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

   1 #include "config.h"
   2 #include <stdarg.h>
   3 #include <string.h>
   4 #include <stdio.h>
   5 #include <stdarg.h>
   6 #include <ctype.h>
   7 #include <err.h>
   8 #include <unistd.h>
   9 #include <poll.h>
  10 #include <errno.h>
  11 #include <sys/types.h>
  12 #include <sys/wait.h>
  13 #include <sys/stat.h>
  14 #include <sys/time.h>
  15 #include <assert.h>
  16 #include <ccan/read_write_all/read_write_all.h>
  17 #include <ccan/failtest/failtest_proto.h>
  18 #include <ccan/failtest/failtest.h>
  19 #include <ccan/build_assert/build_assert.h>
  20
  21 enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
  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 static unsigned int probe_count = 0;
  52
  53 static struct write_call *child_writes = NULL;
  54 static unsigned int child_writes_num = 0;
  55
  56 static pid_t lock_owner;
  57 static struct lock_info *locks = NULL;
  58 static unsigned int lock_num = 0;
  59
  60 static const char info_to_arg[] = "mceoxprwf";
  61
  62 /* Dummy call used for failtest_undo wrappers. */
  63 static struct failtest_call unrecorded_call;
  64
  65 static struct failtest_call *add_history_(enum failtest_call_type type,
  66                                           const char *file,
  67                                           unsigned int line,
  68                                           const void *elem,
  69                                           size_t elem_size)
  70 {
  71         /* NULL file is how we suppress failure. */
  72         if (!file)
  73                 return &unrecorded_call;
  74
  75         history = realloc(history, (history_num + 1) * sizeof(*history));
  76         history[history_num].type = type;
  77         history[history_num].file = file;
  78         history[history_num].line = line;
  79         history[history_num].cleanup = NULL;
  80         memcpy(&history[history_num].u, elem, elem_size);
  81         return &history[history_num++];
  82 }
  83
  84 #define add_history(type, file, line, elem) \
  85         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
  86
  87 /* We do a fake call inside a sizeof(), to check types. */
  88 #define set_cleanup(call, clean, type)                  \
  89         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL),1), (clean))
  90
  91 static bool read_write_info(int fd)
  92 {
  93         struct write_call *w;
  94         char *buf;
  95
  96         /* We don't need all of this, but it's simple. */
  97         child_writes = realloc(child_writes,
  98                                (child_writes_num+1) * sizeof(child_writes[0]));
  99         w = &child_writes[child_writes_num];
 100         if (!read_all(fd, w, sizeof(*w)))
 101                 return false;
 102
 103         w->buf = buf = malloc(w->count);
 104         if (!read_all(fd, buf, w->count))
 105                 return false;
 106
 107         child_writes_num++;
 108         return true;
 109 }
 110
 111 static char *failpath_string(void)
 112 {
 113         unsigned int i;
 114         char *ret = malloc(history_num + 1);
 115
 116         for (i = 0; i < history_num; i++) {
 117                 ret[i] = info_to_arg[history[i].type];
 118                 if (history[i].fail)
 119                         ret[i] = toupper(ret[i]);
 120         }
 121         ret[i] = '\0';
 122         return ret;
 123 }
 124
 125 static void tell_parent(enum info_type type)
 126 {
 127         if (control_fd != -1)
 128                 write_all(control_fd, &type, sizeof(type));
 129 }
 130
 131 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 132 {
 133         va_list ap;
 134         char *path = failpath_string();
 135
 136         va_start(ap, fmt);
 137         vfprintf(stderr, fmt, ap);
 138         va_end(ap);
 139
 140         fprintf(stderr, "%.*s", (int)outlen, out);
 141         printf("To reproduce: --failpath=%s\n", path);
 142         free(path);
 143         tell_parent(FAILURE);
 144         exit(1);
 145 }
 146
 147 static void trace(const char *fmt, ...)
