drivers/md/dm-verity-target.c

   1 /*
   2  * Copyright (C) 2012 Red Hat, Inc.
   3  *
   4  * Author: Mikulas Patocka <mpatocka@redhat.com>
   5  *
   6  * Based on Chromium dm-verity driver (C) 2011 The Chromium OS Authors
   7  *
   8  * This file is released under the GPLv2.
   9  *
  10  * In the file "/sys/module/dm_verity/parameters/prefetch_cluster" you can set
  11  * default prefetch value. Data are read in "prefetch_cluster" chunks from the
  12  * hash device. Setting this greatly improves performance when data and hash
  13  * are on the same disk on different partitions on devices with poor random
  14  * access behavior.
  15  */
  16
  17 #include "dm-verity.h"
  18 #include "dm-verity-fec.h"
  19
  20 #include <linux/module.h>
  21 #include <linux/reboot.h>
  22
  23 #define DM_MSG_PREFIX                   "verity"
  24
  25 #define DM_VERITY_ENV_LENGTH            42
  26 #define DM_VERITY_ENV_VAR_NAME          "DM_VERITY_ERR_BLOCK_NR"
  27
  28 #define DM_VERITY_DEFAULT_PREFETCH_SIZE 262144
  29
  30 #define DM_VERITY_MAX_CORRUPTED_ERRS    100
  31
  32 #define DM_VERITY_OPT_LOGGING           "ignore_corruption"
  33 #define DM_VERITY_OPT_RESTART           "restart_on_corruption"
  34 #define DM_VERITY_OPT_IGN_ZEROES        "ignore_zero_blocks"
  35 #define DM_VERITY_OPT_AT_MOST_ONCE      "check_at_most_once"
  36
  37 #define DM_VERITY_OPTS_MAX              (3 + DM_VERITY_OPTS_FEC)
  38
  39 static unsigned dm_verity_prefetch_cluster = DM_VERITY_DEFAULT_PREFETCH_SIZE;
  40
  41 module_param_named(prefetch_cluster, dm_verity_prefetch_cluster, uint, S_IRUGO | S_IWUSR);
  42
  43 struct dm_verity_prefetch_work {
  44         struct work_struct work;
  45         struct dm_verity *v;
  46         sector_t block;
  47         unsigned n_blocks;
  48 };
  49
  50 /*
  51  * Auxiliary structure appended to each dm-bufio buffer. If the value
  52  * hash_verified is nonzero, hash of the block has been verified.
  53  *
  54  * The variable hash_verified is set to 0 when allocating the buffer, then
  55  * it can be changed to 1 and it is never reset to 0 again.
  56  *
  57  * There is no lock around this value, a race condition can at worst cause
  58  * that multiple processes verify the hash of the same buffer simultaneously
  59  * and write 1 to hash_verified simultaneously.
  60  * This condition is harmless, so we don't need locking.
  61  */
  62 struct buffer_aux {
  63         int hash_verified;
  64 };
  65
  66 /*
  67  * Initialize struct buffer_aux for a freshly created buffer.
  68  */
  69 static void dm_bufio_alloc_callback(struct dm_buffer *buf)
  70 {
  71         struct buffer_aux *aux = dm_bufio_get_aux_data(buf);
  72
  73         aux->hash_verified = 0;
  74 }
  75
  76 /*
  77  * Translate input sector number to the sector number on the target device.
  78  */
  79 static sector_t verity_map_sector(struct dm_verity *v, sector_t bi_sector)
  80 {
  81         return v->data_start + dm_target_offset(v->ti, bi_sector);
  82 }
  83
  84 /*
  85  * Return hash position of a specified block at a specified tree level
  86  * (0 is the lowest level).
  87  * The lowest "hash_per_block_bits"-bits of the result denote hash position
  88  * inside a hash block. The remaining bits denote location of the hash block.
  89  */
  90 static sector_t verity_position_at_level(struct dm_verity *v, sector_t block,
  91                                          int level)
  92 {
  93         return block >> (level * v->hash_per_block_bits);
  94 }
  95
  96 static int verity_hash_update(struct dm_verity *v, struct ahash_request *req,
  97                                 const u8 *data, size_t len,
  98                                 struct crypto_wait *wait)
  99 {
 100         struct scatterlist sg;
 101
 102         if (likely(!is_vmalloc_addr(data))) {
 103                 sg_init_one(&sg, data, len);
 104                 ahash_request_set_crypt(req, &sg, NULL, len);
 105                 return crypto_wait_req(crypto_ahash_update(req), wait);
 106         } else {
 107                 do {
 108                         int r;
 109                         size_t this_step = min_t(size_t, len, PAGE_SIZE - offset_in_page(data));
 110                         flush_kernel_vmap_range((void *)data, this_step);
 111                         sg_init_table(&sg, 1);
 112                         sg_set_page(&sg, vmalloc_to_page(data), this_step, offset_in_page(data));
 113                         ahash_request_set_crypt(req, &sg, NULL, this_step);
 114                         r = crypto_wait_req(crypto_ahash_update(req), wait);
 115                         if (unlikely(r))
 116                                 return r;
 117                         data += this_step;
 118                         len -= this_step;
 119                 } while (len);
 120                 return 0;
 121         }
 122 }
 123
 124 /*
 125  * Wrapper for crypto_ahash_init, which handles verity salting.
