src/core/mount-setup.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #include <errno.h>
   4 #include <ftw.h>
   5 #include <stdlib.h>
   6 #include <sys/mount.h>
   7 #include <sys/statvfs.h>
   8 #include <unistd.h>
   9
  10 #include "alloc-util.h"
  11 #include "bus-util.h"
  12 #include "cgroup-util.h"
  13 #include "dev-setup.h"
  14 #include "dirent-util.h"
  15 #include "efivars.h"
  16 #include "fd-util.h"
  17 #include "fileio.h"
  18 #include "fs-util.h"
  19 #include "label.h"
  20 #include "log.h"
  21 #include "macro.h"
  22 #include "missing.h"
  23 #include "mkdir.h"
  24 #include "mount-setup.h"
  25 #include "mountpoint-util.h"
  26 #include "path-util.h"
  27 #include "set.h"
  28 #include "smack-util.h"
  29 #include "strv.h"
  30 #include "user-util.h"
  31 #include "util.h"
  32 #include "virt.h"
  33
  34 typedef enum MountMode {
  35         MNT_NONE           = 0,
  36         MNT_FATAL          = 1 << 0,
  37         MNT_IN_CONTAINER   = 1 << 1,
  38         MNT_CHECK_WRITABLE = 1 << 2,
  39 } MountMode;
  40
  41 typedef struct MountPoint {
  42         const char *what;
  43         const char *where;
  44         const char *type;
  45         const char *options;
  46         unsigned long flags;
  47         bool (*condition_fn)(void);
  48         MountMode mode;
  49 } MountPoint;
  50
  51 /* The first three entries we might need before SELinux is up. The
  52  * fourth (securityfs) is needed by IMA to load a custom policy. The
  53  * other ones we can delay until SELinux and IMA are loaded. When
  54  * SMACK is enabled we need smackfs, too, so it's a fifth one. */
  55 #if ENABLE_SMACK
  56 #define N_EARLY_MOUNT 5
  57 #else
  58 #define N_EARLY_MOUNT 4
  59 #endif
  60
  61 static const MountPoint mount_table[] = {
  62         { "sysfs",       "/sys",                      "sysfs",      NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
  63           NULL,          MNT_FATAL|MNT_IN_CONTAINER },
  64         { "proc",        "/proc",                     "proc",       NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
  65           NULL,          MNT_FATAL|MNT_IN_CONTAINER },
  66         { "devtmpfs",    "/dev",                      "devtmpfs",   "mode=755",                MS_NOSUID|MS_STRICTATIME,
  67           NULL,          MNT_FATAL|MNT_IN_CONTAINER },
  68         { "securityfs",  "/sys/kernel/security",      "securityfs", NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
  69           NULL,          MNT_NONE                   },
  70 #if ENABLE_SMACK
  71         { "smackfs",     "/sys/fs/smackfs",           "smackfs",    "smackfsdef=*",            MS_NOSUID|MS_NOEXEC|MS_NODEV,
  72           mac_smack_use, MNT_FATAL                  },
  73         { "tmpfs",       "/dev/shm",                  "tmpfs",      "mode=1777,smackfsroot=*", MS_NOSUID|MS_NODEV|MS_STRICTATIME,
  74           mac_smack_use, MNT_FATAL                  },
  75 #endif
  76         { "tmpfs",       "/dev/shm",                  "tmpfs",      "mode=1777",               MS_NOSUID|MS_NODEV|MS_STRICTATIME,
  77           NULL,          MNT_FATAL|MNT_IN_CONTAINER },
  78         { "devpts",      "/dev/pts",                  "devpts",     "mode=620,gid=" STRINGIFY(TTY_GID), MS_NOSUID|MS_NOEXEC,
  79           NULL,          MNT_IN_CONTAINER           },
  80 #if ENABLE_SMACK
  81         { "tmpfs",       "/run",                      "tmpfs",      "mode=755,smackfsroot=*",  MS_NOSUID|MS_NODEV|MS_STRICTATIME,
  82           mac_smack_use, MNT_FATAL                  },
  83 #endif
  84         { "tmpfs",       "/run",                      "tmpfs",      "mode=755",                MS_NOSUID|MS_NODEV|MS_STRICTATIME,
  85           NULL,          MNT_FATAL|MNT_IN_CONTAINER },
  86         { "cgroup2",     "/sys/fs/cgroup",            "cgroup2",    "nsdelegate",              MS_NOSUID|MS_NOEXEC|MS_NODEV,
  87           cg_is_unified_wanted, MNT_IN_CONTAINER|MNT_CHECK_WRITABLE },
  88         { "cgroup2",     "/sys/fs/cgroup",            "cgroup2",    NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
  89           cg_is_unified_wanted, MNT_IN_CONTAINER|MNT_CHECK_WRITABLE },
