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                 offset += sizeof(unsigned long);
 200         }
 201
 202         return ret;
 203 }
 204
 205 int syscall_exit_define_fields(struct ftrace_event_call *call)
 206 {
 207         struct syscall_trace_exit trace;
 208         int ret;
 209
 210         ret = trace_define_common_fields(call);
 211         if (ret)
 212                 return ret;
 213
 214         ret = trace_define_field(call, SYSCALL_FIELD(unsigned long, ret), 0);
 215
 216         return ret;
 217 }
 218
 219 void ftrace_syscall_enter(struct pt_regs *regs, long id)
 220 {
 221         struct syscall_trace_enter *entry;
 222         struct syscall_metadata *sys_data;
 223         struct ring_buffer_event *event;
 224         int size;
 225         int syscall_nr;
 226
 227         syscall_nr = syscall_get_nr(current, regs);
 228         if (!test_bit(syscall_nr, enabled_enter_syscalls))
 229                 return;
 230
 231         sys_data = syscall_nr_to_meta(syscall_nr);
 232         if (!sys_data)
 233                 return;
 234
 235         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 236
 237         event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
 238                                                         0, 0);
 239         if (!event)
 240                 return;
 241
 242         entry = ring_buffer_event_data(event);
 243         entry->nr = syscall_nr;
 244         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 245
 246         if (!filter_current_check_discard(sys_data->enter_event, entry, event))
 247                 trace_current_buffer_unlock_commit(event, 0, 0);
 248 }
 249
 250 void ftrace_syscall_exit(struct pt_regs *regs, long ret)
 251 {
 252         struct syscall_trace_exit *entry;
 253         struct syscall_metadata *sys_data;
 254         struct ring_buffer_event *event;
 255         int syscall_nr;
 256
 257         syscall_nr = syscall_get_nr(current, regs);
 258         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 259                 return;
 260
 261         sys_data = syscall_nr_to_meta(syscall_nr);
 262         if (!sys_data)
 263                 return;
 264
 265         event = trace_current_buffer_lock_reserve(sys_data->exit_id,
 266                                 sizeof(*entry), 0, 0);
 267         if (!event)
 268                 return;
 269
 270         entry = ring_buffer_event_data(event);
 271         entry->nr = syscall_nr;
 272         entry->ret = syscall_get_return_value(current, regs);
 273
 274         if (!filter_current_check_discard(sys_data->exit_event, entry, event))
 275                 trace_current_buffer_unlock_commit(event, 0, 0);
 276 }
 277
 278 int reg_event_syscall_enter(void *ptr)
 279 {
 280         int ret = 0;
 281         int num;
 282         char *name;
 283
 284         name = (char *)ptr;
 285         num = syscall_name_to_nr(name);
 286         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 287                 return -ENOSYS;
 288         mutex_lock(&syscall_trace_lock);
 289         if (!sys_refcount_enter)
 290                 ret = register_trace_sys_enter(ftrace_syscall_enter);
 291         if (ret) {
 292                 pr_info("event trace: Could not activate"
 293                                 "syscall entry trace point");
 294         } else {
 295                 set_bit(num, enabled_enter_syscalls);
 296                 sys_refcount_enter++;
 297         }
 298         mutex_unlock(&syscall_trace_lock);
 299         return ret;
 300 }
 301
 302 void unreg_event_syscall_enter(void *ptr)
 303 {
 304         int num;
 305         char *name;
 306
 307         name = (char *)ptr;
 308         num = syscall_name_to_nr(name);
 309         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 310                 return;
 311         mutex_lock(&syscall_trace_lock);
 312         sys_refcount_enter--;
 313         clear_bit(num, enabled_enter_syscalls);
 314         if (!sys_refcount_enter)
 315                 unregister_trace_sys_enter(ftrace_syscall_enter);
 316         mutex_unlock(&syscall_trace_lock);
 317 }
 318
 319 int reg_event_syscall_exit(void *ptr)
 320 {
 321         int ret = 0;
 322         int num;
 323         char *name;
 324
 325         name = (char *)ptr;
 326         num = syscall_name_to_nr(name);
 327         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 328                 return -ENOSYS;
 329         mutex_lock(&syscall_trace_lock);
 330         if (!sys_refcount_exit)
 331                 ret = register_trace_sys_exit(ftrace_syscall_exit);
 332         if (ret) {
 333                 pr_info("event trace: Could not activate"
 334                                 "syscall exit trace point");
 335         } else {
 336                 set_bit(num, enabled_exit_syscalls);
 337                 sys_refcount_exit++;
 338         }
 339         mutex_unlock(&syscall_trace_lock);
 340         return ret;
 341 }
 342
 343 void unreg_event_syscall_exit(void *ptr)
 344 {
 345         int num;
 346         char *name;
 347
 348         name = (char *)ptr;
 349         num = syscall_name_to_nr(name);
 350         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 351                 return;
 352         mutex_lock(&syscall_trace_lock);
 353         sys_refcount_exit--;
 354         clear_bit(num, enabled_exit_syscalls);
 355         if (!sys_refcount_exit)
 356                 unregister_trace_sys_exit(ftrace_syscall_exit);
