Merge branch 'core-rcu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
authorLinus Torvalds <torvalds@linux-foundation.org>
Fri, 20 May 2011 01:14:34 +0000 (18:14 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 20 May 2011 01:14:34 +0000 (18:14 -0700)
* 'core-rcu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (78 commits)
  Revert "rcu: Decrease memory-barrier usage based on semi-formal proof"
  net,rcu: convert call_rcu(prl_entry_destroy_rcu) to kfree
  batman,rcu: convert call_rcu(softif_neigh_free_rcu) to kfree_rcu
  batman,rcu: convert call_rcu(neigh_node_free_rcu) to kfree()
  batman,rcu: convert call_rcu(gw_node_free_rcu) to kfree_rcu
  net,rcu: convert call_rcu(kfree_tid_tx) to kfree_rcu()
  net,rcu: convert call_rcu(xt_osf_finger_free_rcu) to kfree_rcu()
  net/mac80211,rcu: convert call_rcu(work_free_rcu) to kfree_rcu()
  net,rcu: convert call_rcu(wq_free_rcu) to kfree_rcu()
  net,rcu: convert call_rcu(phonet_device_rcu_free) to kfree_rcu()
  perf,rcu: convert call_rcu(swevent_hlist_release_rcu) to kfree_rcu()
  perf,rcu: convert call_rcu(free_ctx) to kfree_rcu()
  net,rcu: convert call_rcu(__nf_ct_ext_free_rcu) to kfree_rcu()
  net,rcu: convert call_rcu(net_generic_release) to kfree_rcu()
  net,rcu: convert call_rcu(netlbl_unlhsh_free_addr6) to kfree_rcu()
  net,rcu: convert call_rcu(netlbl_unlhsh_free_addr4) to kfree_rcu()
  security,rcu: convert call_rcu(sel_netif_free) to kfree_rcu()
  net,rcu: convert call_rcu(xps_dev_maps_release) to kfree_rcu()
  net,rcu: convert call_rcu(xps_map_release) to kfree_rcu()
  net,rcu: convert call_rcu(rps_map_release) to kfree_rcu()
  ...

1  2 
init/Kconfig
kernel/events/core.c
lib/Kconfig.debug

diff --cc init/Kconfig
Simple merge
index 0fc34a370ba41be068f160e07875abd0bf1bdbdc,0000000000000000000000000000000000000000..c09767f7db3e20e4a510c7a51fba7d662e967659
mode 100644,000000..100644
--- /dev/null
@@@ -1,7455 -1,0 +1,7439 @@@
- static void free_ctx(struct rcu_head *head)
- {
-       struct perf_event_context *ctx;
-       ctx = container_of(head, struct perf_event_context, rcu_head);
-       kfree(ctx);
- }
 +/*
 + * Performance events core code:
 + *
 + *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
 + *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
 + *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
 + *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
 + *
 + * For licensing details see kernel-base/COPYING
 + */
 +
 +#include <linux/fs.h>
 +#include <linux/mm.h>
 +#include <linux/cpu.h>
 +#include <linux/smp.h>
 +#include <linux/idr.h>
 +#include <linux/file.h>
 +#include <linux/poll.h>
 +#include <linux/slab.h>
 +#include <linux/hash.h>
 +#include <linux/sysfs.h>
 +#include <linux/dcache.h>
 +#include <linux/percpu.h>
 +#include <linux/ptrace.h>
 +#include <linux/reboot.h>
 +#include <linux/vmstat.h>
 +#include <linux/device.h>
 +#include <linux/vmalloc.h>
 +#include <linux/hardirq.h>
 +#include <linux/rculist.h>
 +#include <linux/uaccess.h>
 +#include <linux/syscalls.h>
 +#include <linux/anon_inodes.h>
 +#include <linux/kernel_stat.h>
 +#include <linux/perf_event.h>
 +#include <linux/ftrace_event.h>
 +#include <linux/hw_breakpoint.h>
 +
 +#include <asm/irq_regs.h>
 +
 +struct remote_function_call {
 +      struct task_struct      *p;
 +      int                     (*func)(void *info);
 +      void                    *info;
 +      int                     ret;
 +};
 +
 +static void remote_function(void *data)
 +{
 +      struct remote_function_call *tfc = data;
 +      struct task_struct *p = tfc->p;
 +
 +      if (p) {
 +              tfc->ret = -EAGAIN;
 +              if (task_cpu(p) != smp_processor_id() || !task_curr(p))
 +                      return;
 +      }
 +
 +      tfc->ret = tfc->func(tfc->info);
 +}
 +
 +/**
 + * task_function_call - call a function on the cpu on which a task runs
 + * @p:                the task to evaluate
 + * @func:     the function to be called
 + * @info:     the function call argument
 + *
 + * Calls the function @func when the task is currently running. This might
 + * be on the current CPU, which just calls the function directly
 + *
 + * returns: @func return value, or
 + *        -ESRCH  - when the process isn't running
 + *        -EAGAIN - when the process moved away
 + */
 +static int
 +task_function_call(struct task_struct *p, int (*func) (void *info), void *info)
 +{
 +      struct remote_function_call data = {
 +              .p      = p,
 +              .func   = func,
 +              .info   = info,
 +              .ret    = -ESRCH, /* No such (running) process */
 +      };
 +
 +      if (task_curr(p))
 +              smp_call_function_single(task_cpu(p), remote_function, &data, 1);
 +
 +      return data.ret;
 +}
 +
 +/**
 + * cpu_function_call - call a function on the cpu
 + * @func:     the function to be called
 + * @info:     the function call argument
 + *
 + * Calls the function @func on the remote cpu.
 + *
 + * returns: @func return value or -ENXIO when the cpu is offline
 + */
 +static int cpu_function_call(int cpu, int (*func) (void *info), void *info)
 +{
 +      struct remote_function_call data = {
 +              .p      = NULL,
 +              .func   = func,
 +              .info   = info,
 +              .ret    = -ENXIO, /* No such CPU */
 +      };
 +
 +      smp_call_function_single(cpu, remote_function, &data, 1);
 +
 +      return data.ret;
 +}
 +
 +#define PERF_FLAG_ALL (PERF_FLAG_FD_NO_GROUP |\
 +                     PERF_FLAG_FD_OUTPUT  |\
 +                     PERF_FLAG_PID_CGROUP)
 +
 +enum event_type_t {
 +      EVENT_FLEXIBLE = 0x1,
 +      EVENT_PINNED = 0x2,
 +      EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
 +};
 +
 +/*
 + * perf_sched_events : >0 events exist
 + * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
 + */
 +struct jump_label_key perf_sched_events __read_mostly;
 +static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
 +
 +static atomic_t nr_mmap_events __read_mostly;
 +static atomic_t nr_comm_events __read_mostly;
 +static atomic_t nr_task_events __read_mostly;
 +
 +static LIST_HEAD(pmus);
 +static DEFINE_MUTEX(pmus_lock);
 +static struct srcu_struct pmus_srcu;
 +
 +/*
 + * perf event paranoia level:
 + *  -1 - not paranoid at all
 + *   0 - disallow raw tracepoint access for unpriv
 + *   1 - disallow cpu events for unpriv
 + *   2 - disallow kernel profiling for unpriv
 + */
 +int sysctl_perf_event_paranoid __read_mostly = 1;
 +
 +/* Minimum for 512 kiB + 1 user control page */
 +int sysctl_perf_event_mlock __read_mostly = 512 + (PAGE_SIZE / 1024); /* 'free' kiB per user */
 +
 +/*
 + * max perf event sample rate
 + */
 +#define DEFAULT_MAX_SAMPLE_RATE 100000
 +int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
 +static int max_samples_per_tick __read_mostly =
 +      DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
 +
 +int perf_proc_update_handler(struct ctl_table *table, int write,
 +              void __user *buffer, size_t *lenp,
 +              loff_t *ppos)
 +{
 +      int ret = proc_dointvec(table, write, buffer, lenp, ppos);
 +
 +      if (ret || !write)
 +              return ret;
 +
 +      max_samples_per_tick = DIV_ROUND_UP(sysctl_perf_event_sample_rate, HZ);
 +
 +      return 0;
 +}
 +
 +static atomic64_t perf_event_id;
 +
 +static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
 +                            enum event_type_t event_type);
 +
 +static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
 +                           enum event_type_t event_type,
 +                           struct task_struct *task);
 +
 +static void update_context_time(struct perf_event_context *ctx);
 +static u64 perf_event_time(struct perf_event *event);
 +
 +void __weak perf_event_print_debug(void)      { }
 +
 +extern __weak const char *perf_pmu_name(void)
 +{
 +      return "pmu";
 +}
 +
 +static inline u64 perf_clock(void)
 +{
 +      return local_clock();
 +}
 +
 +static inline struct perf_cpu_context *
 +__get_cpu_context(struct perf_event_context *ctx)
 +{
 +      return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
 +}
 +
 +#ifdef CONFIG_CGROUP_PERF
 +
 +/*
 + * Must ensure cgroup is pinned (css_get) before calling
 + * this function. In other words, we cannot call this function
 + * if there is no cgroup event for the current CPU context.
 + */
 +static inline struct perf_cgroup *
 +perf_cgroup_from_task(struct task_struct *task)
 +{
 +      return container_of(task_subsys_state(task, perf_subsys_id),
 +                      struct perf_cgroup, css);
 +}
 +
 +static inline bool
 +perf_cgroup_match(struct perf_event *event)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +      struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 +
 +      return !event->cgrp || event->cgrp == cpuctx->cgrp;
 +}
 +
 +static inline void perf_get_cgroup(struct perf_event *event)
 +{
 +      css_get(&event->cgrp->css);
 +}
 +
 +static inline void perf_put_cgroup(struct perf_event *event)
 +{
 +      css_put(&event->cgrp->css);
 +}
 +
 +static inline void perf_detach_cgroup(struct perf_event *event)
 +{
 +      perf_put_cgroup(event);
 +      event->cgrp = NULL;
 +}
 +
 +static inline int is_cgroup_event(struct perf_event *event)
 +{
 +      return event->cgrp != NULL;
 +}
 +
 +static inline u64 perf_cgroup_event_time(struct perf_event *event)
 +{
 +      struct perf_cgroup_info *t;
 +
 +      t = per_cpu_ptr(event->cgrp->info, event->cpu);
 +      return t->time;
 +}
 +
 +static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
 +{
 +      struct perf_cgroup_info *info;
 +      u64 now;
 +
 +      now = perf_clock();
 +
 +      info = this_cpu_ptr(cgrp->info);
 +
 +      info->time += now - info->timestamp;
 +      info->timestamp = now;
 +}
 +
 +static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
 +{
 +      struct perf_cgroup *cgrp_out = cpuctx->cgrp;
 +      if (cgrp_out)
 +              __update_cgrp_time(cgrp_out);
 +}
 +
 +static inline void update_cgrp_time_from_event(struct perf_event *event)
 +{
 +      struct perf_cgroup *cgrp;
 +
 +      /*
 +       * ensure we access cgroup data only when needed and
 +       * when we know the cgroup is pinned (css_get)
 +       */
 +      if (!is_cgroup_event(event))
 +              return;
 +
 +      cgrp = perf_cgroup_from_task(current);
 +      /*
 +       * Do not update time when cgroup is not active
 +       */
 +      if (cgrp == event->cgrp)
 +              __update_cgrp_time(event->cgrp);
 +}
 +
 +static inline void
 +perf_cgroup_set_timestamp(struct task_struct *task,
 +                        struct perf_event_context *ctx)
 +{
 +      struct perf_cgroup *cgrp;
 +      struct perf_cgroup_info *info;
 +
 +      /*
 +       * ctx->lock held by caller
 +       * ensure we do not access cgroup data
 +       * unless we have the cgroup pinned (css_get)
 +       */
 +      if (!task || !ctx->nr_cgroups)
 +              return;
 +
 +      cgrp = perf_cgroup_from_task(task);
 +      info = this_cpu_ptr(cgrp->info);
 +      info->timestamp = ctx->timestamp;
 +}
 +
 +#define PERF_CGROUP_SWOUT     0x1 /* cgroup switch out every event */
 +#define PERF_CGROUP_SWIN      0x2 /* cgroup switch in events based on task */
 +
 +/*
 + * reschedule events based on the cgroup constraint of task.
 + *
 + * mode SWOUT : schedule out everything
 + * mode SWIN : schedule in based on cgroup for next
 + */
 +void perf_cgroup_switch(struct task_struct *task, int mode)
 +{
 +      struct perf_cpu_context *cpuctx;
 +      struct pmu *pmu;
 +      unsigned long flags;
 +
 +      /*
 +       * disable interrupts to avoid geting nr_cgroup
 +       * changes via __perf_event_disable(). Also
 +       * avoids preemption.
 +       */
 +      local_irq_save(flags);
 +
 +      /*
 +       * we reschedule only in the presence of cgroup
 +       * constrained events.
 +       */
 +      rcu_read_lock();
 +
 +      list_for_each_entry_rcu(pmu, &pmus, entry) {
 +
 +              cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
 +
 +              perf_pmu_disable(cpuctx->ctx.pmu);
 +
 +              /*
 +               * perf_cgroup_events says at least one
 +               * context on this CPU has cgroup events.
 +               *
 +               * ctx->nr_cgroups reports the number of cgroup
 +               * events for a context.
 +               */
 +              if (cpuctx->ctx.nr_cgroups > 0) {
 +
 +                      if (mode & PERF_CGROUP_SWOUT) {
 +                              cpu_ctx_sched_out(cpuctx, EVENT_ALL);
 +                              /*
 +                               * must not be done before ctxswout due
 +                               * to event_filter_match() in event_sched_out()
 +                               */
 +                              cpuctx->cgrp = NULL;
 +                      }
 +
 +                      if (mode & PERF_CGROUP_SWIN) {
 +                              WARN_ON_ONCE(cpuctx->cgrp);
 +                              /* set cgrp before ctxsw in to
 +                               * allow event_filter_match() to not
 +                               * have to pass task around
 +                               */
 +                              cpuctx->cgrp = perf_cgroup_from_task(task);
 +                              cpu_ctx_sched_in(cpuctx, EVENT_ALL, task);
 +                      }
 +              }
 +
 +              perf_pmu_enable(cpuctx->ctx.pmu);
 +      }
 +
 +      rcu_read_unlock();
 +
 +      local_irq_restore(flags);
 +}
 +
 +static inline void perf_cgroup_sched_out(struct task_struct *task)
 +{
 +      perf_cgroup_switch(task, PERF_CGROUP_SWOUT);
 +}
 +
 +static inline void perf_cgroup_sched_in(struct task_struct *task)
 +{
 +      perf_cgroup_switch(task, PERF_CGROUP_SWIN);
 +}
 +
 +static inline int perf_cgroup_connect(int fd, struct perf_event *event,
 +                                    struct perf_event_attr *attr,
 +                                    struct perf_event *group_leader)
 +{
 +      struct perf_cgroup *cgrp;
 +      struct cgroup_subsys_state *css;
 +      struct file *file;
 +      int ret = 0, fput_needed;
 +
 +      file = fget_light(fd, &fput_needed);
 +      if (!file)
 +              return -EBADF;
 +
 +      css = cgroup_css_from_dir(file, perf_subsys_id);
 +      if (IS_ERR(css)) {
 +              ret = PTR_ERR(css);
 +              goto out;
 +      }
 +
 +      cgrp = container_of(css, struct perf_cgroup, css);
 +      event->cgrp = cgrp;
 +
 +      /* must be done before we fput() the file */
 +      perf_get_cgroup(event);
 +
 +      /*
 +       * all events in a group must monitor
 +       * the same cgroup because a task belongs
 +       * to only one perf cgroup at a time
 +       */
 +      if (group_leader && group_leader->cgrp != cgrp) {
 +              perf_detach_cgroup(event);
 +              ret = -EINVAL;
 +      }
 +out:
 +      fput_light(file, fput_needed);
 +      return ret;
 +}
 +
 +static inline void
 +perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
 +{
 +      struct perf_cgroup_info *t;
 +      t = per_cpu_ptr(event->cgrp->info, event->cpu);
 +      event->shadow_ctx_time = now - t->timestamp;
 +}
 +
 +static inline void
 +perf_cgroup_defer_enabled(struct perf_event *event)
 +{
 +      /*
 +       * when the current task's perf cgroup does not match
 +       * the event's, we need to remember to call the
 +       * perf_mark_enable() function the first time a task with
 +       * a matching perf cgroup is scheduled in.
 +       */
 +      if (is_cgroup_event(event) && !perf_cgroup_match(event))
 +              event->cgrp_defer_enabled = 1;
 +}
 +
 +static inline void
 +perf_cgroup_mark_enabled(struct perf_event *event,
 +                       struct perf_event_context *ctx)
 +{
 +      struct perf_event *sub;
 +      u64 tstamp = perf_event_time(event);
 +
 +      if (!event->cgrp_defer_enabled)
 +              return;
 +
 +      event->cgrp_defer_enabled = 0;
 +
 +      event->tstamp_enabled = tstamp - event->total_time_enabled;
 +      list_for_each_entry(sub, &event->sibling_list, group_entry) {
 +              if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
 +                      sub->tstamp_enabled = tstamp - sub->total_time_enabled;
 +                      sub->cgrp_defer_enabled = 0;
 +              }
 +      }
 +}
 +#else /* !CONFIG_CGROUP_PERF */
 +
 +static inline bool
 +perf_cgroup_match(struct perf_event *event)
 +{
 +      return true;
 +}
 +
 +static inline void perf_detach_cgroup(struct perf_event *event)
 +{}
 +
 +static inline int is_cgroup_event(struct perf_event *event)
 +{
 +      return 0;
 +}
 +
 +static inline u64 perf_cgroup_event_cgrp_time(struct perf_event *event)
 +{
 +      return 0;
 +}
 +
 +static inline void update_cgrp_time_from_event(struct perf_event *event)
 +{
 +}
 +
 +static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
 +{
 +}
 +
 +static inline void perf_cgroup_sched_out(struct task_struct *task)
 +{
 +}
 +
 +static inline void perf_cgroup_sched_in(struct task_struct *task)
 +{
 +}
 +
 +static inline int perf_cgroup_connect(pid_t pid, struct perf_event *event,
 +                                    struct perf_event_attr *attr,
 +                                    struct perf_event *group_leader)
 +{
 +      return -EINVAL;
 +}
 +
 +static inline void
 +perf_cgroup_set_timestamp(struct task_struct *task,
 +                        struct perf_event_context *ctx)
 +{
 +}
 +
 +void
 +perf_cgroup_switch(struct task_struct *task, struct task_struct *next)
 +{
 +}
 +
 +static inline void
 +perf_cgroup_set_shadow_time(struct perf_event *event, u64 now)
 +{
 +}
 +
 +static inline u64 perf_cgroup_event_time(struct perf_event *event)
 +{
 +      return 0;
 +}
 +
 +static inline void
 +perf_cgroup_defer_enabled(struct perf_event *event)
 +{
 +}
 +
 +static inline void
 +perf_cgroup_mark_enabled(struct perf_event *event,
 +                       struct perf_event_context *ctx)
 +{
 +}
 +#endif
 +
 +void perf_pmu_disable(struct pmu *pmu)
 +{
 +      int *count = this_cpu_ptr(pmu->pmu_disable_count);
 +      if (!(*count)++)
 +              pmu->pmu_disable(pmu);
 +}
 +
 +void perf_pmu_enable(struct pmu *pmu)
 +{
 +      int *count = this_cpu_ptr(pmu->pmu_disable_count);
 +      if (!--(*count))
 +              pmu->pmu_enable(pmu);
 +}
 +
 +static DEFINE_PER_CPU(struct list_head, rotation_list);
 +
 +/*
 + * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
 + * because they're strictly cpu affine and rotate_start is called with IRQs
 + * disabled, while rotate_context is called from IRQ context.
 + */
 +static void perf_pmu_rotate_start(struct pmu *pmu)
 +{
 +      struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
 +      struct list_head *head = &__get_cpu_var(rotation_list);
 +
 +      WARN_ON(!irqs_disabled());
 +
 +      if (list_empty(&cpuctx->rotation_list))
 +              list_add(&cpuctx->rotation_list, head);
 +}
 +
 +static void get_ctx(struct perf_event_context *ctx)
 +{
 +      WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
 +}
 +
-               call_rcu(&ctx->rcu_head, free_ctx);
 +static void put_ctx(struct perf_event_context *ctx)
 +{
 +      if (atomic_dec_and_test(&ctx->refcount)) {
 +              if (ctx->parent_ctx)
 +                      put_ctx(ctx->parent_ctx);
 +              if (ctx->task)
 +                      put_task_struct(ctx->task);
- static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
- {
-       struct swevent_hlist *hlist;
-       hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
-       kfree(hlist);
- }
++              kfree_rcu(ctx, rcu_head);
 +      }
 +}
 +
 +static void unclone_ctx(struct perf_event_context *ctx)
 +{
 +      if (ctx->parent_ctx) {
 +              put_ctx(ctx->parent_ctx);
 +              ctx->parent_ctx = NULL;
 +      }
 +}
 +
 +static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
 +{
 +      /*
 +       * only top level events have the pid namespace they were created in
 +       */
 +      if (event->parent)
 +              event = event->parent;
 +
 +      return task_tgid_nr_ns(p, event->ns);
 +}
 +
 +static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
 +{
 +      /*
 +       * only top level events have the pid namespace they were created in
 +       */
 +      if (event->parent)
 +              event = event->parent;
 +
 +      return task_pid_nr_ns(p, event->ns);
 +}
 +
 +/*
 + * If we inherit events we want to return the parent event id
 + * to userspace.
 + */
 +static u64 primary_event_id(struct perf_event *event)
 +{
 +      u64 id = event->id;
 +
 +      if (event->parent)
 +              id = event->parent->id;
 +
 +      return id;
 +}
 +
 +/*
 + * Get the perf_event_context for a task and lock it.
 + * This has to cope with with the fact that until it is locked,
 + * the context could get moved to another task.
 + */
 +static struct perf_event_context *
 +perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
 +{
 +      struct perf_event_context *ctx;
 +
 +      rcu_read_lock();
 +retry:
 +      ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
 +      if (ctx) {
 +              /*
 +               * If this context is a clone of another, it might
 +               * get swapped for another underneath us by
 +               * perf_event_task_sched_out, though the
 +               * rcu_read_lock() protects us from any context
 +               * getting freed.  Lock the context and check if it
 +               * got swapped before we could get the lock, and retry
 +               * if so.  If we locked the right context, then it
 +               * can't get swapped on us any more.
 +               */
 +              raw_spin_lock_irqsave(&ctx->lock, *flags);
 +              if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
 +                      raw_spin_unlock_irqrestore(&ctx->lock, *flags);
 +                      goto retry;
 +              }
 +
 +              if (!atomic_inc_not_zero(&ctx->refcount)) {
 +                      raw_spin_unlock_irqrestore(&ctx->lock, *flags);
 +                      ctx = NULL;
 +              }
 +      }
 +      rcu_read_unlock();
 +      return ctx;
 +}
 +
 +/*
 + * Get the context for a task and increment its pin_count so it
 + * can't get swapped to another task.  This also increments its
 + * reference count so that the context can't get freed.
 + */
 +static struct perf_event_context *
 +perf_pin_task_context(struct task_struct *task, int ctxn)
 +{
 +      struct perf_event_context *ctx;
 +      unsigned long flags;
 +
 +      ctx = perf_lock_task_context(task, ctxn, &flags);
 +      if (ctx) {
 +              ++ctx->pin_count;
 +              raw_spin_unlock_irqrestore(&ctx->lock, flags);
 +      }
 +      return ctx;
 +}
 +
 +static void perf_unpin_context(struct perf_event_context *ctx)
 +{
 +      unsigned long flags;
 +
 +      raw_spin_lock_irqsave(&ctx->lock, flags);
 +      --ctx->pin_count;
 +      raw_spin_unlock_irqrestore(&ctx->lock, flags);
 +}
 +
 +/*
 + * Update the record of the current time in a context.
 + */
 +static void update_context_time(struct perf_event_context *ctx)
 +{
 +      u64 now = perf_clock();
 +
 +      ctx->time += now - ctx->timestamp;
 +      ctx->timestamp = now;
 +}
 +
 +static u64 perf_event_time(struct perf_event *event)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +
 +      if (is_cgroup_event(event))
 +              return perf_cgroup_event_time(event);
 +
 +      return ctx ? ctx->time : 0;
 +}
 +
 +/*
 + * Update the total_time_enabled and total_time_running fields for a event.
 + */
 +static void update_event_times(struct perf_event *event)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +      u64 run_end;
 +
 +      if (event->state < PERF_EVENT_STATE_INACTIVE ||
 +          event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
 +              return;
 +      /*
 +       * in cgroup mode, time_enabled represents
 +       * the time the event was enabled AND active
 +       * tasks were in the monitored cgroup. This is
 +       * independent of the activity of the context as
 +       * there may be a mix of cgroup and non-cgroup events.
 +       *
 +       * That is why we treat cgroup events differently
 +       * here.
 +       */
 +      if (is_cgroup_event(event))
 +              run_end = perf_event_time(event);
 +      else if (ctx->is_active)
 +              run_end = ctx->time;
 +      else
 +              run_end = event->tstamp_stopped;
 +
 +      event->total_time_enabled = run_end - event->tstamp_enabled;
 +
 +      if (event->state == PERF_EVENT_STATE_INACTIVE)
 +              run_end = event->tstamp_stopped;
 +      else
 +              run_end = perf_event_time(event);
 +
 +      event->total_time_running = run_end - event->tstamp_running;
 +
 +}
 +
 +/*
 + * Update total_time_enabled and total_time_running for all events in a group.
 + */
 +static void update_group_times(struct perf_event *leader)
 +{
 +      struct perf_event *event;
 +
 +      update_event_times(leader);
 +      list_for_each_entry(event, &leader->sibling_list, group_entry)
 +              update_event_times(event);
 +}
 +
 +static struct list_head *
 +ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
 +{
 +      if (event->attr.pinned)
 +              return &ctx->pinned_groups;
 +      else
 +              return &ctx->flexible_groups;
 +}
 +
 +/*
 + * Add a event from the lists for its context.
 + * Must be called with ctx->mutex and ctx->lock held.
 + */
 +static void
 +list_add_event(struct perf_event *event, struct perf_event_context *ctx)
 +{
 +      WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
 +      event->attach_state |= PERF_ATTACH_CONTEXT;
 +
 +      /*
 +       * If we're a stand alone event or group leader, we go to the context
 +       * list, group events are kept attached to the group so that
 +       * perf_group_detach can, at all times, locate all siblings.
 +       */
 +      if (event->group_leader == event) {
 +              struct list_head *list;
 +
 +              if (is_software_event(event))
 +                      event->group_flags |= PERF_GROUP_SOFTWARE;
 +
 +              list = ctx_group_list(event, ctx);
 +              list_add_tail(&event->group_entry, list);
 +      }
 +
 +      if (is_cgroup_event(event))
 +              ctx->nr_cgroups++;
 +
 +      list_add_rcu(&event->event_entry, &ctx->event_list);
 +      if (!ctx->nr_events)
 +              perf_pmu_rotate_start(ctx->pmu);
 +      ctx->nr_events++;
 +      if (event->attr.inherit_stat)
 +              ctx->nr_stat++;
 +}
 +
 +/*
 + * Called at perf_event creation and when events are attached/detached from a
 + * group.
