kernel/events/callchain.c

   1 /*
   2  * Performance events callchain code, extracted from core.c:
   3  *
   4  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   5  *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
   6  *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
   7  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   8  *
   9  * For licensing details see kernel-base/COPYING
  10  */
  11
  12 #include <linux/perf_event.h>
  13 #include <linux/slab.h>
  14 #include "internal.h"
  15
  16 struct callchain_cpus_entries {
  17         struct rcu_head                 rcu_head;
  18         struct perf_callchain_entry     *cpu_entries[0];
  19 };
  20
  21 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
  22 static atomic_t nr_callchain_events;
  23 static DEFINE_MUTEX(callchain_mutex);
  24 static struct callchain_cpus_entries *callchain_cpus_entries;
  25
  26
  27 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
  28                                   struct pt_regs *regs)
  29 {
  30 }
  31
  32 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
  33                                 struct pt_regs *regs)
  34 {
  35 }
  36
  37 static void release_callchain_buffers_rcu(struct rcu_head *head)
  38 {
  39         struct callchain_cpus_entries *entries;
  40         int cpu;
  41
  42         entries = container_of(head, struct callchain_cpus_entries, rcu_head);
  43
  44         for_each_possible_cpu(cpu)
  45                 kfree(entries->cpu_entries[cpu]);
  46
  47         kfree(entries);
  48 }
  49
  50 static void release_callchain_buffers(void)
  51 {
  52         struct callchain_cpus_entries *entries;
  53
  54         entries = callchain_cpus_entries;
  55         rcu_assign_pointer(callchain_cpus_entries, NULL);
  56         call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
  57 }
  58
  59 static int alloc_callchain_buffers(void)
  60 {
  61         int cpu;
  62         int size;
  63         struct callchain_cpus_entries *entries;
  64
  65         /*
  66          * We can't use the percpu allocation API for data that can be
  67          * accessed from NMI. Use a temporary manual per cpu allocation
  68          * until that gets sorted out.
  69          */
  70         size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
  71
  72         entries = kzalloc(size, GFP_KERNEL);
  73         if (!entries)
  74                 return -ENOMEM;
  75
  76         size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
  77
  78         for_each_possible_cpu(cpu) {
  79                 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
  80                                                          cpu_to_node(cpu));
  81                 if (!entries->cpu_entries[cpu])
  82                         goto fail;
  83         }
  84
  85         rcu_assign_pointer(callchain_cpus_entries, entries);
  86
  87         return 0;
  88
  89 fail:
  90         for_each_possible_cpu(cpu)
  91                 kfree(entries->cpu_entries[cpu]);
  92         kfree(entries);
  93
  94         return -ENOMEM;
  95 }
  96
  97 int get_callchain_buffers(void)
  98 {
  99         int err = 0;
 100         int count;
 101
 102         mutex_lock(&callchain_mutex);
 103
 104         count = atomic_inc_return(&nr_callchain_events);
 105         if (WARN_ON_ONCE(count < 1)) {
 106                 err = -EINVAL;
 107                 goto exit;
 108         }
 109
 110         if (count > 1) {
 111                 /* If the allocation failed, give up */
 112                 if (!callchain_cpus_entries)
 113                         err = -ENOMEM;
 114                 goto exit;
 115         }
 116
 117         err = alloc_callchain_buffers();
 118         if (err)
 119                 release_callchain_buffers();
 120 exit:
 121         mutex_unlock(&callchain_mutex);
 122
 123         return err;
 124 }
 125
 126 void put_callchain_buffers(void)
 127 {
 128         if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
 129                 release_callchain_buffers();
 130                 mutex_unlock(&callchain_mutex);
 131         }
 132 }
 133
 134 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
 135 {
 136         int cpu;
 137         struct callchain_cpus_entries *entries;
 138
 139         *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
 140         if (*rctx == -1)
 141                 return NULL;
 142
 143         entries = rcu_dereference(callchain_cpus_entries);
 144         if (!entries)
 145                 return NULL;
 146
 147         cpu = smp_processor_id();
 148
 149         return &entries->cpu_entries[cpu][*rctx];
 150 }
 151
 152 static void
 153 put_callchain_entry(int rctx)
 154 {
 155         put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
 156 }
 157
 158 struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
 159 {
 160         int rctx;
 161         struct perf_callchain_entry *entry;
 162
 163
 164         entry = get_callchain_entry(&rctx);
 165         if (rctx == -1)
 166                 return NULL;
 167
 168         if (!entry)
 169                 goto exit_put;
 170
 171         entry->nr = 0;
 172
 173         if (!user_mode(regs)) {
 174                 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
 175                 perf_callchain_kernel(entry, regs);
 176                 if (current->mm)
 177                         regs = task_pt_regs(current);
 178                 else
 179                         regs = NULL;
 180         }
 181
 182         if (regs) {
 183                 perf_callchain_store(entry, PERF_CONTEXT_USER);
 184                 perf_callchain_user(entry, regs);
 185         }
 186
 187 exit_put:
 188         put_callchain_entry(rctx);
 189
 190         return entry;
 191 }