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

   1 #include "config.h"
   2 #include <stdarg.h>
   3 #include <string.h>
   4 #include <stdio.h>
   5 #include <stdarg.h>
   6 #include <ctype.h>
   7 #include <err.h>
   8 #include <unistd.h>
   9 #include <poll.h>
  10 #include <errno.h>
  11 #include <sys/types.h>
  12 #include <sys/wait.h>
  13 #include <sys/stat.h>
  14 #include <sys/time.h>
  15 #include <signal.h>
  16 #include <assert.h>
  17 #include <ccan/read_write_all/read_write_all.h>
  18 #include <ccan/failtest/failtest_proto.h>
  19 #include <ccan/failtest/failtest.h>
  20 #include <ccan/build_assert/build_assert.h>
  21
  22 enum failtest_result (*failtest_hook)(struct failtest_call *, unsigned);
  23
  24 static int tracefd = -1;
  25
  26 unsigned int failtest_timeout_ms = 20000;
  27
  28 const char *failpath;
  29 const char *debugpath;
  30
  31 enum info_type {
  32         WRITE,
  33         RELEASE_LOCKS,
  34         FAILURE,
  35         SUCCESS,
  36         UNEXPECTED
  37 };
  38
  39 struct lock_info {
  40         int fd;
  41         /* end is inclusive: you can't have a 0-byte lock. */
  42         off_t start, end;
  43         int type;
  44 };
  45
  46 bool (*failtest_exit_check)(struct failtest_call *history, unsigned num);
  47
  48 static struct failtest_call *history = NULL;
  49 static unsigned int history_num = 0;
  50 static int control_fd = -1;
  51 static struct timeval start;
  52 static unsigned int probe_count = 0;
  53
  54 static struct write_call *child_writes = NULL;
  55 static unsigned int child_writes_num = 0;
  56
  57 static pid_t lock_owner;
  58 static struct lock_info *locks = NULL;
  59 static unsigned int lock_num = 0;
  60
  61 static pid_t orig_pid;
  62
  63 static const char info_to_arg[] = "mceoxprwf";
  64
  65 /* Dummy call used for failtest_undo wrappers. */
  66 static struct failtest_call unrecorded_call;
  67
  68 static struct failtest_call *add_history_(enum failtest_call_type type,
  69                                           const char *file,
  70                                           unsigned int line,
  71                                           const void *elem,
  72                                           size_t elem_size)
  73 {
  74         /* NULL file is how we suppress failure. */
  75         if (!file)
  76                 return &unrecorded_call;
  77
  78         history = realloc(history, (history_num + 1) * sizeof(*history));
  79         history[history_num].type = type;
  80         history[history_num].file = file;
  81         history[history_num].line = line;
  82         history[history_num].cleanup = NULL;
  83         memcpy(&history[history_num].u, elem, elem_size);
  84         return &history[history_num++];
  85 }
  86
  87 #define add_history(type, file, line, elem) \
  88         add_history_((type), (file), (line), (elem), sizeof(*(elem)))
  89
  90 /* We do a fake call inside a sizeof(), to check types. */
  91 #define set_cleanup(call, clean, type)                  \
  92         (call)->cleanup = (void *)((void)sizeof(clean((type *)NULL),1), (clean))
  93
  94 static bool read_write_info(int fd)
  95 {
  96         struct write_call *w;
  97         char *buf;
  98
  99         /* We don't need all of this, but it's simple. */
 100         child_writes = realloc(child_writes,
 101                                (child_writes_num+1) * sizeof(child_writes[0]));
 102         w = &child_writes[child_writes_num];
 103         if (!read_all(fd, w, sizeof(*w)))
 104                 return false;
 105
 106         w->buf = buf = malloc(w->count);
 107         if (!read_all(fd, buf, w->count))
 108                 return false;
 109
 110         child_writes_num++;
 111         return true;
 112 }
 113
 114 static char *failpath_string(void)
 115 {
 116         unsigned int i;
 117         char *ret = malloc(history_num + 1);
 118
 119         for (i = 0; i < history_num; i++) {
 120                 ret[i] = info_to_arg[history[i].type];
 121                 if (history[i].fail)
 122                         ret[i] = toupper(ret[i]);
 123         }
 124         ret[i] = '\0';
 125         return ret;
 126 }
 127
 128 static void tell_parent(enum info_type type)
 129 {
 130         if (control_fd != -1)
 131                 write_all(control_fd, &type, sizeof(type));
 132 }
 133
 134 static void child_fail(const char *out, size_t outlen, const char *fmt, ...)
