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 DEFINE_PER_CPU(struct pt_regs, perf_trace_regs);
  13 EXPORT_PER_CPU_SYMBOL_GPL(perf_trace_regs);
  14
  15 EXPORT_SYMBOL_GPL(perf_arch_fetch_caller_regs);
  16
  17 static char *perf_trace_buf;
  18 static char *perf_trace_buf_nmi;
  19
  20 typedef typeof(char [PERF_MAX_TRACE_SIZE]) perf_trace_t ;
  21
  22 /* Count the events in use (per event id, not per instance) */
  23 static int      total_ref_count;
  24
  25 static int perf_trace_event_enable(struct ftrace_event_call *event)
  26 {
  27         char *buf;
  28         int ret = -ENOMEM;
  29
  30         if (event->perf_refcount++ > 0)
  31                 return 0;
  32
  33         if (!total_ref_count) {
  34                 buf = (char *)alloc_percpu(perf_trace_t);
  35                 if (!buf)
  36                         goto fail_buf;
  37
  38                 rcu_assign_pointer(perf_trace_buf, buf);
  39
  40                 buf = (char *)alloc_percpu(perf_trace_t);
  41                 if (!buf)
  42                         goto fail_buf_nmi;
  43
  44                 rcu_assign_pointer(perf_trace_buf_nmi, buf);
  45         }
  46
  47         ret = event->perf_event_enable(event);
  48         if (!ret) {
  49                 total_ref_count++;
  50                 return 0;
  51         }
  52
  53 fail_buf_nmi:
  54         if (!total_ref_count) {
  55                 free_percpu(perf_trace_buf_nmi);
  56                 free_percpu(perf_trace_buf);
  57                 perf_trace_buf_nmi = NULL;
  58                 perf_trace_buf = NULL;
  59         }
  60 fail_buf:
  61         event->perf_refcount--;
  62
  63         return ret;
  64 }
  65
  66 int perf_trace_enable(int event_id)
  67 {
  68         struct ftrace_event_call *event;
  69         int ret = -EINVAL;
  70
  71         mutex_lock(&event_mutex);
  72         list_for_each_entry(event, &ftrace_events, list) {
  73                 if (event->id == event_id && event->perf_event_enable &&
  74                     try_module_get(event->mod)) {
  75                         ret = perf_trace_event_enable(event);
  76                         break;
  77                 }
  78         }
  79         mutex_unlock(&event_mutex);
  80
  81         return ret;
  82 }
  83
  84 static void perf_trace_event_disable(struct ftrace_event_call *event)
  85 {
  86         char *buf, *nmi_buf;
  87
  88         if (--event->perf_refcount > 0)
  89                 return;
  90
  91         event->perf_event_disable(event);
  92
  93         if (!--total_ref_count) {
  94                 buf = perf_trace_buf;
  95                 rcu_assign_pointer(perf_trace_buf, NULL);
  96
  97                 nmi_buf = perf_trace_buf_nmi;
  98                 rcu_assign_pointer(perf_trace_buf_nmi, NULL);
  99
 100                 /*
 101                  * Ensure every events in profiling have finished before
 102                  * releasing the buffers
 103                  */
 104                 synchronize_sched();
 105
 106                 free_percpu(buf);
 107                 free_percpu(nmi_buf);
 108         }
 109 }
 110
 111 void perf_trace_disable(int event_id)
 112 {
 113         struct ftrace_event_call *event;
 114
 115         mutex_lock(&event_mutex);
 116         list_for_each_entry(event, &ftrace_events, list) {
 117                 if (event->id == event_id) {
 118                         perf_trace_event_disable(event);
 119                         module_put(event->mod);
 120                         break;
 121                 }
 122         }
 123         mutex_unlock(&event_mutex);
 124 }
 125
 126 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
 127                                        int *rctxp, unsigned long *irq_flags)
 128 {
 129         struct trace_entry *entry;
 130         char *trace_buf, *raw_data;
 131         int pc, cpu;
 132
 133         pc = preempt_count();
 134
 135         /* Protect the per cpu buffer, begin the rcu read side */
 136         local_irq_save(*irq_flags);
 137
 138         *rctxp = perf_swevent_get_recursion_context();
 139         if (*rctxp < 0)
 140                 goto err_recursion;
 141
 142         cpu = smp_processor_id();
 143
 144         if (in_nmi())
 145                 trace_buf = rcu_dereference_sched(perf_trace_buf_nmi);
 146         else
 147                 trace_buf = rcu_dereference_sched(perf_trace_buf);
 148
 149         if (!trace_buf)
 150                 goto err;
 151
 152         raw_data = per_cpu_ptr(trace_buf, cpu);
 153
 154         /* zero the dead bytes from align to not leak stack to user */
 155         *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 156
 157         entry = (struct trace_entry *)raw_data;
 158         tracing_generic_entry_update(entry, *irq_flags, pc);
 159         entry->type = type;
 160
 161         return raw_data;
 162 err:
 163         perf_swevent_put_recursion_context(*rctxp);
 164 err_recursion:
 165         local_irq_restore(*irq_flags);
 166         return NULL;
 167 }
 168 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);