 + */
 +static void perf_event__read_size(struct perf_event *event)
 +{
 +      int entry = sizeof(u64); /* value */
 +      int size = 0;
 +      int nr = 1;
 +
 +      if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 +              size += sizeof(u64);
 +
 +      if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 +              size += sizeof(u64);
 +
 +      if (event->attr.read_format & PERF_FORMAT_ID)
 +              entry += sizeof(u64);
 +
 +      if (event->attr.read_format & PERF_FORMAT_GROUP) {
 +              nr += event->group_leader->nr_siblings;
 +              size += sizeof(u64);
 +      }
 +
 +      size += entry * nr;
 +      event->read_size = size;
 +}
 +
 +static void perf_event__header_size(struct perf_event *event)
 +{
 +      struct perf_sample_data *data;
 +      u64 sample_type = event->attr.sample_type;
 +      u16 size = 0;
 +
 +      perf_event__read_size(event);
 +
 +      if (sample_type & PERF_SAMPLE_IP)
 +              size += sizeof(data->ip);
 +
 +      if (sample_type & PERF_SAMPLE_ADDR)
 +              size += sizeof(data->addr);
 +
 +      if (sample_type & PERF_SAMPLE_PERIOD)
 +              size += sizeof(data->period);
 +
 +      if (sample_type & PERF_SAMPLE_READ)
 +              size += event->read_size;
 +
 +      event->header_size = size;
 +}
 +
 +static void perf_event__id_header_size(struct perf_event *event)
 +{
 +      struct perf_sample_data *data;
 +      u64 sample_type = event->attr.sample_type;
 +      u16 size = 0;
 +
 +      if (sample_type & PERF_SAMPLE_TID)
 +              size += sizeof(data->tid_entry);
 +
 +      if (sample_type & PERF_SAMPLE_TIME)
 +              size += sizeof(data->time);
 +
 +      if (sample_type & PERF_SAMPLE_ID)
 +              size += sizeof(data->id);
 +
 +      if (sample_type & PERF_SAMPLE_STREAM_ID)
 +              size += sizeof(data->stream_id);
 +
 +      if (sample_type & PERF_SAMPLE_CPU)
 +              size += sizeof(data->cpu_entry);
 +
 +      event->id_header_size = size;
 +}
 +
 +static void perf_group_attach(struct perf_event *event)
 +{
 +      struct perf_event *group_leader = event->group_leader, *pos;
 +
 +      /*
 +       * We can have double attach due to group movement in perf_event_open.
 +       */
 +      if (event->attach_state & PERF_ATTACH_GROUP)
 +              return;
 +
 +      event->attach_state |= PERF_ATTACH_GROUP;
 +
 +      if (group_leader == event)
 +              return;
 +
 +      if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
 +                      !is_software_event(event))
 +              group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
 +
 +      list_add_tail(&event->group_entry, &group_leader->sibling_list);
 +      group_leader->nr_siblings++;
 +
 +      perf_event__header_size(group_leader);
 +
 +      list_for_each_entry(pos, &group_leader->sibling_list, group_entry)
 +              perf_event__header_size(pos);
 +}
 +
 +/*
 + * Remove a event from the lists for its context.
 + * Must be called with ctx->mutex and ctx->lock held.
 + */
 +static void
 +list_del_event(struct perf_event *event, struct perf_event_context *ctx)
 +{
 +      struct perf_cpu_context *cpuctx;
 +      /*
 +       * We can have double detach due to exit/hot-unplug + close.
 +       */
 +      if (!(event->attach_state & PERF_ATTACH_CONTEXT))
 +              return;
 +
 +      event->attach_state &= ~PERF_ATTACH_CONTEXT;
 +
 +      if (is_cgroup_event(event)) {
 +              ctx->nr_cgroups--;
 +              cpuctx = __get_cpu_context(ctx);
 +              /*
 +               * if there are no more cgroup events
 +               * then cler cgrp to avoid stale pointer
 +               * in update_cgrp_time_from_cpuctx()
 +               */
 +              if (!ctx->nr_cgroups)
 +                      cpuctx->cgrp = NULL;
 +      }
 +
 +      ctx->nr_events--;
 +      if (event->attr.inherit_stat)
 +              ctx->nr_stat--;
 +
 +      list_del_rcu(&event->event_entry);
 +
 +      if (event->group_leader == event)
 +              list_del_init(&event->group_entry);
 +
 +      update_group_times(event);
 +
 +      /*
 +       * If event was in error state, then keep it
 +       * that way, otherwise bogus counts will be
 +       * returned on read(). The only way to get out
 +       * of error state is by explicit re-enabling
 +       * of the event
 +       */
 +      if (event->state > PERF_EVENT_STATE_OFF)
 +              event->state = PERF_EVENT_STATE_OFF;
 +}
 +
 +static void perf_group_detach(struct perf_event *event)
 +{
 +      struct perf_event *sibling, *tmp;
 +      struct list_head *list = NULL;
 +
 +      /*
 +       * We can have double detach due to exit/hot-unplug + close.
 +       */
 +      if (!(event->attach_state & PERF_ATTACH_GROUP))
 +              return;
 +
 +      event->attach_state &= ~PERF_ATTACH_GROUP;
 +
 +      /*
 +       * If this is a sibling, remove it from its group.
 +       */
 +      if (event->group_leader != event) {
 +              list_del_init(&event->group_entry);
 +              event->group_leader->nr_siblings--;
 +              goto out;
 +      }
 +
 +      if (!list_empty(&event->group_entry))
 +              list = &event->group_entry;
 +
 +      /*
 +       * If this was a group event with sibling events then
 +       * upgrade the siblings to singleton events by adding them
 +       * to whatever list we are on.
 +       */
 +      list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
 +              if (list)
 +                      list_move_tail(&sibling->group_entry, list);
 +              sibling->group_leader = sibling;
 +
 +              /* Inherit group flags from the previous leader */
 +              sibling->group_flags = event->group_flags;
 +      }
 +
 +out:
 +      perf_event__header_size(event->group_leader);
 +
 +      list_for_each_entry(tmp, &event->group_leader->sibling_list, group_entry)
 +              perf_event__header_size(tmp);
 +}
 +
 +static inline int
 +event_filter_match(struct perf_event *event)
 +{
 +      return (event->cpu == -1 || event->cpu == smp_processor_id())
 +          && perf_cgroup_match(event);
 +}
 +
 +static void
 +event_sched_out(struct perf_event *event,
 +                struct perf_cpu_context *cpuctx,
 +                struct perf_event_context *ctx)
 +{
 +      u64 tstamp = perf_event_time(event);
 +      u64 delta;
 +      /*
 +       * An event which could not be activated because of
 +       * filter mismatch still needs to have its timings
 +       * maintained, otherwise bogus information is return
 +       * via read() for time_enabled, time_running:
 +       */
 +      if (event->state == PERF_EVENT_STATE_INACTIVE
 +          && !event_filter_match(event)) {
 +              delta = tstamp - event->tstamp_stopped;
 +              event->tstamp_running += delta;
 +              event->tstamp_stopped = tstamp;
 +      }
 +
 +      if (event->state != PERF_EVENT_STATE_ACTIVE)
 +              return;
 +
 +      event->state = PERF_EVENT_STATE_INACTIVE;
 +      if (event->pending_disable) {
 +              event->pending_disable = 0;
 +              event->state = PERF_EVENT_STATE_OFF;
 +      }
 +      event->tstamp_stopped = tstamp;
 +      event->pmu->del(event, 0);
 +      event->oncpu = -1;
 +
 +      if (!is_software_event(event))
 +              cpuctx->active_oncpu--;
 +      ctx->nr_active--;
 +      if (event->attr.exclusive || !cpuctx->active_oncpu)
 +              cpuctx->exclusive = 0;
 +}
 +
 +static void
 +group_sched_out(struct perf_event *group_event,
 +              struct perf_cpu_context *cpuctx,
 +              struct perf_event_context *ctx)
 +{
 +      struct perf_event *event;
 +      int state = group_event->state;
 +
 +      event_sched_out(group_event, cpuctx, ctx);
 +
 +      /*
 +       * Schedule out siblings (if any):
 +       */
 +      list_for_each_entry(event, &group_event->sibling_list, group_entry)
 +              event_sched_out(event, cpuctx, ctx);
 +
 +      if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
 +              cpuctx->exclusive = 0;
 +}
 +
 +/*
 + * Cross CPU call to remove a performance event
 + *
 + * We disable the event on the hardware level first. After that we
 + * remove it from the context list.
 + */
 +static int __perf_remove_from_context(void *info)
 +{
 +      struct perf_event *event = info;
 +      struct perf_event_context *ctx = event->ctx;
 +      struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 +
 +      raw_spin_lock(&ctx->lock);
 +      event_sched_out(event, cpuctx, ctx);
 +      list_del_event(event, ctx);
 +      raw_spin_unlock(&ctx->lock);
 +
 +      return 0;
 +}
 +
 +
 +/*
 + * Remove the event from a task's (or a CPU's) list of events.
 + *
 + * CPU events are removed with a smp call. For task events we only
 + * call when the task is on a CPU.
 + *
 + * If event->ctx is a cloned context, callers must make sure that
 + * every task struct that event->ctx->task could possibly point to
 + * remains valid.  This is OK when called from perf_release since
 + * that only calls us on the top-level context, which can't be a clone.
 + * When called from perf_event_exit_task, it's OK because the
 + * context has been detached from its task.
 + */
 +static void perf_remove_from_context(struct perf_event *event)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +      struct task_struct *task = ctx->task;
 +
 +      lockdep_assert_held(&ctx->mutex);
 +
 +      if (!task) {
 +              /*
 +               * Per cpu events are removed via an smp call and
 +               * the removal is always successful.
 +               */
 +              cpu_function_call(event->cpu, __perf_remove_from_context, event);
 +              return;
 +      }
 +
 +retry:
 +      if (!task_function_call(task, __perf_remove_from_context, event))
 +              return;
 +
 +      raw_spin_lock_irq(&ctx->lock);
 +      /*
 +       * If we failed to find a running task, but find the context active now
 +       * that we've acquired the ctx->lock, retry.
 +       */
 +      if (ctx->is_active) {
 +              raw_spin_unlock_irq(&ctx->lock);
 +              goto retry;
 +      }
 +
 +      /*
 +       * Since the task isn't running, its safe to remove the event, us
 +       * holding the ctx->lock ensures the task won't get scheduled in.
 +       */
 +      list_del_event(event, ctx);
 +      raw_spin_unlock_irq(&ctx->lock);
 +}
 +
 +/*
 + * Cross CPU call to disable a performance event
 + */
 +static int __perf_event_disable(void *info)
 +{
 +      struct perf_event *event = info;
 +      struct perf_event_context *ctx = event->ctx;
 +      struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 +
 +      /*
 +       * If this is a per-task event, need to check whether this
 +       * event's task is the current task on this cpu.
 +       *
 +       * Can trigger due to concurrent perf_event_context_sched_out()
 +       * flipping contexts around.
 +       */
 +      if (ctx->task && cpuctx->task_ctx != ctx)
 +              return -EINVAL;
 +
 +      raw_spin_lock(&ctx->lock);
 +
 +      /*
 +       * If the event is on, turn it off.
 +       * If it is in error state, leave it in error state.
 +       */
 +      if (event->state >= PERF_EVENT_STATE_INACTIVE) {
 +              update_context_time(ctx);
 +              update_cgrp_time_from_event(event);
 +              update_group_times(event);
 +              if (event == event->group_leader)
 +                      group_sched_out(event, cpuctx, ctx);
 +              else
 +                      event_sched_out(event, cpuctx, ctx);
 +              event->state = PERF_EVENT_STATE_OFF;
 +      }
 +
 +      raw_spin_unlock(&ctx->lock);
 +
 +      return 0;
 +}
 +
 +/*
 + * Disable a event.
 + *
 + * If event->ctx is a cloned context, callers must make sure that
 + * every task struct that event->ctx->task could possibly point to
 + * remains valid.  This condition is satisifed when called through
 + * perf_event_for_each_child or perf_event_for_each because they
 + * hold the top-level event's child_mutex, so any descendant that
 + * goes to exit will block in sync_child_event.
 + * When called from perf_pending_event it's OK because event->ctx
 + * is the current context on this CPU and preemption is disabled,
 + * hence we can't get into perf_event_task_sched_out for this context.
 + */
 +void perf_event_disable(struct perf_event *event)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +      struct task_struct *task = ctx->task;
 +
 +      if (!task) {
 +              /*
 +               * Disable the event on the cpu that it's on
 +               */
 +              cpu_function_call(event->cpu, __perf_event_disable, event);
 +              return;
 +      }
 +
 +retry:
 +      if (!task_function_call(task, __perf_event_disable, event))
 +              return;
 +
 +      raw_spin_lock_irq(&ctx->lock);
 +      /*
 +       * If the event is still active, we need to retry the cross-call.
 +       */
 +      if (event->state == PERF_EVENT_STATE_ACTIVE) {
 +              raw_spin_unlock_irq(&ctx->lock);
 +              /*
 +               * Reload the task pointer, it might have been changed by
 +               * a concurrent perf_event_context_sched_out().
 +               */
 +              task = ctx->task;
 +              goto retry;
 +      }
 +
 +      /*
 +       * Since we have the lock this context can't be scheduled
 +       * in, so we can change the state safely.
 +       */
 +      if (event->state == PERF_EVENT_STATE_INACTIVE) {
 +              update_group_times(event);
 +              event->state = PERF_EVENT_STATE_OFF;
 +      }
 +      raw_spin_unlock_irq(&ctx->lock);
 +}
 +
 +static void perf_set_shadow_time(struct perf_event *event,
 +                               struct perf_event_context *ctx,
 +                               u64 tstamp)
 +{
 +      /*
 +       * use the correct time source for the time snapshot
 +       *
 +       * We could get by without this by leveraging the
 +       * fact that to get to this function, the caller
 +       * has most likely already called update_context_time()
 +       * and update_cgrp_time_xx() and thus both timestamp
 +       * are identical (or very close). Given that tstamp is,
 +       * already adjusted for cgroup, we could say that:
 +       *    tstamp - ctx->timestamp
 +       * is equivalent to
 +       *    tstamp - cgrp->timestamp.
 +       *
 +       * Then, in perf_output_read(), the calculation would
 +       * work with no changes because:
 +       * - event is guaranteed scheduled in
 +       * - no scheduled out in between
 +       * - thus the timestamp would be the same
 +       *
 +       * But this is a bit hairy.
 +       *
 +       * So instead, we have an explicit cgroup call to remain
 +       * within the time time source all along. We believe it
 +       * is cleaner and simpler to understand.
 +       */
 +      if (is_cgroup_event(event))
 +              perf_cgroup_set_shadow_time(event, tstamp);
 +      else
 +              event->shadow_ctx_time = tstamp - ctx->timestamp;
 +}
 +
 +#define MAX_INTERRUPTS (~0ULL)
 +
 +static void perf_log_throttle(struct perf_event *event, int enable);
 +
 +static int
 +event_sched_in(struct perf_event *event,
 +               struct perf_cpu_context *cpuctx,
 +               struct perf_event_context *ctx)
 +{
 +      u64 tstamp = perf_event_time(event);
 +
 +      if (event->state <= PERF_EVENT_STATE_OFF)
 +              return 0;
 +
 +      event->state = PERF_EVENT_STATE_ACTIVE;
 +      event->oncpu = smp_processor_id();
 +
 +      /*
 +       * Unthrottle events, since we scheduled we might have missed several
 +       * ticks already, also for a heavily scheduling task there is little
 +       * guarantee it'll get a tick in a timely manner.
 +       */
 +      if (unlikely(event->hw.interrupts == MAX_INTERRUPTS)) {
 +              perf_log_throttle(event, 1);
 +              event->hw.interrupts = 0;
 +      }
 +
 +      /*
 +       * The new state must be visible before we turn it on in the hardware:
 +       */
 +      smp_wmb();
 +
 +      if (event->pmu->add(event, PERF_EF_START)) {
 +              event->state = PERF_EVENT_STATE_INACTIVE;
 +              event->oncpu = -1;
 +              return -EAGAIN;
 +      }
 +
 +      event->tstamp_running += tstamp - event->tstamp_stopped;
 +
 +      perf_set_shadow_time(event, ctx, tstamp);
 +
 +      if (!is_software_event(event))
 +              cpuctx->active_oncpu++;
 +      ctx->nr_active++;
 +
 +      if (event->attr.exclusive)
 +              cpuctx->exclusive = 1;
 +
 +      return 0;
 +}
 +
 +static int
 +group_sched_in(struct perf_event *group_event,
 +             struct perf_cpu_context *cpuctx,
 +             struct perf_event_context *ctx)
 +{
 +      struct perf_event *event, *partial_group = NULL;
 +      struct pmu *pmu = group_event->pmu;
 +      u64 now = ctx->time;
 +      bool simulate = false;
 +
 +      if (group_event->state == PERF_EVENT_STATE_OFF)
 +              return 0;
 +
 +      pmu->start_txn(pmu);
 +
 +      if (event_sched_in(group_event, cpuctx, ctx)) {
 +              pmu->cancel_txn(pmu);
 +              return -EAGAIN;
 +      }
 +
 +      /*
 +       * Schedule in siblings as one group (if any):
 +       */
 +      list_for_each_entry(event, &group_event->sibling_list, group_entry) {
 +              if (event_sched_in(event, cpuctx, ctx)) {
 +                      partial_group = event;
 +                      goto group_error;
 +              }
 +      }
 +
 +      if (!pmu->commit_txn(pmu))
 +              return 0;
 +
 +group_error:
 +      /*
 +       * Groups can be scheduled in as one unit only, so undo any
 +       * partial group before returning:
 +       * The events up to the failed event are scheduled out normally,
 +       * tstamp_stopped will be updated.
 +       *
 +       * The failed events and the remaining siblings need to have
 +       * their timings updated as if they had gone thru event_sched_in()
 +       * and event_sched_out(). This is required to get consistent timings
 +       * across the group. This also takes care of the case where the group
 +       * could never be scheduled by ensuring tstamp_stopped is set to mark
 +       * the time the event was actually stopped, such that time delta
 +       * calculation in update_event_times() is correct.
 +       */
 +      list_for_each_entry(event, &group_event->sibling_list, group_entry) {
 +              if (event == partial_group)
 +                      simulate = true;
 +
 +              if (simulate) {
 +                      event->tstamp_running += now - event->tstamp_stopped;
 +                      event->tstamp_stopped = now;
 +              } else {
 +                      event_sched_out(event, cpuctx, ctx);
 +              }
 +      }
 +      event_sched_out(group_event, cpuctx, ctx);
 +
 +      pmu->cancel_txn(pmu);
 +
 +      return -EAGAIN;
 +}
 +
 +/*
 + * Work out whether we can put this event group on the CPU now.
 + */
 +static int group_can_go_on(struct perf_event *event,
 +                         struct perf_cpu_context *cpuctx,
 +                         int can_add_hw)
 +{
 +      /*
 +       * Groups consisting entirely of software events can always go on.
 +       */
 +      if (event->group_flags & PERF_GROUP_SOFTWARE)
 +              return 1;
 +      /*
 +       * If an exclusive group is already on, no other hardware
 +       * events can go on.
 +       */
 +      if (cpuctx->exclusive)
 +              return 0;
 +      /*
 +       * If this group is exclusive and there are already
 +       * events on the CPU, it can't go on.
 +       */
 +      if (event->attr.exclusive && cpuctx->active_oncpu)
 +              return 0;
 +      /*
 +       * Otherwise, try to add it if all previous groups were able
 +       * to go on.
 +       */
 +      return can_add_hw;
 +}
 +
 +static void add_event_to_ctx(struct perf_event *event,
 +                             struct perf_event_context *ctx)
 +{
 +      u64 tstamp = perf_event_time(event);
 +
 +      list_add_event(event, ctx);
 +      perf_group_attach(event);
 +      event->tstamp_enabled = tstamp;
 +      event->tstamp_running = tstamp;
 +      event->tstamp_stopped = tstamp;
 +}
 +
 +static void perf_event_context_sched_in(struct perf_event_context *ctx,
 +                                      struct task_struct *tsk);
 +
 +/*
 + * Cross CPU call to install and enable a performance event
 + *
 + * Must be called with ctx->mutex held
 + */
 +static int  __perf_install_in_context(void *info)
 +{
 +      struct perf_event *event = info;
 +      struct perf_event_context *ctx = event->ctx;
 +      struct perf_event *leader = event->group_leader;
 +      struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 +      int err;
 +
 +      /*
 +       * In case we're installing a new context to an already running task,
 +       * could also happen before perf_event_task_sched_in() on architectures
 +       * which do context switches with IRQs enabled.
 +       */
 +      if (ctx->task && !cpuctx->task_ctx)
 +              perf_event_context_sched_in(ctx, ctx->task);
 +
 +      raw_spin_lock(&ctx->lock);
 +      ctx->is_active = 1;
 +      update_context_time(ctx);
 +      /*
 +       * update cgrp time only if current cgrp
 +       * matches event->cgrp. Must be done before
 +       * calling add_event_to_ctx()
 +       */
 +      update_cgrp_time_from_event(event);
 +
 +      add_event_to_ctx(event, ctx);
 +
 +      if (!event_filter_match(event))
 +              goto unlock;
 +
 +      /*
 +       * Don't put the event on if it is disabled or if
 +       * it is in a group and the group isn't on.
 +       */
 +      if (event->state != PERF_EVENT_STATE_INACTIVE ||
 +          (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
 +              goto unlock;
 +
 +      /*
 +       * An exclusive event can't go on if there are already active
 +       * hardware events, and no hardware event can go on if there
 +       * is already an exclusive event on.
 +       */
 +      if (!group_can_go_on(event, cpuctx, 1))
 +              err = -EEXIST;
 +      else
 +              err = event_sched_in(event, cpuctx, ctx);
 +
 +      if (err) {
 +              /*
 +               * This event couldn't go on.  If it is in a group
 +               * then we have to pull the whole group off.
 +               * If the event group is pinned then put it in error state.
 +               */
 +              if (leader != event)
 +                      group_sched_out(leader, cpuctx, ctx);
 +              if (leader->attr.pinned) {
 +                      update_group_times(leader);
 +                      leader->state = PERF_EVENT_STATE_ERROR;
 +              }
 +      }
 +
 +unlock:
 +      raw_spin_unlock(&ctx->lock);
 +
 +      return 0;
 +}
 +
 +/*
 + * Attach a performance event to a context
 + *
 + * First we add the event to the list with the hardware enable bit
 + * in event->hw_config cleared.
 + *
 + * If the event is attached to a task which is on a CPU we use a smp
 + * call to enable it in the task context. The task might have been
 + * scheduled away, but we check this in the smp call again.
 + */
 +static void
 +perf_install_in_context(struct perf_event_context *ctx,
 +                      struct perf_event *event,
 +                      int cpu)
 +{
 +      struct task_struct *task = ctx->task;
 +
 +      lockdep_assert_held(&ctx->mutex);
 +
 +      event->ctx = ctx;
 +
 +      if (!task) {
 +              /*
 +               * Per cpu events are installed via an smp call and
 +               * the install is always successful.
 +               */
 +              cpu_function_call(cpu, __perf_install_in_context, event);
 +              return;
 +      }
 +
 +retry:
 +      if (!task_function_call(task, __perf_install_in_context, event))
 +              return;
 +
 +      raw_spin_lock_irq(&ctx->lock);
 +      /*
 +       * If we failed to find a running task, but find the context active now
 +       * that we've acquired the ctx->lock, retry.
 +       */
 +      if (ctx->is_active) {
 +              raw_spin_unlock_irq(&ctx->lock);
 +              goto retry;
 +      }
 +
 +      /*
 +       * Since the task isn't running, its safe to add the event, us holding
 +       * the ctx->lock ensures the task won't get scheduled in.
 +       */
 +      add_event_to_ctx(event, ctx);
 +      raw_spin_unlock_irq(&ctx->lock);
 +}
 +
 +/*
 + * Put a event into inactive state and update time fields.
 + * Enabling the leader of a group effectively enables all
 + * the group members that aren't explicitly disabled, so we
 + * have to update their ->tstamp_enabled also.
 + * Note: this works for group members as well as group leaders
 + * since the non-leader members' sibling_lists will be empty.
 + */
 +static void __perf_event_mark_enabled(struct perf_event *event,
 +                                      struct perf_event_context *ctx)
 +{
 +      struct perf_event *sub;
 +      u64 tstamp = perf_event_time(event);
 +
 +      event->state = PERF_EVENT_STATE_INACTIVE;
 +      event->tstamp_enabled = tstamp - event->total_time_enabled;
 +      list_for_each_entry(sub, &event->sibling_list, group_entry) {
 +              if (sub->state >= PERF_EVENT_STATE_INACTIVE)
 +                      sub->tstamp_enabled = tstamp - sub->total_time_enabled;
 +      }
 +}
 +
 +/*
 + * Cross CPU call to enable a performance event
 + */
 +static int __perf_event_enable(void *info)
 +{
 +      struct perf_event *event = info;
 +      struct perf_event_context *ctx = event->ctx;
 +      struct perf_event *leader = event->group_leader;
 +      struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 +      int err;
 +
 +      if (WARN_ON_ONCE(!ctx->is_active))
 +              return -EINVAL;
 +
 +      raw_spin_lock(&ctx->lock);
 +      update_context_time(ctx);
 +
 +      if (event->state >= PERF_EVENT_STATE_INACTIVE)
 +              goto unlock;
 +
 +      /*
 +       * set current task's cgroup time reference point
 +       */
 +      perf_cgroup_set_timestamp(current, ctx);
 +
 +      __perf_event_mark_enabled(event, ctx);
 +
 +      if (!event_filter_match(event)) {
 +              if (is_cgroup_event(event))
 +                      perf_cgroup_defer_enabled(event);
 +              goto unlock;
 +      }
 +
 +      /*
 +       * If the event is in a group and isn't the group leader,
 +       * then don't put it on unless the group is on.
 +       */
 +      if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
 +              goto unlock;
 +
 +      if (!group_can_go_on(event, cpuctx, 1)) {
 +              err = -EEXIST;
 +      } else {
 +              if (event == leader)
 +                      err = group_sched_in(event, cpuctx, ctx);
 +              else
 +                      err = event_sched_in(event, cpuctx, ctx);
 +      }
 +
 +      if (err) {
 +              /*
 +               * If this event can't go on and it's part of a
 +               * group, then the whole group has to come off.
 +               */
 +              if (leader != event)
 +                      group_sched_out(leader, cpuctx, ctx);
 +              if (leader->attr.pinned) {
 +                      update_group_times(leader);
 +                      leader->state = PERF_EVENT_STATE_ERROR;
 +              }
 +      }
 +
 +unlock:
 +      raw_spin_unlock(&ctx->lock);
 +
 +      return 0;
 +}
 +
 +/*
 + * Enable a event.
 + *
 + * If event->ctx is a cloned context, callers must make sure that
 + * every task struct that event->ctx->task could possibly point to
 + * remains valid.  This condition is satisfied when called through
 + * perf_event_for_each_child or perf_event_for_each as described
 + * for perf_event_disable.
 + */
 +void perf_event_enable(struct perf_event *event)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +      struct task_struct *task = ctx->task;
 +
 +      if (!task) {
 +              /*
 +               * Enable the event on the cpu that it's on
 +               */
 +              cpu_function_call(event->cpu, __perf_event_enable, event);
 +              return;
 +      }
 +
 +      raw_spin_lock_irq(&ctx->lock);
 +      if (event->state >= PERF_EVENT_STATE_INACTIVE)
 +              goto out;
 +
 +      /*
 +       * If the event is in error state, clear that first.
 +       * That way, if we see the event in error state below, we
 +       * know that it has gone back into error state, as distinct
 +       * from the task having been scheduled away before the
 +       * cross-call arrived.
 +       */
 +      if (event->state == PERF_EVENT_STATE_ERROR)
 +              event->state = PERF_EVENT_STATE_OFF;
 +
 +retry:
 +      if (!ctx->is_active) {
 +              __perf_event_mark_enabled(event, ctx);
 +              goto out;
 +      }
 +
 +      raw_spin_unlock_irq(&ctx->lock);
 +
 +      if (!task_function_call(task, __perf_event_enable, event))
 +              return;
 +
 +      raw_spin_lock_irq(&ctx->lock);
 +
 +      /*
 +       * If the context is active and the event is still off,
 +       * we need to retry the cross-call.
 +       */
 +      if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF) {
 +              /*
 +               * task could have been flipped by a concurrent
 +               * perf_event_context_sched_out()
 +               */
 +              task = ctx->task;
 +              goto retry;
 +      }
 +
 +out:
 +      raw_spin_unlock_irq(&ctx->lock);
 +}
 +
 +static int perf_event_refresh(struct perf_event *event, int refresh)
 +{
 +      /*
 +       * not supported on inherited events
 +       */
 +      if (event->attr.inherit || !is_sampling_event(event))
 +              return -EINVAL;
 +
 +      atomic_add(refresh, &event->event_limit);
 +      perf_event_enable(event);
 +
 +      return 0;
 +}
 +
 +static void ctx_sched_out(struct perf_event_context *ctx,
 +                        struct perf_cpu_context *cpuctx,
 +                        enum event_type_t event_type)
 +{
 +      struct perf_event *event;
 +
 +      raw_spin_lock(&ctx->lock);
 +      perf_pmu_disable(ctx->pmu);
 +      ctx->is_active = 0;
 +      if (likely(!ctx->nr_events))
 +              goto out;
 +      update_context_time(ctx);
 +      update_cgrp_time_from_cpuctx(cpuctx);
 +
 +      if (!ctx->nr_active)
 +              goto out;
 +
 +      if (event_type & EVENT_PINNED) {
 +              list_for_each_entry(event, &ctx->pinned_groups, group_entry)
 +                      group_sched_out(event, cpuctx, ctx);
 +      }
 +
 +      if (event_type & EVENT_FLEXIBLE) {
 +              list_for_each_entry(event, &ctx->flexible_groups, group_entry)
 +                      group_sched_out(event, cpuctx, ctx);
 +      }
 +out:
 +      perf_pmu_enable(ctx->pmu);
 +      raw_spin_unlock(&ctx->lock);
 +}
 +
 +/*
 + * Test whether two contexts are equivalent, i.e. whether they
 + * have both been cloned from the same version of the same context
 + * and they both have the same number of enabled events.
