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