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 <linux/unaligned.h>
   8 #include <trace/events/erofs.h>
   9
  10 struct z_erofs_maprecorder {
  11         struct inode *inode;
  12         struct erofs_map_blocks *map;
  13         unsigned long lcn;
  14         /* compression extent information gathered */
  15         u8  type, headtype;
  16         u16 clusterofs;
  17         u16 delta[2];
  18         erofs_blk_t pblk, compressedblks;
  19         erofs_off_t nextpackoff;
  20         bool partialref;
  21 };
  22
  23 static int z_erofs_load_full_lcluster(struct z_erofs_maprecorder *m,
  24                                       unsigned long lcn)
  25 {
  26         struct inode *const inode = m->inode;
  27         struct erofs_inode *const vi = EROFS_I(inode);
  28         const erofs_off_t pos = Z_EROFS_FULL_INDEX_ALIGN(erofs_iloc(inode) +
  29                         vi->inode_isize + vi->xattr_isize) +
  30                         lcn * sizeof(struct z_erofs_lcluster_index);
  31         struct z_erofs_lcluster_index *di;
  32         unsigned int advise;
  33
  34         di = erofs_read_metabuf(&m->map->buf, inode->i_sb, pos, EROFS_KMAP);
  35         if (IS_ERR(di))
  36                 return PTR_ERR(di);
  37         m->lcn = lcn;
  38         m->nextpackoff = pos + sizeof(struct z_erofs_lcluster_index);
  39
  40         advise = le16_to_cpu(di->di_advise);
  41         m->type = advise & Z_EROFS_LI_LCLUSTER_TYPE_MASK;
  42         if (m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
  43                 m->clusterofs = 1 << vi->z_logical_clusterbits;
  44                 m->delta[0] = le16_to_cpu(di->di_u.delta[0]);
  45                 if (m->delta[0] & Z_EROFS_LI_D0_CBLKCNT) {
  46                         if (!(vi->z_advise & (Z_EROFS_ADVISE_BIG_PCLUSTER_1 |
  47                                         Z_EROFS_ADVISE_BIG_PCLUSTER_2))) {
  48                                 DBG_BUGON(1);
  49                                 return -EFSCORRUPTED;
  50                         }
  51                         m->compressedblks = m->delta[0] & ~Z_EROFS_LI_D0_CBLKCNT;
  52                         m->delta[0] = 1;
  53                 }
  54                 m->delta[1] = le16_to_cpu(di->di_u.delta[1]);
  55         } else {
  56                 m->partialref = !!(advise & Z_EROFS_LI_PARTIAL_REF);
  57                 m->clusterofs = le16_to_cpu(di->di_clusterofs);
  58                 if (m->clusterofs >= 1 << vi->z_logical_clusterbits) {
  59                         DBG_BUGON(1);
  60                         return -EFSCORRUPTED;
  61                 }
  62                 m->pblk = le32_to_cpu(di->di_u.blkaddr);
  63         }
  64         return 0;
  65 }
  66
  67 static unsigned int decode_compactedbits(unsigned int lobits,
  68                                          u8 *in, unsigned int pos, u8 *type)
  69 {
  70         const unsigned int v = get_unaligned_le32(in + pos / 8) >> (pos & 7);
  71         const unsigned int lo = v & ((1 << lobits) - 1);
  72
  73         *type = (v >> lobits) & 3;
  74         return lo;
  75 }
  76
  77 static int get_compacted_la_distance(unsigned int lobits,
  78                                      unsigned int encodebits,
  79                                      unsigned int vcnt, u8 *in, int i)
  80 {
  81         unsigned int lo, d1 = 0;
  82         u8 type;
  83
  84         DBG_BUGON(i >= vcnt);
  85
  86         do {
  87                 lo = decode_compactedbits(lobits, in, encodebits * i, &type);
  88
  89                 if (type != Z_EROFS_LCLUSTER_TYPE_NONHEAD)
  90                         return d1;
  91                 ++d1;
  92         } while (++i < vcnt);
  93
  94         /* vcnt - 1 (Z_EROFS_LCLUSTER_TYPE_NONHEAD) item */
  95         if (!(lo & Z_EROFS_LI_D0_CBLKCNT))
  96                 d1 += lo - 1;
  97         return d1;
  98 }
  99
 100 static int z_erofs_load_compact_lcluster(struct z_erofs_maprecorder *m,
 101                                          unsigned long lcn, bool lookahead)