  90         { "tmpfs",       "/sys/fs/cgroup",            "tmpfs",      "mode=755",                MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME,
  91           cg_is_legacy_wanted, MNT_FATAL|MNT_IN_CONTAINER },
  92         { "cgroup2",     "/sys/fs/cgroup/unified",    "cgroup2",    "nsdelegate",              MS_NOSUID|MS_NOEXEC|MS_NODEV,
  93           cg_is_hybrid_wanted, MNT_IN_CONTAINER|MNT_CHECK_WRITABLE },
  94         { "cgroup2",     "/sys/fs/cgroup/unified",    "cgroup2",    NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
  95           cg_is_hybrid_wanted, MNT_IN_CONTAINER|MNT_CHECK_WRITABLE },
  96         { "cgroup",      "/sys/fs/cgroup/systemd",    "cgroup",     "none,name=systemd,xattr", MS_NOSUID|MS_NOEXEC|MS_NODEV,
  97           cg_is_legacy_wanted, MNT_IN_CONTAINER     },
  98         { "cgroup",      "/sys/fs/cgroup/systemd",    "cgroup",     "none,name=systemd",       MS_NOSUID|MS_NOEXEC|MS_NODEV,
  99           cg_is_legacy_wanted, MNT_FATAL|MNT_IN_CONTAINER },
 100         { "pstore",      "/sys/fs/pstore",            "pstore",     NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
 101           NULL,          MNT_NONE                   },
 102 #if ENABLE_EFI
 103         { "efivarfs",    "/sys/firmware/efi/efivars", "efivarfs",   NULL,                      MS_NOSUID|MS_NOEXEC|MS_NODEV,
 104           is_efi_boot,   MNT_NONE                   },
 105 #endif
 106         { "bpf",         "/sys/fs/bpf",               "bpf",        "mode=700",                MS_NOSUID|MS_NOEXEC|MS_NODEV,
 107           NULL,          MNT_NONE,                  },
 108 };
 109
 110 /* These are API file systems that might be mounted by other software,
 111  * we just list them here so that we know that we should ignore them */
 112
 113 static const char ignore_paths[] =
 114         /* SELinux file systems */
 115         "/sys/fs/selinux\0"
 116         /* Container bind mounts */
 117         "/proc/sys\0"
 118         "/dev/console\0"
 119         "/proc/kmsg\0";
 120
 121 bool mount_point_is_api(const char *path) {
 122         unsigned i;
 123
 124         /* Checks if this mount point is considered "API", and hence
 125          * should be ignored */
 126
 127         for (i = 0; i < ELEMENTSOF(mount_table); i ++)
 128                 if (path_equal(path, mount_table[i].where))
 129                         return true;
 130
 131         return path_startswith(path, "/sys/fs/cgroup/");
 132 }
 133
 134 bool mount_point_ignore(const char *path) {
 135         const char *i;
 136
 137         NULSTR_FOREACH(i, ignore_paths)
 138                 if (path_equal(path, i))
 139                         return true;
 140
 141         return false;
 142 }
 143
 144 static int mount_one(const MountPoint *p, bool relabel) {
 145         int r, priority;
 146
 147         assert(p);
 148
 149         priority = (p->mode & MNT_FATAL) ? LOG_ERR : LOG_DEBUG;
 150
 151         if (p->condition_fn && !p->condition_fn())
 152                 return 0;
 153
 154         /* Relabel first, just in case */
 155         if (relabel)
 156                 (void) label_fix(p->where, LABEL_IGNORE_ENOENT|LABEL_IGNORE_EROFS);
 157
 158         r = path_is_mount_point(p->where, NULL, AT_SYMLINK_FOLLOW);
 159         if (r < 0 && r != -ENOENT) {
 160                 log_full_errno(priority, r, "Failed to determine whether %s is a mount point: %m", p->where);
 161                 return (p->mode & MNT_FATAL) ? r : 0;
 162         }
 163         if (r > 0)
 164                 return 0;
 165
 166         /* Skip securityfs in a container */
 167         if (!(p->mode & MNT_IN_CONTAINER) && detect_container() > 0)
 168                 return 0;
 169
 170         /* The access mode here doesn't really matter too much, since
 171          * the mounted file system will take precedence anyway. */
 172         if (relabel)
 173                 (void) mkdir_p_label(p->where, 0755);
 174         else
 175                 (void) mkdir_p(p->where, 0755);
 176
 177         log_debug("Mounting %s to %s of type %s with options %s.",
 178                   p->what,
 179                   p->where,
 180                   p->type,