 148 {
 149         va_list ap;
 150
 151         if (tracefd == -1)
 152                 return;
 153
 154         va_start(ap, fmt);
 155         vdprintf(tracefd, fmt, ap);
 156         va_end(ap);
 157 }
 158
 159 static pid_t child;
 160
 161 static void hand_down(int signum)
 162 {
 163         kill(child, signum);
 164 }
 165
 166 static void release_locks(void)
 167 {
 168         /* Locks were never acquired/reacquired? */
 169         if (lock_owner == 0)
 170                 return;
 171
 172         /* We own them?  Release them all. */
 173         if (lock_owner == getpid()) {
 174                 unsigned int i;
 175                 struct flock fl;
 176                 fl.l_type = F_UNLCK;
 177                 fl.l_whence = SEEK_SET;
 178                 fl.l_start = 0;
 179                 fl.l_len = 0;
 180
 181                 for (i = 0; i < lock_num; i++)
 182                         fcntl(locks[i].fd, F_SETLK, &fl);
 183         } else {
 184                 /* Our parent must have them; pass request up. */
 185                 enum info_type type = RELEASE_LOCKS;
 186                 assert(control_fd != -1);
 187                 write_all(control_fd, &type, sizeof(type));
 188         }
 189         lock_owner = 0;
 190 }
 191
 192 /* off_t is a signed type.  Getting its max is non-trivial. */
 193 static off_t off_max(void)
 194 {
 195         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
 196         if (sizeof(off_t) == 4)
 197                 return (off_t)0x7FFFFFF;
 198         else
 199                 return (off_t)0x7FFFFFFFFFFFFFFULL;
 200 }
 201
 202 static void get_locks(void)
 203 {
 204         unsigned int i;
 205         struct flock fl;
 206
 207         if (lock_owner == getpid())
 208                 return;
 209
 210         if (lock_owner != 0) {
 211                 enum info_type type = RELEASE_LOCKS;
 212                 assert(control_fd != -1);
 213                 write_all(control_fd, &type, sizeof(type));
 214         }
 215
 216         fl.l_whence = SEEK_SET;
 217
 218         for (i = 0; i < lock_num; i++) {
 219                 fl.l_type = locks[i].type;
 220                 fl.l_start = locks[i].start;
 221                 if (locks[i].end == off_max())
 222                         fl.l_len = 0;
 223                 else
 224                         fl.l_len = locks[i].end - locks[i].start + 1;
 225
 226                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
 227                         abort();
 228         }
 229         lock_owner = getpid();
 230 }
 231
 232 struct saved_file {
 233         struct saved_file *next;
 234         int fd;
 235         void *contents;
 236         off_t off, len;
 237 };
 238
 239 static struct saved_file *save_file(struct saved_file *next, int fd)
 240 {
 241         struct saved_file *s = malloc(sizeof(*s));
 242
 243         s->next = next;
 244         s->fd = fd;
 245         s->off = lseek(fd, 0, SEEK_CUR);
 246         /* Special file?  Erk... */
 247         assert(s->off != -1);
 248         s->len = lseek(fd, 0, SEEK_END);
 249         lseek(fd, 0, SEEK_SET);
 250         s->contents = malloc(s->len);
 251         read(fd, s->contents, s->len);
 252         lseek(fd, s->off, SEEK_SET);
 253         return s;
 254 }
 255
 256 /* We have little choice but to save and restore open files: mmap means we
 257  * can really intercept changes in the child.
