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

   1 #include <stdarg.h>
   2 #include <string.h>
   3 #include <stdio.h>
   4 #include <stdarg.h>
   5 #include <ctype.h>
   6 #include <err.h>
   7 #include <unistd.h>
   8 #include <poll.h>
   9 #include <errno.h>
  10 #include <sys/types.h>
  11 #include <sys/wait.h>
  12 #include <sys/stat.h>
  13 #include <sys/time.h>
  14 #include <assert.h>
  15 #include <ccan/read_write_all/read_write_all.h>
  16 #include <ccan/failtest/failtest_proto.h>
  17 #include <ccan/failtest/failtest.h>
  18 #include <ccan/build_assert/build_assert.h>
  19
  20 enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
  21
  22 static int tracefd = -1;
  23
  24 unsigned int failtest_timeout_ms = 20000;
  25
  26 const char *failpath;
  27 const char *debugpath;
  28
  29 enum info_type {
  30         WRITE,
  31         RELEASE_LOCKS,
  32         FAILURE,
  33         SUCCESS,
  34         UNEXPECTED
  35 };
  36
  37 struct lock_info {
  38         int fd;
  39         /* end is inclusive: you can't have a 0-byte lock. */
  40         off_t start, end;
  41         int type;
  42 };
  43
  44 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
  45
  46 static struct failtest_call *history = NULL;
  47 static unsigned int history_num = 0;
  48 static int control_fd = -1;
  49 static struct timeval start;
  50 static unsigned int probe_count = 0;
  51
  52 static struct write_call *child_writes = NULL;
  53 static unsigned int child_writes_num = 0;
  54
  55 static pid_t lock_owner;
  56 static struct lock_info *locks = NULL;
  57 static unsigned int lock_num = 0;
  58
  59 static const char info_to_arg[] = "mceoxprwf";
  60
  61 /* Dummy call used for failtest_undo wrappers. */
  62 static struct failtest_call unrecorded_call;
  63
  64 static struct failtest_call *add_history_(enum failtest_call_type type,
  65                                           const char *file,
  66                                           unsigned int line,
  67                                           const void *elem,
  68                                           size_t elem_size)
  69 {
  70         /* NULL file is how we suppress failure. */
  71         if (!file)
  72                 return &unrecorded_call;
  73
  74         history = realloc(history, (history_num + 1) * sizeof(*history));
  75         history[history_num].type = type;
  76         history[history_num].file = file;
  77         history[history_num].line = line;
  78         history[history_num].cleanup = NULL;
  79         memcpy(&history[history_num].u, elem, elem_size);
  80         return &history[history_num++];
  81 }
  82
  83 #define add_history(type, file, line, elem) \
  84         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
  85
  86 #define set_cleanup(call, clean, type)                  \
  87         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL)), (clean))
  88
  89 static bool read_write_info(int fd)
  90 {
  91         struct write_call *w;
  92         char *buf;
  93
  94         /* We don't need all of this, but it's simple. */
  95         child_writes = realloc(child_writes,
  96                                (child_writes_num+1) * sizeof(child_writes[0]));
  97         w = &child_writes[child_writes_num];
  98         if (!read_all(fd, w, sizeof(*w)))
  99                 return false;
 100
 101         w->buf = buf = malloc(w->count);
 102         if (!read_all(fd, buf, w->count))
 103                 return false;
 104
 105         child_writes_num++;
 106         return true;
 107 }
 108
 109 static char *failpath_string(void)
 110 {
 111         unsigned int i;
 112         char *ret = malloc(history_num + 1);
 113
 114         for (i = 0; i < history_num; i++) {
 115                 ret[i] = info_to_arg[history[i].type];
 116                 if (history[i].fail)
 117                         ret[i] = toupper(ret[i]);
 118         }
 119         ret[i] = '\0';
 120         return ret;
 121 }
 122
 123 static void tell_parent(enum info_type type)
 124 {
 125         if (control_fd != -1)
