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 <err.h>
   9 #include <unistd.h>
  10 #include <poll.h>
  11 #include <errno.h>
  12 #include <sys/types.h>
  13 #include <sys/wait.h>
  14 #include <sys/stat.h>
  15 #include <sys/time.h>
  16 #include <sys/mman.h>
  17 #include <signal.h>
  18 #include <assert.h>
  19 #include <ccan/time/time.h>
  20 #include <ccan/read_write_all/read_write_all.h>
  21 #include <ccan/failtest/failtest_proto.h>
  22 #include <ccan/build_assert/build_assert.h>
  23 #include <ccan/hash/hash.h>
  24 #include <ccan/htable/htable_type.h>
  25 #include <ccan/str/str.h>
  26 #include <ccan/compiler/compiler.h>
  27
  28 enum failtest_result (*failtest_hook)(struct tlist_calls *);
  29
  30 static int tracefd = -1;
  31 static int traceindent = 0;
  32 static int warnfd;
  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 timeval 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
 199         for (i = FD_SETSIZE - 1; i >= 0; i--) {
 200                 if (fcntl(i, F_GETFL) == -1 && errno == EBADF) {
 201                         if (dup2(fd, i) == -1)
 202                                 err(1, "Failed to dup fd %i to %i", fd, i);
 203                         close(fd);
 204                         return i;
 205                 }
 206         }
 207         /* Nothing?  Really?  Er... ok? */
 208         return fd;
 209 }
 210
 211 static bool read_write_info(int fd)
 212 {
 213         struct write_call *w;
 214         char *buf;
 215
 216         /* We don't need all of this, but it's simple. */
 217         child_writes = realloc(child_writes,
 218                                (child_writes_num+1) * sizeof(child_writes[0]));
 219         w = &child_writes[child_writes_num];
 220         if (!read_all(fd, w, sizeof(*w)))
 221                 return false;
 222
 223         w->buf = buf = malloc(w->count);
 224         if (!read_all(fd, buf, w->count))
 225                 return false;
 226
 227         child_writes_num++;
 228         return true;
 229 }
 230
 231 static char *failpath_string(void)
 232 {
 233         struct failtest_call *i;
 234         char *ret = strdup("");
 235         unsigned len = 0;
 236
 237         /* Inefficient, but who cares? */
 238         tlist_for_each(&history, i, list) {
 239                 ret = realloc(ret, len + 2);
 240                 ret[len] = info_to_arg[i->type];
 241                 if (i->fail)
 242                         ret[len] = toupper(ret[len]);
 243                 ret[++len] = '\0';
 244         }
 245         return ret;
 246 }
 247
 248 static void warn_via_fd(int e, const char *fmt, va_list ap)
 249 {
 250         char *p = failpath_string();
 251
 252         vdprintf(warnfd, fmt, ap);
 253         if (e != -1)
 254                 dprintf(warnfd, ": %s", strerror(e));
 255         dprintf(warnfd, " [%s]\n", p);
 256         free(p);
 257 }
 258
 259 static void fwarn(const char *fmt, ...)
 260 {
 261         va_list ap;
 262         int e = errno;
 263
 264         va_start(ap, fmt);
 265         warn_via_fd(e, fmt, ap);
 266         va_end(ap);
 267 }
 268
 269
 270 static void fwarnx(const char *fmt, ...)
 271 {
 272         va_list ap;
 273
 274         va_start(ap, fmt);
 275         warn_via_fd(-1, fmt, ap);
 276         va_end(ap);
 277 }
 278
 279 static void tell_parent(enum info_type type)
 280 {
 281         if (control_fd != -1)
 282                 write_all(control_fd, &type, sizeof(type));
 283 }
 284
 285 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 286 {
 287         va_list ap;
 288         char *path = failpath_string();
 289
 290         va_start(ap, fmt);
 291         vfprintf(stderr, fmt, ap);
 292         va_end(ap);
 293
 294         fprintf(stderr, "%.*s", (int)outlen, out);
 295         printf("To reproduce: --failpath=%s\n", path);
 296         free(path);
 297         tell_parent(FAILURE);
 298         exit(1);
 299 }
 300
 301 static void PRINTF_FMT(1, 2) trace(const char *fmt, ...)
 302 {
 303         va_list ap;
 304         unsigned int i;
 305
 306         if (tracefd == -1)
 307                 return;
 308
 309         for (i = 0; i < traceindent; i++)
 310                 dprintf(tracefd, "  ");
 311
 312         dprintf(tracefd, "%u: ", getpid());
 313         va_start(ap, fmt);
 314         vdprintf(tracefd, fmt, ap);
 315         va_end(ap);
 316 }
 317
 318 static pid_t child;
 319
 320 static void hand_down(int signum)
 321 {
 322         kill(child, signum);
 323 }
 324
 325 static void release_locks(void)
 326 {
 327         /* Locks were never acquired/reacquired? */
 328         if (lock_owner == 0)
 329                 return;
 330
 331         /* We own them?  Release them all. */
 332         if (lock_owner == getpid()) {
 333                 unsigned int i;
 334                 struct flock fl;
 335                 fl.l_type = F_UNLCK;
 336                 fl.l_whence = SEEK_SET;
 337                 fl.l_start = 0;
 338                 fl.l_len = 0;
 339
 340                 trace("Releasing %u locks\n", lock_num);
 341                 for (i = 0; i < lock_num; i++)
 342                         fcntl(locks[i].fd, F_SETLK, &fl);
 343         } else {
 344                 /* Our parent must have them; pass request up. */
 345                 enum info_type type = RELEASE_LOCKS;
 346                 assert(control_fd != -1);
 347                 write_all(control_fd, &type, sizeof(type));
 348         }
 349         lock_owner = 0;
 350 }
 351
 352 /* off_t is a signed type.  Getting its max is non-trivial. */
 353 static off_t off_max(void)
 354 {
 355         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
 356         if (sizeof(off_t) == 4)
 357                 return (off_t)0x7FFFFFF;
 358         else
 359                 return (off_t)0x7FFFFFFFFFFFFFFULL;
 360 }
 361
 362 static void get_locks(void)
 363 {
 364         unsigned int i;
 365         struct flock fl;
 366
 367         if (lock_owner == getpid())
 368                 return;
 369
 370         if (lock_owner != 0) {
 371                 enum info_type type = RELEASE_LOCKS;
 372                 assert(control_fd != -1);
 373                 trace("Asking parent to release locks\n");
 374                 write_all(control_fd, &type, sizeof(type));
 375         }
 376
 377         fl.l_whence = SEEK_SET;
 378
 379         for (i = 0; i < lock_num; i++) {
 380                 fl.l_type = locks[i].type;
 381                 fl.l_start = locks[i].start;
 382                 if (locks[i].end == off_max())
 383                         fl.l_len = 0;
 384                 else
 385                         fl.l_len = locks[i].end - locks[i].start + 1;
 386
 387                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
 388                         abort();
 389         }
 390         trace("Acquired %u locks\n", lock_num);
 391         lock_owner = getpid();
 392 }
 393
 394
 395 static struct contents_saved *save_contents(const char *filename,
 396                                             int fd, size_t count, off_t off,
 397                                             const char *why)
 398 {
 399         struct contents_saved *s = malloc(sizeof(*s) + count);
 400         ssize_t ret;
 401
 402         s->off = off;
 403
 404         ret = pread(fd, s->contents, count, off);
 405         if (ret < 0) {
 406                 fwarn("failtest_write: failed to save old contents!");
 407                 s->count = 0;
 408         } else
 409                 s->count = ret;
 410
 411         /* Use lseek to get the size of file, but we have to restore
 412          * file offset */
 413         off = lseek(fd, 0, SEEK_CUR);
 414         s->old_len = lseek(fd, 0, SEEK_END);
