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 /* Option to display the parameters types */
  17 enum {
  18         TRACE_SYSCALLS_OPT_TYPES = 0x1,
  19 };
  20
  21 static struct tracer_opt syscalls_opts[] = {
  22         { TRACER_OPT(syscall_arg_type, TRACE_SYSCALLS_OPT_TYPES) },
  23         { }
  24 };
  25
  26 static struct tracer_flags syscalls_flags = {
  27         .val = 0, /* By default: no parameters types */
  28         .opts = syscalls_opts
  29 };
  30
  31 enum print_line_t
  32 print_syscall_enter(struct trace_iterator *iter, int flags)
  33 {
  34         struct trace_seq *s = &iter->seq;
  35         struct trace_entry *ent = iter->ent;
  36         struct syscall_trace_enter *trace;
  37         struct syscall_metadata *entry;
  38         int i, ret, syscall;
  39
  40         trace = (typeof(trace))ent;
  41         syscall = trace->nr;
  42         entry = syscall_nr_to_meta(syscall);
  43
  44         if (!entry)
  45                 goto end;
  46
  47         if (entry->enter_id != ent->type) {
  48                 WARN_ON_ONCE(1);
  49                 goto end;
  50         }
  51
  52         ret = trace_seq_printf(s, "%s(", entry->name);
  53         if (!ret)
  54                 return TRACE_TYPE_PARTIAL_LINE;
  55
  56         for (i = 0; i < entry->nb_args; i++) {
  57                 /* parameter types */
  58                 if (syscalls_flags.val & TRACE_SYSCALLS_OPT_TYPES) {
  59                         ret = trace_seq_printf(s, "%s ", entry->types[i]);
  60                         if (!ret)
  61                                 return TRACE_TYPE_PARTIAL_LINE;
  62                 }
  63                 /* parameter values */
  64                 ret = trace_seq_printf(s, "%s: %lx%s ", entry->args[i],
  65                                        trace->args[i],
  66                                        i == entry->nb_args - 1 ? ")" : ",");
  67                 if (!ret)
  68                         return TRACE_TYPE_PARTIAL_LINE;
  69         }
  70
  71 end:
  72         trace_seq_printf(s, "\n");
  73         return TRACE_TYPE_HANDLED;
  74 }
  75
  76 enum print_line_t
  77 print_syscall_exit(struct trace_iterator *iter, int flags)
  78 {
  79         struct trace_seq *s = &iter->seq;
  80         struct trace_entry *ent = iter->ent;
  81         struct syscall_trace_exit *trace;
  82         int syscall;
  83         struct syscall_metadata *entry;
  84         int ret;
  85
  86         trace = (typeof(trace))ent;
  87         syscall = trace->nr;
  88         entry = syscall_nr_to_meta(syscall);
  89
  90         if (!entry) {
  91                 trace_seq_printf(s, "\n");
  92                 return TRACE_TYPE_HANDLED;
  93         }
  94
  95         if (entry->exit_id != ent->type) {
  96                 WARN_ON_ONCE(1);
  97                 return TRACE_TYPE_UNHANDLED;
  98         }
  99
 100         ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
 101                                 trace->ret);
 102         if (!ret)
 103                 return TRACE_TYPE_PARTIAL_LINE;
 104
 105         return TRACE_TYPE_HANDLED;
 106 }
 107
 108 int ftrace_format_syscall(struct ftrace_event_call *call, struct trace_seq *s)
 109 {
 110         int i;
 111         int nr;
 112         int ret = 0;
 113         struct syscall_metadata *entry;
 114         int offset = sizeof(struct trace_entry);
 115
 116         nr = syscall_name_to_nr((char *)call->data);
 117         entry = syscall_nr_to_meta(nr);
 118
 119         if (!entry)
 120                 return ret;
 121
 122         for (i = 0; i < entry->nb_args; i++) {
 123                 ret = trace_seq_printf(s, "\tfield:%s %s;", entry->types[i],
 124                                         entry->args[i]);
 125                 if (!ret)
 126                         return 0;
 127                 ret = trace_seq_printf(s, "\toffset:%d;\tsize:%lu;\n", offset,
 128                                        sizeof(unsigned long));
 129                 if (!ret)
 130                         return 0;
 131                 offset += sizeof(unsigned long);
 132         }
 133
 134         trace_seq_printf(s, "\nprint fmt: \"");
 135         for (i = 0; i < entry->nb_args; i++) {
 136                 ret = trace_seq_printf(s, "%s: 0x%%0%lulx%s", entry->args[i],
 137                                         sizeof(unsigned long),
 138                                         i == entry->nb_args - 1 ? "\", " : ", ");
 139                 if (!ret)
 140                         return 0;
 141         }
 142
 143         for (i = 0; i < entry->nb_args; i++) {
 144                 ret = trace_seq_printf(s, "((unsigned long)(REC->%s))%s",
 145                                         entry->args[i],
 146                                         i == entry->nb_args - 1 ? "\n" : ", ");
 147                 if (!ret)
 148                         return 0;
 149         }
 150
 151         return ret;
 152 }
 153
 154 void ftrace_syscall_enter(struct pt_regs *regs, long id)
 155 {
 156         struct syscall_trace_enter *entry;
 157         struct syscall_metadata *sys_data;
 158         struct ring_buffer_event *event;