 126                 write_all(control_fd, &type, sizeof(type));
 127 }
 128
 129 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 130 {
 131         va_list ap;
 132         char *path = failpath_string();
 133
 134         va_start(ap, fmt);
 135         vfprintf(stderr, fmt, ap);
 136         va_end(ap);
 137
 138         fprintf(stderr, "%.*s", (int)outlen, out);
 139         printf("To reproduce: --failpath=%s\n", path);
 140         free(path);
 141         tell_parent(FAILURE);
 142         exit(1);
 143 }
 144
 145 static pid_t child;
 146
 147 static void hand_down(int signal)
 148 {
 149         kill(child, signal);
 150 }
 151
 152 static void release_locks(void)
 153 {
 154         /* Locks were never acquired/reacquired? */
 155         if (lock_owner == 0)
 156                 return;
 157
 158         /* We own them?  Release them all. */
 159         if (lock_owner == getpid()) {
 160                 unsigned int i;
 161                 struct flock fl;
 162                 fl.l_type = F_UNLCK;
 163                 fl.l_whence = SEEK_SET;
 164                 fl.l_start = 0;
 165                 fl.l_len = 0;
 166
 167                 for (i = 0; i < lock_num; i++)
 168                         fcntl(locks[i].fd, F_SETLK, &fl);
 169         } else {
 170                 /* Our parent must have them; pass request up. */
 171                 enum info_type type = RELEASE_LOCKS;
 172                 assert(control_fd != -1);
 173                 write_all(control_fd, &type, sizeof(type));
 174         }
 175         lock_owner = 0;
 176 }
 177
 178 /* off_t is a signed type.  Getting its max is non-trivial. */
 179 static off_t off_max(void)
 180 {
 181         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
 182         if (sizeof(off_t) == 4)
 183                 return (off_t)0x7FFFFFF;
 184         else
 185                 return (off_t)0x7FFFFFFFFFFFFFFULL;
 186 }
 187
 188 static void get_locks(void)
 189 {
 190         unsigned int i;
 191         struct flock fl;
 192
 193         if (lock_owner == getpid())
 194                 return;
 195
 196         if (lock_owner != 0) {
 197                 enum info_type type = RELEASE_LOCKS;
 198                 assert(control_fd != -1);
 199                 write_all(control_fd, &type, sizeof(type));
 200         }
 201
 202         fl.l_whence = SEEK_SET;
 203
 204         for (i = 0; i < lock_num; i++) {
 205                 fl.l_type = locks[i].type;
 206                 fl.l_start = locks[i].start;
 207                 if (locks[i].end == off_max())
 208                         fl.l_len = 0;
 209                 else
 210                         fl.l_len = locks[i].end - locks[i].start + 1;
 211
 212                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
 213                         abort();
 214         }
 215         lock_owner = getpid();
 216 }
 217
 218 static void trace_str(const char *str)
 219 {
 220         ssize_t ret;
 221
 222         while ((ret = write(tracefd, str, strlen(str))) <= 0) {
 223                 str += ret;
 224                 if (!*str)
 225                         return;
 226         }
 227         err(1, "Writing trace.");
 228 }
 229
 230 struct saved_file {
 231         struct saved_file *next;
 232         int fd;
 233         void *contents;
 234         off_t off, len;
 235 };
 236
 237 static struct saved_file *save_file(struct saved_file *next, int fd)
 238 {
 239         struct saved_file *s = malloc(sizeof(*s));
 240
 241         s->next = next;
 242         s->fd = fd;
 243         s->off = lseek(fd, 0, SEEK_CUR);
 244         /* Special file?  Erk... */
 245         assert(s->off != -1);
 246         s->len = lseek(fd, 0, SEEK_END);
 247         lseek(fd, 0, SEEK_SET);
 248         s->contents = malloc(s->len);
 249         read(fd, s->contents, s->len);
 250         lseek(fd, s->off, SEEK_SET);
 251         return s;
 252 }
 253
 254 /* We have little choice but to save and restore open files: mmap means we
 255  * can really intercept changes in the child.
