sys/fs/ext2fs/ext2_htree.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2010, 2012 Zheng Liu <lz@freebsd.org>
   5  * Copyright (c) 2012, Vyacheslav Matyushin
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  * $FreeBSD$
  30  */
  31
  32 #include <sys/param.h>
  33 #include <sys/endian.h>
  34 #include <sys/systm.h>
  35 #include <sys/namei.h>
  36 #include <sys/bio.h>
  37 #include <sys/buf.h>
  38 #include <sys/endian.h>
  39 #include <sys/mount.h>
  40 #include <sys/vnode.h>
  41 #include <sys/malloc.h>
  42 #include <sys/dirent.h>
  43 #include <sys/sdt.h>
  44 #include <sys/sysctl.h>
  45
  46 #include <ufs/ufs/dir.h>
  47
  48 #include <fs/ext2fs/fs.h>
  49 #include <fs/ext2fs/inode.h>
  50 #include <fs/ext2fs/ext2_mount.h>
  51 #include <fs/ext2fs/ext2fs.h>
  52 #include <fs/ext2fs/fs.h>
  53 #include <fs/ext2fs/ext2_extern.h>
  54 #include <fs/ext2fs/ext2_dinode.h>
  55 #include <fs/ext2fs/ext2_dir.h>
  56 #include <fs/ext2fs/htree.h>
  57
  58 SDT_PROVIDER_DECLARE(ext2fs);
  59 /*
  60  * ext2fs trace probe:
  61  * arg0: verbosity. Higher numbers give more verbose messages
  62  * arg1: Textual message
  63  */
  64 SDT_PROBE_DEFINE2(ext2fs, , trace, htree, "int", "char*");
  65
  66 static void     ext2_append_entry(char *block, uint32_t blksize,
  67                     struct ext2fs_direct_2 *last_entry,
  68                     struct ext2fs_direct_2 *new_entry, int csum_size);
  69 static int      ext2_htree_append_block(struct vnode *vp, char *data,
  70                     struct componentname *cnp, uint32_t blksize);
  71 static int      ext2_htree_check_next(struct inode *ip, uint32_t hash,
  72                     const char *name, struct ext2fs_htree_lookup_info *info);
  73 static int      ext2_htree_cmp_sort_entry(const void *e1, const void *e2);
  74 static int      ext2_htree_find_leaf(struct inode *ip, const char *name,
  75                     int namelen, uint32_t *hash, uint8_t *hash_version,
  76                     struct ext2fs_htree_lookup_info *info);
  77 static uint32_t ext2_htree_get_block(struct ext2fs_htree_entry *ep);
  78 static uint16_t ext2_htree_get_count(struct ext2fs_htree_entry *ep);
  79 static uint32_t ext2_htree_get_hash(struct ext2fs_htree_entry *ep);
  80 static uint16_t ext2_htree_get_limit(struct ext2fs_htree_entry *ep);
  81 static void     ext2_htree_insert_entry_to_level(struct ext2fs_htree_lookup_level *level,
  82                     uint32_t hash, uint32_t blk);
  83 static void     ext2_htree_insert_entry(struct ext2fs_htree_lookup_info *info,
  84                     uint32_t hash, uint32_t blk);
  85 static uint32_t ext2_htree_node_limit(struct inode *ip);
  86 static void     ext2_htree_set_block(struct ext2fs_htree_entry *ep,
  87                     uint32_t blk);
  88 static void     ext2_htree_set_count(struct ext2fs_htree_entry *ep,
  89                     uint16_t cnt);
  90 static void     ext2_htree_set_hash(struct ext2fs_htree_entry *ep,
  91                     uint32_t hash);
  92 static void     ext2_htree_set_limit(struct ext2fs_htree_entry *ep,
  93                     uint16_t limit);
  94 static int      ext2_htree_split_dirblock(struct inode *ip,
  95                     char *block1, char *block2, uint32_t blksize,
  96                     uint32_t *hash_seed, uint8_t hash_version,
  97                     uint32_t *split_hash, struct  ext2fs_direct_2 *entry);
  98 static void     ext2_htree_release(struct ext2fs_htree_lookup_info *info);
  99 static uint32_t ext2_htree_root_limit(struct inode *ip, int len);
 100 static int      ext2_htree_writebuf(struct inode *ip,
 101                     struct ext2fs_htree_lookup_info *info);
 102
 103 int
 104 ext2_htree_has_idx(struct inode *ip)
 105 {
 106         if (EXT2_HAS_COMPAT_FEATURE(ip->i_e2fs, EXT2F_COMPAT_DIRHASHINDEX) &&
 107             ip->i_flag & IN_E3INDEX)
 108                 return (1);
 109         else
 110                 return (0);
 111 }
 112
 113 static int
 114 ext2_htree_check_next(struct inode *ip, uint32_t hash, const char *name,
 115     struct ext2fs_htree_lookup_info *info)
 116 {
 117         struct vnode *vp = ITOV(ip);
 118         struct ext2fs_htree_lookup_level *level;
 119         struct buf *bp;
 120         uint32_t next_hash;
 121         int idx = info->h_levels_num - 1;
 122         int levels = 0;
