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
  29 enum info_type {
  30         WRITE,
  31         RELEASE_LOCKS,
  32         FAILURE,
  33         SUCCESS,
  34         UNEXPECTED
  35 };
  36
  37 struct write_info_hdr {
  38         size_t len;
  39         off_t offset;
  40         int fd;
  41 };
  42
  43 struct fd_orig {
  44         int fd;
  45         off_t offset;
  46         size_t size;
  47         bool dupped;
  48 };
  49
  50 struct write_info {
  51         struct write_info_hdr hdr;
  52         char *data;
  53         size_t oldlen;
  54         char *olddata;
  55 };
  56
  57 struct lock_info {
  58         int fd;
  59         /* end is inclusive: you can't have a 0-byte lock. */
  60         off_t start, end;
  61         int type;
  62 };
  63
  64 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
  65
  66 static struct failtest_call *history = NULL;
  67 static unsigned int history_num = 0;
  68 static int control_fd = -1;
  69 static struct timeval start;
  70
  71 static struct write_info *writes = NULL;
  72 static unsigned int writes_num = 0;
  73
  74 static struct write_info *child_writes = NULL;
  75 static unsigned int child_writes_num = 0;
  76
  77 static struct fd_orig *fd_orig = NULL;
  78 static unsigned int fd_orig_num = 0;
  79
  80 static pid_t lock_owner;
  81 static struct lock_info *locks = NULL;
  82 static unsigned int lock_num = 0;
  83
  84 static const char info_to_arg[] = "mceoprwf";
  85
  86 /* Dummy call used for failtest_undo wrappers. */
  87 static struct failtest_call unrecorded_call;
  88
  89 static struct failtest_call *add_history_(enum failtest_call_type type,
  90                                           const char *file,
  91                                           unsigned int line,
  92                                           const void *elem,
  93                                           size_t elem_size)
  94 {
  95         /* NULL file is how we suppress failure. */
  96         if (!file)
  97                 return &unrecorded_call;
  98
  99         history = realloc(history, (history_num + 1) * sizeof(*history));
 100         history[history_num].type = type;
 101         history[history_num].file = file;
 102         history[history_num].line = line;
 103         memcpy(&history[history_num].u, elem, elem_size);
 104         return &history[history_num++];
 105 }
 106
 107 #define add_history(type, file, line, elem) \
 108         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
 109
 110 static void save_fd_orig(int fd)
 111 {
 112         unsigned int i;
 113
 114         for (i = 0; i < fd_orig_num; i++)
 115                 if (fd_orig[i].fd == fd)
 116                         return;
 117
 118         fd_orig = realloc(fd_orig, (fd_orig_num + 1) * sizeof(*fd_orig));
 119         fd_orig[fd_orig_num].fd = fd;
 120         fd_orig[fd_orig_num].dupped = false;
 121         fd_orig[fd_orig_num].offset = lseek(fd, 0, SEEK_CUR);
 122         fd_orig[fd_orig_num].size = lseek(fd, 0, SEEK_END);
 123         lseek(fd, fd_orig[fd_orig_num].offset, SEEK_SET);
 124         fd_orig_num++;
 125 }
 126
 127 bool failtest_default_hook(struct failtest_call *history, unsigned num)
 128 {
 129         return true;
 130 }
 131
 132 static bool read_write_info(int fd)
 133 {
 134         struct write_info_hdr hdr;
 135
 136         if (!read_all(fd, &hdr, sizeof(hdr)))
 137                 return false;
 138
 139         child_writes = realloc(child_writes,
 140                                (child_writes_num+1) * sizeof(child_writes[0]));
 141         child_writes[child_writes_num].hdr = hdr;
 142         child_writes[child_writes_num].data = malloc(hdr.len);
 143         if (!read_all(fd, child_writes[child_writes_num].data, hdr.len))
 144                 return false;
 145
 146         child_writes_num++;
 147         return true;
 148 }
 149
 150 static char *failpath_string(void)
 151 {
 152         unsigned int i;
 153         char *ret = malloc(history_num + 1);
 154
 155         for (i = 0; i < history_num; i++) {
 156                 ret[i] = info_to_arg[history[i].type];
 157                 if (history[i].fail)
 158                         ret[i] = toupper(ret[i]);
 159         }
 160         ret[i] = '\0';
 161         return ret;
 162 }
 163
 164 static void tell_parent(enum info_type type)
 165 {
 166         if (control_fd != -1)
 167                 write_all(control_fd, &type, sizeof(type));
 168 }
 169
