4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see [sun.com URL with a
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 * Top-level entry points into iam module
39 * Author: Wang Di <wangdi@clusterfs.com>
40 * Author: Nikita Danilov <nikita@clusterfs.com>
44 * iam: big theory statement.
46 * iam (Index Access Module) is a module providing abstraction of persistent
47 * transactional container on top of generalized ldiskfs htree.
51 * - key, pointer, and record size specifiable per container.
53 * - trees taller than 2 index levels.
55 * - read/write to existing ldiskfs htree directories as iam containers.
57 * iam container is a tree, consisting of leaf nodes containing keys and
58 * records stored in this container, and index nodes, containing keys and
59 * pointers to leaf or index nodes.
61 * iam does not work with keys directly, instead it calls user-supplied key
62 * comparison function (->dpo_keycmp()).
64 * Pointers are (currently) interpreted as logical offsets (measured in
65 * blocksful) within underlying flat file on top of which iam tree lives.
69 * iam mostly tries to reuse existing htree formats.
71 * Format of index node:
73 * +-----+-------+-------+-------+------+-------+------------+
74 * | | count | | | | | |
75 * | gap | / | entry | entry | .... | entry | free space |
76 * | | limit | | | | | |
77 * +-----+-------+-------+-------+------+-------+------------+
79 * gap this part of node is never accessed by iam code. It
80 * exists for binary compatibility with ldiskfs htree (that,
81 * in turn, stores fake struct ext2_dirent for ext2
82 * compatibility), and to keep some unspecified per-node
83 * data. Gap can be different for root and non-root index
84 * nodes. Gap size can be specified for each container
85 * (gap of 0 is allowed).
87 * count/limit current number of entries in this node, and the maximal
88 * number of entries that can fit into node. count/limit
89 * has the same size as entry, and is itself counted in
92 * entry index entry: consists of a key immediately followed by
93 * a pointer to a child node. Size of a key and size of a
94 * pointer depends on container. Entry has neither
95 * alignment nor padding.
97 * free space portion of node new entries are added to
99 * Entries in index node are sorted by their key value.
101 * Format of a leaf node is not specified. Generic iam code accesses leaf
102 * nodes through ->id_leaf methods in struct iam_descr.
104 * The IAM root block is a special node, which contains the IAM descriptor.
105 * It is on disk format:
107 * +---------+-------+--------+---------+-------+------+-------+------------+
108 * |IAM desc | count | idle | | | | | |
109 * |(fix/var)| / | blocks | padding | entry | .... | entry | free space |
110 * | | limit | | | | | | |
111 * +---------+-------+--------+---------+-------+------+-------+------------+
113 * The padding length is calculated with the parameters in the IAM descriptor.
115 * The field "idle_blocks" is used to record empty leaf nodes, which have not
116 * been released but all contained entries in them have been removed. Usually,
117 * the idle blocks in the IAM should be reused when need to allocate new leaf
118 * nodes for new entries, it depends on the IAM hash functions to map the new
119 * entries to these idle blocks. Unfortunately, it is not easy to design some
120 * hash functions for such clever mapping, especially considering the insert/
121 * lookup performance.
123 * So the IAM recycles the empty leaf nodes, and put them into a per-file based
124 * idle blocks pool. If need some new leaf node, it will try to take idle block
125 * from such pool with priority, in spite of how the IAM hash functions to map
128 * The idle blocks pool is organized as a series of tables, and each table
129 * can be described as following (on-disk format):
131 * +---------+---------+---------+---------+------+---------+-------+
132 * | magic | count | next | logic | | logic | free |
133 * |(16 bits)|(16 bits)| table | blk # | .... | blk # | space |
134 * | | |(32 bits)|(32 bits)| |(32 bits)| |
135 * +---------+---------+---------+---------+------+---------+-------+
137 * The logic blk# for the first table is stored in the root node "idle_blocks".
141 #include <linux/module.h>
142 #include <linux/fs.h>
143 #include <linux/pagemap.h>
144 #include <linux/time.h>
145 #include <linux/fcntl.h>
146 #include <linux/stat.h>
147 #include <linux/string.h>
148 #include <linux/quotaops.h>
149 #include <linux/buffer_head.h>
151 #include <ldiskfs/ldiskfs.h>
152 #include <ldiskfs/xattr.h>
155 #include "osd_internal.h"
157 #include <ldiskfs/acl.h>
160 * List of all registered formats.
162 * No locking. Callers synchronize.
164 static struct list_head iam_formats = LIST_HEAD_INIT(iam_formats);
166 void iam_format_register(struct iam_format *fmt)
168 list_add(&fmt->if_linkage, &iam_formats);
171 static struct buffer_head *
172 iam_load_idle_blocks(struct iam_container *c, iam_ptr_t blk)
174 struct inode *inode = c->ic_object;
175 struct iam_idle_head *head;
176 struct buffer_head *bh;
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;
189 err = err ? err : -EIO;
193 head = (struct iam_idle_head *)(bh->b_data);
194 if (le16_to_cpu(head->iih_magic) != IAM_IDLE_HEADER_MAGIC) {
195 CERROR("%.16s: invalid idle block head, blk = %u, magic = %d\n",
196 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk,
197 le16_to_cpu(head->iih_magic));
199 c->ic_idle_failed = 1;
200 return ERR_PTR(-EBADF);
207 * Determine format of given container. This is done by scanning list of
208 * registered formats and calling ->if_guess() method of each in turn.
210 static int iam_format_guess(struct iam_container *c)
213 struct iam_format *fmt;
216 * XXX temporary initialization hook.
219 static int initialized = 0;
222 iam_lvar_format_init();
223 iam_lfix_format_init();
229 list_for_each_entry(fmt, &iam_formats, if_linkage) {
230 result = fmt->if_guess(c);
236 struct buffer_head *bh;
239 LASSERT(c->ic_root_bh != NULL);
241 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
242 c->ic_descr->id_root_gap +
243 sizeof(struct dx_countlimit));
244 mutex_lock(&c->ic_idle_mutex);
245 bh = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
246 if (bh != NULL && IS_ERR(bh))
247 result = PTR_ERR(bh);
250 mutex_unlock(&c->ic_idle_mutex);
257 * Initialize container @c.
259 int iam_container_init(struct iam_container *c,
260 struct iam_descr *descr, struct inode *inode)
262 memset(c, 0, sizeof *c);
264 c->ic_object = inode;
265 init_rwsem(&c->ic_sem);
266 dynlock_init(&c->ic_tree_lock);
267 mutex_init(&c->ic_idle_mutex);
272 * Determine container format.
274 int iam_container_setup(struct iam_container *c)
276 return iam_format_guess(c);
280 * Finalize container @c, release all resources.
282 void iam_container_fini(struct iam_container *c)
284 brelse(c->ic_idle_bh);
285 c->ic_idle_bh = NULL;
286 brelse(c->ic_root_bh);
287 c->ic_root_bh = NULL;
290 void iam_path_init(struct iam_path *path, struct iam_container *c,
291 struct iam_path_descr *pd)
293 memset(path, 0, sizeof *path);
294 path->ip_container = c;
295 path->ip_frame = path->ip_frames;
297 path->ip_leaf.il_path = path;
300 static void iam_leaf_fini(struct iam_leaf *leaf);
302 void iam_path_release(struct iam_path *path)
306 for (i = 0; i < ARRAY_SIZE(path->ip_frames); i++) {
307 if (path->ip_frames[i].bh != NULL) {
308 path->ip_frames[i].at_shifted = 0;
309 brelse(path->ip_frames[i].bh);
310 path->ip_frames[i].bh = NULL;
315 void iam_path_fini(struct iam_path *path)
317 iam_leaf_fini(&path->ip_leaf);
318 iam_path_release(path);
322 void iam_path_compat_init(struct iam_path_compat *path, struct inode *inode)
326 path->ipc_hinfo = &path->ipc_hinfo_area;
327 for (i = 0; i < ARRAY_SIZE(path->ipc_scratch); ++i)
328 path->ipc_descr.ipd_key_scratch[i] =
329 (struct iam_ikey *)&path->ipc_scratch[i];
331 iam_path_init(&path->ipc_path, &path->ipc_container, &path->ipc_descr);
334 void iam_path_compat_fini(struct iam_path_compat *path)
336 iam_path_fini(&path->ipc_path);
340 * Helper function initializing iam_path_descr and its key scratch area.
342 struct iam_path_descr *iam_ipd_alloc(void *area, int keysize)
344 struct iam_path_descr *ipd;
350 for (i = 0; i < ARRAY_SIZE(ipd->ipd_key_scratch); ++i, karea += keysize)
351 ipd->ipd_key_scratch[i] = karea;
355 void iam_ipd_free(struct iam_path_descr *ipd)
359 int iam_node_read(struct iam_container *c, iam_ptr_t ptr,
360 handle_t *h, struct buffer_head **bh)
364 /* NB: it can be called by iam_lfix_guess() which is still at
365 * very early stage, c->ic_root_bh and c->ic_descr->id_ops
366 * haven't been intialized yet.
