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
  40 #include "log.h"
  41 #include "util.h"
  42 #include "missing.h"
  43 #include "cgroup-util.h"
  44 #include "sd-daemon.h"
  45 #include "strv.h"
  46
  47 static char *arg_directory = NULL;
  48
  49 static int help(void) {
  50
  51         printf("%s [OPTIONS...] [PATH] [ARGUMENTS...]\n\n"
  52                "Spawn a minimal namespace container for debugging, testing and building.\n\n"
  53                "  -h --help            Show this help\n"
  54                "  -D --directory=NAME  Root directory for the container\n",
  55                program_invocation_short_name);
  56
  57         return 0;
  58 }
  59
  60 static int parse_argv(int argc, char *argv[]) {
  61
  62         static const struct option options[] = {
  63                 { "help",      no_argument,       NULL, 'h' },
  64                 { "directory", required_argument, NULL, 'D' },
  65                 { NULL,        0,                 NULL, 0   }
  66         };
  67
  68         int c;
  69
  70         assert(argc >= 0);
  71         assert(argv);
  72
  73         while ((c = getopt_long(argc, argv, "+hD:", options, NULL)) >= 0) {
  74
  75                 switch (c) {
  76
  77                 case 'h':
  78                         help();
  79                         return 0;
  80
  81                 case 'D':
  82                         free(arg_directory);
  83                         if (!(arg_directory = strdup(optarg))) {
  84                                 log_error("Failed to duplicate root directory.");
  85                                 return -ENOMEM;
  86                         }
  87
  88                         break;
  89
  90                 case '?':
  91                         return -EINVAL;
  92
  93                 default:
  94                         log_error("Unknown option code %c", c);
  95                         return -EINVAL;
  96                 }
  97         }
  98
  99         return 1;
 100 }
 101
 102 static int mount_all(const char *dest) {
 103
 104         typedef struct MountPoint {
 105                 const char *what;
 106                 const char *where;
 107                 const char *type;
 108                 const char *options;
 109                 unsigned long flags;
 110                 bool fatal;
 111         } MountPoint;
 112
 113         static const MountPoint mount_table[] = {
 114                 { "proc",      "/proc",     "proc",      NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV, true },
 115                 { "/proc/sys", "/proc/sys", "bind",      NULL,        MS_BIND, true },                      /* Bind mount first */
 116                 { "/proc/sys", "/proc/sys", "bind",      NULL,        MS_BIND|MS_RDONLY|MS_REMOUNT, true }, /* Then, make it r/o */
 117                 { "sysfs",     "/sys",      "sysfs",     NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, true },
 118                 { "tmpfs",     "/dev",      "tmpfs",     "mode=755",  MS_NOSUID, true },
 119                 { "/dev/pts",  "/dev/pts",  "bind",      NULL,        MS_BIND, true },
 120                 { "tmpfs",     "/dev/.run", "tmpfs",     "mode=755",  MS_NOSUID|MS_NOEXEC|MS_NODEV, true },
 121 #ifdef HAVE_SELINUX
 122                 { "selinux",   "/selinux",  "selinuxfs", NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, false },
 123 #endif
 124         };
 125
 126         unsigned k;
 127         int r = 0;
 128
 129         for (k = 0; k < ELEMENTSOF(mount_table); k++) {
 130                 char *where;
 131                 int t;
 132
 133                 if (asprintf(&where, "%s/%s", dest, mount_table[k].where) < 0) {
 134                         log_error("Out of memory");
 135
 136                         if (r == 0)
 137                                 r = -ENOMEM;
 138
 139                         break;
 140                 }
 141
 142                 if ((t = path_is_mount_point(where)) < 0) {
 143                         log_error("Failed to detect whether %s is a mount point: %s", where, strerror(-t));
 144                         free(where);
 145
 146                         if (r == 0)
 147                                 r = t;
 148
 149                         continue;
 150                 }
 151
 152                 mkdir_p(where, 0755);
 153
 154                 if (mount(mount_table[k].what,
 155                           where,