 + * If the number of enabled events is the same, then the set
 + * of enabled events should be the same, because these are both
 + * inherited contexts, therefore we can't access individual events
 + * in them directly with an fd; we can only enable/disable all
 + * events via prctl, or enable/disable all events in a family
 + * via ioctl, which will have the same effect on both contexts.
 + */
 +static int context_equiv(struct perf_event_context *ctx1,
 +                       struct perf_event_context *ctx2)
 +{
 +      return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
 +              && ctx1->parent_gen == ctx2->parent_gen
 +              && !ctx1->pin_count && !ctx2->pin_count;
 +}
 +
 +static void __perf_event_sync_stat(struct perf_event *event,
 +                                   struct perf_event *next_event)
 +{
 +      u64 value;
 +
 +      if (!event->attr.inherit_stat)
 +              return;
 +
 +      /*
 +       * Update the event value, we cannot use perf_event_read()
 +       * because we're in the middle of a context switch and have IRQs
 +       * disabled, which upsets smp_call_function_single(), however
 +       * we know the event must be on the current CPU, therefore we
 +       * don't need to use it.
 +       */
 +      switch (event->state) {
 +      case PERF_EVENT_STATE_ACTIVE:
 +              event->pmu->read(event);
 +              /* fall-through */
 +
 +      case PERF_EVENT_STATE_INACTIVE:
 +              update_event_times(event);
 +              break;
 +
 +      default:
 +              break;
 +      }
 +
 +      /*
 +       * In order to keep per-task stats reliable we need to flip the event
 +       * values when we flip the contexts.
 +       */
 +      value = local64_read(&next_event->count);
 +      value = local64_xchg(&event->count, value);
 +      local64_set(&next_event->count, value);
 +
 +      swap(event->total_time_enabled, next_event->total_time_enabled);
 +      swap(event->total_time_running, next_event->total_time_running);
 +
 +      /*
 +       * Since we swizzled the values, update the user visible data too.
 +       */
 +      perf_event_update_userpage(event);
 +      perf_event_update_userpage(next_event);
 +}
 +
 +#define list_next_entry(pos, member) \
 +      list_entry(pos->member.next, typeof(*pos), member)
 +
 +static void perf_event_sync_stat(struct perf_event_context *ctx,
 +                                 struct perf_event_context *next_ctx)
 +{
 +      struct perf_event *event, *next_event;
 +
 +      if (!ctx->nr_stat)
 +              return;
 +
 +      update_context_time(ctx);
 +
 +      event = list_first_entry(&ctx->event_list,
 +                                 struct perf_event, event_entry);
 +
 +      next_event = list_first_entry(&next_ctx->event_list,
 +                                      struct perf_event, event_entry);
 +
 +      while (&event->event_entry != &ctx->event_list &&
 +             &next_event->event_entry != &next_ctx->event_list) {
 +
 +              __perf_event_sync_stat(event, next_event);
 +
 +              event = list_next_entry(event, event_entry);
 +              next_event = list_next_entry(next_event, event_entry);
 +      }
 +}
 +
 +static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
 +                                       struct task_struct *next)
 +{
 +      struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
 +      struct perf_event_context *next_ctx;
 +      struct perf_event_context *parent;
 +      struct perf_cpu_context *cpuctx;
 +      int do_switch = 1;
 +
 +      if (likely(!ctx))
 +              return;
 +
 +      cpuctx = __get_cpu_context(ctx);
 +      if (!cpuctx->task_ctx)
 +              return;
 +
 +      rcu_read_lock();
 +      parent = rcu_dereference(ctx->parent_ctx);
 +      next_ctx = next->perf_event_ctxp[ctxn];
 +      if (parent && next_ctx &&
 +          rcu_dereference(next_ctx->parent_ctx) == parent) {
 +              /*
 +               * Looks like the two contexts are clones, so we might be
 +               * able to optimize the context switch.  We lock both
 +               * contexts and check that they are clones under the
 +               * lock (including re-checking that neither has been
 +               * uncloned in the meantime).  It doesn't matter which
 +               * order we take the locks because no other cpu could
 +               * be trying to lock both of these tasks.
 +               */
 +              raw_spin_lock(&ctx->lock);
 +              raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
 +              if (context_equiv(ctx, next_ctx)) {
 +                      /*
 +                       * XXX do we need a memory barrier of sorts
 +                       * wrt to rcu_dereference() of perf_event_ctxp
 +                       */
 +                      task->perf_event_ctxp[ctxn] = next_ctx;
 +                      next->perf_event_ctxp[ctxn] = ctx;
 +                      ctx->task = next;
 +                      next_ctx->task = task;
 +                      do_switch = 0;
 +
 +                      perf_event_sync_stat(ctx, next_ctx);
 +              }
 +              raw_spin_unlock(&next_ctx->lock);
 +              raw_spin_unlock(&ctx->lock);
 +      }
 +      rcu_read_unlock();
 +
 +      if (do_switch) {
 +              ctx_sched_out(ctx, cpuctx, EVENT_ALL);
 +              cpuctx->task_ctx = NULL;
 +      }
 +}
 +
 +#define for_each_task_context_nr(ctxn)                                        \
 +      for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
 +
 +/*
 + * Called from scheduler to remove the events of the current task,
 + * with interrupts disabled.
 + *
 + * We stop each event and update the event value in event->count.
 + *
 + * This does not protect us against NMI, but disable()
 + * sets the disabled bit in the control field of event _before_
 + * accessing the event control register. If a NMI hits, then it will
 + * not restart the event.
 + */
 +void __perf_event_task_sched_out(struct task_struct *task,
 +                               struct task_struct *next)
 +{
 +      int ctxn;
 +
 +      for_each_task_context_nr(ctxn)
 +              perf_event_context_sched_out(task, ctxn, next);
 +
 +      /*
 +       * if cgroup events exist on this CPU, then we need
 +       * to check if we have to switch out PMU state.
 +       * cgroup event are system-wide mode only
 +       */
 +      if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
 +              perf_cgroup_sched_out(task);
 +}
 +
 +static void task_ctx_sched_out(struct perf_event_context *ctx,
 +                             enum event_type_t event_type)
 +{
 +      struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 +
 +      if (!cpuctx->task_ctx)
 +              return;
 +
 +      if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
 +              return;
 +
 +      ctx_sched_out(ctx, cpuctx, event_type);
 +      cpuctx->task_ctx = NULL;
 +}
 +
 +/*
 + * Called with IRQs disabled
 + */
 +static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
 +                            enum event_type_t event_type)
 +{
 +      ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
 +}
 +
 +static void
 +ctx_pinned_sched_in(struct perf_event_context *ctx,
 +                  struct perf_cpu_context *cpuctx)
 +{
 +      struct perf_event *event;
 +
 +      list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
 +              if (event->state <= PERF_EVENT_STATE_OFF)
 +                      continue;
 +              if (!event_filter_match(event))
 +                      continue;
 +
 +              /* may need to reset tstamp_enabled */
 +              if (is_cgroup_event(event))
 +                      perf_cgroup_mark_enabled(event, ctx);
 +
 +              if (group_can_go_on(event, cpuctx, 1))
 +                      group_sched_in(event, cpuctx, ctx);
 +
 +              /*
 +               * If this pinned group hasn't been scheduled,
 +               * put it in error state.
 +               */
 +              if (event->state == PERF_EVENT_STATE_INACTIVE) {
 +                      update_group_times(event);
 +                      event->state = PERF_EVENT_STATE_ERROR;
 +              }
 +      }
 +}
 +
 +static void
 +ctx_flexible_sched_in(struct perf_event_context *ctx,
 +                    struct perf_cpu_context *cpuctx)
 +{
 +      struct perf_event *event;
 +      int can_add_hw = 1;
 +
 +      list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
 +              /* Ignore events in OFF or ERROR state */
 +              if (event->state <= PERF_EVENT_STATE_OFF)
 +                      continue;
 +              /*
 +               * Listen to the 'cpu' scheduling filter constraint
 +               * of events:
 +               */
 +              if (!event_filter_match(event))
 +                      continue;
 +
 +              /* may need to reset tstamp_enabled */
 +              if (is_cgroup_event(event))
 +                      perf_cgroup_mark_enabled(event, ctx);
 +
 +              if (group_can_go_on(event, cpuctx, can_add_hw)) {
 +                      if (group_sched_in(event, cpuctx, ctx))
 +                              can_add_hw = 0;
 +              }
 +      }
 +}
 +
 +static void
 +ctx_sched_in(struct perf_event_context *ctx,
 +           struct perf_cpu_context *cpuctx,
 +           enum event_type_t event_type,
 +           struct task_struct *task)
 +{
 +      u64 now;
 +
 +      raw_spin_lock(&ctx->lock);
 +      ctx->is_active = 1;
 +      if (likely(!ctx->nr_events))
 +              goto out;
 +
 +      now = perf_clock();
 +      ctx->timestamp = now;
 +      perf_cgroup_set_timestamp(task, ctx);
 +      /*
 +       * First go through the list and put on any pinned groups
 +       * in order to give them the best chance of going on.
 +       */
 +      if (event_type & EVENT_PINNED)
 +              ctx_pinned_sched_in(ctx, cpuctx);
 +
 +      /* Then walk through the lower prio flexible groups */
 +      if (event_type & EVENT_FLEXIBLE)
 +              ctx_flexible_sched_in(ctx, cpuctx);
 +
 +out:
 +      raw_spin_unlock(&ctx->lock);
 +}
 +
 +static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
 +                           enum event_type_t event_type,
 +                           struct task_struct *task)
 +{
 +      struct perf_event_context *ctx = &cpuctx->ctx;
 +
 +      ctx_sched_in(ctx, cpuctx, event_type, task);
 +}
 +
 +static void task_ctx_sched_in(struct perf_event_context *ctx,
 +                            enum event_type_t event_type)
 +{
 +      struct perf_cpu_context *cpuctx;
 +
 +      cpuctx = __get_cpu_context(ctx);
 +      if (cpuctx->task_ctx == ctx)
 +              return;
 +
 +      ctx_sched_in(ctx, cpuctx, event_type, NULL);
 +      cpuctx->task_ctx = ctx;
 +}
 +
 +static void perf_event_context_sched_in(struct perf_event_context *ctx,
 +                                      struct task_struct *task)
 +{
 +      struct perf_cpu_context *cpuctx;
 +
 +      cpuctx = __get_cpu_context(ctx);
 +      if (cpuctx->task_ctx == ctx)
 +              return;
 +
 +      perf_pmu_disable(ctx->pmu);
 +      /*
 +       * We want to keep the following priority order:
 +       * cpu pinned (that don't need to move), task pinned,
 +       * cpu flexible, task flexible.
 +       */
 +      cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
 +
 +      ctx_sched_in(ctx, cpuctx, EVENT_PINNED, task);
 +      cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, task);
 +      ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
 +
 +      cpuctx->task_ctx = ctx;
 +
 +      /*
 +       * Since these rotations are per-cpu, we need to ensure the
 +       * cpu-context we got scheduled on is actually rotating.
 +       */
 +      perf_pmu_rotate_start(ctx->pmu);
 +      perf_pmu_enable(ctx->pmu);
 +}
 +
 +/*
 + * Called from scheduler to add the events of the current task
 + * with interrupts disabled.
 + *
 + * We restore the event value and then enable it.
 + *
 + * This does not protect us against NMI, but enable()
 + * sets the enabled bit in the control field of event _before_
 + * accessing the event control register. If a NMI hits, then it will
 + * keep the event running.
 + */
 +void __perf_event_task_sched_in(struct task_struct *task)
 +{
 +      struct perf_event_context *ctx;
 +      int ctxn;
 +
 +      for_each_task_context_nr(ctxn) {
 +              ctx = task->perf_event_ctxp[ctxn];
 +              if (likely(!ctx))
 +                      continue;
 +
 +              perf_event_context_sched_in(ctx, task);
 +      }
 +      /*
 +       * if cgroup events exist on this CPU, then we need
 +       * to check if we have to switch in PMU state.
 +       * cgroup event are system-wide mode only
 +       */
 +      if (atomic_read(&__get_cpu_var(perf_cgroup_events)))
 +              perf_cgroup_sched_in(task);
 +}
 +
 +static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
 +{
 +      u64 frequency = event->attr.sample_freq;
 +      u64 sec = NSEC_PER_SEC;
 +      u64 divisor, dividend;
 +
 +      int count_fls, nsec_fls, frequency_fls, sec_fls;
 +
 +      count_fls = fls64(count);
 +      nsec_fls = fls64(nsec);
 +      frequency_fls = fls64(frequency);
 +      sec_fls = 30;
 +
 +      /*
 +       * We got @count in @nsec, with a target of sample_freq HZ
 +       * the target period becomes:
 +       *
 +       *             @count * 10^9
 +       * period = -------------------
 +       *          @nsec * sample_freq
 +       *
 +       */
 +
 +      /*
 +       * Reduce accuracy by one bit such that @a and @b converge
 +       * to a similar magnitude.
 +       */
 +#define REDUCE_FLS(a, b)              \
 +do {                                  \
 +      if (a##_fls > b##_fls) {        \
 +              a >>= 1;                \
 +              a##_fls--;              \
 +      } else {                        \
 +              b >>= 1;                \
 +              b##_fls--;              \
 +      }                               \
 +} while (0)
 +
 +      /*
 +       * Reduce accuracy until either term fits in a u64, then proceed with
 +       * the other, so that finally we can do a u64/u64 division.
 +       */
 +      while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
 +              REDUCE_FLS(nsec, frequency);
 +              REDUCE_FLS(sec, count);
 +      }
 +
 +      if (count_fls + sec_fls > 64) {
 +              divisor = nsec * frequency;
 +
 +              while (count_fls + sec_fls > 64) {
 +                      REDUCE_FLS(count, sec);
 +                      divisor >>= 1;
 +              }
 +
 +              dividend = count * sec;
 +      } else {
 +              dividend = count * sec;
 +
 +              while (nsec_fls + frequency_fls > 64) {
 +                      REDUCE_FLS(nsec, frequency);
 +                      dividend >>= 1;
 +              }
 +
 +              divisor = nsec * frequency;
 +      }
 +
 +      if (!divisor)
 +              return dividend;
 +
 +      return div64_u64(dividend, divisor);
 +}
 +
 +static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
 +{
 +      struct hw_perf_event *hwc = &event->hw;
 +      s64 period, sample_period;
 +      s64 delta;
 +
 +      period = perf_calculate_period(event, nsec, count);
 +
 +      delta = (s64)(period - hwc->sample_period);
 +      delta = (delta + 7) / 8; /* low pass filter */
 +
 +      sample_period = hwc->sample_period + delta;
 +
 +      if (!sample_period)
 +              sample_period = 1;
 +
 +      hwc->sample_period = sample_period;
 +
 +      if (local64_read(&hwc->period_left) > 8*sample_period) {
 +              event->pmu->stop(event, PERF_EF_UPDATE);
 +              local64_set(&hwc->period_left, 0);
 +              event->pmu->start(event, PERF_EF_RELOAD);
 +      }
 +}
 +
 +static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
 +{
 +      struct perf_event *event;
 +      struct hw_perf_event *hwc;
 +      u64 interrupts, now;
 +      s64 delta;
 +
 +      raw_spin_lock(&ctx->lock);
 +      list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
 +              if (event->state != PERF_EVENT_STATE_ACTIVE)
 +                      continue;
 +
 +              if (!event_filter_match(event))
 +                      continue;
 +
 +              hwc = &event->hw;
 +
 +              interrupts = hwc->interrupts;
 +              hwc->interrupts = 0;
 +
 +              /*
 +               * unthrottle events on the tick
 +               */
 +              if (interrupts == MAX_INTERRUPTS) {
 +                      perf_log_throttle(event, 1);
 +                      event->pmu->start(event, 0);
 +              }
 +
 +              if (!event->attr.freq || !event->attr.sample_freq)
 +                      continue;
 +
 +              event->pmu->read(event);
 +              now = local64_read(&event->count);
 +              delta = now - hwc->freq_count_stamp;
 +              hwc->freq_count_stamp = now;
 +
 +              if (delta > 0)
 +                      perf_adjust_period(event, period, delta);
 +      }
 +      raw_spin_unlock(&ctx->lock);
 +}
 +
 +/*
 + * Round-robin a context's events:
 + */
 +static void rotate_ctx(struct perf_event_context *ctx)
 +{
 +      raw_spin_lock(&ctx->lock);
 +
 +      /*
 +       * Rotate the first entry last of non-pinned groups. Rotation might be
 +       * disabled by the inheritance code.
 +       */
 +      if (!ctx->rotate_disable)
 +              list_rotate_left(&ctx->flexible_groups);
 +
 +      raw_spin_unlock(&ctx->lock);
 +}
 +
 +/*
 + * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
 + * because they're strictly cpu affine and rotate_start is called with IRQs
 + * disabled, while rotate_context is called from IRQ context.
 + */
 +static void perf_rotate_context(struct perf_cpu_context *cpuctx)
 +{
 +      u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
 +      struct perf_event_context *ctx = NULL;
 +      int rotate = 0, remove = 1;
 +
 +      if (cpuctx->ctx.nr_events) {
 +              remove = 0;
 +              if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
 +                      rotate = 1;
 +      }
 +
 +      ctx = cpuctx->task_ctx;
 +      if (ctx && ctx->nr_events) {
 +              remove = 0;
 +              if (ctx->nr_events != ctx->nr_active)
 +                      rotate = 1;
 +      }
 +
 +      perf_pmu_disable(cpuctx->ctx.pmu);
 +      perf_ctx_adjust_freq(&cpuctx->ctx, interval);
 +      if (ctx)
 +              perf_ctx_adjust_freq(ctx, interval);
 +
 +      if (!rotate)
 +              goto done;
 +
 +      cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
 +      if (ctx)
 +              task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
 +
 +      rotate_ctx(&cpuctx->ctx);
 +      if (ctx)
 +              rotate_ctx(ctx);
 +
 +      cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE, current);
 +      if (ctx)
 +              task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
 +
 +done:
 +      if (remove)
 +              list_del_init(&cpuctx->rotation_list);
 +
 +      perf_pmu_enable(cpuctx->ctx.pmu);
 +}
 +
 +void perf_event_task_tick(void)
 +{
 +      struct list_head *head = &__get_cpu_var(rotation_list);
 +      struct perf_cpu_context *cpuctx, *tmp;
 +
 +      WARN_ON(!irqs_disabled());
 +
 +      list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
 +              if (cpuctx->jiffies_interval == 1 ||
 +                              !(jiffies % cpuctx->jiffies_interval))
 +                      perf_rotate_context(cpuctx);
 +      }
 +}
 +
 +static int event_enable_on_exec(struct perf_event *event,
 +                              struct perf_event_context *ctx)
 +{
 +      if (!event->attr.enable_on_exec)
 +              return 0;
 +
 +      event->attr.enable_on_exec = 0;
 +      if (event->state >= PERF_EVENT_STATE_INACTIVE)
 +              return 0;
 +
 +      __perf_event_mark_enabled(event, ctx);
 +
 +      return 1;
 +}
 +
 +/*
 + * Enable all of a task's events that have been marked enable-on-exec.
 + * This expects task == current.
 + */
 +static void perf_event_enable_on_exec(struct perf_event_context *ctx)
 +{
 +      struct perf_event *event;
 +      unsigned long flags;
 +      int enabled = 0;
 +      int ret;
 +
 +      local_irq_save(flags);
 +      if (!ctx || !ctx->nr_events)
 +              goto out;
 +
 +      /*
 +       * We must ctxsw out cgroup events to avoid conflict
 +       * when invoking perf_task_event_sched_in() later on
 +       * in this function. Otherwise we end up trying to
 +       * ctxswin cgroup events which are already scheduled
 +       * in.
 +       */
 +      perf_cgroup_sched_out(current);
 +      task_ctx_sched_out(ctx, EVENT_ALL);
 +
 +      raw_spin_lock(&ctx->lock);
 +
 +      list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
 +              ret = event_enable_on_exec(event, ctx);
 +              if (ret)
 +                      enabled = 1;
 +      }
 +
 +      list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
 +              ret = event_enable_on_exec(event, ctx);
 +              if (ret)
 +                      enabled = 1;
 +      }
 +
 +      /*
 +       * Unclone this context if we enabled any event.
 +       */
 +      if (enabled)
 +              unclone_ctx(ctx);
 +
 +      raw_spin_unlock(&ctx->lock);
 +
 +      /*
 +       * Also calls ctxswin for cgroup events, if any:
 +       */
 +      perf_event_context_sched_in(ctx, ctx->task);
 +out:
 +      local_irq_restore(flags);
 +}
 +
 +/*
 + * Cross CPU call to read the hardware event
 + */
 +static void __perf_event_read(void *info)
 +{
 +      struct perf_event *event = info;
 +      struct perf_event_context *ctx = event->ctx;
 +      struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
 +
 +      /*
 +       * If this is a task context, we need to check whether it is
 +       * the current task context of this cpu.  If not it has been
 +       * scheduled out before the smp call arrived.  In that case
 +       * event->count would have been updated to a recent sample
 +       * when the event was scheduled out.
 +       */
 +      if (ctx->task && cpuctx->task_ctx != ctx)
 +              return;
 +
 +      raw_spin_lock(&ctx->lock);
 +      if (ctx->is_active) {
 +              update_context_time(ctx);
 +              update_cgrp_time_from_event(event);
 +      }
 +      update_event_times(event);
 +      if (event->state == PERF_EVENT_STATE_ACTIVE)
 +              event->pmu->read(event);
 +      raw_spin_unlock(&ctx->lock);
 +}
 +
 +static inline u64 perf_event_count(struct perf_event *event)
 +{
 +      return local64_read(&event->count) + atomic64_read(&event->child_count);
 +}
 +
 +static u64 perf_event_read(struct perf_event *event)
 +{
 +      /*
 +       * If event is enabled and currently active on a CPU, update the
 +       * value in the event structure:
 +       */
 +      if (event->state == PERF_EVENT_STATE_ACTIVE) {
 +              smp_call_function_single(event->oncpu,
 +                                       __perf_event_read, event, 1);
 +      } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
 +              struct perf_event_context *ctx = event->ctx;
 +              unsigned long flags;
 +
 +              raw_spin_lock_irqsave(&ctx->lock, flags);
 +              /*
 +               * may read while context is not active
 +               * (e.g., thread is blocked), in that case
 +               * we cannot update context time
 +               */
 +              if (ctx->is_active) {
 +                      update_context_time(ctx);
 +                      update_cgrp_time_from_event(event);
 +              }
 +              update_event_times(event);
 +              raw_spin_unlock_irqrestore(&ctx->lock, flags);
 +      }
 +
 +      return perf_event_count(event);
 +}
 +
 +/*
 + * Callchain support
 + */
 +
 +struct callchain_cpus_entries {
 +      struct rcu_head                 rcu_head;
 +      struct perf_callchain_entry     *cpu_entries[0];
 +};
 +
 +static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
 +static atomic_t nr_callchain_events;
 +static DEFINE_MUTEX(callchain_mutex);
 +struct callchain_cpus_entries *callchain_cpus_entries;
 +
 +
 +__weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
 +                                struct pt_regs *regs)
 +{
 +}
 +
 +__weak void perf_callchain_user(struct perf_callchain_entry *entry,
 +                              struct pt_regs *regs)
 +{
 +}
 +
 +static void release_callchain_buffers_rcu(struct rcu_head *head)
 +{
 +      struct callchain_cpus_entries *entries;
 +      int cpu;
 +
 +      entries = container_of(head, struct callchain_cpus_entries, rcu_head);
 +
 +      for_each_possible_cpu(cpu)
 +              kfree(entries->cpu_entries[cpu]);
 +
 +      kfree(entries);
 +}
 +
 +static void release_callchain_buffers(void)
 +{
 +      struct callchain_cpus_entries *entries;
 +
 +      entries = callchain_cpus_entries;
 +      rcu_assign_pointer(callchain_cpus_entries, NULL);
 +      call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
 +}
 +
 +static int alloc_callchain_buffers(void)
 +{
 +      int cpu;
 +      int size;
 +      struct callchain_cpus_entries *entries;
 +
 +      /*
 +       * We can't use the percpu allocation API for data that can be
 +       * accessed from NMI. Use a temporary manual per cpu allocation
 +       * until that gets sorted out.
 +       */
 +      size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
 +
 +      entries = kzalloc(size, GFP_KERNEL);
 +      if (!entries)
 +              return -ENOMEM;
 +
 +      size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
 +
 +      for_each_possible_cpu(cpu) {
 +              entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
 +                                                       cpu_to_node(cpu));
 +              if (!entries->cpu_entries[cpu])
 +                      goto fail;
 +      }
 +
 +      rcu_assign_pointer(callchain_cpus_entries, entries);
 +
 +      return 0;
 +
 +fail:
 +      for_each_possible_cpu(cpu)
 +              kfree(entries->cpu_entries[cpu]);
 +      kfree(entries);
 +
 +      return -ENOMEM;
 +}
 +
 +static int get_callchain_buffers(void)
 +{
 +      int err = 0;
 +      int count;
 +
 +      mutex_lock(&callchain_mutex);
 +
 +      count = atomic_inc_return(&nr_callchain_events);
 +      if (WARN_ON_ONCE(count < 1)) {
 +              err = -EINVAL;
 +              goto exit;
 +      }
 +
 +      if (count > 1) {
 +              /* If the allocation failed, give up */
 +              if (!callchain_cpus_entries)
 +                      err = -ENOMEM;
 +              goto exit;
 +      }
 +
 +      err = alloc_callchain_buffers();
 +      if (err)
 +              release_callchain_buffers();
 +exit:
 +      mutex_unlock(&callchain_mutex);
 +
 +      return err;
 +}
 +
 +static void put_callchain_buffers(void)
 +{
 +      if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
 +              release_callchain_buffers();
 +              mutex_unlock(&callchain_mutex);
 +      }
 +}
 +
 +static int get_recursion_context(int *recursion)
 +{
 +      int rctx;
 +
 +      if (in_nmi())
 +              rctx = 3;
 +      else if (in_irq())
 +              rctx = 2;
 +      else if (in_softirq())
 +              rctx = 1;
 +      else
 +              rctx = 0;
 +
 +      if (recursion[rctx])
 +              return -1;
 +
 +      recursion[rctx]++;
 +      barrier();
 +
 +      return rctx;
 +}
 +
 +static inline void put_recursion_context(int *recursion, int rctx)
 +{
 +      barrier();
 +      recursion[rctx]--;
 +}
 +
 +static struct perf_callchain_entry *get_callchain_entry(int *rctx)
 +{
 +      int cpu;
 +      struct callchain_cpus_entries *entries;
 +
 +      *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
 +      if (*rctx == -1)
 +              return NULL;
 +
 +      entries = rcu_dereference(callchain_cpus_entries);
 +      if (!entries)
 +              return NULL;
 +
 +      cpu = smp_processor_id();
 +
 +      return &entries->cpu_entries[cpu][*rctx];
 +}
 +
 +static void
 +put_callchain_entry(int rctx)
 +{
 +      put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
 +}
 +
 +static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
 +{
 +      int rctx;
 +      struct perf_callchain_entry *entry;
 +
 +
 +      entry = get_callchain_entry(&rctx);
 +      if (rctx == -1)
 +              return NULL;
 +
 +      if (!entry)
 +              goto exit_put;
 +
 +      entry->nr = 0;
 +
 +      if (!user_mode(regs)) {
 +              perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
 +              perf_callchain_kernel(entry, regs);
 +              if (current->mm)
 +                      regs = task_pt_regs(current);
 +              else
 +                      regs = NULL;
 +      }
 +
 +      if (regs) {
 +              perf_callchain_store(entry, PERF_CONTEXT_USER);
 +              perf_callchain_user(entry, regs);
 +      }
 +
 +exit_put:
 +      put_callchain_entry(rctx);
 +
 +      return entry;
 +}
 +
 +/*
 + * Initialize the perf_event context in a task_struct:
 + */
 +static void __perf_event_init_context(struct perf_event_context *ctx)
 +{
 +      raw_spin_lock_init(&ctx->lock);
 +      mutex_init(&ctx->mutex);
 +      INIT_LIST_HEAD(&ctx->pinned_groups);
 +      INIT_LIST_HEAD(&ctx->flexible_groups);
 +      INIT_LIST_HEAD(&ctx->event_list);
 +      atomic_set(&ctx->refcount, 1);
 +}
 +
 +static struct perf_event_context *
 +alloc_perf_context(struct pmu *pmu, struct task_struct *task)
 +{
 +      struct perf_event_context *ctx;
 +
 +      ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
 +      if (!ctx)
 +              return NULL;
 +
 +      __perf_event_init_context(ctx);
 +      if (task) {
 +              ctx->task = task;
 +              get_task_struct(task);
 +      }
 +      ctx->pmu = pmu;
 +
 +      return ctx;
 +}
 +
 +static struct task_struct *
 +find_lively_task_by_vpid(pid_t vpid)
 +{
 +      struct task_struct *task;
 +      int err;
 +
 +      rcu_read_lock();
 +      if (!vpid)
 +              task = current;
 +      else
 +              task = find_task_by_vpid(vpid);
 +      if (task)
 +              get_task_struct(task);
 +      rcu_read_unlock();
 +
 +      if (!task)
 +              return ERR_PTR(-ESRCH);
 +
 +      /* Reuse ptrace permission checks for now. */
 +      err = -EACCES;
 +      if (!ptrace_may_access(task, PTRACE_MODE_READ))
 +              goto errout;
 +
 +      return task;
 +errout:
 +      put_task_struct(task);
 +      return ERR_PTR(err);
 +
 +}
 +
 +/*
 + * Returns a matching context with refcount and pincount.