 135 {
 136         va_list ap;
 137         char *path = failpath_string();
 138
 139         va_start(ap, fmt);
 140         vfprintf(stderr, fmt, ap);
 141         va_end(ap);
 142
 143         fprintf(stderr, "%.*s", (int)outlen, out);
 144         printf("To reproduce: --failpath=%s\n", path);
 145         free(path);
 146         tell_parent(FAILURE);
 147         exit(1);
 148 }
 149
 150 static void trace(const char *fmt, ...)
 151 {
 152         va_list ap;
 153
 154         if (tracefd == -1)
 155                 return;
 156
 157         va_start(ap, fmt);
 158         vdprintf(tracefd, fmt, ap);
 159         va_end(ap);
 160 }
 161
 162 static pid_t child;
 163
 164 static void hand_down(int signum)
 165 {
 166         kill(child, signum);
 167 }
 168
 169 static void release_locks(void)
 170 {
 171         /* Locks were never acquired/reacquired? */
 172         if (lock_owner == 0)
 173                 return;
 174
 175         /* We own them?  Release them all. */
 176         if (lock_owner == getpid()) {
 177                 unsigned int i;
 178                 struct flock fl;
 179                 fl.l_type = F_UNLCK;
 180                 fl.l_whence = SEEK_SET;
 181                 fl.l_start = 0;
 182                 fl.l_len = 0;
 183
 184                 for (i = 0; i < lock_num; i++)
 185                         fcntl(locks[i].fd, F_SETLK, &fl);
 186         } else {
 187                 /* Our parent must have them; pass request up. */
 188                 enum info_type type = RELEASE_LOCKS;
 189                 assert(control_fd != -1);
 190                 write_all(control_fd, &type, sizeof(type));
 191         }
 192         lock_owner = 0;
 193 }
 194
 195 /* off_t is a signed type.  Getting its max is non-trivial. */
 196 static off_t off_max(void)
 197 {
 198         BUILD_ASSERT(sizeof(off_t) == 4 || sizeof(off_t) == 8);
 199         if (sizeof(off_t) == 4)
 200                 return (off_t)0x7FFFFFF;
 201         else
 202                 return (off_t)0x7FFFFFFFFFFFFFFULL;
 203 }
 204
 205 static void get_locks(void)
 206 {
 207         unsigned int i;
 208         struct flock fl;
 209
 210         if (lock_owner == getpid())
 211                 return;
 212
 213         if (lock_owner != 0) {
 214                 enum info_type type = RELEASE_LOCKS;
 215                 assert(control_fd != -1);
 216                 write_all(control_fd, &type, sizeof(type));
 217         }
 218
 219         fl.l_whence = SEEK_SET;
 220
 221         for (i = 0; i < lock_num; i++) {
 222                 fl.l_type = locks[i].type;
 223                 fl.l_start = locks[i].start;
 224                 if (locks[i].end == off_max())
 225                         fl.l_len = 0;
 226                 else
 227                         fl.l_len = locks[i].end - locks[i].start + 1;
 228
 229                 if (fcntl(locks[i].fd, F_SETLKW, &fl) != 0)
 230                         abort();
 231         }
 232         lock_owner = getpid();
 233 }
 234
 235 struct saved_file {
 236         struct saved_file *next;
 237         int fd;
 238         void *contents;
 239         off_t off, len;
 240 };
 241
 242 static struct saved_file *save_file(struct saved_file *next, int fd)
 243 {
 244         struct saved_file *s = malloc(sizeof(*s));
 245
 246         s->next = next;
 247         s->fd = fd;
 248         s->off = lseek(fd, 0, SEEK_CUR);
 249         /* Special file?  Erk... */
 250         assert(s->off != -1);
 251         s->len = lseek(fd, 0, SEEK_END);
 252         lseek(fd, 0, SEEK_SET);
 253         s->contents = malloc(s->len);
 254         if (read(fd, s->contents, s->len) != s->len)
 255                 err(1, "Failed to save %zu bytes", (size_t)s->len);
 256         lseek(fd, s->off, SEEK_SET);
 257         return s;
 258 }
 259
 260 /* We have little choice but to save and restore open files: mmap means we
 261  * can really intercept changes in the child.
