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

   1 /* Licensed under LGPL - see LICENSE file for details */
   2 #include <ccan/failtest/failtest.h>
   3 #include <stdarg.h>
   4 #include <string.h>
   5 #include <stdio.h>
   6 #include <stdarg.h>
   7 #include <ctype.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 <sys/mman.h>
  16 #include <sys/resource.h>
  17 #include <signal.h>
  18 #include <assert.h>
  19 #include <ccan/err/err.h>
  20 #include <ccan/time/time.h>
  21 #include <ccan/read_write_all/read_write_all.h>
  22 #include <ccan/failtest/failtest_proto.h>
  23 #include <ccan/build_assert/build_assert.h>
  24 #include <ccan/hash/hash.h>
  25 #include <ccan/htable/htable_type.h>
  26 #include <ccan/str/str.h>
  27 #include <ccan/compiler/compiler.h>
  28
  29 enum failtest_result (*failtest_hook)(struct tlist_calls *);
  30
  31 static FILE *tracef = NULL, *warnf;
  32 static int traceindent = 0;
  33
  34 unsigned int failtest_timeout_ms = 20000;
  35
  36 const char *failpath;
  37 const char *debugpath;
  38
  39 enum info_type {
  40         WRITE,
  41         RELEASE_LOCKS,
  42         FAILURE,
  43         SUCCESS,
  44         UNEXPECTED
  45 };
  46
  47 struct lock_info {
  48         int fd;
  49         /* end is inclusive: you can't have a 0-byte lock. */
  50         off_t start, end;
  51         int type;
  52 };
  53
  54 /* We hash the call location together with its backtrace. */
  55 static size_t hash_call(const struct failtest_call *call)
  56 {
  57         return hash(call->file, strlen(call->file),
  58                     hash(&call->line, 1,
  59                          hash(call->backtrace, call->backtrace_num,
  60                               call->type)));
  61 }
  62
  63 static bool call_eq(const struct failtest_call *call1,
  64                     const struct failtest_call *call2)
  65 {
  66         unsigned int i;
  67
  68         if (strcmp(call1->file, call2->file) != 0
  69             || call1->line != call2->line
  70             || call1->type != call2->type
  71             || call1->backtrace_num != call2->backtrace_num)
  72                 return false;
  73
  74         for (i = 0; i < call1->backtrace_num; i++)
  75                 if (call1->backtrace[i] != call2->backtrace[i])
  76                         return false;
  77
  78         return true;
  79 }
  80
  81 /* Defines struct failtable. */
  82 HTABLE_DEFINE_TYPE(struct failtest_call, (struct failtest_call *), hash_call,
  83                    call_eq, failtable);
  84
  85 bool (*failtest_exit_check)(struct tlist_calls *history);
  86
  87 /* The entire history of all calls. */
  88 static struct tlist_calls history = TLIST_INIT(history);
  89 /* If we're a child, the fd two write control info to the parent. */
  90 static int control_fd = -1;
  91 /* If we're a child, this is the first call we did ourselves. */
  92 static struct failtest_call *our_history_start = NULL;
  93 /* For printing runtime with --trace. */
  94 static struct timeabs start;
  95 /* Set when failtest_hook returns FAIL_PROBE */
  96 static bool probing = false;
  97 /* Table to track duplicates. */
  98 static struct failtable failtable;
  99
 100 /* Array of writes which our child did.  We report them on failure. */
 101 static struct write_call *child_writes = NULL;
 102 static unsigned int child_writes_num = 0;
 103
 104 /* fcntl locking info. */
 105 static pid_t lock_owner;
 106 static struct lock_info *locks = NULL;
 107 static unsigned int lock_num = 0;
 108
 109 /* Our original pid, which we return to anyone who asks. */
 110 static pid_t orig_pid;
 111
 112 /* Mapping from failtest_type to char. */
 113 static const char info_to_arg[] = "mceoxprwfal";
 114
 115 /* Dummy call used for failtest_undo wrappers. */
 116 static struct failtest_call unrecorded_call;
 117
 118 struct contents_saved {
 119         size_t count;
 120         off_t off;
 121         off_t old_len;
 122         char contents[1];
 123 };
 124
 125 /* File contents, saved in this child only. */
 126 struct saved_mmapped_file {
 127         struct saved_mmapped_file *next;
 128         struct failtest_call *opener;
 129         struct contents_saved *s;
 130 };
 131
 132 static struct saved_mmapped_file *saved_mmapped_files;
 133
 134 #if HAVE_BACKTRACE
 135 #include <execinfo.h>
 136
 137 static void **get_backtrace(unsigned int *num)
 138 {
 139         static unsigned int max_back = 100;
 140         void **ret;
 141
 142 again:
 143         ret = malloc(max_back * sizeof(void *));
 144         *num = backtrace(ret, max_back);
 145         if (*num == max_back) {
 146                 free(ret);
 147                 max_back *= 2;
 148                 goto again;
 149         }
 150         return ret;
 151 }
 152 #else
 153 /* This will test slightly less, since will consider all of the same
 154  * calls as identical.  But, it's slightly faster! */
 155 static void **get_backtrace(unsigned int *num)
 156 {
 157         *num = 0;
 158         return NULL;
 159 }
 160 #endif /* HAVE_BACKTRACE */
 161
 162 static struct failtest_call *add_history_(enum failtest_call_type type,
 163                                           bool can_leak,
 164                                           const char *file,
 165                                           unsigned int line,
 166                                           const void *elem,
 167                                           size_t elem_size)
 168 {
 169         struct failtest_call *call;
 170
 171         /* NULL file is how we suppress failure. */
 172         if (!file)
 173                 return &unrecorded_call;
 174
 175         call = malloc(sizeof *call);
 176         call->type = type;
 177         call->can_leak = can_leak;
 178         call->file = file;
 179         call->line = line;
 180         call->cleanup = NULL;
 181         call->backtrace = get_backtrace(&call->backtrace_num);
 182         memcpy(&call->u, elem, elem_size);
 183         tlist_add_tail(&history, call, list);
 184         return call;
 185 }
 186
 187 #define add_history(type, can_leak, file, line, elem)           \
 188         add_history_((type), (can_leak), (file), (line), (elem), sizeof(*(elem)))
 189
 190 /* We do a fake call inside a sizeof(), to check types. */
 191 #define set_cleanup(call, clean, type)                  \
 192         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL, false),1), (clean))
 193
 194 /* Dup the fd to a high value (out of the way I hope!), and close the old fd. */
 195 static int move_fd_to_high(int fd)
 196 {
 197         int i;
 198         struct rlimit lim;
 199         int max;
 200
 201         if (getrlimit(RLIMIT_NOFILE, &lim) == 0) {
 202                 max = lim.rlim_cur;
 203                 printf("Max is %i\n", max);
 204         } else
 205                 max = FD_SETSIZE;
 206
 207         for (i = max - 1; i > fd; i--) {
 208                 if (fcntl(i, F_GETFL) == -1 && errno == EBADF) {
 209                         if (dup2(fd, i) == -1) {
 210                                 warn("Failed to dup fd %i to %i", fd, i);
 211                                 continue;
 212                         }
 213                         close(fd);
 214                         return i;
 215                 }
 216         }
 217         /* Nothing?  Really?  Er... ok? */
 218         return fd;
 219 }
 220
 221 static bool read_write_info(int fd)
 222 {
 223         struct write_call *w;
 224         char *buf;
 225
 226         /* We don't need all of this, but it's simple. */
 227         child_writes = realloc(child_writes,
 228                                (child_writes_num+1) * sizeof(child_writes[0]));
 229         w = &child_writes[child_writes_num];
 230         if (!read_all(fd, w, sizeof(*w)))
 231                 return false;
 232
 233         w->buf = buf = malloc(w->count);
 234         if (!read_all(fd, buf, w->count))
 235                 return false;
 236
 237         child_writes_num++;
 238         return true;
 239 }
 240
 241 static char *failpath_string(void)
 242 {
 243         struct failtest_call *i;
 244         char *ret = strdup("");
 245         unsigned len = 0;
 246
 247         /* Inefficient, but who cares? */
 248         tlist_for_each(&history, i, list) {
 249                 ret = realloc(ret, len + 2);
 250                 ret[len] = info_to_arg[i->type];
 251                 if (i->fail)
 252                         ret[len] = toupper(ret[len]);
 253                 ret[++len] = '\0';
 254         }
 255         return ret;
 256 }
 257
 258 static void do_warn(int e, const char *fmt, va_list ap)
 259 {
 260         char *p = failpath_string();
 261
 262         vfprintf(warnf, fmt, ap);
 263         if (e != -1)
 264                 fprintf(warnf, ": %s", strerror(e));
 265         fprintf(warnf, " [%s]\n", p);
 266         free(p);
 267 }
 268
 269 static void fwarn(const char *fmt, ...)
