src/basic/fd-util.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2 /***
   3   This file is part of systemd.
   4
   5   Copyright 2010 Lennart Poettering
   6 ***/
   7
   8 #include <errno.h>
   9 #include <fcntl.h>
  10 #include <sys/resource.h>
  11 #include <sys/socket.h>
  12 #include <sys/stat.h>
  13 #include <unistd.h>
  14
  15 #include "alloc-util.h"
  16 #include "copy.h"
  17 #include "dirent-util.h"
  18 #include "fd-util.h"
  19 #include "fileio.h"
  20 #include "fs-util.h"
  21 #include "io-util.h"
  22 #include "macro.h"
  23 #include "memfd-util.h"
  24 #include "missing.h"
  25 #include "parse-util.h"
  26 #include "path-util.h"
  27 #include "process-util.h"
  28 #include "socket-util.h"
  29 #include "stdio-util.h"
  30 #include "util.h"
  31
  32 int close_nointr(int fd) {
  33         assert(fd >= 0);
  34
  35         if (close(fd) >= 0)
  36                 return 0;
  37
  38         /*
  39          * Just ignore EINTR; a retry loop is the wrong thing to do on
  40          * Linux.
  41          *
  42          * http://lkml.indiana.edu/hypermail/linux/kernel/0509.1/0877.html
  43          * https://bugzilla.gnome.org/show_bug.cgi?id=682819
  44          * http://utcc.utoronto.ca/~cks/space/blog/unix/CloseEINTR
  45          * https://sites.google.com/site/michaelsafyan/software-engineering/checkforeintrwheninvokingclosethinkagain
  46          */
  47         if (errno == EINTR)
  48                 return 0;
  49
  50         return -errno;
  51 }
  52
  53 int safe_close(int fd) {
  54
  55         /*
  56          * Like close_nointr() but cannot fail. Guarantees errno is
  57          * unchanged. Is a NOP with negative fds passed, and returns
  58          * -1, so that it can be used in this syntax:
  59          *
  60          * fd = safe_close(fd);
  61          */
  62
  63         if (fd >= 0) {
  64                 PROTECT_ERRNO;
  65
  66                 /* The kernel might return pretty much any error code
  67                  * via close(), but the fd will be closed anyway. The
  68                  * only condition we want to check for here is whether
  69                  * the fd was invalid at all... */
  70
  71                 assert_se(close_nointr(fd) != -EBADF);
  72         }
  73
  74         return -1;
  75 }
  76
  77 void safe_close_pair(int p[]) {
  78         assert(p);
  79
  80         if (p[0] == p[1]) {
  81                 /* Special case pairs which use the same fd in both
  82                  * directions... */
  83                 p[0] = p[1] = safe_close(p[0]);
  84                 return;
  85         }
  86
  87         p[0] = safe_close(p[0]);
  88         p[1] = safe_close(p[1]);
  89 }
  90
  91 void close_many(const int fds[], size_t n_fd) {
  92         size_t i;
  93
  94         assert(fds || n_fd <= 0);
  95
  96         for (i = 0; i < n_fd; i++)
  97                 safe_close(fds[i]);
  98 }
  99
 100 int fclose_nointr(FILE *f) {
 101         assert(f);
 102
 103         /* Same as close_nointr(), but for fclose() */
 104
 105         if (fclose(f) == 0)
 106                 return 0;
 107
 108         if (errno == EINTR)
 109                 return 0;
 110
 111         return -errno;
 112 }
 113
 114 FILE* safe_fclose(FILE *f) {
 115
 116         /* Same as safe_close(), but for fclose() */
 117
 118         if (f) {
 119                 PROTECT_ERRNO;
 120
 121                 assert_se(fclose_nointr(f) != EBADF);
 122         }
 123
 124         return NULL;
 125 }
 126
 127 DIR* safe_closedir(DIR *d) {
 128
 129         if (d) {
 130                 PROTECT_ERRNO;
 131
 132                 assert_se(closedir(d) >= 0 || errno != EBADF);
 133         }
 134
 135         return NULL;
 136 }
 137
 138 int fd_nonblock(int fd, bool nonblock) {
 139         int flags, nflags;
 140
 141         assert(fd >= 0);
 142
 143         flags = fcntl(fd, F_GETFL, 0);
 144         if (flags < 0)
 145                 return -errno;
 146
 147         if (nonblock)
 148                 nflags = flags | O_NONBLOCK;
 149         else
 150                 nflags = flags & ~O_NONBLOCK;
 151
 152         if (nflags == flags)
 153                 return 0;
 154
 155         if (fcntl(fd, F_SETFL, nflags) < 0)
 156                 return -errno;
 157
 158         return 0;
 159 }
 160
 161 int fd_cloexec(int fd, bool cloexec) {
 162         int flags, nflags;
 163
 164         assert(fd >= 0);
 165
 166         flags = fcntl(fd, F_GETFD, 0);
 167         if (flags < 0)
 168                 return -errno;
 169
 170         if (cloexec)
 171                 nflags = flags | FD_CLOEXEC;
 172         else
 173                 nflags = flags & ~FD_CLOEXEC;