367 * Also, we don't have this for IAM dir.
369 if (c->ic_root_bh != NULL &&
370 c->ic_descr->id_ops->id_root_ptr(c) == ptr) {
371 get_bh(c->ic_root_bh);
376 *bh = ldiskfs_bread(h, c->ic_object, (int)ptr, 0, &result);
378 result = result ? result : -EIO;
383 * Return pointer to current leaf record. Pointer is valid while corresponding
384 * leaf node is locked and pinned.
386 static struct iam_rec *iam_leaf_rec(const struct iam_leaf *leaf)
388 return iam_leaf_ops(leaf)->rec(leaf);
392 * Return pointer to the current leaf key. This function returns pointer to
393 * the key stored in node.
395 * Caller should assume that returned pointer is only valid while leaf node is
398 static struct iam_key *iam_leaf_key(const struct iam_leaf *leaf)
400 return iam_leaf_ops(leaf)->key(leaf);
403 static int iam_leaf_key_size(const struct iam_leaf *leaf)
405 return iam_leaf_ops(leaf)->key_size(leaf);
408 static struct iam_ikey *iam_leaf_ikey(const struct iam_leaf *leaf,
409 struct iam_ikey *key)
411 return iam_leaf_ops(leaf)->ikey(leaf, key);
414 static int iam_leaf_keycmp(const struct iam_leaf *leaf,
415 const struct iam_key *key)
417 return iam_leaf_ops(leaf)->key_cmp(leaf, key);
420 static int iam_leaf_keyeq(const struct iam_leaf *leaf,
421 const struct iam_key *key)
423 return iam_leaf_ops(leaf)->key_eq(leaf, key);
426 #if LDISKFS_INVARIANT_ON
427 static int iam_leaf_check(struct iam_leaf *leaf);
428 extern int dx_node_check(struct iam_path *p, struct iam_frame *f);
430 static int iam_path_check(struct iam_path *p)
435 struct iam_descr *param;
438 param = iam_path_descr(p);
439 for (i = 0; result && i < ARRAY_SIZE(p->ip_frames); ++i) {
440 f = &p->ip_frames[i];
442 result = dx_node_check(p, f);
444 result = !param->id_ops->id_node_check(p, f);
447 if (result && p->ip_leaf.il_bh != NULL)
448 result = iam_leaf_check(&p->ip_leaf);
450 ldiskfs_std_error(iam_path_obj(p)->i_sb, result);
456 static int iam_leaf_load(struct iam_path *path)
460 struct iam_container *c;
461 struct buffer_head *bh;
462 struct iam_leaf *leaf;
463 struct iam_descr *descr;
465 c = path->ip_container;
466 leaf = &path->ip_leaf;
467 descr = iam_path_descr(path);
468 block = path->ip_frame->leaf;
471 printk(KERN_EMERG "wrong leaf: %lu %d [%p %p %p]\n",
472 (long unsigned)path->ip_frame->leaf,
473 dx_get_count(dx_node_get_entries(path, path->ip_frame)),
474 path->ip_frames[0].bh, path->ip_frames[1].bh,
475 path->ip_frames[2].bh);
477 err = descr->id_ops->id_node_read(c, block, NULL, &bh);
480 leaf->il_curidx = block;
481 err = iam_leaf_ops(leaf)->init(leaf);
482 assert_inv(ergo(err == 0, iam_leaf_check(leaf)));
487 static void iam_unlock_htree(struct iam_container *ic,
488 struct dynlock_handle *lh)
491 dynlock_unlock(&ic->ic_tree_lock, lh);
495 static void iam_leaf_unlock(struct iam_leaf *leaf)
497 if (leaf->il_lock != NULL) {
498 iam_unlock_htree(iam_leaf_container(leaf),
501 leaf->il_lock = NULL;
505 static void iam_leaf_fini(struct iam_leaf *leaf)
507 if (leaf->il_path != NULL) {
508 iam_leaf_unlock(leaf);
509 assert_inv(ergo(leaf->il_bh != NULL, iam_leaf_check(leaf)));
510 iam_leaf_ops(leaf)->fini(leaf);
519 static void iam_leaf_start(struct iam_leaf *folio)
521 iam_leaf_ops(folio)->start(folio);
524 void iam_leaf_next(struct iam_leaf *folio)
526 iam_leaf_ops(folio)->next(folio);
529 static void iam_leaf_rec_add(struct iam_leaf *leaf, const struct iam_key *key,
530 const struct iam_rec *rec)
532 iam_leaf_ops(leaf)->rec_add(leaf, key, rec);
535 static void iam_rec_del(struct iam_leaf *leaf, int shift)
537 iam_leaf_ops(leaf)->rec_del(leaf, shift);
540 int iam_leaf_at_end(const struct iam_leaf *leaf)
542 return iam_leaf_ops(leaf)->at_end(leaf);
545 static void iam_leaf_split(struct iam_leaf *l, struct buffer_head **bh,
548 iam_leaf_ops(l)->split(l, bh, nr);
551 static inline int iam_leaf_empty(struct iam_leaf *l)
553 return iam_leaf_ops(l)->leaf_empty(l);
556 int iam_leaf_can_add(const struct iam_leaf *l,
557 const struct iam_key *k, const struct iam_rec *r)
559 return iam_leaf_ops(l)->can_add(l, k, r);
562 #if LDISKFS_INVARIANT_ON
563 static int iam_leaf_check(struct iam_leaf *leaf)
567 struct iam_lentry *orig;
568 struct iam_path *path;
569 struct iam_container *bag;
576 path = iam_leaf_path(leaf);
577 bag = iam_leaf_container(leaf);
579 result = iam_leaf_ops(leaf)->init(leaf);
584 iam_leaf_start(leaf);
585 k0 = iam_path_ikey(path, 0);
586 k1 = iam_path_ikey(path, 1);
587 while (!iam_leaf_at_end(leaf)) {
588 iam_ikeycpy(bag, k0, k1);
589 iam_ikeycpy(bag, k1, iam_leaf_ikey(leaf, k1));
590 if (!first && iam_ikeycmp(bag, k0, k1) > 0) {
602 static int iam_txn_dirty(handle_t *handle,
603 struct iam_path *path, struct buffer_head *bh)
607 result = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
609 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
613 static int iam_txn_add(handle_t *handle,
614 struct iam_path *path, struct buffer_head *bh)
618 result = ldiskfs_journal_get_write_access(handle, bh);
620 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
624 /***********************************************************************/
625 /* iterator interface */
626 /***********************************************************************/
628 static enum iam_it_state it_state(const struct iam_iterator *it)
634 * Helper function returning scratch key.
636 static struct iam_container *iam_it_container(const struct iam_iterator *it)
638 return it->ii_path.ip_container;
641 static inline int it_keycmp(const struct iam_iterator *it,
642 const struct iam_key *k)
644 return iam_leaf_keycmp(&it->ii_path.ip_leaf, k);
647 static inline int it_keyeq(const struct iam_iterator *it,
648 const struct iam_key *k)
650 return iam_leaf_keyeq(&it->ii_path.ip_leaf, k);
653 static int it_ikeycmp(const struct iam_iterator *it, const struct iam_ikey *ik)
655 return iam_ikeycmp(it->ii_path.ip_container,
656 iam_leaf_ikey(&it->ii_path.ip_leaf,
657 iam_path_ikey(&it->ii_path, 0)), ik);
660 static inline int it_at_rec(const struct iam_iterator *it)
662 return !iam_leaf_at_end(&it->ii_path.ip_leaf);
665 static inline int it_before(const struct iam_iterator *it)
667 return it_state(it) == IAM_IT_SKEWED && it_at_rec(it);
671 * Helper wrapper around iam_it_get(): returns 0 (success) only when record
672 * with exactly the same key as asked is found.
674 static int iam_it_get_exact(struct iam_iterator *it, const struct iam_key *k)
678 result = iam_it_get(it, k);
681 else if (result == 0)
683 * Return -ENOENT if cursor is located above record with a key
684 * different from one specified, or in the empty leaf.
686 * XXX returning -ENOENT only works if iam_it_get() never
687 * returns -ENOENT as a legitimate error.
693 void iam_container_write_lock(struct iam_container *ic)
695 down_write(&ic->ic_sem);
698 void iam_container_write_unlock(struct iam_container *ic)
700 up_write(&ic->ic_sem);
703 void iam_container_read_lock(struct iam_container *ic)
705 down_read(&ic->ic_sem);
708 void iam_container_read_unlock(struct iam_container *ic)
710 up_read(&ic->ic_sem);
714 * Initialize iterator to IAM_IT_DETACHED state.
716 * postcondition: it_state(it) == IAM_IT_DETACHED
718 int iam_it_init(struct iam_iterator *it, struct iam_container *c, __u32 flags,
719 struct iam_path_descr *pd)
721 memset(it, 0, sizeof *it);
722 it->ii_flags = flags;
723 it->ii_state = IAM_IT_DETACHED;
724 iam_path_init(&it->ii_path, c, pd);
729 * Finalize iterator and release all resources.
