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
   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 #include "trace_probe.h"
  12
  13 static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
  14
  15 /*
  16  * Force it to be aligned to unsigned long to avoid misaligned accesses
  17  * suprises
  18  */
  19 typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
  20         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_perm(struct trace_event_call *tp_event,
  26                                  struct perf_event *p_event)
  27 {
  28         if (tp_event->perf_perm) {
  29                 int ret = tp_event->perf_perm(tp_event, p_event);
  30                 if (ret)
  31                         return ret;
  32         }
  33
  34         /*
  35          * We checked and allowed to create parent,
  36          * allow children without checking.
  37          */
  38         if (p_event->parent)
  39                 return 0;
  40
  41         /*
  42          * It's ok to check current process (owner) permissions in here,
  43          * because code below is called only via perf_event_open syscall.
  44          */
  45
  46         /* The ftrace function trace is allowed only for root. */
  47         if (ftrace_event_is_function(tp_event)) {
  48                 if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  49                         return -EPERM;
  50
  51                 if (!is_sampling_event(p_event))
  52                         return 0;
  53
  54                 /*
  55                  * We don't allow user space callchains for  function trace
  56                  * event, due to issues with page faults while tracing page
  57                  * fault handler and its overall trickiness nature.
  58                  */
  59                 if (!p_event->attr.exclude_callchain_user)
  60                         return -EINVAL;
  61
  62                 /*
  63                  * Same reason to disable user stack dump as for user space
  64                  * callchains above.
  65                  */
  66                 if (p_event->attr.sample_type & PERF_SAMPLE_STACK_USER)
  67                         return -EINVAL;
  68         }
  69
  70         /* No tracing, just counting, so no obvious leak */
  71         if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
  72                 return 0;
  73
  74         /* Some events are ok to be traced by non-root users... */
  75         if (p_event->attach_state == PERF_ATTACH_TASK) {
  76                 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY)
  77                         return 0;
  78         }
  79
  80         /*
  81          * ...otherwise raw tracepoint data can be a severe data leak,
  82          * only allow root to have these.
  83          */
  84         if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN))
  85                 return -EPERM;
  86
  87         return 0;
  88 }
  89
  90 static int perf_trace_event_reg(struct trace_event_call *tp_event,
  91                                 struct perf_event *p_event)
  92 {
  93         struct hlist_head __percpu *list;
  94         int ret = -ENOMEM;
  95         int cpu;
  96
  97         p_event->tp_event = tp_event;
  98         if (tp_event->perf_refcount++ > 0)
  99                 return 0;
 100
 101         list = alloc_percpu(struct hlist_head);
 102         if (!list)
 103                 goto fail;
 104
 105         for_each_possible_cpu(cpu)
 106                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
 107
 108         tp_event->perf_events = list;
 109
 110         if (!total_ref_count) {
 111                 char __percpu *buf;
 112                 int i;
 113
 114                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 115                         buf = (char __percpu *)alloc_percpu(perf_trace_t);
 116                         if (!buf)
 117                                 goto fail;
 118
 119                         perf_trace_buf[i] = buf;
 120                 }
 121         }
 122
 123         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
 124         if (ret)
 125                 goto fail;
 126
 127         total_ref_count++;
 128         return 0;
 129
 130 fail:
 131         if (!total_ref_count) {
 132                 int i;
 133
 134                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 135                         free_percpu(perf_trace_buf[i]);
 136                         perf_trace_buf[i] = NULL;
 137                 }
 138         }
 139
 140         if (!--tp_event->perf_refcount) {
 141                 free_percpu(tp_event->perf_events);
 142                 tp_event->perf_events = NULL;
 143         }
 144
 145         return ret;
 146 }
 147
 148 static void perf_trace_event_unreg(struct perf_event *p_event)
 149 {
 150         struct trace_event_call *tp_event = p_event->tp_event;
 151         int i;
 152
 153         if (--tp_event->perf_refcount > 0)
 154                 goto out;
 155
 156         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
 157
 158         /*
 159          * Ensure our callback won't be called anymore. The buffers
 160          * will be freed after that.
