4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see [sun.com URL with a
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * Top-level entry points into iam module
39 * Author: Wang Di <wangdi@clusterfs.com>
40 * Author: Nikita Danilov <nikita@clusterfs.com>
44 * iam: big theory statement.
46 * iam (Index Access Module) is a module providing abstraction of persistent
47 * transactional container on top of generalized ldiskfs htree.
51 * - key, pointer, and record size specifiable per container.
53 * - trees taller than 2 index levels.
55 * - read/write to existing ldiskfs htree directories as iam containers.
57 * iam container is a tree, consisting of leaf nodes containing keys and
58 * records stored in this container, and index nodes, containing keys and
59 * pointers to leaf or index nodes.
61 * iam does not work with keys directly, instead it calls user-supplied key
62 * comparison function (->dpo_keycmp()).
64 * Pointers are (currently) interpreted as logical offsets (measured in
65 * blocksful) within underlying flat file on top of which iam tree lives.
69 * iam mostly tries to reuse existing htree formats.
71 * Format of index node:
73 * +-----+-------+-------+-------+------+-------+------------+
74 * | | count | | | | | |
75 * | gap | / | entry | entry | .... | entry | free space |
76 * | | limit | | | | | |
77 * +-----+-------+-------+-------+------+-------+------------+
79 * gap this part of node is never accessed by iam code. It
80 * exists for binary compatibility with ldiskfs htree (that,
81 * in turn, stores fake struct ext2_dirent for ext2
82 * compatibility), and to keep some unspecified per-node
83 * data. Gap can be different for root and non-root index
84 * nodes. Gap size can be specified for each container
85 * (gap of 0 is allowed).
87 * count/limit current number of entries in this node, and the maximal
88 * number of entries that can fit into node. count/limit
89 * has the same size as entry, and is itself counted in
92 * entry index entry: consists of a key immediately followed by
93 * a pointer to a child node. Size of a key and size of a
94 * pointer depends on container. Entry has neither
95 * alignment nor padding.
97 * free space portion of node new entries are added to
99 * Entries in index node are sorted by their key value.
101 * Format of a leaf node is not specified. Generic iam code accesses leaf
102 * nodes through ->id_leaf methods in struct iam_descr.
104 * The IAM root block is a special node, which contains the IAM descriptor.
105 * It is on disk format:
107 * +---------+-------+--------+---------+-------+------+-------+------------+
108 * |IAM desc | count | idle | | | | | |
109 * |(fix/var)| / | blocks | padding | entry | .... | entry | free space |
110 * | | limit | | | | | | |
111 * +---------+-------+--------+---------+-------+------+-------+------------+
113 * The padding length is calculated with the parameters in the IAM descriptor.
115 * The field "idle_blocks" is used to record empty leaf nodes, which have not
116 * been released but all contained entries in them have been removed. Usually,
117 * the idle blocks in the IAM should be reused when need to allocate new leaf
118 * nodes for new entries, it depends on the IAM hash functions to map the new
119 * entries to these idle blocks. Unfortunately, it is not easy to design some
120 * hash functions for such clever mapping, especially considering the insert/
121 * lookup performance.
123 * So the IAM recycles the empty leaf nodes, and put them into a per-file based
124 * idle blocks pool. If need some new leaf node, it will try to take idle block
125 * from such pool with priority, in spite of how the IAM hash functions to map
128 * The idle blocks pool is organized as a series of tables, and each table
129 * can be described as following (on-disk format):
131 * +---------+---------+---------+---------+------+---------+-------+
132 * | magic | count | next | logic | | logic | free |
133 * |(16 bits)|(16 bits)| table | blk # | .... | blk # | space |
134 * | | |(32 bits)|(32 bits)| |(32 bits)| |
135 * +---------+---------+---------+---------+------+---------+-------+
137 * The logic blk# for the first table is stored in the root node "idle_blocks".
141 #include <linux/module.h>
142 #include <linux/fs.h>
143 #include <linux/pagemap.h>
144 #include <linux/time.h>
145 #include <linux/fcntl.h>
146 #include <linux/stat.h>
147 #include <linux/string.h>
148 #include <linux/quotaops.h>
149 #include <linux/buffer_head.h>
151 #include <ldiskfs/ldiskfs.h>
152 #include <ldiskfs/xattr.h>
155 #include "osd_internal.h"
157 #include <ldiskfs/acl.h>
160 * List of all registered formats.
162 * No locking. Callers synchronize.
164 static struct list_head iam_formats = LIST_HEAD_INIT(iam_formats);
166 void iam_format_register(struct iam_format *fmt)
168 list_add(&fmt->if_linkage, &iam_formats);
171 static struct buffer_head *
172 iam_load_idle_blocks(struct iam_container *c, iam_ptr_t blk)
174 struct inode *inode = c->ic_object;
175 struct iam_idle_head *head;
176 struct buffer_head *bh;
178 LASSERT(mutex_is_locked(&c->ic_idle_mutex));
183 bh = __ldiskfs_bread(NULL, inode, blk, 0);
184 if (IS_ERR_OR_NULL(bh)) {
185 CERROR("%.16s: cannot load idle blocks, blk = %u, err = %ld\n",
186 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk,
187 bh ? PTR_ERR(bh) : -EIO);
188 c->ic_idle_failed = 1;
194 head = (struct iam_idle_head *)(bh->b_data);
195 if (le16_to_cpu(head->iih_magic) != IAM_IDLE_HEADER_MAGIC) {
196 CERROR("%.16s: invalid idle block head, blk = %u, magic = %d\n",
197 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk,
198 le16_to_cpu(head->iih_magic));
200 c->ic_idle_failed = 1;
201 return ERR_PTR(-EBADF);
208 * Determine format of given container. This is done by scanning list of
209 * registered formats and calling ->if_guess() method of each in turn.
211 static int iam_format_guess(struct iam_container *c)
214 struct iam_format *fmt;
217 * XXX temporary initialization hook.
220 static int initialized = 0;
223 iam_lvar_format_init();
224 iam_lfix_format_init();
230 list_for_each_entry(fmt, &iam_formats, if_linkage) {
231 result = fmt->if_guess(c);
237 struct buffer_head *bh;
240 LASSERT(c->ic_root_bh != NULL);
242 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
243 c->ic_descr->id_root_gap +
244 sizeof(struct dx_countlimit));
245 mutex_lock(&c->ic_idle_mutex);
246 bh = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
247 if (bh != NULL && IS_ERR(bh))
248 result = PTR_ERR(bh);
251 mutex_unlock(&c->ic_idle_mutex);
258 * Initialize container @c.
260 int iam_container_init(struct iam_container *c,
261 struct iam_descr *descr, struct inode *inode)
263 memset(c, 0, sizeof *c);
265 c->ic_object = inode;
266 init_rwsem(&c->ic_sem);
267 dynlock_init(&c->ic_tree_lock);
268 mutex_init(&c->ic_idle_mutex);
273 * Determine container format.
275 int iam_container_setup(struct iam_container *c)
277 return iam_format_guess(c);
281 * Finalize container @c, release all resources.
283 void iam_container_fini(struct iam_container *c)
285 brelse(c->ic_idle_bh);
286 c->ic_idle_bh = NULL;
287 brelse(c->ic_root_bh);
288 c->ic_root_bh = NULL;
291 void iam_path_init(struct iam_path *path, struct iam_container *c,
292 struct iam_path_descr *pd)
294 memset(path, 0, sizeof *path);
295 path->ip_container = c;
296 path->ip_frame = path->ip_frames;
298 path->ip_leaf.il_path = path;
301 static void iam_leaf_fini(struct iam_leaf *leaf);
303 void iam_path_release(struct iam_path *path)
307 for (i = 0; i < ARRAY_SIZE(path->ip_frames); i++) {
308 if (path->ip_frames[i].bh != NULL) {
309 path->ip_frames[i].at_shifted = 0;
310 brelse(path->ip_frames[i].bh);
311 path->ip_frames[i].bh = NULL;
316 void iam_path_fini(struct iam_path *path)
318 iam_leaf_fini(&path->ip_leaf);
319 iam_path_release(path);
323 void iam_path_compat_init(struct iam_path_compat *path, struct inode *inode)
327 path->ipc_hinfo = &path->ipc_hinfo_area;
328 for (i = 0; i < ARRAY_SIZE(path->ipc_scratch); ++i)
329 path->ipc_descr.ipd_key_scratch[i] =
330 (struct iam_ikey *)&path->ipc_scratch[i];
332 iam_path_init(&path->ipc_path, &path->ipc_container, &path->ipc_descr);
335 void iam_path_compat_fini(struct iam_path_compat *path)
337 iam_path_fini(&path->ipc_path);
341 * Helper function initializing iam_path_descr and its key scratch area.
