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                 if (tolower(*failpath) != info_to_arg[call->type])
 284                         errx(1, "Failpath expected '%c' got '%c'\n",
 285                              info_to_arg[call->type], *failpath);
 286                 call->fail = isupper(*(failpath++));
 287                 return call->fail;
 288         }
 289
 290         if (!failtest_hook(history, history_num)) {
 291                 call->fail = false;
 292                 return false;
 293         }
 294
 295         /* We're going to fail in the child. */
 296         call->fail = true;
 297         if (pipe(control) != 0 || pipe(output) != 0)
 298                 err(1, "opening pipe");
 299
 300         /* Prevent double-printing (in child and parent) */
 301         fflush(stdout);
 302         child = fork();
 303         if (child == -1)
 304                 err(1, "forking failed");
 305
 306         if (child == 0) {
 307                 if (tracefd != -1) {
 308                         struct timeval now;
 309                         char str[50], *p;
 310                         gettimeofday(&now, NULL);
 311                         if (now.tv_usec < start.tv_usec) {
 312                                 now.tv_sec--;
 313                                 now.tv_usec += 1000000;
 314                         }
 315                         now.tv_usec -= start.tv_usec;
 316                         now.tv_sec -= start.tv_sec;
 317                         sprintf(str, "%u (%u.%02u): ", getpid(),
 318                                 (int)now.tv_sec, (int)now.tv_usec / 10000);
 319                         trace_str(str);
 320                         p = failpath_string();
 321                         trace_str(p);
 322                         free(p);
 323                         trace_str("(");
 324                         p = strchr(history[history_num-1].file, '/');
 325                         if (p)
 326                                 trace_str(p+1);
 327                         else
 328                                 trace_str(history[history_num-1].file);
 329                         sprintf(str, ":%u)\n", history[history_num-1].line);
 330                         trace_str(str);
 331                 }
 332                 close(control[0]);
 333                 close(output[0]);
 334                 dup2(output[1], STDOUT_FILENO);
 335                 dup2(output[1], STDERR_FILENO);
 336                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
 337                         close(output[1]);
 338                 control_fd = control[1];
 339                 return true;
 340         }
 341
 342         signal(SIGUSR1, hand_down);
 343
 344         close(control[1]);
 345         close(output[1]);
 346
 347         /* We grab output so we can display it; we grab writes so we
 348          * can compare. */
 349         do {
 350                 struct pollfd pfd[2];
 351                 int ret;
 352
 353                 pfd[0].fd = output[0];
 354                 pfd[0].events = POLLIN|POLLHUP;
 355                 pfd[1].fd = control[0];
 356                 pfd[1].events = POLLIN|POLLHUP;
 357
 358                 if (type == SUCCESS)
 359                         ret = poll(pfd, 1, failtest_timeout_ms);
 360                 else
 361                         ret = poll(pfd, 2, failtest_timeout_ms);
 362
 363                 if (ret <= 0)
 364                         hand_down(SIGUSR1);
 365
 366                 if (pfd[0].revents & POLLIN) {
 367                         ssize_t len;
 368
 369                         out = realloc(out, outlen + 8192);
 370                         len = read(output[0], out + outlen, 8192);
 371                         outlen += len;
 372                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 373                         if (read_all(control[0], &type, sizeof(type))) {
 374                                 if (type == WRITE) {
 375                                         if (!read_write_info(control[0]))
 376                                                 break;
 377                                 } else if (type == RELEASE_LOCKS) {
 378                                         release_locks();
 379                                         /* FIXME: Tell them we're done... */
 380                                 }
 381                         }
 382                 } else if (pfd[0].revents & POLLHUP) {
 383                         break;
 384                 }
 385         } while (type != FAILURE);
 386
 387         close(output[0]);
 388         close(control[0]);
 389         waitpid(child, &status, 0);