 256  *
 257  * We could do non-mmap'ed files on demand, however. */
 258 static struct saved_file *save_files(void)
 259 {
 260         struct saved_file *files = NULL;
 261         int i;
 262
 263         /* Figure out the set of live fds. */
 264         for (i = history_num - 2; i >= 0; i--) {
 265                 if (history[i].type == FAILTEST_OPEN) {
 266                         int fd = history[i].u.open.ret;
 267                         /* Only do successful, writable fds. */
 268                         if (fd < 0)
 269                                 continue;
 270
 271                         /* If it was closed, cleanup == NULL. */
 272                         if (!history[i].cleanup)
 273                                 continue;
 274
 275                         if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
 276                                 files = save_file(files, fd);
 277                         } else if ((history[i].u.open.flags & O_WRONLY)
 278                                    == O_WRONLY) {
 279                                 /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
 280                                 abort();
 281                         }
 282                 }
 283         }
 284
 285         return files;
 286 }
 287
 288 static void restore_files(struct saved_file *s)
 289 {
 290         while (s) {
 291                 struct saved_file *next = s->next;
 292
 293                 lseek(s->fd, 0, SEEK_SET);
 294                 write(s->fd, s->contents, s->len);
 295                 ftruncate(s->fd, s->len);
 296                 free(s->contents);
 297                 lseek(s->fd, s->off, SEEK_SET);
 298                 free(s);
 299                 s = next;
 300         }
 301 }
 302
 303 /* Free up memory, so valgrind doesn't report leaks. */
 304 static void free_everything(void)
 305 {
 306         unsigned int i;
 307
 308         /* We don't do this in cleanup: needed even for failed opens. */
 309         for (i = 0; i < history_num; i++) {
 310                 if (history[i].type == FAILTEST_OPEN)
 311                         free((char *)history[i].u.open.pathname);
 312         }
 313         free(history);
 314 }
 315
 316 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
 317 {
 318         int i;
 319
 320         if (forced_cleanup)
 321                 history_num--;
 322
 323         /* Cleanup everything, in reverse order. */
 324         for (i = history_num - 1; i >= 0; i--)
 325                 if (history[i].cleanup)
 326                         history[i].cleanup(&history[i].u);
 327
 328         free_everything();
 329
 330         if (control_fd == -1)
 331                 exit(status);
 332
 333         tell_parent(SUCCESS);
 334         exit(0);
 335 }
 336
 337 static bool should_fail(struct failtest_call *call)
 338 {
 339         int status;
 340         int control[2], output[2];
 341         enum info_type type = UNEXPECTED;
 342         char *out = NULL;
 343         size_t outlen = 0;
 344         struct saved_file *files;
 345
 346         /* Are we probing? */
 347         if (probe_count && --probe_count == 0)
 348                 failtest_cleanup(true, 0);
 349
 350         if (call == &unrecorded_call)
 351                 return false;
 352
 353         if (failpath) {
 354                 /* + means continue after end, like normal. */
 355                 if (*failpath == '+')
 356                         failpath = NULL;
 357                 else {
 358                         if (tolower(*failpath) != info_to_arg[call->type])
 359                                 errx(1, "Failpath expected '%c' got '%c'\n",
 360                                      info_to_arg[call->type], *failpath);
 361                         call->fail = isupper(*(failpath++));
 362                         return call->fail;
 363                 }
 364         }
 365
 366         /* Attach debugger if they asked for it. */
 367         if (debugpath && history_num == strlen(debugpath)) {
 368                 unsigned int i;
 369
 370                 for (i = 0; i < history_num; i++) {
 371                         char c = info_to_arg[history[i].type];
 372                         if (history[i].fail)
 373                                 c = toupper(c);
 374                         if (c != debugpath[i])
 375                                 break;
 376                 }
 377                 if (i == history_num) {
 378                         char str[80];
 379
 380                         /* Don't timeout. */
 381                         signal(SIGUSR1, SIG_IGN);
 382                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
 383                                 getpid(), getpid());
 384                         system(str);
 385                         sleep(5);
 386                 }
 387         }
 388
 389         if (failtest_hook) {
 390                 switch (failtest_hook(history, history_num)) {
 391                 case FAIL_OK:
 392                         break;
 393                 case FAIL_DONT_FAIL:
 394                         call->fail = false;
 395                         return false;
 396                 case FAIL_PROBE:
 397                         /* Already down probe path?  Stop now. */
 398                         if (probe_count)
 399                                 failtest_cleanup(true, 0);
 400                         /* FIXME: We should run *parent* and run probe until
 401                          * calls match up again. */
 402                         probe_count = 3;
 403                         break;
 404                 default:
 405                         abort();
 406                 }
 407         }
 408
 409         files = save_files();
 410
 411         /* We're going to fail in the child. */
 412         call->fail = true;
 413         if (pipe(control) != 0 || pipe(output) != 0)
 414                 err(1, "opening pipe");
 415
 416         /* Prevent double-printing (in child and parent) */
 417         fflush(stdout);
 418         child = fork();
 419         if (child == -1)
 420                 err(1, "forking failed");
 421
 422         if (child == 0) {
 423                 if (tracefd != -1) {
 424                         struct timeval now;
 425                         char str[50], *p;
 426                         gettimeofday(&now, NULL);
 427                         if (now.tv_usec < start.tv_usec) {
 428                                 now.tv_sec--;
 429                                 now.tv_usec += 1000000;
 430                         }
 431                         now.tv_usec -= start.tv_usec;
 432                         now.tv_sec -= start.tv_sec;
 433                         sprintf(str, "%u (%u.%02u): ", getpid(),
 434                                 (int)now.tv_sec, (int)now.tv_usec / 10000);
 435                         trace_str(str);
 436                         p = failpath_string();
 437                         trace_str(p);
 438                         free(p);
 439                         trace_str("(");
 440                         p = strchr(history[history_num-1].file, '/');
 441                         if (p)
 442                                 trace_str(p+1);
 443                         else
 444                                 trace_str(history[history_num-1].file);
 445                         sprintf(str, ":%u)\n", history[history_num-1].line);
 446                         trace_str(str);
 447                 }
 448                 close(control[0]);
 449                 close(output[0]);
 450                 dup2(output[1], STDOUT_FILENO);
 451                 dup2(output[1], STDERR_FILENO);
 452                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
 453                         close(output[1]);
 454                 control_fd = control[1];
 455                 return true;
 456         }
 457
 458         signal(SIGUSR1, hand_down);
 459
 460         close(control[1]);
 461         close(output[1]);
 462
 463         /* We grab output so we can display it; we grab writes so we
 464          * can compare. */
 465         do {
 466                 struct pollfd pfd[2];
 467                 int ret;
 468
 469                 pfd[0].fd = output[0];
 470                 pfd[0].events = POLLIN|POLLHUP;
 471                 pfd[1].fd = control[0];
 472                 pfd[1].events = POLLIN|POLLHUP;
 473
 474                 if (type == SUCCESS)
 475                         ret = poll(pfd, 1, failtest_timeout_ms);
 476                 else
 477                         ret = poll(pfd, 2, failtest_timeout_ms);
 478
 479                 if (ret <= 0)
 480                         hand_down(SIGUSR1);
 481
 482                 if (pfd[0].revents & POLLIN) {
 483                         ssize_t len;
 484
 485                         out = realloc(out, outlen + 8192);
 486                         len = read(output[0], out + outlen, 8192);
 487                         outlen += len;
 488                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 489                         if (read_all(control[0], &type, sizeof(type))) {
 490                                 if (type == WRITE) {
 491                                         if (!read_write_info(control[0]))
 492                                                 break;
 493                                 } else if (type == RELEASE_LOCKS) {
 494                                         release_locks();
 495                                         /* FIXME: Tell them we're done... */
 496                                 }
 497                         }
 498                 } else if (pfd[0].revents & POLLHUP) {
 499                         break;
 500                 }
 501         } while (type != FAILURE);
 502
 503         close(output[0]);
 504         close(control[0]);
 505         waitpid(child, &status, 0);
 506         if (!WIFEXITED(status)) {
 507                 if (WTERMSIG(status) == SIGUSR1)
