src/nspawn.c

   1 /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
   2
   3 /***
   4   This file is part of systemd.
   5
   6   Copyright 2010 Lennart Poettering
   7
   8   systemd is free software; you can redistribute it and/or modify it
   9   under the terms of the GNU General Public License as published by
  10   the Free Software Foundation; either version 2 of the License, or
  11   (at your option) any later version.
  12
  13   systemd is distributed in the hope that it will be useful, but
  14   WITHOUT ANY WARRANTY; without even the implied warranty of
  15   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  16   General Public License for more details.
  17
  18   You should have received a copy of the GNU General Public License
  19   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  20 ***/
  21
  22 #include <signal.h>
  23 #include <sched.h>
  24 #include <unistd.h>
  25 #include <sys/types.h>
  26 #include <sys/syscall.h>
  27 #include <sys/mount.h>
  28 #include <sys/wait.h>
  29 #include <stdlib.h>
  30 #include <string.h>
  31 #include <stdio.h>
  32 #include <errno.h>
  33 #include <sys/prctl.h>
  34 #include <sys/capability.h>
  35 #include <getopt.h>
  36 #include <sys/epoll.h>
  37 #include <termios.h>
  38 #include <sys/signalfd.h>
  39 #include <grp.h>
  40
  41 #include "log.h"
  42 #include "util.h"
  43 #include "missing.h"
  44 #include "cgroup-util.h"
  45 #include "sd-daemon.h"
  46 #include "strv.h"
  47
  48 static char *arg_directory = NULL;
  49 static char *arg_user = NULL;
  50
  51 static int help(void) {
  52
  53         printf("%s [OPTIONS...] [PATH] [ARGUMENTS...]\n\n"
  54                "Spawn a minimal namespace container for debugging, testing and building.\n\n"
  55                "  -h --help            Show this help\n"
  56                "  -D --directory=NAME  Root directory for the container\n"
  57                "  -u --user=USER       Run the command under specified user or uid\n",
  58                program_invocation_short_name);
  59
  60         return 0;
  61 }
  62
  63 static int parse_argv(int argc, char *argv[]) {
  64
  65         static const struct option options[] = {
  66                 { "help",      no_argument,       NULL, 'h' },
  67                 { "directory", required_argument, NULL, 'D' },
  68                 { "user",      optional_argument, NULL, 'u' },
  69                 { NULL,        0,                 NULL, 0   }
  70         };
  71
  72         int c;
  73
  74         assert(argc >= 0);
  75         assert(argv);
  76
  77         while ((c = getopt_long(argc, argv, "+hD:u:", options, NULL)) >= 0) {
  78
  79                 switch (c) {
  80
  81                 case 'h':
  82                         help();
  83                         return 0;
  84
  85                 case 'D':
  86                         free(arg_directory);
  87                         if (!(arg_directory = strdup(optarg))) {
  88                                 log_error("Failed to duplicate root directory.");
  89                                 return -ENOMEM;
  90                         }
  91
  92                         break;
  93
  94                 case 'u':
  95                         free(arg_user);
  96                         if (!(arg_user = strdup(optarg))) {
  97                                 log_error("Failed to duplicate user name.");
  98                                 return -ENOMEM;
  99                         }
 100
 101                         break;
 102
 103                 case '?':
 104                         return -EINVAL;
 105
 106                 default:
 107                         log_error("Unknown option code %c", c);
 108                         return -EINVAL;
 109                 }
 110         }
 111
 112         return 1;
 113 }
 114
 115 static int mount_all(const char *dest) {
 116
 117         typedef struct MountPoint {
 118                 const char *what;
 119                 const char *where;
 120                 const char *type;
 121                 const char *options;
 122                 unsigned long flags;
 123                 bool fatal;
 124         } MountPoint;
 125
 126         static const MountPoint mount_table[] = {
 127                 { "proc",      "/proc",     "proc",  NULL,       MS_NOSUID|MS_NOEXEC|MS_NODEV, true  },
 128                 { "/proc/sys", "/proc/sys", "bind",  NULL,       MS_BIND, true                       },   /* Bind mount first */
 129                 { "/proc/sys", "/proc/sys", "bind",  NULL,       MS_BIND|MS_RDONLY|MS_REMOUNT, true  },   /* Then, make it r/o */
 130                 { "/sys",      "/sys",      "bind",  NULL,       MS_BIND,                      true  },   /* Bind mount first */
 131                 { "/sys",      "/sys",      "bind",  NULL,       MS_BIND|MS_RDONLY|MS_REMOUNT, true  },   /* Then, make it r/o */
