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 #include <linux/fs.h>
  41
  42 #include "log.h"
  43 #include "util.h"
  44 #include "missing.h"
  45 #include "cgroup-util.h"
  46 #include "sd-daemon.h"
  47 #include "strv.h"
  48 #include "loopback-setup.h"
  49
  50 static char *arg_directory = NULL;
  51 static char *arg_user = NULL;
  52 static bool arg_private_network = false;
  53
  54 static int help(void) {
  55
  56         printf("%s [OPTIONS...] [PATH] [ARGUMENTS...]\n\n"
  57                "Spawn a minimal namespace container for debugging, testing and building.\n\n"
  58                "  -h --help            Show this help\n"
  59                "  -D --directory=NAME  Root directory for the container\n"
  60                "  -u --user=USER       Run the command under specified user or uid\n"
  61                "     --private-network Disable network in container\n",
  62                program_invocation_short_name);
  63
  64         return 0;
  65 }
  66
  67 static int parse_argv(int argc, char *argv[]) {
  68
  69         enum {
  70                 ARG_PRIVATE_NETWORK = 0x100
  71         };
  72
  73         static const struct option options[] = {
  74                 { "help",            no_argument,       NULL, 'h'                 },
  75                 { "directory",       required_argument, NULL, 'D'                 },
  76                 { "user",            required_argument, NULL, 'u'                 },
  77                 { "private-network", no_argument,       NULL, ARG_PRIVATE_NETWORK },
  78                 { NULL,              0,                 NULL, 0                   }
  79         };
  80
  81         int c;
  82
  83         assert(argc >= 0);
  84         assert(argv);
  85
  86         while ((c = getopt_long(argc, argv, "+hD:u:", options, NULL)) >= 0) {
  87
  88                 switch (c) {
  89
  90                 case 'h':
  91                         help();
  92                         return 0;
  93
  94                 case 'D':
  95                         free(arg_directory);
  96                         if (!(arg_directory = strdup(optarg))) {
  97                                 log_error("Failed to duplicate root directory.");
  98                                 return -ENOMEM;
  99                         }
 100
 101                         break;
 102
 103                 case 'u':
 104                         free(arg_user);
 105                         if (!(arg_user = strdup(optarg))) {
 106                                 log_error("Failed to duplicate user name.");
 107                                 return -ENOMEM;
 108                         }
 109
 110                         break;
 111
 112                 case ARG_PRIVATE_NETWORK:
 113                         arg_private_network = true;
 114                         break;
 115
 116                 case '?':
 117                         return -EINVAL;
 118
 119                 default:
 120                         log_error("Unknown option code %c", c);
 121                         return -EINVAL;
 122                 }
 123         }
 124
 125         return 1;
 126 }
 127
 128 static int mount_all(const char *dest) {
 129
 130         typedef struct MountPoint {
 131                 const char *what;
 132                 const char *where;
 133                 const char *type;
 134                 const char *options;
 135                 unsigned long flags;
 136                 bool fatal;
 137         } MountPoint;
 138
 139         static const MountPoint mount_table[] = {
 140                 { "proc",      "/proc",     "proc",  NULL,       MS_NOSUID|MS_NOEXEC|MS_NODEV, true  },
 141                 { "/proc/sys", "/proc/sys", "bind",  NULL,       MS_BIND, true                       },   /* Bind mount first */
 142                 { "/proc/sys", "/proc/sys", "bind",  NULL,       MS_BIND|MS_RDONLY|MS_REMOUNT, true  },   /* Then, make it r/o */
 143                 { "/sys",      "/sys",      "bind",  NULL,       MS_BIND,                      true  },   /* Bind mount first */
 144                 { "/sys",      "/sys",      "bind",  NULL,       MS_BIND|MS_RDONLY|MS_REMOUNT, true  },   /* Then, make it r/o */
 145                 { "tmpfs",     "/dev",      "tmpfs", "mode=755", MS_NOSUID,                    true  },
 146                 { "/dev/pts",  "/dev/pts",  "bind",  NULL,       MS_BIND,                      true  },
 147                 { "tmpfs",     "/run",      "tmpfs", "mode=755", MS_NOSUID|MS_NODEV,           true  },
 148 #ifdef HAVE_SELINUX
 149                 { "/sys/fs/selinux", "/sys/fs/selinux", "bind", NULL, MS_BIND,                      false },  /* Bind mount first */