 161          */
 162         tracepoint_synchronize_unregister();
 163
 164         free_percpu(tp_event->perf_events);
 165         tp_event->perf_events = NULL;
 166
 167         if (!--total_ref_count) {
 168                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 169                         free_percpu(perf_trace_buf[i]);
 170                         perf_trace_buf[i] = NULL;
 171                 }
 172         }
 173 out:
 174         module_put(tp_event->mod);
 175 }
 176
 177 static int perf_trace_event_open(struct perf_event *p_event)
 178 {
 179         struct trace_event_call *tp_event = p_event->tp_event;
 180         return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event);
 181 }
 182
 183 static void perf_trace_event_close(struct perf_event *p_event)
 184 {
 185         struct trace_event_call *tp_event = p_event->tp_event;
 186         tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
 187 }
 188
 189 static int perf_trace_event_init(struct trace_event_call *tp_event,
 190                                  struct perf_event *p_event)
 191 {
 192         int ret;
 193
 194         ret = perf_trace_event_perm(tp_event, p_event);
 195         if (ret)
 196                 return ret;
 197
 198         ret = perf_trace_event_reg(tp_event, p_event);
 199         if (ret)
 200                 return ret;
 201
 202         ret = perf_trace_event_open(p_event);
 203         if (ret) {
 204                 perf_trace_event_unreg(p_event);
 205                 return ret;
 206         }
 207
 208         return 0;
 209 }
 210
 211 int perf_trace_init(struct perf_event *p_event)
 212 {
 213         struct trace_event_call *tp_event;
 214         u64 event_id = p_event->attr.config;
 215         int ret = -EINVAL;
 216
 217         mutex_lock(&event_mutex);
 218         list_for_each_entry(tp_event, &ftrace_events, list) {
 219                 if (tp_event->event.type == event_id &&
 220                     tp_event->class && tp_event->class->reg &&
 221                     try_module_get(tp_event->mod)) {
 222                         ret = perf_trace_event_init(tp_event, p_event);
 223                         if (ret)
 224                                 module_put(tp_event->mod);
 225                         break;
 226                 }
 227         }
 228         mutex_unlock(&event_mutex);
 229
 230         return ret;
 231 }
 232
 233 void perf_trace_destroy(struct perf_event *p_event)
 234 {
 235         mutex_lock(&event_mutex);
 236         perf_trace_event_close(p_event);
 237         perf_trace_event_unreg(p_event);
 238         mutex_unlock(&event_mutex);
 239 }
 240
 241 #ifdef CONFIG_KPROBE_EVENTS
 242 int perf_kprobe_init(struct perf_event *p_event, bool is_retprobe)
 243 {
 244         int ret;
 245         char *func = NULL;
 246         struct trace_event_call *tp_event;
 247
 248         if (p_event->attr.kprobe_func) {
 249                 func = kzalloc(KSYM_NAME_LEN, GFP_KERNEL);
 250                 if (!func)
 251                         return -ENOMEM;
 252                 ret = strncpy_from_user(
 253                         func, u64_to_user_ptr(p_event->attr.kprobe_func),
 254                         KSYM_NAME_LEN);
 255                 if (ret < 0)
 256                         goto out;
 257
 258                 if (func[0] == '\0') {
 259                         kfree(func);
 260                         func = NULL;
 261                 }
 262         }
 263
 264         tp_event = create_local_trace_kprobe(
 265                 func, (void *)(unsigned long)(p_event->attr.kprobe_addr),
 266                 p_event->attr.probe_offset, is_retprobe);
 267         if (IS_ERR(tp_event)) {
 268                 ret = PTR_ERR(tp_event);
