git.ozlabs.org Git - ccan/blob - ccan/failtest/failtest.c

   1 /* Licensed under LGPL - see LICENSE file for details */
   2 #include <ccan/failtest/failtest.h>
   3 #include <stdarg.h>
   4 #include <string.h>
   5 #include <stdio.h>
   6 #include <stdarg.h>
   7 #include <ctype.h>
   8 #include <err.h>
   9 #include <unistd.h>
  10 #include <poll.h>
  11 #include <errno.h>
  12 #include <sys/types.h>
  13 #include <sys/wait.h>
  14 #include <sys/stat.h>
  15 #include <sys/time.h>
  16 #include <signal.h>
  17 #include <assert.h>
  18 #include <ccan/time/time.h>
  19 #include <ccan/read_write_all/read_write_all.h>
  20 #include <ccan/failtest/failtest_proto.h>
  21 #include <ccan/build_assert/build_assert.h>
  22 #include <ccan/hash/hash.h>
  23 #include <ccan/htable/htable_type.h>
  24 #include <ccan/str/str.h>
  25
  26 enum failtest_result (*failtest_hook)(struct tlist_calls *);
  27
  28 static int tracefd = -1;
  29 static int warnfd;
  30
  31 unsigned int failtest_timeout_ms = 20000;
  32
  33 const char *failpath;
  34 const char *debugpath;
  35
  36 enum info_type {
  37         WRITE,
  38         RELEASE_LOCKS,
  39         FAILURE,
  40         SUCCESS,
  41         UNEXPECTED
  42 };
  43
  44 struct lock_info {
  45         int fd;
  46         /* end is inclusive: you can't have a 0-byte lock. */
  47         off_t start, end;
  48         int type;
  49 };
  50
  51 /* We hash the call location together with its backtrace. */
  52 static size_t hash_call(const struct failtest_call *call)
  53 {
  54         return hash(call->file, strlen(call->file),
  55                     hash(&call->line, 1,
  56                          hash(call->backtrace, call->backtrace_num,
  57                               call->type)));
  58 }
  59
  60 static bool call_eq(const struct failtest_call *call1,
  61                     const struct failtest_call *call2)
  62 {
  63         unsigned int i;
  64
  65         if (strcmp(call1->file, call2->file) != 0
  66             || call1->line != call2->line
  67             || call1->type != call2->type
  68             || call1->backtrace_num != call2->backtrace_num)
  69                 return false;
  70
  71         for (i = 0; i < call1->backtrace_num; i++)
  72                 if (call1->backtrace[i] != call2->backtrace[i])
  73                         return false;
  74
  75         return true;
  76 }
  77
  78 /* Defines struct failtable. */
  79 HTABLE_DEFINE_TYPE(struct failtest_call, (struct failtest_call *), hash_call,
  80                    call_eq, failtable);
  81
  82 bool (*failtest_exit_check)(struct tlist_calls *history);
  83
  84 static struct tlist_calls history = TLIST_INIT(history);
  85 static int control_fd = -1;
  86 static struct timeval start;
  87 static bool probing = false;
  88 static struct failtable failtable;
  89
  90 static struct write_call *child_writes = NULL;
  91 static unsigned int child_writes_num = 0;
  92
  93 static pid_t lock_owner;
  94 static struct lock_info *locks = NULL;
  95 static unsigned int lock_num = 0;
  96
  97 static pid_t orig_pid;
  98
  99 static const char info_to_arg[] = "mceoxprwf";
 100
 101 /* Dummy call used for failtest_undo wrappers. */
 102 static struct failtest_call unrecorded_call;
 103
 104 #if HAVE_BACKTRACE
 105 #include <execinfo.h>
 106
 107 static void **get_backtrace(unsigned int *num)
 108 {
 109         static unsigned int max_back = 100;
 110         void **ret;
 111
 112 again:
 113         ret = malloc(max_back * sizeof(void *));
 114         *num = backtrace(ret, max_back);
 115         if (*num == max_back) {
 116                 free(ret);
 117                 max_back *= 2;
 118                 goto again;
 119         }
 120         return ret;
 121 }
 122 #else
 123 /* This will test slightly less, since will consider all of the same
 124  * calls as identical.  But, it's slightly faster! */
 125 static void **get_backtrace(unsigned int *num)
 126 {
 127         *num = 0;
 128         return NULL;
 129 }
 130 #endif /* HAVE_BACKTRACE */
 131
 132 static struct failtest_call *add_history_(enum failtest_call_type type,
 133                                           const char *file,
 134                                           unsigned int line,
