kernel/trace/trace_syscalls.c

   1 #include <trace/syscall.h>
   2 #include <trace/events/syscalls.h>
   3 #include <linux/slab.h>
   4 #include <linux/kernel.h>
   5 #include <linux/ftrace.h>
   6 #include <linux/perf_event.h>
   7 #include <asm/syscall.h>
   8
   9 #include "trace_output.h"
  10 #include "trace.h"
  11
  12 static DEFINE_MUTEX(syscall_trace_lock);
  13 static int sys_refcount_enter;
  14 static int sys_refcount_exit;
  15 static DECLARE_BITMAP(enabled_enter_syscalls, NR_syscalls);
  16 static DECLARE_BITMAP(enabled_exit_syscalls, NR_syscalls);
  17
  18 static int syscall_enter_register(struct ftrace_event_call *event,
  19                                  enum trace_reg type);
  20 static int syscall_exit_register(struct ftrace_event_call *event,
  21                                  enum trace_reg type);
  22
  23 struct ftrace_event_class event_class_syscall_enter = {
  24         .system                 = "syscalls",
  25         .reg                    = syscall_enter_register
  26 };
  27
  28 struct ftrace_event_class event_class_syscall_exit = {
  29         .system                 = "syscalls",
  30         .reg                    = syscall_exit_register
  31 };
  32
  33 extern unsigned long __start_syscalls_metadata[];
  34 extern unsigned long __stop_syscalls_metadata[];
  35
  36 static struct syscall_metadata **syscalls_metadata;
  37
  38 static struct syscall_metadata *find_syscall_meta(unsigned long syscall)
  39 {
  40         struct syscall_metadata *start;
  41         struct syscall_metadata *stop;
  42         char str[KSYM_SYMBOL_LEN];
  43
  44
  45         start = (struct syscall_metadata *)__start_syscalls_metadata;
  46         stop = (struct syscall_metadata *)__stop_syscalls_metadata;
  47         kallsyms_lookup(syscall, NULL, NULL, NULL, str);
  48
  49         for ( ; start < stop; start++) {
  50                 /*
  51                  * Only compare after the "sys" prefix. Archs that use
  52                  * syscall wrappers may have syscalls symbols aliases prefixed
  53                  * with "SyS" instead of "sys", leading to an unwanted
  54                  * mismatch.
  55                  */
  56                 if (start->name && !strcmp(start->name + 3, str + 3))
  57                         return start;
  58         }
  59         return NULL;
  60 }
  61
  62 static struct syscall_metadata *syscall_nr_to_meta(int nr)
  63 {
  64         if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
  65                 return NULL;
  66
  67         return syscalls_metadata[nr];
  68 }
  69
  70 enum print_line_t
  71 print_syscall_enter(struct trace_iterator *iter, int flags)
  72 {
  73         struct trace_seq *s = &iter->seq;
  74         struct trace_entry *ent = iter->ent;
  75         struct syscall_trace_enter *trace;
  76         struct syscall_metadata *entry;
  77         int i, ret, syscall;
  78
  79         trace = (typeof(trace))ent;
  80         syscall = trace->nr;
  81         entry = syscall_nr_to_meta(syscall);
  82
  83         if (!entry)
  84                 goto end;
  85
  86         if (entry->enter_event->id != ent->type) {
  87                 WARN_ON_ONCE(1);
  88                 goto end;
  89         }
  90
  91         ret = trace_seq_printf(s, "%s(", entry->name);
  92         if (!ret)
  93                 return TRACE_TYPE_PARTIAL_LINE;
  94
  95         for (i = 0; i < entry->nb_args; i++) {
  96                 /* parameter types */
  97                 if (trace_flags & TRACE_ITER_VERBOSE) {
  98                         ret = trace_seq_printf(s, "%s ", entry->types[i]);
  99                         if (!ret)
 100                                 return TRACE_TYPE_PARTIAL_LINE;
 101                 }
 102                 /* parameter values */
 103                 ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
 104                                        trace->args[i],
 105                                        i == entry->nb_args - 1 ? "" : ", ");
 106                 if (!ret)
 107                         return TRACE_TYPE_PARTIAL_LINE;
 108         }
 109
 110         ret = trace_seq_putc(s, ')');
 111         if (!ret)
 112                 return TRACE_TYPE_PARTIAL_LINE;
 113
 114 end:
 115         ret =  trace_seq_putc(s, '\n');
 116         if (!ret)
 117                 return TRACE_TYPE_PARTIAL_LINE;
 118
 119         return TRACE_TYPE_HANDLED;
 120 }
 121
 122 enum print_line_t
 123 print_syscall_exit(struct trace_iterator *iter, int flags)
 124 {
 125         struct trace_seq *s = &iter->seq;
 126         struct trace_entry *ent = iter->ent;
 127         struct syscall_trace_exit *trace;
 128         int syscall;
 129         struct syscall_metadata *entry;
 130         int ret;
 131
 132         trace = (typeof(trace))ent;
 133         syscall = trace->nr;