 150                 { "/sys/fs/selinux", "/sys/fs/selinux", "bind", NULL, MS_BIND|MS_RDONLY|MS_REMOUNT, false },  /* Then, make it r/o */
 151 #endif
 152         };
 153
 154         unsigned k;
 155         int r = 0;
 156         char *where;
 157
 158         for (k = 0; k < ELEMENTSOF(mount_table); k++) {
 159                 int t;
 160
 161                 if (asprintf(&where, "%s/%s", dest, mount_table[k].where) < 0) {
 162                         log_error("Out of memory");
 163
 164                         if (r == 0)
 165                                 r = -ENOMEM;
 166
 167                         break;
 168                 }
 169
 170                 if ((t = path_is_mount_point(where, false)) < 0) {
 171                         log_error("Failed to detect whether %s is a mount point: %s", where, strerror(-t));
 172                         free(where);
 173
 174                         if (r == 0)
 175                                 r = t;
 176
 177                         continue;
 178                 }
 179
 180                 mkdir_p(where, 0755);
 181
 182                 if (mount(mount_table[k].what,
 183                           where,
 184                           mount_table[k].type,
 185                           mount_table[k].flags,
 186                           mount_table[k].options) < 0 &&
 187                     mount_table[k].fatal) {
 188
 189                         log_error("mount(%s) failed: %m", where);
 190
 191                         if (r == 0)
 192                                 r = -errno;
 193                 }
 194
 195                 free(where);
 196         }
 197
 198         /* Fix the timezone, if possible */
 199         if (asprintf(&where, "%s/%s", dest, "/etc/localtime") >= 0) {
 200
 201                 if (mount("/etc/localtime", where, "bind", MS_BIND, NULL) >= 0)
 202                         mount("/etc/localtime", where, "bind", MS_BIND|MS_REMOUNT|MS_RDONLY, NULL);
 203
 204                 free(where);
 205         }
 206
 207         return r;
 208 }
 209
 210 static int copy_devnodes(const char *dest, const char *console) {
 211
 212         static const char devnodes[] =
 213                 "null\0"
 214                 "zero\0"
 215                 "full\0"
 216                 "random\0"
 217                 "urandom\0"
 218                 "tty\0"
 219                 "ptmx\0"
 220                 "kmsg\0"
 221                 "rtc0\0";
 222
 223         const char *d;
 224         int r = 0, k;
 225         mode_t u;
 226         struct stat st;
 227         char *from = NULL, *to = NULL;
 228
 229         assert(dest);
 230         assert(console);
 231
 232         u = umask(0000);
 233
 234         NULSTR_FOREACH(d, devnodes) {
 235                 from = to = NULL;
 236
 237                 asprintf(&from, "/dev/%s", d);
 238                 asprintf(&to, "%s/dev/%s", dest, d);
 239
 240                 if (!from || !to) {
 241                         log_error("Failed to allocate devnode path");
 242
 243                         free(from);
 244                         free(to);
 245
 246                         from = to = NULL;
 247
 248                         if (r == 0)
 249                                 r = -ENOMEM;
 250
 251                         break;
 252                 }
 253
 254                 if (stat(from, &st) < 0) {
 255
 256                         if (errno != ENOENT) {
 257                                 log_error("Failed to stat %s: %m", from);
 258                                 if (r == 0)
 259                                         r = -errno;
 260                         }
 261
 262                 } else if (!S_ISCHR(st.st_mode) && !S_ISBLK(st.st_mode)) {
 263
 264                         log_error("%s is not a char or block device, cannot copy.", from);
 265                         if (r == 0)
 266                                 r = -EIO;
 267
 268                 } else if (mknod(to, st.st_mode, st.st_rdev) < 0) {
 269
 270                         log_error("mknod(%s) failed: %m", dest);
 271                         if (r == 0)
 272                                 r = -errno;
 273                 }
 274
 275                 free(from);
 276                 free(to);
 277         }
 278
 279         if (stat(console, &st) < 0) {
 280
 281                 log_error("Failed to stat %s: %m", console);
 282                 if (r == 0)
 283                         r = -errno;
 284
 285                 goto finish;
 286
 287         } else if (!S_ISCHR(st.st_mode)) {
 288
 289                 log_error("/dev/console is not a char device.");
 290                 if (r == 0)
 291                         r = -EIO;
 292
 293                 goto finish;