 102 {
 103         struct inode *const inode = m->inode;
 104         struct erofs_inode *const vi = EROFS_I(inode);
 105         const erofs_off_t ebase = sizeof(struct z_erofs_map_header) +
 106                 ALIGN(erofs_iloc(inode) + vi->inode_isize + vi->xattr_isize, 8);
 107         const unsigned int lclusterbits = vi->z_logical_clusterbits;
 108         const unsigned int totalidx = erofs_iblks(inode);
 109         unsigned int compacted_4b_initial, compacted_2b, amortizedshift;
 110         unsigned int vcnt, lo, lobits, encodebits, nblk, bytes;
 111         bool big_pcluster = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1;
 112         erofs_off_t pos;
 113         u8 *in, type;
 114         int i;
 115
 116         if (lcn >= totalidx || lclusterbits > 14)
 117                 return -EINVAL;
 118
 119         m->lcn = lcn;
 120         /* used to align to 32-byte (compacted_2b) alignment */
 121         compacted_4b_initial = ((32 - ebase % 32) / 4) & 7;
 122         compacted_2b = 0;
 123         if ((vi->z_advise & Z_EROFS_ADVISE_COMPACTED_2B) &&
 124             compacted_4b_initial < totalidx)
 125                 compacted_2b = rounddown(totalidx - compacted_4b_initial, 16);
 126
 127         pos = ebase;
 128         amortizedshift = 2;     /* compact_4b */
 129         if (lcn >= compacted_4b_initial) {
 130                 pos += compacted_4b_initial * 4;
 131                 lcn -= compacted_4b_initial;
 132                 if (lcn < compacted_2b) {
 133                         amortizedshift = 1;
 134                 } else {
 135                         pos += compacted_2b * 2;
 136                         lcn -= compacted_2b;
 137                 }
 138         }
 139         pos += lcn * (1 << amortizedshift);
 140
 141         /* figure out the lcluster count in this pack */
 142         if (1 << amortizedshift == 4 && lclusterbits <= 14)
 143                 vcnt = 2;
 144         else if (1 << amortizedshift == 2 && lclusterbits <= 12)
 145                 vcnt = 16;
 146         else
 147                 return -EOPNOTSUPP;
 148
 149         in = erofs_read_metabuf(&m->map->buf, m->inode->i_sb, pos, EROFS_KMAP);
 150         if (IS_ERR(in))
 151                 return PTR_ERR(in);
 152
 153         /* it doesn't equal to round_up(..) */
 154         m->nextpackoff = round_down(pos, vcnt << amortizedshift) +
 155                          (vcnt << amortizedshift);
 156         lobits = max(lclusterbits, ilog2(Z_EROFS_LI_D0_CBLKCNT) + 1U);
 157         encodebits = ((vcnt << amortizedshift) - sizeof(__le32)) * 8 / vcnt;
 158         bytes = pos & ((vcnt << amortizedshift) - 1);
 159         in -= bytes;
 160         i = bytes >> amortizedshift;
 161
 162         lo = decode_compactedbits(lobits, in, encodebits * i, &type);
 163         m->type = type;
 164         if (type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
 165                 m->clusterofs = 1 << lclusterbits;
 166
 167                 /* figure out lookahead_distance: delta[1] if needed */
 168                 if (lookahead)
 169                         m->delta[1] = get_compacted_la_distance(lobits,
 170                                                 encodebits, vcnt, in, i);
 171                 if (lo & Z_EROFS_LI_D0_CBLKCNT) {
 172                         if (!big_pcluster) {
 173                                 DBG_BUGON(1);
 174                                 return -EFSCORRUPTED;
 175                         }
 176                         m->compressedblks = lo & ~Z_EROFS_LI_D0_CBLKCNT;
 177                         m->delta[0] = 1;
 178                         return 0;
 179                 } else if (i + 1 != (int)vcnt) {
 180                         m->delta[0] = lo;
 181                         return 0;
 182                 }
 183                 /*
 184                  * since the last lcluster in the pack is special,
 185                  * of which lo saves delta[1] rather than delta[0].
 186                  * Hence, get delta[0] by the previous lcluster indirectly.
