3752         u64 overflow;
3753 
3754         data->period = event->hw.last_period;
3755         overflow = perf_swevent_set_period(event);
3756 
3757         if (hwc->interrupts == MAX_INTERRUPTS)
3758                 return;
3759 
3760         for (; overflow; overflow--) {
3761                 if (__perf_event_overflow(event, nmi, throttle,
3762                                             data, regs)) {
3763                         /*
3764                          * We inhibit the overflow from happening when
3765                          * hwc->interrupts == MAX_INTERRUPTS.
3766                          */
3767                         break;
3768                 }
3769                 throttle = 1;
3770         }
3771 }
3772 
3773 static void perf_swevent_unthrottle(struct perf_event *event)
3774 {
3775         /*
3776          * Nothing to do, we already reset hwc->interrupts.
3777          */
3778 }
3779 
3780 static void perf_swevent_add(struct perf_event *event, u64 nr,
3781                                int nmi, struct perf_sample_data *data,
3782                                struct pt_regs *regs)
3783 {
3784         struct hw_perf_event *hwc = &event->hw;
3785 
3786         atomic64_add(nr, &event->count);
3787 
3788         if (!hwc->sample_period)
3789                 return;
3790 
3791         if (!regs)
3792                 return;
3793 
3794         if (!atomic64_add_negative(nr, &hwc->period_left))
3795                 perf_swevent_overflow(event, nmi, data, regs);
3796 }
3797 
3798 static int perf_swevent_is_counting(struct perf_event *event)
3799 {
3800         /*
3801          * The event is active, we're good!
3802          */
3803         if (event->state == PERF_EVENT_STATE_ACTIVE)
3804                 return 1;
3805 
3806         /*
3807          * The event is off/error, not counting.
3808          */
3809         if (event->state != PERF_EVENT_STATE_INACTIVE)
3810                 return 0;
3811 
3812         /*
3813          * The event is inactive, if the context is active
3814          * we're part of a group that didn't make it on the 'pmu',
3815          * not counting.
3816          */
3817         if (event->ctx->is_active)
3818                 return 0;
3819 
3820         /*
3821          * We're inactive and the context is too, this means the
3822          * task is scheduled out, we're counting events that happen
3823          * to us, like migration events.
3824          */
3825         return 1;
3826 }
3827 
3828 static int perf_swevent_match(struct perf_event *event,
3829                                 enum perf_type_id type,
3830                                 u32 event_id, struct pt_regs *regs)
3831 {
3832         if (event->cpu != -1 && event->cpu != smp_processor_id())
3833                 return 0;
3834 
3835         if (!perf_swevent_is_counting(event))
3836                 return 0;
3837 
3838         if (event->attr.type != type)
3839                 return 0;
3840         if (event->attr.config != event_id)
3841                 return 0;
3842 
3843         if (regs) {
3844                 if (event->attr.exclude_user && user_mode(regs))
3845                         return 0;
3846 
3847                 if (event->attr.exclude_kernel && !user_mode(regs))
3848                         return 0;
3849         }
3850 
3851         return 1;
3852 }
3853 
3854 static void perf_swevent_ctx_event(struct perf_event_context *ctx,
3855                                      enum perf_type_id type,
3856                                      u32 event_id, u64 nr, int nmi,
3857                                      struct perf_sample_data *data,
3858                                      struct pt_regs *regs)
3859 {
3860         struct perf_event *event;
3861 
3862         if (system_state != SYSTEM_RUNNING || list_empty(&ctx->event_list))
3863                 return;
3864 
3865         rcu_read_lock();
3866         list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
3867                 if (perf_swevent_match(event, type, event_id, regs))
3868                         perf_swevent_add(event, nr, nmi, data, regs);
3869         }
3870         rcu_read_unlock();
3871 }
3872 
3873 static int *perf_swevent_recursion_context(struct perf_cpu_context *cpuctx)
3874 {
3875         if (in_nmi())
3876                 return &cpuctx->recursion[3];
3877 
3878         if (in_irq())
3879                 return &cpuctx->recursion[2];
3880 
3881         if (in_softirq())
3882                 return &cpuctx->recursion[1];
3883 
3884         return &cpuctx->recursion[0];
3885 }
3886 
3887 static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
3888                                     u64 nr, int nmi,
3889                                     struct perf_sample_data *data,
3890                                     struct pt_regs *regs)
3891 {
3892         struct perf_cpu_context *cpuctx = &get_cpu_var(perf_cpu_context);
3893         int *recursion = perf_swevent_recursion_context(cpuctx);
3894         struct perf_event_context *ctx;
3895 
3896         if (*recursion)
3897                 goto out;
3898 
3899         (*recursion)++;
3900         barrier();
3901 
3902         perf_swevent_ctx_event(&cpuctx->ctx, type, event_id,
3903                                  nr, nmi, data, regs);
3904         rcu_read_lock();
3905         /*
3906          * doesn't really matter which of the child contexts the
3907          * events ends up in.
