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         if (tp_event->class->reg)
  58                 ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER);
  59         else
  60                 ret = tracepoint_probe_register(tp_event->name,
  61                                                 tp_event->class->perf_probe,
  62                                                 tp_event);
  63
  64         if (ret)
  65                 goto fail;
  66
  67         total_ref_count++;
  68         return 0;
  69
  70 fail:
  71         if (!total_ref_count) {
  72                 int i;
  73
  74                 for (i = 0; i < 4; i++) {
  75                         free_percpu(perf_trace_buf[i]);
  76                         perf_trace_buf[i] = NULL;
  77                 }
  78         }
  79
  80         if (!--tp_event->perf_refcount) {
  81                 free_percpu(tp_event->perf_events);
  82                 tp_event->perf_events = NULL;
  83         }
  84
  85         return ret;
  86 }
  87
  88 int perf_trace_init(struct perf_event *p_event)
  89 {
  90         struct ftrace_event_call *tp_event;
  91         int event_id = p_event->attr.config;
  92         int ret = -EINVAL;
  93
  94         mutex_lock(&event_mutex);
  95         list_for_each_entry(tp_event, &ftrace_events, list) {
  96                 if (tp_event->event.type == event_id &&
  97                     tp_event->class && tp_event->class->perf_probe &&
  98                     try_module_get(tp_event->mod)) {
  99                         ret = perf_trace_event_init(tp_event, p_event);
 100                         break;
 101                 }
 102         }
 103         mutex_unlock(&event_mutex);
 104
 105         return ret;
 106 }
 107
 108 int perf_trace_enable(struct perf_event *p_event)
 109 {
 110         struct ftrace_event_call *tp_event = p_event->tp_event;
 111         struct hlist_head *list;
 112
 113         list = tp_event->perf_events;
 114         if (WARN_ON_ONCE(!list))
 115                 return -EINVAL;
 116
 117         list = this_cpu_ptr(list);
 118         hlist_add_head_rcu(&p_event->hlist_entry, list);
 119
 120         return 0;
 121 }
 122
 123 void perf_trace_disable(struct perf_event *p_event)
 124 {
 125         hlist_del_rcu(&p_event->hlist_entry);
 126 }
 127
 128 void perf_trace_destroy(struct perf_event *p_event)
 129 {
 130         struct ftrace_event_call *tp_event = p_event->tp_event;
 131         int i;
 132
 133         mutex_lock(&event_mutex);
 134         if (--tp_event->perf_refcount > 0)
 135                 goto out;
 136
 137         if (tp_event->class->reg)
 138                 tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER);
 139         else
 140                 tracepoint_probe_unregister(tp_event->name,
 141                                             tp_event->class->perf_probe,
 142                                             tp_event);
 143
 144         /*
 145          * Ensure our callback won't be called anymore. See
 146          * tracepoint_probe_unregister() and __DO_TRACE().
 147          */
 148         synchronize_sched();
 149
 150         free_percpu(tp_event->perf_events);
 151         tp_event->perf_events = NULL;
 152
 153         if (!--total_ref_count) {
 154                 for (i = 0; i < 4; i++) {
 155                         free_percpu(perf_trace_buf[i]);
 156                         perf_trace_buf[i] = NULL;
 157                 }
 158         }
 159 out:
 160         mutex_unlock(&event_mutex);
 161 }
 162
 163 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
 164                                        struct pt_regs *regs, int *rctxp)
 165 {
 166         struct trace_entry *entry;
 167         unsigned long flags;
 168         char *raw_data;
 169         int pc;
 170
 171         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 172
 173         pc = preempt_count();
 174
 175         *rctxp = perf_swevent_get_recursion_context();
 176         if (*rctxp < 0)
 177                 return NULL;
 178
 179         raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
 180
 181         /* zero the dead bytes from align to not leak stack to user */
 182         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 183
 184         entry = (struct trace_entry *)raw_data;
 185         local_save_flags(flags);
 186         tracing_generic_entry_update(entry, flags, pc);
 187         entry->type = type;
 188
 189         return raw_data;
 190 }
 191 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);