 134         entry = syscall_nr_to_meta(syscall);
 135
 136         if (!entry) {
 137                 trace_seq_printf(s, "\n");
 138                 return TRACE_TYPE_HANDLED;
 139         }
 140
 141         if (entry->exit_event->id != ent->type) {
 142                 WARN_ON_ONCE(1);
 143                 return TRACE_TYPE_UNHANDLED;
 144         }
 145
 146         ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
 147                                 trace->ret);
 148         if (!ret)
 149                 return TRACE_TYPE_PARTIAL_LINE;
 150
 151         return TRACE_TYPE_HANDLED;
 152 }
 153
 154 extern char *__bad_type_size(void);
 155
 156 #define SYSCALL_FIELD(type, name)                                       \
 157         sizeof(type) != sizeof(trace.name) ?                            \
 158                 __bad_type_size() :                                     \
 159                 #type, #name, offsetof(typeof(trace), name),            \
 160                 sizeof(trace.name), is_signed_type(type)
 161
 162 static
 163 int  __set_enter_print_fmt(struct syscall_metadata *entry, char *buf, int len)
 164 {
 165         int i;
 166         int pos = 0;
 167
 168         /* When len=0, we just calculate the needed length */
 169 #define LEN_OR_ZERO (len ? len - pos : 0)
 170
 171         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
 172         for (i = 0; i < entry->nb_args; i++) {
 173                 pos += snprintf(buf + pos, LEN_OR_ZERO, "%s: 0x%%0%zulx%s",
 174                                 entry->args[i], sizeof(unsigned long),
 175                                 i == entry->nb_args - 1 ? "" : ", ");
 176         }
 177         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
 178
 179         for (i = 0; i < entry->nb_args; i++) {
 180                 pos += snprintf(buf + pos, LEN_OR_ZERO,
 181                                 ", ((unsigned long)(REC->%s))", entry->args[i]);
 182         }
 183
 184 #undef LEN_OR_ZERO
 185
 186         /* return the length of print_fmt */
 187         return pos;
 188 }
 189
 190 static int set_syscall_print_fmt(struct ftrace_event_call *call)
 191 {
 192         char *print_fmt;
 193         int len;
 194         struct syscall_metadata *entry = call->data;
 195
 196         if (entry->enter_event != call) {
 197                 call->print_fmt = "\"0x%lx\", REC->ret";
 198                 return 0;
 199         }
 200
 201         /* First: called with 0 length to calculate the needed length */
 202         len = __set_enter_print_fmt(entry, NULL, 0);
 203
 204         print_fmt = kmalloc(len + 1, GFP_KERNEL);
 205         if (!print_fmt)
 206                 return -ENOMEM;
 207
 208         /* Second: actually write the @print_fmt */
 209         __set_enter_print_fmt(entry, print_fmt, len + 1);
 210         call->print_fmt = print_fmt;
 211
 212         return 0;
 213 }
 214
 215 static void free_syscall_print_fmt(struct ftrace_event_call *call)
 216 {
 217         struct syscall_metadata *entry = call->data;
 218
 219         if (entry->enter_event == call)
 220                 kfree(call->print_fmt);
 221 }
 222
 223 int syscall_enter_define_fields(struct ftrace_event_call *call)
 224 {
 225         struct syscall_trace_enter trace;
 226         struct syscall_metadata *meta = call->data;
 227         int ret;
 228         int i;
 229         int offset = offsetof(typeof(trace), args);
 230
 231         ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
 232         if (ret)
 233                 return ret;
 234
 235         for (i = 0; i < meta->nb_args; i++) {
 236                 ret = trace_define_field(call, meta->types[i],
 237                                          meta->args[i], offset,
 238                                          sizeof(unsigned long), 0,
 239                                          FILTER_OTHER);
 240                 offset += sizeof(unsigned long);
 241         }
 242
 243         return ret;
 244 }
 245
 246 int syscall_exit_define_fields(struct ftrace_event_call *call)
 247 {
 248         struct syscall_trace_exit trace;
 249         int ret;
 250
 251         ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
 252         if (ret)
 253                 return ret;
 254
 255         ret = trace_define_field(call, SYSCALL_FIELD(long, ret),
 256                                  FILTER_OTHER);
 257
 258         return ret;
 259 }
 260
 261 void ftrace_syscall_enter(void *ignore, struct pt_regs *regs, long id)
 262 {
 263         struct syscall_trace_enter *entry;
 264         struct syscall_metadata *sys_data;
 265         struct ring_buffer_event *event;
 266         struct ring_buffer *buffer;
 267         int size;
 268         int syscall_nr;
 269
 270         syscall_nr = syscall_get_nr(current, regs);
 271         if (syscall_nr < 0)