 174
 175         if (nflags == flags)
 176                 return 0;
 177
 178         if (fcntl(fd, F_SETFD, nflags) < 0)
 179                 return -errno;
 180
 181         return 0;
 182 }
 183
 184 _pure_ static bool fd_in_set(int fd, const int fdset[], size_t n_fdset) {
 185         size_t i;
 186
 187         assert(n_fdset == 0 || fdset);
 188
 189         for (i = 0; i < n_fdset; i++)
 190                 if (fdset[i] == fd)
 191                         return true;
 192
 193         return false;
 194 }
 195
 196 int close_all_fds(const int except[], size_t n_except) {
 197         _cleanup_closedir_ DIR *d = NULL;
 198         struct dirent *de;
 199         int r = 0;
 200
 201         assert(n_except == 0 || except);
 202
 203         d = opendir("/proc/self/fd");
 204         if (!d) {
 205                 struct rlimit rl;
 206                 int fd, max_fd;
 207
 208                 /* When /proc isn't available (for example in chroots) the fallback is brute forcing through the fd
 209                  * table */
 210
 211                 assert_se(getrlimit(RLIMIT_NOFILE, &rl) >= 0);
 212
 213                 if (rl.rlim_max == 0)
 214                         return -EINVAL;
 215
 216                 /* Let's take special care if the resource limit is set to unlimited, or actually larger than the range
 217                  * of 'int'. Let's avoid implicit overflows. */
 218                 max_fd = (rl.rlim_max == RLIM_INFINITY || rl.rlim_max > INT_MAX) ? INT_MAX : (int) (rl.rlim_max - 1);
 219
 220                 for (fd = 3; fd >= 0; fd = fd < max_fd ? fd + 1 : -1) {
 221                         int q;
 222
 223                         if (fd_in_set(fd, except, n_except))
 224                                 continue;
 225
 226                         q = close_nointr(fd);
 227                         if (q < 0 && q != -EBADF && r >= 0)
 228                                 r = q;
 229                 }
 230
 231                 return r;
 232         }
 233
 234         FOREACH_DIRENT(de, d, return -errno) {
 235                 int fd = -1, q;
 236
 237                 if (safe_atoi(de->d_name, &fd) < 0)
 238                         /* Let's better ignore this, just in case */
 239                         continue;
 240
 241                 if (fd < 3)
 242                         continue;
 243
 244                 if (fd == dirfd(d))
 245                         continue;
 246
 247                 if (fd_in_set(fd, except, n_except))
 248                         continue;
 249
 250                 q = close_nointr(fd);
 251                 if (q < 0 && q != -EBADF && r >= 0) /* Valgrind has its own FD and doesn't want to have it closed */
 252                         r = q;
 253         }
 254
 255         return r;
 256 }
 257
 258 int same_fd(int a, int b) {
 259         struct stat sta, stb;
 260         pid_t pid;
 261         int r, fa, fb;
 262
 263         assert(a >= 0);
 264         assert(b >= 0);
 265
 266         /* Compares two file descriptors. Note that semantics are
 267          * quite different depending on whether we have kcmp() or we
 268          * don't. If we have kcmp() this will only return true for
 269          * dup()ed file descriptors, but not otherwise. If we don't
 270          * have kcmp() this will also return true for two fds of the same
 271          * file, created by separate open() calls. Since we use this
 272          * call mostly for filtering out duplicates in the fd store
 273          * this difference hopefully doesn't matter too much. */
 274
 275         if (a == b)
 276                 return true;
 277
 278         /* Try to use kcmp() if we have it. */
 279         pid = getpid_cached();
 280         r = kcmp(pid, pid, KCMP_FILE, a, b);
 281         if (r == 0)
 282                 return true;
 283         if (r > 0)
 284                 return false;
 285         if (errno != ENOSYS)
 286                 return -errno;
 287
 288         /* We don't have kcmp(), use fstat() instead. */
 289         if (fstat(a, &sta) < 0)
 290                 return -errno;
 291
 292         if (fstat(b, &stb) < 0)
 293                 return -errno;
 294
 295         if ((sta.st_mode & S_IFMT) != (stb.st_mode & S_IFMT))
 296                 return false;
 297
 298         /* We consider all device fds different, since two device fds
 299          * might refer to quite different device contexts even though
 300          * they share the same inode and backing dev_t. */
 301
 302         if (S_ISCHR(sta.st_mode) || S_ISBLK(sta.st_mode))
 303                 return false;
 304
 305         if (sta.st_dev != stb.st_dev || sta.st_ino != stb.st_ino)
