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