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
  37 #include "log.h"
  38 #include "util.h"
  39 #include "missing.h"
  40
  41 static char *arg_directory = NULL;
  42
  43 static int help(void) {
  44
  45         printf("%s [OPTIONS...] [PATH] [ARGUMENTS...]\n\n"
  46                "Spawn a minimal namespace container for debugging, testing and building.\n\n"
  47                "  -h --help            Show this help\n"
  48                "  -D --directory=NAME  Root directory for the container\n",
  49                program_invocation_short_name);
  50
  51         return 0;
  52 }
  53
  54 static int parse_argv(int argc, char *argv[]) {
  55
  56         static const struct option options[] = {
  57                 { "help",      no_argument,       NULL, 'h' },
  58                 { "directory", required_argument, NULL, 'D' },
  59                 { NULL,        0,                 NULL, 0   }
  60         };
  61
  62         int c;
  63
  64         assert(argc >= 0);
  65         assert(argv);
  66
  67         while ((c = getopt_long(argc, argv, "+hD:", options, NULL)) >= 0) {
  68
  69                 switch (c) {
  70
  71                 case 'h':
  72                         help();
  73                         return 0;
  74
  75                 case 'D':
  76                         free(arg_directory);
  77                         if (!(arg_directory = strdup(optarg))) {
  78                                 log_error("Failed to duplicate root directory.");
  79                                 return -ENOMEM;
  80                         }
  81
  82                         break;
  83
  84                 case '?':
  85                         return -EINVAL;
  86
  87                 default:
  88                         log_error("Unknown option code %c", c);
  89                         return -EINVAL;
  90                 }
  91         }
  92
  93         return 1;
  94 }
  95
  96 static int mount_all(const char *dest) {
  97
  98         typedef struct MountPoint {
  99                 const char *what;
 100                 const char *where;
 101                 const char *type;
 102                 const char *options;
 103                 unsigned long flags;
 104         } MountPoint;
 105
 106         static const MountPoint mount_table[] = {
 107                 { "proc",      "/proc",     "proc",      NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV },
 108                 { "/proc/sys", "/proc/sys", "bind",      NULL,        MS_BIND },                      /* Bind mount first */
 109                 { "/proc/sys", "/proc/sys", "bind",      NULL,        MS_BIND|MS_RDONLY|MS_REMOUNT }, /* Then, make it r/o */
 110                 { "sysfs",     "/sys",      "sysfs",     NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY },
 111                 { "tmpfs",     "/dev",      "tmpfs",     "mode=755",  MS_NOSUID },
 112                 { "/dev/pts",  "/dev/pts",  "bind",      NULL,        MS_BIND },
 113                 { "tmpfs",     "/dev/.run", "tmpfs",     "mode=755",  MS_NOSUID|MS_NOEXEC|MS_NODEV },
 114 #ifdef HAVE_SELINUX
 115                 { "selinux",   "/selinux",  "selinuxfs", NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY },
 116 #endif
 117         };
 118
 119         unsigned k;
 120         int r = 0;
 121
 122         for (k = 0; k < ELEMENTSOF(mount_table); k++) {
 123                 char *where;
 124                 int t;
 125
 126                 if (asprintf(&where, "%s/%s", dest, mount_table[k].where) < 0) {
 127                         log_error("Out of memory");
 128
 129                         if (r == 0)
 130                                 r = -ENOMEM;
 131
 132                         break;
 133                 }
 134
 135                 if ((t = path_is_mount_point(where)) < 0) {
 136                         log_error("Failed to detect whether %s is a mount point: %s", where, strerror(-t));
 137                         free(where);
 138
 139                         if (r == 0)
 140                                 r = t;
 141
 142                         continue;
 143                 }
 144
 145                 mkdir_p(where, 0755);
 146
 147                 if (mount(mount_table[k].what,
 148                           where,
 149                           mount_table[k].type,
 150                           mount_table[k].flags,
 151                           mount_table[k].options) < 0) {
 152
 153                         log_error("mount(%s) failed: %m", where);
 154
 155                         if (r == 0)
 156                                 r = -errno;
 157                 }
 158
 159                 free(where);
 160         }
 161
 162         return r;
 163 }
 164
 165 static int copy_devnodes(const char *dest) {
 166
 167         static const char devnodes[] =
 168                 "null\0"
 169                 "zero\0"
 170                 "full\0"