 135                                           const void *elem,
 136                                           size_t elem_size)
 137 {
 138         struct failtest_call *call;
 139
 140         /* NULL file is how we suppress failure. */
 141         if (!file)
 142                 return &unrecorded_call;
 143
 144         call = malloc(sizeof *call);
 145         call->type = type;
 146         call->file = file;
 147         call->line = line;
 148         call->cleanup = NULL;
 149         call->backtrace = get_backtrace(&call->backtrace_num);
 150         memcpy(&call->u, elem, elem_size);
 151         tlist_add_tail(&history, call, list);
 152         return call;
 153 }
 154
 155 #define add_history(type, file, line, elem) \
 156         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
 157
 158 /* We do a fake call inside a sizeof(), to check types. */
 159 #define set_cleanup(call, clean, type)                  \
 160         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL),1), (clean))
 161
 162
 163 /* Dup the fd to a high value (out of the way I hope!), and close the old fd. */
 164 static int move_fd_to_high(int fd)
 165 {
 166         int i;
 167
 168         for (i = FD_SETSIZE - 1; i >= 0; i--) {
 169                 if (fcntl(i, F_GETFL) == -1 && errno == EBADF) {
 170                         if (dup2(fd, i) == -1)
 171                                 err(1, "Failed to dup fd %i to %i", fd, i);
 172                         close(fd);
 173                         return i;
 174                 }
 175         }
 176         /* Nothing?  Really?  Er... ok? */
 177         return fd;
 178 }
 179
 180 static bool read_write_info(int fd)
 181 {
 182         struct write_call *w;
 183         char *buf;
 184
 185         /* We don't need all of this, but it's simple. */
 186         child_writes = realloc(child_writes,
 187                                (child_writes_num+1) * sizeof(child_writes[0]));
 188         w = &child_writes[child_writes_num];
 189         if (!read_all(fd, w, sizeof(*w)))
 190                 return false;
 191
 192         w->buf = buf = malloc(w->count);
 193         if (!read_all(fd, buf, w->count))
 194                 return false;
 195
 196         child_writes_num++;
 197         return true;
 198 }
 199
 200 static char *failpath_string(void)
 201 {
 202         struct failtest_call *i;
 203         char *ret = strdup("");
 204         unsigned len = 0;
 205
 206         /* Inefficient, but who cares? */
 207         tlist_for_each(&history, i, list) {
 208                 ret = realloc(ret, len + 2);
 209                 ret[len] = info_to_arg[i->type];
 210                 if (i->fail)
 211                         ret[len] = toupper(ret[len]);
 212                 ret[++len] = '\0';
 213         }
 214         return ret;
 215 }
 216
 217 static void warn_via_fd(int e, const char *fmt, va_list ap)
 218 {
 219         char *p = failpath_string();
 220
 221         vdprintf(warnfd, fmt, ap);
 222         if (e != -1)
 223                 dprintf(warnfd, ": %s", strerror(e));
 224         dprintf(warnfd, " [%s]\n", p);
 225         free(p);
 226 }
 227
 228 static void fwarn(const char *fmt, ...)
 229 {
 230         va_list ap;
 231         int e = errno;
 232
 233         va_start(ap, fmt);
 234         warn_via_fd(e, fmt, ap);
 235         va_end(ap);
 236 }
 237
 238
 239 static void fwarnx(const char *fmt, ...)
 240 {
 241         va_list ap;
 242
 243         va_start(ap, fmt);
 244         warn_via_fd(-1, fmt, ap);
 245         va_end(ap);
 246 }
 247
 248 static void tell_parent(enum info_type type)
 249 {
 250         if (control_fd != -1)
 251                 write_all(control_fd, &type, sizeof(type));
 252 }
 253
 254 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 255 {
 256         va_list ap;
 257         char *path = failpath_string();
 258
 259         va_start(ap, fmt);
 260         vfprintf(stderr, fmt, ap);
 261         va_end(ap);
 262
 263         fprintf(stderr, "%.*s", (int)outlen, out);
 264         printf("To reproduce: --failpath=%s\n", path);
 265         free(path);
 266         tell_parent(FAILURE);
 267         exit(1);
 268 }
 269
 270 static void trace(const char *fmt, ...)