 258  *
 259  * We could do non-mmap'ed files on demand, however. */
 260 static struct saved_file *save_files(void)
 261 {
 262         struct saved_file *files = NULL;
 263         int i;
 264
 265         /* Figure out the set of live fds. */
 266         for (i = history_num - 2; i >= 0; i--) {
 267                 if (history[i].type == FAILTEST_OPEN) {
 268                         int fd = history[i].u.open.ret;
 269                         /* Only do successful, writable fds. */
 270                         if (fd < 0)
 271                                 continue;
 272
 273                         /* If it was closed, cleanup == NULL. */
 274                         if (!history[i].cleanup)
 275                                 continue;
 276
 277                         if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
 278                                 files = save_file(files, fd);
 279                         } else if ((history[i].u.open.flags & O_WRONLY)
 280                                    == O_WRONLY) {
 281                                 /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
 282                                 abort();
 283                         }
 284                 }
 285         }
 286
 287         return files;
 288 }
 289
 290 static void restore_files(struct saved_file *s)
 291 {
 292         while (s) {
 293                 struct saved_file *next = s->next;
 294
 295                 lseek(s->fd, 0, SEEK_SET);
 296                 write(s->fd, s->contents, s->len);
 297                 ftruncate(s->fd, s->len);
 298                 free(s->contents);
 299                 lseek(s->fd, s->off, SEEK_SET);
 300                 free(s);
 301                 s = next;
 302         }
 303 }
 304
 305 /* Free up memory, so valgrind doesn't report leaks. */
 306 static void free_everything(void)
 307 {
 308         unsigned int i;
 309
 310         /* We don't do this in cleanup: needed even for failed opens. */
 311         for (i = 0; i < history_num; i++) {
 312                 if (history[i].type == FAILTEST_OPEN)
 313                         free((char *)history[i].u.open.pathname);
 314         }
 315         free(history);
 316 }
 317
 318 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
 319 {
 320         int i;
 321
 322         /* For children, we don't care if they "failed" the testing. */
 323         if (control_fd != -1)
 324                 status = 0;
 325
 326         if (forced_cleanup)
 327                 history_num--;
 328
 329         /* Cleanup everything, in reverse order. */
 330         for (i = history_num - 1; i >= 0; i--) {
 331                 if (!history[i].cleanup)
 332                         continue;
 333                 if (!forced_cleanup) {
 334                         printf("Leak at %s:%u: --failpath=%s\n",
 335                                history[i].file, history[i].line,
 336                                failpath_string());
 337                         status = 1;
 338                 }
 339                 history[i].cleanup(&history[i].u);
 340         }
 341
 342         free_everything();
 343         tell_parent(SUCCESS);
 344         exit(status);
 345 }
 346
 347 static bool should_fail(struct failtest_call *call)
 348 {
 349         int status;
 350         int control[2], output[2];
 351         enum info_type type = UNEXPECTED;
 352         char *out = NULL;
 353         size_t outlen = 0;
 354         struct saved_file *files;
 355
 356         /* Are we probing? */
 357         if (probe_count && --probe_count == 0)
 358                 failtest_cleanup(true, 0);
 359
 360         if (call == &unrecorded_call)
 361                 return false;
 362
 363         if (failpath) {
 364                 /* + means continue after end, like normal. */
 365                 if (*failpath == '+')
 366                         failpath = NULL;
 367                 else {
 368                         if (tolower((unsigned char)*failpath)
 369                             != info_to_arg[call->type])
 370                                 errx(1, "Failpath expected '%c' got '%c'\n",
 371                                      info_to_arg[call->type], *failpath);
 372                         call->fail = isupper((unsigned char)*(failpath++));
 373                         return call->fail;
 374                 }
 375         }
 376
 377         /* Attach debugger if they asked for it. */
 378         if (debugpath && history_num == strlen(debugpath)) {
 379                 unsigned int i;
 380
 381                 for (i = 0; i < history_num; i++) {
 382                         unsigned char c = info_to_arg[history[i].type];
 383                         if (history[i].fail)
 384                                 c = toupper(c);
 385                         if (c != debugpath[i])
 386                                 break;
 387                 }
 388                 if (i == history_num) {
 389                         char str[80];
 390
 391                         /* Don't timeout. */
 392                         signal(SIGUSR1, SIG_IGN);
 393                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
