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.
55 /* Schedule a dir to be rebuilt during pass 3A. */
56 void e2fsck_rehash_dir_later(e2fsck_t ctx, ext2_ino_t ino)
58 if (!ctx->dirs_to_hash)
59 ext2fs_u32_list_create(&ctx->dirs_to_hash, 50);
60 if (ctx->dirs_to_hash)
61 ext2fs_u32_list_add(ctx->dirs_to_hash, ino);
64 /* Ask if a dir will be rebuilt during pass 3A. */
65 int e2fsck_dir_will_be_rehashed(e2fsck_t ctx, ext2_ino_t ino)
67 if (ctx->options & E2F_OPT_COMPRESS_DIRS)
69 if (!ctx->dirs_to_hash)
71 return ext2fs_u32_list_test(ctx->dirs_to_hash, ino);
74 struct fill_dir_struct {
76 struct ext2_inode *inode;
79 struct hash_entry *harray;
80 int max_array, num_array;
89 ext2_dirhash_t minor_hash;
91 struct ext2_dir_entry *dir;
98 ext2_dirhash_t *hashes;
101 static int fill_dir_block(ext2_filsys fs,
103 e2_blkcnt_t blockcnt,
104 blk64_t ref_block EXT2FS_ATTR((unused)),
105 int ref_offset EXT2FS_ATTR((unused)),
108 struct fill_dir_struct *fd = (struct fill_dir_struct *) priv_data;
109 struct hash_entry *new_array, *ent;
110 struct ext2_dir_entry *dirent;
112 unsigned int offset, dir_offset, rec_len;
118 offset = blockcnt * fs->blocksize;
119 if (offset + fs->blocksize > fd->inode->i_size) {
120 fd->err = EXT2_ET_DIR_CORRUPTED;
123 dir = (fd->buf+offset);
124 if (HOLE_BLKADDR(*block_nr)) {
125 memset(dir, 0, fs->blocksize);
126 dirent = (struct ext2_dir_entry *) dir;
127 (void) ext2fs_set_rec_len(fs, fs->blocksize, dirent);
129 fs->flags |= EXT2_FLAG_IGNORE_CSUM_ERRORS;
130 fd->err = ext2fs_read_dir_block4(fs, *block_nr, dir, 0,
132 fs->flags &= ~EXT2_FLAG_IGNORE_CSUM_ERRORS;
136 hash_alg = fs->super->s_def_hash_version;
137 if ((hash_alg <= EXT2_HASH_TEA) &&
138 (fs->super->s_flags & EXT2_FLAGS_UNSIGNED_HASH))
140 /* While the directory block is "hot", index it. */
142 while (dir_offset < fs->blocksize) {
143 dirent = (struct ext2_dir_entry *) (dir + dir_offset);
144 (void) ext2fs_get_rec_len(fs, dirent, &rec_len);
145 if (((dir_offset + rec_len) > fs->blocksize) ||
147 ((rec_len % 4) != 0) ||
148 (((dirent->name_len & 0xFF)+8) > rec_len)) {
149 fd->err = EXT2_ET_DIR_CORRUPTED;
152 dir_offset += rec_len;
153 if (dirent->inode == 0)
155 if (!fd->compress && ((dirent->name_len&0xFF) == 1) &&
156 (dirent->name[0] == '.'))