 156                           mount_table[k].type,
 157                           mount_table[k].flags,
 158                           mount_table[k].options) < 0 &&
 159                     mount_table[k].fatal) {
 160
 161                         log_error("mount(%s) failed: %m", where);
 162
 163                         if (r == 0)
 164                                 r = -errno;
 165                 }
 166
 167                 free(where);
 168         }
 169
 170         return r;
 171 }
 172
 173 static int copy_devnodes(const char *dest, const char *console) {
 174
 175         static const char devnodes[] =
 176                 "null\0"
 177                 "zero\0"
 178                 "full\0"
 179                 "random\0"
 180                 "urandom\0"
 181                 "tty\0"
 182                 "ptmx\0"
 183                 "kmsg\0"
 184                 "rtc0\0";
 185
 186         const char *d;
 187         int r = 0, k;
 188         mode_t u;
 189         struct stat st;
 190         char *from = NULL, *to = NULL;
 191
 192         assert(dest);
 193         assert(console);
 194
 195         u = umask(0000);
 196
 197         NULSTR_FOREACH(d, devnodes) {
 198                 from = to = NULL;
 199
 200                 asprintf(&from, "/dev/%s", d);
 201                 asprintf(&to, "%s/dev/%s", dest, d);
 202
 203                 if (!from || !to) {
 204                         log_error("Failed to allocate devnode path");
 205
 206                         free(from);
 207                         free(to);
 208
 209                         from = to = NULL;
 210
 211                         if (r == 0)
 212                                 r = -ENOMEM;
 213
 214                         break;
 215                 }
 216
 217                 if (stat(from, &st) < 0) {
 218
 219                         if (errno != ENOENT) {
 220                                 log_error("Failed to stat %s: %m", from);
 221                                 if (r == 0)
 222                                         r = -errno;
 223                         }
 224
 225                 } else if (!S_ISCHR(st.st_mode) && !S_ISBLK(st.st_mode)) {
 226
 227                         log_error("%s is not a char or block device, cannot copy.", from);
 228                         if (r == 0)
 229                                 r = -EIO;
 230
 231                 } else if (mknod(to, st.st_mode, st.st_rdev) < 0) {
 232
 233                         log_error("mknod(%s) failed: %m", dest);
 234                         if (r == 0)
 235                                 r = -errno;
 236                 }
 237
 238                 free(from);
 239                 free(to);
 240         }
 241
 242         if (stat(console, &st) < 0) {
 243
 244                 log_error("Failed to stat %s: %m", console);
 245                 if (r == 0)
 246                         r = -errno;
 247
 248                 goto finish;
 249
 250         } else if (!S_ISCHR(st.st_mode)) {
 251
 252                 log_error("/dev/console is not a char device.");
 253                 if (r == 0)
 254                         r = -EIO;
 255
 256                 goto finish;
 257         }
 258
 259         if (asprintf(&to, "%s/dev/console", dest) < 0) {
 260
 261                 log_error("Out of memory");
 262                 if (r == 0)
 263                         r = -ENOMEM;
 264
 265                  goto finish;
 266         }
 267
 268         /* We need to bind mount the right tty to /dev/console since
 269          * ptys can only exist on pts file systems. To have something
 270          * to bind mount things on we create a device node first, that
 271          * has the right major/minor (note that the major minor
 272          * doesn't actually matter here, since we mount it over
 273          * anyway). */
 274
 275         if (mknod(to, (st.st_mode & ~07777) | 0600, st.st_rdev) < 0)
 276                 log_error("mknod for /dev/console failed: %m");
 277
 278         if (mount(console, to, "bind", MS_BIND, NULL) < 0) {
 279                 log_error("bind mount for /dev/console failed: %m");
 280
 281                 if (r == 0)
 282                         r = -errno;
 283         }
 284
 285         free(to);
 286
 287         if ((k = chmod_and_chown(console, 0600, 0, 0)) < 0) {
 288                 log_error("Failed to correct access mode for TTY: %s", strerror(-k));