 271 {
 272         va_list ap;
 273
 274         if (tracefd == -1)
 275                 return;
 276
 277         va_start(ap, fmt);
 278         vdprintf(tracefd, fmt, ap);
 279         va_end(ap);
 280 }
 281
 282 static pid_t child;
 283
 284 static void hand_down(int signum)
 285 {
 286         kill(child, signum);
 287 }
 288
 289 static void release_locks(void)
 290 {
 291         /* Locks were never acquired/reacquired? */
 292         if (lock_owner == 0)
 293                 return;
 294
 295         /* We own them?  Release them all. */
 296         if (lock_owner == getpid()) {
 297                 unsigned int i;
 298                 struct flock fl;
 299                 fl.l_type = F_UNLCK;
 300                 fl.l_whence = SEEK_SET;
 301                 fl.l_start = 0;
 302                 fl.l_len = 0;
 303
 304                 for (i = 0; i < lock_num; i++)
 305                         fcntl(locks[i].fd, F_SETLK, &fl);
 306         } else {
 307                 /* Our parent must have them; pass request up. */
 308                 enum info_type type = RELEASE_LOCKS;
 309                 assert(control_fd != -1);
 310                 write_all(control_fd, &type, sizeof(type));
 311         }
 312         lock_owner = 0;
 313 }
 314
 315 /* off_t is a signed type.  Getting its max is non-trivial. */
 316 static off_t off_max(void)
 317 {
 318         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
 319         if (sizeof(off_t) == 4)
 320                 return (off_t)0x7FFFFFF;
 321         else
 322                 return (off_t)0x7FFFFFFFFFFFFFFULL;
 323 }
 324
 325 static void get_locks(void)
 326 {
 327         unsigned int i;
 328         struct flock fl;
 329
 330         if (lock_owner == getpid())
 331                 return;
 332
 333         if (lock_owner != 0) {
 334                 enum info_type type = RELEASE_LOCKS;
 335                 assert(control_fd != -1);
 336                 write_all(control_fd, &type, sizeof(type));
 337         }
 338
 339         fl.l_whence = SEEK_SET;
 340
 341         for (i = 0; i < lock_num; i++) {
 342                 fl.l_type = locks[i].type;
 343                 fl.l_start = locks[i].start;
 344                 if (locks[i].end == off_max())
 345                         fl.l_len = 0;
 346                 else
 347                         fl.l_len = locks[i].end - locks[i].start + 1;
 348
 349                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
 350                         abort();
 351         }
 352         lock_owner = getpid();
 353 }
 354
 355 struct saved_file {
 356         struct saved_file *next;
 357         int fd;
 358         void *contents;
 359         off_t off, len;
 360 };
 361
 362 static struct saved_file *save_file(struct saved_file *next, int fd)
 363 {
 364         struct saved_file *s = malloc(sizeof(*s));
 365
 366         s->next = next;
 367         s->fd = fd;
 368         s->off = lseek(fd, 0, SEEK_CUR);
 369         /* Special file?  Erk... */
 370         assert(s->off != -1);
 371         s->len = lseek(fd, 0, SEEK_END);
 372         lseek(fd, 0, SEEK_SET);
 373         s->contents = malloc(s->len);
 374         if (read(fd, s->contents, s->len) != s->len)
 375                 err(1, "Failed to save %zu bytes", (size_t)s->len);
 376         lseek(fd, s->off, SEEK_SET);
 377         return s;
 378 }
 379
 380 /* We have little choice but to save and restore open files: mmap means we
 381  * can really intercept changes in the child.
 382  *
 383  * We could do non-mmap'ed files on demand, however. */
 384 static struct saved_file *save_files(void)
 385 {
 386         struct saved_file *files = NULL;
 387         struct failtest_call *i;
 388
 389         /* Figure out the set of live fds. */
 390         tlist_for_each_rev(&history, i, list) {
 391                 if (i->type == FAILTEST_OPEN) {
 392                         int fd = i->u.open.ret;
 393                         /* Only do successful, writable fds. */
 394                         if (fd < 0)
 395                                 continue;
 396
 397                         /* If it was closed, cleanup == NULL. */
 398                         if (!i->cleanup)
 399                                 continue;
 400
 401                         if ((i->u.open.flags & O_RDWR) == O_RDWR) {
 402                                 files = save_file(files, fd);
 403                         } else if ((i->u.open.flags & O_WRONLY)
 404                                    == O_WRONLY) {
 405                                 /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
 406                                 abort();
 407                         }
 408                 }
 409         }
 410
 411         return files;
 412 }
 413
 414 static void restore_files(struct saved_file *s)
 415 {
 416         while (s) {
 417                 struct saved_file *next = s->next;
 418
 419                 lseek(s->fd, 0, SEEK_SET);
 420                 if (write(s->fd, s->contents, s->len) != s->len)
 421                         err(1, "Failed to restore %zu bytes", (size_t)s->len);
 422                 if (ftruncate(s->fd, s->len) != 0)
 423                         err(1, "Failed to trim file to length %zu",
 424                             (size_t)s->len);
 425                 free(s->contents);
 426                 lseek(s->fd, s->off, SEEK_SET);
 427                 free(s);
 428                 s = next;
 429         }
 430 }
 431
 432 static void free_files(struct saved_file *s)
 433 {
 434         while (s) {
 435                 struct saved_file *next = s->next;
 436                 free(s->contents);
 437                 free(s);
 438                 s = next;
 439         }
 440 }
 441
 442 static void free_call(struct failtest_call *call)
 443 {
 444         /* We don't do this in cleanup: needed even for failed opens. */
 445         if (call->type == FAILTEST_OPEN)
 446                 free((char *)call->u.open.pathname);
