src/shared/dissect-image.c

   1 /***
   2   This file is part of systemd.
   3
   4   Copyright 2016 Lennart Poettering
   5
   6   systemd is free software; you can redistribute it and/or modify it
   7   under the terms of the GNU Lesser General Public License as published by
   8   the Free Software Foundation; either version 2.1 of the License, or
   9   (at your option) any later version.
  10
  11   systemd is distributed in the hope that it will be useful, but
  12   WITHOUT ANY WARRANTY; without even the implied warranty of
  13   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14   Lesser General Public License for more details.
  15
  16   You should have received a copy of the GNU Lesser General Public License
  17   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  18 ***/
  19
  20 #if HAVE_LIBCRYPTSETUP
  21 #include <libcryptsetup.h>
  22 #endif
  23 #include <sys/mount.h>
  24
  25 #include "architecture.h"
  26 #include "ask-password-api.h"
  27 #include "blkid-util.h"
  28 #include "dissect-image.h"
  29 #include "fd-util.h"
  30 #include "fileio.h"
  31 #include "fs-util.h"
  32 #include "gpt.h"
  33 #include "hexdecoct.h"
  34 #include "linux-3.13/dm-ioctl.h"
  35 #include "mount-util.h"
  36 #include "path-util.h"
  37 #include "stat-util.h"
  38 #include "stdio-util.h"
  39 #include "string-table.h"
  40 #include "string-util.h"
  41 #include "strv.h"
  42 #include "udev-util.h"
  43 #include "xattr-util.h"
  44
  45 _unused_ static int probe_filesystem(const char *node, char **ret_fstype) {
  46 #if HAVE_BLKID
  47         _cleanup_blkid_free_probe_ blkid_probe b = NULL;
  48         const char *fstype;
  49         int r;
  50
  51         b = blkid_new_probe_from_filename(node);
  52         if (!b)
  53                 return -ENOMEM;
  54
  55         blkid_probe_enable_superblocks(b, 1);
  56         blkid_probe_set_superblocks_flags(b, BLKID_SUBLKS_TYPE);
  57
  58         errno = 0;
  59         r = blkid_do_safeprobe(b);
  60         if (IN_SET(r, -2, 1)) {
  61                 log_debug("Failed to identify any partition type on partition %s", node);
  62                 goto not_found;
  63         }
  64         if (r != 0)
  65                 return -errno ?: -EIO;
  66
  67         (void) blkid_probe_lookup_value(b, "TYPE", &fstype, NULL);
  68
  69         if (fstype) {
  70                 char *t;
  71
  72                 t = strdup(fstype);
  73                 if (!t)
  74                         return -ENOMEM;
  75
  76                 *ret_fstype = t;
  77                 return 1;
  78         }
  79
  80 not_found:
  81         *ret_fstype = NULL;
  82         return 0;
  83 #else
  84         return -EOPNOTSUPP;
  85 #endif
  86 }
  87
  88 int dissect_image(int fd, const void *root_hash, size_t root_hash_size, DissectImageFlags flags, DissectedImage **ret) {
  89
  90 #if HAVE_BLKID
  91         sd_id128_t root_uuid = SD_ID128_NULL, verity_uuid = SD_ID128_NULL;
  92         _cleanup_udev_enumerate_unref_ struct udev_enumerate *e = NULL;
  93         bool is_gpt, is_mbr, generic_rw, multiple_generic = false;
  94         _cleanup_udev_device_unref_ struct udev_device *d = NULL;
  95         _cleanup_(dissected_image_unrefp) DissectedImage *m = NULL;
  96         _cleanup_blkid_free_probe_ blkid_probe b = NULL;
  97         _cleanup_udev_unref_ struct udev *udev = NULL;
  98         _cleanup_free_ char *generic_node = NULL;
  99         sd_id128_t generic_uuid = SD_ID128_NULL;
 100         const char *pttype = NULL;
 101         struct udev_list_entry *first, *item;
 102         blkid_partlist pl;
 103         int r, generic_nr;
 104         struct stat st;
 105         unsigned i;
 106
 107         assert(fd >= 0);
 108         assert(ret);
 109         assert(root_hash || root_hash_size == 0);
 110
 111         /* Probes a disk image, and returns information about what it found in *ret.
 112          *
 113          * Returns -ENOPKG if no suitable partition table or file system could be found.