 306                 return false;
 307
 308         /* The fds refer to the same inode on disk, let's also check
 309          * if they have the same fd flags. This is useful to
 310          * distinguish the read and write side of a pipe created with
 311          * pipe(). */
 312         fa = fcntl(a, F_GETFL);
 313         if (fa < 0)
 314                 return -errno;
 315
 316         fb = fcntl(b, F_GETFL);
 317         if (fb < 0)
 318                 return -errno;
 319
 320         return fa == fb;
 321 }
 322
 323 void cmsg_close_all(struct msghdr *mh) {
 324         struct cmsghdr *cmsg;
 325
 326         assert(mh);
 327
 328         CMSG_FOREACH(cmsg, mh)
 329                 if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS)
 330                         close_many((int*) CMSG_DATA(cmsg), (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int));
 331 }
 332
 333 bool fdname_is_valid(const char *s) {
 334         const char *p;
 335
 336         /* Validates a name for $LISTEN_FDNAMES. We basically allow
 337          * everything ASCII that's not a control character. Also, as
 338          * special exception the ":" character is not allowed, as we
 339          * use that as field separator in $LISTEN_FDNAMES.
 340          *
 341          * Note that the empty string is explicitly allowed
 342          * here. However, we limit the length of the names to 255
 343          * characters. */
 344
 345         if (!s)
 346                 return false;
 347
 348         for (p = s; *p; p++) {
 349                 if (*p < ' ')
 350                         return false;
 351                 if (*p >= 127)
 352                         return false;
 353                 if (*p == ':')
 354                         return false;
 355         }
 356
 357         return p - s < 256;
 358 }
 359
 360 int fd_get_path(int fd, char **ret) {
 361         _cleanup_close_ int dir = -1;
 362         char fdname[DECIMAL_STR_MAX(int)];
 363         int r;
 364
 365         dir = open("/proc/self/fd/", O_CLOEXEC | O_DIRECTORY | O_PATH);
 366         if (dir < 0)
 367                 /* /proc is not available or not set up properly, we're most likely
 368                  * in some chroot environment. */
 369                 return errno == ENOENT ? -EOPNOTSUPP : -errno;
 370
 371         xsprintf(fdname, "%i", fd);
 372
 373         r = readlinkat_malloc(dir, fdname, ret);
 374         if (r == -ENOENT)
 375                 /* If the file doesn't exist the fd is invalid */
 376                 return -EBADF;
 377
 378         return r;
 379 }
 380
 381 int move_fd(int from, int to, int cloexec) {
 382         int r;
 383
 384         /* Move fd 'from' to 'to', make sure FD_CLOEXEC remains equal if requested, and release the old fd. If
 385          * 'cloexec' is passed as -1, the original FD_CLOEXEC is inherited for the new fd. If it is 0, it is turned
 386          * off, if it is > 0 it is turned on. */
 387
 388         if (from < 0)
 389                 return -EBADF;
 390         if (to < 0)
 391                 return -EBADF;
 392
 393         if (from == to) {
 394
 395                 if (cloexec >= 0) {
 396                         r = fd_cloexec(to, cloexec);
 397                         if (r < 0)
 398                                 return r;
 399                 }
 400
 401                 return to;
 402         }
 403
 404         if (cloexec < 0) {
 405                 int fl;
 406
 407                 fl = fcntl(from, F_GETFD, 0);
 408                 if (fl < 0)
 409                         return -errno;
 410
 411                 cloexec = !!(fl & FD_CLOEXEC);
 412         }
 413
 414         r = dup3(from, to, cloexec ? O_CLOEXEC : 0);
 415         if (r < 0)
 416                 return -errno;
 417
 418         assert(r == to);
 419
 420         safe_close(from);
 421
 422         return to;
 423 }
 424
 425 int acquire_data_fd(const void *data, size_t size, unsigned flags) {
 426
 427         _cleanup_close_pair_ int pipefds[2] = { -1, -1 };
 428         char pattern[] = "/dev/shm/data-fd-XXXXXX";
 429         _cleanup_close_ int fd = -1;
 430         int isz = 0, r;
 431         ssize_t n;
 432         off_t f;
 433
 434         assert(data || size == 0);
 435
 436         /* Acquire a read-only file descriptor that when read from returns the specified data. This is much more
 437          * complex than I wish it was. But here's why:
 438          *
 439          * a) First we try to use memfds. They are the best option, as we can seal them nicely to make them
 440          *    read-only. Unfortunately they require kernel 3.17, and – at the time of writing – we still support 3.14.
