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_counter.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 enum print_line_t
  18 print_syscall_enter(struct trace_iterator *iter, int flags)
  19 {
  20         struct trace_seq *s = &iter->seq;
  21         struct trace_entry *ent = iter->ent;
  22         struct syscall_trace_enter *trace;
  23         struct syscall_metadata *entry;
  24         int i, ret, syscall;
  25
  26         trace = (typeof(trace))ent;
  27         syscall = trace->nr;
  28         entry = syscall_nr_to_meta(syscall);
  29
  30         if (!entry)
  31                 goto end;
  32
  33         if (entry->enter_id != ent->type) {
  34                 WARN_ON_ONCE(1);
  35                 goto end;
  36         }
  37
  38         ret = trace_seq_printf(s, "%s(", entry->name);
  39         if (!ret)
  40                 return TRACE_TYPE_PARTIAL_LINE;
  41
  42         for (i = 0; i < entry->nb_args; i++) {
  43                 /* parameter types */
  44                 if (trace_flags & TRACE_ITER_VERBOSE) {
  45                         ret = trace_seq_printf(s, "%s ", entry->types[i]);
  46                         if (!ret)
  47                                 return TRACE_TYPE_PARTIAL_LINE;
  48                 }
  49                 /* parameter values */
  50                 ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
  51                                        trace->args[i],
  52                                        i == entry->nb_args - 1 ? "" : ", ");
  53                 if (!ret)
  54                         return TRACE_TYPE_PARTIAL_LINE;
  55         }
  56
  57         ret = trace_seq_putc(s, ')');
  58         if (!ret)
  59                 return TRACE_TYPE_PARTIAL_LINE;
  60
  61 end:
  62         ret =  trace_seq_putc(s, '\n');
  63         if (!ret)
  64                 return TRACE_TYPE_PARTIAL_LINE;
  65
  66         return TRACE_TYPE_HANDLED;
  67 }
  68
  69 enum print_line_t
  70 print_syscall_exit(struct trace_iterator *iter, int flags)
  71 {
  72         struct trace_seq *s = &iter->seq;
  73         struct trace_entry *ent = iter->ent;
  74         struct syscall_trace_exit *trace;
  75         int syscall;
  76         struct syscall_metadata *entry;
  77         int ret;
  78
  79         trace = (typeof(trace))ent;
  80         syscall = trace->nr;
  81         entry = syscall_nr_to_meta(syscall);
  82
  83         if (!entry) {
  84                 trace_seq_printf(s, "\n");
  85                 return TRACE_TYPE_HANDLED;
  86         }
  87
  88         if (entry->exit_id != ent->type) {
  89                 WARN_ON_ONCE(1);
  90                 return TRACE_TYPE_UNHANDLED;
  91         }
  92
  93         ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
  94                                 trace->ret);
  95         if (!ret)
  96                 return TRACE_TYPE_PARTIAL_LINE;
  97
  98         return TRACE_TYPE_HANDLED;
  99 }
 100
 101 extern char *__bad_type_size(void);
 102
 103 #define SYSCALL_FIELD(type, name)                                       \
 104         sizeof(type) != sizeof(trace.name) ?                            \
 105                 __bad_type_size() :                                     \
 106                 #type, #name, offsetof(typeof(trace), name), sizeof(trace.name)
 107
 108 int syscall_enter_format(struct ftrace_event_call *call, struct trace_seq *s)
 109 {
 110         int i;
 111         int nr;
 112         int ret;
 113         struct syscall_metadata *entry;
 114         struct syscall_trace_enter trace;
 115         int offset = offsetof(struct syscall_trace_enter, args);
 116
 117         nr = syscall_name_to_nr(call->data);
 118         entry = syscall_nr_to_meta(nr);
 119
 120         if (!entry)
 121                 return 0;
 122
 123         ret = trace_seq_printf(s, "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
 124                                SYSCALL_FIELD(int, nr));
 125         if (!ret)
 126                 return 0;
 127
 128         for (i = 0; i < entry->nb_args; i++) {
 129                 ret = trace_seq_printf(s, "\tfield:%s %s;", entry->types[i],
 130                                         entry->args[i]);
 131                 if (!ret)
 132                         return 0;
 133                 ret = trace_seq_printf(s, "\toffset:%d;\tsize:%zu;\n", offset,
 134                                        sizeof(unsigned long));
 135                 if (!ret)