3908          */
3909         ctx = rcu_dereference(current->perf_event_ctxp);
3910         if (ctx)
3911                 perf_swevent_ctx_event(ctx, type, event_id, nr, nmi, data, regs);
3912         rcu_read_unlock();
3913 
3914         barrier();
3915         (*recursion)--;
3916 
3917 out:
3918         put_cpu_var(perf_cpu_context);
3919 }
3920 
3921 void __perf_sw_event(u32 event_id, u64 nr, int nmi,
3922                             struct pt_regs *regs, u64 addr)
3923 {
3924         struct perf_sample_data data = {
3925                 .addr = addr,
3926         };
3927 
3928         do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi,
3929                                 &data, regs);
3930 }
3931 
3932 static void perf_swevent_read(struct perf_event *event)
3933 {
3934 }
3935 
3936 static int perf_swevent_enable(struct perf_event *event)
3937 {
3938         struct hw_perf_event *hwc = &event->hw;
3939 
3940         if (hwc->sample_period) {
3941                 hwc->last_period = hwc->sample_period;
3942                 perf_swevent_set_period(event);
3943         }
3944         return 0;
3945 }
3946 
3947 static void perf_swevent_disable(struct perf_event *event)
3948 {
3949 }
3950 
3951 static const struct pmu perf_ops_generic = {
3952         .enable         = perf_swevent_enable,
3953         .disable        = perf_swevent_disable,
3954         .read           = perf_swevent_read,
3955         .unthrottle     = perf_swevent_unthrottle,
3956 };
3957 
3958 /*
3959  * hrtimer based swevent callback
3960  */
3961 
3962 static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
3963 {
3964         enum hrtimer_restart ret = HRTIMER_RESTART;
3965         struct perf_sample_data data;
3966         struct pt_regs *regs;
3967         struct perf_event *event;
3968         u64 period;
3969 
3970         event   = container_of(hrtimer, struct perf_event, hw.hrtimer);
3971         event->pmu->read(event);
3972 
3973         data.addr = 0;
3974         data.period = event->hw.last_period;
3975         regs = get_irq_regs();
3976         /*
3977          * In case we exclude kernel IPs or are somehow not in interrupt
3978          * context, provide the next best thing, the user IP.
3979          */
3980         if ((event->attr.exclude_kernel || !regs) &&
3981                         !event->attr.exclude_user)
3982                 regs = task_pt_regs(current);
3983 
3984         if (regs) {
3985                 if (!(event->attr.exclude_idle && current->pid == 0))
3986                         if (perf_event_overflow(event, 0, &data, regs))
3987                                 ret = HRTIMER_NORESTART;
3988         }
3989 
3990         period = max_t(u64, 10000, event->hw.sample_period);
3991         hrtimer_forward_now(hrtimer, ns_to_ktime(period));
3992 
3993         return ret;
3994 }
3995 
3996 static void perf_swevent_start_hrtimer(struct perf_event *event)
3997 {
3998         struct hw_perf_event *hwc = &event->hw;
3999 
4000         hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4001         hwc->hrtimer.function = perf_swevent_hrtimer;
4002         if (hwc->sample_period) {
4003                 u64 period;
4004 
4005                 if (hwc->remaining) {
4006                         if (hwc->remaining < 0)
4007                                 period = 10000;
4008                         else
4009                                 period = hwc->remaining;
4010                         hwc->remaining = 0;
4011                 } else {
4012                         period = max_t(u64, 10000, hwc->sample_period);
4013                 }
4014                 __hrtimer_start_range_ns(&hwc->hrtimer,
4015                                 ns_to_ktime(period), 0,
4016                                 HRTIMER_MODE_REL, 0);
4017         }
4018 }
4019 
4020 static void perf_swevent_cancel_hrtimer(struct perf_event *event)
4021 {
4022         struct hw_perf_event *hwc = &event->hw;
4023 
4024         if (hwc->sample_period) {
4025                 ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
4026                 hwc->remaining = ktime_to_ns(remaining);
4027 
4028                 hrtimer_cancel(&hwc->hrtimer);
4029         }
4030 }
4031 
4032 /*
4033  * Software event: cpu wall time clock
4034  */
4035 
4036 static void cpu_clock_perf_event_update(struct perf_event *event)
4037 {
4038         int cpu = raw_smp_processor_id();
4039         s64 prev;
4040         u64 now;
4041 
4042         now = cpu_clock(cpu);
4043         prev = atomic64_read(&event->hw.prev_count);
4044         atomic64_set(&event->hw.prev_count, now);
4045         atomic64_add(now - prev, &event->count);
4046 }
4047 
4048 static int cpu_clock_perf_event_enable(struct perf_event *event)
4049 {
4050         struct hw_perf_event *hwc = &event->hw;
4051         int cpu = raw_smp_processor_id();
4052 
4053         atomic64_set(&hwc->prev_count, cpu_clock(cpu));
4054         perf_swevent_start_hrtimer(event);
4055 
4056         return 0;
4057 }
4058 
4059 static void cpu_clock_perf_event_disable(struct perf_event *event)
4060 {
4061         perf_swevent_cancel_hrtimer(event);
4062         cpu_clock_perf_event_update(event);
4063 }
4064 
4065 static void cpu_clock_perf_event_read(struct perf_event *event)
4066 {
4067         cpu_clock_perf_event_update(event);
4068 }
4069 
4070 static const struct pmu perf_ops_cpu_clock = {
4071         .enable         = cpu_clock_perf_event_enable,
4072         .disable        = cpu_clock_perf_event_disable,
4073         .read           = cpu_clock_perf_event_read,
4074 };
4075 
4076 /*
4077  * Software event: task time clock
4078  */
4079 
4080 static void task_clock_perf_event_update(struct perf_event *event, u64 now)
4081 {
4082         u64 prev;
4083         s64 delta;
4084 
4085         prev = atomic64_xchg(&event->hw.prev_count, now);
4086         delta = now - prev;
4087         atomic64_add(delta, &event->count);
4088 }
4089 
4090 static int task_clock_perf_event_enable(struct perf_event *event)
4091 {
4092         struct hw_perf_event *hwc = &event->hw;
4093         u64 now;
4094 
4095         now = event->ctx->time;
4096 
4097         atomic64_set(&hwc->prev_count, now);
4098 
4099         perf_swevent_start_hrtimer(event);
4100 
4101         return 0;
4102 }
4103 
4104 static void task_clock_perf_event_disable(struct perf_event *event)
4105 {
4106         perf_swevent_cancel_hrtimer(event);
4107         task_clock_perf_event_update(event, event->ctx->time);
4108 
4109 }
4110 
4111 static void task_clock_perf_event_read(struct perf_event *event)
4112 {
4113         u64 time;
4114 
4115         if (!in_nmi()) {
4116                 update_context_time(event->ctx);
4117                 time = event->ctx->time;
4118         } else {
4119                 u64 now = perf_clock();
4120                 u64 delta = now - event->ctx->timestamp;
4121                 time = event->ctx->time + delta;
4122         }
4123 
4124         task_clock_perf_event_update(event, time);
4125 }
4126 
4127 static const struct pmu perf_ops_task_clock = {
4128         .enable         = task_clock_perf_event_enable,
4129         .disable        = task_clock_perf_event_disable,
4130         .read           = task_clock_perf_event_read,
4131 };
4132 
4133 #ifdef CONFIG_EVENT_PROFILE
4134 void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
4135                           int entry_size)
4136 {
4137         struct perf_raw_record raw = {
4138                 .size = entry_size,
4139                 .data = record,
4140         };
4141 
4142         struct perf_sample_data data = {
4143                 .addr = addr,
4144                 .raw = &raw,
4145         };
4146 
4147         struct pt_regs *regs = get_irq_regs();
4148 
4149         if (!regs)
4150                 regs = task_pt_regs(current);
4151 
4152         do_perf_sw_event(PERF_TYPE_TRACEPOINT, event_id, count, 1,
4153                                 &data, regs);
4154 }
4155 EXPORT_SYMBOL_GPL(perf_tp_event);
4156 
4157 extern int ftrace_profile_enable(int);
4158 extern void ftrace_profile_disable(int);
4159 
4160 static void tp_perf_event_destroy(struct perf_event *event)
4161 {
4162         ftrace_profile_disable(event->attr.config);
4163 }
4164 
4165 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4166 {
4167         /*
4168          * Raw tracepoint data is a severe data leak, only allow root to
4169          * have these.
