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;
 527
 528                 /* Pretend this last call matches whatever path wanted:
 529                  * keeps valgrind happy. */
 530                 call->fail = cisupper(debugpath[strlen(debugpath)-1]);
 531                 path = failpath_string();
 532
 533                 if (streq(path, debugpath)) {
 534                         char str[80];
 535
 536                         /* Don't timeout. */
 537                         signal(SIGUSR1, SIG_IGN);
 538                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
 539                                 getpid(), getpid());
 540                         if (system(str) == 0)
 541                                 sleep(5);
 542                 } else {
 543                         /* Ignore last character: could be upper or lower. */
 544                         path[strlen(path)-1] = '\0';
 545                         if (!strstarts(debugpath, path)) {
 546                                 fprintf(stderr,
 547                                         "--debugpath not followed: %s\n", path);
 548                                 debugpath = NULL;
 549                         }
 550                 }
 551                 free(path);
 552         }
 553
 554         /* Are we probing?  If so, we never fail twice. */
 555         if (probing)
 556                 return call->fail = false;
 557
 558         /* Don't more than once in the same place. */
 559         dup = failtable_get(&failtable, call);
 560         if (dup)
 561                 return call->fail = false;
 562
 563         if (failtest_hook) {
 564                 switch (failtest_hook(&history)) {
 565                 case FAIL_OK:
 566                         break;
 567                 case FAIL_PROBE:
 568                         probing = true;
 569                         break;
 570                 case FAIL_DONT_FAIL:
 571                         call->fail = false;
 572                         return false;
 573                 default:
 574                         abort();
 575                 }
 576         }
 577
 578         /* Add it to our table of calls. */
 579         failtable_add(&failtable, call);
 580
 581         files = save_files();
 582
 583         /* We're going to fail in the child. */
 584         call->fail = true;
 585         if (pipe(control) != 0 || pipe(output) != 0)
 586                 err(1, "opening pipe");
 587
 588         /* Prevent double-printing (in child and parent) */
 589         fflush(stdout);
 590         child = fork();
 591         if (child == -1)
 592                 err(1, "forking failed");
 593
 594         if (child == 0) {
 595                 if (tracefd != -1) {
 596                         struct timeval diff;
 597                         const char *p;
 598                         char *failpath;
 599                         struct failtest_call *c;
 600
 601                         c = tlist_tail(&history, struct failtest_call, list);
 602                         diff = time_sub(time_now(), start);
 603                         failpath = failpath_string();
 604                         trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
 605                               (int)diff.tv_sec, (int)diff.tv_usec / 10000,
 606                               failpath);
 607                         free(failpath);
 608                         p = strrchr(c->file, '/');
 609                         if (p)
 610                                 trace("%s", p+1);
 611                         else
 612                                 trace("%s", c->file);
 613                         trace(":%u)\n", c->line);
 614                 }
 615                 close(control[0]);
 616                 close(output[0]);
 617                 dup2(output[1], STDOUT_FILENO);
 618                 dup2(output[1], STDERR_FILENO);
 619                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
 620                         close(output[1]);
 621                 control_fd = move_fd_to_high(control[1]);
 622                 /* Valgrind spots the leak if we don't free these. */
 623                 free_files(files);
 624                 return true;
 625         }
 626
 627         signal(SIGUSR1, hand_down);
 628
 629         close(control[1]);
 630         close(output[1]);
 631
 632         /* We grab output so we can display it; we grab writes so we
 633          * can compare. */
 634         do {
 635                 struct pollfd pfd[2];
 636                 int ret;
 637