 269                 goto out;
 270         }
 271
 272         ret = perf_trace_event_init(tp_event, p_event);
 273         if (ret)
 274                 destroy_local_trace_kprobe(tp_event);
 275 out:
 276         kfree(func);
 277         return ret;
 278 }
 279
 280 void perf_kprobe_destroy(struct perf_event *p_event)
 281 {
 282         perf_trace_event_close(p_event);
 283         perf_trace_event_unreg(p_event);
 284
 285         destroy_local_trace_kprobe(p_event->tp_event);
 286 }
 287 #endif /* CONFIG_KPROBE_EVENTS */
 288
 289 #ifdef CONFIG_UPROBE_EVENTS
 290 int perf_uprobe_init(struct perf_event *p_event, bool is_retprobe)
 291 {
 292         int ret;
 293         char *path = NULL;
 294         struct trace_event_call *tp_event;
 295
 296         if (!p_event->attr.uprobe_path)
 297                 return -EINVAL;
 298         path = kzalloc(PATH_MAX, GFP_KERNEL);
 299         if (!path)
 300                 return -ENOMEM;
 301         ret = strncpy_from_user(
 302                 path, u64_to_user_ptr(p_event->attr.uprobe_path), PATH_MAX);
 303         if (ret < 0)
 304                 goto out;
 305         if (path[0] == '\0') {
 306                 ret = -EINVAL;
 307                 goto out;
 308         }
 309
 310         tp_event = create_local_trace_uprobe(
 311                 path, p_event->attr.probe_offset, is_retprobe);
 312         if (IS_ERR(tp_event)) {
 313                 ret = PTR_ERR(tp_event);
 314                 goto out;
 315         }
 316
 317         /*
 318          * local trace_uprobe need to hold event_mutex to call
 319          * uprobe_buffer_enable() and uprobe_buffer_disable().
 320          * event_mutex is not required for local trace_kprobes.
 321          */
 322         mutex_lock(&event_mutex);
 323         ret = perf_trace_event_init(tp_event, p_event);
 324         if (ret)
 325                 destroy_local_trace_uprobe(tp_event);
 326         mutex_unlock(&event_mutex);
 327 out:
 328         kfree(path);
 329         return ret;
 330 }
 331
 332 void perf_uprobe_destroy(struct perf_event *p_event)
 333 {
 334         mutex_lock(&event_mutex);
 335         perf_trace_event_close(p_event);
 336         perf_trace_event_unreg(p_event);
 337         mutex_unlock(&event_mutex);
 338         destroy_local_trace_uprobe(p_event->tp_event);
 339 }
 340 #endif /* CONFIG_UPROBE_EVENTS */
 341
 342 int perf_trace_add(struct perf_event *p_event, int flags)
 343 {
 344         struct trace_event_call *tp_event = p_event->tp_event;
 345
 346         if (!(flags & PERF_EF_START))
 347                 p_event->hw.state = PERF_HES_STOPPED;
 348
 349         /*
 350          * If TRACE_REG_PERF_ADD returns false; no custom action was performed
 351          * and we need to take the default action of enqueueing our event on
 352          * the right per-cpu hlist.
 353          */
 354         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event)) {
 355                 struct hlist_head __percpu *pcpu_list;
 356                 struct hlist_head *list;
 357
 358                 pcpu_list = tp_event->perf_events;
 359                 if (WARN_ON_ONCE(!pcpu_list))
 360                         return -EINVAL;
 361
 362                 list = this_cpu_ptr(pcpu_list);
 363                 hlist_add_head_rcu(&p_event->hlist_entry, list);
 364         }
 365
 366         return 0;
 367 }
 368
 369 void perf_trace_del(struct perf_event *p_event, int flags)
 370 {
 371         struct trace_event_call *tp_event = p_event->tp_event;
 372
 373         /*
 374          * If TRACE_REG_PERF_DEL returns false; no custom action was performed
 375          * and we need to take the default action of dequeueing our event from
 376          * the right per-cpu hlist.