 126  */
 127 static int verity_hash_init(struct dm_verity *v, struct ahash_request *req,
 128                                 struct crypto_wait *wait)
 129 {
 130         int r;
 131
 132         ahash_request_set_tfm(req, v->tfm);
 133         ahash_request_set_callback(req, CRYPTO_TFM_REQ_MAY_SLEEP |
 134                                         CRYPTO_TFM_REQ_MAY_BACKLOG,
 135                                         crypto_req_done, (void *)wait);
 136         crypto_init_wait(wait);
 137
 138         r = crypto_wait_req(crypto_ahash_init(req), wait);
 139
 140         if (unlikely(r < 0)) {
 141                 DMERR("crypto_ahash_init failed: %d", r);
 142                 return r;
 143         }
 144
 145         if (likely(v->salt_size && (v->version >= 1)))
 146                 r = verity_hash_update(v, req, v->salt, v->salt_size, wait);
 147
 148         return r;
 149 }
 150
 151 static int verity_hash_final(struct dm_verity *v, struct ahash_request *req,
 152                              u8 *digest, struct crypto_wait *wait)
 153 {
 154         int r;
 155
 156         if (unlikely(v->salt_size && (!v->version))) {
 157                 r = verity_hash_update(v, req, v->salt, v->salt_size, wait);
 158
 159                 if (r < 0) {
 160                         DMERR("verity_hash_final failed updating salt: %d", r);
 161                         goto out;
 162                 }
 163         }
 164
 165         ahash_request_set_crypt(req, NULL, digest, 0);
 166         r = crypto_wait_req(crypto_ahash_final(req), wait);
 167 out:
 168         return r;
 169 }
 170
 171 int verity_hash(struct dm_verity *v, struct ahash_request *req,
 172                 const u8 *data, size_t len, u8 *digest)
 173 {
 174         int r;
 175         struct crypto_wait wait;
 176
 177         r = verity_hash_init(v, req, &wait);
 178         if (unlikely(r < 0))
 179                 goto out;
 180
 181         r = verity_hash_update(v, req, data, len, &wait);
 182         if (unlikely(r < 0))
 183                 goto out;
 184
 185         r = verity_hash_final(v, req, digest, &wait);
 186
 187 out:
 188         return r;
 189 }
 190
 191 static void verity_hash_at_level(struct dm_verity *v, sector_t block, int level,
 192                                  sector_t *hash_block, unsigned *offset)
 193 {
 194         sector_t position = verity_position_at_level(v, block, level);
 195         unsigned idx;
 196
 197         *hash_block = v->hash_level_block[level] + (position >> v->hash_per_block_bits);
 198
 199         if (!offset)
 200                 return;
 201
 202         idx = position & ((1 << v->hash_per_block_bits) - 1);
 203         if (!v->version)
 204                 *offset = idx * v->digest_size;
 205         else
 206                 *offset = idx << (v->hash_dev_block_bits - v->hash_per_block_bits);
 207 }
 208
 209 /*
 210  * Handle verification errors.
 211  */
 212 static int verity_handle_err(struct dm_verity *v, enum verity_block_type type,
 213                              unsigned long long block)
 214 {
 215         char verity_env[DM_VERITY_ENV_LENGTH];
 216         char *envp[] = { verity_env, NULL };
 217         const char *type_str = "";
 218         struct mapped_device *md = dm_table_get_md(v->ti->table);
 219
 220         /* Corruption should be visible in device status in all modes */
 221         v->hash_failed = 1;
 222
 223         if (v->corrupted_errs >= DM_VERITY_MAX_CORRUPTED_ERRS)
 224                 goto out;
 225
 226         v->corrupted_errs++;
 227
 228         switch (type) {
 229         case DM_VERITY_BLOCK_TYPE_DATA:
 230                 type_str = "data";
 231                 break;
 232         case DM_VERITY_BLOCK_TYPE_METADATA:
 233                 type_str = "metadata";
 234                 break;
 235         default:
 236                 BUG();
 237         }
 238
 239         DMERR_LIMIT("%s: %s block %llu is corrupted", v->data_dev->name,
 240                     type_str, block);
 241
 242         if (v->corrupted_errs == DM_VERITY_MAX_CORRUPTED_ERRS)
 243                 DMERR("%s: reached maximum errors", v->data_dev->name);
 244
 245         snprintf(verity_env, DM_VERITY_ENV_LENGTH, "%s=%d,%llu",
 246                 DM_VERITY_ENV_VAR_NAME, type, block);
 247
 248         kobject_uevent_env(&disk_to_dev(dm_disk(md))->kobj, KOBJ_CHANGE, envp);
 249
 250 out:
 251         if (v->mode == DM_VERITY_MODE_LOGGING)
 252                 return 0;
 253
 254         if (v->mode == DM_VERITY_MODE_RESTART)
 255                 kernel_restart("dm-verity device corrupted");
 256
 257         return 1;
 258 }
 259
 260 /*
 261  * Verify hash of a metadata block pertaining to the specified data block
 262  * ("block" argument) at a specified level ("level" argument).
 263  *
 264  * On successful return, verity_io_want_digest(v, io) contains the hash value
 265  * for a lower tree level or for the data block (if we're at the lowest level).
 266  *
 267  * If "skip_unverified" is true, unverified buffer is skipped and 1 is returned.
 268  * If "skip_unverified" is false, unverified buffer is hashed and verified
 269  * against current value of verity_io_want_digest(v, io).
