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>
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;
179 LASSERT(mutex_is_locked(&c->ic_idle_mutex));
184 bh = ldiskfs_bread(NULL, inode, blk, 0, &err);
186 CERROR("%.16s: cannot load idle blocks, blk = %u, err = %d\n",
187 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk, err);
188 c->ic_idle_failed = 1;
192 head = (struct iam_idle_head *)(bh->b_data);
193 if (le16_to_cpu(head->iih_magic) != IAM_IDLE_HEADER_MAGIC) {
194 CERROR("%.16s: invalid idle block head, blk = %u, magic = %d\n",
195 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk,
196 le16_to_cpu(head->iih_magic));
198 c->ic_idle_failed = 1;
199 return ERR_PTR(-EBADF);
206 * Determine format of given container. This is done by scanning list of
207 * registered formats and calling ->if_guess() method of each in turn.
209 static int iam_format_guess(struct iam_container *c)
212 struct iam_format *fmt;
215 * XXX temporary initialization hook.
218 static int initialized = 0;
221 iam_lvar_format_init();
222 iam_lfix_format_init();
228 list_for_each_entry(fmt, &iam_formats, if_linkage) {
229 result = fmt->if_guess(c);
235 struct buffer_head *bh;
238 LASSERT(c->ic_root_bh != NULL);
240 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
241 c->ic_descr->id_root_gap +
242 sizeof(struct dx_countlimit));
243 mutex_lock(&c->ic_idle_mutex);
244 bh = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
245 if (bh != NULL && IS_ERR(bh))
246 result = PTR_ERR(bh);
249 mutex_unlock(&c->ic_idle_mutex);
256 * Initialize container @c.
258 int iam_container_init(struct iam_container *c,
259 struct iam_descr *descr, struct inode *inode)
261 memset(c, 0, sizeof *c);
263 c->ic_object = inode;
264 init_rwsem(&c->ic_sem);
265 dynlock_init(&c->ic_tree_lock);
266 mutex_init(&c->ic_idle_mutex);
271 * Determine container format.
273 int iam_container_setup(struct iam_container *c)
275 return iam_format_guess(c);
279 * Finalize container @c, release all resources.
281 void iam_container_fini(struct iam_container *c)
283 brelse(c->ic_idle_bh);
284 c->ic_idle_bh = NULL;
285 brelse(c->ic_root_bh);
286 c->ic_root_bh = NULL;
289 void iam_path_init(struct iam_path *path, struct iam_container *c,
290 struct iam_path_descr *pd)
292 memset(path, 0, sizeof *path);
293 path->ip_container = c;
294 path->ip_frame = path->ip_frames;
296 path->ip_leaf.il_path = path;
299 static void iam_leaf_fini(struct iam_leaf *leaf);
301 void iam_path_release(struct iam_path *path)
305 for (i = 0; i < ARRAY_SIZE(path->ip_frames); i++) {
306 if (path->ip_frames[i].bh != NULL) {
307 path->ip_frames[i].at_shifted = 0;
308 brelse(path->ip_frames[i].bh);
309 path->ip_frames[i].bh = NULL;
314 void iam_path_fini(struct iam_path *path)
316 iam_leaf_fini(&path->ip_leaf);
317 iam_path_release(path);
321 void iam_path_compat_init(struct iam_path_compat *path, struct inode *inode)
325 path->ipc_hinfo = &path->ipc_hinfo_area;
326 for (i = 0; i < ARRAY_SIZE(path->ipc_scratch); ++i)
327 path->ipc_descr.ipd_key_scratch[i] =
328 (struct iam_ikey *)&path->ipc_scratch[i];
330 iam_path_init(&path->ipc_path, &path->ipc_container, &path->ipc_descr);
333 void iam_path_compat_fini(struct iam_path_compat *path)
335 iam_path_fini(&path->ipc_path);
339 * Helper function initializing iam_path_descr and its key scratch area.
341 struct iam_path_descr *iam_ipd_alloc(void *area, int keysize)
343 struct iam_path_descr *ipd;
349 for (i = 0; i < ARRAY_SIZE(ipd->ipd_key_scratch); ++i, karea += keysize)
350 ipd->ipd_key_scratch[i] = karea;
354 void iam_ipd_free(struct iam_path_descr *ipd)
358 int iam_node_read(struct iam_container *c, iam_ptr_t ptr,
359 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, &result);
382 * Return pointer to current leaf record. Pointer is valid while corresponding
383 * leaf node is locked and pinned.
385 static struct iam_rec *iam_leaf_rec(const struct iam_leaf *leaf)
387 return iam_leaf_ops(leaf)->rec(leaf);
391 * Return pointer to the current leaf key. This function returns pointer to
392 * the key stored in node.
394 * Caller should assume that returned pointer is only valid while leaf node is
397 static struct iam_key *iam_leaf_key(const struct iam_leaf *leaf)
399 return iam_leaf_ops(leaf)->key(leaf);
402 static int iam_leaf_key_size(const struct iam_leaf *leaf)
404 return iam_leaf_ops(leaf)->key_size(leaf);
407 static struct iam_ikey *iam_leaf_ikey(const struct iam_leaf *leaf,
408 struct iam_ikey *key)
410 return iam_leaf_ops(leaf)->ikey(leaf, key);
413 static int iam_leaf_keycmp(const struct iam_leaf *leaf,
414 const struct iam_key *key)
416 return iam_leaf_ops(leaf)->key_cmp(leaf, key);
419 static int iam_leaf_keyeq(const struct iam_leaf *leaf,
420 const struct iam_key *key)
422 return iam_leaf_ops(leaf)->key_eq(leaf, key);
425 #if LDISKFS_INVARIANT_ON
426 static int iam_leaf_check(struct iam_leaf *leaf);
427 extern int dx_node_check(struct iam_path *p, struct iam_frame *f);
429 static int iam_path_check(struct iam_path *p)
434 struct iam_descr *param;
437 param = iam_path_descr(p);
438 for (i = 0; result && i < ARRAY_SIZE(p->ip_frames); ++i) {
439 f = &p->ip_frames[i];
441 result = dx_node_check(p, f);
443 result = !param->id_ops->id_node_check(p, f);
446 if (result && p->ip_leaf.il_bh != NULL)
447 result = iam_leaf_check(&p->ip_leaf);
449 ldiskfs_std_error(iam_path_obj(p)->i_sb, result);
455 static int iam_leaf_load(struct iam_path *path)
459 struct iam_container *c;
460 struct buffer_head *bh;
461 struct iam_leaf *leaf;
462 struct iam_descr *descr;
464 c = path->ip_container;
465 leaf = &path->ip_leaf;
466 descr = iam_path_descr(path);
467 block = path->ip_frame->leaf;
470 printk(KERN_EMERG "wrong leaf: %lu %d [%p %p %p]\n",
471 (long unsigned)path->ip_frame->leaf,
472 dx_get_count(dx_node_get_entries(path, path->ip_frame)),
473 path->ip_frames[0].bh, path->ip_frames[1].bh,
474 path->ip_frames[2].bh);
476 err = descr->id_ops->id_node_read(c, block, NULL, &bh);
479 leaf->il_curidx = block;
480 err = iam_leaf_ops(leaf)->init(leaf);
481 assert_inv(ergo(err == 0, iam_leaf_check(leaf)));
486 static void iam_unlock_htree(struct iam_container *ic,
487 struct dynlock_handle *lh)
490 dynlock_unlock(&ic->ic_tree_lock, lh);
494 static void iam_leaf_unlock(struct iam_leaf *leaf)
496 if (leaf->il_lock != NULL) {
497 iam_unlock_htree(iam_leaf_container(leaf),
500 leaf->il_lock = NULL;
504 static void iam_leaf_fini(struct iam_leaf *leaf)
506 if (leaf->il_path != NULL) {
507 iam_leaf_unlock(leaf);
508 assert_inv(ergo(leaf->il_bh != NULL, iam_leaf_check(leaf)));
509 iam_leaf_ops(leaf)->fini(leaf);
518 static void iam_leaf_start(struct iam_leaf *folio)
520 iam_leaf_ops(folio)->start(folio);
523 void iam_leaf_next(struct iam_leaf *folio)
525 iam_leaf_ops(folio)->next(folio);
528 static void iam_leaf_rec_add(struct iam_leaf *leaf, const struct iam_key *key,
529 const struct iam_rec *rec)
531 iam_leaf_ops(leaf)->rec_add(leaf, key, rec);
534 static void iam_rec_del(struct iam_leaf *leaf, int shift)
536 iam_leaf_ops(leaf)->rec_del(leaf, shift);
539 int iam_leaf_at_end(const struct iam_leaf *leaf)
541 return iam_leaf_ops(leaf)->at_end(leaf);
544 static void iam_leaf_split(struct iam_leaf *l, struct buffer_head **bh,
547 iam_leaf_ops(l)->split(l, bh, nr);
550 static inline int iam_leaf_empty(struct iam_leaf *l)
552 return iam_leaf_ops(l)->leaf_empty(l);
555 int iam_leaf_can_add(const struct iam_leaf *l,
556 const struct iam_key *k, const struct iam_rec *r)
558 return iam_leaf_ops(l)->can_add(l, k, r);
561 #if LDISKFS_INVARIANT_ON
562 static int iam_leaf_check(struct iam_leaf *leaf)
566 struct iam_lentry *orig;
567 struct iam_path *path;
568 struct iam_container *bag;
575 path = iam_leaf_path(leaf);
576 bag = iam_leaf_container(leaf);
578 result = iam_leaf_ops(leaf)->init(leaf);
583 iam_leaf_start(leaf);
584 k0 = iam_path_ikey(path, 0);
585 k1 = iam_path_ikey(path, 1);
586 while (!iam_leaf_at_end(leaf)) {
587 iam_ikeycpy(bag, k0, k1);
588 iam_ikeycpy(bag, k1, iam_leaf_ikey(leaf, k1));
589 if (!first && iam_ikeycmp(bag, k0, k1) > 0) {
601 static int iam_txn_dirty(handle_t *handle,
602 struct iam_path *path, struct buffer_head *bh)
606 result = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
608 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
612 static int iam_txn_add(handle_t *handle,
613 struct iam_path *path, struct buffer_head *bh)
617 result = ldiskfs_journal_get_write_access(handle, bh);
619 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
623 /***********************************************************************/
624 /* iterator interface */
625 /***********************************************************************/
627 static enum iam_it_state it_state(const struct iam_iterator *it)
633 * Helper function returning scratch key.