 170 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 171 {
 172         va_list ap;
 173         char *path = failpath_string();
 174
 175         va_start(ap, fmt);
 176         vfprintf(stderr, fmt, ap);
 177         va_end(ap);
 178
 179         fprintf(stderr, "%.*s", (int)outlen, out);
 180         printf("To reproduce: --failpath=%s\n", path);
 181         free(path);
 182         tell_parent(FAILURE);
 183         exit(1);
 184 }
 185
 186 static pid_t child;
 187
 188 static void hand_down(int signal)
 189 {
 190         kill(child, signal);
 191 }
 192
 193 static void release_locks(void)
 194 {
 195         /* Locks were never acquired/reacquired? */
 196         if (lock_owner == 0)
 197                 return;
 198
 199         /* We own them?  Release them all. */
 200         if (lock_owner == getpid()) {
 201                 unsigned int i;
 202                 struct flock fl;
 203                 fl.l_type = F_UNLCK;
 204                 fl.l_whence = SEEK_SET;
 205                 fl.l_start = 0;
 206                 fl.l_len = 0;
 207
 208                 for (i = 0; i < lock_num; i++)
 209                         fcntl(locks[i].fd, F_SETLK, &fl);
 210         } else {
 211                 /* Our parent must have them; pass request up. */
 212                 enum info_type type = RELEASE_LOCKS;
 213                 assert(control_fd != -1);
 214                 write_all(control_fd, &type, sizeof(type));
 215         }
 216         lock_owner = 0;
 217 }
 218
 219 /* off_t is a signed type.  Getting its max is non-trivial. */
 220 static off_t off_max(void)
 221 {
 222         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
 223         if (sizeof(off_t) == 4)
 224                 return (off_t)0x7FFFFFF;
 225         else
 226                 return (off_t)0x7FFFFFFFFFFFFFFULL;
 227 }
 228
 229 static void get_locks(void)
 230 {
 231         unsigned int i;
 232         struct flock fl;
 233
 234         if (lock_owner == getpid())
 235                 return;
 236
 237         if (lock_owner != 0) {
 238                 enum info_type type = RELEASE_LOCKS;
 239                 assert(control_fd != -1);
 240                 write_all(control_fd, &type, sizeof(type));
 241         }
 242
 243         fl.l_whence = SEEK_SET;
 244
 245         for (i = 0; i < lock_num; i++) {
 246                 fl.l_type = locks[i].type;
 247                 fl.l_start = locks[i].start;
 248                 if (locks[i].end == off_max())
 249                         fl.l_len = 0;
 250                 else
 251                         fl.l_len = locks[i].end - locks[i].start + 1;
 252
 253                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
 254                         abort();
 255         }
 256         lock_owner = getpid();
 257 }
 258
 259 static void trace_str(const char *str)
 260 {
 261         ssize_t ret;
 262
 263         while ((ret = write(tracefd, str, strlen(str))) <= 0) {
 264                 str += ret;
 265                 if (!*str)
 266                         return;
 267         }
 268         err(1, "Writing trace.");
 269 }
 270
 271 static bool should_fail(struct failtest_call *call)
 272 {
 273         int status;
 274         int control[2], output[2];
 275         enum info_type type = UNEXPECTED;
 276         char *out = NULL;
 277         size_t outlen = 0;
 278
 279         if (call == &unrecorded_call)
 280                 return false;
 281
 282         if (failpath) {
 283                 /* + means continue after end, like normal. */
 284                 if (*failpath == '+')
 285                         failpath = NULL;
 286                 else {
 287                         if (tolower(*failpath) != info_to_arg[call->type])
 288                                 errx(1, "Failpath expected '%c' got '%c'\n",
 289                                      info_to_arg[call->type], *failpath);
 290                         call->fail = isupper(*(failpath++));
 291                         return call->fail;
 292                 }
 293         }
 294
 295         if (!failtest_hook(history, history_num)) {
 296                 call->fail = false;
 297                 return false;
 298         }
 299
 300         /* We're going to fail in the child. */
 301         call->fail = true;
 302         if (pipe(control) != 0 || pipe(output) != 0)
 303                 err(1, "opening pipe");
 304
 305         /* Prevent double-printing (in child and parent) */
 306         fflush(stdout);