731 * precondition: it_state(it) == IAM_IT_DETACHED
733 void iam_it_fini(struct iam_iterator *it)
735 assert_corr(it_state(it) == IAM_IT_DETACHED);
736 iam_path_fini(&it->ii_path);
740 * this locking primitives are used to protect parts
741 * of dir's htree. protection unit is block: leaf or index
743 static struct dynlock_handle *iam_lock_htree(struct iam_container *ic,
745 enum dynlock_type lt)
747 return dynlock_lock(&ic->ic_tree_lock, value, lt, GFP_NOFS);
750 static int iam_index_lock(struct iam_path *path, struct dynlock_handle **lh)
754 for (f = path->ip_frame; f >= path->ip_frames; --f, ++lh) {
756 *lh = iam_lock_htree(path->ip_container, f->curidx, DLT_READ);
764 * Fast check for frame consistency.
766 static int iam_check_fast(struct iam_path *path, struct iam_frame *frame)
768 struct iam_container *bag;
769 struct iam_entry *next;
770 struct iam_entry *last;
771 struct iam_entry *entries;
772 struct iam_entry *at;
774 bag = path->ip_container;
776 entries = frame->entries;
777 last = iam_entry_shift(path, entries, dx_get_count(entries) - 1);
779 if (unlikely(at > last))
782 if (unlikely(dx_get_block(path, at) != frame->leaf))
785 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, at),
786 path->ip_ikey_target) > 0))
789 next = iam_entry_shift(path, at, +1);
791 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, next),
792 path->ip_ikey_target) <= 0))
798 int dx_index_is_compat(struct iam_path *path)
800 return iam_path_descr(path) == NULL;
806 * search position of specified hash in index
810 static struct iam_entry *iam_find_position(struct iam_path *path,
811 struct iam_frame *frame)
818 count = dx_get_count(frame->entries);
819 assert_corr(count && count <= dx_get_limit(frame->entries));
820 p = iam_entry_shift(path, frame->entries,
821 dx_index_is_compat(path) ? 1 : 2);
822 q = iam_entry_shift(path, frame->entries, count - 1);
824 m = iam_entry_shift(path, p, iam_entry_diff(path, q, p) / 2);
825 if (iam_ikeycmp(path->ip_container, iam_ikey_at(path, m),
826 path->ip_ikey_target) > 0)
827 q = iam_entry_shift(path, m, -1);
829 p = iam_entry_shift(path, m, +1);
831 return iam_entry_shift(path, p, -1);
836 static iam_ptr_t iam_find_ptr(struct iam_path *path, struct iam_frame *frame)
838 return dx_get_block(path, iam_find_position(path, frame));
841 void iam_insert_key(struct iam_path *path, struct iam_frame *frame,
842 const struct iam_ikey *key, iam_ptr_t ptr)
844 struct iam_entry *entries = frame->entries;
845 struct iam_entry *new = iam_entry_shift(path, frame->at, +1);
846 int count = dx_get_count(entries);
849 * Unfortunately we cannot assert this, as this function is sometimes
850 * called by VFS under i_sem and without pdirops lock.
852 assert_corr(1 || iam_frame_is_locked(path, frame));
853 assert_corr(count < dx_get_limit(entries));
854 assert_corr(frame->at < iam_entry_shift(path, entries, count));
855 assert_inv(dx_node_check(path, frame));
857 memmove(iam_entry_shift(path, new, 1), new,
858 (char *)iam_entry_shift(path, entries, count) - (char *)new);
859 dx_set_ikey(path, new, key);
860 dx_set_block(path, new, ptr);
861 dx_set_count(entries, count + 1);
862 assert_inv(dx_node_check(path, frame));
865 void iam_insert_key_lock(struct iam_path *path, struct iam_frame *frame,
866 const struct iam_ikey *key, iam_ptr_t ptr)
868 iam_lock_bh(frame->bh);
869 iam_insert_key(path, frame, key, ptr);
870 iam_unlock_bh(frame->bh);
873 * returns 0 if path was unchanged, -EAGAIN otherwise.
875 static int iam_check_path(struct iam_path *path, struct iam_frame *frame)
879 iam_lock_bh(frame->bh);
880 equal = iam_check_fast(path, frame) == 0 ||
881 frame->leaf == iam_find_ptr(path, frame);
882 DX_DEVAL(iam_lock_stats.dls_bh_again += !equal);
883 iam_unlock_bh(frame->bh);
885 return equal ? 0 : -EAGAIN;
888 static int iam_lookup_try(struct iam_path *path)
894 struct iam_descr *param;
895 struct iam_frame *frame;
896 struct iam_container *c;
898 param = iam_path_descr(path);
899 c = path->ip_container;
901 ptr = param->id_ops->id_root_ptr(c);
902 for (frame = path->ip_frames, i = 0; i <= path->ip_indirect;
904 err = param->id_ops->id_node_read(c, (iam_ptr_t)ptr, NULL,
908 iam_lock_bh(frame->bh);
910 * node must be initialized under bh lock because concurrent
911 * creation procedure may change it and iam_lookup_try() will
912 * see obsolete tree height. -bzzz
917 if (LDISKFS_INVARIANT_ON) {
918 err = param->id_ops->id_node_check(path, frame);
923 err = param->id_ops->id_node_load(path, frame);
927 assert_inv(dx_node_check(path, frame));
929 * splitting may change root index block and move hash we're
930 * looking for into another index block so, we have to check
931 * this situation and repeat from begining if path got changed
935 err = iam_check_path(path, frame - 1);
940 frame->at = iam_find_position(path, frame);
942 frame->leaf = ptr = dx_get_block(path, frame->at);
944 iam_unlock_bh(frame->bh);
948 iam_unlock_bh(frame->bh);
949 path->ip_frame = --frame;
953 static int __iam_path_lookup(struct iam_path *path)
958 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++ i)
959 assert(path->ip_frames[i].bh == NULL);
962 err = iam_lookup_try(path);
966 } while (err == -EAGAIN);
972 * returns 0 if path was unchanged, -EAGAIN otherwise.
974 static int iam_check_full_path(struct iam_path *path, int search)
976 struct iam_frame *bottom;
977 struct iam_frame *scan;
983 for (bottom = path->ip_frames, i = 0;
984 i < DX_MAX_TREE_HEIGHT && bottom->bh != NULL; ++bottom, ++i) {
985 ; /* find last filled in frame */
989 * Lock frames, bottom to top.
991 for (scan = bottom - 1; scan >= path->ip_frames; --scan)
992 iam_lock_bh(scan->bh);
994 * Check them top to bottom.
997 for (scan = path->ip_frames; scan < bottom; ++scan) {
998 struct iam_entry *pos;
1001 if (iam_check_fast(path, scan) == 0)
1004 pos = iam_find_position(path, scan);
1005 if (scan->leaf != dx_get_block(path, pos)) {
1011 pos = iam_entry_shift(path, scan->entries,
1012 dx_get_count(scan->entries) - 1);
1013 if (scan->at > pos ||
1014 scan->leaf != dx_get_block(path, scan->at)) {
1022 * Unlock top to bottom.
1024 for (scan = path->ip_frames; scan < bottom; ++scan)
1025 iam_unlock_bh(scan->bh);
1026 DX_DEVAL(iam_lock_stats.dls_bh_full_again += !!result);
1027 do_corr(schedule());
1034 * Performs path lookup and returns with found leaf (if any) locked by htree
1037 static int iam_lookup_lock(struct iam_path *path,
1038 struct dynlock_handle **dl, enum dynlock_type lt)
1042 while ((result = __iam_path_lookup(path)) == 0) {
1043 do_corr(schedule());
1044 *dl = iam_lock_htree(path->ip_container, path->ip_frame->leaf,
1047 iam_path_fini(path);
1051 do_corr(schedule());
1053 * while locking leaf we just found may get split so we need
1054 * to check this -bzzz
1056 if (iam_check_full_path(path, 1) == 0)
1058 iam_unlock_htree(path->ip_container, *dl);
1060 iam_path_fini(path);
1065 * Performs tree top-to-bottom traversal starting from root, and loads leaf
1068 static int iam_path_lookup(struct iam_path *path, int index)
1070 struct iam_container *c;
1071 struct iam_leaf *leaf;
1074 c = path->ip_container;
1075 leaf = &path->ip_leaf;
1076 result = iam_lookup_lock(path, &leaf->il_lock, DLT_WRITE);
1077 assert_inv(iam_path_check(path));
1078 do_corr(schedule());
1080 result = iam_leaf_load(path);
1081 assert_inv(ergo(result == 0, iam_leaf_check(leaf)));
1083 do_corr(schedule());
1085 result = iam_leaf_ops(leaf)->
1086 ilookup(leaf, path->ip_ikey_target);
1088 result = iam_leaf_ops(leaf)->
1089 lookup(leaf, path->ip_key_target);
1090 do_corr(schedule());
1093 iam_leaf_unlock(leaf);
1099 * Common part of iam_it_{i,}get().