 377          */
 378         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event))
 379                 hlist_del_rcu(&p_event->hlist_entry);
 380 }
 381
 382 void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
 383 {
 384         char *raw_data;
 385         int rctx;
 386
 387         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 388
 389         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 390                       "perf buffer not large enough"))
 391                 return NULL;
 392
 393         *rctxp = rctx = perf_swevent_get_recursion_context();
 394         if (rctx < 0)
 395                 return NULL;
 396
 397         if (regs)
 398                 *regs = this_cpu_ptr(&__perf_regs[rctx]);
 399         raw_data = this_cpu_ptr(perf_trace_buf[rctx]);
 400
 401         /* zero the dead bytes from align to not leak stack to user */
 402         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 403         return raw_data;
 404 }
 405 EXPORT_SYMBOL_GPL(perf_trace_buf_alloc);
 406 NOKPROBE_SYMBOL(perf_trace_buf_alloc);
 407
 408 void perf_trace_buf_update(void *record, u16 type)
 409 {
 410         struct trace_entry *entry = record;
 411         int pc = preempt_count();
 412         unsigned long flags;
 413
 414         local_save_flags(flags);
 415         tracing_generic_entry_update(entry, flags, pc);
 416         entry->type = type;
 417 }
 418 NOKPROBE_SYMBOL(perf_trace_buf_update);
 419
 420 #ifdef CONFIG_FUNCTION_TRACER
 421 static void
 422 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 423                           struct ftrace_ops *ops, struct pt_regs *pt_regs)
 424 {
 425         struct ftrace_entry *entry;
 426         struct perf_event *event;
 427         struct hlist_head head;
 428         struct pt_regs regs;
 429         int rctx;
 430
 431         if ((unsigned long)ops->private != smp_processor_id())
 432                 return;
 433
 434         event = container_of(ops, struct perf_event, ftrace_ops);
 435
 436         /*
 437          * @event->hlist entry is NULL (per INIT_HLIST_NODE), and all
 438          * the perf code does is hlist_for_each_entry_rcu(), so we can
 439          * get away with simply setting the @head.first pointer in order
 440          * to create a singular list.
 441          */
 442         head.first = &event->hlist_entry;
 443
 444 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 445                     sizeof(u64)) - sizeof(u32))
 446
 447         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 448
 449         memset(&regs, 0, sizeof(regs));
 450         perf_fetch_caller_regs(&regs);
 451
 452         entry = perf_trace_buf_alloc(ENTRY_SIZE, NULL, &rctx);
 453         if (!entry)
 454                 return;
 455
 456         entry->ip = ip;
 457         entry->parent_ip = parent_ip;
 458         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
 459                               1, &regs, &head, NULL);
 460
 461 #undef ENTRY_SIZE
 462 }
 463
 464 static int perf_ftrace_function_register(struct perf_event *event)
 465 {
 466         struct ftrace_ops *ops = &event->ftrace_ops;
 467
 468         ops->flags   = FTRACE_OPS_FL_RCU;
 469         ops->func    = perf_ftrace_function_call;
 470         ops->private = (void *)(unsigned long)nr_cpu_ids;
 471
 472         return register_ftrace_function(ops);
 473 }
 474
 475 static int perf_ftrace_function_unregister(struct perf_event *event)
 476 {
 477         struct ftrace_ops *ops = &event->ftrace_ops;
 478         int ret = unregister_ftrace_function(ops);
 479         ftrace_free_filter(ops);
 480         return ret;
 481 }
 482
 483 int perf_ftrace_event_register(struct trace_event_call *call,
 484                                enum trace_reg type, void *data)
 485 {
 486         struct perf_event *event = data;
 487
 488         switch (type) {
 489         case TRACE_REG_REGISTER:
 490         case TRACE_REG_UNREGISTER:
 491                 break;
 492         case TRACE_REG_PERF_REGISTER:
 493         case TRACE_REG_PERF_UNREGISTER:
 494                 return 0;
 495         case TRACE_REG_PERF_OPEN:
 496                 return perf_ftrace_function_register(data);
 497         case TRACE_REG_PERF_CLOSE:
 498                 return perf_ftrace_function_unregister(data);
 499         case TRACE_REG_PERF_ADD:
 500                 event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();
 501                 return 1;
 502         case TRACE_REG_PERF_DEL:
 503                 event->ftrace_ops.private = (void *)(unsigned long)nr_cpu_ids;
 504                 return 1;
 505         }
 506
 507         return -EINVAL;
 508 }
 509 #endif /* CONFIG_FUNCTION_TRACER */