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, 2014, 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>
150 #include "osd_internal.h"
156 * List of all registered formats.
158 * No locking. Callers synchronize.
160 static struct list_head iam_formats = LIST_HEAD_INIT(iam_formats);
162 void iam_format_register(struct iam_format *fmt)
164 list_add(&fmt->if_linkage, &iam_formats);
167 static struct buffer_head *
168 iam_load_idle_blocks(struct iam_container *c, iam_ptr_t blk)
170 struct inode *inode = c->ic_object;
171 struct iam_idle_head *head;
172 struct buffer_head *bh;
175 LASSERT(mutex_is_locked(&c->ic_idle_mutex));
180 bh = ldiskfs_bread(NULL, inode, blk, 0, &err);
182 CERROR("%.16s: cannot load idle blocks, blk = %u, err = %d\n",
183 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk, err);
184 c->ic_idle_failed = 1;
188 head = (struct iam_idle_head *)(bh->b_data);
189 if (le16_to_cpu(head->iih_magic) != IAM_IDLE_HEADER_MAGIC) {
190 CERROR("%.16s: invalid idle block head, blk = %u, magic = %d\n",
191 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk,
192 le16_to_cpu(head->iih_magic));
194 c->ic_idle_failed = 1;
195 return ERR_PTR(-EBADF);
202 * Determine format of given container. This is done by scanning list of
203 * registered formats and calling ->if_guess() method of each in turn.
205 static int iam_format_guess(struct iam_container *c)
208 struct iam_format *fmt;
211 * XXX temporary initialization hook.
214 static int initialized = 0;
217 iam_lvar_format_init();
218 iam_lfix_format_init();
224 list_for_each_entry(fmt, &iam_formats, if_linkage) {
225 result = fmt->if_guess(c);
231 struct buffer_head *bh;
234 LASSERT(c->ic_root_bh != NULL);
236 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
237 c->ic_descr->id_root_gap +
238 sizeof(struct dx_countlimit));
239 mutex_lock(&c->ic_idle_mutex);
240 bh = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
241 if (bh != NULL && IS_ERR(bh))
242 result = PTR_ERR(bh);
245 mutex_unlock(&c->ic_idle_mutex);
252 * Initialize container @c.
254 int iam_container_init(struct iam_container *c,
255 struct iam_descr *descr, struct inode *inode)
257 memset(c, 0, sizeof *c);
259 c->ic_object = inode;
260 init_rwsem(&c->ic_sem);
261 dynlock_init(&c->ic_tree_lock);
262 mutex_init(&c->ic_idle_mutex);
267 * Determine container format.
269 int iam_container_setup(struct iam_container *c)
271 return iam_format_guess(c);
275 * Finalize container @c, release all resources.
277 void iam_container_fini(struct iam_container *c)
279 brelse(c->ic_idle_bh);
280 c->ic_idle_bh = NULL;
281 brelse(c->ic_root_bh);
282 c->ic_root_bh = NULL;
285 void iam_path_init(struct iam_path *path, struct iam_container *c,
286 struct iam_path_descr *pd)
288 memset(path, 0, sizeof *path);
289 path->ip_container = c;
290 path->ip_frame = path->ip_frames;
292 path->ip_leaf.il_path = path;
295 static void iam_leaf_fini(struct iam_leaf *leaf);
297 void iam_path_release(struct iam_path *path)
301 for (i = 0; i < ARRAY_SIZE(path->ip_frames); i++) {
302 if (path->ip_frames[i].bh != NULL) {
303 path->ip_frames[i].at_shifted = 0;
304 brelse(path->ip_frames[i].bh);
305 path->ip_frames[i].bh = NULL;
310 void iam_path_fini(struct iam_path *path)
312 iam_leaf_fini(&path->ip_leaf);
313 iam_path_release(path);
317 void iam_path_compat_init(struct iam_path_compat *path, struct inode *inode)
321 path->ipc_hinfo = &path->ipc_hinfo_area;
322 for (i = 0; i < ARRAY_SIZE(path->ipc_scratch); ++i)
323 path->ipc_descr.ipd_key_scratch[i] =
324 (struct iam_ikey *)&path->ipc_scratch[i];
326 iam_path_init(&path->ipc_path, &path->ipc_container, &path->ipc_descr);
329 void iam_path_compat_fini(struct iam_path_compat *path)
331 iam_path_fini(&path->ipc_path);
335 * Helper function initializing iam_path_descr and its key scratch area.
337 struct iam_path_descr *iam_ipd_alloc(void *area, int keysize)
339 struct iam_path_descr *ipd;
345 for (i = 0; i < ARRAY_SIZE(ipd->ipd_key_scratch); ++i, karea += keysize)
346 ipd->ipd_key_scratch[i] = karea;
350 void iam_ipd_free(struct iam_path_descr *ipd)
354 int iam_node_read(struct iam_container *c, iam_ptr_t ptr,
355 handle_t *h, struct buffer_head **bh)
359 /* NB: it can be called by iam_lfix_guess() which is still at
360 * very early stage, c->ic_root_bh and c->ic_descr->id_ops
361 * haven't been intialized yet.
362 * Also, we don't have this for IAM dir.
364 if (c->ic_root_bh != NULL &&
365 c->ic_descr->id_ops->id_root_ptr(c) == ptr) {
366 get_bh(c->ic_root_bh);
371 *bh = ldiskfs_bread(h, c->ic_object, (int)ptr, 0, &result);
378 * Return pointer to current leaf record. Pointer is valid while corresponding
379 * leaf node is locked and pinned.
381 static struct iam_rec *iam_leaf_rec(const struct iam_leaf *leaf)
383 return iam_leaf_ops(leaf)->rec(leaf);
387 * Return pointer to the current leaf key. This function returns pointer to
388 * the key stored in node.
390 * Caller should assume that returned pointer is only valid while leaf node is
393 static struct iam_key *iam_leaf_key(const struct iam_leaf *leaf)
395 return iam_leaf_ops(leaf)->key(leaf);
398 static int iam_leaf_key_size(const struct iam_leaf *leaf)
400 return iam_leaf_ops(leaf)->key_size(leaf);
403 static struct iam_ikey *iam_leaf_ikey(const struct iam_leaf *leaf,
404 struct iam_ikey *key)
406 return iam_leaf_ops(leaf)->ikey(leaf, key);
409 static int iam_leaf_keycmp(const struct iam_leaf *leaf,
410 const struct iam_key *key)
412 return iam_leaf_ops(leaf)->key_cmp(leaf, key);
415 static int iam_leaf_keyeq(const struct iam_leaf *leaf,
416 const struct iam_key *key)
418 return iam_leaf_ops(leaf)->key_eq(leaf, key);
421 #if LDISKFS_INVARIANT_ON
422 static int iam_leaf_check(struct iam_leaf *leaf);
423 extern int dx_node_check(struct iam_path *p, struct iam_frame *f);
425 static int iam_path_check(struct iam_path *p)
430 struct iam_descr *param;
433 param = iam_path_descr(p);
434 for (i = 0; result && i < ARRAY_SIZE(p->ip_frames); ++i) {
435 f = &p->ip_frames[i];
437 result = dx_node_check(p, f);
439 result = !param->id_ops->id_node_check(p, f);
442 if (result && p->ip_leaf.il_bh != NULL)
443 result = iam_leaf_check(&p->ip_leaf);
445 ldiskfs_std_error(iam_path_obj(p)->i_sb, result);
451 static int iam_leaf_load(struct iam_path *path)
455 struct iam_container *c;
456 struct buffer_head *bh;
457 struct iam_leaf *leaf;
458 struct iam_descr *descr;
460 c = path->ip_container;
461 leaf = &path->ip_leaf;
462 descr = iam_path_descr(path);
463 block = path->ip_frame->leaf;
466 printk(KERN_EMERG "wrong leaf: %lu %d [%p %p %p]\n",
467 (long unsigned)path->ip_frame->leaf,
468 dx_get_count(dx_node_get_entries(path, path->ip_frame)),
469 path->ip_frames[0].bh, path->ip_frames[1].bh,
470 path->ip_frames[2].bh);
472 err = descr->id_ops->id_node_read(c, block, NULL, &bh);
475 leaf->il_curidx = block;
476 err = iam_leaf_ops(leaf)->init(leaf);
477 assert_inv(ergo(err == 0, iam_leaf_check(leaf)));
482 static void iam_unlock_htree(struct iam_container *ic,
483 struct dynlock_handle *lh)
486 dynlock_unlock(&ic->ic_tree_lock, lh);
490 static void iam_leaf_unlock(struct iam_leaf *leaf)
492 if (leaf->il_lock != NULL) {
493 iam_unlock_htree(iam_leaf_container(leaf),
496 leaf->il_lock = NULL;
