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
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 * Top-level entry points into iam module
35 * Author: Wang Di <wangdi@clusterfs.com>
36 * Author: Nikita Danilov <nikita@clusterfs.com>
40 * iam: big theory statement.
42 * iam (Index Access Module) is a module providing abstraction of persistent
43 * transactional container on top of generalized ldiskfs htree.
47 * - key, pointer, and record size specifiable per container.
49 * - trees taller than 2 index levels.
51 * - read/write to existing ldiskfs htree directories as iam containers.
53 * iam container is a tree, consisting of leaf nodes containing keys and
54 * records stored in this container, and index nodes, containing keys and
55 * pointers to leaf or index nodes.
57 * iam does not work with keys directly, instead it calls user-supplied key
58 * comparison function (->dpo_keycmp()).
60 * Pointers are (currently) interpreted as logical offsets (measured in
61 * blocksful) within underlying flat file on top of which iam tree lives.
65 * iam mostly tries to reuse existing htree formats.
67 * Format of index node:
69 * +-----+-------+-------+-------+------+-------+------------+
70 * | | count | | | | | |
71 * | gap | / | entry | entry | .... | entry | free space |
72 * | | limit | | | | | |
73 * +-----+-------+-------+-------+------+-------+------------+
75 * gap this part of node is never accessed by iam code. It
76 * exists for binary compatibility with ldiskfs htree (that,
77 * in turn, stores fake struct ext2_dirent for ext2
78 * compatibility), and to keep some unspecified per-node
79 * data. Gap can be different for root and non-root index
80 * nodes. Gap size can be specified for each container
81 * (gap of 0 is allowed).
83 * count/limit current number of entries in this node, and the maximal
84 * number of entries that can fit into node. count/limit
85 * has the same size as entry, and is itself counted in
88 * entry index entry: consists of a key immediately followed by
89 * a pointer to a child node. Size of a key and size of a
90 * pointer depends on container. Entry has neither
91 * alignment nor padding.
93 * free space portion of node new entries are added to
95 * Entries in index node are sorted by their key value.
97 * Format of a leaf node is not specified. Generic iam code accesses leaf
98 * nodes through ->id_leaf methods in struct iam_descr.
100 * The IAM root block is a special node, which contains the IAM descriptor.
101 * It is on disk format:
103 * +---------+-------+--------+---------+-------+------+-------+------------+
104 * |IAM desc | count | idle | | | | | |
105 * |(fix/var)| / | blocks | padding | entry | .... | entry | free space |
106 * | | limit | | | | | | |
107 * +---------+-------+--------+---------+-------+------+-------+------------+
109 * The padding length is calculated with the parameters in the IAM descriptor.
111 * The field "idle_blocks" is used to record empty leaf nodes, which have not
112 * been released but all contained entries in them have been removed. Usually,
113 * the idle blocks in the IAM should be reused when need to allocate new leaf
114 * nodes for new entries, it depends on the IAM hash functions to map the new
115 * entries to these idle blocks. Unfortunately, it is not easy to design some
116 * hash functions for such clever mapping, especially considering the insert/
117 * lookup performance.
119 * So the IAM recycles the empty leaf nodes, and put them into a per-file based
120 * idle blocks pool. If need some new leaf node, it will try to take idle block
121 * from such pool with priority, in spite of how the IAM hash functions to map
124 * The idle blocks pool is organized as a series of tables, and each table
125 * can be described as following (on-disk format):
127 * +---------+---------+---------+---------+------+---------+-------+
128 * | magic | count | next | logic | | logic | free |
129 * |(16 bits)|(16 bits)| table | blk # | .... | blk # | space |
130 * | | |(32 bits)|(32 bits)| |(32 bits)| |
131 * +---------+---------+---------+---------+------+---------+-------+
133 * The logic blk# for the first table is stored in the root node "idle_blocks".
137 #include <linux/module.h>
138 #include <linux/fs.h>
139 #include <linux/pagemap.h>
140 #include <linux/time.h>
141 #include <linux/fcntl.h>
142 #include <linux/stat.h>
143 #include <linux/string.h>
144 #include <linux/quotaops.h>
145 #include <linux/buffer_head.h>
147 #include <ldiskfs/ldiskfs.h>
148 #include <ldiskfs/xattr.h>
151 #include "osd_internal.h"
153 #include <ldiskfs/acl.h>
156 * List of all registered formats.