 638                 pfd[0].fd = output[0];
 639                 pfd[0].events = POLLIN|POLLHUP;
 640                 pfd[1].fd = control[0];
 641                 pfd[1].events = POLLIN|POLLHUP;
 642
 643                 if (type == SUCCESS)
 644                         ret = poll(pfd, 1, failtest_timeout_ms);
 645                 else
 646                         ret = poll(pfd, 2, failtest_timeout_ms);
 647
 648                 if (ret == 0)
 649                         hand_down(SIGUSR1);
 650                 if (ret < 0) {
 651                         if (errno == EINTR)
 652                                 continue;
 653                         err(1, "Poll returned %i", ret);
 654                 }
 655
 656                 if (pfd[0].revents & POLLIN) {
 657                         ssize_t len;
 658
 659                         out = realloc(out, outlen + 8192);
 660                         len = read(output[0], out + outlen, 8192);
 661                         outlen += len;
 662                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 663                         if (read_all(control[0], &type, sizeof(type))) {
 664                                 if (type == WRITE) {
 665                                         if (!read_write_info(control[0]))
 666                                                 break;
 667                                 } else if (type == RELEASE_LOCKS) {
 668                                         release_locks();
 669                                         /* FIXME: Tell them we're done... */
 670                                 }
 671                         }
 672                 } else if (pfd[0].revents & POLLHUP) {
 673                         break;
 674                 }
 675         } while (type != FAILURE);
 676
 677         close(output[0]);
 678         close(control[0]);
 679         waitpid(child, &status, 0);
 680         if (!WIFEXITED(status)) {
 681                 if (WTERMSIG(status) == SIGUSR1)
 682                         child_fail(out, outlen, "Timed out");
 683                 else
 684                         child_fail(out, outlen, "Killed by signal %u: ",
 685                                    WTERMSIG(status));
 686         }
 687         /* Child printed failure already, just pass up exit code. */
 688         if (type == FAILURE) {
 689                 fprintf(stderr, "%.*s", (int)outlen, out);
 690                 tell_parent(type);
 691                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 692         }
 693         if (WEXITSTATUS(status) != 0)
 694                 child_fail(out, outlen, "Exited with status %i: ",
 695                            WEXITSTATUS(status));
 696
 697         free(out);
 698         signal(SIGUSR1, SIG_DFL);
 699
 700         restore_files(files);
 701
 702         /* Only child does probe. */
 703         probing = false;
 704
 705         /* We continue onwards without failing. */
 706         call->fail = false;
 707         return false;
 708 }
 709
 710 static void cleanup_calloc(struct calloc_call *call)
 711 {
 712         free(call->ret);
 713 }
 714
 715 void *failtest_calloc(size_t nmemb, size_t size,
 716                       const char *file, unsigned line)
 717 {
 718         struct failtest_call *p;
 719         struct calloc_call call;
 720         call.nmemb = nmemb;
 721         call.size = size;
 722         p = add_history(FAILTEST_CALLOC, file, line, &call);
 723
 724         if (should_fail(p)) {
 725                 p->u.calloc.ret = NULL;
 726                 p->error = ENOMEM;
 727         } else {
 728                 p->u.calloc.ret = calloc(nmemb, size);
 729                 set_cleanup(p, cleanup_calloc, struct calloc_call);
 730         }
 731         errno = p->error;
 732         return p->u.calloc.ret;
 733 }
 734
 735 static void cleanup_malloc(struct malloc_call *call)
 736 {
 737         free(call->ret);
 738 }
 739
 740 void *failtest_malloc(size_t size, const char *file, unsigned line)
 741 {
 742         struct failtest_call *p;
 743         struct malloc_call call;
 744         call.size = size;
 745
 746         p = add_history(FAILTEST_MALLOC, file, line, &call);
 747         if (should_fail(p)) {
 748                 p->u.malloc.ret = NULL;
 749                 p->error = ENOMEM;
 750         } else {
 751                 p->u.malloc.ret = malloc(size);
 752                 set_cleanup(p, cleanup_malloc, struct malloc_call);
 753         }
 754         errno = p->error;
 755         return p->u.malloc.ret;
 756 }
 757
