2 * fs/logfs/dir.c - directory-related code
4 * As should be obvious for Linux kernel code, license is GPLv2
6 * Copyright (c) 2005-2008 Joern Engel <joern@logfs.org>
9 #include <linux/slab.h>
12 * Atomic dir operations
14 * Directory operations are by default not atomic. Dentries and Inodes are
15 * created/removed/altered in separate operations. Therefore we need to do
16 * a small amount of journaling.
18 * Create, link, mkdir, mknod and symlink all share the same function to do
19 * the work: __logfs_create. This function works in two atomic steps:
20 * 1. allocate inode (remember in journal)
21 * 2. allocate dentry (clear journal)
23 * As we can only get interrupted between the two, when the inode we just
24 * created is simply stored in the anchor. On next mount, if we were
25 * interrupted, we delete the inode. From a users point of view the
26 * operation never happened.
28 * Unlink and rmdir also share the same function: unlink. Again, this
29 * function works in two atomic steps
30 * 1. remove dentry (remember inode in journal)
31 * 2. unlink inode (clear journal)
33 * And again, on the next mount, if we were interrupted, we delete the inode.
34 * From a users point of view the operation succeeded.
36 * Rename is the real pain to deal with, harder than all the other methods
37 * combined. Depending on the circumstances we can run into three cases.
38 * A "target rename" where the target dentry already existed, a "local
39 * rename" where both parent directories are identical or a "cross-directory
40 * rename" in the remaining case.
42 * Local rename is atomic, as the old dentry is simply rewritten with a new
45 * Cross-directory rename works in two steps, similar to __logfs_create and
47 * 1. Write new dentry (remember old dentry in journal)
48 * 2. Remove old dentry (clear journal)
50 * Here we remember a dentry instead of an inode. On next mount, if we were
51 * interrupted, we delete the dentry. From a users point of view, the
52 * operation succeeded.
54 * Target rename works in three atomic steps:
55 * 1. Attach old inode to new dentry (remember old dentry and new inode)
56 * 2. Remove old dentry (still remember the new inode)
57 * 3. Remove victim inode
59 * Here we remember both an inode an a dentry. If we get interrupted
60 * between steps 1 and 2, we delete both the dentry and the inode. If
61 * we get interrupted between steps 2 and 3, we delete just the inode.
62 * In either case, the remaining objects are deleted on next mount. From
63 * a users point of view, the operation succeeded.
66 static int write_dir(struct inode *dir, struct logfs_disk_dentry *dd,
69 return logfs_inode_write(dir, dd, sizeof(*dd), pos, WF_LOCK, NULL);
72 static int write_inode(struct inode *inode)
74 return __logfs_write_inode(inode, NULL, WF_LOCK);
77 static s64 dir_seek_data(struct inode *inode, s64 pos)
79 s64 new_pos = logfs_seek_data(inode, pos);
81 return max(pos, new_pos - 1);
84 static int beyond_eof(struct inode *inode, loff_t bix)
86 loff_t pos = bix << inode->i_sb->s_blocksize_bits;
87 return pos >= i_size_read(inode);
91 * Prime value was chosen to be roughly 256 + 26. r5 hash uses 11,
92 * so short names (len <= 9) don't even occupy the complete 32bit name
93 * space. A prime >256 ensures short names quickly spread the 32bit
94 * name space. Add about 26 for the estimated amount of information
95 * of each character and pick a prime nearby, preferably a bit-sparse
98 static u32 hash_32(const char *s, int len, u32 seed)
103 for (i = 0; i < len; i++)
104 hash = hash * 293 + s[i];
109 * We have to satisfy several conflicting requirements here. Small
110 * directories should stay fairly compact and not require too many
111 * indirect blocks. The number of possible locations for a given hash
112 * should be small to make lookup() fast. And we should try hard not
113 * to overflow the 32bit name space or nfs and 32bit host systems will
116 * So we use the following scheme. First we reduce the hash to 0..15
117 * and try a direct block. If that is occupied we reduce the hash to
118 * 16..255 and try an indirect block. Same for 2x and 3x indirect
119 * blocks. Lastly we reduce the hash to 0x800_0000 .. 0xffff_ffff,
120 * but use buckets containing eight entries instead of a single one.