635 static struct iam_container *iam_it_container(const struct iam_iterator *it)
637 return it->ii_path.ip_container;
640 static inline int it_keycmp(const struct iam_iterator *it,
641 const struct iam_key *k)
643 return iam_leaf_keycmp(&it->ii_path.ip_leaf, k);
646 static inline int it_keyeq(const struct iam_iterator *it,
647 const struct iam_key *k)
649 return iam_leaf_keyeq(&it->ii_path.ip_leaf, k);
652 static int it_ikeycmp(const struct iam_iterator *it, const struct iam_ikey *ik)
654 return iam_ikeycmp(it->ii_path.ip_container,
655 iam_leaf_ikey(&it->ii_path.ip_leaf,
656 iam_path_ikey(&it->ii_path, 0)), ik);
659 static inline int it_at_rec(const struct iam_iterator *it)
661 return !iam_leaf_at_end(&it->ii_path.ip_leaf);
664 static inline int it_before(const struct iam_iterator *it)
666 return it_state(it) == IAM_IT_SKEWED && it_at_rec(it);
670 * Helper wrapper around iam_it_get(): returns 0 (success) only when record
671 * with exactly the same key as asked is found.
673 static int iam_it_get_exact(struct iam_iterator *it, const struct iam_key *k)
677 result = iam_it_get(it, k);
680 else if (result == 0)
682 * Return -ENOENT if cursor is located above record with a key
683 * different from one specified, or in the empty leaf.
685 * XXX returning -ENOENT only works if iam_it_get() never
686 * returns -ENOENT as a legitimate error.
692 void iam_container_write_lock(struct iam_container *ic)
694 down_write(&ic->ic_sem);
697 void iam_container_write_unlock(struct iam_container *ic)
699 up_write(&ic->ic_sem);
702 void iam_container_read_lock(struct iam_container *ic)
704 down_read(&ic->ic_sem);
707 void iam_container_read_unlock(struct iam_container *ic)
709 up_read(&ic->ic_sem);
713 * Initialize iterator to IAM_IT_DETACHED state.
715 * postcondition: it_state(it) == IAM_IT_DETACHED
717 int iam_it_init(struct iam_iterator *it, struct iam_container *c, __u32 flags,
718 struct iam_path_descr *pd)
720 memset(it, 0, sizeof *it);
721 it->ii_flags = flags;
722 it->ii_state = IAM_IT_DETACHED;
723 iam_path_init(&it->ii_path, c, pd);
728 * Finalize iterator and release all resources.
730 * precondition: it_state(it) == IAM_IT_DETACHED
732 void iam_it_fini(struct iam_iterator *it)
734 assert_corr(it_state(it) == IAM_IT_DETACHED);
735 iam_path_fini(&it->ii_path);
739 * this locking primitives are used to protect parts
740 * of dir's htree. protection unit is block: leaf or index
742 static struct dynlock_handle *iam_lock_htree(struct iam_container *ic,
744 enum dynlock_type lt)
746 return dynlock_lock(&ic->ic_tree_lock, value, lt, GFP_NOFS);
749 static int iam_index_lock(struct iam_path *path, struct dynlock_handle **lh)
753 for (f = path->ip_frame; f >= path->ip_frames; --f, ++lh) {
755 *lh = iam_lock_htree(path->ip_container, f->curidx, DLT_READ);
763 * Fast check for frame consistency.
765 static int iam_check_fast(struct iam_path *path, struct iam_frame *frame)
767 struct iam_container *bag;
768 struct iam_entry *next;
769 struct iam_entry *last;
770 struct iam_entry *entries;
771 struct iam_entry *at;
773 bag = path->ip_container;
775 entries = frame->entries;
776 last = iam_entry_shift(path, entries, dx_get_count(entries) - 1);
778 if (unlikely(at > last))
781 if (unlikely(dx_get_block(path, at) != frame->leaf))
784 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, at),
785 path->ip_ikey_target) > 0))
788 next = iam_entry_shift(path, at, +1);
790 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, next),
791 path->ip_ikey_target) <= 0))
797 int dx_index_is_compat(struct iam_path *path)
799 return iam_path_descr(path) == NULL;
805 * search position of specified hash in index
809 static struct iam_entry *iam_find_position(struct iam_path *path,
810 struct iam_frame *frame)
817 count = dx_get_count(frame->entries);
818 assert_corr(count && count <= dx_get_limit(frame->entries));
819 p = iam_entry_shift(path, frame->entries,
820 dx_index_is_compat(path) ? 1 : 2);
821 q = iam_entry_shift(path, frame->entries, count - 1);
823 m = iam_entry_shift(path, p, iam_entry_diff(path, q, p) / 2);
824 if (iam_ikeycmp(path->ip_container, iam_ikey_at(path, m),
825 path->ip_ikey_target) > 0)
826 q = iam_entry_shift(path, m, -1);
828 p = iam_entry_shift(path, m, +1);
830 return iam_entry_shift(path, p, -1);
835 static iam_ptr_t iam_find_ptr(struct iam_path *path, struct iam_frame *frame)
837 return dx_get_block(path, iam_find_position(path, frame));
840 void iam_insert_key(struct iam_path *path, struct iam_frame *frame,
841 const struct iam_ikey *key, iam_ptr_t ptr)
843 struct iam_entry *entries = frame->entries;
844 struct iam_entry *new = iam_entry_shift(path, frame->at, +1);
845 int count = dx_get_count(entries);
848 * Unfortunately we cannot assert this, as this function is sometimes
849 * called by VFS under i_sem and without pdirops lock.
851 assert_corr(1 || iam_frame_is_locked(path, frame));
852 assert_corr(count < dx_get_limit(entries));
853 assert_corr(frame->at < iam_entry_shift(path, entries, count));
854 assert_inv(dx_node_check(path, frame));
856 memmove(iam_entry_shift(path, new, 1), new,
857 (char *)iam_entry_shift(path, entries, count) - (char *)new);
858 dx_set_ikey(path, new, key);
859 dx_set_block(path, new, ptr);
860 dx_set_count(entries, count + 1);
861 assert_inv(dx_node_check(path, frame));
864 void iam_insert_key_lock(struct iam_path *path, struct iam_frame *frame,
865 const struct iam_ikey *key, iam_ptr_t ptr)
867 iam_lock_bh(frame->bh);
868 iam_insert_key(path, frame, key, ptr);
869 iam_unlock_bh(frame->bh);
872 * returns 0 if path was unchanged, -EAGAIN otherwise.
874 static int iam_check_path(struct iam_path *path, struct iam_frame *frame)
878 iam_lock_bh(frame->bh);
879 equal = iam_check_fast(path, frame) == 0 ||
880 frame->leaf == iam_find_ptr(path, frame);
881 DX_DEVAL(iam_lock_stats.dls_bh_again += !equal);
882 iam_unlock_bh(frame->bh);
884 return equal ? 0 : -EAGAIN;
887 static int iam_lookup_try(struct iam_path *path)
893 struct iam_descr *param;
894 struct iam_frame *frame;
895 struct iam_container *c;
897 param = iam_path_descr(path);
898 c = path->ip_container;
900 ptr = param->id_ops->id_root_ptr(c);
901 for (frame = path->ip_frames, i = 0; i <= path->ip_indirect;
903 err = param->id_ops->id_node_read(c, (iam_ptr_t)ptr, NULL,
907 iam_lock_bh(frame->bh);
909 * node must be initialized under bh lock because concurrent
910 * creation procedure may change it and iam_lookup_try() will
911 * see obsolete tree height. -bzzz
916 if (LDISKFS_INVARIANT_ON) {
917 err = param->id_ops->id_node_check(path, frame);
922 err = param->id_ops->id_node_load(path, frame);
926 assert_inv(dx_node_check(path, frame));
928 * splitting may change root index block and move hash we're
929 * looking for into another index block so, we have to check
930 * this situation and repeat from begining if path got changed
934 err = iam_check_path(path, frame - 1);
939 frame->at = iam_find_position(path, frame);
941 frame->leaf = ptr = dx_get_block(path, frame->at);
943 iam_unlock_bh(frame->bh);
947 iam_unlock_bh(frame->bh);
948 path->ip_frame = --frame;
952 static int __iam_path_lookup(struct iam_path *path)
957 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++ i)
958 assert(path->ip_frames[i].bh == NULL);
961 err = iam_lookup_try(path);
965 } while (err == -EAGAIN);
971 * returns 0 if path was unchanged, -EAGAIN otherwise.
973 static int iam_check_full_path(struct iam_path *path, int search)
975 struct iam_frame *bottom;
976 struct iam_frame *scan;
982 for (bottom = path->ip_frames, i = 0;
983 i < DX_MAX_TREE_HEIGHT && bottom->bh != NULL; ++bottom, ++i) {
984 ; /* find last filled in frame */
988 * Lock frames, bottom to top.
990 for (scan = bottom - 1; scan >= path->ip_frames; --scan)
991 iam_lock_bh(scan->bh);
993 * Check them top to bottom.
996 for (scan = path->ip_frames; scan < bottom; ++scan) {
997 struct iam_entry *pos;
1000 if (iam_check_fast(path, scan) == 0)
1003 pos = iam_find_position(path, scan);
1004 if (scan->leaf != dx_get_block(path, pos)) {
1010 pos = iam_entry_shift(path, scan->entries,
1011 dx_get_count(scan->entries) - 1);
1012 if (scan->at > pos ||
1013 scan->leaf != dx_get_block(path, scan->at)) {
1021 * Unlock top to bottom.
1023 for (scan = path->ip_frames; scan < bottom; ++scan)
1024 iam_unlock_bh(scan->bh);
1025 DX_DEVAL(iam_lock_stats.dls_bh_full_again += !!result);
1026 do_corr(schedule());
1033 * Performs path lookup and returns with found leaf (if any) locked by htree
1036 static int iam_lookup_lock(struct iam_path *path,
1037 struct dynlock_handle **dl, enum dynlock_type lt)
1041 while ((result = __iam_path_lookup(path)) == 0) {
1042 do_corr(schedule());
1043 *dl = iam_lock_htree(path->ip_container, path->ip_frame->leaf,
1046 iam_path_fini(path);
1050 do_corr(schedule());
1052 * while locking leaf we just found may get split so we need
1053 * to check this -bzzz
1055 if (iam_check_full_path(path, 1) == 0)
1057 iam_unlock_htree(path->ip_container, *dl);
1059 iam_path_fini(path);
1064 * Performs tree top-to-bottom traversal starting from root, and loads leaf
1067 static int iam_path_lookup(struct iam_path *path, int index)
1069 struct iam_container *c;
1070 struct iam_leaf *leaf;
1073 c = path->ip_container;
1074 leaf = &path->ip_leaf;
1075 result = iam_lookup_lock(path, &leaf->il_lock, DLT_WRITE);
1076 assert_inv(iam_path_check(path));
1077 do_corr(schedule());
1079 result = iam_leaf_load(path);
1080 assert_inv(ergo(result == 0, iam_leaf_check(leaf)));
1082 do_corr(schedule());
1084 result = iam_leaf_ops(leaf)->
1085 ilookup(leaf, path->ip_ikey_target);
1087 result = iam_leaf_ops(leaf)->
1088 lookup(leaf, path->ip_key_target);
1089 do_corr(schedule());
1092 iam_leaf_unlock(leaf);
1098 * Common part of iam_it_{i,}get().
1100 static int __iam_it_get(struct iam_iterator *it, int index)
1103 assert_corr(it_state(it) == IAM_IT_DETACHED);
1105 result = iam_path_lookup(&it->ii_path, index);
1109 collision = result & IAM_LOOKUP_LAST;
1110 switch (result & ~IAM_LOOKUP_LAST) {
1111 case IAM_LOOKUP_EXACT:
1113 it->ii_state = IAM_IT_ATTACHED;
1117 it->ii_state = IAM_IT_ATTACHED;
1119 case IAM_LOOKUP_BEFORE:
1120 case IAM_LOOKUP_EMPTY:
1122 it->ii_state = IAM_IT_SKEWED;
1127 result |= collision;
1130 * See iam_it_get_exact() for explanation.
1132 assert_corr(result != -ENOENT);
1137 * Correct hash, but not the same key was found, iterate through hash
1138 * collision chain, looking for correct record.
1140 static int iam_it_collision(struct iam_iterator *it)
1144 assert(ergo(it_at_rec(it), !it_keyeq(it, it->ii_path.ip_key_target)));
1146 while ((result = iam_it_next(it)) == 0) {
1147 do_corr(schedule());
1148 if (it_ikeycmp(it, it->ii_path.ip_ikey_target) != 0)
1150 if (it_keyeq(it, it->ii_path.ip_key_target))
1157 * Attach iterator. After successful completion, @it points to record with
1158 * least key not larger than @k.
1160 * Return value: 0: positioned on existing record,
1161 * +ve: exact position found,
1164 * precondition: it_state(it) == IAM_IT_DETACHED
1165 * postcondition: ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1166 * it_keycmp(it, k) <= 0)
1168 int iam_it_get(struct iam_iterator *it, const struct iam_key *k)
1171 assert_corr(it_state(it) == IAM_IT_DETACHED);
1173 it->ii_path.ip_ikey_target = NULL;
1174 it->ii_path.ip_key_target = k;
1176 result = __iam_it_get(it, 0);
1178 if (result == IAM_LOOKUP_LAST) {
1179 result = iam_it_collision(it);
1183 result = __iam_it_get(it, 0);
1188 result &= ~IAM_LOOKUP_LAST;
1190 assert_corr(ergo(result > 0, it_keycmp(it, k) == 0));
1191 assert_corr(ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1192 it_keycmp(it, k) <= 0));
1197 * Attach iterator by index key.
1199 static int iam_it_iget(struct iam_iterator *it, const struct iam_ikey *k)
1201 assert_corr(it_state(it) == IAM_IT_DETACHED);
1203 it->ii_path.ip_ikey_target = k;
1204 return __iam_it_get(it, 1) & ~IAM_LOOKUP_LAST;
1208 * Attach iterator, and assure it points to the record (not skewed).
1210 * Return value: 0: positioned on existing record,
1211 * +ve: exact position found,
1214 * precondition: it_state(it) == IAM_IT_DETACHED &&
1215 * !(it->ii_flags&IAM_IT_WRITE)
1216 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED)
1218 int iam_it_get_at(struct iam_iterator *it, const struct iam_key *k)
1221 assert_corr(it_state(it) == IAM_IT_DETACHED &&
1222 !(it->ii_flags&IAM_IT_WRITE));
1223 result = iam_it_get(it, k);
1225 if (it_state(it) != IAM_IT_ATTACHED) {
1226 assert_corr(it_state(it) == IAM_IT_SKEWED);
1227 result = iam_it_next(it);
1230 assert_corr(ergo(result >= 0, it_state(it) == IAM_IT_ATTACHED));
1235 * Duplicates iterator.
1237 * postcondition: it_state(dst) == it_state(src) &&
1238 * iam_it_container(dst) == iam_it_container(src) &&
1239 * dst->ii_flags = src->ii_flags &&
1240 * ergo(it_state(src) == IAM_IT_ATTACHED,
1241 * iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1242 * iam_it_key_get(dst) == iam_it_key_get(src))
1244 void iam_it_dup(struct iam_iterator *dst, const struct iam_iterator *src)
1246 dst->ii_flags = src->ii_flags;
1247 dst->ii_state = src->ii_state;
1248 /* XXX not yet. iam_path_dup(&dst->ii_path, &src->ii_path); */
1250 * XXX: duplicate lock.
1252 assert_corr(it_state(dst) == it_state(src));
1253 assert_corr(iam_it_container(dst) == iam_it_container(src));
1254 assert_corr(dst->ii_flags = src->ii_flags);
1255 assert_corr(ergo(it_state(src) == IAM_IT_ATTACHED,
1256 iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1257 iam_it_key_get(dst) == iam_it_key_get(src)));
1262 * Detach iterator. Does nothing it detached state.
1264 * postcondition: it_state(it) == IAM_IT_DETACHED
1266 void iam_it_put(struct iam_iterator *it)
1268 if (it->ii_state != IAM_IT_DETACHED) {
1269 it->ii_state = IAM_IT_DETACHED;
1270 iam_leaf_fini(&it->ii_path.ip_leaf);
1274 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1275 struct iam_ikey *ikey);
1279 * This function increments the frame pointer to search the next leaf
1280 * block, and reads in the necessary intervening nodes if the search
1281 * should be necessary. Whether or not the search is necessary is
1282 * controlled by the hash parameter. If the hash value is even, then
1283 * the search is only continued if the next block starts with that
1284 * hash value. This is used if we are searching for a specific file.
1286 * If the hash value is HASH_NB_ALWAYS, then always go to the next block.
1288 * This function returns 1 if the caller should continue to search,
1289 * or 0 if it should not. If there is an error reading one of the
1290 * index blocks, it will a negative error code.
1292 * If start_hash is non-null, it will be filled in with the starting
1293 * hash of the next page.
1295 static int iam_htree_advance(struct inode *dir, __u32 hash,
1296 struct iam_path *path, __u32 *start_hash,
1299 struct iam_frame *p;
1300 struct buffer_head *bh;
1301 int err, num_frames = 0;
1306 * Find the next leaf page by incrementing the frame pointer.
1307 * If we run out of entries in the interior node, loop around and
1308 * increment pointer in the parent node. When we break out of
1309 * this loop, num_frames indicates the number of interior
1310 * nodes need to be read.
1313 do_corr(schedule());
1318 p->at = iam_entry_shift(path, p->at, +1);
1319 if (p->at < iam_entry_shift(path, p->entries,
1320 dx_get_count(p->entries))) {
1321 p->leaf = dx_get_block(path, p->at);
1322 iam_unlock_bh(p->bh);
1325 iam_unlock_bh(p->bh);
1326 if (p == path->ip_frames)
1337 * If the hash is 1, then continue only if the next page has a
1338 * continuation hash of any value. This is used for readdir
1339 * handling. Otherwise, check to see if the hash matches the
1340 * desired contiuation hash. If it doesn't, return since
1341 * there's no point to read in the successive index pages.
1343 dx_get_ikey(path, p->at, (struct iam_ikey *)&bhash);
1345 *start_hash = bhash;
1346 if ((hash & 1) == 0) {
1347 if ((bhash & ~1) != hash)
1352 * If the hash is HASH_NB_ALWAYS, we always go to the next
1353 * block so no check is necessary
1355 while (num_frames--) {
1358 do_corr(schedule());
1360 idx = p->leaf = dx_get_block(path, p->at);
1361 iam_unlock_bh(p->bh);
1362 err = iam_path_descr(path)->id_ops->
1363 id_node_read(path->ip_container, idx, NULL, &bh);
1365 return err; /* Failure */
1368 assert_corr(p->bh != bh);
1370 p->entries = dx_node_get_entries(path, p);
1371 p->at = iam_entry_shift(path, p->entries, !compat);
1372 assert_corr(p->curidx != idx);
1375 assert_corr(p->leaf != dx_get_block(path, p->at));
1376 p->leaf = dx_get_block(path, p->at);
1377 iam_unlock_bh(p->bh);
1378 assert_inv(dx_node_check(path, p));
1384 static inline int iam_index_advance(struct iam_path *path)
1386 return iam_htree_advance(iam_path_obj(path), 0, path, NULL, 0);
1389 static void iam_unlock_array(struct iam_container *ic,
1390 struct dynlock_handle **lh)
1394 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++i, ++lh) {
1396 iam_unlock_htree(ic, *lh);
1402 * Advance index part of @path to point to the next leaf. Returns 1 on
1403 * success, 0, when end of container was reached. Leaf node is locked.
1405 int iam_index_next(struct iam_container *c, struct iam_path *path)
1408 struct dynlock_handle *lh[DX_MAX_TREE_HEIGHT] = { NULL, };
1410 struct inode *object;
1413 * Locking for iam_index_next()... is to be described.
1416 object = c->ic_object;
1417 cursor = path->ip_frame->leaf;
1420 result = iam_index_lock(path, lh);
1421 do_corr(schedule());
1425 result = iam_check_full_path(path, 0);
1426 if (result == 0 && cursor == path->ip_frame->leaf) {
1427 result = iam_index_advance(path);
1429 assert_corr(result == 0 ||
1430 cursor != path->ip_frame->leaf);
1434 iam_unlock_array(c, lh);
1436 iam_path_release(path);
1437 do_corr(schedule());
1439 result = __iam_path_lookup(path);
1443 while (path->ip_frame->leaf != cursor) {
1444 do_corr(schedule());
1446 result = iam_index_lock(path, lh);
1447 do_corr(schedule());
1451 result = iam_check_full_path(path, 0);
1455 result = iam_index_advance(path);
1457 CERROR("cannot find cursor : %u\n",
1463 result = iam_check_full_path(path, 0);
1466 iam_unlock_array(c, lh);
1468 } while (result == -EAGAIN);
1472 iam_unlock_array(c, lh);
1477 * Move iterator one record right.
1479 * Return value: 0: success,
1480 * +1: end of container reached
1483 * precondition: (it_state(it) == IAM_IT_ATTACHED ||
1484 * it_state(it) == IAM_IT_SKEWED) && it->ii_flags&IAM_IT_MOVE
1485 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED) &&
1486 * ergo(result > 0, it_state(it) == IAM_IT_DETACHED)
1488 int iam_it_next(struct iam_iterator *it)
1491 struct iam_path *path;
1492 struct iam_leaf *leaf;
1493 do_corr(struct iam_ikey *ik_orig);
1495 /* assert_corr(it->ii_flags&IAM_IT_MOVE); */
1496 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1497 it_state(it) == IAM_IT_SKEWED);
1499 path = &it->ii_path;
1500 leaf = &path->ip_leaf;
1502 assert_corr(iam_leaf_is_locked(leaf));
1505 do_corr(ik_orig = it_at_rec(it) ?
1506 iam_it_ikey_get(it, iam_path_ikey(path, 2)) : NULL);
1507 if (it_before(it)) {
1508 assert_corr(!iam_leaf_at_end(leaf));
1509 it->ii_state = IAM_IT_ATTACHED;
1511 if (!iam_leaf_at_end(leaf))
1512 /* advance within leaf node */
1513 iam_leaf_next(leaf);
1515 * multiple iterations may be necessary due to empty leaves.