4170          */
4171         if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
4172                         perf_paranoid_tracepoint_raw() &&
4173                         !capable(CAP_SYS_ADMIN))
4174                 return ERR_PTR(-EPERM);
4175 
4176         if (ftrace_profile_enable(event->attr.config))
4177                 return NULL;
4178 
4179         event->destroy = tp_perf_event_destroy;
4180 
4181         return &perf_ops_generic;
4182 }
4183 #else
4184 static const struct pmu *tp_perf_event_init(struct perf_event *event)
4185 {
4186         return NULL;
4187 }
4188 #endif
4189 
4190 atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
4191 
4192 static void sw_perf_event_destroy(struct perf_event *event)
4193 {
4194         u64 event_id = event->attr.config;
4195 
4196         WARN_ON(event->parent);
4197 
4198         atomic_dec(&perf_swevent_enabled[event_id]);
4199 }
4200 
4201 static const struct pmu *sw_perf_event_init(struct perf_event *event)
4202 {
4203         const struct pmu *pmu = NULL;
4204         u64 event_id = event->attr.config;
4205 
4206         /*
4207          * Software events (currently) can't in general distinguish
4208          * between user, kernel and hypervisor events.
4209          * However, context switches and cpu migrations are considered
4210          * to be kernel events, and page faults are never hypervisor
4211          * events.
4212          */
4213         switch (event_id) {
4214         case PERF_COUNT_SW_CPU_CLOCK:
4215                 pmu = &perf_ops_cpu_clock;
4216 
4217                 break;
4218         case PERF_COUNT_SW_TASK_CLOCK:
4219                 /*
4220                  * If the user instantiates this as a per-cpu event,
4221                  * use the cpu_clock event instead.
4222                  */
4223                 if (event->ctx->task)
4224                         pmu = &perf_ops_task_clock;
4225                 else
4226                         pmu = &perf_ops_cpu_clock;
4227 
4228                 break;
4229         case PERF_COUNT_SW_PAGE_FAULTS:
4230         case PERF_COUNT_SW_PAGE_FAULTS_MIN:
4231         case PERF_COUNT_SW_PAGE_FAULTS_MAJ:
4232         case PERF_COUNT_SW_CONTEXT_SWITCHES:
4233         case PERF_COUNT_SW_CPU_MIGRATIONS:
4234                 if (!event->parent) {
4235                         atomic_inc(&perf_swevent_enabled[event_id]);
4236                         event->destroy = sw_perf_event_destroy;
4237                 }
4238                 pmu = &perf_ops_generic;
4239                 break;
4240         }
4241 
4242         return pmu;
4243 }
4244 
4245 /*
4246  * Allocate and initialize a event structure
4247  */
4248 static struct perf_event *
4249 perf_event_alloc(struct perf_event_attr *attr,
4250                    int cpu,
4251                    struct perf_event_context *ctx,
4252                    struct perf_event *group_leader,
4253                    struct perf_event *parent_event,
4254                    gfp_t gfpflags)
4255 {
4256         const struct pmu *pmu;
4257         struct perf_event *event;
4258         struct hw_perf_event *hwc;
4259         long err;
4260 
4261         event = kzalloc(sizeof(*event), gfpflags);
4262         if (!event)
4263                 return ERR_PTR(-ENOMEM);
4264 
4265         /*
4266          * Single events are their own group leaders, with an
4267          * empty sibling list:
4268          */
4269         if (!group_leader)
4270                 group_leader = event;
4271 
4272         mutex_init(&event->child_mutex);
4273         INIT_LIST_HEAD(&event->child_list);
4274 
4275         INIT_LIST_HEAD(&event->group_entry);
4276         INIT_LIST_HEAD(&event->event_entry);
4277         INIT_LIST_HEAD(&event->sibling_list);
4278         init_waitqueue_head(&event->waitq);
4279 
4280         mutex_init(&event->mmap_mutex);
4281 
4282         event->cpu              = cpu;
4283         event->attr             = *attr;
4284         event->group_leader     = group_leader;
4285         event->pmu              = NULL;
4286         event->ctx              = ctx;
4287         event->oncpu            = -1;
4288 
4289         event->parent           = parent_event;
4290 
4291         event->ns               = get_pid_ns(current->nsproxy->pid_ns);
4292         event->id               = atomic64_inc_return(&perf_event_id);
4293 
4294         event->state            = PERF_EVENT_STATE_INACTIVE;
4295 
4296         if (attr->disabled)
4297                 event->state = PERF_EVENT_STATE_OFF;
4298 
4299         pmu = NULL;
4300 
4301         hwc = &event->hw;
4302         hwc->sample_period = attr->sample_period;
4303         if (attr->freq && attr->sample_freq)
4304                 hwc->sample_period = 1;
4305         hwc->last_period = hwc->sample_period;
4306 
4307         atomic64_set(&hwc->period_left, hwc->sample_period);
4308 
4309         /*
4310          * we currently do not support PERF_FORMAT_GROUP on inherited events
4311          */
4312         if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
4313                 goto done;
4314 
4315         switch (attr->type) {
4316         case PERF_TYPE_RAW:
4317         case PERF_TYPE_HARDWARE:
4318         case PERF_TYPE_HW_CACHE:
4319                 pmu = hw_perf_event_init(event);
