kernel/trace/trace_syscalls.c

   1 #include <trace/syscall.h>
   2 #include <linux/kernel.h>
   3 #include <linux/ftrace.h>
   4 #include <linux/perf_counter.h>
   5 #include <asm/syscall.h>
   6
   7 #include "trace_output.h"
   8 #include "trace.h"
   9
  10 static DEFINE_MUTEX(syscall_trace_lock);
  11 static int sys_refcount_enter;
  12 static int sys_refcount_exit;
  13 static DECLARE_BITMAP(enabled_enter_syscalls, FTRACE_SYSCALL_MAX);
  14 static DECLARE_BITMAP(enabled_exit_syscalls, FTRACE_SYSCALL_MAX);
  15
  16 enum print_line_t
  17 print_syscall_enter(struct trace_iterator *iter, int flags)
  18 {
  19         struct trace_seq *s = &iter->seq;
  20         struct trace_entry *ent = iter->ent;
  21         struct syscall_trace_enter *trace;
  22         struct syscall_metadata *entry;
  23         int i, ret, syscall;
  24
  25         trace = (typeof(trace))ent;
  26         syscall = trace->nr;
  27         entry = syscall_nr_to_meta(syscall);
  28
  29         if (!entry)
  30                 goto end;
  31
  32         if (entry->enter_id != ent->type) {
  33                 WARN_ON_ONCE(1);
  34                 goto end;
  35         }
  36
  37         ret = trace_seq_printf(s, "%s(", entry->name);
  38         if (!ret)
  39                 return TRACE_TYPE_PARTIAL_LINE;
  40
  41         for (i = 0; i < entry->nb_args; i++) {
  42                 /* parameter types */
  43                 if (trace_flags & TRACE_ITER_VERBOSE) {
  44                         ret = trace_seq_printf(s, "%s ", entry->types[i]);
  45                         if (!ret)
  46                                 return TRACE_TYPE_PARTIAL_LINE;
  47                 }
  48                 /* parameter values */
  49                 ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
  50                                        trace->args[i],
  51                                        i == entry->nb_args - 1 ? "" : ", ");
  52                 if (!ret)
  53                         return TRACE_TYPE_PARTIAL_LINE;
  54         }
  55
  56         ret = trace_seq_putc(s, ')');
  57         if (!ret)
  58                 return TRACE_TYPE_PARTIAL_LINE;
  59
  60 end:
  61         ret =  trace_seq_putc(s, '\n');
  62         if (!ret)
  63                 return TRACE_TYPE_PARTIAL_LINE;
  64
  65         return TRACE_TYPE_HANDLED;
  66 }
  67
  68 enum print_line_t
  69 print_syscall_exit(struct trace_iterator *iter, int flags)
  70 {
  71         struct trace_seq *s = &iter->seq;
  72         struct trace_entry *ent = iter->ent;
  73         struct syscall_trace_exit *trace;
  74         int syscall;
  75         struct syscall_metadata *entry;
  76         int ret;
  77
  78         trace = (typeof(trace))ent;
  79         syscall = trace->nr;
  80         entry = syscall_nr_to_meta(syscall);
  81
  82         if (!entry) {
  83                 trace_seq_printf(s, "\n");
  84                 return TRACE_TYPE_HANDLED;
  85         }
  86
  87         if (entry->exit_id != ent->type) {
  88                 WARN_ON_ONCE(1);
  89                 return TRACE_TYPE_UNHANDLED;
  90         }
  91
  92         ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
  93                                 trace->ret);
  94         if (!ret)
  95                 return TRACE_TYPE_PARTIAL_LINE;
  96
  97         return TRACE_TYPE_HANDLED;
  98 }
  99
 100 extern char *__bad_type_size(void);
 101
 102 #define SYSCALL_FIELD(type, name)                                       \
 103         sizeof(type) != sizeof(trace.name) ?                            \
 104                 __bad_type_size() :                                     \
 105                 #type, #name, offsetof(typeof(trace), name), sizeof(trace.name)
 106
 107 int syscall_enter_format(struct ftrace_event_call *call, struct trace_seq *s)
 108 {
 109         int i;
 110         int nr;
 111         int ret;
 112         struct syscall_metadata *entry;
 113         struct syscall_trace_enter trace;
 114         int offset = offsetof(struct syscall_trace_enter, args);
 115
 116         nr = syscall_name_to_nr(call->data);
 117         entry = syscall_nr_to_meta(nr);
 118
 119         if (!entry)
 120                 return 0;
 121
 122         ret = trace_seq_printf(s, "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
 123                                SYSCALL_FIELD(int, nr));
 124         if (!ret)
 125                 return 0;
 126
 127         for (i = 0; i < entry->nb_args; i++) {