 171                 "random\0"
 172                 "urandom\0"
 173                 "tty\0"
 174                 "ptmx\0"
 175                 "kmsg\0"
 176                 "rtc0\0";
 177
 178         const char *d;
 179         int r = 0, k;
 180         char *tty = NULL;
 181         dev_t tty_devnum;
 182         mode_t u;
 183
 184         u = umask(0000);
 185
 186         NULSTR_FOREACH(d, devnodes) {
 187                 char *from = NULL, *to = NULL;
 188                 struct stat st;
 189
 190                 asprintf(&from, "/dev/%s", d);
 191                 asprintf(&to, "%s/dev/%s", dest, d);
 192
 193                 if (!from || !to) {
 194                         log_error("Failed to allocate devnode path");
 195
 196                         free(from);
 197                         free(to);
 198
 199                         if (r == 0)
 200                                 r = -ENOMEM;
 201
 202                         break;
 203                 }
 204
 205                 if (stat(from, &st) < 0) {
 206
 207                         if (errno != ENOENT) {
 208                                 log_error("Failed to stat %s: %m", from);
 209
 210                                 if (r == 0)
 211                                         r = -errno;
 212                         }
 213
 214                 } else {
 215                         if (mknod(to, st.st_mode, st.st_rdev) < 0) {
 216                                 log_error("mknod(%s) failed: %m", dest);
 217
 218                                 if (r == 0)
 219                                         r = -errno;
 220                         }
 221                 }
 222
 223                 free(from);
 224                 free(to);
 225         }
 226
 227         if ((k = get_ctty(&tty, &tty_devnum)) < 0) {
 228                 log_error("Failed to determine controlling tty: %s", strerror(-k));
 229
 230                 if (r == 0)
 231                         r = k;
 232         } else {
 233                 char *from = NULL, *to = NULL;
 234
 235                 asprintf(&from, "/dev/%s", tty);
 236                 asprintf(&to, "%s/dev/console", dest);
 237
 238                 if (!from || !to) {
 239                         log_error("Out of memory");
 240
 241                         if (r == 0)
 242                                 r = k;
 243                 } else {
 244                         /* We need to bind mount our own tty on
 245                          * /dev/console, since ptys cannot be used
 246                          * unless on a devpts file system. But to bind
 247                          * mount it we first have to create a device
 248                          * node where we can bind mount it on. This is
 249                          * kinda ugly since the TTY will very likely
 250                          * be owned by a user/group that does not
 251                          * exist in the container. */
 252
 253                         if (mknod(to, S_IFCHR|0600, tty_devnum) < 0) {
 254                                 log_error("mknod for /dev/console failed: %m");
 255
 256                                 if (r == 0)
 257                                         r = -errno;
 258                         }
 259
 260                         if (mount(from, to, "bind", MS_BIND|MS_RDONLY, NULL) < 0) {
 261                                 log_error("bind mount for /dev/console failed: %m");
 262
 263                                 if (r == 0)
 264                                         r = -errno;
 265                         }
 266                 }
 267
 268                 free(from);
 269                 free(to);
 270         }
 271
 272         free(tty);
 273
 274         umask(u);
 275
 276         return r;
 277 }
 278
 279 static int drop_capabilities(void) {
 280         static const unsigned long retain[] = {
 281                 CAP_CHOWN,
 282                 CAP_DAC_OVERRIDE,
 283                 CAP_DAC_READ_SEARCH,
 284                 CAP_FOWNER,
 285                 CAP_FSETID,
 286                 CAP_IPC_OWNER,
 287                 CAP_KILL,
 288                 CAP_LEASE,
 289                 CAP_LINUX_IMMUTABLE,
 290                 CAP_NET_BIND_SERVICE,
 291                 CAP_NET_BROADCAST,
 292                 CAP_NET_RAW,
 293                 CAP_SETGID,
 294                 CAP_SETFCAP,
 295                 CAP_SETPCAP,
 296                 CAP_SETUID,
 297                 CAP_SYS_ADMIN,
 298                 CAP_SYS_CHROOT,
 299                 CAP_SYS_NICE,
 300                 CAP_SYS_PTRACE,
 301                 CAP_SYS_TTY_CONFIG
 302         };
 303
 304         unsigned long l;
 305
 306         for (l = 0; l <= MAX(63LU, (unsigned long) CAP_LAST_CAP); l ++) {
 307                 unsigned i;
 308
 309                 for (i = 0; i < ELEMENTSOF(retain); i++)
 310                         if (retain[i] == l)