4320                 break;
4321 
4322         case PERF_TYPE_SOFTWARE:
4323                 pmu = sw_perf_event_init(event);
4324                 break;
4325 
4326         case PERF_TYPE_TRACEPOINT:
4327                 pmu = tp_perf_event_init(event);
4328                 break;
4329 
4330         default:
4331                 break;
4332         }
4333 done:
4334         err = 0;
4335         if (!pmu)
4336                 err = -EINVAL;
4337         else if (IS_ERR(pmu))
4338                 err = PTR_ERR(pmu);
4339 
4340         if (err) {
4341                 if (event->ns)
4342                         put_pid_ns(event->ns);
4343                 kfree(event);
4344                 return ERR_PTR(err);
4345         }
4346 
4347         event->pmu = pmu;
4348 
4349         if (!event->parent) {
4350                 atomic_inc(&nr_events);
4351                 if (event->attr.mmap)
4352                         atomic_inc(&nr_mmap_events);
4353                 if (event->attr.comm)
4354                         atomic_inc(&nr_comm_events);
4355                 if (event->attr.task)
4356                         atomic_inc(&nr_task_events);
4357         }
4358 
4359         return event;
4360 }
4361 
4362 static int perf_copy_attr(struct perf_event_attr __user *uattr,
4363                           struct perf_event_attr *attr)
4364 {
4365         u32 size;
4366         int ret;
4367 
4368         if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
4369                 return -EFAULT;
4370 
4371         /*
4372          * zero the full structure, so that a short copy will be nice.
4373          */
4374         memset(attr, 0, sizeof(*attr));
4375 
4376         ret = get_user(size, &uattr->size);
4377         if (ret)
4378                 return ret;
4379 
4380         if (size > PAGE_SIZE)   /* silly large */
4381                 goto err_size;
4382 
4383         if (!size)              /* abi compat */
4384                 size = PERF_ATTR_SIZE_VER0;
4385 
4386         if (size < PERF_ATTR_SIZE_VER0)
4387                 goto err_size;
4388 
4389         /*
4390          * If we're handed a bigger struct than we know of,
4391          * ensure all the unknown bits are 0 - i.e. new
4392          * user-space does not rely on any kernel feature
4393          * extensions we dont know about yet.
4394          */
4395         if (size > sizeof(*attr)) {
4396                 unsigned char __user *addr;
4397                 unsigned char __user *end;
4398                 unsigned char val;
4399 
4400                 addr = (void __user *)uattr + sizeof(*attr);
4401                 end  = (void __user *)uattr + size;
4402 
4403                 for (; addr < end; addr++) {
4404                         ret = get_user(val, addr);
4405                         if (ret)
4406                                 return ret;
4407                         if (val)
4408                                 goto err_size;
4409                 }
4410                 size = sizeof(*attr);
4411         }
4412 
4413         ret = copy_from_user(attr, uattr, size);
4414         if (ret)
4415                 return -EFAULT;
4416 
4417         /*
4418          * If the type exists, the corresponding creation will verify
4419          * the attr->config.
4420          */
4421         if (attr->type >= PERF_TYPE_MAX)
4422                 return -EINVAL;
4423 
4424         if (attr->__reserved_1 || attr->__reserved_2 || attr->__reserved_3)
4425                 return -EINVAL;
4426 
4427         if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
4428                 return -EINVAL;
4429 
4430         if (attr->read_format & ~(PERF_FORMAT_MAX-1))
4431                 return -EINVAL;
4432 
4433 out:
4434         return ret;
4435 
4436 err_size:
4437         put_user(sizeof(*attr), &uattr->size);
4438         ret = -E2BIG;
4439         goto out;
4440 }
4441 
4442 int perf_event_set_output(struct perf_event *event, int output_fd)
4443 {
4444         struct perf_event *output_event = NULL;
4445         struct file *output_file = NULL;
4446         struct perf_event *old_output;
4447         int fput_needed = 0;
4448         int ret = -EINVAL;
4449 
4450         if (!output_fd)
4451                 goto set;
4452 
4453         output_file = fget_light(output_fd, &fput_needed);
4454         if (!output_file)
4455                 return -EBADF;
4456 
4457         if (output_file->f_op != &perf_fops)
4458                 goto out;
4459 
4460         output_event = output_file->private_data;
4461 
4462         /* Don't chain output fds */
4463         if (output_event->output)
4464                 goto out;
4465 
4466         /* Don't set an output fd when we already have an output channel */
4467         if (event->data)
4468                 goto out;
4469 
4470         atomic_long_inc(&output_file->f_count);
4471 
4472 set:
4473         mutex_lock(&event->mmap_mutex);
4474         old_output = event->output;
4475         rcu_assign_pointer(event->output, output_event);
4476         mutex_unlock(&event->mmap_mutex);
4477 
4478         if (old_output) {
4479                 /*
4480                  * we need to make sure no existing perf_output_*()
4481                  * is still referencing this event.
