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 == KSYM_NAME_LEN)
 256                         ret = -E2BIG;
 257                 if (ret < 0)
 258                         goto out;
 259
 260                 if (func[0] == '\0') {
 261                         kfree(func);
 262                         func = NULL;
 263                 }
 264         }
 265
 266         tp_event = create_local_trace_kprobe(
 267                 func, (void *)(unsigned long)(p_event->attr.kprobe_addr),
 268                 p_event->attr.probe_offset, is_retprobe);
 269         if (IS_ERR(tp_event)) {
 270                 ret = PTR_ERR(tp_event);
 271                 goto out;
 272         }
 273
 274         ret = perf_trace_event_init(tp_event, p_event);
 275         if (ret)
 276                 destroy_local_trace_kprobe(tp_event);
 277 out:
 278         kfree(func);
 279         return ret;
 280 }
 281
 282 void perf_kprobe_destroy(struct perf_event *p_event)
 283 {
 284         perf_trace_event_close(p_event);
 285         perf_trace_event_unreg(p_event);
 286
 287         destroy_local_trace_kprobe(p_event->tp_event);
 288 }
 289 #endif /* CONFIG_KPROBE_EVENTS */
 290
 291 #ifdef CONFIG_UPROBE_EVENTS
 292 int perf_uprobe_init(struct perf_event *p_event, bool is_retprobe)
 293 {
 294         int ret;
 295         char *path = NULL;
 296         struct trace_event_call *tp_event;
 297
 298         if (!p_event->attr.uprobe_path)
 299                 return -EINVAL;
 300         path = kzalloc(PATH_MAX, GFP_KERNEL);
 301         if (!path)
 302                 return -ENOMEM;
 303         ret = strncpy_from_user(
 304                 path, u64_to_user_ptr(p_event->attr.uprobe_path), PATH_MAX);
 305         if (ret == PATH_MAX)
 306                 return -E2BIG;
 307         if (ret < 0)
 308                 goto out;
 309         if (path[0] == '\0') {
 310                 ret = -EINVAL;
 311                 goto out;
 312         }
 313
 314         tp_event = create_local_trace_uprobe(
 315                 path, p_event->attr.probe_offset, is_retprobe);
 316         if (IS_ERR(tp_event)) {
 317                 ret = PTR_ERR(tp_event);
 318                 goto out;
 319         }
 320
 321         /*
 322          * local trace_uprobe need to hold event_mutex to call
 323          * uprobe_buffer_enable() and uprobe_buffer_disable().
 324          * event_mutex is not required for local trace_kprobes.
 325          */
 326         mutex_lock(&event_mutex);
 327         ret = perf_trace_event_init(tp_event, p_event);
 328         if (ret)
 329                 destroy_local_trace_uprobe(tp_event);
 330         mutex_unlock(&event_mutex);
 331 out:
 332         kfree(path);
 333         return ret;
 334 }
 335
 336 void perf_uprobe_destroy(struct perf_event *p_event)
 337 {
 338         mutex_lock(&event_mutex);
 339         perf_trace_event_close(p_event);
 340         perf_trace_event_unreg(p_event);
 341         mutex_unlock(&event_mutex);
 342         destroy_local_trace_uprobe(p_event->tp_event);
 343 }
 344 #endif /* CONFIG_UPROBE_EVENTS */
 345
 346 int perf_trace_add(struct perf_event *p_event, int flags)
 347 {
 348         struct trace_event_call *tp_event = p_event->tp_event;
 349
 350         if (!(flags & PERF_EF_START))
 351                 p_event->hw.state = PERF_HES_STOPPED;
 352
 353         /*
 354          * If TRACE_REG_PERF_ADD returns false; no custom action was performed
 355          * and we need to take the default action of enqueueing our event on
 356          * the right per-cpu hlist.
 357          */
 358         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event)) {
 359                 struct hlist_head __percpu *pcpu_list;
 360                 struct hlist_head *list;
 361
 362                 pcpu_list = tp_event->perf_events;
 363                 if (WARN_ON_ONCE(!pcpu_list))
 364                         return -EINVAL;
 365
 366                 list = this_cpu_ptr(pcpu_list);
 367                 hlist_add_head_rcu(&p_event->hlist_entry, list);
 368         }
 369
 370         return 0;
 371 }
 372
 373 void perf_trace_del(struct perf_event *p_event, int flags)
 374 {
 375         struct trace_event_call *tp_event = p_event->tp_event;
 376
 377         /*
 378          * If TRACE_REG_PERF_DEL returns false; no custom action was performed
 379          * and we need to take the default action of dequeueing our event from
 380          * the right per-cpu hlist.