 270 {
 271         va_list ap;
 272         int e = errno;
 273
 274         va_start(ap, fmt);
 275         do_warn(e, fmt, ap);
 276         va_end(ap);
 277 }
 278
 279
 280 static void fwarnx(const char *fmt, ...)
 281 {
 282         va_list ap;
 283
 284         va_start(ap, fmt);
 285         do_warn(-1, fmt, ap);
 286         va_end(ap);
 287 }
 288
 289 static void tell_parent(enum info_type type)
 290 {
 291         if (control_fd != -1)
 292                 write_all(control_fd, &type, sizeof(type));
 293 }
 294
 295 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 296 {
 297         va_list ap;
 298         char *path = failpath_string();
 299
 300         va_start(ap, fmt);
 301         vfprintf(stderr, fmt, ap);
 302         va_end(ap);
 303
 304         fprintf(stderr, "%.*s", (int)outlen, out);
 305         printf("To reproduce: --failpath=%s\n", path);
 306         free(path);
 307         tell_parent(FAILURE);
 308         exit(1);
 309 }
 310
 311 static void PRINTF_FMT(1, 2) trace(const char *fmt, ...)
 312 {
 313         va_list ap;
 314         unsigned int i;
 315         char *p;
 316         static int idx;
 317
 318         if (!tracef)
 319                 return;
 320
 321         for (i = 0; i < traceindent; i++)
 322                 fprintf(tracef, "  ");
 323
 324         p = failpath_string();
 325         fprintf(tracef, "%i: %u: %s ", idx++, getpid(), p);
 326         va_start(ap, fmt);
 327         vfprintf(tracef, fmt, ap);
 328         va_end(ap);
 329         free(p);
 330 }
 331
 332 static pid_t child;
 333
 334 static void hand_down(int signum)
 335 {
 336         kill(child, signum);
 337 }
 338
 339 static void release_locks(void)
 340 {
 341         /* Locks were never acquired/reacquired? */
 342         if (lock_owner == 0)
 343                 return;
 344
 345         /* We own them?  Release them all. */
 346         if (lock_owner == getpid()) {
 347                 unsigned int i;
 348                 struct flock fl;
 349                 fl.l_type = F_UNLCK;
 350                 fl.l_whence = SEEK_SET;
 351                 fl.l_start = 0;
 352                 fl.l_len = 0;
 353
 354                 trace("Releasing %u locks\n", lock_num);
 355                 for (i = 0; i < lock_num; i++)
 356                         fcntl(locks[i].fd, F_SETLK, &fl);
 357         } else {
 358                 /* Our parent must have them; pass request up. */
 359                 enum info_type type = RELEASE_LOCKS;
 360                 assert(control_fd != -1);
 361                 write_all(control_fd, &type, sizeof(type));
 362         }
 363         lock_owner = 0;
 364 }
 365
 366 /* off_t is a signed type.  Getting its max is non-trivial. */
 367 static off_t off_max(void)
 368 {
 369         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
 370         if (sizeof(off_t) == 4)
 371                 return (off_t)0x7FFFFFF;
 372         else
 373                 return (off_t)0x7FFFFFFFFFFFFFFULL;
 374 }
 375
 376 static void get_locks(void)
 377 {
 378         unsigned int i;
 379         struct flock fl;
 380
 381         if (lock_owner == getpid())
 382                 return;
 383
 384         if (lock_owner != 0) {
 385                 enum info_type type = RELEASE_LOCKS;
 386                 assert(control_fd != -1);
 387                 trace("Asking parent to release locks\n");
 388                 write_all(control_fd, &type, sizeof(type));
 389         }
 390
 391         fl.l_whence = SEEK_SET;
 392
 393         for (i = 0; i < lock_num; i++) {
 394                 fl.l_type = locks[i].type;
 395                 fl.l_start = locks[i].start;
 396                 if (locks[i].end == off_max())
 397                         fl.l_len = 0;
 398                 else
 399                         fl.l_len = locks[i].end - locks[i].start + 1;
 400
 401                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
 402                         abort();
 403         }
 404         trace("Acquired %u locks\n", lock_num);
 405         lock_owner = getpid();
 406 }
 407
 408
 409 static struct contents_saved *save_contents(const char *filename,
 410                                             int fd, size_t count, off_t off,
 411                                             const char *why)
 412 {
 413         struct contents_saved *s = malloc(sizeof(*s) + count);
 414         ssize_t ret;
 415
 416         s->off = off;
 417
 418         ret = pread(fd, s->contents, count, off);
 419         if (ret < 0) {
 420                 fwarn("failtest_write: failed to save old contents!");
 421                 s->count = 0;
 422         } else
 423                 s->count = ret;
 424
 425         /* Use lseek to get the size of file, but we have to restore
 426          * file offset */
 427         off = lseek(fd, 0, SEEK_CUR);
 428         s->old_len = lseek(fd, 0, SEEK_END);
 429         lseek(fd, off, SEEK_SET);
 430
 431         trace("Saving %p %s %zu@%llu after %s (filelength %llu) via fd %i\n",
 432               s, filename, s->count, (long long)s->off, why,
 433               (long long)s->old_len, fd);
 434         return s;
 435 }
 436
 437 static void restore_contents(struct failtest_call *opener,
 438                              struct contents_saved *s,
 439                              bool restore_offset,
 440                              const char *caller)
 441 {
 442         int fd;
 443
 444         /* The top parent doesn't need to restore. */
 445         if (control_fd == -1)
 446                 return;
 447
 448         /* Has the fd been closed? */
 449         if (opener->u.open.closed) {
 450                 /* Reopen, replace fd, close silently as we clean up. */
 451                 fd = open(opener->u.open.pathname, O_RDWR);
 452                 if (fd < 0) {
 453                         fwarn("failtest: could not reopen %s to clean up %s!",
 454                               opener->u.open.pathname, caller);
 455                         return;
 456                 }
 457                 /* Make it clearly distinguisable from a "normal" fd. */
 458                 fd = move_fd_to_high(fd);
 459                 trace("Reopening %s to restore it (was fd %i, now %i)\n",
 460                       opener->u.open.pathname, opener->u.open.ret, fd);
 461                 opener->u.open.ret = fd;
 462                 opener->u.open.closed = false;
 463         }
 464         fd = opener->u.open.ret;
 465
 466         trace("Restoring %p %s %zu@%llu after %s (filelength %llu) via fd %i\n",
 467               s, opener->u.open.pathname, s->count, (long long)s->off, caller,
 468               (long long)s->old_len, fd);
 469         if (pwrite(fd, s->contents, s->count, s->off) != s->count) {
 470                 fwarn("failtest: write failed cleaning up %s for %s!",
 471                       opener->u.open.pathname, caller);
 472         }
 473
 474         if (ftruncate(fd, s->old_len) != 0) {
 475                 fwarn("failtest_write: truncate failed cleaning up %s for %s!",
 476                       opener->u.open.pathname, caller);
 477         }
 478
 479         if (restore_offset) {
 480                 trace("Restoring offset of fd %i to %llu\n",
 481                       fd, (long long)s->off);
 482                 lseek(fd, s->off, SEEK_SET);
 483         }
 484 }
 485
 486 /* We save/restore most things on demand, but always do mmaped files. */
 487 static void save_mmapped_files(void)
 488 {
 489         struct failtest_call *i;
 490         trace("Saving mmapped files in child\n");
 491
 492         tlist_for_each_rev(&history, i, list) {
 493                 struct mmap_call *m = &i->u.mmap;
 494                 struct saved_mmapped_file *s;
 495
 496                 if (i->type != FAILTEST_MMAP)
 497                         continue;
 498
 499                 /* FIXME: We only handle mmapped files where fd is still open. */
 500                 if (m->opener->u.open.closed)
 501                         continue;
 502
 503                 s = malloc(sizeof *s);
 504                 s->s = save_contents(m->opener->u.open.pathname,
 505                                      m->fd, m->length, m->offset,
 506                                      "mmapped file before fork");
 507                 s->opener = m->opener;
 508                 s->next = saved_mmapped_files;
 509                 saved_mmapped_files = s;
 510         }
 511 }
 512
 513 static void free_mmapped_files(bool restore)
 514 {
 515         trace("%s mmapped files in child\n",
 516               restore ? "Restoring" : "Discarding");
 517         while (saved_mmapped_files) {
 518                 struct saved_mmapped_file *next = saved_mmapped_files->next;
 519                 if (restore)
