2 * rehash.c --- rebuild hash tree directories
4 * Copyright (C) 2002 Theodore Ts'o
7 * This file may be redistributed under the terms of the GNU Public
11 * This algorithm is designed for simplicity of implementation and to
12 * pack the directory as much as possible. It however requires twice
13 * as much memory as the size of the directory. The maximum size
14 * directory supported using a 4k blocksize is roughly a gigabyte, and
15 * so there may very well be problems with machines that don't have
16 * virtual memory, and obscenely large directories.
18 * An alternate algorithm which is much more disk intensive could be
19 * written, and probably will need to be written in the future. The
20 * design goals of such an algorithm are: (a) use (roughly) constant
21 * amounts of memory, no matter how large the directory, (b) the
22 * directory must be safe at all times, even if e2fsck is interrupted
23 * in the middle, (c) we must use minimal amounts of extra disk
24 * blocks. This pretty much requires an incremental approach, where
25 * we are reading from one part of the directory, and inserting into
26 * the front half. So the algorithm will have to keep track of a
27 * moving block boundary between the new tree and the old tree, and
28 * files will need to be moved from the old directory and inserted
29 * into the new tree. If the new directory requires space which isn't
30 * yet available, blocks from the beginning part of the old directory
31 * may need to be moved to the end of the directory to make room for
34 * --------------------------------------------------------
35 * | new tree | | old tree |
36 * --------------------------------------------------------
40 * This is going to be a pain in the tuckus to implement, and will
41 * require a lot more disk accesses. So I'm going to skip it for now;
42 * it's only really going to be an issue for really, really big
43 * filesystems (when we reach the level of tens of millions of files
44 * in a single directory). It will probably be easier to simply
45 * require that e2fsck use VM first.
54 struct fill_dir_struct {
56 struct ext2_inode *inode;
59 struct hash_entry *harray;
60 int max_array, num_array;
68 ext2_dirhash_t minor_hash;
70 struct ext2_dir_entry *dir;
77 ext2_dirhash_t *hashes;
80 static int fill_dir_block(ext2_filsys fs,
83 blk_t ref_block EXT2FS_ATTR((unused)),
84 int ref_offset EXT2FS_ATTR((unused)),
87 struct fill_dir_struct *fd = (struct fill_dir_struct *) priv_data;
88 struct hash_entry *new_array, *ent;
89 struct ext2_dir_entry *dirent;
91 unsigned int offset, dir_offset;
97 offset = blockcnt * fs->blocksize;
98 if (offset + fs->blocksize > fd->inode->i_size) {
99 fd->err = EXT2_ET_DIR_CORRUPTED;
102 dir = (fd->buf+offset);
103 if (HOLE_BLKADDR(*block_nr)) {
104 memset(dir, 0, fs->blocksize);
105 dirent = (struct ext2_dir_entry *) dir;
106 dirent->rec_len = fs->blocksize;
108 fd->err = ext2fs_read_dir_block(fs, *block_nr, dir);
112 hash_alg = fs->super->s_def_hash_version;
113 if ((hash_alg <= EXT2_HASH_TEA) &&
114 (fs->super->s_flags & EXT2_FLAGS_UNSIGNED_HASH))
116 /* While the directory block is "hot", index it. */
118 while (dir_offset < fs->blocksize) {
119 dirent = (struct ext2_dir_entry *) (dir + dir_offset);
120 if (((dir_offset + dirent->rec_len) > fs->blocksize) ||
121 (dirent->rec_len < 8) ||
122 ((dirent->rec_len % 4) != 0) ||
123 (((dirent->name_len & 0xFF)+8) > dirent->rec_len)) {
124 fd->err = EXT2_ET_DIR_CORRUPTED;
127 dir_offset += dirent->rec_len;
128 if (dirent->inode == 0)
130 if (!fd->compress && ((dirent->name_len&0xFF) == 1) &&
131 (dirent->name[0] == '.'))
133 if (!fd->compress && ((dirent->name_len&0xFF) == 2) &&
134 (dirent->name[0] == '.') && (dirent->name[1] == '.')) {
135 fd->parent = dirent->inode;
138 if (fd->num_array >= fd->max_array) {
139 new_array = realloc(fd->harray,
140 sizeof(struct hash_entry) * (fd->max_array+500));
145 fd->harray = new_array;
146 fd->max_array += 500;
148 ent = fd->harray + fd->num_array++;
150 fd->dir_size += EXT2_DIR_REC_LEN(dirent->name_len & 0xFF);
151 ent->ino = dirent->inode;
153 ent->hash = ent->minor_hash = 0;
155 fd->err = ext2fs_dirhash(hash_alg, dirent->name,
156 dirent->name_len & 0xFF,
157 fs->super->s_hash_seed,
158 &ent->hash, &ent->minor_hash);
167 /* Used for sorting the hash entry */
168 static EXT2_QSORT_TYPE ino_cmp(const void *a, const void *b)
170 const struct hash_entry *he_a = (const struct hash_entry *) a;
171 const struct hash_entry *he_b = (const struct hash_entry *) b;
173 return (he_a->ino - he_b->ino);
176 /* Used for sorting the hash entry */
177 static EXT2_QSORT_TYPE name_cmp(const void *a, const void *b)
179 const struct hash_entry *he_a = (const struct hash_entry *) a;
180 const struct hash_entry *he_b = (const struct hash_entry *) b;
184 min_len = he_a->dir->name_len;
185 if (min_len > he_b->dir->name_len)
186 min_len = he_b->dir->name_len;
188 ret = strncmp(he_a->dir->name, he_b->dir->name, min_len);
190 if (he_a->dir->name_len > he_b->dir->name_len)
192 else if (he_a->dir->name_len < he_b->dir->name_len)
195 ret = he_b->dir->inode - he_a->dir->inode;
200 /* Used for sorting the hash entry */
201 static EXT2_QSORT_TYPE hash_cmp(const void *a, const void *b)
203 const struct hash_entry *he_a = (const struct hash_entry *) a;
204 const struct hash_entry *he_b = (const struct hash_entry *) b;
207 if (he_a->hash > he_b->hash)
209 else if (he_a->hash < he_b->hash)
212 if (he_a->minor_hash > he_b->minor_hash)
214 else if (he_a->minor_hash < he_b->minor_hash)
217 ret = name_cmp(a, b);
222 static errcode_t alloc_size_dir(ext2_filsys fs, struct out_dir *outdir,
228 new_mem = realloc(outdir->buf, blocks * fs->blocksize);
231 outdir->buf = new_mem;
232 new_mem = realloc(outdir->hashes,
233 blocks * sizeof(ext2_dirhash_t));
236 outdir->hashes = new_mem;
238 outdir->buf = malloc(blocks * fs->blocksize);
239 outdir->hashes = malloc(blocks * sizeof(ext2_dirhash_t));
242 outdir->max = blocks;
246 static void free_out_dir(struct out_dir *outdir)
251 free(outdir->hashes);
256 static errcode_t get_next_block(ext2_filsys fs, struct out_dir *outdir,
261 if (outdir->num >= outdir->max) {
262 retval = alloc_size_dir(fs, outdir, outdir->max + 50);
266 *ret = outdir->buf + (outdir->num++ * fs->blocksize);
267 memset(*ret, 0, fs->blocksize);
272 * This function is used to make a unique filename. We do this by
273 * appending ~0, and then incrementing the number. However, we cannot
274 * expand the length of the filename beyond the padding available in
275 * the directory entry.
277 static void mutate_name(char *str, __u16 *len)
280 __u16 l = *len & 0xFF, h = *len & 0xff00;
283 * First check to see if it looks the name has been mutated
286 for (i = l-1; i > 0; i--) {
287 if (!isdigit(str[i]))
290 if ((i == l-1) || (str[i] != '~')) {
300 for (i = l-1; i >= 0; i--) {
301 if (isdigit(str[i])) {
313 else if (str[0] == 'Z') {
331 static int duplicate_search_and_fix(e2fsck_t ctx, ext2_filsys fs,
333 struct fill_dir_struct *fd)
335 struct problem_context pctx;
336 struct hash_entry *ent, *prev;
343 clear_problem_context(&pctx);
346 hash_alg = fs->super->s_def_hash_version;
347 if ((hash_alg <= EXT2_HASH_TEA) &&
348 (fs->super->s_flags & EXT2_FLAGS_UNSIGNED_HASH))
351 for (i=1; i < fd->num_array; i++) {
352 ent = fd->harray + i;
354 if (!ent->dir->inode ||
355 ((ent->dir->name_len & 0xFF) !=
356 (prev->dir->name_len & 0xFF)) ||
357 (strncmp(ent->dir->name, prev->dir->name,
358 ent->dir->name_len & 0xFF)))
360 pctx.dirent = ent->dir;
361 if ((ent->dir->inode == prev->dir->inode) &&
362 fix_problem(ctx, PR_2_DUPLICATE_DIRENT, &pctx)) {
363 e2fsck_adjust_inode_count(ctx, ent->dir->inode, -1);
368 memcpy(new_name, ent->dir->name, ent->dir->name_len & 0xFF);
369 new_len = ent->dir->name_len;
370 mutate_name(new_name, &new_len);
371 for (j=0; j < fd->num_array; j++) {
373 ((ent->dir->name_len & 0xFF) !=
374 (fd->harray[j].dir->name_len & 0xFF)) ||
375 (strncmp(new_name, fd->harray[j].dir->name,
378 mutate_name(new_name, &new_len);
382 new_name[new_len & 0xFF] = 0;
384 if (fix_problem(ctx, PR_2_NON_UNIQUE_FILE, &pctx)) {
385 memcpy(ent->dir->name, new_name, new_len & 0xFF);
386 ent->dir->name_len = new_len;
387 ext2fs_dirhash(hash_alg, ent->dir->name,
388 ent->dir->name_len & 0xFF,
389 fs->super->s_hash_seed,
390 &ent->hash, &ent->minor_hash);
398 static errcode_t copy_dir_entries(ext2_filsys fs,
399 struct fill_dir_struct *fd,
400 struct out_dir *outdir)
404 struct hash_entry *ent;
405 struct ext2_dir_entry *dirent;
406 int i, rec_len, left;
407 ext2_dirhash_t prev_hash;
411 retval = alloc_size_dir(fs, outdir,
412 (fd->dir_size / fs->blocksize) + 2);
415 outdir->num = fd->compress ? 0 : 1;
417 outdir->hashes[0] = 0;
419 if ((retval = get_next_block(fs, outdir, &block_start)))
421 dirent = (struct ext2_dir_entry *) block_start;
422 left = fs->blocksize;
423 for (i=0; i < fd->num_array; i++) {
424 ent = fd->harray + i;
425 if (ent->dir->inode == 0)
427 rec_len = EXT2_DIR_REC_LEN(ent->dir->name_len & 0xFF);
428 if (rec_len > left) {
430 dirent->rec_len += left;
431 if ((retval = get_next_block(fs, outdir,
436 left = fs->blocksize - offset;
437 dirent = (struct ext2_dir_entry *) (block_start + offset);
439 if (ent->hash == prev_hash)
440 outdir->hashes[outdir->num-1] = ent->hash | 1;
442 outdir->hashes[outdir->num-1] = ent->hash;
444 dirent->inode = ent->dir->inode;
445 dirent->name_len = ent->dir->name_len;
446 dirent->rec_len = rec_len;
447 memcpy(dirent->name, ent->dir->name, dirent->name_len & 0xFF);
451 dirent->rec_len += left;
455 prev_hash = ent->hash;
458 dirent->rec_len += left;
464 static struct ext2_dx_root_info *set_root_node(ext2_filsys fs, char *buf,
465 ext2_ino_t ino, ext2_ino_t parent)
467 struct ext2_dir_entry *dir;
468 struct ext2_dx_root_info *root;
469 struct ext2_dx_countlimit *limits;
472 if (fs->super->s_feature_incompat & EXT2_FEATURE_INCOMPAT_FILETYPE)
473 filetype = EXT2_FT_DIR << 8;
475 memset(buf, 0, fs->blocksize);
476 dir = (struct ext2_dir_entry *) buf;
479 dir->name_len = 1 | filetype;
481 dir = (struct ext2_dir_entry *) (buf + 12);
485 dir->name_len = 2 | filetype;
486 dir->rec_len = fs->blocksize - 12;
488 root = (struct ext2_dx_root_info *) (buf+24);
489 root->reserved_zero = 0;
490 root->hash_version = fs->super->s_def_hash_version;
491 root->info_length = 8;
492 root->indirect_levels = 0;
493 root->unused_flags = 0;
495 limits = (struct ext2_dx_countlimit *) (buf+32);
496 limits->limit = (fs->blocksize - 32) / sizeof(struct ext2_dx_entry);
503 static struct ext2_dx_entry *set_int_node(ext2_filsys fs, char *buf)
505 struct ext2_dir_entry *dir;
506 struct ext2_dx_countlimit *limits;
508 memset(buf, 0, fs->blocksize);
509 dir = (struct ext2_dir_entry *) buf;
511 dir->rec_len = fs->blocksize;
513 limits = (struct ext2_dx_countlimit *) (buf+8);
514 limits->limit = (fs->blocksize - 8) / sizeof(struct ext2_dx_entry);
517 return (struct ext2_dx_entry *) limits;
521 * This function takes the leaf nodes which have been written in
522 * outdir, and populates the root node and any necessary interior nodes.
524 static errcode_t calculate_tree(ext2_filsys fs,
525 struct out_dir *outdir,
529 struct ext2_dx_root_info *root_info;
530 struct ext2_dx_entry *root, *dx_ent = 0;
531 struct ext2_dx_countlimit *root_limit, *limit;
534 int i, c1, c2, nblks;
535 int limit_offset, root_offset;
537 root_info = set_root_node(fs, outdir->buf, ino, parent);
538 root_offset = limit_offset = ((char *) root_info - outdir->buf) +
539 root_info->info_length;
540 root_limit = (struct ext2_dx_countlimit *) (outdir->buf + limit_offset);
541 c1 = root_limit->limit;
544 /* Write out the pointer blocks */
546 /* Just write out the root block, and we're done */
547 root = (struct ext2_dx_entry *) (outdir->buf + root_offset);
548 for (i=1; i < nblks; i++) {
549 root->block = ext2fs_cpu_to_le32(i);
552 ext2fs_cpu_to_le32(outdir->hashes[i]);
559 root_info->indirect_levels = 1;
560 for (i=1; i < nblks; i++) {
565 limit->limit = limit->count =
566 ext2fs_cpu_to_le16(limit->limit);
567 root = (struct ext2_dx_entry *)
568 (outdir->buf + root_offset);
569 root->block = ext2fs_cpu_to_le32(outdir->num);
572 ext2fs_cpu_to_le32(outdir->hashes[i]);
573 if ((retval = get_next_block(fs, outdir,
576 dx_ent = set_int_node(fs, block_start);
577 limit = (struct ext2_dx_countlimit *) dx_ent;
579 root_offset += sizeof(struct ext2_dx_entry);
582 dx_ent->block = ext2fs_cpu_to_le32(i);
583 if (c2 != limit->limit)
585 ext2fs_cpu_to_le32(outdir->hashes[i]);
589 limit->count = ext2fs_cpu_to_le16(limit->limit - c2);
590 limit->limit = ext2fs_cpu_to_le16(limit->limit);
592 root_limit = (struct ext2_dx_countlimit *) (outdir->buf + limit_offset);
593 root_limit->count = ext2fs_cpu_to_le16(root_limit->limit - c1);
594 root_limit->limit = ext2fs_cpu_to_le16(root_limit->limit);
599 struct write_dir_struct {
600 struct out_dir *outdir;
607 * Helper function which writes out a directory block.
609 static int write_dir_block(ext2_filsys fs,
611 e2_blkcnt_t blockcnt,
612 blk_t ref_block EXT2FS_ATTR((unused)),
613 int ref_offset EXT2FS_ATTR((unused)),
616 struct write_dir_struct *wd = (struct write_dir_struct *) priv_data;
622 if (blockcnt >= wd->outdir->num) {
623 e2fsck_read_bitmaps(wd->ctx);
625 ext2fs_unmark_block_bitmap(wd->ctx->block_found_map, blk);
626 ext2fs_block_alloc_stats(fs, blk, -1);
629 return BLOCK_CHANGED;
634 dir = wd->outdir->buf + (blockcnt * fs->blocksize);
635 wd->err = ext2fs_write_dir_block(fs, *block_nr, dir);
641 static errcode_t write_directory(e2fsck_t ctx, ext2_filsys fs,
642 struct out_dir *outdir,
643 ext2_ino_t ino, int compress)
645 struct write_dir_struct wd;
647 struct ext2_inode inode;
649 retval = e2fsck_expand_directory(ctx, ino, -1, outdir->num);
658 retval = ext2fs_block_iterate2(fs, ino, 0, 0,
659 write_dir_block, &wd);
665 e2fsck_read_inode(ctx, ino, &inode, "rehash_dir");
667 inode.i_flags &= ~EXT2_INDEX_FL;
669 inode.i_flags |= EXT2_INDEX_FL;
670 inode.i_size = outdir->num * fs->blocksize;
671 ext2fs_iblk_sub_blocks(fs, &inode, wd.cleared);
672 e2fsck_write_inode(ctx, ino, &inode, "rehash_dir");
677 errcode_t e2fsck_rehash_dir(e2fsck_t ctx, ext2_ino_t ino)
679 ext2_filsys fs = ctx->fs;
681 struct ext2_inode inode;
683 struct fill_dir_struct fd;
684 struct out_dir outdir;
686 outdir.max = outdir.num = 0;
689 e2fsck_read_inode(ctx, ino, &inode, "rehash_dir");
693 dir_buf = malloc(inode.i_size);
697 fd.max_array = inode.i_size / 32;
699 fd.harray = malloc(fd.max_array * sizeof(struct hash_entry));
709 if (!(fs->super->s_feature_compat & EXT2_FEATURE_COMPAT_DIR_INDEX) ||
710 (inode.i_size / fs->blocksize) < 2)
714 /* Read in the entire directory into memory */
715 retval = ext2fs_block_iterate2(fs, ino, 0, 0,
716 fill_dir_block, &fd);
723 printf("%d entries (%d bytes) found in inode %d\n",
724 fd.num_array, fd.dir_size, ino);
730 qsort(fd.harray+2, fd.num_array-2,
731 sizeof(struct hash_entry), ino_cmp);
733 qsort(fd.harray, fd.num_array,
734 sizeof(struct hash_entry), hash_cmp);
737 * Look for duplicates
739 if (duplicate_search_and_fix(ctx, fs, ino, &fd))
742 if (ctx->options & E2F_OPT_NO) {
748 * Copy the directory entries. In a htree directory these
749 * will become the leaf nodes.