 123
 124         do {
 125                 level = &info->h_levels[idx];
 126                 level->h_entry++;
 127                 if (level->h_entry < level->h_entries +
 128                     ext2_htree_get_count(level->h_entries))
 129                         break;
 130                 if (idx == 0)
 131                         return (0);
 132                 idx--;
 133                 levels++;
 134         } while (1);
 135
 136         next_hash = ext2_htree_get_hash(level->h_entry);
 137         if ((hash & 1) == 0) {
 138                 if (hash != (next_hash & ~1))
 139                         return (0);
 140         }
 141
 142         while (levels > 0) {
 143                 levels--;
 144                 if (ext2_blkatoff(vp, ext2_htree_get_block(level->h_entry) *
 145                     ip->i_e2fs->e2fs_bsize, NULL, &bp) != 0)
 146                         return (0);
 147                 level = &info->h_levels[idx + 1];
 148                 brelse(level->h_bp);
 149                 level->h_bp = bp;
 150                 level->h_entry = level->h_entries =
 151                     ((struct ext2fs_htree_node *)bp->b_data)->h_entries;
 152         }
 153
 154         return (1);
 155 }
 156
 157 static uint32_t
 158 ext2_htree_get_block(struct ext2fs_htree_entry *ep)
 159 {
 160         return (ep->h_blk & 0x00FFFFFF);
 161 }
 162
 163 static void
 164 ext2_htree_set_block(struct ext2fs_htree_entry *ep, uint32_t blk)
 165 {
 166         ep->h_blk = blk;
 167 }
 168
 169 static uint16_t
 170 ext2_htree_get_count(struct ext2fs_htree_entry *ep)
 171 {
 172         return (((struct ext2fs_htree_count *)(ep))->h_entries_num);
 173 }
 174
 175 static void
 176 ext2_htree_set_count(struct ext2fs_htree_entry *ep, uint16_t cnt)
 177 {
 178         ((struct ext2fs_htree_count *)(ep))->h_entries_num = cnt;
 179 }
 180
 181 static uint32_t
 182 ext2_htree_get_hash(struct ext2fs_htree_entry *ep)
 183 {
 184         return (ep->h_hash);
 185 }
 186
 187 static uint16_t
 188 ext2_htree_get_limit(struct ext2fs_htree_entry *ep)
 189 {
 190         return (((struct ext2fs_htree_count *)(ep))->h_entries_max);
 191 }
 192
 193 static void
 194 ext2_htree_set_hash(struct ext2fs_htree_entry *ep, uint32_t hash)
 195 {
 196         ep->h_hash = hash;
 197 }
 198
 199 static void
 200 ext2_htree_set_limit(struct ext2fs_htree_entry *ep, uint16_t limit)
 201 {
 202         ((struct ext2fs_htree_count *)(ep))->h_entries_max = limit;
 203 }
 204
 205 static void
 206 ext2_htree_release(struct ext2fs_htree_lookup_info *info)
 207 {
 208         u_int i;
 209
 210         for (i = 0; i < info->h_levels_num; i++) {
 211                 struct buf *bp = info->h_levels[i].h_bp;
 212
 213                 if (bp != NULL)
 214                         brelse(bp);
 215         }
 216 }
 217
 218 static uint32_t
 219 ext2_htree_root_limit(struct inode *ip, int len)
 220 {
 221         struct m_ext2fs *fs;
 222         uint32_t space;
 223
 224         fs = ip->i_e2fs;
 225         space = ip->i_e2fs->e2fs_bsize - EXT2_DIR_REC_LEN(1) -
 226             EXT2_DIR_REC_LEN(2) - len;
 227
 228         if (EXT2_HAS_RO_COMPAT_FEATURE(fs, EXT2F_ROCOMPAT_METADATA_CKSUM))
 229                 space -= sizeof(struct ext2fs_htree_tail);
 230
 231         return (space / sizeof(struct ext2fs_htree_entry));
 232 }
 233
 234 static uint32_t
 235 ext2_htree_node_limit(struct inode *ip)
 236 {
 237         struct m_ext2fs *fs;
 238         uint32_t space;
 239
 240         fs = ip->i_e2fs;
 241         space = fs->e2fs_bsize - EXT2_DIR_REC_LEN(0);
 242
 243         if (EXT2_HAS_RO_COMPAT_FEATURE(fs, EXT2F_ROCOMPAT_METADATA_CKSUM))
 244                 space -= sizeof(struct ext2fs_htree_tail);
 245
 246         return (space / sizeof(struct ext2fs_htree_entry));
 247 }
 248
 249 static int
 250 ext2_htree_find_leaf(struct inode *ip, const char *name, int namelen,
 251     uint32_t *hash, uint8_t *hash_ver,
 252     struct ext2fs_htree_lookup_info *info)
 253 {
 254         struct vnode *vp;
 255         struct ext2fs *fs;
 256         struct m_ext2fs *m_fs;
 257         struct buf *bp = NULL;
 258         struct ext2fs_htree_root *rootp;
 259         struct ext2fs_htree_entry *entp, *start, *end, *middle, *found;
 260         struct ext2fs_htree_lookup_level *level_info;
 261         uint32_t hash_major = 0, hash_minor = 0;