1101 static int __iam_it_get(struct iam_iterator *it, int index)
1104 assert_corr(it_state(it) == IAM_IT_DETACHED);
1106 result = iam_path_lookup(&it->ii_path, index);
1110 collision = result & IAM_LOOKUP_LAST;
1111 switch (result & ~IAM_LOOKUP_LAST) {
1112 case IAM_LOOKUP_EXACT:
1114 it->ii_state = IAM_IT_ATTACHED;
1118 it->ii_state = IAM_IT_ATTACHED;
1120 case IAM_LOOKUP_BEFORE:
1121 case IAM_LOOKUP_EMPTY:
1123 it->ii_state = IAM_IT_SKEWED;
1128 result |= collision;
1131 * See iam_it_get_exact() for explanation.
1133 assert_corr(result != -ENOENT);
1138 * Correct hash, but not the same key was found, iterate through hash
1139 * collision chain, looking for correct record.
1141 static int iam_it_collision(struct iam_iterator *it)
1145 assert(ergo(it_at_rec(it), !it_keyeq(it, it->ii_path.ip_key_target)));
1147 while ((result = iam_it_next(it)) == 0) {
1148 do_corr(schedule());
1149 if (it_ikeycmp(it, it->ii_path.ip_ikey_target) != 0)
1151 if (it_keyeq(it, it->ii_path.ip_key_target))
1158 * Attach iterator. After successful completion, @it points to record with
1159 * least key not larger than @k.
1161 * Return value: 0: positioned on existing record,
1162 * +ve: exact position found,
1165 * precondition: it_state(it) == IAM_IT_DETACHED
1166 * postcondition: ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1167 * it_keycmp(it, k) <= 0)
1169 int iam_it_get(struct iam_iterator *it, const struct iam_key *k)
1172 assert_corr(it_state(it) == IAM_IT_DETACHED);
1174 it->ii_path.ip_ikey_target = NULL;
1175 it->ii_path.ip_key_target = k;
1177 result = __iam_it_get(it, 0);
1179 if (result == IAM_LOOKUP_LAST) {
1180 result = iam_it_collision(it);
1184 result = __iam_it_get(it, 0);
1189 result &= ~IAM_LOOKUP_LAST;
1191 assert_corr(ergo(result > 0, it_keycmp(it, k) == 0));
1192 assert_corr(ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1193 it_keycmp(it, k) <= 0));
1198 * Attach iterator by index key.
1200 static int iam_it_iget(struct iam_iterator *it, const struct iam_ikey *k)
1202 assert_corr(it_state(it) == IAM_IT_DETACHED);
1204 it->ii_path.ip_ikey_target = k;
1205 return __iam_it_get(it, 1) & ~IAM_LOOKUP_LAST;
1209 * Attach iterator, and assure it points to the record (not skewed).
1211 * Return value: 0: positioned on existing record,
1212 * +ve: exact position found,
1215 * precondition: it_state(it) == IAM_IT_DETACHED &&
1216 * !(it->ii_flags&IAM_IT_WRITE)
1217 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED)
1219 int iam_it_get_at(struct iam_iterator *it, const struct iam_key *k)
1222 assert_corr(it_state(it) == IAM_IT_DETACHED &&
1223 !(it->ii_flags&IAM_IT_WRITE));
1224 result = iam_it_get(it, k);
1226 if (it_state(it) != IAM_IT_ATTACHED) {
1227 assert_corr(it_state(it) == IAM_IT_SKEWED);
1228 result = iam_it_next(it);
1231 assert_corr(ergo(result >= 0, it_state(it) == IAM_IT_ATTACHED));
1236 * Duplicates iterator.
1238 * postcondition: it_state(dst) == it_state(src) &&
1239 * iam_it_container(dst) == iam_it_container(src) &&
1240 * dst->ii_flags = src->ii_flags &&
1241 * ergo(it_state(src) == IAM_IT_ATTACHED,
1242 * iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1243 * iam_it_key_get(dst) == iam_it_key_get(src))
1245 void iam_it_dup(struct iam_iterator *dst, const struct iam_iterator *src)
1247 dst->ii_flags = src->ii_flags;
1248 dst->ii_state = src->ii_state;
1249 /* XXX not yet. iam_path_dup(&dst->ii_path, &src->ii_path); */
1251 * XXX: duplicate lock.
1253 assert_corr(it_state(dst) == it_state(src));
1254 assert_corr(iam_it_container(dst) == iam_it_container(src));
1255 assert_corr(dst->ii_flags = src->ii_flags);
1256 assert_corr(ergo(it_state(src) == IAM_IT_ATTACHED,
1257 iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1258 iam_it_key_get(dst) == iam_it_key_get(src)));
1263 * Detach iterator. Does nothing it detached state.
1265 * postcondition: it_state(it) == IAM_IT_DETACHED
1267 void iam_it_put(struct iam_iterator *it)
1269 if (it->ii_state != IAM_IT_DETACHED) {
1270 it->ii_state = IAM_IT_DETACHED;
1271 iam_leaf_fini(&it->ii_path.ip_leaf);
1275 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1276 struct iam_ikey *ikey);
1280 * This function increments the frame pointer to search the next leaf
1281 * block, and reads in the necessary intervening nodes if the search
1282 * should be necessary. Whether or not the search is necessary is
1283 * controlled by the hash parameter. If the hash value is even, then
1284 * the search is only continued if the next block starts with that
1285 * hash value. This is used if we are searching for a specific file.
1287 * If the hash value is HASH_NB_ALWAYS, then always go to the next block.
1289 * This function returns 1 if the caller should continue to search,
1290 * or 0 if it should not. If there is an error reading one of the
1291 * index blocks, it will a negative error code.
1293 * If start_hash is non-null, it will be filled in with the starting
1294 * hash of the next page.
1296 static int iam_htree_advance(struct inode *dir, __u32 hash,
1297 struct iam_path *path, __u32 *start_hash,
1300 struct iam_frame *p;
1301 struct buffer_head *bh;
1302 int err, num_frames = 0;
1307 * Find the next leaf page by incrementing the frame pointer.
1308 * If we run out of entries in the interior node, loop around and
1309 * increment pointer in the parent node. When we break out of
1310 * this loop, num_frames indicates the number of interior
1311 * nodes need to be read.
1314 do_corr(schedule());
1319 p->at = iam_entry_shift(path, p->at, +1);
1320 if (p->at < iam_entry_shift(path, p->entries,
1321 dx_get_count(p->entries))) {
1322 p->leaf = dx_get_block(path, p->at);
1323 iam_unlock_bh(p->bh);
1326 iam_unlock_bh(p->bh);
1327 if (p == path->ip_frames)
1338 * If the hash is 1, then continue only if the next page has a
1339 * continuation hash of any value. This is used for readdir
1340 * handling. Otherwise, check to see if the hash matches the
1341 * desired contiuation hash. If it doesn't, return since
1342 * there's no point to read in the successive index pages.
1344 dx_get_ikey(path, p->at, (struct iam_ikey *)&bhash);
1346 *start_hash = bhash;
1347 if ((hash & 1) == 0) {
1348 if ((bhash & ~1) != hash)
1353 * If the hash is HASH_NB_ALWAYS, we always go to the next
1354 * block so no check is necessary
1356 while (num_frames--) {
1359 do_corr(schedule());
1361 idx = p->leaf = dx_get_block(path, p->at);
1362 iam_unlock_bh(p->bh);
1363 err = iam_path_descr(path)->id_ops->
1364 id_node_read(path->ip_container, idx, NULL, &bh);
1366 return err; /* Failure */
1369 assert_corr(p->bh != bh);
1371 p->entries = dx_node_get_entries(path, p);
1372 p->at = iam_entry_shift(path, p->entries, !compat);
1373 assert_corr(p->curidx != idx);
1376 assert_corr(p->leaf != dx_get_block(path, p->at));
1377 p->leaf = dx_get_block(path, p->at);
1378 iam_unlock_bh(p->bh);
1379 assert_inv(dx_node_check(path, p));
1385 static inline int iam_index_advance(struct iam_path *path)
1387 return iam_htree_advance(iam_path_obj(path), 0, path, NULL, 0);
1390 static void iam_unlock_array(struct iam_container *ic,
1391 struct dynlock_handle **lh)
1395 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++i, ++lh) {
1397 iam_unlock_htree(ic, *lh);
1403 * Advance index part of @path to point to the next leaf. Returns 1 on
1404 * success, 0, when end of container was reached. Leaf node is locked.
1406 int iam_index_next(struct iam_container *c, struct iam_path *path)
1409 struct dynlock_handle *lh[DX_MAX_TREE_HEIGHT] = { NULL, };
1411 struct inode *object;
1414 * Locking for iam_index_next()... is to be described.
