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 /*
  21  * Force it to be aligned to unsigned long to avoid misaligned accesses
  22  * suprises
  23  */
  24 typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
  25         perf_trace_t;
  26
  27 /* Count the events in use (per event id, not per instance) */
  28 static int      total_ref_count;
  29
  30 static int perf_trace_event_enable(struct ftrace_event_call *event)
  31 {
  32         char *buf;
  33         int ret = -ENOMEM;
  34
  35         if (event->perf_refcount++ > 0)
  36                 return 0;
  37
  38         if (!total_ref_count) {
  39                 buf = (char *)alloc_percpu(perf_trace_t);
  40                 if (!buf)
  41                         goto fail_buf;
  42
  43                 rcu_assign_pointer(perf_trace_buf, buf);
  44
  45                 buf = (char *)alloc_percpu(perf_trace_t);
  46                 if (!buf)
  47                         goto fail_buf_nmi;
  48
  49                 rcu_assign_pointer(perf_trace_buf_nmi, buf);
  50         }
  51
  52         if (event->class->reg)
  53                 ret = event->class->reg(event, TRACE_REG_PERF_REGISTER);
  54         else
  55                 ret = tracepoint_probe_register(event->name,
  56                                                 event->class->perf_probe,
  57                                                 event);
  58         if (!ret) {
  59                 total_ref_count++;
  60                 return 0;
  61         }
  62
  63 fail_buf_nmi:
  64         if (!total_ref_count) {
  65                 free_percpu(perf_trace_buf_nmi);
  66                 free_percpu(perf_trace_buf);
  67                 perf_trace_buf_nmi = NULL;
  68                 perf_trace_buf = NULL;
  69         }
  70 fail_buf:
  71         event->perf_refcount--;
  72
  73         return ret;
  74 }
  75
  76 int perf_trace_enable(int event_id)
  77 {
  78         struct ftrace_event_call *event;
  79         int ret = -EINVAL;
  80
  81         mutex_lock(&event_mutex);
  82         list_for_each_entry(event, &ftrace_events, list) {
  83                 if (event->event.type == event_id &&
  84                     event->class && event->class->perf_probe &&
  85                     try_module_get(event->mod)) {
  86                         ret = perf_trace_event_enable(event);
  87                         break;
  88                 }
  89         }
  90         mutex_unlock(&event_mutex);
  91
  92         return ret;
  93 }
  94
  95 static void perf_trace_event_disable(struct ftrace_event_call *event)
  96 {
  97         char *buf, *nmi_buf;
  98
  99         if (--event->perf_refcount > 0)
 100                 return;
 101
 102         if (event->class->reg)
 103                 event->class->reg(event, TRACE_REG_PERF_UNREGISTER);
 104         else
 105                 tracepoint_probe_unregister(event->name, event->class->perf_probe, event);
 106
 107         if (!--total_ref_count) {
 108                 buf = perf_trace_buf;
 109                 rcu_assign_pointer(perf_trace_buf, NULL);
 110
 111                 nmi_buf = perf_trace_buf_nmi;
 112                 rcu_assign_pointer(perf_trace_buf_nmi, NULL);
 113
 114                 /*
 115                  * Ensure every events in profiling have finished before
 116                  * releasing the buffers
 117                  */
 118                 synchronize_sched();
 119
 120                 free_percpu(buf);
 121                 free_percpu(nmi_buf);
 122         }
 123 }
 124
 125 void perf_trace_disable(int event_id)
 126 {
 127         struct ftrace_event_call *event;
 128
 129         mutex_lock(&event_mutex);
 130         list_for_each_entry(event, &ftrace_events, list) {
 131                 if (event->event.type == event_id) {
 132                         perf_trace_event_disable(event);
 133                         module_put(event->mod);
 134                         break;
 135                 }
 136         }
 137         mutex_unlock(&event_mutex);
 138 }
 139
 140 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
 141                                        int *rctxp, unsigned long *irq_flags)
 142 {
 143         struct trace_entry *entry;
 144         char *trace_buf, *raw_data;
 145         int pc, cpu;
 146
 147         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 148
 149         pc = preempt_count();
 150
 151         /* Protect the per cpu buffer, begin the rcu read side */
 152         local_irq_save(*irq_flags);
 153
 154         *rctxp = perf_swevent_get_recursion_context();
 155         if (*rctxp < 0)
 156                 goto err_recursion;
 157
 158         cpu = smp_processor_id();
 159
 160         if (in_nmi())
 161                 trace_buf = rcu_dereference_sched(perf_trace_buf_nmi);
 162         else
 163                 trace_buf = rcu_dereference_sched(perf_trace_buf);
 164
 165         if (!trace_buf)
 166                 goto err;
 167
 168         raw_data = per_cpu_ptr(trace_buf, cpu);
 169
 170         /* zero the dead bytes from align to not leak stack to user */
 171         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 172
 173         entry = (struct trace_entry *)raw_data;
 174         tracing_generic_entry_update(entry, *irq_flags, pc);
 175         entry->type = type;
 176
 177         return raw_data;
 178 err:
 179         perf_swevent_put_recursion_context(*rctxp);
 180 err_recursion:
 181         local_irq_restore(*irq_flags);
 182         return NULL;
 183 }
 184 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);