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, FTRACE_SYSCALL_MAX);
  15 static DECLARE_BITMAP(enabled_exit_syscalls, FTRACE_SYSCALL_MAX);
  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 (!test_bit(syscall_nr, enabled_enter_syscalls))
 231                 return;
 232
 233         sys_data = syscall_nr_to_meta(syscall_nr);
 234         if (!sys_data)
 235                 return;
 236
 237         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 238
 239         event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
 240                                                         0, 0);
 241         if (!event)
 242                 return;
 243
 244         entry = ring_buffer_event_data(event);
 245         entry->nr = syscall_nr;
 246         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 247
 248         if (!filter_current_check_discard(sys_data->enter_event, entry, event))
 249                 trace_current_buffer_unlock_commit(event, 0, 0);
 250 }
 251
 252 void ftrace_syscall_exit(struct pt_regs *regs, long ret)
 253 {
 254         struct syscall_trace_exit *entry;
 255         struct syscall_metadata *sys_data;
 256         struct ring_buffer_event *event;
 257         int syscall_nr;
 258
 259         syscall_nr = syscall_get_nr(current, regs);
 260         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 261                 return;
 262
 263         sys_data = syscall_nr_to_meta(syscall_nr);
 264         if (!sys_data)
 265                 return;
 266
 267         event = trace_current_buffer_lock_reserve(sys_data->exit_id,
 268                                 sizeof(*entry), 0, 0);
 269         if (!event)
 270                 return;
 271
 272         entry = ring_buffer_event_data(event);
 273         entry->nr = syscall_nr;
 274         entry->ret = syscall_get_return_value(current, regs);
 275
 276         if (!filter_current_check_discard(sys_data->exit_event, entry, event))
 277                 trace_current_buffer_unlock_commit(event, 0, 0);
 278 }
 279
 280 int reg_event_syscall_enter(void *ptr)
 281 {
 282         int ret = 0;
 283         int num;
 284         char *name;
 285
 286         name = (char *)ptr;
 287         num = syscall_name_to_nr(name);
 288         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 289                 return -ENOSYS;
 290         mutex_lock(&syscall_trace_lock);
 291         if (!sys_refcount_enter)
 292                 ret = register_trace_sys_enter(ftrace_syscall_enter);
 293         if (ret) {
 294                 pr_info("event trace: Could not activate"
 295                                 "syscall entry trace point");
 296         } else {
 297                 set_bit(num, enabled_enter_syscalls);
 298                 sys_refcount_enter++;
 299         }
 300         mutex_unlock(&syscall_trace_lock);
 301         return ret;
 302 }
 303
 304 void unreg_event_syscall_enter(void *ptr)
 305 {
 306         int num;
 307         char *name;
 308
 309         name = (char *)ptr;
 310         num = syscall_name_to_nr(name);
 311         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 312                 return;
 313         mutex_lock(&syscall_trace_lock);
 314         sys_refcount_enter--;
 315         clear_bit(num, enabled_enter_syscalls);
 316         if (!sys_refcount_enter)
 317                 unregister_trace_sys_enter(ftrace_syscall_enter);
 318         mutex_unlock(&syscall_trace_lock);
 319 }
 320
 321 int reg_event_syscall_exit(void *ptr)
 322 {
 323         int ret = 0;
 324         int num;
 325         char *name;
 326
 327         name = (char *)ptr;
 328         num = syscall_name_to_nr(name);
 329         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 330                 return -ENOSYS;
 331         mutex_lock(&syscall_trace_lock);
 332         if (!sys_refcount_exit)
 333                 ret = register_trace_sys_exit(ftrace_syscall_exit);
 334         if (ret) {
 335                 pr_info("event trace: Could not activate"
 336                                 "syscall exit trace point");
 337         } else {
 338                 set_bit(num, enabled_exit_syscalls);
 339                 sys_refcount_exit++;
 340         }
 341         mutex_unlock(&syscall_trace_lock);
 342         return ret;
 343 }
 344
 345 void unreg_event_syscall_exit(void *ptr)
 346 {
 347         int num;
 348         char *name;
 349
 350         name = (char *)ptr;
 351         num = syscall_name_to_nr(name);
 352         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 353                 return;
 354         mutex_lock(&syscall_trace_lock);
 355         sys_refcount_exit--;
 356         clear_bit(num, enabled_exit_syscalls);
 357         if (!sys_refcount_exit)
 358                 unregister_trace_sys_exit(ftrace_syscall_exit);
 359         mutex_unlock(&syscall_trace_lock);
 360 }
 361
 362 struct trace_event event_syscall_enter = {
 363         .trace                  = print_syscall_enter,
 364 };
 365
 366 struct trace_event event_syscall_exit = {
 367         .trace                  = print_syscall_exit,
 368 };
 369
 370 #ifdef CONFIG_EVENT_PROFILE
 371