 394                                 getpid(), getpid());
 395                         system(str);
 396                         sleep(5);
 397                 }
 398         }
 399
 400         if (failtest_hook) {
 401                 switch (failtest_hook(history, history_num)) {
 402                 case FAIL_OK:
 403                         break;
 404                 case FAIL_DONT_FAIL:
 405                         call->fail = false;
 406                         return false;
 407                 case FAIL_PROBE:
 408                         /* Already down probe path?  Stop now. */
 409                         if (probe_count)
 410                                 failtest_cleanup(true, 0);
 411                         /* FIXME: We should run *parent* and run probe until
 412                          * calls match up again. */
 413                         probe_count = 3;
 414                         break;
 415                 default:
 416                         abort();
 417                 }
 418         }
 419
 420         files = save_files();
 421
 422         /* We're going to fail in the child. */
 423         call->fail = true;
 424         if (pipe(control) != 0 || pipe(output) != 0)
 425                 err(1, "opening pipe");
 426
 427         /* Prevent double-printing (in child and parent) */
 428         fflush(stdout);
 429         child = fork();
 430         if (child == -1)
 431                 err(1, "forking failed");
 432
 433         if (child == 0) {
 434                 if (tracefd != -1) {
 435                         struct timeval now;
 436                         const char *p;
 437                         gettimeofday(&now, NULL);
 438                         if (now.tv_usec < start.tv_usec) {
 439                                 now.tv_sec--;
 440                                 now.tv_usec += 1000000;
 441                         }
 442                         now.tv_usec -= start.tv_usec;
 443                         now.tv_sec -= start.tv_sec;
 444                         p = failpath_string();
 445                         trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
 446                               (int)now.tv_sec, (int)now.tv_usec / 10000, p);
 447                         free((char *)p);
 448                         p = strrchr(history[history_num-1].file, '/');
 449                         if (p)
 450                                 trace("%s", p+1);
 451                         else
 452                                 trace("%s", history[history_num-1].file);
 453                         trace(":%u)\n", history[history_num-1].line);
 454                 }
 455                 close(control[0]);
 456                 close(output[0]);
 457                 dup2(output[1], STDOUT_FILENO);
 458                 dup2(output[1], STDERR_FILENO);
 459                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
 460                         close(output[1]);
 461                 control_fd = control[1];
 462                 return true;
 463         }
 464
 465         signal(SIGUSR1, hand_down);
 466
 467         close(control[1]);
 468         close(output[1]);
 469
 470         /* We grab output so we can display it; we grab writes so we
 471          * can compare. */
 472         do {
 473                 struct pollfd pfd[2];
 474                 int ret;
 475
 476                 pfd[0].fd = output[0];
 477                 pfd[0].events = POLLIN|POLLHUP;
 478                 pfd[1].fd = control[0];
 479                 pfd[1].events = POLLIN|POLLHUP;
 480
 481                 if (type == SUCCESS)
 482                         ret = poll(pfd, 1, failtest_timeout_ms);
 483                 else
 484                         ret = poll(pfd, 2, failtest_timeout_ms);
 485
 486                 if (ret == 0)
 487                         hand_down(SIGUSR1);
 488                 if (ret < 0) {
 489                         if (errno == EINTR)
 490                                 continue;
 491                         err(1, "Poll returned %i", ret);
 492                 }
 493
 494                 if (pfd[0].revents & POLLIN) {
 495                         ssize_t len;
 496
 497                         out = realloc(out, outlen + 8192);
 498                         len = read(output[0], out + outlen, 8192);
 499                         outlen += len;
 500                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 501                         if (read_all(control[0], &type, sizeof(type))) {
 502                                 if (type == WRITE) {
 503                                         if (!read_write_info(control[0]))
 504                                                 break;
 505                                 } else if (type == RELEASE_LOCKS) {
 506                                         release_locks();
 507                                         /* FIXME: Tell them we're done... */
 508                                 }
 509                         }
 510                 } else if (pfd[0].revents & POLLHUP) {
 511                         break;
 512                 }
 513         } while (type != FAILURE);