 181                   strna(p->options));
 182
 183         if (mount(p->what,
 184                   p->where,
 185                   p->type,
 186                   p->flags,
 187                   p->options) < 0) {
 188                 log_full_errno(priority, errno, "Failed to mount %s at %s: %m", p->type, p->where);
 189                 return (p->mode & MNT_FATAL) ? -errno : 0;
 190         }
 191
 192         /* Relabel again, since we now mounted something fresh here */
 193         if (relabel)
 194                 (void) label_fix(p->where, 0);
 195
 196         if (p->mode & MNT_CHECK_WRITABLE) {
 197                 if (access(p->where, W_OK) < 0) {
 198                         r = -errno;
 199
 200                         (void) umount(p->where);
 201                         (void) rmdir(p->where);
 202
 203                         log_full_errno(priority, r, "Mount point %s not writable after mounting: %m", p->where);
 204                         return (p->mode & MNT_FATAL) ? r : 0;
 205                 }
 206         }
 207
 208         return 1;
 209 }
 210
 211 static int mount_points_setup(unsigned n, bool loaded_policy) {
 212         unsigned i;
 213         int r = 0;
 214
 215         for (i = 0; i < n; i ++) {
 216                 int j;
 217
 218                 j = mount_one(mount_table + i, loaded_policy);
 219                 if (j != 0 && r >= 0)
 220                         r = j;
 221         }
 222
 223         return r;
 224 }
 225
 226 int mount_setup_early(void) {
 227         assert_cc(N_EARLY_MOUNT <= ELEMENTSOF(mount_table));
 228
 229         /* Do a minimal mount of /proc and friends to enable the most
 230          * basic stuff, such as SELinux */
 231         return mount_points_setup(N_EARLY_MOUNT, false);
 232 }
 233
 234 static const char *join_with(const char *controller) {
 235
 236         static const char* const pairs[] = {
 237                 "cpu", "cpuacct",
 238                 "net_cls", "net_prio",
 239                 NULL
 240         };
 241
 242         const char *const *x, *const *y;
 243
 244         assert(controller);
 245
 246         /* This will lookup which controller to mount another controller with. Input is a controller name, and output
 247          * is the other controller name. The function works both ways: you can input one and get the other, and input
 248          * the other to get the one. */
 249
 250         STRV_FOREACH_PAIR(x, y, pairs) {
 251                 if (streq(controller, *x))
 252                         return *y;
 253                 if (streq(controller, *y))
 254                         return *x;
 255         }
 256
 257         return NULL;
 258 }
 259
 260 static int symlink_controller(const char *target, const char *alias) {
 261         const char *a;
 262         int r;
 263
 264         assert(target);
 265         assert(alias);
 266
 267         a = strjoina("/sys/fs/cgroup/", alias);
 268
 269         r = symlink_idempotent(target, a, false);
 270         if (r < 0)
 271                 return log_error_errno(r, "Failed to create symlink %s: %m", a);
 272
 273 #ifdef SMACK_RUN_LABEL
 274         const char *p;
 275
 276         p = strjoina("/sys/fs/cgroup/", target);
 277
 278         r = mac_smack_copy(a, p);
 279         if (r < 0 && r != -EOPNOTSUPP)
 280                 return log_error_errno(r, "Failed to copy smack label from %s to %s: %m", p, a);
 281 #endif
 282
 283         return 0;
 284 }
 285
 286 int mount_cgroup_controllers(void) {
 287         _cleanup_set_free_free_ Set *controllers = NULL;
 288         int r;
 289
 290         if (!cg_is_legacy_wanted())
 291                 return 0;
 292
 293         /* Mount all available cgroup controllers that are built into the kernel. */
 294         r = cg_kernel_controllers(&controllers);
 295         if (r < 0)
 296                 return log_error_errno(r, "Failed to enumerate cgroup controllers: %m");
 297
 298         for (;;) {
 299                 _cleanup_free_ char *options = NULL, *controller = NULL, *where = NULL;
 300                 const char *other_controller;
 301                 MountPoint p = {
 302                         .what = "cgroup",
