kernel/trace/trace_event_perf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * trace event based perf event profiling/tracing
   4  *
   5  * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra
   6  * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com>
   7  */
   8
   9 #include <linux/module.h>
  10 #include <linux/kprobes.h>
  11 #include <linux/security.h>
  12 #include "trace.h"
  13 #include "trace_probe.h"
  14
  15 static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
  16
  17 /*
  18  * Force it to be aligned to unsigned long to avoid misaligned accesses
  19  * surprises
  20  */
  21 typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
  22         perf_trace_t;
  23
  24 /* Count the events in use (per event id, not per instance) */
  25 static int      total_ref_count;
  26
  27 static int perf_trace_event_perm(struct trace_event_call *tp_event,
  28                                  struct perf_event *p_event)
  29 {
  30         int ret;
  31
  32         if (tp_event->perf_perm) {
  33                 ret = tp_event->perf_perm(tp_event, p_event);
  34                 if (ret)
  35                         return ret;
  36         }
  37
  38         /*
  39          * We checked and allowed to create parent,
  40          * allow children without checking.
  41          */
  42         if (p_event->parent)
  43                 return 0;
  44
  45         /*
  46          * It's ok to check current process (owner) permissions in here,
  47          * because code below is called only via perf_event_open syscall.
  48          */
  49
  50         /* The ftrace function trace is allowed only for root. */
  51         if (ftrace_event_is_function(tp_event)) {
  52                 ret = perf_allow_tracepoint(&p_event->attr);
  53                 if (ret)
  54                         return ret;
  55
  56                 if (!is_sampling_event(p_event))
  57                         return 0;
  58
  59                 /*
  60                  * We don't allow user space callchains for  function trace
  61                  * event, due to issues with page faults while tracing page
  62                  * fault handler and its overall trickiness nature.
  63                  */
  64                 if (!p_event->attr.exclude_callchain_user)
  65                         return -EINVAL;
  66
  67                 /*
  68                  * Same reason to disable user stack dump as for user space
  69                  * callchains above.
  70                  */
  71                 if (p_event->attr.sample_type & PERF_SAMPLE_STACK_USER)
  72                         return -EINVAL;
  73         }
  74
  75         /* No tracing, just counting, so no obvious leak */
  76         if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW))
  77                 return 0;
  78
  79         /* Some events are ok to be traced by non-root users... */
  80         if (p_event->attach_state == PERF_ATTACH_TASK) {
  81                 if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY)
  82                         return 0;
  83         }
  84
  85         /*
  86          * ...otherwise raw tracepoint data can be a severe data leak,
  87          * only allow root to have these.
  88          */
  89         ret = perf_allow_tracepoint(&p_event->attr);
  90         if (ret)
  91                 return ret;
  92
  93         return 0;
  94 }
  95
  96 static int perf_trace_event_reg(struct trace_event_call *tp_event,
  97                                 struct perf_event *p_event)
  98 {
  99         struct hlist_head __percpu *list;
 100         int ret = -ENOMEM;
 101         int cpu;
 102
 103         p_event->tp_event = tp_event;
 104         if (tp_event->perf_refcount++ > 0)
 105                 return 0;
 106
 107         list = alloc_percpu(struct hlist_head);
 108         if (!list)
 109                 goto fail;
 110
 111         for_each_possible_cpu(cpu)
 112                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
 113
 114         tp_event->perf_events = list;
 115
 116         if (!total_ref_count) {
 117                 char __percpu *buf;
 118                 int i;
 119
 120                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 121                         buf = (char __percpu *)alloc_percpu(perf_trace_t);
 122                         if (!buf)
 123                                 goto fail;
 124
 125                         perf_trace_buf[i] = buf;
 126                 }
 127         }
 128
 129         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER, NULL);
 130         if (ret)
 131                 goto fail;
 132
 133         total_ref_count++;
 134         return 0;
 135
 136 fail:
 137         if (!total_ref_count) {
 138                 int i;
 139
 140                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 141                         free_percpu(perf_trace_buf[i]);
 142                         perf_trace_buf[i] = NULL;
 143                 }
 144         }
 145
 146         if (!--tp_event->perf_refcount) {
 147                 free_percpu(tp_event->perf_events);
 148                 tp_event->perf_events = NULL;
 149         }
 150
 151         return ret;
 152 }
 153
 154 static void perf_trace_event_unreg(struct perf_event *p_event)
 155 {
 156         struct trace_event_call *tp_event = p_event->tp_event;
 157         int i;
 158
 159         if (--tp_event->perf_refcount > 0)
 160                 goto out;
 161
 162         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER, NULL);
 163
 164         /*
 165          * Ensure our callback won't be called anymore. The buffers
 166          * will be freed after that.