 514
 515         close(output[0]);
 516         close(control[0]);
 517         waitpid(child, &status, 0);
 518         if (!WIFEXITED(status)) {
 519                 if (WTERMSIG(status) == SIGUSR1)
 520                         child_fail(out, outlen, "Timed out");
 521                 else
 522                         child_fail(out, outlen, "Killed by signal %u: ",
 523                                    WTERMSIG(status));
 524         }
 525         /* Child printed failure already, just pass up exit code. */
 526         if (type == FAILURE) {
 527                 fprintf(stderr, "%.*s", (int)outlen, out);
 528                 tell_parent(type);
 529                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 530         }
 531         if (WEXITSTATUS(status) != 0)
 532                 child_fail(out, outlen, "Exited with status %i: ",
 533                            WEXITSTATUS(status));
 534
 535         free(out);
 536         signal(SIGUSR1, SIG_DFL);
 537
 538         restore_files(files);
 539
 540         /* We continue onwards without failing. */
 541         call->fail = false;
 542         return false;
 543 }
 544
 545 static void cleanup_calloc(struct calloc_call *call)
 546 {
 547         free(call->ret);
 548 }
 549
 550 void *failtest_calloc(size_t nmemb, size_t size,
 551                       const char *file, unsigned line)
 552 {
 553         struct failtest_call *p;
 554         struct calloc_call call;
 555         call.nmemb = nmemb;
 556         call.size = size;
 557         p = add_history(FAILTEST_CALLOC, file, line, &call);
 558
 559         if (should_fail(p)) {
 560                 p->u.calloc.ret = NULL;
 561                 p->error = ENOMEM;
 562         } else {
 563                 p->u.calloc.ret = calloc(nmemb, size);
 564                 set_cleanup(p, cleanup_calloc, struct calloc_call);
 565         }
 566         errno = p->error;
 567         return p->u.calloc.ret;
 568 }
 569
 570 static void cleanup_malloc(struct malloc_call *call)
 571 {
 572         free(call->ret);
 573 }
 574
 575 void *failtest_malloc(size_t size, const char *file, unsigned line)
 576 {
 577         struct failtest_call *p;
 578         struct malloc_call call;
 579         call.size = size;
 580
 581         p = add_history(FAILTEST_MALLOC, file, line, &call);
 582         if (should_fail(p)) {
 583                 p->u.calloc.ret = NULL;
 584                 p->error = ENOMEM;
 585         } else {
 586                 p->u.calloc.ret = malloc(size);
 587                 set_cleanup(p, cleanup_malloc, struct malloc_call);
 588         }
 589         errno = p->error;
 590         return p->u.calloc.ret;
 591 }
 592
 593 static void cleanup_realloc(struct realloc_call *call)
 594 {
 595         free(call->ret);
 596 }
 597
 598 /* Walk back and find out if we got this ptr from a previous routine. */
 599 static void fixup_ptr_history(void *ptr, unsigned int last)
 600 {
 601         int i;
 602
 603         /* Start at end of history, work back. */
 604         for (i = last - 1; i >= 0; i--) {
 605                 switch (history[i].type) {
 606                 case FAILTEST_REALLOC:
 607                         if (history[i].u.realloc.ret == ptr) {
 608                                 history[i].cleanup = NULL;
 609                                 return;
 610                         }
 611                         break;
 612                 case FAILTEST_MALLOC:
 613                         if (history[i].u.malloc.ret == ptr) {
 614                                 history[i].cleanup = NULL;
 615                                 return;
 616                         }
 617                         break;
 618                 case FAILTEST_CALLOC:
 619                         if (history[i].u.calloc.ret == ptr) {
 620                                 history[i].cleanup = NULL;
 621                                 return;
 622                         }
 623                         break;
 624                 default:
 625                         break;
 626                 }
 627         }
 628 }
 629
 630 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 631 {
 632         struct failtest_call *p;
 633         struct realloc_call call;
 634         call.size = size;
 635         p = add_history(FAILTEST_REALLOC, file, line, &call);
 636
 637         /* FIXME: Try one child moving allocation, one not. */
 638         if (should_fail(p)) {
 639                 p->u.realloc.ret = NULL;
 640                 p->error = ENOMEM;
 641         } else {
 642                 fixup_ptr_history(ptr, history_num-1);
 643                 p->u.realloc.ret = realloc(ptr, size);
 644                 set_cleanup(p, cleanup_realloc, struct realloc_call);
 645         }
 646         errno = p->error;
 647         return p->u.realloc.ret;
 648 }
 649
 650 void failtest_free(void *ptr)
 651 {
 652         fixup_ptr_history(ptr, history_num);
 653         free(ptr);
 654 }
 655
 656 static void cleanup_open(struct open_call *call)
 657 {
 658         close(call->ret);
 659 }
 660
 661 int failtest_open(const char *pathname,
 662                   const char *file, unsigned line, ...)