 187                  */
 188                 lo = decode_compactedbits(lobits, in,
 189                                           encodebits * (i - 1), &type);
 190                 if (type != Z_EROFS_LCLUSTER_TYPE_NONHEAD)
 191                         lo = 0;
 192                 else if (lo & Z_EROFS_LI_D0_CBLKCNT)
 193                         lo = 1;
 194                 m->delta[0] = lo + 1;
 195                 return 0;
 196         }
 197         m->clusterofs = lo;
 198         m->delta[0] = 0;
 199         /* figout out blkaddr (pblk) for HEAD lclusters */
 200         if (!big_pcluster) {
 201                 nblk = 1;
 202                 while (i > 0) {
 203                         --i;
 204                         lo = decode_compactedbits(lobits, in,
 205                                                   encodebits * i, &type);
 206                         if (type == Z_EROFS_LCLUSTER_TYPE_NONHEAD)
 207                                 i -= lo;
 208
 209                         if (i >= 0)
 210                                 ++nblk;
 211                 }
 212         } else {
 213                 nblk = 0;
 214                 while (i > 0) {
 215                         --i;
 216                         lo = decode_compactedbits(lobits, in,
 217                                                   encodebits * i, &type);
 218                         if (type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
 219                                 if (lo & Z_EROFS_LI_D0_CBLKCNT) {
 220                                         --i;
 221                                         nblk += lo & ~Z_EROFS_LI_D0_CBLKCNT;
 222                                         continue;
 223                                 }
 224                                 /* bigpcluster shouldn't have plain d0 == 1 */
 225                                 if (lo <= 1) {
 226                                         DBG_BUGON(1);
 227                                         return -EFSCORRUPTED;
 228                                 }
 229                                 i -= lo - 2;
 230                                 continue;
 231                         }
 232                         ++nblk;
 233                 }
 234         }
 235         in += (vcnt << amortizedshift) - sizeof(__le32);
 236         m->pblk = le32_to_cpu(*(__le32 *)in) + nblk;
 237         return 0;
 238 }
 239
 240 static int z_erofs_load_lcluster_from_disk(struct z_erofs_maprecorder *m,
 241                                            unsigned int lcn, bool lookahead)
 242 {
 243         switch (EROFS_I(m->inode)->datalayout) {
 244         case EROFS_INODE_COMPRESSED_FULL:
 245                 return z_erofs_load_full_lcluster(m, lcn);
 246         case EROFS_INODE_COMPRESSED_COMPACT:
 247                 return z_erofs_load_compact_lcluster(m, lcn, lookahead);
 248         default:
 249                 return -EINVAL;
 250         }
 251 }
 252
 253 static int z_erofs_extent_lookback(struct z_erofs_maprecorder *m,
 254                                    unsigned int lookback_distance)
 255 {
 256         struct super_block *sb = m->inode->i_sb;
 257         struct erofs_inode *const vi = EROFS_I(m->inode);
 258         const unsigned int lclusterbits = vi->z_logical_clusterbits;
 259
 260         while (m->lcn >= lookback_distance) {
 261                 unsigned long lcn = m->lcn - lookback_distance;
 262                 int err;
 263
 264                 err = z_erofs_load_lcluster_from_disk(m, lcn, false);
 265                 if (err)
 266                         return err;
 267
 268                 switch (m->type) {
 269                 case Z_EROFS_LCLUSTER_TYPE_NONHEAD:
 270                         lookback_distance = m->delta[0];
 271                         if (!lookback_distance)
 272                                 goto err_bogus;
 273                         continue;
 274                 case Z_EROFS_LCLUSTER_TYPE_PLAIN:
 275                 case Z_EROFS_LCLUSTER_TYPE_HEAD1:
 276                 case Z_EROFS_LCLUSTER_TYPE_HEAD2:
 277                         m->headtype = m->type;
 278                         m->map->m_la = (lcn << lclusterbits) | m->clusterofs;
 279                         return 0;
 280                 default:
 281                         erofs_err(sb, "unknown type %u @ lcn %lu of nid %llu",
 282                                   m->type, lcn, vi->nid);
 283                         DBG_BUGON(1);
 284                         return -EOPNOTSUPP;
 285                 }
 286         }
 287 err_bogus:
 288         erofs_err(sb, "bogus lookback distance %u @ lcn %lu of nid %llu",
 289                   lookback_distance, m->lcn, vi->nid);
 290         DBG_BUGON(1);
 291         return -EFSCORRUPTED;
 292 }
 293
 294 static int z_erofs_get_extent_compressedlen(struct z_erofs_maprecorder *m,
 295                                             unsigned int initial_lcn)
 296 {
 297         struct inode *inode = m->inode;
 298         struct super_block *sb = inode->i_sb;
 299         struct erofs_inode *vi = EROFS_I(inode);
 300         bool bigpcl1 = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1;
 301         bool bigpcl2 = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_2;
 302         unsigned long lcn = m->lcn + 1;
 303         int err;
 304
 305         DBG_BUGON(m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD);
 306         DBG_BUGON(m->type != m->headtype);
 307
 308         if ((m->headtype == Z_EROFS_LCLUSTER_TYPE_HEAD1 && !bigpcl1) ||
 309             ((m->headtype == Z_EROFS_LCLUSTER_TYPE_PLAIN ||
 310               m->headtype == Z_EROFS_LCLUSTER_TYPE_HEAD2) && !bigpcl2) ||
 311             (lcn << vi->z_logical_clusterbits) >= inode->i_size)
 312                 m->compressedblks = 1;
 313
 314         if (m->compressedblks)
 315                 goto out;
 316
 317         err = z_erofs_load_lcluster_from_disk(m, lcn, false);
 318         if (err)
 319                 return err;
 320
 321         /*
 322          * If the 1st NONHEAD lcluster has already been handled initially w/o
 323          * valid compressedblks, which means at least it mustn't be CBLKCNT, or
 324          * an internal implemenatation error is detected.
 325          *
 326          * The following code can also handle it properly anyway, but let's
 327          * BUG_ON in the debugging mode only for developers to notice that.
 328          */
 329         DBG_BUGON(lcn == initial_lcn &&
 330                   m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD);
 331
 332         switch (m->type) {
 333         case Z_EROFS_LCLUSTER_TYPE_PLAIN:
 334         case Z_EROFS_LCLUSTER_TYPE_HEAD1:
 335         case Z_EROFS_LCLUSTER_TYPE_HEAD2:
 336                 /*
 337                  * if the 1st NONHEAD lcluster is actually PLAIN or HEAD type
 338                  * rather than CBLKCNT, it's a 1 block-sized pcluster.
 339                  */
 340                 m->compressedblks = 1;
 341                 break;
 342         case Z_EROFS_LCLUSTER_TYPE_NONHEAD:
 343                 if (m->delta[0] != 1)
 344                         goto err_bonus_cblkcnt;
 345                 if (m->compressedblks)
 346                         break;
 347                 fallthrough;
 348         default:
 349                 erofs_err(sb, "cannot found CBLKCNT @ lcn %lu of nid %llu", lcn,
 350                           vi->nid);
 351                 DBG_BUGON(1);
 352                 return -EFSCORRUPTED;
 353         }
 354 out:
 355         m->map->m_plen = erofs_pos(sb, m->compressedblks);
 356         return 0;
 357 err_bonus_cblkcnt:
 358         erofs_err(sb, "bogus CBLKCNT @ lcn %lu of nid %llu", lcn, vi->nid);
 359         DBG_BUGON(1);
 360         return -EFSCORRUPTED;
 361 }
 362
 363 static int z_erofs_get_extent_decompressedlen(struct z_erofs_maprecorder *m)