 508                         child_fail(out, outlen, "Timed out");
 509                 else
 510                         child_fail(out, outlen, "Killed by signal %u: ",
 511                                    WTERMSIG(status));
 512         }
 513         /* Child printed failure already, just pass up exit code. */
 514         if (type == FAILURE) {
 515                 fprintf(stderr, "%.*s", (int)outlen, out);
 516                 tell_parent(type);
 517                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 518         }
 519         if (WEXITSTATUS(status) != 0)
 520                 child_fail(out, outlen, "Exited with status %i: ",
 521                            WEXITSTATUS(status));
 522
 523         free(out);
 524         signal(SIGUSR1, SIG_DFL);
 525
 526         restore_files(files);
 527
 528         /* We continue onwards without failing. */
 529         call->fail = false;
 530         return false;
 531 }
 532
 533 static void cleanup_calloc(struct calloc_call *call)
 534 {
 535         free(call->ret);
 536 }
 537
 538 void *failtest_calloc(size_t nmemb, size_t size,
 539                       const char *file, unsigned line)
 540 {
 541         struct failtest_call *p;
 542         struct calloc_call call;
 543         call.nmemb = nmemb;
 544         call.size = size;
 545         p = add_history(FAILTEST_CALLOC, file, line, &call);
 546
 547         if (should_fail(p)) {
 548                 p->u.calloc.ret = NULL;
 549                 p->error = ENOMEM;
 550         } else {
 551                 p->u.calloc.ret = calloc(nmemb, size);
 552                 set_cleanup(p, cleanup_calloc, struct calloc_call);
 553         }
 554         errno = p->error;
 555         return p->u.calloc.ret;
 556 }
 557
 558 static void cleanup_malloc(struct malloc_call *call)
 559 {
 560         free(call->ret);
 561 }
 562
 563 void *failtest_malloc(size_t size, const char *file, unsigned line)
 564 {
 565         struct failtest_call *p;
 566         struct malloc_call call;
 567         call.size = size;
 568
 569         p = add_history(FAILTEST_MALLOC, file, line, &call);
 570         if (should_fail(p)) {
 571                 p->u.calloc.ret = NULL;
 572                 p->error = ENOMEM;
 573         } else {
 574                 p->u.calloc.ret = malloc(size);
 575                 set_cleanup(p, cleanup_malloc, struct malloc_call);
 576         }
 577         errno = p->error;
 578         return p->u.calloc.ret;
 579 }
 580
 581 static void cleanup_realloc(struct realloc_call *call)
 582 {
 583         free(call->ret);
 584 }
 585
 586 /* Walk back and find out if we got this ptr from a previous routine. */
 587 static void fixup_ptr_history(void *ptr, unsigned int last)
 588 {
 589         int i;
 590
 591         /* Start at end of history, work back. */
 592         for (i = last - 1; i >= 0; i--) {
 593                 switch (history[i].type) {
 594                 case FAILTEST_REALLOC:
 595                         if (history[i].u.realloc.ret == ptr) {
 596                                 history[i].cleanup = NULL;
 597                                 return;
 598                         }
 599                         break;
 600                 case FAILTEST_MALLOC:
 601                         if (history[i].u.malloc.ret == ptr) {
 602                                 history[i].cleanup = NULL;
 603                                 return;
 604                         }
 605                         break;
 606                 case FAILTEST_CALLOC:
 607                         if (history[i].u.calloc.ret == ptr) {
 608                                 history[i].cleanup = NULL;
 609                                 return;
 610                         }
 611                         break;
 612                 default:
 613                         break;
 614                 }
 615         }
 616 }
 617
 618 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 619 {
 620         struct failtest_call *p;
 621         struct realloc_call call;
 622         call.size = size;
 623         p = add_history(FAILTEST_REALLOC, file, line, &call);
 624
 625         /* FIXME: Try one child moving allocation, one not. */
 626         if (should_fail(p)) {
 627                 p->u.realloc.ret = NULL;
 628                 p->error = ENOMEM;
 629         } else {
 630                 fixup_ptr_history(ptr, history_num-1);
 631                 p->u.realloc.ret = realloc(ptr, size);
 632                 set_cleanup(p, cleanup_realloc, struct realloc_call);
 633         }
 634         errno = p->error;
 635         return p->u.realloc.ret;
 636 }
 637
 638 void failtest_free(void *ptr)
 639 {
 640         fixup_ptr_history(ptr, history_num);
 641         free(ptr);
 642 }
 643
 644 static void cleanup_open(struct open_call *call)
 645 {
 646         close(call->ret);
 647 }
 648
 649 int failtest_open(const char *pathname,
 650                   const char *file, unsigned line, ...)