 415         lseek(fd, off, SEEK_SET);
 416
 417         trace("Saving %p %s %zu@%llu after %s (filelength %llu) via fd %i\n",
 418               s, filename, s->count, (long long)s->off, why,
 419               (long long)s->old_len, fd);
 420         return s;
 421 }
 422
 423 static void restore_contents(struct failtest_call *opener,
 424                              struct contents_saved *s,
 425                              bool restore_offset,
 426                              const char *caller)
 427 {
 428         int fd;
 429
 430         /* The top parent doesn't need to restore. */
 431         if (control_fd == -1)
 432                 return;
 433
 434         /* Has the fd been closed? */
 435         if (opener->u.open.closed) {
 436                 /* Reopen, replace fd, close silently as we clean up. */
 437                 fd = open(opener->u.open.pathname, O_RDWR);
 438                 if (fd < 0) {
 439                         fwarn("failtest: could not reopen %s to clean up %s!",
 440                               opener->u.open.pathname, caller);
 441                         return;
 442                 }
 443                 /* Make it clearly distinguisable from a "normal" fd. */
 444                 fd = move_fd_to_high(fd);
 445                 trace("Reopening %s to restore it (was fd %i, now %i)\n",
 446                       opener->u.open.pathname, opener->u.open.ret, fd);
 447                 opener->u.open.ret = fd;
 448                 opener->u.open.closed = false;
 449         }
 450         fd = opener->u.open.ret;
 451
 452         trace("Restoring %p %s %zu@%llu after %s (filelength %llu) via fd %i\n",
 453               s, opener->u.open.pathname, s->count, (long long)s->off, caller,
 454               (long long)s->old_len, fd);
 455         if (pwrite(fd, s->contents, s->count, s->off) != s->count) {
 456                 fwarn("failtest: write failed cleaning up %s for %s!",
 457                       opener->u.open.pathname, caller);
 458         }
 459
 460         if (ftruncate(fd, s->old_len) != 0) {
 461                 fwarn("failtest_write: truncate failed cleaning up %s for %s!",
 462                       opener->u.open.pathname, caller);
 463         }
 464
 465         if (restore_offset) {
 466                 trace("Restoring offset of fd %i to %llu\n",
 467                       fd, (long long)s->off);
 468                 lseek(fd, s->off, SEEK_SET);
 469         }
 470 }
 471
 472 /* We save/restore most things on demand, but always do mmaped files. */
 473 static void save_mmapped_files(void)
 474 {
 475         struct failtest_call *i;
 476         trace("Saving mmapped files in child\n");
 477
 478         tlist_for_each_rev(&history, i, list) {
 479                 struct mmap_call *m = &i->u.mmap;
 480                 struct saved_mmapped_file *s;
 481
 482                 if (i->type != FAILTEST_MMAP)
 483                         continue;
 484
 485                 /* FIXME: We only handle mmapped files where fd is still open. */
 486                 if (m->opener->u.open.closed)
 487                         continue;
 488
 489                 s = malloc(sizeof *s);
 490                 s->s = save_contents(m->opener->u.open.pathname,
 491                                      m->fd, m->length, m->offset,
 492                                      "mmapped file before fork");
 493                 s->opener = m->opener;
 494                 s->next = saved_mmapped_files;
 495                 saved_mmapped_files = s;
 496         }
 497 }
 498
 499 static void free_mmapped_files(bool restore)
 500 {
 501         trace("%s mmapped files in child\n",
 502               restore ? "Restoring" : "Discarding");
 503         while (saved_mmapped_files) {
 504                 struct saved_mmapped_file *next = saved_mmapped_files->next;
 505                 if (restore)
 506                         restore_contents(saved_mmapped_files->opener,
 507                                          saved_mmapped_files->s, false,
 508                                          "saved mmap");
 509                 free(saved_mmapped_files->s);
 510                 free(saved_mmapped_files);
 511                 saved_mmapped_files = next;
 512         }
 513 }
 514
 515 /* Returns a FAILTEST_OPEN, FAILTEST_PIPE or NULL. */
 516 static struct failtest_call *opener_of(int fd)
 517 {
 518         struct failtest_call *i;
 519
 520         /* Don't get confused and match genuinely failed opens. */
 521         if (fd < 0)
 522                 return NULL;
 523
 524         /* Figure out the set of live fds. */
 525         tlist_for_each_rev(&history, i, list) {
 526                 if (i->fail)
 527                         continue;
 528                 switch (i->type) {
 529                 case FAILTEST_CLOSE:
 530                         if (i->u.close.fd == fd) {
 531                                 return NULL;
 532                         }
 533                         break;
 534                 case FAILTEST_OPEN:
 535                         if (i->u.open.ret == fd) {
 536                                 if (i->u.open.closed)
 537                                         return NULL;
 538                                 return i;
 539                         }
 540                         break;
 541                 case FAILTEST_PIPE:
 542                         if (i->u.pipe.fds[0] == fd || i->u.pipe.fds[1] == fd) {
 543                                 return i;
 544                         }
 545                         break;
 546                 default:
 547                         break;
 548                 }
 549         }
 550
 551         /* FIXME: socket, dup, etc are untracked! */
 552         return NULL;
 553 }
 554
 555 static void free_call(struct failtest_call *call)
 556 {
 557         /* We don't do this in cleanup: needed even for failed opens. */
 558         if (call->type == FAILTEST_OPEN)
 559                 free((char *)call->u.open.pathname);
 560         free(call->backtrace);
 561         tlist_del_from(&history, call, list);
 562         free(call);
 563 }
 564
 565 /* Free up memory, so valgrind doesn't report leaks. */
 566 static void free_everything(void)
 567 {
 568         struct failtest_call *i;
 569
 570         while ((i = tlist_top(&history, struct failtest_call, list)) != NULL)
 571                 free_call(i);
 572
 573         failtable_clear(&failtable);
 574 }
 575
 576 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
 577 {
 578         struct failtest_call *i;
 579         bool restore = true;
 580
 581         /* For children, we don't care if they "failed" the testing. */
 582         if (control_fd != -1)
 583                 status = 0;
 584         else
 585                 /* We don't restore contents for original parent. */
 586                 restore = false;
 587
 588         /* Cleanup everything, in reverse order. */
 589         tlist_for_each_rev(&history, i, list) {
 590                 /* Don't restore things our parent did. */
 591                 if (i == our_history_start)
 592                         restore = false;
 593
 594                 if (i->fail)
 595                         continue;
 596
 597                 if (i->cleanup)
 598                         i->cleanup(&i->u, restore);
 599
 600                 /* But their program shouldn't leak, even on failure. */
 601                 if (!forced_cleanup && i->can_leak) {
 602                         printf("Leak at %s:%u: --failpath=%s\n",
 603                                i->file, i->line, failpath_string());
 604                         status = 1;
 605                 }
 606         }
 607
