src/shared/dissect-image.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #if HAVE_VALGRIND_MEMCHECK_H
   4 #include <valgrind/memcheck.h>
   5 #endif
   6
   7 #include <linux/dm-ioctl.h>
   8 #include <linux/loop.h>
   9 #include <sys/mount.h>
  10 #include <sys/prctl.h>
  11 #include <sys/wait.h>
  12
  13 #include "sd-device.h"
  14 #include "sd-id128.h"
  15
  16 #include "architecture.h"
  17 #include "ask-password-api.h"
  18 #include "blkid-util.h"
  19 #include "blockdev-util.h"
  20 #include "copy.h"
  21 #include "crypt-util.h"
  22 #include "def.h"
  23 #include "device-nodes.h"
  24 #include "device-util.h"
  25 #include "dissect-image.h"
  26 #include "dm-util.h"
  27 #include "env-file.h"
  28 #include "fd-util.h"
  29 #include "fileio.h"
  30 #include "fs-util.h"
  31 #include "fsck-util.h"
  32 #include "gpt.h"
  33 #include "hexdecoct.h"
  34 #include "hostname-util.h"
  35 #include "id128-util.h"
  36 #include "mount-util.h"
  37 #include "mountpoint-util.h"
  38 #include "nulstr-util.h"
  39 #include "os-util.h"
  40 #include "path-util.h"
  41 #include "process-util.h"
  42 #include "raw-clone.h"
  43 #include "signal-util.h"
  44 #include "stat-util.h"
  45 #include "stdio-util.h"
  46 #include "string-table.h"
  47 #include "string-util.h"
  48 #include "strv.h"
  49 #include "tmpfile-util.h"
  50 #include "udev-util.h"
  51 #include "user-util.h"
  52 #include "xattr-util.h"
  53
  54 int probe_filesystem(const char *node, char **ret_fstype) {
  55         /* Try to find device content type and return it in *ret_fstype. If nothing is found,
  56          * 0/NULL will be returned. -EUCLEAN will be returned for ambiguous results, and an
  57          * different error otherwise. */
  58
  59 #if HAVE_BLKID
  60         _cleanup_(blkid_free_probep) blkid_probe b = NULL;
  61         const char *fstype;
  62         int r;
  63
  64         errno = 0;
  65         b = blkid_new_probe_from_filename(node);
  66         if (!b)
  67                 return errno_or_else(ENOMEM);
  68
  69         blkid_probe_enable_superblocks(b, 1);
  70         blkid_probe_set_superblocks_flags(b, BLKID_SUBLKS_TYPE);
  71
  72         errno = 0;
  73         r = blkid_do_safeprobe(b);
  74         if (r == 1) {
  75                 log_debug("No type detected on partition %s", node);
  76                 goto not_found;
  77         }
  78         if (r == -2)
  79                 return log_debug_errno(SYNTHETIC_ERRNO(EUCLEAN),
  80                                        "Results ambiguous for partition %s", node);
  81         if (r != 0)
  82                 return errno_or_else(EIO);
  83
  84         (void) blkid_probe_lookup_value(b, "TYPE", &fstype, NULL);
  85
  86         if (fstype) {
  87                 char *t;
  88
  89                 t = strdup(fstype);
  90                 if (!t)
  91                         return -ENOMEM;
  92
  93                 *ret_fstype = t;
  94                 return 1;
  95         }
  96
  97 not_found:
  98         *ret_fstype = NULL;
  99         return 0;
 100 #else
 101         return -EOPNOTSUPP;
 102 #endif
 103 }
 104
 105 #if HAVE_BLKID
 106 /* Detect RPMB and Boot partitions, which are not listed by blkid.
 107  * See https://github.com/systemd/systemd/issues/5806. */
 108 static bool device_is_mmc_special_partition(sd_device *d) {
 109         const char *sysname;
 110
 111         assert(d);
 112
 113         if (sd_device_get_sysname(d, &sysname) < 0)
 114                 return false;
 115
 116         return startswith(sysname, "mmcblk") &&
 117                 (endswith(sysname, "rpmb") || endswith(sysname, "boot0") || endswith(sysname, "boot1"));
 118 }
 119
 120 static bool device_is_block(sd_device *d) {
 121         const char *ss;
 122
 123         assert(d);
 124
 125         if (sd_device_get_subsystem(d, &ss) < 0)
 126                 return false;
 127
 128         return streq(ss, "block");
 129 }
 130
 131 static int enumerator_for_parent(sd_device *d, sd_device_enumerator **ret) {
 132         _cleanup_(sd_device_enumerator_unrefp) sd_device_enumerator *e = NULL;
 133         int r;
 134
 135         assert(d);
 136         assert(ret);
 137
 138         r = sd_device_enumerator_new(&e);
 139         if (r < 0)
 140                 return r;
 141
 142         r = sd_device_enumerator_allow_uninitialized(e);
 143         if (r < 0)
 144                 return r;
 145
 146         r = sd_device_enumerator_add_match_parent(e, d);
 147         if (r < 0)
 148                 return r;
 149
 150         *ret = TAKE_PTR(e);
 151         return 0;
 152 }
 153
 154 /* how many times to wait for the device nodes to appear */
 155 #define N_DEVICE_NODE_LIST_ATTEMPTS 10
 156
 157 static int wait_for_partitions_to_appear(
 158                 int fd,
 159                 sd_device *d,
 160                 unsigned num_partitions,
 161                 DissectImageFlags flags,
 162                 sd_device_enumerator **ret_enumerator) {
 163
 164         _cleanup_(sd_device_enumerator_unrefp) sd_device_enumerator *e = NULL;
 165         sd_device *q;
 166         unsigned n;
 167         int r;
 168
 169         assert(fd >= 0);
 170         assert(d);
 171         assert(ret_enumerator);
 172
 173         r = enumerator_for_parent(d, &e);
 174         if (r < 0)
 175                 return r;
 176
 177         /* Count the partitions enumerated by the kernel */
 178         n = 0;
 179         FOREACH_DEVICE(e, q) {
 180                 if (sd_device_get_devnum(q, NULL) < 0)
 181                         continue;
 182                 if (!device_is_block(q))
 183                         continue;
 184                 if (device_is_mmc_special_partition(q))
 185                         continue;
 186
 187                 if (!FLAGS_SET(flags, DISSECT_IMAGE_NO_UDEV)) {
 188                         r = device_wait_for_initialization(q, "block", USEC_INFINITY, NULL);
 189                         if (r < 0)
 190                                 return r;
 191                 }
 192
 193                 n++;
 194         }
 195
 196         if (n == num_partitions + 1) {
 197                 *ret_enumerator = TAKE_PTR(e);
 198                 return 0; /* success! */
 199         }
 200         if (n > num_partitions + 1)
 201                 return log_debug_errno(SYNTHETIC_ERRNO(EIO),
 202                                        "blkid and kernel partition lists do not match.");
 203
 204         /* The kernel has probed fewer partitions than blkid? Maybe the kernel prober is still running or it
 205          * got EBUSY because udev already opened the device. Let's reprobe the device, which is a synchronous
 206          * call that waits until probing is complete. */
 207
 208         for (unsigned j = 0; ; j++) {
 209                 if (j++ > 20)
 210                         return -EBUSY;
 211
 212                 if (ioctl(fd, BLKRRPART, 0) >= 0)
 213                         break;
 214                 r = -errno;
 215                 if (r == -EINVAL) {
 216                         struct loop_info64 info;
 217
 218                         /* If we are running on a loop device that has partition scanning off, return
 219                          * an explicit recognizable error about this, so that callers can generate a
 220                          * proper message explaining the situation. */
 221
 222                         if (ioctl(fd, LOOP_GET_STATUS64, &info) >= 0) {
 223 #if HAVE_VALGRIND_MEMCHECK_H
 224                                 /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
 225                                 VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info));
 226 #endif
 227
 228                                 if ((info.lo_flags & LO_FLAGS_PARTSCAN) == 0)
 229                                         return log_debug_errno(EPROTONOSUPPORT,
 230                                                                "Device is a loop device and partition scanning is off!");
 231                         }
 232                 }
 233                 if (r != -EBUSY)
 234                         return r;
 235
 236                 /* If something else has the device open, such as an udev rule, the ioctl will return
 237                  * EBUSY. Since there's no way to wait until it isn't busy anymore, let's just wait a bit,
 238                  * and try again.