 294         }
 295
 296         if (asprintf(&to, "%s/dev/console", dest) < 0) {
 297
 298                 log_error("Out of memory");
 299                 if (r == 0)
 300                         r = -ENOMEM;
 301
 302                  goto finish;
 303         }
 304
 305         /* We need to bind mount the right tty to /dev/console since
 306          * ptys can only exist on pts file systems. To have something
 307          * to bind mount things on we create a device node first, that
 308          * has the right major/minor (note that the major minor
 309          * doesn't actually matter here, since we mount it over
 310          * anyway). */
 311
 312         if (mknod(to, (st.st_mode & ~07777) | 0600, st.st_rdev) < 0)
 313                 log_error("mknod for /dev/console failed: %m");
 314
 315         if (mount(console, to, "bind", MS_BIND, NULL) < 0) {
 316                 log_error("bind mount for /dev/console failed: %m");
 317
 318                 if (r == 0)
 319                         r = -errno;
 320         }
 321
 322         free(to);
 323
 324         if ((k = chmod_and_chown(console, 0600, 0, 0)) < 0) {
 325                 log_error("Failed to correct access mode for TTY: %s", strerror(-k));
 326
 327                 if (r == 0)
 328                         r = k;
 329         }
 330
 331 finish:
 332         umask(u);
 333
 334         return r;
 335 }
 336
 337 static int drop_capabilities(void) {
 338         static const unsigned long retain[] = {
 339                 CAP_CHOWN,
 340                 CAP_DAC_OVERRIDE,
 341                 CAP_DAC_READ_SEARCH,
 342                 CAP_FOWNER,
 343                 CAP_FSETID,
 344                 CAP_IPC_OWNER,
 345                 CAP_KILL,
 346                 CAP_LEASE,
 347                 CAP_LINUX_IMMUTABLE,
 348                 CAP_NET_BIND_SERVICE,
 349                 CAP_NET_BROADCAST,
 350                 CAP_NET_RAW,
 351                 CAP_SETGID,
 352                 CAP_SETFCAP,
 353                 CAP_SETPCAP,
 354                 CAP_SETUID,
 355                 CAP_SYS_ADMIN,
 356                 CAP_SYS_CHROOT,
 357                 CAP_SYS_NICE,
 358                 CAP_SYS_PTRACE,
 359                 CAP_SYS_TTY_CONFIG
 360         };
 361
 362         unsigned long l;
 363
 364         for (l = 0; l <= cap_last_cap(); l++) {
 365                 unsigned i;
 366
 367                 for (i = 0; i < ELEMENTSOF(retain); i++)
 368                         if (retain[i] == l)
 369                                 break;
 370
 371                 if (i < ELEMENTSOF(retain))
 372                         continue;
 373
 374                 if (prctl(PR_CAPBSET_DROP, l) < 0) {
 375                         log_error("PR_CAPBSET_DROP failed: %m");
 376                         return -errno;
 377                 }
 378         }
 379
 380         return 0;
 381 }
 382
 383 static int is_os_tree(const char *path) {
 384         int r;
 385         char *p;
 386         /* We use /bin/sh as flag file if something is an OS */
 387
 388         if (asprintf(&p, "%s/bin/sh", path) < 0)
 389                 return -ENOMEM;
 390
 391         r = access(p, F_OK);
 392         free(p);
 393
 394         return r < 0 ? 0 : 1;
 395 }
 396
 397 #define BUFFER_SIZE 1024
 398
 399 static int process_pty(int master, sigset_t *mask) {
 400
 401         char in_buffer[BUFFER_SIZE], out_buffer[BUFFER_SIZE];
 402         size_t in_buffer_full = 0, out_buffer_full = 0;
 403         struct epoll_event stdin_ev, stdout_ev, master_ev, signal_ev;
 404         bool stdin_readable = false, stdout_writable = false, master_readable = false, master_writable = false;
 405         int ep = -1, signal_fd = -1, r;
 406
 407         fd_nonblock(STDIN_FILENO, 1);
 408         fd_nonblock(STDOUT_FILENO, 1);
 409         fd_nonblock(master, 1);
 410
 411         if ((signal_fd = signalfd(-1, mask, SFD_NONBLOCK|SFD_CLOEXEC)) < 0) {
 412                 log_error("signalfd(): %m");
 413                 r = -errno;
 414                 goto finish;
 415         }
 416
 417         if ((ep = epoll_create1(EPOLL_CLOEXEC)) < 0) {
 418                 log_error("Failed to create epoll: %m");
 419                 r = -errno;
 420                 goto finish;
 421         }
 422
 423         zero(stdin_ev);
 424         stdin_ev.events = EPOLLIN|EPOLLET;
 425         stdin_ev.data.fd = STDIN_FILENO;
 426
 427         zero(stdout_ev);
 428         stdout_ev.events = EPOLLOUT|EPOLLET;
 429         stdout_ev.data.fd = STDOUT_FILENO;
 430
 431         zero(master_ev);
 432         master_ev.events = EPOLLIN|EPOLLOUT|EPOLLET;
 433         master_ev.data.fd = master;
 434
 435         zero(signal_ev);