 441          *
 442          * b) Then, we try classic pipes. They are the second best options, as we can close the writing side, retaining
 443          *    a nicely read-only fd in the reading side. However, they are by default quite small, and unprivileged
 444          *    clients can only bump their size to a system-wide limit, which might be quite low.
 445          *
 446          * c) Then, we try an O_TMPFILE file in /dev/shm (that dir is the only suitable one known to exist from
 447          *    earliest boot on). To make it read-only we open the fd a second time with O_RDONLY via
 448          *    /proc/self/<fd>. Unfortunately O_TMPFILE is not available on older kernels on tmpfs.
 449          *
 450          * d) Finally, we try creating a regular file in /dev/shm, which we then delete.
 451          *
 452          * It sucks a bit that depending on the situation we return very different objects here, but that's Linux I
 453          * figure. */
 454
 455         if (size == 0 && ((flags & ACQUIRE_NO_DEV_NULL) == 0)) {
 456                 /* As a special case, return /dev/null if we have been called for an empty data block */
 457                 r = open("/dev/null", O_RDONLY|O_CLOEXEC|O_NOCTTY);
 458                 if (r < 0)
 459                         return -errno;
 460
 461                 return r;
 462         }
 463
 464         if ((flags & ACQUIRE_NO_MEMFD) == 0) {
 465                 fd = memfd_new("data-fd");
 466                 if (fd < 0)
 467                         goto try_pipe;
 468
 469                 n = write(fd, data, size);
 470                 if (n < 0)
 471                         return -errno;
 472                 if ((size_t) n != size)
 473                         return -EIO;
 474
 475                 f = lseek(fd, 0, SEEK_SET);
 476                 if (f != 0)
 477                         return -errno;
 478
 479                 r = memfd_set_sealed(fd);
 480                 if (r < 0)
 481                         return r;
 482
 483                 return TAKE_FD(fd);
 484         }
 485
 486 try_pipe:
 487         if ((flags & ACQUIRE_NO_PIPE) == 0) {
 488                 if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
 489                         return -errno;
 490
 491                 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
 492                 if (isz < 0)
 493                         return -errno;
 494
 495                 if ((size_t) isz < size) {
 496                         isz = (int) size;
 497                         if (isz < 0 || (size_t) isz != size)
 498                                 return -E2BIG;
 499
 500                         /* Try to bump the pipe size */
 501                         (void) fcntl(pipefds[1], F_SETPIPE_SZ, isz);
 502
 503                         /* See if that worked */
 504                         isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
 505                         if (isz < 0)
 506                                 return -errno;
 507
 508                         if ((size_t) isz < size)
 509                                 goto try_dev_shm;
 510                 }
 511
 512                 n = write(pipefds[1], data, size);
 513                 if (n < 0)
 514                         return -errno;
 515                 if ((size_t) n != size)
 516                         return -EIO;
 517
 518                 (void) fd_nonblock(pipefds[0], false);
 519
 520                 return TAKE_FD(pipefds[0]);
 521         }
 522
 523 try_dev_shm:
 524         if ((flags & ACQUIRE_NO_TMPFILE) == 0) {
 525                 fd = open("/dev/shm", O_RDWR|O_TMPFILE|O_CLOEXEC, 0500);
 526                 if (fd < 0)
 527                         goto try_dev_shm_without_o_tmpfile;
 528
 529                 n = write(fd, data, size);
 530                 if (n < 0)
 531                         return -errno;
 532                 if ((size_t) n != size)
 533                         return -EIO;
 534
 535                 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
 536                 return fd_reopen(fd, O_RDONLY|O_CLOEXEC);
 537         }
 538
 539 try_dev_shm_without_o_tmpfile:
 540         if ((flags & ACQUIRE_NO_REGULAR) == 0) {
 541                 fd = mkostemp_safe(pattern);
 542                 if (fd < 0)
 543                         return fd;
 544
 545                 n = write(fd, data, size);
 546                 if (n < 0) {
 547                         r = -errno;
 548                         goto unlink_and_return;
 549                 }
 550                 if ((size_t) n != size) {
 551                         r = -EIO;
 552                         goto unlink_and_return;
 553                 }
 554
 555                 /* Let's reopen the thing, in order to get an O_RDONLY fd for the original O_RDWR one */
 556                 r = open(pattern, O_RDONLY|O_CLOEXEC);
 557                 if (r < 0)
 558                         r = -errno;
 559