 447         free(call->backtrace);
 448         tlist_del_from(&history, call, list);
 449         free(call);
 450 }
 451
 452 /* Free up memory, so valgrind doesn't report leaks. */
 453 static void free_everything(void)
 454 {
 455         struct failtest_call *i;
 456
 457         while ((i = tlist_top(&history, struct failtest_call, list)) != NULL)
 458                 free_call(i);
 459
 460         failtable_clear(&failtable);
 461 }
 462
 463 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
 464 {
 465         struct failtest_call *i;
 466
 467         /* For children, we don't care if they "failed" the testing. */
 468         if (control_fd != -1)
 469                 status = 0;
 470
 471         if (forced_cleanup) {
 472                 /* We didn't actually do final operation: remove it. */
 473                 i = tlist_tail(&history, struct failtest_call, list);
 474                 free_call(i);
 475         }
 476
 477         /* Cleanup everything, in reverse order. */
 478         tlist_for_each_rev(&history, i, list) {
 479                 if (!i->cleanup)
 480                         continue;
 481                 if (!forced_cleanup) {
 482                         printf("Leak at %s:%u: --failpath=%s\n",
 483                                i->file, i->line, failpath_string());
 484                         status = 1;
 485                 }
 486                 i->cleanup(&i->u);
 487         }
 488
 489         free_everything();
 490         tell_parent(SUCCESS);
 491         exit(status);
 492 }
 493
 494 static bool should_fail(struct failtest_call *call)
 495 {
 496         int status;
 497         int control[2], output[2];
 498         enum info_type type = UNEXPECTED;
 499         char *out = NULL;
 500         size_t outlen = 0;
 501         struct saved_file *files;
 502         struct failtest_call *dup;
 503
 504         if (call == &unrecorded_call)
 505                 return false;
 506
 507         if (failpath) {
 508                 /* + means continue after end, like normal. */
 509                 if (*failpath == '+')
 510                         failpath = NULL;
 511                 else if (*failpath == '\0') {
 512                         /* Continue, but don't inject errors. */
 513                         return call->fail = false;
 514                 } else {
 515                         if (tolower((unsigned char)*failpath)
 516                             != info_to_arg[call->type])
 517                                 errx(1, "Failpath expected '%c' got '%c'\n",
 518                                      info_to_arg[call->type], *failpath);
 519                         call->fail = cisupper(*(failpath++));
 520                         return call->fail;
 521                 }
 522         }
 523
 524         /* Attach debugger if they asked for it. */
 525         if (debugpath) {
 526                 char *path = failpath_string();
 527
 528                 if (streq(path, debugpath)) {
 529                         char str[80];
 530
 531                         /* Don't timeout. */
 532                         signal(SIGUSR1, SIG_IGN);
 533                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
 534                                 getpid(), getpid());
 535                         if (system(str) == 0)
 536                                 sleep(5);
 537                 } else if (!strstarts(path, debugpath)) {
 538                         fprintf(stderr, "--debugpath not followed: %s\n", path);
 539                         debugpath = NULL;
 540                 }
 541                 free(path);
 542         }
 543
 544         /* Are we probing?  If so, we never fail twice. */
 545         if (probing)
 546                 return call->fail = false;
 547
 548         /* Don't more than once in the same place. */
 549         dup = failtable_get(&failtable, call);
 550         if (dup)
 551                 return call->fail = false;
 552
 553         if (failtest_hook) {
 554                 switch (failtest_hook(&history)) {
 555                 case FAIL_OK:
 556                         break;
 557                 case FAIL_PROBE:
 558                         probing = true;
 559                         break;
 560                 case FAIL_DONT_FAIL:
 561                         call->fail = false;
 562                         return false;
 563                 default:
 564                         abort();
 565                 }
 566         }
 567
 568         /* Add it to our table of calls. */
 569         failtable_add(&failtable, call);
 570
 571         files = save_files();
 572
 573         /* We're going to fail in the child. */
 574         call->fail = true;
 575         if (pipe(control) != 0 || pipe(output) != 0)
 576                 err(1, "opening pipe");
 577
 578         /* Prevent double-printing (in child and parent) */
 579         fflush(stdout);
 580         child = fork();
 581         if (child == -1)
 582                 err(1, "forking failed");
 583
 584         if (child == 0) {
 585                 if (tracefd != -1) {
 586                         struct timeval diff;
 587                         const char *p;
 588                         char *failpath;
 589                         struct failtest_call *c;
 590
 591                         c = tlist_tail(&history, struct failtest_call, list);
 592                         diff = time_sub(time_now(), start);
 593                         failpath = failpath_string();
 594                         trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
 595                               (int)diff.tv_sec, (int)diff.tv_usec / 10000,
 596                               failpath);
 597                         free(failpath);
 598                         p = strrchr(c->file, '/');
 599                         if (p)
 600                                 trace("%s", p+1);
 601                         else