 289
 290                 if (r == 0)
 291                         r = k;
 292         }
 293
 294 finish:
 295
 296         umask(u);
 297
 298         return r;
 299 }
 300
 301 static int drop_capabilities(void) {
 302         static const unsigned long retain[] = {
 303                 CAP_CHOWN,
 304                 CAP_DAC_OVERRIDE,
 305                 CAP_DAC_READ_SEARCH,
 306                 CAP_FOWNER,
 307                 CAP_FSETID,
 308                 CAP_IPC_OWNER,
 309                 CAP_KILL,
 310                 CAP_LEASE,
 311                 CAP_LINUX_IMMUTABLE,
 312                 CAP_NET_BIND_SERVICE,
 313                 CAP_NET_BROADCAST,
 314                 CAP_NET_RAW,
 315                 CAP_SETGID,
 316                 CAP_SETFCAP,
 317                 CAP_SETPCAP,
 318                 CAP_SETUID,
 319                 CAP_SYS_ADMIN,
 320                 CAP_SYS_CHROOT,
 321                 CAP_SYS_NICE,
 322                 CAP_SYS_PTRACE,
 323                 CAP_SYS_TTY_CONFIG
 324         };
 325
 326         unsigned long l;
 327
 328         for (l = 0; l <= MAX(63LU, (unsigned long) CAP_LAST_CAP); l ++) {
 329                 unsigned i;
 330
 331                 for (i = 0; i < ELEMENTSOF(retain); i++)
 332                         if (retain[i] == l)
 333                                 break;
 334
 335                 if (i < ELEMENTSOF(retain))
 336                         continue;
 337
 338                 if (prctl(PR_CAPBSET_DROP, l) < 0) {
 339
 340                         /* If this capability is not known, EINVAL
 341                          * will be returned, let's ignore this. */
 342                         if (errno == EINVAL)
 343                                 continue;
 344
 345                         log_error("PR_CAPBSET_DROP failed: %m");
 346                         return -errno;
 347                 }
 348         }
 349
 350         return 0;
 351 }
 352
 353 static int is_os_tree(const char *path) {
 354         int r;
 355         char *p;
 356         /* We use /bin/sh as flag file if something is an OS */
 357
 358         if (asprintf(&p, "%s/bin/sh", path) < 0)
 359                 return -ENOMEM;
 360
 361         r = access(p, F_OK);
 362         free(p);
 363
 364         return r < 0 ? 0 : 1;
 365 }
 366
 367 #define BUFFER_SIZE 1024
 368
 369 static int process_pty(int master, sigset_t *mask) {
 370         char in_buffer[BUFFER_SIZE], out_buffer[BUFFER_SIZE];
 371         size_t in_buffer_full = 0, out_buffer_full = 0;
 372         struct epoll_event stdin_ev, stdout_ev, master_ev, signal_ev;
 373         bool stdin_readable = false, stdout_writable = false, master_readable = false, master_writable = false;
 374         bool stdin_rhup = false, stdout_whup = false, master_rhup = false, master_whup = false;
 375         int ep = -1, signal_fd = -1, r;
 376
 377         fd_nonblock(STDIN_FILENO, 1);
 378         fd_nonblock(STDOUT_FILENO, 1);
 379         fd_nonblock(master, 1);
 380
 381         if ((signal_fd = signalfd(-1, mask, SFD_NONBLOCK|SFD_CLOEXEC)) < 0) {
 382                 log_error("signalfd(): %m");
 383                 r = -errno;
 384                 goto finish;
 385         }
 386
 387         if ((ep = epoll_create1(EPOLL_CLOEXEC)) < 0) {
 388                 log_error("Failed to create epoll: %m");
 389                 r = -errno;
 390                 goto finish;
 391         }
 392
 393         zero(stdin_ev);
 394         stdin_ev.events = EPOLLIN|EPOLLET;
 395         stdin_ev.data.fd = STDIN_FILENO;
 396
 397         zero(stdout_ev);
 398         stdout_ev.events = EPOLLOUT|EPOLLET;
 399         stdout_ev.data.fd = STDOUT_FILENO;
 400
 401         zero(master_ev);
 402         master_ev.events = EPOLLIN|EPOLLOUT|EPOLLET;
 403         master_ev.data.fd = master;
 404
 405         zero(signal_ev);
 406         signal_ev.events = EPOLLIN;
 407         signal_ev.data.fd = signal_fd;
 408
 409         if (epoll_ctl(ep, EPOLL_CTL_ADD, STDIN_FILENO, &stdin_ev) < 0 ||
 410             epoll_ctl(ep, EPOLL_CTL_ADD, STDOUT_FILENO, &stdout_ev) < 0 ||
 411             epoll_ctl(ep, EPOLL_CTL_ADD, master, &master_ev) < 0 ||
 412             epoll_ctl(ep, EPOLL_CTL_ADD, signal_fd, &signal_ev) < 0) {
 413                 log_error("Failed to regiser fds in epoll: %m");
 414                 r = -errno;
 415                 goto finish;
 416         }
 417
 418         do {
