kernel/trace/trace_syscalls.c

   1 #include <trace/syscall.h>
   2 #include <trace/events/syscalls.h>
   3 #include <linux/kernel.h>
   4 #include <linux/ftrace.h>
   5 #include <linux/perf_event.h>
   6 #include <asm/syscall.h>
   7
   8 #include "trace_output.h"
   9 #include "trace.h"
  10
  11 static DEFINE_MUTEX(syscall_trace_lock);
  12 static int sys_refcount_enter;
  13 static int sys_refcount_exit;
  14 static DECLARE_BITMAP(enabled_enter_syscalls, NR_syscalls);
  15 static DECLARE_BITMAP(enabled_exit_syscalls, NR_syscalls);
  16
  17 extern unsigned long __start_syscalls_metadata[];
  18 extern unsigned long __stop_syscalls_metadata[];
  19
  20 static struct syscall_metadata **syscalls_metadata;
  21
  22 static struct syscall_metadata *find_syscall_meta(unsigned long syscall)
  23 {
  24         struct syscall_metadata *start;
  25         struct syscall_metadata *stop;
  26         char str[KSYM_SYMBOL_LEN];
  27
  28
  29         start = (struct syscall_metadata *)__start_syscalls_metadata;
  30         stop = (struct syscall_metadata *)__stop_syscalls_metadata;
  31         kallsyms_lookup(syscall, NULL, NULL, NULL, str);
  32
  33         for ( ; start < stop; start++) {
  34                 /*
  35                  * Only compare after the "sys" prefix. Archs that use
  36                  * syscall wrappers may have syscalls symbols aliases prefixed
  37                  * with "SyS" instead of "sys", leading to an unwanted
  38                  * mismatch.
  39                  */
  40                 if (start->name && !strcmp(start->name + 3, str + 3))
  41                         return start;
  42         }
  43         return NULL;
  44 }
  45
  46 static struct syscall_metadata *syscall_nr_to_meta(int nr)
  47 {
  48         if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
  49                 return NULL;
  50
  51         return syscalls_metadata[nr];
  52 }
  53
  54 int syscall_name_to_nr(char *name)
  55 {
  56         int i;
  57
  58         if (!syscalls_metadata)
  59                 return -1;
  60
  61         for (i = 0; i < NR_syscalls; i++) {
  62                 if (syscalls_metadata[i]) {
  63                         if (!strcmp(syscalls_metadata[i]->name, name))
  64                                 return i;
  65                 }
  66         }
  67         return -1;
  68 }
  69
  70 void set_syscall_enter_id(int num, int id)
  71 {
  72         syscalls_metadata[num]->enter_id = id;
  73 }
  74
  75 void set_syscall_exit_id(int num, int id)
  76 {
  77         syscalls_metadata[num]->exit_id = id;
  78 }
  79
  80 enum print_line_t
  81 print_syscall_enter(struct trace_iterator *iter, int flags)
  82 {
  83         struct trace_seq *s = &iter->seq;
  84         struct trace_entry *ent = iter->ent;
  85         struct syscall_trace_enter *trace;
  86         struct syscall_metadata *entry;
  87         int i, ret, syscall;
  88
  89         trace = (typeof(trace))ent;
  90         syscall = trace->nr;
  91         entry = syscall_nr_to_meta(syscall);
  92
  93         if (!entry)
  94                 goto end;
  95
  96         if (entry->enter_id != ent->type) {
  97                 WARN_ON_ONCE(1);
  98                 goto end;
  99         }
 100
 101         ret = trace_seq_printf(s, "%s(", entry->name);
 102         if (!ret)
 103                 return TRACE_TYPE_PARTIAL_LINE;
 104
 105         for (i = 0; i < entry->nb_args; i++) {
 106                 /* parameter types */
 107                 if (trace_flags & TRACE_ITER_VERBOSE) {
 108                         ret = trace_seq_printf(s, "%s ", entry->types[i]);
 109                         if (!ret)
 110                                 return TRACE_TYPE_PARTIAL_LINE;
 111                 }
 112                 /* parameter values */
 113                 ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
 114                                        trace->args[i],
 115                                        i == entry->nb_args - 1 ? "" : ", ");