 364 {
 365         struct inode *inode = m->inode;
 366         struct erofs_inode *vi = EROFS_I(inode);
 367         struct erofs_map_blocks *map = m->map;
 368         unsigned int lclusterbits = vi->z_logical_clusterbits;
 369         u64 lcn = m->lcn, headlcn = map->m_la >> lclusterbits;
 370         int err;
 371
 372         while (1) {
 373                 /* handle the last EOF pcluster (no next HEAD lcluster) */
 374                 if ((lcn << lclusterbits) >= inode->i_size) {
 375                         map->m_llen = inode->i_size - map->m_la;
 376                         return 0;
 377                 }
 378
 379                 err = z_erofs_load_lcluster_from_disk(m, lcn, true);
 380                 if (err)
 381                         return err;
 382
 383                 if (m->type == Z_EROFS_LCLUSTER_TYPE_NONHEAD) {
 384                         /* work around invalid d1 generated by pre-1.0 mkfs */
 385                         if (unlikely(!m->delta[1])) {
 386                                 m->delta[1] = 1;
 387                                 DBG_BUGON(1);
 388                         }
 389                 } else if (m->type == Z_EROFS_LCLUSTER_TYPE_PLAIN ||
 390                            m->type == Z_EROFS_LCLUSTER_TYPE_HEAD1 ||
 391                            m->type == Z_EROFS_LCLUSTER_TYPE_HEAD2) {
 392                         if (lcn != headlcn)
 393                                 break;  /* ends at the next HEAD lcluster */
 394                         m->delta[1] = 1;
 395                 } else {
 396                         erofs_err(inode->i_sb, "unknown type %u @ lcn %llu of nid %llu",
 397                                   m->type, lcn, vi->nid);
 398                         DBG_BUGON(1);
 399                         return -EOPNOTSUPP;
 400                 }
 401                 lcn += m->delta[1];
 402         }
 403         map->m_llen = (lcn << lclusterbits) + m->clusterofs - map->m_la;
 404         return 0;
 405 }
 406
 407 static int z_erofs_do_map_blocks(struct inode *inode,
 408                                  struct erofs_map_blocks *map, int flags)
 409 {
 410         struct erofs_inode *const vi = EROFS_I(inode);
 411         bool ztailpacking = vi->z_advise & Z_EROFS_ADVISE_INLINE_PCLUSTER;
 412         bool fragment = vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
 413         struct z_erofs_maprecorder m = {
 414                 .inode = inode,
 415                 .map = map,
 416         };
 417         int err = 0;
 418         unsigned int lclusterbits, endoff, afmt;
 419         unsigned long initial_lcn;
 420         unsigned long long ofs, end;
 421
 422         lclusterbits = vi->z_logical_clusterbits;
 423         ofs = flags & EROFS_GET_BLOCKS_FINDTAIL ? inode->i_size - 1 : map->m_la;
 424         initial_lcn = ofs >> lclusterbits;
 425         endoff = ofs & ((1 << lclusterbits) - 1);
 426
 427         err = z_erofs_load_lcluster_from_disk(&m, initial_lcn, false);
 428         if (err)
 429                 goto unmap_out;
 430
 431         if (ztailpacking && (flags & EROFS_GET_BLOCKS_FINDTAIL))
 432                 vi->z_idataoff = m.nextpackoff;
 433
 434         map->m_flags = EROFS_MAP_MAPPED | EROFS_MAP_ENCODED;
 435         end = (m.lcn + 1ULL) << lclusterbits;
 436
 437         switch (m.type) {
 438         case Z_EROFS_LCLUSTER_TYPE_PLAIN:
 439         case Z_EROFS_LCLUSTER_TYPE_HEAD1:
 440         case Z_EROFS_LCLUSTER_TYPE_HEAD2:
 441                 if (endoff >= m.clusterofs) {
 442                         m.headtype = m.type;
 443                         map->m_la = (m.lcn << lclusterbits) | m.clusterofs;
 444                         /*
 445                          * For ztailpacking files, in order to inline data more
 446                          * effectively, special EOF lclusters are now supported
 447                          * which can have three parts at most.