 159         int size;
 160         int syscall_nr;
 161
 162         syscall_nr = syscall_get_nr(current, regs);
 163         if (!test_bit(syscall_nr, enabled_enter_syscalls))
 164                 return;
 165
 166         sys_data = syscall_nr_to_meta(syscall_nr);
 167         if (!sys_data)
 168                 return;
 169
 170         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 171
 172         event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
 173                                                         0, 0);
 174         if (!event)
 175                 return;
 176
 177         entry = ring_buffer_event_data(event);
 178         entry->nr = syscall_nr;
 179         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 180
 181         trace_current_buffer_unlock_commit(event, 0, 0);
 182         trace_wake_up();
 183 }
 184
 185 void ftrace_syscall_exit(struct pt_regs *regs, long ret)
 186 {
 187         struct syscall_trace_exit *entry;
 188         struct syscall_metadata *sys_data;
 189         struct ring_buffer_event *event;
 190         int syscall_nr;
 191
 192         syscall_nr = syscall_get_nr(current, regs);
 193         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 194                 return;
 195
 196         sys_data = syscall_nr_to_meta(syscall_nr);
 197         if (!sys_data)
 198                 return;
 199
 200         event = trace_current_buffer_lock_reserve(sys_data->exit_id,
 201                                 sizeof(*entry), 0, 0);
 202         if (!event)
 203                 return;
 204
 205         entry = ring_buffer_event_data(event);
 206         entry->nr = syscall_nr;
 207         entry->ret = syscall_get_return_value(current, regs);
 208
 209         trace_current_buffer_unlock_commit(event, 0, 0);
 210         trace_wake_up();
 211 }
 212
 213 int reg_event_syscall_enter(void *ptr)
 214 {
 215         int ret = 0;
 216         int num;
 217         char *name;
 218
 219         name = (char *)ptr;
 220         num = syscall_name_to_nr(name);
 221         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 222                 return -ENOSYS;
 223         mutex_lock(&syscall_trace_lock);
 224         if (!sys_refcount_enter)
 225                 ret = register_trace_syscall_enter(ftrace_syscall_enter);
 226         if (ret) {
 227                 pr_info("event trace: Could not activate"
 228                                 "syscall entry trace point");
 229         } else {
 230                 set_bit(num, enabled_enter_syscalls);
 231                 sys_refcount_enter++;
 232         }
 233         mutex_unlock(&syscall_trace_lock);
 234         return ret;
 235 }
 236
 237 void unreg_event_syscall_enter(void *ptr)
 238 {
 239         int num;
 240         char *name;
 241
 242         name = (char *)ptr;
 243         num = syscall_name_to_nr(name);
 244         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 245                 return;
 246         mutex_lock(&syscall_trace_lock);
 247         sys_refcount_enter--;
 248         clear_bit(num, enabled_enter_syscalls);
 249         if (!sys_refcount_enter)
 250                 unregister_trace_syscall_enter(ftrace_syscall_enter);
 251         mutex_unlock(&syscall_trace_lock);
 252 }
 253
 254 int reg_event_syscall_exit(void *ptr)
 255 {
 256         int ret = 0;
 257         int num;
 258         char *name;
 259
 260         name = (char *)ptr;
 261         num = syscall_name_to_nr(name);
 262         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 263                 return -ENOSYS;
 264         mutex_lock(&syscall_trace_lock);
 265         if (!sys_refcount_exit)
 266                 ret = register_trace_syscall_exit(ftrace_syscall_exit);
 267         if (ret) {
 268                 pr_info("event trace: Could not activate"
 269                                 "syscall exit trace point");
 270         } else {
 271                 set_bit(num, enabled_exit_syscalls);
 272                 sys_refcount_exit++;
 273         }
 274         mutex_unlock(&syscall_trace_lock);
 275         return ret;
 276 }
 277
 278 void unreg_event_syscall_exit(void *ptr)
 279 {
 280         int num;
 281         char *name;
 282
 283         name = (char *)ptr;
 284         num = syscall_name_to_nr(name);
 285         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 286                 return;
 287         mutex_lock(&syscall_trace_lock);
 288         sys_refcount_exit--;
 289         clear_bit(num, enabled_exit_syscalls);
 290         if (!sys_refcount_exit)
 291                 unregister_trace_syscall_exit(ftrace_syscall_exit);
 292         mutex_unlock(&syscall_trace_lock);
 293 }
 294
 295 struct trace_event event_syscall_enter = {
 296         .trace                  = print_syscall_enter,
 297 };
 298
 299 struct trace_event event_syscall_exit = {
 300         .trace                  = print_syscall_exit,
 301 };
 302
 303 #ifdef CONFIG_EVENT_PROFILE
 304
 305 struct syscall_enter_record {
 306         struct trace_entry      entry;
 307         unsigned long           args[0];
 308 };
 309
 310 struct syscall_exit_record {
 311         struct trace_entry      entry;