 520                         restore_contents(saved_mmapped_files->opener,
 521                                          saved_mmapped_files->s, false,
 522                                          "saved mmap");
 523                 free(saved_mmapped_files->s);
 524                 free(saved_mmapped_files);
 525                 saved_mmapped_files = next;
 526         }
 527 }
 528
 529 /* Returns a FAILTEST_OPEN, FAILTEST_PIPE or NULL. */
 530 static struct failtest_call *opener_of(int fd)
 531 {
 532         struct failtest_call *i;
 533
 534         /* Don't get confused and match genuinely failed opens. */
 535         if (fd < 0)
 536                 return NULL;
 537
 538         /* Figure out the set of live fds. */
 539         tlist_for_each_rev(&history, i, list) {
 540                 if (i->fail)
 541                         continue;
 542                 switch (i->type) {
 543                 case FAILTEST_CLOSE:
 544                         if (i->u.close.fd == fd) {
 545                                 return NULL;
 546                         }
 547                         break;
 548                 case FAILTEST_OPEN:
 549                         if (i->u.open.ret == fd) {
 550                                 if (i->u.open.closed)
 551                                         return NULL;
 552                                 return i;
 553                         }
 554                         break;
 555                 case FAILTEST_PIPE:
 556                         if (i->u.pipe.fds[0] == fd || i->u.pipe.fds[1] == fd) {
 557                                 return i;
 558                         }
 559                         break;
 560                 default:
 561                         break;
 562                 }
 563         }
 564
 565         /* FIXME: socket, dup, etc are untracked! */
 566         return NULL;
 567 }
 568
 569 static void free_call(struct failtest_call *call)
 570 {
 571         /* We don't do this in cleanup: needed even for failed opens. */
 572         if (call->type == FAILTEST_OPEN)
 573                 free((char *)call->u.open.pathname);
 574         free(call->backtrace);
 575         tlist_del_from(&history, call, list);
 576         free(call);
 577 }
 578
 579 /* Free up memory, so valgrind doesn't report leaks. */
 580 static void free_everything(void)
 581 {
 582         struct failtest_call *i;
 583
 584         while ((i = tlist_top(&history, list)) != NULL)
 585                 free_call(i);
 586
 587         failtable_clear(&failtable);
 588 }
 589
 590 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
 591 {
 592         struct failtest_call *i;
 593         bool restore = true;
 594
 595         /* For children, we don't care if they "failed" the testing. */
 596         if (control_fd != -1)
 597                 status = 0;
 598         else
 599                 /* We don't restore contents for original parent. */
 600                 restore = false;
 601
 602         /* Cleanup everything, in reverse order. */
 603         tlist_for_each_rev(&history, i, list) {
 604                 /* Don't restore things our parent did. */
 605                 if (i == our_history_start)
 606                         restore = false;
 607
 608                 if (i->fail)
 609                         continue;
 610
 611                 if (i->cleanup)
 612                         i->cleanup(&i->u, restore);
 613
 614                 /* But their program shouldn't leak, even on failure. */
 615                 if (!forced_cleanup && i->can_leak) {
 616                         char *p = failpath_string();
 617                         printf("Leak at %s:%u: --failpath=%s\n",
 618                                i->file, i->line, p);
 619                         free(p);
 620                         status = 1;
 621                 }
 622         }
 623
 624         /* Put back mmaped files the way our parent (if any) expects. */
 625         free_mmapped_files(true);
 626
 627         free_everything();
 628         if (status == 0)
 629                 tell_parent(SUCCESS);
 630         else
 631                 tell_parent(FAILURE);
 632         exit(status);
 633 }
 634
 635 static bool following_path(void)
 636 {
 637         if (!failpath)
 638                 return false;
 639         /* + means continue after end, like normal. */
 640         if (*failpath == '+') {
 641                 failpath = NULL;
 642                 return false;
 643         }
 644         return true;
 645 }
 646
 647 static bool follow_path(struct failtest_call *call)
 648 {
 649         if (*failpath == '\0') {
 650                 /* Continue, but don't inject errors. */
 651                 return call->fail = false;
 652         }
 653
 654         if (tolower((unsigned char)*failpath) != info_to_arg[call->type])
 655                 errx(1, "Failpath expected '%s' got '%c'\n",
 656                      failpath, info_to_arg[call->type]);
 657         call->fail = cisupper(*(failpath++));
 658                         if (call->fail)
 659                                 call->can_leak = false;
 660         return call->fail;
 661 }
 662
 663 static bool should_fail(struct failtest_call *call)
 664 {
 665         int status;
 666         int control[2], output[2];
 667         enum info_type type = UNEXPECTED;
 668         char *out = NULL;
 669         size_t outlen = 0;
 670         struct failtest_call *dup;
 671
 672         if (call == &unrecorded_call)
 673                 return false;
 674
 675         if (following_path())
 676                 return follow_path(call);
 677
 678         /* Attach debugger if they asked for it. */
 679         if (debugpath) {
 680                 char *path;
 681
 682                 /* Pretend this last call matches whatever path wanted:
 683                  * keeps valgrind happy. */
 684                 call->fail = cisupper(debugpath[strlen(debugpath)-1]);
 685                 path = failpath_string();
 686
 687                 if (streq(path, debugpath)) {
 688                         char str[80];
 689
 690                         /* Don't timeout. */
 691                         signal(SIGUSR1, SIG_IGN);
 692                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
 693                                 getpid(), getpid());
 694                         if (system(str) == 0)
 695                                 sleep(5);
 696                 } else {
 697                         /* Ignore last character: could be upper or lower. */
 698                         path[strlen(path)-1] = '\0';
 699                         if (!strstarts(debugpath, path)) {
 700                                 fprintf(stderr,
 701                                         "--debugpath not followed: %s\n", path);
 702                                 debugpath = NULL;
 703                         }
 704                 }
 705                 free(path);
 706         }
 707
 708         /* Are we probing?  If so, we never fail twice. */
 709         if (probing) {
 710                 trace("Not failing %c due to FAIL_PROBE return\n",
 711                       info_to_arg[call->type]);
 712                 return call->fail = false;
 713         }
 714
 715         /* Don't fail more than once in the same place. */
 716         dup = failtable_get(&failtable, call);
 717         if (dup) {
 718                 trace("Not failing %c due to duplicate\n",
 719                       info_to_arg[call->type]);
 720                 return call->fail = false;
 721         }
 722
 723         if (failtest_hook) {
 724                 switch (failtest_hook(&history)) {
 725                 case FAIL_OK:
 726                         break;
 727                 case FAIL_PROBE:
 728                         probing = true;
 729                         break;
 730                 case FAIL_DONT_FAIL:
 731                         trace("Not failing %c due to failhook return\n",
 732                               info_to_arg[call->type]);
 733                         call->fail = false;
 734                         return false;
 735                 default:
 736                         abort();
 737                 }
 738         }
 739
 740         /* Add it to our table of calls. */
 741         failtable_add(&failtable, call);
 742
 743         /* We're going to fail in the child. */
 744         call->fail = true;
 745         if (pipe(control) != 0 || pipe(output) != 0)
 746                 err(1, "opening pipe");
 747
 748         /* Move out the way, to high fds. */
 749         control[0] = move_fd_to_high(control[0]);
 750         control[1] = move_fd_to_high(control[1]);
 751         output[0] = move_fd_to_high(output[0]);