500 static void iam_leaf_fini(struct iam_leaf *leaf)
502 if (leaf->il_path != NULL) {
503 iam_leaf_unlock(leaf);
504 assert_inv(ergo(leaf->il_bh != NULL, iam_leaf_check(leaf)));
505 iam_leaf_ops(leaf)->fini(leaf);
514 static void iam_leaf_start(struct iam_leaf *folio)
516 iam_leaf_ops(folio)->start(folio);
519 void iam_leaf_next(struct iam_leaf *folio)
521 iam_leaf_ops(folio)->next(folio);
524 static void iam_leaf_rec_add(struct iam_leaf *leaf, const struct iam_key *key,
525 const struct iam_rec *rec)
527 iam_leaf_ops(leaf)->rec_add(leaf, key, rec);
530 static void iam_rec_del(struct iam_leaf *leaf, int shift)
532 iam_leaf_ops(leaf)->rec_del(leaf, shift);
535 int iam_leaf_at_end(const struct iam_leaf *leaf)
537 return iam_leaf_ops(leaf)->at_end(leaf);
540 static void iam_leaf_split(struct iam_leaf *l, struct buffer_head **bh,
543 iam_leaf_ops(l)->split(l, bh, nr);
546 static inline int iam_leaf_empty(struct iam_leaf *l)
548 return iam_leaf_ops(l)->leaf_empty(l);
551 int iam_leaf_can_add(const struct iam_leaf *l,
552 const struct iam_key *k, const struct iam_rec *r)
554 return iam_leaf_ops(l)->can_add(l, k, r);
557 #if LDISKFS_INVARIANT_ON
558 static int iam_leaf_check(struct iam_leaf *leaf)
562 struct iam_lentry *orig;
563 struct iam_path *path;
564 struct iam_container *bag;
571 path = iam_leaf_path(leaf);
572 bag = iam_leaf_container(leaf);
574 result = iam_leaf_ops(leaf)->init(leaf);
579 iam_leaf_start(leaf);
580 k0 = iam_path_ikey(path, 0);
581 k1 = iam_path_ikey(path, 1);
582 while (!iam_leaf_at_end(leaf)) {
583 iam_ikeycpy(bag, k0, k1);
584 iam_ikeycpy(bag, k1, iam_leaf_ikey(leaf, k1));
585 if (!first && iam_ikeycmp(bag, k0, k1) > 0) {
597 static int iam_txn_dirty(handle_t *handle,
598 struct iam_path *path, struct buffer_head *bh)
602 result = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
604 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
608 static int iam_txn_add(handle_t *handle,
609 struct iam_path *path, struct buffer_head *bh)
613 result = ldiskfs_journal_get_write_access(handle, bh);
615 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
619 /***********************************************************************/
620 /* iterator interface */
621 /***********************************************************************/
623 static enum iam_it_state it_state(const struct iam_iterator *it)
629 * Helper function returning scratch key.
631 static struct iam_container *iam_it_container(const struct iam_iterator *it)
633 return it->ii_path.ip_container;
636 static inline int it_keycmp(const struct iam_iterator *it,
637 const struct iam_key *k)
639 return iam_leaf_keycmp(&it->ii_path.ip_leaf, k);
642 static inline int it_keyeq(const struct iam_iterator *it,
643 const struct iam_key *k)
645 return iam_leaf_keyeq(&it->ii_path.ip_leaf, k);
648 static int it_ikeycmp(const struct iam_iterator *it, const struct iam_ikey *ik)
650 return iam_ikeycmp(it->ii_path.ip_container,
651 iam_leaf_ikey(&it->ii_path.ip_leaf,
652 iam_path_ikey(&it->ii_path, 0)), ik);
655 static inline int it_at_rec(const struct iam_iterator *it)
657 return !iam_leaf_at_end(&it->ii_path.ip_leaf);
660 static inline int it_before(const struct iam_iterator *it)
662 return it_state(it) == IAM_IT_SKEWED && it_at_rec(it);
666 * Helper wrapper around iam_it_get(): returns 0 (success) only when record
667 * with exactly the same key as asked is found.
669 static int iam_it_get_exact(struct iam_iterator *it, const struct iam_key *k)
673 result = iam_it_get(it, k);
676 else if (result == 0)
678 * Return -ENOENT if cursor is located above record with a key
679 * different from one specified, or in the empty leaf.
681 * XXX returning -ENOENT only works if iam_it_get() never
682 * returns -ENOENT as a legitimate error.
688 void iam_container_write_lock(struct iam_container *ic)
690 down_write(&ic->ic_sem);
693 void iam_container_write_unlock(struct iam_container *ic)
695 up_write(&ic->ic_sem);
698 void iam_container_read_lock(struct iam_container *ic)
700 down_read(&ic->ic_sem);
703 void iam_container_read_unlock(struct iam_container *ic)
705 up_read(&ic->ic_sem);
709 * Initialize iterator to IAM_IT_DETACHED state.
711 * postcondition: it_state(it) == IAM_IT_DETACHED
713 int iam_it_init(struct iam_iterator *it, struct iam_container *c, __u32 flags,
714 struct iam_path_descr *pd)
716 memset(it, 0, sizeof *it);
717 it->ii_flags = flags;
718 it->ii_state = IAM_IT_DETACHED;
719 iam_path_init(&it->ii_path, c, pd);
724 * Finalize iterator and release all resources.
726 * precondition: it_state(it) == IAM_IT_DETACHED
728 void iam_it_fini(struct iam_iterator *it)
730 assert_corr(it_state(it) == IAM_IT_DETACHED);
731 iam_path_fini(&it->ii_path);
735 * this locking primitives are used to protect parts
736 * of dir's htree. protection unit is block: leaf or index
738 static struct dynlock_handle *iam_lock_htree(struct iam_container *ic,
740 enum dynlock_type lt)
742 return dynlock_lock(&ic->ic_tree_lock, value, lt, GFP_NOFS);
745 static int iam_index_lock(struct iam_path *path, struct dynlock_handle **lh)
749 for (f = path->ip_frame; f >= path->ip_frames; --f, ++lh) {
751 *lh = iam_lock_htree(path->ip_container, f->curidx, DLT_READ);
759 * Fast check for frame consistency.
761 static int iam_check_fast(struct iam_path *path, struct iam_frame *frame)
763 struct iam_container *bag;
764 struct iam_entry *next;
765 struct iam_entry *last;
766 struct iam_entry *entries;
767 struct iam_entry *at;
769 bag = path->ip_container;
771 entries = frame->entries;
772 last = iam_entry_shift(path, entries, dx_get_count(entries) - 1);
774 if (unlikely(at > last))
777 if (unlikely(dx_get_block(path, at) != frame->leaf))
780 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, at),
781 path->ip_ikey_target) > 0))
784 next = iam_entry_shift(path, at, +1);
786 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, next),
787 path->ip_ikey_target) <= 0))
793 int dx_index_is_compat(struct iam_path *path)
795 return iam_path_descr(path) == NULL;
801 * search position of specified hash in index
805 static struct iam_entry *iam_find_position(struct iam_path *path,
806 struct iam_frame *frame)
813 count = dx_get_count(frame->entries);
814 assert_corr(count && count <= dx_get_limit(frame->entries));
815 p = iam_entry_shift(path, frame->entries,
816 dx_index_is_compat(path) ? 1 : 2);
817 q = iam_entry_shift(path, frame->entries, count - 1);
819 m = iam_entry_shift(path, p, iam_entry_diff(path, q, p) / 2);
820 if (iam_ikeycmp(path->ip_container, iam_ikey_at(path, m),
821 path->ip_ikey_target) > 0)
822 q = iam_entry_shift(path, m, -1);
824 p = iam_entry_shift(path, m, +1);
826 return iam_entry_shift(path, p, -1);
831 static iam_ptr_t iam_find_ptr(struct iam_path *path, struct iam_frame *frame)
833 return dx_get_block(path, iam_find_position(path, frame));
836 void iam_insert_key(struct iam_path *path, struct iam_frame *frame,
837 const struct iam_ikey *key, iam_ptr_t ptr)
839 struct iam_entry *entries = frame->entries;
840 struct iam_entry *new = iam_entry_shift(path, frame->at, +1);
841 int count = dx_get_count(entries);
844 * Unfortunately we cannot assert this, as this function is sometimes
845 * called by VFS under i_sem and without pdirops lock.