 270  */
 271 static int verity_verify_level(struct dm_verity *v, struct dm_verity_io *io,
 272                                sector_t block, int level, bool skip_unverified,
 273                                u8 *want_digest)
 274 {
 275         struct dm_buffer *buf;
 276         struct buffer_aux *aux;
 277         u8 *data;
 278         int r;
 279         sector_t hash_block;
 280         unsigned offset;
 281
 282         verity_hash_at_level(v, block, level, &hash_block, &offset);
 283
 284         data = dm_bufio_read(v->bufio, hash_block, &buf);
 285         if (IS_ERR(data))
 286                 return PTR_ERR(data);
 287
 288         aux = dm_bufio_get_aux_data(buf);
 289
 290         if (!aux->hash_verified) {
 291                 if (skip_unverified) {
 292                         r = 1;
 293                         goto release_ret_r;
 294                 }
 295
 296                 r = verity_hash(v, verity_io_hash_req(v, io),
 297                                 data, 1 << v->hash_dev_block_bits,
 298                                 verity_io_real_digest(v, io));
 299                 if (unlikely(r < 0))
 300                         goto release_ret_r;
 301
 302                 if (likely(memcmp(verity_io_real_digest(v, io), want_digest,
 303                                   v->digest_size) == 0))
 304                         aux->hash_verified = 1;
 305                 else if (verity_fec_decode(v, io,
 306                                            DM_VERITY_BLOCK_TYPE_METADATA,
 307                                            hash_block, data, NULL) == 0)
 308                         aux->hash_verified = 1;
 309                 else if (verity_handle_err(v,
 310                                            DM_VERITY_BLOCK_TYPE_METADATA,
 311                                            hash_block)) {
 312                         r = -EIO;
 313                         goto release_ret_r;
 314                 }
 315         }
 316
 317         data += offset;
 318         memcpy(want_digest, data, v->digest_size);
 319         r = 0;
 320
 321 release_ret_r:
 322         dm_bufio_release(buf);
 323         return r;
 324 }
 325
 326 /*
 327  * Find a hash for a given block, write it to digest and verify the integrity
 328  * of the hash tree if necessary.
 329  */
 330 int verity_hash_for_block(struct dm_verity *v, struct dm_verity_io *io,
 331                           sector_t block, u8 *digest, bool *is_zero)
 332 {
 333         int r = 0, i;
 334
 335         if (likely(v->levels)) {
 336                 /*
 337                  * First, we try to get the requested hash for
 338                  * the current block. If the hash block itself is
 339                  * verified, zero is returned. If it isn't, this
 340                  * function returns 1 and we fall back to whole
 341                  * chain verification.
 342                  */
 343                 r = verity_verify_level(v, io, block, 0, true, digest);
 344                 if (likely(r <= 0))
 345                         goto out;
 346         }
 347
 348         memcpy(digest, v->root_digest, v->digest_size);
 349
 350         for (i = v->levels - 1; i >= 0; i--) {
 351                 r = verity_verify_level(v, io, block, i, false, digest);
 352                 if (unlikely(r))
 353                         goto out;
 354         }
 355 out:
 356         if (!r && v->zero_digest)
 357                 *is_zero = !memcmp(v->zero_digest, digest, v->digest_size);
 358         else
 359                 *is_zero = false;
 360
 361         return r;
 362 }
 363
 364 /*
 365  * Calculates the digest for the given bio
 366  */
 367 static int verity_for_io_block(struct dm_verity *v, struct dm_verity_io *io,
 368                                struct bvec_iter *iter, struct crypto_wait *wait)
 369 {
 370         unsigned int todo = 1 << v->data_dev_block_bits;
 371         struct bio *bio = dm_bio_from_per_bio_data(io, v->ti->per_io_data_size);
 372         struct scatterlist sg;
 373         struct ahash_request *req = verity_io_hash_req(v, io);
 374
 375         do {
 376                 int r;
 377                 unsigned int len;
 378                 struct bio_vec bv = bio_iter_iovec(bio, *iter);
 379
 380                 sg_init_table(&sg, 1);
 381
 382                 len = bv.bv_len;
 383
 384                 if (likely(len >= todo))
 385                         len = todo;
 386                 /*
 387                  * Operating on a single page at a time looks suboptimal
 388                  * until you consider the typical block size is 4,096B.
 389                  * Going through this loops twice should be very rare.
 390                  */
 391                 sg_set_page(&sg, bv.bv_page, len, bv.bv_offset);
 392                 ahash_request_set_crypt(req, &sg, NULL, len);
 393                 r = crypto_wait_req(crypto_ahash_update(req), wait);
 394
 395                 if (unlikely(r < 0)) {
 396                         DMERR("verity_for_io_block crypto op failed: %d", r);
 397                         return r;
 398                 }
 399
 400                 bio_advance_iter(bio, iter, len);
 401                 todo -= len;
 402         } while (todo);
 403
 404         return 0;
 405 }
 406
 407 /*
 408  * Calls function process for 1 << v->data_dev_block_bits bytes in the bio_vec
 409  * starting from iter.
 410  */
 411 int verity_for_bv_block(struct dm_verity *v, struct dm_verity_io *io,
 412                         struct bvec_iter *iter,
 413                         int (*process)(struct dm_verity *v,
 414                                        struct dm_verity_io *io, u8 *data,
 415                                        size_t len))
 416 {
 417         unsigned todo = 1 << v->data_dev_block_bits;
 418         struct bio *bio = dm_bio_from_per_bio_data(io, v->ti->per_io_data_size);
 419
 420         do {
 421                 int r;
 422                 u8 *page;
 423                 unsigned len;
 424                 struct bio_vec bv = bio_iter_iovec(bio, *iter);
 425
 426                 page = kmap_atomic(bv.bv_page);
 427                 len = bv.bv_len;
 428
 429                 if (likely(len >= todo))
 430                         len = todo;
 431
 432                 r = process(v, io, page + bv.bv_offset, len);
 433                 kunmap_atomic(page);
 434
 435                 if (r < 0)
 436                         return r;
 437
 438                 bio_advance_iter(bio, iter, len);
 439                 todo -= len;
 440         } while (todo);
 441
 442         return 0;
 443 }
 444
 445 static int verity_bv_zero(struct dm_verity *v, struct dm_verity_io *io,
 446                           u8 *data, size_t len)
 447 {
 448         memset(data, 0, len);
 449         return 0;
 450 }
 451
 452 /*
 453  * Moves the bio iter one data block forward.
 454  */
 455 static inline void verity_bv_skip_block(struct dm_verity *v,
 456                                         struct dm_verity_io *io,
 457                                         struct bvec_iter *iter)
 458 {
 459         struct bio *bio = dm_bio_from_per_bio_data(io, v->ti->per_io_data_size);
 460
 461         bio_advance_iter(bio, iter, 1 << v->data_dev_block_bits);
 462 }
 463
 464 /*
 465  * Verify one "dm_verity_io" structure.
 466  */
 467 static int verity_verify_io(struct dm_verity_io *io)
 468 {
 469         bool is_zero;
 470         struct dm_verity *v = io->v;
 471         struct bvec_iter start;
 472         unsigned b;
 473         struct crypto_wait wait;
 474         struct bio *bio = dm_bio_from_per_bio_data(io, v->ti->per_io_data_size);
 475
 476         for (b = 0; b < io->n_blocks; b++) {
 477                 int r;
 478                 sector_t cur_block = io->block + b;
 479                 struct ahash_request *req = verity_io_hash_req(v, io);
 480
 481                 if (v->validated_blocks && bio->bi_status == BLK_STS_OK &&
 482                     likely(test_bit(cur_block, v->validated_blocks))) {
 483                         verity_bv_skip_block(v, io, &io->iter);
 484                         continue;
 485                 }
 486
 487                 r = verity_hash_for_block(v, io, cur_block,
 488                                           verity_io_want_digest(v, io),
 489                                           &is_zero);
 490                 if (unlikely(r < 0))
 491                         return r;
 492
 493                 if (is_zero) {
 494                         /*
 495                          * If we expect a zero block, don't validate, just
 496                          * return zeros.
 497                          */
 498                         r = verity_for_bv_block(v, io, &io->iter,
 499                                                 verity_bv_zero);
 500                         if (unlikely(r < 0))
 501                                 return r;
 502
 503                         continue;
 504                 }
 505
 506                 r = verity_hash_init(v, req, &wait);
 507                 if (unlikely(r < 0))
 508                         return r;
 509
 510                 start = io->iter;
 511                 r = verity_for_io_block(v, io, &io->iter, &wait);
 512                 if (unlikely(r < 0))
 513                         return r;
 514
 515                 r = verity_hash_final(v, req, verity_io_real_digest(v, io),
 516                                         &wait);
 517                 if (unlikely(r < 0))
 518                         return r;
 519
 520                 if (likely(memcmp(verity_io_real_digest(v, io),
 521                                   verity_io_want_digest(v, io), v->digest_size) == 0)) {
 522                         if (v->validated_blocks)
 523                                 set_bit(cur_block, v->validated_blocks);
 524                         continue;
 525                 }
 526                 else if (verity_fec_decode(v, io, DM_VERITY_BLOCK_TYPE_DATA,
 527                                            cur_block, NULL, &start) == 0)
 528                         continue;
 529                 else {
 530                         if (bio->bi_status) {
 531                                 /*
 532                                  * Error correction failed; Just return error
 533                                  */
 534                                 return -EIO;
 535                         }
 536                         if (verity_handle_err(v, DM_VERITY_BLOCK_TYPE_DATA,
 537                                               cur_block))
 538                                 return -EIO;
 539                 }
 540         }
 541
 542         return 0;
 543 }
 544
 545 /*
 546  * Skip verity work in response to I/O error when system is shutting down.
 547  */
 548 static inline bool verity_is_system_shutting_down(void)
 549 {
 550         return system_state == SYSTEM_HALT || system_state == SYSTEM_POWER_OFF
 551                 || system_state == SYSTEM_RESTART;
 552 }
 553
 554 /*
 555  * End one "io" structure with a given error.
 556  */
 557 static void verity_finish_io(struct dm_verity_io *io, blk_status_t status)
 558 {
 559         struct dm_verity *v = io->v;
 560         struct bio *bio = dm_bio_from_per_bio_data(io, v->ti->per_io_data_size);
 561
 562         bio->bi_end_io = io->orig_bi_end_io;
 563         bio->bi_status = status;
 564
 565         verity_fec_finish_io(io);
 566
 567         bio_endio(bio);
 568 }
 569
 570 static void verity_work(struct work_struct *w)
 571 {
 572         struct dm_verity_io *io = container_of(w, struct dm_verity_io, work);
 573
 574         verity_finish_io(io, errno_to_blk_status(verity_verify_io(io)));
 575 }
 576
 577 static void verity_end_io(struct bio *bio)
 578 {
 579         struct dm_verity_io *io = bio->bi_private;
 580
 581         if (bio->bi_status &&
 582             (!verity_fec_is_enabled(io->v) ||
 583              verity_is_system_shutting_down() ||
 584              (bio->bi_opf & REQ_RAHEAD))) {
 585                 verity_finish_io(io, bio->bi_status);
 586                 return;
 587         }
 588
 589         INIT_WORK(&io->work, verity_work);
 590         queue_work(io->v->verify_wq, &io->work);
 591 }
 592
 593 /*
 594  * Prefetch buffers for the specified io.
 595  * The root buffer is not prefetched, it is assumed that it will be cached
 596  * all the time.
 597  */
 598 static void verity_prefetch_io(struct work_struct *work)
 599 {
 600         struct dm_verity_prefetch_work *pw =
 601                 container_of(work, struct dm_verity_prefetch_work, work);
 602         struct dm_verity *v = pw->v;
 603         int i;
 604
 605         for (i = v->levels - 2; i >= 0; i--) {
 606                 sector_t hash_block_start;
 607                 sector_t hash_block_end;
 608                 verity_hash_at_level(v, pw->block, i, &hash_block_start, NULL);
 609                 verity_hash_at_level(v, pw->block + pw->n_blocks - 1, i, &hash_block_end, NULL);
 610                 if (!i) {
 611                         unsigned cluster = READ_ONCE(dm_verity_prefetch_cluster);
 612
 613                         cluster >>= v->data_dev_block_bits;
 614                         if (unlikely(!cluster))
 615                                 goto no_prefetch_cluster;
 616
 617                         if (unlikely(cluster & (cluster - 1)))
 618                                 cluster = 1 << __fls(cluster);
 619
 620                         hash_block_start &= ~(sector_t)(cluster - 1);
 621                         hash_block_end |= cluster - 1;
 622                         if (unlikely(hash_block_end >= v->hash_blocks))
 623                                 hash_block_end = v->hash_blocks - 1;
 624                 }
 625 no_prefetch_cluster:
 626                 dm_bufio_prefetch(v->bufio, hash_block_start,
 627                                   hash_block_end - hash_block_start + 1);
 628         }
 629
 630         kfree(pw);
 631 }
 632
 633 static void verity_submit_prefetch(struct dm_verity *v, struct dm_verity_io *io)
 634 {
 635         struct dm_verity_prefetch_work *pw;
 636
 637         pw = kmalloc(sizeof(struct dm_verity_prefetch_work),
 638                 GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
 639
 640         if (!pw)
 641                 return;
 642
 643         INIT_WORK(&pw->work, verity_prefetch_io);
 644         pw->v = v;
 645         pw->block = io->block;
 646         pw->n_blocks = io->n_blocks;
 647         queue_work(v->verify_wq, &pw->work);
 648 }
 649
 650 /*
 651  * Bio map function. It allocates dm_verity_io structure and bio vector and
 652  * fills them. Then it issues prefetches and the I/O.
 653  */
 654 static int verity_map(struct dm_target *ti, struct bio *bio)
 655 {
 656         struct dm_verity *v = ti->private;
 657         struct dm_verity_io *io;
 658
 659         bio_set_dev(bio, v->data_dev->bdev);
 660         bio->bi_iter.bi_sector = verity_map_sector(v, bio->bi_iter.bi_sector);
 661
 662         if (((unsigned)bio->bi_iter.bi_sector | bio_sectors(bio)) &
 663             ((1 << (v->data_dev_block_bits - SECTOR_SHIFT)) - 1)) {
 664                 DMERR_LIMIT("unaligned io");
 665                 return DM_MAPIO_KILL;
 666         }
 667
 668         if (bio_end_sector(bio) >>
 669             (v->data_dev_block_bits - SECTOR_SHIFT) > v->data_blocks) {
 670                 DMERR_LIMIT("io out of range");
 671                 return DM_MAPIO_KILL;
 672         }
 673
 674         if (bio_data_dir(bio) == WRITE)
 675                 return DM_MAPIO_KILL;
 676
 677         io = dm_per_bio_data(bio, ti->per_io_data_size);
 678         io->v = v;
 679         io->orig_bi_end_io = bio->bi_end_io;
 680         io->block = bio->bi_iter.bi_sector >> (v->data_dev_block_bits - SECTOR_SHIFT);
 681         io->n_blocks = bio->bi_iter.bi_size >> v->data_dev_block_bits;
 682
 683         bio->bi_end_io = verity_end_io;
 684         bio->bi_private = io;
 685         io->iter = bio->bi_iter;
 686
 687         verity_fec_init_io(io);
 688
 689         verity_submit_prefetch(v, io);
 690
 691         generic_make_request(bio);
 692
 693         return DM_MAPIO_SUBMITTED;
 694 }
 695
 696 /*
 697  * Status: V (valid) or C (corruption found)
 698  */
 699 static void verity_status(struct dm_target *ti, status_type_t type,
 700                           unsigned status_flags, char *result, unsigned maxlen)
 701 {
 702         struct dm_verity *v = ti->private;
 703         unsigned args = 0;
 704         unsigned sz = 0;
 705         unsigned x;
 706
 707         switch (type) {
 708         case STATUSTYPE_INFO:
 709                 DMEMIT("%c", v->hash_failed ? 'C' : 'V');
 710                 break;
 711         case STATUSTYPE_TABLE:
 712                 DMEMIT("%u %s %s %u %u %llu %llu %s ",
 713                         v->version,
 714                         v->data_dev->name,
 715                         v->hash_dev->name,
 716                         1 << v->data_dev_block_bits,
 717                         1 << v->hash_dev_block_bits,
 718                         (unsigned long long)v->data_blocks,
 719                         (unsigned long long)v->hash_start,
 720                         v->alg_name
 721                         );
 722                 for (x = 0; x < v->digest_size; x++)
 723                         DMEMIT("%02x", v->root_digest[x]);
 724                 DMEMIT(" ");
 725                 if (!v->salt_size)
 726                         DMEMIT("-");
 727                 else
 728                         for (x = 0; x < v->salt_size; x++)
 729                                 DMEMIT("%02x", v->salt[x]);
 730                 if (v->mode != DM_VERITY_MODE_EIO)
 731                         args++;
 732                 if (verity_fec_is_enabled(v))
 733                         args += DM_VERITY_OPTS_FEC;
 734                 if (v->zero_digest)
 735                         args++;
 736                 if (v->validated_blocks)
 737                         args++;
 738                 if (!args)
 739                         return;
 740                 DMEMIT(" %u", args);
 741                 if (v->mode != DM_VERITY_MODE_EIO) {
 742                         DMEMIT(" ");
 743                         switch (v->mode) {
 744                         case DM_VERITY_MODE_LOGGING:
 745                                 DMEMIT(DM_VERITY_OPT_LOGGING);
 746                                 break;
 747                         case DM_VERITY_MODE_RESTART:
 748                                 DMEMIT(DM_VERITY_OPT_RESTART);
 749                                 break;
 750                         default:
 751                                 BUG();
 752                         }
 753                 }
 754                 if (v->zero_digest)
 755                         DMEMIT(" " DM_VERITY_OPT_IGN_ZEROES);
 756                 if (v->validated_blocks)
 757                         DMEMIT(" " DM_VERITY_OPT_AT_MOST_ONCE);
 758                 sz = verity_fec_status_table(v, sz, result, maxlen);
 759                 break;
 760         }
 761 }
 762
 763 static int verity_prepare_ioctl(struct dm_target *ti, struct block_device **bdev)
 764 {
 765         struct dm_verity *v = ti->private;
 766
 767         *bdev = v->data_dev->bdev;
 768
 769         if (v->data_start ||
 770             ti->len != i_size_read(v->data_dev->bdev->bd_inode) >> SECTOR_SHIFT)
 771                 return 1;
 772         return 0;
 773 }
 774
 775 static int verity_iterate_devices(struct dm_target *ti,
 776                                   iterate_devices_callout_fn fn, void *data)
 777 {
 778         struct dm_verity *v = ti->private;
 779
 780         return fn(ti, v->data_dev, v->data_start, ti->len, data);
 781 }
 782
 783 static void verity_io_hints(struct dm_target *ti, struct queue_limits *limits)
 784 {
 785         struct dm_verity *v = ti->private;
 786
 787         if (limits->logical_block_size < 1 << v->data_dev_block_bits)
 788                 limits->logical_block_size = 1 << v->data_dev_block_bits;
 789
 790         if (limits->physical_block_size < 1 << v->data_dev_block_bits)
 791                 limits->physical_block_size = 1 << v->data_dev_block_bits;
 792
 793         blk_limits_io_min(limits, limits->logical_block_size);
 794 }
 795
 796 static void verity_dtr(struct dm_target *ti)
 797 {
 798         struct dm_verity *v = ti->private;
 799
 800         if (v->verify_wq)
 801                 destroy_workqueue(v->verify_wq);
 802
 803         if (v->bufio)
 804                 dm_bufio_client_destroy(v->bufio);
 805
 806         kvfree(v->validated_blocks);
 807         kfree(v->salt);
 808         kfree(v->root_digest);
 809         kfree(v->zero_digest);
 810
 811         if (v->tfm)
 812                 crypto_free_ahash(v->tfm);
 813
 814         kfree(v->alg_name);
 815
 816         if (v->hash_dev)
 817                 dm_put_device(ti, v->hash_dev);
 818
 819         if (v->data_dev)
 820                 dm_put_device(ti, v->data_dev);
 821
 822         verity_fec_dtr(v);
 823
 824         kfree(v);
 825 }
 826
 827 static int verity_alloc_most_once(struct dm_verity *v)
 828 {
 829         struct dm_target *ti = v->ti;
 830
 831         /* the bitset can only handle INT_MAX blocks */
 832         if (v->data_blocks > INT_MAX) {
 833                 ti->error = "device too large to use check_at_most_once";
 834                 return -E2BIG;
 835         }
 836
 837         v->validated_blocks = kvcalloc(BITS_TO_LONGS(v->data_blocks),
 838                                        sizeof(unsigned long),
 839                                        GFP_KERNEL);
 840         if (!v->validated_blocks) {
 841                 ti->error = "failed to allocate bitset for check_at_most_once";
 842                 return -ENOMEM;
 843         }
 844
 845         return 0;
 846 }
 847
 848 static int verity_alloc_zero_digest(struct dm_verity *v)
 849 {
 850         int r = -ENOMEM;
 851         struct ahash_request *req;
 852         u8 *zero_data;
 853
 854         v->zero_digest = kmalloc(v->digest_size, GFP_KERNEL);
 855
 856         if (!v->zero_digest)
 857                 return r;
 858
 859         req = kmalloc(v->ahash_reqsize, GFP_KERNEL);
 860
 861         if (!req)
 862                 return r; /* verity_dtr will free zero_digest */
 863
 864         zero_data = kzalloc(1 << v->data_dev_block_bits, GFP_KERNEL);
 865
 866         if (!zero_data)
 867                 goto out;
 868
 869         r = verity_hash(v, req, zero_data, 1 << v->data_dev_block_bits,
 870                         v->zero_digest);
 871
 872 out:
 873         kfree(req);
 874         kfree(zero_data);
 875
 876         return r;
 877 }
 878
 879 static int verity_parse_opt_args(struct dm_arg_set *as, struct dm_verity *v)
 880 {
 881         int r;
 882         unsigned argc;
 883         struct dm_target *ti = v->ti;
 884         const char *arg_name;
 885
 886         static const struct dm_arg _args[] = {
 887                 {0, DM_VERITY_OPTS_MAX, "Invalid number of feature args"},
 888         };
 889
 890         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 891         if (r)
 892                 return -EINVAL;
 893
 894         if (!argc)
 895                 return 0;
 896
 897         do {
 898                 arg_name = dm_shift_arg(as);
 899                 argc--;
 900
 901                 if (!strcasecmp(arg_name, DM_VERITY_OPT_LOGGING)) {
 902                         v->mode = DM_VERITY_MODE_LOGGING;
 903                         continue;
 904
 905                 } else if (!strcasecmp(arg_name, DM_VERITY_OPT_RESTART)) {
 906                         v->mode = DM_VERITY_MODE_RESTART;
 907                         continue;
 908
 909                 } else if (!strcasecmp(arg_name, DM_VERITY_OPT_IGN_ZEROES)) {
 910                         r = verity_alloc_zero_digest(v);
 911                         if (r) {
 912                                 ti->error = "Cannot allocate zero digest";
 913                                 return r;
 914                         }
 915                         continue;
 916
 917                 } else if (!strcasecmp(arg_name, DM_VERITY_OPT_AT_MOST_ONCE)) {
 918                         r = verity_alloc_most_once(v);
 919                         if (r)
 920                                 return r;
 921                         continue;
 922
 923                 } else if (verity_is_fec_opt_arg(arg_name)) {
 924                         r = verity_fec_parse_opt_args(as, v, &argc, arg_name);
 925                         if (r)
 926                                 return r;
 927                         continue;
 928                 }
 929
 930                 ti->error = "Unrecognized verity feature request";
 931                 return -EINVAL;
 932         } while (argc && !r);
 933
 934         return r;
 935 }
 936
 937 /*
 938  * Target parameters:
 939  *      <version>       The current format is version 1.