 752         output[1] = move_fd_to_high(output[1]);
 753
 754         /* Prevent double-printing (in child and parent) */
 755         fflush(stdout);
 756         fflush(warnf);
 757         if (tracef)
 758                 fflush(tracef);
 759         child = fork();
 760         if (child == -1)
 761                 err(1, "forking failed");
 762
 763         if (child == 0) {
 764                 traceindent++;
 765                 if (tracef) {
 766                         struct timerel diff;
 767                         const char *p;
 768                         char *failpath;
 769                         struct failtest_call *c;
 770
 771                         c = tlist_tail(&history, list);
 772                         diff = time_between(time_now(), start);
 773                         failpath = failpath_string();
 774                         p = strrchr(c->file, '/');
 775                         if (p)
 776                                 p++;
 777                         else
 778                                 p = c->file;
 779                         trace("%u->%u (%u.%02u): %s (%s:%u)\n",
 780                               getppid(), getpid(),
 781                               (int)diff.ts.tv_sec, (int)diff.ts.tv_nsec / 10000000,
 782                               failpath, p, c->line);
 783                         free(failpath);
 784                 }
 785                 /* From here on, we have to clean up! */
 786                 our_history_start = tlist_tail(&history, list);
 787                 close(control[0]);
 788                 close(output[0]);
 789                 /* Don't swallow stderr if we're tracing. */
 790                 if (!tracef) {
 791                         dup2(output[1], STDOUT_FILENO);
 792                         dup2(output[1], STDERR_FILENO);
 793                         if (output[1] != STDOUT_FILENO
 794                             && output[1] != STDERR_FILENO)
 795                                 close(output[1]);
 796                 }
 797                 control_fd = move_fd_to_high(control[1]);
 798
 799                 /* Forget any of our parent's saved files. */
 800                 free_mmapped_files(false);
 801
 802                 /* Now, save any files we need to. */
 803                 save_mmapped_files();
 804
 805                 /* Failed calls can't leak. */
 806                 call->can_leak = false;
 807
 808                 return true;
 809         }
 810
 811         signal(SIGUSR1, hand_down);
 812
 813         close(control[1]);
 814         close(output[1]);
 815
 816         /* We grab output so we can display it; we grab writes so we
 817          * can compare. */
 818         do {
 819                 struct pollfd pfd[2];
 820                 int ret;
 821
 822                 pfd[0].fd = output[0];
 823                 pfd[0].events = POLLIN|POLLHUP;
 824                 pfd[1].fd = control[0];
 825                 pfd[1].events = POLLIN|POLLHUP;
 826
 827                 if (type == SUCCESS)
 828                         ret = poll(pfd, 1, failtest_timeout_ms);
 829                 else
 830                         ret = poll(pfd, 2, failtest_timeout_ms);
 831
 832                 if (ret == 0)
 833                         hand_down(SIGUSR1);
 834                 if (ret < 0) {
 835                         if (errno == EINTR)
 836                                 continue;
 837                         err(1, "Poll returned %i", ret);
 838                 }
 839
 840                 if (pfd[0].revents & POLLIN) {
 841                         ssize_t len;
 842
 843                         out = realloc(out, outlen + 8192);
 844                         len = read(output[0], out + outlen, 8192);
 845                         outlen += len;
 846                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 847                         if (read_all(control[0], &type, sizeof(type))) {
 848                                 if (type == WRITE) {
 849                                         if (!read_write_info(control[0]))
 850                                                 break;
 851                                 } else if (type == RELEASE_LOCKS) {
 852                                         release_locks();
 853                                         /* FIXME: Tell them we're done... */
 854                                 }
 855                         }
 856                 } else if (pfd[0].revents & POLLHUP) {
 857                         break;
 858                 }
 859         } while (type != FAILURE);
 860
 861         close(output[0]);
 862         close(control[0]);
 863         waitpid(child, &status, 0);
 864         if (!WIFEXITED(status)) {
 865                 if (WTERMSIG(status) == SIGUSR1)
 866                         child_fail(out, outlen, "Timed out");
 867                 else
 868                         child_fail(out, outlen, "Killed by signal %u: ",
 869                                    WTERMSIG(status));
 870         }
 871         /* Child printed failure already, just pass up exit code. */
 872         if (type == FAILURE) {
 873                 fprintf(stderr, "%.*s", (int)outlen, out);
 874                 tell_parent(type);
 875                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 876         }
 877         if (WEXITSTATUS(status) != 0)
 878                 child_fail(out, outlen, "Exited with status %i: ",
 879                            WEXITSTATUS(status));
 880
 881         free(out);
 882         signal(SIGUSR1, SIG_DFL);
 883
 884         /* Only child does probe. */
 885         probing = false;
 886
 887         /* We continue onwards without failing. */
 888         call->fail = false;
 889         return false;
 890 }
 891
 892 static void cleanup_calloc(struct calloc_call *call, bool restore)
 893 {
 894         trace("undoing calloc %p\n", call->ret);
 895         free(call->ret);
 896 }
 897
 898 void *failtest_calloc(size_t nmemb, size_t size,
 899                       const char *file, unsigned line)
 900 {
 901         struct failtest_call *p;
 902         struct calloc_call call;
 903         call.nmemb = nmemb;
 904         call.size = size;
 905         p = add_history(FAILTEST_CALLOC, true, file, line, &call);
 906
 907         if (should_fail(p)) {
 908                 p->u.calloc.ret = NULL;
 909                 p->error = ENOMEM;
 910         } else {
 911                 p->u.calloc.ret = calloc(nmemb, size);
 912                 set_cleanup(p, cleanup_calloc, struct calloc_call);
 913         }
 914         trace("calloc %zu x %zu %s:%u -> %p\n",
 915               nmemb, size, file, line, p->u.calloc.ret);
 916         errno = p->error;
 917         return p->u.calloc.ret;
 918 }
 919
 920 static void cleanup_malloc(struct malloc_call *call, bool restore)
 921 {
 922         trace("undoing malloc %p\n", call->ret);
 923         free(call->ret);
 924 }
 925
 926 void *failtest_malloc(size_t size, const char *file, unsigned line)
 927 {
 928         struct failtest_call *p;
 929         struct malloc_call call;
 930         call.size = size;
 931
 932         p = add_history(FAILTEST_MALLOC, true, file, line, &call);
 933         if (should_fail(p)) {
 934                 p->u.malloc.ret = NULL;
 935                 p->error = ENOMEM;
 936         } else {
 937                 p->u.malloc.ret = malloc(size);
 938                 set_cleanup(p, cleanup_malloc, struct malloc_call);
 939         }
 940         trace("malloc %zu %s:%u -> %p\n",
 941               size, file, line, p->u.malloc.ret);
 942         errno = p->error;
 943         return p->u.malloc.ret;
 944 }
 945
 946 static void cleanup_realloc(struct realloc_call *call, bool restore)
 947 {
 948         trace("undoing realloc %p\n", call->ret);
 949         free(call->ret);
 950 }
 951
 952 /* Walk back and find out if we got this ptr from a previous routine. */
 953 static void fixup_ptr_history(void *ptr, const char *why)
 954 {
 955         struct failtest_call *i;
 956
 957         /* Start at end of history, work back. */
 958         tlist_for_each_rev(&history, i, list) {
 959                 switch (i->type) {
 960                 case FAILTEST_REALLOC:
 961                         if (i->u.realloc.ret == ptr) {
 962                                 trace("found realloc %p %s:%u matching %s\n",
 963                                       ptr, i->file, i->line, why);
 964                                 i->cleanup = NULL;
 965                                 i->can_leak = false;
 966                                 return;
 967                         }
 968                         break;
 969                 case FAILTEST_MALLOC:
 970                         if (i->u.malloc.ret == ptr) {
