fs/erofs/zmap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2018-2019 HUAWEI, Inc.
   4  *             https://www.huawei.com/
   5  */
   6 #include "internal.h"
   7 #include <asm/unaligned.h>
   8 #include <trace/events/erofs.h>
   9
  10 static int z_erofs_do_map_blocks(struct inode *inode,
  11                                  struct erofs_map_blocks *map,
  12                                  int flags);
  13
  14 int z_erofs_fill_inode(struct inode *inode)
  15 {
  16         struct erofs_inode *const vi = EROFS_I(inode);
  17         struct erofs_sb_info *sbi = EROFS_SB(inode->i_sb);
  18
  19         if (!erofs_sb_has_big_pcluster(sbi) &&
  20             !erofs_sb_has_ztailpacking(sbi) && !erofs_sb_has_fragments(sbi) &&
  21             vi->datalayout == EROFS_INODE_FLAT_COMPRESSION_LEGACY) {
  22                 vi->z_advise = 0;
  23                 vi->z_algorithmtype[0] = 0;
  24                 vi->z_algorithmtype[1] = 0;
  25                 vi->z_logical_clusterbits = LOG_BLOCK_SIZE;
  26                 set_bit(EROFS_I_Z_INITED_BIT, &vi->flags);
  27         }
  28         inode->i_mapping->a_ops = &z_erofs_aops;
  29         return 0;
  30 }
  31
  32 static int z_erofs_fill_inode_lazy(struct inode *inode)
  33 {
  34         struct erofs_inode *const vi = EROFS_I(inode);
  35         struct super_block *const sb = inode->i_sb;
  36         int err, headnr;
  37         erofs_off_t pos;
  38         struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
  39         void *kaddr;
  40         struct z_erofs_map_header *h;
  41
  42         if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags)) {
  43                 /*
  44                  * paired with smp_mb() at the end of the function to ensure
  45                  * fields will only be observed after the bit is set.
  46                  */
  47                 smp_mb();
  48                 return 0;
  49         }
  50
  51         if (wait_on_bit_lock(&vi->flags, EROFS_I_BL_Z_BIT, TASK_KILLABLE))
  52                 return -ERESTARTSYS;
  53
  54         err = 0;
  55         if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags))
  56                 goto out_unlock;
  57
  58         pos = ALIGN(erofs_iloc(inode) + vi->inode_isize + vi->xattr_isize, 8);
  59         kaddr = erofs_read_metabuf(&buf, sb, erofs_blknr(pos), EROFS_KMAP);
  60         if (IS_ERR(kaddr)) {
  61                 err = PTR_ERR(kaddr);
  62                 goto out_unlock;
  63         }
  64
  65         h = kaddr + erofs_blkoff(pos);
  66         /*
  67          * if the highest bit of the 8-byte map header is set, the whole file
  68          * is stored in the packed inode. The rest bits keeps z_fragmentoff.
  69          */
  70         if (h->h_clusterbits >> Z_EROFS_FRAGMENT_INODE_BIT) {
  71                 vi->z_advise = Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
  72                 vi->z_fragmentoff = le64_to_cpu(*(__le64 *)h) ^ (1ULL << 63);
  73                 vi->z_tailextent_headlcn = 0;
  74                 goto done;
  75         }
  76         vi->z_advise = le16_to_cpu(h->h_advise);
  77         vi->z_algorithmtype[0] = h->h_algorithmtype & 15;
  78         vi->z_algorithmtype[1] = h->h_algorithmtype >> 4;
  79
  80         headnr = 0;
  81         if (vi->z_algorithmtype[0] >= Z_EROFS_COMPRESSION_MAX ||
  82             vi->z_algorithmtype[++headnr] >= Z_EROFS_COMPRESSION_MAX) {
  83                 erofs_err(sb, "unknown HEAD%u format %u for nid %llu, please upgrade kernel",
  84                           headnr + 1, vi->z_algorithmtype[headnr], vi->nid);
  85                 err = -EOPNOTSUPP;
  86                 goto out_put_metabuf;
  87         }
  88
  89         vi->z_logical_clusterbits = LOG_BLOCK_SIZE + (h->h_clusterbits & 7);
  90         if (!erofs_sb_has_big_pcluster(EROFS_SB(sb)) &&
  91             vi->z_advise & (Z_EROFS_ADVISE_BIG_PCLUSTER_1 |
  92                             Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
  93                 erofs_err(sb, "per-inode big pcluster without sb feature for nid %llu",
  94                           vi->nid);
  95                 err = -EFSCORRUPTED;
  96                 goto out_put_metabuf;
  97         }
  98         if (vi->datalayout == EROFS_INODE_FLAT_COMPRESSION &&
  99             !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1) ^
 100             !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
 101                 erofs_err(sb, "big pcluster head1/2 of compact indexes should be consistent for nid %llu",
 102                           vi->nid);
 103                 err = -EFSCORRUPTED;
 104                 goto out_put_metabuf;
 105         }
 106
 107         if (vi->z_advise & Z_EROFS_ADVISE_INLINE_PCLUSTER) {
 108                 struct erofs_map_blocks map = {
 109                         .buf = __EROFS_BUF_INITIALIZER
 110                 };
 111
 112                 vi->z_idata_size = le16_to_cpu(h->h_idata_size);
 113                 err = z_erofs_do_map_blocks(inode, &map,
 114                                             EROFS_GET_BLOCKS_FINDTAIL);
 115                 erofs_put_metabuf(&map.buf);
 116
 117                 if (!map.m_plen ||
 118                     erofs_blkoff(map.m_pa) + map.m_plen > EROFS_BLKSIZ) {
 119                         erofs_err(sb, "invalid tail-packing pclustersize %llu",
 120                                   map.m_plen);
 121                         err = -EFSCORRUPTED;
 122                 }
 123                 if (err < 0)
 124                         goto out_put_metabuf;
 125         }
 126
 127         if (vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER &&
 128             !(h->h_clusterbits >> Z_EROFS_FRAGMENT_INODE_BIT)) {
 129                 struct erofs_map_blocks map = {
 130                         .buf = __EROFS_BUF_INITIALIZER
 131                 };
 132
 133                 vi->z_fragmentoff = le32_to_cpu(h->h_fragmentoff);
 134                 err = z_erofs_do_map_blocks(inode, &map,
 135                                             EROFS_GET_BLOCKS_FINDTAIL);
 136                 erofs_put_metabuf(&map.buf);
 137                 if (err < 0)
 138                         goto out_put_metabuf;
 139         }
 140 done:
 141         /* paired with smp_mb() at the beginning of the function */
 142         smp_mb();
 143         set_bit(EROFS_I_Z_INITED_BIT, &vi->flags);
 144 out_put_metabuf:
 145         erofs_put_metabuf(&buf);
 146 out_unlock:
 147         clear_and_wake_up_bit(EROFS_I_BL_Z_BIT, &vi->flags);
 148         return err;
 149 }
 150
 151 struct z_erofs_maprecorder {
 152         struct inode *inode;
 153         struct erofs_map_blocks *map;
 154         void *kaddr;
 155
 156         unsigned long lcn;
 157         /* compression extent information gathered */
 158         u8  type, headtype;
 159         u16 clusterofs;
 160         u16 delta[2];
 161         erofs_blk_t pblk, compressedblks;
 162         erofs_off_t nextpackoff;
 163         bool partialref;
 164 };
 165
 166 static int legacy_load_cluster_from_disk(struct z_erofs_maprecorder *m,
 167                                          unsigned long lcn)
 168 {
 169         struct inode *const inode = m->inode;
 170         struct erofs_inode *const vi = EROFS_I(inode);
 171         const erofs_off_t pos =
 172                 Z_EROFS_VLE_LEGACY_INDEX_ALIGN(erofs_iloc(inode) +
 173                                 vi->inode_isize + vi->xattr_isize) +
 174                 lcn * sizeof(struct z_erofs_vle_decompressed_index);
 175         struct z_erofs_vle_decompressed_index *di;
 176         unsigned int advise, type;
 177
 178         m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
 179                                       erofs_blknr(pos), EROFS_KMAP_ATOMIC);
 180         if (IS_ERR(m->kaddr))
 181                 return PTR_ERR(m->kaddr);
 182
 183         m->nextpackoff = pos + sizeof(struct z_erofs_vle_decompressed_index);
 184         m->lcn = lcn;
 185         di = m->kaddr + erofs_blkoff(pos);
 186
 187         advise = le16_to_cpu(di->di_advise);
 188         type = (advise >> Z_EROFS_VLE_DI_CLUSTER_TYPE_BIT) &
 189                 ((1 << Z_EROFS_VLE_DI_CLUSTER_TYPE_BITS) - 1);
 190         switch (type) {
 191         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
 192                 m->clusterofs = 1 << vi->z_logical_clusterbits;
 193                 m->delta[0] = le16_to_cpu(di->di_u.delta[0]);
 194                 if (m->delta[0] & Z_EROFS_VLE_DI_D0_CBLKCNT) {
 195                         if (!(vi->z_advise & (Z_EROFS_ADVISE_BIG_PCLUSTER_1 |
 196                                         Z_EROFS_ADVISE_BIG_PCLUSTER_2))) {
 197                                 DBG_BUGON(1);
 198                                 return -EFSCORRUPTED;
 199                         }
 200                         m->compressedblks = m->delta[0] &
 201                                 ~Z_EROFS_VLE_DI_D0_CBLKCNT;
 202                         m->delta[0] = 1;
 203                 }
 204                 m->delta[1] = le16_to_cpu(di->di_u.delta[1]);
 205                 break;
 206         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
 207         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD1:
 208         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD2:
 209                 if (advise & Z_EROFS_VLE_DI_PARTIAL_REF)
 210                         m->partialref = true;
 211                 m->clusterofs = le16_to_cpu(di->di_clusterofs);
 212                 if (m->clusterofs >= 1 << vi->z_logical_clusterbits) {
 213                         DBG_BUGON(1);
 214                         return -EFSCORRUPTED;
 215                 }
 216                 m->pblk = le32_to_cpu(di->di_u.blkaddr);
 217                 break;
 218         default:
 219                 DBG_BUGON(1);
 220                 return -EOPNOTSUPP;
 221         }
 222         m->type = type;
 223         return 0;
 224 }
 225
 226 static unsigned int decode_compactedbits(unsigned int lobits,
 227                                          unsigned int lomask,
 228                                          u8 *in, unsigned int pos, u8 *type)
 229 {
 230         const unsigned int v = get_unaligned_le32(in + pos / 8) >> (pos & 7);
 231         const unsigned int lo = v & lomask;
 232
 233         *type = (v >> lobits) & 3;
 234         return lo;
 235 }
 236
 237 static int get_compacted_la_distance(unsigned int lclusterbits,
 238                                      unsigned int encodebits,
 239                                      unsigned int vcnt, u8 *in, int i)
 240 {
 241         const unsigned int lomask = (1 << lclusterbits) - 1;
 242         unsigned int lo, d1 = 0;
 243         u8 type;
 244
 245         DBG_BUGON(i >= vcnt);
 246
 247         do {
 248                 lo = decode_compactedbits(lclusterbits, lomask,
 249                                           in, encodebits * i, &type);
 250
 251                 if (type != Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD)
 252                         return d1;
 253                 ++d1;
 254         } while (++i < vcnt);
 255
 256         /* vcnt - 1 (Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD) item */
 257         if (!(lo & Z_EROFS_VLE_DI_D0_CBLKCNT))
 258                 d1 += lo - 1;
 259         return d1;
 260 }
 261
 262 static int unpack_compacted_index(struct z_erofs_maprecorder *m,
 263                                   unsigned int amortizedshift,
 264                                   erofs_off_t pos, bool lookahead)
 265 {
 266         struct erofs_inode *const vi = EROFS_I(m->inode);
 267         const unsigned int lclusterbits = vi->z_logical_clusterbits;
 268         const unsigned int lomask = (1 << lclusterbits) - 1;
 269         unsigned int vcnt, base, lo, encodebits, nblk, eofs;
 270         int i;
 271         u8 *in, type;
 272         bool big_pcluster;
 273
 274         if (1 << amortizedshift == 4 && lclusterbits <= 14)
 275                 vcnt = 2;
 276         else if (1 << amortizedshift == 2 && lclusterbits == 12)
 277                 vcnt = 16;
 278         else
 279                 return -EOPNOTSUPP;
 280
 281         /* it doesn't equal to round_up(..) */
 282         m->nextpackoff = round_down(pos, vcnt << amortizedshift) +
 283                          (vcnt << amortizedshift);
 284         big_pcluster = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1;
 285         encodebits = ((vcnt << amortizedshift) - sizeof(__le32)) * 8 / vcnt;
 286         eofs = erofs_blkoff(pos);
 287         base = round_down(eofs, vcnt << amortizedshift);
 288         in = m->kaddr + base;
 289
 290         i = (eofs - base) >> amortizedshift;
 291
 292         lo = decode_compactedbits(lclusterbits, lomask,
 293                                   in, encodebits * i, &type);
 294         m->type = type;
 295         if (type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD) {
 296                 m->clusterofs = 1 << lclusterbits;
 297
 298                 /* figure out lookahead_distance: delta[1] if needed */
 299                 if (lookahead)
 300                         m->delta[1] = get_compacted_la_distance(lclusterbits,
 301                                                 encodebits, vcnt, in, i);
 302                 if (lo & Z_EROFS_VLE_DI_D0_CBLKCNT) {
 303                         if (!big_pcluster) {
 304                                 DBG_BUGON(1);
 305                                 return -EFSCORRUPTED;
 306                         }
 307                         m->compressedblks = lo & ~Z_EROFS_VLE_DI_D0_CBLKCNT;
 308                         m->delta[0] = 1;
 309                         return 0;
 310                 } else if (i + 1 != (int)vcnt) {
 311                         m->delta[0] = lo;
 312                         return 0;
 313                 }
 314                 /*
 315                  * since the last lcluster in the pack is special,
 316                  * of which lo saves delta[1] rather than delta[0].
 317                  * Hence, get delta[0] by the previous lcluster indirectly.
 318                  */
 319                 lo = decode_compactedbits(lclusterbits, lomask,
 320                                           in, encodebits * (i - 1), &type);
 321                 if (type != Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD)
 322                         lo = 0;
 323                 else if (lo & Z_EROFS_VLE_DI_D0_CBLKCNT)
 324                         lo = 1;
 325                 m->delta[0] = lo + 1;
 326                 return 0;
 327         }
 328         m->clusterofs = lo;
 329         m->delta[0] = 0;
 330         /* figout out blkaddr (pblk) for HEAD lclusters */
 331         if (!big_pcluster) {
 332                 nblk = 1;
 333                 while (i > 0) {
 334                         --i;
 335                         lo = decode_compactedbits(lclusterbits, lomask,
 336                                                   in, encodebits * i, &type);
 337                         if (type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD)
 338                                 i -= lo;
 339
 340                         if (i >= 0)
 341                                 ++nblk;
 342                 }
 343         } else {
 344                 nblk = 0;
 345                 while (i > 0) {
 346                         --i;
 347                         lo = decode_compactedbits(lclusterbits, lomask,
 348                                                   in, encodebits * i, &type);
 349                         if (type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD) {
 350                                 if (lo & Z_EROFS_VLE_DI_D0_CBLKCNT) {
 351                                         --i;
 352                                         nblk += lo & ~Z_EROFS_VLE_DI_D0_CBLKCNT;
 353                                         continue;
 354                                 }
 355                                 /* bigpcluster shouldn't have plain d0 == 1 */
 356                                 if (lo <= 1) {
 357                                         DBG_BUGON(1);
 358                                         return -EFSCORRUPTED;
 359                                 }
 360                                 i -= lo - 2;
 361                                 continue;
 362                         }
 363                         ++nblk;
 364                 }
 365         }
 366         in += (vcnt << amortizedshift) - sizeof(__le32);
 367         m->pblk = le32_to_cpu(*(__le32 *)in) + nblk;
 368         return 0;
 369 }
 370
 371 static int compacted_load_cluster_from_disk(struct z_erofs_maprecorder *m,
 372                                             unsigned long lcn, bool lookahead)
 373 {
 374         struct inode *const inode = m->inode;
 375         struct erofs_inode *const vi = EROFS_I(inode);
 376         const erofs_off_t ebase = sizeof(struct z_erofs_map_header) +
 377                 ALIGN(erofs_iloc(inode) + vi->inode_isize + vi->xattr_isize, 8);
 378         const unsigned int totalidx = DIV_ROUND_UP(inode->i_size, EROFS_BLKSIZ);
 379         unsigned int compacted_4b_initial, compacted_2b;
 380         unsigned int amortizedshift;
 381         erofs_off_t pos;
 382
 383         if (lcn >= totalidx)
 384                 return -EINVAL;
 385
 386         m->lcn = lcn;
 387         /* used to align to 32-byte (compacted_2b) alignment */
 388         compacted_4b_initial = (32 - ebase % 32) / 4;
 389         if (compacted_4b_initial == 32 / 4)
 390                 compacted_4b_initial = 0;
 391
 392         if ((vi->z_advise & Z_EROFS_ADVISE_COMPACTED_2B) &&
 393             compacted_4b_initial < totalidx)
 394                 compacted_2b = rounddown(totalidx - compacted_4b_initial, 16);
 395         else
 396                 compacted_2b = 0;
 397
 398         pos = ebase;
 399         if (lcn < compacted_4b_initial) {
 400                 amortizedshift = 2;
 401                 goto out;
 402         }
 403         pos += compacted_4b_initial * 4;
 404         lcn -= compacted_4b_initial;
 405
 406         if (lcn < compacted_2b) {
 407                 amortizedshift = 1;
 408                 goto out;
 409         }
 410         pos += compacted_2b * 2;
 411         lcn -= compacted_2b;
 412         amortizedshift = 2;
 413 out:
 414         pos += lcn * (1 << amortizedshift);
 415         m->kaddr = erofs_read_metabuf(&m->map->buf, inode->i_sb,
 416                                       erofs_blknr(pos), EROFS_KMAP_ATOMIC);
 417         if (IS_ERR(m->kaddr))
 418                 return PTR_ERR(m->kaddr);
 419         return unpack_compacted_index(m, amortizedshift, pos, lookahead);
 420 }
 421
 422 static int z_erofs_load_cluster_from_disk(struct z_erofs_maprecorder *m,
 423                                           unsigned int lcn, bool lookahead)
 424 {
 425         const unsigned int datamode = EROFS_I(m->inode)->datalayout;
 426
 427         if (datamode == EROFS_INODE_FLAT_COMPRESSION_LEGACY)
 428                 return legacy_load_cluster_from_disk(m, lcn);
 429
 430         if (datamode == EROFS_INODE_FLAT_COMPRESSION)
 431                 return compacted_load_cluster_from_disk(m, lcn, lookahead);
 432
 433         return -EINVAL;
 434 }
 435
 436 static int z_erofs_extent_lookback(struct z_erofs_maprecorder *m,
 437                                    unsigned int lookback_distance)
 438 {
 439         struct erofs_inode *const vi = EROFS_I(m->inode);
 440         const unsigned int lclusterbits = vi->z_logical_clusterbits;
 441
 442         while (m->lcn >= lookback_distance) {
 443                 unsigned long lcn = m->lcn - lookback_distance;
 444                 int err;
 445
 446                 /* load extent head logical cluster if needed */
 447                 err = z_erofs_load_cluster_from_disk(m, lcn, false);
 448                 if (err)
 449                         return err;
 450
 451                 switch (m->type) {
 452                 case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
 453                         if (!m->delta[0]) {
 454                                 erofs_err(m->inode->i_sb,
 455                                           "invalid lookback distance 0 @ nid %llu",
 456                                           vi->nid);
 457                                 DBG_BUGON(1);
 458                                 return -EFSCORRUPTED;
 459                         }
 460                         lookback_distance = m->delta[0];
 461                         continue;
 462                 case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
 463                 case Z_EROFS_VLE_CLUSTER_TYPE_HEAD1:
 464                 case Z_EROFS_VLE_CLUSTER_TYPE_HEAD2:
 465                         m->headtype = m->type;
 466                         m->map->m_la = (lcn << lclusterbits) | m->clusterofs;
 467                         return 0;
 468                 default:
 469                         erofs_err(m->inode->i_sb,
 470                                   "unknown type %u @ lcn %lu of nid %llu",
 471                                   m->type, lcn, vi->nid);
 472                         DBG_BUGON(1);
 473                         return -EOPNOTSUPP;
 474                 }
 475         }
 476
 477         erofs_err(m->inode->i_sb, "bogus lookback distance @ nid %llu",
 478                   vi->nid);
 479         DBG_BUGON(1);
 480         return -EFSCORRUPTED;
 481 }
 482
 483 static int z_erofs_get_extent_compressedlen(struct z_erofs_maprecorder *m,
 484                                             unsigned int initial_lcn)
 485 {
 486         struct erofs_inode *const vi = EROFS_I(m->inode);
 487         struct erofs_map_blocks *const map = m->map;
 488         const unsigned int lclusterbits = vi->z_logical_clusterbits;
 489         unsigned long lcn;
 490         int err;
 491
 492         DBG_BUGON(m->type != Z_EROFS_VLE_CLUSTER_TYPE_PLAIN &&
 493                   m->type != Z_EROFS_VLE_CLUSTER_TYPE_HEAD1 &&
 494                   m->type != Z_EROFS_VLE_CLUSTER_TYPE_HEAD2);
 495         DBG_BUGON(m->type != m->headtype);
 496
 497         if (m->headtype == Z_EROFS_VLE_CLUSTER_TYPE_PLAIN ||
 498             ((m->headtype == Z_EROFS_VLE_CLUSTER_TYPE_HEAD1) &&
 499              !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1)) ||
 500             ((m->headtype == Z_EROFS_VLE_CLUSTER_TYPE_HEAD2) &&
 501              !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_2))) {
 502                 map->m_plen = 1ULL << lclusterbits;
 503                 return 0;
 504         }
 505         lcn = m->lcn + 1;
 506         if (m->compressedblks)
 507                 goto out;
 508
 509         err = z_erofs_load_cluster_from_disk(m, lcn, false);
 510         if (err)
 511                 return err;
 512
 513         /*
 514          * If the 1st NONHEAD lcluster has already been handled initially w/o
 515          * valid compressedblks, which means at least it mustn't be CBLKCNT, or
 516          * an internal implemenatation error is detected.
 517          *
 518          * The following code can also handle it properly anyway, but let's
 519          * BUG_ON in the debugging mode only for developers to notice that.
 520          */
 521         DBG_BUGON(lcn == initial_lcn &&
 522                   m->type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD);
 523
 524         switch (m->type) {
 525         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
 526         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD1:
 527         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD2:
 528                 /*
 529                  * if the 1st NONHEAD lcluster is actually PLAIN or HEAD type
 530                  * rather than CBLKCNT, it's a 1 lcluster-sized pcluster.
 531                  */
 532                 m->compressedblks = 1 << (lclusterbits - LOG_BLOCK_SIZE);
 533                 break;
 534         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
 535                 if (m->delta[0] != 1)
 536                         goto err_bonus_cblkcnt;
 537                 if (m->compressedblks)
 538                         break;
 539                 fallthrough;
 540         default:
 541                 erofs_err(m->inode->i_sb,
 542                           "cannot found CBLKCNT @ lcn %lu of nid %llu",
 543                           lcn, vi->nid);
 544                 DBG_BUGON(1);
 545                 return -EFSCORRUPTED;
 546         }
 547 out:
 548         map->m_plen = (u64)m->compressedblks << LOG_BLOCK_SIZE;
 549         return 0;
 550 err_bonus_cblkcnt:
 551         erofs_err(m->inode->i_sb,
 552                   "bogus CBLKCNT @ lcn %lu of nid %llu",
 553                   lcn, vi->nid);
 554         DBG_BUGON(1);
 555         return -EFSCORRUPTED;
 556 }
 557
 558 static int z_erofs_get_extent_decompressedlen(struct z_erofs_maprecorder *m)
 559 {
 560         struct inode *inode = m->inode;
 561         struct erofs_inode *vi = EROFS_I(inode);
 562         struct erofs_map_blocks *map = m->map;
 563         unsigned int lclusterbits = vi->z_logical_clusterbits;
 564         u64 lcn = m->lcn, headlcn = map->m_la >> lclusterbits;
 565         int err;
 566
 567         do {
 568                 /* handle the last EOF pcluster (no next HEAD lcluster) */
 569                 if ((lcn << lclusterbits) >= inode->i_size) {
 570                         map->m_llen = inode->i_size - map->m_la;
 571                         return 0;
 572                 }
 573
 574                 err = z_erofs_load_cluster_from_disk(m, lcn, true);
 575                 if (err)
 576                         return err;
 577
 578                 if (m->type == Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD) {
 579                         DBG_BUGON(!m->delta[1] &&
 580                                   m->clusterofs != 1 << lclusterbits);
 581                 } else if (m->type == Z_EROFS_VLE_CLUSTER_TYPE_PLAIN ||
 582                            m->type == Z_EROFS_VLE_CLUSTER_TYPE_HEAD1 ||
 583                            m->type == Z_EROFS_VLE_CLUSTER_TYPE_HEAD2) {
 584                         /* go on until the next HEAD lcluster */
 585                         if (lcn != headlcn)
 586                                 break;
 587                         m->delta[1] = 1;
 588                 } else {
 589                         erofs_err(inode->i_sb, "unknown type %u @ lcn %llu of nid %llu",
 590                                   m->type, lcn, vi->nid);
 591                         DBG_BUGON(1);
 592                         return -EOPNOTSUPP;
 593                 }
 594                 lcn += m->delta[1];
 595         } while (m->delta[1]);
 596
 597         map->m_llen = (lcn << lclusterbits) + m->clusterofs - map->m_la;
 598         return 0;
 599 }
 600
 601 static int z_erofs_do_map_blocks(struct inode *inode,
 602                                  struct erofs_map_blocks *map,
 603                                  int flags)
 604 {
 605         struct erofs_inode *const vi = EROFS_I(inode);
 606         bool ztailpacking = vi->z_advise & Z_EROFS_ADVISE_INLINE_PCLUSTER;
 607         bool fragment = vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
 608         struct z_erofs_maprecorder m = {
 609                 .inode = inode,
 610                 .map = map,
 611         };
 612         int err = 0;
 613         unsigned int lclusterbits, endoff;
 614         unsigned long initial_lcn;
 615         unsigned long long ofs, end;
 616
 617         lclusterbits = vi->z_logical_clusterbits;
 618         ofs = flags & EROFS_GET_BLOCKS_FINDTAIL ? inode->i_size - 1 : map->m_la;
 619         initial_lcn = ofs >> lclusterbits;
 620         endoff = ofs & ((1 << lclusterbits) - 1);
 621
 622         err = z_erofs_load_cluster_from_disk(&m, initial_lcn, false);
 623         if (err)
 624                 goto unmap_out;
 625
 626         if (ztailpacking && (flags & EROFS_GET_BLOCKS_FINDTAIL))
 627                 vi->z_idataoff = m.nextpackoff;
 628
 629         map->m_flags = EROFS_MAP_MAPPED | EROFS_MAP_ENCODED;
 630         end = (m.lcn + 1ULL) << lclusterbits;
 631
 632         switch (m.type) {
 633         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
 634         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD1:
 635         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD2:
 636                 if (endoff >= m.clusterofs) {
 637                         m.headtype = m.type;
 638                         map->m_la = (m.lcn << lclusterbits) | m.clusterofs;
 639                         /*
 640                          * For ztailpacking files, in order to inline data more
 641                          * effectively, special EOF lclusters are now supported
 642                          * which can have three parts at most.