158 * No locking. Callers synchronize.
160 static struct list_head iam_formats = LIST_HEAD_INIT(iam_formats);
162 void iam_format_register(struct iam_format *fmt)
164 list_add(&fmt->if_linkage, &iam_formats);
167 static struct buffer_head *
168 iam_load_idle_blocks(struct iam_container *c, iam_ptr_t blk)
170 struct inode *inode = c->ic_object;
171 struct iam_idle_head *head;
172 struct buffer_head *bh;
174 LASSERT(mutex_is_locked(&c->ic_idle_mutex));
179 bh = __ldiskfs_bread(NULL, inode, blk, 0);
180 if (IS_ERR_OR_NULL(bh)) {
181 CERROR("%s: cannot load idle blocks, blk = %u, err = %ld\n",
182 osd_ino2name(inode), blk, bh ? PTR_ERR(bh) : -EIO);
183 c->ic_idle_failed = 1;
189 head = (struct iam_idle_head *)(bh->b_data);
190 if (le16_to_cpu(head->iih_magic) != IAM_IDLE_HEADER_MAGIC) {
191 CERROR("%s: invalid idle block head, blk = %u, magic = %d\n",
192 osd_ino2name(inode), blk, le16_to_cpu(head->iih_magic));
194 c->ic_idle_failed = 1;
195 return ERR_PTR(-EBADF);
202 * Determine format of given container. This is done by scanning list of
203 * registered formats and calling ->if_guess() method of each in turn.
205 static int iam_format_guess(struct iam_container *c)
208 struct iam_format *fmt;
211 * XXX temporary initialization hook.
214 static int initialized = 0;
217 iam_lvar_format_init();
218 iam_lfix_format_init();
224 list_for_each_entry(fmt, &iam_formats, if_linkage) {
225 result = fmt->if_guess(c);
231 struct buffer_head *bh;
234 LASSERT(c->ic_root_bh != NULL);
236 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
237 c->ic_descr->id_root_gap +
238 sizeof(struct dx_countlimit));
239 mutex_lock(&c->ic_idle_mutex);
240 bh = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
241 if (bh != NULL && IS_ERR(bh))
242 result = PTR_ERR(bh);
245 mutex_unlock(&c->ic_idle_mutex);
252 * Initialize container @c.
254 int iam_container_init(struct iam_container *c,
255 struct iam_descr *descr, struct inode *inode)
257 memset(c, 0, sizeof *c);
259 c->ic_object = inode;
260 init_rwsem(&c->ic_sem);
261 dynlock_init(&c->ic_tree_lock);
262 mutex_init(&c->ic_idle_mutex);
267 * Determine container format.
269 int iam_container_setup(struct iam_container *c)
271 return iam_format_guess(c);
275 * Finalize container @c, release all resources.
277 void iam_container_fini(struct iam_container *c)
279 brelse(c->ic_idle_bh);
280 c->ic_idle_bh = NULL;
281 brelse(c->ic_root_bh);
282 c->ic_root_bh = NULL;
285 void iam_path_init(struct iam_path *path, struct iam_container *c,
286 struct iam_path_descr *pd)
288 memset(path, 0, sizeof *path);
289 path->ip_container = c;
290 path->ip_frame = path->ip_frames;
292 path->ip_leaf.il_path = path;
295 static void iam_leaf_fini(struct iam_leaf *leaf);
297 void iam_path_release(struct iam_path *path)
301 for (i = 0; i < ARRAY_SIZE(path->ip_frames); i++) {
302 if (path->ip_frames[i].bh != NULL) {
303 path->ip_frames[i].at_shifted = 0;
304 brelse(path->ip_frames[i].bh);
305 path->ip_frames[i].bh = NULL;
310 void iam_path_fini(struct iam_path *path)
312 iam_leaf_fini(&path->ip_leaf);
313 iam_path_release(path);
317 void iam_path_compat_init(struct iam_path_compat *path, struct inode *inode)
321 path->ipc_hinfo = &path->ipc_hinfo_area;
322 for (i = 0; i < ARRAY_SIZE(path->ipc_scratch); ++i)
323 path->ipc_descr.ipd_key_scratch[i] =
324 (struct iam_ikey *)&path->ipc_scratch[i];
326 iam_path_init(&path->ipc_path, &path->ipc_container, &path->ipc_descr);
329 void iam_path_compat_fini(struct iam_path_compat *path)
331 iam_path_fini(&path->ipc_path);
335 * Helper function initializing iam_path_descr and its key scratch area.
337 struct iam_path_descr *iam_ipd_alloc(void *area, int keysize)
339 struct iam_path_descr *ipd;
345 for (i = 0; i < ARRAY_SIZE(ipd->ipd_key_scratch); ++i, karea += keysize)
346 ipd->ipd_key_scratch[i] = karea;
350 void iam_ipd_free(struct iam_path_descr *ipd)
354 int iam_node_read(struct iam_container *c, iam_ptr_t ptr,
355 handle_t *h, struct buffer_head **bh)
357 /* NB: it can be called by iam_lfix_guess() which is still at
358 * very early stage, c->ic_root_bh and c->ic_descr->id_ops
359 * haven't been intialized yet.
360 * Also, we don't have this for IAM dir.
362 if (c->ic_root_bh != NULL &&
363 c->ic_descr->id_ops->id_root_ptr(c) == ptr) {
364 get_bh(c->ic_root_bh);
369 *bh = __ldiskfs_bread(h, c->ic_object, (int)ptr, 0);
380 * Return pointer to current leaf record. Pointer is valid while corresponding
381 * leaf node is locked and pinned.
383 static struct iam_rec *iam_leaf_rec(const struct iam_leaf *leaf)
385 return iam_leaf_ops(leaf)->rec(leaf);
389 * Return pointer to the current leaf key. This function returns pointer to
390 * the key stored in node.
392 * Caller should assume that returned pointer is only valid while leaf node is
395 static struct iam_key *iam_leaf_key(const struct iam_leaf *leaf)
397 return iam_leaf_ops(leaf)->key(leaf);
400 static int iam_leaf_key_size(const struct iam_leaf *leaf)
402 return iam_leaf_ops(leaf)->key_size(leaf);
405 static struct iam_ikey *iam_leaf_ikey(const struct iam_leaf *leaf,
406 struct iam_ikey *key)
408 return iam_leaf_ops(leaf)->ikey(leaf, key);
411 static int iam_leaf_keycmp(const struct iam_leaf *leaf,
412 const struct iam_key *key)
414 return iam_leaf_ops(leaf)->key_cmp(leaf, key);
417 static int iam_leaf_keyeq(const struct iam_leaf *leaf,
418 const struct iam_key *key)
420 return iam_leaf_ops(leaf)->key_eq(leaf, key);
423 #if LDISKFS_INVARIANT_ON
424 extern int dx_node_check(struct iam_path *p, struct iam_frame *f);
426 static int iam_path_check(struct iam_path *p)
431 struct iam_descr *param;
434 param = iam_path_descr(p);
435 for (i = 0; result && i < ARRAY_SIZE(p->ip_frames); ++i) {
436 f = &p->ip_frames[i];
438 result = dx_node_check(p, f);
440 result = !param->id_ops->id_node_check(p, f);
443 if (result && p->ip_leaf.il_bh != NULL)
446 ldiskfs_std_error(iam_path_obj(p)->i_sb, result);
452 static int iam_leaf_load(struct iam_path *path)
456 struct iam_container *c;
457 struct buffer_head *bh;
458 struct iam_leaf *leaf;
459 struct iam_descr *descr;
461 c = path->ip_container;
462 leaf = &path->ip_leaf;
463 descr = iam_path_descr(path);
464 block = path->ip_frame->leaf;
467 printk(KERN_EMERG "wrong leaf: %lu %d [%p %p %p]\n",
468 (long unsigned)path->ip_frame->leaf,
469 dx_get_count(dx_node_get_entries(path, path->ip_frame)),
470 path->ip_frames[0].bh, path->ip_frames[1].bh,
471 path->ip_frames[2].bh);
473 err = descr->id_ops->id_node_read(c, block, NULL, &bh);
476 leaf->il_curidx = block;
477 err = iam_leaf_ops(leaf)->init(leaf);
482 static void iam_unlock_htree(struct iam_container *ic,
483 struct dynlock_handle *lh)
486 dynlock_unlock(&ic->ic_tree_lock, lh);
490 static void iam_leaf_unlock(struct iam_leaf *leaf)