4482                  */
4483                 synchronize_rcu();
4484                 fput(old_output->filp);
4485         }
4486 
4487         ret = 0;
4488 out:
4489         fput_light(output_file, fput_needed);
4490         return ret;
4491 }
4492 
4493 /**
4494  * sys_perf_event_open - open a performance event, associate it to a task/cpu
4495  *
4496  * @attr_uptr:  event_id type attributes for monitoring/sampling
4497  * @pid:                target pid
4498  * @cpu:                target cpu
4499  * @group_fd:           group leader event fd
4500  */
4501 SYSCALL_DEFINE5(perf_event_open,
4502                 struct perf_event_attr __user *, attr_uptr,
4503                 pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
4504 {
4505         struct perf_event *event, *group_leader;
4506         struct perf_event_attr attr;
4507         struct perf_event_context *ctx;
4508         struct file *event_file = NULL;
4509         struct file *group_file = NULL;
4510         int event_fd;
4511         int fput_needed = 0;
4512         int err;
4513 
4514         /* for future expandability... */
4515         if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
4516                 return -EINVAL;
4517 
4518         err = perf_copy_attr(attr_uptr, &attr);
4519         if (err)
4520                 return err;
4521 
4522         if (!attr.exclude_kernel) {
4523                 if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
4524                         return -EACCES;
4525         }
4526 
4527         if (attr.freq) {
4528                 if (attr.sample_freq > sysctl_perf_event_sample_rate)
4529                         return -EINVAL;
4530         }
4531 
4532         event_fd = get_unused_fd_flags(O_RDWR);
4533         if (event_fd < 0)
4534                 return event_fd;
4535 
4536         /*
4537          * Get the target context (task or percpu):
4538          */
4539         ctx = find_get_context(pid, cpu);
4540         if (IS_ERR(ctx)) {
4541                 err = PTR_ERR(ctx);
4542                 goto err_fd;
4543         }
4544 
4545         /*
4546          * Look up the group leader (we will attach this event to it):
4547          */
4548         group_leader = NULL;
4549         if (group_fd != -1 && !(flags & PERF_FLAG_FD_NO_GROUP)) {
4550                 err = -EINVAL;
4551                 group_file = fget_light(group_fd, &fput_needed);
4552                 if (!group_file)
4553                         goto err_put_context;
4554                 if (group_file->f_op != &perf_fops)
4555                         goto err_put_context;
4556 
4557                 group_leader = group_file->private_data;
4558                 /*
4559                  * Do not allow a recursive hierarchy (this new sibling
4560                  * becoming part of another group-sibling):
4561                  */
4562                 if (group_leader->group_leader != group_leader)
4563                         goto err_put_context;
4564                 /*
4565                  * Do not allow to attach to a group in a different
4566                  * task or CPU context:
4567                  */
4568                 if (group_leader->ctx != ctx)
4569                         goto err_put_context;
4570                 /*
4571                  * Only a group leader can be exclusive or pinned
4572                  */
4573                 if (attr.exclusive || attr.pinned)
4574                         goto err_put_context;
4575         }
4576 
4577         event = perf_event_alloc(&attr, cpu, ctx, group_leader,
4578                                      NULL, GFP_KERNEL);
4579         err = PTR_ERR(event);
4580         if (IS_ERR(event))
4581                 goto err_put_context;
4582 
4583         event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
4584         if (IS_ERR(event_file)) {
4585                 err = PTR_ERR(event_file);
4586                 goto err_free_put_context;
4587         }
4588 
4589         if (flags & PERF_FLAG_FD_OUTPUT) {
4590                 err = perf_event_set_output(event, group_fd);
4591                 if (err)
4592                         goto err_fput_free_put_context;
4593         }
4594 
4595         event->filp = event_file;
4596         WARN_ON_ONCE(ctx->parent_ctx);
4597         mutex_lock(&ctx->mutex);
4598         perf_install_in_context(ctx, event, cpu);
4599         ++ctx->generation;
4600         mutex_unlock(&ctx->mutex);
4601 
4602         event->owner = current;
4603         get_task_struct(current);
4604         mutex_lock(&current->perf_event_mutex);
4605         list_add_tail(&event->owner_entry, &current->perf_event_list);
4606         mutex_unlock(&current->perf_event_mutex);
4607 
4608         fput_light(group_file, fput_needed);
4609         fd_install(event_fd, event_file);
4610         return event_fd;
4611 
4612 err_fput_free_put_context:
4613         fput(event_file);
4614 err_free_put_context:
4615         free_event(event);
4616 err_put_context:
4617         fput_light(group_file, fput_needed);
4618         put_ctx(ctx);
4619 err_fd:
4620         put_unused_fd(event_fd);
4621         return err;
4622 }
4623 
4624 /*
4625  * inherit a event from parent task to child task:
4626  */
4627 static struct perf_event *
4628 inherit_event(struct perf_event *parent_event,
4629               struct task_struct *parent,
4630               struct perf_event_context *parent_ctx,
4631               struct task_struct *child,
4632               struct perf_event *group_leader,
4633               struct perf_event_context *child_ctx)
4634 {
4635         struct perf_event *child_event;
4636 
4637         /*
4638          * Instead of creating recursive hierarchies of events,
4639          * we link inherited events back to the original parent,
4640          * which has a filp for sure, which we use as the reference
4641          * count:
4642          */
4643         if (parent_event->parent)
4644                 parent_event = parent_event->parent;
4645 
4646         child_event = perf_event_alloc(&parent_event->attr,
4647                                            parent_event->cpu, child_ctx,
4648                                            group_leader, parent_event,
4649                                            GFP_KERNEL);
4650         if (IS_ERR(child_event))
4651                 return child_event;
4652         get_ctx(child_ctx);
4653 
4654         /*
4655          * Make the child state follow the state of the parent event,
4656          * not its attr.disabled bit.  We hold the parent's mutex,
4657          * so we won't race with perf_event_{en, dis}able_family.
