kernel/trace/trace_syscalls.c

   1 #include <trace/syscall.h>
   2 #include <trace/events/syscalls.h>
   3 #include <linux/slab.h>
   4 #include <linux/kernel.h>
   5 #include <linux/ftrace.h>
   6 #include <linux/perf_event.h>
   7 #include <asm/syscall.h>
   8
   9 #include "trace_output.h"
  10 #include "trace.h"
  11
  12 static DEFINE_MUTEX(syscall_trace_lock);
  13 static int sys_refcount_enter;
  14 static int sys_refcount_exit;
  15 static DECLARE_BITMAP(enabled_enter_syscalls, NR_syscalls);
  16 static DECLARE_BITMAP(enabled_exit_syscalls, NR_syscalls);
  17
  18 struct ftrace_event_class event_class_syscalls = {
  19         .system                 = "syscalls"
  20 };
  21
  22 extern unsigned long __start_syscalls_metadata[];
  23 extern unsigned long __stop_syscalls_metadata[];
  24
  25 static struct syscall_metadata **syscalls_metadata;
  26
  27 static struct syscall_metadata *find_syscall_meta(unsigned long syscall)
  28 {
  29         struct syscall_metadata *start;
  30         struct syscall_metadata *stop;
  31         char str[KSYM_SYMBOL_LEN];
  32
  33
  34         start = (struct syscall_metadata *)__start_syscalls_metadata;
  35         stop = (struct syscall_metadata *)__stop_syscalls_metadata;
  36         kallsyms_lookup(syscall, NULL, NULL, NULL, str);
  37
  38         for ( ; start < stop; start++) {
  39                 /*
  40                  * Only compare after the "sys" prefix. Archs that use
  41                  * syscall wrappers may have syscalls symbols aliases prefixed
  42                  * with "SyS" instead of "sys", leading to an unwanted
  43                  * mismatch.
  44                  */
  45                 if (start->name && !strcmp(start->name + 3, str + 3))
  46                         return start;
  47         }
  48         return NULL;
  49 }
  50
  51 static struct syscall_metadata *syscall_nr_to_meta(int nr)
  52 {
  53         if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
  54                 return NULL;
  55
  56         return syscalls_metadata[nr];
  57 }
  58
  59 enum print_line_t
  60 print_syscall_enter(struct trace_iterator *iter, int flags)
  61 {
  62         struct trace_seq *s = &iter->seq;
  63         struct trace_entry *ent = iter->ent;
  64         struct syscall_trace_enter *trace;
  65         struct syscall_metadata *entry;
  66         int i, ret, syscall;
  67
  68         trace = (typeof(trace))ent;
  69         syscall = trace->nr;
  70         entry = syscall_nr_to_meta(syscall);
  71
  72         if (!entry)
  73                 goto end;
  74
  75         if (entry->enter_event->id != ent->type) {
  76                 WARN_ON_ONCE(1);
  77                 goto end;
  78         }
  79
  80         ret = trace_seq_printf(s, "%s(", entry->name);
  81         if (!ret)
  82                 return TRACE_TYPE_PARTIAL_LINE;
  83
  84         for (i = 0; i < entry->nb_args; i++) {
  85                 /* parameter types */
  86                 if (trace_flags & TRACE_ITER_VERBOSE) {
  87                         ret = trace_seq_printf(s, "%s ", entry->types[i]);
  88                         if (!ret)
  89                                 return TRACE_TYPE_PARTIAL_LINE;
  90                 }
  91                 /* parameter values */
  92                 ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
  93                                        trace->args[i],
  94                                        i == entry->nb_args - 1 ? "" : ", ");
  95                 if (!ret)
  96                         return TRACE_TYPE_PARTIAL_LINE;
  97         }
  98
  99         ret = trace_seq_putc(s, ')');
 100         if (!ret)
 101                 return TRACE_TYPE_PARTIAL_LINE;
 102
 103 end:
 104         ret =  trace_seq_putc(s, '\n');
 105         if (!ret)
 106                 return TRACE_TYPE_PARTIAL_LINE;
 107
 108         return TRACE_TYPE_HANDLED;
 109 }
 110
 111 enum print_line_t
 112 print_syscall_exit(struct trace_iterator *iter, int flags)
 113 {
 114         struct trace_seq *s = &iter->seq;
 115         struct trace_entry *ent = iter->ent;
 116         struct syscall_trace_exit *trace;