343 struct iam_path_descr *iam_ipd_alloc(void *area, int keysize)
345 struct iam_path_descr *ipd;
351 for (i = 0; i < ARRAY_SIZE(ipd->ipd_key_scratch); ++i, karea += keysize)
352 ipd->ipd_key_scratch[i] = karea;
356 void iam_ipd_free(struct iam_path_descr *ipd)
360 int iam_node_read(struct iam_container *c, iam_ptr_t ptr,
361 handle_t *h, struct buffer_head **bh)
363 /* NB: it can be called by iam_lfix_guess() which is still at
364 * very early stage, c->ic_root_bh and c->ic_descr->id_ops
365 * haven't been intialized yet.
366 * Also, we don't have this for IAM dir.
368 if (c->ic_root_bh != NULL &&
369 c->ic_descr->id_ops->id_root_ptr(c) == ptr) {
370 get_bh(c->ic_root_bh);
375 *bh = __ldiskfs_bread(h, c->ic_object, (int)ptr, 0);
386 * Return pointer to current leaf record. Pointer is valid while corresponding
387 * leaf node is locked and pinned.
389 static struct iam_rec *iam_leaf_rec(const struct iam_leaf *leaf)
391 return iam_leaf_ops(leaf)->rec(leaf);
395 * Return pointer to the current leaf key. This function returns pointer to
396 * the key stored in node.
398 * Caller should assume that returned pointer is only valid while leaf node is
401 static struct iam_key *iam_leaf_key(const struct iam_leaf *leaf)
403 return iam_leaf_ops(leaf)->key(leaf);
406 static int iam_leaf_key_size(const struct iam_leaf *leaf)
408 return iam_leaf_ops(leaf)->key_size(leaf);
411 static struct iam_ikey *iam_leaf_ikey(const struct iam_leaf *leaf,
412 struct iam_ikey *key)
414 return iam_leaf_ops(leaf)->ikey(leaf, key);
417 static int iam_leaf_keycmp(const struct iam_leaf *leaf,
418 const struct iam_key *key)
420 return iam_leaf_ops(leaf)->key_cmp(leaf, key);
423 static int iam_leaf_keyeq(const struct iam_leaf *leaf,
424 const struct iam_key *key)
426 return iam_leaf_ops(leaf)->key_eq(leaf, key);
429 #if LDISKFS_INVARIANT_ON
430 extern int dx_node_check(struct iam_path *p, struct iam_frame *f);
432 static int iam_path_check(struct iam_path *p)
437 struct iam_descr *param;
440 param = iam_path_descr(p);
441 for (i = 0; result && i < ARRAY_SIZE(p->ip_frames); ++i) {
442 f = &p->ip_frames[i];
444 result = dx_node_check(p, f);
446 result = !param->id_ops->id_node_check(p, f);
449 if (result && p->ip_leaf.il_bh != NULL)
452 ldiskfs_std_error(iam_path_obj(p)->i_sb, result);
458 static int iam_leaf_load(struct iam_path *path)
462 struct iam_container *c;
463 struct buffer_head *bh;
464 struct iam_leaf *leaf;
465 struct iam_descr *descr;
467 c = path->ip_container;
468 leaf = &path->ip_leaf;
469 descr = iam_path_descr(path);
470 block = path->ip_frame->leaf;
473 printk(KERN_EMERG "wrong leaf: %lu %d [%p %p %p]\n",
474 (long unsigned)path->ip_frame->leaf,
475 dx_get_count(dx_node_get_entries(path, path->ip_frame)),
476 path->ip_frames[0].bh, path->ip_frames[1].bh,
477 path->ip_frames[2].bh);
479 err = descr->id_ops->id_node_read(c, block, NULL, &bh);
482 leaf->il_curidx = block;
483 err = iam_leaf_ops(leaf)->init(leaf);
488 static void iam_unlock_htree(struct iam_container *ic,
489 struct dynlock_handle *lh)
492 dynlock_unlock(&ic->ic_tree_lock, lh);
496 static void iam_leaf_unlock(struct iam_leaf *leaf)
498 if (leaf->il_lock != NULL) {
499 iam_unlock_htree(iam_leaf_container(leaf),
502 leaf->il_lock = NULL;
506 static void iam_leaf_fini(struct iam_leaf *leaf)
508 if (leaf->il_path != NULL) {
509 iam_leaf_unlock(leaf);
510 iam_leaf_ops(leaf)->fini(leaf);
519 static void iam_leaf_start(struct iam_leaf *folio)
521 iam_leaf_ops(folio)->start(folio);
524 void iam_leaf_next(struct iam_leaf *folio)
526 iam_leaf_ops(folio)->next(folio);
529 static void iam_leaf_rec_add(struct iam_leaf *leaf, const struct iam_key *key,
530 const struct iam_rec *rec)
532 iam_leaf_ops(leaf)->rec_add(leaf, key, rec);
535 static void iam_rec_del(struct iam_leaf *leaf, int shift)
537 iam_leaf_ops(leaf)->rec_del(leaf, shift);
540 int iam_leaf_at_end(const struct iam_leaf *leaf)
542 return iam_leaf_ops(leaf)->at_end(leaf);
545 static void iam_leaf_split(struct iam_leaf *l, struct buffer_head **bh,
548 iam_leaf_ops(l)->split(l, bh, nr);
551 static inline int iam_leaf_empty(struct iam_leaf *l)
553 return iam_leaf_ops(l)->leaf_empty(l);
556 int iam_leaf_can_add(const struct iam_leaf *l,
557 const struct iam_key *k, const struct iam_rec *r)
559 return iam_leaf_ops(l)->can_add(l, k, r);
562 static int iam_txn_dirty(handle_t *handle,
563 struct iam_path *path, struct buffer_head *bh)
567 result = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
569 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
573 static int iam_txn_add(handle_t *handle,
574 struct iam_path *path, struct buffer_head *bh)
578 result = ldiskfs_journal_get_write_access(handle, bh);
580 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
584 /***********************************************************************/
585 /* iterator interface */
586 /***********************************************************************/
588 static enum iam_it_state it_state(const struct iam_iterator *it)
594 * Helper function returning scratch key.
596 static struct iam_container *iam_it_container(const struct iam_iterator *it)
598 return it->ii_path.ip_container;
601 static inline int it_keycmp(const struct iam_iterator *it,
602 const struct iam_key *k)
604 return iam_leaf_keycmp(&it->ii_path.ip_leaf, k);
607 static inline int it_keyeq(const struct iam_iterator *it,
608 const struct iam_key *k)
610 return iam_leaf_keyeq(&it->ii_path.ip_leaf, k);
613 static int it_ikeycmp(const struct iam_iterator *it, const struct iam_ikey *ik)
615 return iam_ikeycmp(it->ii_path.ip_container,
616 iam_leaf_ikey(&it->ii_path.ip_leaf,
617 iam_path_ikey(&it->ii_path, 0)), ik);
620 static inline int it_at_rec(const struct iam_iterator *it)
622 return !iam_leaf_at_end(&it->ii_path.ip_leaf);
625 static inline int it_before(const struct iam_iterator *it)
627 return it_state(it) == IAM_IT_SKEWED && it_at_rec(it);
631 * Helper wrapper around iam_it_get(): returns 0 (success) only when record
632 * with exactly the same key as asked is found.
634 static int iam_it_get_exact(struct iam_iterator *it, const struct iam_key *k)
638 result = iam_it_get(it, k);
641 else if (result == 0)
643 * Return -ENOENT if cursor is located above record with a key
644 * different from one specified, or in the empty leaf.
646 * XXX returning -ENOENT only works if iam_it_get() never
647 * returns -ENOENT as a legitimate error.
653 void iam_container_write_lock(struct iam_container *ic)
655 down_write(&ic->ic_sem);
658 void iam_container_write_unlock(struct iam_container *ic)
660 up_write(&ic->ic_sem);
663 void iam_container_read_lock(struct iam_container *ic)
665 down_read(&ic->ic_sem);
668 void iam_container_read_unlock(struct iam_container *ic)
670 up_read(&ic->ic_sem);
674 * Initialize iterator to IAM_IT_DETACHED state.
676 * postcondition: it_state(it) == IAM_IT_DETACHED
678 int iam_it_init(struct iam_iterator *it, struct iam_container *c, __u32 flags,
679 struct iam_path_descr *pd)
681 memset(it, 0, sizeof *it);
682 it->ii_flags = flags;
683 it->ii_state = IAM_IT_DETACHED;
684 iam_path_init(&it->ii_path, c, pd);
689 * Finalize iterator and release all resources.