4658          */
4659         if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
4660                 child_event->state = PERF_EVENT_STATE_INACTIVE;
4661         else
4662                 child_event->state = PERF_EVENT_STATE_OFF;
4663 
4664         if (parent_event->attr.freq)
4665                 child_event->hw.sample_period = parent_event->hw.sample_period;
4666 
4667         /*
4668          * Link it up in the child's context:
4669          */
4670         add_event_to_ctx(child_event, child_ctx);
4671 
4672         /*
4673          * Get a reference to the parent filp - we will fput it
4674          * when the child event exits. This is safe to do because
4675          * we are in the parent and we know that the filp still
4676          * exists and has a nonzero count:
4677          */
4678         atomic_long_inc(&parent_event->filp->f_count);
4679 
4680         /*
4681          * Link this into the parent event's child list
4682          */
4683         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4684         mutex_lock(&parent_event->child_mutex);
4685         list_add_tail(&child_event->child_list, &parent_event->child_list);
4686         mutex_unlock(&parent_event->child_mutex);
4687 
4688         return child_event;
4689 }
4690 
4691 static int inherit_group(struct perf_event *parent_event,
4692               struct task_struct *parent,
4693               struct perf_event_context *parent_ctx,
4694               struct task_struct *child,
4695               struct perf_event_context *child_ctx)
4696 {
4697         struct perf_event *leader;
4698         struct perf_event *sub;
4699         struct perf_event *child_ctr;
4700 
4701         leader = inherit_event(parent_event, parent, parent_ctx,
4702                                  child, NULL, child_ctx);
4703         if (IS_ERR(leader))
4704                 return PTR_ERR(leader);
4705         list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
4706                 child_ctr = inherit_event(sub, parent, parent_ctx,
4707                                             child, leader, child_ctx);
4708                 if (IS_ERR(child_ctr))
4709                         return PTR_ERR(child_ctr);
4710         }
4711         return 0;
4712 }
4713 
4714 static void sync_child_event(struct perf_event *child_event,
4715                                struct task_struct *child)
4716 {
4717         struct perf_event *parent_event = child_event->parent;
4718         u64 child_val;
4719 
4720         if (child_event->attr.inherit_stat)
4721                 perf_event_read_event(child_event, child);
4722 
4723         child_val = atomic64_read(&child_event->count);
4724 
4725         /*
4726          * Add back the child's count to the parent's count:
4727          */
4728         atomic64_add(child_val, &parent_event->count);
4729         atomic64_add(child_event->total_time_enabled,
4730                      &parent_event->child_total_time_enabled);
4731         atomic64_add(child_event->total_time_running,
4732                      &parent_event->child_total_time_running);
4733 
4734         /*
4735          * Remove this event from the parent's list
4736          */
4737         WARN_ON_ONCE(parent_event->ctx->parent_ctx);
4738         mutex_lock(&parent_event->child_mutex);
4739         list_del_init(&child_event->child_list);
4740         mutex_unlock(&parent_event->child_mutex);
4741 
4742         /*
4743          * Release the parent event, if this was the last
4744          * reference to it.
4745          */
4746         fput(parent_event->filp);
4747 }
4748 
4749 static void
4750 __perf_event_exit_task(struct perf_event *child_event,
4751                          struct perf_event_context *child_ctx,
4752                          struct task_struct *child)
4753 {
4754         struct perf_event *parent_event;
4755 
4756         update_event_times(child_event);
4757         perf_event_remove_from_context(child_event);
4758 
4759         parent_event = child_event->parent;
4760         /*
4761          * It can happen that parent exits first, and has events
4762          * that are still around due to the child reference. These
4763          * events need to be zapped - but otherwise linger.
4764          */
4765         if (parent_event) {
4766                 sync_child_event(child_event, child);
4767                 free_event(child_event);
4768         }
4769 }
4770 
4771 /*
4772  * When a child task exits, feed back event values to parent events.
4773  */
4774 void perf_event_exit_task(struct task_struct *child)
4775 {
4776         struct perf_event *child_event, *tmp;
4777         struct perf_event_context *child_ctx;
4778         unsigned long flags;
4779 
4780         if (likely(!child->perf_event_ctxp)) {
4781                 perf_event_task(child, NULL, 0);
4782                 return;
4783         }
4784 
4785         local_irq_save(flags);
4786         /*
4787          * We can't reschedule here because interrupts are disabled,
4788          * and either child is current or it is a task that can't be
4789          * scheduled, so we are now safe from rescheduling changing
4790          * our context.
4791          */
4792         child_ctx = child->perf_event_ctxp;
4793         __perf_event_task_sched_out(child_ctx);
4794 
4795         /*
4796          * Take the context lock here so that if find_get_context is
4797          * reading child->perf_event_ctxp, we wait until it has
4798          * incremented the context's refcount before we do put_ctx below.
4799          */
4800         spin_lock(&child_ctx->lock);
4801         child->perf_event_ctxp = NULL;
4802         /*
4803          * If this context is a clone; unclone it so it can't get
4804          * swapped to another process while we're removing all
4805          * the events from it.
4806          */
4807         unclone_ctx(child_ctx);
4808         spin_unlock_irqrestore(&child_ctx->lock, flags);
4809 
4810         /*
4811          * Report the task dead after unscheduling the events so that we
4812          * won't get any samples after PERF_RECORD_EXIT. We can however still
4813          * get a few PERF_RECORD_READ events.
4814          */
4815         perf_event_task(child, child_ctx, 0);
4816 
4817         /*
4818          * We can recurse on the same lock type through:
4819          *
4820          *   __perf_event_exit_task()
4821          *     sync_child_event()
4822          *       fput(parent_event->filp)
4823          *         perf_release()
4824          *           mutex_lock(&ctx->mutex)
4825          *
4826          * But since its the parent context it won't be the same instance.