 381          */
 382         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event))
 383                 hlist_del_rcu(&p_event->hlist_entry);
 384 }
 385
 386 void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
 387 {
 388         char *raw_data;
 389         int rctx;
 390
 391         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 392
 393         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 394                       "perf buffer not large enough"))
 395                 return NULL;
 396
 397         *rctxp = rctx = perf_swevent_get_recursion_context();
 398         if (rctx < 0)
 399                 return NULL;
 400
 401         if (regs)
 402                 *regs = this_cpu_ptr(&__perf_regs[rctx]);
 403         raw_data = this_cpu_ptr(perf_trace_buf[rctx]);
 404
 405         /* zero the dead bytes from align to not leak stack to user */
 406         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 407         return raw_data;
 408 }
 409 EXPORT_SYMBOL_GPL(perf_trace_buf_alloc);
 410 NOKPROBE_SYMBOL(perf_trace_buf_alloc);
 411
 412 void perf_trace_buf_update(void *record, u16 type)
 413 {
 414         struct trace_entry *entry = record;
 415         int pc = preempt_count();
 416         unsigned long flags;
 417
 418         local_save_flags(flags);
 419         tracing_generic_entry_update(entry, flags, pc);
 420         entry->type = type;
 421 }
 422 NOKPROBE_SYMBOL(perf_trace_buf_update);
 423
 424 #ifdef CONFIG_FUNCTION_TRACER
 425 static void
 426 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 427                           struct ftrace_ops *ops, struct pt_regs *pt_regs)
 428 {
 429         struct ftrace_entry *entry;
 430         struct perf_event *event;
 431         struct hlist_head head;
 432         struct pt_regs regs;
 433         int rctx;
 434
 435         if ((unsigned long)ops->private != smp_processor_id())
 436                 return;
 437
 438         event = container_of(ops, struct perf_event, ftrace_ops);
 439
 440         /*
 441          * @event->hlist entry is NULL (per INIT_HLIST_NODE), and all
 442          * the perf code does is hlist_for_each_entry_rcu(), so we can
 443          * get away with simply setting the @head.first pointer in order
 444          * to create a singular list.
 445          */
 446         head.first = &event->hlist_entry;
 447
 448 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 449                     sizeof(u64)) - sizeof(u32))
 450
 451         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 452
 453         memset(&regs, 0, sizeof(regs));
 454         perf_fetch_caller_regs(&regs);
 455
 456         entry = perf_trace_buf_alloc(ENTRY_SIZE, NULL, &rctx);
 457         if (!entry)
 458                 return;
 459
 460         entry->ip = ip;
 461         entry->parent_ip = parent_ip;
 462         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
 463                               1, &regs, &head, NULL);
 464
 465 #undef ENTRY_SIZE
 466 }
 467
 468 static int perf_ftrace_function_register(struct perf_event *event)
 469 {
 470         struct ftrace_ops *ops = &event->ftrace_ops;
 471
 472         ops->flags   = FTRACE_OPS_FL_RCU;
 473         ops->func    = perf_ftrace_function_call;
 474         ops->private = (void *)(unsigned long)nr_cpu_ids;
 475
 476         return register_ftrace_function(ops);
 477 }
 478
 479 static int perf_ftrace_function_unregister(struct perf_event *event)
 480 {
 481         struct ftrace_ops *ops = &event->ftrace_ops;
 482         int ret = unregister_ftrace_function(ops);
 483         ftrace_free_filter(ops);
 484         return ret;
 485 }
 486
 487 int perf_ftrace_event_register(struct trace_event_call *call,
 488                                enum trace_reg type, void *data)
 489 {
 490         struct perf_event *event = data;
 491
 492         switch (type) {
 493         case TRACE_REG_REGISTER:
 494         case TRACE_REG_UNREGISTER:
 495                 break;
 496         case TRACE_REG_PERF_REGISTER:
 497         case TRACE_REG_PERF_UNREGISTER:
 498                 return 0;
 499         case TRACE_REG_PERF_OPEN:
 500                 return perf_ftrace_function_register(data);
 501         case TRACE_REG_PERF_CLOSE:
 502                 return perf_ftrace_function_unregister(data);
 503         case TRACE_REG_PERF_ADD:
 504                 event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();
 505                 return 1;
 506         case TRACE_REG_PERF_DEL:
 507                 event->ftrace_ops.private = (void *)(unsigned long)nr_cpu_ids;
 508                 return 1;
 509         }
 510
 511         return -EINVAL;
 512 }
 513 #endif /* CONFIG_FUNCTION_TRACER */