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