751 retval = copy_dir_entries(fs, &fd, &outdir);
755 free(dir_buf); dir_buf = 0;
758 /* Calculate the interior nodes */
759 retval = calculate_tree(fs, &outdir, ino, fd.parent);
764 retval = write_directory(ctx, fs, &outdir, ino, fd.compress);
774 free_out_dir(&outdir);
778 void e2fsck_rehash_directories(e2fsck_t ctx)
780 struct problem_context pctx;
781 #ifdef RESOURCE_TRACK
782 struct resource_track rtrack;
784 struct dir_info *dir;
785 ext2_u32_iterate iter;
786 struct dir_info_iter * dirinfo_iter = 0;
789 int cur, max, all_dirs, dir_index, first = 1;
791 #ifdef RESOURCE_TRACK
792 init_resource_track(&rtrack, ctx->fs->io);
795 all_dirs = ctx->options & E2F_OPT_COMPRESS_DIRS;
797 if (!ctx->dirs_to_hash && !all_dirs)
800 e2fsck_get_lost_and_found(ctx, 0);
802 clear_problem_context(&pctx);
804 dir_index = ctx->fs->super->s_feature_compat & EXT2_FEATURE_COMPAT_DIR_INDEX;
807 dirinfo_iter = e2fsck_dir_info_iter_begin(ctx);
808 max = e2fsck_get_num_dirinfo(ctx);
810 retval = ext2fs_u32_list_iterate_begin(ctx->dirs_to_hash,
813 pctx.errcode = retval;
814 fix_problem(ctx, PR_3A_OPTIMIZE_ITER, &pctx);
817 max = ext2fs_u32_list_count(ctx->dirs_to_hash);
821 if ((dir = e2fsck_dir_info_iter(ctx,
826 if (!ext2fs_u32_list_iterate(iter, &ino))
829 if (ino == ctx->lost_and_found)
833 fix_problem(ctx, PR_3A_PASS_HEADER, &pctx);
837 fix_problem(ctx, PR_3A_OPTIMIZE_DIR, &pctx);
839 pctx.errcode = e2fsck_rehash_dir(ctx, ino);
841 end_problem_latch(ctx, PR_LATCH_OPTIMIZE_DIR);
842 fix_problem(ctx, PR_3A_OPTIMIZE_DIR_ERR, &pctx);
844 if (ctx->progress && !ctx->progress_fd)
845 e2fsck_simple_progress(ctx, "Rebuilding directory",
846 100.0 * (float) (++cur) / (float) max, ino);
848 end_problem_latch(ctx, PR_LATCH_OPTIMIZE_DIR);
850 e2fsck_dir_info_iter_end(ctx, dirinfo_iter);
852 ext2fs_u32_list_iterate_end(iter);
854 if (ctx->dirs_to_hash)
855 ext2fs_u32_list_free(ctx->dirs_to_hash);
856 ctx->dirs_to_hash = 0;
858 #ifdef RESOURCE_TRACK
859 if (ctx->options & E2F_OPT_TIME2) {
860 e2fsck_clear_progbar(ctx);
861 print_resource_track("Pass 3A", &rtrack, ctx->fs->io);