158 if (!fd->compress && ((dirent->name_len&0xFF) == 2) &&
159 (dirent->name[0] == '.') && (dirent->name[1] == '.')) {
160 fd->parent = dirent->inode;
163 if (fd->num_array >= fd->max_array) {
164 new_array = realloc(fd->harray,
165 sizeof(struct hash_entry) * (fd->max_array+500));
170 fd->harray = new_array;
171 fd->max_array += 500;
173 ent = fd->harray + fd->num_array++;
175 fd->dir_size += EXT2_DIR_REC_LEN(dirent->name_len & 0xFF);
176 ent->ino = dirent->inode;
178 ent->hash = ent->minor_hash = 0;
180 fd->err = ext2fs_dirhash(hash_alg, dirent->name,
181 dirent->name_len & 0xFF,
182 fs->super->s_hash_seed,
183 &ent->hash, &ent->minor_hash);
192 /* Used for sorting the hash entry */
193 static EXT2_QSORT_TYPE ino_cmp(const void *a, const void *b)
195 const struct hash_entry *he_a = (const struct hash_entry *) a;
196 const struct hash_entry *he_b = (const struct hash_entry *) b;
198 return (he_a->ino - he_b->ino);
201 /* Used for sorting the hash entry */
202 static EXT2_QSORT_TYPE name_cmp(const void *a, const void *b)
204 const struct hash_entry *he_a = (const struct hash_entry *) a;
205 const struct hash_entry *he_b = (const struct hash_entry *) b;
209 min_len = he_a->dir->name_len;
210 if (min_len > he_b->dir->name_len)
211 min_len = he_b->dir->name_len;
213 ret = strncmp(he_a->dir->name, he_b->dir->name, min_len);
215 if (he_a->dir->name_len > he_b->dir->name_len)
217 else if (he_a->dir->name_len < he_b->dir->name_len)
220 ret = he_b->dir->inode - he_a->dir->inode;
225 /* Used for sorting the hash entry */
226 static EXT2_QSORT_TYPE hash_cmp(const void *a, const void *b)
228 const struct hash_entry *he_a = (const struct hash_entry *) a;
229 const struct hash_entry *he_b = (const struct hash_entry *) b;
232 if (he_a->hash > he_b->hash)
234 else if (he_a->hash < he_b->hash)
237 if (he_a->minor_hash > he_b->minor_hash)
239 else if (he_a->minor_hash < he_b->minor_hash)
242 ret = name_cmp(a, b);
247 static errcode_t alloc_size_dir(ext2_filsys fs, struct out_dir *outdir,
253 new_mem = realloc(outdir->buf, blocks * fs->blocksize);
256 outdir->buf = new_mem;
257 new_mem = realloc(outdir->hashes,
258 blocks * sizeof(ext2_dirhash_t));
261 outdir->hashes = new_mem;
263 outdir->buf = malloc(blocks * fs->blocksize);
264 outdir->hashes = malloc(blocks * sizeof(ext2_dirhash_t));
267 outdir->max = blocks;
271 static void free_out_dir(struct out_dir *outdir)
274 free(outdir->hashes);
279 static errcode_t get_next_block(ext2_filsys fs, struct out_dir *outdir,
284 if (outdir->num >= outdir->max) {
285 retval = alloc_size_dir(fs, outdir, outdir->max + 50);
289 *ret = outdir->buf + (outdir->num++ * fs->blocksize);
290 memset(*ret, 0, fs->blocksize);
295 * This function is used to make a unique filename. We do this by
296 * appending ~0, and then incrementing the number. However, we cannot
297 * expand the length of the filename beyond the padding available in
298 * the directory entry.