 307         child = fork();
 308         if (child == -1)
 309                 err(1, "forking failed");
 310
 311         if (child == 0) {
 312                 if (tracefd != -1) {
 313                         struct timeval now;
 314                         char str[50], *p;
 315                         gettimeofday(&now, NULL);
 316                         if (now.tv_usec < start.tv_usec) {
 317                                 now.tv_sec--;
 318                                 now.tv_usec += 1000000;
 319                         }
 320                         now.tv_usec -= start.tv_usec;
 321                         now.tv_sec -= start.tv_sec;
 322                         sprintf(str, "%u (%u.%02u): ", getpid(),
 323                                 (int)now.tv_sec, (int)now.tv_usec / 10000);
 324                         trace_str(str);
 325                         p = failpath_string();
 326                         trace_str(p);
 327                         free(p);
 328                         trace_str("(");
 329                         p = strchr(history[history_num-1].file, '/');
 330                         if (p)
 331                                 trace_str(p+1);
 332                         else
 333                                 trace_str(history[history_num-1].file);
 334                         sprintf(str, ":%u)\n", history[history_num-1].line);
 335                         trace_str(str);
 336                 }
 337                 close(control[0]);
 338                 close(output[0]);
 339                 dup2(output[1], STDOUT_FILENO);
 340                 dup2(output[1], STDERR_FILENO);
 341                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
 342                         close(output[1]);
 343                 control_fd = control[1];
 344                 return true;
 345         }
 346
 347         signal(SIGUSR1, hand_down);
 348
 349         close(control[1]);
 350         close(output[1]);
 351
 352         /* We grab output so we can display it; we grab writes so we
 353          * can compare. */
 354         do {
 355                 struct pollfd pfd[2];
 356                 int ret;
 357
 358                 pfd[0].fd = output[0];
 359                 pfd[0].events = POLLIN|POLLHUP;
 360                 pfd[1].fd = control[0];
 361                 pfd[1].events = POLLIN|POLLHUP;
 362
 363                 if (type == SUCCESS)
 364                         ret = poll(pfd, 1, failtest_timeout_ms);
 365                 else
 366                         ret = poll(pfd, 2, failtest_timeout_ms);
 367
 368                 if (ret <= 0)
 369                         hand_down(SIGUSR1);
 370
 371                 if (pfd[0].revents & POLLIN) {
 372                         ssize_t len;
 373
 374                         out = realloc(out, outlen + 8192);
 375                         len = read(output[0], out + outlen, 8192);
 376                         outlen += len;
 377                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 378                         if (read_all(control[0], &type, sizeof(type))) {
 379                                 if (type == WRITE) {
 380                                         if (!read_write_info(control[0]))
 381                                                 break;
 382                                 } else if (type == RELEASE_LOCKS) {
 383                                         release_locks();
 384                                         /* FIXME: Tell them we're done... */
 385                                 }
 386                         }
 387                 } else if (pfd[0].revents & POLLHUP) {
 388                         break;
 389                 }
 390         } while (type != FAILURE);
 391
 392         close(output[0]);
 393         close(control[0]);
 394         waitpid(child, &status, 0);
 395         if (!WIFEXITED(status))
 396                 child_fail(out, outlen, "Killed by signal %u: ",
 397                            WTERMSIG(status));
 398         /* Child printed failure already, just pass up exit code. */
 399         if (type == FAILURE) {
 400                 fprintf(stderr, "%.*s", (int)outlen, out);
 401                 tell_parent(type);
 402                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 403         }
 404         if (WEXITSTATUS(status) != 0)
 405                 child_fail(out, outlen, "Exited with status %i: ",
 406                            WEXITSTATUS(status));
 407