 + */
 +static struct perf_event_context *
 +find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
 +{
 +      struct perf_event_context *ctx;
 +      struct perf_cpu_context *cpuctx;
 +      unsigned long flags;
 +      int ctxn, err;
 +
 +      if (!task) {
 +              /* Must be root to operate on a CPU event: */
 +              if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
 +                      return ERR_PTR(-EACCES);
 +
 +              /*
 +               * We could be clever and allow to attach a event to an
 +               * offline CPU and activate it when the CPU comes up, but
 +               * that's for later.
 +               */
 +              if (!cpu_online(cpu))
 +                      return ERR_PTR(-ENODEV);
 +
 +              cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
 +              ctx = &cpuctx->ctx;
 +              get_ctx(ctx);
 +              ++ctx->pin_count;
 +
 +              return ctx;
 +      }
 +
 +      err = -EINVAL;
 +      ctxn = pmu->task_ctx_nr;
 +      if (ctxn < 0)
 +              goto errout;
 +
 +retry:
 +      ctx = perf_lock_task_context(task, ctxn, &flags);
 +      if (ctx) {
 +              unclone_ctx(ctx);
 +              ++ctx->pin_count;
 +              raw_spin_unlock_irqrestore(&ctx->lock, flags);
 +      }
 +
 +      if (!ctx) {
 +              ctx = alloc_perf_context(pmu, task);
 +              err = -ENOMEM;
 +              if (!ctx)
 +                      goto errout;
 +
 +              get_ctx(ctx);
 +
 +              err = 0;
 +              mutex_lock(&task->perf_event_mutex);
 +              /*
 +               * If it has already passed perf_event_exit_task().
 +               * we must see PF_EXITING, it takes this mutex too.
 +               */
 +              if (task->flags & PF_EXITING)
 +                      err = -ESRCH;
 +              else if (task->perf_event_ctxp[ctxn])
 +                      err = -EAGAIN;
 +              else {
 +                      ++ctx->pin_count;
 +                      rcu_assign_pointer(task->perf_event_ctxp[ctxn], ctx);
 +              }
 +              mutex_unlock(&task->perf_event_mutex);
 +
 +              if (unlikely(err)) {
 +                      put_task_struct(task);
 +                      kfree(ctx);
 +
 +                      if (err == -EAGAIN)
 +                              goto retry;
 +                      goto errout;
 +              }
 +      }
 +
 +      return ctx;
 +
 +errout:
 +      return ERR_PTR(err);
 +}
 +
 +static void perf_event_free_filter(struct perf_event *event);
 +
 +static void free_event_rcu(struct rcu_head *head)
 +{
 +      struct perf_event *event;
 +
 +      event = container_of(head, struct perf_event, rcu_head);
 +      if (event->ns)
 +              put_pid_ns(event->ns);
 +      perf_event_free_filter(event);
 +      kfree(event);
 +}
 +
 +static void perf_buffer_put(struct perf_buffer *buffer);
 +
 +static void free_event(struct perf_event *event)
 +{
 +      irq_work_sync(&event->pending);
 +
 +      if (!event->parent) {
 +              if (event->attach_state & PERF_ATTACH_TASK)
 +                      jump_label_dec(&perf_sched_events);
 +              if (event->attr.mmap || event->attr.mmap_data)
 +                      atomic_dec(&nr_mmap_events);
 +              if (event->attr.comm)
 +                      atomic_dec(&nr_comm_events);
 +              if (event->attr.task)
 +                      atomic_dec(&nr_task_events);
 +              if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
 +                      put_callchain_buffers();
 +              if (is_cgroup_event(event)) {
 +                      atomic_dec(&per_cpu(perf_cgroup_events, event->cpu));
 +                      jump_label_dec(&perf_sched_events);
 +              }
 +      }
 +
 +      if (event->buffer) {
 +              perf_buffer_put(event->buffer);
 +              event->buffer = NULL;
 +      }
 +
 +      if (is_cgroup_event(event))
 +              perf_detach_cgroup(event);
 +
 +      if (event->destroy)
 +              event->destroy(event);
 +
 +      if (event->ctx)
 +              put_ctx(event->ctx);
 +
 +      call_rcu(&event->rcu_head, free_event_rcu);
 +}
 +
 +int perf_event_release_kernel(struct perf_event *event)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +
 +      /*
 +       * Remove from the PMU, can't get re-enabled since we got
 +       * here because the last ref went.
 +       */
 +      perf_event_disable(event);
 +
 +      WARN_ON_ONCE(ctx->parent_ctx);
 +      /*
 +       * There are two ways this annotation is useful:
 +       *
 +       *  1) there is a lock recursion from perf_event_exit_task
 +       *     see the comment there.
 +       *
 +       *  2) there is a lock-inversion with mmap_sem through
 +       *     perf_event_read_group(), which takes faults while
 +       *     holding ctx->mutex, however this is called after
 +       *     the last filedesc died, so there is no possibility
 +       *     to trigger the AB-BA case.
 +       */
 +      mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
 +      raw_spin_lock_irq(&ctx->lock);
 +      perf_group_detach(event);
 +      list_del_event(event, ctx);
 +      raw_spin_unlock_irq(&ctx->lock);
 +      mutex_unlock(&ctx->mutex);
 +
 +      free_event(event);
 +
 +      return 0;
 +}
 +EXPORT_SYMBOL_GPL(perf_event_release_kernel);
 +
 +/*
 + * Called when the last reference to the file is gone.
 + */
 +static int perf_release(struct inode *inode, struct file *file)
 +{
 +      struct perf_event *event = file->private_data;
 +      struct task_struct *owner;
 +
 +      file->private_data = NULL;
 +
 +      rcu_read_lock();
 +      owner = ACCESS_ONCE(event->owner);
 +      /*
 +       * Matches the smp_wmb() in perf_event_exit_task(). If we observe
 +       * !owner it means the list deletion is complete and we can indeed
 +       * free this event, otherwise we need to serialize on
 +       * owner->perf_event_mutex.
 +       */
 +      smp_read_barrier_depends();
 +      if (owner) {
 +              /*
 +               * Since delayed_put_task_struct() also drops the last
 +               * task reference we can safely take a new reference
 +               * while holding the rcu_read_lock().
 +               */
 +              get_task_struct(owner);
 +      }
 +      rcu_read_unlock();
 +
 +      if (owner) {
 +              mutex_lock(&owner->perf_event_mutex);
 +              /*
 +               * We have to re-check the event->owner field, if it is cleared
 +               * we raced with perf_event_exit_task(), acquiring the mutex
 +               * ensured they're done, and we can proceed with freeing the
 +               * event.
 +               */
 +              if (event->owner)
 +                      list_del_init(&event->owner_entry);
 +              mutex_unlock(&owner->perf_event_mutex);
 +              put_task_struct(owner);
 +      }
 +
 +      return perf_event_release_kernel(event);
 +}
 +
 +u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
 +{
 +      struct perf_event *child;
 +      u64 total = 0;
 +
 +      *enabled = 0;
 +      *running = 0;
 +
 +      mutex_lock(&event->child_mutex);
 +      total += perf_event_read(event);
 +      *enabled += event->total_time_enabled +
 +                      atomic64_read(&event->child_total_time_enabled);
 +      *running += event->total_time_running +
 +                      atomic64_read(&event->child_total_time_running);
 +
 +      list_for_each_entry(child, &event->child_list, child_list) {
 +              total += perf_event_read(child);
 +              *enabled += child->total_time_enabled;
 +              *running += child->total_time_running;
 +      }
 +      mutex_unlock(&event->child_mutex);
 +
 +      return total;
 +}
 +EXPORT_SYMBOL_GPL(perf_event_read_value);
 +
 +static int perf_event_read_group(struct perf_event *event,
 +                                 u64 read_format, char __user *buf)
 +{
 +      struct perf_event *leader = event->group_leader, *sub;
 +      int n = 0, size = 0, ret = -EFAULT;
 +      struct perf_event_context *ctx = leader->ctx;
 +      u64 values[5];
 +      u64 count, enabled, running;
 +
 +      mutex_lock(&ctx->mutex);
 +      count = perf_event_read_value(leader, &enabled, &running);
 +
 +      values[n++] = 1 + leader->nr_siblings;
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 +              values[n++] = enabled;
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 +              values[n++] = running;
 +      values[n++] = count;
 +      if (read_format & PERF_FORMAT_ID)
 +              values[n++] = primary_event_id(leader);
 +
 +      size = n * sizeof(u64);
 +
 +      if (copy_to_user(buf, values, size))
 +              goto unlock;
 +
 +      ret = size;
 +
 +      list_for_each_entry(sub, &leader->sibling_list, group_entry) {
 +              n = 0;
 +
 +              values[n++] = perf_event_read_value(sub, &enabled, &running);
 +              if (read_format & PERF_FORMAT_ID)
 +                      values[n++] = primary_event_id(sub);
 +
 +              size = n * sizeof(u64);
 +
 +              if (copy_to_user(buf + ret, values, size)) {
 +                      ret = -EFAULT;
 +                      goto unlock;
 +              }
 +
 +              ret += size;
 +      }
 +unlock:
 +      mutex_unlock(&ctx->mutex);
 +
 +      return ret;
 +}
 +
 +static int perf_event_read_one(struct perf_event *event,
 +                               u64 read_format, char __user *buf)
 +{
 +      u64 enabled, running;
 +      u64 values[4];
 +      int n = 0;
 +
 +      values[n++] = perf_event_read_value(event, &enabled, &running);
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 +              values[n++] = enabled;
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 +              values[n++] = running;
 +      if (read_format & PERF_FORMAT_ID)
 +              values[n++] = primary_event_id(event);
 +
 +      if (copy_to_user(buf, values, n * sizeof(u64)))
 +              return -EFAULT;
 +
 +      return n * sizeof(u64);
 +}
 +
 +/*
 + * Read the performance event - simple non blocking version for now
 + */
 +static ssize_t
 +perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
 +{
 +      u64 read_format = event->attr.read_format;
 +      int ret;
 +
 +      /*
 +       * Return end-of-file for a read on a event that is in
 +       * error state (i.e. because it was pinned but it couldn't be
 +       * scheduled on to the CPU at some point).
 +       */
 +      if (event->state == PERF_EVENT_STATE_ERROR)
 +              return 0;
 +
 +      if (count < event->read_size)
 +              return -ENOSPC;
 +
 +      WARN_ON_ONCE(event->ctx->parent_ctx);
 +      if (read_format & PERF_FORMAT_GROUP)
 +              ret = perf_event_read_group(event, read_format, buf);
 +      else
 +              ret = perf_event_read_one(event, read_format, buf);
 +
 +      return ret;
 +}
 +
 +static ssize_t
 +perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 +{
 +      struct perf_event *event = file->private_data;
 +
 +      return perf_read_hw(event, buf, count);
 +}
 +
 +static unsigned int perf_poll(struct file *file, poll_table *wait)
 +{
 +      struct perf_event *event = file->private_data;
 +      struct perf_buffer *buffer;
 +      unsigned int events = POLL_HUP;
 +
 +      rcu_read_lock();
 +      buffer = rcu_dereference(event->buffer);
 +      if (buffer)
 +              events = atomic_xchg(&buffer->poll, 0);
 +      rcu_read_unlock();
 +
 +      poll_wait(file, &event->waitq, wait);
 +
 +      return events;
 +}
 +
 +static void perf_event_reset(struct perf_event *event)
 +{
 +      (void)perf_event_read(event);
 +      local64_set(&event->count, 0);
 +      perf_event_update_userpage(event);
 +}
 +
 +/*
 + * Holding the top-level event's child_mutex means that any
 + * descendant process that has inherited this event will block
 + * in sync_child_event if it goes to exit, thus satisfying the
 + * task existence requirements of perf_event_enable/disable.
 + */
 +static void perf_event_for_each_child(struct perf_event *event,
 +                                      void (*func)(struct perf_event *))
 +{
 +      struct perf_event *child;
 +
 +      WARN_ON_ONCE(event->ctx->parent_ctx);
 +      mutex_lock(&event->child_mutex);
 +      func(event);
 +      list_for_each_entry(child, &event->child_list, child_list)
 +              func(child);
 +      mutex_unlock(&event->child_mutex);
 +}
 +
 +static void perf_event_for_each(struct perf_event *event,
 +                                void (*func)(struct perf_event *))
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +      struct perf_event *sibling;
 +
 +      WARN_ON_ONCE(ctx->parent_ctx);
 +      mutex_lock(&ctx->mutex);
 +      event = event->group_leader;
 +
 +      perf_event_for_each_child(event, func);
 +      func(event);
 +      list_for_each_entry(sibling, &event->sibling_list, group_entry)
 +              perf_event_for_each_child(event, func);
 +      mutex_unlock(&ctx->mutex);
 +}
 +
 +static int perf_event_period(struct perf_event *event, u64 __user *arg)
 +{
 +      struct perf_event_context *ctx = event->ctx;
 +      int ret = 0;
 +      u64 value;
 +
 +      if (!is_sampling_event(event))
 +              return -EINVAL;
 +
 +      if (copy_from_user(&value, arg, sizeof(value)))
 +              return -EFAULT;
 +
 +      if (!value)
 +              return -EINVAL;
 +
 +      raw_spin_lock_irq(&ctx->lock);
 +      if (event->attr.freq) {
 +              if (value > sysctl_perf_event_sample_rate) {
 +                      ret = -EINVAL;
 +                      goto unlock;
 +              }
 +
 +              event->attr.sample_freq = value;
 +      } else {
 +              event->attr.sample_period = value;
 +              event->hw.sample_period = value;
 +      }
 +unlock:
 +      raw_spin_unlock_irq(&ctx->lock);
 +
 +      return ret;
 +}
 +
 +static const struct file_operations perf_fops;
 +
 +static struct perf_event *perf_fget_light(int fd, int *fput_needed)
 +{
 +      struct file *file;
 +
 +      file = fget_light(fd, fput_needed);
 +      if (!file)
 +              return ERR_PTR(-EBADF);
 +
 +      if (file->f_op != &perf_fops) {
 +              fput_light(file, *fput_needed);
 +              *fput_needed = 0;
 +              return ERR_PTR(-EBADF);
 +      }
 +
 +      return file->private_data;
 +}
 +
 +static int perf_event_set_output(struct perf_event *event,
 +                               struct perf_event *output_event);
 +static int perf_event_set_filter(struct perf_event *event, void __user *arg);
 +
 +static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 +{
 +      struct perf_event *event = file->private_data;
 +      void (*func)(struct perf_event *);
 +      u32 flags = arg;
 +
 +      switch (cmd) {
 +      case PERF_EVENT_IOC_ENABLE:
 +              func = perf_event_enable;
 +              break;
 +      case PERF_EVENT_IOC_DISABLE:
 +              func = perf_event_disable;
 +              break;
 +      case PERF_EVENT_IOC_RESET:
 +              func = perf_event_reset;
 +              break;
 +
 +      case PERF_EVENT_IOC_REFRESH:
 +              return perf_event_refresh(event, arg);
 +
 +      case PERF_EVENT_IOC_PERIOD:
 +              return perf_event_period(event, (u64 __user *)arg);
 +
 +      case PERF_EVENT_IOC_SET_OUTPUT:
 +      {
 +              struct perf_event *output_event = NULL;
 +              int fput_needed = 0;
 +              int ret;
 +
 +              if (arg != -1) {
 +                      output_event = perf_fget_light(arg, &fput_needed);
 +                      if (IS_ERR(output_event))
 +                              return PTR_ERR(output_event);
 +              }
 +
 +              ret = perf_event_set_output(event, output_event);
 +              if (output_event)
 +                      fput_light(output_event->filp, fput_needed);
 +
 +              return ret;
 +      }
 +
 +      case PERF_EVENT_IOC_SET_FILTER:
 +              return perf_event_set_filter(event, (void __user *)arg);
 +
 +      default:
 +              return -ENOTTY;
 +      }
 +
 +      if (flags & PERF_IOC_FLAG_GROUP)
 +              perf_event_for_each(event, func);
 +      else
 +              perf_event_for_each_child(event, func);
 +
 +      return 0;
 +}
 +
 +int perf_event_task_enable(void)
 +{
 +      struct perf_event *event;
 +
 +      mutex_lock(&current->perf_event_mutex);
 +      list_for_each_entry(event, &current->perf_event_list, owner_entry)
 +              perf_event_for_each_child(event, perf_event_enable);
 +      mutex_unlock(&current->perf_event_mutex);
 +
 +      return 0;
 +}
 +
 +int perf_event_task_disable(void)
 +{
 +      struct perf_event *event;
 +
 +      mutex_lock(&current->perf_event_mutex);
 +      list_for_each_entry(event, &current->perf_event_list, owner_entry)
 +              perf_event_for_each_child(event, perf_event_disable);
 +      mutex_unlock(&current->perf_event_mutex);
 +
 +      return 0;
 +}
 +
 +#ifndef PERF_EVENT_INDEX_OFFSET
 +# define PERF_EVENT_INDEX_OFFSET 0
 +#endif
 +
 +static int perf_event_index(struct perf_event *event)
 +{
 +      if (event->hw.state & PERF_HES_STOPPED)
 +              return 0;
 +
 +      if (event->state != PERF_EVENT_STATE_ACTIVE)
 +              return 0;
 +
 +      return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
 +}
 +
 +/*
 + * Callers need to ensure there can be no nesting of this function, otherwise
 + * the seqlock logic goes bad. We can not serialize this because the arch
 + * code calls this from NMI context.
 + */
 +void perf_event_update_userpage(struct perf_event *event)
 +{
 +      struct perf_event_mmap_page *userpg;
 +      struct perf_buffer *buffer;
 +
 +      rcu_read_lock();
 +      buffer = rcu_dereference(event->buffer);
 +      if (!buffer)
 +              goto unlock;
 +
 +      userpg = buffer->user_page;
 +
 +      /*
 +       * Disable preemption so as to not let the corresponding user-space
 +       * spin too long if we get preempted.
 +       */
 +      preempt_disable();
 +      ++userpg->lock;
 +      barrier();
 +      userpg->index = perf_event_index(event);
 +      userpg->offset = perf_event_count(event);
 +      if (event->state == PERF_EVENT_STATE_ACTIVE)
 +              userpg->offset -= local64_read(&event->hw.prev_count);
 +
 +      userpg->time_enabled = event->total_time_enabled +
 +                      atomic64_read(&event->child_total_time_enabled);
 +
 +      userpg->time_running = event->total_time_running +
 +                      atomic64_read(&event->child_total_time_running);
 +
 +      barrier();
 +      ++userpg->lock;
 +      preempt_enable();
 +unlock:
 +      rcu_read_unlock();
 +}
 +
 +static unsigned long perf_data_size(struct perf_buffer *buffer);
 +
 +static void
 +perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
 +{
 +      long max_size = perf_data_size(buffer);
 +
 +      if (watermark)
 +              buffer->watermark = min(max_size, watermark);
 +
 +      if (!buffer->watermark)
 +              buffer->watermark = max_size / 2;
 +
 +      if (flags & PERF_BUFFER_WRITABLE)
 +              buffer->writable = 1;
 +
 +      atomic_set(&buffer->refcount, 1);
 +}
 +
 +#ifndef CONFIG_PERF_USE_VMALLOC
 +
 +/*
 + * Back perf_mmap() with regular GFP_KERNEL-0 pages.
 + */
 +
 +static struct page *
 +perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
 +{
 +      if (pgoff > buffer->nr_pages)
 +              return NULL;
 +
 +      if (pgoff == 0)
 +              return virt_to_page(buffer->user_page);
 +
 +      return virt_to_page(buffer->data_pages[pgoff - 1]);
 +}
 +
 +static void *perf_mmap_alloc_page(int cpu)
 +{
 +      struct page *page;
 +      int node;
 +
 +      node = (cpu == -1) ? cpu : cpu_to_node(cpu);
 +      page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
 +      if (!page)
 +              return NULL;
 +
 +      return page_address(page);
 +}
 +
 +static struct perf_buffer *
 +perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
 +{
 +      struct perf_buffer *buffer;
 +      unsigned long size;
 +      int i;
 +
 +      size = sizeof(struct perf_buffer);
 +      size += nr_pages * sizeof(void *);
 +
 +      buffer = kzalloc(size, GFP_KERNEL);
 +      if (!buffer)
 +              goto fail;
 +
 +      buffer->user_page = perf_mmap_alloc_page(cpu);
 +      if (!buffer->user_page)
 +              goto fail_user_page;
 +
 +      for (i = 0; i < nr_pages; i++) {
 +              buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
 +              if (!buffer->data_pages[i])
 +                      goto fail_data_pages;
 +      }
 +
 +      buffer->nr_pages = nr_pages;
 +
 +      perf_buffer_init(buffer, watermark, flags);
 +
 +      return buffer;
 +
 +fail_data_pages:
 +      for (i--; i >= 0; i--)
 +              free_page((unsigned long)buffer->data_pages[i]);
 +
 +      free_page((unsigned long)buffer->user_page);
 +
 +fail_user_page:
 +      kfree(buffer);
 +
 +fail:
 +      return NULL;
 +}
 +
 +static void perf_mmap_free_page(unsigned long addr)
 +{
 +      struct page *page = virt_to_page((void *)addr);
 +
 +      page->mapping = NULL;
 +      __free_page(page);
 +}
 +
 +static void perf_buffer_free(struct perf_buffer *buffer)
 +{
 +      int i;
 +
 +      perf_mmap_free_page((unsigned long)buffer->user_page);
 +      for (i = 0; i < buffer->nr_pages; i++)
 +              perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
 +      kfree(buffer);
 +}
 +
 +static inline int page_order(struct perf_buffer *buffer)
 +{
 +      return 0;
 +}
 +
 +#else
 +
 +/*
 + * Back perf_mmap() with vmalloc memory.
 + *
 + * Required for architectures that have d-cache aliasing issues.
 + */
 +
 +static inline int page_order(struct perf_buffer *buffer)
 +{
 +      return buffer->page_order;
 +}
 +
 +static struct page *
 +perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
 +{
 +      if (pgoff > (1UL << page_order(buffer)))
 +              return NULL;
 +
 +      return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
 +}
 +
 +static void perf_mmap_unmark_page(void *addr)
 +{
 +      struct page *page = vmalloc_to_page(addr);
 +
 +      page->mapping = NULL;
 +}
 +
 +static void perf_buffer_free_work(struct work_struct *work)
 +{
 +      struct perf_buffer *buffer;
 +      void *base;
 +      int i, nr;
 +
 +      buffer = container_of(work, struct perf_buffer, work);
 +      nr = 1 << page_order(buffer);
 +
 +      base = buffer->user_page;
 +      for (i = 0; i < nr + 1; i++)
 +              perf_mmap_unmark_page(base + (i * PAGE_SIZE));
 +
 +      vfree(base);
 +      kfree(buffer);
 +}
 +
 +static void perf_buffer_free(struct perf_buffer *buffer)
 +{
 +      schedule_work(&buffer->work);
 +}
 +
 +static struct perf_buffer *
 +perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
 +{
 +      struct perf_buffer *buffer;
 +      unsigned long size;
 +      void *all_buf;
 +
 +      size = sizeof(struct perf_buffer);
 +      size += sizeof(void *);
 +
 +      buffer = kzalloc(size, GFP_KERNEL);
 +      if (!buffer)
 +              goto fail;
 +
 +      INIT_WORK(&buffer->work, perf_buffer_free_work);
 +
 +      all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
 +      if (!all_buf)
 +              goto fail_all_buf;
 +
 +      buffer->user_page = all_buf;
 +      buffer->data_pages[0] = all_buf + PAGE_SIZE;
 +      buffer->page_order = ilog2(nr_pages);
 +      buffer->nr_pages = 1;
 +
 +      perf_buffer_init(buffer, watermark, flags);
 +
 +      return buffer;
 +
 +fail_all_buf:
 +      kfree(buffer);
 +
 +fail:
 +      return NULL;
 +}
 +
 +#endif
 +
 +static unsigned long perf_data_size(struct perf_buffer *buffer)
 +{
 +      return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
 +}
 +
 +static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 +{
 +      struct perf_event *event = vma->vm_file->private_data;
 +      struct perf_buffer *buffer;
 +      int ret = VM_FAULT_SIGBUS;
 +
 +      if (vmf->flags & FAULT_FLAG_MKWRITE) {
 +              if (vmf->pgoff == 0)
 +                      ret = 0;
 +              return ret;
 +      }
 +
 +      rcu_read_lock();
 +      buffer = rcu_dereference(event->buffer);
 +      if (!buffer)
 +              goto unlock;
 +
 +      if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
 +              goto unlock;
 +
 +      vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
 +      if (!vmf->page)
 +              goto unlock;
 +
 +      get_page(vmf->page);
 +      vmf->page->mapping = vma->vm_file->f_mapping;
 +      vmf->page->index   = vmf->pgoff;
 +
 +      ret = 0;
 +unlock:
 +      rcu_read_unlock();
 +
 +      return ret;
 +}
 +
 +static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
 +{
 +      struct perf_buffer *buffer;
 +
 +      buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
 +      perf_buffer_free(buffer);
 +}
 +
 +static struct perf_buffer *perf_buffer_get(struct perf_event *event)
 +{
 +      struct perf_buffer *buffer;
 +
 +      rcu_read_lock();
 +      buffer = rcu_dereference(event->buffer);
 +      if (buffer) {
 +              if (!atomic_inc_not_zero(&buffer->refcount))
 +                      buffer = NULL;
 +      }
 +      rcu_read_unlock();
 +
 +      return buffer;
 +}
 +
 +static void perf_buffer_put(struct perf_buffer *buffer)
 +{
 +      if (!atomic_dec_and_test(&buffer->refcount))
 +              return;
 +
 +      call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
 +}
 +
 +static void perf_mmap_open(struct vm_area_struct *vma)
 +{
 +      struct perf_event *event = vma->vm_file->private_data;
 +
 +      atomic_inc(&event->mmap_count);
 +}
 +
 +static void perf_mmap_close(struct vm_area_struct *vma)
 +{
 +      struct perf_event *event = vma->vm_file->private_data;
 +
 +      if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
 +              unsigned long size = perf_data_size(event->buffer);
 +              struct user_struct *user = event->mmap_user;
 +              struct perf_buffer *buffer = event->buffer;
 +
 +              atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
 +              vma->vm_mm->locked_vm -= event->mmap_locked;
 +              rcu_assign_pointer(event->buffer, NULL);
 +              mutex_unlock(&event->mmap_mutex);
 +
 +              perf_buffer_put(buffer);
 +              free_uid(user);
 +      }
 +}
 +
 +static const struct vm_operations_struct perf_mmap_vmops = {
 +      .open           = perf_mmap_open,
 +      .close          = perf_mmap_close,
 +      .fault          = perf_mmap_fault,
 +      .page_mkwrite   = perf_mmap_fault,
 +};
 +
 +static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 +{
 +      struct perf_event *event = file->private_data;
 +      unsigned long user_locked, user_lock_limit;
 +      struct user_struct *user = current_user();
 +      unsigned long locked, lock_limit;
 +      struct perf_buffer *buffer;
 +      unsigned long vma_size;
 +      unsigned long nr_pages;
 +      long user_extra, extra;
 +      int ret = 0, flags = 0;
 +
 +      /*
 +       * Don't allow mmap() of inherited per-task counters. This would
 +       * create a performance issue due to all children writing to the
 +       * same buffer.