 117         int syscall;
 118         struct syscall_metadata *entry;
 119         int ret;
 120
 121         trace = (typeof(trace))ent;
 122         syscall = trace->nr;
 123         entry = syscall_nr_to_meta(syscall);
 124
 125         if (!entry) {
 126                 trace_seq_printf(s, "\n");
 127                 return TRACE_TYPE_HANDLED;
 128         }
 129
 130         if (entry->exit_event->id != ent->type) {
 131                 WARN_ON_ONCE(1);
 132                 return TRACE_TYPE_UNHANDLED;
 133         }
 134
 135         ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
 136                                 trace->ret);
 137         if (!ret)
 138                 return TRACE_TYPE_PARTIAL_LINE;
 139
 140         return TRACE_TYPE_HANDLED;
 141 }
 142
 143 extern char *__bad_type_size(void);
 144
 145 #define SYSCALL_FIELD(type, name)                                       \
 146         sizeof(type) != sizeof(trace.name) ?                            \
 147                 __bad_type_size() :                                     \
 148                 #type, #name, offsetof(typeof(trace), name),            \
 149                 sizeof(trace.name), is_signed_type(type)
 150
 151 static
 152 int  __set_enter_print_fmt(struct syscall_metadata *entry, char *buf, int len)
 153 {
 154         int i;
 155         int pos = 0;
 156
 157         /* When len=0, we just calculate the needed length */
 158 #define LEN_OR_ZERO (len ? len - pos : 0)
 159
 160         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
 161         for (i = 0; i < entry->nb_args; i++) {
 162                 pos += snprintf(buf + pos, LEN_OR_ZERO, "%s: 0x%%0%zulx%s",
 163                                 entry->args[i], sizeof(unsigned long),
 164                                 i == entry->nb_args - 1 ? "" : ", ");
 165         }
 166         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
 167
 168         for (i = 0; i < entry->nb_args; i++) {
 169                 pos += snprintf(buf + pos, LEN_OR_ZERO,
 170                                 ", ((unsigned long)(REC->%s))", entry->args[i]);
 171         }
 172
 173 #undef LEN_OR_ZERO
 174
 175         /* return the length of print_fmt */
 176         return pos;
 177 }
 178
 179 static int set_syscall_print_fmt(struct ftrace_event_call *call)
 180 {
 181         char *print_fmt;
 182         int len;
 183         struct syscall_metadata *entry = call->data;
 184
 185         if (entry->enter_event != call) {
 186                 call->print_fmt = "\"0x%lx\", REC->ret";
 187                 return 0;
 188         }
 189
 190         /* First: called with 0 length to calculate the needed length */
 191         len = __set_enter_print_fmt(entry, NULL, 0);
 192
 193         print_fmt = kmalloc(len + 1, GFP_KERNEL);
 194         if (!print_fmt)
 195                 return -ENOMEM;
 196
 197         /* Second: actually write the @print_fmt */
 198         __set_enter_print_fmt(entry, print_fmt, len + 1);
 199         call->print_fmt = print_fmt;
 200
 201         return 0;
 202 }
 203
 204 static void free_syscall_print_fmt(struct ftrace_event_call *call)
 205 {
 206         struct syscall_metadata *entry = call->data;
 207
 208         if (entry->enter_event == call)
 209                 kfree(call->print_fmt);
 210 }
 211
 212 int syscall_enter_define_fields(struct ftrace_event_call *call)
 213 {
 214         struct syscall_trace_enter trace;
 215         struct syscall_metadata *meta = call->data;
 216         int ret;
 217         int i;
 218         int offset = offsetof(typeof(trace), args);
 219
 220         ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
 221         if (ret)
 222                 return ret;
 223
 224         for (i = 0; i < meta->nb_args; i++) {
 225                 ret = trace_define_field(call, meta->types[i],
 226                                          meta->args[i], offset,
 227                                          sizeof(unsigned long), 0,
 228                                          FILTER_OTHER);
 229                 offset += sizeof(unsigned long);
 230         }
 231
 232         return ret;