 608         /* Put back mmaped files the way our parent (if any) expects. */
 609         free_mmapped_files(true);
 610
 611         free_everything();
 612         if (status == 0)
 613                 tell_parent(SUCCESS);
 614         else
 615                 tell_parent(FAILURE);
 616         exit(status);
 617 }
 618
 619 static bool following_path(void)
 620 {
 621         if (!failpath)
 622                 return false;
 623         /* + means continue after end, like normal. */
 624         if (*failpath == '+') {
 625                 failpath = NULL;
 626                 return false;
 627         }
 628         return true;
 629 }
 630
 631 static bool follow_path(struct failtest_call *call)
 632 {
 633         if (*failpath == '\0') {
 634                 /* Continue, but don't inject errors. */
 635                 return call->fail = false;
 636         }
 637
 638         if (tolower((unsigned char)*failpath) != info_to_arg[call->type])
 639                 errx(1, "Failpath expected '%s' got '%c'\n",
 640                      failpath, info_to_arg[call->type]);
 641         call->fail = cisupper(*(failpath++));
 642                         if (call->fail)
 643                                 call->can_leak = false;
 644         return call->fail;
 645 }
 646
 647 static bool should_fail(struct failtest_call *call)
 648 {
 649         int status;
 650         int control[2], output[2];
 651         enum info_type type = UNEXPECTED;
 652         char *out = NULL;
 653         size_t outlen = 0;
 654         struct failtest_call *dup;
 655
 656         if (call == &unrecorded_call)
 657                 return false;
 658
 659         if (following_path())
 660                 return follow_path(call);
 661
 662         /* Attach debugger if they asked for it. */
 663         if (debugpath) {
 664                 char *path;
 665
 666                 /* Pretend this last call matches whatever path wanted:
 667                  * keeps valgrind happy. */
 668                 call->fail = cisupper(debugpath[strlen(debugpath)-1]);
 669                 path = failpath_string();
 670
 671                 if (streq(path, debugpath)) {
 672                         char str[80];
 673
 674                         /* Don't timeout. */
 675                         signal(SIGUSR1, SIG_IGN);
 676                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
 677                                 getpid(), getpid());
 678                         if (system(str) == 0)
 679                                 sleep(5);
 680                 } else {
 681                         /* Ignore last character: could be upper or lower. */
 682                         path[strlen(path)-1] = '\0';
 683                         if (!strstarts(debugpath, path)) {
 684                                 fprintf(stderr,
 685                                         "--debugpath not followed: %s\n", path);
 686                                 debugpath = NULL;
 687                         }
 688                 }
 689                 free(path);
 690         }
 691
 692         /* Are we probing?  If so, we never fail twice. */
 693         if (probing) {
 694                 trace("Not failing %c due to FAIL_PROBE return\n",
 695                       info_to_arg[call->type]);
 696                 return call->fail = false;
 697         }
 698
 699         /* Don't fail more than once in the same place. */
 700         dup = failtable_get(&failtable, call);
 701         if (dup) {
 702                 trace("Not failing %c due to duplicate\n",
 703                       info_to_arg[call->type]);
 704                 return call->fail = false;
 705         }
 706
 707         if (failtest_hook) {
 708                 switch (failtest_hook(&history)) {
 709                 case FAIL_OK:
 710                         break;
 711                 case FAIL_PROBE:
 712                         probing = true;
 713                         break;
 714                 case FAIL_DONT_FAIL:
 715                         trace("Not failing %c due to failhook return\n",
 716                               info_to_arg[call->type]);
 717                         call->fail = false;
 718                         return false;
 719                 default:
 720                         abort();
 721                 }
 722         }
 723
 724         /* Add it to our table of calls. */
 725         failtable_add(&failtable, call);
 726
 727         /* We're going to fail in the child. */
 728         call->fail = true;
 729         if (pipe(control) != 0 || pipe(output) != 0)
 730                 err(1, "opening pipe");
 731
 732         /* Move out the way, to high fds. */
 733         control[0] = move_fd_to_high(control[0]);
 734         control[1] = move_fd_to_high(control[1]);
 735         output[0] = move_fd_to_high(output[0]);
 736         output[1] = move_fd_to_high(output[1]);
 737
 738         /* Prevent double-printing (in child and parent) */
 739         fflush(stdout);
 740         child = fork();
 741         if (child == -1)
 742                 err(1, "forking failed");
 743
 744         if (child == 0) {
 745                 traceindent++;
 746                 if (tracefd != -1) {
 747                         struct timeval diff;
 748                         const char *p;
 749                         char *failpath;
 750                         struct failtest_call *c;
 751
 752                         c = tlist_tail(&history, struct failtest_call, list);
 753                         diff = time_sub(time_now(), start);
 754                         failpath = failpath_string();
 755                         trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
 756                               (int)diff.tv_sec, (int)diff.tv_usec / 10000,
 757                               failpath);
 758                         free(failpath);
 759                         p = strrchr(c->file, '/');
 760                         if (p)
 761                                 trace("%s", p+1);
 762                         else
 763                                 trace("%s", c->file);
 764                         trace(":%u)\n", c->line);
 765                 }
 766                 /* From here on, we have to clean up! */
 767                 our_history_start = tlist_tail(&history, struct failtest_call,
 768                                                list);
 769                 close(control[0]);
 770                 close(output[0]);
 771                 /* Don't swallow stderr if we're tracing. */
 772                 if (tracefd != -1) {
 773                         dup2(output[1], STDOUT_FILENO);
 774                         dup2(output[1], STDERR_FILENO);
 775                         if (output[1] != STDOUT_FILENO
 776                             && output[1] != STDERR_FILENO)
 777                                 close(output[1]);
 778                 }
 779                 control_fd = move_fd_to_high(control[1]);
 780
 781                 /* Forget any of our parent's saved files. */
 782                 free_mmapped_files(false);
 783
 784                 /* Now, save any files we need to. */
 785                 save_mmapped_files();
 786
 787                 /* Failed calls can't leak. */
 788                 call->can_leak = false;
 789
 790                 return true;
 791         }
 792
 793         signal(SIGUSR1, hand_down);
 794
 795         close(control[1]);
 796         close(output[1]);
 797
 798         /* We grab output so we can display it; we grab writes so we
 799          * can compare. */
 800         do {
 801                 struct pollfd pfd[2];
 802                 int ret;
 803
 804                 pfd[0].fd = output[0];