492 if (leaf->il_lock != NULL) {
493 iam_unlock_htree(iam_leaf_container(leaf),
496 leaf->il_lock = NULL;
500 static void iam_leaf_fini(struct iam_leaf *leaf)
502 if (leaf->il_path != NULL) {
503 iam_leaf_unlock(leaf);
504 iam_leaf_ops(leaf)->fini(leaf);
513 static void iam_leaf_start(struct iam_leaf *folio)
515 iam_leaf_ops(folio)->start(folio);
518 void iam_leaf_next(struct iam_leaf *folio)
520 iam_leaf_ops(folio)->next(folio);
523 static void iam_leaf_rec_add(struct iam_leaf *leaf, const struct iam_key *key,
524 const struct iam_rec *rec)
526 iam_leaf_ops(leaf)->rec_add(leaf, key, rec);
529 static void iam_rec_del(struct iam_leaf *leaf, int shift)
531 iam_leaf_ops(leaf)->rec_del(leaf, shift);
534 int iam_leaf_at_end(const struct iam_leaf *leaf)
536 return iam_leaf_ops(leaf)->at_end(leaf);
539 static void iam_leaf_split(struct iam_leaf *l, struct buffer_head **bh,
542 iam_leaf_ops(l)->split(l, bh, nr);
545 static inline int iam_leaf_empty(struct iam_leaf *l)
547 return iam_leaf_ops(l)->leaf_empty(l);
550 int iam_leaf_can_add(const struct iam_leaf *l,
551 const struct iam_key *k, const struct iam_rec *r)
553 return iam_leaf_ops(l)->can_add(l, k, r);
556 static int iam_txn_dirty(handle_t *handle,
557 struct iam_path *path, struct buffer_head *bh)
561 result = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
563 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
567 static int iam_txn_add(handle_t *handle,
568 struct iam_path *path, struct buffer_head *bh)
572 result = ldiskfs_journal_get_write_access(handle, bh);
574 ldiskfs_std_error(iam_path_obj(path)->i_sb, result);
578 /***********************************************************************/
579 /* iterator interface */
580 /***********************************************************************/
582 static enum iam_it_state it_state(const struct iam_iterator *it)
588 * Helper function returning scratch key.
590 static struct iam_container *iam_it_container(const struct iam_iterator *it)
592 return it->ii_path.ip_container;
595 static inline int it_keycmp(const struct iam_iterator *it,
596 const struct iam_key *k)
598 return iam_leaf_keycmp(&it->ii_path.ip_leaf, k);
601 static inline int it_keyeq(const struct iam_iterator *it,
602 const struct iam_key *k)
604 return iam_leaf_keyeq(&it->ii_path.ip_leaf, k);
607 static int it_ikeycmp(const struct iam_iterator *it, const struct iam_ikey *ik)
609 return iam_ikeycmp(it->ii_path.ip_container,
610 iam_leaf_ikey(&it->ii_path.ip_leaf,
611 iam_path_ikey(&it->ii_path, 0)), ik);
614 static inline int it_at_rec(const struct iam_iterator *it)
616 return !iam_leaf_at_end(&it->ii_path.ip_leaf);
619 static inline int it_before(const struct iam_iterator *it)
621 return it_state(it) == IAM_IT_SKEWED && it_at_rec(it);
625 * Helper wrapper around iam_it_get(): returns 0 (success) only when record
626 * with exactly the same key as asked is found.
628 static int iam_it_get_exact(struct iam_iterator *it, const struct iam_key *k)
632 result = iam_it_get(it, k);
635 else if (result == 0)
637 * Return -ENOENT if cursor is located above record with a key
638 * different from one specified, or in the empty leaf.