 940  *                      Vsn 0 is compatible with original Chromium OS releases.
 941  *      <data device>
 942  *      <hash device>
 943  *      <data block size>
 944  *      <hash block size>
 945  *      <the number of data blocks>
 946  *      <hash start block>
 947  *      <algorithm>
 948  *      <digest>
 949  *      <salt>          Hex string or "-" if no salt.
 950  */
 951 static int verity_ctr(struct dm_target *ti, unsigned argc, char **argv)
 952 {
 953         struct dm_verity *v;
 954         struct dm_arg_set as;
 955         unsigned int num;
 956         unsigned long long num_ll;
 957         int r;
 958         int i;
 959         sector_t hash_position;
 960         char dummy;
 961
 962         v = kzalloc(sizeof(struct dm_verity), GFP_KERNEL);
 963         if (!v) {
 964                 ti->error = "Cannot allocate verity structure";
 965                 return -ENOMEM;
 966         }
 967         ti->private = v;
 968         v->ti = ti;
 969
 970         r = verity_fec_ctr_alloc(v);
 971         if (r)
 972                 goto bad;
 973
 974         if ((dm_table_get_mode(ti->table) & ~FMODE_READ)) {
 975                 ti->error = "Device must be readonly";
 976                 r = -EINVAL;
 977                 goto bad;
 978         }
 979
 980         if (argc < 10) {
 981                 ti->error = "Not enough arguments";
 982                 r = -EINVAL;
 983                 goto bad;
 984         }
 985
 986         if (sscanf(argv[0], "%u%c", &num, &dummy) != 1 ||
 987             num > 1) {
 988                 ti->error = "Invalid version";
 989                 r = -EINVAL;
 990                 goto bad;
 991         }
 992         v->version = num;
 993
 994         r = dm_get_device(ti, argv[1], FMODE_READ, &v->data_dev);
 995         if (r) {
 996                 ti->error = "Data device lookup failed";
 997                 goto bad;
 998         }
 999
1000         r = dm_get_device(ti, argv[2], FMODE_READ, &v->hash_dev);
1001         if (r) {
1002                 ti->error = "Hash device lookup failed";
1003                 goto bad;
1004         }
1005
1006         if (sscanf(argv[3], "%u%c", &num, &dummy) != 1 ||
1007             !num || (num & (num - 1)) ||
1008             num < bdev_logical_block_size(v->data_dev->bdev) ||
1009             num > PAGE_SIZE) {
1010                 ti->error = "Invalid data device block size";
1011                 r = -EINVAL;
1012                 goto bad;
1013         }
1014         v->data_dev_block_bits = __ffs(num);
1015
1016         if (sscanf(argv[4], "%u%c", &num, &dummy) != 1 ||
1017             !num || (num & (num - 1)) ||
1018             num < bdev_logical_block_size(v->hash_dev->bdev) ||
1019             num > INT_MAX) {
1020                 ti->error = "Invalid hash device block size";
1021                 r = -EINVAL;
1022                 goto bad;
1023         }
1024         v->hash_dev_block_bits = __ffs(num);
1025
1026         if (sscanf(argv[5], "%llu%c", &num_ll, &dummy) != 1 ||
1027             (sector_t)(num_ll << (v->data_dev_block_bits - SECTOR_SHIFT))
1028             >> (v->data_dev_block_bits - SECTOR_SHIFT) != num_ll) {
1029                 ti->error = "Invalid data blocks";
1030                 r = -EINVAL;
1031                 goto bad;
1032         }
1033         v->data_blocks = num_ll;
1034
1035         if (ti->len > (v->data_blocks << (v->data_dev_block_bits - SECTOR_SHIFT))) {
1036                 ti->error = "Data device is too small";
1037                 r = -EINVAL;
1038                 goto bad;
1039         }
1040
1041         if (sscanf(argv[6], "%llu%c", &num_ll, &dummy) != 1 ||
1042             (sector_t)(num_ll << (v->hash_dev_block_bits - SECTOR_SHIFT))
1043             >> (v->hash_dev_block_bits - SECTOR_SHIFT) != num_ll) {
1044                 ti->error = "Invalid hash start";
1045                 r = -EINVAL;
1046                 goto bad;
1047         }
1048         v->hash_start = num_ll;
1049
1050         v->alg_name = kstrdup(argv[7], GFP_KERNEL);
1051         if (!v->alg_name) {
1052                 ti->error = "Cannot allocate algorithm name";
1053                 r = -ENOMEM;
1054                 goto bad;
1055         }
1056
1057         v->tfm = crypto_alloc_ahash(v->alg_name, 0, 0);
1058         if (IS_ERR(v->tfm)) {
1059                 ti->error = "Cannot initialize hash function";
1060                 r = PTR_ERR(v->tfm);
1061                 v->tfm = NULL;
1062                 goto bad;
1063         }
1064         v->digest_size = crypto_ahash_digestsize(v->tfm);
1065         if ((1 << v->hash_dev_block_bits) < v->digest_size * 2) {
1066                 ti->error = "Digest size too big";
1067                 r = -EINVAL;
1068                 goto bad;
1069         }
1070         v->ahash_reqsize = sizeof(struct ahash_request) +
1071                 crypto_ahash_reqsize(v->tfm);
1072
1073         v->root_digest = kmalloc(v->digest_size, GFP_KERNEL);
1074         if (!v->root_digest) {
1075                 ti->error = "Cannot allocate root digest";
1076                 r = -ENOMEM;
1077                 goto bad;
1078         }
1079         if (strlen(argv[8]) != v->digest_size * 2 ||
1080             hex2bin(v->root_digest, argv[8], v->digest_size)) {