 758 static void cleanup_realloc(struct realloc_call *call)
 759 {
 760         free(call->ret);
 761 }
 762
 763 /* Walk back and find out if we got this ptr from a previous routine. */
 764 static void fixup_ptr_history(void *ptr)
 765 {
 766         struct failtest_call *i;
 767
 768         /* Start at end of history, work back. */
 769         tlist_for_each_rev(&history, i, list) {
 770                 switch (i->type) {
 771                 case FAILTEST_REALLOC:
 772                         if (i->u.realloc.ret == ptr) {
 773                                 i->cleanup = NULL;
 774                                 return;
 775                         }
 776                         break;
 777                 case FAILTEST_MALLOC:
 778                         if (i->u.malloc.ret == ptr) {
 779                                 i->cleanup = NULL;
 780                                 return;
 781                         }
 782                         break;
 783                 case FAILTEST_CALLOC:
 784                         if (i->u.calloc.ret == ptr) {
 785                                 i->cleanup = NULL;
 786                                 return;
 787                         }
 788                         break;
 789                 default:
 790                         break;
 791                 }
 792         }
 793 }
 794
 795 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 796 {
 797         struct failtest_call *p;
 798         struct realloc_call call;
 799         call.size = size;
 800         p = add_history(FAILTEST_REALLOC, file, line, &call);
 801
 802         /* FIXME: Try one child moving allocation, one not. */
 803         if (should_fail(p)) {
 804                 p->u.realloc.ret = NULL;
 805                 p->error = ENOMEM;
 806         } else {
 807                 /* Don't catch this one in the history fixup... */
 808                 p->u.realloc.ret = NULL;
 809                 fixup_ptr_history(ptr);
 810                 p->u.realloc.ret = realloc(ptr, size);
 811                 set_cleanup(p, cleanup_realloc, struct realloc_call);
 812         }
 813         errno = p->error;
 814         return p->u.realloc.ret;
 815 }
 816
 817 void failtest_free(void *ptr)
 818 {
 819         fixup_ptr_history(ptr);
 820         free(ptr);
 821 }
 822
 823 static void cleanup_open(struct open_call *call)
 824 {
 825         close(call->ret);
 826 }
 827
 828 int failtest_open(const char *pathname,
 829                   const char *file, unsigned line, ...)
 830 {
 831         struct failtest_call *p;
 832         struct open_call call;
 833         va_list ap;
 834
 835         call.pathname = strdup(pathname);
 836         va_start(ap, line);
 837         call.flags = va_arg(ap, int);
 838         if (call.flags & O_CREAT) {
 839                 call.mode = va_arg(ap, int);
 840                 va_end(ap);
 841         }
 842         p = add_history(FAILTEST_OPEN, file, line, &call);
 843         /* Avoid memory leak! */
 844         if (p == &unrecorded_call)
 845                 free((char *)call.pathname);
 846         p->u.open.ret = open(pathname, call.flags, call.mode);
 847
 848         if (p->u.open.ret == -1) {
 849                 p->fail = false;
 850                 p->error = errno;
 851         } else if (should_fail(p)) {
 852                 close(p->u.open.ret);
 853                 p->u.open.ret = -1;
 854                 /* FIXME: Play with error codes? */
 855                 p->error = EACCES;
 856         } else {
 857                 set_cleanup(p, cleanup_open, struct open_call);
 858         }
 859         errno = p->error;
 860         return p->u.open.ret;
 861 }
 862
 863 static void cleanup_pipe(struct pipe_call *call)
 864 {
 865         if (!call->closed[0])
 866                 close(call->fds[0]);
 867         if (!call->closed[1])
 868                 close(call->fds[1]);
 869 }
 870
 871 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
 872 {
 873         struct failtest_call *p;
 874         struct pipe_call call;
 875
 876         p = add_history(FAILTEST_PIPE, file, line, &call);
 877         if (should_fail(p)) {
 878                 p->u.open.ret = -1;
 879                 /* FIXME: Play with error codes? */
 880                 p->error = EMFILE;
 881         } else {
 882                 p->u.pipe.ret = pipe(p->u.pipe.fds);
 883                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
 884                 set_cleanup(p, cleanup_pipe, struct pipe_call);
 885         }
 886         /* This causes valgrind to notice if they use pipefd[] after failure */