 132                 { "tmpfs",     "/dev",      "tmpfs", "mode=755", MS_NOSUID,                    true  },
 133                 { "/dev/pts",  "/dev/pts",  "bind",  NULL,       MS_BIND,                      true  },
 134                 { "tmpfs",     "/run",      "tmpfs", "mode=755", MS_NOSUID|MS_NODEV,           true  },
 135 #ifdef HAVE_SELINUX
 136                 { "/sys/fs/selinux", "/sys/fs/selinux", "bind", NULL, MS_BIND,                      false },  /* Bind mount first */
 137                 { "/sys/fs/selinux", "/sys/fs/selinux", "bind", NULL, MS_BIND|MS_RDONLY|MS_REMOUNT, false },  /* Then, make it r/o */
 138 #endif
 139         };
 140
 141         unsigned k;
 142         int r = 0;
 143         char *where;
 144
 145         for (k = 0; k < ELEMENTSOF(mount_table); k++) {
 146                 int t;
 147
 148                 if (asprintf(&where, "%s/%s", dest, mount_table[k].where) < 0) {
 149                         log_error("Out of memory");
 150
 151                         if (r == 0)
 152                                 r = -ENOMEM;
 153
 154                         break;
 155                 }
 156
 157                 if ((t = path_is_mount_point(where)) < 0) {
 158                         log_error("Failed to detect whether %s is a mount point: %s", where, strerror(-t));
 159                         free(where);
 160
 161                         if (r == 0)
 162                                 r = t;
 163
 164                         continue;
 165                 }
 166
 167                 mkdir_p(where, 0755);
 168
 169                 if (mount(mount_table[k].what,
 170                           where,
 171                           mount_table[k].type,
 172                           mount_table[k].flags,
 173                           mount_table[k].options) < 0 &&
 174                     mount_table[k].fatal) {
 175
 176                         log_error("mount(%s) failed: %m", where);
 177
 178                         if (r == 0)
 179                                 r = -errno;
 180                 }
 181
 182                 free(where);
 183         }
 184
 185         /* Fix the timezone, if possible */
 186         if (asprintf(&where, "%s/%s", dest, "/etc/localtime") >= 0) {
 187                 mount("/etc/localtime", where, "bind", MS_BIND, NULL);
 188                 mount("/etc/localtime", where, "bind", MS_BIND|MS_REMOUNT|MS_RDONLY, NULL);
 189                 free(where);
 190         }
 191
 192         return r;
 193 }
 194
 195 static int copy_devnodes(const char *dest, const char *console) {
 196
 197         static const char devnodes[] =
 198                 "null\0"
 199                 "zero\0"
 200                 "full\0"
 201                 "random\0"
 202                 "urandom\0"
 203                 "tty\0"
 204                 "ptmx\0"
 205                 "kmsg\0"
 206                 "rtc0\0";
 207
 208         const char *d;
 209         int r = 0, k;
 210         mode_t u;
 211         struct stat st;
 212         char *from = NULL, *to = NULL;
 213
 214         assert(dest);
 215         assert(console);
 216
 217         u = umask(0000);
 218
 219         NULSTR_FOREACH(d, devnodes) {
 220                 from = to = NULL;
 221
 222                 asprintf(&from, "/dev/%s", d);
 223                 asprintf(&to, "%s/dev/%s", dest, d);
 224
 225                 if (!from || !to) {
 226                         log_error("Failed to allocate devnode path");
 227
 228                         free(from);
 229                         free(to);
 230
 231                         from = to = NULL;
 232
 233                         if (r == 0)
 234                                 r = -ENOMEM;
 235
 236                         break;
 237                 }
 238
 239                 if (stat(from, &st) < 0) {
 240
 241                         if (errno != ENOENT) {
 242                                 log_error("Failed to stat %s: %m", from);
 243                                 if (r == 0)
 244                                         r = -errno;
 245                         }
 246
 247                 } else if (!S_ISCHR(st.st_mode) && !S_ISBLK(st.st_mode)) {
 248
 249                         log_error("%s is not a char or block device, cannot copy.", from);
 250                         if (r == 0)
 251                                 r = -EIO;
 252
 253                 } else if (mknod(to, st.st_mode, st.st_rdev) < 0) {
 254
 255                         log_error("mknod(%s) failed: %m", dest);
 256                         if (r == 0)
 257                                 r = -errno;
 258                 }
 259
 260                 free(from);
 261                 free(to);
 262         }
 263
 264         if (stat(console, &st) < 0) {