 663 {
 664         struct failtest_call *p;
 665         struct open_call call;
 666         va_list ap;
 667
 668         call.pathname = strdup(pathname);
 669         va_start(ap, line);
 670         call.flags = va_arg(ap, int);
 671         if (call.flags & O_CREAT) {
 672                 call.mode = va_arg(ap, mode_t);
 673                 va_end(ap);
 674         }
 675         p = add_history(FAILTEST_OPEN, file, line, &call);
 676         /* Avoid memory leak! */
 677         if (p == &unrecorded_call)
 678                 free((char *)call.pathname);
 679         p->u.open.ret = open(pathname, call.flags, call.mode);
 680
 681         if (!failpath && p->u.open.ret == -1) {
 682                 p->fail = false;
 683                 p->error = errno;
 684         } else if (should_fail(p)) {
 685                 close(p->u.open.ret);
 686                 p->u.open.ret = -1;
 687                 /* FIXME: Play with error codes? */
 688                 p->error = EACCES;
 689         } else {
 690                 set_cleanup(p, cleanup_open, struct open_call);
 691         }
 692         errno = p->error;
 693         return p->u.open.ret;
 694 }
 695
 696 static void cleanup_pipe(struct pipe_call *call)
 697 {
 698         if (!call->closed[0])
 699                 close(call->fds[0]);
 700         if (!call->closed[1])
 701                 close(call->fds[1]);
 702 }
 703
 704 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
 705 {
 706         struct failtest_call *p;
 707         struct pipe_call call;
 708
 709         p = add_history(FAILTEST_PIPE, file, line, &call);
 710         if (should_fail(p)) {
 711                 p->u.open.ret = -1;
 712                 /* FIXME: Play with error codes? */
 713                 p->error = EMFILE;
 714         } else {
 715                 p->u.pipe.ret = pipe(p->u.pipe.fds);
 716                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
 717                 set_cleanup(p, cleanup_pipe, struct pipe_call);
 718         }
 719         /* This causes valgrind to notice if they use pipefd[] after failure */
 720         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
 721         errno = p->error;
 722         return p->u.pipe.ret;
 723 }
 724
 725 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
 726                        const char *file, unsigned line)
 727 {
 728         struct failtest_call *p;
 729         struct read_call call;
 730         call.fd = fd;
 731         call.buf = buf;
 732         call.count = count;
 733         call.off = off;
 734         p = add_history(FAILTEST_READ, file, line, &call);
 735
 736         /* FIXME: Try partial read returns. */
 737         if (should_fail(p)) {
 738                 p->u.read.ret = -1;
 739                 p->error = EIO;
 740         } else {
 741                 p->u.read.ret = pread(fd, buf, count, off);
 742         }
 743         errno = p->error;
 744         return p->u.read.ret;
 745 }
 746
 747 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
 748                         const char *file, unsigned line)
 749 {
 750         struct failtest_call *p;
 751         struct write_call call;
 752
 753         call.fd = fd;
 754         call.buf = buf;
 755         call.count = count;
 756         call.off = off;
 757         p = add_history(FAILTEST_WRITE, file, line, &call);
 758
 759         /* If we're a child, we need to make sure we write the same thing
 760          * to non-files as the parent does, so tell it. */
 761         if (control_fd != -1 && off == (off_t)-1) {
 762                 enum info_type type = WRITE;
 763
 764                 write_all(control_fd, &type, sizeof(type));
 765                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
 766                 write_all(control_fd, buf, count);
 767         }
 768
 769         /* FIXME: Try partial write returns. */
 770         if (should_fail(p)) {
 771                 p->u.write.ret = -1;
 772                 p->error = EIO;
 773         } else {
 774                 /* FIXME: We assume same write order in parent and child */
 775                 if (off == (off_t)-1 && child_writes_num != 0) {
 776                         if (child_writes[0].fd != fd)
 777                                 errx(1, "Child wrote to fd %u, not %u?",
 778                                      child_writes[0].fd, fd);
 779                         if (child_writes[0].off != p->u.write.off)
 780                                 errx(1, "Child wrote to offset %zu, not %zu?",
 781                                      (size_t)child_writes[0].off,
 782                                      (size_t)p->u.write.off);
 783                         if (child_writes[0].count != count)
 784                                 errx(1, "Child wrote length %zu, not %zu?",
 785                                      child_writes[0].count, count);
 786                         if (memcmp(child_writes[0].buf, buf, count)) {
 787                                 child_fail(NULL, 0,
 788                                            "Child wrote differently to"
 789                                            " fd %u than we did!\n", fd);