 651 {
 652         struct failtest_call *p;
 653         struct open_call call;
 654         va_list ap;
 655
 656         call.pathname = strdup(pathname);
 657         va_start(ap, line);
 658         call.flags = va_arg(ap, int);
 659         if (call.flags & O_CREAT) {
 660                 call.mode = va_arg(ap, mode_t);
 661                 va_end(ap);
 662         }
 663         p = add_history(FAILTEST_OPEN, file, line, &call);
 664         /* Avoid memory leak! */
 665         if (p == &unrecorded_call)
 666                 free((char *)call.pathname);
 667         p->u.open.ret = open(pathname, call.flags, call.mode);
 668
 669         if (!failpath && p->u.open.ret == -1) {
 670                 p->fail = false;
 671                 p->error = errno;
 672         } else if (should_fail(p)) {
 673                 close(p->u.open.ret);
 674                 p->u.open.ret = -1;
 675                 /* FIXME: Play with error codes? */
 676                 p->error = EACCES;
 677         } else {
 678                 set_cleanup(p, cleanup_open, struct open_call);
 679         }
 680         errno = p->error;
 681         return p->u.open.ret;
 682 }
 683
 684 static void cleanup_pipe(struct pipe_call *call)
 685 {
 686         if (!call->closed[0])
 687                 close(call->fds[0]);
 688         if (!call->closed[1])
 689                 close(call->fds[1]);
 690 }
 691
 692 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
 693 {
 694         struct failtest_call *p;
 695         struct pipe_call call;
 696
 697         p = add_history(FAILTEST_PIPE, file, line, &call);
 698         if (should_fail(p)) {
 699                 p->u.open.ret = -1;
 700                 /* FIXME: Play with error codes? */
 701                 p->error = EMFILE;
 702         } else {
 703                 p->u.pipe.ret = pipe(p->u.pipe.fds);
 704                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
 705                 set_cleanup(p, cleanup_pipe, struct pipe_call);
 706         }
 707         /* This causes valgrind to notice if they use pipefd[] after failure */
 708         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
 709         errno = p->error;
 710         return p->u.pipe.ret;
 711 }
 712
 713 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
 714                        const char *file, unsigned line)
 715 {
 716         struct failtest_call *p;
 717         struct read_call call;
 718         call.fd = fd;
 719         call.buf = buf;
 720         call.count = count;
 721         call.off = off;
 722         p = add_history(FAILTEST_READ, file, line, &call);
 723
 724         /* FIXME: Try partial read returns. */
 725         if (should_fail(p)) {
 726                 p->u.read.ret = -1;
 727                 p->error = EIO;
 728         } else {
 729                 p->u.read.ret = pread(fd, buf, count, off);
 730         }
 731         errno = p->error;
 732         return p->u.read.ret;
 733 }
 734
 735 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
 736                         const char *file, unsigned line)
 737 {
 738         struct failtest_call *p;
 739         struct write_call call;
 740
 741         call.fd = fd;
 742         call.buf = buf;
 743         call.count = count;
 744         call.off = off;
 745         p = add_history(FAILTEST_WRITE, file, line, &call);
 746
 747         /* If we're a child, we need to make sure we write the same thing
 748          * to non-files as the parent does, so tell it. */
 749         if (control_fd != -1 && off == (off_t)-1) {
 750                 enum info_type type = WRITE;
 751
 752                 write_all(control_fd, &type, sizeof(type));
 753                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
 754                 write_all(control_fd, buf, count);
 755         }
 756
 757         /* FIXME: Try partial write returns. */
 758         if (should_fail(p)) {
 759                 p->u.write.ret = -1;
 760                 p->error = EIO;
 761         } else {
 762                 /* FIXME: We assume same write order in parent and child */
 763                 if (off == (off_t)-1 && child_writes_num != 0) {
 764                         if (child_writes[0].fd != fd)
 765                                 errx(1, "Child wrote to fd %u, not %u?",
 766                                      child_writes[0].fd, fd);
 767                         if (child_writes[0].off != p->u.write.off)
 768                                 errx(1, "Child wrote to offset %zu, not %zu?",
 769                                      (size_t)child_writes[0].off,
 770                                      (size_t)p->u.write.off);
 771                         if (child_writes[0].count != count)
 772                                 errx(1, "Child wrote length %zu, not %zu?",
 773                                      child_writes[0].count, count);
 774                         if (memcmp(child_writes[0].buf, buf, count)) {
 775                                 child_fail(NULL, 0,
 776                                            "Child wrote differently to"
 777                                            " fd %u than we did!\n", fd);
