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         if (tp_event->perf_perm) {
  28                 int ret = tp_event->perf_perm(tp_event, p_event);
  29                 if (ret)
  30                         return ret;
  31         }
  32
  33         /* The ftrace function trace is allowed only for root. */
  34         if (ftrace_event_is_function(tp_event) &&
  35             perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  36                 return -EPERM;
  37
  38         /* No tracing, just counting, so no obvious leak */
  39         if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
  40                 return 0;
  41
  42         /* Some events are ok to be traced by non-root users... */
  43         if (p_event->attach_state == PERF_ATTACH_TASK) {
  44                 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY)
  45                         return 0;
  46         }
  47
  48         /*
  49          * ...otherwise raw tracepoint data can be a severe data leak,
  50          * only allow root to have these.
  51          */
  52         if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  53                 return -EPERM;
  54
  55         return 0;
  56 }
  57
  58 static int perf_trace_event_reg(struct ftrace_event_call *tp_event,
  59                                 struct perf_event *p_event)
  60 {
  61         struct hlist_head __percpu *list;
  62         int ret = -ENOMEM;
  63         int cpu;
  64
  65         p_event->tp_event = tp_event;
  66         if (tp_event->perf_refcount++ > 0)
  67                 return 0;
  68
  69         list = alloc_percpu(struct hlist_head);
  70         if (!list)
  71                 goto fail;
  72
  73         for_each_possible_cpu(cpu)
  74                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
  75
  76         tp_event->perf_events = list;
  77
  78         if (!total_ref_count) {
  79                 char __percpu *buf;
  80                 int i;
  81
  82                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
  83                         buf = (char __percpu *)alloc_percpu(perf_trace_t);
  84                         if (!buf)
  85                                 goto fail;
  86
  87                         perf_trace_buf[i] = buf;
  88                 }
  89         }
  90
  91         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
  92         if (ret)
  93                 goto fail;
  94
  95         total_ref_count++;
  96         return 0;
  97
  98 fail:
  99         if (!total_ref_count) {
 100                 int i;
 101
 102                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 103                         free_percpu(perf_trace_buf[i]);
 104                         perf_trace_buf[i] = NULL;
 105                 }
 106         }
 107
 108         if (!--tp_event->perf_refcount) {
 109                 free_percpu(tp_event->perf_events);
 110                 tp_event->perf_events = NULL;
 111         }
 112
 113         return ret;
 114 }
 115
 116 static void perf_trace_event_unreg(struct perf_event *p_event)
 117 {
 118         struct ftrace_event_call *tp_event = p_event->tp_event;
 119         int i;
 120
 121         if (--tp_event->perf_refcount > 0)
 122                 goto out;
 123
 124         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
 125
 126         /*
 127          * Ensure our callback won't be called anymore. The buffers
 128          * will be freed after that.
 129          */
 130         tracepoint_synchronize_unregister();
 131
 132         free_percpu(tp_event->perf_events);
 133         tp_event->perf_events = NULL;
 134
 135         if (!--total_ref_count) {
 136                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 137                         free_percpu(perf_trace_buf[i]);
 138                         perf_trace_buf[i] = NULL;
 139                 }
 140         }
 141 out:
 142         module_put(tp_event->mod);
 143 }
 144
 145 static int perf_trace_event_open(struct perf_event *p_event)
 146 {
 147         struct ftrace_event_call *tp_event = p_event->tp_event;
 148         return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event);
 149 }
 150
 151 static void perf_trace_event_close(struct perf_event *p_event)
 152 {
 153         struct ftrace_event_call *tp_event = p_event->tp_event;
 154         tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
 155 }
 156
 157 static int perf_trace_event_init(struct ftrace_event_call *tp_event,
 158                                  struct perf_event *p_event)
 159 {
 160         int ret;
 161
 162         ret = perf_trace_event_perm(tp_event, p_event);
 163         if (ret)
 164                 return ret;
 165
 166         ret = perf_trace_event_reg(tp_event, p_event);
 167         if (ret)
 168                 return ret;
 169
 170         ret = perf_trace_event_open(p_event);
 171         if (ret) {
 172                 perf_trace_event_unreg(p_event);
 173                 return ret;
 174         }
 175
 176         return 0;
 177 }
 178
 179 int perf_trace_init(struct perf_event *p_event)
 180 {
 181         struct ftrace_event_call *tp_event;
 182         u64 event_id = p_event->attr.config;
 183         int ret = -EINVAL;
 184
 185         mutex_lock(&event_mutex);
 186         list_for_each_entry(tp_event, &ftrace_events, list) {
 187                 if (tp_event->event.type == event_id &&
 188                     tp_event->class && tp_event->class->reg &&
 189                     try_module_get(tp_event->mod)) {