 303                         .type = "cgroup",
 304                         .flags = MS_NOSUID|MS_NOEXEC|MS_NODEV,
 305                         .mode = MNT_IN_CONTAINER,
 306                 };
 307
 308                 controller = set_steal_first(controllers);
 309                 if (!controller)
 310                         break;
 311
 312                 /* Check if we shall mount this together with another controller */
 313                 other_controller = join_with(controller);
 314                 if (other_controller) {
 315                         _cleanup_free_ char *c = NULL;
 316
 317                         /* Check if the other controller is actually available in the kernel too */
 318                         c = set_remove(controllers, other_controller);
 319                         if (c) {
 320
 321                                 /* Join the two controllers into one string, and maintain a stable ordering */
 322                                 if (strcmp(controller, other_controller) < 0)
 323                                         options = strjoin(controller, ",", other_controller);
 324                                 else
 325                                         options = strjoin(other_controller, ",", controller);
 326                                 if (!options)
 327                                         return log_oom();
 328                         }
 329                 }
 330
 331                 /* The simple case, where there's only one controller to mount together */
 332                 if (!options)
 333                         options = TAKE_PTR(controller);
 334
 335                 where = strappend("/sys/fs/cgroup/", options);
 336                 if (!where)
 337                         return log_oom();
 338
 339                 p.where = where;
 340                 p.options = options;
 341
 342                 r = mount_one(&p, true);
 343                 if (r < 0)
 344                         return r;
 345
 346                 /* Create symlinks from the individual controller names, in case we have a joined mount */
 347                 if (controller)
 348                         (void) symlink_controller(options, controller);
 349                 if (other_controller)
 350                         (void) symlink_controller(options, other_controller);
 351         }
 352
 353         /* Now that we mounted everything, let's make the tmpfs the cgroup file systems are mounted into read-only. */
 354         (void) mount("tmpfs", "/sys/fs/cgroup", "tmpfs", MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME|MS_RDONLY, "mode=755");
 355
 356         return 0;
 357 }
 358
 359 #if HAVE_SELINUX || ENABLE_SMACK
 360 static int nftw_cb(
 361                 const char *fpath,
 362                 const struct stat *sb,
 363                 int tflag,
 364                 struct FTW *ftwbuf) {
 365
 366         /* No need to label /dev twice in a row... */
 367         if (_unlikely_(ftwbuf->level == 0))
 368                 return FTW_CONTINUE;
 369
 370         (void) label_fix(fpath, 0);
 371
 372         /* /run/initramfs is static data and big, no need to
 373          * dynamically relabel its contents at boot... */
 374         if (_unlikely_(ftwbuf->level == 1 &&
 375                       tflag == FTW_D &&
 376                       streq(fpath, "/run/initramfs")))
 377                 return FTW_SKIP_SUBTREE;
 378
 379         return FTW_CONTINUE;
 380 };
 381
 382 static int relabel_cgroup_filesystems(void) {
 383         int r;
 384         struct statfs st;
 385
 386         r = cg_all_unified();
 387         if (r == 0) {
 388                 /* Temporarily remount the root cgroup filesystem to give it a proper label. Do this
 389                    only when the filesystem has been already populated by a previous instance of systemd
 390                    running from initrd. Otherwise don't remount anything and leave the filesystem read-write
 391                    for the cgroup filesystems to be mounted inside. */
 392                 if (statfs("/sys/fs/cgroup", &st) < 0)