 448                          */
 449                         if (ztailpacking && end > inode->i_size)
 450                                 end = inode->i_size;
 451                         break;
 452                 }
 453                 /* m.lcn should be >= 1 if endoff < m.clusterofs */
 454                 if (!m.lcn) {
 455                         erofs_err(inode->i_sb,
 456                                   "invalid logical cluster 0 at nid %llu",
 457                                   vi->nid);
 458                         err = -EFSCORRUPTED;
 459                         goto unmap_out;
 460                 }
 461                 end = (m.lcn << lclusterbits) | m.clusterofs;
 462                 map->m_flags |= EROFS_MAP_FULL_MAPPED;
 463                 m.delta[0] = 1;
 464                 fallthrough;
 465         case Z_EROFS_LCLUSTER_TYPE_NONHEAD:
 466                 /* get the corresponding first chunk */
 467                 err = z_erofs_extent_lookback(&m, m.delta[0]);
 468                 if (err)
 469                         goto unmap_out;
 470                 break;
 471         default:
 472                 erofs_err(inode->i_sb,
 473                           "unknown type %u @ offset %llu of nid %llu",
 474                           m.type, ofs, vi->nid);
 475                 err = -EOPNOTSUPP;
 476                 goto unmap_out;
 477         }
 478         if (m.partialref)
 479                 map->m_flags |= EROFS_MAP_PARTIAL_REF;
 480         map->m_llen = end - map->m_la;
 481
 482         if (flags & EROFS_GET_BLOCKS_FINDTAIL) {
 483                 vi->z_tailextent_headlcn = m.lcn;
 484                 /* for non-compact indexes, fragmentoff is 64 bits */
 485                 if (fragment && vi->datalayout == EROFS_INODE_COMPRESSED_FULL)
 486                         vi->z_fragmentoff |= (u64)m.pblk << 32;
 487         }
 488         if (ztailpacking && m.lcn == vi->z_tailextent_headlcn) {
 489                 map->m_flags |= EROFS_MAP_META;
 490                 map->m_pa = vi->z_idataoff;
 491                 map->m_plen = vi->z_idata_size;
 492         } else if (fragment && m.lcn == vi->z_tailextent_headlcn) {
 493                 map->m_flags |= EROFS_MAP_FRAGMENT;
 494         } else {
 495                 map->m_pa = erofs_pos(inode->i_sb, m.pblk);
 496                 err = z_erofs_get_extent_compressedlen(&m, initial_lcn);
 497                 if (err)
 498                         goto unmap_out;
 499         }
 500
 501         if (m.headtype == Z_EROFS_LCLUSTER_TYPE_PLAIN) {
 502                 if (map->m_llen > map->m_plen) {
 503                         DBG_BUGON(1);
 504                         err = -EFSCORRUPTED;
 505                         goto unmap_out;
 506                 }
 507                 afmt = vi->z_advise & Z_EROFS_ADVISE_INTERLACED_PCLUSTER ?
 508                         Z_EROFS_COMPRESSION_INTERLACED :
 509                         Z_EROFS_COMPRESSION_SHIFTED;
 510         } else {
 511                 afmt = m.headtype == Z_EROFS_LCLUSTER_TYPE_HEAD2 ?
 512                         vi->z_algorithmtype[1] : vi->z_algorithmtype[0];
 513                 if (!(EROFS_I_SB(inode)->available_compr_algs & (1 << afmt))) {
 514                         erofs_err(inode->i_sb, "inconsistent algorithmtype %u for nid %llu",
 515                                   afmt, vi->nid);
 516                         err = -EFSCORRUPTED;
 517                         goto unmap_out;
 518                 }
 519         }
 520         map->m_algorithmformat = afmt;
 521
 522         if ((flags & EROFS_GET_BLOCKS_FIEMAP) ||
 523             ((flags & EROFS_GET_BLOCKS_READMORE) &&
 524              (map->m_algorithmformat == Z_EROFS_COMPRESSION_LZMA ||
 525               map->m_algorithmformat == Z_EROFS_COMPRESSION_DEFLATE ||
 526               map->m_algorithmformat == Z_EROFS_COMPRESSION_ZSTD) &&
 527               map->m_llen >= i_blocksize(inode))) {
 528                 err = z_erofs_get_extent_decompressedlen(&m);
 529                 if (!err)
 530                         map->m_flags |= EROFS_MAP_FULL_MAPPED;
 531         }
 532
 533 unmap_out:
 534         erofs_unmap_metabuf(&m.map->buf);
 535         return err;
 536 }
 537
 538 static int z_erofs_fill_inode_lazy(struct inode *inode)
 539 {
 540         struct erofs_inode *const vi = EROFS_I(inode);
 541         struct super_block *const sb = inode->i_sb;
 542         int err, headnr;
 543         erofs_off_t pos;
 544         struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
 545         struct z_erofs_map_header *h;
 546
 547         if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags)) {
 548                 /*
 549                  * paired with smp_mb() at the end of the function to ensure
 550                  * fields will only be observed after the bit is set.