 239                  *
 240                  * This is really something they should fix in the kernel! */
 241                 (void) usleep(50 * USEC_PER_MSEC);
 242
 243         }
 244
 245         return -EAGAIN; /* no success yet, try again */
 246 }
 247
 248 static int loop_wait_for_partitions_to_appear(
 249                 int fd,
 250                 sd_device *d,
 251                 unsigned num_partitions,
 252                 DissectImageFlags flags,
 253                 sd_device_enumerator **ret_enumerator) {
 254         _cleanup_(sd_device_unrefp) sd_device *device = NULL;
 255         int r;
 256
 257         assert(fd >= 0);
 258         assert(d);
 259         assert(ret_enumerator);
 260
 261         log_debug("Waiting for device (parent + %d partitions) to appear...", num_partitions);
 262
 263         if (!FLAGS_SET(flags, DISSECT_IMAGE_NO_UDEV)) {
 264                 r = device_wait_for_initialization(d, "block", USEC_INFINITY, &device);
 265                 if (r < 0)
 266                         return r;
 267         } else
 268                 device = sd_device_ref(d);
 269
 270         for (unsigned i = 0; i < N_DEVICE_NODE_LIST_ATTEMPTS; i++) {
 271                 r = wait_for_partitions_to_appear(fd, device, num_partitions, flags, ret_enumerator);
 272                 if (r != -EAGAIN)
 273                         return r;
 274         }
 275
 276         return log_debug_errno(SYNTHETIC_ERRNO(ENXIO),
 277                                "Kernel partitions dit not appear within %d attempts",
 278                                N_DEVICE_NODE_LIST_ATTEMPTS);
 279 }
 280
 281 static void check_partition_flags(
 282                 const char *node,
 283                 unsigned long long pflags,
 284                 unsigned long long supported) {
 285
 286         assert(node);
 287
 288         /* Mask away all flags supported by this partition's type and the three flags the UEFI spec defines generically */
 289         pflags &= ~(supported | GPT_FLAG_REQUIRED_PARTITION | GPT_FLAG_NO_BLOCK_IO_PROTOCOL | GPT_FLAG_LEGACY_BIOS_BOOTABLE);
 290
 291         if (pflags == 0)
 292                 return;
 293
 294         /* If there are other bits set, then log about it, to make things discoverable */
 295         for (unsigned i = 0; i < sizeof(pflags) * 8; i++) {
 296                 unsigned long long bit = 1ULL << i;
 297                 if (!FLAGS_SET(pflags, bit))
 298                         continue;
 299
 300                 log_debug("Unexpected partition flag %llu set on %s!", bit, node);
 301         }
 302 }
 303
 304 #endif
 305
 306 int dissect_image(
 307                 int fd,
 308                 const void *root_hash,
 309                 size_t root_hash_size,
 310                 const char *verity_data,
 311                 DissectImageFlags flags,
 312                 DissectedImage **ret) {
 313
 314 #if HAVE_BLKID
 315         sd_id128_t root_uuid = SD_ID128_NULL, verity_uuid = SD_ID128_NULL;
 316         _cleanup_(sd_device_enumerator_unrefp) sd_device_enumerator *e = NULL;
 317         bool is_gpt, is_mbr, generic_rw, multiple_generic = false;
 318         _cleanup_(sd_device_unrefp) sd_device *d = NULL;
 319         _cleanup_(dissected_image_unrefp) DissectedImage *m = NULL;
 320         _cleanup_(blkid_free_probep) blkid_probe b = NULL;
 321         _cleanup_free_ char *generic_node = NULL;
 322         sd_id128_t generic_uuid = SD_ID128_NULL;
 323         const char *pttype = NULL;
 324         blkid_partlist pl;
 325         int r, generic_nr;
 326         struct stat st;
 327         sd_device *q;
 328         unsigned i;
 329
 330         assert(fd >= 0);
 331         assert(ret);
 332         assert(root_hash || root_hash_size == 0);
 333         assert(!((flags & DISSECT_IMAGE_GPT_ONLY) && (flags & DISSECT_IMAGE_NO_PARTITION_TABLE)));
 334
 335         /* Probes a disk image, and returns information about what it found in *ret.
 336          *
 337          * Returns -ENOPKG if no suitable partition table or file system could be found.
 338          * Returns -EADDRNOTAVAIL if a root hash was specified but no matching root/verity partitions found. */
 339
 340         if (root_hash) {
 341                 /* If a root hash is supplied, then we use the root partition that has a UUID that match the first
 342                  * 128bit of the root hash. And we use the verity partition that has a UUID that match the final
 343                  * 128bit. */
 344
 345                 if (root_hash_size < sizeof(sd_id128_t))
 346                         return -EINVAL;
 347
 348                 memcpy(&root_uuid, root_hash, sizeof(sd_id128_t));
 349                 memcpy(&verity_uuid, (const uint8_t*) root_hash + root_hash_size - sizeof(sd_id128_t), sizeof(sd_id128_t));
 350
 351                 if (sd_id128_is_null(root_uuid))
 352                         return -EINVAL;
 353                 if (sd_id128_is_null(verity_uuid))
 354                         return -EINVAL;
 355         }
 356
 357         if (fstat(fd, &st) < 0)
 358                 return -errno;
 359
 360         if (!S_ISBLK(st.st_mode))
 361                 return -ENOTBLK;
 362
 363         b = blkid_new_probe();
 364         if (!b)
 365                 return -ENOMEM;
 366
 367         errno = 0;
 368         r = blkid_probe_set_device(b, fd, 0, 0);
 369         if (r != 0)
 370                 return errno_or_else(ENOMEM);
 371
 372         if ((flags & DISSECT_IMAGE_GPT_ONLY) == 0) {
 373                 /* Look for file system superblocks, unless we only shall look for GPT partition tables */
 374                 blkid_probe_enable_superblocks(b, 1);
 375                 blkid_probe_set_superblocks_flags(b, BLKID_SUBLKS_TYPE|BLKID_SUBLKS_USAGE);
 376         }
 377
 378         blkid_probe_enable_partitions(b, 1);
 379         blkid_probe_set_partitions_flags(b, BLKID_PARTS_ENTRY_DETAILS);
 380
 381         errno = 0;
 382         r = blkid_do_safeprobe(b);
 383         if (IN_SET(r, -2, 1))
 384                 return log_debug_errno(SYNTHETIC_ERRNO(ENOPKG), "Failed to identify any partition table.");
 385         if (r != 0)
 386                 return errno_or_else(EIO);
 387
 388         m = new0(DissectedImage, 1);
 389         if (!m)
 390                 return -ENOMEM;
 391
 392         r = sd_device_new_from_devnum(&d, 'b', st.st_rdev);
 393         if (r < 0)
 394                 return r;
 395
 396         if ((!(flags & DISSECT_IMAGE_GPT_ONLY) &&
 397             (flags & DISSECT_IMAGE_REQUIRE_ROOT)) ||
 398             (flags & DISSECT_IMAGE_NO_PARTITION_TABLE)) {
 399                 const char *usage = NULL;
 400
 401                 (void) blkid_probe_lookup_value(b, "USAGE", &usage, NULL);
 402                 if (STRPTR_IN_SET(usage, "filesystem", "crypto")) {
 403                         _cleanup_free_ char *t = NULL, *n = NULL;
 404                         const char *fstype = NULL;
 405
 406                         /* OK, we have found a file system, that's our root partition then. */
 407                         (void) blkid_probe_lookup_value(b, "TYPE", &fstype, NULL);
 408
 409                         if (fstype) {
 410                                 t = strdup(fstype);
 411                                 if (!t)
 412                                         return -ENOMEM;
 413                         }
 414
 415                         r = device_path_make_major_minor(st.st_mode, st.st_rdev, &n);
 416                         if (r < 0)
 417                                 return r;
 418
 419                         m->single_file_system = true;
 420                         m->verity = root_hash && verity_data;
 421                         m->can_verity = !!verity_data;
 422
 423                         m->partitions[PARTITION_ROOT] = (DissectedPartition) {
 424                                 .found = true,
 425                                 .rw = !m->verity,
 426                                 .partno = -1,
 427                                 .architecture = _ARCHITECTURE_INVALID,
 428                                 .fstype = TAKE_PTR(t),
 429                                 .node = TAKE_PTR(n),
 430                         };
 431
 432                         m->encrypted = streq_ptr(fstype, "crypto_LUKS");
 433
 434                         /* Even on a single partition we need to wait for udev to create the
 435                          * /dev/block/X:Y symlink to /dev/loopZ */
 436                         r = loop_wait_for_partitions_to_appear(fd, d, 0, flags, &e);
 437                         if (r < 0)
 438                                 return r;
 439                         *ret = TAKE_PTR(m);
 440
 441                         return 0;
 442                 }
 443         }
