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