 778                         }
 779                         free((char *)child_writes[0].buf);
 780                         child_writes_num--;
 781                         memmove(&child_writes[0], &child_writes[1],
 782                                 sizeof(child_writes[0]) * child_writes_num);
 783
 784                         /* Is this is a socket or pipe, child wrote it
 785                            already. */
 786                         if (p->u.write.off == (off_t)-1) {
 787                                 p->u.write.ret = count;
 788                                 errno = p->error;
 789                                 return p->u.write.ret;
 790                         }
 791                 }
 792                 p->u.write.ret = pwrite(fd, buf, count, off);
 793         }
 794         errno = p->error;
 795         return p->u.write.ret;
 796 }
 797
 798 ssize_t failtest_read(int fd, void *buf, size_t count,
 799                       const char *file, unsigned line)
 800 {
 801         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 802                               file, line);
 803 }
 804
 805 ssize_t failtest_write(int fd, const void *buf, size_t count,
 806                        const char *file, unsigned line)
 807 {
 808         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 809                                file, line);
 810 }
 811
 812 static struct lock_info *WARN_UNUSED_RESULT
 813 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 814 {
 815         unsigned int i;
 816         struct lock_info *l;
 817
 818         for (i = 0; i < lock_num; i++) {
 819                 l = &locks[i];
 820
 821                 if (l->fd != fd)
 822                         continue;
 823                 /* Four cases we care about:
 824                  * Start overlap:
 825                  *      l =    |      |
 826                  *      new = |   |
 827                  * Mid overlap:
 828                  *      l =    |      |
 829                  *      new =    |  |
 830                  * End overlap:
 831                  *      l =    |      |
 832                  *      new =      |    |
 833                  * Total overlap:
 834                  *      l =    |      |
 835                  *      new = |         |
 836                  */
 837                 if (start > l->start && end < l->end) {
 838                         /* Mid overlap: trim entry, add new one. */
 839                         off_t new_start, new_end;
 840                         new_start = end + 1;
 841                         new_end = l->end;
 842                         l->end = start - 1;
 843                         locks = add_lock(locks,
 844                                          fd, new_start, new_end, l->type);
 845                         l = &locks[i];
 846                 } else if (start <= l->start && end >= l->end) {
 847                         /* Total overlap: eliminate entry. */
 848                         l->end = 0;
 849                         l->start = 1;
 850                 } else if (end >= l->start && end < l->end) {
 851                         /* Start overlap: trim entry. */
 852                         l->start = end + 1;
 853                 } else if (start > l->start && start <= l->end) {
 854                         /* End overlap: trim entry. */
 855                         l->end = start-1;
 856                 }
 857                 /* Nothing left?  Remove it. */
 858                 if (l->end < l->start) {
 859                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
 860                         i--;
 861                 }
 862         }
 863
 864         if (type != F_UNLCK) {
 865                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
 866                 l = &locks[lock_num++];
 867                 l->fd = fd;
 868                 l->start = start;
 869                 l->end = end;
 870                 l->type = type;
 871         }
 872         return locks;
 873 }
 874
 875 /* We trap this so we can record it: we don't fail it. */
 876 int failtest_close(int fd, const char *file, unsigned line)
 877 {
 878         int i;
 879         struct close_call call;
 880         struct failtest_call *p;
 881
 882         call.fd = fd;
 883         p = add_history(FAILTEST_CLOSE, file, line, &call);
 884         p->fail = false;
 885
 886         /* Consume close from failpath. */
 887         if (failpath)
 888                 if (should_fail(p))
 889                         abort();
 890
 891         if (fd < 0)
 892                 return close(fd);
 893
 894         /* Trace history to find source of fd. */
 895         for (i = history_num-1; i >= 0; i--) {
 896                 switch (history[i].type) {
 897                 case FAILTEST_PIPE:
 898                         /* From a pipe? */
 899                         if (history[i].u.pipe.fds[0] == fd) {
 900                                 assert(!history[i].u.pipe.closed[0]);
 901                                 history[i].u.pipe.closed[0] = true;
 902                                 if (history[i].u.pipe.closed[1])