 167          */
 168         tracepoint_synchronize_unregister();
 169
 170         free_percpu(tp_event->perf_events);
 171         tp_event->perf_events = NULL;
 172
 173         if (!--total_ref_count) {
 174                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
 175                         free_percpu(perf_trace_buf[i]);
 176                         perf_trace_buf[i] = NULL;
 177                 }
 178         }
 179 out:
 180         module_put(tp_event->mod);
 181 }
 182
 183 static int perf_trace_event_open(struct perf_event *p_event)
 184 {
 185         struct trace_event_call *tp_event = p_event->tp_event;
 186         return tp_event->class->reg(tp_event, TRACE_REG_PERF_OPEN, p_event);
 187 }
 188
 189 static void perf_trace_event_close(struct perf_event *p_event)
 190 {
 191         struct trace_event_call *tp_event = p_event->tp_event;
 192         tp_event->class->reg(tp_event, TRACE_REG_PERF_CLOSE, p_event);
 193 }
 194
 195 static int perf_trace_event_init(struct trace_event_call *tp_event,
 196                                  struct perf_event *p_event)
 197 {
 198         int ret;
 199
 200         ret = perf_trace_event_perm(tp_event, p_event);
 201         if (ret)
 202                 return ret;
 203
 204         ret = perf_trace_event_reg(tp_event, p_event);
 205         if (ret)
 206                 return ret;
 207
 208         ret = perf_trace_event_open(p_event);
 209         if (ret) {
 210                 perf_trace_event_unreg(p_event);
 211                 return ret;
 212         }
 213
 214         return 0;
 215 }
 216
 217 int perf_trace_init(struct perf_event *p_event)
 218 {
 219         struct trace_event_call *tp_event;
 220         u64 event_id = p_event->attr.config;
 221         int ret = -EINVAL;
 222
 223         mutex_lock(&event_mutex);
 224         list_for_each_entry(tp_event, &ftrace_events, list) {
 225                 if (tp_event->event.type == event_id &&
 226                     tp_event->class && tp_event->class->reg &&
 227                     try_module_get(tp_event->mod)) {
 228                         ret = perf_trace_event_init(tp_event, p_event);
 229                         if (ret)
 230                                 module_put(tp_event->mod);
 231                         break;
 232                 }
 233         }
 234         mutex_unlock(&event_mutex);
 235
 236         return ret;
 237 }
 238
 239 void perf_trace_destroy(struct perf_event *p_event)
 240 {
 241         mutex_lock(&event_mutex);
 242         perf_trace_event_close(p_event);
 243         perf_trace_event_unreg(p_event);
 244         mutex_unlock(&event_mutex);
 245 }
 246
 247 #ifdef CONFIG_KPROBE_EVENTS
 248 int perf_kprobe_init(struct perf_event *p_event, bool is_retprobe)
 249 {
 250         int ret;
 251         char *func = NULL;
 252         struct trace_event_call *tp_event;
 253
 254         if (p_event->attr.kprobe_func) {
 255                 func = kzalloc(KSYM_NAME_LEN, GFP_KERNEL);
 256                 if (!func)
 257                         return -ENOMEM;
 258                 ret = strncpy_from_user(
 259                         func, u64_to_user_ptr(p_event->attr.kprobe_func),
 260                         KSYM_NAME_LEN);
 261                 if (ret == KSYM_NAME_LEN)
 262                         ret = -E2BIG;
 263                 if (ret < 0)
 264                         goto out;
 265
 266                 if (func[0] == '\0') {
 267                         kfree(func);
 268                         func = NULL;
 269                 }
 270         }
 271
 272         tp_event = create_local_trace_kprobe(
 273                 func, (void *)(unsigned long)(p_event->attr.kprobe_addr),
 274                 p_event->attr.probe_offset, is_retprobe);
 275         if (IS_ERR(tp_event)) {
 276                 ret = PTR_ERR(tp_event);
 277                 goto out;
 278         }
 279
 280         mutex_lock(&event_mutex);