 436         signal_ev.events = EPOLLIN;
 437         signal_ev.data.fd = signal_fd;
 438
 439         if (epoll_ctl(ep, EPOLL_CTL_ADD, STDIN_FILENO, &stdin_ev) < 0 ||
 440             epoll_ctl(ep, EPOLL_CTL_ADD, STDOUT_FILENO, &stdout_ev) < 0 ||
 441             epoll_ctl(ep, EPOLL_CTL_ADD, master, &master_ev) < 0 ||
 442             epoll_ctl(ep, EPOLL_CTL_ADD, signal_fd, &signal_ev) < 0) {
 443                 log_error("Failed to regiser fds in epoll: %m");
 444                 r = -errno;
 445                 goto finish;
 446         }
 447
 448         for (;;) {
 449                 struct epoll_event ev[16];
 450                 ssize_t k;
 451                 int i, nfds;
 452
 453                 if ((nfds = epoll_wait(ep, ev, ELEMENTSOF(ev), -1)) < 0) {
 454
 455                         if (errno == EINTR || errno == EAGAIN)
 456                                 continue;
 457
 458                         log_error("epoll_wait(): %m");
 459                         r = -errno;
 460                         goto finish;
 461                 }
 462
 463                 assert(nfds >= 1);
 464
 465                 for (i = 0; i < nfds; i++) {
 466                         if (ev[i].data.fd == STDIN_FILENO) {
 467
 468                                 if (ev[i].events & (EPOLLIN|EPOLLHUP))
 469                                         stdin_readable = true;
 470
 471                         } else if (ev[i].data.fd == STDOUT_FILENO) {
 472
 473                                 if (ev[i].events & (EPOLLOUT|EPOLLHUP))
 474                                         stdout_writable = true;
 475
 476                         } else if (ev[i].data.fd == master) {
 477
 478                                 if (ev[i].events & (EPOLLIN|EPOLLHUP))
 479                                         master_readable = true;
 480
 481                                 if (ev[i].events & (EPOLLOUT|EPOLLHUP))
 482                                         master_writable = true;
 483
 484                         } else if (ev[i].data.fd == signal_fd) {
 485                                 struct signalfd_siginfo sfsi;
 486                                 ssize_t n;
 487
 488                                 if ((n = read(signal_fd, &sfsi, sizeof(sfsi))) != sizeof(sfsi)) {
 489
 490                                         if (n >= 0) {
 491                                                 log_error("Failed to read from signalfd: invalid block size");
 492                                                 r = -EIO;
 493                                                 goto finish;
 494                                         }
 495
 496                                         if (errno != EINTR && errno != EAGAIN) {
 497                                                 log_error("Failed to read from signalfd: %m");
 498                                                 r = -errno;
 499                                                 goto finish;
 500                                         }
 501                                 } else {
 502
 503                                         if (sfsi.ssi_signo == SIGWINCH) {
 504                                                 struct winsize ws;
 505
 506                                                 /* The window size changed, let's forward that. */
 507                                                 if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 508                                                         ioctl(master, TIOCSWINSZ, &ws);
 509                                         } else {
 510                                                 r = 0;
 511                                                 goto finish;
 512                                         }
 513                                 }
 514                         }
 515                 }
 516
 517                 while ((stdin_readable && in_buffer_full <= 0) ||
 518                        (master_writable && in_buffer_full > 0) ||
 519                        (master_readable && out_buffer_full <= 0) ||
 520                        (stdout_writable && out_buffer_full > 0)) {
 521
 522                         if (stdin_readable && in_buffer_full < BUFFER_SIZE) {
 523
 524                                 if ((k = read(STDIN_FILENO, in_buffer + in_buffer_full, BUFFER_SIZE - in_buffer_full)) < 0) {
 525
 526                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 527                                                 stdin_readable = false;
 528                                         else {
 529                                                 log_error("read(): %m");
 530                                                 r = -errno;
 531                                                 goto finish;