300 static void mutate_name(char *str, __u16 *len)
303 __u16 l = *len & 0xFF, h = *len & 0xff00;
306 * First check to see if it looks the name has been mutated
309 for (i = l-1; i > 0; i--) {
310 if (!isdigit(str[i]))
313 if ((i == l-1) || (str[i] != '~')) {
323 for (i = l-1; i >= 0; i--) {
324 if (isdigit(str[i])) {
336 else if (str[0] == 'Z') {
354 static int duplicate_search_and_fix(e2fsck_t ctx, ext2_filsys fs,
356 struct fill_dir_struct *fd)
358 struct problem_context pctx;
359 struct hash_entry *ent, *prev;
366 clear_problem_context(&pctx);
369 hash_alg = fs->super->s_def_hash_version;
370 if ((hash_alg <= EXT2_HASH_TEA) &&
371 (fs->super->s_flags & EXT2_FLAGS_UNSIGNED_HASH))
374 for (i=1; i < fd->num_array; i++) {
375 ent = fd->harray + i;
377 if (!ent->dir->inode ||
378 ((ent->dir->name_len & 0xFF) !=
379 (prev->dir->name_len & 0xFF)) ||
380 (strncmp(ent->dir->name, prev->dir->name,
381 ent->dir->name_len & 0xFF)))
383 pctx.dirent = ent->dir;
384 if ((ent->dir->inode == prev->dir->inode) &&
385 fix_problem(ctx, PR_2_DUPLICATE_DIRENT, &pctx)) {
386 e2fsck_adjust_inode_count(ctx, ent->dir->inode, -1);
391 memcpy(new_name, ent->dir->name, ent->dir->name_len & 0xFF);
392 new_len = ent->dir->name_len;
393 mutate_name(new_name, &new_len);
394 for (j=0; j < fd->num_array; j++) {
397 (fd->harray[j].dir->name_len & 0xFF)) ||
398 (strncmp(new_name, fd->harray[j].dir->name,
401 mutate_name(new_name, &new_len);
405 new_name[new_len & 0xFF] = 0;
407 if (fix_problem(ctx, PR_2_NON_UNIQUE_FILE, &pctx)) {
408 memcpy(ent->dir->name, new_name, new_len & 0xFF);
409 ent->dir->name_len = new_len;
410 ext2fs_dirhash(hash_alg, ent->dir->name,
411 ent->dir->name_len & 0xFF,
412 fs->super->s_hash_seed,
413 &ent->hash, &ent->minor_hash);
421 static errcode_t copy_dir_entries(e2fsck_t ctx,
422 struct fill_dir_struct *fd,
423 struct out_dir *outdir,
426 ext2_filsys fs = ctx->fs;
429 struct hash_entry *ent;
430 struct ext2_dir_entry *dirent;
431 unsigned int rec_len, prev_rec_len;
433 ext2_dirhash_t prev_hash;
436 struct ext2_dir_entry_tail *t;
438 if (ctx->htree_slack_percentage == 255) {
439 profile_get_uint(ctx->profile, "options",
440 "indexed_dir_slack_percentage",
442 &ctx->htree_slack_percentage);
443 if (ctx->htree_slack_percentage > 100)
444 ctx->htree_slack_percentage = 20;
447 if (EXT2_HAS_RO_COMPAT_FEATURE(fs->super,
448 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
449 csum_size = sizeof(struct ext2_dir_entry_tail);
452 retval = alloc_size_dir(fs, outdir,
453 (fd->dir_size / fs->blocksize) + 2);
456 outdir->num = fd->compress ? 0 : 1;
458 outdir->hashes[0] = 0;
460 if ((retval = get_next_block(fs, outdir, &block_start)))
462 dirent = (struct ext2_dir_entry *) block_start;
465 left = fs->blocksize - csum_size;
466 slack = fd->compress ? 12 :
467 ((fs->blocksize - csum_size) * ctx->htree_slack_percentage)/100;
470 for (i = 0; i < fd->num_array; i++) {