 281         ret = perf_trace_event_init(tp_event, p_event);
 282         if (ret)
 283                 destroy_local_trace_kprobe(tp_event);
 284         mutex_unlock(&event_mutex);
 285 out:
 286         kfree(func);
 287         return ret;
 288 }
 289
 290 void perf_kprobe_destroy(struct perf_event *p_event)
 291 {
 292         mutex_lock(&event_mutex);
 293         perf_trace_event_close(p_event);
 294         perf_trace_event_unreg(p_event);
 295         mutex_unlock(&event_mutex);
 296
 297         destroy_local_trace_kprobe(p_event->tp_event);
 298 }
 299 #endif /* CONFIG_KPROBE_EVENTS */
 300
 301 #ifdef CONFIG_UPROBE_EVENTS
 302 int perf_uprobe_init(struct perf_event *p_event,
 303                      unsigned long ref_ctr_offset, bool is_retprobe)
 304 {
 305         int ret;
 306         char *path = NULL;
 307         struct trace_event_call *tp_event;
 308
 309         if (!p_event->attr.uprobe_path)
 310                 return -EINVAL;
 311
 312         path = strndup_user(u64_to_user_ptr(p_event->attr.uprobe_path),
 313                             PATH_MAX);
 314         if (IS_ERR(path)) {
 315                 ret = PTR_ERR(path);
 316                 return (ret == -EINVAL) ? -E2BIG : ret;
 317         }
 318         if (path[0] == '\0') {
 319                 ret = -EINVAL;
 320                 goto out;
 321         }
 322
 323         tp_event = create_local_trace_uprobe(path, p_event->attr.probe_offset,
 324                                              ref_ctr_offset, is_retprobe);
 325         if (IS_ERR(tp_event)) {
 326                 ret = PTR_ERR(tp_event);
 327                 goto out;
 328         }
 329
 330         /*
 331          * local trace_uprobe need to hold event_mutex to call
 332          * uprobe_buffer_enable() and uprobe_buffer_disable().
 333          * event_mutex is not required for local trace_kprobes.
 334          */
 335         mutex_lock(&event_mutex);
 336         ret = perf_trace_event_init(tp_event, p_event);
 337         if (ret)
 338                 destroy_local_trace_uprobe(tp_event);
 339         mutex_unlock(&event_mutex);
 340 out:
 341         kfree(path);
 342         return ret;
 343 }
 344
 345 void perf_uprobe_destroy(struct perf_event *p_event)
 346 {
 347         mutex_lock(&event_mutex);
 348         perf_trace_event_close(p_event);
 349         perf_trace_event_unreg(p_event);
 350         mutex_unlock(&event_mutex);
 351         destroy_local_trace_uprobe(p_event->tp_event);
 352 }
 353 #endif /* CONFIG_UPROBE_EVENTS */
 354
 355 int perf_trace_add(struct perf_event *p_event, int flags)
 356 {
 357         struct trace_event_call *tp_event = p_event->tp_event;
 358
 359         if (!(flags & PERF_EF_START))
 360                 p_event->hw.state = PERF_HES_STOPPED;
 361
 362         /*
 363          * If TRACE_REG_PERF_ADD returns false; no custom action was performed
 364          * and we need to take the default action of enqueueing our event on
 365          * the right per-cpu hlist.
 366          */
 367         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_ADD, p_event)) {
 368                 struct hlist_head __percpu *pcpu_list;
 369                 struct hlist_head *list;
 370
 371                 pcpu_list = tp_event->perf_events;
 372                 if (WARN_ON_ONCE(!pcpu_list))
 373                         return -EINVAL;
 374
 375                 list = this_cpu_ptr(pcpu_list);
 376                 hlist_add_head_rcu(&p_event->hlist_entry, list);
 377         }
 378
 379         return 0;
 380 }
 381
 382 void perf_trace_del(struct perf_event *p_event, int flags)
 383 {
 384         struct trace_event_call *tp_event = p_event->tp_event;
 385
 386         /*
 387          * If TRACE_REG_PERF_DEL returns false; no custom action was performed
 388          * and we need to take the default action of dequeueing our event from
 389          * the right per-cpu hlist.