 136                         return 0;
 137                 offset += sizeof(unsigned long);
 138         }
 139
 140         trace_seq_puts(s, "\nprint fmt: \"");
 141         for (i = 0; i < entry->nb_args; i++) {
 142                 ret = trace_seq_printf(s, "%s: 0x%%0%zulx%s", entry->args[i],
 143                                         sizeof(unsigned long),
 144                                         i == entry->nb_args - 1 ? "" : ", ");
 145                 if (!ret)
 146                         return 0;
 147         }
 148         trace_seq_putc(s, '"');
 149
 150         for (i = 0; i < entry->nb_args; i++) {
 151                 ret = trace_seq_printf(s, ", ((unsigned long)(REC->%s))",
 152                                        entry->args[i]);
 153                 if (!ret)
 154                         return 0;
 155         }
 156
 157         return trace_seq_putc(s, '\n');
 158 }
 159
 160 int syscall_exit_format(struct ftrace_event_call *call, struct trace_seq *s)
 161 {
 162         int ret;
 163         struct syscall_trace_exit trace;
 164
 165         ret = trace_seq_printf(s,
 166                                "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n"
 167                                "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
 168                                SYSCALL_FIELD(int, nr),
 169                                SYSCALL_FIELD(unsigned long, ret));
 170         if (!ret)
 171                 return 0;
 172
 173         return trace_seq_printf(s, "\nprint fmt: \"0x%%lx\", REC->ret\n");
 174 }
 175
 176 int syscall_enter_define_fields(struct ftrace_event_call *call)
 177 {
 178         struct syscall_trace_enter trace;
 179         struct syscall_metadata *meta;
 180         int ret;
 181         int nr;
 182         int i;
 183         int offset = offsetof(typeof(trace), args);
 184
 185         nr = syscall_name_to_nr(call->data);
 186         meta = syscall_nr_to_meta(nr);
 187
 188         if (!meta)
 189                 return 0;
 190
 191         ret = trace_define_common_fields(call);
 192         if (ret)
 193                 return ret;
 194
 195         for (i = 0; i < meta->nb_args; i++) {
 196                 ret = trace_define_field(call, meta->types[i],
 197                                          meta->args[i], offset,
 198                                          sizeof(unsigned long), 0,
 199                                          FILTER_OTHER);
 200                 offset += sizeof(unsigned long);
 201         }
 202
 203         return ret;
 204 }
 205
 206 int syscall_exit_define_fields(struct ftrace_event_call *call)
 207 {
 208         struct syscall_trace_exit trace;
 209         int ret;
 210
 211         ret = trace_define_common_fields(call);
 212         if (ret)
 213                 return ret;
 214
 215         ret = trace_define_field(call, SYSCALL_FIELD(unsigned long, ret), 0,
 216                                  FILTER_OTHER);
 217
 218         return ret;
 219 }
 220
 221 void ftrace_syscall_enter(struct pt_regs *regs, long id)
 222 {
 223         struct syscall_trace_enter *entry;
 224         struct syscall_metadata *sys_data;
 225         struct ring_buffer_event *event;
 226         int size;
 227         int syscall_nr;
 228
 229         syscall_nr = syscall_get_nr(current, regs);
 230         if (syscall_nr < 0)
 231                 return;
 232         if (!test_bit(syscall_nr, enabled_enter_syscalls))
 233                 return;
 234
 235         sys_data = syscall_nr_to_meta(syscall_nr);
 236         if (!sys_data)
 237                 return;
 238
 239         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 240
 241         event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
 242                                                         0, 0);
 243         if (!event)
 244                 return;
 245
 246         entry = ring_buffer_event_data(event);
 247         entry->nr = syscall_nr;
 248         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 249
 250         if (!filter_current_check_discard(sys_data->enter_event, entry, event))
 251                 trace_current_buffer_unlock_commit(event, 0, 0);
 252 }
 253
 254 void ftrace_syscall_exit(struct pt_regs *regs, long ret)
 255 {
 256         struct syscall_trace_exit *entry;
 257         struct syscall_metadata *sys_data;
 258         struct ring_buffer_event *event;
 259         int syscall_nr;
 260
 261         syscall_nr = syscall_get_nr(current, regs);