 116                 if (!ret)
 117                         return TRACE_TYPE_PARTIAL_LINE;
 118         }
 119
 120         ret = trace_seq_putc(s, ')');
 121         if (!ret)
 122                 return TRACE_TYPE_PARTIAL_LINE;
 123
 124 end:
 125         ret =  trace_seq_putc(s, '\n');
 126         if (!ret)
 127                 return TRACE_TYPE_PARTIAL_LINE;
 128
 129         return TRACE_TYPE_HANDLED;
 130 }
 131
 132 enum print_line_t
 133 print_syscall_exit(struct trace_iterator *iter, int flags)
 134 {
 135         struct trace_seq *s = &iter->seq;
 136         struct trace_entry *ent = iter->ent;
 137         struct syscall_trace_exit *trace;
 138         int syscall;
 139         struct syscall_metadata *entry;
 140         int ret;
 141
 142         trace = (typeof(trace))ent;
 143         syscall = trace->nr;
 144         entry = syscall_nr_to_meta(syscall);
 145
 146         if (!entry) {
 147                 trace_seq_printf(s, "\n");
 148                 return TRACE_TYPE_HANDLED;
 149         }
 150
 151         if (entry->exit_id != ent->type) {
 152                 WARN_ON_ONCE(1);
 153                 return TRACE_TYPE_UNHANDLED;
 154         }
 155
 156         ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
 157                                 trace->ret);
 158         if (!ret)
 159                 return TRACE_TYPE_PARTIAL_LINE;
 160
 161         return TRACE_TYPE_HANDLED;
 162 }
 163
 164 extern char *__bad_type_size(void);
 165
 166 #define SYSCALL_FIELD(type, name)                                       \
 167         sizeof(type) != sizeof(trace.name) ?                            \
 168                 __bad_type_size() :                                     \
 169                 #type, #name, offsetof(typeof(trace), name),            \
 170                 sizeof(trace.name), is_signed_type(type)
 171
 172 int syscall_enter_format(struct ftrace_event_call *call, struct trace_seq *s)
 173 {
 174         int i;
 175         int nr;
 176         int ret;
 177         struct syscall_metadata *entry;
 178         struct syscall_trace_enter trace;
 179         int offset = offsetof(struct syscall_trace_enter, args);
 180
 181         nr = syscall_name_to_nr(call->data);
 182         entry = syscall_nr_to_meta(nr);
 183
 184         if (!entry)
 185                 return 0;
 186
 187         ret = trace_seq_printf(s, "\tfield:%s %s;\toffset:%zu;\tsize:%zu;"
 188                                "\tsigned:%u;\n",
 189                                SYSCALL_FIELD(int, nr));
 190         if (!ret)
 191                 return 0;
 192
 193         for (i = 0; i < entry->nb_args; i++) {
 194                 ret = trace_seq_printf(s, "\tfield:%s %s;", entry->types[i],
 195                                         entry->args[i]);
 196                 if (!ret)
 197                         return 0;
 198                 ret = trace_seq_printf(s, "\toffset:%d;\tsize:%zu;"
 199                                        "\tsigned:%u;\n", offset,
 200                                        sizeof(unsigned long),
 201                                        is_signed_type(unsigned long));
 202                 if (!ret)
 203                         return 0;
 204                 offset += sizeof(unsigned long);
 205         }
 206
 207         trace_seq_puts(s, "\nprint fmt: \"");
 208         for (i = 0; i < entry->nb_args; i++) {
 209                 ret = trace_seq_printf(s, "%s: 0x%%0%zulx%s", entry->args[i],
 210                                         sizeof(unsigned long),
 211                                         i == entry->nb_args - 1 ? "" : ", ");
 212                 if (!ret)
 213                         return 0;
 214         }
 215         trace_seq_putc(s, '"');
 216
 217         for (i = 0; i < entry->nb_args; i++) {
 218                 ret = trace_seq_printf(s, ", ((unsigned long)(REC->%s))",
 219                                        entry->args[i]);
 220                 if (!ret)
 221                         return 0;
 222         }
 223