 532                                         }
 533                                 } else
 534                                         in_buffer_full += (size_t) k;
 535                         }
 536
 537                         if (master_writable && in_buffer_full > 0) {
 538
 539                                 if ((k = write(master, in_buffer, in_buffer_full)) < 0) {
 540
 541                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 542                                                 master_writable = false;
 543                                         else {
 544                                                 log_error("write(): %m");
 545                                                 r = -errno;
 546                                                 goto finish;
 547                                         }
 548
 549                                 } else {
 550                                         assert(in_buffer_full >= (size_t) k);
 551                                         memmove(in_buffer, in_buffer + k, in_buffer_full - k);
 552                                         in_buffer_full -= k;
 553                                 }
 554                         }
 555
 556                         if (master_readable && out_buffer_full < BUFFER_SIZE) {
 557
 558                                 if ((k = read(master, out_buffer + out_buffer_full, BUFFER_SIZE - out_buffer_full)) < 0) {
 559
 560                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 561                                                 master_readable = false;
 562                                         else {
 563                                                 log_error("read(): %m");
 564                                                 r = -errno;
 565                                                 goto finish;
 566                                         }
 567                                 }  else
 568                                         out_buffer_full += (size_t) k;
 569                         }
 570
 571                         if (stdout_writable && out_buffer_full > 0) {
 572
 573                                 if ((k = write(STDOUT_FILENO, out_buffer, out_buffer_full)) < 0) {
 574
 575                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 576                                                 stdout_writable = false;
 577                                         else {
 578                                                 log_error("write(): %m");
 579                                                 r = -errno;
 580                                                 goto finish;
 581                                         }
 582
 583                                 } else {
 584                                         assert(out_buffer_full >= (size_t) k);
 585                                         memmove(out_buffer, out_buffer + k, out_buffer_full - k);
 586                                         out_buffer_full -= k;
 587                                 }
 588                         }
 589                 }
 590         }
 591
 592 finish:
 593         if (ep >= 0)
 594                 close_nointr_nofail(ep);
 595
 596         if (signal_fd >= 0)
 597                 close_nointr_nofail(signal_fd);
 598
 599         return r;
 600 }
 601
 602 int main(int argc, char *argv[]) {
 603         pid_t pid = 0;
 604         int r = EXIT_FAILURE, k;
 605         char *oldcg = NULL, *newcg = NULL;
 606         int master = -1;
 607         const char *console = NULL;
 608         struct termios saved_attr, raw_attr;
 609         sigset_t mask;
 610         bool saved_attr_valid = false;
 611         struct winsize ws;
 612
 613         log_parse_environment();
 614         log_open();
 615
 616         if ((r = parse_argv(argc, argv)) <= 0)
 617                 goto finish;
 618
 619         if (arg_directory) {
 620                 char *p;
 621
 622                 p = path_make_absolute_cwd(arg_directory);
 623                 free(arg_directory);
 624                 arg_directory = p;
 625         } else
 626                 arg_directory = get_current_dir_name();
 627
 628         if (!arg_directory) {
 629                 log_error("Failed to determine path");
 630                 goto finish;
 631         }
 632
 633         path_kill_slashes(arg_directory);
 634
 635         if (geteuid() != 0) {
 636                 log_error("Need to be root.");
 637                 goto finish;
 638         }
 639
 640         if (sd_booted() <= 0) {
 641                 log_error("Not running on a systemd system.");
 642                 goto finish;
 643         }
 644