 444
 445         (void) blkid_probe_lookup_value(b, "PTTYPE", &pttype, NULL);
 446         if (!pttype)
 447                 return -ENOPKG;
 448
 449         is_gpt = streq_ptr(pttype, "gpt");
 450         is_mbr = streq_ptr(pttype, "dos");
 451
 452         if (!is_gpt && ((flags & DISSECT_IMAGE_GPT_ONLY) || !is_mbr))
 453                 return -ENOPKG;
 454
 455         errno = 0;
 456         pl = blkid_probe_get_partitions(b);
 457         if (!pl)
 458                 return errno_or_else(ENOMEM);
 459
 460         r = loop_wait_for_partitions_to_appear(fd, d, blkid_partlist_numof_partitions(pl), flags, &e);
 461         if (r < 0)
 462                 return r;
 463
 464         FOREACH_DEVICE(e, q) {
 465                 unsigned long long pflags;
 466                 blkid_partition pp;
 467                 const char *node;
 468                 dev_t qn;
 469                 int nr;
 470
 471                 r = sd_device_get_devnum(q, &qn);
 472                 if (r < 0)
 473                         continue;
 474
 475                 if (st.st_rdev == qn)
 476                         continue;
 477
 478                 if (!device_is_block(q))
 479                         continue;
 480
 481                 if (device_is_mmc_special_partition(q))
 482                         continue;
 483
 484                 r = sd_device_get_devname(q, &node);
 485                 if (r < 0)
 486                         continue;
 487
 488                 pp = blkid_partlist_devno_to_partition(pl, qn);
 489                 if (!pp)
 490                         continue;
 491
 492                 pflags = blkid_partition_get_flags(pp);
 493
 494                 nr = blkid_partition_get_partno(pp);
 495                 if (nr < 0)
 496                         continue;
 497
 498                 if (is_gpt) {
 499                         int designator = _PARTITION_DESIGNATOR_INVALID, architecture = _ARCHITECTURE_INVALID;
 500                         const char *stype, *sid, *fstype = NULL;
 501                         sd_id128_t type_id, id;
 502                         bool rw = true;
 503
 504                         sid = blkid_partition_get_uuid(pp);
 505                         if (!sid)
 506                                 continue;
 507                         if (sd_id128_from_string(sid, &id) < 0)
 508                                 continue;
 509
 510                         stype = blkid_partition_get_type_string(pp);
 511                         if (!stype)
 512                                 continue;
 513                         if (sd_id128_from_string(stype, &type_id) < 0)
 514                                 continue;
 515
 516                         if (sd_id128_equal(type_id, GPT_HOME)) {
 517
 518                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 519
 520                                 if (pflags & GPT_FLAG_NO_AUTO)
 521                                         continue;
 522
 523                                 designator = PARTITION_HOME;
 524                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 525                         } else if (sd_id128_equal(type_id, GPT_SRV)) {
 526
 527                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 528
 529                                 if (pflags & GPT_FLAG_NO_AUTO)
 530                                         continue;
 531
 532                                 designator = PARTITION_SRV;
 533                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 534                         } else if (sd_id128_equal(type_id, GPT_ESP)) {
 535
 536                                 /* Note that we don't check the GPT_FLAG_NO_AUTO flag for the ESP, as it is not defined
 537                                  * there. We instead check the GPT_FLAG_NO_BLOCK_IO_PROTOCOL, as recommended by the
 538                                  * UEFI spec (See "12.3.3 Number and Location of System Partitions"). */
 539
 540                                 if (pflags & GPT_FLAG_NO_BLOCK_IO_PROTOCOL)
 541                                         continue;
 542
 543                                 designator = PARTITION_ESP;
 544                                 fstype = "vfat";
 545
 546                         } else if (sd_id128_equal(type_id, GPT_XBOOTLDR)) {
 547
 548                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 549
 550                                 if (pflags & GPT_FLAG_NO_AUTO)
 551                                         continue;
 552
 553                                 designator = PARTITION_XBOOTLDR;
 554                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 555                         }
 556 #ifdef GPT_ROOT_NATIVE
 557                         else if (sd_id128_equal(type_id, GPT_ROOT_NATIVE)) {
 558
 559                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 560
 561                                 if (pflags & GPT_FLAG_NO_AUTO)
 562                                         continue;
 563
 564                                 /* If a root ID is specified, ignore everything but the root id */
 565                                 if (!sd_id128_is_null(root_uuid) && !sd_id128_equal(root_uuid, id))
 566                                         continue;
 567
 568                                 designator = PARTITION_ROOT;
 569                                 architecture = native_architecture();
 570                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 571                         } else if (sd_id128_equal(type_id, GPT_ROOT_NATIVE_VERITY)) {
 572
 573                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 574
 575                                 if (pflags & GPT_FLAG_NO_AUTO)
 576                                         continue;
 577
 578                                 m->can_verity = true;
 579
 580                                 /* Ignore verity unless a root hash is specified */
 581                                 if (sd_id128_is_null(verity_uuid) || !sd_id128_equal(verity_uuid, id))
 582                                         continue;
 583
 584                                 designator = PARTITION_ROOT_VERITY;
 585                                 fstype = "DM_verity_hash";
 586                                 architecture = native_architecture();
 587                                 rw = false;
 588                         }
 589 #endif
 590 #ifdef GPT_ROOT_SECONDARY
 591                         else if (sd_id128_equal(type_id, GPT_ROOT_SECONDARY)) {
 592
 593                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 594
 595                                 if (pflags & GPT_FLAG_NO_AUTO)
 596                                         continue;
 597
 598                                 /* If a root ID is specified, ignore everything but the root id */
 599                                 if (!sd_id128_is_null(root_uuid) && !sd_id128_equal(root_uuid, id))
 600                                         continue;
 601
 602                                 designator = PARTITION_ROOT_SECONDARY;
 603                                 architecture = SECONDARY_ARCHITECTURE;
 604                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 605                         } else if (sd_id128_equal(type_id, GPT_ROOT_SECONDARY_VERITY)) {
 606
 607                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 608
 609                                 if (pflags & GPT_FLAG_NO_AUTO)
 610                                         continue;
 611
 612                                 m->can_verity = true;
 613
 614                                 /* Ignore verity unless root has is specified */
 615                                 if (sd_id128_is_null(verity_uuid) || !sd_id128_equal(verity_uuid, id))
 616                                         continue;
 617
 618                                 designator = PARTITION_ROOT_SECONDARY_VERITY;
 619                                 fstype = "DM_verity_hash";
 620                                 architecture = SECONDARY_ARCHITECTURE;
 621                                 rw = false;
 622                         }
 623 #endif
 624                         else if (sd_id128_equal(type_id, GPT_SWAP)) {
 625
 626                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO);
 627
 628                                 if (pflags & GPT_FLAG_NO_AUTO)
 629                                         continue;
 630
 631                                 designator = PARTITION_SWAP;
 632                                 fstype = "swap";
 633                         } else if (sd_id128_equal(type_id, GPT_LINUX_GENERIC)) {
 634
 635                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 636
 637                                 if (pflags & GPT_FLAG_NO_AUTO)