 408         free(out);
 409         signal(SIGUSR1, SIG_DFL);
 410
 411         /* We continue onwards without failing. */
 412         call->fail = false;
 413         return false;
 414 }
 415
 416 void *failtest_calloc(size_t nmemb, size_t size,
 417                       const char *file, unsigned line)
 418 {
 419         struct failtest_call *p;
 420         struct calloc_call call;
 421         call.nmemb = nmemb;
 422         call.size = size;
 423         p = add_history(FAILTEST_CALLOC, file, line, &call);
 424
 425         if (should_fail(p)) {
 426                 p->u.calloc.ret = NULL;
 427                 p->error = ENOMEM;
 428         } else {
 429                 p->u.calloc.ret = calloc(nmemb, size);
 430         }
 431         errno = p->error;
 432         return p->u.calloc.ret;
 433 }
 434
 435 void *failtest_malloc(size_t size, const char *file, unsigned line)
 436 {
 437         struct failtest_call *p;
 438         struct malloc_call call;
 439         call.size = size;
 440
 441         p = add_history(FAILTEST_MALLOC, file, line, &call);
 442         if (should_fail(p)) {
 443                 p->u.calloc.ret = NULL;
 444                 p->error = ENOMEM;
 445         } else {
 446                 p->u.calloc.ret = malloc(size);
 447         }
 448         errno = p->error;
 449         return p->u.calloc.ret;
 450 }
 451
 452 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 453 {
 454         struct failtest_call *p;
 455         struct realloc_call call;
 456         call.size = size;
 457         p = add_history(FAILTEST_REALLOC, file, line, &call);
 458
 459         /* FIXME: Try one child moving allocation, one not. */
 460         if (should_fail(p)) {
 461                 p->u.realloc.ret = NULL;
 462                 p->error = ENOMEM;
 463         } else {
 464                 p->u.realloc.ret = realloc(ptr, size);
 465         }
 466         errno = p->error;
 467         return p->u.realloc.ret;
 468 }
 469
 470 int failtest_open(const char *pathname,
 471                   const char *file, unsigned line, ...)
 472 {
 473         struct failtest_call *p;
 474         struct open_call call;
 475         va_list ap;
 476
 477         call.pathname = strdup(pathname);
 478         va_start(ap, line);
 479         call.flags = va_arg(ap, int);
 480         if (call.flags & O_CREAT) {
 481                 call.mode = va_arg(ap, mode_t);
 482                 va_end(ap);
 483         }
 484         p = add_history(FAILTEST_OPEN, file, line, &call);
 485         /* Avoid memory leak! */
 486         if (p == &unrecorded_call)
 487                 free((char *)call.pathname);
 488         if (should_fail(p)) {
 489                 p->u.open.ret = -1;
 490                 /* FIXME: Play with error codes? */
 491                 p->error = EACCES;
 492         } else {
 493                 p->u.open.ret = open(pathname, call.flags, call.mode);
 494         }
 495         errno = p->error;
 496         return p->u.open.ret;
 497 }
 498
 499 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
 500 {
 501         struct failtest_call *p;
 502         struct pipe_call call;
 503
 504         p = add_history(FAILTEST_PIPE, file, line, &call);
 505         if (should_fail(p)) {
 506                 p->u.open.ret = -1;
 507                 /* FIXME: Play with error codes? */
 508                 p->error = EMFILE;
 509         } else {
 510                 p->u.pipe.ret = pipe(p->u.pipe.fds);
 511         }
 512         /* This causes valgrind to notice if they use pipefd[] after failure */
 513         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
 514         errno = p->error;
 515         return p->u.pipe.ret;
 516 }
 517
 518 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
 519                        const char *file, unsigned line)
 520 {
 521         struct failtest_call *p;
 522         struct read_call call;
 523         call.fd = fd;
 524         call.buf = buf;
 525         call.count = count;
 526         call.off = off;
 527         p = add_history(FAILTEST_READ, file, line, &call);
 528
 529         /* This is going to change seek offset, so save it. */
 530         if (control_fd != -1)
 531                 save_fd_orig(fd);
 532
 533         /* FIXME: Try partial read returns. */
 534         if (should_fail(p)) {
 535                 p->u.read.ret = -1;
 536                 p->error = EIO;
 537         } else {
 538                 p->u.read.ret = pread(fd, buf, count, off);