 887         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
 888         errno = p->error;
 889         return p->u.pipe.ret;
 890 }
 891
 892 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
 893                        const char *file, unsigned line)
 894 {
 895         struct failtest_call *p;
 896         struct read_call call;
 897         call.fd = fd;
 898         call.buf = buf;
 899         call.count = count;
 900         call.off = off;
 901         p = add_history(FAILTEST_READ, file, line, &call);
 902
 903         /* FIXME: Try partial read returns. */
 904         if (should_fail(p)) {
 905                 p->u.read.ret = -1;
 906                 p->error = EIO;
 907         } else {
 908                 p->u.read.ret = pread(fd, buf, count, off);
 909         }
 910         errno = p->error;
 911         return p->u.read.ret;
 912 }
 913
 914 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
 915                         const char *file, unsigned line)
 916 {
 917         struct failtest_call *p;
 918         struct write_call call;
 919
 920         call.fd = fd;
 921         call.buf = buf;
 922         call.count = count;
 923         call.off = off;
 924         p = add_history(FAILTEST_WRITE, file, line, &call);
 925
 926         /* If we're a child, we need to make sure we write the same thing
 927          * to non-files as the parent does, so tell it. */
 928         if (control_fd != -1 && off == (off_t)-1) {
 929                 enum info_type type = WRITE;
 930
 931                 write_all(control_fd, &type, sizeof(type));
 932                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
 933                 write_all(control_fd, buf, count);
 934         }
 935
 936         /* FIXME: Try partial write returns. */
 937         if (should_fail(p)) {
 938                 p->u.write.ret = -1;
 939                 p->error = EIO;
 940         } else {
 941                 /* FIXME: We assume same write order in parent and child */
 942                 if (off == (off_t)-1 && child_writes_num != 0) {
 943                         if (child_writes[0].fd != fd)
 944                                 errx(1, "Child wrote to fd %u, not %u?",
 945                                      child_writes[0].fd, fd);
 946                         if (child_writes[0].off != p->u.write.off)
 947                                 errx(1, "Child wrote to offset %zu, not %zu?",
 948                                      (size_t)child_writes[0].off,
 949                                      (size_t)p->u.write.off);
 950                         if (child_writes[0].count != count)
 951                                 errx(1, "Child wrote length %zu, not %zu?",
 952                                      child_writes[0].count, count);
 953                         if (memcmp(child_writes[0].buf, buf, count)) {
 954                                 child_fail(NULL, 0,
 955                                            "Child wrote differently to"
 956                                            " fd %u than we did!\n", fd);
 957                         }
 958                         free((char *)child_writes[0].buf);
 959                         child_writes_num--;
 960                         memmove(&child_writes[0], &child_writes[1],
 961                                 sizeof(child_writes[0]) * child_writes_num);
 962
 963                         /* Is this is a socket or pipe, child wrote it
 964                            already. */
 965                         if (p->u.write.off == (off_t)-1) {
 966                                 p->u.write.ret = count;
 967                                 errno = p->error;
 968                                 return p->u.write.ret;
 969                         }
 970                 }
 971                 p->u.write.ret = pwrite(fd, buf, count, off);
 972         }
 973         errno = p->error;
 974         return p->u.write.ret;
 975 }
 976
 977 ssize_t failtest_read(int fd, void *buf, size_t count,
 978                       const char *file, unsigned line)
 979 {
 980         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 981                               file, line);
 982 }
 983
 984 ssize_t failtest_write(int fd, const void *buf, size_t count,
 985                        const char *file, unsigned line)
 986 {
 987         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 988                                file, line);
 989 }
 990
 991 static struct lock_info *WARN_UNUSED_RESULT
 992 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 993 {