122 * Using 16 entries should allow for a reasonable amount of hash
123 * collisions, so the 32bit name space can be packed fairly tight
124 * before overflowing. Oh and currently we don't overflow but return
127 * How likely are collisions? Doing the appropriate math is beyond me
128 * and the Bronstein textbook. But running a test program to brute
129 * force collisions for a couple of days showed that on average the
130 * first collision occurs after 598M entries, with 290M being the
131 * smallest result. Obviously 21 entries could already cause a
132 * collision if all entries are carefully chosen.
134 static pgoff_t hash_index(u32 hash, int round)
136 u32 i0_blocks = I0_BLOCKS;
137 u32 i1_blocks = I1_BLOCKS;
138 u32 i2_blocks = I2_BLOCKS;
139 u32 i3_blocks = I3_BLOCKS;
143 return hash % i0_blocks;
145 return i0_blocks + hash % (i1_blocks - i0_blocks);
147 return i1_blocks + hash % (i2_blocks - i1_blocks);
149 return i2_blocks + hash % (i3_blocks - i2_blocks);
151 return i3_blocks + 16 * (hash % (((1<<31) - i3_blocks) / 16))
157 static struct page *logfs_get_dd_page(struct inode *dir, struct dentry *dentry)
159 struct qstr *name = &dentry->d_name;
161 struct logfs_disk_dentry *dd;
162 u32 hash = hash_32(name->name, name->len, 0);
166 if (name->len > LOGFS_MAX_NAMELEN)
167 return ERR_PTR(-ENAMETOOLONG);
169 for (round = 0; round < 20; round++) {
170 index = hash_index(hash, round);
172 if (beyond_eof(dir, index))
174 if (!logfs_exist_block(dir, index))
176 page = read_cache_page(dir->i_mapping, index,
177 (filler_t *)logfs_readpage, NULL);
180 dd = kmap_atomic(page);
181 BUG_ON(dd->namelen == 0);
183 if (name->len != be16_to_cpu(dd->namelen) ||
184 memcmp(name->name, dd->name, name->len)) {
186 page_cache_release(page);
196 static int logfs_remove_inode(struct inode *inode)
201 ret = write_inode(inode);
202 LOGFS_BUG_ON(ret, inode->i_sb);
206 static void abort_transaction(struct inode *inode, struct logfs_transaction *ta)
208 if (logfs_inode(inode)->li_block)
209 logfs_inode(inode)->li_block->ta = NULL;
213 static int logfs_unlink(struct inode *dir, struct dentry *dentry)
215 struct logfs_super *super = logfs_super(dir->i_sb);
216 struct inode *inode = d_inode(dentry);
217 struct logfs_transaction *ta;
222 ta = kzalloc(sizeof(*ta), GFP_KERNEL);
226 ta->state = UNLINK_1;
227 ta->ino = inode->i_ino;
229 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
231 page = logfs_get_dd_page(dir, dentry);
238 return PTR_ERR(page);
241 page_cache_release(page);
243 mutex_lock(&super->s_dirop_mutex);
244 logfs_add_transaction(dir, ta);
246 ret = logfs_delete(dir, index, NULL);
248 ret = write_inode(dir);
251 abort_transaction(dir, ta);
252 printk(KERN_ERR"LOGFS: unable to delete inode\n");
256 ta->state = UNLINK_2;
257 logfs_add_transaction(inode, ta);
258 ret = logfs_remove_inode(inode);
260 mutex_unlock(&super->s_dirop_mutex);
264 static inline int logfs_empty_dir(struct inode *dir)
268 data = logfs_seek_data(dir, 0) << dir->i_sb->s_blocksize_bits;
269 return data >= i_size_read(dir);
272 static int logfs_rmdir(struct inode *dir, struct dentry *dentry)
274 struct inode *inode = d_inode(dentry);
276 if (!logfs_empty_dir(inode))
279 return logfs_unlink(dir, dentry);
282 /* FIXME: readdir currently has it's own dir_walk code. I don't see a good
283 * way to combine the two copies */
284 static int logfs_readdir(struct file *file, struct dir_context *ctx)
286 struct inode *dir = file_inode(file);