847 assert_corr(1 || iam_frame_is_locked(path, frame));
848 assert_corr(count < dx_get_limit(entries));
849 assert_corr(frame->at < iam_entry_shift(path, entries, count));
850 assert_inv(dx_node_check(path, frame));
852 memmove(iam_entry_shift(path, new, 1), new,
853 (char *)iam_entry_shift(path, entries, count) - (char *)new);
854 dx_set_ikey(path, new, key);
855 dx_set_block(path, new, ptr);
856 dx_set_count(entries, count + 1);
857 assert_inv(dx_node_check(path, frame));
860 void iam_insert_key_lock(struct iam_path *path, struct iam_frame *frame,
861 const struct iam_ikey *key, iam_ptr_t ptr)
863 iam_lock_bh(frame->bh);
864 iam_insert_key(path, frame, key, ptr);
865 iam_unlock_bh(frame->bh);
868 * returns 0 if path was unchanged, -EAGAIN otherwise.
870 static int iam_check_path(struct iam_path *path, struct iam_frame *frame)
874 iam_lock_bh(frame->bh);
875 equal = iam_check_fast(path, frame) == 0 ||
876 frame->leaf == iam_find_ptr(path, frame);
877 DX_DEVAL(iam_lock_stats.dls_bh_again += !equal);
878 iam_unlock_bh(frame->bh);
880 return equal ? 0 : -EAGAIN;
883 static int iam_lookup_try(struct iam_path *path)
889 struct iam_descr *param;
890 struct iam_frame *frame;
891 struct iam_container *c;
893 param = iam_path_descr(path);
894 c = path->ip_container;
896 ptr = param->id_ops->id_root_ptr(c);
897 for (frame = path->ip_frames, i = 0; i <= path->ip_indirect;
899 err = param->id_ops->id_node_read(c, (iam_ptr_t)ptr, NULL,
903 iam_lock_bh(frame->bh);
905 * node must be initialized under bh lock because concurrent
906 * creation procedure may change it and iam_lookup_try() will
907 * see obsolete tree height. -bzzz
912 if (LDISKFS_INVARIANT_ON) {
913 err = param->id_ops->id_node_check(path, frame);
918 err = param->id_ops->id_node_load(path, frame);
922 assert_inv(dx_node_check(path, frame));
924 * splitting may change root index block and move hash we're
925 * looking for into another index block so, we have to check
926 * this situation and repeat from begining if path got changed
930 err = iam_check_path(path, frame - 1);
935 frame->at = iam_find_position(path, frame);
937 frame->leaf = ptr = dx_get_block(path, frame->at);
939 iam_unlock_bh(frame->bh);
943 iam_unlock_bh(frame->bh);
944 path->ip_frame = --frame;
948 static int __iam_path_lookup(struct iam_path *path)
953 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++ i)
954 assert(path->ip_frames[i].bh == NULL);
957 err = iam_lookup_try(path);
961 } while (err == -EAGAIN);
967 * returns 0 if path was unchanged, -EAGAIN otherwise.
969 static int iam_check_full_path(struct iam_path *path, int search)
971 struct iam_frame *bottom;
972 struct iam_frame *scan;
978 for (bottom = path->ip_frames, i = 0;
979 i < DX_MAX_TREE_HEIGHT && bottom->bh != NULL; ++bottom, ++i) {
980 ; /* find last filled in frame */
984 * Lock frames, bottom to top.
986 for (scan = bottom - 1; scan >= path->ip_frames; --scan)
987 iam_lock_bh(scan->bh);
989 * Check them top to bottom.
992 for (scan = path->ip_frames; scan < bottom; ++scan) {
993 struct iam_entry *pos;
996 if (iam_check_fast(path, scan) == 0)
999 pos = iam_find_position(path, scan);
1000 if (scan->leaf != dx_get_block(path, pos)) {
1006 pos = iam_entry_shift(path, scan->entries,
1007 dx_get_count(scan->entries) - 1);
1008 if (scan->at > pos ||
1009 scan->leaf != dx_get_block(path, scan->at)) {
1017 * Unlock top to bottom.
1019 for (scan = path->ip_frames; scan < bottom; ++scan)
1020 iam_unlock_bh(scan->bh);
1021 DX_DEVAL(iam_lock_stats.dls_bh_full_again += !!result);
1022 do_corr(schedule());
1029 * Performs path lookup and returns with found leaf (if any) locked by htree
1032 static int iam_lookup_lock(struct iam_path *path,
1033 struct dynlock_handle **dl, enum dynlock_type lt)
1037 while ((result = __iam_path_lookup(path)) == 0) {
1038 do_corr(schedule());
1039 *dl = iam_lock_htree(path->ip_container, path->ip_frame->leaf,
1042 iam_path_fini(path);
1046 do_corr(schedule());
1048 * while locking leaf we just found may get split so we need
1049 * to check this -bzzz
1051 if (iam_check_full_path(path, 1) == 0)
1053 iam_unlock_htree(path->ip_container, *dl);
1055 iam_path_fini(path);
1060 * Performs tree top-to-bottom traversal starting from root, and loads leaf
1063 static int iam_path_lookup(struct iam_path *path, int index)
1065 struct iam_container *c;
1066 struct iam_leaf *leaf;
1069 c = path->ip_container;
1070 leaf = &path->ip_leaf;
1071 result = iam_lookup_lock(path, &leaf->il_lock, DLT_WRITE);
1072 assert_inv(iam_path_check(path));
1073 do_corr(schedule());
1075 result = iam_leaf_load(path);
1076 assert_inv(ergo(result == 0, iam_leaf_check(leaf)));
1078 do_corr(schedule());
1080 result = iam_leaf_ops(leaf)->
1081 ilookup(leaf, path->ip_ikey_target);
1083 result = iam_leaf_ops(leaf)->
1084 lookup(leaf, path->ip_key_target);
1085 do_corr(schedule());
1088 iam_leaf_unlock(leaf);
1094 * Common part of iam_it_{i,}get().
1096 static int __iam_it_get(struct iam_iterator *it, int index)
1099 assert_corr(it_state(it) == IAM_IT_DETACHED);
1101 result = iam_path_lookup(&it->ii_path, index);
1105 collision = result & IAM_LOOKUP_LAST;
1106 switch (result & ~IAM_LOOKUP_LAST) {
1107 case IAM_LOOKUP_EXACT:
1109 it->ii_state = IAM_IT_ATTACHED;
1113 it->ii_state = IAM_IT_ATTACHED;
1115 case IAM_LOOKUP_BEFORE:
1116 case IAM_LOOKUP_EMPTY:
1118 it->ii_state = IAM_IT_SKEWED;
1123 result |= collision;
1126 * See iam_it_get_exact() for explanation.
1128 assert_corr(result != -ENOENT);
1133 * Correct hash, but not the same key was found, iterate through hash
1134 * collision chain, looking for correct record.
1136 static int iam_it_collision(struct iam_iterator *it)
1140 assert(ergo(it_at_rec(it), !it_keyeq(it, it->ii_path.ip_key_target)));
1142 while ((result = iam_it_next(it)) == 0) {
1143 do_corr(schedule());
1144 if (it_ikeycmp(it, it->ii_path.ip_ikey_target) != 0)
1146 if (it_keyeq(it, it->ii_path.ip_key_target))
1153 * Attach iterator. After successful completion, @it points to record with
1154 * least key not larger than @k.
1156 * Return value: 0: positioned on existing record,
1157 * +ve: exact position found,
1160 * precondition: it_state(it) == IAM_IT_DETACHED
1161 * postcondition: ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1162 * it_keycmp(it, k) <= 0)
1164 int iam_it_get(struct iam_iterator *it, const struct iam_key *k)
1167 assert_corr(it_state(it) == IAM_IT_DETACHED);
1169 it->ii_path.ip_ikey_target = NULL;
1170 it->ii_path.ip_key_target = k;
1172 result = __iam_it_get(it, 0);
1174 if (result == IAM_LOOKUP_LAST) {
1175 result = iam_it_collision(it);
1179 result = __iam_it_get(it, 0);
1184 result &= ~IAM_LOOKUP_LAST;
1186 assert_corr(ergo(result > 0, it_keycmp(it, k) == 0));
1187 assert_corr(ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1188 it_keycmp(it, k) <= 0));
1193 * Attach iterator by index key.
1195 static int iam_it_iget(struct iam_iterator *it, const struct iam_ikey *k)
1197 assert_corr(it_state(it) == IAM_IT_DETACHED);
1199 it->ii_path.ip_ikey_target = k;
1200 return __iam_it_get(it, 1) & ~IAM_LOOKUP_LAST;
1204 * Attach iterator, and assure it points to the record (not skewed).
1206 * Return value: 0: positioned on existing record,
1207 * +ve: exact position found,
1210 * precondition: it_state(it) == IAM_IT_DETACHED &&
1211 * !(it->ii_flags&IAM_IT_WRITE)
1212 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED)
1214 int iam_it_get_at(struct iam_iterator *it, const struct iam_key *k)
1217 assert_corr(it_state(it) == IAM_IT_DETACHED &&
1218 !(it->ii_flags&IAM_IT_WRITE));
1219 result = iam_it_get(it, k);
1221 if (it_state(it) != IAM_IT_ATTACHED) {
1222 assert_corr(it_state(it) == IAM_IT_SKEWED);
1223 result = iam_it_next(it);
1226 assert_corr(ergo(result >= 0, it_state(it) == IAM_IT_ATTACHED));
1231 * Duplicates iterator.
1233 * postcondition: it_state(dst) == it_state(src) &&
1234 * iam_it_container(dst) == iam_it_container(src) &&
1235 * dst->ii_flags = src->ii_flags &&
1236 * ergo(it_state(src) == IAM_IT_ATTACHED,
1237 * iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1238 * iam_it_key_get(dst) == iam_it_key_get(src))
1240 void iam_it_dup(struct iam_iterator *dst, const struct iam_iterator *src)
1242 dst->ii_flags = src->ii_flags;
1243 dst->ii_state = src->ii_state;
1244 /* XXX not yet. iam_path_dup(&dst->ii_path, &src->ii_path); */
1246 * XXX: duplicate lock.
1248 assert_corr(it_state(dst) == it_state(src));
1249 assert_corr(iam_it_container(dst) == iam_it_container(src));
1250 assert_corr(dst->ii_flags = src->ii_flags);
1251 assert_corr(ergo(it_state(src) == IAM_IT_ATTACHED,
1252 iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1253 iam_it_key_get(dst) == iam_it_key_get(src)));
1258 * Detach iterator. Does nothing it detached state.
1260 * postcondition: it_state(it) == IAM_IT_DETACHED
1262 void iam_it_put(struct iam_iterator *it)
1264 if (it->ii_state != IAM_IT_DETACHED) {
1265 it->ii_state = IAM_IT_DETACHED;
1266 iam_leaf_fini(&it->ii_path.ip_leaf);
1270 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1271 struct iam_ikey *ikey);
1275 * This function increments the frame pointer to search the next leaf
1276 * block, and reads in the necessary intervening nodes if the search
1277 * should be necessary. Whether or not the search is necessary is
1278 * controlled by the hash parameter. If the hash value is even, then
1279 * the search is only continued if the next block starts with that
1280 * hash value. This is used if we are searching for a specific file.
1282 * If the hash value is HASH_NB_ALWAYS, then always go to the next block.
1284 * This function returns 1 if the caller should continue to search,
1285 * or 0 if it should not. If there is an error reading one of the
1286 * index blocks, it will a negative error code.
1288 * If start_hash is non-null, it will be filled in with the starting
1289 * hash of the next page.
1291 static int iam_htree_advance(struct inode *dir, __u32 hash,
1292 struct iam_path *path, __u32 *start_hash,
1295 struct iam_frame *p;
1296 struct buffer_head *bh;
1297 int err, num_frames = 0;
1302 * Find the next leaf page by incrementing the frame pointer.
1303 * If we run out of entries in the interior node, loop around and
1304 * increment pointer in the parent node. When we break out of
1305 * this loop, num_frames indicates the number of interior
1306 * nodes need to be read.
1309 do_corr(schedule());
1314 p->at = iam_entry_shift(path, p->at, +1);
1315 if (p->at < iam_entry_shift(path, p->entries,
1316 dx_get_count(p->entries))) {
1317 p->leaf = dx_get_block(path, p->at);
1318 iam_unlock_bh(p->bh);
1321 iam_unlock_bh(p->bh);
1322 if (p == path->ip_frames)
1333 * If the hash is 1, then continue only if the next page has a
1334 * continuation hash of any value. This is used for readdir
1335 * handling. Otherwise, check to see if the hash matches the
1336 * desired contiuation hash. If it doesn't, return since
1337 * there's no point to read in the successive index pages.
1339 dx_get_ikey(path, p->at, (struct iam_ikey *)&bhash);
1341 *start_hash = bhash;
1342 if ((hash & 1) == 0) {
1343 if ((bhash & ~1) != hash)
1348 * If the hash is HASH_NB_ALWAYS, we always go to the next
1349 * block so no check is necessary
1351 while (num_frames--) {
1354 do_corr(schedule());
1356 idx = p->leaf = dx_get_block(path, p->at);
1357 iam_unlock_bh(p->bh);
1358 err = iam_path_descr(path)->id_ops->
1359 id_node_read(path->ip_container, idx, NULL, &bh);
1361 return err; /* Failure */
1364 assert_corr(p->bh != bh);
1366 p->entries = dx_node_get_entries(path, p);
1367 p->at = iam_entry_shift(path, p->entries, !compat);
1368 assert_corr(p->curidx != idx);
1371 assert_corr(p->leaf != dx_get_block(path, p->at));
1372 p->leaf = dx_get_block(path, p->at);
1373 iam_unlock_bh(p->bh);
1374 assert_inv(dx_node_check(path, p));
1380 static inline int iam_index_advance(struct iam_path *path)
1382 return iam_htree_advance(iam_path_obj(path), 0, path, NULL, 0);
1385 static void iam_unlock_array(struct iam_container *ic,
1386 struct dynlock_handle **lh)
1390 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++i, ++lh) {
1392 iam_unlock_htree(ic, *lh);
1398 * Advance index part of @path to point to the next leaf. Returns 1 on
1399 * success, 0, when end of container was reached. Leaf node is locked.
1401 int iam_index_next(struct iam_container *c, struct iam_path *path)
1404 struct dynlock_handle *lh[DX_MAX_TREE_HEIGHT] = { NULL, };
1406 struct inode *object;
1409 * Locking for iam_index_next()... is to be described.
1412 object = c->ic_object;
1413 cursor = path->ip_frame->leaf;
1416 result = iam_index_lock(path, lh);
1417 do_corr(schedule());
1421 result = iam_check_full_path(path, 0);
1422 if (result == 0 && cursor == path->ip_frame->leaf) {
1423 result = iam_index_advance(path);
1425 assert_corr(result == 0 ||
1426 cursor != path->ip_frame->leaf);
1430 iam_unlock_array(c, lh);
1432 iam_path_release(path);
1433 do_corr(schedule());
1435 result = __iam_path_lookup(path);
1439 while (path->ip_frame->leaf != cursor) {
1440 do_corr(schedule());
1442 result = iam_index_lock(path, lh);
1443 do_corr(schedule());
1447 result = iam_check_full_path(path, 0);
1451 result = iam_index_advance(path);
1453 CERROR("cannot find cursor : %u\n",
1459 result = iam_check_full_path(path, 0);
1462 iam_unlock_array(c, lh);
1464 } while (result == -EAGAIN);
1468 iam_unlock_array(c, lh);
1473 * Move iterator one record right.
1475 * Return value: 0: success,
1476 * +1: end of container reached
1479 * precondition: (it_state(it) == IAM_IT_ATTACHED ||
1480 * it_state(it) == IAM_IT_SKEWED) && it->ii_flags&IAM_IT_MOVE
1481 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED) &&
1482 * ergo(result > 0, it_state(it) == IAM_IT_DETACHED)
1484 int iam_it_next(struct iam_iterator *it)
1487 struct iam_path *path;
1488 struct iam_leaf *leaf;
1489 do_corr(struct iam_ikey *ik_orig);
1491 /* assert_corr(it->ii_flags&IAM_IT_MOVE); */
1492 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1493 it_state(it) == IAM_IT_SKEWED);
1495 path = &it->ii_path;
1496 leaf = &path->ip_leaf;
1498 assert_corr(iam_leaf_is_locked(leaf));
1501 do_corr(ik_orig = it_at_rec(it) ?