 419                 struct epoll_event ev[16];
 420                 ssize_t k;
 421                 int i, nfds;
 422
 423                 if ((nfds = epoll_wait(ep, ev, ELEMENTSOF(ev), -1)) < 0) {
 424
 425                         if (errno == EINTR || errno == EAGAIN)
 426                                 continue;
 427
 428                         log_error("epoll_wait(): %m");
 429                         r = -errno;
 430                         goto finish;
 431                 }
 432
 433                 assert(nfds >= 1);
 434
 435                 for (i = 0; i < nfds; i++) {
 436                         if (ev[i].data.fd == STDIN_FILENO) {
 437
 438                                 if (!stdin_rhup && (ev[i].events & (EPOLLHUP|EPOLLIN)))
 439                                         stdin_readable = true;
 440
 441                         } else if (ev[i].data.fd == STDOUT_FILENO) {
 442
 443                                 if (ev[i].events & EPOLLHUP) {
 444                                         stdout_writable = false;
 445                                         stdout_whup = true;
 446                                 }
 447
 448                                 if (!stdout_whup && (ev[i].events & EPOLLOUT))
 449                                         stdout_writable = true;
 450
 451                         } else if (ev[i].data.fd == master) {
 452
 453                                 /* We don't connect EPOLLHUP to
 454                                  * master_whup here, since EPOLLHUP
 455                                  * can happen when noone has opened
 456                                  * the other side */
 457
 458                                 if (!master_rhup && (ev[i].events & (EPOLLHUP|EPOLLIN)))
 459                                         master_readable = true;
 460
 461                                 if (!master_whup && (ev[i].events & EPOLLOUT))
 462                                         master_writable = true;
 463
 464                         } else if (ev[i].data.fd == signal_fd) {
 465                                 struct signalfd_siginfo sfsi;
 466                                 ssize_t n;
 467
 468                                 if ((n = read(signal_fd, &sfsi, sizeof(sfsi))) != sizeof(sfsi)) {
 469
 470                                         if (n >= 0) {
 471                                                 r = -EIO;
 472                                                 goto finish;
 473                                         }
 474
 475                                         if (errno != EINTR && errno != EAGAIN) {
 476                                                 r = -errno;
 477                                                 goto finish;
 478                                         }
 479                                 } else {
 480
 481                                         if (sfsi.ssi_signo == SIGWINCH) {
 482                                                 struct winsize ws;
 483
 484                                                 /* The window size changed, let's forward that. */
 485
 486                                                 if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 487                                                         ioctl(master, TIOCSWINSZ, &ws);
 488                                         } else {
 489                                                 r = -EINTR;
 490                                                 goto finish;
 491                                         }
 492                                 }
 493                         }
 494                 }
 495
 496                 while ((stdin_readable && in_buffer_full <= 0) ||
 497                        (master_writable && in_buffer_full > 0) ||
 498                        (master_readable && out_buffer_full <= 0) ||
 499                        (stdout_writable && out_buffer_full > 0)) {
 500
 501                         if (stdin_readable && in_buffer_full < BUFFER_SIZE) {
 502
 503                                 if ((k = read(STDIN_FILENO, in_buffer + in_buffer_full, BUFFER_SIZE - in_buffer_full)) < 0) {
 504
 505                                         if (errno == EAGAIN)
 506                                                 stdin_readable = false;
 507                                         else if (errno == EPIPE || errno == ECONNRESET || errno == EIO)
 508                                                 k = 0;
