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