 805                 pfd[0].events = POLLIN|POLLHUP;
 806                 pfd[1].fd = control[0];
 807                 pfd[1].events = POLLIN|POLLHUP;
 808
 809                 if (type == SUCCESS)
 810                         ret = poll(pfd, 1, failtest_timeout_ms);
 811                 else
 812                         ret = poll(pfd, 2, failtest_timeout_ms);
 813
 814                 if (ret == 0)
 815                         hand_down(SIGUSR1);
 816                 if (ret < 0) {
 817                         if (errno == EINTR)
 818                                 continue;
 819                         err(1, "Poll returned %i", ret);
 820                 }
 821
 822                 if (pfd[0].revents & POLLIN) {
 823                         ssize_t len;
 824
 825                         out = realloc(out, outlen + 8192);
 826                         len = read(output[0], out + outlen, 8192);
 827                         outlen += len;
 828                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 829                         if (read_all(control[0], &type, sizeof(type))) {
 830                                 if (type == WRITE) {
 831                                         if (!read_write_info(control[0]))
 832                                                 break;
 833                                 } else if (type == RELEASE_LOCKS) {
 834                                         release_locks();
 835                                         /* FIXME: Tell them we're done... */
 836                                 }
 837                         }
 838                 } else if (pfd[0].revents & POLLHUP) {
 839                         break;
 840                 }
 841         } while (type != FAILURE);
 842
 843         close(output[0]);
 844         close(control[0]);
 845         waitpid(child, &status, 0);
 846         if (!WIFEXITED(status)) {
 847                 if (WTERMSIG(status) == SIGUSR1)
 848                         child_fail(out, outlen, "Timed out");
 849                 else
 850                         child_fail(out, outlen, "Killed by signal %u: ",
 851                                    WTERMSIG(status));
 852         }
 853         /* Child printed failure already, just pass up exit code. */
 854         if (type == FAILURE) {
 855                 fprintf(stderr, "%.*s", (int)outlen, out);
 856                 tell_parent(type);
 857                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 858         }
 859         if (WEXITSTATUS(status) != 0)
 860                 child_fail(out, outlen, "Exited with status %i: ",
 861                            WEXITSTATUS(status));
 862
 863         free(out);
 864         signal(SIGUSR1, SIG_DFL);
 865
 866         /* Only child does probe. */
 867         probing = false;
 868
 869         /* We continue onwards without failing. */
 870         call->fail = false;
 871         return false;
 872 }
 873
 874 static void cleanup_calloc(struct calloc_call *call, bool restore)
 875 {
 876         trace("undoing calloc %p\n", call->ret);
 877         free(call->ret);
 878 }
 879
 880 void *failtest_calloc(size_t nmemb, size_t size,
 881                       const char *file, unsigned line)
 882 {
 883         struct failtest_call *p;
 884         struct calloc_call call;
 885         call.nmemb = nmemb;
 886         call.size = size;
 887         p = add_history(FAILTEST_CALLOC, true, file, line, &call);
 888
 889         if (should_fail(p)) {
 890                 p->u.calloc.ret = NULL;
 891                 p->error = ENOMEM;
 892         } else {
 893                 p->u.calloc.ret = calloc(nmemb, size);
 894                 set_cleanup(p, cleanup_calloc, struct calloc_call);
 895         }
 896         trace("calloc %zu x %zu %s:%u -> %p\n",
 897               nmemb, size, file, line, p->u.calloc.ret);
 898         errno = p->error;
 899         return p->u.calloc.ret;
 900 }
 901
 902 static void cleanup_malloc(struct malloc_call *call, bool restore)
 903 {
 904         trace("undoing malloc %p\n", call->ret);
 905         free(call->ret);
 906 }
 907
 908 void *failtest_malloc(size_t size, const char *file, unsigned line)
 909 {
 910         struct failtest_call *p;
 911         struct malloc_call call;
 912         call.size = size;
 913
 914         p = add_history(FAILTEST_MALLOC, true, file, line, &call);
 915         if (should_fail(p)) {
 916                 p->u.malloc.ret = NULL;
 917                 p->error = ENOMEM;
 918         } else {
 919                 p->u.malloc.ret = malloc(size);
 920                 set_cleanup(p, cleanup_malloc, struct malloc_call);
 921         }
 922         trace("malloc %zu %s:%u -> %p\n",
 923               size, file, line, p->u.malloc.ret);
 924         errno = p->error;
 925         return p->u.malloc.ret;
 926 }
 927
 928 static void cleanup_realloc(struct realloc_call *call, bool restore)
 929 {
 930         trace("undoing realloc %p\n", call->ret);
 931         free(call->ret);
 932 }
 933
 934 /* Walk back and find out if we got this ptr from a previous routine. */
 935 static void fixup_ptr_history(void *ptr, const char *why)
 936 {
 937         struct failtest_call *i;
 938
 939         /* Start at end of history, work back. */
 940         tlist_for_each_rev(&history, i, list) {
 941                 switch (i->type) {
 942                 case FAILTEST_REALLOC:
 943                         if (i->u.realloc.ret == ptr) {
 944                                 trace("found realloc %p %s:%u matching %s\n",
 945                                       ptr, i->file, i->line, why);
 946                                 i->cleanup = NULL;
 947                                 i->can_leak = false;
 948                                 return;
 949                         }
 950                         break;
 951                 case FAILTEST_MALLOC:
 952                         if (i->u.malloc.ret == ptr) {
 953                                 trace("found malloc %p %s:%u matching %s\n",
 954                                       ptr, i->file, i->line, why);
 955                                 i->cleanup = NULL;
 956                                 i->can_leak = false;
 957                                 return;
 958                         }
 959                         break;
 960                 case FAILTEST_CALLOC:
 961                         if (i->u.calloc.ret == ptr) {
 962                                 trace("found calloc %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                 default:
 970                         break;
 971                 }
 972         }
 973         trace("Did not find %p matching %s\n", ptr, why);
 974 }
 975
 976 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 977 {
 978         struct failtest_call *p;
 979         struct realloc_call call;
 980         call.size = size;
 981         p = add_history(FAILTEST_REALLOC, true, file, line, &call);
 982
 983         /* FIXME: Try one child moving allocation, one not. */
 984         if (should_fail(p)) {
 985                 p->u.realloc.ret = NULL;
 986                 p->error = ENOMEM;
 987         } else {
 988                 /* Don't catch this one in the history fixup... */
 989                 p->u.realloc.ret = NULL;
 990                 fixup_ptr_history(ptr, "realloc");
 991                 p->u.realloc.ret = realloc(ptr, size);
 992                 set_cleanup(p, cleanup_realloc, struct realloc_call);
 993         }
 994         trace("realloc %p %s:%u -> %p\n",
 995               ptr, file, line, p->u.realloc.ret);
 996         errno = p->error;
 997         return p->u.realloc.ret;
 998 }
 999
1000 /* FIXME: Record free, so we can terminate fixup_ptr_history correctly.