 224         return trace_seq_putc(s, '\n');
 225 }
 226
 227 int syscall_exit_format(struct ftrace_event_call *call, struct trace_seq *s)
 228 {
 229         int ret;
 230         struct syscall_trace_exit trace;
 231
 232         ret = trace_seq_printf(s,
 233                                "\tfield:%s %s;\toffset:%zu;\tsize:%zu;"
 234                                "\tsigned:%u;\n"
 235                                "\tfield:%s %s;\toffset:%zu;\tsize:%zu;"
 236                                "\tsigned:%u;\n",
 237                                SYSCALL_FIELD(int, nr),
 238                                SYSCALL_FIELD(long, ret));
 239         if (!ret)
 240                 return 0;
 241
 242         return trace_seq_printf(s, "\nprint fmt: \"0x%%lx\", REC->ret\n");
 243 }
 244
 245 int syscall_enter_define_fields(struct ftrace_event_call *call)
 246 {
 247         struct syscall_trace_enter trace;
 248         struct syscall_metadata *meta;
 249         int ret;
 250         int nr;
 251         int i;
 252         int offset = offsetof(typeof(trace), args);
 253
 254         nr = syscall_name_to_nr(call->data);
 255         meta = syscall_nr_to_meta(nr);
 256
 257         if (!meta)
 258                 return 0;
 259
 260         ret = trace_define_common_fields(call);
 261         if (ret)
 262                 return ret;
 263
 264         ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
 265         if (ret)
 266                 return ret;
 267
 268         for (i = 0; i < meta->nb_args; i++) {
 269                 ret = trace_define_field(call, meta->types[i],
 270                                          meta->args[i], offset,
 271                                          sizeof(unsigned long), 0,
 272                                          FILTER_OTHER);
 273                 offset += sizeof(unsigned long);
 274         }
 275
 276         return ret;
 277 }
 278
 279 int syscall_exit_define_fields(struct ftrace_event_call *call)
 280 {
 281         struct syscall_trace_exit trace;
 282         int ret;
 283
 284         ret = trace_define_common_fields(call);
 285         if (ret)
 286                 return ret;
 287
 288         ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
 289         if (ret)
 290                 return ret;
 291
 292         ret = trace_define_field(call, SYSCALL_FIELD(long, ret),
 293                                  FILTER_OTHER);
 294
 295         return ret;
 296 }
 297
 298 void ftrace_syscall_enter(struct pt_regs *regs, long id)
 299 {
 300         struct syscall_trace_enter *entry;
 301         struct syscall_metadata *sys_data;
 302         struct ring_buffer_event *event;
 303         struct ring_buffer *buffer;
 304         int size;
 305         int syscall_nr;
 306
 307         syscall_nr = syscall_get_nr(current, regs);
 308         if (syscall_nr < 0)
 309                 return;
 310         if (!test_bit(syscall_nr, enabled_enter_syscalls))
 311                 return;
 312
 313         sys_data = syscall_nr_to_meta(syscall_nr);
 314         if (!sys_data)
 315                 return;
 316
 317         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 318
 319         event = trace_current_buffer_lock_reserve(&buffer, sys_data->enter_id,
 320                                                   size, 0, 0);
 321         if (!event)
 322                 return;
 323
 324         entry = ring_buffer_event_data(event);
 325         entry->nr = syscall_nr;
 326         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 327
 328         if (!filter_current_check_discard(buffer, sys_data->enter_event,
 329                                           entry, event))
 330                 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
 331 }
 332
 333 void ftrace_syscall_exit(struct pt_regs *regs, long ret)
 334 {
 335         struct syscall_trace_exit *entry;
 336         struct syscall_metadata *sys_data;
 337         struct ring_buffer_event *event;
 338         struct ring_buffer *buffer;