 645         if (path_equal(arg_directory, "/")) {
 646                 log_error("Spawning container on root directory not supported.");
 647                 goto finish;
 648         }
 649
 650         if (is_os_tree(arg_directory) <= 0) {
 651                 log_error("Directory %s doesn't look like an OS root directory. Refusing.", arg_directory);
 652                 goto finish;
 653         }
 654
 655         if ((k = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 0, &oldcg)) < 0) {
 656                 log_error("Failed to determine current cgroup: %s", strerror(-k));
 657                 goto finish;
 658         }
 659
 660         if (asprintf(&newcg, "%s/nspawn-%lu", oldcg, (unsigned long) getpid()) < 0) {
 661                 log_error("Failed to allocate cgroup path.");
 662                 goto finish;
 663         }
 664
 665         if ((k = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, newcg, 0)) < 0)  {
 666                 log_error("Failed to create cgroup: %s", strerror(-k));
 667                 goto finish;
 668         }
 669
 670         if ((master = posix_openpt(O_RDWR|O_NOCTTY|O_CLOEXEC|O_NDELAY)) < 0) {
 671                 log_error("Failed to acquire pseudo tty: %m");
 672                 goto finish;
 673         }
 674
 675         if (!(console = ptsname(master))) {
 676                 log_error("Failed to determine tty name: %m");
 677                 goto finish;
 678         }
 679
 680         log_info("Spawning namespace container on %s (console is %s).", arg_directory, console);
 681
 682         if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 683                 ioctl(master, TIOCSWINSZ, &ws);
 684
 685         if (unlockpt(master) < 0) {
 686                 log_error("Failed to unlock tty: %m");
 687                 goto finish;
 688         }
 689
 690         if (tcgetattr(STDIN_FILENO, &saved_attr) < 0) {
 691                 log_error("Failed to get terminal attributes: %m");
 692                 goto finish;
 693         }
 694
 695         saved_attr_valid = true;
 696
 697         raw_attr = saved_attr;
 698         cfmakeraw(&raw_attr);
 699         raw_attr.c_lflag &= ~ECHO;
 700
 701         if (tcsetattr(STDIN_FILENO, TCSANOW, &raw_attr) < 0) {
 702                 log_error("Failed to set terminal attributes: %m");
 703                 goto finish;
 704         }
 705
 706         assert_se(sigemptyset(&mask) == 0);
 707         sigset_add_many(&mask, SIGCHLD, SIGWINCH, SIGTERM, SIGINT, -1);
 708         assert_se(sigprocmask(SIG_BLOCK, &mask, NULL) == 0);
 709
 710         if ((pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS|(arg_private_network ? CLONE_NEWNET : 0), NULL)) < 0) {
 711                 log_error("clone() failed: %m");
 712                 goto finish;
 713         }
 714
 715         if (pid == 0) {
 716                 /* child */
 717
 718                 const char *hn;
 719                 const char *home = NULL;
 720                 uid_t uid = (uid_t) -1;
 721                 gid_t gid = (gid_t) -1;
 722                 const char *envp[] = {
 723                         "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
 724                         "container=systemd-nspawn", /* LXC sets container=lxc, so follow the scheme here */
 725                         NULL, /* TERM */
 726                         NULL, /* HOME */
 727                         NULL, /* USER */
 728                         NULL, /* LOGNAME */
 729                         NULL
 730                 };
 731
 732                 envp[2] = strv_find_prefix(environ, "TERM=");
 733
 734                 close_nointr_nofail(master);
 735
 736                 close_nointr(STDIN_FILENO);
 737                 close_nointr(STDOUT_FILENO);
 738                 close_nointr(STDERR_FILENO);
 739
 740                 close_all_fds(NULL, 0);
 741
 742                 reset_all_signal_handlers();
 743
 744                 assert_se(sigemptyset(&mask) == 0);
 745                 assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
 746
 747                 if (setsid() < 0)
 748                         goto child_fail;
 749
 750                 if (prctl(PR_SET_PDEATHSIG, SIGKILL) < 0)
 751                         goto child_fail;
 752
 753                 /* Mark / as private, in case somebody marked it shared */
 754                 if (mount(NULL, "/", NULL, MS_PRIVATE|MS_REC, NULL) < 0)
 755                         goto child_fail;
 756
 757                 if (mount_all(arg_directory) < 0)
 758                         goto child_fail;
 759
 760                 if (copy_devnodes(arg_directory, console) < 0)
 761                         goto child_fail;
 762
 763                 if (chdir(arg_directory) < 0) {