 262         uint32_t levels, cnt;
 263         uint8_t hash_version;
 264
 265         if (name == NULL || info == NULL)
 266                 return (-1);
 267
 268         vp = ITOV(ip);
 269         fs = ip->i_e2fs->e2fs;
 270         m_fs = ip->i_e2fs;
 271
 272         if (ext2_blkatoff(vp, 0, NULL, &bp) != 0)
 273                 return (-1);
 274
 275         info->h_levels_num = 1;
 276         info->h_levels[0].h_bp = bp;
 277         rootp = (struct ext2fs_htree_root *)bp->b_data;
 278         if (rootp->h_info.h_hash_version != EXT2_HTREE_LEGACY &&
 279             rootp->h_info.h_hash_version != EXT2_HTREE_HALF_MD4 &&
 280             rootp->h_info.h_hash_version != EXT2_HTREE_TEA)
 281                 goto error;
 282
 283         hash_version = rootp->h_info.h_hash_version;
 284         if (hash_version <= EXT2_HTREE_TEA)
 285                 hash_version += m_fs->e2fs_uhash;
 286         *hash_ver = hash_version;
 287
 288         ext2_htree_hash(name, namelen, fs->e3fs_hash_seed,
 289             hash_version, &hash_major, &hash_minor);
 290         *hash = hash_major;
 291
 292         if ((levels = rootp->h_info.h_ind_levels) > 1)
 293                 goto error;
 294
 295         entp = (struct ext2fs_htree_entry *)(((char *)&rootp->h_info) +
 296             rootp->h_info.h_info_len);
 297
 298         if (ext2_htree_get_limit(entp) !=
 299             ext2_htree_root_limit(ip, rootp->h_info.h_info_len))
 300                 goto error;
 301
 302         while (1) {
 303                 cnt = ext2_htree_get_count(entp);
 304                 if (cnt == 0 || cnt > ext2_htree_get_limit(entp))
 305                         goto error;
 306
 307                 start = entp + 1;
 308                 end = entp + cnt - 1;
 309                 while (start <= end) {
 310                         middle = start + (end - start) / 2;
 311                         if (ext2_htree_get_hash(middle) > hash_major)
 312                                 end = middle - 1;
 313                         else
 314                                 start = middle + 1;
 315                 }
 316                 found = start - 1;
 317
 318                 level_info = &(info->h_levels[info->h_levels_num - 1]);
 319                 level_info->h_bp = bp;
 320                 level_info->h_entries = entp;
 321                 level_info->h_entry = found;
 322                 if (levels == 0)
 323                         return (0);
 324                 levels--;
 325                 if (ext2_blkatoff(vp,
 326                     ext2_htree_get_block(found) * m_fs->e2fs_bsize,
 327                     NULL, &bp) != 0)
 328                         goto error;
 329                 entp = ((struct ext2fs_htree_node *)bp->b_data)->h_entries;
 330                 info->h_levels_num++;
 331                 info->h_levels[info->h_levels_num - 1].h_bp = bp;
 332         }
 333
 334 error:
 335         ext2_htree_release(info);
 336         return (-1);
 337 }
 338
 339 /*
 340  * Try to lookup a directory entry in HTree index
 341  */
 342 int
 343 ext2_htree_lookup(struct inode *ip, const char *name, int namelen,
 344     struct buf **bpp, int *entryoffp, doff_t *offp,
 345     doff_t *prevoffp, doff_t *endusefulp,
 346     struct ext2fs_searchslot *ss)
 347 {
 348         struct vnode *vp;
 349         struct ext2fs_htree_lookup_info info;
 350         struct ext2fs_htree_entry *leaf_node;
 351         struct m_ext2fs *m_fs;
 352         struct buf *bp;
 353         uint32_t blk;
 354         uint32_t dirhash;
 355         uint32_t bsize;
 356         uint8_t hash_version;
 357         int search_next;
 358         int found = 0;
 359
 360         m_fs = ip->i_e2fs;
 361         bsize = m_fs->e2fs_bsize;
 362         vp = ITOV(ip);
 363
 364         /* TODO: print error msg because we don't lookup '.' and '..' */
 365
 366         memset(&info, 0, sizeof(info));
 367         if (ext2_htree_find_leaf(ip, name, namelen, &dirhash,
 368             &hash_version, &info))
 369                 return (-1);
 370
 371         do {
 372                 leaf_node = info.h_levels[info.h_levels_num - 1].h_entry;
 373                 blk = ext2_htree_get_block(leaf_node);
 374                 if (ext2_blkatoff(vp, blk * bsize, NULL, &bp) != 0) {
 375                         ext2_htree_release(&info);
 376                         return (-1);
 377                 }
 378
 379                 *offp = blk * bsize;
 380                 *entryoffp = 0;
 381                 *prevoffp = blk * bsize;
 382                 *endusefulp = blk * bsize;
 383
 384                 if (ss->slotstatus == NONE) {
 385                         ss->slotoffset = -1;
 386                         ss->slotfreespace = 0;
 387                 }
 388
 389                 if (ext2_search_dirblock(ip, bp->b_data, &found,
 390                     name, namelen, entryoffp, offp, prevoffp,
 391                     endusefulp, ss) != 0) {
 392                         brelse(bp);
 393                         ext2_htree_release(&info);
 394                         return (-1);
 395                 }
 396
 397                 if (found) {
 398                         *bpp = bp;
 399                         ext2_htree_release(&info);
 400                         return (0);
 401                 }
 402
 403                 brelse(bp);
 404                 search_next = ext2_htree_check_next(ip, dirhash, name, &info);
 405         } while (search_next);
 406
 407         ext2_htree_release(&info);
 408         return (ENOENT);
 409 }
 410
 411 static int
 412 ext2_htree_append_block(struct vnode *vp, char *data,
 413     struct componentname *cnp, uint32_t blksize)
 414 {
 415         struct iovec aiov;
 416         struct uio auio;
 417         struct inode *dp = VTOI(vp);
 418         uint64_t cursize, newsize;
 419         int error;
 420
 421         cursize = roundup(dp->i_size, blksize);
 422         newsize = cursize + blksize;
 423
 424         auio.uio_offset = cursize;
 425         auio.uio_resid = blksize;
 426         aiov.iov_len = blksize;
 427         aiov.iov_base = data;
 428         auio.uio_iov = &aiov;
 429         auio.uio_iovcnt = 1;
 430         auio.uio_rw = UIO_WRITE;
 431         auio.uio_segflg = UIO_SYSSPACE;
 432         error = VOP_WRITE(vp, &auio, IO_SYNC, cnp->cn_cred);
 433         if (!error)
 434                 dp->i_size = newsize;
 435
 436         return (error);
 437 }
 438
 439 static int
 440 ext2_htree_writebuf(struct inode* ip, struct ext2fs_htree_lookup_info *info)
 441 {
 442         int i, error;
 443
 444         for (i = 0; i < info->h_levels_num; i++) {
 445                 struct buf *bp = info->h_levels[i].h_bp;
 446                 ext2_dx_csum_set(ip, (struct ext2fs_direct_2 *)bp->b_data);
 447                 error = bwrite(bp);
 448                 if (error)
 449                         return (error);
 450         }
 451
 452         return (0);
 453 }
 454
 455 static void
 456 ext2_htree_insert_entry_to_level(struct ext2fs_htree_lookup_level *level,
 457     uint32_t hash, uint32_t blk)
 458 {
 459         struct ext2fs_htree_entry *target;
 460         int entries_num;
 461
 462         target = level->h_entry + 1;
 463         entries_num = ext2_htree_get_count(level->h_entries);
 464
 465         memmove(target + 1, target, (char *)(level->h_entries + entries_num) -
 466             (char *)target);
 467         ext2_htree_set_block(target, blk);
 468         ext2_htree_set_hash(target, hash);
 469         ext2_htree_set_count(level->h_entries, entries_num + 1);
 470 }
 471
 472 /*
 473  * Insert an index entry to the index node.
 474  */
 475 static void
 476 ext2_htree_insert_entry(struct ext2fs_htree_lookup_info *info,
 477     uint32_t hash, uint32_t blk)
 478 {
 479         struct ext2fs_htree_lookup_level *level;
 480
 481         level = &info->h_levels[info->h_levels_num - 1];
 482         ext2_htree_insert_entry_to_level(level, hash, blk);
 483 }
 484
 485 /*
 486  * Compare two entry sort descriptors by name hash value.
 487  * This is used together with qsort.
 488  */
 489 static int
 490 ext2_htree_cmp_sort_entry(const void *e1, const void *e2)
 491 {
 492         const struct ext2fs_htree_sort_entry *entry1, *entry2;
 493
 494         entry1 = (const struct ext2fs_htree_sort_entry *)e1;
 495         entry2 = (const struct ext2fs_htree_sort_entry *)e2;
 496
 497         if (entry1->h_hash < entry2->h_hash)
 498                 return (-1);
 499         if (entry1->h_hash > entry2->h_hash)
 500                 return (1);
 501         return (0);
 502 }
 503
 504 /*
 505  * Append an entry to the end of the directory block.
 506  */
 507 static void
 508 ext2_append_entry(char *block, uint32_t blksize,
 509     struct ext2fs_direct_2 *last_entry,
 510     struct ext2fs_direct_2 *new_entry, int csum_size)
 511 {
 512         uint16_t entry_len;
 513
 514         entry_len = EXT2_DIR_REC_LEN(last_entry->e2d_namlen);
 515         last_entry->e2d_reclen = entry_len;
 516         last_entry = (struct ext2fs_direct_2 *)((char *)last_entry + entry_len);
 517         new_entry->e2d_reclen = block + blksize - (char *)last_entry - csum_size;
 518         memcpy(last_entry, new_entry, EXT2_DIR_REC_LEN(new_entry->e2d_namlen));
 519 }
 520
 521 /*
 522  * Move half of entries from the old directory block to the new one.
 523  */
 524 static int
 525 ext2_htree_split_dirblock(struct inode *ip, char *block1, char *block2,
 526     uint32_t blksize, uint32_t *hash_seed, uint8_t hash_version,
 527     uint32_t *split_hash, struct ext2fs_direct_2 *entry)
 528 {
 529         struct m_ext2fs *fs;
 530         int entry_cnt = 0;
 531         int size = 0, csum_size = 0;
 532         int i, k;
 533         uint32_t offset;
 534         uint16_t entry_len = 0;
 535         uint32_t entry_hash;
 536         struct ext2fs_direct_2 *ep, *last;
 537         char *dest;
 538         struct ext2fs_htree_sort_entry *sort_info;
 539
 540         fs = ip->i_e2fs;
 541         ep = (struct ext2fs_direct_2 *)block1;
 542         dest = block2;
 543         sort_info = (struct ext2fs_htree_sort_entry *)
 544             ((char *)block2 + blksize);
 545
 546         if (EXT2_HAS_RO_COMPAT_FEATURE(fs, EXT2F_ROCOMPAT_METADATA_CKSUM))
 547                 csum_size = sizeof(struct ext2fs_direct_tail);
 548
 549         /*
 550          * Calculate name hash value for the entry which is to be added.
 551          */
 552         ext2_htree_hash(entry->e2d_name, entry->e2d_namlen, hash_seed,
 553             hash_version, &entry_hash, NULL);
 554
 555         /*
 556          * Fill in directory entry sort descriptors.
 557          */
 558         while ((char *)ep < block1 + blksize - csum_size) {
 559                 if (ep->e2d_ino && ep->e2d_namlen) {
 560                         entry_cnt++;
 561                         sort_info--;
 562                         sort_info->h_size = ep->e2d_reclen;
 563                         sort_info->h_offset = (char *)ep - block1;
 564                         ext2_htree_hash(ep->e2d_name, ep->e2d_namlen,
 565                             hash_seed, hash_version,
 566                             &sort_info->h_hash, NULL);
 567                 }
 568                 ep = (struct ext2fs_direct_2 *)
 569                     ((char *)ep + ep->e2d_reclen);
 570         }
 571
 572         /*
 573          * Sort directory entry descriptors by name hash value.
 574          */
 575         qsort(sort_info, entry_cnt, sizeof(struct ext2fs_htree_sort_entry),
 576             ext2_htree_cmp_sort_entry);
 577
 578         /*
 579          * Count the number of entries to move to directory block 2.
 580          */
 581         for (i = entry_cnt - 1; i >= 0; i--) {
 582                 if (sort_info[i].h_size + size > blksize / 2)
 583                         break;
 584                 size += sort_info[i].h_size;
 585         }
 586
 587         *split_hash = sort_info[i + 1].h_hash;
 588
 589         /*
 590          * Set collision bit.
 591          */
 592         if (*split_hash == sort_info[i].h_hash)
 593                 *split_hash += 1;
 594
 595         /*
 596          * Move half of directory entries from block 1 to block 2.
 597          */
 598         for (k = i + 1; k < entry_cnt; k++) {
 599                 ep = (struct ext2fs_direct_2 *)((char *)block1 +
 600                     sort_info[k].h_offset);
 601                 entry_len = EXT2_DIR_REC_LEN(ep->e2d_namlen);
 602                 memcpy(dest, ep, entry_len);
 603                 ((struct ext2fs_direct_2 *)dest)->e2d_reclen = entry_len;
 604                 /* Mark directory entry as unused. */
 605                 ep->e2d_ino = 0;
 606                 dest += entry_len;
 607         }
 608         dest -= entry_len;
 609
 610         /* Shrink directory entries in block 1. */
 611         last = (struct ext2fs_direct_2 *)block1;
 612         entry_len = 0;
 613         for (offset = 0; offset < blksize - csum_size; ) {
 614                 ep = (struct ext2fs_direct_2 *)(block1 + offset);
 615                 offset += ep->e2d_reclen;
 616                 if (ep->e2d_ino) {
 617                         last = (struct ext2fs_direct_2 *)
 618                             ((char *)last + entry_len);
 619                         entry_len = EXT2_DIR_REC_LEN(ep->e2d_namlen);
 620                         memcpy((void *)last, (void *)ep, entry_len);
 621                         last->e2d_reclen = entry_len;
 622                 }
 623         }
 624
 625         if (entry_hash >= *split_hash) {
 626                 /* Add entry to block 2. */
 627                 ext2_append_entry(block2, blksize,
 628                     (struct ext2fs_direct_2 *)dest, entry, csum_size);
 629
 630                 /* Adjust length field of last entry of block 1. */
 631                 last->e2d_reclen = block1 + blksize - (char *)last - csum_size;
 632         } else {
 633                 /* Add entry to block 1. */
 634                 ext2_append_entry(block1, blksize, last, entry, csum_size);
 635
 636                 /* Adjust length field of last entry of block 2. */
 637                 ((struct ext2fs_direct_2 *)dest)->e2d_reclen =
 638                     block2 + blksize - dest - csum_size;
 639         }
 640
 641         if (csum_size) {
 642                 ext2_init_dirent_tail(EXT2_DIRENT_TAIL(block1, blksize));
 643                 ext2_init_dirent_tail(EXT2_DIRENT_TAIL(block2, blksize));
 644         }
 645
 646         return (0);
 647 }
 648
 649 /*
 650  * Create an HTree index for a directory
 651  */
 652 int
 653 ext2_htree_create_index(struct vnode *vp, struct componentname *cnp,
 654     struct ext2fs_direct_2 *new_entry)
 655 {
 656         struct buf *bp = NULL;
 657         struct inode *dp;
 658         struct ext2fs *fs;
 659         struct m_ext2fs *m_fs;
 660         struct ext2fs_direct_2 *ep, *dotdot;
 661         struct ext2fs_htree_root *root;
 662         struct ext2fs_htree_lookup_info info;
 663         uint32_t blksize, dirlen, split_hash;
 664         uint8_t hash_version;
 665         char *buf1 = NULL;
 666         char *buf2 = NULL;
 667         int error = 0;
 668
 669         dp = VTOI(vp);
 670         fs = dp->i_e2fs->e2fs;
 671         m_fs = dp->i_e2fs;
 672         blksize = m_fs->e2fs_bsize;
 673
 674         buf1 = malloc(blksize, M_TEMP, M_WAITOK | M_ZERO);
 675         buf2 = malloc(blksize, M_TEMP, M_WAITOK | M_ZERO);
 676
 677         if ((error = ext2_blkatoff(vp, 0, NULL, &bp)) != 0)
 678                 goto out;
 679
 680         root = (struct ext2fs_htree_root *)bp->b_data;
 681         dotdot = (struct ext2fs_direct_2 *)((char *)&(root->h_dotdot));
 682         ep = (struct ext2fs_direct_2 *)((char *)dotdot + dotdot->e2d_reclen);
 683         dirlen = (char *)root + blksize - (char *)ep;
 684         memcpy(buf1, ep, dirlen);
 685         ep = (struct ext2fs_direct_2 *)buf1;
 686         while ((char *)ep < buf1 + dirlen)
 687                 ep = (struct ext2fs_direct_2 *)
 688                     ((char *)ep + ep->e2d_reclen);
 689         ep->e2d_reclen = buf1 + blksize - (char *)ep;
 690
 691         dp->i_flag |= IN_E3INDEX;
 692
 693         /*
 694          * Initialize index root.
 695          */
 696         dotdot->e2d_reclen = blksize - EXT2_DIR_REC_LEN(1);
 697         memset(&root->h_info, 0, sizeof(root->h_info));
 698         root->h_info.h_hash_version = fs->e3fs_def_hash_version;
 699         root->h_info.h_info_len = sizeof(root->h_info);
 700         ext2_htree_set_block(root->h_entries, 1);
 701         ext2_htree_set_count(root->h_entries, 1);
 702         ext2_htree_set_limit(root->h_entries,
 703             ext2_htree_root_limit(dp, sizeof(root->h_info)));
 704
 705         memset(&info, 0, sizeof(info));
 706         info.h_levels_num = 1;
 707         info.h_levels[0].h_entries = root->h_entries;
 708         info.h_levels[0].h_entry = root->h_entries;
 709
 710         hash_version = root->h_info.h_hash_version;
 711         if (hash_version <= EXT2_HTREE_TEA)
 712                 hash_version += m_fs->e2fs_uhash;
 713         ext2_htree_split_dirblock(dp, buf1, buf2, blksize, fs->e3fs_hash_seed,
 714             hash_version, &split_hash, new_entry);
 715         ext2_htree_insert_entry(&info, split_hash, 2);
 716
 717         /*
 718          * Write directory block 0.