 272                 return;
 273         if (!test_bit(syscall_nr, enabled_enter_syscalls))
 274                 return;
 275
 276         sys_data = syscall_nr_to_meta(syscall_nr);
 277         if (!sys_data)
 278                 return;
 279
 280         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 281
 282         event = trace_current_buffer_lock_reserve(&buffer,
 283                         sys_data->enter_event->id, size, 0, 0);
 284         if (!event)
 285                 return;
 286
 287         entry = ring_buffer_event_data(event);
 288         entry->nr = syscall_nr;
 289         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 290
 291         if (!filter_current_check_discard(buffer, sys_data->enter_event,
 292                                           entry, event))
 293                 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
 294 }
 295
 296 void ftrace_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
 297 {
 298         struct syscall_trace_exit *entry;
 299         struct syscall_metadata *sys_data;
 300         struct ring_buffer_event *event;
 301         struct ring_buffer *buffer;
 302         int syscall_nr;
 303
 304         syscall_nr = syscall_get_nr(current, regs);
 305         if (syscall_nr < 0)
 306                 return;
 307         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 308                 return;
 309
 310         sys_data = syscall_nr_to_meta(syscall_nr);
 311         if (!sys_data)
 312                 return;
 313
 314         event = trace_current_buffer_lock_reserve(&buffer,
 315                         sys_data->exit_event->id, sizeof(*entry), 0, 0);
 316         if (!event)
 317                 return;
 318
 319         entry = ring_buffer_event_data(event);
 320         entry->nr = syscall_nr;
 321         entry->ret = syscall_get_return_value(current, regs);
 322
 323         if (!filter_current_check_discard(buffer, sys_data->exit_event,
 324                                           entry, event))
 325                 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
 326 }
 327
 328 int reg_event_syscall_enter(struct ftrace_event_call *call)
 329 {
 330         int ret = 0;
 331         int num;
 332
 333         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 334         if (num < 0 || num >= NR_syscalls)
 335                 return -ENOSYS;
 336         mutex_lock(&syscall_trace_lock);
 337         if (!sys_refcount_enter)
 338                 ret = register_trace_sys_enter(ftrace_syscall_enter, NULL);
 339         if (!ret) {
 340                 set_bit(num, enabled_enter_syscalls);
 341                 sys_refcount_enter++;
 342         }
 343         mutex_unlock(&syscall_trace_lock);
 344         return ret;
 345 }
 346
 347 void unreg_event_syscall_enter(struct ftrace_event_call *call)
 348 {
 349         int num;
 350
 351         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 352         if (num < 0 || num >= NR_syscalls)
 353                 return;
 354         mutex_lock(&syscall_trace_lock);
 355         sys_refcount_enter--;
 356         clear_bit(num, enabled_enter_syscalls);
 357         if (!sys_refcount_enter)
 358                 unregister_trace_sys_enter(ftrace_syscall_enter, NULL);
 359         mutex_unlock(&syscall_trace_lock);
 360 }
 361
 362 int reg_event_syscall_exit(struct ftrace_event_call *call)
 363 {
 364         int ret = 0;
 365         int num;
 366
 367         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 368         if (num < 0 || num >= NR_syscalls)
 369                 return -ENOSYS;
 370         mutex_lock(&syscall_trace_lock);
 371         if (!sys_refcount_exit)
 372                 ret = register_trace_sys_exit(ftrace_syscall_exit, NULL);
 373         if (!ret) {
 374                 set_bit(num, enabled_exit_syscalls);
 375                 sys_refcount_exit++;
 376         }
 377         mutex_unlock(&syscall_trace_lock);
 378         return ret;
 379 }
 380
 381 void unreg_event_syscall_exit(struct ftrace_event_call *call)
 382 {
 383         int num;
 384
 385         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 386         if (num < 0 || num >= NR_syscalls)
 387                 return;
 388         mutex_lock(&syscall_trace_lock);
 389         sys_refcount_exit--;
 390         clear_bit(num, enabled_exit_syscalls);
 391         if (!sys_refcount_exit)
 392                 unregister_trace_sys_exit(ftrace_syscall_exit, NULL);
 393         mutex_unlock(&syscall_trace_lock);
 394 }
 395
 396 int init_syscall_trace(struct ftrace_event_call *call)
 397 {
 398         int id;
 399
 400         if (set_syscall_print_fmt(call) < 0)
 401                 return -ENOMEM;
 402
 403         id = trace_event_raw_init(call);
 404
 405         if (id < 0) {
 406                 free_syscall_print_fmt(call);