 262  *
 263  * We could do non-mmap'ed files on demand, however. */
 264 static struct saved_file *save_files(void)
 265 {
 266         struct saved_file *files = NULL;
 267         int i;
 268
 269         /* Figure out the set of live fds. */
 270         for (i = history_num - 2; i >= 0; i--) {
 271                 if (history[i].type == FAILTEST_OPEN) {
 272                         int fd = history[i].u.open.ret;
 273                         /* Only do successful, writable fds. */
 274                         if (fd < 0)
 275                                 continue;
 276
 277                         /* If it was closed, cleanup == NULL. */
 278                         if (!history[i].cleanup)
 279                                 continue;
 280
 281                         if ((history[i].u.open.flags & O_RDWR) == O_RDWR) {
 282                                 files = save_file(files, fd);
 283                         } else if ((history[i].u.open.flags & O_WRONLY)
 284                                    == O_WRONLY) {
 285                                 /* FIXME: Handle O_WRONLY.  Open with O_RDWR? */
 286                                 abort();
 287                         }
 288                 }
 289         }
 290
 291         return files;
 292 }
 293
 294 static void restore_files(struct saved_file *s)
 295 {
 296         while (s) {
 297                 struct saved_file *next = s->next;
 298
 299                 lseek(s->fd, 0, SEEK_SET);
 300                 if (write(s->fd, s->contents, s->len) != s->len)
 301                         err(1, "Failed to restore %zu bytes", (size_t)s->len);
 302                 if (ftruncate(s->fd, s->len) != 0)
 303                         err(1, "Failed to trim file to length %zu",
 304                             (size_t)s->len);
 305                 free(s->contents);
 306                 lseek(s->fd, s->off, SEEK_SET);
 307                 free(s);
 308                 s = next;
 309         }
 310 }
 311
 312 /* Free up memory, so valgrind doesn't report leaks. */
 313 static void free_everything(void)
 314 {
 315         unsigned int i;
 316
 317         /* We don't do this in cleanup: needed even for failed opens. */
 318         for (i = 0; i < history_num; i++) {
 319                 if (history[i].type == FAILTEST_OPEN)
 320                         free((char *)history[i].u.open.pathname);
 321         }
 322         free(history);
 323 }
 324
 325 static NORETURN void failtest_cleanup(bool forced_cleanup, int status)
 326 {
 327         int i;
 328
 329         /* For children, we don't care if they "failed" the testing. */
 330         if (control_fd != -1)
 331                 status = 0;
 332
 333         if (forced_cleanup)
 334                 history_num--;
 335
 336         /* Cleanup everything, in reverse order. */
 337         for (i = history_num - 1; i >= 0; i--) {
 338                 if (!history[i].cleanup)
 339                         continue;
 340                 if (!forced_cleanup) {
 341                         printf("Leak at %s:%u: --failpath=%s\n",
 342                                history[i].file, history[i].line,
 343                                failpath_string());
 344                         status = 1;
 345                 }
 346                 history[i].cleanup(&history[i].u);
 347         }
 348
 349         free_everything();
 350         tell_parent(SUCCESS);
 351         exit(status);
 352 }
 353
 354 static bool should_fail(struct failtest_call *call)
 355 {
 356         int status;
 357         int control[2], output[2];
 358         enum info_type type = UNEXPECTED;
 359         char *out = NULL;
 360         size_t outlen = 0;
 361         struct saved_file *files;
 362
 363         /* Are we probing? */
 364         if (probe_count && --probe_count == 0)
 365                 failtest_cleanup(true, 0);
 366
 367         if (call == &unrecorded_call)
 368                 return false;
 369
 370         if (failpath) {
 371                 /* + means continue after end, like normal. */
 372                 if (*failpath == '+')
 373                         failpath = NULL;
 374                 else if (*failpath == '\0') {
 375                         /* Continue, but don't inject errors. */
 376                         return call->fail = false;
 377                 } else {
 378                         if (tolower((unsigned char)*failpath)
 379                             != info_to_arg[call->type])
 380                                 errx(1, "Failpath expected '%c' got '%c'\n",
 381                                      info_to_arg[call->type], *failpath);
 382                         call->fail = isupper((unsigned char)*(failpath++));
 383                         return call->fail;
 384                 }
 385         }
 386
 387         /* Attach debugger if they asked for it. */
 388         if (debugpath && history_num == strlen(debugpath)) {
 389                 unsigned int i;
 390
 391                 for (i = 0; i < history_num; i++) {
 392                         unsigned char c = info_to_arg[history[i].type];
 393                         if (history[i].fail)
 394                                 c = toupper(c);
 395                         if (c != debugpath[i])
 396                                 break;
 397                 }