471 ent = fd->harray + i;
472 if (ent->dir->inode == 0)
474 rec_len = EXT2_DIR_REC_LEN(ent->dir->name_len & 0xFF);
475 if (rec_len > left) {
477 left += prev_rec_len;
478 retval = ext2fs_set_rec_len(fs, left, dirent);
483 t = EXT2_DIRENT_TAIL(block_start,
485 ext2fs_initialize_dirent_tail(fs, t);
487 if ((retval = get_next_block(fs, outdir,
492 left = (fs->blocksize - csum_size) - offset;
493 dirent = (struct ext2_dir_entry *) (block_start + offset);
495 if (ent->hash == prev_hash)
496 outdir->hashes[outdir->num-1] = ent->hash | 1;
498 outdir->hashes[outdir->num-1] = ent->hash;
500 dirent->inode = ent->dir->inode;
501 dirent->name_len = ent->dir->name_len;
502 retval = ext2fs_set_rec_len(fs, rec_len, dirent);
505 prev_rec_len = rec_len;
506 memcpy(dirent->name, ent->dir->name, dirent->name_len & 0xFF);
510 prev_rec_len += left;
511 retval = ext2fs_set_rec_len(fs, prev_rec_len, dirent);
517 prev_hash = ent->hash;
520 retval = ext2fs_set_rec_len(fs, rec_len + left, dirent);
522 t = EXT2_DIRENT_TAIL(block_start, fs->blocksize);
523 ext2fs_initialize_dirent_tail(fs, t);
530 static struct ext2_dx_root_info *set_root_node(ext2_filsys fs, char *buf,
531 ext2_ino_t ino, ext2_ino_t parent)
533 struct ext2_dir_entry *dir;
534 struct ext2_dx_root_info *root;
535 struct ext2_dx_countlimit *limits;
539 if (fs->super->s_feature_incompat & EXT2_FEATURE_INCOMPAT_FILETYPE)
540 filetype = EXT2_FT_DIR << 8;
542 memset(buf, 0, fs->blocksize);
543 dir = (struct ext2_dir_entry *) buf;
546 dir->name_len = 1 | filetype;
548 dir = (struct ext2_dir_entry *) (buf + 12);
552 dir->name_len = 2 | filetype;
553 dir->rec_len = fs->blocksize - 12;
555 root = (struct ext2_dx_root_info *) (buf+24);
556 root->reserved_zero = 0;
557 root->hash_version = fs->super->s_def_hash_version;
558 root->info_length = 8;
559 root->indirect_levels = 0;
560 root->unused_flags = 0;
562 if (EXT2_HAS_RO_COMPAT_FEATURE(fs->super,
563 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
564 csum_size = sizeof(struct ext2_dx_tail);
566 limits = (struct ext2_dx_countlimit *) (buf+32);
567 limits->limit = (fs->blocksize - (32 + csum_size)) /
568 sizeof(struct ext2_dx_entry);
575 static struct ext2_dx_entry *set_int_node(ext2_filsys fs, char *buf)
577 struct ext2_dir_entry *dir;
578 struct ext2_dx_countlimit *limits;
581 memset(buf, 0, fs->blocksize);
582 dir = (struct ext2_dir_entry *) buf;
584 (void) ext2fs_set_rec_len(fs, fs->blocksize, dir);
586 if (EXT2_HAS_RO_COMPAT_FEATURE(fs->super,
587 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
588 csum_size = sizeof(struct ext2_dx_tail);
590 limits = (struct ext2_dx_countlimit *) (buf+8);
591 limits->limit = (fs->blocksize - (8 + csum_size)) /
592 sizeof(struct ext2_dx_entry);
595 return (struct ext2_dx_entry *) limits;
599 * This function takes the leaf nodes which have been written in
600 * outdir, and populates the root node and any necessary interior nodes.
602 static errcode_t calculate_tree(ext2_filsys fs,
603 struct out_dir *outdir,
607 struct ext2_dx_root_info *root_info;
608 struct ext2_dx_entry *root, *dx_ent = 0;
609 struct ext2_dx_countlimit *root_limit, *limit;
612 int i, c1, c2, nblks;
613 int limit_offset, root_offset;
615 root_info = set_root_node(fs, outdir->buf, ino, parent);
616 root_offset = limit_offset = ((char *) root_info - outdir->buf) +
617 root_info->info_length;
618 root_limit = (struct ext2_dx_countlimit *) (outdir->buf + limit_offset);
619 c1 = root_limit->limit;
622 /* Write out the pointer blocks */
624 /* Just write out the root block, and we're done */
625 root = (struct ext2_dx_entry *) (outdir->buf + root_offset);
626 for (i=1; i < nblks; i++) {
627 root->block = ext2fs_cpu_to_le32(i);
630 ext2fs_cpu_to_le32(outdir->hashes[i]);
637 root_info->indirect_levels = 1;
638 for (i=1; i < nblks; i++) {
643 limit->limit = limit->count =
644 ext2fs_cpu_to_le16(limit->limit);
645 root = (struct ext2_dx_entry *)
646 (outdir->buf + root_offset);
647 root->block = ext2fs_cpu_to_le32(outdir->num);
650 ext2fs_cpu_to_le32(outdir->hashes[i]);
651 if ((retval = get_next_block(fs, outdir,
654 dx_ent = set_int_node(fs, block_start);
655 limit = (struct ext2_dx_countlimit *) dx_ent;
657 root_offset += sizeof(struct ext2_dx_entry);
660 dx_ent->block = ext2fs_cpu_to_le32(i);
661 if (c2 != limit->limit)
663 ext2fs_cpu_to_le32(outdir->hashes[i]);
667 limit->count = ext2fs_cpu_to_le16(limit->limit - c2);
668 limit->limit = ext2fs_cpu_to_le16(limit->limit);
670 root_limit = (struct ext2_dx_countlimit *) (outdir->buf + limit_offset);
671 root_limit->count = ext2fs_cpu_to_le16(root_limit->limit - c1);
672 root_limit->limit = ext2fs_cpu_to_le16(root_limit->limit);
677 struct write_dir_struct {
678 struct out_dir *outdir;
686 * Helper function which writes out a directory block.
688 static int write_dir_block(ext2_filsys fs,
690 e2_blkcnt_t blockcnt,
691 blk64_t ref_block EXT2FS_ATTR((unused)),
692 int ref_offset EXT2FS_ATTR((unused)),
695 struct write_dir_struct *wd = (struct write_dir_struct *) priv_data;
703 if (blockcnt < wd->outdir->num)
704 dir = wd->outdir->buf + (blockcnt * fs->blocksize);
705 else if (wd->ctx->lost_and_found == wd->dir) {
706 /* Don't release any extra directory blocks for lost+found */
707 wd->err = ext2fs_new_dir_block(fs, 0, 0, &buf);
713 /* We don't need this block, so release it */
714 e2fsck_read_bitmaps(wd->ctx);
716 ext2fs_unmark_block_bitmap2(wd->ctx->block_found_map, blk);
717 ext2fs_block_alloc_stats2(fs, blk, -1);
720 return BLOCK_CHANGED;
723 wd->err = ext2fs_write_dir_block4(fs, *block_nr, dir, 0, wd->dir);
725 ext2fs_free_mem(&buf);
732 static errcode_t write_directory(e2fsck_t ctx, ext2_filsys fs,
733 struct out_dir *outdir,
734 ext2_ino_t ino, int compress)
736 struct write_dir_struct wd;
738 struct ext2_inode inode;
740 retval = e2fsck_expand_directory(ctx, ino, -1, outdir->num);
750 retval = ext2fs_block_iterate3(fs, ino, 0, 0,
751 write_dir_block, &wd);
757 e2fsck_read_inode(ctx, ino, &inode, "rehash_dir");
759 inode.i_flags &= ~EXT2_INDEX_FL;
761 inode.i_flags |= EXT2_INDEX_FL;
762 inode.i_size = outdir->num * fs->blocksize;
763 ext2fs_iblk_sub_blocks(fs, &inode, wd.cleared);
764 e2fsck_write_inode(ctx, ino, &inode, "rehash_dir");
769 errcode_t e2fsck_rehash_dir(e2fsck_t ctx, ext2_ino_t ino)
771 ext2_filsys fs = ctx->fs;
773 struct ext2_inode inode;
775 struct fill_dir_struct fd;
776 struct out_dir outdir;
778 outdir.max = outdir.num = 0;
781 e2fsck_read_inode(ctx, ino, &inode, "rehash_dir");
785 dir_buf = malloc(inode.i_size);
789 fd.max_array = inode.i_size / 32;
791 fd.harray = malloc(fd.max_array * sizeof(struct hash_entry));
802 if (!(fs->super->s_feature_compat & EXT2_FEATURE_COMPAT_DIR_INDEX) ||
803 (inode.i_size / fs->blocksize) < 2)
808 /* Read in the entire directory into memory */
809 retval = ext2fs_block_iterate3(fs, ino, 0, 0,
810 fill_dir_block, &fd);
817 * If the entries read are less than a block, then don't index
820 if (!fd.compress && (fd.dir_size < (fs->blocksize - 24))) {
828 printf("%d entries (%d bytes) found in inode %d\n",
829 fd.num_array, fd.dir_size, ino);
835 qsort(fd.harray+2, fd.num_array-2, sizeof(struct hash_entry),
838 qsort(fd.harray, fd.num_array, sizeof(struct hash_entry),
842 * Look for duplicates
844 if (duplicate_search_and_fix(ctx, fs, ino, &fd))
847 if (ctx->options & E2F_OPT_NO) {
852 /* Sort non-hashed directories by inode number */
854 qsort(fd.harray+2, fd.num_array-2,
855 sizeof(struct hash_entry), ino_cmp);
858 * Copy the directory entries. In a htree directory these
859 * will become the leaf nodes.