 560         unlink_and_return:
 561                 (void) unlink(pattern);
 562                 return r;
 563         }
 564
 565         return -EOPNOTSUPP;
 566 }
 567
 568 /* When the data is smaller or equal to 64K, try to place the copy in a memfd/pipe */
 569 #define DATA_FD_MEMORY_LIMIT (64U*1024U)
 570
 571 /* If memfd/pipe didn't work out, then let's use a file in /tmp up to a size of 1M. If it's large than that use /var/tmp instead. */
 572 #define DATA_FD_TMP_LIMIT (1024U*1024U)
 573
 574 int fd_duplicate_data_fd(int fd) {
 575
 576         _cleanup_close_ int copy_fd = -1, tmp_fd = -1;
 577         _cleanup_free_ void *remains = NULL;
 578         size_t remains_size = 0;
 579         const char *td;
 580         struct stat st;
 581         int r;
 582
 583         /* Creates a 'data' fd from the specified source fd, containing all the same data in a read-only fashion, but
 584          * independent of it (i.e. the source fd can be closed and unmounted after this call succeeded). Tries to be
 585          * somewhat smart about where to place the data. In the best case uses a memfd(). If memfd() are not supported
 586          * uses a pipe instead. For larger data will use an unlinked file in /tmp, and for even larger data one in
 587          * /var/tmp. */
 588
 589         if (fstat(fd, &st) < 0)
 590                 return -errno;
 591
 592         /* For now, let's only accept regular files, sockets, pipes and char devices */
 593         if (S_ISDIR(st.st_mode))
 594                 return -EISDIR;
 595         if (S_ISLNK(st.st_mode))
 596                 return -ELOOP;
 597         if (!S_ISREG(st.st_mode) && !S_ISSOCK(st.st_mode) && !S_ISFIFO(st.st_mode) && !S_ISCHR(st.st_mode))
 598                 return -EBADFD;
 599
 600         /* If we have reason to believe the data is bounded in size, then let's use memfds or pipes as backing fd. Note
 601          * that we use the reported regular file size only as a hint, given that there are plenty special files in
 602          * /proc and /sys which report a zero file size but can be read from. */
 603
 604         if (!S_ISREG(st.st_mode) || st.st_size < DATA_FD_MEMORY_LIMIT) {
 605
 606                 /* Try a memfd first */
 607                 copy_fd = memfd_new("data-fd");
 608                 if (copy_fd >= 0) {
 609                         off_t f;
 610
 611                         r = copy_bytes(fd, copy_fd, DATA_FD_MEMORY_LIMIT, 0);
 612                         if (r < 0)
 613                                 return r;
 614
 615                         f = lseek(copy_fd, 0, SEEK_SET);
 616                         if (f != 0)
 617                                 return -errno;
 618
 619                         if (r == 0) {
 620                                 /* Did it fit into the limit? If so, we are done. */
 621                                 r = memfd_set_sealed(copy_fd);
 622                                 if (r < 0)
 623                                         return r;
 624
 625                                 return TAKE_FD(copy_fd);
 626                         }
 627
 628                         /* Hmm, pity, this didn't fit. Let's fall back to /tmp then, see below */
 629
 630                 } else {
 631                         _cleanup_(close_pairp) int pipefds[2] = { -1, -1 };
 632                         int isz;
 633
 634                         /* If memfds aren't available, use a pipe. Set O_NONBLOCK so that we will get EAGAIN rather
 635                          * then block indefinitely when we hit the pipe size limit */
 636
 637                         if (pipe2(pipefds, O_CLOEXEC|O_NONBLOCK) < 0)
 638                                 return -errno;
 639
 640                         isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
 641                         if (isz < 0)
 642                                 return -errno;
 643
 644                         /* Try to enlarge the pipe size if necessary */
 645                         if ((size_t) isz < DATA_FD_MEMORY_LIMIT) {
 646
 647                                 (void) fcntl(pipefds[1], F_SETPIPE_SZ, DATA_FD_MEMORY_LIMIT);
 648
 649                                 isz = fcntl(pipefds[1], F_GETPIPE_SZ, 0);
 650                                 if (isz < 0)
 651                                         return -errno;
 652                         }
 653
 654                         if ((size_t) isz >= DATA_FD_MEMORY_LIMIT) {
 655
 656                                 r = copy_bytes_full(fd, pipefds[1], DATA_FD_MEMORY_LIMIT, 0, &remains, &remains_size);
 657                                 if (r < 0 && r != -EAGAIN)
 658                                         return r; /* If we get EAGAIN it could be because of the source or because of
 659                                                    * the destination fd, we can't know, as sendfile() and friends won't
 660                                                    * tell us. Hence, treat this as reason to fall back, just to be