 602                                 trace("%s", c->file);
 603                         trace(":%u)\n", c->line);
 604                 }
 605                 close(control[0]);
 606                 close(output[0]);
 607                 dup2(output[1], STDOUT_FILENO);
 608                 dup2(output[1], STDERR_FILENO);
 609                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
 610                         close(output[1]);
 611                 control_fd = move_fd_to_high(control[1]);
 612                 /* Valgrind spots the leak if we don't free these. */
 613                 free_files(files);
 614                 return true;
 615         }
 616
 617         signal(SIGUSR1, hand_down);
 618
 619         close(control[1]);
 620         close(output[1]);
 621
 622         /* We grab output so we can display it; we grab writes so we
 623          * can compare. */
 624         do {
 625                 struct pollfd pfd[2];
 626                 int ret;
 627
 628                 pfd[0].fd = output[0];
 629                 pfd[0].events = POLLIN|POLLHUP;
 630                 pfd[1].fd = control[0];
 631                 pfd[1].events = POLLIN|POLLHUP;
 632
 633                 if (type == SUCCESS)
 634                         ret = poll(pfd, 1, failtest_timeout_ms);
 635                 else
 636                         ret = poll(pfd, 2, failtest_timeout_ms);
 637
 638                 if (ret == 0)
 639                         hand_down(SIGUSR1);
 640                 if (ret < 0) {
 641                         if (errno == EINTR)
 642                                 continue;
 643                         err(1, "Poll returned %i", ret);
 644                 }
 645
 646                 if (pfd[0].revents & POLLIN) {
 647                         ssize_t len;
 648
 649                         out = realloc(out, outlen + 8192);
 650                         len = read(output[0], out + outlen, 8192);
 651                         outlen += len;
 652                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 653                         if (read_all(control[0], &type, sizeof(type))) {
 654                                 if (type == WRITE) {
 655                                         if (!read_write_info(control[0]))
 656                                                 break;
 657                                 } else if (type == RELEASE_LOCKS) {
 658                                         release_locks();
 659                                         /* FIXME: Tell them we're done... */
 660                                 }
 661                         }
 662                 } else if (pfd[0].revents & POLLHUP) {
 663                         break;
 664                 }
 665         } while (type != FAILURE);
 666
 667         close(output[0]);
 668         close(control[0]);
 669         waitpid(child, &status, 0);
 670         if (!WIFEXITED(status)) {
 671                 if (WTERMSIG(status) == SIGUSR1)
 672                         child_fail(out, outlen, "Timed out");
 673                 else
 674                         child_fail(out, outlen, "Killed by signal %u: ",
 675                                    WTERMSIG(status));
 676         }
 677         /* Child printed failure already, just pass up exit code. */
 678         if (type == FAILURE) {
 679                 fprintf(stderr, "%.*s", (int)outlen, out);
 680                 tell_parent(type);
 681                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 682         }
 683         if (WEXITSTATUS(status) != 0)
 684                 child_fail(out, outlen, "Exited with status %i: ",
 685                            WEXITSTATUS(status));
 686
 687         free(out);
 688         signal(SIGUSR1, SIG_DFL);
 689
 690         restore_files(files);
 691
 692         /* Only child does probe. */
 693         probing = false;
 694
 695         /* We continue onwards without failing. */
 696         call->fail = false;
 697         return false;
 698 }
 699
 700 static void cleanup_calloc(struct calloc_call *call)
 701 {
 702         free(call->ret);
 703 }
 704
 705 void *failtest_calloc(size_t nmemb, size_t size,
 706                       const char *file, unsigned line)
 707 {
 708         struct failtest_call *p;
 709         struct calloc_call call;
 710         call.nmemb = nmemb;
 711         call.size = size;
 712         p = add_history(FAILTEST_CALLOC, file, line, &call);
 713
 714         if (should_fail(p)) {
 715                 p->u.calloc.ret = NULL;
 716                 p->error = ENOMEM;
 717         } else {
 718                 p->u.calloc.ret = calloc(nmemb, size);
 719                 set_cleanup(p, cleanup_calloc, struct calloc_call);
 720         }
 721         errno = p->error;
 722         return p->u.calloc.ret;
 723 }
 724
 725 static void cleanup_malloc(struct malloc_call *call)
 726 {
 727         free(call->ret);
 728 }
 729
 730 void *failtest_malloc(size_t size, const char *file, unsigned line)
 731 {
 732         struct failtest_call *p;
 733         struct malloc_call call;
 734         call.size = size;
 735
 736         p = add_history(FAILTEST_MALLOC, file, line, &call);
 737         if (should_fail(p)) {
 738                 p->u.malloc.ret = NULL;
 739                 p->error = ENOMEM;
 740         } else {
 741                 p->u.malloc.ret = malloc(size);
 742                 set_cleanup(p, cleanup_malloc, struct malloc_call);
 743         }
 744         errno = p->error;
 745         return p->u.malloc.ret;
 746 }
 747
 748 static void cleanup_realloc(struct realloc_call *call)
 749 {
 750         free(call->ret);
 751 }
 752
 753 /* Walk back and find out if we got this ptr from a previous routine. */
 754 static void fixup_ptr_history(void *ptr)
 755 {
 756         struct failtest_call *i;
 757
 758         /* Start at end of history, work back. */
 759         tlist_for_each_rev(&history, i, list) {
 760                 switch (i->type) {