 398                 if (i == history_num) {
 399                         char str[80];
 400
 401                         /* Don't timeout. */
 402                         signal(SIGUSR1, SIG_IGN);
 403                         sprintf(str, "xterm -e gdb /proc/%d/exe %d &",
 404                                 getpid(), getpid());
 405                         if (system(str) == 0)
 406                                 sleep(5);
 407                 }
 408         }
 409
 410         if (failtest_hook) {
 411                 switch (failtest_hook(history, history_num)) {
 412                 case FAIL_OK:
 413                         break;
 414                 case FAIL_DONT_FAIL:
 415                         call->fail = false;
 416                         return false;
 417                 case FAIL_PROBE:
 418                         /* Already down probe path?  Stop now. */
 419                         if (probe_count)
 420                                 failtest_cleanup(true, 0);
 421                         /* FIXME: We should run *parent* and run probe until
 422                          * calls match up again. */
 423                         probe_count = 3;
 424                         break;
 425                 default:
 426                         abort();
 427                 }
 428         }
 429
 430         files = save_files();
 431
 432         /* We're going to fail in the child. */
 433         call->fail = true;
 434         if (pipe(control) != 0 || pipe(output) != 0)
 435                 err(1, "opening pipe");
 436
 437         /* Prevent double-printing (in child and parent) */
 438         fflush(stdout);
 439         child = fork();
 440         if (child == -1)
 441                 err(1, "forking failed");
 442
 443         if (child == 0) {
 444                 if (tracefd != -1) {
 445                         struct timeval now;
 446                         const char *p;
 447                         gettimeofday(&now, NULL);
 448                         if (now.tv_usec < start.tv_usec) {
 449                                 now.tv_sec--;
 450                                 now.tv_usec += 1000000;
 451                         }
 452                         now.tv_usec -= start.tv_usec;
 453                         now.tv_sec -= start.tv_sec;
 454                         p = failpath_string();
 455                         trace("%u->%u (%u.%02u): %s (", getppid(), getpid(),
 456                               (int)now.tv_sec, (int)now.tv_usec / 10000, p);
 457                         free((char *)p);
 458                         p = strrchr(history[history_num-1].file, '/');
 459                         if (p)
 460                                 trace("%s", p+1);
 461                         else
 462                                 trace("%s", history[history_num-1].file);
 463                         trace(":%u)\n", history[history_num-1].line);
 464                 }
 465                 close(control[0]);
 466                 close(output[0]);
 467                 dup2(output[1], STDOUT_FILENO);
 468                 dup2(output[1], STDERR_FILENO);
 469                 if (output[1] != STDOUT_FILENO && output[1] != STDERR_FILENO)
 470                         close(output[1]);
 471                 control_fd = control[1];
 472                 return true;
 473         }
 474
 475         signal(SIGUSR1, hand_down);
 476
 477         close(control[1]);
 478         close(output[1]);
 479
 480         /* We grab output so we can display it; we grab writes so we
 481          * can compare. */
 482         do {
 483                 struct pollfd pfd[2];
 484                 int ret;
 485
 486                 pfd[0].fd = output[0];
 487                 pfd[0].events = POLLIN|POLLHUP;
 488                 pfd[1].fd = control[0];
 489                 pfd[1].events = POLLIN|POLLHUP;
 490
 491                 if (type == SUCCESS)
 492                         ret = poll(pfd, 1, failtest_timeout_ms);
 493                 else
 494                         ret = poll(pfd, 2, failtest_timeout_ms);
 495
 496                 if (ret == 0)
 497                         hand_down(SIGUSR1);
 498                 if (ret < 0) {
 499                         if (errno == EINTR)
 500                                 continue;
 501                         err(1, "Poll returned %i", ret);
 502                 }
 503
 504                 if (pfd[0].revents & POLLIN) {
 505                         ssize_t len;
 506
 507                         out = realloc(out, outlen + 8192);
 508                         len = read(output[0], out + outlen, 8192);
 509                         outlen += len;
 510                 } else if (type != SUCCESS && (pfd[1].revents & POLLIN)) {
 511                         if (read_all(control[0], &type, sizeof(type))) {
 512                                 if (type == WRITE) {
 513                                         if (!read_write_info(control[0]))
 514                                                 break;
 515                                 } else if (type == RELEASE_LOCKS) {
 516                                         release_locks();
 517                                         /* FIXME: Tell them we're done... */
 518                                 }
 519                         }
 520                 } else if (pfd[0].revents & POLLHUP) {