 638                                         continue;
 639
 640                                 if (generic_node)
 641                                         multiple_generic = true;
 642                                 else {
 643                                         generic_nr = nr;
 644                                         generic_rw = !(pflags & GPT_FLAG_READ_ONLY);
 645                                         generic_uuid = id;
 646                                         generic_node = strdup(node);
 647                                         if (!generic_node)
 648                                                 return -ENOMEM;
 649                                 }
 650
 651                         } else if (sd_id128_equal(type_id, GPT_TMP)) {
 652
 653                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 654
 655                                 if (pflags & GPT_FLAG_NO_AUTO)
 656                                         continue;
 657
 658                                 designator = PARTITION_TMP;
 659                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 660
 661                         } else if (sd_id128_equal(type_id, GPT_VAR)) {
 662
 663                                 check_partition_flags(node, pflags, GPT_FLAG_NO_AUTO|GPT_FLAG_READ_ONLY);
 664
 665                                 if (pflags & GPT_FLAG_NO_AUTO)
 666                                         continue;
 667
 668                                 if (!FLAGS_SET(flags, DISSECT_IMAGE_RELAX_VAR_CHECK)) {
 669                                         sd_id128_t var_uuid;
 670
 671                                         /* For /var we insist that the uuid of the partition matches the
 672                                          * HMAC-SHA256 of the /var GPT partition type uuid, keyed by machine
 673                                          * ID. Why? Unlike the other partitions /var is inherently
 674                                          * installation specific, hence we need to be careful not to mount it
 675                                          * in the wrong installation. By hashing the partition UUID from
 676                                          * /etc/machine-id we can securely bind the partition to the
 677                                          * installation. */
 678
 679                                         r = sd_id128_get_machine_app_specific(GPT_VAR, &var_uuid);
 680                                         if (r < 0)
 681                                                 return r;
 682
 683                                         if (!sd_id128_equal(var_uuid, id)) {
 684                                                 log_debug("Found a /var/ partition, but its UUID didn't match our expectations, ignoring.");
 685                                                 continue;
 686                                         }
 687                                 }
 688
 689                                 designator = PARTITION_VAR;
 690                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 691                         }
 692
 693                         if (designator != _PARTITION_DESIGNATOR_INVALID) {
 694                                 _cleanup_free_ char *t = NULL, *n = NULL;
 695
 696                                 /* First one wins */
 697                                 if (m->partitions[designator].found)
 698                                         continue;
 699
 700                                 if (fstype) {
 701                                         t = strdup(fstype);
 702                                         if (!t)
 703                                                 return -ENOMEM;
 704                                 }
 705
 706                                 n = strdup(node);
 707                                 if (!n)
 708                                         return -ENOMEM;
 709
 710                                 m->partitions[designator] = (DissectedPartition) {
 711                                         .found = true,
 712                                         .partno = nr,
 713                                         .rw = rw,
 714                                         .architecture = architecture,
 715                                         .node = TAKE_PTR(n),
 716                                         .fstype = TAKE_PTR(t),
 717                                         .uuid = id,
 718                                 };
 719                         }
 720
 721                 } else if (is_mbr) {
 722
 723                         switch (blkid_partition_get_type(pp)) {
 724
 725                         case 0x83: /* Linux partition */
 726
 727                                 if (pflags != 0x80) /* Bootable flag */
 728                                         continue;
 729
 730                                 if (generic_node)
 731                                         multiple_generic = true;
 732                                 else {
 733                                         generic_nr = nr;
 734                                         generic_rw = true;
 735                                         generic_node = strdup(node);
 736                                         if (!generic_node)
 737                                                 return -ENOMEM;
 738                                 }
 739
 740                                 break;
 741
 742                         case 0xEA: { /* Boot Loader Spec extended $BOOT partition */
 743                                 _cleanup_free_ char *n = NULL;
 744                                 sd_id128_t id = SD_ID128_NULL;
 745                                 const char *sid;
 746
 747                                 /* First one wins */
 748                                 if (m->partitions[PARTITION_XBOOTLDR].found)
 749                                         continue;
 750
 751                                 sid = blkid_partition_get_uuid(pp);
 752                                 if (sid)
 753                                         (void) sd_id128_from_string(sid, &id);
 754
 755                                 n = strdup(node);
 756                                 if (!n)
 757                                         return -ENOMEM;
 758
 759                                 m->partitions[PARTITION_XBOOTLDR] = (DissectedPartition) {
 760                                         .found = true,
 761                                         .partno = nr,
 762                                         .rw = true,
 763                                         .architecture = _ARCHITECTURE_INVALID,
 764                                         .node = TAKE_PTR(n),
 765                                         .uuid = id,
 766                                 };
 767
 768                                 break;
 769                         }}
 770                 }
 771         }
 772
 773         if (!m->partitions[PARTITION_ROOT].found) {
 774                 /* No root partition found? Then let's see if ther's one for the secondary architecture. And if not
 775                  * either, then check if there's a single generic one, and use that. */
 776
 777                 if (m->partitions[PARTITION_ROOT_VERITY].found)
 778                         return -EADDRNOTAVAIL;
 779
 780                 if (m->partitions[PARTITION_ROOT_SECONDARY].found) {
 781                         m->partitions[PARTITION_ROOT] = m->partitions[PARTITION_ROOT_SECONDARY];
 782                         zero(m->partitions[PARTITION_ROOT_SECONDARY]);
 783
 784                         m->partitions[PARTITION_ROOT_VERITY] = m->partitions[PARTITION_ROOT_SECONDARY_VERITY];
 785                         zero(m->partitions[PARTITION_ROOT_SECONDARY_VERITY]);
 786
 787                 } else if (flags & DISSECT_IMAGE_REQUIRE_ROOT) {
 788
 789                         /* If the root has was set, then we won't fallback to a generic node, because the root hash
 790                          * decides */
 791                         if (root_hash)
 792                                 return -EADDRNOTAVAIL;
 793
 794                         /* If we didn't find a generic node, then we can't fix this up either */
 795                         if (!generic_node)
 796                                 return -ENXIO;
 797
 798                         /* If we didn't find a properly marked root partition, but we did find a single suitable
 799                          * generic Linux partition, then use this as root partition, if the caller asked for it. */
 800                         if (multiple_generic)
 801                                 return -ENOTUNIQ;
 802
 803                         m->partitions[PARTITION_ROOT] = (DissectedPartition) {