 265
 266                 log_error("Failed to stat %s: %m", console);
 267                 if (r == 0)
 268                         r = -errno;
 269
 270                 goto finish;
 271
 272         } else if (!S_ISCHR(st.st_mode)) {
 273
 274                 log_error("/dev/console is not a char device.");
 275                 if (r == 0)
 276                         r = -EIO;
 277
 278                 goto finish;
 279         }
 280
 281         if (asprintf(&to, "%s/dev/console", dest) < 0) {
 282
 283                 log_error("Out of memory");
 284                 if (r == 0)
 285                         r = -ENOMEM;
 286
 287                  goto finish;
 288         }
 289
 290         /* We need to bind mount the right tty to /dev/console since
 291          * ptys can only exist on pts file systems. To have something
 292          * to bind mount things on we create a device node first, that
 293          * has the right major/minor (note that the major minor
 294          * doesn't actually matter here, since we mount it over
 295          * anyway). */
 296
 297         if (mknod(to, (st.st_mode & ~07777) | 0600, st.st_rdev) < 0)
 298                 log_error("mknod for /dev/console failed: %m");
 299
 300         if (mount(console, to, "bind", MS_BIND, NULL) < 0) {
 301                 log_error("bind mount for /dev/console failed: %m");
 302
 303                 if (r == 0)
 304                         r = -errno;
 305         }
 306
 307         free(to);
 308
 309         if ((k = chmod_and_chown(console, 0600, 0, 0)) < 0) {
 310                 log_error("Failed to correct access mode for TTY: %s", strerror(-k));
 311
 312                 if (r == 0)
 313                         r = k;
 314         }
 315
 316 finish:
 317         umask(u);
 318
 319         return r;
 320 }
 321
 322 static int drop_capabilities(void) {
 323         static const unsigned long retain[] = {
 324                 CAP_CHOWN,
 325                 CAP_DAC_OVERRIDE,
 326                 CAP_DAC_READ_SEARCH,
 327                 CAP_FOWNER,
 328                 CAP_FSETID,
 329                 CAP_IPC_OWNER,
 330                 CAP_KILL,
 331                 CAP_LEASE,
 332                 CAP_LINUX_IMMUTABLE,
 333                 CAP_NET_BIND_SERVICE,
 334                 CAP_NET_BROADCAST,
 335                 CAP_NET_RAW,
 336                 CAP_SETGID,
 337                 CAP_SETFCAP,
 338                 CAP_SETPCAP,
 339                 CAP_SETUID,
 340                 CAP_SYS_ADMIN,
 341                 CAP_SYS_CHROOT,
 342                 CAP_SYS_NICE,
 343                 CAP_SYS_PTRACE,
 344                 CAP_SYS_TTY_CONFIG
 345         };
 346
 347         unsigned long l;
 348
 349         for (l = 0; l <= MAX(63LU, (unsigned long) CAP_LAST_CAP); l++) {
 350                 unsigned i;
 351
 352                 for (i = 0; i < ELEMENTSOF(retain); i++)
 353                         if (retain[i] == l)
 354                                 break;
 355
 356                 if (i < ELEMENTSOF(retain))
 357                         continue;
 358
 359                 if (prctl(PR_CAPBSET_DROP, l) < 0) {
 360
 361                         /* If this capability is not known, EINVAL
 362                          * will be returned, let's ignore this. */
 363                         if (errno == EINVAL)
 364                                 break;
 365
 366                         log_error("PR_CAPBSET_DROP failed: %m");
 367                         return -errno;
 368                 }
 369         }
 370
 371         return 0;
 372 }
 373
 374 static int is_os_tree(const char *path) {
 375         int r;
 376         char *p;
 377         /* We use /bin/sh as flag file if something is an OS */
 378
 379         if (asprintf(&p, "%s/bin/sh", path) < 0)
 380                 return -ENOMEM;
 381
 382         r = access(p, F_OK);
 383         free(p);
 384
 385         return r < 0 ? 0 : 1;
 386 }
 387
 388 #define BUFFER_SIZE 1024
 389
 390 static int process_pty(int master, sigset_t *mask) {
 391
 392         char in_buffer[BUFFER_SIZE], out_buffer[BUFFER_SIZE];
 393         size_t in_buffer_full = 0, out_buffer_full = 0;
 394         struct epoll_event stdin_ev, stdout_ev, master_ev, signal_ev;
 395         bool stdin_readable = false, stdout_writable = false, master_readable = false, master_writable = false;
 396         int ep = -1, signal_fd = -1, r;
 397
 398         fd_nonblock(STDIN_FILENO, 1);
 399         fd_nonblock(STDOUT_FILENO, 1);
 400         fd_nonblock(master, 1);
 401
 402         if ((signal_fd = signalfd(-1, mask, SFD_NONBLOCK|SFD_CLOEXEC)) < 0) {
 403                 log_error("signalfd(): %m");
 404                 r = -errno;
 405                 goto finish;
 406         }
 407
 408         if ((ep = epoll_create1(EPOLL_CLOEXEC)) < 0) {
