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 <= MAX(63LU, (unsigned long) 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
 376                         /* If this capability is not known, EINVAL
 377                          * will be returned, let's ignore this. */
 378                         if (errno == EINVAL)
 379                                 break;
 380
 381                         log_error("PR_CAPBSET_DROP failed: %m");
 382                         return -errno;
 383                 }
 384         }
 385
 386         return 0;
 387 }
 388
 389 static int is_os_tree(const char *path) {
 390         int r;
 391         char *p;
 392         /* We use /bin/sh as flag file if something is an OS */
 393
 394         if (asprintf(&p, "%s/bin/sh", path) < 0)
 395                 return -ENOMEM;
 396
 397         r = access(p, F_OK);
 398         free(p);
 399
 400         return r < 0 ? 0 : 1;
 401 }
 402
 403 #define BUFFER_SIZE 1024
 404
 405 static int process_pty(int master, sigset_t *mask) {
 406
 407         char in_buffer[BUFFER_SIZE], out_buffer[BUFFER_SIZE];
 408         size_t in_buffer_full = 0, out_buffer_full = 0;
 409         struct epoll_event stdin_ev, stdout_ev, master_ev, signal_ev;
 410         bool stdin_readable = false, stdout_writable = false, master_readable = false, master_writable = false;
 411         int ep = -1, signal_fd = -1, r;
 412
 413         fd_nonblock(STDIN_FILENO, 1);
 414         fd_nonblock(STDOUT_FILENO, 1);
 415         fd_nonblock(master, 1);
 416
 417         if ((signal_fd = signalfd(-1, mask, SFD_NONBLOCK|SFD_CLOEXEC)) < 0) {
 418                 log_error("signalfd(): %m");
 419                 r = -errno;
 420                 goto finish;
 421         }
 422
 423         if ((ep = epoll_create1(EPOLL_CLOEXEC)) < 0) {
 424                 log_error("Failed to create epoll: %m");
 425                 r = -errno;
 426                 goto finish;
 427         }
 428
 429         zero(stdin_ev);
 430         stdin_ev.events = EPOLLIN|EPOLLET;
 431         stdin_ev.data.fd = STDIN_FILENO;
 432
 433         zero(stdout_ev);
 434         stdout_ev.events = EPOLLOUT|EPOLLET;
 435         stdout_ev.data.fd = STDOUT_FILENO;
 436
 437         zero(master_ev);
 438         master_ev.events = EPOLLIN|EPOLLOUT|EPOLLET;
 439         master_ev.data.fd = master;
 440
 441         zero(signal_ev);
 442         signal_ev.events = EPOLLIN;
 443         signal_ev.data.fd = signal_fd;
 444
 445         if (epoll_ctl(ep, EPOLL_CTL_ADD, STDIN_FILENO, &stdin_ev) < 0 ||
 446             epoll_ctl(ep, EPOLL_CTL_ADD, STDOUT_FILENO, &stdout_ev) < 0 ||
 447             epoll_ctl(ep, EPOLL_CTL_ADD, master, &master_ev) < 0 ||
 448             epoll_ctl(ep, EPOLL_CTL_ADD, signal_fd, &signal_ev) < 0) {
 449                 log_error("Failed to regiser fds in epoll: %m");
 450                 r = -errno;
 451                 goto finish;
 452         }
 453
 454         for (;;) {
 455                 struct epoll_event ev[16];
 456                 ssize_t k;
 457                 int i, nfds;
 458
 459                 if ((nfds = epoll_wait(ep, ev, ELEMENTSOF(ev), -1)) < 0) {
 460
 461                         if (errno == EINTR || errno == EAGAIN)
 462                                 continue;
 463
 464                         log_error("epoll_wait(): %m");
 465                         r = -errno;
 466                         goto finish;
 467                 }
 468
 469                 assert(nfds >= 1);
 470
 471                 for (i = 0; i < nfds; i++) {
 472                         if (ev[i].data.fd == STDIN_FILENO) {
 473
 474                                 if (ev[i].events & (EPOLLIN|EPOLLHUP))
 475                                         stdin_readable = true;
 476
 477                         } else if (ev[i].data.fd == STDOUT_FILENO) {
 478
 479                                 if (ev[i].events & (EPOLLOUT|EPOLLHUP))
 480                                         stdout_writable = true;
 481
 482                         } else if (ev[i].data.fd == master) {
 483
 484                                 if (ev[i].events & (EPOLLIN|EPOLLHUP))
 485                                         master_readable = true;
 486
 487                                 if (ev[i].events & (EPOLLOUT|EPOLLHUP))
 488                                         master_writable = true;
 489
 490                         } else if (ev[i].data.fd == signal_fd) {
 491                                 struct signalfd_siginfo sfsi;
 492                                 ssize_t n;
 493
 494                                 if ((n = read(signal_fd, &sfsi, sizeof(sfsi))) != sizeof(sfsi)) {
 495
 496                                         if (n >= 0) {
 497                                                 log_error("Failed to read from signalfd: invalid block size");
 498                                                 r = -EIO;
 499                                                 goto finish;
 500                                         }
 501
 502                                         if (errno != EINTR && errno != EAGAIN) {
 503                                                 log_error("Failed to read from signalfd: %m");
 504                                                 r = -errno;
 505                                                 goto finish;