 339         int syscall_nr;
 340
 341         syscall_nr = syscall_get_nr(current, regs);
 342         if (syscall_nr < 0)
 343                 return;
 344         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 345                 return;
 346
 347         sys_data = syscall_nr_to_meta(syscall_nr);
 348         if (!sys_data)
 349                 return;
 350
 351         event = trace_current_buffer_lock_reserve(&buffer, sys_data->exit_id,
 352                                 sizeof(*entry), 0, 0);
 353         if (!event)
 354                 return;
 355
 356         entry = ring_buffer_event_data(event);
 357         entry->nr = syscall_nr;
 358         entry->ret = syscall_get_return_value(current, regs);
 359
 360         if (!filter_current_check_discard(buffer, sys_data->exit_event,
 361                                           entry, event))
 362                 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
 363 }
 364
 365 int reg_event_syscall_enter(struct ftrace_event_call *call)
 366 {
 367         int ret = 0;
 368         int num;
 369         char *name;
 370
 371         name = (char *)call->data;
 372         num = syscall_name_to_nr(name);
 373         if (num < 0 || num >= NR_syscalls)
 374                 return -ENOSYS;
 375         mutex_lock(&syscall_trace_lock);
 376         if (!sys_refcount_enter)
 377                 ret = register_trace_sys_enter(ftrace_syscall_enter);
 378         if (ret) {
 379                 pr_info("event trace: Could not activate"
 380                                 "syscall entry trace point");
 381         } else {
 382                 set_bit(num, enabled_enter_syscalls);
 383                 sys_refcount_enter++;
 384         }
 385         mutex_unlock(&syscall_trace_lock);
 386         return ret;
 387 }
 388
 389 void unreg_event_syscall_enter(struct ftrace_event_call *call)
 390 {
 391         int num;
 392         char *name;
 393
 394         name = (char *)call->data;
 395         num = syscall_name_to_nr(name);
 396         if (num < 0 || num >= NR_syscalls)
 397                 return;
 398         mutex_lock(&syscall_trace_lock);
 399         sys_refcount_enter--;
 400         clear_bit(num, enabled_enter_syscalls);
 401         if (!sys_refcount_enter)
 402                 unregister_trace_sys_enter(ftrace_syscall_enter);
 403         mutex_unlock(&syscall_trace_lock);
 404 }
 405
 406 int reg_event_syscall_exit(struct ftrace_event_call *call)
 407 {
 408         int ret = 0;
 409         int num;
 410         char *name;
 411
 412         name = call->data;
 413         num = syscall_name_to_nr(name);
 414         if (num < 0 || num >= NR_syscalls)
 415                 return -ENOSYS;
 416         mutex_lock(&syscall_trace_lock);
 417         if (!sys_refcount_exit)
 418                 ret = register_trace_sys_exit(ftrace_syscall_exit);
 419         if (ret) {
 420                 pr_info("event trace: Could not activate"
 421                                 "syscall exit trace point");
 422         } else {
 423                 set_bit(num, enabled_exit_syscalls);
 424                 sys_refcount_exit++;
 425         }
 426         mutex_unlock(&syscall_trace_lock);
 427         return ret;
 428 }
 429
 430 void unreg_event_syscall_exit(struct ftrace_event_call *call)
 431 {
 432         int num;
 433         char *name;
 434
 435         name = call->data;
 436         num = syscall_name_to_nr(name);
 437         if (num < 0 || num >= NR_syscalls)
 438                 return;
 439         mutex_lock(&syscall_trace_lock);
 440         sys_refcount_exit--;
 441         clear_bit(num, enabled_exit_syscalls);
 442         if (!sys_refcount_exit)
 443                 unregister_trace_sys_exit(ftrace_syscall_exit);
 444         mutex_unlock(&syscall_trace_lock);
 445 }
 446
 447 struct trace_event event_syscall_enter = {
 448         .trace                  = print_syscall_enter,
 449 };
 450
 451 struct trace_event event_syscall_exit = {
 452         .trace                  = print_syscall_exit,