 114          * Returns -EADDRNOTAVAIL if a root hash was specified but no matching root/verity partitions found. */
 115
 116         if (root_hash) {
 117                 /* If a root hash is supplied, then we use the root partition that has a UUID that match the first
 118                  * 128bit of the root hash. And we use the verity partition that has a UUID that match the final
 119                  * 128bit. */
 120
 121                 if (root_hash_size < sizeof(sd_id128_t))
 122                         return -EINVAL;
 123
 124                 memcpy(&root_uuid, root_hash, sizeof(sd_id128_t));
 125                 memcpy(&verity_uuid, (const uint8_t*) root_hash + root_hash_size - sizeof(sd_id128_t), sizeof(sd_id128_t));
 126
 127                 if (sd_id128_is_null(root_uuid))
 128                         return -EINVAL;
 129                 if (sd_id128_is_null(verity_uuid))
 130                         return -EINVAL;
 131         }
 132
 133         if (fstat(fd, &st) < 0)
 134                 return -errno;
 135
 136         if (!S_ISBLK(st.st_mode))
 137                 return -ENOTBLK;
 138
 139         b = blkid_new_probe();
 140         if (!b)
 141                 return -ENOMEM;
 142
 143         errno = 0;
 144         r = blkid_probe_set_device(b, fd, 0, 0);
 145         if (r != 0)
 146                 return -errno ?: -ENOMEM;
 147
 148         if ((flags & DISSECT_IMAGE_GPT_ONLY) == 0) {
 149                 /* Look for file system superblocks, unless we only shall look for GPT partition tables */
 150                 blkid_probe_enable_superblocks(b, 1);
 151                 blkid_probe_set_superblocks_flags(b, BLKID_SUBLKS_TYPE|BLKID_SUBLKS_USAGE);
 152         }
 153
 154         blkid_probe_enable_partitions(b, 1);
 155         blkid_probe_set_partitions_flags(b, BLKID_PARTS_ENTRY_DETAILS);
 156
 157         errno = 0;
 158         r = blkid_do_safeprobe(b);
 159         if (IN_SET(r, -2, 1)) {
 160                 log_debug("Failed to identify any partition table.");
 161                 return -ENOPKG;
 162         }
 163         if (r != 0)
 164                 return -errno ?: -EIO;
 165
 166         m = new0(DissectedImage, 1);
 167         if (!m)
 168                 return -ENOMEM;
 169
 170         if (!(flags & DISSECT_IMAGE_GPT_ONLY) &&
 171             (flags & DISSECT_IMAGE_REQUIRE_ROOT)) {
 172                 const char *usage = NULL;
 173
 174                 (void) blkid_probe_lookup_value(b, "USAGE", &usage, NULL);
 175                 if (STRPTR_IN_SET(usage, "filesystem", "crypto")) {
 176                         _cleanup_free_ char *t = NULL, *n = NULL;
 177                         const char *fstype = NULL;
 178
 179                         /* OK, we have found a file system, that's our root partition then. */
 180                         (void) blkid_probe_lookup_value(b, "TYPE", &fstype, NULL);
 181
 182                         if (fstype) {
 183                                 t = strdup(fstype);
 184                                 if (!t)
 185                                         return -ENOMEM;
 186                         }
 187
 188                         if (asprintf(&n, "/dev/block/%u:%u", major(st.st_rdev), minor(st.st_rdev)) < 0)
 189                                 return -ENOMEM;
 190
 191                         m->partitions[PARTITION_ROOT] = (DissectedPartition) {
 192                                 .found = true,
 193                                 .rw = true,
 194                                 .partno = -1,
 195                                 .architecture = _ARCHITECTURE_INVALID,
 196                                 .fstype = t,
 197                                 .node = n,
 198                         };
 199
 200                         t = n = NULL;
 201
 202                         m->encrypted = streq(fstype, "crypto_LUKS");
 203
 204                         *ret = m;
 205                         m = NULL;
 206
 207                         return 0;
 208                 }
 209         }
 210
 211         (void) blkid_probe_lookup_value(b, "PTTYPE", &pttype, NULL);
 212         if (!pttype)
 213                 return -ENOPKG;
 214
 215         is_gpt = streq_ptr(pttype, "gpt");
 216         is_mbr = streq_ptr(pttype, "dos");
 217
 218         if (!is_gpt && ((flags & DISSECT_IMAGE_GPT_ONLY) || !is_mbr))
 219                 return -ENOPKG;
 220
 221         errno = 0;
 222         pl = blkid_probe_get_partitions(b);
 223         if (!pl)
 224                 return -errno ?: -ENOMEM;
 225
 226         udev = udev_new();
 227         if (!udev)
 228                 return -errno;
 229
 230         d = udev_device_new_from_devnum(udev, 'b', st.st_rdev);
 231         if (!d)
 232                 return -ENOMEM;
 233
 234         for (i = 0;; i++) {
 235                 int n, z;
 236
 237                 if (i >= 10) {
 238                         log_debug("Kernel partitions never appeared.");
 239                         return -ENXIO;
 240                 }
 241
 242                 e = udev_enumerate_new(udev);
 243                 if (!e)
 244                         return -errno;
 245
 246                 r = udev_enumerate_add_match_parent(e, d);
 247                 if (r < 0)
 248                         return r;
 249
 250                 r = udev_enumerate_scan_devices(e);
 251                 if (r < 0)
 252                         return r;
 253
 254                 /* Count the partitions enumerated by the kernel */
 255                 n = 0;
 256                 first = udev_enumerate_get_list_entry(e);
 257                 udev_list_entry_foreach(item, first)
 258                         n++;
 259
 260                 /* Count the partitions enumerated by blkid */
 261                 z = blkid_partlist_numof_partitions(pl);
 262                 if (n == z + 1)
 263                         break;
 264                 if (n > z + 1) {
 265                         log_debug("blkid and kernel partition list do not match.");
 266                         return -EIO;
 267                 }
 268                 if (n < z + 1) {
 269                         unsigned j = 0;
 270
 271                         /* The kernel has probed fewer partitions than blkid? Maybe the kernel prober is still running
 272                          * or it got EBUSY because udev already opened the device. Let's reprobe the device, which is a
 273                          * synchronous call that waits until probing is complete. */
 274
 275                         for (;;) {
 276                                 if (j++ > 20)
 277                                         return -EBUSY;
 278
 279                                 if (ioctl(fd, BLKRRPART, 0) < 0) {
 280                                         r = -errno;
 281
 282                                         if (r == -EINVAL) {
 283                                                 struct loop_info64 info;
 284
 285                                                 /* If we are running on a loop device that has partition scanning off,
 286                                                  * return an explicit recognizable error about this, so that callers
 287                                                  * can generate a proper message explaining the situation. */
 288
 289                                                 if (ioctl(fd, LOOP_GET_STATUS64, &info) >= 0 && (info.lo_flags & LO_FLAGS_PARTSCAN) == 0) {
 290                                                         log_debug("Device is loop device and partition scanning is off!");
 291                                                         return -EPROTONOSUPPORT;
 292                                                 }
 293                                         }
 294                                         if (r != -EBUSY)
 295                                                 return r;
 296                                 } else
 297                                         break;
 298
 299                                 /* If something else has the device open, such as an udev rule, the ioctl will return
 300                                  * EBUSY. Since there's no way to wait until it isn't busy anymore, let's just wait a
 301                                  * bit, and try again.