 764                         log_error("chdir(%s) failed: %m", arg_directory);
 765                         goto child_fail;
 766                 }
 767
 768                 if (open_terminal("dev/console", O_RDWR) != STDIN_FILENO ||
 769                     dup2(STDIN_FILENO, STDOUT_FILENO) != STDOUT_FILENO ||
 770                     dup2(STDIN_FILENO, STDERR_FILENO) != STDERR_FILENO)
 771                         goto child_fail;
 772
 773                 if (mount(arg_directory, "/", "bind", MS_BIND|MS_MOVE, NULL) < 0) {
 774                         log_error("mount(MS_MOVE) failed: %m");
 775                         goto child_fail;
 776                 }
 777
 778                 if (chroot(".") < 0) {
 779                         log_error("chroot() failed: %m");
 780                         goto child_fail;
 781                 }
 782
 783                 if (chdir("/") < 0) {
 784                         log_error("chdir() failed: %m");
 785                         goto child_fail;
 786                 }
 787
 788                 umask(0022);
 789
 790                 loopback_setup();
 791
 792                 if (drop_capabilities() < 0)
 793                         goto child_fail;
 794
 795                 if (arg_user) {
 796
 797                         if (get_user_creds((const char**)&arg_user, &uid, &gid, &home) < 0) {
 798                                 log_error("get_user_creds() failed: %m");
 799                                 goto child_fail;
 800                         }
 801
 802                         if (mkdir_parents(home, 0775) < 0) {
 803                                 log_error("mkdir_parents() failed: %m");
 804                                 goto child_fail;
 805                         }
 806
 807                         if (safe_mkdir(home, 0775, uid, gid) < 0) {
 808                                 log_error("safe_mkdir() failed: %m");
 809                                 goto child_fail;
 810                         }
 811
 812                         if (initgroups((const char*)arg_user, gid) < 0) {
 813                                 log_error("initgroups() failed: %m");
 814                                 goto child_fail;
 815                         }
 816
 817                         if (setresgid(gid, gid, gid) < 0) {
 818                                 log_error("setregid() failed: %m");
 819                                 goto child_fail;
 820                         }
 821
 822                         if (setresuid(uid, uid, uid) < 0) {
 823                                 log_error("setreuid() failed: %m");
 824                                 goto child_fail;
 825                         }
 826                 }
 827
 828                 if ((asprintf((char**)(envp + 3), "HOME=%s", home? home: "/root") < 0) ||
 829                     (asprintf((char**)(envp + 4), "USER=%s", arg_user? arg_user : "root") < 0) ||
 830                     (asprintf((char**)(envp + 5), "LOGNAME=%s", arg_user? arg_user : "root") < 0)) {
 831                     log_error("Out of memory");
 832                     goto child_fail;
 833                 }
 834
 835                 if ((hn = file_name_from_path(arg_directory)))
 836                         sethostname(hn, strlen(hn));
 837
 838                 if (argc > optind)
 839                         execvpe(argv[optind], argv + optind, (char**) envp);
 840                 else {
 841                         chdir(home ? home : "/root");
 842                         execle("/bin/bash", "-bash", NULL, (char**) envp);
 843                 }
 844
 845                 log_error("execv() failed: %m");
 846
 847         child_fail:
 848                 _exit(EXIT_FAILURE);
 849         }
 850
 851         if (process_pty(master, &mask) < 0)
 852                 goto finish;
 853
 854         if (saved_attr_valid) {
 855                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 856                 saved_attr_valid = false;
 857         }
 858
 859         r = wait_for_terminate_and_warn(argc > optind ? argv[optind] : "bash", pid);
 860
 861         if (r < 0)
 862                 r = EXIT_FAILURE;
 863
 864 finish:
 865         if (saved_attr_valid)
 866                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 867
 868         if (master >= 0)
 869                 close_nointr_nofail(master);
 870
 871         if (oldcg)
 872                 cg_attach(SYSTEMD_CGROUP_CONTROLLER, oldcg, 0);
 873
 874         if (newcg)
 875                 cg_kill_recursive_and_wait(SYSTEMD_CGROUP_CONTROLLER, newcg, true);
 876
 877         free(arg_directory);
 878         free(oldcg);
 879         free(newcg);
 880
 881         return r;
 882 }