 409                 log_error("Failed to create epoll: %m");
 410                 r = -errno;
 411                 goto finish;
 412         }
 413
 414         zero(stdin_ev);
 415         stdin_ev.events = EPOLLIN|EPOLLET;
 416         stdin_ev.data.fd = STDIN_FILENO;
 417
 418         zero(stdout_ev);
 419         stdout_ev.events = EPOLLOUT|EPOLLET;
 420         stdout_ev.data.fd = STDOUT_FILENO;
 421
 422         zero(master_ev);
 423         master_ev.events = EPOLLIN|EPOLLOUT|EPOLLET;
 424         master_ev.data.fd = master;
 425
 426         zero(signal_ev);
 427         signal_ev.events = EPOLLIN;
 428         signal_ev.data.fd = signal_fd;
 429
 430         if (epoll_ctl(ep, EPOLL_CTL_ADD, STDIN_FILENO, &stdin_ev) < 0 ||
 431             epoll_ctl(ep, EPOLL_CTL_ADD, STDOUT_FILENO, &stdout_ev) < 0 ||
 432             epoll_ctl(ep, EPOLL_CTL_ADD, master, &master_ev) < 0 ||
 433             epoll_ctl(ep, EPOLL_CTL_ADD, signal_fd, &signal_ev) < 0) {
 434                 log_error("Failed to regiser fds in epoll: %m");
 435                 r = -errno;
 436                 goto finish;
 437         }
 438
 439         for (;;) {
 440                 struct epoll_event ev[16];
 441                 ssize_t k;
 442                 int i, nfds;
 443
 444                 if ((nfds = epoll_wait(ep, ev, ELEMENTSOF(ev), -1)) < 0) {
 445
 446                         if (errno == EINTR || errno == EAGAIN)
 447                                 continue;
 448
 449                         log_error("epoll_wait(): %m");
 450                         r = -errno;
 451                         goto finish;
 452                 }
 453
 454                 assert(nfds >= 1);
 455
 456                 for (i = 0; i < nfds; i++) {
 457                         if (ev[i].data.fd == STDIN_FILENO) {
 458
 459                                 if (ev[i].events & (EPOLLIN|EPOLLHUP))
 460                                         stdin_readable = true;
 461
 462                         } else if (ev[i].data.fd == STDOUT_FILENO) {
 463
 464                                 if (ev[i].events & (EPOLLOUT|EPOLLHUP))
 465                                         stdout_writable = true;
 466
 467                         } else if (ev[i].data.fd == master) {
 468
 469                                 if (ev[i].events & (EPOLLIN|EPOLLHUP))
 470                                         master_readable = true;
 471
 472                                 if (ev[i].events & (EPOLLOUT|EPOLLHUP))
 473                                         master_writable = true;
 474
 475                         } else if (ev[i].data.fd == signal_fd) {
 476                                 struct signalfd_siginfo sfsi;
 477                                 ssize_t n;
 478
 479                                 if ((n = read(signal_fd, &sfsi, sizeof(sfsi))) != sizeof(sfsi)) {
 480
 481                                         if (n >= 0) {
 482                                                 log_error("Failed to read from signalfd: invalid block size");
 483                                                 r = -EIO;
 484                                                 goto finish;
 485                                         }
 486
 487                                         if (errno != EINTR && errno != EAGAIN) {
 488                                                 log_error("Failed to read from signalfd: %m");
 489                                                 r = -errno;
 490                                                 goto finish;
 491                                         }
 492                                 } else {
 493
 494                                         if (sfsi.ssi_signo == SIGWINCH) {
 495                                                 struct winsize ws;
 496
 497                                                 /* The window size changed, let's forward that. */
 498                                                 if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 499                                                         ioctl(master, TIOCSWINSZ, &ws);
 500                                         } else {
 501                                                 r = 0;
 502                                                 goto finish;
 503                                         }
 504                                 }
 505                         }
 506                 }
 507
 508                 while ((stdin_readable && in_buffer_full <= 0) ||
 509                        (master_writable && in_buffer_full > 0) ||
 510                        (master_readable && out_buffer_full <= 0) ||
 511                        (stdout_writable && out_buffer_full > 0)) {
 512
 513                         if (stdin_readable && in_buffer_full < BUFFER_SIZE) {
 514