 372 static DECLARE_BITMAP(enabled_prof_enter_syscalls, FTRACE_SYSCALL_MAX);
 373 static DECLARE_BITMAP(enabled_prof_exit_syscalls, FTRACE_SYSCALL_MAX);
 374 static int sys_prof_refcount_enter;
 375 static int sys_prof_refcount_exit;
 376
 377 static void prof_syscall_enter(struct pt_regs *regs, long id)
 378 {
 379         struct syscall_trace_enter *rec;
 380         struct syscall_metadata *sys_data;
 381         int syscall_nr;
 382         int size;
 383
 384         syscall_nr = syscall_get_nr(current, regs);
 385         if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
 386                 return;
 387
 388         sys_data = syscall_nr_to_meta(syscall_nr);
 389         if (!sys_data)
 390                 return;
 391
 392         /* get the size after alignment with the u32 buffer size field */
 393         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 394         size = ALIGN(size + sizeof(u32), sizeof(u64));
 395         size -= sizeof(u32);
 396
 397         do {
 398                 char raw_data[size];
 399
 400                 /* zero the dead bytes from align to not leak stack to user */
 401                 *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 402
 403                 rec = (struct syscall_trace_enter *) raw_data;
 404                 tracing_generic_entry_update(&rec->ent, 0, 0);
 405                 rec->ent.type = sys_data->enter_id;
 406                 rec->nr = syscall_nr;
 407                 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 408                                        (unsigned long *)&rec->args);
 409                 perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
 410         } while(0);
 411 }
 412
 413 int reg_prof_syscall_enter(char *name)
 414 {
 415         int ret = 0;
 416         int num;
 417
 418         num = syscall_name_to_nr(name);
 419         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 420                 return -ENOSYS;
 421
 422         mutex_lock(&syscall_trace_lock);
 423         if (!sys_prof_refcount_enter)
 424                 ret = register_trace_sys_enter(prof_syscall_enter);
 425         if (ret) {
 426                 pr_info("event trace: Could not activate"
 427                                 "syscall entry trace point");
 428         } else {
 429                 set_bit(num, enabled_prof_enter_syscalls);
 430                 sys_prof_refcount_enter++;
 431         }
 432         mutex_unlock(&syscall_trace_lock);
 433         return ret;
 434 }
 435
 436 void unreg_prof_syscall_enter(char *name)
 437 {
 438         int num;
 439
 440         num = syscall_name_to_nr(name);
 441         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 442                 return;
 443
 444         mutex_lock(&syscall_trace_lock);
 445         sys_prof_refcount_enter--;
 446         clear_bit(num, enabled_prof_enter_syscalls);
 447         if (!sys_prof_refcount_enter)
 448                 unregister_trace_sys_enter(prof_syscall_enter);
 449         mutex_unlock(&syscall_trace_lock);
 450 }
 451
 452 static void prof_syscall_exit(struct pt_regs *regs, long ret)
 453 {
 454         struct syscall_metadata *sys_data;
 455         struct syscall_trace_exit rec;
 456         int syscall_nr;
 457
 458         syscall_nr = syscall_get_nr(current, regs);
 459         if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
 460                 return;
 461
 462         sys_data = syscall_nr_to_meta(syscall_nr);
 463         if (!sys_data)
 464                 return;
 465
 466         tracing_generic_entry_update(&rec.ent, 0, 0);
 467         rec.ent.type = sys_data->exit_id;
 468         rec.nr = syscall_nr;
 469         rec.ret = syscall_get_return_value(current, regs);
 470
 471         perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
 472 }
 473
 474 int reg_prof_syscall_exit(char *name)
 475 {
 476         int ret = 0;
 477         int num;
 478
 479         num = syscall_name_to_nr(name);
 480         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 481                 return -ENOSYS;
 482
 483         mutex_lock(&syscall_trace_lock);
 484         if (!sys_prof_refcount_exit)
 485                 ret = register_trace_sys_exit(prof_syscall_exit);
 486         if (ret) {
 487                 pr_info("event trace: Could not activate"
 488                                 "syscall entry trace point");
 489         } else {
 490                 set_bit(num, enabled_prof_exit_syscalls);
 491                 sys_prof_refcount_exit++;
 492         }
 493         mutex_unlock(&syscall_trace_lock);
 494         return ret;
 495 }
 496
 497 void unreg_prof_syscall_exit(char *name)
 498 {
 499         int num;
 500
 501         num = syscall_name_to_nr(name);
 502         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 503                 return;
 504
 505         mutex_lock(&syscall_trace_lock);
 506         sys_prof_refcount_exit--;
 507         clear_bit(num, enabled_prof_exit_syscalls);
 508         if (!sys_prof_refcount_exit)
 509                 unregister_trace_sys_exit(prof_syscall_exit);
 510         mutex_unlock(&syscall_trace_lock);
 511 }
 512
 513 #endif
 514
 515