 128                 ret = trace_seq_printf(s, "\tfield:%s %s;", entry->types[i],
 129                                         entry->args[i]);
 130                 if (!ret)
 131                         return 0;
 132                 ret = trace_seq_printf(s, "\toffset:%d;\tsize:%zu;\n", offset,
 133                                        sizeof(unsigned long));
 134                 if (!ret)
 135                         return 0;
 136                 offset += sizeof(unsigned long);
 137         }
 138
 139         trace_seq_puts(s, "\nprint fmt: \"");
 140         for (i = 0; i < entry->nb_args; i++) {
 141                 ret = trace_seq_printf(s, "%s: 0x%%0%zulx%s", entry->args[i],
 142                                         sizeof(unsigned long),
 143                                         i == entry->nb_args - 1 ? "" : ", ");
 144                 if (!ret)
 145                         return 0;
 146         }
 147         trace_seq_putc(s, '"');
 148
 149         for (i = 0; i < entry->nb_args; i++) {
 150                 ret = trace_seq_printf(s, ", ((unsigned long)(REC->%s))",
 151                                        entry->args[i]);
 152                 if (!ret)
 153                         return 0;
 154         }
 155
 156         return trace_seq_putc(s, '\n');
 157 }
 158
 159 int syscall_exit_format(struct ftrace_event_call *call, struct trace_seq *s)
 160 {
 161         int ret;
 162         struct syscall_trace_exit trace;
 163
 164         ret = trace_seq_printf(s,
 165                                "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n"
 166                                "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
 167                                SYSCALL_FIELD(int, nr),
 168                                SYSCALL_FIELD(unsigned long, ret));
 169         if (!ret)
 170                 return 0;
 171
 172         return trace_seq_printf(s, "\nprint fmt: \"0x%%lx\", REC->ret\n");
 173 }
 174
 175 int syscall_enter_define_fields(struct ftrace_event_call *call)
 176 {
 177         struct syscall_trace_enter trace;
 178         struct syscall_metadata *meta;
 179         int ret;
 180         int nr;
 181         int i;
 182         int offset = offsetof(typeof(trace), args);
 183
 184         nr = syscall_name_to_nr(call->data);
 185         meta = syscall_nr_to_meta(nr);
 186
 187         if (!meta)
 188                 return 0;
 189
 190         ret = trace_define_common_fields(call);
 191         if (ret)
 192                 return ret;
 193
 194         for (i = 0; i < meta->nb_args; i++) {
 195                 ret = trace_define_field(call, meta->types[i],
 196                                          meta->args[i], offset,
 197                                          sizeof(unsigned long), 0,
 198                                          FILTER_OTHER);
 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                                  FILTER_OTHER);
 216
 217         return ret;
 218 }
 219
 220 void ftrace_syscall_enter(struct pt_regs *regs, long id)
 221 {
 222         struct syscall_trace_enter *entry;
 223         struct syscall_metadata *sys_data;
 224         struct ring_buffer_event *event;
 225         int size;
 226         int syscall_nr;
 227
 228         syscall_nr = syscall_get_nr(current, regs);
 229         if (!test_bit(syscall_nr, enabled_enter_syscalls))
 230                 return;
 231
 232         sys_data = syscall_nr_to_meta(syscall_nr);
 233         if (!sys_data)
 234                 return;
 235
 236         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 237
 238         event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
 239                                                         0, 0);
 240         if (!event)
 241                 return;
 242
 243         entry = ring_buffer_event_data(event);
 244         entry->nr = syscall_nr;
 245         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 246
 247         if (!filter_current_check_discard(sys_data->enter_event, entry, event))
 248                 trace_current_buffer_unlock_commit(event, 0, 0);
 249 }
 250
 251 void ftrace_syscall_exit(struct pt_regs *regs, long ret)
 252 {
 253         struct syscall_trace_exit *entry;
 254         struct syscall_metadata *sys_data;
 255         struct ring_buffer_event *event;
 256         int syscall_nr;
 257
 258         syscall_nr = syscall_get_nr(current, regs);
 259         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 260                 return;