 262         if (syscall_nr < 0)
 263                 return;
 264         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 265                 return;
 266
 267         sys_data = syscall_nr_to_meta(syscall_nr);
 268         if (!sys_data)
 269                 return;
 270
 271         event = trace_current_buffer_lock_reserve(sys_data->exit_id,
 272                                 sizeof(*entry), 0, 0);
 273         if (!event)
 274                 return;
 275
 276         entry = ring_buffer_event_data(event);
 277         entry->nr = syscall_nr;
 278         entry->ret = syscall_get_return_value(current, regs);
 279
 280         if (!filter_current_check_discard(sys_data->exit_event, entry, event))
 281                 trace_current_buffer_unlock_commit(event, 0, 0);
 282 }
 283
 284 int reg_event_syscall_enter(void *ptr)
 285 {
 286         int ret = 0;
 287         int num;
 288         char *name;
 289
 290         name = (char *)ptr;
 291         num = syscall_name_to_nr(name);
 292         if (num < 0 || num >= NR_syscalls)
 293                 return -ENOSYS;
 294         mutex_lock(&syscall_trace_lock);
 295         if (!sys_refcount_enter)
 296                 ret = register_trace_sys_enter(ftrace_syscall_enter);
 297         if (ret) {
 298                 pr_info("event trace: Could not activate"
 299                                 "syscall entry trace point");
 300         } else {
 301                 set_bit(num, enabled_enter_syscalls);
 302                 sys_refcount_enter++;
 303         }
 304         mutex_unlock(&syscall_trace_lock);
 305         return ret;
 306 }
 307
 308 void unreg_event_syscall_enter(void *ptr)
 309 {
 310         int num;
 311         char *name;
 312
 313         name = (char *)ptr;
 314         num = syscall_name_to_nr(name);
 315         if (num < 0 || num >= NR_syscalls)
 316                 return;
 317         mutex_lock(&syscall_trace_lock);
 318         sys_refcount_enter--;
 319         clear_bit(num, enabled_enter_syscalls);
 320         if (!sys_refcount_enter)
 321                 unregister_trace_sys_enter(ftrace_syscall_enter);
 322         mutex_unlock(&syscall_trace_lock);
 323 }
 324
 325 int reg_event_syscall_exit(void *ptr)
 326 {
 327         int ret = 0;
 328         int num;
 329         char *name;
 330
 331         name = (char *)ptr;
 332         num = syscall_name_to_nr(name);
 333         if (num < 0 || num >= NR_syscalls)
 334                 return -ENOSYS;
 335         mutex_lock(&syscall_trace_lock);
 336         if (!sys_refcount_exit)
 337                 ret = register_trace_sys_exit(ftrace_syscall_exit);
 338         if (ret) {
 339                 pr_info("event trace: Could not activate"
 340                                 "syscall exit trace point");
 341         } else {
 342                 set_bit(num, enabled_exit_syscalls);
 343                 sys_refcount_exit++;
 344         }
 345         mutex_unlock(&syscall_trace_lock);
 346         return ret;
 347 }
 348
 349 void unreg_event_syscall_exit(void *ptr)
 350 {
 351         int num;
 352         char *name;
 353
 354         name = (char *)ptr;
 355         num = syscall_name_to_nr(name);
 356         if (num < 0 || num >= NR_syscalls)
 357                 return;
 358         mutex_lock(&syscall_trace_lock);
 359         sys_refcount_exit--;
 360         clear_bit(num, enabled_exit_syscalls);
 361         if (!sys_refcount_exit)
 362                 unregister_trace_sys_exit(ftrace_syscall_exit);
 363         mutex_unlock(&syscall_trace_lock);
 364 }
 365
 366 struct trace_event event_syscall_enter = {
 367         .trace                  = print_syscall_enter,
 368 };
 369
 370 struct trace_event event_syscall_exit = {
 371         .trace                  = print_syscall_exit,
 372 };
 373
 374 #ifdef CONFIG_EVENT_PROFILE
 375
 376 static DECLARE_BITMAP(enabled_prof_enter_syscalls, NR_syscalls);
 377 static DECLARE_BITMAP(enabled_prof_exit_syscalls, NR_syscalls);
 378 static int sys_prof_refcount_enter;
 379 static int sys_prof_refcount_exit;
 380
 381 static void prof_syscall_enter(struct pt_regs *regs, long id)
 382 {
 383         struct syscall_trace_enter *rec;
 384         struct syscall_metadata *sys_data;
 385         int syscall_nr;
 386         int size;
 387
 388         syscall_nr = syscall_get_nr(current, regs);
 389         if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
 390                 return;
 391
 392         sys_data = syscall_nr_to_meta(syscall_nr);
 393         if (!sys_data)
 394                 return;
 395
 396         /* get the size after alignment with the u32 buffer size field */
 397         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 398         size = ALIGN(size + sizeof(u32), sizeof(u64));
 399         size -= sizeof(u32);
 400
 401         do {
 402                 char raw_data[size];
 403
 404                 /* zero the dead bytes from align to not leak stack to user */
 405                 *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 406
 407                 rec = (struct syscall_trace_enter *) raw_data;
 408                 tracing_generic_entry_update(&rec->ent, 0, 0);
 409                 rec->ent.type = sys_data->enter_id;
 410                 rec->nr = syscall_nr;
 411                 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 412                                        (unsigned long *)&rec->args);
 413                 perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
 414         } while(0);
 415 }
 416
 417 int reg_prof_syscall_enter(char *name)
 418 {
 419         int ret = 0;
 420         int num;
 421
 422         num = syscall_name_to_nr(name);
 423         if (num < 0 || num >= NR_syscalls)
 424                 return -ENOSYS;
 425
 426         mutex_lock(&syscall_trace_lock);
 427         if (!sys_prof_refcount_enter)
 428                 ret = register_trace_sys_enter(prof_syscall_enter);
 429         if (ret) {
 430                 pr_info("event trace: Could not activate"
 431                                 "syscall entry trace point");
 432         } else {
 433                 set_bit(num, enabled_prof_enter_syscalls);
 434                 sys_prof_refcount_enter++;
 435         }
 436         mutex_unlock(&syscall_trace_lock);
 437         return ret;
 438 }
 439
 440 void unreg_prof_syscall_enter(char *name)
 441 {
 442         int num;
 443
 444         num = syscall_name_to_nr(name);
 445         if (num < 0 || num >= NR_syscalls)
 446                 return;
 447
 448         mutex_lock(&syscall_trace_lock);
 449         sys_prof_refcount_enter--;
 450         clear_bit(num, enabled_prof_enter_syscalls);
 451         if (!sys_prof_refcount_enter)
 452                 unregister_trace_sys_enter(prof_syscall_enter);
 453         mutex_unlock(&syscall_trace_lock);
 454 }
 455
 456 static void prof_syscall_exit(struct pt_regs *regs, long ret)
 457 {
 458         struct syscall_metadata *sys_data;
 459         struct syscall_trace_exit rec;
 460         int syscall_nr;
 461
 462         syscall_nr = syscall_get_nr(current, regs);
 463         if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
 464                 return;
 465
 466         sys_data = syscall_nr_to_meta(syscall_nr);
 467         if (!sys_data)
 468                 return;
 469
 470         tracing_generic_entry_update(&rec.ent, 0, 0);
 471         rec.ent.type = sys_data->exit_id;
 472         rec.nr = syscall_nr;
 473         rec.ret = syscall_get_return_value(current, regs);
 474
 475         perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
 476 }
 477
 478 int reg_prof_syscall_exit(char *name)
 479 {
 480         int ret = 0;
 481         int num;
 482
 483         num = syscall_name_to_nr(name);
 484         if (num < 0 || num >= NR_syscalls)
 485                 return -ENOSYS;
 486
 487         mutex_lock(&syscall_trace_lock);
 488         if (!sys_prof_refcount_exit)
 489                 ret = register_trace_sys_exit(prof_syscall_exit);
 490         if (ret) {
 491                 pr_info("event trace: Could not activate"
 492                                 "syscall entry trace point");
 493         } else {
 494                 set_bit(num, enabled_prof_exit_syscalls);
 495                 sys_prof_refcount_exit++;
 496         }
 497         mutex_unlock(&syscall_trace_lock);
 498         return ret;
 499 }
 500
 501 void unreg_prof_syscall_exit(char *name)
 502 {
 503         int num;
 504
 505         num = syscall_name_to_nr(name);
 506         if (num < 0 || num >= NR_syscalls)
 507                 return;
 508
 509         mutex_lock(&syscall_trace_lock);
 510         sys_prof_refcount_exit--;
 511         clear_bit(num, enabled_prof_exit_syscalls);
 512         if (!sys_prof_refcount_exit)
 513                 unregister_trace_sys_exit(prof_syscall_exit);
 514         mutex_unlock(&syscall_trace_lock);
 515 }
 516
 517 #endif
 518
 519