4827          */
4828         mutex_lock_nested(&child_ctx->mutex, SINGLE_DEPTH_NESTING);
4829 
4830 again:
4831         list_for_each_entry_safe(child_event, tmp, &child_ctx->group_list,
4832                                  group_entry)
4833                 __perf_event_exit_task(child_event, child_ctx, child);
4834 
4835         /*
4836          * If the last event was a group event, it will have appended all
4837          * its siblings to the list, but we obtained 'tmp' before that which
4838          * will still point to the list head terminating the iteration.
4839          */
4840         if (!list_empty(&child_ctx->group_list))
4841                 goto again;
4842 
4843         mutex_unlock(&child_ctx->mutex);
4844 
4845         put_ctx(child_ctx);
4846 }
4847 
4848 /*
4849  * free an unexposed, unused context as created by inheritance by
4850  * init_task below, used by fork() in case of fail.
4851  */
4852 void perf_event_free_task(struct task_struct *task)
4853 {
4854         struct perf_event_context *ctx = task->perf_event_ctxp;
4855         struct perf_event *event, *tmp;
4856 
4857         if (!ctx)
4858                 return;
4859 
4860         mutex_lock(&ctx->mutex);
4861 again:
4862         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry) {
4863                 struct perf_event *parent = event->parent;
4864 
4865                 if (WARN_ON_ONCE(!parent))
4866                         continue;
4867 
4868                 mutex_lock(&parent->child_mutex);
4869                 list_del_init(&event->child_list);
4870                 mutex_unlock(&parent->child_mutex);
4871 
4872                 fput(parent->filp);
4873 
4874                 list_del_event(event, ctx);
4875                 free_event(event);
4876         }
4877 
4878         if (!list_empty(&ctx->group_list))
4879                 goto again;
4880 
4881         mutex_unlock(&ctx->mutex);
4882 
4883         put_ctx(ctx);
4884 }
4885 
4886 /*
4887  * Initialize the perf_event context in task_struct
4888  */
4889 int perf_event_init_task(struct task_struct *child)
4890 {
4891         struct perf_event_context *child_ctx, *parent_ctx;
4892         struct perf_event_context *cloned_ctx;
4893         struct perf_event *event;
4894         struct task_struct *parent = current;
4895         int inherited_all = 1;
4896         int ret = 0;
4897 
4898         child->perf_event_ctxp = NULL;
4899 
4900         mutex_init(&child->perf_event_mutex);
4901         INIT_LIST_HEAD(&child->perf_event_list);
4902 
4903         if (likely(!parent->perf_event_ctxp))
4904                 return 0;
4905 
4906         /*
4907          * This is executed from the parent task context, so inherit
4908          * events that have been marked for cloning.
4909          * First allocate and initialize a context for the child.
4910          */
4911 
4912         child_ctx = kmalloc(sizeof(struct perf_event_context), GFP_KERNEL);
4913         if (!child_ctx)
4914                 return -ENOMEM;
4915 
4916         __perf_event_init_context(child_ctx, child);
4917         child->perf_event_ctxp = child_ctx;
4918         get_task_struct(child);
4919 
4920         /*
4921          * If the parent's context is a clone, pin it so it won't get
4922          * swapped under us.
4923          */
4924         parent_ctx = perf_pin_task_context(parent);
4925 
4926         /*
4927          * No need to check if parent_ctx != NULL here; since we saw
4928          * it non-NULL earlier, the only reason for it to become NULL
4929          * is if we exit, and since we're currently in the middle of
4930          * a fork we can't be exiting at the same time.
4931          */
4932 
4933         /*
4934          * Lock the parent list. No need to lock the child - not PID
4935          * hashed yet and not running, so nobody can access it.
4936          */
4937         mutex_lock(&parent_ctx->mutex);
4938 
4939         /*
4940          * We dont have to disable NMIs - we are only looking at
4941          * the list, not manipulating it:
4942          */
4943         list_for_each_entry(event, &parent_ctx->group_list, group_entry) {
4944 
4945                 if (!event->attr.inherit) {
4946                         inherited_all = 0;
4947                         continue;
4948                 }
4949 
4950                 ret = inherit_group(event, parent, parent_ctx,
4951                                              child, child_ctx);
4952                 if (ret) {
4953                         inherited_all = 0;
4954                         break;
4955                 }
4956         }
4957 
4958         if (inherited_all) {
4959                 /*
4960                  * Mark the child context as a clone of the parent
4961                  * context, or of whatever the parent is a clone of.
4962                  * Note that if the parent is a clone, it could get
4963                  * uncloned at any point, but that doesn't matter
4964                  * because the list of events and the generation
4965                  * count can't have changed since we took the mutex.