 804                                 .found = true,
 805                                 .rw = generic_rw,
 806                                 .partno = generic_nr,
 807                                 .architecture = _ARCHITECTURE_INVALID,
 808                                 .node = TAKE_PTR(generic_node),
 809                                 .uuid = generic_uuid,
 810                         };
 811                 }
 812         }
 813
 814         if (root_hash) {
 815                 if (!m->partitions[PARTITION_ROOT_VERITY].found || !m->partitions[PARTITION_ROOT].found)
 816                         return -EADDRNOTAVAIL;
 817
 818                 /* If we found the primary root with the hash, then we definitely want to suppress any secondary root
 819                  * (which would be weird, after all the root hash should only be assigned to one pair of
 820                  * partitions... */
 821                 m->partitions[PARTITION_ROOT_SECONDARY].found = false;
 822                 m->partitions[PARTITION_ROOT_SECONDARY_VERITY].found = false;
 823
 824                 /* If we found a verity setup, then the root partition is necessarily read-only. */
 825                 m->partitions[PARTITION_ROOT].rw = false;
 826
 827                 m->verity = true;
 828         }
 829
 830         blkid_free_probe(b);
 831         b = NULL;
 832
 833         /* Fill in file system types if we don't know them yet. */
 834         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
 835                 DissectedPartition *p = m->partitions + i;
 836
 837                 if (!p->found)
 838                         continue;
 839
 840                 if (!p->fstype && p->node) {
 841                         r = probe_filesystem(p->node, &p->fstype);
 842                         if (r < 0 && r != -EUCLEAN)
 843                                 return r;
 844                 }
 845
 846                 if (streq_ptr(p->fstype, "crypto_LUKS"))
 847                         m->encrypted = true;
 848
 849                 if (p->fstype && fstype_is_ro(p->fstype))
 850                         p->rw = false;
 851         }
 852
 853         *ret = TAKE_PTR(m);
 854
 855         return 0;
 856 #else
 857         return -EOPNOTSUPP;
 858 #endif
 859 }
 860
 861 DissectedImage* dissected_image_unref(DissectedImage *m) {
 862         unsigned i;
 863
 864         if (!m)
 865                 return NULL;
 866
 867         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
 868                 free(m->partitions[i].fstype);
 869                 free(m->partitions[i].node);
 870                 free(m->partitions[i].decrypted_fstype);
 871                 free(m->partitions[i].decrypted_node);
 872         }
 873
 874         free(m->hostname);
 875         strv_free(m->machine_info);
 876         strv_free(m->os_release);
 877
 878         return mfree(m);
 879 }
 880
 881 static int is_loop_device(const char *path) {
 882         char s[SYS_BLOCK_PATH_MAX("/../loop/")];
 883         struct stat st;
 884
 885         assert(path);
 886
 887         if (stat(path, &st) < 0)
 888                 return -errno;
 889
 890         if (!S_ISBLK(st.st_mode))
 891                 return -ENOTBLK;
 892
 893         xsprintf_sys_block_path(s, "/loop/", st.st_dev);
 894         if (access(s, F_OK) < 0) {
 895                 if (errno != ENOENT)
 896                         return -errno;
 897
 898                 /* The device itself isn't a loop device, but maybe it's a partition and its parent is? */
 899                 xsprintf_sys_block_path(s, "/../loop/", st.st_dev);
 900                 if (access(s, F_OK) < 0)
 901                         return errno == ENOENT ? false : -errno;
 902         }
 903
 904         return true;
 905 }
 906
 907 static int run_fsck(const char *node, const char *fstype) {
 908         int r, exit_status;
 909         pid_t pid;
 910
 911         assert(node);
 912         assert(fstype);
 913
 914         r = fsck_exists(fstype);
 915         if (r < 0) {
 916                 log_debug_errno(r, "Couldn't determine whether fsck for %s exists, proceeding anyway.", fstype);
 917                 return 0;
 918         }
 919         if (r == 0) {
 920                 log_debug("Not checking partition %s, as fsck for %s does not exist.", node, fstype);
 921                 return 0;
 922         }
 923
 924         r = safe_fork("(fsck)", FORK_RESET_SIGNALS|FORK_CLOSE_ALL_FDS|FORK_RLIMIT_NOFILE_SAFE|FORK_DEATHSIG|FORK_NULL_STDIO, &pid);
 925         if (r < 0)
 926                 return log_debug_errno(r, "Failed to fork off fsck: %m");
 927         if (r == 0) {
 928                 /* Child */
 929                 execl("/sbin/fsck", "/sbin/fsck", "-aT", node, NULL);
 930                 log_debug_errno(errno, "Failed to execl() fsck: %m");
 931                 _exit(FSCK_OPERATIONAL_ERROR);
 932         }
 933
 934         exit_status = wait_for_terminate_and_check("fsck", pid, 0);
 935         if (exit_status < 0)
 936                 return log_debug_errno(exit_status, "Failed to fork off /sbin/fsck: %m");
 937
 938         if ((exit_status & ~FSCK_ERROR_CORRECTED) != FSCK_SUCCESS) {
 939                 log_debug("fsck failed with exit status %i.", exit_status);
 940
 941                 if ((exit_status & (FSCK_SYSTEM_SHOULD_REBOOT|FSCK_ERRORS_LEFT_UNCORRECTED)) != 0)
 942                         return log_debug_errno(SYNTHETIC_ERRNO(EUCLEAN), "File system is corrupted, refusing.");
 943
 944                 log_debug("Ignoring fsck error.");
 945         }
 946
 947         return 0;
 948 }
 949
 950 static int mount_partition(
 951                 DissectedPartition *m,
 952                 const char *where,
 953                 const char *directory,
 954                 uid_t uid_shift,
 955                 DissectImageFlags flags) {
 956
 957         _cleanup_free_ char *chased = NULL, *options = NULL;
 958         const char *p, *node, *fstype;
 959         bool rw;
 960         int r;
 961
 962         assert(m);
 963         assert(where);
 964
 965         node = m->decrypted_node ?: m->node;
 966         fstype = m->decrypted_fstype ?: m->fstype;
 967
 968         if (!m->found || !node || !fstype)
 969                 return 0;
 970
 971         /* Stacked encryption? Yuck */
 972         if (streq_ptr(fstype, "crypto_LUKS"))
 973                 return -ELOOP;
 974
 975         rw = m->rw && !(flags & DISSECT_IMAGE_READ_ONLY);
 976
 977         if (FLAGS_SET(flags, DISSECT_IMAGE_FSCK) && rw) {
 978                 r = run_fsck(node, fstype);
 979                 if (r < 0)
 980                         return r;
 981         }
 982
 983         if (directory) {
 984                 r = chase_symlinks(directory, where, CHASE_PREFIX_ROOT, &chased, NULL);
 985                 if (r < 0)
 986                         return r;
 987
 988                 p = chased;
 989         } else
 990                 p = where;
 991
 992         /* If requested, turn on discard support. */
 993         if (fstype_can_discard(fstype) &&
 994             ((flags & DISSECT_IMAGE_DISCARD) ||
 995              ((flags & DISSECT_IMAGE_DISCARD_ON_LOOP) && is_loop_device(m->node)))) {
 996                 options = strdup("discard");
 997                 if (!options)
 998                         return -ENOMEM;
 999         }
1000
1001         if (uid_is_valid(uid_shift) && uid_shift != 0 && fstype_can_uid_gid(fstype)) {
1002                 _cleanup_free_ char *uid_option = NULL;
1003
1004                 if (asprintf(&uid_option, "uid=" UID_FMT ",gid=" GID_FMT, uid_shift, (gid_t) uid_shift) < 0)
1005                         return -ENOMEM;
1006
1007                 if (!strextend_with_separator(&options, ",", uid_option, NULL))
1008                         return -ENOMEM;
1009         }
1010
1011         r = mount_verbose(LOG_DEBUG, node, p, fstype, MS_NODEV|(rw ? 0 : MS_RDONLY), options);
1012         if (r < 0)
1013                 return r;
1014
1015         return 1;
1016 }
1017
1018 int dissected_image_mount(DissectedImage *m, const char *where, uid_t uid_shift, DissectImageFlags flags) {
1019         int r, boot_mounted;
1020
1021         assert(m);
1022         assert(where);
1023
1024         if (!m->partitions[PARTITION_ROOT].found)
1025                 return -ENXIO;
1026
1027         if ((flags & DISSECT_IMAGE_MOUNT_NON_ROOT_ONLY) == 0) {
1028                 r = mount_partition(m->partitions + PARTITION_ROOT, where, NULL, uid_shift, flags);
1029                 if (r < 0)
1030                         return r;
1031
1032                 if (flags & DISSECT_IMAGE_VALIDATE_OS) {
1033                         r = path_is_os_tree(where);
1034                         if (r < 0)
1035                                 return r;