 539         }
 540         errno = p->error;
 541         return p->u.read.ret;
 542 }
 543
 544 static struct write_info *new_write(void)
 545 {
 546         writes = realloc(writes, (writes_num + 1) * sizeof(*writes));
 547         return &writes[writes_num++];
 548 }
 549
 550 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
 551                         const char *file, unsigned line)
 552 {
 553         struct failtest_call *p;
 554         struct write_call call;
 555         off_t offset;
 556
 557         call.fd = fd;
 558         call.buf = buf;
 559         call.count = count;
 560         call.off = off;
 561         p = add_history(FAILTEST_WRITE, file, line, &call);
 562
 563         offset = lseek(fd, 0, SEEK_CUR);
 564
 565         /* If we're a child, save contents and tell parent about write. */
 566         if (control_fd != -1) {
 567                 struct write_info *winfo = new_write();
 568                 enum info_type type = WRITE;
 569
 570                 save_fd_orig(fd);
 571
 572                 winfo->hdr.len = count;
 573                 winfo->hdr.fd = fd;
 574                 winfo->data = malloc(count);
 575                 memcpy(winfo->data, buf, count);
 576                 winfo->hdr.offset = offset;
 577                 if (winfo->hdr.offset != (off_t)-1) {
 578                         lseek(fd, offset, SEEK_SET);
 579                         winfo->olddata = malloc(count);
 580                         winfo->oldlen = read(fd, winfo->olddata, count);
 581                         if (winfo->oldlen == -1)
 582                                 winfo->oldlen = 0;
 583                 }
 584                 write_all(control_fd, &type, sizeof(type));
 585                 write_all(control_fd, &winfo->hdr, sizeof(winfo->hdr));
 586                 write_all(control_fd, winfo->data, count);
 587         }
 588
 589         /* FIXME: Try partial write returns. */
 590         if (should_fail(p)) {
 591                 p->u.write.ret = -1;
 592                 p->error = EIO;
 593         } else {
 594                 /* FIXME: We assume same write order in parent and child */
 595                 if (child_writes_num != 0) {
 596                         if (child_writes[0].hdr.fd != fd)
 597                                 errx(1, "Child wrote to fd %u, not %u?",
 598                                      child_writes[0].hdr.fd, fd);
 599                         if (child_writes[0].hdr.offset != offset)
 600                                 errx(1, "Child wrote to offset %zu, not %zu?",
 601                                      (size_t)child_writes[0].hdr.offset,
 602                                      (size_t)offset);
 603                         if (child_writes[0].hdr.len != count)
 604                                 errx(1, "Child wrote length %zu, not %zu?",
 605                                      child_writes[0].hdr.len, count);
 606                         if (memcmp(child_writes[0].data, buf, count)) {
 607                                 child_fail(NULL, 0,
 608                                            "Child wrote differently to"
 609                                            " fd %u than we did!\n", fd);
 610                         }
 611                         free(child_writes[0].data);
 612                         child_writes_num--;
 613                         memmove(&child_writes[0], &child_writes[1],
 614                                 sizeof(child_writes[0]) * child_writes_num);
 615
 616                         /* Is this is a socket or pipe, child wrote it
 617                            already. */
 618                         if (offset == (off_t)-1) {
 619                                 p->u.write.ret = count;
 620                                 errno = p->error;
 621                                 return p->u.write.ret;
 622                         }
 623                 }
 624                 p->u.write.ret = pwrite(fd, buf, count, off);
 625         }
 626         errno = p->error;
 627         return p->u.write.ret;
 628 }
 629
 630 ssize_t failtest_read(int fd, void *buf, size_t count,
 631                       const char *file, unsigned line)
 632 {
 633         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 634                               file, line);
 635 }
 636
 637 ssize_t failtest_write(int fd, const void *buf, size_t count,
 638                        const char *file, unsigned line)
 639 {