 506                                         }
 507                                 } else {
 508
 509                                         if (sfsi.ssi_signo == SIGWINCH) {
 510                                                 struct winsize ws;
 511
 512                                                 /* The window size changed, let's forward that. */
 513                                                 if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 514                                                         ioctl(master, TIOCSWINSZ, &ws);
 515                                         } else {
 516                                                 r = 0;
 517                                                 goto finish;
 518                                         }
 519                                 }
 520                         }
 521                 }
 522
 523                 while ((stdin_readable && in_buffer_full <= 0) ||
 524                        (master_writable && in_buffer_full > 0) ||
 525                        (master_readable && out_buffer_full <= 0) ||
 526                        (stdout_writable && out_buffer_full > 0)) {
 527
 528                         if (stdin_readable && in_buffer_full < BUFFER_SIZE) {
 529
 530                                 if ((k = read(STDIN_FILENO, in_buffer + in_buffer_full, BUFFER_SIZE - in_buffer_full)) < 0) {
 531
 532                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 533                                                 stdin_readable = false;
 534                                         else {
 535                                                 log_error("read(): %m");
 536                                                 r = -errno;
 537                                                 goto finish;
 538                                         }
 539                                 } else
 540                                         in_buffer_full += (size_t) k;
 541                         }
 542
 543                         if (master_writable && in_buffer_full > 0) {
 544
 545                                 if ((k = write(master, in_buffer, in_buffer_full)) < 0) {
 546
 547                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 548                                                 master_writable = false;
 549                                         else {
 550                                                 log_error("write(): %m");
 551                                                 r = -errno;
 552                                                 goto finish;
 553                                         }
 554
 555                                 } else {
 556                                         assert(in_buffer_full >= (size_t) k);
 557                                         memmove(in_buffer, in_buffer + k, in_buffer_full - k);
 558                                         in_buffer_full -= k;
 559                                 }
 560                         }
 561
 562                         if (master_readable && out_buffer_full < BUFFER_SIZE) {
 563
 564                                 if ((k = read(master, out_buffer + out_buffer_full, BUFFER_SIZE - out_buffer_full)) < 0) {
 565
 566                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 567                                                 master_readable = false;
 568                                         else {
 569                                                 log_error("read(): %m");
 570                                                 r = -errno;
 571                                                 goto finish;
 572                                         }
 573                                 }  else
 574                                         out_buffer_full += (size_t) k;
 575                         }
 576
 577                         if (stdout_writable && out_buffer_full > 0) {
 578
 579                                 if ((k = write(STDOUT_FILENO, out_buffer, out_buffer_full)) < 0) {
 580
 581                                         if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
 582                                                 stdout_writable = false;
 583                                         else {
 584                                                 log_error("write(): %m");
 585                                                 r = -errno;
 586                                                 goto finish;
 587                                         }
 588
 589                                 } else {
 590                                         assert(out_buffer_full >= (size_t) k);
 591                                         memmove(out_buffer, out_buffer + k, out_buffer_full - k);
 592                                         out_buffer_full -= k;
 593                                 }
 594                         }
 595                 }
 596         }
 597
 598 finish:
 599         if (ep >= 0)
 600                 close_nointr_nofail(ep);
 601
 602         if (signal_fd >= 0)
 603                 close_nointr_nofail(signal_fd);
 604
 605         return r;
 606 }
 607
 608 int main(int argc, char *argv[]) {
 609         pid_t pid = 0;
 610         int r = EXIT_FAILURE, k;
 611         char *oldcg = NULL, *newcg = NULL;