 302                                  *
 303                                  * This is really something they should fix in the kernel! */
 304
 305                                 (void) usleep(50 * USEC_PER_MSEC);
 306                         }
 307                 }
 308
 309                 e = udev_enumerate_unref(e);
 310         }
 311
 312         first = udev_enumerate_get_list_entry(e);
 313         udev_list_entry_foreach(item, first) {
 314                 _cleanup_udev_device_unref_ struct udev_device *q;
 315                 unsigned long long pflags;
 316                 blkid_partition pp;
 317                 const char *node, *sysname;
 318                 dev_t qn;
 319                 int nr;
 320
 321                 q = udev_device_new_from_syspath(udev, udev_list_entry_get_name(item));
 322                 if (!q)
 323                         return -errno;
 324
 325                 qn = udev_device_get_devnum(q);
 326                 if (major(qn) == 0)
 327                         continue;
 328
 329                 if (st.st_rdev == qn)
 330                         continue;
 331
 332                 /* Filter out weird MMC RPMB partitions, which cannot reasonably be read, see
 333                  * https://github.com/systemd/systemd/issues/5806 */
 334                 sysname = udev_device_get_sysname(q);
 335                 if (sysname && startswith(sysname, "mmcblk") && endswith(sysname, "rpmb"))
 336                         continue;
 337
 338                 node = udev_device_get_devnode(q);
 339                 if (!node)
 340                         continue;
 341
 342                 pp = blkid_partlist_devno_to_partition(pl, qn);
 343                 if (!pp)
 344                         continue;
 345
 346                 pflags = blkid_partition_get_flags(pp);
 347
 348                 nr = blkid_partition_get_partno(pp);
 349                 if (nr < 0)
 350                         continue;
 351
 352                 if (is_gpt) {
 353                         int designator = _PARTITION_DESIGNATOR_INVALID, architecture = _ARCHITECTURE_INVALID;
 354                         const char *stype, *sid, *fstype = NULL;
 355                         sd_id128_t type_id, id;
 356                         bool rw = true;
 357
 358                         sid = blkid_partition_get_uuid(pp);
 359                         if (!sid)
 360                                 continue;
 361                         if (sd_id128_from_string(sid, &id) < 0)
 362                                 continue;
 363
 364                         stype = blkid_partition_get_type_string(pp);
 365                         if (!stype)
 366                                 continue;
 367                         if (sd_id128_from_string(stype, &type_id) < 0)
 368                                 continue;
 369
 370                         if (sd_id128_equal(type_id, GPT_HOME)) {
 371
 372                                 if (pflags & GPT_FLAG_NO_AUTO)
 373                                         continue;
 374
 375                                 designator = PARTITION_HOME;
 376                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 377                         } else if (sd_id128_equal(type_id, GPT_SRV)) {
 378
 379                                 if (pflags & GPT_FLAG_NO_AUTO)
 380                                         continue;
 381
 382                                 designator = PARTITION_SRV;
 383                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 384                         } else if (sd_id128_equal(type_id, GPT_ESP)) {
 385
 386                                 /* Note that we don't check the GPT_FLAG_NO_AUTO flag for the ESP, as it is not defined
 387                                  * there. We instead check the GPT_FLAG_NO_BLOCK_IO_PROTOCOL, as recommended by the
 388                                  * UEFI spec (See "12.3.3 Number and Location of System Partitions"). */
 389
 390                                 if (pflags & GPT_FLAG_NO_BLOCK_IO_PROTOCOL)
 391                                         continue;
 392
 393                                 designator = PARTITION_ESP;
 394                                 fstype = "vfat";
 395                         }
 396 #ifdef GPT_ROOT_NATIVE
 397                         else if (sd_id128_equal(type_id, GPT_ROOT_NATIVE)) {
 398
 399                                 if (pflags & GPT_FLAG_NO_AUTO)
 400                                         continue;
 401
 402                                 /* If a root ID is specified, ignore everything but the root id */
 403                                 if (!sd_id128_is_null(root_uuid) && !sd_id128_equal(root_uuid, id))
 404                                         continue;
 405
 406                                 designator = PARTITION_ROOT;
 407                                 architecture = native_architecture();
 408                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 409                         } else if (sd_id128_equal(type_id, GPT_ROOT_NATIVE_VERITY)) {
 410
 411                                 if (pflags & GPT_FLAG_NO_AUTO)
 412                                         continue;
 413
 414                                 m->can_verity = true;
 415
 416                                 /* Ignore verity unless a root hash is specified */
 417                                 if (sd_id128_is_null(verity_uuid) || !sd_id128_equal(verity_uuid, id))
 418                                         continue;
 419
 420                                 designator = PARTITION_ROOT_VERITY;
 421                                 fstype = "DM_verity_hash";
 422                                 architecture = native_architecture();