 312         unsigned long           ret;
 313 };
 314
 315 static DECLARE_BITMAP(enabled_prof_enter_syscalls, FTRACE_SYSCALL_MAX);
 316 static DECLARE_BITMAP(enabled_prof_exit_syscalls, FTRACE_SYSCALL_MAX);
 317 static int sys_prof_refcount_enter;
 318 static int sys_prof_refcount_exit;
 319
 320 static void prof_syscall_enter(struct pt_regs *regs, long id)
 321 {
 322         struct syscall_enter_record *rec;
 323         struct syscall_metadata *sys_data;
 324         int syscall_nr;
 325         int size;
 326
 327         syscall_nr = syscall_get_nr(current, regs);
 328         if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
 329                 return;
 330
 331         sys_data = syscall_nr_to_meta(syscall_nr);
 332         if (!sys_data)
 333                 return;
 334
 335         /* get the size after alignment with the u32 buffer size field */
 336         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 337         size = ALIGN(size + sizeof(u32), sizeof(u64));
 338         size -= sizeof(u32);
 339
 340         do {
 341                 char raw_data[size];
 342
 343                 /* zero the dead bytes from align to not leak stack to user */
 344                 *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
 345
 346                 rec = (struct syscall_enter_record *) raw_data;
 347                 tracing_generic_entry_update(&rec->entry, 0, 0);
 348                 rec->entry.type = sys_data->enter_id;
 349                 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 350                                        (unsigned long *)&rec->args);
 351                 perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
 352         } while(0);
 353 }
 354
 355 int reg_prof_syscall_enter(char *name)
 356 {
 357         int ret = 0;
 358         int num;
 359
 360         num = syscall_name_to_nr(name);
 361         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 362                 return -ENOSYS;
 363
 364         mutex_lock(&syscall_trace_lock);
 365         if (!sys_prof_refcount_enter)
 366                 ret = register_trace_syscall_enter(prof_syscall_enter);
 367         if (ret) {
 368                 pr_info("event trace: Could not activate"
 369                                 "syscall entry trace point");
 370         } else {
 371                 set_bit(num, enabled_prof_enter_syscalls);
 372                 sys_prof_refcount_enter++;
 373         }
 374         mutex_unlock(&syscall_trace_lock);
 375         return ret;
 376 }
 377
 378 void unreg_prof_syscall_enter(char *name)
 379 {
 380         int num;
 381
 382         num = syscall_name_to_nr(name);
 383         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 384                 return;
 385
 386         mutex_lock(&syscall_trace_lock);
 387         sys_prof_refcount_enter--;
 388         clear_bit(num, enabled_prof_enter_syscalls);
 389         if (!sys_prof_refcount_enter)
 390                 unregister_trace_syscall_enter(prof_syscall_enter);
 391         mutex_unlock(&syscall_trace_lock);
 392 }
 393
 394 static void prof_syscall_exit(struct pt_regs *regs, long ret)
 395 {
 396         struct syscall_metadata *sys_data;
 397         struct syscall_exit_record rec;
 398         int syscall_nr;
 399
 400         syscall_nr = syscall_get_nr(current, regs);
 401         if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
 402                 return;
 403
 404         sys_data = syscall_nr_to_meta(syscall_nr);
 405         if (!sys_data)
 406                 return;
 407
 408         tracing_generic_entry_update(&rec.entry, 0, 0);
 409         rec.entry.type = sys_data->exit_id;
 410         rec.ret = syscall_get_return_value(current, regs);
 411
 412         perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
 413 }
 414
 415 int reg_prof_syscall_exit(char *name)
 416 {
 417         int ret = 0;
 418         int num;
 419
 420         num = syscall_name_to_nr(name);
 421         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 422                 return -ENOSYS;
 423
 424         mutex_lock(&syscall_trace_lock);
 425         if (!sys_prof_refcount_exit)
 426                 ret = register_trace_syscall_exit(prof_syscall_exit);
 427         if (ret) {
 428                 pr_info("event trace: Could not activate"
 429                                 "syscall entry trace point");
 430         } else {
 431                 set_bit(num, enabled_prof_exit_syscalls);
 432                 sys_prof_refcount_exit++;
 433         }
 434         mutex_unlock(&syscall_trace_lock);
 435         return ret;
 436 }
 437
 438 void unreg_prof_syscall_exit(char *name)
 439 {
 440         int num;
 441
 442         num = syscall_name_to_nr(name);
 443         if (num < 0 || num >= FTRACE_SYSCALL_MAX)
 444                 return;
 445
 446         mutex_lock(&syscall_trace_lock);
 447         sys_prof_refcount_exit--;
 448         clear_bit(num, enabled_prof_exit_syscalls);
 449         if (!sys_prof_refcount_exit)
 450                 unregister_trace_syscall_exit(prof_syscall_exit);
 451         mutex_unlock(&syscall_trace_lock);
 452 }
 453
 454 #endif
 455
 456