 661                                                    * sure. */
 662                                 if (r == 0) {
 663                                         /* Everything fit in, yay! */
 664                                         (void) fd_nonblock(pipefds[0], false);
 665
 666                                         return TAKE_FD(pipefds[0]);
 667                                 }
 668
 669                                 /* Things didn't fit in. But we read data into the pipe, let's remember that, so that
 670                                  * when writing the new file we incorporate this first. */
 671                                 copy_fd = TAKE_FD(pipefds[0]);
 672                         }
 673                 }
 674         }
 675
 676         /* If we have reason to believe this will fit fine in /tmp, then use that as first fallback. */
 677         if ((!S_ISREG(st.st_mode) || st.st_size < DATA_FD_TMP_LIMIT) &&
 678             (DATA_FD_MEMORY_LIMIT + remains_size) < DATA_FD_TMP_LIMIT) {
 679                 off_t f;
 680
 681                 tmp_fd = open_tmpfile_unlinkable(NULL /* NULL as directory means /tmp */, O_RDWR|O_CLOEXEC);
 682                 if (tmp_fd < 0)
 683                         return tmp_fd;
 684
 685                 if (copy_fd >= 0) {
 686                         /* If we tried a memfd/pipe first and it ended up being too large, then copy this into the
 687                          * temporary file first. */
 688
 689                         r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, 0);
 690                         if (r < 0)
 691                                 return r;
 692
 693                         assert(r == 0);
 694                 }
 695
 696                 if (remains_size > 0) {
 697                         /* If there were remaining bytes (i.e. read into memory, but not written out yet) from the
 698                          * failed copy operation, let's flush them out next. */
 699
 700                         r = loop_write(tmp_fd, remains, remains_size, false);
 701                         if (r < 0)
 702                                 return r;
 703                 }
 704
 705                 r = copy_bytes(fd, tmp_fd, DATA_FD_TMP_LIMIT - DATA_FD_MEMORY_LIMIT - remains_size, COPY_REFLINK);
 706                 if (r < 0)
 707                         return r;
 708                 if (r == 0)
 709                         goto finish;  /* Yay, it fit in */
 710
 711                 /* It didn't fit in. Let's not forget to use what we already used */
 712                 f = lseek(tmp_fd, 0, SEEK_SET);
 713                 if (f != 0)
 714                         return -errno;
 715
 716                 safe_close(copy_fd);
 717                 copy_fd = TAKE_FD(tmp_fd);
 718
 719                 remains = mfree(remains);
 720                 remains_size = 0;
 721         }
 722
 723         /* As last fallback use /var/tmp */
 724         r = var_tmp_dir(&td);
 725         if (r < 0)
 726                 return r;
 727
 728         tmp_fd = open_tmpfile_unlinkable(td, O_RDWR|O_CLOEXEC);
 729         if (tmp_fd < 0)
 730                 return tmp_fd;
 731
 732         if (copy_fd >= 0) {
 733                 /* If we tried a memfd/pipe first, or a file in /tmp, and it ended up being too large, than copy this
 734                  * into the temporary file first. */
 735                 r = copy_bytes(copy_fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
 736                 if (r < 0)
 737                         return r;
 738
 739                 assert(r == 0);
 740         }
 741
 742         if (remains_size > 0) {
 743                 /* Then, copy in any read but not yet written bytes. */
 744                 r = loop_write(tmp_fd, remains, remains_size, false);
 745                 if (r < 0)
 746                         return r;
 747         }
 748
 749         /* Copy in the rest */
 750         r = copy_bytes(fd, tmp_fd, UINT64_MAX, COPY_REFLINK);
 751         if (r < 0)
 752                 return r;
 753
 754         assert(r == 0);
 755
 756 finish:
 757         /* Now convert the O_RDWR file descriptor into an O_RDONLY one (and as side effect seek to the beginning of the
 758          * file again */
 759
 760         return fd_reopen(tmp_fd, O_RDONLY|O_CLOEXEC);
 761 }
 762
 763 int fd_move_above_stdio(int fd) {
 764         int flags, copy;
 765         PROTECT_ERRNO;
 766
 767         /* Moves the specified file descriptor if possible out of the range [0…2], i.e. the range of
 768          * stdin/stdout/stderr. If it can't be moved outside of this range the original file descriptor is
 769          * returned. This call is supposed to be used for long-lasting file descriptors we allocate in our code that
 770          * might get loaded into foreign code, and where we want ensure our fds are unlikely used accidentally as
 771          * stdin/stdout/stderr of unrelated code.