 261
 262         sys_data = syscall_nr_to_meta(syscall_nr);
 263         if (!sys_data)
 264                 return;
 265
 266         event = trace_current_buffer_lock_reserve(sys_data->exit_id,
 267                                 sizeof(*entry), 0, 0);
 268         if (!event)
 269                 return;
 270
 271         entry = ring_buffer_event_data(event);
 272         entry->nr = syscall_nr;
 273         entry->ret = syscall_get_return_value(current, regs);
 274
 275         if (!filter_current_check_discard(sys_data->exit_event, entry, event))
 276                 trace_current_buffer_unlock_commit(event, 0, 0);
 277 }
 278
 279 int reg_event_syscall_enter(void *ptr)
 280 {
 281         int ret = 0;
 282         int num;
 283         char *name;
 284
 285         name = (char *)ptr;
 286         num = syscall_name_to_nr(name);
 287         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 288                 return -ENOSYS;
 289         mutex_lock(&syscall_trace_lock);
 290         if (!sys_refcount_enter)
 291                 ret = register_trace_syscall_enter(ftrace_syscall_enter);
 292         if (ret) {
 293                 pr_info("event trace: Could not activate"
 294                                 "syscall entry trace point");
 295         } else {
 296                 set_bit(num, enabled_enter_syscalls);
 297                 sys_refcount_enter++;
 298         }
 299         mutex_unlock(&syscall_trace_lock);
 300         return ret;
 301 }
 302
 303 void unreg_event_syscall_enter(void *ptr)
 304 {
 305         int num;
 306         char *name;
 307
 308         name = (char *)ptr;
 309         num = syscall_name_to_nr(name);
 310         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 311                 return;
 312         mutex_lock(&syscall_trace_lock);
 313         sys_refcount_enter--;
 314         clear_bit(num, enabled_enter_syscalls);
 315         if (!sys_refcount_enter)
 316                 unregister_trace_syscall_enter(ftrace_syscall_enter);
 317         mutex_unlock(&syscall_trace_lock);
 318 }
 319
 320 int reg_event_syscall_exit(void *ptr)
 321 {
 322         int ret = 0;
 323         int num;
 324         char *name;
 325
 326         name = (char *)ptr;
 327         num = syscall_name_to_nr(name);
 328         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 329                 return -ENOSYS;
 330         mutex_lock(&syscall_trace_lock);
 331         if (!sys_refcount_exit)
 332                 ret = register_trace_syscall_exit(ftrace_syscall_exit);
 333         if (ret) {
 334                 pr_info("event trace: Could not activate"
 335                                 "syscall exit trace point");
 336         } else {
 337                 set_bit(num, enabled_exit_syscalls);
 338                 sys_refcount_exit++;
 339         }
 340         mutex_unlock(&syscall_trace_lock);
 341         return ret;
 342 }
 343
 344 void unreg_event_syscall_exit(void *ptr)
 345 {
 346         int num;
 347         char *name;
 348
 349         name = (char *)ptr;
 350         num = syscall_name_to_nr(name);
 351         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 352                 return;
 353         mutex_lock(&syscall_trace_lock);
 354         sys_refcount_exit--;
 355         clear_bit(num, enabled_exit_syscalls);
 356         if (!sys_refcount_exit)
 357                 unregister_trace_syscall_exit(ftrace_syscall_exit);
 358         mutex_unlock(&syscall_trace_lock);
 359 }
 360
 361 struct trace_event event_syscall_enter = {
 362         .trace                  = print_syscall_enter,
 363 };
 364
 365 struct trace_event event_syscall_exit = {
 366         .trace                  = print_syscall_exit,
 367 };
 368
 369 #ifdef CONFIG_EVENT_PROFILE
 370
 371 static DECLARE_BITMAP(enabled_prof_enter_syscalls, FTRACE_SYSCALL_MAX);
 372 static DECLARE_BITMAP(enabled_prof_exit_syscalls, FTRACE_SYSCALL_MAX);
 373 static int sys_prof_refcount_enter;
 374 static int sys_prof_refcount_exit;
 375
 376 static void prof_syscall_enter(struct pt_regs *regs, long id)
 377 {
 378         struct syscall_trace_enter *rec;
 379         struct syscall_metadata *sys_data;
 380         int syscall_nr;
 381         int size;
 382
 383         syscall_nr = syscall_get_nr(current, regs);
 384         if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
 385                 return;
 386
 387         sys_data = syscall_nr_to_meta(syscall_nr);
 388         if (!sys_data)