 515                                 if ((k = read(STDIN_FILENO, in_buffer + in_buffer_full, BUFFER_SIZE - in_buffer_full)) < 0) {
 516
 517                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 518                                                 stdin_readable = false;
 519                                         else {
 520                                                 log_error("read(): %m");
 521                                                 r = -errno;
 522                                                 goto finish;
 523                                         }
 524                                 } else
 525                                         in_buffer_full += (size_t) k;
 526                         }
 527
 528                         if (master_writable && in_buffer_full > 0) {
 529
 530                                 if ((k = write(master, in_buffer, in_buffer_full)) < 0) {
 531
 532                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 533                                                 master_writable = false;
 534                                         else {
 535                                                 log_error("write(): %m");
 536                                                 r = -errno;
 537                                                 goto finish;
 538                                         }
 539
 540                                 } else {
 541                                         assert(in_buffer_full >= (size_t) k);
 542                                         memmove(in_buffer, in_buffer + k, in_buffer_full - k);
 543                                         in_buffer_full -= k;
 544                                 }
 545                         }
 546
 547                         if (master_readable && out_buffer_full < BUFFER_SIZE) {
 548
 549                                 if ((k = read(master, out_buffer + out_buffer_full, BUFFER_SIZE - out_buffer_full)) < 0) {
 550
 551                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 552                                                 master_readable = false;
 553                                         else {
 554                                                 log_error("read(): %m");
 555                                                 r = -errno;
 556                                                 goto finish;
 557                                         }
 558                                 }  else
 559                                         out_buffer_full += (size_t) k;
 560                         }
 561
 562                         if (stdout_writable && out_buffer_full > 0) {
 563
 564                                 if ((k = write(STDOUT_FILENO, out_buffer, out_buffer_full)) < 0) {
 565
 566                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 567                                                 stdout_writable = false;
 568                                         else {
 569                                                 log_error("write(): %m");
 570                                                 r = -errno;
 571                                                 goto finish;
 572                                         }
 573
 574                                 } else {
 575                                         assert(out_buffer_full >= (size_t) k);
 576                                         memmove(out_buffer, out_buffer + k, out_buffer_full - k);
 577                                         out_buffer_full -= k;
 578                                 }
 579                         }
 580                 }
 581         }
 582
 583 finish:
 584         if (ep >= 0)
 585                 close_nointr_nofail(ep);
 586
 587         if (signal_fd >= 0)
 588                 close_nointr_nofail(signal_fd);
 589
 590         return r;
 591 }
 592
 593 int main(int argc, char *argv[]) {
 594         pid_t pid = 0;
 595         int r = EXIT_FAILURE, k;
 596         char *oldcg = NULL, *newcg = NULL;
 597         int master = -1;
 598         const char *console = NULL;
 599         struct termios saved_attr, raw_attr;
 600         sigset_t mask;
 601         bool saved_attr_valid = false;
 602         struct winsize ws;
 603
 604         log_parse_environment();
 605         log_open();
 606
 607         if ((r = parse_argv(argc, argv)) <= 0)
 608                 goto finish;
 609
 610         if (arg_directory) {
 611                 char *p;
 612
 613                 p = path_make_absolute_cwd(arg_directory);
 614                 free(arg_directory);
 615                 arg_directory = p;
 616         } else
 617                 arg_directory = get_current_dir_name();
 618
 619         if (!arg_directory) {
 620                 log_error("Failed to determine path");
 621                 goto finish;