289 struct logfs_disk_dentry *dd;
294 if (!dir_emit_dots(file, ctx))
299 for (;; pos++, ctx->pos++) {
301 if (beyond_eof(dir, pos))
303 if (!logfs_exist_block(dir, pos)) {
305 pos = dir_seek_data(dir, pos);
308 page = read_cache_page(dir->i_mapping, pos,
309 (filler_t *)logfs_readpage, NULL);
311 return PTR_ERR(page);
313 BUG_ON(dd->namelen == 0);
315 full = !dir_emit(ctx, (char *)dd->name,
316 be16_to_cpu(dd->namelen),
317 be64_to_cpu(dd->ino), dd->type);
319 page_cache_release(page);
326 static void logfs_set_name(struct logfs_disk_dentry *dd, struct qstr *name)
328 dd->namelen = cpu_to_be16(name->len);
329 memcpy(dd->name, name->name, name->len);
332 static struct dentry *logfs_lookup(struct inode *dir, struct dentry *dentry,
336 struct logfs_disk_dentry *dd;
341 page = logfs_get_dd_page(dir, dentry);
343 return ERR_CAST(page);
349 dd = kmap_atomic(page);
350 ino = be64_to_cpu(dd->ino);
352 page_cache_release(page);
354 inode = logfs_iget(dir->i_sb, ino);
356 printk(KERN_ERR"LogFS: Cannot read inode #%llx for dentry (%lx, %lx)n",
357 ino, dir->i_ino, index);
358 return d_splice_alias(inode, dentry);
361 static void grow_dir(struct inode *dir, loff_t index)
363 index = (index + 1) << dir->i_sb->s_blocksize_bits;
364 if (i_size_read(dir) < index)
365 i_size_write(dir, index);
368 static int logfs_write_dir(struct inode *dir, struct dentry *dentry,
372 struct logfs_disk_dentry *dd;
373 u32 hash = hash_32(dentry->d_name.name, dentry->d_name.len, 0);
377 for (round = 0; round < 20; round++) {
378 index = hash_index(hash, round);
380 if (logfs_exist_block(dir, index))
382 page = find_or_create_page(dir->i_mapping, index, GFP_KERNEL);
386 dd = kmap_atomic(page);
387 memset(dd, 0, sizeof(*dd));
388 dd->ino = cpu_to_be64(inode->i_ino);
389 dd->type = logfs_type(inode);
390 logfs_set_name(dd, &dentry->d_name);
393 err = logfs_write_buf(dir, page, WF_LOCK);
395 page_cache_release(page);
397 grow_dir(dir, index);
400 /* FIXME: Is there a better return value? In most cases neither
401 * the filesystem nor the directory are full. But we have had
402 * too many collisions for this particular hash and no fallback.
407 static int __logfs_create(struct inode *dir, struct dentry *dentry,
408 struct inode *inode, const char *dest, long destlen)
410 struct logfs_super *super = logfs_super(dir->i_sb);
411 struct logfs_inode *li = logfs_inode(inode);
412 struct logfs_transaction *ta;
415 ta = kzalloc(sizeof(*ta), GFP_KERNEL);
422 ta->state = CREATE_1;
423 ta->ino = inode->i_ino;
424 mutex_lock(&super->s_dirop_mutex);
425 logfs_add_transaction(inode, ta);
429 ret = logfs_inode_write(inode, dest, destlen, 0, WF_LOCK, NULL);
431 ret = write_inode(inode);
433 /* creat/mkdir/mknod */
434 ret = write_inode(inode);
437 abort_transaction(inode, ta);
438 li->li_flags |= LOGFS_IF_STILLBORN;
439 /* FIXME: truncate symlink */
445 ta->state = CREATE_2;
446 logfs_add_transaction(dir, ta);
447 ret = logfs_write_dir(dir, dentry, inode);
450 ret = write_inode(dir);
453 logfs_del_transaction(dir, ta);
454 ta->state = CREATE_2;
455 logfs_add_transaction(inode, ta);
456 logfs_remove_inode(inode);
460 d_instantiate(dentry, inode);
462 mutex_unlock(&super->s_dirop_mutex);
466 static int logfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
471 * FIXME: why do we have to fill in S_IFDIR, while the mode is
472 * correct for mknod, creat, etc.? Smells like the vfs *should*
473 * do it for us but for some reason fails to do so.