 640         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 641                                file, line);
 642 }
 643
 644 static struct lock_info *WARN_UNUSED_RESULT
 645 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 646 {
 647         unsigned int i;
 648         struct lock_info *l;
 649
 650         for (i = 0; i < lock_num; i++) {
 651                 l = &locks[i];
 652
 653                 if (l->fd != fd)
 654                         continue;
 655                 /* Four cases we care about:
 656                  * Start overlap:
 657                  *      l =    |      |
 658                  *      new = |   |
 659                  * Mid overlap:
 660                  *      l =    |      |
 661                  *      new =    |  |
 662                  * End overlap:
 663                  *      l =    |      |
 664                  *      new =      |    |
 665                  * Total overlap:
 666                  *      l =    |      |
 667                  *      new = |         |
 668                  */
 669                 if (start > l->start && end < l->end) {
 670                         /* Mid overlap: trim entry, add new one. */
 671                         off_t new_start, new_end;
 672                         new_start = end + 1;
 673                         new_end = l->end;
 674                         l->end = start - 1;
 675                         locks = add_lock(locks,
 676                                          fd, new_start, new_end, l->type);
 677                         l = &locks[i];
 678                 } else if (start <= l->start && end >= l->end) {
 679                         /* Total overlap: eliminate entry. */
 680                         l->end = 0;
 681                         l->start = 1;
 682                 } else if (end >= l->start && end < l->end) {
 683                         /* Start overlap: trim entry. */
 684                         l->start = end + 1;
 685                 } else if (start > l->start && start <= l->end) {
 686                         /* End overlap: trim entry. */
 687                         l->end = start-1;
 688                 }
 689                 /* Nothing left?  Remove it. */
 690                 if (l->end < l->start) {
 691                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
 692                         i--;
 693                 }
 694         }
 695
 696         if (type != F_UNLCK) {
 697                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
 698                 l = &locks[lock_num++];
 699                 l->fd = fd;
 700                 l->start = start;
 701                 l->end = end;
 702                 l->type = type;
 703         }
 704         return locks;
 705 }
 706
 707 /* We only trap this so we can dup fds in case we need to restore. */
 708 int failtest_close(int fd)
 709 {
 710         unsigned int i;
 711         int newfd = -1;
 712
 713         for (i = 0; i < fd_orig_num; i++) {
 714                 if (fd_orig[i].fd == fd) {
 715                         fd_orig[i].fd = newfd = dup(fd);
 716                         fd_orig[i].dupped = true;
 717                 }
 718         }
 719
 720         for (i = 0; i < writes_num; i++) {
 721                 if (writes[i].hdr.fd == fd)
 722                         writes[i].hdr.fd = newfd;
 723         }
 724
 725         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
 726         return close(fd);
 727 }
 728
 729 /* Zero length means "to end of file" */
 730 static off_t end_of(off_t start, off_t len)
 731 {
 732         if (len == 0)
 733                 return off_max();
 734         return start + len - 1;
 735 }
 736
 737 /* FIXME: This only handles locks, really. */
 738 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
 739 {
 740         struct failtest_call *p;
 741         struct fcntl_call call;
 742         va_list ap;
 743
 744         call.fd = fd;
 745         call.cmd = cmd;
 746
 747         /* Argument extraction. */
 748         switch (cmd) {
 749         case F_SETFL:
 750         case F_SETFD:
 751                 va_start(ap, cmd);
 752                 call.arg.l = va_arg(ap, long);
 753                 va_end(ap);
 754                 return fcntl(fd, cmd, call.arg.l);
 755         case F_GETFD:
 756         case F_GETFL:
 757                 return fcntl(fd, cmd);
 758         case F_GETLK:
 759                 get_locks();
 760                 va_start(ap, cmd);
 761                 call.arg.fl = *va_arg(ap, struct flock *);