 +       */
 +      if (event->cpu == -1 && event->attr.inherit)
 +              return -EINVAL;
 +
 +      if (!(vma->vm_flags & VM_SHARED))
 +              return -EINVAL;
 +
 +      vma_size = vma->vm_end - vma->vm_start;
 +      nr_pages = (vma_size / PAGE_SIZE) - 1;
 +
 +      /*
 +       * If we have buffer pages ensure they're a power-of-two number, so we
 +       * can do bitmasks instead of modulo.
 +       */
 +      if (nr_pages != 0 && !is_power_of_2(nr_pages))
 +              return -EINVAL;
 +
 +      if (vma_size != PAGE_SIZE * (1 + nr_pages))
 +              return -EINVAL;
 +
 +      if (vma->vm_pgoff != 0)
 +              return -EINVAL;
 +
 +      WARN_ON_ONCE(event->ctx->parent_ctx);
 +      mutex_lock(&event->mmap_mutex);
 +      if (event->buffer) {
 +              if (event->buffer->nr_pages == nr_pages)
 +                      atomic_inc(&event->buffer->refcount);
 +              else
 +                      ret = -EINVAL;
 +              goto unlock;
 +      }
 +
 +      user_extra = nr_pages + 1;
 +      user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
 +
 +      /*
 +       * Increase the limit linearly with more CPUs:
 +       */
 +      user_lock_limit *= num_online_cpus();
 +
 +      user_locked = atomic_long_read(&user->locked_vm) + user_extra;
 +
 +      extra = 0;
 +      if (user_locked > user_lock_limit)
 +              extra = user_locked - user_lock_limit;
 +
 +      lock_limit = rlimit(RLIMIT_MEMLOCK);
 +      lock_limit >>= PAGE_SHIFT;
 +      locked = vma->vm_mm->locked_vm + extra;
 +
 +      if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
 +              !capable(CAP_IPC_LOCK)) {
 +              ret = -EPERM;
 +              goto unlock;
 +      }
 +
 +      WARN_ON(event->buffer);
 +
 +      if (vma->vm_flags & VM_WRITE)
 +              flags |= PERF_BUFFER_WRITABLE;
 +
 +      buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
 +                                 event->cpu, flags);
 +      if (!buffer) {
 +              ret = -ENOMEM;
 +              goto unlock;
 +      }
 +      rcu_assign_pointer(event->buffer, buffer);
 +
 +      atomic_long_add(user_extra, &user->locked_vm);
 +      event->mmap_locked = extra;
 +      event->mmap_user = get_current_user();
 +      vma->vm_mm->locked_vm += event->mmap_locked;
 +
 +unlock:
 +      if (!ret)
 +              atomic_inc(&event->mmap_count);
 +      mutex_unlock(&event->mmap_mutex);
 +
 +      vma->vm_flags |= VM_RESERVED;
 +      vma->vm_ops = &perf_mmap_vmops;
 +
 +      return ret;
 +}
 +
 +static int perf_fasync(int fd, struct file *filp, int on)
 +{
 +      struct inode *inode = filp->f_path.dentry->d_inode;
 +      struct perf_event *event = filp->private_data;
 +      int retval;
 +
 +      mutex_lock(&inode->i_mutex);
 +      retval = fasync_helper(fd, filp, on, &event->fasync);
 +      mutex_unlock(&inode->i_mutex);
 +
 +      if (retval < 0)
 +              return retval;
 +
 +      return 0;
 +}
 +
 +static const struct file_operations perf_fops = {
 +      .llseek                 = no_llseek,
 +      .release                = perf_release,
 +      .read                   = perf_read,
 +      .poll                   = perf_poll,
 +      .unlocked_ioctl         = perf_ioctl,
 +      .compat_ioctl           = perf_ioctl,
 +      .mmap                   = perf_mmap,
 +      .fasync                 = perf_fasync,
 +};
 +
 +/*
 + * Perf event wakeup
 + *
 + * If there's data, ensure we set the poll() state and publish everything
 + * to user-space before waking everybody up.
 + */
 +
 +void perf_event_wakeup(struct perf_event *event)
 +{
 +      wake_up_all(&event->waitq);
 +
 +      if (event->pending_kill) {
 +              kill_fasync(&event->fasync, SIGIO, event->pending_kill);
 +              event->pending_kill = 0;
 +      }
 +}
 +
 +static void perf_pending_event(struct irq_work *entry)
 +{
 +      struct perf_event *event = container_of(entry,
 +                      struct perf_event, pending);
 +
 +      if (event->pending_disable) {
 +              event->pending_disable = 0;
 +              __perf_event_disable(event);
 +      }
 +
 +      if (event->pending_wakeup) {
 +              event->pending_wakeup = 0;
 +              perf_event_wakeup(event);
 +      }
 +}
 +
 +/*
 + * We assume there is only KVM supporting the callbacks.
 + * Later on, we might change it to a list if there is
 + * another virtualization implementation supporting the callbacks.
 + */
 +struct perf_guest_info_callbacks *perf_guest_cbs;
 +
 +int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
 +{
 +      perf_guest_cbs = cbs;
 +      return 0;
 +}
 +EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
 +
 +int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
 +{
 +      perf_guest_cbs = NULL;
 +      return 0;
 +}
 +EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
 +
 +/*
 + * Output
 + */
 +static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
 +                            unsigned long offset, unsigned long head)
 +{
 +      unsigned long mask;
 +
 +      if (!buffer->writable)
 +              return true;
 +
 +      mask = perf_data_size(buffer) - 1;
 +
 +      offset = (offset - tail) & mask;
 +      head   = (head   - tail) & mask;
 +
 +      if ((int)(head - offset) < 0)
 +              return false;
 +
 +      return true;
 +}
 +
 +static void perf_output_wakeup(struct perf_output_handle *handle)
 +{
 +      atomic_set(&handle->buffer->poll, POLL_IN);
 +
 +      if (handle->nmi) {
 +              handle->event->pending_wakeup = 1;
 +              irq_work_queue(&handle->event->pending);
 +      } else
 +              perf_event_wakeup(handle->event);
 +}
 +
 +/*
 + * We need to ensure a later event_id doesn't publish a head when a former
 + * event isn't done writing. However since we need to deal with NMIs we
 + * cannot fully serialize things.
 + *
 + * We only publish the head (and generate a wakeup) when the outer-most
 + * event completes.
 + */
 +static void perf_output_get_handle(struct perf_output_handle *handle)
 +{
 +      struct perf_buffer *buffer = handle->buffer;
 +
 +      preempt_disable();
 +      local_inc(&buffer->nest);
 +      handle->wakeup = local_read(&buffer->wakeup);
 +}
 +
 +static void perf_output_put_handle(struct perf_output_handle *handle)
 +{
 +      struct perf_buffer *buffer = handle->buffer;
 +      unsigned long head;
 +
 +again:
 +      head = local_read(&buffer->head);
 +
 +      /*
 +       * IRQ/NMI can happen here, which means we can miss a head update.
 +       */
 +
 +      if (!local_dec_and_test(&buffer->nest))
 +              goto out;
 +
 +      /*
 +       * Publish the known good head. Rely on the full barrier implied
 +       * by atomic_dec_and_test() order the buffer->head read and this
 +       * write.
 +       */
 +      buffer->user_page->data_head = head;
 +
 +      /*
 +       * Now check if we missed an update, rely on the (compiler)
 +       * barrier in atomic_dec_and_test() to re-read buffer->head.
 +       */
 +      if (unlikely(head != local_read(&buffer->head))) {
 +              local_inc(&buffer->nest);
 +              goto again;
 +      }
 +
 +      if (handle->wakeup != local_read(&buffer->wakeup))
 +              perf_output_wakeup(handle);
 +
 +out:
 +      preempt_enable();
 +}
 +
 +__always_inline void perf_output_copy(struct perf_output_handle *handle,
 +                    const void *buf, unsigned int len)
 +{
 +      do {
 +              unsigned long size = min_t(unsigned long, handle->size, len);
 +
 +              memcpy(handle->addr, buf, size);
 +
 +              len -= size;
 +              handle->addr += size;
 +              buf += size;
 +              handle->size -= size;
 +              if (!handle->size) {
 +                      struct perf_buffer *buffer = handle->buffer;
 +
 +                      handle->page++;
 +                      handle->page &= buffer->nr_pages - 1;
 +                      handle->addr = buffer->data_pages[handle->page];
 +                      handle->size = PAGE_SIZE << page_order(buffer);
 +              }
 +      } while (len);
 +}
 +
 +static void __perf_event_header__init_id(struct perf_event_header *header,
 +                                       struct perf_sample_data *data,
 +                                       struct perf_event *event)
 +{
 +      u64 sample_type = event->attr.sample_type;
 +
 +      data->type = sample_type;
 +      header->size += event->id_header_size;
 +
 +      if (sample_type & PERF_SAMPLE_TID) {
 +              /* namespace issues */
 +              data->tid_entry.pid = perf_event_pid(event, current);
 +              data->tid_entry.tid = perf_event_tid(event, current);
 +      }
 +
 +      if (sample_type & PERF_SAMPLE_TIME)
 +              data->time = perf_clock();
 +
 +      if (sample_type & PERF_SAMPLE_ID)
 +              data->id = primary_event_id(event);
 +
 +      if (sample_type & PERF_SAMPLE_STREAM_ID)
 +              data->stream_id = event->id;
 +
 +      if (sample_type & PERF_SAMPLE_CPU) {
 +              data->cpu_entry.cpu      = raw_smp_processor_id();
 +              data->cpu_entry.reserved = 0;
 +      }
 +}
 +
 +static void perf_event_header__init_id(struct perf_event_header *header,
 +                                     struct perf_sample_data *data,
 +                                     struct perf_event *event)
 +{
 +      if (event->attr.sample_id_all)
 +              __perf_event_header__init_id(header, data, event);
 +}
 +
 +static void __perf_event__output_id_sample(struct perf_output_handle *handle,
 +                                         struct perf_sample_data *data)
 +{
 +      u64 sample_type = data->type;
 +
 +      if (sample_type & PERF_SAMPLE_TID)
 +              perf_output_put(handle, data->tid_entry);
 +
 +      if (sample_type & PERF_SAMPLE_TIME)
 +              perf_output_put(handle, data->time);
 +
 +      if (sample_type & PERF_SAMPLE_ID)
 +              perf_output_put(handle, data->id);
 +
 +      if (sample_type & PERF_SAMPLE_STREAM_ID)
 +              perf_output_put(handle, data->stream_id);
 +
 +      if (sample_type & PERF_SAMPLE_CPU)
 +              perf_output_put(handle, data->cpu_entry);
 +}
 +
 +static void perf_event__output_id_sample(struct perf_event *event,
 +                                       struct perf_output_handle *handle,
 +                                       struct perf_sample_data *sample)
 +{
 +      if (event->attr.sample_id_all)
 +              __perf_event__output_id_sample(handle, sample);
 +}
 +
 +int perf_output_begin(struct perf_output_handle *handle,
 +                    struct perf_event *event, unsigned int size,
 +                    int nmi, int sample)
 +{
 +      struct perf_buffer *buffer;
 +      unsigned long tail, offset, head;
 +      int have_lost;
 +      struct perf_sample_data sample_data;
 +      struct {
 +              struct perf_event_header header;
 +              u64                      id;
 +              u64                      lost;
 +      } lost_event;
 +
 +      rcu_read_lock();
 +      /*
 +       * For inherited events we send all the output towards the parent.
 +       */
 +      if (event->parent)
 +              event = event->parent;
 +
 +      buffer = rcu_dereference(event->buffer);
 +      if (!buffer)
 +              goto out;
 +
 +      handle->buffer  = buffer;
 +      handle->event   = event;
 +      handle->nmi     = nmi;
 +      handle->sample  = sample;
 +
 +      if (!buffer->nr_pages)
 +              goto out;
 +
 +      have_lost = local_read(&buffer->lost);
 +      if (have_lost) {
 +              lost_event.header.size = sizeof(lost_event);
 +              perf_event_header__init_id(&lost_event.header, &sample_data,
 +                                         event);
 +              size += lost_event.header.size;
 +      }
 +
 +      perf_output_get_handle(handle);
 +
 +      do {
 +              /*
 +               * Userspace could choose to issue a mb() before updating the
 +               * tail pointer. So that all reads will be completed before the
 +               * write is issued.
 +               */
 +              tail = ACCESS_ONCE(buffer->user_page->data_tail);
 +              smp_rmb();
 +              offset = head = local_read(&buffer->head);
 +              head += size;
 +              if (unlikely(!perf_output_space(buffer, tail, offset, head)))
 +                      goto fail;
 +      } while (local_cmpxchg(&buffer->head, offset, head) != offset);
 +
 +      if (head - local_read(&buffer->wakeup) > buffer->watermark)
 +              local_add(buffer->watermark, &buffer->wakeup);
 +
 +      handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
 +      handle->page &= buffer->nr_pages - 1;
 +      handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
 +      handle->addr = buffer->data_pages[handle->page];
 +      handle->addr += handle->size;
 +      handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
 +
 +      if (have_lost) {
 +              lost_event.header.type = PERF_RECORD_LOST;
 +              lost_event.header.misc = 0;
 +              lost_event.id          = event->id;
 +              lost_event.lost        = local_xchg(&buffer->lost, 0);
 +
 +              perf_output_put(handle, lost_event);
 +              perf_event__output_id_sample(event, handle, &sample_data);
 +      }
 +
 +      return 0;
 +
 +fail:
 +      local_inc(&buffer->lost);
 +      perf_output_put_handle(handle);
 +out:
 +      rcu_read_unlock();
 +
 +      return -ENOSPC;
 +}
 +
 +void perf_output_end(struct perf_output_handle *handle)
 +{
 +      struct perf_event *event = handle->event;
 +      struct perf_buffer *buffer = handle->buffer;
 +
 +      int wakeup_events = event->attr.wakeup_events;
 +
 +      if (handle->sample && wakeup_events) {
 +              int events = local_inc_return(&buffer->events);
 +              if (events >= wakeup_events) {
 +                      local_sub(wakeup_events, &buffer->events);
 +                      local_inc(&buffer->wakeup);
 +              }
 +      }
 +
 +      perf_output_put_handle(handle);
 +      rcu_read_unlock();
 +}
 +
 +static void perf_output_read_one(struct perf_output_handle *handle,
 +                               struct perf_event *event,
 +                               u64 enabled, u64 running)
 +{
 +      u64 read_format = event->attr.read_format;
 +      u64 values[4];
 +      int n = 0;
 +
 +      values[n++] = perf_event_count(event);
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
 +              values[n++] = enabled +
 +                      atomic64_read(&event->child_total_time_enabled);
 +      }
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
 +              values[n++] = running +
 +                      atomic64_read(&event->child_total_time_running);
 +      }
 +      if (read_format & PERF_FORMAT_ID)
 +              values[n++] = primary_event_id(event);
 +
 +      perf_output_copy(handle, values, n * sizeof(u64));
 +}
 +
 +/*
 + * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
 + */
 +static void perf_output_read_group(struct perf_output_handle *handle,
 +                          struct perf_event *event,
 +                          u64 enabled, u64 running)
 +{
 +      struct perf_event *leader = event->group_leader, *sub;
 +      u64 read_format = event->attr.read_format;
 +      u64 values[5];
 +      int n = 0;
 +
 +      values[n++] = 1 + leader->nr_siblings;
 +
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 +              values[n++] = enabled;
 +
 +      if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 +              values[n++] = running;
 +
 +      if (leader != event)
 +              leader->pmu->read(leader);
 +
 +      values[n++] = perf_event_count(leader);
 +      if (read_format & PERF_FORMAT_ID)
 +              values[n++] = primary_event_id(leader);
 +
 +      perf_output_copy(handle, values, n * sizeof(u64));
 +
 +      list_for_each_entry(sub, &leader->sibling_list, group_entry) {
 +              n = 0;
 +
 +              if (sub != event)
 +                      sub->pmu->read(sub);
 +
 +              values[n++] = perf_event_count(sub);
 +              if (read_format & PERF_FORMAT_ID)
 +                      values[n++] = primary_event_id(sub);
 +
 +              perf_output_copy(handle, values, n * sizeof(u64));
 +      }
 +}
 +
 +#define PERF_FORMAT_TOTAL_TIMES (PERF_FORMAT_TOTAL_TIME_ENABLED|\
 +                               PERF_FORMAT_TOTAL_TIME_RUNNING)
 +
 +static void perf_output_read(struct perf_output_handle *handle,
 +                           struct perf_event *event)
 +{
 +      u64 enabled = 0, running = 0, now, ctx_time;
 +      u64 read_format = event->attr.read_format;
 +
 +      /*
 +       * compute total_time_enabled, total_time_running
 +       * based on snapshot values taken when the event
 +       * was last scheduled in.
 +       *
 +       * we cannot simply called update_context_time()
 +       * because of locking issue as we are called in
 +       * NMI context
 +       */
 +      if (read_format & PERF_FORMAT_TOTAL_TIMES) {
 +              now = perf_clock();
 +              ctx_time = event->shadow_ctx_time + now;
 +              enabled = ctx_time - event->tstamp_enabled;
 +              running = ctx_time - event->tstamp_running;
 +      }
 +
 +      if (event->attr.read_format & PERF_FORMAT_GROUP)
 +              perf_output_read_group(handle, event, enabled, running);
 +      else
 +              perf_output_read_one(handle, event, enabled, running);
 +}
 +
 +void perf_output_sample(struct perf_output_handle *handle,
 +                      struct perf_event_header *header,
 +                      struct perf_sample_data *data,
 +                      struct perf_event *event)
 +{
 +      u64 sample_type = data->type;
 +
 +      perf_output_put(handle, *header);
 +
 +      if (sample_type & PERF_SAMPLE_IP)
 +              perf_output_put(handle, data->ip);
 +
 +      if (sample_type & PERF_SAMPLE_TID)
 +              perf_output_put(handle, data->tid_entry);
 +
 +      if (sample_type & PERF_SAMPLE_TIME)
 +              perf_output_put(handle, data->time);
 +
 +      if (sample_type & PERF_SAMPLE_ADDR)
 +              perf_output_put(handle, data->addr);
 +
 +      if (sample_type & PERF_SAMPLE_ID)
 +              perf_output_put(handle, data->id);
 +
 +      if (sample_type & PERF_SAMPLE_STREAM_ID)
 +              perf_output_put(handle, data->stream_id);
 +
 +      if (sample_type & PERF_SAMPLE_CPU)
 +              perf_output_put(handle, data->cpu_entry);
 +
 +      if (sample_type & PERF_SAMPLE_PERIOD)
 +              perf_output_put(handle, data->period);
 +
 +      if (sample_type & PERF_SAMPLE_READ)
 +              perf_output_read(handle, event);
 +
 +      if (sample_type & PERF_SAMPLE_CALLCHAIN) {
 +              if (data->callchain) {
 +                      int size = 1;
 +
 +                      if (data->callchain)
 +                              size += data->callchain->nr;
 +
 +                      size *= sizeof(u64);
 +
 +                      perf_output_copy(handle, data->callchain, size);
 +              } else {
 +                      u64 nr = 0;
 +                      perf_output_put(handle, nr);
 +              }
 +      }
 +
 +      if (sample_type & PERF_SAMPLE_RAW) {
 +              if (data->raw) {
 +                      perf_output_put(handle, data->raw->size);
 +                      perf_output_copy(handle, data->raw->data,
 +                                       data->raw->size);
 +              } else {
 +                      struct {
 +                              u32     size;
 +                              u32     data;
 +                      } raw = {
 +                              .size = sizeof(u32),
 +                              .data = 0,
 +                      };
 +                      perf_output_put(handle, raw);
 +              }
 +      }
 +}
 +
 +void perf_prepare_sample(struct perf_event_header *header,
 +                       struct perf_sample_data *data,
 +                       struct perf_event *event,
 +                       struct pt_regs *regs)
 +{
 +      u64 sample_type = event->attr.sample_type;
 +
 +      header->type = PERF_RECORD_SAMPLE;
 +      header->size = sizeof(*header) + event->header_size;
 +
 +      header->misc = 0;
 +      header->misc |= perf_misc_flags(regs);
 +
 +      __perf_event_header__init_id(header, data, event);
 +
 +      if (sample_type & PERF_SAMPLE_IP)
 +              data->ip = perf_instruction_pointer(regs);
 +
 +      if (sample_type & PERF_SAMPLE_CALLCHAIN) {
 +              int size = 1;
 +
 +              data->callchain = perf_callchain(regs);
 +
 +              if (data->callchain)
 +                      size += data->callchain->nr;
 +
 +              header->size += size * sizeof(u64);
 +      }
 +
 +      if (sample_type & PERF_SAMPLE_RAW) {
 +              int size = sizeof(u32);
 +
 +              if (data->raw)
 +                      size += data->raw->size;
 +              else
 +                      size += sizeof(u32);
 +
 +              WARN_ON_ONCE(size & (sizeof(u64)-1));
 +              header->size += size;
 +      }
 +}
 +
 +static void perf_event_output(struct perf_event *event, int nmi,
 +                              struct perf_sample_data *data,
 +                              struct pt_regs *regs)
 +{
 +      struct perf_output_handle handle;
 +      struct perf_event_header header;
 +
 +      /* protect the callchain buffers */
 +      rcu_read_lock();
 +
 +      perf_prepare_sample(&header, data, event, regs);
 +
 +      if (perf_output_begin(&handle, event, header.size, nmi, 1))
 +              goto exit;
 +
 +      perf_output_sample(&handle, &header, data, event);
 +
 +      perf_output_end(&handle);
 +
 +exit:
 +      rcu_read_unlock();
 +}
 +
 +/*
 + * read event_id
 + */
 +
 +struct perf_read_event {
 +      struct perf_event_header        header;
 +
 +      u32                             pid;
 +      u32                             tid;
 +};
 +
 +static void
 +perf_event_read_event(struct perf_event *event,
 +                      struct task_struct *task)
 +{
 +      struct perf_output_handle handle;
 +      struct perf_sample_data sample;
 +      struct perf_read_event read_event = {
 +              .header = {
 +                      .type = PERF_RECORD_READ,
 +                      .misc = 0,
 +                      .size = sizeof(read_event) + event->read_size,
 +              },
 +              .pid = perf_event_pid(event, task),
 +              .tid = perf_event_tid(event, task),
 +      };
 +      int ret;
 +
 +      perf_event_header__init_id(&read_event.header, &sample, event);
 +      ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
 +      if (ret)
 +              return;
 +
 +      perf_output_put(&handle, read_event);
 +      perf_output_read(&handle, event);
 +      perf_event__output_id_sample(event, &handle, &sample);
 +
 +      perf_output_end(&handle);
 +}
 +
 +/*
 + * task tracking -- fork/exit
 + *
 + * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
 + */
 +
 +struct perf_task_event {
 +      struct task_struct              *task;
 +      struct perf_event_context       *task_ctx;
 +
 +      struct {
 +              struct perf_event_header        header;
 +
 +              u32                             pid;
 +              u32                             ppid;
 +              u32                             tid;
 +              u32                             ptid;
 +              u64                             time;
 +      } event_id;
 +};
 +
 +static void perf_event_task_output(struct perf_event *event,
 +                                   struct perf_task_event *task_event)
 +{
 +      struct perf_output_handle handle;
 +      struct perf_sample_data sample;
 +      struct task_struct *task = task_event->task;
 +      int ret, size = task_event->event_id.header.size;
 +
 +      perf_event_header__init_id(&task_event->event_id.header, &sample, event);
 +
 +      ret = perf_output_begin(&handle, event,
 +                              task_event->event_id.header.size, 0, 0);
 +      if (ret)
 +              goto out;
 +
 +      task_event->event_id.pid = perf_event_pid(event, task);
 +      task_event->event_id.ppid = perf_event_pid(event, current);
 +
 +      task_event->event_id.tid = perf_event_tid(event, task);
 +      task_event->event_id.ptid = perf_event_tid(event, current);
 +
 +      perf_output_put(&handle, task_event->event_id);
 +
 +      perf_event__output_id_sample(event, &handle, &sample);
 +
 +      perf_output_end(&handle);
 +out:
 +      task_event->event_id.header.size = size;
 +}
 +
 +static int perf_event_task_match(struct perf_event *event)
 +{
 +      if (event->state < PERF_EVENT_STATE_INACTIVE)
 +              return 0;
 +
 +      if (!event_filter_match(event))
 +              return 0;
 +
 +      if (event->attr.comm || event->attr.mmap ||
 +          event->attr.mmap_data || event->attr.task)
 +              return 1;
 +
 +      return 0;
 +}
 +
 +static void perf_event_task_ctx(struct perf_event_context *ctx,
 +                                struct perf_task_event *task_event)
 +{
 +      struct perf_event *event;
 +
 +      list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
 +              if (perf_event_task_match(event))
 +                      perf_event_task_output(event, task_event);
 +      }
 +}
 +
 +static void perf_event_task_event(struct perf_task_event *task_event)
 +{
 +      struct perf_cpu_context *cpuctx;
 +      struct perf_event_context *ctx;
 +      struct pmu *pmu;
 +      int ctxn;
 +
 +      rcu_read_lock();
 +      list_for_each_entry_rcu(pmu, &pmus, entry) {
 +              cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
 +              if (cpuctx->active_pmu != pmu)
 +                      goto next;
 +              perf_event_task_ctx(&cpuctx->ctx, task_event);
 +
 +              ctx = task_event->task_ctx;
 +              if (!ctx) {
 +                      ctxn = pmu->task_ctx_nr;
 +                      if (ctxn < 0)
 +                              goto next;
 +                      ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
 +              }
 +              if (ctx)
 +                      perf_event_task_ctx(ctx, task_event);
 +next:
 +              put_cpu_ptr(pmu->pmu_cpu_context);
 +      }
 +      rcu_read_unlock();
 +}
 +
 +static void perf_event_task(struct task_struct *task,
 +                            struct perf_event_context *task_ctx,
 +                            int new)
 +{
 +      struct perf_task_event task_event;
 +
 +      if (!atomic_read(&nr_comm_events) &&
 +          !atomic_read(&nr_mmap_events) &&
 +          !atomic_read(&nr_task_events))
 +              return;
 +
 +      task_event = (struct perf_task_event){
 +              .task     = task,
 +              .task_ctx = task_ctx,
 +              .event_id    = {
 +                      .header = {
 +                              .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
 +                              .misc = 0,
 +                              .size = sizeof(task_event.event_id),
 +                      },
 +                      /* .pid  */
 +                      /* .ppid */
 +                      /* .tid  */
 +                      /* .ptid */
 +                      .time = perf_clock(),
 +              },
 +      };
 +
 +      perf_event_task_event(&task_event);
 +}
 +
 +void perf_event_fork(struct task_struct *task)
 +{
 +      perf_event_task(task, NULL, 1);
 +}
 +
 +/*
 + * comm tracking
 + */
 +
 +struct perf_comm_event {
 +      struct task_struct      *task;
 +      char                    *comm;
 +      int                     comm_size;
 +
 +      struct {
 +              struct perf_event_header        header;
 +
 +              u32                             pid;
 +              u32                             tid;
 +      } event_id;
 +};
 +
 +static void perf_event_comm_output(struct perf_event *event,
 +                                   struct perf_comm_event *comm_event)
 +{
 +      struct perf_output_handle handle;
 +      struct perf_sample_data sample;
 +      int size = comm_event->event_id.header.size;
 +      int ret;
 +
 +      perf_event_header__init_id(&comm_event->event_id.header, &sample, event);
 +      ret = perf_output_begin(&handle, event,
 +                              comm_event->event_id.header.size, 0, 0);
 +
 +      if (ret)
 +              goto out;
 +
 +      comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
 +      comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
 +
 +      perf_output_put(&handle, comm_event->event_id);
 +      perf_output_copy(&handle, comm_event->comm,
 +                                 comm_event->comm_size);
 +
 +      perf_event__output_id_sample(event, &handle, &sample);
 +
 +      perf_output_end(&handle);
 +out:
 +      comm_event->event_id.header.size = size;
 +}
 +
 +static int perf_event_comm_match(struct perf_event *event)
 +{
 +      if (event->state < PERF_EVENT_STATE_INACTIVE)
 +              return 0;
 +
 +      if (!event_filter_match(event))
 +              return 0;
 +
 +      if (event->attr.comm)
 +              return 1;
 +
 +      return 0;
 +}
 +
 +static void perf_event_comm_ctx(struct perf_event_context *ctx,
 +                                struct perf_comm_event *comm_event)
 +{
 +      struct perf_event *event;
 +
 +      list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
 +              if (perf_event_comm_match(event))
 +                      perf_event_comm_output(event, comm_event);
 +      }
 +}
 +
 +static void perf_event_comm_event(struct perf_comm_event *comm_event)
 +{
 +      struct perf_cpu_context *cpuctx;
 +      struct perf_event_context *ctx;
 +      char comm[TASK_COMM_LEN];
 +      unsigned int size;
 +      struct pmu *pmu;
 +      int ctxn;
 +
 +      memset(comm, 0, sizeof(comm));
 +      strlcpy(comm, comm_event->task->comm, sizeof(comm));
 +      size = ALIGN(strlen(comm)+1, sizeof(u64));
 +
 +      comm_event->comm = comm;
 +      comm_event->comm_size = size;
 +
 +      comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
 +      rcu_read_lock();
 +      list_for_each_entry_rcu(pmu, &pmus, entry) {
 +              cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
 +              if (cpuctx->active_pmu != pmu)
 +                      goto next;
 +              perf_event_comm_ctx(&cpuctx->ctx, comm_event);
 +
 +              ctxn = pmu->task_ctx_nr;
 +              if (ctxn < 0)
 +                      goto next;
 +
 +              ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
 +              if (ctx)
 +                      perf_event_comm_ctx(ctx, comm_event);
 +next:
 +              put_cpu_ptr(pmu->pmu_cpu_context);
 +      }
 +      rcu_read_unlock();
 +}
 +
 +void perf_event_comm(struct task_struct *task)
 +{
 +      struct perf_comm_event comm_event;
 +      struct perf_event_context *ctx;
 +      int ctxn;
 +
 +      for_each_task_context_nr(ctxn) {
 +              ctx = task->perf_event_ctxp[ctxn];
 +              if (!ctx)
 +                      continue;
 +
 +              perf_event_enable_on_exec(ctx);
 +      }
 +
 +      if (!atomic_read(&nr_comm_events))
 +              return;
 +
 +      comm_event = (struct perf_comm_event){
 +              .task   = task,
 +              /* .comm      */
 +              /* .comm_size */
 +              .event_id  = {
 +                      .header = {
 +                              .type = PERF_RECORD_COMM,
 +                              .misc = 0,
 +                              /* .size */
 +                      },
 +                      /* .pid */
 +                      /* .tid */
 +              },
 +      };
 +
 +      perf_event_comm_event(&comm_event);
 +}
 +
 +/*
 + * mmap tracking
 + */
 +
 +struct perf_mmap_event {
 +      struct vm_area_struct   *vma;
 +
 +      const char              *file_name;
 +      int                     file_size;
 +
 +      struct {
 +              struct perf_event_header        header;
 +
 +              u32                             pid;
 +              u32                             tid;
 +              u64                             start;
 +              u64                             len;
 +              u64                             pgoff;
 +      } event_id;
 +};
 +
 +static void perf_event_mmap_output(struct perf_event *event,
 +                                   struct perf_mmap_event *mmap_event)
 +{
 +      struct perf_output_handle handle;
 +      struct perf_sample_data sample;
 +      int size = mmap_event->event_id.header.size;
 +      int ret;
 +
 +      perf_event_header__init_id(&mmap_event->event_id.header, &sample, event);
 +      ret = perf_output_begin(&handle, event,
 +                              mmap_event->event_id.header.size, 0, 0);
 +      if (ret)
 +              goto out;
 +
 +      mmap_event->event_id.pid = perf_event_pid(event, current);
 +      mmap_event->event_id.tid = perf_event_tid(event, current);
 +
 +      perf_output_put(&handle, mmap_event->event_id);
 +      perf_output_copy(&handle, mmap_event->file_name,
 +                                 mmap_event->file_size);
 +
 +      perf_event__output_id_sample(event, &handle, &sample);
 +
 +      perf_output_end(&handle);
 +out:
 +      mmap_event->event_id.header.size = size;
 +}
 +
 +static int perf_event_mmap_match(struct perf_event *event,
 +                                 struct perf_mmap_event *mmap_event,
 +                                 int executable)
 +{
 +      if (event->state < PERF_EVENT_STATE_INACTIVE)
 +              return 0;
 +
 +      if (!event_filter_match(event))
 +              return 0;
 +
 +      if ((!executable && event->attr.mmap_data) ||
 +          (executable && event->attr.mmap))
 +              return 1;
 +
 +      return 0;
 +}
 +
 +static void perf_event_mmap_ctx(struct perf_event_context *ctx,
 +                                struct perf_mmap_event *mmap_event,
 +                                int executable)
 +{
 +      struct perf_event *event;
 +
 +      list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
 +              if (perf_event_mmap_match(event, mmap_event, executable))
 +                      perf_event_mmap_output(event, mmap_event);
 +      }
 +}
 +
 +static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
 +{
 +      struct perf_cpu_context *cpuctx;
 +      struct perf_event_context *ctx;
 +      struct vm_area_struct *vma = mmap_event->vma;
 +      struct file *file = vma->vm_file;
 +      unsigned int size;
 +      char tmp[16];
 +      char *buf = NULL;
 +      const char *name;
 +      struct pmu *pmu;
 +      int ctxn;
 +
 +      memset(tmp, 0, sizeof(tmp));
 +
 +      if (file) {
 +              /*
 +               * d_path works from the end of the buffer backwards, so we
 +               * need to add enough zero bytes after the string to handle
 +               * the 64bit alignment we do later.