 994         unsigned int i;
 995         struct lock_info *l;
 996
 997         for (i = 0; i < lock_num; i++) {
 998                 l = &locks[i];
 999
1000                 if (l->fd != fd)
1001                         continue;
1002                 /* Four cases we care about:
1003                  * Start overlap:
1004                  *      l =    |      |
1005                  *      new = |   |
1006                  * Mid overlap:
1007                  *      l =    |      |
1008                  *      new =    |  |
1009                  * End overlap:
1010                  *      l =    |      |
1011                  *      new =      |    |
1012                  * Total overlap:
1013                  *      l =    |      |
1014                  *      new = |         |
1015                  */
1016                 if (start > l->start && end < l->end) {
1017                         /* Mid overlap: trim entry, add new one. */
1018                         off_t new_start, new_end;
1019                         new_start = end + 1;
1020                         new_end = l->end;
1021                         l->end = start - 1;
1022                         locks = add_lock(locks,
1023                                          fd, new_start, new_end, l->type);
1024                         l = &locks[i];
1025                 } else if (start <= l->start && end >= l->end) {
1026                         /* Total overlap: eliminate entry. */
1027                         l->end = 0;
1028                         l->start = 1;
1029                 } else if (end >= l->start && end < l->end) {
1030                         /* Start overlap: trim entry. */
1031                         l->start = end + 1;
1032                 } else if (start > l->start && start <= l->end) {
1033                         /* End overlap: trim entry. */
1034                         l->end = start-1;
1035                 }
1036                 /* Nothing left?  Remove it. */
1037                 if (l->end < l->start) {
1038                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
1039                         i--;
1040                 }
1041         }
1042
1043         if (type != F_UNLCK) {
1044                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
1045                 l = &locks[lock_num++];
1046                 l->fd = fd;
1047                 l->start = start;
1048                 l->end = end;
1049                 l->type = type;
1050         }
1051         return locks;
1052 }
1053
1054 /* We trap this so we can record it: we don't fail it. */
1055 int failtest_close(int fd, const char *file, unsigned line)
1056 {
1057         struct failtest_call *i;
1058         struct close_call call;
1059         struct failtest_call *p;
1060
1061         call.fd = fd;
1062         p = add_history(FAILTEST_CLOSE, file, line, &call);
1063         p->fail = false;
1064
1065         /* Consume close from failpath. */
1066         if (failpath)
1067                 if (should_fail(p))
1068                         abort();
1069
1070         if (fd < 0)
1071                 return close(fd);
1072
1073         /* Trace history to find source of fd. */
1074         tlist_for_each_rev(&history, i, list) {
1075                 switch (i->type) {
1076                 case FAILTEST_PIPE:
1077                         /* From a pipe? */
1078                         if (i->u.pipe.fds[0] == fd) {
1079                                 assert(!i->u.pipe.closed[0]);
1080                                 i->u.pipe.closed[0] = true;
1081                                 if (i->u.pipe.closed[1])
1082                                         i->cleanup = NULL;
1083                                 goto out;
1084                         }
1085                         if (i->u.pipe.fds[1] == fd) {
1086                                 assert(!i->u.pipe.closed[1]);
1087                                 i->u.pipe.closed[1] = true;
1088                                 if (i->u.pipe.closed[0])
1089                                         i->cleanup = NULL;
1090                                 goto out;
1091                         }
1092                         break;
1093                 case FAILTEST_OPEN:
1094                         if (i->u.open.ret == fd) {
1095                                 assert((void *)i->cleanup
1096                                        == (void *)cleanup_open);
1097                                 i->cleanup = NULL;
1098                                 goto out;
1099                         }