640 * XXX returning -ENOENT only works if iam_it_get() never
641 * returns -ENOENT as a legitimate error.
647 void iam_container_write_lock(struct iam_container *ic)
649 down_write(&ic->ic_sem);
652 void iam_container_write_unlock(struct iam_container *ic)
654 up_write(&ic->ic_sem);
657 void iam_container_read_lock(struct iam_container *ic)
659 down_read(&ic->ic_sem);
662 void iam_container_read_unlock(struct iam_container *ic)
664 up_read(&ic->ic_sem);
668 * Initialize iterator to IAM_IT_DETACHED state.
670 * postcondition: it_state(it) == IAM_IT_DETACHED
672 int iam_it_init(struct iam_iterator *it, struct iam_container *c, __u32 flags,
673 struct iam_path_descr *pd)
675 memset(it, 0, sizeof *it);
676 it->ii_flags = flags;
677 it->ii_state = IAM_IT_DETACHED;
678 iam_path_init(&it->ii_path, c, pd);
683 * Finalize iterator and release all resources.
685 * precondition: it_state(it) == IAM_IT_DETACHED
687 void iam_it_fini(struct iam_iterator *it)
689 assert_corr(it_state(it) == IAM_IT_DETACHED);
690 iam_path_fini(&it->ii_path);
694 * this locking primitives are used to protect parts
695 * of dir's htree. protection unit is block: leaf or index
697 static struct dynlock_handle *iam_lock_htree(struct iam_container *ic,
699 enum dynlock_type lt)
701 return dynlock_lock(&ic->ic_tree_lock, value, lt, GFP_NOFS);
704 static int iam_index_lock(struct iam_path *path, struct dynlock_handle **lh)
708 for (f = path->ip_frame; f >= path->ip_frames; --f, ++lh) {
710 *lh = iam_lock_htree(path->ip_container, f->curidx, DLT_READ);
718 * Fast check for frame consistency.
720 static int iam_check_fast(struct iam_path *path, struct iam_frame *frame)
722 struct iam_container *bag;
723 struct iam_entry *next;
724 struct iam_entry *last;
725 struct iam_entry *entries;
726 struct iam_entry *at;
728 bag = path->ip_container;
730 entries = frame->entries;
731 last = iam_entry_shift(path, entries, dx_get_count(entries) - 1);
733 if (unlikely(at > last))
736 if (unlikely(dx_get_block(path, at) != frame->leaf))
739 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, at),
740 path->ip_ikey_target) > 0))
743 next = iam_entry_shift(path, at, +1);
745 if (unlikely(iam_ikeycmp(bag, iam_ikey_at(path, next),
746 path->ip_ikey_target) <= 0))
752 int dx_index_is_compat(struct iam_path *path)
754 return iam_path_descr(path) == NULL;
760 * search position of specified hash in index
764 static struct iam_entry *iam_find_position(struct iam_path *path,
765 struct iam_frame *frame)
772 count = dx_get_count(frame->entries);
773 assert_corr(count && count <= dx_get_limit(frame->entries));
774 p = iam_entry_shift(path, frame->entries,
775 dx_index_is_compat(path) ? 1 : 2);
776 q = iam_entry_shift(path, frame->entries, count - 1);
778 m = iam_entry_shift(path, p, iam_entry_diff(path, q, p) / 2);
779 if (iam_ikeycmp(path->ip_container, iam_ikey_at(path, m),
780 path->ip_ikey_target) > 0)
781 q = iam_entry_shift(path, m, -1);
783 p = iam_entry_shift(path, m, +1);
785 return iam_entry_shift(path, p, -1);
790 static iam_ptr_t iam_find_ptr(struct iam_path *path, struct iam_frame *frame)
792 return dx_get_block(path, iam_find_position(path, frame));
795 void iam_insert_key(struct iam_path *path, struct iam_frame *frame,
796 const struct iam_ikey *key, iam_ptr_t ptr)
798 struct iam_entry *entries = frame->entries;
799 struct iam_entry *new = iam_entry_shift(path, frame->at, +1);
800 int count = dx_get_count(entries);
803 * Unfortunately we cannot assert this, as this function is sometimes
804 * called by VFS under i_sem and without pdirops lock.
