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