 509                                         else {
 510                                                 log_error("read(): %m");
 511                                                 goto finish;
 512                                         }
 513                                 } else
 514                                         in_buffer_full += (size_t) k;
 515
 516                                 if (k == 0) {
 517                                         stdin_rhup = true;
 518                                         stdin_readable = false;
 519                                         shutdown(STDIN_FILENO, SHUT_RD);
 520                                 }
 521                         }
 522
 523                         if (master_writable && in_buffer_full > 0) {
 524
 525                                 if ((k = write(master, in_buffer, in_buffer_full)) < 0) {
 526
 527                                         if (errno == EAGAIN)
 528                                                 master_writable = false;
 529                                         else if (errno == EPIPE || errno == ECONNRESET || errno == EIO) {
 530                                                 master_whup = true;
 531                                                 master_writable = false;
 532                                         } else {
 533                                                 log_error("write(): %m");
 534                                                 goto finish;
 535                                         }
 536
 537                                 } else {
 538                                         assert(in_buffer_full >= (size_t) k);
 539                                         memmove(in_buffer, in_buffer + k, in_buffer_full - k);
 540                                         in_buffer_full -= k;
 541                                 }
 542                         }
 543
 544                         if (master_readable && out_buffer_full < BUFFER_SIZE) {
 545
 546                                 if ((k = read(master, out_buffer + out_buffer_full, BUFFER_SIZE - out_buffer_full)) < 0) {
 547
 548                                         if (errno == EAGAIN)
 549                                                 master_readable = false;
 550                                         else if (errno == EPIPE || errno == ECONNRESET || errno == EIO)
 551                                                 k = 0;
 552                                         else {
 553                                                 log_error("read(): %m");
 554                                                 goto finish;
 555                                         }
 556                                 }  else
 557                                         out_buffer_full += (size_t) k;
 558
 559                                 if (k == 0) {
 560                                         master_rhup = true;
 561                                         master_readable = false;
 562                                 }
 563                         }
 564
 565                         if (stdout_writable && out_buffer_full > 0) {
 566
 567                                 if ((k = write(STDOUT_FILENO, out_buffer, out_buffer_full)) < 0) {
 568
 569                                         if (errno == EAGAIN)
 570                                                 stdout_writable = false;
 571                                         else if (errno == EPIPE || errno == ECONNRESET || errno == EIO) {
 572                                                 stdout_whup = true;
 573                                                 stdout_writable = false;
 574                                         } else {
 575                                                 log_error("write(): %m");
 576                                                 goto finish;
 577                                         }
 578
 579                                 } else {
 580                                         assert(out_buffer_full >= (size_t) k);
 581                                         memmove(out_buffer, out_buffer + k, out_buffer_full - k);
 582                                         out_buffer_full -= k;
 583                                 }
 584                         }
 585                 }
 586
 587                 if (stdin_rhup && in_buffer_full <= 0 && !master_whup) {
 588                         master_whup = true;
