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                 bool fatal;
 105         } MountPoint;
 106
 107         static const MountPoint mount_table[] = {
 108                 { "proc",      "/proc",     "proc",      NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV, true },
 109                 { "/proc/sys", "/proc/sys", "bind",      NULL,        MS_BIND, true },                      /* Bind mount first */
 110                 { "/proc/sys", "/proc/sys", "bind",      NULL,        MS_BIND|MS_RDONLY|MS_REMOUNT, true }, /* Then, make it r/o */
 111                 { "sysfs",     "/sys",      "sysfs",     NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, true },
 112                 { "tmpfs",     "/dev",      "tmpfs",     "mode=755",  MS_NOSUID, true },
 113                 { "/dev/pts",  "/dev/pts",  "bind",      NULL,        MS_BIND, true },
 114                 { "tmpfs",     "/dev/.run", "tmpfs",     "mode=755",  MS_NOSUID|MS_NOEXEC|MS_NODEV, true },
 115 #ifdef HAVE_SELINUX
 116                 { "selinux",   "/selinux",  "selinuxfs", NULL,        MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, false },
 117 #endif
 118         };
 119
 120         unsigned k;
 121         int r = 0;
 122
 123         for (k = 0; k < ELEMENTSOF(mount_table); k++) {
 124                 char *where;
 125                 int t;
 126
 127                 if (asprintf(&where, "%s/%s", dest, mount_table[k].where) < 0) {
 128                         log_error("Out of memory");
 129
 130                         if (r == 0)
 131                                 r = -ENOMEM;
 132
 133                         break;
 134                 }
 135
 136                 if ((t = path_is_mount_point(where)) < 0) {
 137                         log_error("Failed to detect whether %s is a mount point: %s", where, strerror(-t));
 138                         free(where);
 139
 140                         if (r == 0)
 141                                 r = t;
 142
 143                         continue;
 144                 }
 145
 146                 mkdir_p(where, 0755);
 147
 148                 if (mount(mount_table[k].what,
 149                           where,
 150                           mount_table[k].type,
 151                           mount_table[k].flags,
 152                           mount_table[k].options) < 0 &&
 153                     mount_table[k].fatal) {
 154
 155                         log_error("mount(%s) failed: %m", where);
 156
 157                         if (r == 0)
 158                                 r = -errno;
 159                 }
 160
 161                 free(where);
 162         }
 163
 164         return r;
 165 }
 166
 167 static int copy_devnodes(const char *dest) {
 168
 169         static const char devnodes[] =
 170                 "null\0"
 171                 "zero\0"
 172                 "full\0"
 173                 "random\0"
 174                 "urandom\0"
 175                 "tty\0"
 176                 "ptmx\0"
 177                 "kmsg\0"
 178                 "rtc0\0";
 179
 180         const char *d;
 181         int r = 0, k;
 182         char *tty = NULL;
 183         dev_t tty_devnum;
 184         mode_t u;
 185
 186         u = umask(0000);
 187
 188         NULSTR_FOREACH(d, devnodes) {
 189                 char *from = NULL, *to = NULL;
 190                 struct stat st;
 191
 192                 asprintf(&from, "/dev/%s", d);
 193                 asprintf(&to, "%s/dev/%s", dest, d);
 194
 195                 if (!from || !to) {
 196                         log_error("Failed to allocate devnode path");
 197
 198                         free(from);
 199                         free(to);
 200
 201                         if (r == 0)
 202                                 r = -ENOMEM;
 203
 204                         break;
 205                 }
 206
 207                 if (stat(from, &st) < 0) {
 208
 209                         if (errno != ENOENT) {
 210                                 log_error("Failed to stat %s: %m", from);
 211
 212                                 if (r == 0)
 213                                         r = -errno;
 214                         }
 215
 216                 } else {
 217                         if (mknod(to, st.st_mode, st.st_rdev) < 0) {
 218                                 log_error("mknod(%s) failed: %m", dest);
 219
 220                                 if (r == 0)
 221                                         r = -errno;
 222                         }
 223                 }
 224
 225                 free(from);
 226                 free(to);
 227         }
 228
 229         if ((k = get_ctty(&tty, &tty_devnum)) < 0) {
 230                 log_error("Failed to determine controlling tty: %s", strerror(-k));
 231
 232                 if (r == 0)
 233                         r = k;
 234         } else {
 235                 char *from = NULL, *to = NULL;
 236
 237                 asprintf(&from, "/dev/%s", tty);
 238                 asprintf(&to, "%s/dev/console", dest);
 239
 240                 if (!from || !to) {
 241                         log_error("Out of memory");
 242
 243                         if (r == 0)
 244                                 r = k;
 245                 } else {
 246                         /* We need to bind mount our own tty on
 247                          * /dev/console, since ptys cannot be used
 248                          * unless on a devpts file system. But to bind
 249                          * mount it we first have to create a device
 250                          * node where we can bind mount it on. This is
 251                          * kinda ugly since the TTY will very likely
 252                          * be owned by a user/group that does not
 253                          * exist in the container. */
 254
 255                         if (mknod(to, S_IFCHR|0600, tty_devnum) < 0) {
 256                                 log_error("mknod for /dev/console failed: %m");
 257
 258                                 if (r == 0)
 259                                         r = -errno;
 260                         }
 261
 262                         if (mount(from, to, "bind", MS_BIND|MS_RDONLY, NULL) < 0) {
 263                                 log_error("bind mount for /dev/console failed: %m");
 264
 265                                 if (r == 0)