 790                         }
 791                         free((char *)child_writes[0].buf);
 792                         child_writes_num--;
 793                         memmove(&child_writes[0], &child_writes[1],
 794                                 sizeof(child_writes[0]) * child_writes_num);
 795
 796                         /* Is this is a socket or pipe, child wrote it
 797                            already. */
 798                         if (p->u.write.off == (off_t)-1) {
 799                                 p->u.write.ret = count;
 800                                 errno = p->error;
 801                                 return p->u.write.ret;
 802                         }
 803                 }
 804                 p->u.write.ret = pwrite(fd, buf, count, off);
 805         }
 806         errno = p->error;
 807         return p->u.write.ret;
 808 }
 809
 810 ssize_t failtest_read(int fd, void *buf, size_t count,
 811                       const char *file, unsigned line)
 812 {
 813         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 814                               file, line);
 815 }
 816
 817 ssize_t failtest_write(int fd, const void *buf, size_t count,
 818                        const char *file, unsigned line)
 819 {
 820         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 821                                file, line);
 822 }
 823
 824 static struct lock_info *WARN_UNUSED_RESULT
 825 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 826 {
 827         unsigned int i;
 828         struct lock_info *l;
 829
 830         for (i = 0; i < lock_num; i++) {
 831                 l = &locks[i];
 832
 833                 if (l->fd != fd)
 834                         continue;
 835                 /* Four cases we care about:
 836                  * Start overlap:
 837                  *      l =    |      |
 838                  *      new = |   |
 839                  * Mid overlap:
 840                  *      l =    |      |
 841                  *      new =    |  |
 842                  * End overlap:
 843                  *      l =    |      |
 844                  *      new =      |    |
 845                  * Total overlap:
 846                  *      l =    |      |
 847                  *      new = |         |
 848                  */
 849                 if (start > l->start && end < l->end) {
 850                         /* Mid overlap: trim entry, add new one. */
 851                         off_t new_start, new_end;
 852                         new_start = end + 1;
 853                         new_end = l->end;
 854                         l->end = start - 1;
 855                         locks = add_lock(locks,
 856                                          fd, new_start, new_end, l->type);
 857                         l = &locks[i];
 858                 } else if (start <= l->start && end >= l->end) {
 859                         /* Total overlap: eliminate entry. */
 860                         l->end = 0;
 861                         l->start = 1;
 862                 } else if (end >= l->start && end < l->end) {
 863                         /* Start overlap: trim entry. */
 864                         l->start = end + 1;
 865                 } else if (start > l->start && start <= l->end) {
 866                         /* End overlap: trim entry. */
 867                         l->end = start-1;
 868                 }
 869                 /* Nothing left?  Remove it. */
 870                 if (l->end < l->start) {
 871                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
 872                         i--;
 873                 }
 874         }
 875
 876         if (type != F_UNLCK) {
 877                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
 878                 l = &locks[lock_num++];
 879                 l->fd = fd;
 880                 l->start = start;
 881                 l->end = end;
 882                 l->type = type;
 883         }
 884         return locks;
 885 }
 886
 887 /* We trap this so we can record it: we don't fail it. */
 888 int failtest_close(int fd, const char *file, unsigned line)
 889 {
 890         int i;
 891         struct close_call call;
 892         struct failtest_call *p;
 893
 894         call.fd = fd;
 895         p = add_history(FAILTEST_CLOSE, file, line, &call);
 896         p->fail = false;
 897
 898         /* Consume close from failpath. */
 899         if (failpath)
 900                 if (should_fail(p))
 901                         abort();
 902
 903         if (fd < 0)
 904                 return close(fd);
 905
 906         /* Trace history to find source of fd. */
 907         for (i = history_num-1; i >= 0; i--) {
 908                 switch (history[i].type) {
 909                 case FAILTEST_PIPE:
 910                         /* From a pipe? */
 911                         if (history[i].u.pipe.fds[0] == fd) {
 912                                 assert(!history[i].u.pipe.closed[0]);
 913                                 history[i].u.pipe.closed[0] = true;
 914                                 if (history[i].u.pipe.closed[1])