 971                                 trace("found malloc %p %s:%u matching %s\n",
 972                                       ptr, i->file, i->line, why);
 973                                 i->cleanup = NULL;
 974                                 i->can_leak = false;
 975                                 return;
 976                         }
 977                         break;
 978                 case FAILTEST_CALLOC:
 979                         if (i->u.calloc.ret == ptr) {
 980                                 trace("found calloc %p %s:%u matching %s\n",
 981                                       ptr, i->file, i->line, why);
 982                                 i->cleanup = NULL;
 983                                 i->can_leak = false;
 984                                 return;
 985                         }
 986                         break;
 987                 default:
 988                         break;
 989                 }
 990         }
 991         trace("Did not find %p matching %s\n", ptr, why);
 992 }
 993
 994 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 995 {
 996         struct failtest_call *p;
 997         struct realloc_call call;
 998         call.size = size;
 999         p = add_history(FAILTEST_REALLOC, true, file, line, &call);
1000
1001         /* FIXME: Try one child moving allocation, one not. */
1002         if (should_fail(p)) {
1003                 p->u.realloc.ret = NULL;
1004                 p->error = ENOMEM;
1005         } else {
1006                 /* Don't catch this one in the history fixup... */
1007                 p->u.realloc.ret = NULL;
1008                 fixup_ptr_history(ptr, "realloc");
1009                 p->u.realloc.ret = realloc(ptr, size);
1010                 set_cleanup(p, cleanup_realloc, struct realloc_call);
1011         }
1012         trace("realloc %p %s:%u -> %p\n",
1013               ptr, file, line, p->u.realloc.ret);
1014         errno = p->error;
1015         return p->u.realloc.ret;
1016 }
1017
1018 /* FIXME: Record free, so we can terminate fixup_ptr_history correctly.
1019  * If there's an alloc we don't see, it could get confusing if it matches
1020  * a previous allocation we did see. */
1021 void failtest_free(void *ptr)
1022 {
1023         fixup_ptr_history(ptr, "free");
1024         trace("free %p\n", ptr);
1025         free(ptr);
1026 }
1027
1028
1029 static struct contents_saved *save_file(const char *pathname)
1030 {
1031         int fd;
1032         struct contents_saved *s;
1033
1034         fd = open(pathname, O_RDONLY);
1035         if (fd < 0)
1036                 return NULL;
1037
1038         s = save_contents(pathname, fd, lseek(fd, 0, SEEK_END), 0,
1039                           "open with O_TRUNC");
1040         close(fd);
1041         return s;
1042 }
1043
1044 /* Optimization: don't create a child for an open which *we know*
1045  * would fail anyway. */
1046 static bool open_would_fail(const char *pathname, int flags)
1047 {
1048         if ((flags & O_ACCMODE) == O_RDONLY)
1049                 return access(pathname, R_OK) != 0;
1050         if (!(flags & O_CREAT)) {
1051                 if ((flags & O_ACCMODE) == O_WRONLY)
1052                         return access(pathname, W_OK) != 0;
1053                 if ((flags & O_ACCMODE) == O_RDWR)
1054                         return access(pathname, W_OK) != 0
1055                                 || access(pathname, R_OK) != 0;
1056         }
1057         /* FIXME: We could check if it exists, for O_CREAT|O_EXCL */
1058         return false;
1059 }
1060
1061 static void cleanup_open(struct open_call *call, bool restore)
1062 {
1063         if (restore && call->saved)
1064                 restore_contents(container_of(call, struct failtest_call,
1065                                               u.open),
1066                                  call->saved, false, "open with O_TRUNC");
1067         if (!call->closed) {
1068                 trace("Cleaning up open %s by closing fd %i\n",
1069                       call->pathname, call->ret);
1070                 close(call->ret);
1071                 call->closed = true;
1072         }
1073         free(call->saved);
1074 }
1075
1076 int failtest_open(const char *pathname,
1077                   const char *file, unsigned line, ...)
1078 {
1079         struct failtest_call *p;
1080         struct open_call call;
1081         va_list ap;
1082
1083         call.pathname = strdup(pathname);
1084         va_start(ap, line);
1085         call.flags = va_arg(ap, int);
1086         call.always_save = false;
1087         call.closed = false;
1088         if (call.flags & O_CREAT) {
1089                 call.mode = va_arg(ap, int);
1090         }
1091         va_end(ap);
1092         p = add_history(FAILTEST_OPEN, true, file, line, &call);
1093         /* Avoid memory leak! */
1094         if (p == &unrecorded_call)
1095                 free((char *)call.pathname);
1096
1097         if (should_fail(p)) {
1098                 /* Don't bother inserting failures that would happen anyway. */
1099                 if (open_would_fail(pathname, call.flags)) {
1100                         trace("Open would have failed anyway: stopping\n");
1101                         failtest_cleanup(true, 0);
1102                 }
1103                 p->u.open.ret = -1;
1104                 /* FIXME: Play with error codes? */
1105                 p->error = EACCES;
1106         } else {
1107                 /* Save the old version if they're truncating it. */
1108                 if (call.flags & O_TRUNC)
1109                         p->u.open.saved = save_file(pathname);
1110                 else
1111                         p->u.open.saved = NULL;
1112                 p->u.open.ret = open(pathname, call.flags, call.mode);
1113                 if (p->u.open.ret == -1) {
1114                         p->u.open.closed = true;
1115                         p->can_leak = false;
1116                 } else {
1117                         set_cleanup(p, cleanup_open, struct open_call);
1118                 }
1119         }
1120         trace("open %s %s:%u -> %i (opener %p)\n",
1121               pathname, file, line, p->u.open.ret, &p->u.open);
1122         errno = p->error;
1123         return p->u.open.ret;
1124 }
1125
1126 static void cleanup_mmap(struct mmap_call *mmap, bool restore)
1127 {
1128         trace("cleaning up mmap @%p (opener %p)\n",
1129               mmap->ret, mmap->opener);
1130         if (restore)
1131                 restore_contents(mmap->opener, mmap->saved, false, "mmap");
1132         free(mmap->saved);
1133 }
1134
1135 void *failtest_mmap(void *addr, size_t length, int prot, int flags,
1136                     int fd, off_t offset, const char *file, unsigned line)
1137 {
1138         struct failtest_call *p;
1139         struct mmap_call call;
1140
1141         call.addr = addr;
1142         call.length = length;
1143         call.prot = prot;
1144         call.flags = flags;
1145         call.offset = offset;
1146         call.fd = fd;
1147         call.opener = opener_of(fd);
1148
1149         /* If we don't know what file it was, don't fail. */
1150         if (!call.opener) {
1151                 if (fd != -1) {
1152                         fwarnx("failtest_mmap: couldn't figure out source for"
1153                                " fd %i at %s:%u", fd, file, line);
1154                 }
1155                 addr = mmap(addr, length, prot, flags, fd, offset);
1156                 trace("mmap of fd %i -> %p (opener = NULL)\n", fd, addr);
1157                 return addr;
1158         }
1159
1160         p = add_history(FAILTEST_MMAP, false, file, line, &call);
1161         if (should_fail(p)) {
1162                 p->u.mmap.ret = MAP_FAILED;
1163                 p->error = ENOMEM;
1164         } else {
1165                 p->u.mmap.ret = mmap(addr, length, prot, flags, fd, offset);
1166                 /* Save contents if we're writing to a normal file */
1167                 if (p->u.mmap.ret != MAP_FAILED
1168                     && (prot & PROT_WRITE)
1169                     && call.opener->type == FAILTEST_OPEN) {
1170                         const char *fname = call.opener->u.open.pathname;
1171                         p->u.mmap.saved = save_contents(fname, fd, length,
1172                                                         offset, "being mmapped");
1173                         set_cleanup(p, cleanup_mmap, struct mmap_call);
1174                 }
1175         }
1176         trace("mmap of fd %i %s:%u -> %p (opener = %p)\n",
1177               fd, file, line, addr, call.opener);
1178         errno = p->error;
1179         return p->u.mmap.ret;
1180 }
1181
1182 /* Since OpenBSD can't handle adding args, we use this file and line.