1502 iam_it_ikey_get(it, iam_path_ikey(path, 2)) : NULL);
1503 if (it_before(it)) {
1504 assert_corr(!iam_leaf_at_end(leaf));
1505 it->ii_state = IAM_IT_ATTACHED;
1507 if (!iam_leaf_at_end(leaf))
1508 /* advance within leaf node */
1509 iam_leaf_next(leaf);
1511 * multiple iterations may be necessary due to empty leaves.
1513 while (result == 0 && iam_leaf_at_end(leaf)) {
1514 do_corr(schedule());
1515 /* advance index portion of the path */
1516 result = iam_index_next(iam_it_container(it), path);
1517 assert_corr(iam_leaf_is_locked(leaf));
1519 struct dynlock_handle *lh;
1520 lh = iam_lock_htree(iam_it_container(it),
1521 path->ip_frame->leaf,
1524 iam_leaf_fini(leaf);
1526 result = iam_leaf_load(path);
1528 iam_leaf_start(leaf);
1531 } else if (result == 0)
1532 /* end of container reached */
1538 it->ii_state = IAM_IT_ATTACHED;
1540 assert_corr(ergo(result == 0, it_state(it) == IAM_IT_ATTACHED));
1541 assert_corr(ergo(result > 0, it_state(it) == IAM_IT_DETACHED));
1542 assert_corr(ergo(result == 0 && ik_orig != NULL,
1543 it_ikeycmp(it, ik_orig) >= 0));
1548 * Return pointer to the record under iterator.
1550 * precondition: it_state(it) == IAM_IT_ATTACHED && it_at_rec(it)
1551 * postcondition: it_state(it) == IAM_IT_ATTACHED
1553 struct iam_rec *iam_it_rec_get(const struct iam_iterator *it)
1555 assert_corr(it_state(it) == IAM_IT_ATTACHED);
1556 assert_corr(it_at_rec(it));
1557 return iam_leaf_rec(&it->ii_path.ip_leaf);
1560 static void iam_it_reccpy(struct iam_iterator *it, const struct iam_rec *r)
1562 struct iam_leaf *folio;
1564 folio = &it->ii_path.ip_leaf;
1565 iam_leaf_ops(folio)->rec_set(folio, r);
1569 * Replace contents of record under iterator.
1571 * precondition: it_state(it) == IAM_IT_ATTACHED &&
1572 * it->ii_flags&IAM_IT_WRITE
1573 * postcondition: it_state(it) == IAM_IT_ATTACHED &&
1574 * ergo(result == 0, !memcmp(iam_it_rec_get(it), r, ...))
1576 int iam_it_rec_set(handle_t *h,
1577 struct iam_iterator *it, const struct iam_rec *r)
1580 struct iam_path *path;
1581 struct buffer_head *bh;
1583 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
1584 it->ii_flags&IAM_IT_WRITE);
1585 assert_corr(it_at_rec(it));
1587 path = &it->ii_path;
1588 bh = path->ip_leaf.il_bh;
1589 result = iam_txn_add(h, path, bh);
1591 iam_it_reccpy(it, r);
1592 result = iam_txn_dirty(h, path, bh);
1598 * Return pointer to the index key under iterator.
1600 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1601 * it_state(it) == IAM_IT_SKEWED
1603 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1604 struct iam_ikey *ikey)
1606 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1607 it_state(it) == IAM_IT_SKEWED);
1608 assert_corr(it_at_rec(it));
1609 return iam_leaf_ikey(&it->ii_path.ip_leaf, ikey);
1613 * Return pointer to the key under iterator.
1615 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1616 * it_state(it) == IAM_IT_SKEWED
1618 struct iam_key *iam_it_key_get(const struct iam_iterator *it)
1620 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1621 it_state(it) == IAM_IT_SKEWED);
1622 assert_corr(it_at_rec(it));
1623 return iam_leaf_key(&it->ii_path.ip_leaf);
1627 * Return size of key under iterator (in bytes)
1629 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1630 * it_state(it) == IAM_IT_SKEWED
1632 int iam_it_key_size(const struct iam_iterator *it)
1634 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1635 it_state(it) == IAM_IT_SKEWED);
1636 assert_corr(it_at_rec(it));
1637 return iam_leaf_key_size(&it->ii_path.ip_leaf);
1640 static struct buffer_head *
1641 iam_new_node(handle_t *h, struct iam_container *c, iam_ptr_t *b, int *e)
1643 struct inode *inode = c->ic_object;
1644 struct buffer_head *bh = NULL;
1645 struct iam_idle_head *head;
1646 struct buffer_head *idle;
1650 if (c->ic_idle_bh == NULL)
1653 mutex_lock(&c->ic_idle_mutex);
1654 if (unlikely(c->ic_idle_bh == NULL)) {
1655 mutex_unlock(&c->ic_idle_mutex);
1659 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
1660 count = le16_to_cpu(head->iih_count);
1662 *e = ldiskfs_journal_get_write_access(h, c->ic_idle_bh);
1667 *b = le32_to_cpu(head->iih_blks[count]);
1668 head->iih_count = cpu_to_le16(count);
1669 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_idle_bh);
1673 mutex_unlock(&c->ic_idle_mutex);
1674 bh = ldiskfs_bread(NULL, inode, *b, 0, e);
1680 /* The block itself which contains the iam_idle_head is
1681 * also an idle block, and can be used as the new node. */
1682 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
1683 c->ic_descr->id_root_gap +
1684 sizeof(struct dx_countlimit));
1685 *e = ldiskfs_journal_get_write_access(h, c->ic_root_bh);
1689 *b = le32_to_cpu(*idle_blocks);
1690 iam_lock_bh(c->ic_root_bh);
1691 *idle_blocks = head->iih_next;
1692 iam_unlock_bh(c->ic_root_bh);
1693 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_root_bh);
1695 iam_lock_bh(c->ic_root_bh);
1696 *idle_blocks = cpu_to_le32(*b);
1697 iam_unlock_bh(c->ic_root_bh);
1702 idle = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
1703 if (idle != NULL && IS_ERR(idle)) {
1705 c->ic_idle_bh = NULL;
1710 c->ic_idle_bh = idle;
1711 mutex_unlock(&c->ic_idle_mutex);
1714 /* get write access for the found buffer head */
1715 *e = ldiskfs_journal_get_write_access(h, bh);
1719 ldiskfs_std_error(inode->i_sb, *e);
1721 /* Clear the reused node as new node does. */
1722 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1723 set_buffer_uptodate(bh);
1728 bh = osd_ldiskfs_append(h, inode, b, e);
1732 mutex_unlock(&c->ic_idle_mutex);
1733 ldiskfs_std_error(inode->i_sb, *e);
1738 * Insertion of new record. Interaction with jbd during non-trivial case (when
1739 * split happens) is as following:
1741 * - new leaf node is involved into transaction by iam_new_node();
1743 * - old leaf node is involved into transaction by iam_add_rec();
1745 * - leaf where insertion point ends in, is marked dirty by iam_add_rec();
1747 * - leaf without insertion point is marked dirty (as @new_leaf) by
1750 * - split index nodes are involved into transaction and marked dirty by
1751 * split_index_node().
1753 * - "safe" index node, which is no split, but where new pointer is inserted
1754 * is involved into transaction and marked dirty by split_index_node().
1756 * - index node where pointer to new leaf is inserted is involved into
1757 * transaction by split_index_node() and marked dirty by iam_add_rec().
1759 * - inode is marked dirty by iam_add_rec().
1763 static int iam_new_leaf(handle_t *handle, struct iam_leaf *leaf)
1767 struct buffer_head *new_leaf;
1768 struct buffer_head *old_leaf;
1769 struct iam_container *c;
1771 struct iam_path *path;
1773 assert_inv(iam_leaf_check(leaf));
1775 c = iam_leaf_container(leaf);
1776 path = leaf->il_path;
1779 new_leaf = iam_new_node(handle, c, &blknr, &err);
1780 do_corr(schedule());
1781 if (new_leaf != NULL) {
1782 struct dynlock_handle *lh;
1784 lh = iam_lock_htree(c, blknr, DLT_WRITE);
1785 do_corr(schedule());
1787 iam_leaf_ops(leaf)->init_new(c, new_leaf);
1788 do_corr(schedule());
1789 old_leaf = leaf->il_bh;
1790 iam_leaf_split(leaf, &new_leaf, blknr);
1791 if (old_leaf != leaf->il_bh) {
1793 * Switched to the new leaf.