691 * precondition: it_state(it) == IAM_IT_DETACHED
693 void iam_it_fini(struct iam_iterator *it)
695 assert_corr(it_state(it) == IAM_IT_DETACHED);
696 iam_path_fini(&it->ii_path);
700 * this locking primitives are used to protect parts
701 * of dir's htree. protection unit is block: leaf or index
703 static struct dynlock_handle *iam_lock_htree(struct iam_container *ic,
705 enum dynlock_type lt)
707 return dynlock_lock(&ic->ic_tree_lock, value, lt, GFP_NOFS);
710 static int iam_index_lock(struct iam_path *path, struct dynlock_handle **lh)
714 for (f = path->ip_frame; f >= path->ip_frames; --f, ++lh) {
716 *lh = iam_lock_htree(path->ip_container, f->curidx, DLT_READ);
724 * Fast check for frame consistency.
726 static int iam_check_fast(struct iam_path *path, struct iam_frame *frame)
728 struct iam_container *bag;
729 struct iam_entry *next;
730 struct iam_entry *last;
731 struct iam_entry *entries;
732 struct iam_entry *at;
734 bag = path->ip_container;
736 entries = frame->entries;
737 last = iam_entry_shift(path, entries, dx_get_count(entries) - 1);
739 if (unlikely(at > last))
742 if (unlikely(dx_get_block(path, at) != frame->leaf))
745 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, at),
746 path->ip_ikey_target) > 0))
749 next = iam_entry_shift(path, at, +1);
751 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, next),
752 path->ip_ikey_target) <= 0))
758 int dx_index_is_compat(struct iam_path *path)
760 return iam_path_descr(path) == NULL;
766 * search position of specified hash in index
770 static struct iam_entry *iam_find_position(struct iam_path *path,
771 struct iam_frame *frame)
778 count = dx_get_count(frame->entries);
779 assert_corr(count && count <= dx_get_limit(frame->entries));
780 p = iam_entry_shift(path, frame->entries,
781 dx_index_is_compat(path) ? 1 : 2);
782 q = iam_entry_shift(path, frame->entries, count - 1);
784 m = iam_entry_shift(path, p, iam_entry_diff(path, q, p) / 2);
785 if (iam_ikeycmp(path->ip_container, iam_ikey_at(path, m),
786 path->ip_ikey_target) > 0)
787 q = iam_entry_shift(path, m, -1);
789 p = iam_entry_shift(path, m, +1);
791 return iam_entry_shift(path, p, -1);
796 static iam_ptr_t iam_find_ptr(struct iam_path *path, struct iam_frame *frame)
798 return dx_get_block(path, iam_find_position(path, frame));
801 void iam_insert_key(struct iam_path *path, struct iam_frame *frame,
802 const struct iam_ikey *key, iam_ptr_t ptr)
804 struct iam_entry *entries = frame->entries;
805 struct iam_entry *new = iam_entry_shift(path, frame->at, +1);
806 int count = dx_get_count(entries);
809 * Unfortunately we cannot assert this, as this function is sometimes
810 * called by VFS under i_sem and without pdirops lock.
812 assert_corr(1 || iam_frame_is_locked(path, frame));
813 assert_corr(count < dx_get_limit(entries));
814 assert_corr(frame->at < iam_entry_shift(path, entries, count));
815 assert_inv(dx_node_check(path, frame));
817 memmove(iam_entry_shift(path, new, 1), new,
818 (char *)iam_entry_shift(path, entries, count) - (char *)new);
819 dx_set_ikey(path, new, key);
820 dx_set_block(path, new, ptr);
821 dx_set_count(entries, count + 1);
822 assert_inv(dx_node_check(path, frame));
825 void iam_insert_key_lock(struct iam_path *path, struct iam_frame *frame,
826 const struct iam_ikey *key, iam_ptr_t ptr)
828 iam_lock_bh(frame->bh);
829 iam_insert_key(path, frame, key, ptr);
830 iam_unlock_bh(frame->bh);
833 * returns 0 if path was unchanged, -EAGAIN otherwise.
835 static int iam_check_path(struct iam_path *path, struct iam_frame *frame)
839 iam_lock_bh(frame->bh);
840 equal = iam_check_fast(path, frame) == 0 ||
841 frame->leaf == iam_find_ptr(path, frame);
842 DX_DEVAL(iam_lock_stats.dls_bh_again += !equal);
843 iam_unlock_bh(frame->bh);
845 return equal ? 0 : -EAGAIN;
848 static int iam_lookup_try(struct iam_path *path)
854 struct iam_descr *param;
855 struct iam_frame *frame;
856 struct iam_container *c;
858 param = iam_path_descr(path);
859 c = path->ip_container;
861 ptr = param->id_ops->id_root_ptr(c);
862 for (frame = path->ip_frames, i = 0; i <= path->ip_indirect;
864 err = param->id_ops->id_node_read(c, (iam_ptr_t)ptr, NULL,
868 iam_lock_bh(frame->bh);
870 * node must be initialized under bh lock because concurrent
871 * creation procedure may change it and iam_lookup_try() will
872 * see obsolete tree height. -bzzz
877 if (LDISKFS_INVARIANT_ON) {
878 err = param->id_ops->id_node_check(path, frame);
883 err = param->id_ops->id_node_load(path, frame);
887 assert_inv(dx_node_check(path, frame));
889 * splitting may change root index block and move hash we're
890 * looking for into another index block so, we have to check
891 * this situation and repeat from begining if path got changed
895 err = iam_check_path(path, frame - 1);
900 frame->at = iam_find_position(path, frame);
902 frame->leaf = ptr = dx_get_block(path, frame->at);
904 iam_unlock_bh(frame->bh);
908 iam_unlock_bh(frame->bh);
909 path->ip_frame = --frame;
913 static int __iam_path_lookup(struct iam_path *path)
918 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++ i)
919 assert(path->ip_frames[i].bh == NULL);
922 err = iam_lookup_try(path);
926 } while (err == -EAGAIN);
932 * returns 0 if path was unchanged, -EAGAIN otherwise.
934 static int iam_check_full_path(struct iam_path *path, int search)
936 struct iam_frame *bottom;
937 struct iam_frame *scan;
943 for (bottom = path->ip_frames, i = 0;
944 i < DX_MAX_TREE_HEIGHT && bottom->bh != NULL; ++bottom, ++i) {
945 ; /* find last filled in frame */
949 * Lock frames, bottom to top.
951 for (scan = bottom - 1; scan >= path->ip_frames; --scan)
952 iam_lock_bh(scan->bh);
954 * Check them top to bottom.
957 for (scan = path->ip_frames; scan < bottom; ++scan) {
958 struct iam_entry *pos;
961 if (iam_check_fast(path, scan) == 0)
964 pos = iam_find_position(path, scan);
965 if (scan->leaf != dx_get_block(path, pos)) {
971 pos = iam_entry_shift(path, scan->entries,
972 dx_get_count(scan->entries) - 1);
973 if (scan->at > pos ||
974 scan->leaf != dx_get_block(path, scan->at)) {
982 * Unlock top to bottom.
984 for (scan = path->ip_frames; scan < bottom; ++scan)
985 iam_unlock_bh(scan->bh);
986 DX_DEVAL(iam_lock_stats.dls_bh_full_again += !!result);
994 * Performs path lookup and returns with found leaf (if any) locked by htree
997 static int iam_lookup_lock(struct iam_path *path,
998 struct dynlock_handle **dl, enum dynlock_type lt)
1002 while ((result = __iam_path_lookup(path)) == 0) {
1003 do_corr(schedule());
1004 *dl = iam_lock_htree(path->ip_container, path->ip_frame->leaf,
1007 iam_path_fini(path);
1011 do_corr(schedule());
1013 * while locking leaf we just found may get split so we need
1014 * to check this -bzzz
1016 if (iam_check_full_path(path, 1) == 0)
1018 iam_unlock_htree(path->ip_container, *dl);
1020 iam_path_fini(path);
1025 * Performs tree top-to-bottom traversal starting from root, and loads leaf
1028 static int iam_path_lookup(struct iam_path *path, int index)
1030 struct iam_container *c;
1031 struct iam_leaf *leaf;
1034 c = path->ip_container;
1035 leaf = &path->ip_leaf;
1036 result = iam_lookup_lock(path, &leaf->il_lock, DLT_WRITE);
1037 assert_inv(iam_path_check(path));
1038 do_corr(schedule());
1040 result = iam_leaf_load(path);
1042 do_corr(schedule());
1044 result = iam_leaf_ops(leaf)->
1045 ilookup(leaf, path->ip_ikey_target);
1047 result = iam_leaf_ops(leaf)->
1048 lookup(leaf, path->ip_key_target);
1049 do_corr(schedule());
1052 iam_leaf_unlock(leaf);
1058 * Common part of iam_it_{i,}get().