 761                 case FAILTEST_REALLOC:
 762                         if (i->u.realloc.ret == ptr) {
 763                                 i->cleanup = NULL;
 764                                 return;
 765                         }
 766                         break;
 767                 case FAILTEST_MALLOC:
 768                         if (i->u.malloc.ret == ptr) {
 769                                 i->cleanup = NULL;
 770                                 return;
 771                         }
 772                         break;
 773                 case FAILTEST_CALLOC:
 774                         if (i->u.calloc.ret == ptr) {
 775                                 i->cleanup = NULL;
 776                                 return;
 777                         }
 778                         break;
 779                 default:
 780                         break;
 781                 }
 782         }
 783 }
 784
 785 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 786 {
 787         struct failtest_call *p;
 788         struct realloc_call call;
 789         call.size = size;
 790         p = add_history(FAILTEST_REALLOC, file, line, &call);
 791
 792         /* FIXME: Try one child moving allocation, one not. */
 793         if (should_fail(p)) {
 794                 p->u.realloc.ret = NULL;
 795                 p->error = ENOMEM;
 796         } else {
 797                 /* Don't catch this one in the history fixup... */
 798                 p->u.realloc.ret = NULL;
 799                 fixup_ptr_history(ptr);
 800                 p->u.realloc.ret = realloc(ptr, size);
 801                 set_cleanup(p, cleanup_realloc, struct realloc_call);
 802         }
 803         errno = p->error;
 804         return p->u.realloc.ret;
 805 }
 806
 807 void failtest_free(void *ptr)
 808 {
 809         fixup_ptr_history(ptr);
 810         free(ptr);
 811 }
 812
 813 static void cleanup_open(struct open_call *call)
 814 {
 815         close(call->ret);
 816 }
 817
 818 int failtest_open(const char *pathname,
 819                   const char *file, unsigned line, ...)
 820 {
 821         struct failtest_call *p;
 822         struct open_call call;
 823         va_list ap;
 824
 825         call.pathname = strdup(pathname);
 826         va_start(ap, line);
 827         call.flags = va_arg(ap, int);
 828         if (call.flags & O_CREAT) {
 829                 call.mode = va_arg(ap, int);
 830                 va_end(ap);
 831         }
 832         p = add_history(FAILTEST_OPEN, file, line, &call);
 833         /* Avoid memory leak! */
 834         if (p == &unrecorded_call)
 835                 free((char *)call.pathname);
 836         p->u.open.ret = open(pathname, call.flags, call.mode);
 837
 838         if (p->u.open.ret == -1) {
 839                 p->fail = false;
 840                 p->error = errno;
 841         } else if (should_fail(p)) {
 842                 close(p->u.open.ret);
 843                 p->u.open.ret = -1;
 844                 /* FIXME: Play with error codes? */
 845                 p->error = EACCES;
 846         } else {
 847                 set_cleanup(p, cleanup_open, struct open_call);
 848         }
 849         errno = p->error;
 850         return p->u.open.ret;
 851 }
 852
 853 static void cleanup_pipe(struct pipe_call *call)
 854 {
 855         if (!call->closed[0])
 856                 close(call->fds[0]);
 857         if (!call->closed[1])
 858                 close(call->fds[1]);
 859 }
 860
 861 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
 862 {
 863         struct failtest_call *p;
 864         struct pipe_call call;
 865
 866         p = add_history(FAILTEST_PIPE, file, line, &call);
 867         if (should_fail(p)) {
 868                 p->u.open.ret = -1;
 869                 /* FIXME: Play with error codes? */
 870                 p->error = EMFILE;
 871         } else {
 872                 p->u.pipe.ret = pipe(p->u.pipe.fds);
 873                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
 874                 set_cleanup(p, cleanup_pipe, struct pipe_call);
 875         }
 876         /* This causes valgrind to notice if they use pipefd[] after failure */
 877         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
 878         errno = p->error;
 879         return p->u.pipe.ret;
 880 }
 881
 882 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
 883                        const char *file, unsigned line)
 884 {
 885         struct failtest_call *p;
 886         struct read_call call;
 887         call.fd = fd;
 888         call.buf = buf;
 889         call.count = count;
 890         call.off = off;
 891         p = add_history(FAILTEST_READ, file, line, &call);
 892
 893         /* FIXME: Try partial read returns. */
 894         if (should_fail(p)) {
 895                 p->u.read.ret = -1;
 896                 p->error = EIO;
 897         } else {
 898                 p->u.read.ret = pread(fd, buf, count, off);
 899         }
 900         errno = p->error;
 901         return p->u.read.ret;
 902 }
 903
 904 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
 905                         const char *file, unsigned line)
 906 {
 907         struct failtest_call *p;
 908         struct write_call call;
 909
 910         call.fd = fd;
 911         call.buf = buf;
 912         call.count = count;
 913         call.off = off;
 914         p = add_history(FAILTEST_WRITE, file, line, &call);
 915
 916         /* If we're a child, we need to make sure we write the same thing
 917          * to non-files as the parent does, so tell it. */
 918         if (control_fd != -1 && off == (off_t)-1) {
 919                 enum info_type type = WRITE;
 920
 921                 write_all(control_fd, &type, sizeof(type));
 922                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
 923                 write_all(control_fd, buf, count);
 924         }
 925
 926         /* FIXME: Try partial write returns. */
