src/basic/fd-util.c

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