 +               */
 +              buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
 +              if (!buf) {
 +                      name = strncpy(tmp, "//enomem", sizeof(tmp));
 +                      goto got_name;
 +              }
 +              name = d_path(&file->f_path, buf, PATH_MAX);
 +              if (IS_ERR(name)) {
 +                      name = strncpy(tmp, "//toolong", sizeof(tmp));
 +                      goto got_name;
 +              }
 +      } else {
 +              if (arch_vma_name(mmap_event->vma)) {
 +                      name = strncpy(tmp, arch_vma_name(mmap_event->vma),
 +                                     sizeof(tmp));
 +                      goto got_name;
 +              }
 +
 +              if (!vma->vm_mm) {
 +                      name = strncpy(tmp, "[vdso]", sizeof(tmp));
 +                      goto got_name;
 +              } else if (vma->vm_start <= vma->vm_mm->start_brk &&
 +                              vma->vm_end >= vma->vm_mm->brk) {
 +                      name = strncpy(tmp, "[heap]", sizeof(tmp));
 +                      goto got_name;
 +              } else if (vma->vm_start <= vma->vm_mm->start_stack &&
 +                              vma->vm_end >= vma->vm_mm->start_stack) {
 +                      name = strncpy(tmp, "[stack]", sizeof(tmp));
 +                      goto got_name;
 +              }
 +
 +              name = strncpy(tmp, "//anon", sizeof(tmp));
 +              goto got_name;
 +      }
 +
 +got_name:
 +      size = ALIGN(strlen(name)+1, sizeof(u64));
 +
 +      mmap_event->file_name = name;
 +      mmap_event->file_size = size;
 +
 +      mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
 +
 +      rcu_read_lock();
 +      list_for_each_entry_rcu(pmu, &pmus, entry) {
 +              cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
 +              if (cpuctx->active_pmu != pmu)
 +                      goto next;
 +              perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
 +                                      vma->vm_flags & VM_EXEC);
 +
 +              ctxn = pmu->task_ctx_nr;
 +              if (ctxn < 0)
 +                      goto next;
 +
 +              ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
 +              if (ctx) {
 +                      perf_event_mmap_ctx(ctx, mmap_event,
 +                                      vma->vm_flags & VM_EXEC);
 +              }
 +next:
 +              put_cpu_ptr(pmu->pmu_cpu_context);
 +      }
 +      rcu_read_unlock();
 +
 +      kfree(buf);
 +}
 +
 +void perf_event_mmap(struct vm_area_struct *vma)
 +{
 +      struct perf_mmap_event mmap_event;
 +
 +      if (!atomic_read(&nr_mmap_events))
 +              return;
 +
 +      mmap_event = (struct perf_mmap_event){
 +              .vma    = vma,
 +              /* .file_name */
 +              /* .file_size */
 +              .event_id  = {
 +                      .header = {
 +                              .type = PERF_RECORD_MMAP,
 +                              .misc = PERF_RECORD_MISC_USER,
 +                              /* .size */
 +                      },
 +                      /* .pid */
 +                      /* .tid */
 +                      .start  = vma->vm_start,
 +                      .len    = vma->vm_end - vma->vm_start,
 +                      .pgoff  = (u64)vma->vm_pgoff << PAGE_SHIFT,
 +              },
 +      };
 +
 +      perf_event_mmap_event(&mmap_event);
 +}
 +
 +/*
 + * IRQ throttle logging
 + */
 +
 +static void perf_log_throttle(struct perf_event *event, int enable)
 +{
 +      struct perf_output_handle handle;
 +      struct perf_sample_data sample;
 +      int ret;
 +
 +      struct {
 +              struct perf_event_header        header;
 +              u64                             time;
 +              u64                             id;
 +              u64                             stream_id;
 +      } throttle_event = {
 +              .header = {
 +                      .type = PERF_RECORD_THROTTLE,
 +                      .misc = 0,
 +                      .size = sizeof(throttle_event),
 +              },
 +              .time           = perf_clock(),
 +              .id             = primary_event_id(event),
 +              .stream_id      = event->id,
 +      };
 +
 +      if (enable)
 +              throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
 +
 +      perf_event_header__init_id(&throttle_event.header, &sample, event);
 +
 +      ret = perf_output_begin(&handle, event,
 +                              throttle_event.header.size, 1, 0);
 +      if (ret)
 +              return;
 +
 +      perf_output_put(&handle, throttle_event);
 +      perf_event__output_id_sample(event, &handle, &sample);
 +      perf_output_end(&handle);
 +}
 +
 +/*
 + * Generic event overflow handling, sampling.
 + */
 +
 +static int __perf_event_overflow(struct perf_event *event, int nmi,
 +                                 int throttle, struct perf_sample_data *data,
 +                                 struct pt_regs *regs)
 +{
 +      int events = atomic_read(&event->event_limit);
 +      struct hw_perf_event *hwc = &event->hw;
 +      int ret = 0;
 +
 +      /*
 +       * Non-sampling counters might still use the PMI to fold short
 +       * hardware counters, ignore those.
 +       */
 +      if (unlikely(!is_sampling_event(event)))
 +              return 0;
 +
 +      if (unlikely(hwc->interrupts >= max_samples_per_tick)) {
 +              if (throttle) {
 +                      hwc->interrupts = MAX_INTERRUPTS;
 +                      perf_log_throttle(event, 0);
 +                      ret = 1;
 +              }
 +      } else
 +              hwc->interrupts++;
 +
 +      if (event->attr.freq) {
 +              u64 now = perf_clock();
 +              s64 delta = now - hwc->freq_time_stamp;
 +
 +              hwc->freq_time_stamp = now;
 +
 +              if (delta > 0 && delta < 2*TICK_NSEC)
 +                      perf_adjust_period(event, delta, hwc->last_period);
 +      }
 +
 +      /*
 +       * XXX event_limit might not quite work as expected on inherited
 +       * events
 +       */
 +
 +      event->pending_kill = POLL_IN;
 +      if (events && atomic_dec_and_test(&event->event_limit)) {
 +              ret = 1;
 +              event->pending_kill = POLL_HUP;
 +              if (nmi) {
 +                      event->pending_disable = 1;
 +                      irq_work_queue(&event->pending);
 +              } else
 +                      perf_event_disable(event);
 +      }
 +
 +      if (event->overflow_handler)
 +              event->overflow_handler(event, nmi, data, regs);
 +      else
 +              perf_event_output(event, nmi, data, regs);
 +
 +      return ret;
 +}
 +
 +int perf_event_overflow(struct perf_event *event, int nmi,
 +                        struct perf_sample_data *data,
 +                        struct pt_regs *regs)
 +{
 +      return __perf_event_overflow(event, nmi, 1, data, regs);
 +}
 +
 +/*
 + * Generic software event infrastructure
 + */
 +
 +struct swevent_htable {
 +      struct swevent_hlist            *swevent_hlist;
 +      struct mutex                    hlist_mutex;
 +      int                             hlist_refcount;
 +
 +      /* Recursion avoidance in each contexts */
 +      int                             recursion[PERF_NR_CONTEXTS];
 +};
 +
 +static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
 +
 +/*
 + * We directly increment event->count and keep a second value in
 + * event->hw.period_left to count intervals. This period event
 + * is kept in the range [-sample_period, 0] so that we can use the
 + * sign as trigger.
 + */
 +
 +static u64 perf_swevent_set_period(struct perf_event *event)
 +{
 +      struct hw_perf_event *hwc = &event->hw;
 +      u64 period = hwc->last_period;
 +      u64 nr, offset;
 +      s64 old, val;
 +
 +      hwc->last_period = hwc->sample_period;
 +
 +again:
 +      old = val = local64_read(&hwc->period_left);
 +      if (val < 0)
 +              return 0;
 +
 +      nr = div64_u64(period + val, period);
 +      offset = nr * period;
 +      val -= offset;
 +      if (local64_cmpxchg(&hwc->period_left, old, val) != old)
 +              goto again;
 +
 +      return nr;
 +}
 +
 +static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
 +                                  int nmi, struct perf_sample_data *data,
 +                                  struct pt_regs *regs)
 +{
 +      struct hw_perf_event *hwc = &event->hw;
 +      int throttle = 0;
 +
 +      data->period = event->hw.last_period;
 +      if (!overflow)
 +              overflow = perf_swevent_set_period(event);
 +
 +      if (hwc->interrupts == MAX_INTERRUPTS)
 +              return;
 +
 +      for (; overflow; overflow--) {
 +              if (__perf_event_overflow(event, nmi, throttle,
 +                                          data, regs)) {
 +                      /*
 +                       * We inhibit the overflow from happening when
 +                       * hwc->interrupts == MAX_INTERRUPTS.
 +                       */
 +                      break;
 +              }
 +              throttle = 1;
 +      }
 +}
 +
 +static void perf_swevent_event(struct perf_event *event, u64 nr,
 +                             int nmi, struct perf_sample_data *data,
 +                             struct pt_regs *regs)
 +{
 +      struct hw_perf_event *hwc = &event->hw;
 +
 +      local64_add(nr, &event->count);
 +
 +      if (!regs)
 +              return;
 +
 +      if (!is_sampling_event(event))
 +              return;
 +
 +      if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
 +              return perf_swevent_overflow(event, 1, nmi, data, regs);
 +
 +      if (local64_add_negative(nr, &hwc->period_left))
 +              return;
 +
 +      perf_swevent_overflow(event, 0, nmi, data, regs);
 +}
 +
 +static int perf_exclude_event(struct perf_event *event,
 +                            struct pt_regs *regs)
 +{
 +      if (event->hw.state & PERF_HES_STOPPED)
 +              return 1;
 +
 +      if (regs) {
 +              if (event->attr.exclude_user && user_mode(regs))
 +                      return 1;
 +
 +              if (event->attr.exclude_kernel && !user_mode(regs))
 +                      return 1;
 +      }
 +
 +      return 0;
 +}
 +
 +static int perf_swevent_match(struct perf_event *event,
 +                              enum perf_type_id type,
 +                              u32 event_id,
 +                              struct perf_sample_data *data,
 +                              struct pt_regs *regs)
 +{
 +      if (event->attr.type != type)
 +              return 0;
 +
 +      if (event->attr.config != event_id)
 +              return 0;
 +
 +      if (perf_exclude_event(event, regs))
 +              return 0;
 +
 +      return 1;
 +}
 +
 +static inline u64 swevent_hash(u64 type, u32 event_id)
 +{
 +      u64 val = event_id | (type << 32);
 +
 +      return hash_64(val, SWEVENT_HLIST_BITS);
 +}
 +
 +static inline struct hlist_head *
 +__find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
 +{
 +      u64 hash = swevent_hash(type, event_id);
 +
 +      return &hlist->heads[hash];
 +}
 +
 +/* For the read side: events when they trigger */
 +static inline struct hlist_head *
 +find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
 +{
 +      struct swevent_hlist *hlist;
 +
 +      hlist = rcu_dereference(swhash->swevent_hlist);
 +      if (!hlist)
 +              return NULL;
 +
 +      return __find_swevent_head(hlist, type, event_id);
 +}
 +
 +/* For the event head insertion and removal in the hlist */
 +static inline struct hlist_head *
 +find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
 +{
 +      struct swevent_hlist *hlist;
 +      u32 event_id = event->attr.config;
 +      u64 type = event->attr.type;
 +
 +      /*
 +       * Event scheduling is always serialized against hlist allocation
 +       * and release. Which makes the protected version suitable here.
 +       * The context lock guarantees that.
 +       */
 +      hlist = rcu_dereference_protected(swhash->swevent_hlist,
 +                                        lockdep_is_held(&event->ctx->lock));
 +      if (!hlist)
 +              return NULL;
 +
 +      return __find_swevent_head(hlist, type, event_id);
 +}
 +
 +static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
 +                                  u64 nr, int nmi,
 +                                  struct perf_sample_data *data,
 +                                  struct pt_regs *regs)
 +{
 +      struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
 +      struct perf_event *event;
 +      struct hlist_node *node;
 +      struct hlist_head *head;
 +
 +      rcu_read_lock();
 +      head = find_swevent_head_rcu(swhash, type, event_id);
 +      if (!head)
 +              goto end;
 +
 +      hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
 +              if (perf_swevent_match(event, type, event_id, data, regs))
 +                      perf_swevent_event(event, nr, nmi, data, regs);
 +      }
 +end:
 +      rcu_read_unlock();
 +}
 +
 +int perf_swevent_get_recursion_context(void)
 +{
 +      struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
 +
 +      return get_recursion_context(swhash->recursion);
 +}
 +EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
 +
 +inline void perf_swevent_put_recursion_context(int rctx)
 +{
 +      struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
 +
 +      put_recursion_context(swhash->recursion, rctx);
 +}
 +
 +void __perf_sw_event(u32 event_id, u64 nr, int nmi,
 +                          struct pt_regs *regs, u64 addr)
 +{
 +      struct perf_sample_data data;
 +      int rctx;
 +
 +      preempt_disable_notrace();
 +      rctx = perf_swevent_get_recursion_context();
 +      if (rctx < 0)
 +              return;
 +
 +      perf_sample_data_init(&data, addr);
 +
 +      do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
 +
 +      perf_swevent_put_recursion_context(rctx);
 +      preempt_enable_notrace();
 +}
 +
 +static void perf_swevent_read(struct perf_event *event)
 +{
 +}
 +
 +static int perf_swevent_add(struct perf_event *event, int flags)
 +{
 +      struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
 +      struct hw_perf_event *hwc = &event->hw;
 +      struct hlist_head *head;
 +
 +      if (is_sampling_event(event)) {
 +              hwc->last_period = hwc->sample_period;
 +              perf_swevent_set_period(event);
 +      }
 +
 +      hwc->state = !(flags & PERF_EF_START);
 +
 +      head = find_swevent_head(swhash, event);
 +      if (WARN_ON_ONCE(!head))
 +              return -EINVAL;
 +
 +      hlist_add_head_rcu(&event->hlist_entry, head);
 +
 +      return 0;
 +}
 +
 +static void perf_swevent_del(struct perf_event *event, int flags)
 +{
 +      hlist_del_rcu(&event->hlist_entry);
 +}
 +
 +static void perf_swevent_start(struct perf_event *event, int flags)
 +{
 +      event->hw.state = 0;
 +}
 +
 +static void perf_swevent_stop(struct perf_event *event, int flags)
 +{
 +      event->hw.state = PERF_HES_STOPPED;
 +}
 +
 +/* Deref the hlist from the update side */
 +static inline struct swevent_hlist *
 +swevent_hlist_deref(struct swevent_htable *swhash)
 +{
 +      return rcu_dereference_protected(swhash->swevent_hlist,
 +                                       lockdep_is_held(&swhash->hlist_mutex));
 +}
 +
-       call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
 +static void swevent_hlist_release(struct swevent_htable *swhash)
 +{
 +      struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
 +
 +      if (!hlist)
 +              return;
 +
 +      rcu_assign_pointer(swhash->swevent_hlist, NULL);
++      kfree_rcu(hlist, rcu_head);
 +}
 +
 +static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
 +{
 +      struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
 +
 +      mutex_lock(&swhash->hlist_mutex);
 +
 +      if (!--swhash->hlist_refcount)
 +              swevent_hlist_release(swhash);
 +
 +      mutex_unlock(&swhash->hlist_mutex);
 +}
 +
 +static void swevent_hlist_put(struct perf_event *event)
 +{
 +      int cpu;
 +
 +      if (event->cpu != -1) {
 +              swevent_hlist_put_cpu(event, event->cpu);
 +              return;
 +      }
 +
 +      for_each_possible_cpu(cpu)
 +              swevent_hlist_put_cpu(event, cpu);
 +}
 +
 +static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
 +{
 +      struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
 +      int err = 0;
 +
 +      mutex_lock(&swhash->hlist_mutex);
 +
 +      if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
 +              struct swevent_hlist *hlist;
 +
 +              hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
 +              if (!hlist) {
 +                      err = -ENOMEM;
 +                      goto exit;
 +              }
 +              rcu_assign_pointer(swhash->swevent_hlist, hlist);
 +      }
 +      swhash->hlist_refcount++;
 +exit:
 +      mutex_unlock(&swhash->hlist_mutex);
 +
 +      return err;
 +}
 +
 +static int swevent_hlist_get(struct perf_event *event)
 +{
 +      int err;
 +      int cpu, failed_cpu;
 +
 +      if (event->cpu != -1)
 +              return swevent_hlist_get_cpu(event, event->cpu);
 +
 +      get_online_cpus();
 +      for_each_possible_cpu(cpu) {
 +              err = swevent_hlist_get_cpu(event, cpu);
 +              if (err) {
 +                      failed_cpu = cpu;
 +                      goto fail;
 +              }
 +      }
 +      put_online_cpus();
 +
 +      return 0;
 +fail:
 +      for_each_possible_cpu(cpu) {
 +              if (cpu == failed_cpu)
 +                      break;
 +              swevent_hlist_put_cpu(event, cpu);
 +      }
 +
 +      put_online_cpus();
 +      return err;
 +}
 +
 +struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
 +
 +static void sw_perf_event_destroy(struct perf_event *event)
 +{
 +      u64 event_id = event->attr.config;
 +
 +      WARN_ON(event->parent);
 +
 +      jump_label_dec(&perf_swevent_enabled[event_id]);
 +      swevent_hlist_put(event);
 +}
 +
 +static int perf_swevent_init(struct perf_event *event)
 +{
 +      int event_id = event->attr.config;
 +
 +      if (event->attr.type != PERF_TYPE_SOFTWARE)
 +              return -ENOENT;
 +
 +      switch (event_id) {
 +      case PERF_COUNT_SW_CPU_CLOCK:
 +      case PERF_COUNT_SW_TASK_CLOCK:
 +              return -ENOENT;
 +
 +      default:
 +              break;
 +      }
 +
 +      if (event_id >= PERF_COUNT_SW_MAX)
 +              return -ENOENT;
 +
 +      if (!event->parent) {
 +              int err;
 +
 +              err = swevent_hlist_get(event);
 +              if (err)
 +                      return err;
 +
 +              jump_label_inc(&perf_swevent_enabled[event_id]);
 +              event->destroy = sw_perf_event_destroy;
 +      }
 +
 +      return 0;
 +}
 +
 +static struct pmu perf_swevent = {
 +      .task_ctx_nr    = perf_sw_context,
 +
 +      .event_init     = perf_swevent_init,
 +      .add            = perf_swevent_add,
 +      .del            = perf_swevent_del,
 +      .start          = perf_swevent_start,
 +      .stop           = perf_swevent_stop,
 +      .read           = perf_swevent_read,
 +};
 +
 +#ifdef CONFIG_EVENT_TRACING
 +
 +static int perf_tp_filter_match(struct perf_event *event,
 +                              struct perf_sample_data *data)
 +{
 +      void *record = data->raw->data;
 +
 +      if (likely(!event->filter) || filter_match_preds(event->filter, record))
 +              return 1;
 +      return 0;
 +}
 +
 +static int perf_tp_event_match(struct perf_event *event,
 +                              struct perf_sample_data *data,
 +                              struct pt_regs *regs)
 +{
 +      if (event->hw.state & PERF_HES_STOPPED)
 +              return 0;
 +      /*
 +       * All tracepoints are from kernel-space.