 390         if (!WIFEXITED(status))
 391                 child_fail(out, outlen, "Killed by signal %u: ",
 392                            WTERMSIG(status));
 393         /* Child printed failure already, just pass up exit code. */
 394         if (type == FAILURE) {
 395                 fprintf(stderr, "%.*s", (int)outlen, out);
 396                 tell_parent(type);
 397                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 398         }
 399         if (WEXITSTATUS(status) != 0)
 400                 child_fail(out, outlen, "Exited with status %i: ",
 401                            WEXITSTATUS(status));
 402
 403         free(out);
 404         signal(SIGUSR1, SIG_DFL);
 405
 406         /* We continue onwards without failing. */
 407         call->fail = false;
 408         return false;
 409 }
 410
 411 void *failtest_calloc(size_t nmemb, size_t size,
 412                       const char *file, unsigned line)
 413 {
 414         struct failtest_call *p;
 415         struct calloc_call call;
 416         call.nmemb = nmemb;
 417         call.size = size;
 418         p = add_history(FAILTEST_CALLOC, file, line, &call);
 419
 420         if (should_fail(p)) {
 421                 p->u.calloc.ret = NULL;
 422                 p->error = ENOMEM;
 423         } else {
 424                 p->u.calloc.ret = calloc(nmemb, size);
 425         }
 426         errno = p->error;
 427         return p->u.calloc.ret;
 428 }
 429
 430 void *failtest_malloc(size_t size, const char *file, unsigned line)
 431 {
 432         struct failtest_call *p;
 433         struct malloc_call call;
 434         call.size = size;
 435
 436         p = add_history(FAILTEST_MALLOC, file, line, &call);
 437         if (should_fail(p)) {
 438                 p->u.calloc.ret = NULL;
 439                 p->error = ENOMEM;
 440         } else {
 441                 p->u.calloc.ret = malloc(size);
 442         }
 443         errno = p->error;
 444         return p->u.calloc.ret;
 445 }
 446
 447 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 448 {
 449         struct failtest_call *p;
 450         struct realloc_call call;
 451         call.size = size;
 452         p = add_history(FAILTEST_REALLOC, file, line, &call);
 453
 454         /* FIXME: Try one child moving allocation, one not. */
 455         if (should_fail(p)) {
 456                 p->u.realloc.ret = NULL;
 457                 p->error = ENOMEM;
 458         } else {
 459                 p->u.realloc.ret = realloc(ptr, size);
 460         }
 461         errno = p->error;
 462         return p->u.realloc.ret;
 463 }
 464
 465 int failtest_open(const char *pathname,
 466                   const char *file, unsigned line, ...)
 467 {
 468         struct failtest_call *p;
 469         struct open_call call;
 470         va_list ap;
 471
 472         call.pathname = strdup(pathname);
 473         va_start(ap, line);
 474         call.flags = va_arg(ap, int);
 475         if (call.flags & O_CREAT) {
 476                 call.mode = va_arg(ap, mode_t);
 477                 va_end(ap);
 478         }
 479         p = add_history(FAILTEST_OPEN, file, line, &call);
 480         /* Avoid memory leak! */
 481         if (p == &unrecorded_call)
 482                 free((char *)call.pathname);
 483         if (should_fail(p)) {
 484                 p->u.open.ret = -1;
 485                 /* FIXME: Play with error codes? */
 486                 p->error = EACCES;
 487         } else {
 488                 p->u.open.ret = open(pathname, call.flags, call.mode);
 489         }
 490         errno = p->error;
 491         return p->u.open.ret;
 492 }
 493
 494 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
 495 {
 496         struct failtest_call *p;
 497         struct pipe_call call;
 498
 499         p = add_history(FAILTEST_PIPE, file, line, &call);
 500         if (should_fail(p)) {
 501                 p->u.open.ret = -1;
 502                 /* FIXME: Play with error codes? */
 503                 p->error = EMFILE;
 504         } else {
 505                 p->u.pipe.ret = pipe(p->u.pipe.fds);
 506         }
 507         /* This causes valgrind to notice if they use pipefd[] after failure */
 508         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
 509         errno = p->error;
 510         return p->u.pipe.ret;
 511 }
 512
 513 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
 514                        const char *file, unsigned line)
 515 {
 516         struct failtest_call *p;
 517         struct read_call call;
 518         call.fd = fd;
 519         call.buf = buf;
 520         call.count = count;