1060 static int __iam_it_get(struct iam_iterator *it, int index)
1063 assert_corr(it_state(it) == IAM_IT_DETACHED);
1065 result = iam_path_lookup(&it->ii_path, index);
1069 collision = result & IAM_LOOKUP_LAST;
1070 switch (result & ~IAM_LOOKUP_LAST) {
1071 case IAM_LOOKUP_EXACT:
1073 it->ii_state = IAM_IT_ATTACHED;
1077 it->ii_state = IAM_IT_ATTACHED;
1079 case IAM_LOOKUP_BEFORE:
1080 case IAM_LOOKUP_EMPTY:
1082 it->ii_state = IAM_IT_SKEWED;
1087 result |= collision;
1090 * See iam_it_get_exact() for explanation.
1092 assert_corr(result != -ENOENT);
1097 * Correct hash, but not the same key was found, iterate through hash
1098 * collision chain, looking for correct record.
1100 static int iam_it_collision(struct iam_iterator *it)
1104 assert(ergo(it_at_rec(it), !it_keyeq(it, it->ii_path.ip_key_target)));
1106 while ((result = iam_it_next(it)) == 0) {
1107 do_corr(schedule());
1108 if (it_ikeycmp(it, it->ii_path.ip_ikey_target) != 0)
1110 if (it_keyeq(it, it->ii_path.ip_key_target))
1117 * Attach iterator. After successful completion, @it points to record with
1118 * least key not larger than @k.
1120 * Return value: 0: positioned on existing record,
1121 * +ve: exact position found,
1124 * precondition: it_state(it) == IAM_IT_DETACHED
1125 * postcondition: ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1126 * it_keycmp(it, k) <= 0)
1128 int iam_it_get(struct iam_iterator *it, const struct iam_key *k)
1131 assert_corr(it_state(it) == IAM_IT_DETACHED);
1133 it->ii_path.ip_ikey_target = NULL;
1134 it->ii_path.ip_key_target = k;
1136 result = __iam_it_get(it, 0);
1138 if (result == IAM_LOOKUP_LAST) {
1139 result = iam_it_collision(it);
1143 result = __iam_it_get(it, 0);
1148 result &= ~IAM_LOOKUP_LAST;
1150 assert_corr(ergo(result > 0, it_keycmp(it, k) == 0));
1151 assert_corr(ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1152 it_keycmp(it, k) <= 0));
1157 * Attach iterator by index key.
1159 static int iam_it_iget(struct iam_iterator *it, const struct iam_ikey *k)
1161 assert_corr(it_state(it) == IAM_IT_DETACHED);
1163 it->ii_path.ip_ikey_target = k;
1164 return __iam_it_get(it, 1) & ~IAM_LOOKUP_LAST;
1168 * Attach iterator, and assure it points to the record (not skewed).
1170 * Return value: 0: positioned on existing record,
1171 * +ve: exact position found,
1174 * precondition: it_state(it) == IAM_IT_DETACHED &&
1175 * !(it->ii_flags&IAM_IT_WRITE)
1176 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED)
1178 int iam_it_get_at(struct iam_iterator *it, const struct iam_key *k)
1181 assert_corr(it_state(it) == IAM_IT_DETACHED &&
1182 !(it->ii_flags&IAM_IT_WRITE));
1183 result = iam_it_get(it, k);
1185 if (it_state(it) != IAM_IT_ATTACHED) {
1186 assert_corr(it_state(it) == IAM_IT_SKEWED);
1187 result = iam_it_next(it);
1190 assert_corr(ergo(result >= 0, it_state(it) == IAM_IT_ATTACHED));
1195 * Duplicates iterator.
1197 * postcondition: it_state(dst) == it_state(src) &&
1198 * iam_it_container(dst) == iam_it_container(src) &&
1199 * dst->ii_flags = src->ii_flags &&
1200 * ergo(it_state(src) == IAM_IT_ATTACHED,
1201 * iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1202 * iam_it_key_get(dst) == iam_it_key_get(src))
1204 void iam_it_dup(struct iam_iterator *dst, const struct iam_iterator *src)
1206 dst->ii_flags = src->ii_flags;
1207 dst->ii_state = src->ii_state;
1208 /* XXX not yet. iam_path_dup(&dst->ii_path, &src->ii_path); */
1210 * XXX: duplicate lock.
1212 assert_corr(it_state(dst) == it_state(src));
1213 assert_corr(iam_it_container(dst) == iam_it_container(src));
1214 assert_corr(dst->ii_flags = src->ii_flags);
1215 assert_corr(ergo(it_state(src) == IAM_IT_ATTACHED,
1216 iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1217 iam_it_key_get(dst) == iam_it_key_get(src)));
1222 * Detach iterator. Does nothing it detached state.
1224 * postcondition: it_state(it) == IAM_IT_DETACHED
1226 void iam_it_put(struct iam_iterator *it)
1228 if (it->ii_state != IAM_IT_DETACHED) {
1229 it->ii_state = IAM_IT_DETACHED;
1230 iam_leaf_fini(&it->ii_path.ip_leaf);
1234 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1235 struct iam_ikey *ikey);
1239 * This function increments the frame pointer to search the next leaf
1240 * block, and reads in the necessary intervening nodes if the search
1241 * should be necessary. Whether or not the search is necessary is
1242 * controlled by the hash parameter. If the hash value is even, then
1243 * the search is only continued if the next block starts with that
1244 * hash value. This is used if we are searching for a specific file.
1246 * If the hash value is HASH_NB_ALWAYS, then always go to the next block.
1248 * This function returns 1 if the caller should continue to search,
1249 * or 0 if it should not. If there is an error reading one of the
1250 * index blocks, it will a negative error code.
1252 * If start_hash is non-null, it will be filled in with the starting
1253 * hash of the next page.
1255 static int iam_htree_advance(struct inode *dir, __u32 hash,
1256 struct iam_path *path, __u32 *start_hash,
1259 struct iam_frame *p;
1260 struct buffer_head *bh;
1261 int err, num_frames = 0;
1266 * Find the next leaf page by incrementing the frame pointer.
1267 * If we run out of entries in the interior node, loop around and
1268 * increment pointer in the parent node. When we break out of
1269 * this loop, num_frames indicates the number of interior
1270 * nodes need to be read.
1273 do_corr(schedule());
1278 p->at = iam_entry_shift(path, p->at, +1);
1279 if (p->at < iam_entry_shift(path, p->entries,
1280 dx_get_count(p->entries))) {
1281 p->leaf = dx_get_block(path, p->at);
1282 iam_unlock_bh(p->bh);
1285 iam_unlock_bh(p->bh);
1286 if (p == path->ip_frames)
1297 * If the hash is 1, then continue only if the next page has a
1298 * continuation hash of any value. This is used for readdir
1299 * handling. Otherwise, check to see if the hash matches the
1300 * desired contiuation hash. If it doesn't, return since
1301 * there's no point to read in the successive index pages.
1303 dx_get_ikey(path, p->at, (struct iam_ikey *)&bhash);
1305 *start_hash = bhash;
1306 if ((hash & 1) == 0) {
1307 if ((bhash & ~1) != hash)
1312 * If the hash is HASH_NB_ALWAYS, we always go to the next
1313 * block so no check is necessary
1315 while (num_frames--) {
1318 do_corr(schedule());
1320 idx = p->leaf = dx_get_block(path, p->at);
1321 iam_unlock_bh(p->bh);
1322 err = iam_path_descr(path)->id_ops->
1323 id_node_read(path->ip_container, idx, NULL, &bh);
1325 return err; /* Failure */
1328 assert_corr(p->bh != bh);
1330 p->entries = dx_node_get_entries(path, p);
1331 p->at = iam_entry_shift(path, p->entries, !compat);
1332 assert_corr(p->curidx != idx);
1335 assert_corr(p->leaf != dx_get_block(path, p->at));
1336 p->leaf = dx_get_block(path, p->at);
1337 iam_unlock_bh(p->bh);
1338 assert_inv(dx_node_check(path, p));
1344 static inline int iam_index_advance(struct iam_path *path)
1346 return iam_htree_advance(iam_path_obj(path), 0, path, NULL, 0);
1349 static void iam_unlock_array(struct iam_container *ic,
1350 struct dynlock_handle **lh)
1354 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++i, ++lh) {
1356 iam_unlock_htree(ic, *lh);
1362 * Advance index part of @path to point to the next leaf. Returns 1 on
1363 * success, 0, when end of container was reached. Leaf node is locked.
1365 int iam_index_next(struct iam_container *c, struct iam_path *path)
1368 struct dynlock_handle *lh[DX_MAX_TREE_HEIGHT] = { NULL, };
1370 struct inode *object;
1373 * Locking for iam_index_next()... is to be described.
1376 object = c->ic_object;
1377 cursor = path->ip_frame->leaf;
1380 result = iam_index_lock(path, lh);
1381 do_corr(schedule());
1385 result = iam_check_full_path(path, 0);
1386 if (result == 0 && cursor == path->ip_frame->leaf) {
1387 result = iam_index_advance(path);
1389 assert_corr(result == 0 ||
1390 cursor != path->ip_frame->leaf);
1394 iam_unlock_array(c, lh);
1396 iam_path_release(path);
1397 do_corr(schedule());
1399 result = __iam_path_lookup(path);
1403 while (path->ip_frame->leaf != cursor) {
1404 do_corr(schedule());
1406 result = iam_index_lock(path, lh);
1407 do_corr(schedule());
1411 result = iam_check_full_path(path, 0);
1415 result = iam_index_advance(path);
1417 CERROR("cannot find cursor : %u\n",
1423 result = iam_check_full_path(path, 0);
1426 iam_unlock_array(c, lh);
1428 } while (result == -EAGAIN);
1432 iam_unlock_array(c, lh);
1437 * Move iterator one record right.
1439 * Return value: 0: success,
1440 * +1: end of container reached
1443 * precondition: (it_state(it) == IAM_IT_ATTACHED ||
1444 * it_state(it) == IAM_IT_SKEWED) && it->ii_flags&IAM_IT_MOVE
1445 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED) &&
1446 * ergo(result > 0, it_state(it) == IAM_IT_DETACHED)
1448 int iam_it_next(struct iam_iterator *it)
1451 struct iam_path *path;
1452 struct iam_leaf *leaf;
1453 do_corr(struct iam_ikey *ik_orig);
1455 /* assert_corr(it->ii_flags&IAM_IT_MOVE); */
1456 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1457 it_state(it) == IAM_IT_SKEWED);
1459 path = &it->ii_path;
1460 leaf = &path->ip_leaf;
1462 assert_corr(iam_leaf_is_locked(leaf));
1465 do_corr(ik_orig = it_at_rec(it) ?
1466 iam_it_ikey_get(it, iam_path_ikey(path, 2)) : NULL);
1467 if (it_before(it)) {
1468 assert_corr(!iam_leaf_at_end(leaf));
1469 it->ii_state = IAM_IT_ATTACHED;
1471 if (!iam_leaf_at_end(leaf))
1472 /* advance within leaf node */
1473 iam_leaf_next(leaf);
1475 * multiple iterations may be necessary due to empty leaves.
1477 while (result == 0 && iam_leaf_at_end(leaf)) {
1478 do_corr(schedule());
1479 /* advance index portion of the path */
1480 result = iam_index_next(iam_it_container(it), path);
1481 assert_corr(iam_leaf_is_locked(leaf));
1483 struct dynlock_handle *lh;
1484 lh = iam_lock_htree(iam_it_container(it),
1485 path->ip_frame->leaf,
1488 iam_leaf_fini(leaf);
1490 result = iam_leaf_load(path);
1492 iam_leaf_start(leaf);
1495 } else if (result == 0)
1496 /* end of container reached */
1502 it->ii_state = IAM_IT_ATTACHED;
1504 assert_corr(ergo(result == 0, it_state(it) == IAM_IT_ATTACHED));
1505 assert_corr(ergo(result > 0, it_state(it) == IAM_IT_DETACHED));
1506 assert_corr(ergo(result == 0 && ik_orig != NULL,
1507 it_ikeycmp(it, ik_orig) >= 0));
1512 * Return pointer to the record under iterator.
1514 * precondition: it_state(it) == IAM_IT_ATTACHED && it_at_rec(it)
1515 * postcondition: it_state(it) == IAM_IT_ATTACHED
1517 struct iam_rec *iam_it_rec_get(const struct iam_iterator *it)
1519 assert_corr(it_state(it) == IAM_IT_ATTACHED);
1520 assert_corr(it_at_rec(it));
1521 return iam_leaf_rec(&it->ii_path.ip_leaf);
1524 static void iam_it_reccpy(struct iam_iterator *it, const struct iam_rec *r)
1526 struct iam_leaf *folio;
1528 folio = &it->ii_path.ip_leaf;
1529 iam_leaf_ops(folio)->rec_set(folio, r);
1533 * Replace contents of record under iterator.
1535 * precondition: it_state(it) == IAM_IT_ATTACHED &&
1536 * it->ii_flags&IAM_IT_WRITE
1537 * postcondition: it_state(it) == IAM_IT_ATTACHED &&
1538 * ergo(result == 0, !memcmp(iam_it_rec_get(it), r, ...))
1540 int iam_it_rec_set(handle_t *h,
1541 struct iam_iterator *it, const struct iam_rec *r)
1544 struct iam_path *path;
1545 struct buffer_head *bh;
1547 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
1548 it->ii_flags&IAM_IT_WRITE);
1549 assert_corr(it_at_rec(it));
1551 path = &it->ii_path;
1552 bh = path->ip_leaf.il_bh;
1553 result = iam_txn_add(h, path, bh);
1555 iam_it_reccpy(it, r);
1556 result = iam_txn_dirty(h, path, bh);
1562 * Return pointer to the index key under iterator.
1564 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1565 * it_state(it) == IAM_IT_SKEWED
1567 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1568 struct iam_ikey *ikey)
1570 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1571 it_state(it) == IAM_IT_SKEWED);
1572 assert_corr(it_at_rec(it));
1573 return iam_leaf_ikey(&it->ii_path.ip_leaf, ikey);
1577 * Return pointer to the key under iterator.
1579 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1580 * it_state(it) == IAM_IT_SKEWED
1582 struct iam_key *iam_it_key_get(const struct iam_iterator *it)
1584 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1585 it_state(it) == IAM_IT_SKEWED);
1586 assert_corr(it_at_rec(it));
1587 return iam_leaf_key(&it->ii_path.ip_leaf);
1591 * Return size of key under iterator (in bytes)
1593 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1594 * it_state(it) == IAM_IT_SKEWED
1596 int iam_it_key_size(const struct iam_iterator *it)
1598 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1599 it_state(it) == IAM_IT_SKEWED);
1600 assert_corr(it_at_rec(it));
1601 return iam_leaf_key_size(&it->ii_path.ip_leaf);
1604 static struct buffer_head *
1605 iam_new_node(handle_t *h, struct iam_container *c, iam_ptr_t *b, int *e)
1607 struct inode *inode = c->ic_object;
1608 struct buffer_head *bh = NULL;
1609 struct iam_idle_head *head;
1610 struct buffer_head *idle;
1614 if (c->ic_idle_bh == NULL)
1617 mutex_lock(&c->ic_idle_mutex);
1618 if (unlikely(c->ic_idle_bh == NULL)) {
1619 mutex_unlock(&c->ic_idle_mutex);
1623 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
1624 count = le16_to_cpu(head->iih_count);
1626 *e = ldiskfs_journal_get_write_access(h, c->ic_idle_bh);
1631 *b = le32_to_cpu(head->iih_blks[count]);
1632 head->iih_count = cpu_to_le16(count);
1633 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_idle_bh);
1637 mutex_unlock(&c->ic_idle_mutex);
1638 bh = __ldiskfs_bread(NULL, inode, *b, 0);
1639 if (IS_ERR_OR_NULL(bh)) {
1649 /* The block itself which contains the iam_idle_head is
1650 * also an idle block, and can be used as the new node. */
1651 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
1652 c->ic_descr->id_root_gap +
1653 sizeof(struct dx_countlimit));
1654 *e = ldiskfs_journal_get_write_access(h, c->ic_root_bh);
1658 *b = le32_to_cpu(*idle_blocks);
1659 iam_lock_bh(c->ic_root_bh);
1660 *idle_blocks = head->iih_next;
1661 iam_unlock_bh(c->ic_root_bh);
1662 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_root_bh);
1664 iam_lock_bh(c->ic_root_bh);
1665 *idle_blocks = cpu_to_le32(*b);
1666 iam_unlock_bh(c->ic_root_bh);
1671 idle = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
1672 if (idle != NULL && IS_ERR(idle)) {
1674 c->ic_idle_bh = NULL;
1679 c->ic_idle_bh = idle;
1680 mutex_unlock(&c->ic_idle_mutex);
1683 /* get write access for the found buffer head */
1684 *e = ldiskfs_journal_get_write_access(h, bh);
1688 ldiskfs_std_error(inode->i_sb, *e);
1690 /* Clear the reused node as new node does. */
1691 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1692 set_buffer_uptodate(bh);
1697 bh = osd_ldiskfs_append(h, inode, b);
1706 mutex_unlock(&c->ic_idle_mutex);
1707 ldiskfs_std_error(inode->i_sb, *e);
1712 * Insertion of new record. Interaction with jbd during non-trivial case (when
1713 * split happens) is as following:
1715 * - new leaf node is involved into transaction by iam_new_node();
1717 * - old leaf node is involved into transaction by iam_add_rec();
1719 * - leaf where insertion point ends in, is marked dirty by iam_add_rec();
1721 * - leaf without insertion point is marked dirty (as @new_leaf) by
1724 * - split index nodes are involved into transaction and marked dirty by
1725 * split_index_node().