 233 }
 234
 235 int syscall_exit_define_fields(struct ftrace_event_call *call)
 236 {
 237         struct syscall_trace_exit trace;
 238         int ret;
 239
 240         ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
 241         if (ret)
 242                 return ret;
 243
 244         ret = trace_define_field(call, SYSCALL_FIELD(long, ret),
 245                                  FILTER_OTHER);
 246
 247         return ret;
 248 }
 249
 250 void ftrace_syscall_enter(void *ignore, struct pt_regs *regs, long id)
 251 {
 252         struct syscall_trace_enter *entry;
 253         struct syscall_metadata *sys_data;
 254         struct ring_buffer_event *event;
 255         struct ring_buffer *buffer;
 256         int size;
 257         int syscall_nr;
 258
 259         syscall_nr = syscall_get_nr(current, regs);
 260         if (syscall_nr < 0)
 261                 return;
 262         if (!test_bit(syscall_nr, enabled_enter_syscalls))
 263                 return;
 264
 265         sys_data = syscall_nr_to_meta(syscall_nr);
 266         if (!sys_data)
 267                 return;
 268
 269         size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
 270
 271         event = trace_current_buffer_lock_reserve(&buffer,
 272                         sys_data->enter_event->id, size, 0, 0);
 273         if (!event)
 274                 return;
 275
 276         entry = ring_buffer_event_data(event);
 277         entry->nr = syscall_nr;
 278         syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
 279
 280         if (!filter_current_check_discard(buffer, sys_data->enter_event,
 281                                           entry, event))
 282                 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
 283 }
 284
 285 void ftrace_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
 286 {
 287         struct syscall_trace_exit *entry;
 288         struct syscall_metadata *sys_data;
 289         struct ring_buffer_event *event;
 290         struct ring_buffer *buffer;
 291         int syscall_nr;
 292
 293         syscall_nr = syscall_get_nr(current, regs);
 294         if (syscall_nr < 0)
 295                 return;
 296         if (!test_bit(syscall_nr, enabled_exit_syscalls))
 297                 return;
 298
 299         sys_data = syscall_nr_to_meta(syscall_nr);
 300         if (!sys_data)
 301                 return;
 302
 303         event = trace_current_buffer_lock_reserve(&buffer,
 304                         sys_data->exit_event->id, sizeof(*entry), 0, 0);
 305         if (!event)
 306                 return;
 307
 308         entry = ring_buffer_event_data(event);
 309         entry->nr = syscall_nr;
 310         entry->ret = syscall_get_return_value(current, regs);
 311
 312         if (!filter_current_check_discard(buffer, sys_data->exit_event,
 313                                           entry, event))
 314                 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
 315 }
 316
 317 int reg_event_syscall_enter(struct ftrace_event_call *call)
 318 {
 319         int ret = 0;
 320         int num;
 321
 322         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 323         if (num < 0 || num >= NR_syscalls)
 324                 return -ENOSYS;
 325         mutex_lock(&syscall_trace_lock);
 326         if (!sys_refcount_enter)
 327                 ret = register_trace_sys_enter(ftrace_syscall_enter, NULL);
 328         if (!ret) {
 329                 set_bit(num, enabled_enter_syscalls);
 330                 sys_refcount_enter++;
 331         }
 332         mutex_unlock(&syscall_trace_lock);
 333         return ret;
 334 }
 335
 336 void unreg_event_syscall_enter(struct ftrace_event_call *call)
 337 {
 338         int num;
 339
 340         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 341         if (num < 0 || num >= NR_syscalls)
 342                 return;
 343         mutex_lock(&syscall_trace_lock);
 344         sys_refcount_enter--;
 345         clear_bit(num, enabled_enter_syscalls);
 346         if (!sys_refcount_enter)
 347                 unregister_trace_sys_enter(ftrace_syscall_enter, NULL);