1081                 ti->error = "Invalid root digest";
1082                 r = -EINVAL;
1083                 goto bad;
1084         }
1085
1086         if (strcmp(argv[9], "-")) {
1087                 v->salt_size = strlen(argv[9]) / 2;
1088                 v->salt = kmalloc(v->salt_size, GFP_KERNEL);
1089                 if (!v->salt) {
1090                         ti->error = "Cannot allocate salt";
1091                         r = -ENOMEM;
1092                         goto bad;
1093                 }
1094                 if (strlen(argv[9]) != v->salt_size * 2 ||
1095                     hex2bin(v->salt, argv[9], v->salt_size)) {
1096                         ti->error = "Invalid salt";
1097                         r = -EINVAL;
1098                         goto bad;
1099                 }
1100         }
1101
1102         argv += 10;
1103         argc -= 10;
1104
1105         /* Optional parameters */
1106         if (argc) {
1107                 as.argc = argc;
1108                 as.argv = argv;
1109
1110                 r = verity_parse_opt_args(&as, v);
1111                 if (r < 0)
1112                         goto bad;
1113         }
1114
1115         v->hash_per_block_bits =
1116                 __fls((1 << v->hash_dev_block_bits) / v->digest_size);
1117
1118         v->levels = 0;
1119         if (v->data_blocks)
1120                 while (v->hash_per_block_bits * v->levels < 64 &&
1121                        (unsigned long long)(v->data_blocks - 1) >>
1122                        (v->hash_per_block_bits * v->levels))
1123                         v->levels++;
1124
1125         if (v->levels > DM_VERITY_MAX_LEVELS) {
1126                 ti->error = "Too many tree levels";
1127                 r = -E2BIG;
1128                 goto bad;
1129         }
1130
1131         hash_position = v->hash_start;
1132         for (i = v->levels - 1; i >= 0; i--) {
1133                 sector_t s;
1134                 v->hash_level_block[i] = hash_position;
1135                 s = (v->data_blocks + ((sector_t)1 << ((i + 1) * v->hash_per_block_bits)) - 1)
1136                                         >> ((i + 1) * v->hash_per_block_bits);
1137                 if (hash_position + s < hash_position) {
1138                         ti->error = "Hash device offset overflow";
1139                         r = -E2BIG;
1140                         goto bad;
1141                 }
1142                 hash_position += s;
1143         }
1144         v->hash_blocks = hash_position;
1145
1146         v->bufio = dm_bufio_client_create(v->hash_dev->bdev,
1147                 1 << v->hash_dev_block_bits, 1, sizeof(struct buffer_aux),
1148                 dm_bufio_alloc_callback, NULL);
1149         if (IS_ERR(v->bufio)) {
1150                 ti->error = "Cannot initialize dm-bufio";
1151                 r = PTR_ERR(v->bufio);
1152                 v->bufio = NULL;
1153                 goto bad;
1154         }
1155
1156         if (dm_bufio_get_device_size(v->bufio) < v->hash_blocks) {
1157                 ti->error = "Hash device is too small";
1158                 r = -E2BIG;
1159                 goto bad;
1160         }
1161
1162         /* WQ_UNBOUND greatly improves performance when running on ramdisk */
1163         v->verify_wq = alloc_workqueue("kverityd", WQ_CPU_INTENSIVE | WQ_MEM_RECLAIM | WQ_UNBOUND, num_online_cpus());
1164         if (!v->verify_wq) {
1165                 ti->error = "Cannot allocate workqueue";
1166                 r = -ENOMEM;
1167                 goto bad;
1168         }
1169
1170         ti->per_io_data_size = sizeof(struct dm_verity_io) +
1171                                 v->ahash_reqsize + v->digest_size * 2;
1172
1173         r = verity_fec_ctr(v);
1174         if (r)
1175                 goto bad;
1176
1177         ti->per_io_data_size = roundup(ti->per_io_data_size,
1178                                        __alignof__(struct dm_verity_io));
1179
1180         return 0;
1181
1182 bad:
1183         verity_dtr(ti);
1184
1185         return r;
1186 }
1187
1188 static struct target_type verity_target = {
1189         .name           = "verity",
1190         .features       = DM_TARGET_IMMUTABLE,
1191         .version        = {1, 4, 0},
1192         .module         = THIS_MODULE,
1193         .ctr            = verity_ctr,
1194         .dtr            = verity_dtr,
1195         .map            = verity_map,
1196         .status         = verity_status,
1197         .prepare_ioctl  = verity_prepare_ioctl,
1198         .iterate_devices = verity_iterate_devices,
1199         .io_hints       = verity_io_hints,
1200 };
1201
1202 static int __init dm_verity_init(void)
1203 {
1204         int r;
1205
1206         r = dm_register_target(&verity_target);
1207         if (r < 0)
1208                 DMERR("register failed %d", r);
1209
1210         return r;
1211 }
1212
1213 static void __exit dm_verity_exit(void)
1214 {
1215         dm_unregister_target(&verity_target);
1216 }
1217
1218 module_init(dm_verity_init);
1219 module_exit(dm_verity_exit);
1220
1221 MODULE_AUTHOR("Mikulas Patocka <mpatocka@redhat.com>");
1222 MODULE_AUTHOR("Mandeep Baines <msb@chromium.org>");
1223 MODULE_AUTHOR("Will Drewry <wad@chromium.org>");
1224 MODULE_DESCRIPTION(DM_NAME " target for transparent disk integrity checking");
1225 MODULE_LICENSE("GPL");