1727 * - "safe" index node, which is no split, but where new pointer is inserted
1728 * is involved into transaction and marked dirty by split_index_node().
1730 * - index node where pointer to new leaf is inserted is involved into
1731 * transaction by split_index_node() and marked dirty by iam_add_rec().
1733 * - inode is marked dirty by iam_add_rec().
1737 static int iam_new_leaf(handle_t *handle, struct iam_leaf *leaf)
1741 struct buffer_head *new_leaf;
1742 struct buffer_head *old_leaf;
1743 struct iam_container *c;
1745 struct iam_path *path;
1747 c = iam_leaf_container(leaf);
1748 path = leaf->il_path;
1751 new_leaf = iam_new_node(handle, c, &blknr, &err);
1752 do_corr(schedule());
1753 if (new_leaf != NULL) {
1754 struct dynlock_handle *lh;
1756 lh = iam_lock_htree(c, blknr, DLT_WRITE);
1757 do_corr(schedule());
1759 iam_leaf_ops(leaf)->init_new(c, new_leaf);
1760 do_corr(schedule());
1761 old_leaf = leaf->il_bh;
1762 iam_leaf_split(leaf, &new_leaf, blknr);
1763 if (old_leaf != leaf->il_bh) {
1765 * Switched to the new leaf.
1767 iam_leaf_unlock(leaf);
1769 path->ip_frame->leaf = blknr;
1771 iam_unlock_htree(path->ip_container, lh);
1772 do_corr(schedule());
1773 err = iam_txn_dirty(handle, path, new_leaf);
1775 err = ldiskfs_mark_inode_dirty(handle, obj);
1776 do_corr(schedule());
1781 assert_inv(iam_path_check(iam_leaf_path(leaf)));
1785 static inline void dx_set_limit(struct iam_entry *entries, unsigned value)
1787 ((struct dx_countlimit *) entries)->limit = cpu_to_le16(value);
1790 static int iam_shift_entries(struct iam_path *path,
1791 struct iam_frame *frame, unsigned count,
1792 struct iam_entry *entries, struct iam_entry *entries2,
1799 struct iam_frame *parent = frame - 1;
1800 struct iam_ikey *pivot = iam_path_ikey(path, 3);
1802 delta = dx_index_is_compat(path) ? 0 : +1;
1804 count1 = count/2 + delta;
1805 count2 = count - count1;
1806 dx_get_ikey(path, iam_entry_shift(path, entries, count1), pivot);
1808 dxtrace(printk("Split index %d/%d\n", count1, count2));
1810 memcpy((char *) iam_entry_shift(path, entries2, delta),
1811 (char *) iam_entry_shift(path, entries, count1),
1812 count2 * iam_entry_size(path));
1814 dx_set_count(entries2, count2 + delta);
1815 dx_set_limit(entries2, dx_node_limit(path));
1818 * NOTE: very subtle piece of code competing dx_probe() may find 2nd
1819 * level index in root index, then we insert new index here and set
1820 * new count in that 2nd level index. so, dx_probe() may see 2nd level
1821 * index w/o hash it looks for. the solution is to check root index
1822 * after we locked just founded 2nd level index -bzzz
1824 iam_insert_key_lock(path, parent, pivot, newblock);
1827 * now old and new 2nd level index blocks contain all pointers, so
1828 * dx_probe() may find it in the both. it's OK -bzzz
1830 iam_lock_bh(frame->bh);
1831 dx_set_count(entries, count1);
1832 iam_unlock_bh(frame->bh);
1835 * now old 2nd level index block points to first half of leafs. it's
1836 * importand that dx_probe() must check root index block for changes
1837 * under dx_lock_bh(frame->bh) -bzzz
1844 int split_index_node(handle_t *handle, struct iam_path *path,
1845 struct dynlock_handle **lh)
1848 struct iam_entry *entries; /* old block contents */
1849 struct iam_entry *entries2; /* new block contents */
1850 struct iam_frame *frame, *safe;
1851 struct buffer_head *bh_new[DX_MAX_TREE_HEIGHT] = {NULL};
1852 u32 newblock[DX_MAX_TREE_HEIGHT] = {0};
1853 struct dynlock_handle *lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1854 struct dynlock_handle *new_lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1855 struct inode *dir = iam_path_obj(path);
1856 struct iam_descr *descr;
1860 descr = iam_path_descr(path);
1862 * Algorithm below depends on this.
1864 assert_corr(dx_root_limit(path) < dx_node_limit(path));
1866 frame = path->ip_frame;
1867 entries = frame->entries;
1870 * Tall-tree handling: we might have to split multiple index blocks
1871 * all the way up to tree root. Tricky point here is error handling:
1872 * to avoid complicated undo/rollback we
1874 * - first allocate all necessary blocks
1876 * - insert pointers into them atomically.
1880 * Locking: leaf is already locked. htree-locks are acquired on all
1881 * index nodes that require split bottom-to-top, on the "safe" node,
1882 * and on all new nodes
1885 dxtrace(printk("using %u of %u node entries\n",
1886 dx_get_count(entries), dx_get_limit(entries)));
1888 /* What levels need split? */
1889 for (nr_splet = 0; frame >= path->ip_frames &&
1890 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1891 --frame, ++nr_splet) {
1892 do_corr(schedule());
1893 if (nr_splet == DX_MAX_TREE_HEIGHT) {
1895 CWARN(dir->i_sb, __FUNCTION__,
1896 "Directory index full!\n");
1906 * Lock all nodes, bottom to top.
1908 for (frame = path->ip_frame, i = nr_splet; i >= 0; --i, --frame) {
1909 do_corr(schedule());
1910 lock[i] = iam_lock_htree(path->ip_container, frame->curidx,
1912 if (lock[i] == NULL) {
1919 * Check for concurrent index modification.
1921 err = iam_check_full_path(path, 1);
1925 * And check that the same number of nodes is to be split.
1927 for (i = 0, frame = path->ip_frame; frame >= path->ip_frames &&
1928 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1932 if (i != nr_splet) {
1937 /* Go back down, allocating blocks, locking them, and adding into
1939 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
1940 bh_new[i] = iam_new_node(handle, path->ip_container,
1941 &newblock[i], &err);
1942 do_corr(schedule());
1944 descr->id_ops->id_node_init(path->ip_container,
1947 new_lock[i] = iam_lock_htree(path->ip_container, newblock[i],
1949 if (new_lock[i] == NULL) {
1953 do_corr(schedule());
1954 BUFFER_TRACE(frame->bh, "get_write_access");
1955 err = ldiskfs_journal_get_write_access(handle, frame->bh);
1959 /* Add "safe" node to transaction too */
1960 if (safe + 1 != path->ip_frames) {
1961 do_corr(schedule());
1962 err = ldiskfs_journal_get_write_access(handle, safe->bh);
1967 /* Go through nodes once more, inserting pointers */
1968 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
1971 struct buffer_head *bh2;
1972 struct buffer_head *bh;
1974 entries = frame->entries;
1975 count = dx_get_count(entries);
1976 idx = iam_entry_diff(path, frame->at, entries);
1979 entries2 = dx_get_entries(path, bh2->b_data, 0);
1982 if (frame == path->ip_frames) {
1983 /* splitting root node. Tricky point:
1985 * In the "normal" B-tree we'd split root *and* add
1986 * new root to the tree with pointers to the old root
1987 * and its sibling (thus introducing two new nodes).