1100                         break;
1101                 default:
1102                         break;
1103                 }
1104         }
1105
1106 out:
1107         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
1108         return close(fd);
1109 }
1110
1111 /* Zero length means "to end of file" */
1112 static off_t end_of(off_t start, off_t len)
1113 {
1114         if (len == 0)
1115                 return off_max();
1116         return start + len - 1;
1117 }
1118
1119 /* FIXME: This only handles locks, really. */
1120 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
1121 {
1122         struct failtest_call *p;
1123         struct fcntl_call call;
1124         va_list ap;
1125
1126         call.fd = fd;
1127         call.cmd = cmd;
1128
1129         /* Argument extraction. */
1130         switch (cmd) {
1131         case F_SETFL:
1132         case F_SETFD:
1133                 va_start(ap, cmd);
1134                 call.arg.l = va_arg(ap, long);
1135                 va_end(ap);
1136                 return fcntl(fd, cmd, call.arg.l);
1137         case F_GETFD:
1138         case F_GETFL:
1139                 return fcntl(fd, cmd);
1140         case F_GETLK:
1141                 get_locks();
1142                 va_start(ap, cmd);
1143                 call.arg.fl = *va_arg(ap, struct flock *);
1144                 va_end(ap);
1145                 return fcntl(fd, cmd, &call.arg.fl);
1146         case F_SETLK:
1147         case F_SETLKW:
1148                 va_start(ap, cmd);
1149                 call.arg.fl = *va_arg(ap, struct flock *);
1150                 va_end(ap);
1151                 break;
1152         default:
1153                 /* This means you need to implement it here. */
1154                 err(1, "failtest: unknown fcntl %u", cmd);
1155         }
1156
1157         p = add_history(FAILTEST_FCNTL, file, line, &call);
1158
1159         if (should_fail(p)) {
1160                 p->u.fcntl.ret = -1;
1161                 if (p->u.fcntl.cmd == F_SETLK)
1162                         p->error = EAGAIN;
1163                 else
1164                         p->error = EDEADLK;
1165         } else {
1166                 get_locks();
1167                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1168                                        &p->u.fcntl.arg.fl);
1169                 if (p->u.fcntl.ret == -1)
1170                         p->error = errno;
1171                 else {
1172                         /* We don't handle anything else yet. */
1173                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1174                         locks = add_lock(locks,
1175                                          p->u.fcntl.fd,
1176                                          p->u.fcntl.arg.fl.l_start,
1177                                          end_of(p->u.fcntl.arg.fl.l_start,
1178                                                 p->u.fcntl.arg.fl.l_len),
1179                                          p->u.fcntl.arg.fl.l_type);
1180                 }
1181         }
1182         errno = p->error;
1183         return p->u.fcntl.ret;
1184 }
1185
1186 pid_t failtest_getpid(const char *file, unsigned line)
1187 {
1188         /* You must call failtest_init first! */
1189         assert(orig_pid);
1190         return orig_pid;
1191 }
1192
1193 void failtest_init(int argc, char *argv[])
1194 {
1195         unsigned int i;
1196
1197         orig_pid = getpid();
1198
1199         warnfd = move_fd_to_high(dup(STDERR_FILENO));
1200         for (i = 1; i < argc; i++) {
1201                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1202                         failpath = argv[i] + strlen("--failpath=");
1203                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1204                         tracefd = warnfd;
1205                         failtest_timeout_ms = -1;
1206                 } else if (!strncmp(argv[i], "--debugpath=",
1207                                     strlen("--debugpath="))) {
1208                         debugpath = argv[i] + strlen("--debugpath=");
1209                 }
1210         }
1211         failtable_init(&failtable);
1212         start = time_now();
1213 }
1214
1215 bool failtest_has_failed(void)
1216 {
1217         return control_fd != -1;
1218 }
1219
1220 void failtest_exit(int status)
1221 {
1222         if (failtest_exit_check) {
1223                 if (!failtest_exit_check(&history))
1224                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1225         }
1226
1227         failtest_cleanup(false, status);
1228 }