 389                 return;
 390
 391         /* get the size after alignment with the u32 buffer size field */
 392         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 393         size = ALIGN(size + sizeof(u32), sizeof(u64));
 394         size -= sizeof(u32);
 395
 396         do {
 397                 char raw_data[size];
 398
 399                 /* zero the dead bytes from align to not leak stack to user */
 400                 *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 401
 402                 rec = (struct syscall_trace_enter *) raw_data;
 403                 tracing_generic_entry_update(&rec->ent, 0, 0);
 404                 rec->ent.type = sys_data->enter_id;
 405                 rec->nr = syscall_nr;
 406                 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 407                                        (unsigned long *)&rec->args);
 408                 perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
 409         } while(0);
 410 }
 411
 412 int reg_prof_syscall_enter(char *name)
 413 {
 414         int ret = 0;
 415         int num;
 416
 417         num = syscall_name_to_nr(name);
 418         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 419                 return -ENOSYS;
 420
 421         mutex_lock(&syscall_trace_lock);
 422         if (!sys_prof_refcount_enter)
 423                 ret = register_trace_syscall_enter(prof_syscall_enter);
 424         if (ret) {
 425                 pr_info("event trace: Could not activate"
 426                                 "syscall entry trace point");
 427         } else {
 428                 set_bit(num, enabled_prof_enter_syscalls);
 429                 sys_prof_refcount_enter++;
 430         }
 431         mutex_unlock(&syscall_trace_lock);
 432         return ret;
 433 }
 434
 435 void unreg_prof_syscall_enter(char *name)
 436 {
 437         int num;
 438
 439         num = syscall_name_to_nr(name);
 440         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 441                 return;
 442
 443         mutex_lock(&syscall_trace_lock);
 444         sys_prof_refcount_enter--;
 445         clear_bit(num, enabled_prof_enter_syscalls);
 446         if (!sys_prof_refcount_enter)
 447                 unregister_trace_syscall_enter(prof_syscall_enter);
 448         mutex_unlock(&syscall_trace_lock);
 449 }
 450
 451 static void prof_syscall_exit(struct pt_regs *regs, long ret)
 452 {
 453         struct syscall_metadata *sys_data;
 454         struct syscall_trace_exit rec;
 455         int syscall_nr;
 456
 457         syscall_nr = syscall_get_nr(current, regs);
 458         if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
 459                 return;
 460
 461         sys_data = syscall_nr_to_meta(syscall_nr);
 462         if (!sys_data)
 463                 return;
 464
 465         tracing_generic_entry_update(&rec.ent, 0, 0);
 466         rec.ent.type = sys_data->exit_id;
 467         rec.nr = syscall_nr;
 468         rec.ret = syscall_get_return_value(current, regs);
 469
 470         perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
 471 }
 472
 473 int reg_prof_syscall_exit(char *name)
 474 {
 475         int ret = 0;
 476         int num;
 477
 478         num = syscall_name_to_nr(name);
 479         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 480                 return -ENOSYS;
 481
 482         mutex_lock(&syscall_trace_lock);
 483         if (!sys_prof_refcount_exit)
 484                 ret = register_trace_syscall_exit(prof_syscall_exit);
 485         if (ret) {
 486                 pr_info("event trace: Could not activate"
 487                                 "syscall entry trace point");
 488         } else {
 489                 set_bit(num, enabled_prof_exit_syscalls);
 490                 sys_prof_refcount_exit++;
 491         }
 492         mutex_unlock(&syscall_trace_lock);
 493         return ret;
 494 }
 495
 496 void unreg_prof_syscall_exit(char *name)
 497 {
 498         int num;
 499
 500         num = syscall_name_to_nr(name);
 501         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 502                 return;
 503
 504         mutex_lock(&syscall_trace_lock);
 505         sys_prof_refcount_exit--;
 506         clear_bit(num, enabled_prof_exit_syscalls);
 507         if (!sys_prof_refcount_exit)
 508                 unregister_trace_syscall_exit(prof_syscall_exit);
 509         mutex_unlock(&syscall_trace_lock);
 510 }
 511
 512 #endif
 513
 514