1036                         if (r == 0)
1037                                 return -EMEDIUMTYPE;
1038                 }
1039         }
1040
1041         if (flags & DISSECT_IMAGE_MOUNT_ROOT_ONLY)
1042                 return 0;
1043
1044         r = mount_partition(m->partitions + PARTITION_HOME, where, "/home", uid_shift, flags);
1045         if (r < 0)
1046                 return r;
1047
1048         r = mount_partition(m->partitions + PARTITION_SRV, where, "/srv", uid_shift, flags);
1049         if (r < 0)
1050                 return r;
1051
1052         r = mount_partition(m->partitions + PARTITION_VAR, where, "/var", uid_shift, flags);
1053         if (r < 0)
1054                 return r;
1055
1056         r = mount_partition(m->partitions + PARTITION_TMP, where, "/var/tmp", uid_shift, flags);
1057         if (r < 0)
1058                 return r;
1059
1060         boot_mounted = mount_partition(m->partitions + PARTITION_XBOOTLDR, where, "/boot", uid_shift, flags);
1061         if (boot_mounted < 0)
1062                 return boot_mounted;
1063
1064         if (m->partitions[PARTITION_ESP].found) {
1065                 /* Mount the ESP to /efi if it exists. If it doesn't exist, use /boot instead, but only if it
1066                  * exists and is empty, and we didn't already mount the XBOOTLDR partition into it. */
1067
1068                 r = chase_symlinks("/efi", where, CHASE_PREFIX_ROOT, NULL, NULL);
1069                 if (r >= 0) {
1070                         r = mount_partition(m->partitions + PARTITION_ESP, where, "/efi", uid_shift, flags);
1071                         if (r < 0)
1072                                 return r;
1073
1074                 } else if (boot_mounted <= 0) {
1075                         _cleanup_free_ char *p = NULL;
1076
1077                         r = chase_symlinks("/boot", where, CHASE_PREFIX_ROOT, &p, NULL);
1078                         if (r >= 0 && dir_is_empty(p) > 0) {
1079                                 r = mount_partition(m->partitions + PARTITION_ESP, where, "/boot", uid_shift, flags);
1080                                 if (r < 0)
1081                                         return r;
1082                         }
1083                 }
1084         }
1085
1086         return 0;
1087 }
1088
1089 #if HAVE_LIBCRYPTSETUP
1090 typedef struct DecryptedPartition {
1091         struct crypt_device *device;
1092         char *name;
1093         bool relinquished;
1094 } DecryptedPartition;
1095
1096 struct DecryptedImage {
1097         DecryptedPartition *decrypted;
1098         size_t n_decrypted;
1099         size_t n_allocated;
1100 };
1101 #endif
1102
1103 DecryptedImage* decrypted_image_unref(DecryptedImage* d) {
1104 #if HAVE_LIBCRYPTSETUP
1105         size_t i;
1106         int r;
1107
1108         if (!d)
1109                 return NULL;
1110
1111         for (i = 0; i < d->n_decrypted; i++) {
1112                 DecryptedPartition *p = d->decrypted + i;
1113
1114                 if (p->device && p->name && !p->relinquished) {
1115                         r = crypt_deactivate(p->device, p->name);
1116                         if (r < 0)
1117                                 log_debug_errno(r, "Failed to deactivate encrypted partition %s", p->name);
1118                 }
1119
1120                 if (p->device)
1121                         crypt_free(p->device);
1122                 free(p->name);
1123         }
1124
1125         free(d);
1126 #endif
1127         return NULL;
1128 }
1129
1130 #if HAVE_LIBCRYPTSETUP
1131
1132 static int make_dm_name_and_node(const void *original_node, const char *suffix, char **ret_name, char **ret_node) {
1133         _cleanup_free_ char *name = NULL, *node = NULL;
1134         const char *base;
1135
1136         assert(original_node);
1137         assert(suffix);
1138         assert(ret_name);
1139         assert(ret_node);
1140
1141         base = strrchr(original_node, '/');
1142         if (!base)
1143                 return -EINVAL;
1144         base++;
1145         if (isempty(base))
1146                 return -EINVAL;
1147
1148         name = strjoin(base, suffix);
1149         if (!name)
1150                 return -ENOMEM;
1151         if (!filename_is_valid(name))
1152                 return -EINVAL;
1153
1154         node = path_join(crypt_get_dir(), name);
1155         if (!node)
1156                 return -ENOMEM;
1157
1158         *ret_name = TAKE_PTR(name);
1159         *ret_node = TAKE_PTR(node);
1160
1161         return 0;
1162 }
1163
1164 static int decrypt_partition(
1165                 DissectedPartition *m,
1166                 const char *passphrase,
1167                 DissectImageFlags flags,
1168                 DecryptedImage *d) {
1169
1170         _cleanup_free_ char *node = NULL, *name = NULL;
1171         _cleanup_(crypt_freep) struct crypt_device *cd = NULL;
1172         int r;
1173
1174         assert(m);
1175         assert(d);
1176
1177         if (!m->found || !m->node || !m->fstype)
1178                 return 0;
1179
1180         if (!streq(m->fstype, "crypto_LUKS"))
1181                 return 0;
1182
1183         if (!passphrase)
1184                 return -ENOKEY;
1185
1186         r = make_dm_name_and_node(m->node, "-decrypted", &name, &node);
1187         if (r < 0)
1188                 return r;
1189
1190         if (!GREEDY_REALLOC0(d->decrypted, d->n_allocated, d->n_decrypted + 1))
1191                 return -ENOMEM;
1192
1193         r = crypt_init(&cd, m->node);
1194         if (r < 0)
1195                 return log_debug_errno(r, "Failed to initialize dm-crypt: %m");
1196
1197         crypt_set_log_callback(cd, cryptsetup_log_glue, NULL);
1198
1199         r = crypt_load(cd, CRYPT_LUKS, NULL);
1200         if (r < 0)
1201                 return log_debug_errno(r, "Failed to load LUKS metadata: %m");
1202
1203         r = crypt_activate_by_passphrase(cd, name, CRYPT_ANY_SLOT, passphrase, strlen(passphrase),
1204                                          ((flags & DISSECT_IMAGE_READ_ONLY) ? CRYPT_ACTIVATE_READONLY : 0) |
1205                                          ((flags & DISSECT_IMAGE_DISCARD_ON_CRYPTO) ? CRYPT_ACTIVATE_ALLOW_DISCARDS : 0));
1206         if (r < 0) {
1207                 log_debug_errno(r, "Failed to activate LUKS device: %m");
1208                 return r == -EPERM ? -EKEYREJECTED : r;
1209         }
1210
1211         d->decrypted[d->n_decrypted].name = TAKE_PTR(name);
1212         d->decrypted[d->n_decrypted].device = TAKE_PTR(cd);
1213         d->n_decrypted++;
1214
1215         m->decrypted_node = TAKE_PTR(node);
1216
1217         return 0;
1218 }
1219
1220 static int verity_partition(
1221                 DissectedPartition *m,
1222                 DissectedPartition *v,
1223                 const void *root_hash,
1224                 size_t root_hash_size,
1225                 const char *verity_data,
1226                 DissectImageFlags flags,
1227                 DecryptedImage *d) {
1228
1229         _cleanup_free_ char *node = NULL, *name = NULL;
1230         _cleanup_(crypt_freep) struct crypt_device *cd = NULL;
1231         int r;
1232
1233         assert(m);
1234         assert(v || verity_data);
1235
1236         if (!root_hash)
1237                 return 0;
1238
1239         if (!m->found || !m->node || !m->fstype)
1240                 return 0;
1241         if (!verity_data) {
1242                 if (!v->found || !v->node || !v->fstype)
1243                         return 0;
1244
1245                 if (!streq(v->fstype, "DM_verity_hash"))
1246                         return 0;
1247         }
1248
1249         r = make_dm_name_and_node(m->node, "-verity", &name, &node);
1250         if (r < 0)
1251                 return r;
1252
1253         if (!GREEDY_REALLOC0(d->decrypted, d->n_allocated, d->n_decrypted + 1))
1254                 return -ENOMEM;
1255
1256         r = crypt_init(&cd, verity_data ?: v->node);
1257         if (r < 0)
1258                 return r;
1259
1260         crypt_set_log_callback(cd, cryptsetup_log_glue, NULL);
1261
1262         r = crypt_load(cd, CRYPT_VERITY, NULL);
1263         if (r < 0)
1264                 return r;
1265
1266         r = crypt_set_data_device(cd, m->node);
1267         if (r < 0)
1268                 return r;
1269
1270         r = crypt_activate_by_volume_key(cd, name, root_hash, root_hash_size, CRYPT_ACTIVATE_READONLY);
1271         if (r < 0)
1272                 return r;
1273
1274         d->decrypted[d->n_decrypted].name = TAKE_PTR(name);
1275         d->decrypted[d->n_decrypted].device = TAKE_PTR(cd);
1276         d->n_decrypted++;
1277
1278         m->decrypted_node = TAKE_PTR(node);
1279
1280         return 0;
1281 }
1282 #endif
1283
1284 int dissected_image_decrypt(