 589                         master_writable = false;
 590                 }
 591
 592                 if (master_rhup && out_buffer_full <= 0 && !stdout_whup) {
 593                         stdout_whup = true;
 594                         stdout_writable = false;
 595                         shutdown(STDOUT_FILENO, SHUT_WR);
 596                 }
 597
 598         } while (!stdout_whup || !master_whup);
 599
 600 finish:
 601         if (ep >= 0)
 602                 close_nointr_nofail(ep);
 603
 604         if (signal_fd >= 0)
 605                 close_nointr_nofail(signal_fd);
 606
 607         return r;
 608 }
 609
 610 int main(int argc, char *argv[]) {
 611         pid_t pid = 0;
 612         int r = EXIT_FAILURE, k;
 613         char *oldcg = NULL, *newcg = NULL;
 614         int master = -1;
 615         const char *console = NULL;
 616         struct termios saved_attr, raw_attr;
 617         sigset_t mask;
 618         bool saved_attr_valid = false;
 619         struct winsize ws;
 620
 621         log_parse_environment();
 622         log_open();
 623
 624         if ((r = parse_argv(argc, argv)) <= 0)
 625                 goto finish;
 626
 627         if (arg_directory) {
 628                 char *p;
 629
 630                 p = path_make_absolute_cwd(arg_directory);
 631                 free(arg_directory);
 632                 arg_directory = p;
 633         } else
 634                 arg_directory = get_current_dir_name();
 635
 636         if (!arg_directory) {
 637                 log_error("Failed to determine path");
 638                 goto finish;
 639         }
 640
 641         path_kill_slashes(arg_directory);
 642
 643         if (geteuid() != 0) {
 644                 log_error("Need to be root.");
 645                 goto finish;
 646         }
 647
 648         if (sd_booted() <= 0) {
 649                 log_error("Not running on a systemd system.");
 650                 goto finish;
 651         }
 652
 653         if (path_equal(arg_directory, "/")) {
 654                 log_error("Spawning container on root directory not supported.");
 655                 goto finish;
 656         }
 657
 658         if (is_os_tree(arg_directory) <= 0) {
 659                 log_error("Directory %s doesn't look like an OS root directory. Refusing.", arg_directory);
 660                 goto finish;
 661         }
 662
 663         if ((k = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 0, &oldcg)) < 0) {
 664                 log_error("Failed to determine current cgroup: %s", strerror(-k));
 665                 goto finish;
 666         }
 667
 668         if (asprintf(&newcg, "%s/nspawn-%lu", oldcg, (unsigned long) getpid()) < 0) {
 669                 log_error("Failed to allocate cgroup path.");
 670                 goto finish;
 671         }
 672
 673         if ((k = cg_create_and_attach(SYSTEMD_CGROUP_CONTROLLER, newcg, 0)) < 0)  {
 674                 log_error("Failed to create cgroup: %s", strerror(-k));
 675                 goto finish;
 676         }
 677
 678         if ((master = posix_openpt(O_RDWR|O_NOCTTY|O_CLOEXEC|O_NDELAY)) < 0) {
 679                 log_error("Failed to acquire pseudo tty: %m");
 680                 goto finish;
 681         }
 682
 683         if (!(console = ptsname(master))) {
 684                 log_error("Failed to determine tty name: %m");
 685                 goto finish;
 686         }
 687
 688         log_info("Spawning namespace container on %s (console is %s).", arg_directory, console);
 689
 690         if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
 691                 ioctl(master, TIOCSWINSZ, &ws);
 692
 693         if (unlockpt(master) < 0) {
 694                 log_error("Failed to unlock tty: %m");
 695                 goto finish;
 696         }
 697
 698         if (tcgetattr(STDIN_FILENO, &saved_attr) < 0) {
 699                 log_error("Failed to get terminal attributes: %m");
 700                 goto finish;
 701         }
 702
 703         saved_attr_valid = true;
 704
 705         raw_attr = saved_attr;
 706         cfmakeraw(&raw_attr);
 707         raw_attr.c_lflag &= ~ECHO;
 708
 709         if (tcsetattr(STDIN_FILENO, TCSANOW, &raw_attr) < 0) {
 710                 log_error("Failed to set terminal attributes: %m");
 711                 goto finish;
 712         }
 713
 714         assert_se(sigemptyset(&mask) == 0);