1417 object = c->ic_object;
1418 cursor = path->ip_frame->leaf;
1421 result = iam_index_lock(path, lh);
1422 do_corr(schedule());
1426 result = iam_check_full_path(path, 0);
1427 if (result == 0 && cursor == path->ip_frame->leaf) {
1428 result = iam_index_advance(path);
1430 assert_corr(result == 0 ||
1431 cursor != path->ip_frame->leaf);
1435 iam_unlock_array(c, lh);
1437 iam_path_release(path);
1438 do_corr(schedule());
1440 result = __iam_path_lookup(path);
1444 while (path->ip_frame->leaf != cursor) {
1445 do_corr(schedule());
1447 result = iam_index_lock(path, lh);
1448 do_corr(schedule());
1452 result = iam_check_full_path(path, 0);
1456 result = iam_index_advance(path);
1458 CERROR("cannot find cursor : %u\n",
1464 result = iam_check_full_path(path, 0);
1467 iam_unlock_array(c, lh);
1469 } while (result == -EAGAIN);
1473 iam_unlock_array(c, lh);
1478 * Move iterator one record right.
1480 * Return value: 0: success,
1481 * +1: end of container reached
1484 * precondition: (it_state(it) == IAM_IT_ATTACHED ||
1485 * it_state(it) == IAM_IT_SKEWED) && it->ii_flags&IAM_IT_MOVE
1486 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED) &&
1487 * ergo(result > 0, it_state(it) == IAM_IT_DETACHED)
1489 int iam_it_next(struct iam_iterator *it)
1492 struct iam_path *path;
1493 struct iam_leaf *leaf;
1494 do_corr(struct iam_ikey *ik_orig);
1496 /* assert_corr(it->ii_flags&IAM_IT_MOVE); */
1497 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1498 it_state(it) == IAM_IT_SKEWED);
1500 path = &it->ii_path;
1501 leaf = &path->ip_leaf;
1503 assert_corr(iam_leaf_is_locked(leaf));
1506 do_corr(ik_orig = it_at_rec(it) ?
1507 iam_it_ikey_get(it, iam_path_ikey(path, 2)) : NULL);
1508 if (it_before(it)) {
1509 assert_corr(!iam_leaf_at_end(leaf));
1510 it->ii_state = IAM_IT_ATTACHED;
1512 if (!iam_leaf_at_end(leaf))
1513 /* advance within leaf node */
1514 iam_leaf_next(leaf);
1516 * multiple iterations may be necessary due to empty leaves.
1518 while (result == 0 && iam_leaf_at_end(leaf)) {
1519 do_corr(schedule());
1520 /* advance index portion of the path */
1521 result = iam_index_next(iam_it_container(it), path);
1522 assert_corr(iam_leaf_is_locked(leaf));
1524 struct dynlock_handle *lh;
1525 lh = iam_lock_htree(iam_it_container(it),
1526 path->ip_frame->leaf,
1529 iam_leaf_fini(leaf);
1531 result = iam_leaf_load(path);
1533 iam_leaf_start(leaf);
1536 } else if (result == 0)
1537 /* end of container reached */
1543 it->ii_state = IAM_IT_ATTACHED;
1545 assert_corr(ergo(result == 0, it_state(it) == IAM_IT_ATTACHED));
1546 assert_corr(ergo(result > 0, it_state(it) == IAM_IT_DETACHED));
1547 assert_corr(ergo(result == 0 && ik_orig != NULL,
1548 it_ikeycmp(it, ik_orig) >= 0));
1553 * Return pointer to the record under iterator.
1555 * precondition: it_state(it) == IAM_IT_ATTACHED && it_at_rec(it)
1556 * postcondition: it_state(it) == IAM_IT_ATTACHED
1558 struct iam_rec *iam_it_rec_get(const struct iam_iterator *it)
1560 assert_corr(it_state(it) == IAM_IT_ATTACHED);
1561 assert_corr(it_at_rec(it));
1562 return iam_leaf_rec(&it->ii_path.ip_leaf);
1565 static void iam_it_reccpy(struct iam_iterator *it, const struct iam_rec *r)
1567 struct iam_leaf *folio;
1569 folio = &it->ii_path.ip_leaf;
1570 iam_leaf_ops(folio)->rec_set(folio, r);
1574 * Replace contents of record under iterator.
1576 * precondition: it_state(it) == IAM_IT_ATTACHED &&
1577 * it->ii_flags&IAM_IT_WRITE
1578 * postcondition: it_state(it) == IAM_IT_ATTACHED &&
1579 * ergo(result == 0, !memcmp(iam_it_rec_get(it), r, ...))
1581 int iam_it_rec_set(handle_t *h,
1582 struct iam_iterator *it, const struct iam_rec *r)
1585 struct iam_path *path;
1586 struct buffer_head *bh;
1588 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
1589 it->ii_flags&IAM_IT_WRITE);
1590 assert_corr(it_at_rec(it));
1592 path = &it->ii_path;
1593 bh = path->ip_leaf.il_bh;
1594 result = iam_txn_add(h, path, bh);
1596 iam_it_reccpy(it, r);
1597 result = iam_txn_dirty(h, path, bh);
1603 * Return pointer to the index key under iterator.
1605 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1606 * it_state(it) == IAM_IT_SKEWED
1608 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1609 struct iam_ikey *ikey)
1611 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1612 it_state(it) == IAM_IT_SKEWED);
1613 assert_corr(it_at_rec(it));
1614 return iam_leaf_ikey(&it->ii_path.ip_leaf, ikey);
1618 * Return pointer to the key under iterator.
1620 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1621 * it_state(it) == IAM_IT_SKEWED
1623 struct iam_key *iam_it_key_get(const struct iam_iterator *it)
1625 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1626 it_state(it) == IAM_IT_SKEWED);
1627 assert_corr(it_at_rec(it));
1628 return iam_leaf_key(&it->ii_path.ip_leaf);
1632 * Return size of key under iterator (in bytes)
1634 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1635 * it_state(it) == IAM_IT_SKEWED
1637 int iam_it_key_size(const struct iam_iterator *it)
1639 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1640 it_state(it) == IAM_IT_SKEWED);
1641 assert_corr(it_at_rec(it));
1642 return iam_leaf_key_size(&it->ii_path.ip_leaf);
1645 static struct buffer_head *
1646 iam_new_node(handle_t *h, struct iam_container *c, iam_ptr_t *b, int *e)
1648 struct inode *inode = c->ic_object;
1649 struct buffer_head *bh = NULL;
1650 struct iam_idle_head *head;
1651 struct buffer_head *idle;
1655 if (c->ic_idle_bh == NULL)
1658 mutex_lock(&c->ic_idle_mutex);
1659 if (unlikely(c->ic_idle_bh == NULL)) {
1660 mutex_unlock(&c->ic_idle_mutex);
1664 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
1665 count = le16_to_cpu(head->iih_count);
1667 *e = ldiskfs_journal_get_write_access(h, c->ic_idle_bh);
1672 *b = le32_to_cpu(head->iih_blks[count]);
1673 head->iih_count = cpu_to_le16(count);
1674 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_idle_bh);
1678 mutex_unlock(&c->ic_idle_mutex);
1679 bh = ldiskfs_bread(NULL, inode, *b, 0, e);
1681 *e = *e ? *e : -EIO;
1687 /* The block itself which contains the iam_idle_head is
1688 * also an idle block, and can be used as the new node. */
1689 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
1690 c->ic_descr->id_root_gap +
1691 sizeof(struct dx_countlimit));
1692 *e = ldiskfs_journal_get_write_access(h, c->ic_root_bh);
1696 *b = le32_to_cpu(*idle_blocks);
1697 iam_lock_bh(c->ic_root_bh);
1698 *idle_blocks = head->iih_next;
1699 iam_unlock_bh(c->ic_root_bh);
1700 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_root_bh);
1702 iam_lock_bh(c->ic_root_bh);
1703 *idle_blocks = cpu_to_le32(*b);
1704 iam_unlock_bh(c->ic_root_bh);
1709 idle = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
1710 if (idle != NULL && IS_ERR(idle)) {
1712 c->ic_idle_bh = NULL;
1717 c->ic_idle_bh = idle;
1718 mutex_unlock(&c->ic_idle_mutex);
1721 /* get write access for the found buffer head */
1722 *e = ldiskfs_journal_get_write_access(h, bh);
1726 ldiskfs_std_error(inode->i_sb, *e);
1728 /* Clear the reused node as new node does. */
1729 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1730 set_buffer_uptodate(bh);
1735 bh = osd_ldiskfs_append(h, inode, b);
1744 mutex_unlock(&c->ic_idle_mutex);
1745 ldiskfs_std_error(inode->i_sb, *e);
1750 * Insertion of new record. Interaction with jbd during non-trivial case (when
1751 * split happens) is as following:
1753 * - new leaf node is involved into transaction by iam_new_node();
1755 * - old leaf node is involved into transaction by iam_add_rec();
1757 * - leaf where insertion point ends in, is marked dirty by iam_add_rec();
1759 * - leaf without insertion point is marked dirty (as @new_leaf) by
1762 * - split index nodes are involved into transaction and marked dirty by
1763 * split_index_node().