475 inode = logfs_new_inode(dir, S_IFDIR | mode);
477 return PTR_ERR(inode);
479 inode->i_op = &logfs_dir_iops;
480 inode->i_fop = &logfs_dir_fops;
482 return __logfs_create(dir, dentry, inode, NULL, 0);
485 static int logfs_create(struct inode *dir, struct dentry *dentry, umode_t mode,
490 inode = logfs_new_inode(dir, mode);
492 return PTR_ERR(inode);
494 inode->i_op = &logfs_reg_iops;
495 inode->i_fop = &logfs_reg_fops;
496 inode->i_mapping->a_ops = &logfs_reg_aops;
498 return __logfs_create(dir, dentry, inode, NULL, 0);
501 static int logfs_mknod(struct inode *dir, struct dentry *dentry, umode_t mode,
506 if (dentry->d_name.len > LOGFS_MAX_NAMELEN)
507 return -ENAMETOOLONG;
509 inode = logfs_new_inode(dir, mode);
511 return PTR_ERR(inode);
513 init_special_inode(inode, mode, rdev);
515 return __logfs_create(dir, dentry, inode, NULL, 0);
518 static int logfs_symlink(struct inode *dir, struct dentry *dentry,
522 size_t destlen = strlen(target) + 1;
524 if (destlen > dir->i_sb->s_blocksize)
525 return -ENAMETOOLONG;
527 inode = logfs_new_inode(dir, S_IFLNK | 0777);
529 return PTR_ERR(inode);
531 inode->i_op = &page_symlink_inode_operations;
532 inode_nohighmem(inode);
533 inode->i_mapping->a_ops = &logfs_reg_aops;
535 return __logfs_create(dir, dentry, inode, target, destlen);
538 static int logfs_link(struct dentry *old_dentry, struct inode *dir,
539 struct dentry *dentry)
541 struct inode *inode = d_inode(old_dentry);
543 inode->i_ctime = dir->i_ctime = dir->i_mtime = CURRENT_TIME;
546 mark_inode_dirty_sync(inode);
548 return __logfs_create(dir, dentry, inode, NULL, 0);
551 static int logfs_get_dd(struct inode *dir, struct dentry *dentry,
552 struct logfs_disk_dentry *dd, loff_t *pos)
557 page = logfs_get_dd_page(dir, dentry);
559 return PTR_ERR(page);
561 map = kmap_atomic(page);
562 memcpy(dd, map, sizeof(*dd));
564 page_cache_release(page);
568 static int logfs_delete_dd(struct inode *dir, loff_t pos)
571 * Getting called with pos somewhere beyond eof is either a goofup
572 * within this file or means someone maliciously edited the
573 * (crc-protected) journal.
575 BUG_ON(beyond_eof(dir, pos));
576 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
577 log_dir(" Delete dentry (%lx, %llx)\n", dir->i_ino, pos);
578 return logfs_delete(dir, pos, NULL);
582 * Cross-directory rename, target does not exist. Just a little nasty.
583 * Create a new dentry in the target dir, then remove the old dentry,
584 * all the while taking care to remember our operation in the journal.
586 static int logfs_rename_cross(struct inode *old_dir, struct dentry *old_dentry,
587 struct inode *new_dir, struct dentry *new_dentry)
589 struct logfs_super *super = logfs_super(old_dir->i_sb);
590 struct logfs_disk_dentry dd;
591 struct logfs_transaction *ta;
595 /* 1. locate source dd */
596 err = logfs_get_dd(old_dir, old_dentry, &dd, &pos);
600 ta = kzalloc(sizeof(*ta), GFP_KERNEL);
604 ta->state = CROSS_RENAME_1;
605 ta->dir = old_dir->i_ino;
608 /* 2. write target dd */
609 mutex_lock(&super->s_dirop_mutex);
610 logfs_add_transaction(new_dir, ta);
611 err = logfs_write_dir(new_dir, new_dentry, d_inode(old_dentry));
613 err = write_inode(new_dir);
616 super->s_rename_dir = 0;
617 super->s_rename_pos = 0;
618 abort_transaction(new_dir, ta);
622 /* 3. remove source dd */
623 ta->state = CROSS_RENAME_2;
624 logfs_add_transaction(old_dir, ta);
625 err = logfs_delete_dd(old_dir, pos);
627 err = write_inode(old_dir);
628 LOGFS_BUG_ON(err, old_dir->i_sb);
630 mutex_unlock(&super->s_dirop_mutex);
634 static int logfs_replace_inode(struct inode *dir, struct dentry *dentry,