 772          *
 773          * Note that this doesn't fix any real bugs, it just makes it less likely that our code will be affected by
 774          * buggy code from others that mindlessly invokes 'fprintf(stderr, …' or similar in places where stderr has
 775          * been closed before.
 776          *
 777          * This function is written in a "best-effort" and "least-impact" style. This means whenever we encounter an
 778          * error we simply return the original file descriptor, and we do not touch errno. */
 779
 780         if (fd < 0 || fd > 2)
 781                 return fd;
 782
 783         flags = fcntl(fd, F_GETFD, 0);
 784         if (flags < 0)
 785                 return fd;
 786
 787         if (flags & FD_CLOEXEC)
 788                 copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 789         else
 790                 copy = fcntl(fd, F_DUPFD, 3);
 791         if (copy < 0)
 792                 return fd;
 793
 794         assert(copy > 2);
 795
 796         (void) close(fd);
 797         return copy;
 798 }
 799
 800 int rearrange_stdio(int original_input_fd, int original_output_fd, int original_error_fd) {
 801
 802         int fd[3] = { /* Put together an array of fds we work on */
 803                 original_input_fd,
 804                 original_output_fd,
 805                 original_error_fd
 806         };
 807
 808         int r, i,
 809                 null_fd = -1,                /* if we open /dev/null, we store the fd to it here */
 810                 copy_fd[3] = { -1, -1, -1 }; /* This contains all fds we duplicate here temporarily, and hence need to close at the end */
 811         bool null_readable, null_writable;
 812
 813         /* Sets up stdin, stdout, stderr with the three file descriptors passed in. If any of the descriptors is
 814          * specified as -1 it will be connected with /dev/null instead. If any of the file descriptors is passed as
 815          * itself (e.g. stdin as STDIN_FILENO) it is left unmodified, but the O_CLOEXEC bit is turned off should it be
 816          * on.
 817          *
 818          * Note that if any of the passed file descriptors are > 2 they will be closed — both on success and on
 819          * failure! Thus, callers should assume that when this function returns the input fds are invalidated.
 820          *
 821          * Note that when this function fails stdin/stdout/stderr might remain half set up!