1989 * In htree it's enough to add one node, because
1990 * capacity of the root node is smaller than that of
1993 struct iam_frame *frames;
1994 struct iam_entry *next;
1996 assert_corr(i == 0);
1998 do_corr(schedule());
2000 frames = path->ip_frames;
2001 memcpy((char *) entries2, (char *) entries,
2002 count * iam_entry_size(path));
2003 dx_set_limit(entries2, dx_node_limit(path));
2006 iam_lock_bh(frame->bh);
2007 next = descr->id_ops->id_root_inc(path->ip_container,
2009 dx_set_block(path, next, newblock[0]);
2010 iam_unlock_bh(frame->bh);
2012 do_corr(schedule());
2013 /* Shift frames in the path */
2014 memmove(frames + 2, frames + 1,
2015 (sizeof path->ip_frames) - 2 * sizeof frames[0]);
2016 /* Add new access path frame */
2017 frames[1].at = iam_entry_shift(path, entries2, idx);
2018 frames[1].entries = entries = entries2;
2020 assert_inv(dx_node_check(path, frame));
2023 assert_inv(dx_node_check(path, frame));
2024 bh_new[0] = NULL; /* buffer head is "consumed" */
2025 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2028 do_corr(schedule());
2030 /* splitting non-root index node. */
2031 struct iam_frame *parent = frame - 1;
2033 do_corr(schedule());
2034 count = iam_shift_entries(path, frame, count,
2035 entries, entries2, newblock[i]);
2036 /* Which index block gets the new entry? */
2038 int d = dx_index_is_compat(path) ? 0 : +1;
2040 frame->at = iam_entry_shift(path, entries2,
2042 frame->entries = entries = entries2;
2043 frame->curidx = newblock[i];
2044 swap(frame->bh, bh2);
2045 assert_corr(lock[i + 1] != NULL);
2046 assert_corr(new_lock[i] != NULL);
2047 swap(lock[i + 1], new_lock[i]);
2049 parent->at = iam_entry_shift(path,
2052 assert_inv(dx_node_check(path, frame));
2053 assert_inv(dx_node_check(path, parent));
2054 dxtrace(dx_show_index ("node", frame->entries));
2055 dxtrace(dx_show_index ("node",
2056 ((struct dx_node *) bh2->b_data)->entries));
2057 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2060 do_corr(schedule());
2061 err = ldiskfs_handle_dirty_metadata(handle, NULL,
2066 do_corr(schedule());
2067 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2072 * This function was called to make insertion of new leaf
2073 * possible. Check that it fulfilled its obligations.
2075 assert_corr(dx_get_count(path->ip_frame->entries) <
2076 dx_get_limit(path->ip_frame->entries));
2077 assert_corr(lock[nr_splet] != NULL);
2078 *lh = lock[nr_splet];
2079 lock[nr_splet] = NULL;
2082 * Log ->i_size modification.
2084 err = ldiskfs_mark_inode_dirty(handle, dir);
2090 ldiskfs_std_error(dir->i_sb, err);
2093 iam_unlock_array(path->ip_container, lock);
2094 iam_unlock_array(path->ip_container, new_lock);
2096 assert_corr(err || iam_frame_is_locked(path, path->ip_frame));
2098 do_corr(schedule());
2099 for (i = 0; i < ARRAY_SIZE(bh_new); ++i) {
2100 if (bh_new[i] != NULL)
2106 static int iam_add_rec(handle_t *handle, struct iam_iterator *it,
2107 struct iam_path *path,
2108 const struct iam_key *k, const struct iam_rec *r)
2111 struct iam_leaf *leaf;
2113 leaf = &path->ip_leaf;
2114 assert_inv(iam_path_check(path));
2115 err = iam_txn_add(handle, path, leaf->il_bh);
2117 do_corr(schedule());
2118 if (!iam_leaf_can_add(leaf, k, r)) {
2119 struct dynlock_handle *lh = NULL;
2122 assert_corr(lh == NULL);
2123 do_corr(schedule());
2124 err = split_index_node(handle, path, &lh);
2125 if (err == -EAGAIN) {
2126 assert_corr(lh == NULL);
2128 iam_path_fini(path);
2129 it->ii_state = IAM_IT_DETACHED;
2131 do_corr(schedule());
2132 err = iam_it_get_exact(it, k);
2134 err = +1; /* repeat split */
2139 assert_inv(iam_path_check(path));
2141 assert_corr(lh != NULL);
2142 do_corr(schedule());
2143 err = iam_new_leaf(handle, leaf);
2145 err = iam_txn_dirty(handle, path,
2146 path->ip_frame->bh);
2148 iam_unlock_htree(path->ip_container, lh);
2149 do_corr(schedule());
2152 iam_leaf_rec_add(leaf, k, r);
2153 err = iam_txn_dirty(handle, path, leaf->il_bh);
2156 assert_inv(iam_path_check(path));
2161 * Insert new record with key @k and contents from @r, shifting records to the
2162 * right. On success, iterator is positioned on the newly inserted record.
2164 * precondition: it->ii_flags&IAM_IT_WRITE &&
2165 * (it_state(it) == IAM_IT_ATTACHED ||
2166 * it_state(it) == IAM_IT_SKEWED) &&
2167 * ergo(it_state(it) == IAM_IT_ATTACHED,
2168 * it_keycmp(it, k) <= 0) &&
2169 * ergo(it_before(it), it_keycmp(it, k) > 0));
2170 * postcondition: ergo(result == 0,
2171 * it_state(it) == IAM_IT_ATTACHED &&
2172 * it_keycmp(it, k) == 0 &&
2173 * !memcmp(iam_it_rec_get(it), r, ...))
2175 int iam_it_rec_insert(handle_t *h, struct iam_iterator *it,
2176 const struct iam_key *k, const struct iam_rec *r)
2179 struct iam_path *path;
2181 path = &it->ii_path;
2183 assert_corr(it->ii_flags&IAM_IT_WRITE);
2184 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2185 it_state(it) == IAM_IT_SKEWED);
2186 assert_corr(ergo(it_state(it) == IAM_IT_ATTACHED,
2187 it_keycmp(it, k) <= 0));
2188 assert_corr(ergo(it_before(it), it_keycmp(it, k) > 0));
2189 result = iam_add_rec(h, it, path, k, r);
2191 it->ii_state = IAM_IT_ATTACHED;
2192 assert_corr(ergo(result == 0,
2193 it_state(it) == IAM_IT_ATTACHED &&
2194 it_keycmp(it, k) == 0));
2198 static inline int iam_idle_blocks_limit(struct inode *inode)
2200 return (inode->i_sb->s_blocksize - sizeof(struct iam_idle_head)) >> 2;
2204 * If the leaf cannnot be recycled, we will lose one block for reusing.
2205 * It is not a serious issue because it almost the same of non-recycle.
2207 static iam_ptr_t iam_index_shrink(handle_t *h, struct iam_path *p,
2208 struct iam_leaf *l, struct buffer_head **bh)
2210 struct iam_container *c = p->ip_container;
2211 struct inode *inode = c->ic_object;
2212 struct iam_frame *frame = p->ip_frame;
2213 struct iam_entry *entries;
2214 struct iam_entry *pos;
2215 struct dynlock_handle *lh;
2219 if (c->ic_idle_failed)
2222 if (unlikely(frame == NULL))
2225 if (!iam_leaf_empty(l))
2228 lh = iam_lock_htree(c, frame->curidx, DLT_WRITE);
2230 CWARN("%.16s: No memory to recycle idle blocks\n",
2231 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name);
2235 rc = iam_txn_add(h, p, frame->bh);
2237 iam_unlock_htree(c, lh);
2241 iam_lock_bh(frame->bh);
2242 entries = frame->entries;
2243 count = dx_get_count(entries);
2244 /* NOT shrink the last entry in the index node, which can be reused
2245 * directly by next new node. */
2247 iam_unlock_bh(frame->bh);
2248 iam_unlock_htree(c, lh);
2252 pos = iam_find_position(p, frame);
2253 /* There may be some new leaf nodes have been added or empty leaf nodes
2254 * have been shrinked during my delete operation.
2256 * If the empty leaf is not under current index node because the index
2257 * node has been split, then just skip the empty leaf, which is rare. */
2258 if (unlikely(frame->leaf != dx_get_block(p, pos))) {
2259 iam_unlock_bh(frame->bh);
2260 iam_unlock_htree(c, lh);
2265 if (frame->at < iam_entry_shift(p, entries, count - 1)) {
2266 struct iam_entry *n = iam_entry_shift(p, frame->at, 1);
2268 memmove(frame->at, n,
2269 (char *)iam_entry_shift(p, entries, count) - (char *)n);
2270 frame->at_shifted = 1;
2272 dx_set_count(entries, count - 1);
2273 iam_unlock_bh(frame->bh);
2274 rc = iam_txn_dirty(h, p, frame->bh);
2275 iam_unlock_htree(c, lh);
2285 iam_install_idle_blocks(handle_t *h, struct iam_path *p, struct buffer_head *bh,
2286 __u32 *idle_blocks, iam_ptr_t blk)
2288 struct iam_container *c = p->ip_container;
2289 struct buffer_head *old = c->ic_idle_bh;
2290 struct iam_idle_head *head;
2293 head = (struct iam_idle_head *)(bh->b_data);
2294 head->iih_magic = cpu_to_le16(IAM_IDLE_HEADER_MAGIC);
2295 head->iih_count = 0;
2296 head->iih_next = *idle_blocks;
2297 /* The bh already get_write_accessed. */
2298 rc = iam_txn_dirty(h, p, bh);
2302 rc = iam_txn_add(h, p, c->ic_root_bh);
2306 iam_lock_bh(c->ic_root_bh);
2307 *idle_blocks = cpu_to_le32(blk);
2308 iam_unlock_bh(c->ic_root_bh);
2309 rc = iam_txn_dirty(h, p, c->ic_root_bh);
2311 /* NOT release old before new assigned. */
2316 iam_lock_bh(c->ic_root_bh);
2317 *idle_blocks = head->iih_next;
2318 iam_unlock_bh(c->ic_root_bh);
2324 * If the leaf cannnot be recycled, we will lose one block for reusing.