 521         call.off = off;
 522         p = add_history(FAILTEST_READ, file, line, &call);
 523
 524         /* This is going to change seek offset, so save it. */
 525         if (control_fd != -1)
 526                 save_fd_orig(fd);
 527
 528         /* FIXME: Try partial read returns. */
 529         if (should_fail(p)) {
 530                 p->u.read.ret = -1;
 531                 p->error = EIO;
 532         } else {
 533                 p->u.read.ret = pread(fd, buf, count, off);
 534         }
 535         errno = p->error;
 536         return p->u.read.ret;
 537 }
 538
 539 static struct write_info *new_write(void)
 540 {
 541         writes = realloc(writes, (writes_num + 1) * sizeof(*writes));
 542         return &writes[writes_num++];
 543 }
 544
 545 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
 546                         const char *file, unsigned line)
 547 {
 548         struct failtest_call *p;
 549         struct write_call call;
 550         off_t offset;
 551
 552         call.fd = fd;
 553         call.buf = buf;
 554         call.count = count;
 555         call.off = off;
 556         p = add_history(FAILTEST_WRITE, file, line, &call);
 557
 558         offset = lseek(fd, 0, SEEK_CUR);
 559
 560         /* If we're a child, save contents and tell parent about write. */
 561         if (control_fd != -1) {
 562                 struct write_info *winfo = new_write();
 563                 enum info_type type = WRITE;
 564
 565                 save_fd_orig(fd);
 566
 567                 winfo->hdr.len = count;
 568                 winfo->hdr.fd = fd;
 569                 winfo->data = malloc(count);
 570                 memcpy(winfo->data, buf, count);
 571                 winfo->hdr.offset = offset;
 572                 if (winfo->hdr.offset != (off_t)-1) {
 573                         lseek(fd, offset, SEEK_SET);
 574                         winfo->olddata = malloc(count);
 575                         winfo->oldlen = read(fd, winfo->olddata, count);
 576                         if (winfo->oldlen == -1)
 577                                 winfo->oldlen = 0;
 578                 }
 579                 write_all(control_fd, &type, sizeof(type));
 580                 write_all(control_fd, &winfo->hdr, sizeof(winfo->hdr));
 581                 write_all(control_fd, winfo->data, count);
 582         }
 583
 584         /* FIXME: Try partial write returns. */
 585         if (should_fail(p)) {
 586                 p->u.write.ret = -1;
 587                 p->error = EIO;
 588         } else {
 589                 /* FIXME: We assume same write order in parent and child */
 590                 if (child_writes_num != 0) {
 591                         if (child_writes[0].hdr.fd != fd)
 592                                 errx(1, "Child wrote to fd %u, not %u?",
 593                                      child_writes[0].hdr.fd, fd);
 594                         if (child_writes[0].hdr.offset != offset)
 595                                 errx(1, "Child wrote to offset %zu, not %zu?",
 596                                      (size_t)child_writes[0].hdr.offset,
 597                                      (size_t)offset);
 598                         if (child_writes[0].hdr.len != count)
 599                                 errx(1, "Child wrote length %zu, not %zu?",
 600                                      child_writes[0].hdr.len, count);
 601                         if (memcmp(child_writes[0].data, buf, count)) {
 602                                 child_fail(NULL, 0,
 603                                            "Child wrote differently to"
 604                                            " fd %u than we did!\n", fd);
 605                         }
 606                         free(child_writes[0].data);
 607                         child_writes_num--;
 608                         memmove(&child_writes[0], &child_writes[1],
 609                                 sizeof(child_writes[0]) * child_writes_num);
 610
 611                         /* Is this is a socket or pipe, child wrote it
 612                            already. */
 613                         if (offset == (off_t)-1) {
 614                                 p->u.write.ret = count;
 615                                 errno = p->error;
 616                                 return p->u.write.ret;
 617                         }
 618                 }
 619                 p->u.write.ret = pwrite(fd, buf, count, off);
 620         }
 621         errno = p->error;
 622         return p->u.write.ret;
 623 }
 624
 625 ssize_t failtest_read(int fd, void *buf, size_t count,