1285                 DissectedImage *m,
1286                 const char *passphrase,
1287                 const void *root_hash,
1288                 size_t root_hash_size,
1289                 const char *verity_data,
1290                 DissectImageFlags flags,
1291                 DecryptedImage **ret) {
1292
1293 #if HAVE_LIBCRYPTSETUP
1294         _cleanup_(decrypted_image_unrefp) DecryptedImage *d = NULL;
1295         unsigned i;
1296         int r;
1297 #endif
1298
1299         assert(m);
1300         assert(root_hash || root_hash_size == 0);
1301
1302         /* Returns:
1303          *
1304          *      = 0           → There was nothing to decrypt
1305          *      > 0           → Decrypted successfully
1306          *      -ENOKEY       → There's something to decrypt but no key was supplied
1307          *      -EKEYREJECTED → Passed key was not correct
1308          */
1309
1310         if (root_hash && root_hash_size < sizeof(sd_id128_t))
1311                 return -EINVAL;
1312
1313         if (!m->encrypted && !m->verity) {
1314                 *ret = NULL;
1315                 return 0;
1316         }
1317
1318 #if HAVE_LIBCRYPTSETUP
1319         d = new0(DecryptedImage, 1);
1320         if (!d)
1321                 return -ENOMEM;
1322
1323         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
1324                 DissectedPartition *p = m->partitions + i;
1325                 int k;
1326
1327                 if (!p->found)
1328                         continue;
1329
1330                 r = decrypt_partition(p, passphrase, flags, d);
1331                 if (r < 0)
1332                         return r;
1333
1334                 k = PARTITION_VERITY_OF(i);
1335                 if (k >= 0) {
1336                         r = verity_partition(p, m->partitions + k, root_hash, root_hash_size, verity_data, flags, d);
1337                         if (r < 0)
1338                                 return r;
1339                 }
1340
1341                 if (!p->decrypted_fstype && p->decrypted_node) {
1342                         r = probe_filesystem(p->decrypted_node, &p->decrypted_fstype);
1343                         if (r < 0 && r != -EUCLEAN)
1344                                 return r;
1345                 }
1346         }
1347
1348         *ret = TAKE_PTR(d);
1349
1350         return 1;
1351 #else
1352         return -EOPNOTSUPP;
1353 #endif
1354 }
1355
1356 int dissected_image_decrypt_interactively(
1357                 DissectedImage *m,
1358                 const char *passphrase,
1359                 const void *root_hash,
1360                 size_t root_hash_size,
1361                 const char *verity_data,
1362                 DissectImageFlags flags,
1363                 DecryptedImage **ret) {
1364
1365         _cleanup_strv_free_erase_ char **z = NULL;
1366         int n = 3, r;
1367
1368         if (passphrase)
1369                 n--;
1370
1371         for (;;) {
1372                 r = dissected_image_decrypt(m, passphrase, root_hash, root_hash_size, verity_data, flags, ret);
1373                 if (r >= 0)
1374                         return r;
1375                 if (r == -EKEYREJECTED)
1376                         log_error_errno(r, "Incorrect passphrase, try again!");
1377                 else if (r != -ENOKEY)
1378                         return log_error_errno(r, "Failed to decrypt image: %m");
1379
1380                 if (--n < 0)
1381                         return log_error_errno(SYNTHETIC_ERRNO(EKEYREJECTED),
1382                                                "Too many retries.");
1383
1384                 z = strv_free(z);
1385
1386                 r = ask_password_auto("Please enter image passphrase:", NULL, "dissect", "dissect", USEC_INFINITY, 0, &z);
1387                 if (r < 0)
1388                         return log_error_errno(r, "Failed to query for passphrase: %m");
1389
1390                 passphrase = z[0];
1391         }
1392 }
1393
1394 int decrypted_image_relinquish(DecryptedImage *d) {
1395
1396 #if HAVE_LIBCRYPTSETUP
1397         size_t i;
1398         int r;
1399 #endif
1400
1401         assert(d);
1402
1403         /* Turns on automatic removal after the last use ended for all DM devices of this image, and sets a boolean so
1404          * that we don't clean it up ourselves either anymore */
1405
1406 #if HAVE_LIBCRYPTSETUP
1407         for (i = 0; i < d->n_decrypted; i++) {
1408                 DecryptedPartition *p = d->decrypted + i;
1409
1410                 if (p->relinquished)
1411                         continue;
1412
1413                 r = dm_deferred_remove(p->name);
1414                 if (r < 0)
1415                         return log_debug_errno(r, "Failed to mark %s for auto-removal: %m", p->name);
1416
1417                 p->relinquished = true;
1418         }
1419 #endif
1420
1421         return 0;
1422 }
1423
1424 int verity_metadata_load(const char *image, void **ret_roothash, size_t *ret_roothash_size, char **ret_verity_data) {
1425         _cleanup_free_ char *verity_filename = NULL;
1426         _cleanup_free_ void *roothash_decoded = NULL;
1427         size_t roothash_decoded_size = 0;
1428         int r;
1429
1430         assert(image);
1431
1432         if (is_device_path(image)) {
1433                 /* If we are asked to load the root hash for a device node, exit early */
1434                 if (ret_roothash)
1435                         *ret_roothash = NULL;
1436                 if (ret_roothash_size)
1437                         *ret_roothash_size = 0;
1438                 if (ret_verity_data)
1439                         *ret_verity_data = NULL;
1440                 return 0;
1441         }
1442
1443         if (ret_verity_data) {
1444                 char *e;
1445
1446                 verity_filename = new(char, strlen(image) + STRLEN(".verity") + 1);
1447                 if (!verity_filename)
1448                         return -ENOMEM;
1449                 strcpy(verity_filename, image);
1450                 e = endswith(verity_filename, ".raw");
1451                 if (e)
1452                         strcpy(e, ".verity");
1453                 else
1454                         strcat(verity_filename, ".verity");
1455
1456                 r = access(verity_filename, F_OK);
1457                 if (r < 0) {
1458                         if (errno != ENOENT)
1459                                 return -errno;
1460                         verity_filename = mfree(verity_filename);
1461                 }
1462         }
1463
1464         if (ret_roothash) {
1465                 _cleanup_free_ char *text = NULL;
1466                 assert(ret_roothash_size);
1467
1468                 r = getxattr_malloc(image, "user.verity.roothash", &text, true);
1469                 if (r < 0) {
1470                         char *fn, *e, *n;
1471
1472                         if (!IN_SET(r, -ENODATA, -EOPNOTSUPP, -ENOENT))
1473                                 return r;
1474
1475                         fn = newa(char, strlen(image) + STRLEN(".roothash") + 1);
1476                         n = stpcpy(fn, image);
1477                         e = endswith(fn, ".raw");
1478                         if (e)
1479                                 n = e;
1480
1481                         strcpy(n, ".roothash");
1482
1483                         r = read_one_line_file(fn, &text);
1484                         if (r < 0 && r != -ENOENT)
1485                                 return r;
1486                 }
1487
1488                 if (text) {
1489                         r = unhexmem(text, strlen(text), &roothash_decoded, &roothash_decoded_size);
1490                         if (r < 0)
1491                                 return r;
1492                         if (roothash_decoded_size < sizeof(sd_id128_t))
1493                                 return -EINVAL;
1494                 }
1495         }
1496
1497         if (ret_roothash) {
1498                 *ret_roothash = TAKE_PTR(roothash_decoded);
1499                 *ret_roothash_size = roothash_decoded_size;
1500         }
1501         if (ret_verity_data)
1502                 *ret_verity_data = TAKE_PTR(verity_filename);
1503
1504         return 1;
1505 }
1506
1507 int dissected_image_acquire_metadata(DissectedImage *m) {
1508
1509         enum {
1510                 META_HOSTNAME,
1511                 META_MACHINE_ID,
1512                 META_MACHINE_INFO,
1513                 META_OS_RELEASE,
1514                 _META_MAX,
1515         };
1516
1517         static const char *const paths[_META_MAX] = {
1518                 [META_HOSTNAME]     = "/etc/hostname\0",
1519                 [META_MACHINE_ID]   = "/etc/machine-id\0",
1520                 [META_MACHINE_INFO] = "/etc/machine-info\0",