 612         int master = -1;
 613         const char *console = NULL;
 614         struct termios saved_attr, raw_attr;
 615         sigset_t mask;
 616         bool saved_attr_valid = false;
 617         struct winsize ws;
 618
 619         log_parse_environment();
 620         log_open();
 621
 622         if ((r = parse_argv(argc, argv)) <= 0)
 623                 goto finish;
 624
 625         if (arg_directory) {
 626                 char *p;
 627
 628                 p = path_make_absolute_cwd(arg_directory);
 629                 free(arg_directory);
 630                 arg_directory = p;
 631         } else
 632                 arg_directory = get_current_dir_name();
 633
 634         if (!arg_directory) {
 635                 log_error("Failed to determine path");
 636                 goto finish;
 637         }
 638
 639         path_kill_slashes(arg_directory);
 640
 641         if (geteuid() != 0) {
 642                 log_error("Need to be root.");
 643                 goto finish;
 644         }
 645
 646         if (sd_booted() <= 0) {
 647                 log_error("Not running on a systemd system.");
 648                 goto finish;
 649         }
 650
 651         if (path_equal(arg_directory, "/")) {
 652                 log_error("Spawning container on root directory not supported.");
 653                 goto finish;
 654         }
 655
 656         if (is_os_tree(arg_directory) <= 0) {
 657                 log_error("Directory %s doesn't look like an OS root directory. Refusing.", arg_directory);
 658                 goto finish;
 659         }
 660
 661         if ((k = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 0, &oldcg)) < 0) {
 662                 log_error("Failed to determine current cgroup: %s", strerror(-k));
 663                 goto finish;
 664         }
 665
 666         if (asprintf(&newcg, "%s/nspawn-%lu", oldcg, (unsigned long) getpid()) < 0) {
 667                 log_error("Failed to allocate cgroup path.");
 668                 goto finish;
 669         }
 670
 671         if ((k = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, newcg, 0)) < 0)  {
 672                 log_error("Failed to create cgroup: %s", strerror(-k));
 673                 goto finish;
 674         }
 675
 676         if ((master = posix_openpt(O_RDWR|O_NOCTTY|O_CLOEXEC|O_NDELAY)) < 0) {
 677                 log_error("Failed to acquire pseudo tty: %m");
 678                 goto finish;
 679         }
 680
 681         if (!(console = ptsname(master))) {
 682                 log_error("Failed to determine tty name: %m");
 683                 goto finish;
 684         }
 685
 686         log_info("Spawning namespace container on %s (console is %s).", arg_directory, console);
 687
 688         if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 689                 ioctl(master, TIOCSWINSZ, &ws);
 690
 691         if (unlockpt(master) < 0) {
 692                 log_error("Failed to unlock tty: %m");
 693                 goto finish;
 694         }
 695
 696         if (tcgetattr(STDIN_FILENO, &saved_attr) < 0) {
 697                 log_error("Failed to get terminal attributes: %m");
 698                 goto finish;
 699         }
 700
 701         saved_attr_valid = true;
 702
 703         raw_attr = saved_attr;
 704         cfmakeraw(&raw_attr);
 705         raw_attr.c_lflag &= ~ECHO;
 706
 707         if (tcsetattr(STDIN_FILENO, TCSANOW, &raw_attr) < 0) {
 708                 log_error("Failed to set terminal attributes: %m");
 709                 goto finish;
 710         }
 711
 712         assert_se(sigemptyset(&mask) == 0);
 713         sigset_add_many(&mask, SIGCHLD, SIGWINCH, SIGTERM, SIGINT, -1);
 714         assert_se(sigprocmask(SIG_BLOCK, &mask, NULL) == 0);
 715
 716         if ((pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS|(arg_private_network ? CLONE_NEWNET : 0), NULL)) < 0) {
 717                 log_error("clone() failed: %m");
 718                 goto finish;
 719         }
 720
 721         if (pid == 0) {
 722                 /* child */
 723
 724                 const char *hn;
 725                 const char *home = NULL;
 726                 uid_t uid = (uid_t) -1;
 727                 gid_t gid = (gid_t) -1;
 728                 const char *envp[] = {
 729                         "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
 730                         "container=systemd-nspawn", /* LXC sets container=lxc, so follow the scheme here */
 731                         NULL, /* TERM */
 732                         NULL, /* HOME */
 733                         NULL, /* USER */
 734                         NULL, /* LOGNAME */
 735                         NULL
 736                 };
 737
 738                 envp[2] = strv_find_prefix(environ, "TERM=");
 739
 740                 close_nointr_nofail(master);
 741
 742                 close_nointr(STDIN_FILENO);
 743                 close_nointr(STDOUT_FILENO);
 744                 close_nointr(STDERR_FILENO);
 745
 746                 close_all_fds(NULL, 0);
 747
 748                 reset_all_signal_handlers();
 749
 750                 assert_se(sigemptyset(&mask) == 0);
 751                 assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
 752
 753                 if (setsid() < 0)
 754                         goto child_fail;