1517 while (result == 0 && iam_leaf_at_end(leaf)) {
1518 do_corr(schedule());
1519 /* advance index portion of the path */
1520 result = iam_index_next(iam_it_container(it), path);
1521 assert_corr(iam_leaf_is_locked(leaf));
1523 struct dynlock_handle *lh;
1524 lh = iam_lock_htree(iam_it_container(it),
1525 path->ip_frame->leaf,
1528 iam_leaf_fini(leaf);
1530 result = iam_leaf_load(path);
1532 iam_leaf_start(leaf);
1535 } else if (result == 0)
1536 /* end of container reached */
1542 it->ii_state = IAM_IT_ATTACHED;
1544 assert_corr(ergo(result == 0, it_state(it) == IAM_IT_ATTACHED));
1545 assert_corr(ergo(result > 0, it_state(it) == IAM_IT_DETACHED));
1546 assert_corr(ergo(result == 0 && ik_orig != NULL,
1547 it_ikeycmp(it, ik_orig) >= 0));
1552 * Return pointer to the record under iterator.
1554 * precondition: it_state(it) == IAM_IT_ATTACHED && it_at_rec(it)
1555 * postcondition: it_state(it) == IAM_IT_ATTACHED
1557 struct iam_rec *iam_it_rec_get(const struct iam_iterator *it)
1559 assert_corr(it_state(it) == IAM_IT_ATTACHED);
1560 assert_corr(it_at_rec(it));
1561 return iam_leaf_rec(&it->ii_path.ip_leaf);
1564 static void iam_it_reccpy(struct iam_iterator *it, const struct iam_rec *r)
1566 struct iam_leaf *folio;
1568 folio = &it->ii_path.ip_leaf;
1569 iam_leaf_ops(folio)->rec_set(folio, r);
1573 * Replace contents of record under iterator.
1575 * precondition: it_state(it) == IAM_IT_ATTACHED &&
1576 * it->ii_flags&IAM_IT_WRITE
1577 * postcondition: it_state(it) == IAM_IT_ATTACHED &&
1578 * ergo(result == 0, !memcmp(iam_it_rec_get(it), r, ...))
1580 int iam_it_rec_set(handle_t *h,
1581 struct iam_iterator *it, const struct iam_rec *r)
1584 struct iam_path *path;
1585 struct buffer_head *bh;
1587 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
1588 it->ii_flags&IAM_IT_WRITE);
1589 assert_corr(it_at_rec(it));
1591 path = &it->ii_path;
1592 bh = path->ip_leaf.il_bh;
1593 result = iam_txn_add(h, path, bh);
1595 iam_it_reccpy(it, r);
1596 result = iam_txn_dirty(h, path, bh);
1602 * Return pointer to the index key under iterator.
1604 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1605 * it_state(it) == IAM_IT_SKEWED
1607 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1608 struct iam_ikey *ikey)
1610 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1611 it_state(it) == IAM_IT_SKEWED);
1612 assert_corr(it_at_rec(it));
1613 return iam_leaf_ikey(&it->ii_path.ip_leaf, ikey);
1617 * Return pointer to the key under iterator.
1619 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1620 * it_state(it) == IAM_IT_SKEWED
1622 struct iam_key *iam_it_key_get(const struct iam_iterator *it)
1624 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1625 it_state(it) == IAM_IT_SKEWED);
1626 assert_corr(it_at_rec(it));
1627 return iam_leaf_key(&it->ii_path.ip_leaf);
1631 * Return size of key under iterator (in bytes)
1633 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1634 * it_state(it) == IAM_IT_SKEWED
1636 int iam_it_key_size(const struct iam_iterator *it)
1638 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1639 it_state(it) == IAM_IT_SKEWED);
1640 assert_corr(it_at_rec(it));
1641 return iam_leaf_key_size(&it->ii_path.ip_leaf);
1644 static struct buffer_head *
1645 iam_new_node(handle_t *h, struct iam_container *c, iam_ptr_t *b, int *e)
1647 struct inode *inode = c->ic_object;
1648 struct buffer_head *bh = NULL;
1649 struct iam_idle_head *head;
1650 struct buffer_head *idle;
1654 if (c->ic_idle_bh == NULL)
1657 mutex_lock(&c->ic_idle_mutex);
1658 if (unlikely(c->ic_idle_bh == NULL)) {
1659 mutex_unlock(&c->ic_idle_mutex);
1663 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
1664 count = le16_to_cpu(head->iih_count);
1666 *e = ldiskfs_journal_get_write_access(h, c->ic_idle_bh);
1671 *b = le32_to_cpu(head->iih_blks[count]);
1672 head->iih_count = cpu_to_le16(count);
1673 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_idle_bh);
1677 mutex_unlock(&c->ic_idle_mutex);
1678 bh = ldiskfs_bread(NULL, inode, *b, 0, e);
1684 /* The block itself which contains the iam_idle_head is
1685 * also an idle block, and can be used as the new node. */
1686 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
1687 c->ic_descr->id_root_gap +
1688 sizeof(struct dx_countlimit));
1689 *e = ldiskfs_journal_get_write_access(h, c->ic_root_bh);
1693 *b = le32_to_cpu(*idle_blocks);
1694 iam_lock_bh(c->ic_root_bh);
1695 *idle_blocks = head->iih_next;
1696 iam_unlock_bh(c->ic_root_bh);
1697 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_root_bh);
1699 iam_lock_bh(c->ic_root_bh);
1700 *idle_blocks = cpu_to_le32(*b);
1701 iam_unlock_bh(c->ic_root_bh);
1706 idle = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
1707 if (idle != NULL && IS_ERR(idle)) {
1709 c->ic_idle_bh = NULL;
1714 c->ic_idle_bh = idle;
1715 mutex_unlock(&c->ic_idle_mutex);
1718 /* get write access for the found buffer head */
1719 *e = ldiskfs_journal_get_write_access(h, bh);
1723 ldiskfs_std_error(inode->i_sb, *e);
1725 /* Clear the reused node as new node does. */
1726 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1727 set_buffer_uptodate(bh);
1732 bh = osd_ldiskfs_append(h, inode, b, e);
1736 mutex_unlock(&c->ic_idle_mutex);
1737 ldiskfs_std_error(inode->i_sb, *e);
1742 * Insertion of new record. Interaction with jbd during non-trivial case (when
1743 * split happens) is as following:
1745 * - new leaf node is involved into transaction by iam_new_node();
1747 * - old leaf node is involved into transaction by iam_add_rec();
1749 * - leaf where insertion point ends in, is marked dirty by iam_add_rec();
1751 * - leaf without insertion point is marked dirty (as @new_leaf) by
1754 * - split index nodes are involved into transaction and marked dirty by
1755 * split_index_node().
1757 * - "safe" index node, which is no split, but where new pointer is inserted
1758 * is involved into transaction and marked dirty by split_index_node().
1760 * - index node where pointer to new leaf is inserted is involved into
1761 * transaction by split_index_node() and marked dirty by iam_add_rec().
1763 * - inode is marked dirty by iam_add_rec().
1767 static int iam_new_leaf(handle_t *handle, struct iam_leaf *leaf)
1771 struct buffer_head *new_leaf;
1772 struct buffer_head *old_leaf;
1773 struct iam_container *c;
1775 struct iam_path *path;
1777 assert_inv(iam_leaf_check(leaf));
1779 c = iam_leaf_container(leaf);
1780 path = leaf->il_path;
1783 new_leaf = iam_new_node(handle, c, &blknr, &err);
1784 do_corr(schedule());
1785 if (new_leaf != NULL) {
1786 struct dynlock_handle *lh;
1788 lh = iam_lock_htree(c, blknr, DLT_WRITE);
1789 do_corr(schedule());
1791 iam_leaf_ops(leaf)->init_new(c, new_leaf);
1792 do_corr(schedule());
1793 old_leaf = leaf->il_bh;
1794 iam_leaf_split(leaf, &new_leaf, blknr);
1795 if (old_leaf != leaf->il_bh) {
1797 * Switched to the new leaf.