1183  * This will make all mmaps look the same, reducing coverage. */
1184 void *failtest_mmap_noloc(void *addr, size_t length, int prot, int flags,
1185                           int fd, off_t offset)
1186 {
1187         return failtest_mmap(addr, length, prot, flags, fd, offset,
1188                              __FILE__, __LINE__);
1189 }
1190
1191 static void cleanup_pipe(struct pipe_call *call, bool restore)
1192 {
1193         trace("cleaning up pipe fd=%i%s,%i%s\n",
1194               call->fds[0], call->closed[0] ? "(already closed)" : "",
1195               call->fds[1], call->closed[1] ? "(already closed)" : "");
1196         if (!call->closed[0])
1197                 close(call->fds[0]);
1198         if (!call->closed[1])
1199                 close(call->fds[1]);
1200 }
1201
1202 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
1203 {
1204         struct failtest_call *p;
1205         struct pipe_call call;
1206
1207         p = add_history(FAILTEST_PIPE, true, file, line, &call);
1208         if (should_fail(p)) {
1209                 p->u.open.ret = -1;
1210                 /* FIXME: Play with error codes? */
1211                 p->error = EMFILE;
1212         } else {
1213                 p->u.pipe.ret = pipe(p->u.pipe.fds);
1214                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
1215                 set_cleanup(p, cleanup_pipe, struct pipe_call);
1216         }
1217
1218         trace("pipe %s:%u -> %i,%i\n", file, line,
1219               p->u.pipe.ret ? -1 : p->u.pipe.fds[0],
1220               p->u.pipe.ret ? -1 : p->u.pipe.fds[1]);
1221
1222         /* This causes valgrind to notice if they use pipefd[] after failure */
1223         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
1224         errno = p->error;
1225         return p->u.pipe.ret;
1226 }
1227
1228 static void cleanup_read(struct read_call *call, bool restore)
1229 {
1230         if (restore) {
1231                 trace("cleaning up read on fd %i: seeking to %llu\n",
1232                       call->fd, (long long)call->off);
1233
1234                 /* Read (not readv!) moves file offset! */
1235                 if (lseek(call->fd, call->off, SEEK_SET) != call->off) {
1236                         fwarn("Restoring lseek pointer failed (read)");
1237                 }
1238         }
1239 }
1240
1241 static ssize_t failtest_add_read(int fd, void *buf, size_t count, off_t off,
1242                                  bool is_pread, const char *file, unsigned line)
1243 {
1244         struct failtest_call *p;
1245         struct read_call call;
1246         call.fd = fd;
1247         call.buf = buf;
1248         call.count = count;
1249         call.off = off;
1250         p = add_history(FAILTEST_READ, false, file, line, &call);
1251
1252         /* FIXME: Try partial read returns. */
1253         if (should_fail(p)) {
1254                 p->u.read.ret = -1;
1255                 p->error = EIO;
1256         } else {
1257                 if (is_pread)
1258                         p->u.read.ret = pread(fd, buf, count, off);
1259                 else {
1260                         p->u.read.ret = read(fd, buf, count);
1261                         if (p->u.read.ret != -1)
1262                                 set_cleanup(p, cleanup_read, struct read_call);
1263                 }
1264         }
1265         trace("%sread %s:%u fd %i %zu@%llu -> %zi\n",
1266               is_pread ? "p" : "", file, line, fd, count, (long long)off,
1267               p->u.read.ret);
1268         errno = p->error;
1269         return p->u.read.ret;
1270 }
1271
1272 static void cleanup_write(struct write_call *write, bool restore)
1273 {
1274         trace("cleaning up write on %s\n", write->opener->u.open.pathname);
1275         if (restore)
1276                 restore_contents(write->opener, write->saved, !write->is_pwrite,
1277                                  "write");
1278         free(write->saved);
1279 }
1280
1281 static ssize_t failtest_add_write(int fd, const void *buf,
1282                                   size_t count, off_t off,
1283                                   bool is_pwrite,
1284                                   const char *file, unsigned line)
1285 {
1286         struct failtest_call *p;
1287         struct write_call call;
1288
1289         call.fd = fd;
1290         call.buf = buf;
1291         call.count = count;
1292         call.off = off;
1293         call.is_pwrite = is_pwrite;
1294         call.opener = opener_of(fd);
1295         p = add_history(FAILTEST_WRITE, false, file, line, &call);
1296
1297         /* If we're a child, we need to make sure we write the same thing
1298          * to non-files as the parent does, so tell it. */
1299         if (control_fd != -1 && off == (off_t)-1) {
1300                 enum info_type type = WRITE;
1301
1302                 write_all(control_fd, &type, sizeof(type));
1303                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
1304                 write_all(control_fd, buf, count);
1305         }
1306
1307         /* FIXME: Try partial write returns. */
1308         if (should_fail(p)) {
1309                 p->u.write.ret = -1;
1310                 p->error = EIO;
1311         } else {
1312                 bool is_file;
1313                 assert(call.opener == p->u.write.opener);
1314
1315                 if (p->u.write.opener) {
1316                         is_file = (p->u.write.opener->type == FAILTEST_OPEN);
1317                 } else {
1318                         /* We can't unwind it, so at least check same
1319                          * in parent and child. */
1320                         is_file = false;
1321                 }
1322
1323                 /* FIXME: We assume same write order in parent and child */
1324                 if (!is_file && child_writes_num != 0) {
1325                         if (child_writes[0].fd != fd)
1326                                 errx(1, "Child wrote to fd %u, not %u?",
1327                                      child_writes[0].fd, fd);
1328                         if (child_writes[0].off != p->u.write.off)
1329                                 errx(1, "Child wrote to offset %zu, not %zu?",
1330                                      (size_t)child_writes[0].off,
1331                                      (size_t)p->u.write.off);
1332                         if (child_writes[0].count != count)
1333                                 errx(1, "Child wrote length %zu, not %zu?",
1334                                      child_writes[0].count, count);
1335                         if (memcmp(child_writes[0].buf, buf, count)) {
1336                                 child_fail(NULL, 0,
1337                                            "Child wrote differently to"
1338                                            " fd %u than we did!\n", fd);
1339                         }
1340                         free((char *)child_writes[0].buf);
1341                         child_writes_num--;
1342                         memmove(&child_writes[0], &child_writes[1],
1343                                 sizeof(child_writes[0]) * child_writes_num);
1344
1345                         /* Child wrote it already. */
1346                         trace("write %s:%i on fd %i already done by child\n",
1347                               file, line, fd);
1348                         p->u.write.ret = count;
1349                         errno = p->error;
1350                         return p->u.write.ret;
1351                 }
1352
1353                 if (is_file) {
1354                         p->u.write.saved = save_contents(call.opener->u.open.pathname,
1355                                                          fd, count, off,
1356                                                          "being overwritten");
1357                         set_cleanup(p, cleanup_write, struct write_call);
1358                 }
1359
1360                 /* Though off is current seek ptr for write case, we need to
1361                  * move it.  write() does that for us. */