 719          */
 720         ext2_dx_csum_set(dp, (struct ext2fs_direct_2 *)bp->b_data);
 721         if (DOINGASYNC(vp)) {
 722                 bdwrite(bp);
 723                 error = 0;
 724         } else {
 725                 error = bwrite(bp);
 726         }
 727         dp->i_flag |= IN_CHANGE | IN_UPDATE;
 728         if (error)
 729                 goto out;
 730
 731         /*
 732          * Write directory block 1.
 733          */
 734         ext2_dirent_csum_set(dp, (struct ext2fs_direct_2 *)buf1);
 735         error = ext2_htree_append_block(vp, buf1, cnp, blksize);
 736         if (error)
 737                 goto out1;
 738
 739         /*
 740          * Write directory block 2.
 741          */
 742         ext2_dirent_csum_set(dp, (struct ext2fs_direct_2 *)buf2);
 743         error = ext2_htree_append_block(vp, buf2, cnp, blksize);
 744
 745         free(buf1, M_TEMP);
 746         free(buf2, M_TEMP);
 747         return (error);
 748 out:
 749         if (bp != NULL)
 750                 brelse(bp);
 751 out1:
 752         free(buf1, M_TEMP);
 753         free(buf2, M_TEMP);
 754         return (error);
 755 }
 756
 757 /*
 758  * Add an entry to the directory using htree index.
 759  */
 760 int
 761 ext2_htree_add_entry(struct vnode *dvp, struct ext2fs_direct_2 *entry,
 762     struct componentname *cnp)
 763 {
 764         struct ext2fs_htree_entry *entries, *leaf_node;
 765         struct ext2fs_htree_lookup_info info;
 766         struct buf *bp = NULL;
 767         struct ext2fs *fs;
 768         struct m_ext2fs *m_fs;
 769         struct inode *ip;
 770         uint16_t ent_num;
 771         uint32_t dirhash, split_hash;
 772         uint32_t blksize, blknum;
 773         uint64_t cursize, dirsize;
 774         uint8_t hash_version;
 775         char *newdirblock = NULL;
 776         char *newidxblock = NULL;
 777         struct ext2fs_htree_node *dst_node;
 778         struct ext2fs_htree_entry *dst_entries;
 779         struct ext2fs_htree_entry *root_entires;
 780         struct buf *dst_bp = NULL;
 781         int error, write_bp = 0, write_dst_bp = 0, write_info = 0;
 782
 783         ip = VTOI(dvp);
 784         m_fs = ip->i_e2fs;
 785         fs = m_fs->e2fs;
 786         blksize = m_fs->e2fs_bsize;
 787
 788         if (ip->i_count != 0)
 789                 return ext2_add_entry(dvp, entry);
 790
 791         /* Target directory block is full, split it */
 792         memset(&info, 0, sizeof(info));
 793         error = ext2_htree_find_leaf(ip, entry->e2d_name, entry->e2d_namlen,
 794             &dirhash, &hash_version, &info);
 795         if (error)
 796                 return (error);
 797
 798         entries = info.h_levels[info.h_levels_num - 1].h_entries;
 799         ent_num = ext2_htree_get_count(entries);
 800         if (ent_num == ext2_htree_get_limit(entries)) {
 801                 /* Split the index node. */
 802                 root_entires = info.h_levels[0].h_entries;
 803                 newidxblock = malloc(blksize, M_TEMP, M_WAITOK | M_ZERO);
 804                 dst_node = (struct ext2fs_htree_node *)newidxblock;
 805                 memset(&dst_node->h_fake_dirent, 0,
 806                     sizeof(dst_node->h_fake_dirent));
 807                 dst_node->h_fake_dirent.e2d_reclen = blksize;
 808
 809                 cursize = roundup(ip->i_size, blksize);
 810                 dirsize = cursize + blksize;
 811                 blknum = dirsize / blksize - 1;
 812                 ext2_dx_csum_set(ip, (struct ext2fs_direct_2 *)newidxblock);
 813                 error = ext2_htree_append_block(dvp, newidxblock,
 814                     cnp, blksize);
 815                 if (error)
 816                         goto finish;
 817                 error = ext2_blkatoff(dvp, cursize, NULL, &dst_bp);
 818                 if (error)
 819                         goto finish;
 820                 dst_node = (struct ext2fs_htree_node *)dst_bp->b_data;
 821                 dst_entries = dst_node->h_entries;
 822
 823                 if (info.h_levels_num == 2) {
 824                         uint16_t src_ent_num, dst_ent_num;
 825
 826                         if (ext2_htree_get_count(root_entires) ==
 827                             ext2_htree_get_limit(root_entires)) {
 828                                 SDT_PROBE2(ext2fs, , trace, htree, 1,
 829                                     "directory index is full");
 830                                 error = EIO;
 831                                 goto finish;
 832                         }
 833
 834                         src_ent_num = ent_num / 2;
 835                         dst_ent_num = ent_num - src_ent_num;
 836                         split_hash = ext2_htree_get_hash(entries + src_ent_num);
 837
 838                         /* Move half of index entries to the new index node */