 762                 va_end(ap);
 763                 return fcntl(fd, cmd, &call.arg.fl);
 764         case F_SETLK:
 765         case F_SETLKW:
 766                 va_start(ap, cmd);
 767                 call.arg.fl = *va_arg(ap, struct flock *);
 768                 va_end(ap);
 769                 break;
 770         default:
 771                 /* This means you need to implement it here. */
 772                 err(1, "failtest: unknown fcntl %u", cmd);
 773         }
 774
 775         p = add_history(FAILTEST_FCNTL, file, line, &call);
 776         get_locks();
 777
 778         if (should_fail(p)) {
 779                 p->u.fcntl.ret = -1;
 780                 if (p->u.fcntl.cmd == F_SETLK)
 781                         p->error = EAGAIN;
 782                 else
 783                         p->error = EDEADLK;
 784         } else {
 785                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
 786                                        &p->u.fcntl.arg.fl);
 787                 if (p->u.fcntl.ret == -1)
 788                         p->error = errno;
 789                 else {
 790                         /* We don't handle anything else yet. */
 791                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
 792                         locks = add_lock(locks,
 793                                          p->u.fcntl.fd,
 794                                          p->u.fcntl.arg.fl.l_start,
 795                                          end_of(p->u.fcntl.arg.fl.l_start,
 796                                                 p->u.fcntl.arg.fl.l_len),
 797                                          p->u.fcntl.arg.fl.l_type);
 798                 }
 799         }
 800         errno = p->error;
 801         return p->u.fcntl.ret;
 802 }
 803
 804 void failtest_init(int argc, char *argv[])
 805 {
 806         unsigned int i;
 807
 808         for (i = 1; i < argc; i++) {
 809                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
 810                         failpath = argv[i] + strlen("--failpath=");
 811                 } else if (strcmp(argv[i], "--tracepath") == 0) {
 812                         tracefd = dup(STDERR_FILENO);
 813                         failtest_timeout_ms = -1;
 814                 }
 815         }
 816         gettimeofday(&start, NULL);
 817 }
 818
 819 /* Free up memory, so valgrind doesn't report leaks. */
 820 static void free_everything(void)
 821 {
 822         unsigned int i;
 823
 824         for (i = 0; i < writes_num; i++) {
 825                 free(writes[i].data);
 826                 if (writes[i].hdr.offset != (off_t)-1)
 827                         free(writes[i].olddata);
 828         }
 829         free(writes);
 830         free(fd_orig);
 831         for (i = 0; i < history_num; i++) {
 832                 if (history[i].type == FAILTEST_OPEN)
 833                         free((char *)history[i].u.open.pathname);
 834         }
 835         free(history);
 836 }
 837
 838 void failtest_exit(int status)
 839 {
 840         unsigned int i;
 841
 842         if (control_fd == -1) {
 843                 free_everything();
 844                 exit(status);
 845         }
 846
 847         if (failtest_exit_check) {
 848                 if (!failtest_exit_check(history, history_num))
 849                         child_fail(NULL, 0, "failtest_exit_check failed\n");
 850         }
 851
 852         /* Restore any stuff we overwrote. */
 853         for (i = 0; i < writes_num; i++) {
 854                 if (writes[i].hdr.offset == (off_t)-1)
 855                         continue;
 856                 if (writes[i].oldlen != 0) {
 857                         lseek(writes[i].hdr.fd, writes[i].hdr.offset,
 858                               SEEK_SET);
 859                         write(writes[i].hdr.fd, writes[i].olddata,
 860                               writes[i].oldlen);
 861                 }
 862         }
 863
 864         /* Fix up fd offsets, restore sizes. */
 865         for (i = 0; i < fd_orig_num; i++) {
 866                 lseek(fd_orig[i].fd, fd_orig[i].offset, SEEK_SET);
 867                 ftruncate(fd_orig[i].fd, fd_orig[i].size);
 868                 /* Free up any file descriptors we dup'ed. */
 869                 if (fd_orig[i].dupped)
 870                         close(fd_orig[i].fd);
 871         }
 872
 873         free_everything();
 874         tell_parent(SUCCESS);
 875         exit(0);
 876 }