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 __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
  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_perm(struct ftrace_event_call *tp_event,
  25                                  struct perf_event *p_event)
  26 {
  27         /* No tracing, just counting, so no obvious leak */
  28         if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
  29                 return 0;
  30
  31         /* Some events are ok to be traced by non-root users... */
  32         if (p_event->attach_state == PERF_ATTACH_TASK) {
  33                 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY)
  34                         return 0;
  35         }
  36
  37         /*
  38          * ...otherwise raw tracepoint data can be a severe data leak,
  39          * only allow root to have these.
  40          */
  41         if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  42                 return -EPERM;
  43
  44         return 0;
  45 }
  46
  47 static int perf_trace_event_init(struct ftrace_event_call *tp_event,
  48                                  struct perf_event *p_event)
  49 {
  50         struct hlist_head __percpu *list;
  51         int ret;
  52         int cpu;
  53
  54         ret = perf_trace_event_perm(tp_event, p_event);
  55         if (ret)
  56                 return ret;
  57
  58         p_event->tp_event = tp_event;
  59         if (tp_event->perf_refcount++ > 0)
  60                 return 0;
  61
  62         ret = -ENOMEM;
  63
  64         list = alloc_percpu(struct hlist_head);
  65         if (!list)
  66                 goto fail;
  67
  68         for_each_possible_cpu(cpu)
  69                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
  70
  71         tp_event->perf_events = list;
  72
  73         if (!total_ref_count) {
  74                 char __percpu *buf;
  75                 int i;
  76
  77                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
  78                         buf = (char __percpu *)alloc_percpu(perf_trace_t);
  79                         if (!buf)
  80                                 goto fail;
  81
  82                         perf_trace_buf[i] = buf;
  83                 }
  84         }
  85
  86         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER);
  87         if (ret)
  88                 goto fail;
  89
  90         total_ref_count++;
  91         return 0;
  92
  93 fail:
  94         if (!total_ref_count) {
  95                 int i;
  96
  97                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
  98                         free_percpu(perf_trace_buf[i]);
  99                         perf_trace_buf[i] = NULL;
 100                 }
 101         }
 102
 103         if (!--tp_event->perf_refcount) {
 104                 free_percpu(tp_event->perf_events);
 105                 tp_event->perf_events = NULL;
 106         }
 107
 108         return ret;
 109 }
 110
 111 int perf_trace_init(struct perf_event *p_event)
 112 {
 113         struct ftrace_event_call *tp_event;
 114         int event_id = p_event->attr.config;
 115         int ret = -EINVAL;
 116
 117         mutex_lock(&event_mutex);
 118         list_for_each_entry(tp_event, &ftrace_events, list) {
 119                 if (tp_event->event.type == event_id &&
 120                     tp_event->class && tp_event->class->reg &&
 121                     try_module_get(tp_event->mod)) {
 122                         ret = perf_trace_event_init(tp_event, p_event);
 123                         if (ret)
 124                                 module_put(tp_event->mod);
 125                         break;
 126                 }
 127         }
 128         mutex_unlock(&event_mutex);
 129
 130         return ret;
 131 }
 132
 133 int perf_trace_add(struct perf_event *p_event, int flags)
 134 {
 135         struct ftrace_event_call *tp_event = p_event->tp_event;
 136         struct hlist_head __percpu *pcpu_list;
 137         struct hlist_head *list;
 138
 139         pcpu_list = tp_event->perf_events;
 140         if (WARN_ON_ONCE(!pcpu_list))
 141                 return -EINVAL;
 142
 143         if (!(flags & PERF_EF_START))
 144                 p_event->hw.state = PERF_HES_STOPPED;
 145
 146         list = this_cpu_ptr(pcpu_list);
 147         hlist_add_head_rcu(&p_event->hlist_entry, list);
 148
 149         return 0;
 150 }
 151
 152 void perf_trace_del(struct perf_event *p_event, int flags)
 153 {
 154         hlist_del_rcu(&p_event->hlist_entry);
 155 }
 156
 157 void perf_trace_destroy(struct perf_event *p_event)
 158 {
 159         struct ftrace_event_call *tp_event = p_event->tp_event;
 160         int i;
 161
 162         mutex_lock(&event_mutex);
 163         if (--tp_event->perf_refcount > 0)
 164                 goto out;
 165
 166         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER);
 167
 168         /*
 169          * Ensure our callback won't be called anymore. The buffers
 170          * will be freed after that.
 171          */
 172         tracepoint_synchronize_unregister();
 173
 174         free_percpu(tp_event->perf_events);
 175         tp_event->perf_events = NULL;
 176
 177         if (!--total_ref_count) {
 178                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 179                         free_percpu(perf_trace_buf[i]);
 180                         perf_trace_buf[i] = NULL;
 181                 }
 182         }
 183 out:
 184         module_put(tp_event->mod);
 185         mutex_unlock(&event_mutex);
 186 }
 187
 188 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
 189                                        struct pt_regs *regs, int *rctxp)
 190 {
 191         struct trace_entry *entry;
 192         unsigned long flags;
 193         char *raw_data;
 194         int pc;
 195
 196         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 197
 198         pc = preempt_count();
 199
 200         *rctxp = perf_swevent_get_recursion_context();
 201         if (*rctxp < 0)
 202                 return NULL;
 203
 204         raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
 205
 206         /* zero the dead bytes from align to not leak stack to user */
 207         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 208
 209         entry = (struct trace_entry *)raw_data;
 210         local_save_flags(flags);
 211         tracing_generic_entry_update(entry, flags, pc);
 212         entry->type = type;
 213
 214         return raw_data;
 215 }
 216 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);