 626                       const char *file, unsigned line)
 627 {
 628         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 629                               file, line);
 630 }
 631
 632 ssize_t failtest_write(int fd, const void *buf, size_t count,
 633                        const char *file, unsigned line)
 634 {
 635         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 636                                file, line);
 637 }
 638
 639 static struct lock_info *WARN_UNUSED_RESULT
 640 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 641 {
 642         unsigned int i;
 643         struct lock_info *l;
 644
 645         for (i = 0; i < lock_num; i++) {
 646                 l = &locks[i];
 647
 648                 if (l->fd != fd)
 649                         continue;
 650                 /* Four cases we care about:
 651                  * Start overlap:
 652                  *      l =    |      |
 653                  *      new = |   |
 654                  * Mid overlap:
 655                  *      l =    |      |
 656                  *      new =    |  |
 657                  * End overlap:
 658                  *      l =    |      |
 659                  *      new =      |    |
 660                  * Total overlap:
 661                  *      l =    |      |
 662                  *      new = |         |
 663                  */
 664                 if (start > l->start && end < l->end) {
 665                         /* Mid overlap: trim entry, add new one. */
 666                         off_t new_start, new_end;
 667                         new_start = end + 1;
 668                         new_end = l->end;
 669                         l->end = start - 1;
 670                         locks = add_lock(locks,
 671                                          fd, new_start, new_end, l->type);
 672                         l = &locks[i];
 673                 } else if (start <= l->start && end >= l->end) {
 674                         /* Total overlap: eliminate entry. */
 675                         l->end = 0;
 676                         l->start = 1;
 677                 } else if (end >= l->start && end < l->end) {
 678                         /* Start overlap: trim entry. */
 679                         l->start = end + 1;
 680                 } else if (start > l->start && start <= l->end) {
 681                         /* End overlap: trim entry. */
 682                         l->end = start-1;
 683                 }
 684                 /* Nothing left?  Remove it. */
 685                 if (l->end < l->start) {
 686                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
 687                         i--;
 688                 }
 689         }
 690
 691         if (type != F_UNLCK) {
 692                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
 693                 l = &locks[lock_num++];
 694                 l->fd = fd;
 695                 l->start = start;
 696                 l->end = end;
 697                 l->type = type;
 698         }
 699         return locks;
 700 }
 701
 702 /* We only trap this so we can dup fds in case we need to restore. */
 703 int failtest_close(int fd)
 704 {
 705         unsigned int i;
 706         int newfd = -1;
 707
 708         for (i = 0; i < fd_orig_num; i++) {
 709                 if (fd_orig[i].fd == fd) {
 710                         fd_orig[i].fd = newfd = dup(fd);
 711                         fd_orig[i].dupped = true;
 712                 }
 713         }
 714
 715         for (i = 0; i < writes_num; i++) {
 716                 if (writes[i].hdr.fd == fd)
 717                         writes[i].hdr.fd = newfd;
 718         }
 719
 720         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
 721         return close(fd);
 722 }
 723
 724 /* Zero length means "to end of file" */
 725 static off_t end_of(off_t start, off_t len)
 726 {
 727         if (len == 0)
 728                 return off_max();
 729         return start + len - 1;
 730 }
 731
 732 /* FIXME: This only handles locks, really. */
 733 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
 734 {
 735         struct failtest_call *p;
 736         struct fcntl_call call;
 737         va_list ap;
 738
 739         call.fd = fd;
 740         call.cmd = cmd;
 741
 742         /* Argument extraction. */
 743         switch (cmd) {
 744         case F_SETFL:
 745         case F_SETFD:
 746                 va_start(ap, cmd);
 747                 call.arg.l = va_arg(ap, long);
 748                 va_end(ap);
 749                 return fcntl(fd, cmd, call.arg.l);
 750         case F_GETFD:
 751         case F_GETFL:
 752                 return fcntl(fd, cmd);
 753         case F_GETLK:
 754                 get_locks();
 755                 va_start(ap, cmd);
 756                 call.arg.fl = *va_arg(ap, struct flock *);
 757                 va_end(ap);
 758                 return fcntl(fd, cmd, &call.arg.fl);
 759         case F_SETLK:
 760         case F_SETLKW:
 761                 va_start(ap, cmd);
 762                 call.arg.fl = *va_arg(ap, struct flock *);
 763                 va_end(ap);
 764                 break;
 765         default:
 766                 /* This means you need to implement it here. */
 767                 err(1, "failtest: unknown fcntl %u", cmd);
 768         }
 769
 770         p = add_history(FAILTEST_FCNTL, file, line, &call);
 771         get_locks();
 772
 773         if (should_fail(p)) {
 774                 p->u.fcntl.ret = -1;
 775                 if (p->u.fcntl.cmd == F_SETLK)
 776                         p->error = EAGAIN;
 777                 else
 778                         p->error = EDEADLK;
 779         } else {
 780                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
 781                                        &p->u.fcntl.arg.fl);
 782                 if (p->u.fcntl.ret == -1)
 783                         p->error = errno;
 784                 else {
 785                         /* We don't handle anything else yet. */
 786                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
 787                         locks = add_lock(locks,
 788                                          p->u.fcntl.fd,
 789                                          p->u.fcntl.arg.fl.l_start,
 790                                          end_of(p->u.fcntl.arg.fl.l_start,
 791                                                 p->u.fcntl.arg.fl.l_len),
 792                                          p->u.fcntl.arg.fl.l_type);
 793                 }
 794         }
 795         errno = p->error;
 796         return p->u.fcntl.ret;
 797 }
 798
 799 void failtest_init(int argc, char *argv[])
 800 {
 801         unsigned int i;
 802
 803         for (i = 1; i < argc; i++) {
 804                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
 805                         failpath = argv[i] + strlen("--failpath=");
 806                 } else if (strcmp(argv[i], "--tracepath") == 0) {
 807                         tracefd = dup(STDERR_FILENO);
 808                         failtest_timeout_ms = -1;
 809                 }
 810         }
 811         gettimeofday(&start, NULL);
 812 }
 813
 814 /* Free up memory, so valgrind doesn't report leaks. */
 815 static void free_everything(void)
 816 {
 817         unsigned int i;
 818
 819         for (i = 0; i < writes_num; i++) {
 820                 free(writes[i].data);
 821                 if (writes[i].hdr.offset != (off_t)-1)
 822                         free(writes[i].olddata);
 823         }
 824         free(writes);
 825         free(fd_orig);
 826         for (i = 0; i < history_num; i++) {
 827                 if (history[i].type == FAILTEST_OPEN)
 828                         free((char *)history[i].u.open.pathname);
 829         }
 830         free(history);
 831 }
 832
 833 void failtest_exit(int status)
 834 {
 835         unsigned int i;
 836
 837         if (control_fd == -1) {
 838                 free_everything();
 839                 exit(status);
 840         }
 841
 842         if (failtest_exit_check) {
 843                 if (!failtest_exit_check(history, history_num))
 844                         child_fail(NULL, 0, "failtest_exit_check failed\n");
 845         }
 846
 847         /* Restore any stuff we overwrote. */
 848         for (i = 0; i < writes_num; i++) {
 849                 if (writes[i].hdr.offset == (off_t)-1)
 850                         continue;
 851                 if (writes[i].oldlen != 0) {
 852                         lseek(writes[i].hdr.fd, writes[i].hdr.offset,
 853                               SEEK_SET);
 854                         write(writes[i].hdr.fd, writes[i].olddata,
 855                               writes[i].oldlen);
 856                 }
 857         }
 858
 859         /* Fix up fd offsets, restore sizes. */
 860         for (i = 0; i < fd_orig_num; i++) {
 861                 lseek(fd_orig[i].fd, fd_orig[i].offset, SEEK_SET);
 862                 ftruncate(fd_orig[i].fd, fd_orig[i].size);
 863                 /* Free up any file descriptors we dup'ed. */
 864                 if (fd_orig[i].dupped)
 865                         close(fd_orig[i].fd);
 866         }
 867
 868         free_everything();
 869         tell_parent(SUCCESS);
 870         exit(0);
 871 }