 521                         break;
 522                 }
 523         } while (type != FAILURE);
 524
 525         close(output[0]);
 526         close(control[0]);
 527         waitpid(child, &status, 0);
 528         if (!WIFEXITED(status)) {
 529                 if (WTERMSIG(status) == SIGUSR1)
 530                         child_fail(out, outlen, "Timed out");
 531                 else
 532                         child_fail(out, outlen, "Killed by signal %u: ",
 533                                    WTERMSIG(status));
 534         }
 535         /* Child printed failure already, just pass up exit code. */
 536         if (type == FAILURE) {
 537                 fprintf(stderr, "%.*s", (int)outlen, out);
 538                 tell_parent(type);
 539                 exit(WEXITSTATUS(status) ? WEXITSTATUS(status) : 1);
 540         }
 541         if (WEXITSTATUS(status) != 0)
 542                 child_fail(out, outlen, "Exited with status %i: ",
 543                            WEXITSTATUS(status));
 544
 545         free(out);
 546         signal(SIGUSR1, SIG_DFL);
 547
 548         restore_files(files);
 549
 550         /* We continue onwards without failing. */
 551         call->fail = false;
 552         return false;
 553 }
 554
 555 static void cleanup_calloc(struct calloc_call *call)
 556 {
 557         free(call->ret);
 558 }
 559
 560 void *failtest_calloc(size_t nmemb, size_t size,
 561                       const char *file, unsigned line)
 562 {
 563         struct failtest_call *p;
 564         struct calloc_call call;
 565         call.nmemb = nmemb;
 566         call.size = size;
 567         p = add_history(FAILTEST_CALLOC, file, line, &call);
 568
 569         if (should_fail(p)) {
 570                 p->u.calloc.ret = NULL;
 571                 p->error = ENOMEM;
 572         } else {
 573                 p->u.calloc.ret = calloc(nmemb, size);
 574                 set_cleanup(p, cleanup_calloc, struct calloc_call);
 575         }
 576         errno = p->error;
 577         return p->u.calloc.ret;
 578 }
 579
 580 static void cleanup_malloc(struct malloc_call *call)
 581 {
 582         free(call->ret);
 583 }
 584
 585 void *failtest_malloc(size_t size, const char *file, unsigned line)
 586 {
 587         struct failtest_call *p;
 588         struct malloc_call call;
 589         call.size = size;
 590
 591         p = add_history(FAILTEST_MALLOC, file, line, &call);
 592         if (should_fail(p)) {
 593                 p->u.calloc.ret = NULL;
 594                 p->error = ENOMEM;
 595         } else {
 596                 p->u.calloc.ret = malloc(size);
 597                 set_cleanup(p, cleanup_malloc, struct malloc_call);
 598         }
 599         errno = p->error;
 600         return p->u.calloc.ret;
 601 }
 602
 603 static void cleanup_realloc(struct realloc_call *call)
 604 {
 605         free(call->ret);
 606 }
 607
 608 /* Walk back and find out if we got this ptr from a previous routine. */
 609 static void fixup_ptr_history(void *ptr, unsigned int last)
 610 {
 611         int i;
 612
 613         /* Start at end of history, work back. */
 614         for (i = last - 1; i >= 0; i--) {
 615                 switch (history[i].type) {
 616                 case FAILTEST_REALLOC:
 617                         if (history[i].u.realloc.ret == ptr) {
 618                                 history[i].cleanup = NULL;
 619                                 return;
 620                         }
 621                         break;
 622                 case FAILTEST_MALLOC:
 623                         if (history[i].u.malloc.ret == ptr) {
 624                                 history[i].cleanup = NULL;
 625                                 return;
 626                         }
 627                         break;
 628                 case FAILTEST_CALLOC:
 629                         if (history[i].u.calloc.ret == ptr) {
 630                                 history[i].cleanup = NULL;
 631                                 return;
 632                         }
 633                         break;
 634                 default:
 635                         break;
 636                 }
 637         }
 638 }
 639
 640 void *failtest_realloc(void *ptr, size_t size, const char *file, unsigned line)
 641 {
 642         struct failtest_call *p;
 643         struct realloc_call call;
 644         call.size = size;
 645         p = add_history(FAILTEST_REALLOC, file, line, &call);
 646
 647         /* FIXME: Try one child moving allocation, one not. */
 648         if (should_fail(p)) {
 649                 p->u.realloc.ret = NULL;
 650                 p->error = ENOMEM;
 651         } else {
 652                 fixup_ptr_history(ptr, history_num-1);
 653                 p->u.realloc.ret = realloc(ptr, size);
 654                 set_cleanup(p, cleanup_realloc, struct realloc_call);
 655         }
 656         errno = p->error;
 657         return p->u.realloc.ret;
 658 }
 659
 660 void failtest_free(void *ptr)
 661 {
 662         fixup_ptr_history(ptr, history_num);
 663         free(ptr);
 664 }
 665
 666 static void cleanup_open(struct open_call *call)
 667 {
 668         close(call->ret);
 669 }
 670
 671 int failtest_open(const char *pathname,
 672                   const char *file, unsigned line, ...)