 927         if (should_fail(p)) {
 928                 p->u.write.ret = -1;
 929                 p->error = EIO;
 930         } else {
 931                 /* FIXME: We assume same write order in parent and child */
 932                 if (off == (off_t)-1 && child_writes_num != 0) {
 933                         if (child_writes[0].fd != fd)
 934                                 errx(1, "Child wrote to fd %u, not %u?",
 935                                      child_writes[0].fd, fd);
 936                         if (child_writes[0].off != p->u.write.off)
 937                                 errx(1, "Child wrote to offset %zu, not %zu?",
 938                                      (size_t)child_writes[0].off,
 939                                      (size_t)p->u.write.off);
 940                         if (child_writes[0].count != count)
 941                                 errx(1, "Child wrote length %zu, not %zu?",
 942                                      child_writes[0].count, count);
 943                         if (memcmp(child_writes[0].buf, buf, count)) {
 944                                 child_fail(NULL, 0,
 945                                            "Child wrote differently to"
 946                                            " fd %u than we did!\n", fd);
 947                         }
 948                         free((char *)child_writes[0].buf);
 949                         child_writes_num--;
 950                         memmove(&child_writes[0], &child_writes[1],
 951                                 sizeof(child_writes[0]) * child_writes_num);
 952
 953                         /* Is this is a socket or pipe, child wrote it
 954                            already. */
 955                         if (p->u.write.off == (off_t)-1) {
 956                                 p->u.write.ret = count;
 957                                 errno = p->error;
 958                                 return p->u.write.ret;
 959                         }
 960                 }
 961                 p->u.write.ret = pwrite(fd, buf, count, off);
 962         }
 963         errno = p->error;
 964         return p->u.write.ret;
 965 }
 966
 967 ssize_t failtest_read(int fd, void *buf, size_t count,
 968                       const char *file, unsigned line)
 969 {
 970         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 971                               file, line);
 972 }
 973
 974 ssize_t failtest_write(int fd, const void *buf, size_t count,
 975                        const char *file, unsigned line)
 976 {
 977         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 978                                file, line);
 979 }
 980
 981 static struct lock_info *WARN_UNUSED_RESULT
 982 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 983 {
 984         unsigned int i;
 985         struct lock_info *l;
 986
 987         for (i = 0; i < lock_num; i++) {
 988                 l = &locks[i];
 989
 990                 if (l->fd != fd)
 991                         continue;
 992                 /* Four cases we care about:
 993                  * Start overlap:
 994                  *      l =    |      |
 995                  *      new = |   |
 996                  * Mid overlap:
 997                  *      l =    |      |
 998                  *      new =    |  |
 999                  * End overlap:
1000                  *      l =    |      |
1001                  *      new =      |    |
1002                  * Total overlap:
1003                  *      l =    |      |
1004                  *      new = |         |
1005                  */
1006                 if (start > l->start && end < l->end) {
1007                         /* Mid overlap: trim entry, add new one. */
1008                         off_t new_start, new_end;
1009                         new_start = end + 1;
1010                         new_end = l->end;
1011                         l->end = start - 1;
1012                         locks = add_lock(locks,
1013                                          fd, new_start, new_end, l->type);
1014                         l = &locks[i];
1015                 } else if (start <= l->start && end >= l->end) {
1016                         /* Total overlap: eliminate entry. */
1017                         l->end = 0;
1018                         l->start = 1;
1019                 } else if (end >= l->start && end < l->end) {
1020                         /* Start overlap: trim entry. */
1021                         l->start = end + 1;
1022                 } else if (start > l->start && start <= l->end) {
1023                         /* End overlap: trim entry. */
1024                         l->end = start-1;
1025                 }
1026                 /* Nothing left?  Remove it. */
1027                 if (l->end < l->start) {
1028                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
1029                         i--;
1030                 }
1031         }
1032
1033         if (type != F_UNLCK) {
1034                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
1035                 l = &locks[lock_num++];
1036                 l->fd = fd;
1037                 l->start = start;
1038                 l->end = end;
1039                 l->type = type;
1040         }
1041         return locks;
1042 }
1043
1044 /* We trap this so we can record it: we don't fail it. */
1045 int failtest_close(int fd, const char *file, unsigned line)
1046 {
1047         struct failtest_call *i;
1048         struct close_call call;
1049         struct failtest_call *p;
1050
1051         call.fd = fd;
1052         p = add_history(FAILTEST_CLOSE, file, line, &call);
1053         p->fail = false;
1054
1055         /* Consume close from failpath. */
1056         if (failpath)
1057                 if (should_fail(p))
1058                         abort();
1059
1060         if (fd < 0)
1061                 return close(fd);
1062
1063         /* Trace history to find source of fd. */