1362                 if (p->u.write.is_pwrite)
1363                         p->u.write.ret = pwrite(fd, buf, count, off);
1364                 else
1365                         p->u.write.ret = write(fd, buf, count);
1366         }
1367         trace("%swrite %s:%i %zu@%llu on fd %i -> %zi\n",
1368               p->u.write.is_pwrite ? "p" : "",
1369               file, line, count, (long long)off, fd, p->u.write.ret);
1370         errno = p->error;
1371         return p->u.write.ret;
1372 }
1373
1374 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t offset,
1375                         const char *file, unsigned line)
1376 {
1377         return failtest_add_write(fd, buf, count, offset, true, file, line);
1378 }
1379
1380 ssize_t failtest_write(int fd, const void *buf, size_t count,
1381                        const char *file, unsigned line)
1382 {
1383         return failtest_add_write(fd, buf, count, lseek(fd, 0, SEEK_CUR), false,
1384                                   file, line);
1385 }
1386
1387 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
1388                        const char *file, unsigned line)
1389 {
1390         return failtest_add_read(fd, buf, count, off, true, file, line);
1391 }
1392
1393 ssize_t failtest_read(int fd, void *buf, size_t count,
1394                       const char *file, unsigned line)
1395 {
1396         return failtest_add_read(fd, buf, count, lseek(fd, 0, SEEK_CUR), false,
1397                                  file, line);
1398 }
1399
1400 static struct lock_info *WARN_UNUSED_RESULT
1401 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
1402 {
1403         unsigned int i;
1404         struct lock_info *l;
1405
1406         for (i = 0; i < lock_num; i++) {
1407                 l = &locks[i];
1408
1409                 if (l->fd != fd)
1410                         continue;
1411                 /* Four cases we care about:
1412                  * Start overlap:
1413                  *      l =    |      |
1414                  *      new = |   |
1415                  * Mid overlap:
1416                  *      l =    |      |
1417                  *      new =    |  |
1418                  * End overlap:
1419                  *      l =    |      |
1420                  *      new =      |    |
1421                  * Total overlap:
1422                  *      l =    |      |
1423                  *      new = |         |
1424                  */
1425                 if (start > l->start && end < l->end) {
1426                         /* Mid overlap: trim entry, add new one. */
1427                         off_t new_start, new_end;
1428                         new_start = end + 1;
1429                         new_end = l->end;
1430                         trace("splitting lock on fd %i from %llu-%llu"
1431                               " to %llu-%llu\n",
1432                               fd, (long long)l->start, (long long)l->end,
1433                               (long long)l->start, (long long)start - 1);
1434                         l->end = start - 1;
1435                         locks = add_lock(locks,
1436                                          fd, new_start, new_end, l->type);
1437                         l = &locks[i];
1438                 } else if (start <= l->start && end >= l->end) {
1439                         /* Total overlap: eliminate entry. */
1440                         trace("erasing lock on fd %i %llu-%llu\n",
1441                               fd, (long long)l->start, (long long)l->end);
1442                         l->end = 0;
1443                         l->start = 1;
1444                 } else if (end >= l->start && end < l->end) {
1445                         trace("trimming lock on fd %i from %llu-%llu"
1446                               " to %llu-%llu\n",
1447                               fd, (long long)l->start, (long long)l->end,
1448                               (long long)end + 1, (long long)l->end);
1449                         /* Start overlap: trim entry. */
1450                         l->start = end + 1;
1451                 } else if (start > l->start && start <= l->end) {
1452                         trace("trimming lock on fd %i from %llu-%llu"
1453                               " to %llu-%llu\n",
1454                               fd, (long long)l->start, (long long)l->end,
1455                               (long long)l->start, (long long)start - 1);
1456                         /* End overlap: trim entry. */
1457                         l->end = start-1;
1458                 }
1459                 /* Nothing left?  Remove it. */
1460                 if (l->end < l->start) {
1461                         trace("forgetting lock on fd %i\n", fd);
1462                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
1463                         i--;
1464                 }
1465         }
1466
1467         if (type != F_UNLCK) {
1468                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
1469                 l = &locks[lock_num++];
1470                 l->fd = fd;
1471                 l->start = start;
1472                 l->end = end;
1473                 l->type = type;
1474                 trace("new lock on fd %i %llu-%llu\n",
1475                       fd, (long long)l->start, (long long)l->end);
1476         }
1477         return locks;
1478 }
1479
1480 /* We trap this so we can record it: we don't fail it. */
1481 int failtest_close(int fd, const char *file, unsigned line)
1482 {
1483         struct close_call call;
1484         struct failtest_call *p, *opener;
1485
1486         /* Do this before we add ourselves to history! */
1487         opener = opener_of(fd);
1488
1489         call.fd = fd;
1490         p = add_history(FAILTEST_CLOSE, false, file, line, &call);
1491         p->fail = false;
1492
1493         /* Consume close from failpath (shouldn't tell us to fail). */
1494         if (following_path()) {
1495                 if (follow_path(p))
1496                         abort();
1497         }
1498
1499         trace("close on fd %i\n", fd);
1500         if (fd < 0)
1501                 return close(fd);
1502
1503         /* Mark opener as not leaking, remove its cleanup function. */
1504         if (opener) {
1505                 trace("close on fd %i found opener %p\n", fd, opener);
1506                 if (opener->type == FAILTEST_PIPE) {
1507                         /* From a pipe? */
1508                         if (opener->u.pipe.fds[0] == fd) {
1509                                 assert(!opener->u.pipe.closed[0]);
1510                                 opener->u.pipe.closed[0] = true;
1511                         } else if (opener->u.pipe.fds[1] == fd) {
1512                                 assert(!opener->u.pipe.closed[1]);
1513                                 opener->u.pipe.closed[1] = true;
1514                         } else
1515                                 abort();
1516                         opener->can_leak = (!opener->u.pipe.closed[0]
1517                                             || !opener->u.pipe.closed[1]);
1518                 } else if (opener->type == FAILTEST_OPEN) {
1519                         opener->u.open.closed = true;
1520                         opener->can_leak = false;
1521                 } else
1522                         abort();
1523         }
1524
1525         /* Restore offset now, in case parent shared (can't do after close!). */
1526         if (control_fd != -1) {
1527                 struct failtest_call *i;
1528
1529                 tlist_for_each_rev(&history, i, list) {
1530                         if (i == our_history_start)
1531                                 break;
1532                         if (i == opener)
1533                                 break;
1534                         if (i->type == FAILTEST_LSEEK && i->u.lseek.fd == fd) {
1535                                 trace("close on fd %i undoes lseek\n", fd);
1536                                 /* This seeks back. */
1537                                 i->cleanup(&i->u, true);