 673 {
 674         struct failtest_call *p;
 675         struct open_call call;
 676         va_list ap;
 677
 678         call.pathname = strdup(pathname);
 679         va_start(ap, line);
 680         call.flags = va_arg(ap, int);
 681         if (call.flags & O_CREAT) {
 682                 call.mode = va_arg(ap, int);
 683                 va_end(ap);
 684         }
 685         p = add_history(FAILTEST_OPEN, file, line, &call);
 686         /* Avoid memory leak! */
 687         if (p == &unrecorded_call)
 688                 free((char *)call.pathname);
 689         p->u.open.ret = open(pathname, call.flags, call.mode);
 690
 691         if (!failpath && p->u.open.ret == -1) {
 692                 p->fail = false;
 693                 p->error = errno;
 694         } else if (should_fail(p)) {
 695                 close(p->u.open.ret);
 696                 p->u.open.ret = -1;
 697                 /* FIXME: Play with error codes? */
 698                 p->error = EACCES;
 699         } else {
 700                 set_cleanup(p, cleanup_open, struct open_call);
 701         }
 702         errno = p->error;
 703         return p->u.open.ret;
 704 }
 705
 706 static void cleanup_pipe(struct pipe_call *call)
 707 {
 708         if (!call->closed[0])
 709                 close(call->fds[0]);
 710         if (!call->closed[1])
 711                 close(call->fds[1]);
 712 }
 713
 714 int failtest_pipe(int pipefd[2], const char *file, unsigned line)
 715 {
 716         struct failtest_call *p;
 717         struct pipe_call call;
 718
 719         p = add_history(FAILTEST_PIPE, file, line, &call);
 720         if (should_fail(p)) {
 721                 p->u.open.ret = -1;
 722                 /* FIXME: Play with error codes? */
 723                 p->error = EMFILE;
 724         } else {
 725                 p->u.pipe.ret = pipe(p->u.pipe.fds);
 726                 p->u.pipe.closed[0] = p->u.pipe.closed[1] = false;
 727                 set_cleanup(p, cleanup_pipe, struct pipe_call);
 728         }
 729         /* This causes valgrind to notice if they use pipefd[] after failure */
 730         memcpy(pipefd, p->u.pipe.fds, sizeof(p->u.pipe.fds));
 731         errno = p->error;
 732         return p->u.pipe.ret;
 733 }
 734
 735 ssize_t failtest_pread(int fd, void *buf, size_t count, off_t off,
 736                        const char *file, unsigned line)
 737 {
 738         struct failtest_call *p;
 739         struct read_call call;
 740         call.fd = fd;
 741         call.buf = buf;
 742         call.count = count;
 743         call.off = off;
 744         p = add_history(FAILTEST_READ, file, line, &call);
 745
 746         /* FIXME: Try partial read returns. */
 747         if (should_fail(p)) {
 748                 p->u.read.ret = -1;
 749                 p->error = EIO;
 750         } else {
 751                 p->u.read.ret = pread(fd, buf, count, off);
 752         }
 753         errno = p->error;
 754         return p->u.read.ret;
 755 }
 756
 757 ssize_t failtest_pwrite(int fd, const void *buf, size_t count, off_t off,
 758                         const char *file, unsigned line)
 759 {
 760         struct failtest_call *p;
 761         struct write_call call;
 762
 763         call.fd = fd;
 764         call.buf = buf;
 765         call.count = count;
 766         call.off = off;
 767         p = add_history(FAILTEST_WRITE, file, line, &call);
 768
 769         /* If we're a child, we need to make sure we write the same thing
 770          * to non-files as the parent does, so tell it. */
 771         if (control_fd != -1 && off == (off_t)-1) {
 772                 enum info_type type = WRITE;
 773
 774                 write_all(control_fd, &type, sizeof(type));
 775                 write_all(control_fd, &p->u.write, sizeof(p->u.write));
 776                 write_all(control_fd, buf, count);
 777         }
 778
 779         /* FIXME: Try partial write returns. */
 780         if (should_fail(p)) {
 781                 p->u.write.ret = -1;
 782                 p->error = EIO;
 783         } else {
 784                 /* FIXME: We assume same write order in parent and child */
 785                 if (off == (off_t)-1 && child_writes_num != 0) {
 786                         if (child_writes[0].fd != fd)
 787                                 errx(1, "Child wrote to fd %u, not %u?",
 788                                      child_writes[0].fd, fd);
 789                         if (child_writes[0].off != p->u.write.off)
 790                                 errx(1, "Child wrote to offset %zu, not %zu?",
 791                                      (size_t)child_writes[0].off,
 792                                      (size_t)p->u.write.off);
 793                         if (child_writes[0].count != count)
 794                                 errx(1, "Child wrote length %zu, not %zu?",
 795                                      child_writes[0].count, count);
 796                         if (memcmp(child_writes[0].buf, buf, count)) {
 797                                 child_fail(NULL, 0,
 798                                            "Child wrote differently to"
 799                                            " fd %u than we did!\n", fd);
 800                         }
 801                         free((char *)child_writes[0].buf);
 802                         child_writes_num--;
 803                         memmove(&child_writes[0], &child_writes[1],
 804                                 sizeof(child_writes[0]) * child_writes_num);
 805
 806                         /* Is this is a socket or pipe, child wrote it
 807                            already. */
 808                         if (p->u.write.off == (off_t)-1) {
 809                                 p->u.write.ret = count;
 810                                 errno = p->error;
 811                                 return p->u.write.ret;
 812                         }
 813                 }
 814                 p->u.write.ret = pwrite(fd, buf, count, off);
 815         }