1765 * - "safe" index node, which is no split, but where new pointer is inserted
1766 * is involved into transaction and marked dirty by split_index_node().
1768 * - index node where pointer to new leaf is inserted is involved into
1769 * transaction by split_index_node() and marked dirty by iam_add_rec().
1771 * - inode is marked dirty by iam_add_rec().
1775 static int iam_new_leaf(handle_t *handle, struct iam_leaf *leaf)
1779 struct buffer_head *new_leaf;
1780 struct buffer_head *old_leaf;
1781 struct iam_container *c;
1783 struct iam_path *path;
1785 assert_inv(iam_leaf_check(leaf));
1787 c = iam_leaf_container(leaf);
1788 path = leaf->il_path;
1791 new_leaf = iam_new_node(handle, c, &blknr, &err);
1792 do_corr(schedule());
1793 if (new_leaf != NULL) {
1794 struct dynlock_handle *lh;
1796 lh = iam_lock_htree(c, blknr, DLT_WRITE);
1797 do_corr(schedule());
1799 iam_leaf_ops(leaf)->init_new(c, new_leaf);
1800 do_corr(schedule());
1801 old_leaf = leaf->il_bh;
1802 iam_leaf_split(leaf, &new_leaf, blknr);
1803 if (old_leaf != leaf->il_bh) {
1805 * Switched to the new leaf.
1807 iam_leaf_unlock(leaf);
1809 path->ip_frame->leaf = blknr;
1811 iam_unlock_htree(path->ip_container, lh);
1812 do_corr(schedule());
1813 err = iam_txn_dirty(handle, path, new_leaf);
1815 err = ldiskfs_mark_inode_dirty(handle, obj);
1816 do_corr(schedule());
1821 assert_inv(iam_leaf_check(leaf));
1822 assert_inv(iam_leaf_check(&iam_leaf_path(leaf)->ip_leaf));
1823 assert_inv(iam_path_check(iam_leaf_path(leaf)));
1827 static inline void dx_set_limit(struct iam_entry *entries, unsigned value)
1829 ((struct dx_countlimit *) entries)->limit = cpu_to_le16(value);
1832 static int iam_shift_entries(struct iam_path *path,
1833 struct iam_frame *frame, unsigned count,
1834 struct iam_entry *entries, struct iam_entry *entries2,
1841 struct iam_frame *parent = frame - 1;
1842 struct iam_ikey *pivot = iam_path_ikey(path, 3);
1844 delta = dx_index_is_compat(path) ? 0 : +1;
1846 count1 = count/2 + delta;
1847 count2 = count - count1;
1848 dx_get_ikey(path, iam_entry_shift(path, entries, count1), pivot);
1850 dxtrace(printk("Split index %d/%d\n", count1, count2));
1852 memcpy((char *) iam_entry_shift(path, entries2, delta),
1853 (char *) iam_entry_shift(path, entries, count1),
1854 count2 * iam_entry_size(path));
1856 dx_set_count(entries2, count2 + delta);
1857 dx_set_limit(entries2, dx_node_limit(path));
1860 * NOTE: very subtle piece of code competing dx_probe() may find 2nd
1861 * level index in root index, then we insert new index here and set
1862 * new count in that 2nd level index. so, dx_probe() may see 2nd level
1863 * index w/o hash it looks for. the solution is to check root index
1864 * after we locked just founded 2nd level index -bzzz
1866 iam_insert_key_lock(path, parent, pivot, newblock);
1869 * now old and new 2nd level index blocks contain all pointers, so
1870 * dx_probe() may find it in the both. it's OK -bzzz
1872 iam_lock_bh(frame->bh);
1873 dx_set_count(entries, count1);
1874 iam_unlock_bh(frame->bh);
1877 * now old 2nd level index block points to first half of leafs. it's
1878 * importand that dx_probe() must check root index block for changes
1879 * under dx_lock_bh(frame->bh) -bzzz
1886 int split_index_node(handle_t *handle, struct iam_path *path,
1887 struct dynlock_handle **lh)
1890 struct iam_entry *entries; /* old block contents */
1891 struct iam_entry *entries2; /* new block contents */
1892 struct iam_frame *frame, *safe;
1893 struct buffer_head *bh_new[DX_MAX_TREE_HEIGHT] = {NULL};
1894 u32 newblock[DX_MAX_TREE_HEIGHT] = {0};
1895 struct dynlock_handle *lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1896 struct dynlock_handle *new_lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1897 struct inode *dir = iam_path_obj(path);
1898 struct iam_descr *descr;
1902 descr = iam_path_descr(path);
1904 * Algorithm below depends on this.
1906 assert_corr(dx_root_limit(path) < dx_node_limit(path));
1908 frame = path->ip_frame;
1909 entries = frame->entries;
1912 * Tall-tree handling: we might have to split multiple index blocks
1913 * all the way up to tree root. Tricky point here is error handling:
1914 * to avoid complicated undo/rollback we
1916 * - first allocate all necessary blocks
1918 * - insert pointers into them atomically.
1922 * Locking: leaf is already locked. htree-locks are acquired on all
1923 * index nodes that require split bottom-to-top, on the "safe" node,
1924 * and on all new nodes
1927 dxtrace(printk("using %u of %u node entries\n",
1928 dx_get_count(entries), dx_get_limit(entries)));
1930 /* What levels need split? */
1931 for (nr_splet = 0; frame >= path->ip_frames &&
1932 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1933 --frame, ++nr_splet) {
1934 do_corr(schedule());
1935 if (nr_splet == DX_MAX_TREE_HEIGHT) {
1937 CWARN(dir->i_sb, __FUNCTION__,
1938 "Directory index full!\n");
1948 * Lock all nodes, bottom to top.
1950 for (frame = path->ip_frame, i = nr_splet; i >= 0; --i, --frame) {
1951 do_corr(schedule());
1952 lock[i] = iam_lock_htree(path->ip_container, frame->curidx,
1954 if (lock[i] == NULL) {
1961 * Check for concurrent index modification.
1963 err = iam_check_full_path(path, 1);
1967 * And check that the same number of nodes is to be split.
1969 for (i = 0, frame = path->ip_frame; frame >= path->ip_frames &&
1970 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1974 if (i != nr_splet) {
1979 /* Go back down, allocating blocks, locking them, and adding into
1981 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
1982 bh_new[i] = iam_new_node(handle, path->ip_container,
1983 &newblock[i], &err);
1984 do_corr(schedule());
1986 descr->id_ops->id_node_init(path->ip_container,
1989 new_lock[i] = iam_lock_htree(path->ip_container, newblock[i],
1991 if (new_lock[i] == NULL) {
1995 do_corr(schedule());
1996 BUFFER_TRACE(frame->bh, "get_write_access");
1997 err = ldiskfs_journal_get_write_access(handle, frame->bh);
2001 /* Add "safe" node to transaction too */
2002 if (safe + 1 != path->ip_frames) {
2003 do_corr(schedule());
2004 err = ldiskfs_journal_get_write_access(handle, safe->bh);
2009 /* Go through nodes once more, inserting pointers */
2010 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
2013 struct buffer_head *bh2;
2014 struct buffer_head *bh;
2016 entries = frame->entries;
2017 count = dx_get_count(entries);
2018 idx = iam_entry_diff(path, frame->at, entries);
2021 entries2 = dx_get_entries(path, bh2->b_data, 0);
2024 if (frame == path->ip_frames) {
2025 /* splitting root node. Tricky point:
2027 * In the "normal" B-tree we'd split root *and* add
2028 * new root to the tree with pointers to the old root
2029 * and its sibling (thus introducing two new nodes).
2031 * In htree it's enough to add one node, because
2032 * capacity of the root node is smaller than that of
2035 struct iam_frame *frames;
2036 struct iam_entry *next;
2038 assert_corr(i == 0);
2040 do_corr(schedule());
2042 frames = path->ip_frames;
2043 memcpy((char *) entries2, (char *) entries,
2044 count * iam_entry_size(path));
2045 dx_set_limit(entries2, dx_node_limit(path));
2048 iam_lock_bh(frame->bh);
2049 next = descr->id_ops->id_root_inc(path->ip_container,
2051 dx_set_block(path, next, newblock[0]);
2052 iam_unlock_bh(frame->bh);
2054 do_corr(schedule());
2055 /* Shift frames in the path */
2056 memmove(frames + 2, frames + 1,
2057 (sizeof path->ip_frames) - 2 * sizeof frames[0]);
2058 /* Add new access path frame */
2059 frames[1].at = iam_entry_shift(path, entries2, idx);
2060 frames[1].entries = entries = entries2;
2062 assert_inv(dx_node_check(path, frame));
2065 assert_inv(dx_node_check(path, frame));
2066 bh_new[0] = NULL; /* buffer head is "consumed" */
2067 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2070 do_corr(schedule());
2072 /* splitting non-root index node. */
2073 struct iam_frame *parent = frame - 1;
2075 do_corr(schedule());
2076 count = iam_shift_entries(path, frame, count,
2077 entries, entries2, newblock[i]);
2078 /* Which index block gets the new entry? */
2080 int d = dx_index_is_compat(path) ? 0 : +1;
2082 frame->at = iam_entry_shift(path, entries2,
2084 frame->entries = entries = entries2;
2085 frame->curidx = newblock[i];
2086 swap(frame->bh, bh2);
2087 assert_corr(lock[i + 1] != NULL);
2088 assert_corr(new_lock[i] != NULL);
2089 swap(lock[i + 1], new_lock[i]);
2091 parent->at = iam_entry_shift(path,
2094 assert_inv(dx_node_check(path, frame));
2095 assert_inv(dx_node_check(path, parent));
2096 dxtrace(dx_show_index ("node", frame->entries));
2097 dxtrace(dx_show_index ("node",
2098 ((struct dx_node *) bh2->b_data)->entries));
2099 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2102 do_corr(schedule());
2103 err = ldiskfs_handle_dirty_metadata(handle, NULL,
2108 do_corr(schedule());
2109 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2114 * This function was called to make insertion of new leaf
2115 * possible. Check that it fulfilled its obligations.
2117 assert_corr(dx_get_count(path->ip_frame->entries) <
2118 dx_get_limit(path->ip_frame->entries));
2119 assert_corr(lock[nr_splet] != NULL);
2120 *lh = lock[nr_splet];
2121 lock[nr_splet] = NULL;
2124 * Log ->i_size modification.
2126 err = ldiskfs_mark_inode_dirty(handle, dir);
2132 ldiskfs_std_error(dir->i_sb, err);
2135 iam_unlock_array(path->ip_container, lock);
2136 iam_unlock_array(path->ip_container, new_lock);
2138 assert_corr(err || iam_frame_is_locked(path, path->ip_frame));
2140 do_corr(schedule());
2141 for (i = 0; i < ARRAY_SIZE(bh_new); ++i) {
2142 if (bh_new[i] != NULL)
2148 static int iam_add_rec(handle_t *handle, struct iam_iterator *it,
2149 struct iam_path *path,
2150 const struct iam_key *k, const struct iam_rec *r)
2153 struct iam_leaf *leaf;
2155 leaf = &path->ip_leaf;
2156 assert_inv(iam_leaf_check(leaf));
2157 assert_inv(iam_path_check(path));
2158 err = iam_txn_add(handle, path, leaf->il_bh);
2160 do_corr(schedule());
2161 if (!iam_leaf_can_add(leaf, k, r)) {
2162 struct dynlock_handle *lh = NULL;
2165 assert_corr(lh == NULL);
2166 do_corr(schedule());
2167 err = split_index_node(handle, path, &lh);
2168 if (err == -EAGAIN) {
2169 assert_corr(lh == NULL);
2171 iam_path_fini(path);
2172 it->ii_state = IAM_IT_DETACHED;
2174 do_corr(schedule());
2175 err = iam_it_get_exact(it, k);
2177 err = +1; /* repeat split */
2182 assert_inv(iam_path_check(path));
2184 assert_corr(lh != NULL);
2185 do_corr(schedule());
2186 err = iam_new_leaf(handle, leaf);
2188 err = iam_txn_dirty(handle, path,
2189 path->ip_frame->bh);
2191 iam_unlock_htree(path->ip_container, lh);
2192 do_corr(schedule());
2195 iam_leaf_rec_add(leaf, k, r);
2196 err = iam_txn_dirty(handle, path, leaf->il_bh);
2199 assert_inv(iam_leaf_check(leaf));
2200 assert_inv(iam_leaf_check(&path->ip_leaf));
2201 assert_inv(iam_path_check(path));
2206 * Insert new record with key @k and contents from @r, shifting records to the
2207 * right. On success, iterator is positioned on the newly inserted record.
2209 * precondition: it->ii_flags&IAM_IT_WRITE &&
2210 * (it_state(it) == IAM_IT_ATTACHED ||
2211 * it_state(it) == IAM_IT_SKEWED) &&
2212 * ergo(it_state(it) == IAM_IT_ATTACHED,
2213 * it_keycmp(it, k) <= 0) &&
2214 * ergo(it_before(it), it_keycmp(it, k) > 0));
2215 * postcondition: ergo(result == 0,
2216 * it_state(it) == IAM_IT_ATTACHED &&
2217 * it_keycmp(it, k) == 0 &&
2218 * !memcmp(iam_it_rec_get(it), r, ...))
2220 int iam_it_rec_insert(handle_t *h, struct iam_iterator *it,
2221 const struct iam_key *k, const struct iam_rec *r)
2224 struct iam_path *path;
2226 path = &it->ii_path;
2228 assert_corr(it->ii_flags&IAM_IT_WRITE);
2229 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2230 it_state(it) == IAM_IT_SKEWED);
2231 assert_corr(ergo(it_state(it) == IAM_IT_ATTACHED,
2232 it_keycmp(it, k) <= 0));
2233 assert_corr(ergo(it_before(it), it_keycmp(it, k) > 0));
2234 result = iam_add_rec(h, it, path, k, r);
2236 it->ii_state = IAM_IT_ATTACHED;
2237 assert_corr(ergo(result == 0,
2238 it_state(it) == IAM_IT_ATTACHED &&
2239 it_keycmp(it, k) == 0));
2243 static inline int iam_idle_blocks_limit(struct inode *inode)
2245 return (inode->i_sb->s_blocksize - sizeof(struct iam_idle_head)) >> 2;
2249 * If the leaf cannnot be recycled, we will lose one block for reusing.
2250 * It is not a serious issue because it almost the same of non-recycle.
2252 static iam_ptr_t iam_index_shrink(handle_t *h, struct iam_path *p,
2253 struct iam_leaf *l, struct buffer_head **bh)
2255 struct iam_container *c = p->ip_container;
2256 struct inode *inode = c->ic_object;
2257 struct iam_frame *frame = p->ip_frame;
2258 struct iam_entry *entries;
2259 struct iam_entry *pos;
2260 struct dynlock_handle *lh;
2264 if (c->ic_idle_failed)
2267 if (unlikely(frame == NULL))
2270 if (!iam_leaf_empty(l))
2273 lh = iam_lock_htree(c, frame->curidx, DLT_WRITE);
2275 CWARN("%.16s: No memory to recycle idle blocks\n",
2276 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name);
2280 rc = iam_txn_add(h, p, frame->bh);
2282 iam_unlock_htree(c, lh);
2286 iam_lock_bh(frame->bh);
2287 entries = frame->entries;
2288 count = dx_get_count(entries);
2289 /* NOT shrink the last entry in the index node, which can be reused
2290 * directly by next new node. */
2292 iam_unlock_bh(frame->bh);
2293 iam_unlock_htree(c, lh);
2297 pos = iam_find_position(p, frame);
2298 /* There may be some new leaf nodes have been added or empty leaf nodes
2299 * have been shrinked during my delete operation.
2301 * If the empty leaf is not under current index node because the index
2302 * node has been split, then just skip the empty leaf, which is rare. */
2303 if (unlikely(frame->leaf != dx_get_block(p, pos))) {
2304 iam_unlock_bh(frame->bh);
2305 iam_unlock_htree(c, lh);
2310 if (frame->at < iam_entry_shift(p, entries, count - 1)) {
2311 struct iam_entry *n = iam_entry_shift(p, frame->at, 1);
2313 memmove(frame->at, n,
2314 (char *)iam_entry_shift(p, entries, count) - (char *)n);
2315 frame->at_shifted = 1;
2317 dx_set_count(entries, count - 1);
2318 iam_unlock_bh(frame->bh);
2319 rc = iam_txn_dirty(h, p, frame->bh);
2320 iam_unlock_htree(c, lh);
2330 iam_install_idle_blocks(handle_t *h, struct iam_path *p, struct buffer_head *bh,
2331 __u32 *idle_blocks, iam_ptr_t blk)
2333 struct iam_container *c = p->ip_container;
2334 struct buffer_head *old = c->ic_idle_bh;
2335 struct iam_idle_head *head;
2338 head = (struct iam_idle_head *)(bh->b_data);
2339 head->iih_magic = cpu_to_le16(IAM_IDLE_HEADER_MAGIC);
2340 head->iih_count = 0;
2341 head->iih_next = *idle_blocks;
2342 /* The bh already get_write_accessed. */
2343 rc = iam_txn_dirty(h, p, bh);
2347 rc = iam_txn_add(h, p, c->ic_root_bh);
2351 iam_lock_bh(c->ic_root_bh);
2352 *idle_blocks = cpu_to_le32(blk);
2353 iam_unlock_bh(c->ic_root_bh);
2354 rc = iam_txn_dirty(h, p, c->ic_root_bh);
2356 /* NOT release old before new assigned. */
2361 iam_lock_bh(c->ic_root_bh);
2362 *idle_blocks = head->iih_next;
2363 iam_unlock_bh(c->ic_root_bh);
2369 * If the leaf cannnot be recycled, we will lose one block for reusing.