 551                  */
 552                 smp_mb();
 553                 return 0;
 554         }
 555
 556         if (wait_on_bit_lock(&vi->flags, EROFS_I_BL_Z_BIT, TASK_KILLABLE))
 557                 return -ERESTARTSYS;
 558
 559         err = 0;
 560         if (test_bit(EROFS_I_Z_INITED_BIT, &vi->flags))
 561                 goto out_unlock;
 562
 563         pos = ALIGN(erofs_iloc(inode) + vi->inode_isize + vi->xattr_isize, 8);
 564         h = erofs_read_metabuf(&buf, sb, pos, EROFS_KMAP);
 565         if (IS_ERR(h)) {
 566                 err = PTR_ERR(h);
 567                 goto out_unlock;
 568         }
 569
 570         /*
 571          * if the highest bit of the 8-byte map header is set, the whole file
 572          * is stored in the packed inode. The rest bits keeps z_fragmentoff.
 573          */
 574         if (h->h_clusterbits >> Z_EROFS_FRAGMENT_INODE_BIT) {
 575                 vi->z_advise = Z_EROFS_ADVISE_FRAGMENT_PCLUSTER;
 576                 vi->z_fragmentoff = le64_to_cpu(*(__le64 *)h) ^ (1ULL << 63);
 577                 vi->z_tailextent_headlcn = 0;
 578                 goto done;
 579         }
 580         vi->z_advise = le16_to_cpu(h->h_advise);
 581         vi->z_algorithmtype[0] = h->h_algorithmtype & 15;
 582         vi->z_algorithmtype[1] = h->h_algorithmtype >> 4;
 583
 584         headnr = 0;
 585         if (vi->z_algorithmtype[0] >= Z_EROFS_COMPRESSION_MAX ||
 586             vi->z_algorithmtype[++headnr] >= Z_EROFS_COMPRESSION_MAX) {
 587                 erofs_err(sb, "unknown HEAD%u format %u for nid %llu, please upgrade kernel",
 588                           headnr + 1, vi->z_algorithmtype[headnr], vi->nid);
 589                 err = -EOPNOTSUPP;
 590                 goto out_put_metabuf;
 591         }
 592
 593         vi->z_logical_clusterbits = sb->s_blocksize_bits + (h->h_clusterbits & 7);
 594         if (!erofs_sb_has_big_pcluster(EROFS_SB(sb)) &&
 595             vi->z_advise & (Z_EROFS_ADVISE_BIG_PCLUSTER_1 |
 596                             Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
 597                 erofs_err(sb, "per-inode big pcluster without sb feature for nid %llu",
 598                           vi->nid);
 599                 err = -EFSCORRUPTED;
 600                 goto out_put_metabuf;
 601         }
 602         if (vi->datalayout == EROFS_INODE_COMPRESSED_COMPACT &&
 603             !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1) ^
 604             !(vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_2)) {
 605                 erofs_err(sb, "big pcluster head1/2 of compact indexes should be consistent for nid %llu",
 606                           vi->nid);
 607                 err = -EFSCORRUPTED;
 608                 goto out_put_metabuf;
 609         }
 610
 611         if (vi->z_advise & Z_EROFS_ADVISE_INLINE_PCLUSTER) {
 612                 struct erofs_map_blocks map = {
 613                         .buf = __EROFS_BUF_INITIALIZER
 614                 };
 615
 616                 vi->z_idata_size = le16_to_cpu(h->h_idata_size);
 617                 err = z_erofs_do_map_blocks(inode, &map,
 618                                             EROFS_GET_BLOCKS_FINDTAIL);
 619                 erofs_put_metabuf(&map.buf);
 620
 621                 if (!map.m_plen ||
 622                     erofs_blkoff(sb, map.m_pa) + map.m_plen > sb->s_blocksize) {
 623                         erofs_err(sb, "invalid tail-packing pclustersize %llu",
 624                                   map.m_plen);
 625                         err = -EFSCORRUPTED;
 626                 }
 627                 if (err < 0)
 628                         goto out_put_metabuf;
 629         }
 630
 631         if (vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER &&
 632             !(h->h_clusterbits >> Z_EROFS_FRAGMENT_INODE_BIT)) {
 633                 struct erofs_map_blocks map = {
 634                         .buf = __EROFS_BUF_INITIALIZER
 635                 };
 636
 637                 vi->z_fragmentoff = le32_to_cpu(h->h_fragmentoff);
 638                 err = z_erofs_do_map_blocks(inode, &map,