861 retval = copy_dir_entries(ctx, &fd, &outdir, ino);
865 free(dir_buf); dir_buf = 0;
868 /* Calculate the interior nodes */
869 retval = calculate_tree(fs, &outdir, ino, fd.parent);
874 retval = write_directory(ctx, fs, &outdir, ino, fd.compress);
882 free_out_dir(&outdir);
886 void e2fsck_rehash_directories(e2fsck_t ctx)
888 struct problem_context pctx;
889 #ifdef RESOURCE_TRACK
890 struct resource_track rtrack;
892 struct dir_info *dir;
893 ext2_u32_iterate iter;
894 struct dir_info_iter * dirinfo_iter = 0;
897 int cur, max, all_dirs, first = 1;
899 init_resource_track(&rtrack, ctx->fs->io);
900 all_dirs = ctx->options & E2F_OPT_COMPRESS_DIRS;
902 if (!ctx->dirs_to_hash && !all_dirs)
905 (void) e2fsck_get_lost_and_found(ctx, 0);
907 clear_problem_context(&pctx);
911 dirinfo_iter = e2fsck_dir_info_iter_begin(ctx);
912 max = e2fsck_get_num_dirinfo(ctx);
914 retval = ext2fs_u32_list_iterate_begin(ctx->dirs_to_hash,
917 pctx.errcode = retval;
918 fix_problem(ctx, PR_3A_OPTIMIZE_ITER, &pctx);
921 max = ext2fs_u32_list_count(ctx->dirs_to_hash);
925 if ((dir = e2fsck_dir_info_iter(ctx,
930 if (!ext2fs_u32_list_iterate(iter, &ino))
936 fix_problem(ctx, PR_3A_PASS_HEADER, &pctx);
940 fix_problem(ctx, PR_3A_OPTIMIZE_DIR, &pctx);
942 pctx.errcode = e2fsck_rehash_dir(ctx, ino);
944 end_problem_latch(ctx, PR_LATCH_OPTIMIZE_DIR);
945 fix_problem(ctx, PR_3A_OPTIMIZE_DIR_ERR, &pctx);
947 if (ctx->progress && !ctx->progress_fd)
948 e2fsck_simple_progress(ctx, "Rebuilding directory",
949 100.0 * (float) (++cur) / (float) max, ino);
951 end_problem_latch(ctx, PR_LATCH_OPTIMIZE_DIR);
953 e2fsck_dir_info_iter_end(ctx, dirinfo_iter);
955 ext2fs_u32_list_iterate_end(iter);
957 if (ctx->dirs_to_hash)
958 ext2fs_u32_list_free(ctx->dirs_to_hash);
959 ctx->dirs_to_hash = 0;
961 print_resource_track(ctx, "Pass 3A", &rtrack, ctx->fs->io);