 311                                 break;
 312
 313                 if (i < ELEMENTSOF(retain))
 314                         continue;
 315
 316                 if (prctl(PR_CAPBSET_DROP, l) < 0) {
 317
 318                         /* If this capability is not known, EINVAL
 319                          * will be returned, let's ignore this. */
 320                         if (errno == EINVAL)
 321                                 continue;
 322
 323                         log_error("PR_CAPBSET_DROP failed: %m");
 324                         return -errno;
 325                 }
 326         }
 327
 328         return 0;
 329 }
 330
 331 static int is_os_tree(const char *path) {
 332         int r;
 333         char *p;
 334         /* We use /bin/sh as flag file if something is an OS */
 335
 336         if (asprintf(&p, "%s/bin/sh", path) < 0)
 337                 return -ENOMEM;
 338
 339         r = access(p, F_OK);
 340         free(p);
 341
 342         return r < 0 ? 0 : 1;
 343 }
 344
 345
 346 int main(int argc, char *argv[]) {
 347         pid_t pid = 0;
 348         int r = EXIT_FAILURE;
 349
 350         log_parse_environment();
 351         log_open();
 352
 353         if ((r = parse_argv(argc, argv)) <= 0)
 354                 goto finish;
 355
 356         if (arg_directory) {
 357                 char *p;
 358
 359                 p = path_make_absolute_cwd(arg_directory);
 360                 free(arg_directory);
 361                 arg_directory = p;
 362         } else
 363                 arg_directory = get_current_dir_name();
 364
 365         if (!arg_directory) {
 366                 log_error("Failed to determine path");
 367                 goto finish;
 368         }
 369
 370         path_kill_slashes(arg_directory);
 371
 372         if (geteuid() != 0) {
 373                 log_error("Need to be root.");
 374                 goto finish;
 375         }
 376
 377         if (path_equal(arg_directory, "/")) {
 378                 log_error("Spawning container on root directory not supported.");
 379                 goto finish;
 380         }
 381
 382         if (is_os_tree(arg_directory) <= 0) {
 383                 log_error("Directory %s doesn't look like an OS root directory. Refusing.", arg_directory);
 384                 goto finish;
 385         }
 386
 387         log_info("Spawning namespace container on %s.", arg_directory);
 388
 389         if ((pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS, NULL)) < 0) {
 390                 log_error("clone() failed: %m");
 391                 goto finish;
 392         }
 393
 394         if (pid == 0) {
 395                 const char *hn;
 396                 const char *envp[] = {
 397                         "HOME=/root",
 398                         "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
 399                         NULL
 400                 };
 401
 402                 /* child */
 403
 404                 if (mount_all(arg_directory) < 0)
 405                         goto child_fail;
 406
 407                 if (copy_devnodes(arg_directory) < 0)
 408                         goto child_fail;
 409
 410                 if (chdir(arg_directory) < 0) {
 411                         log_error("chdir(%s) failed: %m", arg_directory);
 412                         goto child_fail;
 413                 }
 414                 if (mount(arg_directory, "/", "bind", MS_BIND|MS_MOVE, NULL) < 0) {
 415                         log_error("mount(MS_MOVE) failed: %m");
 416                         goto child_fail;
 417                 }
 418
 419                 if (chroot(".") < 0) {
 420                         log_error("chroot() failed: %m");
 421                         goto child_fail;
 422                 }
 423
 424                 if (chdir("/") < 0) {
 425                         log_error("chdir() failed: %m");
 426                         goto child_fail;
 427                 }
 428
 429                 if (drop_capabilities() < 0)
 430                         goto child_fail;
 431
 432                 if ((hn = file_name_from_path(arg_directory)))
 433                         sethostname(hn, strlen(hn));
 434
 435                 if (argc > optind)
 436                         execvpe(argv[optind], argv + optind, (char**) envp);
 437                 else {
 438                         chdir("/root");
 439                         execle("/bin/bash", "-bash", NULL, (char**) envp);
 440                 }
 441
 442                 log_error("execv() failed: %m");
 443
 444         child_fail:
 445                 _exit(EXIT_FAILURE);
 446         }
 447
 448         r = wait_for_terminate_and_warn(argc > optind ? argv[optind] : "bash", pid);
 449
 450         if (r < 0)
 451                 r = EXIT_FAILURE;
 452
 453 finish:
 454         free(arg_directory);
 455
 456         if (pid > 0)
 457                 kill(pid, SIGTERM);
 458
 459         return r;
 460 }