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