 393                         return log_error_errno(errno, "Failed to determine mount flags for /sys/fs/cgroup: %m");
 394
 395                 if (st.f_flags & ST_RDONLY)
 396                         (void) mount(NULL, "/sys/fs/cgroup", NULL, MS_REMOUNT, NULL);
 397
 398                 (void) label_fix("/sys/fs/cgroup", 0);
 399                 (void) nftw("/sys/fs/cgroup", nftw_cb, 64, FTW_MOUNT|FTW_PHYS|FTW_ACTIONRETVAL);
 400
 401                 if (st.f_flags & ST_RDONLY)
 402                         (void) mount(NULL, "/sys/fs/cgroup", NULL, MS_REMOUNT|MS_RDONLY, NULL);
 403
 404         } else if (r < 0)
 405                 return log_error_errno(r, "Failed to determine whether we are in all unified mode: %m");
 406
 407         return 0;
 408 }
 409
 410 static int relabel_extra(void) {
 411         _cleanup_closedir_ DIR *d = NULL;
 412         int r, c = 0;
 413
 414         /* Support for relabelling additional files or directories after loading the policy. For this, code in the
 415          * initrd simply has to drop in *.relabel files into /run/systemd/relabel-extra.d/. We'll read all such files
 416          * expecting one absolute path by line and will relabel each (and everyone below that in case the path refers
 417          * to a directory). These drop-in files are supposed to be absolutely minimal, and do not understand comments
 418          * and such. After the operation succeeded the files are removed, and the drop-in directory as well, if
 419          * possible.
 420          */
 421
 422         d = opendir("/run/systemd/relabel-extra.d/");
 423         if (!d) {
 424                 if (errno == ENOENT)
 425                         return 0;
 426
 427                 return log_warning_errno(errno, "Failed to open /run/systemd/relabel-extra.d/, ignoring: %m");
 428         }
 429
 430         for (;;) {
 431                 _cleanup_fclose_ FILE *f = NULL;
 432                 _cleanup_close_ int fd = -1;
 433                 struct dirent *de;
 434
 435                 errno = 0;
 436                 de = readdir_no_dot(d);
 437                 if (!de) {
 438                         if (errno != 0)
 439                                 return log_error_errno(errno, "Failed read directory /run/systemd/relabel-extra.d/, ignoring: %m");
 440                         break;
 441                 }
 442
 443                 if (hidden_or_backup_file(de->d_name))
 444                         continue;
 445
 446                 if (!endswith(de->d_name, ".relabel"))
 447                         continue;
 448
 449                 if (!IN_SET(de->d_type, DT_REG, DT_UNKNOWN))
 450                         continue;
 451
 452                 fd = openat(dirfd(d), de->d_name, O_RDONLY|O_CLOEXEC|O_NONBLOCK);
 453                 if (fd < 0) {
 454                         log_warning_errno(errno, "Failed to open /run/systemd/relabel-extra.d/%s, ignoring: %m", de->d_name);
 455                         continue;
 456                 }
 457
 458                 f = fdopen(fd, "r");
 459                 if (!f) {
 460                         log_warning_errno(errno, "Failed to convert file descriptor into file object, ignoring: %m");
 461                         continue;
 462                 }
 463                 TAKE_FD(fd);
 464
 465                 for (;;) {
 466                         _cleanup_free_ char *line = NULL;
 467
 468                         r = read_line(f, LONG_LINE_MAX, &line);
 469                         if (r < 0) {
 470                                 log_warning_errno(r, "Failed to read from /run/systemd/relabel-extra.d/%s, ignoring: %m", de->d_name);
 471                                 break;
 472                         }
 473                         if (r == 0) /* EOF */
 474                                 break;
 475
 476                         path_simplify(line, true);
 477
 478                         if (!path_is_normalized(line)) {
 479                                 log_warning("Path to relabel is not normalized, ignoring: %s", line);
 480                                 continue;
 481                         }
 482
 483                         if (!path_is_absolute(line)) {