 639                                             EROFS_GET_BLOCKS_FINDTAIL);
 640                 erofs_put_metabuf(&map.buf);
 641                 if (err < 0)
 642                         goto out_put_metabuf;
 643         }
 644 done:
 645         /* paired with smp_mb() at the beginning of the function */
 646         smp_mb();
 647         set_bit(EROFS_I_Z_INITED_BIT, &vi->flags);
 648 out_put_metabuf:
 649         erofs_put_metabuf(&buf);
 650 out_unlock:
 651         clear_and_wake_up_bit(EROFS_I_BL_Z_BIT, &vi->flags);
 652         return err;
 653 }
 654
 655 int z_erofs_map_blocks_iter(struct inode *inode, struct erofs_map_blocks *map,
 656                             int flags)
 657 {
 658         struct erofs_inode *const vi = EROFS_I(inode);
 659         int err = 0;
 660
 661         trace_erofs_map_blocks_enter(inode, map, flags);
 662         if (map->m_la >= inode->i_size) {       /* post-EOF unmapped extent */
 663                 map->m_llen = map->m_la + 1 - inode->i_size;
 664                 map->m_la = inode->i_size;
 665                 map->m_flags = 0;
 666         } else {
 667                 err = z_erofs_fill_inode_lazy(inode);
 668                 if (!err) {
 669                         if ((vi->z_advise & Z_EROFS_ADVISE_FRAGMENT_PCLUSTER) &&
 670                             !vi->z_tailextent_headlcn) {
 671                                 map->m_la = 0;
 672                                 map->m_llen = inode->i_size;
 673                                 map->m_flags = EROFS_MAP_MAPPED |
 674                                         EROFS_MAP_FULL_MAPPED | EROFS_MAP_FRAGMENT;
 675                         } else {
 676                                 err = z_erofs_do_map_blocks(inode, map, flags);
 677                         }
 678                 }
 679                 if (!err && (map->m_flags & EROFS_MAP_ENCODED) &&
 680                     unlikely(map->m_plen > Z_EROFS_PCLUSTER_MAX_SIZE ||
 681                              map->m_llen > Z_EROFS_PCLUSTER_MAX_DSIZE))
 682                         err = -EOPNOTSUPP;
 683                 if (err)
 684                         map->m_llen = 0;
 685         }
 686         trace_erofs_map_blocks_exit(inode, map, flags, err);
 687         return err;
 688 }
 689
 690 static int z_erofs_iomap_begin_report(struct inode *inode, loff_t offset,
 691                                 loff_t length, unsigned int flags,
 692                                 struct iomap *iomap, struct iomap *srcmap)
 693 {
 694         int ret;
 695         struct erofs_map_blocks map = { .m_la = offset };
 696
 697         ret = z_erofs_map_blocks_iter(inode, &map, EROFS_GET_BLOCKS_FIEMAP);
 698         erofs_put_metabuf(&map.buf);
 699         if (ret < 0)
 700                 return ret;
 701
 702         iomap->bdev = inode->i_sb->s_bdev;
 703         iomap->offset = map.m_la;
 704         iomap->length = map.m_llen;
 705         if (map.m_flags & EROFS_MAP_MAPPED) {
 706                 iomap->type = IOMAP_MAPPED;
 707                 iomap->addr = map.m_flags & EROFS_MAP_FRAGMENT ?
 708                               IOMAP_NULL_ADDR : map.m_pa;
 709         } else {
 710                 iomap->type = IOMAP_HOLE;
 711                 iomap->addr = IOMAP_NULL_ADDR;
 712                 /*
 713                  * No strict rule on how to describe extents for post EOF, yet
 714                  * we need to do like below. Otherwise, iomap itself will get
 715                  * into an endless loop on post EOF.
 716                  *
 717                  * Calculate the effective offset by subtracting extent start
 718                  * (map.m_la) from the requested offset, and add it to length.
 719                  * (NB: offset >= map.m_la always)
 720                  */
 721                 if (iomap->offset >= inode->i_size)
 722                         iomap->length = length + offset - map.m_la;
 723         }
 724         iomap->flags = 0;
 725         return 0;
 726 }
 727
 728 const struct iomap_ops z_erofs_iomap_report_ops = {
 729         .iomap_begin = z_erofs_iomap_begin_report,
 730 };