2325 * It is not a serious issue because it almost the same of non-recycle.
2327 static void iam_recycle_leaf(handle_t *h, struct iam_path *p,
2328 struct buffer_head *bh, iam_ptr_t blk)
2330 struct iam_container *c = p->ip_container;
2331 struct inode *inode = c->ic_object;
2332 struct iam_idle_head *head;
2337 mutex_lock(&c->ic_idle_mutex);
2338 if (unlikely(c->ic_idle_failed)) {
2343 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
2344 c->ic_descr->id_root_gap +
2345 sizeof(struct dx_countlimit));
2346 /* It is the first idle block. */
2347 if (c->ic_idle_bh == NULL) {
2348 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2352 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
2353 count = le16_to_cpu(head->iih_count);
2354 /* Current ic_idle_bh is full, to be replaced by the leaf. */
2355 if (count == iam_idle_blocks_limit(inode)) {
2356 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2360 /* Just add to ic_idle_bh. */
2361 rc = iam_txn_add(h, p, c->ic_idle_bh);
2365 head->iih_blks[count] = cpu_to_le32(blk);
2366 head->iih_count = cpu_to_le16(count + 1);
2367 rc = iam_txn_dirty(h, p, c->ic_idle_bh);
2370 mutex_unlock(&c->ic_idle_mutex);
2372 CWARN("%.16s: idle blocks failed, will lose the blk %u\n",
2373 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk);
2377 * Delete record under iterator.
2379 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2380 * it->ii_flags&IAM_IT_WRITE &&
2382 * postcondition: it_state(it) == IAM_IT_ATTACHED ||
2383 * it_state(it) == IAM_IT_DETACHED
2385 int iam_it_rec_delete(handle_t *h, struct iam_iterator *it)
2388 struct iam_leaf *leaf;
2389 struct iam_path *path;
2391 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
2392 it->ii_flags&IAM_IT_WRITE);
2393 assert_corr(it_at_rec(it));
2395 path = &it->ii_path;
2396 leaf = &path->ip_leaf;
2398 assert_inv(iam_path_check(path));
2400 result = iam_txn_add(h, path, leaf->il_bh);
2402 * no compaction for now.
2405 iam_rec_del(leaf, it->ii_flags&IAM_IT_MOVE);
2406 result = iam_txn_dirty(h, path, leaf->il_bh);
2407 if (result == 0 && iam_leaf_at_end(leaf)) {
2408 struct buffer_head *bh = NULL;
2411 blk = iam_index_shrink(h, path, leaf, &bh);
2412 if (it->ii_flags & IAM_IT_MOVE) {
2413 result = iam_it_next(it);
2419 iam_recycle_leaf(h, path, bh, blk);
2424 assert_inv(iam_path_check(path));
2425 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2426 it_state(it) == IAM_IT_DETACHED);
2431 * Convert iterator to cookie.
2433 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2434 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2435 * postcondition: it_state(it) == IAM_IT_ATTACHED
2437 iam_pos_t iam_it_store(const struct iam_iterator *it)
2441 assert_corr(it_state(it) == IAM_IT_ATTACHED);
2442 assert_corr(it_at_rec(it));
2443 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <=
2447 return *(iam_pos_t *)iam_it_ikey_get(it, (void *)&result);
2451 * Restore iterator from cookie.
2453 * precondition: it_state(it) == IAM_IT_DETACHED && it->ii_flags&IAM_IT_MOVE &&
2454 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2455 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED &&
2456 * iam_it_store(it) == pos)
2458 int iam_it_load(struct iam_iterator *it, iam_pos_t pos)
2460 assert_corr(it_state(it) == IAM_IT_DETACHED &&
2461 it->ii_flags&IAM_IT_MOVE);
2462 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <= sizeof pos);
2463 return iam_it_iget(it, (struct iam_ikey *)&pos);
2466 /***********************************************************************/
2468 /***********************************************************************/
2470 static inline int ptr_inside(void *base, size_t size, void *ptr)
2472 return (base <= ptr) && (ptr < base + size);
2475 static int iam_frame_invariant(struct iam_frame *f)
2479 f->bh->b_data != NULL &&
2480 ptr_inside(f->bh->b_data, f->bh->b_size, f->entries) &&
2481 ptr_inside(f->bh->b_data, f->bh->b_size, f->at) &&
2482 f->entries <= f->at);
2485 static int iam_leaf_invariant(struct iam_leaf *l)
2489 l->il_bh->b_data != NULL &&
2490 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_entries) &&
2491 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_at) &&
2492 l->il_entries <= l->il_at;
2495 static int iam_path_invariant(struct iam_path *p)
2499 if (p->ip_container == NULL ||
2500 p->ip_indirect < 0 || p->ip_indirect > DX_MAX_TREE_HEIGHT - 1 ||
2501 p->ip_frame != p->ip_frames + p->ip_indirect ||
2502 !iam_leaf_invariant(&p->ip_leaf))
2504 for (i = 0; i < ARRAY_SIZE(p->ip_frames); ++i) {
2505 if (i <= p->ip_indirect) {
2506 if (!iam_frame_invariant(&p->ip_frames[i]))
2513 int iam_it_invariant(struct iam_iterator *it)
2516 (it->ii_state == IAM_IT_DETACHED ||
2517 it->ii_state == IAM_IT_ATTACHED ||
2518 it->ii_state == IAM_IT_SKEWED) &&
2519 !(it->ii_flags & ~(IAM_IT_MOVE | IAM_IT_WRITE)) &&
2520 ergo(it->ii_state == IAM_IT_ATTACHED ||
2521 it->ii_state == IAM_IT_SKEWED,
2522 iam_path_invariant(&it->ii_path) &&
2523 equi(it_at_rec(it), it->ii_state == IAM_IT_SKEWED));
2527 * Search container @c for record with key @k. If record is found, its data
2528 * are moved into @r.
2530 * Return values: 0: found, -ENOENT: not-found, -ve: error
2532 int iam_lookup(struct iam_container *c, const struct iam_key *k,
2533 struct iam_rec *r, struct iam_path_descr *pd)
2535 struct iam_iterator it;
2538 iam_it_init(&it, c, 0, pd);
2540 result = iam_it_get_exact(&it, k);
2543 * record with required key found, copy it into user buffer
2545 iam_reccpy(&it.ii_path.ip_leaf, r);
2552 * Insert new record @r with key @k into container @c (within context of
2555 * Return values: 0: success, -ve: error, including -EEXIST when record with
2556 * given key is already present.
2558 * postcondition: ergo(result == 0 || result == -EEXIST,
2559 * iam_lookup(c, k, r2) > 0;
2561 int iam_insert(handle_t *h, struct iam_container *c, const struct iam_key *k,
2562 const struct iam_rec *r, struct iam_path_descr *pd)
2564 struct iam_iterator it;
2567 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2569 result = iam_it_get_exact(&it, k);
2570 if (result == -ENOENT)
2571 result = iam_it_rec_insert(h, &it, k, r);
2572 else if (result == 0)
2580 * Update record with the key @k in container @c (within context of
2581 * transaction @h), new record is given by @r.
2583 * Return values: +1: skip because of the same rec value, 0: success,
2584 * -ve: error, including -ENOENT if no record with the given key found.
2586 int iam_update(handle_t *h, struct iam_container *c, const struct iam_key *k,
2587 const struct iam_rec *r, struct iam_path_descr *pd)
2589 struct iam_iterator it;
2590 struct iam_leaf *folio;
2593 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2595 result = iam_it_get_exact(&it, k);
2597 folio = &it.ii_path.ip_leaf;
2598 result = iam_leaf_ops(folio)->rec_eq(folio, r);
2600 iam_it_rec_set(h, &it, r);
2610 * Delete existing record with key @k.
2612 * Return values: 0: success, -ENOENT: not-found, -ve: other error.
2614 * postcondition: ergo(result == 0 || result == -ENOENT,
2615 * !iam_lookup(c, k, *));
2617 int iam_delete(handle_t *h, struct iam_container *c, const struct iam_key *k,
2618 struct iam_path_descr *pd)
2620 struct iam_iterator it;
2623 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2625 result = iam_it_get_exact(&it, k);
2627 iam_it_rec_delete(h, &it);
2633 int iam_root_limit(int rootgap, int blocksize, int size)
2638 limit = (blocksize - rootgap) / size;
2639 nlimit = blocksize / size;
2640 if (limit == nlimit)