 643                          */
 644                         if (ztailpacking && end > inode->i_size)
 645                                 end = inode->i_size;
 646                         break;
 647                 }
 648                 /* m.lcn should be >= 1 if endoff < m.clusterofs */
 649                 if (!m.lcn) {
 650                         erofs_err(inode->i_sb,
 651                                   "invalid logical cluster 0 at nid %llu",
 652                                   vi->nid);
 653                         err = -EFSCORRUPTED;
 654                         goto unmap_out;
 655                 }
 656                 end = (m.lcn << lclusterbits) | m.clusterofs;
 657                 map->m_flags |= EROFS_MAP_FULL_MAPPED;
 658                 m.delta[0] = 1;
 659                 fallthrough;
 660         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
 661                 /* get the corresponding first chunk */
 662                 err = z_erofs_extent_lookback(&m, m.delta[0]);
 663                 if (err)
 664                         goto unmap_out;
 665                 break;
 666         default:
 667                 erofs_err(inode->i_sb,
 668                           "unknown type %u @ offset %llu of nid %llu",
 669                           m.type, ofs, vi->nid);
 670                 err = -EOPNOTSUPP;
 671                 goto unmap_out;
 672         }
 673         if (m.partialref)
 674                 map->m_flags |= EROFS_MAP_PARTIAL_REF;
 675         map->m_llen = end - map->m_la;
 676
 677         if (flags & EROFS_GET_BLOCKS_FINDTAIL) {
 678                 vi->z_tailextent_headlcn = m.lcn;
 679                 /* for non-compact indexes, fragmentoff is 64 bits */
 680                 if (fragment &&
 681                     vi->datalayout == EROFS_INODE_FLAT_COMPRESSION_LEGACY)
 682                         vi->z_fragmentoff |= (u64)m.pblk << 32;
 683         }
 684         if (ztailpacking && m.lcn == vi->z_tailextent_headlcn) {
 685                 map->m_flags |= EROFS_MAP_META;
 686                 map->m_pa = vi->z_idataoff;
 687                 map->m_plen = vi->z_idata_size;
 688         } else if (fragment && m.lcn == vi->z_tailextent_headlcn) {
 689                 map->m_flags |= EROFS_MAP_FRAGMENT;
 690         } else {
 691                 map->m_pa = blknr_to_addr(m.pblk);
 692                 err = z_erofs_get_extent_compressedlen(&m, initial_lcn);
 693                 if (err)
 694                         goto unmap_out;
 695         }
 696
 697         if (m.headtype == Z_EROFS_VLE_CLUSTER_TYPE_PLAIN) {
 698                 if (map->m_llen > map->m_plen) {
 699                         DBG_BUGON(1);
 700                         err = -EFSCORRUPTED;
 701                         goto unmap_out;
 702                 }
 703                 if (vi->z_advise & Z_EROFS_ADVISE_INTERLACED_PCLUSTER)
 704                         map->m_algorithmformat =
 705                                 Z_EROFS_COMPRESSION_INTERLACED;
 706                 else
 707                         map->m_algorithmformat =
 708                                 Z_EROFS_COMPRESSION_SHIFTED;
 709         } else if (m.headtype == Z_EROFS_VLE_CLUSTER_TYPE_HEAD2) {
 710                 map->m_algorithmformat = vi->z_algorithmtype[1];
 711         } else {
 712                 map->m_algorithmformat = vi->z_algorithmtype[0];
 713         }
 714
 715         if ((flags & EROFS_GET_BLOCKS_FIEMAP) ||
 716             ((flags & EROFS_GET_BLOCKS_READMORE) &&
 717              map->m_algorithmformat == Z_EROFS_COMPRESSION_LZMA &&
 718              map->m_llen >= EROFS_BLKSIZ)) {
 719                 err = z_erofs_get_extent_decompressedlen(&m);