 +       */
 +      if (event->attr.exclude_kernel)
 +              return 0;
 +
 +      if (!perf_tp_filter_match(event, data))
 +              return 0;
 +
 +      return 1;
 +}
 +
 +void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
 +                 struct pt_regs *regs, struct hlist_head *head, int rctx)
 +{
 +      struct perf_sample_data data;
 +      struct perf_event *event;
 +      struct hlist_node *node;
 +
 +      struct perf_raw_record raw = {
 +              .size = entry_size,
 +              .data = record,
 +      };
 +
 +      perf_sample_data_init(&data, addr);
 +      data.raw = &raw;
 +
 +      hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
 +              if (perf_tp_event_match(event, &data, regs))
 +                      perf_swevent_event(event, count, 1, &data, regs);
 +      }
 +
 +      perf_swevent_put_recursion_context(rctx);
 +}
 +EXPORT_SYMBOL_GPL(perf_tp_event);
 +
 +static void tp_perf_event_destroy(struct perf_event *event)
 +{
 +      perf_trace_destroy(event);
 +}
 +
 +static int perf_tp_event_init(struct perf_event *event)
 +{
 +      int err;
 +
 +      if (event->attr.type != PERF_TYPE_TRACEPOINT)
 +              return -ENOENT;
 +
 +      err = perf_trace_init(event);
 +      if (err)
 +              return err;
 +
 +      event->destroy = tp_perf_event_destroy;
 +
 +      return 0;
 +}
 +
 +static struct pmu perf_tracepoint = {
 +      .task_ctx_nr    = perf_sw_context,
 +
 +      .event_init     = perf_tp_event_init,
 +      .add            = perf_trace_add,
 +      .del            = perf_trace_del,
 +      .start          = perf_swevent_start,
 +      .stop           = perf_swevent_stop,
 +      .read           = perf_swevent_read,
 +};
 +
 +static inline void perf_tp_register(void)
 +{
 +      perf_pmu_register(&perf_tracepoint, "tracepoint", PERF_TYPE_TRACEPOINT);
 +}
 +
 +static int perf_event_set_filter(struct perf_event *event, void __user *arg)
 +{
 +      char *filter_str;
 +      int ret;
 +
 +      if (event->attr.type != PERF_TYPE_TRACEPOINT)
 +              return -EINVAL;
 +
 +      filter_str = strndup_user(arg, PAGE_SIZE);
 +      if (IS_ERR(filter_str))
 +              return PTR_ERR(filter_str);
 +
 +      ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
 +
 +      kfree(filter_str);
 +      return ret;
 +}
 +
 +static void perf_event_free_filter(struct perf_event *event)
 +{
 +      ftrace_profile_free_filter(event);
 +}
 +
 +#else
 +
 +static inline void perf_tp_register(void)
 +{
 +}
 +
 +static int perf_event_set_filter(struct perf_event *event, void __user *arg)
 +{
 +      return -ENOENT;
 +}
 +
 +static void perf_event_free_filter(struct perf_event *event)
 +{
 +}
 +
 +#endif /* CONFIG_EVENT_TRACING */
 +
 +#ifdef CONFIG_HAVE_HW_BREAKPOINT
 +void perf_bp_event(struct perf_event *bp, void *data)
 +{
 +      struct perf_sample_data sample;
 +      struct pt_regs *regs = data;
 +
 +      perf_sample_data_init(&sample, bp->attr.bp_addr);
 +
 +      if (!bp->hw.state && !perf_exclude_event(bp, regs))
 +              perf_swevent_event(bp, 1, 1, &sample, regs);
 +}
 +#endif
 +
 +/*
 + * hrtimer based swevent callback
 + */
 +
 +static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
 +{
 +      enum hrtimer_restart ret = HRTIMER_RESTART;
 +      struct perf_sample_data data;
 +      struct pt_regs *regs;
 +      struct perf_event *event;
 +      u64 period;
 +
 +      event = container_of(hrtimer, struct perf_event, hw.hrtimer);
 +
 +      if (event->state != PERF_EVENT_STATE_ACTIVE)
 +              return HRTIMER_NORESTART;
 +
 +      event->pmu->read(event);
 +
 +      perf_sample_data_init(&data, 0);
 +      data.period = event->hw.last_period;
 +      regs = get_irq_regs();
 +
 +      if (regs && !perf_exclude_event(event, regs)) {
 +              if (!(event->attr.exclude_idle && current->pid == 0))
 +                      if (perf_event_overflow(event, 0, &data, regs))
 +                              ret = HRTIMER_NORESTART;
 +      }
 +
 +      period = max_t(u64, 10000, event->hw.sample_period);
 +      hrtimer_forward_now(hrtimer, ns_to_ktime(period));
 +
 +      return ret;
 +}
 +
 +static void perf_swevent_start_hrtimer(struct perf_event *event)
 +{
 +      struct hw_perf_event *hwc = &event->hw;
 +      s64 period;
 +
 +      if (!is_sampling_event(event))
 +              return;
 +
 +      period = local64_read(&hwc->period_left);
 +      if (period) {
 +              if (period < 0)
 +                      period = 10000;
 +
 +              local64_set(&hwc->period_left, 0);
 +      } else {
 +              period = max_t(u64, 10000, hwc->sample_period);
 +      }
 +      __hrtimer_start_range_ns(&hwc->hrtimer,
 +                              ns_to_ktime(period), 0,
 +                              HRTIMER_MODE_REL_PINNED, 0);
 +}
 +
 +static void perf_swevent_cancel_hrtimer(struct perf_event *event)
 +{
 +      struct hw_perf_event *hwc = &event->hw;
 +
 +      if (is_sampling_event(event)) {
 +              ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
 +              local64_set(&hwc->period_left, ktime_to_ns(remaining));
 +
 +              hrtimer_cancel(&hwc->hrtimer);
 +      }
 +}
 +
 +static void perf_swevent_init_hrtimer(struct perf_event *event)
 +{
 +      struct hw_perf_event *hwc = &event->hw;
 +
 +      if (!is_sampling_event(event))
 +              return;
 +
 +      hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 +      hwc->hrtimer.function = perf_swevent_hrtimer;
 +
 +      /*
 +       * Since hrtimers have a fixed rate, we can do a static freq->period
 +       * mapping and avoid the whole period adjust feedback stuff.
 +       */
 +      if (event->attr.freq) {
 +              long freq = event->attr.sample_freq;
 +
 +              event->attr.sample_period = NSEC_PER_SEC / freq;
 +              hwc->sample_period = event->attr.sample_period;
 +              local64_set(&hwc->period_left, hwc->sample_period);
 +              event->attr.freq = 0;
 +      }
 +}
 +
 +/*
 + * Software event: cpu wall time clock
 + */
 +
 +static void cpu_clock_event_update(struct perf_event *event)
 +{
 +      s64 prev;
 +      u64 now;
 +
 +      now = local_clock();
 +      prev = local64_xchg(&event->hw.prev_count, now);
 +      local64_add(now - prev, &event->count);
 +}
 +
 +static void cpu_clock_event_start(struct perf_event *event, int flags)
 +{
 +      local64_set(&event->hw.prev_count, local_clock());
 +      perf_swevent_start_hrtimer(event);
 +}
 +
 +static void cpu_clock_event_stop(struct perf_event *event, int flags)
 +{
 +      perf_swevent_cancel_hrtimer(event);
 +      cpu_clock_event_update(event);
 +}
 +
 +static int cpu_clock_event_add(struct perf_event *event, int flags)
 +{
 +      if (flags & PERF_EF_START)
 +              cpu_clock_event_start(event, flags);
 +
 +      return 0;
 +}
 +
 +static void cpu_clock_event_del(struct perf_event *event, int flags)
 +{
 +      cpu_clock_event_stop(event, flags);
 +}
 +
 +static void cpu_clock_event_read(struct perf_event *event)
 +{
 +      cpu_clock_event_update(event);
 +}
 +
 +static int cpu_clock_event_init(struct perf_event *event)
 +{
 +      if (event->attr.type != PERF_TYPE_SOFTWARE)
 +              return -ENOENT;
 +
 +      if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
 +              return -ENOENT;
 +
 +      perf_swevent_init_hrtimer(event);
 +
 +      return 0;
 +}
 +
 +static struct pmu perf_cpu_clock = {
 +      .task_ctx_nr    = perf_sw_context,
 +
 +      .event_init     = cpu_clock_event_init,
 +      .add            = cpu_clock_event_add,
 +      .del            = cpu_clock_event_del,
 +      .start          = cpu_clock_event_start,
 +      .stop           = cpu_clock_event_stop,
 +      .read           = cpu_clock_event_read,
 +};
 +
 +/*
 + * Software event: task time clock
 + */
 +
 +static void task_clock_event_update(struct perf_event *event, u64 now)
 +{
 +      u64 prev;
 +      s64 delta;
 +
 +      prev = local64_xchg(&event->hw.prev_count, now);
 +      delta = now - prev;
 +      local64_add(delta, &event->count);
 +}
 +
 +static void task_clock_event_start(struct perf_event *event, int flags)
 +{
 +      local64_set(&event->hw.prev_count, event->ctx->time);
 +      perf_swevent_start_hrtimer(event);
 +}
 +
 +static void task_clock_event_stop(struct perf_event *event, int flags)
 +{
 +      perf_swevent_cancel_hrtimer(event);
 +      task_clock_event_update(event, event->ctx->time);
 +}
 +
 +static int task_clock_event_add(struct perf_event *event, int flags)
 +{
 +      if (flags & PERF_EF_START)
 +              task_clock_event_start(event, flags);
 +
 +      return 0;
 +}
 +
 +static void task_clock_event_del(struct perf_event *event, int flags)
 +{
 +      task_clock_event_stop(event, PERF_EF_UPDATE);
 +}
 +
 +static void task_clock_event_read(struct perf_event *event)
 +{
 +      u64 now = perf_clock();
 +      u64 delta = now - event->ctx->timestamp;
 +      u64 time = event->ctx->time + delta;
 +
 +      task_clock_event_update(event, time);
 +}
 +
 +static int task_clock_event_init(struct perf_event *event)
 +{
 +      if (event->attr.type != PERF_TYPE_SOFTWARE)
 +              return -ENOENT;
 +
 +      if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
 +              return -ENOENT;
 +
 +      perf_swevent_init_hrtimer(event);
 +
 +      return 0;
 +}
 +
 +static struct pmu perf_task_clock = {
 +      .task_ctx_nr    = perf_sw_context,
 +
 +      .event_init     = task_clock_event_init,
 +      .add            = task_clock_event_add,
 +      .del            = task_clock_event_del,
 +      .start          = task_clock_event_start,
 +      .stop           = task_clock_event_stop,
 +      .read           = task_clock_event_read,
 +};
 +
 +static void perf_pmu_nop_void(struct pmu *pmu)
 +{
 +}
 +
 +static int perf_pmu_nop_int(struct pmu *pmu)
 +{
 +      return 0;
 +}
 +
 +static void perf_pmu_start_txn(struct pmu *pmu)
 +{
 +      perf_pmu_disable(pmu);
 +}
 +
 +static int perf_pmu_commit_txn(struct pmu *pmu)
 +{
 +      perf_pmu_enable(pmu);
 +      return 0;
 +}
 +
 +static void perf_pmu_cancel_txn(struct pmu *pmu)
 +{
 +      perf_pmu_enable(pmu);
 +}
 +
 +/*
 + * Ensures all contexts with the same task_ctx_nr have the same
 + * pmu_cpu_context too.
 + */
 +static void *find_pmu_context(int ctxn)
 +{
 +      struct pmu *pmu;
 +
 +      if (ctxn < 0)
 +              return NULL;
 +
 +      list_for_each_entry(pmu, &pmus, entry) {
 +              if (pmu->task_ctx_nr == ctxn)
 +                      return pmu->pmu_cpu_context;
 +      }
 +
 +      return NULL;
 +}
 +
 +static void update_pmu_context(struct pmu *pmu, struct pmu *old_pmu)
 +{
 +      int cpu;
 +
 +      for_each_possible_cpu(cpu) {
 +              struct perf_cpu_context *cpuctx;
 +
 +              cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
 +
 +              if (cpuctx->active_pmu == old_pmu)
 +                      cpuctx->active_pmu = pmu;
 +      }
 +}
 +
 +static void free_pmu_context(struct pmu *pmu)
 +{
 +      struct pmu *i;
 +
 +      mutex_lock(&pmus_lock);
 +      /*
 +       * Like a real lame refcount.
 +       */
 +      list_for_each_entry(i, &pmus, entry) {
 +              if (i->pmu_cpu_context == pmu->pmu_cpu_context) {
 +                      update_pmu_context(i, pmu);
 +                      goto out;
 +              }
 +      }
 +
 +      free_percpu(pmu->pmu_cpu_context);
 +out:
 +      mutex_unlock(&pmus_lock);
 +}
 +static struct idr pmu_idr;
 +
 +static ssize_t
 +type_show(struct device *dev, struct device_attribute *attr, char *page)
 +{
 +      struct pmu *pmu = dev_get_drvdata(dev);
 +
 +      return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
 +}
 +
 +static struct device_attribute pmu_dev_attrs[] = {
 +       __ATTR_RO(type),
 +       __ATTR_NULL,
 +};
 +
 +static int pmu_bus_running;
 +static struct bus_type pmu_bus = {
 +      .name           = "event_source",
 +      .dev_attrs      = pmu_dev_attrs,
 +};
 +
 +static void pmu_dev_release(struct device *dev)
 +{
 +      kfree(dev);
 +}
 +
 +static int pmu_dev_alloc(struct pmu *pmu)
 +{
 +      int ret = -ENOMEM;
 +
 +      pmu->dev = kzalloc(sizeof(struct device), GFP_KERNEL);
 +      if (!pmu->dev)
 +              goto out;
 +
 +      device_initialize(pmu->dev);
 +      ret = dev_set_name(pmu->dev, "%s", pmu->name);
 +      if (ret)
 +              goto free_dev;
 +
 +      dev_set_drvdata(pmu->dev, pmu);
 +      pmu->dev->bus = &pmu_bus;
 +      pmu->dev->release = pmu_dev_release;
 +      ret = device_add(pmu->dev);
 +      if (ret)
 +              goto free_dev;
 +
 +out:
 +      return ret;
 +
 +free_dev:
 +      put_device(pmu->dev);
 +      goto out;
 +}
 +
 +static struct lock_class_key cpuctx_mutex;
 +
 +int perf_pmu_register(struct pmu *pmu, char *name, int type)
 +{
 +      int cpu, ret;
 +
 +      mutex_lock(&pmus_lock);
 +      ret = -ENOMEM;
 +      pmu->pmu_disable_count = alloc_percpu(int);
 +      if (!pmu->pmu_disable_count)
 +              goto unlock;
 +
 +      pmu->type = -1;
 +      if (!name)
 +              goto skip_type;
 +      pmu->name = name;
 +
 +      if (type < 0) {
 +              int err = idr_pre_get(&pmu_idr, GFP_KERNEL);
 +              if (!err)
 +                      goto free_pdc;
 +
 +              err = idr_get_new_above(&pmu_idr, pmu, PERF_TYPE_MAX, &type);
 +              if (err) {
 +                      ret = err;
 +                      goto free_pdc;
 +              }
 +      }
 +      pmu->type = type;
 +
 +      if (pmu_bus_running) {
 +              ret = pmu_dev_alloc(pmu);
 +              if (ret)
 +                      goto free_idr;
 +      }
 +
 +skip_type:
 +      pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
 +      if (pmu->pmu_cpu_context)
 +              goto got_cpu_context;
 +
 +      pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
 +      if (!pmu->pmu_cpu_context)
 +              goto free_dev;
 +
 +      for_each_possible_cpu(cpu) {
 +              struct perf_cpu_context *cpuctx;
 +
 +              cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
 +              __perf_event_init_context(&cpuctx->ctx);
 +              lockdep_set_class(&cpuctx->ctx.mutex, &cpuctx_mutex);
 +              cpuctx->ctx.type = cpu_context;
 +              cpuctx->ctx.pmu = pmu;
 +              cpuctx->jiffies_interval = 1;
 +              INIT_LIST_HEAD(&cpuctx->rotation_list);
 +              cpuctx->active_pmu = pmu;
 +      }
 +
 +got_cpu_context:
 +      if (!pmu->start_txn) {
 +              if (pmu->pmu_enable) {
 +                      /*
 +                       * If we have pmu_enable/pmu_disable calls, install
 +                       * transaction stubs that use that to try and batch
 +                       * hardware accesses.
 +                       */
 +                      pmu->start_txn  = perf_pmu_start_txn;
 +                      pmu->commit_txn = perf_pmu_commit_txn;
 +                      pmu->cancel_txn = perf_pmu_cancel_txn;
 +              } else {
 +                      pmu->start_txn  = perf_pmu_nop_void;
 +                      pmu->commit_txn = perf_pmu_nop_int;
 +                      pmu->cancel_txn = perf_pmu_nop_void;
 +              }
 +      }
 +
 +      if (!pmu->pmu_enable) {
 +              pmu->pmu_enable  = perf_pmu_nop_void;
 +              pmu->pmu_disable = perf_pmu_nop_void;
 +      }
 +
 +      list_add_rcu(&pmu->entry, &pmus);
 +      ret = 0;
 +unlock:
 +      mutex_unlock(&pmus_lock);
 +
 +      return ret;
 +
 +free_dev:
 +      device_del(pmu->dev);
 +      put_device(pmu->dev);
 +
 +free_idr:
 +      if (pmu->type >= PERF_TYPE_MAX)
 +              idr_remove(&pmu_idr, pmu->type);
 +
 +free_pdc:
 +      free_percpu(pmu->pmu_disable_count);
 +      goto unlock;
 +}
 +
 +void perf_pmu_unregister(struct pmu *pmu)
 +{
 +      mutex_lock(&pmus_lock);
 +      list_del_rcu(&pmu->entry);
 +      mutex_unlock(&pmus_lock);
 +
 +      /*
 +       * We dereference the pmu list under both SRCU and regular RCU, so
 +       * synchronize against both of those.
 +       */
 +      synchronize_srcu(&pmus_srcu);
 +      synchronize_rcu();
 +
 +      free_percpu(pmu->pmu_disable_count);
 +      if (pmu->type >= PERF_TYPE_MAX)
 +              idr_remove(&pmu_idr, pmu->type);
 +      device_del(pmu->dev);
 +      put_device(pmu->dev);
 +      free_pmu_context(pmu);
 +}
 +
 +struct pmu *perf_init_event(struct perf_event *event)
 +{
 +      struct pmu *pmu = NULL;
 +      int idx;
 +      int ret;
 +
 +      idx = srcu_read_lock(&pmus_srcu);
 +
 +      rcu_read_lock();
 +      pmu = idr_find(&pmu_idr, event->attr.type);
 +      rcu_read_unlock();
 +      if (pmu) {
 +              ret = pmu->event_init(event);
 +              if (ret)
 +                      pmu = ERR_PTR(ret);
 +              goto unlock;
 +      }
 +
 +      list_for_each_entry_rcu(pmu, &pmus, entry) {
 +              ret = pmu->event_init(event);
 +              if (!ret)
 +                      goto unlock;
 +
 +              if (ret != -ENOENT) {
 +                      pmu = ERR_PTR(ret);
 +                      goto unlock;
 +              }
 +      }
 +      pmu = ERR_PTR(-ENOENT);
 +unlock:
 +      srcu_read_unlock(&pmus_srcu, idx);
 +
 +      return pmu;
 +}
 +
 +/*
 + * Allocate and initialize a event structure
 + */
 +static struct perf_event *
 +perf_event_alloc(struct perf_event_attr *attr, int cpu,
 +               struct task_struct *task,
 +               struct perf_event *group_leader,
 +               struct perf_event *parent_event,
 +               perf_overflow_handler_t overflow_handler)
 +{
 +      struct pmu *pmu;
 +      struct perf_event *event;
 +      struct hw_perf_event *hwc;
 +      long err;
 +
 +      if ((unsigned)cpu >= nr_cpu_ids) {
 +              if (!task || cpu != -1)
 +                      return ERR_PTR(-EINVAL);
 +      }
 +
 +      event = kzalloc(sizeof(*event), GFP_KERNEL);
 +      if (!event)
 +              return ERR_PTR(-ENOMEM);
 +
 +      /*
 +       * Single events are their own group leaders, with an
 +       * empty sibling list:
 +       */
 +      if (!group_leader)
 +              group_leader = event;
 +
 +      mutex_init(&event->child_mutex);
 +      INIT_LIST_HEAD(&event->child_list);
 +
 +      INIT_LIST_HEAD(&event->group_entry);
 +      INIT_LIST_HEAD(&event->event_entry);
 +      INIT_LIST_HEAD(&event->sibling_list);
 +      init_waitqueue_head(&event->waitq);
 +      init_irq_work(&event->pending, perf_pending_event);
 +
 +      mutex_init(&event->mmap_mutex);
 +
 +      event->cpu              = cpu;
 +      event->attr             = *attr;
 +      event->group_leader     = group_leader;
 +      event->pmu              = NULL;
 +      event->oncpu            = -1;
 +
 +      event->parent           = parent_event;
 +
 +      event->ns               = get_pid_ns(current->nsproxy->pid_ns);
 +      event->id               = atomic64_inc_return(&perf_event_id);
 +
 +      event->state            = PERF_EVENT_STATE_INACTIVE;
 +
 +      if (task) {
 +              event->attach_state = PERF_ATTACH_TASK;
 +#ifdef CONFIG_HAVE_HW_BREAKPOINT
 +              /*
 +               * hw_breakpoint is a bit difficult here..
 +               */
 +              if (attr->type == PERF_TYPE_BREAKPOINT)
 +                      event->hw.bp_target = task;
 +#endif
 +      }
 +
 +      if (!overflow_handler && parent_event)
 +              overflow_handler = parent_event->overflow_handler;
 +
 +      event->overflow_handler = overflow_handler;
 +
 +      if (attr->disabled)
 +              event->state = PERF_EVENT_STATE_OFF;
 +
 +      pmu = NULL;
 +
 +      hwc = &event->hw;
 +      hwc->sample_period = attr->sample_period;
 +      if (attr->freq && attr->sample_freq)
 +              hwc->sample_period = 1;
 +      hwc->last_period = hwc->sample_period;
 +
 +      local64_set(&hwc->period_left, hwc->sample_period);
 +
 +      /*
 +       * we currently do not support PERF_FORMAT_GROUP on inherited events
 +       */
 +      if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
 +              goto done;
 +
 +      pmu = perf_init_event(event);
 +
 +done:
 +      err = 0;
 +      if (!pmu)
 +              err = -EINVAL;
 +      else if (IS_ERR(pmu))
 +              err = PTR_ERR(pmu);
 +
 +      if (err) {
 +              if (event->ns)
 +                      put_pid_ns(event->ns);
 +              kfree(event);
 +              return ERR_PTR(err);
 +      }
 +
 +      event->pmu = pmu;
 +
 +      if (!event->parent) {
 +              if (event->attach_state & PERF_ATTACH_TASK)
 +                      jump_label_inc(&perf_sched_events);
 +              if (event->attr.mmap || event->attr.mmap_data)
 +                      atomic_inc(&nr_mmap_events);
 +              if (event->attr.comm)
 +                      atomic_inc(&nr_comm_events);
 +              if (event->attr.task)
 +                      atomic_inc(&nr_task_events);
 +              if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
 +                      err = get_callchain_buffers();
 +                      if (err) {
 +                              free_event(event);
 +                              return ERR_PTR(err);
 +                      }
 +              }
 +      }
 +
 +      return event;
 +}
 +
 +static int perf_copy_attr(struct perf_event_attr __user *uattr,
 +                        struct perf_event_attr *attr)
 +{
 +      u32 size;
 +      int ret;
 +
 +      if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
 +              return -EFAULT;
 +
 +      /*
 +       * zero the full structure, so that a short copy will be nice.
 +       */
 +      memset(attr, 0, sizeof(*attr));
 +
 +      ret = get_user(size, &uattr->size);
 +      if (ret)
 +              return ret;
 +
 +      if (size > PAGE_SIZE)   /* silly large */
 +              goto err_size;
 +
 +      if (!size)              /* abi compat */
 +              size = PERF_ATTR_SIZE_VER0;
 +
 +      if (size < PERF_ATTR_SIZE_VER0)
 +              goto err_size;
 +
 +      /*
 +       * If we're handed a bigger struct than we know of,
 +       * ensure all the unknown bits are 0 - i.e. new
 +       * user-space does not rely on any kernel feature
 +       * extensions we dont know about yet.
 +       */
 +      if (size > sizeof(*attr)) {
 +              unsigned char __user *addr;
 +              unsigned char __user *end;
 +              unsigned char val;
 +
 +              addr = (void __user *)uattr + sizeof(*attr);
 +              end  = (void __user *)uattr + size;
 +
 +              for (; addr < end; addr++) {
 +                      ret = get_user(val, addr);
 +                      if (ret)
 +                              return ret;
 +                      if (val)
 +                              goto err_size;
 +              }
 +              size = sizeof(*attr);
 +      }
 +
 +      ret = copy_from_user(attr, uattr, size);
 +      if (ret)
 +              return -EFAULT;
 +
 +      /*
 +       * If the type exists, the corresponding creation will verify
 +       * the attr->config.
 +       */
 +      if (attr->type >= PERF_TYPE_MAX)
 +              return -EINVAL;
 +
 +      if (attr->__reserved_1)
 +              return -EINVAL;
 +
 +      if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
 +              return -EINVAL;
 +
 +      if (attr->read_format & ~(PERF_FORMAT_MAX-1))
 +              return -EINVAL;
 +
 +out:
 +      return ret;
 +
 +err_size:
 +      put_user(sizeof(*attr), &uattr->size);
 +      ret = -E2BIG;
 +      goto out;
 +}
 +
 +static int
 +perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
 +{
 +      struct perf_buffer *buffer = NULL, *old_buffer = NULL;
 +      int ret = -EINVAL;
 +
 +      if (!output_event)
 +              goto set;
 +
 +      /* don't allow circular references */
 +      if (event == output_event)
 +              goto out;
 +
 +      /*
 +       * Don't allow cross-cpu buffers
 +       */
 +      if (output_event->cpu != event->cpu)
 +              goto out;
 +
 +      /*
 +       * If its not a per-cpu buffer, it must be the same task.
 +       */
 +      if (output_event->cpu == -1 && output_event->ctx != event->ctx)
 +              goto out;
 +
 +set:
 +      mutex_lock(&event->mmap_mutex);
 +      /* Can't redirect output if we've got an active mmap() */
 +      if (atomic_read(&event->mmap_count))
 +              goto unlock;
 +
 +      if (output_event) {
 +              /* get the buffer we want to redirect to */
 +              buffer = perf_buffer_get(output_event);
 +              if (!buffer)
 +                      goto unlock;
 +      }
 +
 +      old_buffer = event->buffer;
 +      rcu_assign_pointer(event->buffer, buffer);
 +      ret = 0;
 +unlock:
 +      mutex_unlock(&event->mmap_mutex);
 +
 +      if (old_buffer)
 +              perf_buffer_put(old_buffer);
 +out:
 +      return ret;
 +}
 +
 +/**
 + * sys_perf_event_open - open a performance event, associate it to a task/cpu
 + *
 + * @attr_uptr:        event_id type attributes for monitoring/sampling
 + * @pid:              target pid
 + * @cpu:              target cpu
 + * @group_fd:         group leader event fd
 + */
 +SYSCALL_DEFINE5(perf_event_open,
 +              struct perf_event_attr __user *, attr_uptr,
 +              pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
 +{
 +      struct perf_event *group_leader = NULL, *output_event = NULL;
 +      struct perf_event *event, *sibling;
 +      struct perf_event_attr attr;
 +      struct perf_event_context *ctx;
 +      struct file *event_file = NULL;
 +      struct file *group_file = NULL;
 +      struct task_struct *task = NULL;
 +      struct pmu *pmu;
 +      int event_fd;
 +      int move_group = 0;
 +      int fput_needed = 0;
 +      int err;
 +
 +      /* for future expandability... */
 +      if (flags & ~PERF_FLAG_ALL)
 +              return -EINVAL;
 +
 +      err = perf_copy_attr(attr_uptr, &attr);
 +      if (err)
 +              return err;
 +
 +      if (!attr.exclude_kernel) {
 +              if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
 +                      return -EACCES;
 +      }
 +
 +      if (attr.freq) {
 +              if (attr.sample_freq > sysctl_perf_event_sample_rate)
 +                      return -EINVAL;
 +      }
 +
 +      /*
 +       * In cgroup mode, the pid argument is used to pass the fd
 +       * opened to the cgroup directory in cgroupfs. The cpu argument
 +       * designates the cpu on which to monitor threads from that
 +       * cgroup.
 +       */
 +      if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
 +              return -EINVAL;
 +
 +      event_fd = get_unused_fd_flags(O_RDWR);
 +      if (event_fd < 0)
 +              return event_fd;
 +
 +      if (group_fd != -1) {
 +              group_leader = perf_fget_light(group_fd, &fput_needed);
 +              if (IS_ERR(group_leader)) {
 +                      err = PTR_ERR(group_leader);
 +                      goto err_fd;
 +              }
 +              group_file = group_leader->filp;
 +              if (flags & PERF_FLAG_FD_OUTPUT)
 +                      output_event = group_leader;
 +              if (flags & PERF_FLAG_FD_NO_GROUP)
 +                      group_leader = NULL;
 +      }
 +
 +      if (pid != -1 && !(flags & PERF_FLAG_PID_CGROUP)) {
 +              task = find_lively_task_by_vpid(pid);
 +              if (IS_ERR(task)) {
 +                      err = PTR_ERR(task);
 +                      goto err_group_fd;
 +              }
 +      }
 +
 +      event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
 +      if (IS_ERR(event)) {
 +              err = PTR_ERR(event);
 +              goto err_task;
 +      }
 +
 +      if (flags & PERF_FLAG_PID_CGROUP) {
 +              err = perf_cgroup_connect(pid, event, &attr, group_leader);
 +              if (err)
 +                      goto err_alloc;
 +              /*
 +               * one more event:
 +               * - that has cgroup constraint on event->cpu
 +               * - that may need work on context switch
 +               */
 +              atomic_inc(&per_cpu(perf_cgroup_events, event->cpu));
 +              jump_label_inc(&perf_sched_events);
 +      }
 +
 +      /*
 +       * Special case software events and allow them to be part of
 +       * any hardware group.
 +       */
 +      pmu = event->pmu;
 +
 +      if (group_leader &&
 +          (is_software_event(event) != is_software_event(group_leader))) {
 +              if (is_software_event(event)) {
 +                      /*
 +                       * If event and group_leader are not both a software
 +                       * event, and event is, then group leader is not.
 +                       *
 +                       * Allow the addition of software events to !software
 +                       * groups, this is safe because software events never
 +                       * fail to schedule.
 +                       */
 +                      pmu = group_leader->pmu;
 +              } else if (is_software_event(group_leader) &&
 +                         (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
 +                      /*
 +                       * In case the group is a pure software group, and we
 +                       * try to add a hardware event, move the whole group to
 +                       * the hardware context.