 816         errno = p->error;
 817         return p->u.write.ret;
 818 }
 819
 820 ssize_t failtest_read(int fd, void *buf, size_t count,
 821                       const char *file, unsigned line)
 822 {
 823         return failtest_pread(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 824                               file, line);
 825 }
 826
 827 ssize_t failtest_write(int fd, const void *buf, size_t count,
 828                        const char *file, unsigned line)
 829 {
 830         return failtest_pwrite(fd, buf, count, lseek(fd, 0, SEEK_CUR),
 831                                file, line);
 832 }
 833
 834 static struct lock_info *WARN_UNUSED_RESULT
 835 add_lock(struct lock_info *locks, int fd, off_t start, off_t end, int type)
 836 {
 837         unsigned int i;
 838         struct lock_info *l;
 839
 840         for (i = 0; i < lock_num; i++) {
 841                 l = &locks[i];
 842
 843                 if (l->fd != fd)
 844                         continue;
 845                 /* Four cases we care about:
 846                  * Start overlap:
 847                  *      l =    |      |
 848                  *      new = |   |
 849                  * Mid overlap:
 850                  *      l =    |      |
 851                  *      new =    |  |
 852                  * End overlap:
 853                  *      l =    |      |
 854                  *      new =      |    |
 855                  * Total overlap:
 856                  *      l =    |      |
 857                  *      new = |         |
 858                  */
 859                 if (start > l->start && end < l->end) {
 860                         /* Mid overlap: trim entry, add new one. */
 861                         off_t new_start, new_end;
 862                         new_start = end + 1;
 863                         new_end = l->end;
 864                         l->end = start - 1;
 865                         locks = add_lock(locks,
 866                                          fd, new_start, new_end, l->type);
 867                         l = &locks[i];
 868                 } else if (start <= l->start && end >= l->end) {
 869                         /* Total overlap: eliminate entry. */
 870                         l->end = 0;
 871                         l->start = 1;
 872                 } else if (end >= l->start && end < l->end) {
 873                         /* Start overlap: trim entry. */
 874                         l->start = end + 1;
 875                 } else if (start > l->start && start <= l->end) {
 876                         /* End overlap: trim entry. */
 877                         l->end = start-1;
 878                 }
 879                 /* Nothing left?  Remove it. */
 880                 if (l->end < l->start) {
 881                         memmove(l, l + 1, (--lock_num - i) * sizeof(l[0]));
 882                         i--;
 883                 }
 884         }
 885
 886         if (type != F_UNLCK) {
 887                 locks = realloc(locks, (lock_num + 1) * sizeof(*locks));
 888                 l = &locks[lock_num++];
 889                 l->fd = fd;
 890                 l->start = start;
 891                 l->end = end;
 892                 l->type = type;
 893         }
 894         return locks;
 895 }
 896
 897 /* We trap this so we can record it: we don't fail it. */
 898 int failtest_close(int fd, const char *file, unsigned line)
 899 {
 900         int i;
 901         struct close_call call;
 902         struct failtest_call *p;
 903
 904         call.fd = fd;
 905         p = add_history(FAILTEST_CLOSE, file, line, &call);
 906         p->fail = false;
 907
 908         /* Consume close from failpath. */
 909         if (failpath)
 910                 if (should_fail(p))
 911                         abort();
 912
 913         if (fd < 0)
 914                 return close(fd);
 915
 916         /* Trace history to find source of fd. */
 917         for (i = history_num-1; i >= 0; i--) {
 918                 switch (history[i].type) {
 919                 case FAILTEST_PIPE:
 920                         /* From a pipe? */
 921                         if (history[i].u.pipe.fds[0] == fd) {
 922                                 assert(!history[i].u.pipe.closed[0]);
 923                                 history[i].u.pipe.closed[0] = true;
 924                                 if (history[i].u.pipe.closed[1])
 925                                         history[i].cleanup = NULL;
 926                                 goto out;
 927                         }
 928                         if (history[i].u.pipe.fds[1] == fd) {
 929                                 assert(!history[i].u.pipe.closed[1]);
 930                                 history[i].u.pipe.closed[1] = true;
 931                                 if (history[i].u.pipe.closed[0])
 932                                         history[i].cleanup = NULL;
 933                                 goto out;
 934                         }
 935                         break;
 936                 case FAILTEST_OPEN:
 937                         if (history[i].u.open.ret == fd) {
 938                                 assert((void *)history[i].cleanup
 939                                        == (void *)cleanup_open);
 940                                 history[i].cleanup = NULL;
 941                                 goto out;
 942                         }
 943                         break;
 944                 default:
 945                         break;
 946                 }
 947         }
 948
 949 out:
 950         locks = add_lock(locks, fd, 0, off_max(), F_UNLCK);
 951         return close(fd);
 952 }
 953
 954 /* Zero length means "to end of file" */
 955 static off_t end_of(off_t start, off_t len)
 956 {
 957         if (len == 0)
 958                 return off_max();
 959         return start + len - 1;
 960 }
 961
 962 /* FIXME: This only handles locks, really. */
 963 int failtest_fcntl(int fd, const char *file, unsigned line, int cmd, ...)