 484                                 log_warning("Path to relabel is not absolute, ignoring: %s", line);
 485                                 continue;
 486                         }
 487
 488                         log_debug("Relabelling additional file/directory '%s'.", line);
 489                         (void) nftw(line, nftw_cb, 64, FTW_MOUNT|FTW_PHYS|FTW_ACTIONRETVAL);
 490                         c++;
 491                 }
 492
 493                 if (unlinkat(dirfd(d), de->d_name, 0) < 0)
 494                         log_warning_errno(errno, "Failed to remove /run/systemd/relabel-extra.d/%s, ignoring: %m", de->d_name);
 495         }
 496
 497         /* Remove when we completing things. */
 498         if (rmdir("/run/systemd/relabel-extra.d") < 0)
 499                 log_warning_errno(errno, "Failed to remove /run/systemd/relabel-extra.d/ directory: %m");
 500
 501         return c;
 502 }
 503 #endif
 504
 505 int mount_setup(bool loaded_policy) {
 506         int r = 0;
 507
 508         r = mount_points_setup(ELEMENTSOF(mount_table), loaded_policy);
 509         if (r < 0)
 510                 return r;
 511
 512 #if HAVE_SELINUX || ENABLE_SMACK
 513         /* Nodes in devtmpfs and /run need to be manually updated for
 514          * the appropriate labels, after mounting. The other virtual
 515          * API file systems like /sys and /proc do not need that, they
 516          * use the same label for all their files. */
 517         if (loaded_policy) {
 518                 usec_t before_relabel, after_relabel;
 519                 char timespan[FORMAT_TIMESPAN_MAX];
 520                 const char *i;
 521                 int n_extra;
 522
 523                 before_relabel = now(CLOCK_MONOTONIC);
 524
 525                 FOREACH_STRING(i, "/dev", "/dev/shm", "/run")
 526                         (void) nftw(i, nftw_cb, 64, FTW_MOUNT|FTW_PHYS|FTW_ACTIONRETVAL);
 527
 528                 (void) relabel_cgroup_filesystems();
 529
 530                 n_extra = relabel_extra();
 531
 532                 after_relabel = now(CLOCK_MONOTONIC);
 533
 534                 log_info("Relabelled /dev, /dev/shm, /run, /sys/fs/cgroup%s in %s.",
 535                          n_extra > 0 ? ", additional files" : "",
 536                          format_timespan(timespan, sizeof(timespan), after_relabel - before_relabel, 0));
 537         }
 538 #endif
 539
 540         /* Create a few default symlinks, which are normally created
 541          * by udevd, but some scripts might need them before we start
 542          * udevd. */
 543         dev_setup(NULL, UID_INVALID, GID_INVALID);
 544
 545         /* Mark the root directory as shared in regards to mount propagation. The kernel defaults to "private", but we
 546          * think it makes more sense to have a default of "shared" so that nspawn and the container tools work out of
 547          * the box. If specific setups need other settings they can reset the propagation mode to private if
 548          * needed. Note that we set this only when we are invoked directly by the kernel. If we are invoked by a
 549          * container manager we assume the container manager knows what it is doing (for example, because it set up
 550          * some directories with different propagation modes). */
 551         if (detect_container() <= 0)
 552                 if (mount(NULL, "/", NULL, MS_REC|MS_SHARED, NULL) < 0)
 553                         log_warning_errno(errno, "Failed to set up the root directory for shared mount propagation: %m");
 554
 555         /* Create a few directories we always want around, Note that sd_booted() checks for /run/systemd/system, so
 556          * this mkdir really needs to stay for good, otherwise software that copied sd-daemon.c into their sources will
 557          * misdetect systemd. */
 558         (void) mkdir_label("/run/systemd", 0755);
 559         (void) mkdir_label("/run/systemd/system", 0755);
 560
 561         /* Also create /run/systemd/inaccessible nodes, so that we always have something to mount inaccessible nodes
 562          * from. */
 563         (void) make_inaccessible_nodes(NULL, UID_INVALID, GID_INVALID);
 564
 565         return 0;
 566 }