 357         mutex_unlock(&syscall_trace_lock);
 358 }
 359
 360 struct trace_event event_syscall_enter = {
 361         .trace                  = print_syscall_enter,
 362 };
 363
 364 struct trace_event event_syscall_exit = {
 365         .trace                  = print_syscall_exit,
 366 };
 367
 368 #ifdef CONFIG_EVENT_PROFILE
 369
 370 static DECLARE_BITMAP(enabled_prof_enter_syscalls, FTRACE_SYSCALL_MAX);
 371 static DECLARE_BITMAP(enabled_prof_exit_syscalls, FTRACE_SYSCALL_MAX);
 372 static int sys_prof_refcount_enter;
 373 static int sys_prof_refcount_exit;
 374
 375 static void prof_syscall_enter(struct pt_regs *regs, long id)
 376 {
 377         struct syscall_trace_enter *rec;
 378         struct syscall_metadata *sys_data;
 379         int syscall_nr;
 380         int size;
 381
 382         syscall_nr = syscall_get_nr(current, regs);
 383         if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
 384                 return;
 385
 386         sys_data = syscall_nr_to_meta(syscall_nr);
 387         if (!sys_data)
 388                 return;
 389
 390         /* get the size after alignment with the u32 buffer size field */
 391         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 392         size = ALIGN(size + sizeof(u32), sizeof(u64));
 393         size -= sizeof(u32);
 394
 395         do {
 396                 char raw_data[size];
 397
 398                 /* zero the dead bytes from align to not leak stack to user */
 399                 *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 400
 401                 rec = (struct syscall_trace_enter *) raw_data;
 402                 tracing_generic_entry_update(&rec->ent, 0, 0);
 403                 rec->ent.type = sys_data->enter_id;
 404                 rec->nr = syscall_nr;
 405                 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 406                                        (unsigned long *)&rec->args);
 407                 perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
 408         } while(0);
 409 }
 410
 411 int reg_prof_syscall_enter(char *name)
 412 {
 413         int ret = 0;
 414         int num;
 415
 416         num = syscall_name_to_nr(name);
 417         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 418                 return -ENOSYS;
 419
 420         mutex_lock(&syscall_trace_lock);
 421         if (!sys_prof_refcount_enter)
 422                 ret = register_trace_sys_enter(prof_syscall_enter);
 423         if (ret) {
 424                 pr_info("event trace: Could not activate"
 425                                 "syscall entry trace point");
 426         } else {
 427                 set_bit(num, enabled_prof_enter_syscalls);
 428                 sys_prof_refcount_enter++;
 429         }
 430         mutex_unlock(&syscall_trace_lock);
 431         return ret;
 432 }
 433
 434 void unreg_prof_syscall_enter(char *name)
 435 {
 436         int num;
 437
 438         num = syscall_name_to_nr(name);
 439         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 440                 return;
 441
 442         mutex_lock(&syscall_trace_lock);
 443         sys_prof_refcount_enter--;
 444         clear_bit(num, enabled_prof_enter_syscalls);
 445         if (!sys_prof_refcount_enter)
 446                 unregister_trace_sys_enter(prof_syscall_enter);
 447         mutex_unlock(&syscall_trace_lock);
 448 }
 449
 450 static void prof_syscall_exit(struct pt_regs *regs, long ret)
 451 {
 452         struct syscall_metadata *sys_data;
 453         struct syscall_trace_exit rec;
 454         int syscall_nr;
 455
 456         syscall_nr = syscall_get_nr(current, regs);
 457         if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
 458                 return;
 459
 460         sys_data = syscall_nr_to_meta(syscall_nr);
 461         if (!sys_data)
 462                 return;
 463
 464         tracing_generic_entry_update(&rec.ent, 0, 0);
 465         rec.ent.type = sys_data->exit_id;
 466         rec.nr = syscall_nr;
 467         rec.ret = syscall_get_return_value(current, regs);
 468
 469         perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
 470 }
 471
 472 int reg_prof_syscall_exit(char *name)
 473 {
 474         int ret = 0;
 475         int num;
 476
 477         num = syscall_name_to_nr(name);
 478         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 479                 return -ENOSYS;
 480
 481         mutex_lock(&syscall_trace_lock);
 482         if (!sys_prof_refcount_exit)
 483                 ret = register_trace_sys_exit(prof_syscall_exit);
 484         if (ret) {
 485                 pr_info("event trace: Could not activate"
 486                                 "syscall entry trace point");
 487         } else {
 488                 set_bit(num, enabled_prof_exit_syscalls);
 489                 sys_prof_refcount_exit++;
 490         }
 491         mutex_unlock(&syscall_trace_lock);
 492         return ret;
 493 }
 494
 495 void unreg_prof_syscall_exit(char *name)
 496 {
 497         int num;
 498
 499         num = syscall_name_to_nr(name);
 500         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 501                 return;
 502
 503         mutex_lock(&syscall_trace_lock);
 504         sys_prof_refcount_exit--;
 505         clear_bit(num, enabled_prof_exit_syscalls);
 506         if (!sys_prof_refcount_exit)
 507                 unregister_trace_sys_exit(prof_syscall_exit);
 508         mutex_unlock(&syscall_trace_lock);
 509 }
 510
 511 #endif
 512
 513