 190                         ret = perf_trace_event_init(tp_event, p_event);
 191                         if (ret)
 192                                 module_put(tp_event->mod);
 193                         break;
 194                 }
 195         }
 196         mutex_unlock(&event_mutex);
 197
 198         return ret;
 199 }
 200
 201 void perf_trace_destroy(struct perf_event *p_event)
 202 {
 203         mutex_lock(&event_mutex);
 204         perf_trace_event_close(p_event);
 205         perf_trace_event_unreg(p_event);
 206         mutex_unlock(&event_mutex);
 207 }
 208
 209 int perf_trace_add(struct perf_event *p_event, int flags)
 210 {
 211         struct ftrace_event_call *tp_event = p_event->tp_event;
 212         struct hlist_head __percpu *pcpu_list;
 213         struct hlist_head *list;
 214
 215         pcpu_list = tp_event->perf_events;
 216         if (WARN_ON_ONCE(!pcpu_list))
 217                 return -EINVAL;
 218
 219         if (!(flags & PERF_EF_START))
 220                 p_event->hw.state = PERF_HES_STOPPED;
 221
 222         list = this_cpu_ptr(pcpu_list);
 223         hlist_add_head_rcu(&p_event->hlist_entry, list);
 224
 225         return tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event);
 226 }
 227
 228 void perf_trace_del(struct perf_event *p_event, int flags)
 229 {
 230         struct ftrace_event_call *tp_event = p_event->tp_event;
 231         hlist_del_rcu(&p_event->hlist_entry);
 232         tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event);
 233 }
 234
 235 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
 236                                        struct pt_regs *regs, int *rctxp)
 237 {
 238         struct trace_entry *entry;
 239         unsigned long flags;
 240         char *raw_data;
 241         int pc;
 242
 243         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 244
 245         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 246                         "perf buffer not large enough"))
 247                 return NULL;
 248
 249         pc = preempt_count();
 250
 251         *rctxp = perf_swevent_get_recursion_context();
 252         if (*rctxp < 0)
 253                 return NULL;
 254
 255         raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
 256
 257         /* zero the dead bytes from align to not leak stack to user */
 258         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 259
 260         entry = (struct trace_entry *)raw_data;
 261         local_save_flags(flags);
 262         tracing_generic_entry_update(entry, flags, pc);
 263         entry->type = type;
 264
 265         return raw_data;
 266 }
 267 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);
 268
 269 #ifdef CONFIG_FUNCTION_TRACER
 270 static void
 271 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 272                           struct ftrace_ops *ops, struct pt_regs *pt_regs)
 273 {
 274         struct ftrace_entry *entry;
 275         struct hlist_head *head;
 276         struct pt_regs regs;
 277         int rctx;
 278
 279         head = this_cpu_ptr(event_function.perf_events);
 280         if (hlist_empty(head))
 281                 return;
 282
 283 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 284                     sizeof(u64)) - sizeof(u32))
 285
 286         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 287
 288         perf_fetch_caller_regs(&regs);
 289
 290         entry = perf_trace_buf_prepare(ENTRY_SIZE, TRACE_FN, NULL, &rctx);
 291         if (!entry)
 292                 return;
 293
 294         entry->ip = ip;
 295         entry->parent_ip = parent_ip;
 296         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, 0,
 297                               1, &regs, head, NULL);
 298
 299 #undef ENTRY_SIZE
 300 }
 301
 302 static int perf_ftrace_function_register(struct perf_event *event)
 303 {
 304         struct ftrace_ops *ops = &event->ftrace_ops;
 305
 306         ops->flags |= FTRACE_OPS_FL_CONTROL;
 307         ops->func = perf_ftrace_function_call;
 308         return register_ftrace_function(ops);
 309 }
 310
 311 static int perf_ftrace_function_unregister(struct perf_event *event)
 312 {
 313         struct ftrace_ops *ops = &event->ftrace_ops;
 314         int ret = unregister_ftrace_function(ops);
 315         ftrace_free_filter(ops);
 316         return ret;
 317 }
 318
 319 static void perf_ftrace_function_enable(struct perf_event *event)
 320 {
 321         ftrace_function_local_enable(&event->ftrace_ops);
 322 }
 323
 324 static void perf_ftrace_function_disable(struct perf_event *event)
 325 {
 326         ftrace_function_local_disable(&event->ftrace_ops);
 327 }
 328
 329 int perf_ftrace_event_register(struct ftrace_event_call *call,
 330                                enum trace_reg type, void *data)
 331 {
 332         switch (type) {
 333         case TRACE_REG_REGISTER:
 334         case TRACE_REG_UNREGISTER:
 335                 break;
 336         case TRACE_REG_PERF_REGISTER:
 337         case TRACE_REG_PERF_UNREGISTER:
 338                 return 0;
 339         case TRACE_REG_PERF_OPEN:
 340                 return perf_ftrace_function_register(data);
 341         case TRACE_REG_PERF_CLOSE:
 342                 return perf_ftrace_function_unregister(data);
 343         case TRACE_REG_PERF_ADD:
 344                 perf_ftrace_function_enable(data);
 345                 return 0;
 346         case TRACE_REG_PERF_DEL:
 347                 perf_ftrace_function_disable(data);
 348                 return 0;
 349         }
 350
 351         return -EINVAL;
 352 }
 353 #endif /* CONFIG_FUNCTION_TRACER */