 348         mutex_unlock(&syscall_trace_lock);
 349 }
 350
 351 int reg_event_syscall_exit(struct ftrace_event_call *call)
 352 {
 353         int ret = 0;
 354         int num;
 355
 356         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 357         if (num < 0 || num >= NR_syscalls)
 358                 return -ENOSYS;
 359         mutex_lock(&syscall_trace_lock);
 360         if (!sys_refcount_exit)
 361                 ret = register_trace_sys_exit(ftrace_syscall_exit, NULL);
 362         if (!ret) {
 363                 set_bit(num, enabled_exit_syscalls);
 364                 sys_refcount_exit++;
 365         }
 366         mutex_unlock(&syscall_trace_lock);
 367         return ret;
 368 }
 369
 370 void unreg_event_syscall_exit(struct ftrace_event_call *call)
 371 {
 372         int num;
 373
 374         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 375         if (num < 0 || num >= NR_syscalls)
 376                 return;
 377         mutex_lock(&syscall_trace_lock);
 378         sys_refcount_exit--;
 379         clear_bit(num, enabled_exit_syscalls);
 380         if (!sys_refcount_exit)
 381                 unregister_trace_sys_exit(ftrace_syscall_exit, NULL);
 382         mutex_unlock(&syscall_trace_lock);
 383 }
 384
 385 int init_syscall_trace(struct ftrace_event_call *call)
 386 {
 387         int id;
 388
 389         if (set_syscall_print_fmt(call) < 0)
 390                 return -ENOMEM;
 391
 392         id = trace_event_raw_init(call);
 393
 394         if (id < 0) {
 395                 free_syscall_print_fmt(call);
 396                 return id;
 397         }
 398
 399         return id;
 400 }
 401
 402 unsigned long __init arch_syscall_addr(int nr)
 403 {
 404         return (unsigned long)sys_call_table[nr];
 405 }
 406
 407 int __init init_ftrace_syscalls(void)
 408 {
 409         struct syscall_metadata *meta;
 410         unsigned long addr;
 411         int i;
 412
 413         syscalls_metadata = kzalloc(sizeof(*syscalls_metadata) *
 414                                         NR_syscalls, GFP_KERNEL);
 415         if (!syscalls_metadata) {
 416                 WARN_ON(1);
 417                 return -ENOMEM;
 418         }
 419
 420         for (i = 0; i < NR_syscalls; i++) {
 421                 addr = arch_syscall_addr(i);
 422                 meta = find_syscall_meta(addr);
 423                 if (!meta)
 424                         continue;
 425
 426                 meta->syscall_nr = i;
 427                 syscalls_metadata[i] = meta;
 428         }
 429
 430         return 0;
 431 }
 432 core_initcall(init_ftrace_syscalls);
 433
 434 #ifdef CONFIG_PERF_EVENTS
 435
 436 static DECLARE_BITMAP(enabled_perf_enter_syscalls, NR_syscalls);
 437 static DECLARE_BITMAP(enabled_perf_exit_syscalls, NR_syscalls);
 438 static int sys_perf_refcount_enter;
 439 static int sys_perf_refcount_exit;
 440
 441 static void perf_syscall_enter(void *ignore, struct pt_regs *regs, long id)
 442 {
 443         struct syscall_metadata *sys_data;
 444         struct syscall_trace_enter *rec;
 445         unsigned long flags;
 446         int syscall_nr;
 447         int rctx;
 448         int size;
 449
 450         syscall_nr = syscall_get_nr(current, regs);
 451         if (!test_bit(syscall_nr, enabled_perf_enter_syscalls))
 452                 return;
 453
 454         sys_data = syscall_nr_to_meta(syscall_nr);
 455         if (!sys_data)
 456                 return;
 457
 458         /* get the size after alignment with the u32 buffer size field */
 459         size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
 460         size = ALIGN(size + sizeof(u32), sizeof(u64));
 461         size -= sizeof(u32);
 462
 463         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 464                       "perf buffer not large enough"))
 465                 return;
 466
 467         rec = (struct syscall_trace_enter *)perf_trace_buf_prepare(size,
 468                                 sys_data->enter_event->id, &rctx, &flags);
 469         if (!rec)
 470                 return;
 471
 472         rec->nr = syscall_nr;