1001  * If there's an alloc we don't see, it could get confusing if it matches
1002  * a previous allocation we did see. */
1003 void failtest_free(void *ptr)
1004 {
1005         fixup_ptr_history(ptr, "free");
1006         trace("free %p\n", ptr);
1007         free(ptr);
1008 }
1009
1010
1011 static struct contents_saved *save_file(const char *pathname)
1012 {
1013         int fd;
1014         struct contents_saved *s;
1015
1016         fd = open(pathname, O_RDONLY);
1017         if (fd < 0)
1018                 return NULL;
1019
1020         s = save_contents(pathname, fd, lseek(fd, 0, SEEK_END), 0,
1021                           "open with O_TRUNC");
1022         close(fd);
1023         return s;
1024 }
1025
1026 /* Optimization: don't create a child for an open which *we know*
1027  * would fail anyway. */
1028 static bool open_would_fail(const char *pathname, int flags)
1029 {
1030         if ((flags & O_ACCMODE) == O_RDONLY)
1031                 return access(pathname, R_OK) != 0;
1032         if (!(flags & O_CREAT)) {
1033                 if ((flags & O_ACCMODE) == O_WRONLY)
1034                         return access(pathname, W_OK) != 0;
1035                 if ((flags & O_ACCMODE) == O_RDWR)
1036                         return access(pathname, W_OK) != 0
1037                                 || access(pathname, R_OK) != 0;
1038         }
1039         /* FIXME: We could check if it exists, for O_CREAT|O_EXCL */
1040         return false;
1041 }
1042
1043 static void cleanup_open(struct open_call *call, bool restore)
1044 {
1045         if (restore && call->saved)
1046                 restore_contents(container_of(call, struct failtest_call,
1047                                               u.open),
1048                                  call->saved, false, "open with O_TRUNC");
1049         if (!call->closed) {
1050                 trace("Cleaning up open %s by closing fd %i\n",
1051                       call->pathname, call->ret);
1052                 close(call->ret);
1053                 call->closed = true;
1054         }
1055         free(call->saved);
1056 }
1057
1058 int failtest_open(const char *pathname,
1059                   const char *file, unsigned line, ...)
1060 {
1061         struct failtest_call *p;
1062         struct open_call call;
1063         va_list ap;
1064
1065         call.pathname = strdup(pathname);
1066         va_start(ap, line);
1067         call.flags = va_arg(ap, int);
1068         call.always_save = false;
1069         call.closed = false;
1070         if (call.flags & O_CREAT) {
1071                 call.mode = va_arg(ap, int);
1072                 va_end(ap);
1073         }
1074         p = add_history(FAILTEST_OPEN, true, file, line, &call);
1075         /* Avoid memory leak! */
1076         if (p == &unrecorded_call)
1077                 free((char *)call.pathname);
1078
1079         if (should_fail(p)) {
1080                 /* Don't bother inserting failures that would happen anyway. */
1081                 if (open_would_fail(pathname, call.flags)) {
1082                         trace("Open would have failed anyway: stopping\n");
1083                         failtest_cleanup(true, 0);
1084                 }
1085                 p->u.open.ret = -1;
1086                 /* FIXME: Play with error codes? */
1087                 p->error = EACCES;
1088         } else {
1089                 /* Save the old version if they're truncating it. */
1090                 if (call.flags & O_TRUNC)
1091                         p->u.open.saved = save_file(pathname);
1092                 else
1093                         p->u.open.saved = NULL;
1094                 p->u.open.ret = open(pathname, call.flags, call.mode);
1095                 if (p->u.open.ret == -1) {
1096                         p->u.open.closed = true;
1097                         p->can_leak = false;
1098                 } else {
1099                         set_cleanup(p, cleanup_open, struct open_call);
1100                 }
1101         }
1102         trace("open %s %s:%u -> %i (opener %p)\n",
1103               pathname, file, line, p->u.open.ret, &p->u.open);
1104         errno = p->error;
1105         return p->u.open.ret;
1106 }
1107
1108 static void cleanup_mmap(struct mmap_call *mmap, bool restore)
1109 {
1110         trace("cleaning up mmap @%p (opener %p)\n",
1111               mmap->ret, mmap->opener);
1112         if (restore)
1113                 restore_contents(mmap->opener, mmap->saved, false, "mmap");
1114         free(mmap->saved);
1115 }
1116
1117 void *failtest_mmap(void *addr, size_t length, int prot, int flags,
1118                     int fd, off_t offset, const char *file, unsigned line)
1119 {
1120         struct failtest_call *p;
1121         struct mmap_call call;
1122
1123         call.addr = addr;
1124         call.length = length;
1125         call.prot = prot;
1126         call.flags = flags;
1127         call.offset = offset;
1128         call.fd = fd;
1129         call.opener = opener_of(fd);
1130
1131         /* If we don't know what file it was, don't fail. */
1132         if (!call.opener) {
1133                 if (fd != -1) {
1134                         fwarnx("failtest_mmap: couldn't figure out source for"
1135                                " fd %i at %s:%u", fd, file, line);
1136                 }
1137                 addr = mmap(addr, length, prot, flags, fd, offset);
1138                 trace("mmap of fd %i -> %p (opener = NULL)\n", fd, addr);
1139                 return addr;
1140         }
1141
1142         p = add_history(FAILTEST_MMAP, false, file, line, &call);
1143         if (should_fail(p)) {
1144                 p->u.mmap.ret = MAP_FAILED;
1145                 p->error = ENOMEM;
1146         } else {
1147                 p->u.mmap.ret = mmap(addr, length, prot, flags, fd, offset);
1148                 /* Save contents if we're writing to a normal file */
1149                 if (p->u.mmap.ret != MAP_FAILED
1150                     && (prot & PROT_WRITE)
1151                     && call.opener->type == FAILTEST_OPEN) {
1152                         const char *fname = call.opener->u.open.pathname;
1153                         p->u.mmap.saved = save_contents(fname, fd, length,
1154                                                         offset, "being mmapped");
1155                         set_cleanup(p, cleanup_mmap, struct mmap_call);
1156                 }
1157         }
1158         trace("mmap of fd %i %s:%u -> %p (opener = %p)\n",
1159               fd, file, line, addr, call.opener);
1160         errno = p->error;
1161         return p->u.mmap.ret;
1162 }
1163
1164 static void cleanup_pipe(struct pipe_call *call, bool restore)
1165 {
1166         trace("cleaning up pipe fd=%i%s,%i%s\n",
1167               call->fds[0], call->closed[0] ? "(already closed)" : "",
1168               call->fds[1], call->closed[1] ? "(already closed)" : "");
1169         if (!call->closed[0])
1170                 close(call->fds[0]);
1171         if (!call->closed[1])
1172                 close(call->fds[1]);
1173 }
1174
1175 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
1176 {
1177         struct failtest_call *p;
1178         struct pipe_call call;
1179
1180         p = add_history(FAILTEST_PIPE, true, file, line, &call);
1181         if (should_fail(p)) {
1182                 p->u.open.ret = -1;
1183                 /* FIXME: Play with error codes? */
1184                 p->error = EMFILE;
1185         } else {
1186                 p->u.pipe.ret = pipe(p->u.pipe.fds);
1187                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