 622         }
 623
 624         path_kill_slashes(arg_directory);
 625
 626         if (geteuid() != 0) {
 627                 log_error("Need to be root.");
 628                 goto finish;
 629         }
 630
 631         if (sd_booted() <= 0) {
 632                 log_error("Not running on a systemd system.");
 633                 goto finish;
 634         }
 635
 636         if (path_equal(arg_directory, "/")) {
 637                 log_error("Spawning container on root directory not supported.");
 638                 goto finish;
 639         }
 640
 641         if (is_os_tree(arg_directory) <= 0) {
 642                 log_error("Directory %s doesn't look like an OS root directory. Refusing.", arg_directory);
 643                 goto finish;
 644         }
 645
 646         if ((k = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 0, &oldcg)) < 0) {
 647                 log_error("Failed to determine current cgroup: %s", strerror(-k));
 648                 goto finish;
 649         }
 650
 651         if (asprintf(&newcg, "%s/nspawn-%lu", oldcg, (unsigned long) getpid()) < 0) {
 652                 log_error("Failed to allocate cgroup path.");
 653                 goto finish;
 654         }
 655
 656         if ((k = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, newcg, 0)) < 0)  {
 657                 log_error("Failed to create cgroup: %s", strerror(-k));
 658                 goto finish;
 659         }
 660
 661         if ((master = posix_openpt(O_RDWR|O_NOCTTY|O_CLOEXEC|O_NDELAY)) < 0) {
 662                 log_error("Failed to acquire pseudo tty: %m");
 663                 goto finish;
 664         }
 665
 666         if (!(console = ptsname(master))) {
 667                 log_error("Failed to determine tty name: %m");
 668                 goto finish;
 669         }
 670
 671         log_info("Spawning namespace container on %s (console is %s).", arg_directory, console);
 672
 673         if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 674                 ioctl(master, TIOCSWINSZ, &ws);
 675
 676         if (unlockpt(master) < 0) {
 677                 log_error("Failed to unlock tty: %m");
 678                 goto finish;
 679         }
 680
 681         if (tcgetattr(STDIN_FILENO, &saved_attr) < 0) {
 682                 log_error("Failed to get terminal attributes: %m");
 683                 goto finish;
 684         }
 685
 686         saved_attr_valid = true;
 687
 688         raw_attr = saved_attr;
 689         cfmakeraw(&raw_attr);
 690         raw_attr.c_lflag &= ~ECHO;
 691
 692         if (tcsetattr(STDIN_FILENO, TCSANOW, &raw_attr) < 0) {
 693                 log_error("Failed to set terminal attributes: %m");
 694                 goto finish;
 695         }
 696
 697         assert_se(sigemptyset(&mask) == 0);
 698         sigset_add_many(&mask, SIGCHLD, SIGWINCH, SIGTERM, SIGINT, -1);
 699         assert_se(sigprocmask(SIG_BLOCK, &mask, NULL) == 0);
 700
 701         if ((pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS, NULL)) < 0) {
 702                 log_error("clone() failed: %m");
 703                 goto finish;
 704         }
 705
 706         if (pid == 0) {
 707                 /* child */
 708
 709                 const char *hn;
 710                 const char *home = NULL;
 711                 uid_t uid = (uid_t) -1;
 712                 gid_t gid = (gid_t) -1;
 713                 const char *envp[] = {
 714                         "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
 715                         NULL, /* TERM */
 716                         NULL, /* HOME */
 717                         NULL, /* USER */
 718                         NULL, /* LOGNAME */
 719                         NULL
 720                 };
 721
 722                 envp[1] = strv_find_prefix(environ, "TERM=");
 723
 724                 close_nointr_nofail(master);
 725
 726                 close_nointr(STDIN_FILENO);
 727                 close_nointr(STDOUT_FILENO);
 728                 close_nointr(STDERR_FILENO);
 729
 730                 close_all_fds(NULL, 0);
 731
 732                 reset_all_signal_handlers();
 733
 734                 assert_se(sigemptyset(&mask) == 0);
 735                 assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
 736
 737                 if (setsid() < 0)
 738                         goto child_fail;
 739
 740                 if (prctl(PR_SET_PDEATHSIG, SIGKILL) < 0)
 741                         goto child_fail;
 742
 743                 /* Mark / as private, in case somebody marked it shared */
 744                 if (mount(NULL, "/", NULL, MS_PRIVATE|MS_REC, NULL) < 0)
 745                         goto child_fail;
 746
 747                 if (mount_all(arg_directory) < 0)
 748                         goto child_fail;
 749