 407                 return id;
 408         }
 409
 410         return id;
 411 }
 412
 413 unsigned long __init arch_syscall_addr(int nr)
 414 {
 415         return (unsigned long)sys_call_table[nr];
 416 }
 417
 418 int __init init_ftrace_syscalls(void)
 419 {
 420         struct syscall_metadata *meta;
 421         unsigned long addr;
 422         int i;
 423
 424         syscalls_metadata = kzalloc(sizeof(*syscalls_metadata) *
 425                                         NR_syscalls, GFP_KERNEL);
 426         if (!syscalls_metadata) {
 427                 WARN_ON(1);
 428                 return -ENOMEM;
 429         }
 430
 431         for (i = 0; i < NR_syscalls; i++) {
 432                 addr = arch_syscall_addr(i);
 433                 meta = find_syscall_meta(addr);
 434                 if (!meta)
 435                         continue;
 436
 437                 meta->syscall_nr = i;
 438                 syscalls_metadata[i] = meta;
 439         }
 440
 441         return 0;
 442 }
 443 core_initcall(init_ftrace_syscalls);
 444
 445 #ifdef CONFIG_PERF_EVENTS
 446
 447 static DECLARE_BITMAP(enabled_perf_enter_syscalls, NR_syscalls);
 448 static DECLARE_BITMAP(enabled_perf_exit_syscalls, NR_syscalls);
 449 static int sys_perf_refcount_enter;
 450 static int sys_perf_refcount_exit;
 451
 452 static void perf_syscall_enter(void *ignore, struct pt_regs *regs, long id)
 453 {
 454         struct syscall_metadata *sys_data;
 455         struct syscall_trace_enter *rec;
 456         unsigned long flags;
 457         int syscall_nr;
 458         int rctx;
 459         int size;
 460
 461         syscall_nr = syscall_get_nr(current, regs);
 462         if (!test_bit(syscall_nr, enabled_perf_enter_syscalls))
 463                 return;
 464
 465         sys_data = syscall_nr_to_meta(syscall_nr);
 466         if (!sys_data)
 467                 return;
 468
 469         /* get the size after alignment with the u32 buffer size field */
 470         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 471         size = ALIGN(size + sizeof(u32), sizeof(u64));
 472         size -= sizeof(u32);
 473
 474         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 475                       "perf buffer not large enough"))
 476                 return;
 477
 478         rec = (struct syscall_trace_enter *)perf_trace_buf_prepare(size,
 479                                 sys_data->enter_event->id, &rctx, &flags);
 480         if (!rec)
 481                 return;
 482
 483         rec->nr = syscall_nr;
 484         syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 485                                (unsigned long *)&rec->args);
 486         perf_trace_buf_submit(rec, size, rctx, 0, 1, flags, regs);
 487 }
 488
 489 int perf_sysenter_enable(struct ftrace_event_call *call)
 490 {
 491         int ret = 0;
 492         int num;
 493
 494         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 495
 496         mutex_lock(&syscall_trace_lock);
 497         if (!sys_perf_refcount_enter)
 498                 ret = register_trace_sys_enter(perf_syscall_enter, NULL);
 499         if (ret) {
 500                 pr_info("event trace: Could not activate"
 501                                 "syscall entry trace point");
 502         } else {
 503                 set_bit(num, enabled_perf_enter_syscalls);
 504                 sys_perf_refcount_enter++;
 505         }
 506         mutex_unlock(&syscall_trace_lock);
 507         return ret;
 508 }
 509
 510 void perf_sysenter_disable(struct ftrace_event_call *call)
 511 {
 512         int num;
 513
 514         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 515
 516         mutex_lock(&syscall_trace_lock);
 517         sys_perf_refcount_enter--;
 518         clear_bit(num, enabled_perf_enter_syscalls);
 519         if (!sys_perf_refcount_enter)
 520                 unregister_trace_sys_enter(perf_syscall_enter, NULL);
 521         mutex_unlock(&syscall_trace_lock);
 522 }
 523
 524 static void perf_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
 525 {
 526         struct syscall_metadata *sys_data;
 527         struct syscall_trace_exit *rec;
 528         unsigned long flags;
 529         int syscall_nr;
 530         int rctx;
 531         int size;
 532
 533         syscall_nr = syscall_get_nr(current, regs);
 534         if (!test_bit(syscall_nr, enabled_perf_exit_syscalls))
 535                 return;
 536
 537         sys_data = syscall_nr_to_meta(syscall_nr);
 538         if (!sys_data)
 539                 return;
 540
 541         /* We can probably do that at build time */
 542         size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
 543         size -= sizeof(u32);
 544
 545         /*
 546          * Impossible, but be paranoid with the future
 547          * How to put this check outside runtime?