1188                 set_cleanup(p, cleanup_pipe, struct pipe_call);
1189         }
1190
1191         trace("pipe %s:%u -> %i,%i\n", file, line,
1192               p->u.pipe.ret ? -1 : p->u.pipe.fds[0],
1193               p->u.pipe.ret ? -1 : p->u.pipe.fds[1]);
1194
1195         /* This causes valgrind to notice if they use pipefd[] after failure */
1196         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
1197         errno = p->error;
1198         return p->u.pipe.ret;
1199 }
1200
1201 static void cleanup_read(struct read_call *call, bool restore)
1202 {
1203         if (restore) {
1204                 trace("cleaning up read on fd %i: seeking to %llu\n",
1205                       call->fd, (long long)call->off);
1206
1207                 /* Read (not readv!) moves file offset! */
1208                 if (lseek(call->fd, call->off, SEEK_SET) != call->off) {
1209                         fwarn("Restoring lseek pointer failed (read)");
1210                 }
1211         }
1212 }
1213
1214 static ssize_t failtest_add_read(int fd, void *buf, size_t count, off_t off,
1215                                  bool is_pread, const char *file, unsigned line)
1216 {
1217         struct failtest_call *p;
1218         struct read_call call;
1219         call.fd = fd;
1220         call.buf = buf;
1221         call.count = count;
1222         call.off = off;
1223         p = add_history(FAILTEST_READ, false, file, line, &call);
1224
1225         /* FIXME: Try partial read returns. */
1226         if (should_fail(p)) {
1227                 p->u.read.ret = -1;
1228                 p->error = EIO;
1229         } else {
1230                 if (is_pread)
1231                         p->u.read.ret = pread(fd, buf, count, off);
1232                 else {
1233                         p->u.read.ret = read(fd, buf, count);
1234                         if (p->u.read.ret != -1)
1235                                 set_cleanup(p, cleanup_read, struct read_call);
1236                 }
1237         }
1238         trace("%sread %s:%u fd %i %zu@%llu -> %i\n",
1239               is_pread ? "p" : "", file, line, fd, count, (long long)off,
1240               p->u.read.ret);
1241         errno = p->error;
1242         return p->u.read.ret;
1243 }
1244
1245 static void cleanup_write(struct write_call *write, bool restore)
1246 {
1247         trace("cleaning up write on %s\n", write->opener->u.open.pathname);
1248         if (restore)
1249                 restore_contents(write->opener, write->saved, !write->is_pwrite,
1250                                  "write");
1251         free(write->saved);
1252 }
1253
1254 static ssize_t failtest_add_write(int fd, const void *buf,
1255                                   size_t count, off_t off,
1256                                   bool is_pwrite,
1257                                   const char *file, unsigned line)
1258 {
1259         struct failtest_call *p;
1260         struct write_call call;
1261
1262         call.fd = fd;
1263         call.buf = buf;
1264         call.count = count;
1265         call.off = off;
1266         call.is_pwrite = is_pwrite;
1267         call.opener = opener_of(fd);
1268         p = add_history(FAILTEST_WRITE, false, file, line, &call);
1269
1270         /* If we're a child, we need to make sure we write the same thing
1271          * to non-files as the parent does, so tell it. */
1272         if (control_fd != -1 && off == (off_t)-1) {
1273                 enum info_type type = WRITE;
1274
1275                 write_all(control_fd, &type, sizeof(type));
1276                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
1277                 write_all(control_fd, buf, count);
1278         }
1279
1280         /* FIXME: Try partial write returns. */
1281         if (should_fail(p)) {
1282                 p->u.write.ret = -1;
1283                 p->error = EIO;
1284         } else {
1285                 bool is_file;
1286                 assert(call.opener == p->u.write.opener);
1287
1288                 if (p->u.write.opener) {
1289                         is_file = (p->u.write.opener->type == FAILTEST_OPEN);
1290                 } else {
1291                         /* We can't unwind it, so at least check same
1292                          * in parent and child. */
1293                         is_file = false;
1294                 }
1295
1296                 /* FIXME: We assume same write order in parent and child */
1297                 if (!is_file && child_writes_num != 0) {
1298                         if (child_writes[0].fd != fd)
1299                                 errx(1, "Child wrote to fd %u, not %u?",
1300                                      child_writes[0].fd, fd);
1301                         if (child_writes[0].off != p->u.write.off)
1302                                 errx(1, "Child wrote to offset %zu, not %zu?",
1303                                      (size_t)child_writes[0].off,
1304                                      (size_t)p->u.write.off);
1305                         if (child_writes[0].count != count)
1306                                 errx(1, "Child wrote length %zu, not %zu?",
1307                                      child_writes[0].count, count);
1308                         if (memcmp(child_writes[0].buf, buf, count)) {
1309                                 child_fail(NULL, 0,
1310                                            "Child wrote differently to"
1311                                            " fd %u than we did!\n", fd);
1312                         }
1313                         free((char *)child_writes[0].buf);
1314                         child_writes_num--;
1315                         memmove(&child_writes[0], &child_writes[1],
1316                                 sizeof(child_writes[0]) * child_writes_num);
1317
1318                         /* Child wrote it already. */
1319                         trace("write %s:%i on fd %i already done by child\n",
1320                               file, line, fd);
1321                         p->u.write.ret = count;
1322                         errno = p->error;
1323                         return p->u.write.ret;
1324                 }
1325
1326                 if (is_file) {
1327                         p->u.write.saved = save_contents(call.opener->u.open.pathname,
1328                                                          fd, count, off,
1329                                                          "being overwritten");
1330                         set_cleanup(p, cleanup_write, struct write_call);
1331                 }
1332
1333                 /* Though off is current seek ptr for write case, we need to
1334                  * move it.  write() does that for us. */
1335                 if (p->u.write.is_pwrite)
1336                         p->u.write.ret = pwrite(fd, buf, count, off);
1337                 else
1338                         p->u.write.ret = write(fd, buf, count);
1339         }
1340         trace("%swrite %s:%i %zu@%llu on fd %i -> %i\n",
1341               p->u.write.is_pwrite ? "p" : "",
1342               file, line, count, (long long)off, fd, p->u.write.ret);
1343         errno = p->error;
1344         return p->u.write.ret;
1345 }
1346
1347 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t offset,
1348                         const char *file, unsigned line)
1349 {
1350         return failtest_add_write(fd, buf, count, offset, true, file, line);
1351 }
1352
1353 ssize_t failtest_write(int fd, const void *buf, size_t count,
1354                        const char *file, unsigned line)
1355 {