1799 iam_leaf_unlock(leaf);
1801 path->ip_frame->leaf = blknr;
1803 iam_unlock_htree(path->ip_container, lh);
1804 do_corr(schedule());
1805 err = iam_txn_dirty(handle, path, new_leaf);
1808 err = ldiskfs_mark_inode_dirty(handle, obj);
1809 do_corr(schedule());
1813 assert_inv(iam_leaf_check(leaf));
1814 assert_inv(iam_leaf_check(&iam_leaf_path(leaf)->ip_leaf));
1815 assert_inv(iam_path_check(iam_leaf_path(leaf)));
1819 static inline void dx_set_limit(struct iam_entry *entries, unsigned value)
1821 ((struct dx_countlimit *) entries)->limit = cpu_to_le16(value);
1824 static int iam_shift_entries(struct iam_path *path,
1825 struct iam_frame *frame, unsigned count,
1826 struct iam_entry *entries, struct iam_entry *entries2,
1833 struct iam_frame *parent = frame - 1;
1834 struct iam_ikey *pivot = iam_path_ikey(path, 3);
1836 delta = dx_index_is_compat(path) ? 0 : +1;
1838 count1 = count/2 + delta;
1839 count2 = count - count1;
1840 dx_get_ikey(path, iam_entry_shift(path, entries, count1), pivot);
1842 dxtrace(printk("Split index %d/%d\n", count1, count2));
1844 memcpy((char *) iam_entry_shift(path, entries2, delta),
1845 (char *) iam_entry_shift(path, entries, count1),
1846 count2 * iam_entry_size(path));
1848 dx_set_count(entries2, count2 + delta);
1849 dx_set_limit(entries2, dx_node_limit(path));
1852 * NOTE: very subtle piece of code competing dx_probe() may find 2nd
1853 * level index in root index, then we insert new index here and set
1854 * new count in that 2nd level index. so, dx_probe() may see 2nd level
1855 * index w/o hash it looks for. the solution is to check root index
1856 * after we locked just founded 2nd level index -bzzz
1858 iam_insert_key_lock(path, parent, pivot, newblock);
1861 * now old and new 2nd level index blocks contain all pointers, so
1862 * dx_probe() may find it in the both. it's OK -bzzz
1864 iam_lock_bh(frame->bh);
1865 dx_set_count(entries, count1);
1866 iam_unlock_bh(frame->bh);
1869 * now old 2nd level index block points to first half of leafs. it's
1870 * importand that dx_probe() must check root index block for changes
1871 * under dx_lock_bh(frame->bh) -bzzz
1878 int split_index_node(handle_t *handle, struct iam_path *path,
1879 struct dynlock_handle **lh)
1882 struct iam_entry *entries; /* old block contents */
1883 struct iam_entry *entries2; /* new block contents */
1884 struct iam_frame *frame, *safe;
1885 struct buffer_head *bh_new[DX_MAX_TREE_HEIGHT] = {NULL};
1886 u32 newblock[DX_MAX_TREE_HEIGHT] = {0};
1887 struct dynlock_handle *lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1888 struct dynlock_handle *new_lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1889 struct inode *dir = iam_path_obj(path);
1890 struct iam_descr *descr;
1894 descr = iam_path_descr(path);
1896 * Algorithm below depends on this.
1898 assert_corr(dx_root_limit(path) < dx_node_limit(path));
1900 frame = path->ip_frame;
1901 entries = frame->entries;
1904 * Tall-tree handling: we might have to split multiple index blocks
1905 * all the way up to tree root. Tricky point here is error handling:
1906 * to avoid complicated undo/rollback we
1908 * - first allocate all necessary blocks
1910 * - insert pointers into them atomically.
1914 * Locking: leaf is already locked. htree-locks are acquired on all
1915 * index nodes that require split bottom-to-top, on the "safe" node,
1916 * and on all new nodes
1919 dxtrace(printk("using %u of %u node entries\n",
1920 dx_get_count(entries), dx_get_limit(entries)));
1922 /* What levels need split? */
1923 for (nr_splet = 0; frame >= path->ip_frames &&
1924 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1925 --frame, ++nr_splet) {
1926 do_corr(schedule());
1927 if (nr_splet == DX_MAX_TREE_HEIGHT) {
1929 CWARN(dir->i_sb, __FUNCTION__,
1930 "Directory index full!\n");
1940 * Lock all nodes, bottom to top.
1942 for (frame = path->ip_frame, i = nr_splet; i >= 0; --i, --frame) {
1943 do_corr(schedule());
1944 lock[i] = iam_lock_htree(path->ip_container, frame->curidx,
1946 if (lock[i] == NULL) {
1953 * Check for concurrent index modification.
1955 err = iam_check_full_path(path, 1);
1959 * And check that the same number of nodes is to be split.
1961 for (i = 0, frame = path->ip_frame; frame >= path->ip_frames &&
1962 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1966 if (i != nr_splet) {
1971 /* Go back down, allocating blocks, locking them, and adding into
1973 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
1974 bh_new[i] = iam_new_node(handle, path->ip_container,
1975 &newblock[i], &err);
1976 do_corr(schedule());
1978 descr->id_ops->id_node_init(path->ip_container,
1981 new_lock[i] = iam_lock_htree(path->ip_container, newblock[i],
1983 if (new_lock[i] == NULL) {
1987 do_corr(schedule());
1988 BUFFER_TRACE(frame->bh, "get_write_access");
1989 err = ldiskfs_journal_get_write_access(handle, frame->bh);
1993 /* Add "safe" node to transaction too */
1994 if (safe + 1 != path->ip_frames) {
1995 do_corr(schedule());
1996 err = ldiskfs_journal_get_write_access(handle, safe->bh);
2001 /* Go through nodes once more, inserting pointers */
2002 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
2005 struct buffer_head *bh2;
2006 struct buffer_head *bh;
2008 entries = frame->entries;
2009 count = dx_get_count(entries);
2010 idx = iam_entry_diff(path, frame->at, entries);
2013 entries2 = dx_get_entries(path, bh2->b_data, 0);
2016 if (frame == path->ip_frames) {
2017 /* splitting root node. Tricky point:
2019 * In the "normal" B-tree we'd split root *and* add
2020 * new root to the tree with pointers to the old root
2021 * and its sibling (thus introducing two new nodes).
2023 * In htree it's enough to add one node, because
2024 * capacity of the root node is smaller than that of
2027 struct iam_frame *frames;
2028 struct iam_entry *next;
2030 assert_corr(i == 0);
2032 do_corr(schedule());
2034 frames = path->ip_frames;
2035 memcpy((char *) entries2, (char *) entries,
2036 count * iam_entry_size(path));
2037 dx_set_limit(entries2, dx_node_limit(path));
2040 iam_lock_bh(frame->bh);
2041 next = descr->id_ops->id_root_inc(path->ip_container,
2043 dx_set_block(path, next, newblock[0]);
2044 iam_unlock_bh(frame->bh);
2046 do_corr(schedule());
2047 /* Shift frames in the path */
2048 memmove(frames + 2, frames + 1,
2049 (sizeof path->ip_frames) - 2 * sizeof frames[0]);
2050 /* Add new access path frame */
2051 frames[1].at = iam_entry_shift(path, entries2, idx);
2052 frames[1].entries = entries = entries2;
2054 assert_inv(dx_node_check(path, frame));
2057 assert_inv(dx_node_check(path, frame));
2058 bh_new[0] = NULL; /* buffer head is "consumed" */
2059 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2062 do_corr(schedule());
2064 /* splitting non-root index node. */
2065 struct iam_frame *parent = frame - 1;
2067 do_corr(schedule());
2068 count = iam_shift_entries(path, frame, count,
2069 entries, entries2, newblock[i]);
2070 /* Which index block gets the new entry? */
2072 int d = dx_index_is_compat(path) ? 0 : +1;
2074 frame->at = iam_entry_shift(path, entries2,
2076 frame->entries = entries = entries2;
2077 frame->curidx = newblock[i];
2078 swap(frame->bh, bh2);
2079 assert_corr(lock[i + 1] != NULL);
2080 assert_corr(new_lock[i] != NULL);
2081 swap(lock[i + 1], new_lock[i]);
2083 parent->at = iam_entry_shift(path,
2086 assert_inv(dx_node_check(path, frame));
2087 assert_inv(dx_node_check(path, parent));
2088 dxtrace(dx_show_index ("node", frame->entries));
2089 dxtrace(dx_show_index ("node",
2090 ((struct dx_node *) bh2->b_data)->entries));
2091 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2094 do_corr(schedule());
2095 err = ldiskfs_handle_dirty_metadata(handle, NULL,
2100 do_corr(schedule());
2101 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2106 * This function was called to make insertion of new leaf
2107 * possible. Check that it fulfilled its obligations.