1795 iam_leaf_unlock(leaf);
1797 path->ip_frame->leaf = blknr;
1799 iam_unlock_htree(path->ip_container, lh);
1800 do_corr(schedule());
1801 err = iam_txn_dirty(handle, path, new_leaf);
1804 err = ldiskfs_mark_inode_dirty(handle, obj);
1805 do_corr(schedule());
1809 assert_inv(iam_leaf_check(leaf));
1810 assert_inv(iam_leaf_check(&iam_leaf_path(leaf)->ip_leaf));
1811 assert_inv(iam_path_check(iam_leaf_path(leaf)));
1815 static inline void dx_set_limit(struct iam_entry *entries, unsigned value)
1817 ((struct dx_countlimit *) entries)->limit = cpu_to_le16(value);
1820 static int iam_shift_entries(struct iam_path *path,
1821 struct iam_frame *frame, unsigned count,
1822 struct iam_entry *entries, struct iam_entry *entries2,
1829 struct iam_frame *parent = frame - 1;
1830 struct iam_ikey *pivot = iam_path_ikey(path, 3);
1832 delta = dx_index_is_compat(path) ? 0 : +1;
1834 count1 = count/2 + delta;
1835 count2 = count - count1;
1836 dx_get_ikey(path, iam_entry_shift(path, entries, count1), pivot);
1838 dxtrace(printk("Split index %d/%d\n", count1, count2));
1840 memcpy((char *) iam_entry_shift(path, entries2, delta),
1841 (char *) iam_entry_shift(path, entries, count1),
1842 count2 * iam_entry_size(path));
1844 dx_set_count(entries2, count2 + delta);
1845 dx_set_limit(entries2, dx_node_limit(path));
1848 * NOTE: very subtle piece of code competing dx_probe() may find 2nd
1849 * level index in root index, then we insert new index here and set
1850 * new count in that 2nd level index. so, dx_probe() may see 2nd level
1851 * index w/o hash it looks for. the solution is to check root index
1852 * after we locked just founded 2nd level index -bzzz
1854 iam_insert_key_lock(path, parent, pivot, newblock);
1857 * now old and new 2nd level index blocks contain all pointers, so
1858 * dx_probe() may find it in the both. it's OK -bzzz
1860 iam_lock_bh(frame->bh);
1861 dx_set_count(entries, count1);
1862 iam_unlock_bh(frame->bh);
1865 * now old 2nd level index block points to first half of leafs. it's
1866 * importand that dx_probe() must check root index block for changes
1867 * under dx_lock_bh(frame->bh) -bzzz
1874 int split_index_node(handle_t *handle, struct iam_path *path,
1875 struct dynlock_handle **lh)
1878 struct iam_entry *entries; /* old block contents */
1879 struct iam_entry *entries2; /* new block contents */
1880 struct iam_frame *frame, *safe;
1881 struct buffer_head *bh_new[DX_MAX_TREE_HEIGHT] = {NULL};
1882 u32 newblock[DX_MAX_TREE_HEIGHT] = {0};
1883 struct dynlock_handle *lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1884 struct dynlock_handle *new_lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1885 struct inode *dir = iam_path_obj(path);
1886 struct iam_descr *descr;
1890 descr = iam_path_descr(path);
1892 * Algorithm below depends on this.
1894 assert_corr(dx_root_limit(path) < dx_node_limit(path));
1896 frame = path->ip_frame;
1897 entries = frame->entries;
1900 * Tall-tree handling: we might have to split multiple index blocks
1901 * all the way up to tree root. Tricky point here is error handling:
1902 * to avoid complicated undo/rollback we
1904 * - first allocate all necessary blocks
1906 * - insert pointers into them atomically.
1910 * Locking: leaf is already locked. htree-locks are acquired on all
1911 * index nodes that require split bottom-to-top, on the "safe" node,
1912 * and on all new nodes
1915 dxtrace(printk("using %u of %u node entries\n",
1916 dx_get_count(entries), dx_get_limit(entries)));
1918 /* What levels need split? */
1919 for (nr_splet = 0; frame >= path->ip_frames &&
1920 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1921 --frame, ++nr_splet) {
1922 do_corr(schedule());
1923 if (nr_splet == DX_MAX_TREE_HEIGHT) {
1925 CWARN(dir->i_sb, __FUNCTION__,
1926 "Directory index full!\n");
1936 * Lock all nodes, bottom to top.
1938 for (frame = path->ip_frame, i = nr_splet; i >= 0; --i, --frame) {
1939 do_corr(schedule());
1940 lock[i] = iam_lock_htree(path->ip_container, frame->curidx,
1942 if (lock[i] == NULL) {
1949 * Check for concurrent index modification.
1951 err = iam_check_full_path(path, 1);
1955 * And check that the same number of nodes is to be split.
1957 for (i = 0, frame = path->ip_frame; frame >= path->ip_frames &&
1958 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1962 if (i != nr_splet) {
1967 /* Go back down, allocating blocks, locking them, and adding into
1969 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
1970 bh_new[i] = iam_new_node(handle, path->ip_container,
1971 &newblock[i], &err);
1972 do_corr(schedule());
1974 descr->id_ops->id_node_init(path->ip_container,
1977 new_lock[i] = iam_lock_htree(path->ip_container, newblock[i],
1979 if (new_lock[i] == NULL) {
1983 do_corr(schedule());
1984 BUFFER_TRACE(frame->bh, "get_write_access");
1985 err = ldiskfs_journal_get_write_access(handle, frame->bh);
1989 /* Add "safe" node to transaction too */
1990 if (safe + 1 != path->ip_frames) {
1991 do_corr(schedule());
1992 err = ldiskfs_journal_get_write_access(handle, safe->bh);
1997 /* Go through nodes once more, inserting pointers */
1998 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
2001 struct buffer_head *bh2;
2002 struct buffer_head *bh;
2004 entries = frame->entries;
2005 count = dx_get_count(entries);
2006 idx = iam_entry_diff(path, frame->at, entries);
2009 entries2 = dx_get_entries(path, bh2->b_data, 0);
2012 if (frame == path->ip_frames) {
2013 /* splitting root node. Tricky point:
2015 * In the "normal" B-tree we'd split root *and* add
2016 * new root to the tree with pointers to the old root
2017 * and its sibling (thus introducing two new nodes).
2019 * In htree it's enough to add one node, because
2020 * capacity of the root node is smaller than that of
2023 struct iam_frame *frames;
2024 struct iam_entry *next;
2026 assert_corr(i == 0);
2028 do_corr(schedule());
2030 frames = path->ip_frames;
2031 memcpy((char *) entries2, (char *) entries,
2032 count * iam_entry_size(path));
2033 dx_set_limit(entries2, dx_node_limit(path));
2036 iam_lock_bh(frame->bh);
2037 next = descr->id_ops->id_root_inc(path->ip_container,
2039 dx_set_block(path, next, newblock[0]);
2040 iam_unlock_bh(frame->bh);
2042 do_corr(schedule());
2043 /* Shift frames in the path */
2044 memmove(frames + 2, frames + 1,
2045 (sizeof path->ip_frames) - 2 * sizeof frames[0]);
2046 /* Add new access path frame */
2047 frames[1].at = iam_entry_shift(path, entries2, idx);
2048 frames[1].entries = entries = entries2;
2050 assert_inv(dx_node_check(path, frame));
2053 assert_inv(dx_node_check(path, frame));
2054 bh_new[0] = NULL; /* buffer head is "consumed" */
2055 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2058 do_corr(schedule());
2060 /* splitting non-root index node. */
2061 struct iam_frame *parent = frame - 1;
2063 do_corr(schedule());
2064 count = iam_shift_entries(path, frame, count,
2065 entries, entries2, newblock[i]);
2066 /* Which index block gets the new entry? */
2068 int d = dx_index_is_compat(path) ? 0 : +1;
2070 frame->at = iam_entry_shift(path, entries2,
2072 frame->entries = entries = entries2;
2073 frame->curidx = newblock[i];
2074 swap(frame->bh, bh2);
2075 assert_corr(lock[i + 1] != NULL);
2076 assert_corr(new_lock[i] != NULL);
2077 swap(lock[i + 1], new_lock[i]);
2079 parent->at = iam_entry_shift(path,
2082 assert_inv(dx_node_check(path, frame));
2083 assert_inv(dx_node_check(path, parent));
2084 dxtrace(dx_show_index ("node", frame->entries));
2085 dxtrace(dx_show_index ("node",
2086 ((struct dx_node *) bh2->b_data)->entries));
2087 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2090 do_corr(schedule());
2091 err = ldiskfs_handle_dirty_metadata(handle, NULL,
2096 do_corr(schedule());
2097 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2102 * This function was called to make insertion of new leaf
2103 * possible. Check that it fulfilled its obligations.