 473         syscall_get_arguments(current, regs, 0, sys_data->nb_args,
 474                                (unsigned long *)&rec->args);
 475         perf_trace_buf_submit(rec, size, rctx, 0, 1, flags, regs);
 476 }
 477
 478 int perf_sysenter_enable(struct ftrace_event_call *call)
 479 {
 480         int ret = 0;
 481         int num;
 482
 483         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 484
 485         mutex_lock(&syscall_trace_lock);
 486         if (!sys_perf_refcount_enter)
 487                 ret = register_trace_sys_enter(perf_syscall_enter, NULL);
 488         if (ret) {
 489                 pr_info("event trace: Could not activate"
 490                                 "syscall entry trace point");
 491         } else {
 492                 set_bit(num, enabled_perf_enter_syscalls);
 493                 sys_perf_refcount_enter++;
 494         }
 495         mutex_unlock(&syscall_trace_lock);
 496         return ret;
 497 }
 498
 499 void perf_sysenter_disable(struct ftrace_event_call *call)
 500 {
 501         int num;
 502
 503         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 504
 505         mutex_lock(&syscall_trace_lock);
 506         sys_perf_refcount_enter--;
 507         clear_bit(num, enabled_perf_enter_syscalls);
 508         if (!sys_perf_refcount_enter)
 509                 unregister_trace_sys_enter(perf_syscall_enter, NULL);
 510         mutex_unlock(&syscall_trace_lock);
 511 }
 512
 513 static void perf_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
 514 {
 515         struct syscall_metadata *sys_data;
 516         struct syscall_trace_exit *rec;
 517         unsigned long flags;
 518         int syscall_nr;
 519         int rctx;
 520         int size;
 521
 522         syscall_nr = syscall_get_nr(current, regs);
 523         if (!test_bit(syscall_nr, enabled_perf_exit_syscalls))
 524                 return;
 525
 526         sys_data = syscall_nr_to_meta(syscall_nr);
 527         if (!sys_data)
 528                 return;
 529
 530         /* We can probably do that at build time */
 531         size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
 532         size -= sizeof(u32);
 533
 534         /*
 535          * Impossible, but be paranoid with the future
 536          * How to put this check outside runtime?
 537          */
 538         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
 539                 "exit event has grown above perf buffer size"))
 540                 return;
 541
 542         rec = (struct syscall_trace_exit *)perf_trace_buf_prepare(size,
 543                                 sys_data->exit_event->id, &rctx, &flags);
 544         if (!rec)
 545                 return;
 546
 547         rec->nr = syscall_nr;
 548         rec->ret = syscall_get_return_value(current, regs);
 549
 550         perf_trace_buf_submit(rec, size, rctx, 0, 1, flags, regs);
 551 }
 552
 553 int perf_sysexit_enable(struct ftrace_event_call *call)
 554 {
 555         int ret = 0;
 556         int num;
 557
 558         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 559
 560         mutex_lock(&syscall_trace_lock);
 561         if (!sys_perf_refcount_exit)
 562                 ret = register_trace_sys_exit(perf_syscall_exit, NULL);
 563         if (ret) {
 564                 pr_info("event trace: Could not activate"
 565                                 "syscall exit trace point");
 566         } else {
 567                 set_bit(num, enabled_perf_exit_syscalls);
 568                 sys_perf_refcount_exit++;
 569         }
 570         mutex_unlock(&syscall_trace_lock);
 571         return ret;
 572 }
 573
 574 void perf_sysexit_disable(struct ftrace_event_call *call)
 575 {
 576         int num;
 577
 578         num = ((struct syscall_metadata *)call->data)->syscall_nr;
 579
 580         mutex_lock(&syscall_trace_lock);
 581         sys_perf_refcount_exit--;
 582         clear_bit(num, enabled_perf_exit_syscalls);
 583         if (!sys_perf_refcount_exit)
 584                 unregister_trace_sys_exit(perf_syscall_exit, NULL);
 585         mutex_unlock(&syscall_trace_lock);
 586 }
 587
 588 #endif /* CONFIG_PERF_EVENTS */
 589