1064         tlist_for_each_rev(&history, i, list) {
1065                 switch (i->type) {
1066                 case FAILTEST_PIPE:
1067                         /* From a pipe? */
1068                         if (i->u.pipe.fds[0] == fd) {
1069                                 assert(!i->u.pipe.closed[0]);
1070                                 i->u.pipe.closed[0] = true;
1071                                 if (i->u.pipe.closed[1])
1072                                         i->cleanup = NULL;
1073                                 goto out;
1074                         }
1075                         if (i->u.pipe.fds[1] == fd) {
1076                                 assert(!i->u.pipe.closed[1]);
1077                                 i->u.pipe.closed[1] = true;
1078                                 if (i->u.pipe.closed[0])
1079                                         i->cleanup = NULL;
1080                                 goto out;
1081                         }
1082                         break;
1083                 case FAILTEST_OPEN:
1084                         if (i->u.open.ret == fd) {
1085                                 assert((void *)i->cleanup
1086                                        == (void *)cleanup_open);
1087                                 i->cleanup = NULL;
1088                                 goto out;
1089                         }
1090                         break;
1091                 default:
1092                         break;
1093                 }
1094         }
1095
1096 out:
1097         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
1098         return close(fd);
1099 }
1100
1101 /* Zero length means "to end of file" */
1102 static off_t end_of(off_t start, off_t len)
1103 {
1104         if (len == 0)
1105                 return off_max();
1106         return start + len - 1;
1107 }
1108
1109 /* FIXME: This only handles locks, really. */
1110 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
1111 {
1112         struct failtest_call *p;
1113         struct fcntl_call call;
1114         va_list ap;
1115
1116         call.fd = fd;
1117         call.cmd = cmd;
1118
1119         /* Argument extraction. */
1120         switch (cmd) {
1121         case F_SETFL:
1122         case F_SETFD:
1123                 va_start(ap, cmd);
1124                 call.arg.l = va_arg(ap, long);
1125                 va_end(ap);
1126                 return fcntl(fd, cmd, call.arg.l);
1127         case F_GETFD:
1128         case F_GETFL:
1129                 return fcntl(fd, cmd);
1130         case F_GETLK:
1131                 get_locks();
1132                 va_start(ap, cmd);
1133                 call.arg.fl = *va_arg(ap, struct flock *);
1134                 va_end(ap);
1135                 return fcntl(fd, cmd, &call.arg.fl);
1136         case F_SETLK:
1137         case F_SETLKW:
1138                 va_start(ap, cmd);
1139                 call.arg.fl = *va_arg(ap, struct flock *);
1140                 va_end(ap);
1141                 break;
1142         default:
1143                 /* This means you need to implement it here. */
1144                 err(1, "failtest: unknown fcntl %u", cmd);
1145         }
1146
1147         p = add_history(FAILTEST_FCNTL, file, line, &call);
1148
1149         if (should_fail(p)) {
1150                 p->u.fcntl.ret = -1;
1151                 if (p->u.fcntl.cmd == F_SETLK)
1152                         p->error = EAGAIN;
1153                 else
1154                         p->error = EDEADLK;
1155         } else {
1156                 get_locks();
1157                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1158                                        &p->u.fcntl.arg.fl);
1159                 if (p->u.fcntl.ret == -1)
1160                         p->error = errno;
1161                 else {
1162                         /* We don't handle anything else yet. */
1163                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1164                         locks = add_lock(locks,
1165                                          p->u.fcntl.fd,
1166                                          p->u.fcntl.arg.fl.l_start,
1167                                          end_of(p->u.fcntl.arg.fl.l_start,
1168                                                 p->u.fcntl.arg.fl.l_len),
1169                                          p->u.fcntl.arg.fl.l_type);
1170                 }
1171         }
1172         errno = p->error;
1173         return p->u.fcntl.ret;
1174 }
1175
1176 pid_t failtest_getpid(const char *file, unsigned line)
1177 {
1178         /* You must call failtest_init first! */
1179         assert(orig_pid);
1180         return orig_pid;
1181 }
1182
1183 void failtest_init(int argc, char *argv[])
1184 {
1185         unsigned int i;
1186
1187         orig_pid = getpid();
1188
1189         warnfd = move_fd_to_high(dup(STDERR_FILENO));
1190         for (i = 1; i < argc; i++) {
1191                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1192                         failpath = argv[i] + strlen("--failpath=");
1193                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1194                         tracefd = warnfd;
1195                         failtest_timeout_ms = -1;
1196                 } else if (!strncmp(argv[i], "--debugpath=",
1197                                     strlen("--debugpath="))) {
1198                         debugpath = argv[i] + strlen("--debugpath=");
1199                 }
1200         }
1201         failtable_init(&failtable);
1202         start = time_now();
1203 }
1204
1205 bool failtest_has_failed(void)
1206 {
1207         return control_fd != -1;
1208 }
1209
1210 void failtest_exit(int status)
1211 {
1212         if (failtest_exit_check) {
1213                 if (!failtest_exit_check(&history))
1214                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1215         }
1216
1217         failtest_cleanup(false, status);
1218 }