4966                  */
4967                 cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
4968                 if (cloned_ctx) {
4969                         child_ctx->parent_ctx = cloned_ctx;
4970                         child_ctx->parent_gen = parent_ctx->parent_gen;
4971                 } else {
4972                         child_ctx->parent_ctx = parent_ctx;
4973                         child_ctx->parent_gen = parent_ctx->generation;
4974                 }
4975                 get_ctx(child_ctx->parent_ctx);
4976         }
4977 
4978         mutex_unlock(&parent_ctx->mutex);
4979 
4980         perf_unpin_context(parent_ctx);
4981 
4982         return ret;
4983 }
4984 
4985 static void __init perf_event_init_all_cpus(void)
4986 {
4987         int cpu;
4988         struct perf_cpu_context *cpuctx;
4989 
4990         for_each_possible_cpu(cpu) {
4991                 cpuctx = &per_cpu(perf_cpu_context, cpu);
4992                 __perf_event_init_context(&cpuctx->ctx, NULL);
4993         }
4994 }
4995 
4996 static void __cpuinit perf_event_init_cpu(int cpu)
4997 {
4998         struct perf_cpu_context *cpuctx;
4999 
5000         cpuctx = &per_cpu(perf_cpu_context, cpu);
5001 
5002         spin_lock(&perf_resource_lock);
5003         cpuctx->max_pertask = perf_max_events - perf_reserved_percpu;
5004         spin_unlock(&perf_resource_lock);
5005 
5006         hw_perf_event_setup(cpu);
5007 }
5008 
5009 #ifdef CONFIG_HOTPLUG_CPU
5010 static void __perf_event_exit_cpu(void *info)
5011 {
5012         struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
5013         struct perf_event_context *ctx = &cpuctx->ctx;
5014         struct perf_event *event, *tmp;
5015 
5016         list_for_each_entry_safe(event, tmp, &ctx->group_list, group_entry)
5017                 __perf_event_remove_from_context(event);
5018 }
5019 static void perf_event_exit_cpu(int cpu)
5020 {
5021         struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu);
5022         struct perf_event_context *ctx = &cpuctx->ctx;
5023 
5024         mutex_lock(&ctx->mutex);
5025         smp_call_function_single(cpu, __perf_event_exit_cpu, NULL, 1);
5026         mutex_unlock(&ctx->mutex);
5027 }
5028 #else
5029 static inline void perf_event_exit_cpu(int cpu) { }
5030 #endif
5031 
5032 static int __cpuinit
5033 perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
5034 {
5035         unsigned int cpu = (long)hcpu;
5036 
5037         switch (action) {
5038 
5039         case CPU_UP_PREPARE:
5040         case CPU_UP_PREPARE_FROZEN:
5041                 perf_event_init_cpu(cpu);
5042                 break;
5043 
5044         case CPU_ONLINE:
5045         case CPU_ONLINE_FROZEN:
5046                 hw_perf_event_setup_online(cpu);
5047                 break;
5048 
5049         case CPU_DOWN_PREPARE:
5050         case CPU_DOWN_PREPARE_FROZEN:
5051                 perf_event_exit_cpu(cpu);
5052                 break;
5053 
5054         default:
5055                 break;
5056         }
5057 
5058         return NOTIFY_OK;
5059 }
5060 
5061 /*
5062  * This has to have a higher priority than migration_notifier in sched.c.
5063  */
5064 static struct notifier_block __cpuinitdata perf_cpu_nb = {
5065         .notifier_call          = perf_cpu_notify,
5066         .priority               = 20,
5067 };
5068 
5069 void __init perf_event_init(void)
5070 {
5071         perf_event_init_all_cpus();
5072         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_UP_PREPARE,
5073                         (void *)(long)smp_processor_id());
5074         perf_cpu_notify(&perf_cpu_nb, (unsigned long)CPU_ONLINE,
5075                         (void *)(long)smp_processor_id());
5076         register_cpu_notifier(&perf_cpu_nb);
5077 }
5078 
5079 static ssize_t perf_show_reserve_percpu(struct sysdev_class *class, char *buf)
5080 {
5081         return sprintf(buf, "%d\n", perf_reserved_percpu);
5082 }
5083 
5084 static ssize_t
5085 perf_set_reserve_percpu(struct sysdev_class *class,
5086                         const char *buf,
5087                         size_t count)
5088 {
5089         struct perf_cpu_context *cpuctx;
5090         unsigned long val;
5091         int err, cpu, mpt;
5092 
5093         err = strict_strtoul(buf, 10, &val);
5094         if (err)
5095                 return err;
5096         if (val > perf_max_events)
5097                 return -EINVAL;
5098 
5099         spin_lock(&perf_resource_lock);
5100         perf_reserved_percpu = val;
5101         for_each_online_cpu(cpu) {
5102                 cpuctx = &per_cpu(perf_cpu_context, cpu);
5103                 spin_lock_irq(&cpuctx->ctx.lock);
5104                 mpt = min(perf_max_events - cpuctx->ctx.nr_events,
5105                           perf_max_events - perf_reserved_percpu);
5106                 cpuctx->max_pertask = mpt;
5107                 spin_unlock_irq(&cpuctx->ctx.lock);
5108         }
5109         spin_unlock(&perf_resource_lock);
5110 
5111         return count;
5112 }
5113 
5114 static ssize_t perf_show_overcommit(struct sysdev_class *class, char *buf)
5115 {
5116         return sprintf(buf, "%d\n", perf_overcommit);
5117 }
5118 
5119 static ssize_t
5120 perf_set_overcommit(struct sysdev_class *class, const char *buf, size_t count)
5121 {
5122         unsigned long val;
5123         int err;
5124 
5125         err = strict_strtoul(buf, 10, &val);
5126         if (err)
5127                 return err;
5128         if (val > 1)
5129                 return -EINVAL;
5130 
5131         spin_lock(&perf_resource_lock);
5132         perf_overcommit = val;
5133         spin_unlock(&perf_resource_lock);
5134 
5135         return count;
5136 }
5137 
5138 static SYSDEV_CLASS_ATTR(
5139                                 reserve_percpu,
5140                                 0644,
5141                                 perf_show_reserve_percpu,
5142                                 perf_set_reserve_percpu
5143                         );
5144 
5145 static SYSDEV_CLASS_ATTR(
5146                                 overcommit,
5147                                 0644,
5148                                 perf_show_overcommit,
5149                                 perf_set_overcommit
5150                         );
5151 
5152 static struct attribute *perfclass_attrs[] = {
5153         &attr_reserve_percpu.attr,
5154         &attr_overcommit.attr,
5155         NULL
5156 };
5157 
5158 static struct attribute_group perfclass_attr_group = {
5159         .attrs                  = perfclass_attrs,
5160         .name                   = "perf_events",
5161 };
5162 
5163 static int __init perf_event_sysfs_init(void)
5164 {
5165         return sysfs_create_group(&cpu_sysdev_class.kset.kobj,
5166                                   &perfclass_attr_group);
5167 }
5168 device_initcall(perf_event_sysfs_init);
5169 

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

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp