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 #ifndef CRYPT_LUKS
  23 #define CRYPT_LUKS NULL
  24 #endif
  25 #endif
  26 #include <sys/mount.h>
  27
  28 #include "architecture.h"
  29 #include "ask-password-api.h"
  30 #include "blkid-util.h"
  31 #include "dissect-image.h"
  32 #include "fd-util.h"
  33 #include "fileio.h"
  34 #include "fs-util.h"
  35 #include "gpt.h"
  36 #include "hexdecoct.h"
  37 #include "linux-3.13/dm-ioctl.h"
  38 #include "mount-util.h"
  39 #include "path-util.h"
  40 #include "stat-util.h"
  41 #include "stdio-util.h"
  42 #include "string-table.h"
  43 #include "string-util.h"
  44 #include "strv.h"
  45 #include "udev-util.h"
  46 #include "xattr-util.h"
  47
  48 _unused_ static int probe_filesystem(const char *node, char **ret_fstype) {
  49 #if HAVE_BLKID
  50         _cleanup_blkid_free_probe_ blkid_probe b = NULL;
  51         const char *fstype;
  52         int r;
  53
  54         b = blkid_new_probe_from_filename(node);
  55         if (!b)
  56                 return -ENOMEM;
  57
  58         blkid_probe_enable_superblocks(b, 1);
  59         blkid_probe_set_superblocks_flags(b, BLKID_SUBLKS_TYPE);
  60
  61         errno = 0;
  62         r = blkid_do_safeprobe(b);
  63         if (IN_SET(r, -2, 1)) {
  64                 log_debug("Failed to identify any partition type on partition %s", node);
  65                 goto not_found;
  66         }
  67         if (r != 0)
  68                 return -errno ?: -EIO;
  69
  70         (void) blkid_probe_lookup_value(b, "TYPE", &fstype, NULL);
  71
  72         if (fstype) {
  73                 char *t;
  74
  75                 t = strdup(fstype);
  76                 if (!t)
  77                         return -ENOMEM;
  78
  79                 *ret_fstype = t;
  80                 return 1;
  81         }
  82
  83 not_found:
  84         *ret_fstype = NULL;
  85         return 0;
  86 #else
  87         return -EOPNOTSUPP;
  88 #endif
  89 }
  90
  91 int dissect_image(int fd, const void *root_hash, size_t root_hash_size, DissectImageFlags flags, DissectedImage **ret) {
  92
  93 #if HAVE_BLKID
  94         sd_id128_t root_uuid = SD_ID128_NULL, verity_uuid = SD_ID128_NULL;
  95         _cleanup_udev_enumerate_unref_ struct udev_enumerate *e = NULL;
  96         bool is_gpt, is_mbr, generic_rw, multiple_generic = false;
  97         _cleanup_udev_device_unref_ struct udev_device *d = NULL;
  98         _cleanup_(dissected_image_unrefp) DissectedImage *m = NULL;
  99         _cleanup_blkid_free_probe_ blkid_probe b = NULL;
 100         _cleanup_udev_unref_ struct udev *udev = NULL;
 101         _cleanup_free_ char *generic_node = NULL;
 102         sd_id128_t generic_uuid = SD_ID128_NULL;
 103         const char *pttype = NULL;
 104         struct udev_list_entry *first, *item;
 105         blkid_partlist pl;
 106         int r, generic_nr;
 107         struct stat st;
 108         unsigned i;
 109
 110         assert(fd >= 0);
 111         assert(ret);
 112         assert(root_hash || root_hash_size == 0);
 113
 114         /* Probes a disk image, and returns information about what it found in *ret.
 115          *
 116          * Returns -ENOPKG if no suitable partition table or file system could be found.
 117          * Returns -EADDRNOTAVAIL if a root hash was specified but no matching root/verity partitions found. */
 118
 119         if (root_hash) {
 120                 /* If a root hash is supplied, then we use the root partition that has a UUID that match the first
 121                  * 128bit of the root hash. And we use the verity partition that has a UUID that match the final
 122                  * 128bit. */
 123
 124                 if (root_hash_size < sizeof(sd_id128_t))
 125                         return -EINVAL;
 126
 127                 memcpy(&root_uuid, root_hash, sizeof(sd_id128_t));
 128                 memcpy(&verity_uuid, (const uint8_t*) root_hash + root_hash_size - sizeof(sd_id128_t), sizeof(sd_id128_t));
 129
 130                 if (sd_id128_is_null(root_uuid))
 131                         return -EINVAL;
 132                 if (sd_id128_is_null(verity_uuid))
 133                         return -EINVAL;
 134         }
 135
 136         if (fstat(fd, &st) < 0)
 137                 return -errno;
 138
 139         if (!S_ISBLK(st.st_mode))
 140                 return -ENOTBLK;
 141
 142         b = blkid_new_probe();
 143         if (!b)
 144                 return -ENOMEM;
 145
 146         errno = 0;
 147         r = blkid_probe_set_device(b, fd, 0, 0);
 148         if (r != 0)
 149                 return -errno ?: -ENOMEM;
 150
 151         if ((flags & DISSECT_IMAGE_GPT_ONLY) == 0) {
 152                 /* Look for file system superblocks, unless we only shall look for GPT partition tables */
 153                 blkid_probe_enable_superblocks(b, 1);
 154                 blkid_probe_set_superblocks_flags(b, BLKID_SUBLKS_TYPE|BLKID_SUBLKS_USAGE);
 155         }
 156
 157         blkid_probe_enable_partitions(b, 1);
 158         blkid_probe_set_partitions_flags(b, BLKID_PARTS_ENTRY_DETAILS);
 159
 160         errno = 0;
 161         r = blkid_do_safeprobe(b);
 162         if (IN_SET(r, -2, 1)) {
 163                 log_debug("Failed to identify any partition table.");
 164                 return -ENOPKG;
 165         }
 166         if (r != 0)
 167                 return -errno ?: -EIO;
 168
 169         m = new0(DissectedImage, 1);
 170         if (!m)
 171                 return -ENOMEM;
 172
 173         if (!(flags & DISSECT_IMAGE_GPT_ONLY) &&
 174             (flags & DISSECT_IMAGE_REQUIRE_ROOT)) {
 175                 const char *usage = NULL;
 176
 177                 (void) blkid_probe_lookup_value(b, "USAGE", &usage, NULL);
 178                 if (STRPTR_IN_SET(usage, "filesystem", "crypto")) {
 179                         _cleanup_free_ char *t = NULL, *n = NULL;
 180                         const char *fstype = NULL;
 181
 182                         /* OK, we have found a file system, that's our root partition then. */
 183                         (void) blkid_probe_lookup_value(b, "TYPE", &fstype, NULL);
 184
 185                         if (fstype) {
 186                                 t = strdup(fstype);
 187                                 if (!t)
 188                                         return -ENOMEM;
 189                         }
 190
 191                         if (asprintf(&n, "/dev/block/%u:%u", major(st.st_rdev), minor(st.st_rdev)) < 0)
 192                                 return -ENOMEM;
 193
 194                         m->partitions[PARTITION_ROOT] = (DissectedPartition) {
 195                                 .found = true,
 196                                 .rw = true,
 197                                 .partno = -1,
 198                                 .architecture = _ARCHITECTURE_INVALID,
 199                                 .fstype = t,
 200                                 .node = n,
 201                         };
 202
 203                         t = n = NULL;
 204
 205                         m->encrypted = streq(fstype, "crypto_LUKS");
 206
 207                         *ret = m;
 208                         m = NULL;
 209
 210                         return 0;
 211                 }
 212         }
 213
 214         (void) blkid_probe_lookup_value(b, "PTTYPE", &pttype, NULL);
 215         if (!pttype)
 216                 return -ENOPKG;
 217
 218         is_gpt = streq_ptr(pttype, "gpt");
 219         is_mbr = streq_ptr(pttype, "dos");
 220
 221         if (!is_gpt && ((flags & DISSECT_IMAGE_GPT_ONLY) || !is_mbr))
 222                 return -ENOPKG;
 223
 224         errno = 0;
 225         pl = blkid_probe_get_partitions(b);
 226         if (!pl)
 227                 return -errno ?: -ENOMEM;
 228
 229         udev = udev_new();
 230         if (!udev)
 231                 return -errno;
 232
 233         d = udev_device_new_from_devnum(udev, 'b', st.st_rdev);
 234         if (!d)
 235                 return -ENOMEM;
 236
 237         for (i = 0;; i++) {
 238                 int n, z;
 239
 240                 if (i >= 10) {
 241                         log_debug("Kernel partitions never appeared.");
 242                         return -ENXIO;
 243                 }
 244
 245                 e = udev_enumerate_new(udev);
 246                 if (!e)
 247                         return -errno;
 248
 249                 r = udev_enumerate_add_match_parent(e, d);
 250                 if (r < 0)
 251                         return r;
 252
 253                 r = udev_enumerate_scan_devices(e);
 254                 if (r < 0)
 255                         return r;
 256
 257                 /* Count the partitions enumerated by the kernel */
 258                 n = 0;
 259                 first = udev_enumerate_get_list_entry(e);
 260                 udev_list_entry_foreach(item, first)
 261                         n++;
 262
 263                 /* Count the partitions enumerated by blkid */
 264                 z = blkid_partlist_numof_partitions(pl);
 265                 if (n == z + 1)
 266                         break;
 267                 if (n > z + 1) {
 268                         log_debug("blkid and kernel partition list do not match.");
 269                         return -EIO;
 270                 }
 271                 if (n < z + 1) {
 272                         unsigned j = 0;
 273
 274                         /* The kernel has probed fewer partitions than blkid? Maybe the kernel prober is still running
 275                          * or it got EBUSY because udev already opened the device. Let's reprobe the device, which is a
 276                          * synchronous call that waits until probing is complete. */
 277
 278                         for (;;) {
 279                                 if (j++ > 20)
 280                                         return -EBUSY;
 281
 282                                 if (ioctl(fd, BLKRRPART, 0) < 0) {
 283                                         r = -errno;
 284
 285                                         if (r == -EINVAL) {
 286                                                 struct loop_info64 info;
 287
 288                                                 /* If we are running on a loop device that has partition scanning off,
 289                                                  * return an explicit recognizable error about this, so that callers
 290                                                  * can generate a proper message explaining the situation. */
 291
 292                                                 if (ioctl(fd, LOOP_GET_STATUS64, &info) >= 0 && (info.lo_flags & LO_FLAGS_PARTSCAN) == 0) {
 293                                                         log_debug("Device is loop device and partition scanning is off!");
 294                                                         return -EPROTONOSUPPORT;
 295                                                 }
 296                                         }
 297                                         if (r != -EBUSY)
 298                                                 return r;
 299                                 } else
 300                                         break;
 301
 302                                 /* If something else has the device open, such as an udev rule, the ioctl will return
 303                                  * EBUSY. Since there's no way to wait until it isn't busy anymore, let's just wait a
 304                                  * bit, and try again.
 305                                  *
 306                                  * This is really something they should fix in the kernel! */
 307
 308                                 (void) usleep(50 * USEC_PER_MSEC);
 309                         }
 310                 }
 311
 312                 e = udev_enumerate_unref(e);
 313         }
 314
 315         first = udev_enumerate_get_list_entry(e);
 316         udev_list_entry_foreach(item, first) {
 317                 _cleanup_udev_device_unref_ struct udev_device *q;
 318                 unsigned long long pflags;
 319                 blkid_partition pp;
 320                 const char *node, *sysname;
 321                 dev_t qn;
 322                 int nr;
 323
 324                 q = udev_device_new_from_syspath(udev, udev_list_entry_get_name(item));
 325                 if (!q)
 326                         return -errno;
 327
 328                 qn = udev_device_get_devnum(q);
 329                 if (major(qn) == 0)
 330                         continue;
 331
 332                 if (st.st_rdev == qn)
 333                         continue;
 334
 335                 /* Filter out weird MMC RPMB partitions, which cannot reasonably be read, see
 336                  * https://github.com/systemd/systemd/issues/5806 */
 337                 sysname = udev_device_get_sysname(q);
 338                 if (sysname && startswith(sysname, "mmcblk") && endswith(sysname, "rpmb"))
 339                         continue;
 340
 341                 node = udev_device_get_devnode(q);
 342                 if (!node)
 343                         continue;
 344
 345                 pp = blkid_partlist_devno_to_partition(pl, qn);
 346                 if (!pp)
 347                         continue;
 348
 349                 pflags = blkid_partition_get_flags(pp);
 350
 351                 nr = blkid_partition_get_partno(pp);
 352                 if (nr < 0)
 353                         continue;
 354
 355                 if (is_gpt) {
 356                         int designator = _PARTITION_DESIGNATOR_INVALID, architecture = _ARCHITECTURE_INVALID;
 357                         const char *stype, *sid, *fstype = NULL;
 358                         sd_id128_t type_id, id;
 359                         bool rw = true;
 360
 361                         sid = blkid_partition_get_uuid(pp);
 362                         if (!sid)
 363                                 continue;
 364                         if (sd_id128_from_string(sid, &id) < 0)
 365                                 continue;
 366
 367                         stype = blkid_partition_get_type_string(pp);
 368                         if (!stype)
 369                                 continue;
 370                         if (sd_id128_from_string(stype, &type_id) < 0)
 371                                 continue;
 372
 373                         if (sd_id128_equal(type_id, GPT_HOME)) {
 374
 375                                 if (pflags & GPT_FLAG_NO_AUTO)
 376                                         continue;
 377
 378                                 designator = PARTITION_HOME;
 379                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 380                         } else if (sd_id128_equal(type_id, GPT_SRV)) {
 381
 382                                 if (pflags & GPT_FLAG_NO_AUTO)
 383                                         continue;
 384
 385                                 designator = PARTITION_SRV;
 386                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 387                         } else if (sd_id128_equal(type_id, GPT_ESP)) {
 388
 389                                 /* Note that we don't check the GPT_FLAG_NO_AUTO flag for the ESP, as it is not defined
 390                                  * there. We instead check the GPT_FLAG_NO_BLOCK_IO_PROTOCOL, as recommended by the
 391                                  * UEFI spec (See "12.3.3 Number and Location of System Partitions"). */
 392
 393                                 if (pflags & GPT_FLAG_NO_BLOCK_IO_PROTOCOL)
 394                                         continue;
 395
 396                                 designator = PARTITION_ESP;
 397                                 fstype = "vfat";
 398                         }
 399 #ifdef GPT_ROOT_NATIVE
 400                         else if (sd_id128_equal(type_id, GPT_ROOT_NATIVE)) {
 401
 402                                 if (pflags & GPT_FLAG_NO_AUTO)
 403                                         continue;
 404
 405                                 /* If a root ID is specified, ignore everything but the root id */
 406                                 if (!sd_id128_is_null(root_uuid) && !sd_id128_equal(root_uuid, id))
 407                                         continue;
 408
 409                                 designator = PARTITION_ROOT;
 410                                 architecture = native_architecture();
 411                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 412                         } else if (sd_id128_equal(type_id, GPT_ROOT_NATIVE_VERITY)) {
 413
 414                                 if (pflags & GPT_FLAG_NO_AUTO)
 415                                         continue;
 416
 417                                 m->can_verity = true;
 418
 419                                 /* Ignore verity unless a root hash is specified */
 420                                 if (sd_id128_is_null(verity_uuid) || !sd_id128_equal(verity_uuid, id))
 421                                         continue;
 422
 423                                 designator = PARTITION_ROOT_VERITY;
 424                                 fstype = "DM_verity_hash";
 425                                 architecture = native_architecture();
 426                                 rw = false;
 427                         }
 428 #endif
 429 #ifdef GPT_ROOT_SECONDARY
 430                         else if (sd_id128_equal(type_id, GPT_ROOT_SECONDARY)) {
 431
 432                                 if (pflags & GPT_FLAG_NO_AUTO)
 433                                         continue;
 434
 435                                 /* If a root ID is specified, ignore everything but the root id */
 436                                 if (!sd_id128_is_null(root_uuid) && !sd_id128_equal(root_uuid, id))
 437                                         continue;
 438
 439                                 designator = PARTITION_ROOT_SECONDARY;
 440                                 architecture = SECONDARY_ARCHITECTURE;
 441                                 rw = !(pflags & GPT_FLAG_READ_ONLY);
 442                         } else if (sd_id128_equal(type_id, GPT_ROOT_SECONDARY_VERITY)) {
 443
 444                                 if (pflags & GPT_FLAG_NO_AUTO)
 445                                         continue;
 446
 447                                 m->can_verity = true;
 448
 449                                 /* Ignore verity unless root has is specified */
 450                                 if (sd_id128_is_null(verity_uuid) || !sd_id128_equal(verity_uuid, id))
 451                                         continue;
 452
 453                                 designator = PARTITION_ROOT_SECONDARY_VERITY;
 454                                 fstype = "DM_verity_hash";
 455                                 architecture = SECONDARY_ARCHITECTURE;
 456                                 rw = false;
 457                         }
 458 #endif
 459                         else if (sd_id128_equal(type_id, GPT_SWAP)) {
 460
 461                                 if (pflags & GPT_FLAG_NO_AUTO)
 462                                         continue;
 463
 464                                 designator = PARTITION_SWAP;
 465                                 fstype = "swap";
 466                         } else if (sd_id128_equal(type_id, GPT_LINUX_GENERIC)) {
 467
 468                                 if (pflags & GPT_FLAG_NO_AUTO)
 469                                         continue;
 470
 471                                 if (generic_node)
 472                                         multiple_generic = true;
 473                                 else {
 474                                         generic_nr = nr;
 475                                         generic_rw = !(pflags & GPT_FLAG_READ_ONLY);
 476                                         generic_uuid = id;
 477                                         generic_node = strdup(node);
 478                                         if (!generic_node)
 479                                                 return -ENOMEM;
 480                                 }
 481                         }
 482
 483                         if (designator != _PARTITION_DESIGNATOR_INVALID) {
 484                                 _cleanup_free_ char *t = NULL, *n = NULL;
 485
 486                                 /* First one wins */
 487                                 if (m->partitions[designator].found)
 488                                         continue;
 489
 490                                 if (fstype) {
 491                                         t = strdup(fstype);
 492                                         if (!t)
 493                                                 return -ENOMEM;
 494                                 }
 495
 496                                 n = strdup(node);
 497                                 if (!n)
 498                                         return -ENOMEM;
 499
 500                                 m->partitions[designator] = (DissectedPartition) {
 501                                         .found = true,
 502                                         .partno = nr,
 503                                         .rw = rw,
 504                                         .architecture = architecture,
 505                                         .node = n,
 506                                         .fstype = t,
 507                                         .uuid = id,
 508                                 };
 509
 510                                 n = t = NULL;
 511                         }
 512
 513                 } else if (is_mbr) {
 514
 515                         if (pflags != 0x80) /* Bootable flag */
 516                                 continue;
 517
 518                         if (blkid_partition_get_type(pp) != 0x83) /* Linux partition */
 519                                 continue;
 520
 521                         if (generic_node)
 522                                 multiple_generic = true;
 523                         else {
 524                                 generic_nr = nr;
 525                                 generic_rw = true;
 526                                 generic_node = strdup(node);
 527                                 if (!generic_node)
 528                                         return -ENOMEM;
 529                         }
 530                 }
 531         }
 532
 533         if (!m->partitions[PARTITION_ROOT].found) {
 534                 /* No root partition found? Then let's see if ther's one for the secondary architecture. And if not
 535                  * either, then check if there's a single generic one, and use that. */
 536
 537                 if (m->partitions[PARTITION_ROOT_VERITY].found)
 538                         return -EADDRNOTAVAIL;
 539
 540                 if (m->partitions[PARTITION_ROOT_SECONDARY].found) {
 541                         m->partitions[PARTITION_ROOT] = m->partitions[PARTITION_ROOT_SECONDARY];
 542                         zero(m->partitions[PARTITION_ROOT_SECONDARY]);
 543
 544                         m->partitions[PARTITION_ROOT_VERITY] = m->partitions[PARTITION_ROOT_SECONDARY_VERITY];
 545                         zero(m->partitions[PARTITION_ROOT_SECONDARY_VERITY]);
 546
 547                 } else if (flags & DISSECT_IMAGE_REQUIRE_ROOT) {
 548
 549                         /* If the root has was set, then we won't fallback to a generic node, because the root hash
 550                          * decides */
 551                         if (root_hash)
 552                                 return -EADDRNOTAVAIL;
 553
 554                         /* If we didn't find a generic node, then we can't fix this up either */
 555                         if (!generic_node)
 556                                 return -ENXIO;
 557
 558                         /* If we didn't find a properly marked root partition, but we did find a single suitable
 559                          * generic Linux partition, then use this as root partition, if the caller asked for it. */
 560                         if (multiple_generic)
 561                                 return -ENOTUNIQ;
 562
 563                         m->partitions[PARTITION_ROOT] = (DissectedPartition) {
 564                                 .found = true,
 565                                 .rw = generic_rw,
 566                                 .partno = generic_nr,
 567                                 .architecture = _ARCHITECTURE_INVALID,
 568                                 .node = generic_node,
 569                                 .uuid = generic_uuid,
 570                         };
 571
 572                         generic_node = NULL;
 573                 }
 574         }
 575
 576         if (root_hash) {
 577                 if (!m->partitions[PARTITION_ROOT_VERITY].found || !m->partitions[PARTITION_ROOT].found)
 578                         return -EADDRNOTAVAIL;
 579
 580                 /* If we found the primary root with the hash, then we definitely want to suppress any secondary root
 581                  * (which would be weird, after all the root hash should only be assigned to one pair of
 582                  * partitions... */
 583                 m->partitions[PARTITION_ROOT_SECONDARY].found = false;
 584                 m->partitions[PARTITION_ROOT_SECONDARY_VERITY].found = false;
 585
 586                 /* If we found a verity setup, then the root partition is necessarily read-only. */
 587                 m->partitions[PARTITION_ROOT].rw = false;
 588
 589                 m->verity = true;
 590         }
 591
 592         blkid_free_probe(b);
 593         b = NULL;
 594
 595         /* Fill in file system types if we don't know them yet. */
 596         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
 597                 DissectedPartition *p = m->partitions + i;
 598
 599                 if (!p->found)
 600                         continue;
 601
 602                 if (!p->fstype && p->node) {
 603                         r = probe_filesystem(p->node, &p->fstype);
 604                         if (r < 0)
 605                                 return r;
 606                 }
 607
 608                 if (streq_ptr(p->fstype, "crypto_LUKS"))
 609                         m->encrypted = true;
 610
 611                 if (p->fstype && fstype_is_ro(p->fstype))
 612                         p->rw = false;
 613         }
 614
 615         *ret = m;
 616         m = NULL;
 617
 618         return 0;
 619 #else
 620         return -EOPNOTSUPP;
 621 #endif
 622 }
 623
 624 DissectedImage* dissected_image_unref(DissectedImage *m) {
 625         unsigned i;
 626
 627         if (!m)
 628                 return NULL;
 629
 630         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
 631                 free(m->partitions[i].fstype);
 632                 free(m->partitions[i].node);
 633                 free(m->partitions[i].decrypted_fstype);
 634                 free(m->partitions[i].decrypted_node);
 635         }
 636
 637         free(m);
 638         return NULL;
 639 }
 640
 641 static int is_loop_device(const char *path) {
 642         char s[strlen("/sys/dev/block/") + DECIMAL_STR_MAX(dev_t) + 1 + DECIMAL_STR_MAX(dev_t) + strlen("/../loop/")];
 643         struct stat st;
 644
 645         assert(path);
 646
 647         if (stat(path, &st) < 0)
 648                 return -errno;
 649
 650         if (!S_ISBLK(st.st_mode))
 651                 return -ENOTBLK;
 652
 653         xsprintf(s, "/sys/dev/block/%u:%u/loop/", major(st.st_rdev), minor(st.st_rdev));
 654         if (access(s, F_OK) < 0) {
 655                 if (errno != ENOENT)
 656                         return -errno;
 657
 658                 /* The device itself isn't a loop device, but maybe it's a partition and its parent is? */
 659                 xsprintf(s, "/sys/dev/block/%u:%u/../loop/", major(st.st_rdev), minor(st.st_rdev));
 660                 if (access(s, F_OK) < 0)
 661                         return errno == ENOENT ? false : -errno;
 662         }
 663
 664         return true;
 665 }
 666
 667 static int mount_partition(
 668                 DissectedPartition *m,
 669                 const char *where,
 670                 const char *directory,
 671                 DissectImageFlags flags) {
 672
 673         const char *p, *options = NULL, *node, *fstype;
 674         _cleanup_free_ char *chased = NULL;
 675         bool rw;
 676         int r;
 677
 678         assert(m);
 679         assert(where);
 680
 681         node = m->decrypted_node ?: m->node;
 682         fstype = m->decrypted_fstype ?: m->fstype;
 683
 684         if (!m->found || !node || !fstype)
 685                 return 0;
 686
 687         /* Stacked encryption? Yuck */
 688         if (streq_ptr(fstype, "crypto_LUKS"))
 689                 return -ELOOP;
 690
 691         rw = m->rw && !(flags & DISSECT_IMAGE_READ_ONLY);
 692
 693         if (directory) {
 694                 r = chase_symlinks(directory, where, CHASE_PREFIX_ROOT, &chased);
 695                 if (r < 0)
 696                         return r;
 697
 698                 p = chased;
 699         } else
 700                 p = where;
 701
 702         /* If requested, turn on discard support. */
 703         if (fstype_can_discard(fstype) &&
 704             ((flags & DISSECT_IMAGE_DISCARD) ||
 705              ((flags & DISSECT_IMAGE_DISCARD_ON_LOOP) && is_loop_device(m->node))))
 706                 options = "discard";
 707
 708         return mount_verbose(LOG_DEBUG, node, p, fstype, MS_NODEV|(rw ? 0 : MS_RDONLY), options);
 709 }
 710
 711 int dissected_image_mount(DissectedImage *m, const char *where, DissectImageFlags flags) {
 712         int r;
 713
 714         assert(m);
 715         assert(where);
 716
 717         if (!m->partitions[PARTITION_ROOT].found)
 718                 return -ENXIO;
 719
 720         r = mount_partition(m->partitions + PARTITION_ROOT, where, NULL, flags);
 721         if (r < 0)
 722                 return r;
 723
 724         r = mount_partition(m->partitions + PARTITION_HOME, where, "/home", flags);
 725         if (r < 0)
 726                 return r;
 727
 728         r = mount_partition(m->partitions + PARTITION_SRV, where, "/srv", flags);
 729         if (r < 0)
 730                 return r;
 731
 732         if (m->partitions[PARTITION_ESP].found) {
 733                 const char *mp;
 734
 735                 /* Mount the ESP to /efi if it exists and is empty. If it doesn't exist, use /boot instead. */
 736
 737                 FOREACH_STRING(mp, "/efi", "/boot") {
 738                         _cleanup_free_ char *p = NULL;
 739
 740                         r = chase_symlinks(mp, where, CHASE_PREFIX_ROOT, &p);
 741                         if (r < 0)
 742                                 continue;
 743
 744                         r = dir_is_empty(p);
 745                         if (r > 0) {
 746                                 r = mount_partition(m->partitions + PARTITION_ESP, where, mp, flags);
 747                                 if (r < 0)
 748                                         return r;
 749                         }
 750                 }
 751         }
 752
 753         return 0;
 754 }
 755
 756 #if HAVE_LIBCRYPTSETUP
 757 typedef struct DecryptedPartition {
 758         struct crypt_device *device;
 759         char *name;
 760         bool relinquished;
 761 } DecryptedPartition;
 762
 763 struct DecryptedImage {
 764         DecryptedPartition *decrypted;
 765         size_t n_decrypted;
 766         size_t n_allocated;
 767 };
 768 #endif
 769
 770 DecryptedImage* decrypted_image_unref(DecryptedImage* d) {
 771 #if HAVE_LIBCRYPTSETUP
 772         size_t i;
 773         int r;
 774
 775         if (!d)
 776                 return NULL;
 777
 778         for (i = 0; i < d->n_decrypted; i++) {
 779                 DecryptedPartition *p = d->decrypted + i;
 780
 781                 if (p->device && p->name && !p->relinquished) {
 782                         r = crypt_deactivate(p->device, p->name);
 783                         if (r < 0)
 784                                 log_debug_errno(r, "Failed to deactivate encrypted partition %s", p->name);
 785                 }
 786
 787                 if (p->device)
 788                         crypt_free(p->device);
 789                 free(p->name);
 790         }
 791
 792         free(d);
 793 #endif
 794         return NULL;
 795 }
 796
 797 #if HAVE_LIBCRYPTSETUP
 798
 799 static int make_dm_name_and_node(const void *original_node, const char *suffix, char **ret_name, char **ret_node) {
 800         _cleanup_free_ char *name = NULL, *node = NULL;
 801         const char *base;
 802
 803         assert(original_node);
 804         assert(suffix);
 805         assert(ret_name);
 806         assert(ret_node);
 807
 808         base = strrchr(original_node, '/');
 809         if (!base)
 810                 return -EINVAL;
 811         base++;
 812         if (isempty(base))
 813                 return -EINVAL;
 814
 815         name = strjoin(base, suffix);
 816         if (!name)
 817                 return -ENOMEM;
 818         if (!filename_is_valid(name))
 819                 return -EINVAL;
 820
 821         node = strjoin(crypt_get_dir(), "/", name);
 822         if (!node)
 823                 return -ENOMEM;
 824
 825         *ret_name = name;
 826         *ret_node = node;
 827
 828         name = node = NULL;
 829         return 0;
 830 }
 831
 832 static int decrypt_partition(
 833                 DissectedPartition *m,
 834                 const char *passphrase,
 835                 DissectImageFlags flags,
 836                 DecryptedImage *d) {
 837
 838         _cleanup_free_ char *node = NULL, *name = NULL;
 839         struct crypt_device *cd;
 840         int r;
 841
 842         assert(m);
 843         assert(d);
 844
 845         if (!m->found || !m->node || !m->fstype)
 846                 return 0;
 847
 848         if (!streq(m->fstype, "crypto_LUKS"))
 849                 return 0;
 850
 851         r = make_dm_name_and_node(m->node, "-decrypted", &name, &node);
 852         if (r < 0)
 853                 return r;
 854
 855         if (!GREEDY_REALLOC0(d->decrypted, d->n_allocated, d->n_decrypted + 1))
 856                 return -ENOMEM;
 857
 858         r = crypt_init(&cd, m->node);
 859         if (r < 0)
 860                 return log_debug_errno(r, "Failed to initialize dm-crypt: %m");
 861
 862         r = crypt_load(cd, CRYPT_LUKS, NULL);
 863         if (r < 0) {
 864                 log_debug_errno(r, "Failed to load LUKS metadata: %m");
 865                 goto fail;
 866         }
 867
 868         r = crypt_activate_by_passphrase(cd, name, CRYPT_ANY_SLOT, passphrase, strlen(passphrase),
 869                                          ((flags & DISSECT_IMAGE_READ_ONLY) ? CRYPT_ACTIVATE_READONLY : 0) |
 870                                          ((flags & DISSECT_IMAGE_DISCARD_ON_CRYPTO) ? CRYPT_ACTIVATE_ALLOW_DISCARDS : 0));
 871         if (r < 0)
 872                 log_debug_errno(r, "Failed to activate LUKS device: %m");
 873         if (r == -EPERM) {
 874                 r = -EKEYREJECTED;
 875                 goto fail;
 876         }
 877         if (r < 0)
 878                 goto fail;
 879
 880         d->decrypted[d->n_decrypted].name = name;
 881         name = NULL;
 882
 883         d->decrypted[d->n_decrypted].device = cd;
 884         d->n_decrypted++;
 885
 886         m->decrypted_node = node;
 887         node = NULL;
 888
 889         return 0;
 890
 891 fail:
 892         crypt_free(cd);
 893         return r;
 894 }
 895
 896 static int verity_partition(
 897                 DissectedPartition *m,
 898                 DissectedPartition *v,
 899                 const void *root_hash,
 900                 size_t root_hash_size,
 901                 DissectImageFlags flags,
 902                 DecryptedImage *d) {
 903
 904         _cleanup_free_ char *node = NULL, *name = NULL;
 905         struct crypt_device *cd;
 906         int r;
 907
 908         assert(m);
 909         assert(v);
 910
 911         if (!root_hash)
 912                 return 0;
 913
 914         if (!m->found || !m->node || !m->fstype)
 915                 return 0;
 916         if (!v->found || !v->node || !v->fstype)
 917                 return 0;
 918
 919         if (!streq(v->fstype, "DM_verity_hash"))
 920                 return 0;
 921
 922         r = make_dm_name_and_node(m->node, "-verity", &name, &node);
 923         if (r < 0)
 924                 return r;
 925
 926         if (!GREEDY_REALLOC0(d->decrypted, d->n_allocated, d->n_decrypted + 1))
 927                 return -ENOMEM;
 928
 929         r = crypt_init(&cd, v->node);
 930         if (r < 0)
 931                 return r;
 932
 933         r = crypt_load(cd, CRYPT_VERITY, NULL);
 934         if (r < 0)
 935                 goto fail;
 936
 937         r = crypt_set_data_device(cd, m->node);
 938         if (r < 0)
 939                 goto fail;
 940
 941         r = crypt_activate_by_volume_key(cd, name, root_hash, root_hash_size, CRYPT_ACTIVATE_READONLY);
 942         if (r < 0)
 943                 goto fail;
 944
 945         d->decrypted[d->n_decrypted].name = name;
 946         name = NULL;
 947
 948         d->decrypted[d->n_decrypted].device = cd;
 949         d->n_decrypted++;
 950
 951         m->decrypted_node = node;
 952         node = NULL;
 953
 954         return 0;
 955
 956 fail:
 957         crypt_free(cd);
 958         return r;
 959 }
 960 #endif
 961
 962 int dissected_image_decrypt(
 963                 DissectedImage *m,
 964                 const char *passphrase,
 965                 const void *root_hash,
 966                 size_t root_hash_size,
 967                 DissectImageFlags flags,
 968                 DecryptedImage **ret) {
 969
 970         _cleanup_(decrypted_image_unrefp) DecryptedImage *d = NULL;
 971 #if HAVE_LIBCRYPTSETUP
 972         unsigned i;
 973         int r;
 974 #endif
 975
 976         assert(m);
 977         assert(root_hash || root_hash_size == 0);
 978
 979         /* Returns:
 980          *
 981          *      = 0           → There was nothing to decrypt
 982          *      > 0           → Decrypted successfully
 983          *      -ENOKEY       → There's something to decrypt but no key was supplied
 984          *      -EKEYREJECTED → Passed key was not correct
 985          */
 986
 987         if (root_hash && root_hash_size < sizeof(sd_id128_t))
 988                 return -EINVAL;
 989
 990         if (!m->encrypted && !m->verity) {
 991                 *ret = NULL;
 992                 return 0;
 993         }
 994
 995 #if HAVE_LIBCRYPTSETUP
 996         if (m->encrypted && !passphrase)
 997                 return -ENOKEY;
 998
 999         d = new0(DecryptedImage, 1);
1000         if (!d)
1001                 return -ENOMEM;
1002
1003         for (i = 0; i < _PARTITION_DESIGNATOR_MAX; i++) {
1004                 DissectedPartition *p = m->partitions + i;
1005                 int k;
1006
1007                 if (!p->found)
1008                         continue;
1009
1010                 r = decrypt_partition(p, passphrase, flags, d);
1011                 if (r < 0)
1012                         return r;
1013
1014                 k = PARTITION_VERITY_OF(i);
1015                 if (k >= 0) {
1016                         r = verity_partition(p, m->partitions + k, root_hash, root_hash_size, flags, d);
1017                         if (r < 0)
1018                                 return r;
1019                 }
1020
1021                 if (!p->decrypted_fstype && p->decrypted_node) {
1022                         r = probe_filesystem(p->decrypted_node, &p->decrypted_fstype);
1023                         if (r < 0)
1024                                 return r;
1025                 }
1026         }
1027
1028         *ret = d;
1029         d = NULL;
1030
1031         return 1;
1032 #else
1033         return -EOPNOTSUPP;