2109 assert_corr(dx_get_count(path->ip_frame->entries) <
2110 dx_get_limit(path->ip_frame->entries));
2111 assert_corr(lock[nr_splet] != NULL);
2112 *lh = lock[nr_splet];
2113 lock[nr_splet] = NULL;
2116 * Log ->i_size modification.
2118 err = ldiskfs_mark_inode_dirty(handle, dir);
2124 ldiskfs_std_error(dir->i_sb, err);
2127 iam_unlock_array(path->ip_container, lock);
2128 iam_unlock_array(path->ip_container, new_lock);
2130 assert_corr(err || iam_frame_is_locked(path, path->ip_frame));
2132 do_corr(schedule());
2133 for (i = 0; i < ARRAY_SIZE(bh_new); ++i) {
2134 if (bh_new[i] != NULL)
2140 static int iam_add_rec(handle_t *handle, struct iam_iterator *it,
2141 struct iam_path *path,
2142 const struct iam_key *k, const struct iam_rec *r)
2145 struct iam_leaf *leaf;
2147 leaf = &path->ip_leaf;
2148 assert_inv(iam_leaf_check(leaf));
2149 assert_inv(iam_path_check(path));
2150 err = iam_txn_add(handle, path, leaf->il_bh);
2152 do_corr(schedule());
2153 if (!iam_leaf_can_add(leaf, k, r)) {
2154 struct dynlock_handle *lh = NULL;
2157 assert_corr(lh == NULL);
2158 do_corr(schedule());
2159 err = split_index_node(handle, path, &lh);
2160 if (err == -EAGAIN) {
2161 assert_corr(lh == NULL);
2163 iam_path_fini(path);
2164 it->ii_state = IAM_IT_DETACHED;
2166 do_corr(schedule());
2167 err = iam_it_get_exact(it, k);
2169 err = +1; /* repeat split */
2174 assert_inv(iam_path_check(path));
2176 assert_corr(lh != NULL);
2177 do_corr(schedule());
2178 err = iam_new_leaf(handle, leaf);
2180 err = iam_txn_dirty(handle, path,
2181 path->ip_frame->bh);
2183 iam_unlock_htree(path->ip_container, lh);
2184 do_corr(schedule());
2187 iam_leaf_rec_add(leaf, k, r);
2188 err = iam_txn_dirty(handle, path, leaf->il_bh);
2191 assert_inv(iam_leaf_check(leaf));
2192 assert_inv(iam_leaf_check(&path->ip_leaf));
2193 assert_inv(iam_path_check(path));
2198 * Insert new record with key @k and contents from @r, shifting records to the
2199 * right. On success, iterator is positioned on the newly inserted record.
2201 * precondition: it->ii_flags&IAM_IT_WRITE &&
2202 * (it_state(it) == IAM_IT_ATTACHED ||
2203 * it_state(it) == IAM_IT_SKEWED) &&
2204 * ergo(it_state(it) == IAM_IT_ATTACHED,
2205 * it_keycmp(it, k) <= 0) &&
2206 * ergo(it_before(it), it_keycmp(it, k) > 0));
2207 * postcondition: ergo(result == 0,
2208 * it_state(it) == IAM_IT_ATTACHED &&
2209 * it_keycmp(it, k) == 0 &&
2210 * !memcmp(iam_it_rec_get(it), r, ...))
2212 int iam_it_rec_insert(handle_t *h, struct iam_iterator *it,
2213 const struct iam_key *k, const struct iam_rec *r)
2216 struct iam_path *path;
2218 path = &it->ii_path;
2220 assert_corr(it->ii_flags&IAM_IT_WRITE);
2221 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2222 it_state(it) == IAM_IT_SKEWED);
2223 assert_corr(ergo(it_state(it) == IAM_IT_ATTACHED,
2224 it_keycmp(it, k) <= 0));
2225 assert_corr(ergo(it_before(it), it_keycmp(it, k) > 0));
2226 result = iam_add_rec(h, it, path, k, r);
2228 it->ii_state = IAM_IT_ATTACHED;
2229 assert_corr(ergo(result == 0,
2230 it_state(it) == IAM_IT_ATTACHED &&
2231 it_keycmp(it, k) == 0));
2235 static inline int iam_idle_blocks_limit(struct inode *inode)
2237 return (inode->i_sb->s_blocksize - sizeof(struct iam_idle_head)) >> 2;
2241 * If the leaf cannnot be recycled, we will lose one block for reusing.
2242 * It is not a serious issue because it almost the same of non-recycle.
2244 static iam_ptr_t iam_index_shrink(handle_t *h, struct iam_path *p,
2245 struct iam_leaf *l, struct buffer_head **bh)
2247 struct iam_container *c = p->ip_container;
2248 struct inode *inode = c->ic_object;
2249 struct iam_frame *frame = p->ip_frame;
2250 struct iam_entry *entries;
2251 struct iam_entry *pos;
2252 struct dynlock_handle *lh;
2256 if (c->ic_idle_failed)
2259 if (unlikely(frame == NULL))
2262 if (!iam_leaf_empty(l))
2265 lh = iam_lock_htree(c, frame->curidx, DLT_WRITE);
2267 CWARN("%.16s: No memory to recycle idle blocks\n",
2268 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name);
2272 rc = iam_txn_add(h, p, frame->bh);
2274 iam_unlock_htree(c, lh);
2278 iam_lock_bh(frame->bh);
2279 entries = frame->entries;
2280 count = dx_get_count(entries);
2281 /* NOT shrink the last entry in the index node, which can be reused
2282 * directly by next new node. */
2284 iam_unlock_bh(frame->bh);
2285 iam_unlock_htree(c, lh);
2289 pos = iam_find_position(p, frame);
2290 /* There may be some new leaf nodes have been added or empty leaf nodes
2291 * have been shrinked during my delete operation.
2293 * If the empty leaf is not under current index node because the index
2294 * node has been split, then just skip the empty leaf, which is rare. */
2295 if (unlikely(frame->leaf != dx_get_block(p, pos))) {
2296 iam_unlock_bh(frame->bh);
2297 iam_unlock_htree(c, lh);
2302 if (frame->at < iam_entry_shift(p, entries, count - 1)) {
2303 struct iam_entry *n = iam_entry_shift(p, frame->at, 1);
2305 memmove(frame->at, n,
2306 (char *)iam_entry_shift(p, entries, count) - (char *)n);
2307 frame->at_shifted = 1;
2309 dx_set_count(entries, count - 1);
2310 iam_unlock_bh(frame->bh);
2311 rc = iam_txn_dirty(h, p, frame->bh);
2312 iam_unlock_htree(c, lh);
2322 iam_install_idle_blocks(handle_t *h, struct iam_path *p, struct buffer_head *bh,
2323 __u32 *idle_blocks, iam_ptr_t blk)
2325 struct iam_container *c = p->ip_container;
2326 struct buffer_head *old = c->ic_idle_bh;
2327 struct iam_idle_head *head;
2330 head = (struct iam_idle_head *)(bh->b_data);
2331 head->iih_magic = cpu_to_le16(IAM_IDLE_HEADER_MAGIC);
2332 head->iih_count = 0;
2333 head->iih_next = *idle_blocks;
2334 /* The bh already get_write_accessed. */
2335 rc = iam_txn_dirty(h, p, bh);
2339 rc = iam_txn_add(h, p, c->ic_root_bh);
2343 iam_lock_bh(c->ic_root_bh);
2344 *idle_blocks = cpu_to_le32(blk);
2345 iam_unlock_bh(c->ic_root_bh);
2346 rc = iam_txn_dirty(h, p, c->ic_root_bh);
2348 /* NOT release old before new assigned. */
2353 iam_lock_bh(c->ic_root_bh);
2354 *idle_blocks = head->iih_next;
2355 iam_unlock_bh(c->ic_root_bh);
2361 * If the leaf cannnot be recycled, we will lose one block for reusing.
2362 * It is not a serious issue because it almost the same of non-recycle.
