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