1034 #endif
1035 }
1036
1037 int dissected_image_decrypt_interactively(
1038                 DissectedImage *m,
1039                 const char *passphrase,
1040                 const void *root_hash,
1041                 size_t root_hash_size,
1042                 DissectImageFlags flags,
1043                 DecryptedImage **ret) {
1044
1045         _cleanup_strv_free_erase_ char **z = NULL;
1046         int n = 3, r;
1047
1048         if (passphrase)
1049                 n--;
1050
1051         for (;;) {
1052                 r = dissected_image_decrypt(m, passphrase, root_hash, root_hash_size, flags, ret);
1053                 if (r >= 0)
1054                         return r;
1055                 if (r == -EKEYREJECTED)
1056                         log_error_errno(r, "Incorrect passphrase, try again!");
1057                 else if (r != -ENOKEY) {
1058                         log_error_errno(r, "Failed to decrypt image: %m");
1059                         return r;
1060                 }
1061
1062                 if (--n < 0) {
1063                         log_error("Too many retries.");
1064                         return -EKEYREJECTED;
1065                 }
1066
1067                 z = strv_free(z);
1068
1069                 r = ask_password_auto("Please enter image passphrase!", NULL, "dissect", "dissect", USEC_INFINITY, 0, &z);
1070                 if (r < 0)
1071                         return log_error_errno(r, "Failed to query for passphrase: %m");
1072
1073                 passphrase = z[0];
1074         }
1075 }
1076
1077 #if HAVE_LIBCRYPTSETUP
1078 static int deferred_remove(DecryptedPartition *p) {
1079
1080         struct dm_ioctl dm = {
1081                 .version = {
1082                         DM_VERSION_MAJOR,
1083                         DM_VERSION_MINOR,
1084                         DM_VERSION_PATCHLEVEL
1085                 },
1086                 .data_size = sizeof(dm),
1087                 .flags = DM_DEFERRED_REMOVE,
1088         };
1089
1090         _cleanup_close_ int fd = -1;
1091
1092         assert(p);
1093
1094         /* Unfortunately, libcryptsetup doesn't provide a proper API for this, hence call the ioctl() directly. */
1095
1096         fd = open("/dev/mapper/control", O_RDWR|O_CLOEXEC);
1097         if (fd < 0)
1098                 return -errno;
1099
1100         strncpy(dm.name, p->name, sizeof(dm.name));
1101
1102         if (ioctl(fd, DM_DEV_REMOVE, &dm))
1103                 return -errno;
1104
1105         return 0;
1106 }
1107 #endif
1108
1109 int decrypted_image_relinquish(DecryptedImage *d) {
1110
1111 #if HAVE_LIBCRYPTSETUP
1112         size_t i;
1113         int r;
1114 #endif
1115
1116         assert(d);
1117
1118         /* Turns on automatic removal after the last use ended for all DM devices of this image, and sets a boolean so
1119          * that we don't clean it up ourselves either anymore */
1120
1121 #if HAVE_LIBCRYPTSETUP
1122         for (i = 0; i < d->n_decrypted; i++) {
1123                 DecryptedPartition *p = d->decrypted + i;
1124
1125                 if (p->relinquished)
1126                         continue;
1127
1128                 r = deferred_remove(p);
1129                 if (r < 0)
1130                         return log_debug_errno(r, "Failed to mark %s for auto-removal: %m", p->name);
1131
1132                 p->relinquished = true;
1133         }
1134 #endif
1135
1136         return 0;
1137 }
1138
1139 int root_hash_load(const char *image, void **ret, size_t *ret_size) {
1140         _cleanup_free_ char *text = NULL;
1141         _cleanup_free_ void *k = NULL;
1142         size_t l;
1143         int r;
1144
1145         assert(image);
1146         assert(ret);
1147         assert(ret_size);
1148
1149         if (is_device_path(image)) {
1150                 /* If we are asked to load the root hash for a device node, exit early */
1151                 *ret = NULL;
1152                 *ret_size = 0;
1153                 return 0;
1154         }
1155
1156         r = getxattr_malloc(image, "user.verity.roothash", &text, true);
1157         if (r < 0) {
1158                 char *fn, *e, *n;
1159
1160                 if (!IN_SET(r, -ENODATA, -EOPNOTSUPP, -ENOENT))
1161                         return r;
1162
1163                 fn = newa(char, strlen(image) + strlen(".roothash") + 1);
1164                 n = stpcpy(fn, image);
1165                 e = endswith(fn, ".raw");
1166                 if (e)
1167                         n = e;
1168
1169                 strcpy(n, ".roothash");
1170
1171                 r = read_one_line_file(fn, &text);
1172                 if (r == -ENOENT) {
1173                         *ret = NULL;
1174                         *ret_size = 0;
1175                         return 0;
1176                 }
1177                 if (r < 0)
1178                         return r;
1179         }
1180
1181         r = unhexmem(text, strlen(text), &k, &l);
1182         if (r < 0)
1183                 return r;
1184         if (l < sizeof(sd_id128_t))
1185                 return -EINVAL;
1186
1187         *ret = k;
1188         *ret_size = l;
1189
1190         k = NULL;
1191
1192         return 1;
1193 }
1194
1195 static const char *const partition_designator_table[] = {
1196         [PARTITION_ROOT] = "root",
1197         [PARTITION_ROOT_SECONDARY] = "root-secondary",
1198         [PARTITION_HOME] = "home",
1199         [PARTITION_SRV] = "srv",
1200         [PARTITION_ESP] = "esp",
1201         [PARTITION_SWAP] = "swap",
1202         [PARTITION_ROOT_VERITY] = "root-verity",
1203         [PARTITION_ROOT_SECONDARY_VERITY] = "root-secondary-verity",
1204 };
1205
1206 DEFINE_STRING_TABLE_LOOKUP(partition_designator, int);