 423                                 rw = false;
 424                         }
 425 #endif
 426 #ifdef GPT_ROOT_SECONDARY
 427                         else if (sd_id128_equal(type_id, GPT_ROOT_SECONDARY)) {
 428
 429                                 if (pflags & GPT_FLAG_NO_AUTO)
 430                                         continue;
 431
 432                                 /* If a root ID is specified, ignore everything but the root id */
 433                                 if (!sd_id128_is_null(root_uuid) && !sd_id128_equal(root_uuid, id))
 434                                         continue;
 435
 436                                 designator = PARTITION_ROOT_SECONDARY;
 437                                 architecture = SECONDARY_ARCHITECTURE;
 438                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 439                         } else if (sd_id128_equal(type_id, GPT_ROOT_SECONDARY_VERITY)) {
 440
 441                                 if (pflags & GPT_FLAG_NO_AUTO)
 442                                         continue;
 443
 444                                 m->can_verity = true;
 445
 446                                 /* Ignore verity unless root has is specified */
 447                                 if (sd_id128_is_null(verity_uuid) || !sd_id128_equal(verity_uuid, id))
 448                                         continue;
 449
 450                                 designator = PARTITION_ROOT_SECONDARY_VERITY;
 451                                 fstype = "DM_verity_hash";
 452                                 architecture = SECONDARY_ARCHITECTURE;
 453                                 rw = false;
 454                         }
 455 #endif
 456                         else if (sd_id128_equal(type_id, GPT_SWAP)) {
 457
 458                                 if (pflags & GPT_FLAG_NO_AUTO)
 459                                         continue;
 460
 461                                 designator = PARTITION_SWAP;
 462                                 fstype = "swap";
 463                         } else if (sd_id128_equal(type_id, GPT_LINUX_GENERIC)) {
 464
 465                                 if (pflags & GPT_FLAG_NO_AUTO)
 466                                         continue;
 467
 468                                 if (generic_node)
 469                                         multiple_generic = true;
 470                                 else {
 471                                         generic_nr = nr;
 472                                         generic_rw = !(pflags & GPT_FLAG_READ_ONLY);
 473                                         generic_uuid = id;
 474                                         generic_node = strdup(node);
 475                                         if (!generic_node)
 476                                                 return -ENOMEM;
 477                                 }
 478                         }
 479
 480                         if (designator != _PARTITION_DESIGNATOR_INVALID) {
 481                                 _cleanup_free_ char *t = NULL, *n = NULL;
 482
 483                                 /* First one wins */
 484                                 if (m->partitions[designator].found)
 485                                         continue;
 486
 487                                 if (fstype) {
 488                                         t = strdup(fstype);
 489                                         if (!t)
 490                                                 return -ENOMEM;
 491                                 }
 492
 493                                 n = strdup(node);
 494                                 if (!n)
 495                                         return -ENOMEM;
 496
 497                                 m->partitions[designator] = (DissectedPartition) {
 498                                         .found = true,
 499                                         .partno = nr,
 500                                         .rw = rw,
 501                                         .architecture = architecture,
 502                                         .node = n,
 503                                         .fstype = t,
 504                                         .uuid = id,
 505                                 };
 506
 507                                 n = t = NULL;
 508                         }
 509
 510                 } else if (is_mbr) {
 511
 512                         if (pflags != 0x80) /* Bootable flag */
 513                                 continue;
 514
 515                         if (blkid_partition_get_type(pp) != 0x83) /* Linux partition */
 516                                 continue;
 517
 518                         if (generic_node)
 519                                 multiple_generic = true;
 520                         else {
 521                                 generic_nr = nr;
 522                                 generic_rw = true;
 523                                 generic_node = strdup(node);
 524                                 if (!generic_node)
 525                                         return -ENOMEM;
 526                         }
 527                 }
 528         }
 529
 530         if (!m->partitions[PARTITION_ROOT].found) {
 531                 /* No root partition found? Then let's see if ther's one for the secondary architecture. And if not
 532                  * either, then check if there's a single generic one, and use that. */
 533
 534                 if (m->partitions[PARTITION_ROOT_VERITY].found)
 535                         return -EADDRNOTAVAIL;
 536
 537                 if (m->partitions[PARTITION_ROOT_SECONDARY].found) {