 720                 if (!err)
 721                         map->m_flags |= EROFS_MAP_FULL_MAPPED;
 722         }
 723
 724 unmap_out:
 725         erofs_unmap_metabuf(&m.map->buf);
 726         erofs_dbg("%s, m_la %llu m_pa %llu m_llen %llu m_plen %llu m_flags 0%o",
 727                   __func__, map->m_la, map->m_pa,
 728                   map->m_llen, map->m_plen, map->m_flags);
 729         return err;
 730 }
 731
 732 int z_erofs_map_blocks_iter(struct inode *inode, struct erofs_map_blocks *map,
 733                             int flags)
 734 {
 735         struct erofs_inode *const vi = EROFS_I(inode);
 736         int err = 0;
 737
 738         trace_z_erofs_map_blocks_iter_enter(inode, map, flags);
 739
 740         /* when trying to read beyond EOF, leave it unmapped */
 741         if (map->m_la >= inode->i_size) {
 742                 map->m_llen = map->m_la + 1 - inode->i_size;
 743                 map->m_la = inode->i_size;
 744                 map->m_flags = 0;
 745                 goto out;
 746         }
 747
 748         err = z_erofs_fill_inode_lazy(inode);
 749         if (err)
 750                 goto out;
 751
 752         if ((vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER) &&
 753             !vi->z_tailextent_headlcn) {
 754                 map->m_la = 0;
 755                 map->m_llen = inode->i_size;
 756                 map->m_flags = EROFS_MAP_MAPPED | EROFS_MAP_FULL_MAPPED |
 757                                 EROFS_MAP_FRAGMENT;
 758                 goto out;
 759         }
 760
 761         err = z_erofs_do_map_blocks(inode, map, flags);
 762 out:
 763         trace_z_erofs_map_blocks_iter_exit(inode, map, flags, err);
 764
 765         /* aggressively BUG_ON iff CONFIG_EROFS_FS_DEBUG is on */
 766         DBG_BUGON(err < 0 && err != -ENOMEM);
 767         return err;
 768 }
 769
 770 static int z_erofs_iomap_begin_report(struct inode *inode, loff_t offset,
 771                                 loff_t length, unsigned int flags,
 772                                 struct iomap *iomap, struct iomap *srcmap)
 773 {
 774         int ret;
 775         struct erofs_map_blocks map = { .m_la = offset };
 776
 777         ret = z_erofs_map_blocks_iter(inode, &map, EROFS_GET_BLOCKS_FIEMAP);
 778         erofs_put_metabuf(&map.buf);
 779         if (ret < 0)
 780                 return ret;
 781
 782         iomap->bdev = inode->i_sb->s_bdev;
 783         iomap->offset = map.m_la;
 784         iomap->length = map.m_llen;
 785         if (map.m_flags & EROFS_MAP_MAPPED) {
 786                 iomap->type = IOMAP_MAPPED;
 787                 iomap->addr = map.m_flags & EROFS_MAP_FRAGMENT ?
 788                               IOMAP_NULL_ADDR : map.m_pa;
 789         } else {
 790                 iomap->type = IOMAP_HOLE;
 791                 iomap->addr = IOMAP_NULL_ADDR;
 792                 /*
 793                  * No strict rule on how to describe extents for post EOF, yet
 794                  * we need to do like below. Otherwise, iomap itself will get
 795                  * into an endless loop on post EOF.
 796                  *
 797                  * Calculate the effective offset by subtracting extent start
 798                  * (map.m_la) from the requested offset, and add it to length.
 799                  * (NB: offset >= map.m_la always)
 800                  */
 801                 if (iomap->offset >= inode->i_size)
 802                         iomap->length = length + offset - map.m_la;
 803         }
 804         iomap->flags = 0;
 805         return 0;
 806 }
 807
 808 const struct iomap_ops z_erofs_iomap_report_ops = {
 809         .iomap_begin = z_erofs_iomap_begin_report,
 810 };