1521                 [META_OS_RELEASE]   = "/etc/os-release\0"
1522                                       "/usr/lib/os-release\0",
1523         };
1524
1525         _cleanup_strv_free_ char **machine_info = NULL, **os_release = NULL;
1526         _cleanup_(rmdir_and_freep) char *t = NULL;
1527         _cleanup_(sigkill_waitp) pid_t child = 0;
1528         sd_id128_t machine_id = SD_ID128_NULL;
1529         _cleanup_free_ char *hostname = NULL;
1530         unsigned n_meta_initialized = 0, k;
1531         int fds[2 * _META_MAX], r;
1532
1533         BLOCK_SIGNALS(SIGCHLD);
1534
1535         assert(m);
1536
1537         for (; n_meta_initialized < _META_MAX; n_meta_initialized ++)
1538                 if (pipe2(fds + 2*n_meta_initialized, O_CLOEXEC) < 0) {
1539                         r = -errno;
1540                         goto finish;
1541                 }
1542
1543         r = mkdtemp_malloc("/tmp/dissect-XXXXXX", &t);
1544         if (r < 0)
1545                 goto finish;
1546
1547         r = safe_fork("(sd-dissect)", FORK_RESET_SIGNALS|FORK_DEATHSIG|FORK_NEW_MOUNTNS|FORK_MOUNTNS_SLAVE, &child);
1548         if (r < 0)
1549                 goto finish;
1550         if (r == 0) {
1551                 r = dissected_image_mount(m, t, UID_INVALID, DISSECT_IMAGE_READ_ONLY|DISSECT_IMAGE_MOUNT_ROOT_ONLY|DISSECT_IMAGE_VALIDATE_OS);
1552                 if (r < 0) {
1553                         log_debug_errno(r, "Failed to mount dissected image: %m");
1554                         _exit(EXIT_FAILURE);
1555                 }
1556
1557                 for (k = 0; k < _META_MAX; k++) {
1558                         _cleanup_close_ int fd = -1;
1559                         const char *p;
1560
1561                         fds[2*k] = safe_close(fds[2*k]);
1562
1563                         NULSTR_FOREACH(p, paths[k]) {
1564                                 fd = chase_symlinks_and_open(p, t, CHASE_PREFIX_ROOT, O_RDONLY|O_CLOEXEC|O_NOCTTY, NULL);
1565                                 if (fd >= 0)
1566                                         break;
1567                         }
1568                         if (fd < 0) {
1569                                 log_debug_errno(fd, "Failed to read %s file of image, ignoring: %m", paths[k]);
1570                                 continue;
1571                         }
1572
1573                         r = copy_bytes(fd, fds[2*k+1], (uint64_t) -1, 0);
1574                         if (r < 0)
1575                                 _exit(EXIT_FAILURE);
1576
1577                         fds[2*k+1] = safe_close(fds[2*k+1]);
1578                 }
1579
1580                 _exit(EXIT_SUCCESS);
1581         }
1582
1583         for (k = 0; k < _META_MAX; k++) {
1584                 _cleanup_fclose_ FILE *f = NULL;
1585
1586                 fds[2*k+1] = safe_close(fds[2*k+1]);
1587
1588                 f = take_fdopen(&fds[2*k], "r");
1589                 if (!f) {
1590                         r = -errno;
1591                         goto finish;
1592                 }
1593
1594                 switch (k) {
1595
1596                 case META_HOSTNAME:
1597                         r = read_etc_hostname_stream(f, &hostname);
1598                         if (r < 0)
1599                                 log_debug_errno(r, "Failed to read /etc/hostname: %m");
1600
1601                         break;
1602
1603                 case META_MACHINE_ID: {
1604                         _cleanup_free_ char *line = NULL;
1605
1606                         r = read_line(f, LONG_LINE_MAX, &line);
1607                         if (r < 0)
1608                                 log_debug_errno(r, "Failed to read /etc/machine-id: %m");
1609                         else if (r == 33) {
1610                                 r = sd_id128_from_string(line, &machine_id);
1611                                 if (r < 0)
1612                                         log_debug_errno(r, "Image contains invalid /etc/machine-id: %s", line);
1613                         } else if (r == 0)
1614                                 log_debug("/etc/machine-id file is empty.");
1615                         else
1616                                 log_debug("/etc/machine-id has unexpected length %i.", r);
1617
1618                         break;
1619                 }
1620
1621                 case META_MACHINE_INFO:
1622                         r = load_env_file_pairs(f, "machine-info", &machine_info);
1623                         if (r < 0)
1624                                 log_debug_errno(r, "Failed to read /etc/machine-info: %m");
1625
1626                         break;
1627
1628                 case META_OS_RELEASE:
1629                         r = load_env_file_pairs(f, "os-release", &os_release);
1630                         if (r < 0)
1631                                 log_debug_errno(r, "Failed to read OS release file: %m");
1632
1633                         break;
1634                 }
1635         }
1636
1637         r = wait_for_terminate_and_check("(sd-dissect)", child, 0);
1638         child = 0;
1639         if (r < 0)
1640                 goto finish;
1641         if (r != EXIT_SUCCESS)
1642                 return -EPROTO;
1643
1644         free_and_replace(m->hostname, hostname);
1645         m->machine_id = machine_id;
1646         strv_free_and_replace(m->machine_info, machine_info);
1647         strv_free_and_replace(m->os_release, os_release);
1648
1649 finish:
1650         for (k = 0; k < n_meta_initialized; k++)
1651                 safe_close_pair(fds + 2*k);
1652
1653         return r;
1654 }
1655
1656 int dissect_image_and_warn(
1657                 int fd,
1658                 const char *name,
1659                 const void *root_hash,
1660                 size_t root_hash_size,
1661                 const char *verity_data,
1662                 DissectImageFlags flags,
1663                 DissectedImage **ret) {
1664
1665         _cleanup_free_ char *buffer = NULL;
1666         int r;
1667
1668         if (!name) {
1669                 r = fd_get_path(fd, &buffer);
1670                 if (r < 0)
1671                         return r;
1672
1673                 name = buffer;
1674         }
1675
1676         r = dissect_image(fd, root_hash, root_hash_size, verity_data, flags, ret);
1677
1678         switch (r) {
1679
1680         case -EOPNOTSUPP:
1681                 return log_error_errno(r, "Dissecting images is not supported, compiled without blkid support.");
1682
1683         case -ENOPKG:
1684                 return log_error_errno(r, "Couldn't identify a suitable partition table or file system in '%s'.", name);
1685
1686         case -EADDRNOTAVAIL:
1687                 return log_error_errno(r, "No root partition for specified root hash found in '%s'.", name);
1688
1689         case -ENOTUNIQ:
1690                 return log_error_errno(r, "Multiple suitable root partitions found in image '%s'.", name);
1691
1692         case -ENXIO:
1693                 return log_error_errno(r, "No suitable root partition found in image '%s'.", name);
1694
1695         case -EPROTONOSUPPORT:
1696                 return log_error_errno(r, "Device '%s' is loopback block device with partition scanning turned off, please turn it on.", name);
1697
1698         default:
1699                 if (r < 0)
1700                         return log_error_errno(r, "Failed to dissect image '%s': %m", name);
1701
1702                 return r;
1703         }
1704 }
1705
1706 bool dissected_image_can_do_verity(const DissectedImage *image, unsigned partition_designator) {
1707         if (image->single_file_system)
1708                 return partition_designator == PARTITION_ROOT && image->can_verity;
1709
1710         return PARTITION_VERITY_OF(partition_designator) >= 0;
1711 }
1712
1713 bool dissected_image_has_verity(const DissectedImage *image, unsigned partition_designator) {
1714         int k;
1715
1716         if (image->single_file_system)
1717                 return partition_designator == PARTITION_ROOT && image->verity;
1718
1719         k = PARTITION_VERITY_OF(partition_designator);
1720         return k >= 0 && image->partitions[k].found;
1721 }
1722
1723 static const char *const partition_designator_table[] = {
1724         [PARTITION_ROOT] = "root",
1725         [PARTITION_ROOT_SECONDARY] = "root-secondary",
1726         [PARTITION_HOME] = "home",
1727         [PARTITION_SRV] = "srv",
1728         [PARTITION_ESP] = "esp",
1729         [PARTITION_XBOOTLDR] = "xbootldr",
1730         [PARTITION_SWAP] = "swap",
1731         [PARTITION_ROOT_VERITY] = "root-verity",
1732         [PARTITION_ROOT_SECONDARY_VERITY] = "root-secondary-verity",
1733         [PARTITION_TMP] = "tmp",
1734         [PARTITION_VAR] = "var",
1735 };
1736
1737 DEFINE_STRING_TABLE_LOOKUP(partition_designator, int);