 538                         m->partitions[PARTITION_ROOT] = m->partitions[PARTITION_ROOT_SECONDARY];
 539                         zero(m->partitions[PARTITION_ROOT_SECONDARY]);
 540
 541                         m->partitions[PARTITION_ROOT_VERITY] = m->partitions[PARTITION_ROOT_SECONDARY_VERITY];
 542                         zero(m->partitions[PARTITION_ROOT_SECONDARY_VERITY]);
 543
 544                 } else if (flags & DISSECT_IMAGE_REQUIRE_ROOT) {
 545
 546                         /* If the root has was set, then we won't fallback to a generic node, because the root hash
 547                          * decides */
 548                         if (root_hash)
 549                                 return -EADDRNOTAVAIL;
 550
 551                         /* If we didn't find a generic node, then we can't fix this up either */
 552                         if (!generic_node)
 553                                 return -ENXIO;
 554
 555                         /* If we didn't find a properly marked root partition, but we did find a single suitable
 556                          * generic Linux partition, then use this as root partition, if the caller asked for it. */
 557                         if (multiple_generic)
 558                                 return -ENOTUNIQ;
 559
 560                         m->partitions[PARTITION_ROOT] = (DissectedPartition) {
 561                                 .found = true,
 562                                 .rw = generic_rw,
 563                                 .partno = generic_nr,
 564                                 .architecture = _ARCHITECTURE_INVALID,
 565                                 .node = generic_node,
 566                                 .uuid = generic_uuid,
 567                         };
 568
 569                         generic_node = NULL;
 570                 }
 571         }
 572
 573         if (root_hash) {
 574                 if (!m->partitions[PARTITION_ROOT_VERITY].found || !m->partitions[PARTITION_ROOT].found)
 575                         return -EADDRNOTAVAIL;
 576
 577                 /* If we found the primary root with the hash, then we definitely want to suppress any secondary root
 578                  * (which would be weird, after all the root hash should only be assigned to one pair of
 579                  * partitions... */
 580                 m->partitions[PARTITION_ROOT_SECONDARY].found = false;
 581                 m->partitions[PARTITION_ROOT_SECONDARY_VERITY].found = false;
 582
 583                 /* If we found a verity setup, then the root partition is necessarily read-only. */
 584                 m->partitions[PARTITION_ROOT].rw = false;
 585
 586                 m->verity = true;
 587         }
 588
 589         blkid_free_probe(b);
 590         b = NULL;
 591
 592         /* Fill in file system types if we don't know them yet. */
 593         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
 594                 DissectedPartition *p = m->partitions + i;
 595
 596                 if (!p->found)
 597                         continue;
 598
 599                 if (!p->fstype && p->node) {
 600                         r = probe_filesystem(p->node, &p->fstype);
 601                         if (r < 0)
 602                                 return r;
 603                 }
 604
 605                 if (streq_ptr(p->fstype, "crypto_LUKS"))
 606                         m->encrypted = true;
 607
 608                 if (p->fstype && fstype_is_ro(p->fstype))
 609                         p->rw = false;
 610         }
 611
 612         *ret = m;
 613         m = NULL;
 614
 615         return 0;
 616 #else
 617         return -EOPNOTSUPP;
 618 #endif
 619 }
 620
 621 DissectedImage* dissected_image_unref(DissectedImage *m) {
 622         unsigned i;
 623
 624         if (!m)
 625                 return NULL;
 626
 627         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
 628                 free(m->partitions[i].fstype);
 629                 free(m->partitions[i].node);
 630                 free(m->partitions[i].decrypted_fstype);
 631                 free(m->partitions[i].decrypted_node);
 632         }
 633
 634         free(m);
 635         return NULL;
 636 }
 637
 638 static int is_loop_device(const char *path) {
 639         char s[strlen("/sys/dev/block/") + DECIMAL_STR_MAX(dev_t) + 1 + DECIMAL_STR_MAX(dev_t) + strlen("/../loop/")];
 640         struct stat st;
 641
 642         assert(path);
 643
 644         if (stat(path, &st) < 0)
 645                 return -errno;
 646
 647         if (!S_ISBLK(st.st_mode))
 648                 return -ENOTBLK;
 649
 650         xsprintf(s, "/sys/dev/block/%u:%u/loop/", major(st.st_rdev), minor(st.st_rdev));
 651         if (access(s, F_OK) < 0) {
 652                 if (errno != ENOENT)
 653                         return -errno;
 654
 655                 /* The device itself isn't a loop device, but maybe it's a partition and its parent is? */
 656                 xsprintf(s, "/sys/dev/block/%u:%u/../loop/", major(st.st_rdev), minor(st.st_rdev));
 657                 if (access(s, F_OK) < 0)
 658                         return errno == ENOENT ? false : -errno;
 659         }
 660
 661         return true;
 662 }
 663
 664 static int mount_partition(
 665                 DissectedPartition *m,
 666                 const char *where,
 667                 const char *directory,
 668                 DissectImageFlags flags) {
 669
 670         const char *p, *options = NULL, *node, *fstype;
 671         _cleanup_free_ char *chased = NULL;
 672         bool rw;
 673         int r;
 674
 675         assert(m);
 676         assert(where);
 677
 678         node = m->decrypted_node ?: m->node;
 679         fstype = m->decrypted_fstype ?: m->fstype;