 266                                         r = -errno;
 267                         }
 268                 }
 269
 270                 free(from);
 271                 free(to);
 272         }
 273
 274         free(tty);
 275
 276         umask(u);
 277
 278         return r;
 279 }
 280
 281 static int drop_capabilities(void) {
 282         static const unsigned long retain[] = {
 283                 CAP_CHOWN,
 284                 CAP_DAC_OVERRIDE,
 285                 CAP_DAC_READ_SEARCH,
 286                 CAP_FOWNER,
 287                 CAP_FSETID,
 288                 CAP_IPC_OWNER,
 289                 CAP_KILL,
 290                 CAP_LEASE,
 291                 CAP_LINUX_IMMUTABLE,
 292                 CAP_NET_BIND_SERVICE,
 293                 CAP_NET_BROADCAST,
 294                 CAP_NET_RAW,
 295                 CAP_SETGID,
 296                 CAP_SETFCAP,
 297                 CAP_SETPCAP,
 298                 CAP_SETUID,
 299                 CAP_SYS_ADMIN,
 300                 CAP_SYS_CHROOT,
 301                 CAP_SYS_NICE,
 302                 CAP_SYS_PTRACE,
 303                 CAP_SYS_TTY_CONFIG
 304         };
 305
 306         unsigned long l;
 307
 308         for (l = 0; l <= MAX(63LU, (unsigned long) CAP_LAST_CAP); l ++) {
 309                 unsigned i;
 310
 311                 for (i = 0; i < ELEMENTSOF(retain); i++)
 312                         if (retain[i] == l)
 313                                 break;
 314
 315                 if (i < ELEMENTSOF(retain))
 316                         continue;
 317
 318                 if (prctl(PR_CAPBSET_DROP, l) < 0) {
 319
 320                         /* If this capability is not known, EINVAL
 321                          * will be returned, let's ignore this. */
 322                         if (errno == EINVAL)
 323                                 continue;
 324
 325                         log_error("PR_CAPBSET_DROP failed: %m");
 326                         return -errno;
 327                 }
 328         }
 329
 330         return 0;
 331 }
 332
 333 static int is_os_tree(const char *path) {
 334         int r;
 335         char *p;
 336         /* We use /bin/sh as flag file if something is an OS */
 337
 338         if (asprintf(&p, "%s/bin/sh", path) < 0)
 339                 return -ENOMEM;
 340
 341         r = access(p, F_OK);
 342         free(p);
 343
 344         return r < 0 ? 0 : 1;
 345 }
 346
 347
 348 int main(int argc, char *argv[]) {
 349         pid_t pid = 0;
 350         int r = EXIT_FAILURE;
 351
 352         log_parse_environment();
 353         log_open();
 354
 355         if ((r = parse_argv(argc, argv)) <= 0)
 356                 goto finish;
 357
 358         if (arg_directory) {
 359                 char *p;
 360
 361                 p = path_make_absolute_cwd(arg_directory);
 362                 free(arg_directory);
 363                 arg_directory = p;
 364         } else
 365                 arg_directory = get_current_dir_name();
 366
 367         if (!arg_directory) {
 368                 log_error("Failed to determine path");
 369                 goto finish;
 370         }
 371
 372         path_kill_slashes(arg_directory);
 373
 374         if (geteuid() != 0) {
 375                 log_error("Need to be root.");
 376                 goto finish;
 377         }
 378
 379         if (path_equal(arg_directory, "/")) {
 380                 log_error("Spawning container on root directory not supported.");
 381                 goto finish;
 382         }
 383
 384         if (is_os_tree(arg_directory) <= 0) {
 385                 log_error("Directory %s doesn't look like an OS root directory. Refusing.", arg_directory);
 386                 goto finish;
 387         }
 388
 389         log_info("Spawning namespace container on %s.", arg_directory);
 390
 391         if ((pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS, NULL)) < 0) {
 392                 log_error("clone() failed: %m");
 393                 goto finish;
 394         }
 395
 396         if (pid == 0) {
 397                 const char *hn;
 398                 const char *envp[] = {
 399                         "HOME=/root",
 400                         "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
 401                         NULL
 402                 };
 403
 404                 /* child */
 405
 406                 if (mount_all(arg_directory) < 0)
 407                         goto child_fail;
 408
 409                 if (copy_devnodes(arg_directory) < 0)
 410                         goto child_fail;
 411
 412                 if (chdir(arg_directory) < 0) {
 413                         log_error("chdir(%s) failed: %m", arg_directory);
 414                         goto child_fail;
 415                 }
 416                 if (mount(arg_directory, "/", "bind", MS_BIND|MS_MOVE, NULL) < 0) {
 417                         log_error("mount(MS_MOVE) failed: %m");
 418                         goto child_fail;
 419                 }
 420
 421                 if (chroot(".") < 0) {
 422                         log_error("chroot() failed: %m");
 423                         goto child_fail;
 424                 }
 425
 426                 if (chdir("/") < 0) {
 427                         log_error("chdir() failed: %m");
 428                         goto child_fail;
 429                 }
 430
 431                 if (drop_capabilities() < 0)
 432                         goto child_fail;
 433
 434                 if ((hn = file_name_from_path(arg_directory)))
 435                         sethostname(hn, strlen(hn));
 436
 437                 if (argc > optind)
 438                         execvpe(argv[optind], argv + optind, (char**) envp);
 439                 else {
 440                         chdir("/root");
 441                         execle("/bin/bash", "-bash", NULL, (char**) envp);
 442                 }
 443
 444                 log_error("execv() failed: %m");
 445
 446         child_fail:
 447                 _exit(EXIT_FAILURE);
 448         }
 449
 450         r = wait_for_terminate_and_warn(argc > optind ? argv[optind] : "bash", pid);
 451
 452         if (r < 0)
 453                 r = EXIT_FAILURE;
 454
 455 finish:
 456         free(arg_directory);
 457
 458         if (pid > 0)
 459                 kill(pid, SIGTERM);
 460
 461         return r;
 462 }