2105 assert_corr(dx_get_count(path->ip_frame->entries) <
2106 dx_get_limit(path->ip_frame->entries));
2107 assert_corr(lock[nr_splet] != NULL);
2108 *lh = lock[nr_splet];
2109 lock[nr_splet] = NULL;
2112 * Log ->i_size modification.
2114 err = ldiskfs_mark_inode_dirty(handle, dir);
2120 ldiskfs_std_error(dir->i_sb, err);
2123 iam_unlock_array(path->ip_container, lock);
2124 iam_unlock_array(path->ip_container, new_lock);
2126 assert_corr(err || iam_frame_is_locked(path, path->ip_frame));
2128 do_corr(schedule());
2129 for (i = 0; i < ARRAY_SIZE(bh_new); ++i) {
2130 if (bh_new[i] != NULL)
2136 static int iam_add_rec(handle_t *handle, struct iam_iterator *it,
2137 struct iam_path *path,
2138 const struct iam_key *k, const struct iam_rec *r)
2141 struct iam_leaf *leaf;
2143 leaf = &path->ip_leaf;
2144 assert_inv(iam_leaf_check(leaf));
2145 assert_inv(iam_path_check(path));
2146 err = iam_txn_add(handle, path, leaf->il_bh);
2148 do_corr(schedule());
2149 if (!iam_leaf_can_add(leaf, k, r)) {
2150 struct dynlock_handle *lh = NULL;
2153 assert_corr(lh == NULL);
2154 do_corr(schedule());
2155 err = split_index_node(handle, path, &lh);
2156 if (err == -EAGAIN) {
2157 assert_corr(lh == NULL);
2159 iam_path_fini(path);
2160 it->ii_state = IAM_IT_DETACHED;
2162 do_corr(schedule());
2163 err = iam_it_get_exact(it, k);
2165 err = +1; /* repeat split */
2170 assert_inv(iam_path_check(path));
2172 assert_corr(lh != NULL);
2173 do_corr(schedule());
2174 err = iam_new_leaf(handle, leaf);
2176 err = iam_txn_dirty(handle, path,
2177 path->ip_frame->bh);
2179 iam_unlock_htree(path->ip_container, lh);
2180 do_corr(schedule());
2183 iam_leaf_rec_add(leaf, k, r);
2184 err = iam_txn_dirty(handle, path, leaf->il_bh);
2187 assert_inv(iam_leaf_check(leaf));
2188 assert_inv(iam_leaf_check(&path->ip_leaf));
2189 assert_inv(iam_path_check(path));
2194 * Insert new record with key @k and contents from @r, shifting records to the
2195 * right. On success, iterator is positioned on the newly inserted record.
2197 * precondition: it->ii_flags&IAM_IT_WRITE &&
2198 * (it_state(it) == IAM_IT_ATTACHED ||
2199 * it_state(it) == IAM_IT_SKEWED) &&
2200 * ergo(it_state(it) == IAM_IT_ATTACHED,
2201 * it_keycmp(it, k) <= 0) &&
2202 * ergo(it_before(it), it_keycmp(it, k) > 0));
2203 * postcondition: ergo(result == 0,
2204 * it_state(it) == IAM_IT_ATTACHED &&
2205 * it_keycmp(it, k) == 0 &&
2206 * !memcmp(iam_it_rec_get(it), r, ...))
2208 int iam_it_rec_insert(handle_t *h, struct iam_iterator *it,
2209 const struct iam_key *k, const struct iam_rec *r)
2212 struct iam_path *path;
2214 path = &it->ii_path;
2216 assert_corr(it->ii_flags&IAM_IT_WRITE);
2217 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2218 it_state(it) == IAM_IT_SKEWED);
2219 assert_corr(ergo(it_state(it) == IAM_IT_ATTACHED,
2220 it_keycmp(it, k) <= 0));
2221 assert_corr(ergo(it_before(it), it_keycmp(it, k) > 0));
2222 result = iam_add_rec(h, it, path, k, r);
2224 it->ii_state = IAM_IT_ATTACHED;
2225 assert_corr(ergo(result == 0,
2226 it_state(it) == IAM_IT_ATTACHED &&
2227 it_keycmp(it, k) == 0));
2231 static inline int iam_idle_blocks_limit(struct inode *inode)
2233 return (inode->i_sb->s_blocksize - sizeof(struct iam_idle_head)) >> 2;
2237 * If the leaf cannnot be recycled, we will lose one block for reusing.
2238 * It is not a serious issue because it almost the same of non-recycle.
2240 static iam_ptr_t iam_index_shrink(handle_t *h, struct iam_path *p,
2241 struct iam_leaf *l, struct buffer_head **bh)
2243 struct iam_container *c = p->ip_container;
2244 struct inode *inode = c->ic_object;
2245 struct iam_frame *frame = p->ip_frame;
2246 struct iam_entry *entries;
2247 struct iam_entry *pos;
2248 struct dynlock_handle *lh;
2252 if (c->ic_idle_failed)
2255 if (unlikely(frame == NULL))
2258 if (!iam_leaf_empty(l))
2261 lh = iam_lock_htree(c, frame->curidx, DLT_WRITE);
2263 CWARN("%.16s: No memory to recycle idle blocks\n",
2264 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name);
2268 rc = iam_txn_add(h, p, frame->bh);
2270 iam_unlock_htree(c, lh);
2274 iam_lock_bh(frame->bh);
2275 entries = frame->entries;
2276 count = dx_get_count(entries);
2277 /* NOT shrink the last entry in the index node, which can be reused
2278 * directly by next new node. */
2280 iam_unlock_bh(frame->bh);
2281 iam_unlock_htree(c, lh);
2285 pos = iam_find_position(p, frame);
2286 /* There may be some new leaf nodes have been added or empty leaf nodes
2287 * have been shrinked during my delete operation.
2289 * If the empty leaf is not under current index node because the index
2290 * node has been split, then just skip the empty leaf, which is rare. */
2291 if (unlikely(frame->leaf != dx_get_block(p, pos))) {
2292 iam_unlock_bh(frame->bh);
2293 iam_unlock_htree(c, lh);
2298 if (frame->at < iam_entry_shift(p, entries, count - 1)) {
2299 struct iam_entry *n = iam_entry_shift(p, frame->at, 1);
2301 memmove(frame->at, n,
2302 (char *)iam_entry_shift(p, entries, count) - (char *)n);
2303 frame->at_shifted = 1;
2305 dx_set_count(entries, count - 1);
2306 iam_unlock_bh(frame->bh);
2307 rc = iam_txn_dirty(h, p, frame->bh);
2308 iam_unlock_htree(c, lh);
2318 iam_install_idle_blocks(handle_t *h, struct iam_path *p, struct buffer_head *bh,
2319 __u32 *idle_blocks, iam_ptr_t blk)
2321 struct iam_container *c = p->ip_container;
2322 struct buffer_head *old = c->ic_idle_bh;
2323 struct iam_idle_head *head;
2326 head = (struct iam_idle_head *)(bh->b_data);
2327 head->iih_magic = cpu_to_le16(IAM_IDLE_HEADER_MAGIC);
2328 head->iih_count = 0;
2329 head->iih_next = *idle_blocks;
2330 /* The bh already get_write_accessed. */
2331 rc = iam_txn_dirty(h, p, bh);
2335 rc = iam_txn_add(h, p, c->ic_root_bh);
2339 iam_lock_bh(c->ic_root_bh);
2340 *idle_blocks = cpu_to_le32(blk);
2341 iam_unlock_bh(c->ic_root_bh);
2342 rc = iam_txn_dirty(h, p, c->ic_root_bh);
2344 /* NOT release old before new assigned. */
2349 iam_lock_bh(c->ic_root_bh);
2350 *idle_blocks = head->iih_next;
2351 iam_unlock_bh(c->ic_root_bh);
2357 * If the leaf cannnot be recycled, we will lose one block for reusing.
2358 * It is not a serious issue because it almost the same of non-recycle.