 680
 681         if (!m->found || !node || !fstype)
 682                 return 0;
 683
 684         /* Stacked encryption? Yuck */
 685         if (streq_ptr(fstype, "crypto_LUKS"))
 686                 return -ELOOP;
 687
 688         rw = m->rw && !(flags & DISSECT_IMAGE_READ_ONLY);
 689
 690         if (directory) {
 691                 r = chase_symlinks(directory, where, CHASE_PREFIX_ROOT, &chased);
 692                 if (r < 0)
 693                         return r;
 694
 695                 p = chased;
 696         } else
 697                 p = where;
 698
 699         /* If requested, turn on discard support. */
 700         if (fstype_can_discard(fstype) &&
 701             ((flags & DISSECT_IMAGE_DISCARD) ||
 702              ((flags & DISSECT_IMAGE_DISCARD_ON_LOOP) && is_loop_device(m->node))))
 703                 options = "discard";
 704
 705         return mount_verbose(LOG_DEBUG, node, p, fstype, MS_NODEV|(rw ? 0 : MS_RDONLY), options);
 706 }
 707
 708 int dissected_image_mount(DissectedImage *m, const char *where, DissectImageFlags flags) {
 709         int r;
 710
 711         assert(m);
 712         assert(where);
 713
 714         if (!m->partitions[PARTITION_ROOT].found)
 715                 return -ENXIO;
 716
 717         r = mount_partition(m->partitions + PARTITION_ROOT, where, NULL, flags);
 718         if (r < 0)
 719                 return r;
 720
 721         r = mount_partition(m->partitions + PARTITION_HOME, where, "/home", flags);
 722         if (r < 0)
 723                 return r;
 724
 725         r = mount_partition(m->partitions + PARTITION_SRV, where, "/srv", flags);
 726         if (r < 0)
 727                 return r;
 728
 729         if (m->partitions[PARTITION_ESP].found) {
 730                 const char *mp;
 731
 732                 /* Mount the ESP to /efi if it exists and is empty. If it doesn't exist, use /boot instead. */
 733
 734                 FOREACH_STRING(mp, "/efi", "/boot") {
 735                         _cleanup_free_ char *p = NULL;
 736
 737                         r = chase_symlinks(mp, where, CHASE_PREFIX_ROOT, &p);
 738                         if (r < 0)
 739                                 continue;
 740
 741                         r = dir_is_empty(p);
 742                         if (r > 0) {
 743                                 r = mount_partition(m->partitions + PARTITION_ESP, where, mp, flags);
 744                                 if (r < 0)
 745                                         return r;
 746                         }
 747                 }
 748         }
 749
 750         return 0;
 751 }
 752
 753 #if HAVE_LIBCRYPTSETUP
 754 typedef struct DecryptedPartition {
 755         struct crypt_device *device;
 756         char *name;
 757         bool relinquished;
 758 } DecryptedPartition;
 759
 760 struct DecryptedImage {
 761         DecryptedPartition *decrypted;
 762         size_t n_decrypted;
 763         size_t n_allocated;
 764 };
 765 #endif
 766
 767 DecryptedImage* decrypted_image_unref(DecryptedImage* d) {
 768 #if HAVE_LIBCRYPTSETUP
 769         size_t i;
 770         int r;
 771
 772         if (!d)
 773                 return NULL;
 774
 775         for (i = 0; i < d->n_decrypted; i++) {
 776                 DecryptedPartition *p = d->decrypted + i;
 777
 778                 if (p->device && p->name && !p->relinquished) {
 779                         r = crypt_deactivate(p->device, p->name);
 780                         if (r < 0)
 781                                 log_debug_errno(r, "Failed to deactivate encrypted partition %s", p->name);
 782                 }
 783
 784                 if (p->device)
 785                         crypt_free(p->device);
 786                 free(p->name);
 787         }
 788
 789         free(d);
 790 #endif
 791         return NULL;
 792 }
 793
 794 #if HAVE_LIBCRYPTSETUP
 795
 796 static int make_dm_name_and_node(const void *original_node, const char *suffix, char **ret_name, char **ret_node) {
 797         _cleanup_free_ char *name = NULL, *node = NULL;
 798         const char *base;
 799
 800         assert(original_node);
 801         assert(suffix);
 802         assert(ret_name);
 803         assert(ret_node);
 804
 805         base = strrchr(original_node, '/');
 806         if (!base)
 807                 return -EINVAL;
 808         base++;
 809         if (isempty(base))
 810                 return -EINVAL;
 811
 812         name = strjoin(base, suffix);
 813         if (!name)
 814                 return -ENOMEM;
 815         if (!filename_is_valid(name))
 816                 return -EINVAL;
 817
 818         node = strjoin(crypt_get_dir(), "/", name);
 819         if (!node)
 820                 return -ENOMEM;
 821
 822         *ret_name = name;
 823         *ret_node = node;
 824
 825         name = node = NULL;
 826         return 0;
 827 }
 828
 829 static int decrypt_partition(
 830                 DissectedPartition *m,
 831                 const char *passphrase,
 832                 DissectImageFlags flags,
 833                 DecryptedImage *d) {
 834
 835         _cleanup_free_ char *node = NULL, *name = NULL;
 836         struct crypt_device *cd;
 837         int r;
 838
 839         assert(m);
 840         assert(d);
 841
 842         if (!m->found || !m->node || !m->fstype)
 843                 return 0;
 844
 845         if (!streq(m->fstype, "crypto_LUKS"))
 846                 return 0;
 847
 848         r = make_dm_name_and_node(m->node, "-decrypted", &name, &node);
 849         if (r < 0)
 850                 return r;
 851
 852         if (!GREEDY_REALLOC0(d->decrypted, d->n_allocated, d->n_decrypted + 1))
 853                 return -ENOMEM;
 854
 855         r = crypt_init(&cd, m->node);
 856         if (r < 0)
 857                 return log_debug_errno(r, "Failed to initialize dm-crypt: %m");
 858
 859         r = crypt_load(cd, CRYPT_LUKS1, NULL);
 860         if (r < 0) {
 861                 log_debug_errno(r, "Failed to load LUKS metadata: %m");
 862                 goto fail;
 863         }
 864
 865         r = crypt_activate_by_passphrase(cd, name, CRYPT_ANY_SLOT, passphrase, strlen(passphrase),
 866                                          ((flags & DISSECT_IMAGE_READ_ONLY) ? CRYPT_ACTIVATE_READONLY : 0) |
 867                                          ((flags & DISSECT_IMAGE_DISCARD_ON_CRYPTO) ? CRYPT_ACTIVATE_ALLOW_DISCARDS : 0));
 868         if (r < 0)
 869                 log_debug_errno(r, "Failed to activate LUKS device: %m");
 870         if (r == -EPERM) {
 871                 r = -EKEYREJECTED;
 872                 goto fail;
 873         }
 874         if (r < 0)
 875                 goto fail;
 876
 877         d->decrypted[d->n_decrypted].name = name;
 878         name = NULL;
 879
 880         d->decrypted[d->n_decrypted].device = cd;
 881         d->n_decrypted++;
 882
 883         m->decrypted_node = node;
 884         node = NULL;
 885
 886         return 0;
 887
 888 fail:
 889         crypt_free(cd);
 890         return r;
 891 }
 892
 893 static int verity_partition(
 894                 DissectedPartition *m,
 895                 DissectedPartition *v,
 896                 const void *root_hash,
 897                 size_t root_hash_size,
 898                 DissectImageFlags flags,
 899                 DecryptedImage *d) {
 900
 901         _cleanup_free_ char *node = NULL, *name = NULL;
 902         struct crypt_device *cd;
 903         int r;
 904
 905         assert(m);
 906         assert(v);
 907
 908         if (!root_hash)
 909                 return 0;
 910
 911         if (!m->found || !m->node || !m->fstype)
 912                 return 0;
 913         if (!v->found || !v->node || !v->fstype)
 914                 return 0;
 915
 916         if (!streq(v->fstype, "DM_verity_hash"))
 917                 return 0;
 918
 919         r = make_dm_name_and_node(m->node, "-verity", &name, &node);
 920         if (r < 0)
 921                 return r;
 922
 923         if (!GREEDY_REALLOC0(d->decrypted, d->n_allocated, d->n_decrypted + 1))
 924                 return -ENOMEM;
 925
 926         r = crypt_init(&cd, v->node);
 927         if (r < 0)
 928                 return r;
 929
 930         r = crypt_load(cd, CRYPT_VERITY, NULL);
 931         if (r < 0)
 932                 goto fail;
 933
 934         r = crypt_set_data_device(cd, m->node);
 935         if (r < 0)
 936                 goto fail;
 937
 938         r = crypt_activate_by_volume_key(cd, name, root_hash, root_hash_size, CRYPT_ACTIVATE_READONLY);
 939         if (r < 0)
 940                 goto fail;
 941
 942         d->decrypted[d->n_decrypted].name = name;
 943         name = NULL;
 944
 945         d->decrypted[d->n_decrypted].device = cd;
 946         d->n_decrypted++;
 947
 948         m->decrypted_node = node;
 949         node = NULL;
 950
 951         return 0;
 952
 953 fail:
 954         crypt_free(cd);
 955         return r;
 956 }
 957 #endif
 958
 959 int dissected_image_decrypt(
 960                 DissectedImage *m,
 961                 const char *passphrase,
 962                 const void *root_hash,
 963                 size_t root_hash_size,
 964                 DissectImageFlags flags,
 965                 DecryptedImage **ret) {
 966
 967         _cleanup_(decrypted_image_unrefp) DecryptedImage *d = NULL;
 968 #if HAVE_LIBCRYPTSETUP
 969         unsigned i;
 970         int r;
 971 #endif
 972
 973         assert(m);
 974         assert(root_hash || root_hash_size == 0);
 975
 976         /* Returns:
 977          *
 978          *      = 0           → There was nothing to decrypt
 979          *      > 0           → Decrypted successfully
 980          *      -ENOKEY       → There's something to decrypt but no key was supplied
 981          *      -EKEYREJECTED → Passed key was not correct
 982          */
 983
 984         if (root_hash && root_hash_size < sizeof(sd_id128_t))
 985                 return -EINVAL;
 986
 987         if (!m->encrypted && !m->verity) {
 988                 *ret = NULL;
 989                 return 0;
 990         }
 991
 992 #if HAVE_LIBCRYPTSETUP
 993         if (m->encrypted && !passphrase)
 994                 return -ENOKEY;
 995
 996         d = new0(DecryptedImage, 1);
 997         if (!d)
 998                 return -ENOMEM;
 999
1000         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
1001                 DissectedPartition *p = m->partitions + i;
1002                 int k;
1003
1004                 if (!p->found)
1005                         continue;
1006
1007                 r = decrypt_partition(p, passphrase, flags, d);
1008                 if (r < 0)
1009                         return r;
1010
1011                 k = PARTITION_VERITY_OF(i);
1012                 if (k >= 0) {
1013                         r = verity_partition(p, m->partitions + k, root_hash, root_hash_size, flags, d);
1014                         if (r < 0)
1015                                 return r;
1016                 }
1017
1018                 if (!p->decrypted_fstype && p->decrypted_node) {
1019                         r = probe_filesystem(p->decrypted_node, &p->decrypted_fstype);
1020                         if (r < 0)
1021                                 return r;
1022                 }
1023         }
1024
1025         *ret = d;
1026         d = NULL;
1027
1028         return 1;
1029 #else
1030         return -EOPNOTSUPP;