1538                                 i->cleanup = NULL;
1539                         } else if (i->type == FAILTEST_WRITE
1540                                    && i->u.write.fd == fd
1541                                    && !i->u.write.is_pwrite) {
1542                                 trace("close on fd %i undoes write"
1543                                       " offset change\n", fd);
1544                                 /* Write (not pwrite!) moves file offset! */
1545                                 if (lseek(fd, i->u.write.off, SEEK_SET)
1546                                     != i->u.write.off) {
1547                                         fwarn("Restoring lseek pointer failed (write)");
1548                                 }
1549                         } else if (i->type == FAILTEST_READ
1550                                    && i->u.read.fd == fd) {
1551                                 /* preads don't *have* cleanups */
1552                                 if (i->cleanup) {
1553                                         trace("close on fd %i undoes read"
1554                                               " offset change\n", fd);
1555                                         /* This seeks back. */
1556                                         i->cleanup(&i->u, true);
1557                                         i->cleanup = NULL;
1558                                 }
1559                         }
1560                 }
1561         }
1562
1563         /* Close unlocks everything. */
1564         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
1565         return close(fd);
1566 }
1567
1568 /* Zero length means "to end of file" */
1569 static off_t end_of(off_t start, off_t len)
1570 {
1571         if (len == 0)
1572                 return off_max();
1573         return start + len - 1;
1574 }
1575
1576 /* FIXME: This only handles locks, really. */
1577 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
1578 {
1579         struct failtest_call *p;
1580         struct fcntl_call call;
1581         va_list ap;
1582
1583         call.fd = fd;
1584         call.cmd = cmd;
1585
1586         /* Argument extraction. */
1587         switch (cmd) {
1588         case F_SETFL:
1589         case F_SETFD:
1590                 va_start(ap, cmd);
1591                 call.arg.l = va_arg(ap, long);
1592                 va_end(ap);
1593                 trace("fcntl on fd %i F_SETFL/F_SETFD\n", fd);
1594                 return fcntl(fd, cmd, call.arg.l);
1595         case F_GETFD:
1596         case F_GETFL:
1597                 trace("fcntl on fd %i F_GETFL/F_GETFD\n", fd);
1598                 return fcntl(fd, cmd);
1599         case F_GETLK:
1600                 trace("fcntl on fd %i F_GETLK\n", fd);
1601                 get_locks();
1602                 va_start(ap, cmd);
1603                 call.arg.fl = *va_arg(ap, struct flock *);
1604                 va_end(ap);
1605                 return fcntl(fd, cmd, &call.arg.fl);
1606         case F_SETLK:
1607         case F_SETLKW:
1608                 trace("fcntl on fd %i F_SETLK%s\n",
1609                       fd, cmd == F_SETLKW ? "W" : "");
1610                 va_start(ap, cmd);
1611                 call.arg.fl = *va_arg(ap, struct flock *);
1612                 va_end(ap);
1613                 break;
1614         default:
1615                 /* This means you need to implement it here. */
1616                 err(1, "failtest: unknown fcntl %u", cmd);
1617         }
1618
1619         p = add_history(FAILTEST_FCNTL, false, file, line, &call);
1620
1621         if (should_fail(p)) {
1622                 p->u.fcntl.ret = -1;
1623                 if (p->u.fcntl.cmd == F_SETLK)
1624                         p->error = EAGAIN;
1625                 else
1626                         p->error = EDEADLK;
1627         } else {
1628                 get_locks();
1629                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1630                                        &p->u.fcntl.arg.fl);
1631                 if (p->u.fcntl.ret == -1)
1632                         p->error = errno;
1633                 else {
1634                         /* We don't handle anything else yet. */
1635                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1636                         locks = add_lock(locks,
1637                                          p->u.fcntl.fd,
1638                                          p->u.fcntl.arg.fl.l_start,
1639                                          end_of(p->u.fcntl.arg.fl.l_start,
1640                                                 p->u.fcntl.arg.fl.l_len),
1641                                          p->u.fcntl.arg.fl.l_type);
1642                 }
1643         }
1644         trace("fcntl on fd %i -> %i\n", fd, p->u.fcntl.ret);
1645         errno = p->error;
1646         return p->u.fcntl.ret;
1647 }
1648
1649 static void cleanup_lseek(struct lseek_call *call, bool restore)
1650 {
1651         if (restore) {
1652                 trace("cleaning up lseek on fd %i -> %llu\n",
1653                       call->fd, (long long)call->old_off);
1654                 if (lseek(call->fd, call->old_off, SEEK_SET) != call->old_off)
1655                         fwarn("Restoring lseek pointer failed");
1656         }
1657 }
1658
1659 /* We trap this so we can undo it: we don't fail it. */
1660 off_t failtest_lseek(int fd, off_t offset, int whence, const char *file,
1661                      unsigned int line)
1662 {
1663         struct failtest_call *p;
1664         struct lseek_call call;
1665         call.fd = fd;
1666         call.offset = offset;
1667         call.whence = whence;
1668         call.old_off = lseek(fd, 0, SEEK_CUR);
1669
1670         p = add_history(FAILTEST_LSEEK, false, file, line, &call);
1671         p->fail = false;
1672
1673         /* Consume lseek from failpath. */
1674         if (failpath)
1675                 if (should_fail(p))
1676                         abort();
1677
1678         p->u.lseek.ret = lseek(fd, offset, whence);
1679
1680         if (p->u.lseek.ret != (off_t)-1)
1681                 set_cleanup(p, cleanup_lseek, struct lseek_call);
1682
1683         trace("lseek %s:%u on fd %i from %llu to %llu%s\n",
1684               file, line, fd, (long long)call.old_off, (long long)offset,
1685               whence == SEEK_CUR ? " (from current off)" :
1686               whence == SEEK_END ? " (from end)" :
1687               whence == SEEK_SET ? "" : " (invalid whence)");
1688         return p->u.lseek.ret;
1689 }
1690
1691
1692 pid_t failtest_getpid(const char *file, unsigned line)
1693 {
1694         /* You must call failtest_init first! */
1695         assert(orig_pid);
1696         return orig_pid;
1697 }
1698
1699 void failtest_init(int argc, char *argv[])
1700 {
1701         unsigned int i;
1702
1703         orig_pid = getpid();
1704
1705         warnf = fdopen(move_fd_to_high(dup(STDERR_FILENO)), "w");
1706         for (i = 1; i < argc; i++) {
1707                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1708                         failpath = argv[i] + strlen("--failpath=");
1709                 } else if (strcmp(argv[i], "--trace") == 0) {
1710                         tracef = warnf;
1711                         failtest_timeout_ms = -1;
1712                 } else if (!strncmp(argv[i], "--debugpath=",
1713                                     strlen("--debugpath="))) {
1714                         debugpath = argv[i] + strlen("--debugpath=");
1715                 }
1716         }
1717         failtable_init(&failtable);
1718         start = time_now();
1719 }
1720
1721 bool failtest_has_failed(void)
1722 {
1723         return control_fd != -1;
1724 }
1725
1726 void failtest_exit(int status)
1727 {
1728         trace("failtest_exit with status %i\n", status);
1729         if (failtest_exit_check) {
1730                 if (!failtest_exit_check(&history))
1731                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1732         }
1733
1734         failtest_cleanup(false, status);
1735 }