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