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