 839                         memcpy(dst_entries, entries + src_ent_num,
 840                             dst_ent_num * sizeof(struct ext2fs_htree_entry));
 841                         ext2_htree_set_count(entries, src_ent_num);
 842                         ext2_htree_set_count(dst_entries, dst_ent_num);
 843                         ext2_htree_set_limit(dst_entries,
 844                             ext2_htree_node_limit(ip));
 845
 846                         if (info.h_levels[1].h_entry >= entries + src_ent_num) {
 847                                 struct buf *tmp = info.h_levels[1].h_bp;
 848
 849                                 info.h_levels[1].h_bp = dst_bp;
 850                                 dst_bp = tmp;
 851
 852                                 info.h_levels[1].h_entry =
 853                                     info.h_levels[1].h_entry -
 854                                     (entries + src_ent_num) +
 855                                     dst_entries;
 856                                 info.h_levels[1].h_entries = dst_entries;
 857                         }
 858                         ext2_htree_insert_entry_to_level(&info.h_levels[0],
 859                             split_hash, blknum);
 860
 861                         /* Write new index node to disk */
 862                         ext2_dx_csum_set(ip,
 863                             (struct ext2fs_direct_2 *)dst_bp->b_data);
 864                         error = bwrite(dst_bp);
 865                         ip->i_flag |= IN_CHANGE | IN_UPDATE;
 866                         if (error)
 867                                 goto finish;
 868                         write_dst_bp = 1;
 869                 } else {
 870                         /* Create second level for htree index */
 871                         struct ext2fs_htree_root *idx_root;
 872
 873                         memcpy(dst_entries, entries,
 874                             ent_num * sizeof(struct ext2fs_htree_entry));
 875                         ext2_htree_set_limit(dst_entries,
 876                             ext2_htree_node_limit(ip));
 877
 878                         idx_root = (struct ext2fs_htree_root *)
 879                             info.h_levels[0].h_bp->b_data;
 880                         idx_root->h_info.h_ind_levels = 1;
 881
 882                         ext2_htree_set_count(entries, 1);
 883                         ext2_htree_set_block(entries, blknum);
 884
 885                         info.h_levels_num = 2;
 886                         info.h_levels[1].h_entries = dst_entries;
 887                         info.h_levels[1].h_entry = info.h_levels[0].h_entry -
 888                             info.h_levels[0].h_entries + dst_entries;
 889                         info.h_levels[1].h_bp = dst_bp;
 890                         dst_bp = NULL;
 891                 }
 892         }
 893
 894         leaf_node = info.h_levels[info.h_levels_num - 1].h_entry;
 895         blknum = ext2_htree_get_block(leaf_node);
 896         error = ext2_blkatoff(dvp, blknum * blksize, NULL, &bp);
 897         if (error)
 898                 goto finish;
 899
 900         /* Split target directory block */
 901         newdirblock = malloc(blksize, M_TEMP, M_WAITOK | M_ZERO);
 902         ext2_htree_split_dirblock(ip, (char *)bp->b_data, newdirblock, blksize,
 903             fs->e3fs_hash_seed, hash_version, &split_hash, entry);
 904         cursize = roundup(ip->i_size, blksize);
 905         dirsize = cursize + blksize;
 906         blknum = dirsize / blksize - 1;
 907
 908         /* Add index entry for the new directory block */
 909         ext2_htree_insert_entry(&info, split_hash, blknum);
 910
 911         /* Write the new directory block to the end of the directory */
 912         ext2_dirent_csum_set(ip, (struct ext2fs_direct_2 *)newdirblock);
 913         error = ext2_htree_append_block(dvp, newdirblock, cnp, blksize);
 914         if (error)
 915                 goto finish;
 916
 917         /* Write the target directory block */
 918         ext2_dirent_csum_set(ip, (struct ext2fs_direct_2 *)bp->b_data);
 919         error = bwrite(bp);
 920         ip->i_flag |= IN_CHANGE | IN_UPDATE;
 921         if (error)
 922                 goto finish;
 923         write_bp = 1;
 924
 925         /* Write the index block */
 926         error = ext2_htree_writebuf(ip, &info);
 927         if (!error)
 928                 write_info = 1;
 929
 930 finish:
 931         if (dst_bp != NULL && !write_dst_bp)
 932                 brelse(dst_bp);
 933         if (bp != NULL && !write_bp)
 934                 brelse(bp);
 935         if (newdirblock != NULL)
 936                 free(newdirblock, M_TEMP);
 937         if (newidxblock != NULL)
 938                 free(newidxblock, M_TEMP);
 939         if (!write_info)
 940                 ext2_htree_release(&info);
 941         return (error);
 942 }