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

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