1356         return failtest_add_write(fd, buf, count, lseek(fd, 0, SEEK_CUR), false,
1357                                   file, line);
1358 }
1359
1360 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
1361                        const char *file, unsigned line)
1362 {
1363         return failtest_add_read(fd, buf, count, off, true, file, line);
1364 }
1365
1366 ssize_t failtest_read(int fd, void *buf, size_t count,
1367                       const char *file, unsigned line)
1368 {
1369         return failtest_add_read(fd, buf, count, lseek(fd, 0, SEEK_CUR), false,
1370                                  file, line);
1371 }
1372
1373 static struct lock_info *WARN_UNUSED_RESULT
1374 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
1375 {
1376         unsigned int i;
1377         struct lock_info *l;
1378
1379         for (i = 0; i < lock_num; i++) {
1380                 l = &locks[i];
1381
1382                 if (l->fd != fd)
1383                         continue;
1384                 /* Four cases we care about:
1385                  * Start overlap:
1386                  *      l =    |      |
1387                  *      new = |   |
1388                  * Mid overlap:
1389                  *      l =    |      |
1390                  *      new =    |  |
1391                  * End overlap:
1392                  *      l =    |      |
1393                  *      new =      |    |
1394                  * Total overlap:
1395                  *      l =    |      |
1396                  *      new = |         |
1397                  */
1398                 if (start > l->start && end < l->end) {
1399                         /* Mid overlap: trim entry, add new one. */
1400                         off_t new_start, new_end;
1401                         new_start = end + 1;
1402                         new_end = l->end;
1403                         trace("splitting lock on fd %i from %llu-%llu"
1404                               " to %llu-%llu\n",
1405                               fd, (long long)l->start, (long long)l->end,
1406                               (long long)l->start, (long long)start - 1);
1407                         l->end = start - 1;
1408                         locks = add_lock(locks,
1409                                          fd, new_start, new_end, l->type);
1410                         l = &locks[i];
1411                 } else if (start <= l->start && end >= l->end) {
1412                         /* Total overlap: eliminate entry. */
1413                         trace("erasing lock on fd %i %llu-%llu\n",
1414                               fd, (long long)l->start, (long long)l->end);
1415                         l->end = 0;
1416                         l->start = 1;
1417                 } else if (end >= l->start && end < l->end) {
1418                         trace("trimming lock on fd %i from %llu-%llu"
1419                               " to %llu-%llu\n",
1420                               fd, (long long)l->start, (long long)l->end,
1421                               (long long)end + 1, (long long)l->end);
1422                         /* Start overlap: trim entry. */
1423                         l->start = end + 1;
1424                 } else if (start > l->start && start <= l->end) {
1425                         trace("trimming lock on fd %i from %llu-%llu"
1426                               " to %llu-%llu\n",
1427                               fd, (long long)l->start, (long long)l->end,
1428                               (long long)l->start, (long long)start - 1);
1429                         /* End overlap: trim entry. */
1430                         l->end = start-1;
1431                 }
1432                 /* Nothing left?  Remove it. */
1433                 if (l->end < l->start) {
1434                         trace("forgetting lock on fd %i\n", fd);
1435                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
1436                         i--;
1437                 }
1438         }
1439
1440         if (type != F_UNLCK) {
1441                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
1442                 l = &locks[lock_num++];
1443                 l->fd = fd;
1444                 l->start = start;
1445                 l->end = end;
1446                 l->type = type;
1447                 trace("new lock on fd %i %llu-%llu\n",
1448                       fd, (long long)l->start, (long long)l->end);
1449         }
1450         return locks;
1451 }
1452
1453 /* We trap this so we can record it: we don't fail it. */
1454 int failtest_close(int fd, const char *file, unsigned line)
1455 {
1456         struct close_call call;
1457         struct failtest_call *p, *opener;
1458
1459         /* Do this before we add ourselves to history! */
1460         opener = opener_of(fd);
1461
1462         call.fd = fd;
1463         p = add_history(FAILTEST_CLOSE, false, file, line, &call);
1464         p->fail = false;
1465
1466         /* Consume close from failpath (shouldn't tell us to fail). */
1467         if (following_path()) {
1468                 if (follow_path(p))
1469                         abort();
1470         }
1471
1472         trace("close on fd %i\n", fd);
1473         if (fd < 0)
1474                 return close(fd);
1475
1476         /* Mark opener as not leaking, remove its cleanup function. */
1477         if (opener) {
1478                 trace("close on fd %i found opener %p\n", fd, opener);
1479                 if (opener->type == FAILTEST_PIPE) {
1480                         /* From a pipe? */
1481                         if (opener->u.pipe.fds[0] == fd) {
1482                                 assert(!opener->u.pipe.closed[0]);
1483                                 opener->u.pipe.closed[0] = true;
1484                         } else if (opener->u.pipe.fds[1] == fd) {
1485                                 assert(!opener->u.pipe.closed[1]);
1486                                 opener->u.pipe.closed[1] = true;
1487                         } else
1488                                 abort();
1489                         opener->can_leak = (!opener->u.pipe.closed[0]
1490                                             || !opener->u.pipe.closed[1]);
1491                 } else if (opener->type == FAILTEST_OPEN) {
1492                         opener->u.open.closed = true;
1493                         opener->can_leak = false;
1494                 } else
1495                         abort();
1496         }
1497
1498         /* Restore offset now, in case parent shared (can't do after close!). */
1499         if (control_fd != -1) {
1500                 struct failtest_call *i;
1501
1502                 tlist_for_each_rev(&history, i, list) {
1503                         if (i == our_history_start)
1504                                 break;
1505                         if (i == opener)
1506                                 break;
1507                         if (i->type == FAILTEST_LSEEK && i->u.lseek.fd == fd) {
1508                                 trace("close on fd %i undoes lseek\n", fd);
1509                                 /* This seeks back. */
1510                                 i->cleanup(&i->u, true);
1511                                 i->cleanup = NULL;
1512                         } else if (i->type == FAILTEST_WRITE
1513                                    && i->u.write.fd == fd
1514                                    && !i->u.write.is_pwrite) {
1515                                 trace("close on fd %i undoes write"
1516                                       " offset change\n", fd);
1517                                 /* Write (not pwrite!) moves file offset! */