 964 {
 965         struct failtest_call *p;
 966         struct fcntl_call call;
 967         va_list ap;
 968
 969         call.fd = fd;
 970         call.cmd = cmd;
 971
 972         /* Argument extraction. */
 973         switch (cmd) {
 974         case F_SETFL:
 975         case F_SETFD:
 976                 va_start(ap, cmd);
 977                 call.arg.l = va_arg(ap, long);
 978                 va_end(ap);
 979                 return fcntl(fd, cmd, call.arg.l);
 980         case F_GETFD:
 981         case F_GETFL:
 982                 return fcntl(fd, cmd);
 983         case F_GETLK:
 984                 get_locks();
 985                 va_start(ap, cmd);
 986                 call.arg.fl = *va_arg(ap, struct flock *);
 987                 va_end(ap);
 988                 return fcntl(fd, cmd, &call.arg.fl);
 989         case F_SETLK:
 990         case F_SETLKW:
 991                 va_start(ap, cmd);
 992                 call.arg.fl = *va_arg(ap, struct flock *);
 993                 va_end(ap);
 994                 break;
 995         default:
 996                 /* This means you need to implement it here. */
 997                 err(1, "failtest: unknown fcntl %u", cmd);
 998         }
 999
1000         p = add_history(FAILTEST_FCNTL, file, line, &call);
1001
1002         if (should_fail(p)) {
1003                 p->u.fcntl.ret = -1;
1004                 if (p->u.fcntl.cmd == F_SETLK)
1005                         p->error = EAGAIN;
1006                 else
1007                         p->error = EDEADLK;
1008         } else {
1009                 get_locks();
1010                 p->u.fcntl.ret = fcntl(p->u.fcntl.fd, p->u.fcntl.cmd,
1011                                        &p->u.fcntl.arg.fl);
1012                 if (p->u.fcntl.ret == -1)
1013                         p->error = errno;
1014                 else {
1015                         /* We don't handle anything else yet. */
1016                         assert(p->u.fcntl.arg.fl.l_whence == SEEK_SET);
1017                         locks = add_lock(locks,
1018                                          p->u.fcntl.fd,
1019                                          p->u.fcntl.arg.fl.l_start,
1020                                          end_of(p->u.fcntl.arg.fl.l_start,
1021                                                 p->u.fcntl.arg.fl.l_len),
1022                                          p->u.fcntl.arg.fl.l_type);
1023                 }
1024         }
1025         errno = p->error;
1026         return p->u.fcntl.ret;
1027 }
1028
1029 pid_t failtest_getpid(const char *file, unsigned line)
1030 {
1031         /* You must call failtest_init first! */
1032         assert(orig_pid);
1033         return orig_pid;
1034 }
1035
1036 void failtest_init(int argc, char *argv[])
1037 {
1038         unsigned int i;
1039
1040         orig_pid = getpid();
1041
1042         for (i = 1; i < argc; i++) {
1043                 if (!strncmp(argv[i], "--failpath=", strlen("--failpath="))) {
1044                         failpath = argv[i] + strlen("--failpath=");
1045                 } else if (strcmp(argv[i], "--tracepath") == 0) {
1046                         tracefd = dup(STDERR_FILENO);
1047                         failtest_timeout_ms = -1;
1048                 } else if (!strncmp(argv[i], "--debugpath=",
1049                                     strlen("--debugpath="))) {
1050                         debugpath = argv[i] + strlen("--debugpath=");
1051                 }
1052         }
1053         gettimeofday(&start, NULL);
1054 }
1055
1056 void failtest_exit(int status)
1057 {
1058         if (failtest_exit_check) {
1059                 if (!failtest_exit_check(history, history_num))
1060                         child_fail(NULL, 0, "failtest_exit_check failed\n");
1061         }
1062
1063         failtest_cleanup(false, status);
1064 }