 822          *
 823          * O_CLOEXEC is turned off for all three file descriptors (which is how it should be for
 824          * stdin/stdout/stderr). */
 825
 826         null_readable = original_input_fd < 0;
 827         null_writable = original_output_fd < 0 || original_error_fd < 0;
 828
 829         /* First step, open /dev/null once, if we need it */
 830         if (null_readable || null_writable) {
 831
 832                 /* Let's open this with O_CLOEXEC first, and convert it to non-O_CLOEXEC when we move the fd to the final position. */
 833                 null_fd = open("/dev/null", (null_readable && null_writable ? O_RDWR :
 834                                              null_readable ? O_RDONLY : O_WRONLY) | O_CLOEXEC);
 835                 if (null_fd < 0) {
 836                         r = -errno;
 837                         goto finish;
 838                 }
 839
 840                 /* If this fd is in the 0…2 range, let's move it out of it */
 841                 if (null_fd < 3) {
 842                         int copy;
 843
 844                         copy = fcntl(null_fd, F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
 845                         if (copy < 0) {
 846                                 r = -errno;
 847                                 goto finish;
 848                         }
 849
 850                         safe_close(null_fd);
 851                         null_fd = copy;
 852                 }
 853         }
 854
 855         /* Let's assemble fd[] with the fds to install in place of stdin/stdout/stderr */
 856         for (i = 0; i < 3; i++) {
 857
 858                 if (fd[i] < 0)
 859                         fd[i] = null_fd;        /* A negative parameter means: connect this one to /dev/null */
 860                 else if (fd[i] != i && fd[i] < 3) {
 861                         /* This fd is in the 0…2 territory, but not at its intended place, move it out of there, so that we can work there. */
 862                         copy_fd[i] = fcntl(fd[i], F_DUPFD_CLOEXEC, 3); /* Duplicate this with O_CLOEXEC set */
 863                         if (copy_fd[i] < 0) {
 864                                 r = -errno;
 865                                 goto finish;
 866                         }
 867
 868                         fd[i] = copy_fd[i];
 869                 }
 870         }
 871
 872         /* At this point we now have the fds to use in fd[], and they are all above the stdio range, so that we
 873          * have freedom to move them around. If the fds already were at the right places then the specific fds are
 874          * -1. Let's now move them to the right places. This is the point of no return. */
 875         for (i = 0; i < 3; i++) {
 876
 877                 if (fd[i] == i) {
 878
 879                         /* fd is already in place, but let's make sure O_CLOEXEC is off */
 880                         r = fd_cloexec(i, false);
 881                         if (r < 0)
 882                                 goto finish;
 883
 884                 } else {
 885                         assert(fd[i] > 2);
 886
 887                         if (dup2(fd[i], i) < 0) { /* Turns off O_CLOEXEC on the new fd. */
 888                                 r = -errno;
 889                                 goto finish;
 890                         }
 891                 }
 892         }
 893
 894         r = 0;
 895
 896 finish:
 897         /* Close the original fds, but only if they were outside of the stdio range. Also, properly check for the same
 898          * fd passed in multiple times. */
 899         safe_close_above_stdio(original_input_fd);
 900         if (original_output_fd != original_input_fd)
 901                 safe_close_above_stdio(original_output_fd);
 902         if (original_error_fd != original_input_fd && original_error_fd != original_output_fd)
 903                 safe_close_above_stdio(original_error_fd);
 904
 905         /* Close the copies we moved > 2 */
 906         for (i = 0; i < 3; i++)
 907                 safe_close(copy_fd[i]);
 908
 909         /* Close our null fd, if it's > 2 */
 910         safe_close_above_stdio(null_fd);
 911
 912         return r;
 913 }
 914
 915 int fd_reopen(int fd, int flags) {
 916         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int)];
 917         int new_fd;
 918
 919         /* Reopens the specified fd with new flags. This is useful for convert an O_PATH fd into a regular one, or to
 920          * turn O_RDWR fds into O_RDONLY fds.
 921          *
 922          * This doesn't work on sockets (since they cannot be open()ed, ever).
 923          *
 924          * This implicitly resets the file read index to 0. */
 925
 926         xsprintf(procfs_path, "/proc/self/fd/%i", fd);
 927         new_fd = open(procfs_path, flags);
 928         if (new_fd < 0)
 929                 return -errno;
 930
 931         return new_fd;
 932 }
 933
 934 int read_nr_open(void) {
 935         _cleanup_free_ char *nr_open = NULL;
 936         int r;
 937
 938         /* Returns the kernel's current fd limit, either by reading it of /proc/sys if that works, or using the
 939          * hard-coded default compiled-in value of current kernels (1M) if not. This call will never fail. */
 940
 941         r = read_one_line_file("/proc/sys/fs/nr_open", &nr_open);
 942         if (r < 0)
 943                 log_debug_errno(r, "Failed to read /proc/sys/fs/nr_open, ignoring: %m");
 944         else {
 945                 int v;
 946
 947                 r = safe_atoi(nr_open, &v);
 948                 if (r < 0)
 949                         log_debug_errno(r, "Failed to parse /proc/sys/fs/nr_open value '%s', ignoring: %m", nr_open);
 950                 else
 951                         return v;
 952         }
 953
 954         /* If we fail, fallback to the hard-coded kernel limit of 1024 * 1024. */
 955         return 1024 * 1024;
 956 }