2364 static void iam_recycle_leaf(handle_t *h, struct iam_path *p,
2365 struct buffer_head *bh, iam_ptr_t blk)
2367 struct iam_container *c = p->ip_container;
2368 struct inode *inode = c->ic_object;
2369 struct iam_idle_head *head;
2374 mutex_lock(&c->ic_idle_mutex);
2375 if (unlikely(c->ic_idle_failed)) {
2380 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
2381 c->ic_descr->id_root_gap +
2382 sizeof(struct dx_countlimit));
2383 /* It is the first idle block. */
2384 if (c->ic_idle_bh == NULL) {
2385 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2389 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
2390 count = le16_to_cpu(head->iih_count);
2391 /* Current ic_idle_bh is full, to be replaced by the leaf. */
2392 if (count == iam_idle_blocks_limit(inode)) {
2393 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2397 /* Just add to ic_idle_bh. */
2398 rc = iam_txn_add(h, p, c->ic_idle_bh);
2402 head->iih_blks[count] = cpu_to_le32(blk);
2403 head->iih_count = cpu_to_le16(count + 1);
2404 rc = iam_txn_dirty(h, p, c->ic_idle_bh);
2407 mutex_unlock(&c->ic_idle_mutex);
2409 CWARN("%.16s: idle blocks failed, will lose the blk %u\n",
2410 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk);
2414 * Delete record under iterator.
2416 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2417 * it->ii_flags&IAM_IT_WRITE &&
2419 * postcondition: it_state(it) == IAM_IT_ATTACHED ||
2420 * it_state(it) == IAM_IT_DETACHED
2422 int iam_it_rec_delete(handle_t *h, struct iam_iterator *it)
2425 struct iam_leaf *leaf;
2426 struct iam_path *path;
2428 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
2429 it->ii_flags&IAM_IT_WRITE);
2430 assert_corr(it_at_rec(it));
2432 path = &it->ii_path;
2433 leaf = &path->ip_leaf;
2435 assert_inv(iam_leaf_check(leaf));
2436 assert_inv(iam_path_check(path));
2438 result = iam_txn_add(h, path, leaf->il_bh);
2440 * no compaction for now.
2443 iam_rec_del(leaf, it->ii_flags&IAM_IT_MOVE);
2444 result = iam_txn_dirty(h, path, leaf->il_bh);
2445 if (result == 0 && iam_leaf_at_end(leaf)) {
2446 struct buffer_head *bh = NULL;
2449 blk = iam_index_shrink(h, path, leaf, &bh);
2450 if (it->ii_flags & IAM_IT_MOVE) {
2451 result = iam_it_next(it);
2457 iam_recycle_leaf(h, path, bh, blk);
2462 assert_inv(iam_leaf_check(leaf));
2463 assert_inv(iam_path_check(path));
2464 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2465 it_state(it) == IAM_IT_DETACHED);
2470 * Convert iterator to cookie.
2472 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2473 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2474 * postcondition: it_state(it) == IAM_IT_ATTACHED
2476 iam_pos_t iam_it_store(const struct iam_iterator *it)
2480 assert_corr(it_state(it) == IAM_IT_ATTACHED);
2481 assert_corr(it_at_rec(it));
2482 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <=
2486 return *(iam_pos_t *)iam_it_ikey_get(it, (void *)&result);
2490 * Restore iterator from cookie.
2492 * precondition: it_state(it) == IAM_IT_DETACHED && it->ii_flags&IAM_IT_MOVE &&
2493 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2494 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED &&
2495 * iam_it_store(it) == pos)
2497 int iam_it_load(struct iam_iterator *it, iam_pos_t pos)
2499 assert_corr(it_state(it) == IAM_IT_DETACHED &&
2500 it->ii_flags&IAM_IT_MOVE);
2501 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <= sizeof pos);
2502 return iam_it_iget(it, (struct iam_ikey *)&pos);
2505 /***********************************************************************/
2507 /***********************************************************************/
2509 static inline int ptr_inside(void *base, size_t size, void *ptr)
2511 return (base <= ptr) && (ptr < base + size);
2514 static int iam_frame_invariant(struct iam_frame *f)
2518 f->bh->b_data != NULL &&
2519 ptr_inside(f->bh->b_data, f->bh->b_size, f->entries) &&
2520 ptr_inside(f->bh->b_data, f->bh->b_size, f->at) &&
2521 f->entries <= f->at);
2524 static int iam_leaf_invariant(struct iam_leaf *l)
2528 l->il_bh->b_data != NULL &&
2529 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_entries) &&
2530 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_at) &&
2531 l->il_entries <= l->il_at;
2534 static int iam_path_invariant(struct iam_path *p)
2538 if (p->ip_container == NULL ||
2539 p->ip_indirect < 0 || p->ip_indirect > DX_MAX_TREE_HEIGHT - 1 ||
2540 p->ip_frame != p->ip_frames + p->ip_indirect ||
2541 !iam_leaf_invariant(&p->ip_leaf))
2543 for (i = 0; i < ARRAY_SIZE(p->ip_frames); ++i) {
2544 if (i <= p->ip_indirect) {
2545 if (!iam_frame_invariant(&p->ip_frames[i]))
2552 int iam_it_invariant(struct iam_iterator *it)
2555 (it->ii_state == IAM_IT_DETACHED ||
2556 it->ii_state == IAM_IT_ATTACHED ||
2557 it->ii_state == IAM_IT_SKEWED) &&
2558 !(it->ii_flags & ~(IAM_IT_MOVE | IAM_IT_WRITE)) &&
2559 ergo(it->ii_state == IAM_IT_ATTACHED ||
2560 it->ii_state == IAM_IT_SKEWED,
2561 iam_path_invariant(&it->ii_path) &&
2562 equi(it_at_rec(it), it->ii_state == IAM_IT_SKEWED));
2566 * Search container @c for record with key @k. If record is found, its data
2567 * are moved into @r.
2569 * Return values: 0: found, -ENOENT: not-found, -ve: error
2571 int iam_lookup(struct iam_container *c, const struct iam_key *k,
2572 struct iam_rec *r, struct iam_path_descr *pd)
2574 struct iam_iterator it;
2577 iam_it_init(&it, c, 0, pd);
2579 result = iam_it_get_exact(&it, k);
2582 * record with required key found, copy it into user buffer
2584 iam_reccpy(&it.ii_path.ip_leaf, r);
2591 * Insert new record @r with key @k into container @c (within context of
2594 * Return values: 0: success, -ve: error, including -EEXIST when record with
2595 * given key is already present.
2597 * postcondition: ergo(result == 0 || result == -EEXIST,
2598 * iam_lookup(c, k, r2) > 0;
2600 int iam_insert(handle_t *h, struct iam_container *c, const struct iam_key *k,
2601 const struct iam_rec *r, struct iam_path_descr *pd)
2603 struct iam_iterator it;
2606 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2608 result = iam_it_get_exact(&it, k);
2609 if (result == -ENOENT)
2610 result = iam_it_rec_insert(h, &it, k, r);
2611 else if (result == 0)
2619 * Update record with the key @k in container @c (within context of
2620 * transaction @h), new record is given by @r.
2622 * Return values: +1: skip because of the same rec value, 0: success,
2623 * -ve: error, including -ENOENT if no record with the given key found.
2625 int iam_update(handle_t *h, struct iam_container *c, const struct iam_key *k,
2626 const struct iam_rec *r, struct iam_path_descr *pd)
2628 struct iam_iterator it;
2629 struct iam_leaf *folio;
2632 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2634 result = iam_it_get_exact(&it, k);
2636 folio = &it.ii_path.ip_leaf;
2637 result = iam_leaf_ops(folio)->rec_eq(folio, r);
2639 iam_it_rec_set(h, &it, r);
2649 * Delete existing record with key @k.
2651 * Return values: 0: success, -ENOENT: not-found, -ve: other error.
2653 * postcondition: ergo(result == 0 || result == -ENOENT,
2654 * !iam_lookup(c, k, *));
2656 int iam_delete(handle_t *h, struct iam_container *c, const struct iam_key *k,
2657 struct iam_path_descr *pd)
2659 struct iam_iterator it;
2662 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2664 result = iam_it_get_exact(&it, k);
2666 iam_it_rec_delete(h, &it);
2672 int iam_root_limit(int rootgap, int blocksize, int size)
2677 limit = (blocksize - rootgap) / size;
2678 nlimit = blocksize / size;
2679 if (limit == nlimit)