635 struct logfs_disk_dentry *dd, struct inode *inode)
640 err = logfs_get_dd(dir, dentry, dd, &pos);
643 dd->ino = cpu_to_be64(inode->i_ino);
644 dd->type = logfs_type(inode);
646 err = write_dir(dir, dd, pos);
649 log_dir("Replace dentry (%lx, %llx) %s -> %llx\n", dir->i_ino, pos,
650 dd->name, be64_to_cpu(dd->ino));
651 return write_inode(dir);
654 /* Target dentry exists - the worst case. We need to attach the source
655 * inode to the target dentry, then remove the orphaned target inode and
658 static int logfs_rename_target(struct inode *old_dir, struct dentry *old_dentry,
659 struct inode *new_dir, struct dentry *new_dentry)
661 struct logfs_super *super = logfs_super(old_dir->i_sb);
662 struct inode *old_inode = d_inode(old_dentry);
663 struct inode *new_inode = d_inode(new_dentry);
664 int isdir = S_ISDIR(old_inode->i_mode);
665 struct logfs_disk_dentry dd;
666 struct logfs_transaction *ta;
670 BUG_ON(isdir != S_ISDIR(new_inode->i_mode));
672 if (!logfs_empty_dir(new_inode))
676 /* 1. locate source dd */
677 err = logfs_get_dd(old_dir, old_dentry, &dd, &pos);
681 ta = kzalloc(sizeof(*ta), GFP_KERNEL);
685 ta->state = TARGET_RENAME_1;
686 ta->dir = old_dir->i_ino;
688 ta->ino = new_inode->i_ino;
690 /* 2. attach source inode to target dd */
691 mutex_lock(&super->s_dirop_mutex);
692 logfs_add_transaction(new_dir, ta);
693 err = logfs_replace_inode(new_dir, new_dentry, &dd, old_inode);
695 super->s_rename_dir = 0;
696 super->s_rename_pos = 0;
697 super->s_victim_ino = 0;
698 abort_transaction(new_dir, ta);
702 /* 3. remove source dd */
703 ta->state = TARGET_RENAME_2;
704 logfs_add_transaction(old_dir, ta);
705 err = logfs_delete_dd(old_dir, pos);
707 err = write_inode(old_dir);
708 LOGFS_BUG_ON(err, old_dir->i_sb);
710 /* 4. remove target inode */
711 ta->state = TARGET_RENAME_3;
712 logfs_add_transaction(new_inode, ta);
713 err = logfs_remove_inode(new_inode);
716 mutex_unlock(&super->s_dirop_mutex);
720 static int logfs_rename(struct inode *old_dir, struct dentry *old_dentry,
721 struct inode *new_dir, struct dentry *new_dentry)
723 if (d_really_is_positive(new_dentry))
724 return logfs_rename_target(old_dir, old_dentry,
725 new_dir, new_dentry);
726 return logfs_rename_cross(old_dir, old_dentry, new_dir, new_dentry);
729 /* No locking done here, as this is called before .get_sb() returns. */
730 int logfs_replay_journal(struct super_block *sb)
732 struct logfs_super *super = logfs_super(sb);
737 if (super->s_victim_ino) {
738 /* delete victim inode */
739 ino = super->s_victim_ino;
740 printk(KERN_INFO"LogFS: delete unmapped inode #%llx\n", ino);
741 inode = logfs_iget(sb, ino);
745 LOGFS_BUG_ON(i_size_read(inode) > 0, sb);
746 super->s_victim_ino = 0;
747 err = logfs_remove_inode(inode);
750 super->s_victim_ino = ino;
754 if (super->s_rename_dir) {
755 /* delete old dd from rename */
756 ino = super->s_rename_dir;
757 pos = super->s_rename_pos;
758 printk(KERN_INFO"LogFS: delete unbacked dentry (%llx, %llx)\n",
760 inode = logfs_iget(sb, ino);
764 super->s_rename_dir = 0;
765 super->s_rename_pos = 0;
766 err = logfs_delete_dd(inode, pos);
769 super->s_rename_dir = ino;
770 super->s_rename_pos = pos;
780 const struct inode_operations logfs_dir_iops = {
781 .create = logfs_create,
783 .lookup = logfs_lookup,
784 .mkdir = logfs_mkdir,
785 .mknod = logfs_mknod,
786 .rename = logfs_rename,
787 .rmdir = logfs_rmdir,
788 .symlink = logfs_symlink,
789 .unlink = logfs_unlink,
791 const struct file_operations logfs_dir_fops = {
792 .fsync = logfs_fsync,
793 .unlocked_ioctl = logfs_ioctl,
794 .iterate = logfs_readdir,
795 .read = generic_read_dir,
796 .llseek = default_llseek,