 750                 if (copy_devnodes(arg_directory, console) < 0)
 751                         goto child_fail;
 752
 753                 if (chdir(arg_directory) < 0) {
 754                         log_error("chdir(%s) failed: %m", arg_directory);
 755                         goto child_fail;
 756                 }
 757
 758                 if (open_terminal("dev/console", O_RDWR) != STDIN_FILENO ||
 759                     dup2(STDIN_FILENO, STDOUT_FILENO) != STDOUT_FILENO ||
 760                     dup2(STDIN_FILENO, STDERR_FILENO) != STDERR_FILENO)
 761                         goto child_fail;
 762
 763                 if (mount(arg_directory, "/", "bind", MS_BIND|MS_MOVE, NULL) < 0) {
 764                         log_error("mount(MS_MOVE) failed: %m");
 765                         goto child_fail;
 766                 }
 767
 768                 if (chroot(".") < 0) {
 769                         log_error("chroot() failed: %m");
 770                         goto child_fail;
 771                 }
 772
 773                 if (chdir("/") < 0) {
 774                         log_error("chdir() failed: %m");
 775                         goto child_fail;
 776                 }
 777
 778                 umask(0022);
 779
 780                 if (drop_capabilities() < 0)
 781                         goto child_fail;
 782
 783                 if (arg_user) {
 784
 785                         if (get_user_creds((const char**)&arg_user, &uid, &gid, &home) < 0) {
 786                                 log_error("get_user_creds() failed: %m");
 787                                 goto child_fail;
 788                         }
 789
 790                         if (mkdir_parents(home, 0775) < 0) {
 791                                 log_error("mkdir_parents() failed: %m");
 792                                 goto child_fail;
 793                         }
 794
 795                         if (safe_mkdir(home, 0775, uid, gid) < 0) {
 796                                 log_error("safe_mkdir() failed: %m");
 797                                 goto child_fail;
 798                         }
 799
 800                         if (initgroups((const char*)arg_user, gid) < 0) {
 801                                 log_error("initgroups() failed: %m");
 802                                 goto child_fail;
 803                         }
 804
 805                         if (setresgid(gid, gid, gid) < 0) {
 806                                 log_error("setregid() failed: %m");
 807                                 goto child_fail;
 808                         }
 809
 810                         if (setresuid(uid, uid, uid) < 0) {
 811                                 log_error("setreuid() failed: %m");
 812                                 goto child_fail;
 813                         }
 814                 }
 815
 816                 if ((asprintf((char**)(envp + 2), "HOME=%s", home? home: "/root") < 0) ||
 817                     (asprintf((char**)(envp + 3), "USER=%s", arg_user? arg_user : "root") < 0) ||
 818                     (asprintf((char**)(envp + 4), "LOGNAME=%s", arg_user? arg_user : "root") < 0)) {
 819                     log_error("Out of memory");
 820                     goto child_fail;
 821                 }
 822
 823                 if ((hn = file_name_from_path(arg_directory)))
 824                         sethostname(hn, strlen(hn));
 825
 826                 if (argc > optind)
 827                         execvpe(argv[optind], argv + optind, (char**) envp);
 828                 else {
 829                         chdir(home ? home : "/root");
 830                         execle("/bin/bash", "-bash", NULL, (char**) envp);
 831                 }
 832
 833                 log_error("execv() failed: %m");
 834
 835         child_fail:
 836                 _exit(EXIT_FAILURE);
 837         }
 838
 839         if (process_pty(master, &mask) < 0)
 840                 goto finish;
 841
 842         if (saved_attr_valid) {
 843                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 844                 saved_attr_valid = false;
 845         }
 846
 847         r = wait_for_terminate_and_warn(argc > optind ? argv[optind] : "bash", pid);
 848
 849         if (r < 0)
 850                 r = EXIT_FAILURE;
 851
 852 finish:
 853         if (saved_attr_valid)
 854                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 855
 856         if (master >= 0)
 857                 close_nointr_nofail(master);
 858
 859         if (oldcg)
 860                 cg_attach(SYSTEMD_CGROUP_CONTROLLER, oldcg, 0);
 861
 862         if (newcg)
 863                 cg_kill_recursive_and_wait(SYSTEMD_CGROUP_CONTROLLER, newcg, true);
 864
 865         free(arg_directory);
 866         free(oldcg);
 867         free(newcg);
 868
 869         return r;
 870 }