 903                                         history[i].cleanup = NULL;
 904                                 goto out;
 905                         }
 906                         if (history[i].u.pipe.fds[1] == fd) {
 907                                 assert(!history[i].u.pipe.closed[1]);
 908                                 history[i].u.pipe.closed[1] = true;
 909                                 if (history[i].u.pipe.closed[0])
 910                                         history[i].cleanup = NULL;
 911                                 goto out;
 912                         }
 913                         break;
 914                 case FAILTEST_OPEN:
 915                         if (history[i].u.open.ret == fd) {
 916                                 assert((void *)history[i].cleanup
 917                                        == (void *)cleanup_open);
 918                                 history[i].cleanup = NULL;
 919                                 goto out;
 920                         }
 921                         break;
 922                 default:
 923                         break;
 924                 }
 925         }
 926
 927 out:
 928         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
 929         return close(fd);
 930 }
 931
 932 /* Zero length means "to end of file" */
 933 static off_t end_of(off_t start, off_t len)
 934 {
 935         if (len == 0)
 936                 return off_max();
 937         return start + len - 1;
 938 }
 939
 940 /* FIXME: This only handles locks, really. */
 941 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
 942 {
 943         struct failtest_call *p;
 944         struct fcntl_call call;
 945         va_list ap;
 946
 947         call.fd = fd;
 948         call.cmd = cmd;
 949
 950         /* Argument extraction. */
 951         switch (cmd) {
 952         case F_SETFL:
 953         case F_SETFD:
 954                 va_start(ap, cmd);
 955                 call.arg.l = va_arg(ap, long);
 956                 va_end(ap);
 957                 return fcntl(fd, cmd, call.arg.l);
 958         case F_GETFD:
 959         case F_GETFL:
 960                 return fcntl(fd, cmd);
 961         case F_GETLK:
 962                 get_locks();
 963                 va_start(ap, cmd);
 964                 call.arg.fl = *va_arg(ap, struct flock *);
 965                 va_end(ap);
 966                 return fcntl(fd, cmd, &call.arg.fl);
 967         case F_SETLK:
 968         case F_SETLKW:
 969                 va_start(ap, cmd);
 970                 call.arg.fl = *va_arg(ap, struct flock *);
 971                 va_end(ap);
 972                 break;
 973         default:
 974                 /* This means you need to implement it here. */
 975                 err(1, "failtest: unknown fcntl %u", cmd);
 976         }
 977
 978         p = add_history(FAILTEST_FCNTL, file, line, &call);
 979         get_locks();
 980
 981         if (should_fail(p)) {
 982                 p->u.fcntl.ret = -1;
 983                 if (p->u.fcntl.cmd == F_SETLK)
 984                         p->error = EAGAIN;
 985                 else
 986                         p->error = EDEADLK;
 987         } else {
 988                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
 989                                        &p->u.fcntl.arg.fl);
 990                 if (p->u.fcntl.ret == -1)
 991                         p->error = errno;
 992                 else {
 993                         /* We don't handle anything else yet. */
 994                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
 995                         locks = add_lock(locks,
 996                                          p->u.fcntl.fd,
 997                                          p->u.fcntl.arg.fl.l_start,
 998                                          end_of(p->u.fcntl.arg.fl.l_start,
 999                                                 p->u.fcntl.arg.fl.l_len),
1000                                          p->u.fcntl.arg.fl.l_type);
1001                 }
1002         }
1003         errno = p->error;
1004         return p->u.fcntl.ret;
1005 }
1006
1007 void failtest_init(int argc, char *argv[])
1008 {
1009         unsigned int i;
1010
1011         for (i = 1; i < argc; i++) {
1012                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1013                         failpath = argv[i] + strlen("--failpath=");
1014                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1015                         tracefd = dup(STDERR_FILENO);
1016                         failtest_timeout_ms = -1;
1017                 } else if (!strncmp(argv[i], "--debugpath=",
1018                                     strlen("--debugpath="))) {
1019                         debugpath = argv[i] + strlen("--debugpath=");
1020                 }
1021         }
1022         gettimeofday(&start, NULL);
1023 }
1024
1025 void failtest_exit(int status)
1026 {
1027         if (failtest_exit_check) {
1028                 if (!failtest_exit_check(history, history_num))
1029                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1030         }
1031
1032         failtest_cleanup(false, status);
1033 }