 390          */
 391         if (!tp_event->class->reg(tp_event, TRACE_REG_PERF_DEL, p_event))
 392                 hlist_del_rcu(&p_event->hlist_entry);
 393 }
 394
 395 void *perf_trace_buf_alloc(int size, struct pt_regs **regs, int *rctxp)
 396 {
 397         char *raw_data;
 398         int rctx;
 399
 400         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
 401
 402         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 403                       "perf buffer not large enough"))
 404                 return NULL;
 405
 406         *rctxp = rctx = perf_swevent_get_recursion_context();
 407         if (rctx < 0)
 408                 return NULL;
 409
 410         if (regs)
 411                 *regs = this_cpu_ptr(&__perf_regs[rctx]);
 412         raw_data = this_cpu_ptr(perf_trace_buf[rctx]);
 413
 414         /* zero the dead bytes from align to not leak stack to user */
 415         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
 416         return raw_data;
 417 }
 418 EXPORT_SYMBOL_GPL(perf_trace_buf_alloc);
 419 NOKPROBE_SYMBOL(perf_trace_buf_alloc);
 420
 421 void perf_trace_buf_update(void *record, u16 type)
 422 {
 423         struct trace_entry *entry = record;
 424
 425         tracing_generic_entry_update(entry, type, tracing_gen_ctx());
 426 }
 427 NOKPROBE_SYMBOL(perf_trace_buf_update);
 428
 429 #ifdef CONFIG_FUNCTION_TRACER
 430 static void
 431 perf_ftrace_function_call(unsigned long ip, unsigned long parent_ip,
 432                           struct ftrace_ops *ops,  struct ftrace_regs *fregs)
 433 {
 434         struct ftrace_entry *entry;
 435         struct perf_event *event;
 436         struct hlist_head head;
 437         struct pt_regs regs;
 438         int rctx;
 439         int bit;
 440
 441         if (!rcu_is_watching())
 442                 return;
 443
 444         if ((unsigned long)ops->private != smp_processor_id())
 445                 return;
 446
 447         bit = ftrace_test_recursion_trylock(ip, parent_ip);
 448         if (bit < 0)
 449                 return;
 450
 451         event = container_of(ops, struct perf_event, ftrace_ops);
 452
 453         /*
 454          * @event->hlist entry is NULL (per INIT_HLIST_NODE), and all
 455          * the perf code does is hlist_for_each_entry_rcu(), so we can
 456          * get away with simply setting the @head.first pointer in order
 457          * to create a singular list.
 458          */
 459         head.first = &event->hlist_entry;
 460
 461 #define ENTRY_SIZE (ALIGN(sizeof(struct ftrace_entry) + sizeof(u32), \
 462                     sizeof(u64)) - sizeof(u32))
 463
 464         BUILD_BUG_ON(ENTRY_SIZE > PERF_MAX_TRACE_SIZE);
 465
 466         memset(&regs, 0, sizeof(regs));
 467         perf_fetch_caller_regs(&regs);
 468
 469         entry = perf_trace_buf_alloc(ENTRY_SIZE, NULL, &rctx);
 470         if (!entry)
 471                 goto out;
 472
 473         entry->ip = ip;
 474         entry->parent_ip = parent_ip;
 475         perf_trace_buf_submit(entry, ENTRY_SIZE, rctx, TRACE_FN,
 476                               1, &regs, &head, NULL);
 477
 478 out:
 479         ftrace_test_recursion_unlock(bit);
 480 #undef ENTRY_SIZE
 481 }
 482
 483 static int perf_ftrace_function_register(struct perf_event *event)
 484 {
 485         struct ftrace_ops *ops = &event->ftrace_ops;
 486
 487         ops->func    = perf_ftrace_function_call;
 488         ops->private = (void *)(unsigned long)nr_cpu_ids;
 489
 490         return register_ftrace_function(ops);
 491 }
 492
 493 static int perf_ftrace_function_unregister(struct perf_event *event)
 494 {
 495         struct ftrace_ops *ops = &event->ftrace_ops;
 496         int ret = unregister_ftrace_function(ops);
 497         ftrace_free_filter(ops);
 498         return ret;
 499 }
 500
 501 int perf_ftrace_event_register(struct trace_event_call *call,
 502                                enum trace_reg type, void *data)
 503 {
 504         struct perf_event *event = data;
 505
 506         switch (type) {
 507         case TRACE_REG_REGISTER:
 508         case TRACE_REG_UNREGISTER:
 509                 break;
 510         case TRACE_REG_PERF_REGISTER:
 511         case TRACE_REG_PERF_UNREGISTER:
 512                 return 0;
 513         case TRACE_REG_PERF_OPEN:
 514                 return perf_ftrace_function_register(data);
 515         case TRACE_REG_PERF_CLOSE:
 516                 return perf_ftrace_function_unregister(data);
 517         case TRACE_REG_PERF_ADD:
 518                 event->ftrace_ops.private = (void *)(unsigned long)smp_processor_id();
 519                 return 1;
 520         case TRACE_REG_PERF_DEL:
 521                 event->ftrace_ops.private = (void *)(unsigned long)nr_cpu_ids;
 522                 return 1;
 523         }
 524
 525         return -EINVAL;
 526 }
 527 #endif /* CONFIG_FUNCTION_TRACER */