 453 };
 454
 455 int __init init_ftrace_syscalls(void)
 456 {
 457         struct syscall_metadata *meta;
 458         unsigned long addr;
 459         int i;
 460
 461         syscalls_metadata = kzalloc(sizeof(*syscalls_metadata) *
 462                                         NR_syscalls, GFP_KERNEL);
 463         if (!syscalls_metadata) {
 464                 WARN_ON(1);
 465                 return -ENOMEM;
 466         }
 467
 468         for (i = 0; i < NR_syscalls; i++) {
 469                 addr = arch_syscall_addr(i);
 470                 meta = find_syscall_meta(addr);
 471                 syscalls_metadata[i] = meta;
 472         }
 473
 474         return 0;
 475 }
 476 core_initcall(init_ftrace_syscalls);
 477
 478 #ifdef CONFIG_EVENT_PROFILE
 479
 480 static DECLARE_BITMAP(enabled_prof_enter_syscalls, NR_syscalls);
 481 static DECLARE_BITMAP(enabled_prof_exit_syscalls, NR_syscalls);
 482 static int sys_prof_refcount_enter;
 483 static int sys_prof_refcount_exit;
 484
 485 static void prof_syscall_enter(struct pt_regs *regs, long id)
 486 {
 487         struct syscall_metadata *sys_data;
 488         struct syscall_trace_enter *rec;
 489         unsigned long flags;
 490         char *trace_buf;
 491         char *raw_data;
 492         int syscall_nr;
 493         int rctx;
 494         int size;
 495         int cpu;
 496
 497         syscall_nr = syscall_get_nr(current, regs);
 498         if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
 499                 return;
 500
 501         sys_data = syscall_nr_to_meta(syscall_nr);
 502         if (!sys_data)
 503                 return;
 504
 505         /* get the size after alignment with the u32 buffer size field */
 506         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 507         size = ALIGN(size + sizeof(u32), sizeof(u64));
 508         size -= sizeof(u32);
 509
 510         if (WARN_ONCE(size > FTRACE_MAX_PROFILE_SIZE,
 511                       "profile buffer not large enough"))
 512                 return;
 513
 514         /* Protect the per cpu buffer, begin the rcu read side */
 515         local_irq_save(flags);
 516
 517         rctx = perf_swevent_get_recursion_context();
 518         if (rctx < 0)
 519                 goto end_recursion;
 520
 521         cpu = smp_processor_id();
 522
 523         trace_buf = rcu_dereference(perf_trace_buf);
 524
 525         if (!trace_buf)
 526                 goto end;
 527
 528         raw_data = per_cpu_ptr(trace_buf, cpu);
 529
 530         /* zero the dead bytes from align to not leak stack to user */
 531         *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 532
 533         rec = (struct syscall_trace_enter *) raw_data;
 534         tracing_generic_entry_update(&rec->ent, 0, 0);
 535         rec->ent.type = sys_data->enter_id;
 536         rec->nr = syscall_nr;
 537         syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 538                                (unsigned long *)&rec->args);
 539         perf_tp_event(sys_data->enter_id, 0, 1, rec, size);
 540
 541 end:
 542         perf_swevent_put_recursion_context(rctx);
 543 end_recursion:
 544         local_irq_restore(flags);
 545 }
 546
 547 int reg_prof_syscall_enter(char *name)
 548 {
 549         int ret = 0;
 550         int num;
 551
 552         num = syscall_name_to_nr(name);
 553         if (num < 0 || num >= NR_syscalls)
 554                 return -ENOSYS;
 555
 556         mutex_lock(&syscall_trace_lock);
 557         if (!sys_prof_refcount_enter)
 558                 ret = register_trace_sys_enter(prof_syscall_enter);
 559         if (ret) {
 560                 pr_info("event trace: Could not activate"
 561                                 "syscall entry trace point");
 562         } else {
 563                 set_bit(num, enabled_prof_enter_syscalls);
 564                 sys_prof_refcount_enter++;
 565         }
 566         mutex_unlock(&syscall_trace_lock);
 567         return ret;
 568 }
 569
 570 void unreg_prof_syscall_enter(char *name)