 715         sigset_add_many(&mask, SIGCHLD, SIGWINCH, SIGTERM, SIGINT, -1);
 716         assert_se(sigprocmask(SIG_BLOCK, &mask, NULL) == 0);
 717
 718         if ((pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS, NULL)) < 0) {
 719                 log_error("clone() failed: %m");
 720                 goto finish;
 721         }
 722
 723         if (pid == 0) {
 724                 /* child */
 725
 726                 const char *hn;
 727                 const char *envp[] = {
 728                         "HOME=/root",
 729                         "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
 730                         NULL,
 731                         NULL
 732                 };
 733
 734                 envp[2] = strv_find_prefix(environ, "TERM=");
 735
 736                 close_nointr_nofail(master);
 737
 738                 close_nointr(STDIN_FILENO);
 739                 close_nointr(STDOUT_FILENO);
 740                 close_nointr(STDERR_FILENO);
 741
 742                 close_all_fds(NULL, 0);
 743
 744                 reset_all_signal_handlers();
 745
 746                 assert_se(sigemptyset(&mask) == 0);
 747                 assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
 748
 749                 if (setsid() < 0)
 750                         goto child_fail;
 751
 752                 if (prctl(PR_SET_PDEATHSIG, SIGKILL) < 0)
 753                         goto child_fail;
 754
 755                 if (mount_all(arg_directory) < 0)
 756                         goto child_fail;
 757
 758                 if (copy_devnodes(arg_directory, console) < 0)
 759                         goto child_fail;
 760
 761                 if (chdir(arg_directory) < 0) {
 762                         log_error("chdir(%s) failed: %m", arg_directory);
 763                         goto child_fail;
 764                 }
 765
 766                 if (open_terminal("dev/console", O_RDWR) != STDIN_FILENO ||
 767                     dup2(STDIN_FILENO, STDOUT_FILENO) != STDOUT_FILENO ||
 768                     dup2(STDIN_FILENO, STDERR_FILENO) != STDERR_FILENO)
 769                         goto child_fail;
 770
 771                 if (mount(arg_directory, "/", "bind", MS_BIND|MS_MOVE, NULL) < 0) {
 772                         log_error("mount(MS_MOVE) failed: %m");
 773                         goto child_fail;
 774                 }
 775
 776                 if (chroot(".") < 0) {
 777                         log_error("chroot() failed: %m");
 778                         goto child_fail;
 779                 }
 780
 781                 if (chdir("/") < 0) {
 782                         log_error("chdir() failed: %m");
 783                         goto child_fail;
 784                 }
 785
 786                 umask(0002);
 787
 788                 if (drop_capabilities() < 0)
 789                         goto child_fail;
 790
 791                 if ((hn = file_name_from_path(arg_directory)))
 792                         sethostname(hn, strlen(hn));
 793
 794                 if (argc > optind)
 795                         execvpe(argv[optind], argv + optind, (char**) envp);
 796                 else {
 797                         chdir("/root");
 798                         execle("/bin/bash", "-bash", NULL, (char**) envp);
 799                 }
 800
 801                 log_error("execv() failed: %m");
 802
 803         child_fail:
 804                 _exit(EXIT_FAILURE);
 805         }
 806
 807         if (process_pty(master, &mask) < 0)
 808                 goto finish;
 809
 810         if (saved_attr_valid) {
 811                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 812                 saved_attr_valid = false;
 813         }
 814
 815         r = wait_for_terminate_and_warn(argc > optind ? argv[optind] : "bash", pid);
 816
 817         if (r < 0)
 818                 r = EXIT_FAILURE;
 819
 820 finish:
 821         if (saved_attr_valid)
 822                 tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
 823
 824         if (master >= 0)
 825                 close_nointr_nofail(master);
 826
 827         if (oldcg)
 828                 cg_attach(SYSTEMD_CGROUP_CONTROLLER, oldcg, 0);
 829
 830         if (newcg)
 831                 cg_kill_recursive_and_wait(SYSTEMD_CGROUP_CONTROLLER, newcg, true);
 832
 833         free(arg_directory);
 834         free(oldcg);
 835         free(newcg);
 836
 837         return r;
 838 }