1031 #endif
1032 }
1033
1034 int dissected_image_decrypt_interactively(
1035                 DissectedImage *m,
1036                 const char *passphrase,
1037                 const void *root_hash,
1038                 size_t root_hash_size,
1039                 DissectImageFlags flags,
1040                 DecryptedImage **ret) {
1041
1042         _cleanup_strv_free_erase_ char **z = NULL;
1043         int n = 3, r;
1044
1045         if (passphrase)
1046                 n--;
1047
1048         for (;;) {
1049                 r = dissected_image_decrypt(m, passphrase, root_hash, root_hash_size, flags, ret);
1050                 if (r >= 0)
1051                         return r;
1052                 if (r == -EKEYREJECTED)
1053                         log_error_errno(r, "Incorrect passphrase, try again!");
1054                 else if (r != -ENOKEY) {
1055                         log_error_errno(r, "Failed to decrypt image: %m");
1056                         return r;
1057                 }
1058
1059                 if (--n < 0) {
1060                         log_error("Too many retries.");
1061                         return -EKEYREJECTED;
1062                 }
1063
1064                 z = strv_free(z);
1065
1066                 r = ask_password_auto("Please enter image passphrase!", NULL, "dissect", "dissect", USEC_INFINITY, 0, &z);
1067                 if (r < 0)
1068                         return log_error_errno(r, "Failed to query for passphrase: %m");
1069
1070                 passphrase = z[0];
1071         }
1072 }
1073
1074 #if HAVE_LIBCRYPTSETUP
1075 static int deferred_remove(DecryptedPartition *p) {
1076
1077         struct dm_ioctl dm = {
1078                 .version = {
1079                         DM_VERSION_MAJOR,
1080                         DM_VERSION_MINOR,
1081                         DM_VERSION_PATCHLEVEL
1082                 },
1083                 .data_size = sizeof(dm),
1084                 .flags = DM_DEFERRED_REMOVE,
1085         };
1086
1087         _cleanup_close_ int fd = -1;
1088
1089         assert(p);
1090
1091         /* Unfortunately, libcryptsetup doesn't provide a proper API for this, hence call the ioctl() directly. */
1092
1093         fd = open("/dev/mapper/control", O_RDWR|O_CLOEXEC);
1094         if (fd < 0)
1095                 return -errno;
1096
1097         strncpy(dm.name, p->name, sizeof(dm.name));
1098
1099         if (ioctl(fd, DM_DEV_REMOVE, &dm))
1100                 return -errno;
1101
1102         return 0;
1103 }
1104 #endif
1105
1106 int decrypted_image_relinquish(DecryptedImage *d) {
1107
1108 #if HAVE_LIBCRYPTSETUP
1109         size_t i;
1110         int r;
1111 #endif
1112
1113         assert(d);
1114
1115         /* Turns on automatic removal after the last use ended for all DM devices of this image, and sets a boolean so
1116          * that we don't clean it up ourselves either anymore */
1117
1118 #if HAVE_LIBCRYPTSETUP
1119         for (i = 0; i < d->n_decrypted; i++) {
1120                 DecryptedPartition *p = d->decrypted + i;
1121
1122                 if (p->relinquished)
1123                         continue;
1124
1125                 r = deferred_remove(p);
1126                 if (r < 0)
1127                         return log_debug_errno(r, "Failed to mark %s for auto-removal: %m", p->name);
1128
1129                 p->relinquished = true;
1130         }
1131 #endif
1132
1133         return 0;
1134 }
1135
1136 int root_hash_load(const char *image, void **ret, size_t *ret_size) {
1137         _cleanup_free_ char *text = NULL;
1138         _cleanup_free_ void *k = NULL;
1139         size_t l;
1140         int r;
1141
1142         assert(image);
1143         assert(ret);
1144         assert(ret_size);
1145
1146         if (is_device_path(image)) {
1147                 /* If we are asked to load the root hash for a device node, exit early */
1148                 *ret = NULL;
1149                 *ret_size = 0;
1150                 return 0;
1151         }
1152
1153         r = getxattr_malloc(image, "user.verity.roothash", &text, true);
1154         if (r < 0) {
1155                 char *fn, *e, *n;
1156
1157                 if (!IN_SET(r, -ENODATA, -EOPNOTSUPP, -ENOENT))
1158                         return r;
1159
1160                 fn = newa(char, strlen(image) + strlen(".roothash") + 1);
1161                 n = stpcpy(fn, image);
1162                 e = endswith(fn, ".raw");
1163                 if (e)
1164                         n = e;
1165
1166                 strcpy(n, ".roothash");
1167
1168                 r = read_one_line_file(fn, &text);
1169                 if (r == -ENOENT) {
1170                         *ret = NULL;
1171                         *ret_size = 0;
1172                         return 0;
1173                 }
1174                 if (r < 0)
1175                         return r;
1176         }
1177
1178         r = unhexmem(text, strlen(text), &k, &l);
1179         if (r < 0)
1180                 return r;
1181         if (l < sizeof(sd_id128_t))
1182                 return -EINVAL;
1183
1184         *ret = k;
1185         *ret_size = l;
1186
1187         k = NULL;
1188
1189         return 1;
1190 }
1191
1192 static const char *const partition_designator_table[] = {
1193         [PARTITION_ROOT] = "root",
1194         [PARTITION_ROOT_SECONDARY] = "root-secondary",
1195         [PARTITION_HOME] = "home",
1196         [PARTITION_SRV] = "srv",
1197         [PARTITION_ESP] = "esp",
1198         [PARTITION_SWAP] = "swap",
1199         [PARTITION_ROOT_VERITY] = "root-verity",
1200         [PARTITION_ROOT_SECONDARY_VERITY] = "root-secondary-verity",
1201 };
1202
1203 DEFINE_STRING_TABLE_LOOKUP(partition_designator, int);