1518                                 if (lseek(fd, i->u.write.off, SEEK_SET)
1519                                     != i->u.write.off) {
1520                                         fwarn("Restoring lseek pointer failed (write)");
1521                                 }
1522                         } else if (i->type == FAILTEST_READ
1523                                    && i->u.read.fd == fd) {
1524                                 /* preads don't *have* cleanups */
1525                                 if (i->cleanup) {
1526                                         trace("close on fd %i undoes read"
1527                                               " offset change\n", fd);
1528                                         /* This seeks back. */
1529                                         i->cleanup(&i->u, true);
1530                                         i->cleanup = NULL;
1531                                 }
1532                         }
1533                 }
1534         }
1535
1536         /* Close unlocks everything. */
1537         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
1538         return close(fd);
1539 }
1540
1541 /* Zero length means "to end of file" */
1542 static off_t end_of(off_t start, off_t len)
1543 {
1544         if (len == 0)
1545                 return off_max();
1546         return start + len - 1;
1547 }
1548
1549 /* FIXME: This only handles locks, really. */
1550 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
1551 {
1552         struct failtest_call *p;
1553         struct fcntl_call call;
1554         va_list ap;
1555
1556         call.fd = fd;
1557         call.cmd = cmd;
1558
1559         /* Argument extraction. */
1560         switch (cmd) {
1561         case F_SETFL:
1562         case F_SETFD:
1563                 va_start(ap, cmd);
1564                 call.arg.l = va_arg(ap, long);
1565                 va_end(ap);
1566                 trace("fcntl on fd %i F_SETFL/F_SETFD\n", fd);
1567                 return fcntl(fd, cmd, call.arg.l);
1568         case F_GETFD:
1569         case F_GETFL:
1570                 trace("fcntl on fd %i F_GETFL/F_GETFD\n", fd);
1571                 return fcntl(fd, cmd);
1572         case F_GETLK:
1573                 trace("fcntl on fd %i F_GETLK\n", fd);
1574                 get_locks();
1575                 va_start(ap, cmd);
1576                 call.arg.fl = *va_arg(ap, struct flock *);
1577                 va_end(ap);
1578                 return fcntl(fd, cmd, &call.arg.fl);
1579         case F_SETLK:
1580         case F_SETLKW:
1581                 trace("fcntl on fd %i F_SETLK%s\n",
1582                       fd, cmd == F_SETLKW ? "W" : "");
1583                 va_start(ap, cmd);
1584                 call.arg.fl = *va_arg(ap, struct flock *);
1585                 va_end(ap);
1586                 break;
1587         default:
1588                 /* This means you need to implement it here. */
1589                 err(1, "failtest: unknown fcntl %u", cmd);
1590         }
1591
1592         p = add_history(FAILTEST_FCNTL, false, file, line, &call);
1593
1594         if (should_fail(p)) {
1595                 p->u.fcntl.ret = -1;
1596                 if (p->u.fcntl.cmd == F_SETLK)
1597                         p->error = EAGAIN;
1598                 else
1599                         p->error = EDEADLK;
1600         } else {
1601                 get_locks();
1602                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1603                                        &p->u.fcntl.arg.fl);
1604                 if (p->u.fcntl.ret == -1)
1605                         p->error = errno;
1606                 else {
1607                         /* We don't handle anything else yet. */
1608                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1609                         locks = add_lock(locks,
1610                                          p->u.fcntl.fd,
1611                                          p->u.fcntl.arg.fl.l_start,
1612                                          end_of(p->u.fcntl.arg.fl.l_start,
1613                                                 p->u.fcntl.arg.fl.l_len),
1614                                          p->u.fcntl.arg.fl.l_type);
1615                 }
1616         }
1617         trace("fcntl on fd %i -> %i\n", fd, p->u.fcntl.ret);
1618         errno = p->error;
1619         return p->u.fcntl.ret;
1620 }
1621
1622 static void cleanup_lseek(struct lseek_call *call, bool restore)
1623 {
1624         if (restore) {
1625                 trace("cleaning up lseek on fd %i -> %llu\n",
1626                       call->fd, (long long)call->old_off);
1627                 if (lseek(call->fd, call->old_off, SEEK_SET) != call->old_off)
1628                         fwarn("Restoring lseek pointer failed");
1629         }
1630 }
1631
1632 /* We trap this so we can undo it: we don't fail it. */
1633 off_t failtest_lseek(int fd, off_t offset, int whence, const char *file,
1634                      unsigned int line)
1635 {
1636         struct failtest_call *p;
1637         struct lseek_call call;
1638         call.fd = fd;
1639         call.offset = offset;
1640         call.whence = whence;
1641         call.old_off = lseek(fd, 0, SEEK_CUR);
1642
1643         p = add_history(FAILTEST_LSEEK, false, file, line, &call);
1644         p->fail = false;
1645
1646         /* Consume lseek from failpath. */
1647         if (failpath)
1648                 if (should_fail(p))
1649                         abort();
1650
1651         p->u.lseek.ret = lseek(fd, offset, whence);
1652
1653         if (p->u.lseek.ret != (off_t)-1)
1654                 set_cleanup(p, cleanup_lseek, struct lseek_call);
1655
1656         trace("lseek %s:%u on fd %i from %llu to %llu%s\n",
1657               file, line, fd, (long long)call.old_off, (long long)offset,
1658               whence == SEEK_CUR ? " (from current off)" :
1659               whence == SEEK_END ? " (from end)" :
1660               whence == SEEK_SET ? "" : " (invalid whence)");
1661         return p->u.lseek.ret;
1662 }
1663
1664
1665 pid_t failtest_getpid(const char *file, unsigned line)
1666 {
1667         /* You must call failtest_init first! */
1668         assert(orig_pid);
1669         return orig_pid;
1670 }
1671
1672 void failtest_init(int argc, char *argv[])
1673 {
1674         unsigned int i;
1675
1676         orig_pid = getpid();
1677
1678         warnfd = move_fd_to_high(dup(STDERR_FILENO));
1679         for (i = 1; i < argc; i++) {
1680                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1681                         failpath = argv[i] + strlen("--failpath=");
1682                 } else if (strcmp(argv[i], "--trace") == 0) {
1683                         tracefd = warnfd;
1684                         failtest_timeout_ms = -1;
1685                 } else if (!strncmp(argv[i], "--debugpath=",
1686                                     strlen("--debugpath="))) {
1687                         debugpath = argv[i] + strlen("--debugpath=");
1688                 }
1689         }
1690         failtable_init(&failtable);
1691         start = time_now();
1692 }
1693
1694 bool failtest_has_failed(void)
1695 {
1696         return control_fd != -1;
1697 }
1698
1699 void failtest_exit(int status)
1700 {
1701         trace("failtest_exit with status %i\n", status);
1702         if (failtest_exit_check) {
1703                 if (!failtest_exit_check(&history))
1704                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1705         }
1706
1707         failtest_cleanup(false, status);
1708 }