kernel/trace/trace_event_perf.c

   1 /*
   2  * trace event based perf event profiling/tracing
   3  *
   4  * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra <pzijlstr@redhat.com>
   5  * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com>
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/kprobes.h>
  10 #include "trace.h"
  11
  12 static char *perf_trace_buf[4];
  13
  14 /*
  15  * Force it to be aligned to unsigned long to avoid misaligned accesses
  16  * suprises
  17  */
  18 typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
  19         perf_trace_t;
  20
  21 /* Count the events in use (per event id, not per instance) */
  22 static int      total_ref_count;
  23
  24 static int perf_trace_event_init(struct ftrace_event_call *tp_event,
  25                                  struct perf_event *p_event)
  26 {
  27         struct hlist_head *list;
  28         int ret = -ENOMEM;
  29         int cpu;
  30
  31         p_event->tp_event = tp_event;
  32         if (tp_event->perf_refcount++ > 0)
  33                 return 0;
  34
  35         list = alloc_percpu(struct hlist_head);
  36         if (!list)
  37                 goto fail;
  38
  39         for_each_possible_cpu(cpu)
  40                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
  41
  42         tp_event->perf_events = list;
  43
  44         if (!total_ref_count) {
  45                 char *buf;
  46                 int i;
  47
  48                 for (i = 0; i < 4; i++) {
  49                         buf = (char *)alloc_percpu(perf_trace_t);
  50                         if (!buf)
  51                                 goto fail;
  52
  53                         perf_trace_buf[i] = buf;
  54                 }
  55         }
  56
  57         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER);
  58         if (ret)
  59                 goto fail;
  60
  61         total_ref_count++;
  62         return 0;
  63
  64 fail:
  65         if (!total_ref_count) {
  66                 int i;
  67
  68                 for (i = 0; i < 4; i++) {
  69                         free_percpu(perf_trace_buf[i]);
  70                         perf_trace_buf[i] = NULL;
  71                 }
  72         }
  73
  74         if (!--tp_event->perf_refcount) {
  75                 free_percpu(tp_event->perf_events);
  76                 tp_event->perf_events = NULL;
  77         }
  78
  79         return ret;
  80 }
  81
  82 int perf_trace_init(struct perf_event *p_event)
  83 {
  84         struct ftrace_event_call *tp_event;
  85         int event_id = p_event->attr.config;
  86         int ret = -EINVAL;
  87
  88         mutex_lock(&event_mutex);
  89         list_for_each_entry(tp_event, &ftrace_events, list) {
  90                 if (tp_event->event.type == event_id &&
  91                     tp_event->class && tp_event->class->reg &&
  92                     try_module_get(tp_event->mod)) {
  93                         ret = perf_trace_event_init(tp_event, p_event);
  94                         if (ret)
  95                                 module_put(tp_event->mod);
  96                         break;
  97                 }
  98         }
  99         mutex_unlock(&event_mutex);
 100
 101         return ret;
 102 }
 103
 104 int perf_trace_enable(struct perf_event *p_event)
 105 {
 106         struct ftrace_event_call *tp_event = p_event->tp_event;
 107         struct hlist_head *list;
 108
 109         list = tp_event->perf_events;
 110         if (WARN_ON_ONCE(!list))
 111                 return -EINVAL;
 112
 113         list = this_cpu_ptr(list);
 114         hlist_add_head_rcu(&p_event->hlist_entry, list);
 115
 116         return 0;
 117 }
 118
 119 void perf_trace_disable(struct perf_event *p_event)
 120 {
 121         hlist_del_rcu(&p_event->hlist_entry);
 122 }
 123
 124 void perf_trace_destroy(struct perf_event *p_event)
 125 {
 126         struct ftrace_event_call *tp_event = p_event->tp_event;
 127         int i;
 128
 129         mutex_lock(&event_mutex);
 130         if (--tp_event->perf_refcount > 0)
 131                 goto out;
 132
 133         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER);
 134
 135         /*
 136          * Ensure our callback won't be called anymore. The buffers
 137          * will be freed after that.
 138          */
 139         tracepoint_synchronize_unregister();
 140
 141         free_percpu(tp_event->perf_events);
 142         tp_event->perf_events = NULL;
 143
 144         if (!--total_ref_count) {
 145                 for (i = 0; i < 4; i++) {
 146                         free_percpu(perf_trace_buf[i]);
 147                         perf_trace_buf[i] = NULL;
 148                 }
 149         }
 150 out:
 151         module_put(tp_event->mod);
 152         mutex_unlock(&event_mutex);
 153 }
 154
 155 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
 156                                        struct pt_regs *regs, int *rctxp)
 157 {
 158         struct trace_entry *entry;
 159         unsigned long flags;
 160         char *raw_data;
 161         int pc;
 162
 163         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 164
 165         pc = preempt_count();
 166
 167         *rctxp = perf_swevent_get_recursion_context();
 168         if (*rctxp < 0)
 169                 return NULL;
 170
 171         raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
 172
 173         /* zero the dead bytes from align to not leak stack to user */
 174         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 175
 176         entry = (struct trace_entry *)raw_data;
 177         local_save_flags(flags);
 178         tracing_generic_entry_update(entry, flags, pc);
 179         entry->type = type;
 180
 181         return raw_data;
 182 }
 183 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);