 571 {
 572         int num;
 573
 574         num = syscall_name_to_nr(name);
 575         if (num < 0 || num >= NR_syscalls)
 576                 return;
 577
 578         mutex_lock(&syscall_trace_lock);
 579         sys_prof_refcount_enter--;
 580         clear_bit(num, enabled_prof_enter_syscalls);
 581         if (!sys_prof_refcount_enter)
 582                 unregister_trace_sys_enter(prof_syscall_enter);
 583         mutex_unlock(&syscall_trace_lock);
 584 }
 585
 586 static void prof_syscall_exit(struct pt_regs *regs, long ret)
 587 {
 588         struct syscall_metadata *sys_data;
 589         struct syscall_trace_exit *rec;
 590         unsigned long flags;
 591         int syscall_nr;
 592         char *trace_buf;
 593         char *raw_data;
 594         int rctx;
 595         int size;
 596         int cpu;
 597
 598         syscall_nr = syscall_get_nr(current, regs);
 599         if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
 600                 return;
 601
 602         sys_data = syscall_nr_to_meta(syscall_nr);
 603         if (!sys_data)
 604                 return;
 605
 606         /* We can probably do that at build time */
 607         size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
 608         size -= sizeof(u32);
 609
 610         /*
 611          * Impossible, but be paranoid with the future
 612          * How to put this check outside runtime?
 613          */
 614         if (WARN_ONCE(size > FTRACE_MAX_PROFILE_SIZE,
 615                 "exit event has grown above profile buffer size"))
 616                 return;
 617
 618         /* Protect the per cpu buffer, begin the rcu read side */
 619         local_irq_save(flags);
 620
 621         rctx = perf_swevent_get_recursion_context();
 622         if (rctx < 0)
 623                 goto end_recursion;
 624
 625         cpu = smp_processor_id();
 626
 627         trace_buf = rcu_dereference(perf_trace_buf);
 628
 629         if (!trace_buf)
 630                 goto end;
 631
 632         raw_data = per_cpu_ptr(trace_buf, cpu);
 633
 634         /* zero the dead bytes from align to not leak stack to user */
 635         *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 636
 637         rec = (struct syscall_trace_exit *)raw_data;
 638
 639         tracing_generic_entry_update(&rec->ent, 0, 0);
 640         rec->ent.type = sys_data->exit_id;
 641         rec->nr = syscall_nr;
 642         rec->ret = syscall_get_return_value(current, regs);
 643
 644         perf_tp_event(sys_data->exit_id, 0, 1, rec, size);
 645
 646 end:
 647         perf_swevent_put_recursion_context(rctx);
 648 end_recursion:
 649         local_irq_restore(flags);
 650 }
 651
 652 int reg_prof_syscall_exit(char *name)
 653 {
 654         int ret = 0;
 655         int num;
 656
 657         num = syscall_name_to_nr(name);
 658         if (num < 0 || num >= NR_syscalls)
 659                 return -ENOSYS;
 660
 661         mutex_lock(&syscall_trace_lock);
 662         if (!sys_prof_refcount_exit)
 663                 ret = register_trace_sys_exit(prof_syscall_exit);
 664         if (ret) {
 665                 pr_info("event trace: Could not activate"
 666                                 "syscall entry trace point");
 667         } else {
 668                 set_bit(num, enabled_prof_exit_syscalls);
 669                 sys_prof_refcount_exit++;
 670         }
 671         mutex_unlock(&syscall_trace_lock);
 672         return ret;
 673 }
 674
 675 void unreg_prof_syscall_exit(char *name)
 676 {
 677         int num;
 678
 679         num = syscall_name_to_nr(name);
 680         if (num < 0 || num >= NR_syscalls)
 681                 return;
 682
 683         mutex_lock(&syscall_trace_lock);
 684         sys_prof_refcount_exit--;
 685         clear_bit(num, enabled_prof_exit_syscalls);
 686         if (!sys_prof_refcount_exit)
 687                 unregister_trace_sys_exit(prof_syscall_exit);
 688         mutex_unlock(&syscall_trace_lock);
 689 }
 690
 691 #endif
 692
 693