2360 static void iam_recycle_leaf(handle_t *h, struct iam_path *p,
2361 struct buffer_head *bh, iam_ptr_t blk)
2363 struct iam_container *c = p->ip_container;
2364 struct inode *inode = c->ic_object;
2365 struct iam_idle_head *head;
2370 mutex_lock(&c->ic_idle_mutex);
2371 if (unlikely(c->ic_idle_failed)) {
2376 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
2377 c->ic_descr->id_root_gap +
2378 sizeof(struct dx_countlimit));
2379 /* It is the first idle block. */
2380 if (c->ic_idle_bh == NULL) {
2381 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2385 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
2386 count = le16_to_cpu(head->iih_count);
2387 /* Current ic_idle_bh is full, to be replaced by the leaf. */
2388 if (count == iam_idle_blocks_limit(inode)) {
2389 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2393 /* Just add to ic_idle_bh. */
2394 rc = iam_txn_add(h, p, c->ic_idle_bh);
2398 head->iih_blks[count] = cpu_to_le32(blk);
2399 head->iih_count = cpu_to_le16(count + 1);
2400 rc = iam_txn_dirty(h, p, c->ic_idle_bh);
2403 mutex_unlock(&c->ic_idle_mutex);
2405 CWARN("%.16s: idle blocks failed, will lose the blk %u\n",
2406 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk);
2410 * Delete record under iterator.
2412 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2413 * it->ii_flags&IAM_IT_WRITE &&
2415 * postcondition: it_state(it) == IAM_IT_ATTACHED ||
2416 * it_state(it) == IAM_IT_DETACHED
2418 int iam_it_rec_delete(handle_t *h, struct iam_iterator *it)
2421 struct iam_leaf *leaf;
2422 struct iam_path *path;
2424 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
2425 it->ii_flags&IAM_IT_WRITE);
2426 assert_corr(it_at_rec(it));
2428 path = &it->ii_path;
2429 leaf = &path->ip_leaf;
2431 assert_inv(iam_leaf_check(leaf));
2432 assert_inv(iam_path_check(path));
2434 result = iam_txn_add(h, path, leaf->il_bh);
2436 * no compaction for now.
2439 iam_rec_del(leaf, it->ii_flags&IAM_IT_MOVE);
2440 result = iam_txn_dirty(h, path, leaf->il_bh);
2441 if (result == 0 && iam_leaf_at_end(leaf)) {
2442 struct buffer_head *bh = NULL;
2445 blk = iam_index_shrink(h, path, leaf, &bh);
2446 if (it->ii_flags & IAM_IT_MOVE) {
2447 result = iam_it_next(it);
2453 iam_recycle_leaf(h, path, bh, blk);
2458 assert_inv(iam_leaf_check(leaf));
2459 assert_inv(iam_path_check(path));
2460 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2461 it_state(it) == IAM_IT_DETACHED);
2466 * Convert iterator to cookie.
2468 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2469 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2470 * postcondition: it_state(it) == IAM_IT_ATTACHED
2472 iam_pos_t iam_it_store(const struct iam_iterator *it)
2476 assert_corr(it_state(it) == IAM_IT_ATTACHED);
2477 assert_corr(it_at_rec(it));
2478 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <=
2482 return *(iam_pos_t *)iam_it_ikey_get(it, (void *)&result);
2486 * Restore iterator from cookie.
2488 * precondition: it_state(it) == IAM_IT_DETACHED && it->ii_flags&IAM_IT_MOVE &&
2489 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2490 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED &&
2491 * iam_it_store(it) == pos)
2493 int iam_it_load(struct iam_iterator *it, iam_pos_t pos)
2495 assert_corr(it_state(it) == IAM_IT_DETACHED &&
2496 it->ii_flags&IAM_IT_MOVE);
2497 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <= sizeof pos);
2498 return iam_it_iget(it, (struct iam_ikey *)&pos);
2501 /***********************************************************************/
2503 /***********************************************************************/
2505 static inline int ptr_inside(void *base, size_t size, void *ptr)
2507 return (base <= ptr) && (ptr < base + size);
2510 static int iam_frame_invariant(struct iam_frame *f)
2514 f->bh->b_data != NULL &&
2515 ptr_inside(f->bh->b_data, f->bh->b_size, f->entries) &&
2516 ptr_inside(f->bh->b_data, f->bh->b_size, f->at) &&
2517 f->entries <= f->at);
2520 static int iam_leaf_invariant(struct iam_leaf *l)
2524 l->il_bh->b_data != NULL &&
2525 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_entries) &&
2526 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_at) &&
2527 l->il_entries <= l->il_at;
2530 static int iam_path_invariant(struct iam_path *p)
2534 if (p->ip_container == NULL ||
2535 p->ip_indirect < 0 || p->ip_indirect > DX_MAX_TREE_HEIGHT - 1 ||
2536 p->ip_frame != p->ip_frames + p->ip_indirect ||
2537 !iam_leaf_invariant(&p->ip_leaf))
2539 for (i = 0; i < ARRAY_SIZE(p->ip_frames); ++i) {
2540 if (i <= p->ip_indirect) {
2541 if (!iam_frame_invariant(&p->ip_frames[i]))
2548 int iam_it_invariant(struct iam_iterator *it)
2551 (it->ii_state == IAM_IT_DETACHED ||
2552 it->ii_state == IAM_IT_ATTACHED ||
2553 it->ii_state == IAM_IT_SKEWED) &&
2554 !(it->ii_flags & ~(IAM_IT_MOVE | IAM_IT_WRITE)) &&
2555 ergo(it->ii_state == IAM_IT_ATTACHED ||
2556 it->ii_state == IAM_IT_SKEWED,
2557 iam_path_invariant(&it->ii_path) &&
2558 equi(it_at_rec(it), it->ii_state == IAM_IT_SKEWED));
2562 * Search container @c for record with key @k. If record is found, its data
2563 * are moved into @r.
2565 * Return values: 0: found, -ENOENT: not-found, -ve: error
2567 int iam_lookup(struct iam_container *c, const struct iam_key *k,
2568 struct iam_rec *r, struct iam_path_descr *pd)
2570 struct iam_iterator it;
2573 iam_it_init(&it, c, 0, pd);
2575 result = iam_it_get_exact(&it, k);
2578 * record with required key found, copy it into user buffer
2580 iam_reccpy(&it.ii_path.ip_leaf, r);
2587 * Insert new record @r with key @k into container @c (within context of
2590 * Return values: 0: success, -ve: error, including -EEXIST when record with
2591 * given key is already present.
2593 * postcondition: ergo(result == 0 || result == -EEXIST,
2594 * iam_lookup(c, k, r2) > 0;
2596 int iam_insert(handle_t *h, struct iam_container *c, const struct iam_key *k,
2597 const struct iam_rec *r, struct iam_path_descr *pd)
2599 struct iam_iterator it;
2602 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2604 result = iam_it_get_exact(&it, k);
2605 if (result == -ENOENT)
2606 result = iam_it_rec_insert(h, &it, k, r);
2607 else if (result == 0)
2615 * Update record with the key @k in container @c (within context of
2616 * transaction @h), new record is given by @r.
2618 * Return values: +1: skip because of the same rec value, 0: success,
2619 * -ve: error, including -ENOENT if no record with the given key found.
2621 int iam_update(handle_t *h, struct iam_container *c, const struct iam_key *k,
2622 const struct iam_rec *r, struct iam_path_descr *pd)
2624 struct iam_iterator it;
2625 struct iam_leaf *folio;
2628 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2630 result = iam_it_get_exact(&it, k);
2632 folio = &it.ii_path.ip_leaf;
2633 result = iam_leaf_ops(folio)->rec_eq(folio, r);
2635 iam_it_rec_set(h, &it, r);
2645 * Delete existing record with key @k.
2647 * Return values: 0: success, -ENOENT: not-found, -ve: other error.
2649 * postcondition: ergo(result == 0 || result == -ENOENT,
2650 * !iam_lookup(c, k, *));
2652 int iam_delete(handle_t *h, struct iam_container *c, const struct iam_key *k,
2653 struct iam_path_descr *pd)
2655 struct iam_iterator it;
2658 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2660 result = iam_it_get_exact(&it, k);
2662 iam_it_rec_delete(h, &it);
2668 int iam_root_limit(int rootgap, int blocksize, int size)
2673 limit = (blocksize - rootgap) / size;
2674 nlimit = blocksize / size;
2675 if (limit == nlimit)