 915                                         history[i].cleanup = NULL;
 916                                 goto out;
 917                         }
 918                         if (history[i].u.pipe.fds[1] == fd) {
 919                                 assert(!history[i].u.pipe.closed[1]);
 920                                 history[i].u.pipe.closed[1] = true;
 921                                 if (history[i].u.pipe.closed[0])
 922                                         history[i].cleanup = NULL;
 923                                 goto out;
 924                         }
 925                         break;
 926                 case FAILTEST_OPEN:
 927                         if (history[i].u.open.ret == fd) {
 928                                 assert((void *)history[i].cleanup
 929                                        == (void *)cleanup_open);
 930                                 history[i].cleanup = NULL;
 931                                 goto out;
 932                         }
 933                         break;
 934                 default:
 935                         break;
 936                 }
 937         }
 938
 939 out:
 940         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
 941         return close(fd);
 942 }
 943
 944 /* Zero length means "to end of file" */
 945 static off_t end_of(off_t start, off_t len)
 946 {
 947         if (len == 0)
 948                 return off_max();
 949         return start + len - 1;
 950 }
 951
 952 /* FIXME: This only handles locks, really. */
 953 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
 954 {
 955         struct failtest_call *p;
 956         struct fcntl_call call;
 957         va_list ap;
 958
 959         call.fd = fd;
 960         call.cmd = cmd;
 961
 962         /* Argument extraction. */
 963         switch (cmd) {
 964         case F_SETFL:
 965         case F_SETFD:
 966                 va_start(ap, cmd);
 967                 call.arg.l = va_arg(ap, long);
 968                 va_end(ap);
 969                 return fcntl(fd, cmd, call.arg.l);
 970         case F_GETFD:
 971         case F_GETFL:
 972                 return fcntl(fd, cmd);
 973         case F_GETLK:
 974                 get_locks();
 975                 va_start(ap, cmd);
 976                 call.arg.fl = *va_arg(ap, struct flock *);
 977                 va_end(ap);
 978                 return fcntl(fd, cmd, &call.arg.fl);
 979         case F_SETLK:
 980         case F_SETLKW:
 981                 va_start(ap, cmd);
 982                 call.arg.fl = *va_arg(ap, struct flock *);
 983                 va_end(ap);
 984                 break;
 985         default:
 986                 /* This means you need to implement it here. */
 987                 err(1, "failtest: unknown fcntl %u", cmd);
 988         }
 989
 990         p = add_history(FAILTEST_FCNTL, file, line, &call);
 991
 992         if (should_fail(p)) {
 993                 p->u.fcntl.ret = -1;
 994                 if (p->u.fcntl.cmd == F_SETLK)
 995                         p->error = EAGAIN;
 996                 else
 997                         p->error = EDEADLK;
 998         } else {
 999                 get_locks();
1000                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1001                                        &p->u.fcntl.arg.fl);
1002                 if (p->u.fcntl.ret == -1)
1003                         p->error = errno;
1004                 else {
1005                         /* We don't handle anything else yet. */
1006                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1007                         locks = add_lock(locks,
1008                                          p->u.fcntl.fd,
1009                                          p->u.fcntl.arg.fl.l_start,
1010                                          end_of(p->u.fcntl.arg.fl.l_start,
1011                                                 p->u.fcntl.arg.fl.l_len),
1012                                          p->u.fcntl.arg.fl.l_type);
1013                 }
1014         }
1015         errno = p->error;
1016         return p->u.fcntl.ret;
1017 }
1018
1019 void failtest_init(int argc, char *argv[])
1020 {
1021         unsigned int i;
1022
1023         for (i = 1; i < argc; i++) {
1024                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1025                         failpath = argv[i] + strlen("--failpath=");
1026                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1027                         tracefd = dup(STDERR_FILENO);
1028                         failtest_timeout_ms = -1;
1029                 } else if (!strncmp(argv[i], "--debugpath=",
1030                                     strlen("--debugpath="))) {
1031                         debugpath = argv[i] + strlen("--debugpath=");
1032                 }
1033         }
1034         gettimeofday(&start, NULL);
1035 }
1036
1037 void failtest_exit(int status)
1038 {
1039         if (failtest_exit_check) {
1040                 if (!failtest_exit_check(history, history_num))
1041                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1042         }
1043
1044         failtest_cleanup(false, status);
1045 }