2370 * It is not a serious issue because it almost the same of non-recycle.
2372 static void iam_recycle_leaf(handle_t *h, struct iam_path *p,
2373 struct buffer_head *bh, iam_ptr_t blk)
2375 struct iam_container *c = p->ip_container;
2376 struct inode *inode = c->ic_object;
2377 struct iam_idle_head *head;
2382 mutex_lock(&c->ic_idle_mutex);
2383 if (unlikely(c->ic_idle_failed)) {
2388 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
2389 c->ic_descr->id_root_gap +
2390 sizeof(struct dx_countlimit));
2391 /* It is the first idle block. */
2392 if (c->ic_idle_bh == NULL) {
2393 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2397 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
2398 count = le16_to_cpu(head->iih_count);
2399 /* Current ic_idle_bh is full, to be replaced by the leaf. */
2400 if (count == iam_idle_blocks_limit(inode)) {
2401 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2405 /* Just add to ic_idle_bh. */
2406 rc = iam_txn_add(h, p, c->ic_idle_bh);
2410 head->iih_blks[count] = cpu_to_le32(blk);
2411 head->iih_count = cpu_to_le16(count + 1);
2412 rc = iam_txn_dirty(h, p, c->ic_idle_bh);
2415 mutex_unlock(&c->ic_idle_mutex);
2417 CWARN("%.16s: idle blocks failed, will lose the blk %u\n",
2418 LDISKFS_SB(inode->i_sb)->s_es->s_volume_name, blk);
2422 * Delete record under iterator.
2424 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2425 * it->ii_flags&IAM_IT_WRITE &&
2427 * postcondition: it_state(it) == IAM_IT_ATTACHED ||
2428 * it_state(it) == IAM_IT_DETACHED
2430 int iam_it_rec_delete(handle_t *h, struct iam_iterator *it)
2433 struct iam_leaf *leaf;
2434 struct iam_path *path;
2436 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
2437 it->ii_flags&IAM_IT_WRITE);
2438 assert_corr(it_at_rec(it));
2440 path = &it->ii_path;
2441 leaf = &path->ip_leaf;
2443 assert_inv(iam_leaf_check(leaf));
2444 assert_inv(iam_path_check(path));
2446 result = iam_txn_add(h, path, leaf->il_bh);
2448 * no compaction for now.
2451 iam_rec_del(leaf, it->ii_flags&IAM_IT_MOVE);
2452 result = iam_txn_dirty(h, path, leaf->il_bh);
2453 if (result == 0 && iam_leaf_at_end(leaf)) {
2454 struct buffer_head *bh = NULL;
2457 blk = iam_index_shrink(h, path, leaf, &bh);
2458 if (it->ii_flags & IAM_IT_MOVE) {
2459 result = iam_it_next(it);
2465 iam_recycle_leaf(h, path, bh, blk);
2470 assert_inv(iam_leaf_check(leaf));
2471 assert_inv(iam_path_check(path));
2472 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2473 it_state(it) == IAM_IT_DETACHED);
2478 * Convert iterator to cookie.
2480 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2481 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2482 * postcondition: it_state(it) == IAM_IT_ATTACHED
2484 iam_pos_t iam_it_store(const struct iam_iterator *it)
2488 assert_corr(it_state(it) == IAM_IT_ATTACHED);
2489 assert_corr(it_at_rec(it));
2490 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <=
2494 return *(iam_pos_t *)iam_it_ikey_get(it, (void *)&result);
2498 * Restore iterator from cookie.
2500 * precondition: it_state(it) == IAM_IT_DETACHED && it->ii_flags&IAM_IT_MOVE &&
2501 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2502 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED &&
2503 * iam_it_store(it) == pos)
2505 int iam_it_load(struct iam_iterator *it, iam_pos_t pos)
2507 assert_corr(it_state(it) == IAM_IT_DETACHED &&
2508 it->ii_flags&IAM_IT_MOVE);
2509 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <= sizeof pos);
2510 return iam_it_iget(it, (struct iam_ikey *)&pos);
2513 /***********************************************************************/
2515 /***********************************************************************/
2517 static inline int ptr_inside(void *base, size_t size, void *ptr)
2519 return (base <= ptr) && (ptr < base + size);
2522 static int iam_frame_invariant(struct iam_frame *f)
2526 f->bh->b_data != NULL &&
2527 ptr_inside(f->bh->b_data, f->bh->b_size, f->entries) &&
2528 ptr_inside(f->bh->b_data, f->bh->b_size, f->at) &&
2529 f->entries <= f->at);
2532 static int iam_leaf_invariant(struct iam_leaf *l)
2536 l->il_bh->b_data != NULL &&
2537 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_entries) &&
2538 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_at) &&
2539 l->il_entries <= l->il_at;
2542 static int iam_path_invariant(struct iam_path *p)
2546 if (p->ip_container == NULL ||
2547 p->ip_indirect < 0 || p->ip_indirect > DX_MAX_TREE_HEIGHT - 1 ||
2548 p->ip_frame != p->ip_frames + p->ip_indirect ||
2549 !iam_leaf_invariant(&p->ip_leaf))
2551 for (i = 0; i < ARRAY_SIZE(p->ip_frames); ++i) {
2552 if (i <= p->ip_indirect) {
2553 if (!iam_frame_invariant(&p->ip_frames[i]))
2560 int iam_it_invariant(struct iam_iterator *it)
2563 (it->ii_state == IAM_IT_DETACHED ||
2564 it->ii_state == IAM_IT_ATTACHED ||
2565 it->ii_state == IAM_IT_SKEWED) &&
2566 !(it->ii_flags & ~(IAM_IT_MOVE | IAM_IT_WRITE)) &&
2567 ergo(it->ii_state == IAM_IT_ATTACHED ||
2568 it->ii_state == IAM_IT_SKEWED,
2569 iam_path_invariant(&it->ii_path) &&
2570 equi(it_at_rec(it), it->ii_state == IAM_IT_SKEWED));
2574 * Search container @c for record with key @k. If record is found, its data
2575 * are moved into @r.
2577 * Return values: 0: found, -ENOENT: not-found, -ve: error
2579 int iam_lookup(struct iam_container *c, const struct iam_key *k,
2580 struct iam_rec *r, struct iam_path_descr *pd)
2582 struct iam_iterator it;
2585 iam_it_init(&it, c, 0, pd);
2587 result = iam_it_get_exact(&it, k);
2590 * record with required key found, copy it into user buffer
2592 iam_reccpy(&it.ii_path.ip_leaf, r);
2599 * Insert new record @r with key @k into container @c (within context of
2602 * Return values: 0: success, -ve: error, including -EEXIST when record with
2603 * given key is already present.
2605 * postcondition: ergo(result == 0 || result == -EEXIST,
2606 * iam_lookup(c, k, r2) > 0;
2608 int iam_insert(handle_t *h, struct iam_container *c, const struct iam_key *k,
2609 const struct iam_rec *r, struct iam_path_descr *pd)
2611 struct iam_iterator it;
2614 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2616 result = iam_it_get_exact(&it, k);
2617 if (result == -ENOENT)
2618 result = iam_it_rec_insert(h, &it, k, r);
2619 else if (result == 0)
2627 * Update record with the key @k in container @c (within context of
2628 * transaction @h), new record is given by @r.
2630 * Return values: +1: skip because of the same rec value, 0: success,
2631 * -ve: error, including -ENOENT if no record with the given key found.
2633 int iam_update(handle_t *h, struct iam_container *c, const struct iam_key *k,
2634 const struct iam_rec *r, struct iam_path_descr *pd)
2636 struct iam_iterator it;
2637 struct iam_leaf *folio;
2640 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2642 result = iam_it_get_exact(&it, k);
2644 folio = &it.ii_path.ip_leaf;
2645 result = iam_leaf_ops(folio)->rec_eq(folio, r);
2647 iam_it_rec_set(h, &it, r);
2657 * Delete existing record with key @k.
2659 * Return values: 0: success, -ENOENT: not-found, -ve: other error.
2661 * postcondition: ergo(result == 0 || result == -ENOENT,
2662 * !iam_lookup(c, k, *));
2664 int iam_delete(handle_t *h, struct iam_container *c, const struct iam_key *k,
2665 struct iam_path_descr *pd)
2667 struct iam_iterator it;
2670 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2672 result = iam_it_get_exact(&it, k);
2674 iam_it_rec_delete(h, &it);
2680 int iam_root_limit(int rootgap, int blocksize, int size)
2685 limit = (blocksize - rootgap) / size;
2686 nlimit = blocksize / size;
2687 if (limit == nlimit)