 +                       */
 +                      move_group = 1;
 +              }
 +      }
 +
 +      /*
 +       * Get the target context (task or percpu):
 +       */
 +      ctx = find_get_context(pmu, task, cpu);
 +      if (IS_ERR(ctx)) {
 +              err = PTR_ERR(ctx);
 +              goto err_alloc;
 +      }
 +
 +      if (task) {
 +              put_task_struct(task);
 +              task = NULL;
 +      }
 +
 +      /*
 +       * Look up the group leader (we will attach this event to it):
 +       */
 +      if (group_leader) {
 +              err = -EINVAL;
 +
 +              /*
 +               * Do not allow a recursive hierarchy (this new sibling
 +               * becoming part of another group-sibling):
 +               */
 +              if (group_leader->group_leader != group_leader)
 +                      goto err_context;
 +              /*
 +               * Do not allow to attach to a group in a different
 +               * task or CPU context:
 +               */
 +              if (move_group) {
 +                      if (group_leader->ctx->type != ctx->type)
 +                              goto err_context;
 +              } else {
 +                      if (group_leader->ctx != ctx)
 +                              goto err_context;
 +              }
 +
 +              /*
 +               * Only a group leader can be exclusive or pinned
 +               */
 +              if (attr.exclusive || attr.pinned)
 +                      goto err_context;
 +      }
 +
 +      if (output_event) {
 +              err = perf_event_set_output(event, output_event);
 +              if (err)
 +                      goto err_context;
 +      }
 +
 +      event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
 +      if (IS_ERR(event_file)) {
 +              err = PTR_ERR(event_file);
 +              goto err_context;
 +      }
 +
 +      if (move_group) {
 +              struct perf_event_context *gctx = group_leader->ctx;
 +
 +              mutex_lock(&gctx->mutex);
 +              perf_remove_from_context(group_leader);
 +              list_for_each_entry(sibling, &group_leader->sibling_list,
 +                                  group_entry) {
 +                      perf_remove_from_context(sibling);
 +                      put_ctx(gctx);
 +              }
 +              mutex_unlock(&gctx->mutex);
 +              put_ctx(gctx);
 +      }
 +
 +      event->filp = event_file;
 +      WARN_ON_ONCE(ctx->parent_ctx);
 +      mutex_lock(&ctx->mutex);
 +
 +      if (move_group) {
 +              perf_install_in_context(ctx, group_leader, cpu);
 +              get_ctx(ctx);
 +              list_for_each_entry(sibling, &group_leader->sibling_list,
 +                                  group_entry) {
 +                      perf_install_in_context(ctx, sibling, cpu);
 +                      get_ctx(ctx);
 +              }
 +      }
 +
 +      perf_install_in_context(ctx, event, cpu);
 +      ++ctx->generation;
 +      perf_unpin_context(ctx);
 +      mutex_unlock(&ctx->mutex);
 +
 +      event->owner = current;
 +
 +      mutex_lock(&current->perf_event_mutex);
 +      list_add_tail(&event->owner_entry, &current->perf_event_list);
 +      mutex_unlock(&current->perf_event_mutex);
 +
 +      /*
 +       * Precalculate sample_data sizes
 +       */
 +      perf_event__header_size(event);
 +      perf_event__id_header_size(event);
 +
 +      /*
 +       * Drop the reference on the group_event after placing the
 +       * new event on the sibling_list. This ensures destruction
 +       * of the group leader will find the pointer to itself in
 +       * perf_group_detach().
 +       */
 +      fput_light(group_file, fput_needed);
 +      fd_install(event_fd, event_file);
 +      return event_fd;
 +
 +err_context:
 +      perf_unpin_context(ctx);
 +      put_ctx(ctx);
 +err_alloc:
 +      free_event(event);
 +err_task:
 +      if (task)
 +              put_task_struct(task);
 +err_group_fd:
 +      fput_light(group_file, fput_needed);
 +err_fd:
 +      put_unused_fd(event_fd);
 +      return err;
 +}
 +
 +/**
 + * perf_event_create_kernel_counter
 + *
 + * @attr: attributes of the counter to create
 + * @cpu: cpu in which the counter is bound
 + * @task: task to profile (NULL for percpu)
 + */
 +struct perf_event *
 +perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
 +                               struct task_struct *task,
 +                               perf_overflow_handler_t overflow_handler)
 +{
 +      struct perf_event_context *ctx;
 +      struct perf_event *event;
 +      int err;
 +
 +      /*
 +       * Get the target context (task or percpu):
 +       */
 +
 +      event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
 +      if (IS_ERR(event)) {
 +              err = PTR_ERR(event);
 +              goto err;
 +      }
 +
 +      ctx = find_get_context(event->pmu, task, cpu);
 +      if (IS_ERR(ctx)) {
 +              err = PTR_ERR(ctx);
 +              goto err_free;
 +      }
 +
 +      event->filp = NULL;
 +      WARN_ON_ONCE(ctx->parent_ctx);
 +      mutex_lock(&ctx->mutex);
 +      perf_install_in_context(ctx, event, cpu);
 +      ++ctx->generation;
 +      perf_unpin_context(ctx);
 +      mutex_unlock(&ctx->mutex);
 +
 +      return event;
 +
 +err_free:
 +      free_event(event);
 +err:
 +      return ERR_PTR(err);
 +}
 +EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
 +
 +static void sync_child_event(struct perf_event *child_event,
 +                             struct task_struct *child)
 +{
 +      struct perf_event *parent_event = child_event->parent;
 +      u64 child_val;
 +
 +      if (child_event->attr.inherit_stat)
 +              perf_event_read_event(child_event, child);
 +
 +      child_val = perf_event_count(child_event);
 +
 +      /*
 +       * Add back the child's count to the parent's count:
 +       */
 +      atomic64_add(child_val, &parent_event->child_count);
 +      atomic64_add(child_event->total_time_enabled,
 +                   &parent_event->child_total_time_enabled);
 +      atomic64_add(child_event->total_time_running,
 +                   &parent_event->child_total_time_running);
 +
 +      /*
 +       * Remove this event from the parent's list
 +       */
 +      WARN_ON_ONCE(parent_event->ctx->parent_ctx);
 +      mutex_lock(&parent_event->child_mutex);
 +      list_del_init(&child_event->child_list);
 +      mutex_unlock(&parent_event->child_mutex);
 +
 +      /*
 +       * Release the parent event, if this was the last
 +       * reference to it.
 +       */
 +      fput(parent_event->filp);
 +}
 +
 +static void
 +__perf_event_exit_task(struct perf_event *child_event,
 +                       struct perf_event_context *child_ctx,
 +                       struct task_struct *child)
 +{
 +      if (child_event->parent) {
 +              raw_spin_lock_irq(&child_ctx->lock);
 +              perf_group_detach(child_event);
 +              raw_spin_unlock_irq(&child_ctx->lock);
 +      }
 +
 +      perf_remove_from_context(child_event);
 +
 +      /*
 +       * It can happen that the parent exits first, and has events
 +       * that are still around due to the child reference. These
 +       * events need to be zapped.
 +       */
 +      if (child_event->parent) {
 +              sync_child_event(child_event, child);
 +              free_event(child_event);
 +      }
 +}
 +
 +static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 +{
 +      struct perf_event *child_event, *tmp;
 +      struct perf_event_context *child_ctx;
 +      unsigned long flags;
 +
 +      if (likely(!child->perf_event_ctxp[ctxn])) {
 +              perf_event_task(child, NULL, 0);
 +              return;
 +      }
 +
 +      local_irq_save(flags);
 +      /*
 +       * We can't reschedule here because interrupts are disabled,
 +       * and either child is current or it is a task that can't be
 +       * scheduled, so we are now safe from rescheduling changing
 +       * our context.
 +       */
 +      child_ctx = rcu_dereference_raw(child->perf_event_ctxp[ctxn]);
 +      task_ctx_sched_out(child_ctx, EVENT_ALL);
 +
 +      /*
 +       * Take the context lock here so that if find_get_context is
 +       * reading child->perf_event_ctxp, we wait until it has
 +       * incremented the context's refcount before we do put_ctx below.
 +       */
 +      raw_spin_lock(&child_ctx->lock);
 +      child->perf_event_ctxp[ctxn] = NULL;
 +      /*
 +       * If this context is a clone; unclone it so it can't get
 +       * swapped to another process while we're removing all
 +       * the events from it.
 +       */
 +      unclone_ctx(child_ctx);
 +      update_context_time(child_ctx);
 +      raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
 +
 +      /*
 +       * Report the task dead after unscheduling the events so that we
 +       * won't get any samples after PERF_RECORD_EXIT. We can however still
 +       * get a few PERF_RECORD_READ events.
 +       */
 +      perf_event_task(child, child_ctx, 0);
 +
 +      /*
 +       * We can recurse on the same lock type through:
 +       *
 +       *   __perf_event_exit_task()
 +       *     sync_child_event()
 +       *       fput(parent_event->filp)
 +       *         perf_release()
 +       *           mutex_lock(&ctx->mutex)
 +       *
 +       * But since its the parent context it won't be the same instance.
 +       */
 +      mutex_lock(&child_ctx->mutex);
 +
 +again:
 +      list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
 +                               group_entry)
 +              __perf_event_exit_task(child_event, child_ctx, child);
 +
 +      list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
 +                               group_entry)
 +              __perf_event_exit_task(child_event, child_ctx, child);
 +
 +      /*
 +       * If the last event was a group event, it will have appended all
 +       * its siblings to the list, but we obtained 'tmp' before that which
 +       * will still point to the list head terminating the iteration.
 +       */
 +      if (!list_empty(&child_ctx->pinned_groups) ||
 +          !list_empty(&child_ctx->flexible_groups))
 +              goto again;
 +
 +      mutex_unlock(&child_ctx->mutex);
 +
 +      put_ctx(child_ctx);
 +}
 +
 +/*
 + * When a child task exits, feed back event values to parent events.
 + */
 +void perf_event_exit_task(struct task_struct *child)
 +{
 +      struct perf_event *event, *tmp;
 +      int ctxn;
 +
 +      mutex_lock(&child->perf_event_mutex);
 +      list_for_each_entry_safe(event, tmp, &child->perf_event_list,
 +                               owner_entry) {
 +              list_del_init(&event->owner_entry);
 +
 +              /*
 +               * Ensure the list deletion is visible before we clear
 +               * the owner, closes a race against perf_release() where
 +               * we need to serialize on the owner->perf_event_mutex.
 +               */
 +              smp_wmb();
 +              event->owner = NULL;
 +      }
 +      mutex_unlock(&child->perf_event_mutex);
 +
 +      for_each_task_context_nr(ctxn)
 +              perf_event_exit_task_context(child, ctxn);
 +}
 +
 +static void perf_free_event(struct perf_event *event,
 +                          struct perf_event_context *ctx)
 +{
 +      struct perf_event *parent = event->parent;
 +
 +      if (WARN_ON_ONCE(!parent))
 +              return;
 +
 +      mutex_lock(&parent->child_mutex);
 +      list_del_init(&event->child_list);
 +      mutex_unlock(&parent->child_mutex);
 +
 +      fput(parent->filp);
 +
 +      perf_group_detach(event);
 +      list_del_event(event, ctx);
 +      free_event(event);
 +}
 +
 +/*
 + * free an unexposed, unused context as created by inheritance by
 + * perf_event_init_task below, used by fork() in case of fail.
 + */
 +void perf_event_free_task(struct task_struct *task)
 +{
 +      struct perf_event_context *ctx;
 +      struct perf_event *event, *tmp;
 +      int ctxn;
 +
 +      for_each_task_context_nr(ctxn) {
 +              ctx = task->perf_event_ctxp[ctxn];
 +              if (!ctx)
 +                      continue;
 +
 +              mutex_lock(&ctx->mutex);
 +again:
 +              list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
 +                              group_entry)
 +                      perf_free_event(event, ctx);
 +
 +              list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
 +                              group_entry)
 +                      perf_free_event(event, ctx);
 +
 +              if (!list_empty(&ctx->pinned_groups) ||
 +                              !list_empty(&ctx->flexible_groups))
 +                      goto again;
 +
 +              mutex_unlock(&ctx->mutex);
 +
 +              put_ctx(ctx);
 +      }
 +}
 +
 +void perf_event_delayed_put(struct task_struct *task)
 +{
 +      int ctxn;
 +
 +      for_each_task_context_nr(ctxn)
 +              WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
 +}
 +
 +/*
 + * inherit a event from parent task to child task:
 + */
 +static struct perf_event *
 +inherit_event(struct perf_event *parent_event,
 +            struct task_struct *parent,
 +            struct perf_event_context *parent_ctx,
 +            struct task_struct *child,
 +            struct perf_event *group_leader,
 +            struct perf_event_context *child_ctx)
 +{
 +      struct perf_event *child_event;
 +      unsigned long flags;
 +
 +      /*
 +       * Instead of creating recursive hierarchies of events,
 +       * we link inherited events back to the original parent,
 +       * which has a filp for sure, which we use as the reference
 +       * count:
 +       */
 +      if (parent_event->parent)
 +              parent_event = parent_event->parent;
 +
 +      child_event = perf_event_alloc(&parent_event->attr,
 +                                         parent_event->cpu,
 +                                         child,
 +                                         group_leader, parent_event,
 +                                         NULL);
 +      if (IS_ERR(child_event))
 +              return child_event;
 +      get_ctx(child_ctx);
 +
 +      /*
 +       * Make the child state follow the state of the parent event,
 +       * not its attr.disabled bit.  We hold the parent's mutex,
 +       * so we won't race with perf_event_{en, dis}able_family.
 +       */
 +      if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
 +              child_event->state = PERF_EVENT_STATE_INACTIVE;
 +      else
 +              child_event->state = PERF_EVENT_STATE_OFF;
 +
 +      if (parent_event->attr.freq) {
 +              u64 sample_period = parent_event->hw.sample_period;
 +              struct hw_perf_event *hwc = &child_event->hw;
 +
 +              hwc->sample_period = sample_period;
 +              hwc->last_period   = sample_period;
 +
 +              local64_set(&hwc->period_left, sample_period);
 +      }
 +
 +      child_event->ctx = child_ctx;
 +      child_event->overflow_handler = parent_event->overflow_handler;
 +
 +      /*
 +       * Precalculate sample_data sizes
 +       */
 +      perf_event__header_size(child_event);
 +      perf_event__id_header_size(child_event);
 +
 +      /*
 +       * Link it up in the child's context:
 +       */
 +      raw_spin_lock_irqsave(&child_ctx->lock, flags);
 +      add_event_to_ctx(child_event, child_ctx);
 +      raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
 +
 +      /*
 +       * Get a reference to the parent filp - we will fput it
 +       * when the child event exits. This is safe to do because
 +       * we are in the parent and we know that the filp still
 +       * exists and has a nonzero count:
 +       */
 +      atomic_long_inc(&parent_event->filp->f_count);
 +
 +      /*
 +       * Link this into the parent event's child list
 +       */
 +      WARN_ON_ONCE(parent_event->ctx->parent_ctx);
 +      mutex_lock(&parent_event->child_mutex);
 +      list_add_tail(&child_event->child_list, &parent_event->child_list);
 +      mutex_unlock(&parent_event->child_mutex);
 +
 +      return child_event;
 +}
 +
 +static int inherit_group(struct perf_event *parent_event,
 +            struct task_struct *parent,
 +            struct perf_event_context *parent_ctx,
 +            struct task_struct *child,
 +            struct perf_event_context *child_ctx)
 +{
 +      struct perf_event *leader;
 +      struct perf_event *sub;
 +      struct perf_event *child_ctr;
 +
 +      leader = inherit_event(parent_event, parent, parent_ctx,
 +                               child, NULL, child_ctx);
 +      if (IS_ERR(leader))
 +              return PTR_ERR(leader);
 +      list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
 +              child_ctr = inherit_event(sub, parent, parent_ctx,
 +                                          child, leader, child_ctx);
 +              if (IS_ERR(child_ctr))
 +                      return PTR_ERR(child_ctr);
 +      }
 +      return 0;
 +}
 +
 +static int
 +inherit_task_group(struct perf_event *event, struct task_struct *parent,
 +                 struct perf_event_context *parent_ctx,
 +                 struct task_struct *child, int ctxn,
 +                 int *inherited_all)
 +{
 +      int ret;
 +      struct perf_event_context *child_ctx;
 +
 +      if (!event->attr.inherit) {
 +              *inherited_all = 0;
 +              return 0;
 +      }
 +
 +      child_ctx = child->perf_event_ctxp[ctxn];
 +      if (!child_ctx) {
 +              /*
 +               * This is executed from the parent task context, so
 +               * inherit events that have been marked for cloning.
 +               * First allocate and initialize a context for the
 +               * child.
 +               */
 +
 +              child_ctx = alloc_perf_context(event->pmu, child);
 +              if (!child_ctx)
 +                      return -ENOMEM;
 +
 +              child->perf_event_ctxp[ctxn] = child_ctx;
 +      }
 +
 +      ret = inherit_group(event, parent, parent_ctx,
 +                          child, child_ctx);
 +
 +      if (ret)
 +              *inherited_all = 0;
 +
 +      return ret;
 +}
 +
 +/*
 + * Initialize the perf_event context in task_struct
 + */
 +int perf_event_init_context(struct task_struct *child, int ctxn)
 +{
 +      struct perf_event_context *child_ctx, *parent_ctx;
 +      struct perf_event_context *cloned_ctx;
 +      struct perf_event *event;
 +      struct task_struct *parent = current;
 +      int inherited_all = 1;
 +      unsigned long flags;
 +      int ret = 0;
 +
 +      if (likely(!parent->perf_event_ctxp[ctxn]))
 +              return 0;
 +
 +      /*
 +       * If the parent's context is a clone, pin it so it won't get
 +       * swapped under us.
 +       */
 +      parent_ctx = perf_pin_task_context(parent, ctxn);
 +
 +      /*
 +       * No need to check if parent_ctx != NULL here; since we saw
 +       * it non-NULL earlier, the only reason for it to become NULL
 +       * is if we exit, and since we're currently in the middle of
 +       * a fork we can't be exiting at the same time.
 +       */
 +
 +      /*
 +       * Lock the parent list. No need to lock the child - not PID
 +       * hashed yet and not running, so nobody can access it.
 +       */
 +      mutex_lock(&parent_ctx->mutex);
 +
 +      /*
 +       * We dont have to disable NMIs - we are only looking at
 +       * the list, not manipulating it:
 +       */
 +      list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
 +              ret = inherit_task_group(event, parent, parent_ctx,
 +                                       child, ctxn, &inherited_all);
 +              if (ret)
 +                      break;
 +      }
 +
 +      /*
 +       * We can't hold ctx->lock when iterating the ->flexible_group list due
 +       * to allocations, but we need to prevent rotation because
 +       * rotate_ctx() will change the list from interrupt context.
 +       */
 +      raw_spin_lock_irqsave(&parent_ctx->lock, flags);
 +      parent_ctx->rotate_disable = 1;
 +      raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
 +
 +      list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
 +              ret = inherit_task_group(event, parent, parent_ctx,
 +                                       child, ctxn, &inherited_all);
 +              if (ret)
 +                      break;
 +      }
 +
 +      raw_spin_lock_irqsave(&parent_ctx->lock, flags);
 +      parent_ctx->rotate_disable = 0;
 +
 +      child_ctx = child->perf_event_ctxp[ctxn];
 +
 +      if (child_ctx && inherited_all) {
 +              /*
 +               * Mark the child context as a clone of the parent
 +               * context, or of whatever the parent is a clone of.
 +               *
 +               * Note that if the parent is a clone, the holding of
 +               * parent_ctx->lock avoids it from being uncloned.
 +               */
 +              cloned_ctx = parent_ctx->parent_ctx;
 +              if (cloned_ctx) {
 +                      child_ctx->parent_ctx = cloned_ctx;
 +                      child_ctx->parent_gen = parent_ctx->parent_gen;
 +              } else {
 +                      child_ctx->parent_ctx = parent_ctx;
 +                      child_ctx->parent_gen = parent_ctx->generation;
 +              }
 +              get_ctx(child_ctx->parent_ctx);
 +      }
 +
 +      raw_spin_unlock_irqrestore(&parent_ctx->lock, flags);
 +      mutex_unlock(&parent_ctx->mutex);
 +
 +      perf_unpin_context(parent_ctx);
 +      put_ctx(parent_ctx);
 +
 +      return ret;
 +}
 +
 +/*
 + * Initialize the perf_event context in task_struct
 + */
 +int perf_event_init_task(struct task_struct *child)
 +{
 +      int ctxn, ret;
 +
 +      memset(child->perf_event_ctxp, 0, sizeof(child->perf_event_ctxp));
 +      mutex_init(&child->perf_event_mutex);
 +      INIT_LIST_HEAD(&child->perf_event_list);
 +
 +      for_each_task_context_nr(ctxn) {
 +              ret = perf_event_init_context(child, ctxn);
 +              if (ret)
 +                      return ret;
 +      }
 +
 +      return 0;
 +}
 +
 +static void __init perf_event_init_all_cpus(void)
 +{
 +      struct swevent_htable *swhash;
 +      int cpu;
 +
 +      for_each_possible_cpu(cpu) {
 +              swhash = &per_cpu(swevent_htable, cpu);
 +              mutex_init(&swhash->hlist_mutex);
 +              INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
 +      }
 +}
 +
 +static void __cpuinit perf_event_init_cpu(int cpu)
 +{
 +      struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
 +
 +      mutex_lock(&swhash->hlist_mutex);
 +      if (swhash->hlist_refcount > 0) {
 +              struct swevent_hlist *hlist;
 +
 +              hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
 +              WARN_ON(!hlist);
 +              rcu_assign_pointer(swhash->swevent_hlist, hlist);
 +      }
 +      mutex_unlock(&swhash->hlist_mutex);
 +}
 +
 +#if defined CONFIG_HOTPLUG_CPU || defined CONFIG_KEXEC
 +static void perf_pmu_rotate_stop(struct pmu *pmu)
 +{
 +      struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
 +
 +      WARN_ON(!irqs_disabled());
 +
 +      list_del_init(&cpuctx->rotation_list);
 +}
 +
 +static void __perf_event_exit_context(void *__info)
 +{
 +      struct perf_event_context *ctx = __info;
 +      struct perf_event *event, *tmp;
 +
 +      perf_pmu_rotate_stop(ctx->pmu);
 +
 +      list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
 +              __perf_remove_from_context(event);
 +      list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
 +              __perf_remove_from_context(event);
 +}
 +
 +static void perf_event_exit_cpu_context(int cpu)
 +{
 +      struct perf_event_context *ctx;
 +      struct pmu *pmu;
 +      int idx;
 +
 +      idx = srcu_read_lock(&pmus_srcu);
 +      list_for_each_entry_rcu(pmu, &pmus, entry) {
 +              ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
 +
 +              mutex_lock(&ctx->mutex);
 +              smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
 +              mutex_unlock(&ctx->mutex);
 +      }
 +      srcu_read_unlock(&pmus_srcu, idx);
 +}
 +
 +static void perf_event_exit_cpu(int cpu)
 +{
 +      struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
 +
 +      mutex_lock(&swhash->hlist_mutex);
 +      swevent_hlist_release(swhash);
 +      mutex_unlock(&swhash->hlist_mutex);
 +
 +      perf_event_exit_cpu_context(cpu);
 +}
 +#else
 +static inline void perf_event_exit_cpu(int cpu) { }
 +#endif
 +
 +static int
 +perf_reboot(struct notifier_block *notifier, unsigned long val, void *v)
 +{
 +      int cpu;
 +
 +      for_each_online_cpu(cpu)
 +              perf_event_exit_cpu(cpu);
 +
 +      return NOTIFY_OK;
 +}
 +
 +/*
 + * Run the perf reboot notifier at the very last possible moment so that
 + * the generic watchdog code runs as long as possible.
 + */
 +static struct notifier_block perf_reboot_notifier = {
 +      .notifier_call = perf_reboot,
 +      .priority = INT_MIN,
 +};
 +
 +static int __cpuinit
 +perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
 +{
 +      unsigned int cpu = (long)hcpu;
 +
 +      switch (action & ~CPU_TASKS_FROZEN) {
 +
 +      case CPU_UP_PREPARE:
 +      case CPU_DOWN_FAILED:
 +              perf_event_init_cpu(cpu);
 +              break;
 +
 +      case CPU_UP_CANCELED:
 +      case CPU_DOWN_PREPARE:
 +              perf_event_exit_cpu(cpu);
 +              break;
 +
 +      default:
 +              break;
 +      }
 +
 +      return NOTIFY_OK;
 +}
 +
 +void __init perf_event_init(void)
 +{
 +      int ret;
 +
 +      idr_init(&pmu_idr);
 +
 +      perf_event_init_all_cpus();
 +      init_srcu_struct(&pmus_srcu);
 +      perf_pmu_register(&perf_swevent, "software", PERF_TYPE_SOFTWARE);
 +      perf_pmu_register(&perf_cpu_clock, NULL, -1);
 +      perf_pmu_register(&perf_task_clock, NULL, -1);
 +      perf_tp_register();
 +      perf_cpu_notifier(perf_cpu_notify);
 +      register_reboot_notifier(&perf_reboot_notifier);
 +
 +      ret = init_hw_breakpoint();
 +      WARN(ret, "hw_breakpoint initialization failed with: %d", ret);
 +}
 +
 +static int __init perf_event_sysfs_init(void)
 +{
 +      struct pmu *pmu;
 +      int ret;
 +
 +      mutex_lock(&pmus_lock);
 +
 +      ret = bus_register(&pmu_bus);
 +      if (ret)
 +              goto unlock;
 +
 +      list_for_each_entry(pmu, &pmus, entry) {
 +              if (!pmu->name || pmu->type < 0)
 +                      continue;
 +
 +              ret = pmu_dev_alloc(pmu);
 +              WARN(ret, "Failed to register pmu: %s, reason %d\n", pmu->name, ret);
 +      }
 +      pmu_bus_running = 1;
 +      ret = 0;
 +
 +unlock:
 +      mutex_unlock(&pmus_lock);
 +
 +      return ret;
 +}
 +device_initcall(perf_event_sysfs_init);
 +
 +#ifdef CONFIG_CGROUP_PERF
 +static struct cgroup_subsys_state *perf_cgroup_create(
 +      struct cgroup_subsys *ss, struct cgroup *cont)
 +{
 +      struct perf_cgroup *jc;
 +
 +      jc = kzalloc(sizeof(*jc), GFP_KERNEL);
 +      if (!jc)
 +              return ERR_PTR(-ENOMEM);
 +
 +      jc->info = alloc_percpu(struct perf_cgroup_info);
 +      if (!jc->info) {
 +              kfree(jc);
 +              return ERR_PTR(-ENOMEM);
 +      }
 +
 +      return &jc->css;
 +}
 +
 +static void perf_cgroup_destroy(struct cgroup_subsys *ss,
 +                              struct cgroup *cont)
 +{
 +      struct perf_cgroup *jc;
 +      jc = container_of(cgroup_subsys_state(cont, perf_subsys_id),
 +                        struct perf_cgroup, css);
 +      free_percpu(jc->info);
 +      kfree(jc);
 +}
 +
 +static int __perf_cgroup_move(void *info)
 +{
 +      struct task_struct *task = info;
 +      perf_cgroup_switch(task, PERF_CGROUP_SWOUT | PERF_CGROUP_SWIN);
 +      return 0;
 +}
 +
 +static void perf_cgroup_move(struct task_struct *task)
 +{
 +      task_function_call(task, __perf_cgroup_move, task);
 +}
 +
 +static void perf_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
 +              struct cgroup *old_cgrp, struct task_struct *task,
 +              bool threadgroup)
 +{
 +      perf_cgroup_move(task);
 +      if (threadgroup) {
 +              struct task_struct *c;
 +              rcu_read_lock();
 +              list_for_each_entry_rcu(c, &task->thread_group, thread_group) {
 +                      perf_cgroup_move(c);
 +              }
 +              rcu_read_unlock();
 +      }
 +}
 +
 +static void perf_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
 +              struct cgroup *old_cgrp, struct task_struct *task)
 +{
 +      /*
 +       * cgroup_exit() is called in the copy_process() failure path.
 +       * Ignore this case since the task hasn't ran yet, this avoids
 +       * trying to poke a half freed task state from generic code.
 +       */
 +      if (!(task->flags & PF_EXITING))
 +              return;
 +
 +      perf_cgroup_move(task);
 +}
 +
 +struct cgroup_subsys perf_subsys = {
 +      .name           = "perf_event",
 +      .subsys_id      = perf_subsys_id,
 +      .create         = perf_cgroup_create,
 +      .destroy        = perf_cgroup_destroy,
 +      .exit           = perf_cgroup_exit,
 +      .attach         = perf_cgroup_attach,
 +};
 +#endif /* CONFIG_CGROUP_PERF */
Simple merge