 548          */
 549         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 550                 "exit event has grown above perf buffer size"))
 551                 return;
 552
 553         rec = (struct syscall_trace_exit *)perf_trace_buf_prepare(size,
 554                                 sys_data->exit_event->id, &rctx, &flags);
 555         if (!rec)
 556                 return;
 557
 558         rec->nr = syscall_nr;
 559         rec->ret = syscall_get_return_value(current, regs);
 560
 561         perf_trace_buf_submit(rec, size, rctx, 0, 1, flags, regs);
 562 }
 563
 564 int perf_sysexit_enable(struct ftrace_event_call *call)
 565 {
 566         int ret = 0;
 567         int num;
 568
 569         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 570
 571         mutex_lock(&syscall_trace_lock);
 572         if (!sys_perf_refcount_exit)
 573                 ret = register_trace_sys_exit(perf_syscall_exit, NULL);
 574         if (ret) {
 575                 pr_info("event trace: Could not activate"
 576                                 "syscall exit trace point");
 577         } else {
 578                 set_bit(num, enabled_perf_exit_syscalls);
 579                 sys_perf_refcount_exit++;
 580         }
 581         mutex_unlock(&syscall_trace_lock);
 582         return ret;
 583 }
 584
 585 void perf_sysexit_disable(struct ftrace_event_call *call)
 586 {
 587         int num;
 588
 589         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 590
 591         mutex_lock(&syscall_trace_lock);
 592         sys_perf_refcount_exit--;
 593         clear_bit(num, enabled_perf_exit_syscalls);
 594         if (!sys_perf_refcount_exit)
 595                 unregister_trace_sys_exit(perf_syscall_exit, NULL);
 596         mutex_unlock(&syscall_trace_lock);
 597 }
 598
 599 #endif /* CONFIG_PERF_EVENTS */
 600
 601 static int syscall_enter_register(struct ftrace_event_call *event,
 602                                  enum trace_reg type)
 603 {
 604         switch (type) {
 605         case TRACE_REG_REGISTER:
 606                 return reg_event_syscall_enter(event);
 607         case TRACE_REG_UNREGISTER:
 608                 unreg_event_syscall_enter(event);
 609                 return 0;
 610
 611 #ifdef CONFIG_PERF_EVENTS
 612         case TRACE_REG_PERF_REGISTER:
 613                 return perf_sysenter_enable(event);
 614         case TRACE_REG_PERF_UNREGISTER:
 615                 perf_sysenter_disable(event);
 616                 return 0;
 617 #endif
 618         }
 619         return 0;
 620 }
 621
 622 static int syscall_exit_register(struct ftrace_event_call *event,
 623                                  enum trace_reg type)
 624 {
 625         switch (type) {
 626         case TRACE_REG_REGISTER:
 627                 return reg_event_syscall_exit(event);
 628         case TRACE_REG_UNREGISTER:
 629                 unreg_event_syscall_exit(event);
 630                 return 0;
 631
 632 #ifdef CONFIG_PERF_EVENTS
 633         case TRACE_REG_PERF_REGISTER:
 634                 return perf_sysexit_enable(event);
 635         case TRACE_REG_PERF_UNREGISTER:
 636                 perf_sysexit_disable(event);
 637                 return 0;
 638 #endif
 639         }
 640         return 0;
 641 }