 755
 756                 if (prctl(PR_SET_PDEATHSIG, SIGKILL) < 0)
 757                         goto child_fail;
 758
 759                 /* Mark / as private, in case somebody marked it shared */
 760                 if (mount(NULL, "/", NULL, MS_PRIVATE|MS_REC, NULL) < 0)
 761                         goto child_fail;
 762
 763                 if (mount_all(arg_directory) < 0)
 764                         goto child_fail;
 765
 766                 if (copy_devnodes(arg_directory, console) < 0)
 767                         goto child_fail;
 768
 769                 if (chdir(arg_directory) < 0) {
 770                         log_error("chdir(%s) failed: %m", arg_directory);
 771                         goto child_fail;
 772                 }
 773
 774                 if (open_terminal("dev/console", O_RDWR) != STDIN_FILENO ||
 775                     dup2(STDIN_FILENO, STDOUT_FILENO) != STDOUT_FILENO ||
 776                     dup2(STDIN_FILENO, STDERR_FILENO) != STDERR_FILENO)
 777                         goto child_fail;
 778
 779                 if (mount(arg_directory, "/", "bind", MS_BIND|MS_MOVE, NULL) < 0) {
 780                         log_error("mount(MS_MOVE) failed: %m");
 781                         goto child_fail;
 782                 }
 783
 784                 if (chroot(".") < 0) {
 785                         log_error("chroot() failed: %m");
 786                         goto child_fail;
 787                 }
 788
 789                 if (chdir("/") < 0) {
 790                         log_error("chdir() failed: %m");
 791                         goto child_fail;
 792                 }
 793
 794                 umask(0022);
 795
 796                 loopback_setup();
 797
 798                 if (drop_capabilities() < 0)
 799                         goto child_fail;
 800
 801                 if (arg_user) {
 802
 803                         if (get_user_creds((const char**)&arg_user, &uid, &gid, &home) < 0) {
 804                                 log_error("get_user_creds() failed: %m");
 805                                 goto child_fail;
 806                         }
 807
 808                         if (mkdir_parents(home, 0775) < 0) {
 809                                 log_error("mkdir_parents() failed: %m");
 810                                 goto child_fail;
 811                         }
 812
 813                         if (safe_mkdir(home, 0775, uid, gid) < 0) {
 814                                 log_error("safe_mkdir() failed: %m");
 815                                 goto child_fail;
 816                         }
 817
 818                         if (initgroups((const char*)arg_user, gid) < 0) {
 819                                 log_error("initgroups() failed: %m");
 820                                 goto child_fail;
 821                         }
 822
 823                         if (setresgid(gid, gid, gid) < 0) {
 824                                 log_error("setregid() failed: %m");
 825                                 goto child_fail;
 826                         }
 827
 828                         if (setresuid(uid, uid, uid) < 0) {
 829                                 log_error("setreuid() failed: %m");
 830                                 goto child_fail;
 831                         }
 832                 }
 833
 834                 if ((asprintf((char**)(envp + 3), "HOME=%s", home? home: "/root") < 0) ||
 835                     (asprintf((char**)(envp + 4), "USER=%s", arg_user? arg_user : "root") < 0) ||
 836                     (asprintf((char**)(envp + 5), "LOGNAME=%s", arg_user? arg_user : "root") < 0)) {
 837                     log_error("Out of memory");
 838                     goto child_fail;
 839                 }
 840
 841                 if ((hn = file_name_from_path(arg_directory)))
 842                         sethostname(hn, strlen(hn));
 843
 844                 if (argc > optind)
 845                         execvpe(argv[optind], argv + optind, (char**) envp);
 846                 else {
 847                         chdir(home ? home : "/root");
 848                         execle("/bin/bash", "-bash", NULL, (char**) envp);
 849                 }
 850
 851                 log_error("execv() failed: %m");
 852
 853         child_fail:
 854                 _exit(EXIT_FAILURE);
 855         }
 856
 857         if (process_pty(master, &mask) < 0)
 858                 goto finish;
 859
 860         if (saved_attr_valid) {
 861                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 862                 saved_attr_valid = false;
 863         }
 864
 865         r = wait_for_terminate_and_warn(argc > optind ? argv[optind] : "bash", pid);
 866
 867         if (r < 0)
 868                 r = EXIT_FAILURE;
 869
 870 finish:
 871         if (saved_attr_valid)
 872                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 873
 874         if (master >= 0)
 875                 close_nointr_nofail(master);
 876
 877         if (oldcg)
 878                 cg_attach(SYSTEMD_CGROUP_CONTROLLER, oldcg, 0);
 879
 880         if (newcg)
 881                 cg_kill_recursive_and_wait(SYSTEMD_CGROUP_CONTROLLER, newcg, true);
 882
 883         free(arg_directory);
 884         free(oldcg);
 885         free(newcg);
 886
 887         return r;
 888 }