806 assert_corr(1 || iam_frame_is_locked(path, frame));
807 assert_corr(count < dx_get_limit(entries));
808 assert_corr(frame->at < iam_entry_shift(path, entries, count));
809 assert_inv(dx_node_check(path, frame));
811 memmove(iam_entry_shift(path, new, 1), new,
812 (char *)iam_entry_shift(path, entries, count) - (char *)new);
813 dx_set_ikey(path, new, key);
814 dx_set_block(path, new, ptr);
815 dx_set_count(entries, count + 1);
816 assert_inv(dx_node_check(path, frame));
819 void iam_insert_key_lock(struct iam_path *path, struct iam_frame *frame,
820 const struct iam_ikey *key, iam_ptr_t ptr)
822 iam_lock_bh(frame->bh);
823 iam_insert_key(path, frame, key, ptr);
824 iam_unlock_bh(frame->bh);
827 * returns 0 if path was unchanged, -EAGAIN otherwise.
829 static int iam_check_path(struct iam_path *path, struct iam_frame *frame)
833 iam_lock_bh(frame->bh);
834 equal = iam_check_fast(path, frame) == 0 ||
835 frame->leaf == iam_find_ptr(path, frame);
836 DX_DEVAL(iam_lock_stats.dls_bh_again += !equal);
837 iam_unlock_bh(frame->bh);
839 return equal ? 0 : -EAGAIN;
842 static int iam_lookup_try(struct iam_path *path)
848 struct iam_descr *param;
849 struct iam_frame *frame;
850 struct iam_container *c;
852 param = iam_path_descr(path);
853 c = path->ip_container;
855 ptr = param->id_ops->id_root_ptr(c);
856 for (frame = path->ip_frames, i = 0; i <= path->ip_indirect;
858 err = param->id_ops->id_node_read(c, (iam_ptr_t)ptr, NULL,
862 iam_lock_bh(frame->bh);
864 * node must be initialized under bh lock because concurrent
865 * creation procedure may change it and iam_lookup_try() will
866 * see obsolete tree height. -bzzz
871 if (LDISKFS_INVARIANT_ON) {
872 err = param->id_ops->id_node_check(path, frame);
877 err = param->id_ops->id_node_load(path, frame);
881 assert_inv(dx_node_check(path, frame));
883 * splitting may change root index block and move hash we're
884 * looking for into another index block so, we have to check
885 * this situation and repeat from begining if path got changed
889 err = iam_check_path(path, frame - 1);
894 frame->at = iam_find_position(path, frame);
896 frame->leaf = ptr = dx_get_block(path, frame->at);
898 iam_unlock_bh(frame->bh);
902 iam_unlock_bh(frame->bh);
903 path->ip_frame = --frame;
907 static int __iam_path_lookup(struct iam_path *path)
912 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++ i)
913 assert(path->ip_frames[i].bh == NULL);
916 err = iam_lookup_try(path);
920 } while (err == -EAGAIN);
926 * returns 0 if path was unchanged, -EAGAIN otherwise.
928 static int iam_check_full_path(struct iam_path *path, int search)
930 struct iam_frame *bottom;
931 struct iam_frame *scan;
937 for (bottom = path->ip_frames, i = 0;
938 i < DX_MAX_TREE_HEIGHT && bottom->bh != NULL; ++bottom, ++i) {
939 ; /* find last filled in frame */
943 * Lock frames, bottom to top.
945 for (scan = bottom - 1; scan >= path->ip_frames; --scan)
946 iam_lock_bh(scan->bh);
948 * Check them top to bottom.
951 for (scan = path->ip_frames; scan < bottom; ++scan) {
952 struct iam_entry *pos;
955 if (iam_check_fast(path, scan) == 0)
958 pos = iam_find_position(path, scan);
959 if (scan->leaf != dx_get_block(path, pos)) {
965 pos = iam_entry_shift(path, scan->entries,
966 dx_get_count(scan->entries) - 1);
967 if (scan->at > pos ||
968 scan->leaf != dx_get_block(path, scan->at)) {
976 * Unlock top to bottom.
978 for (scan = path->ip_frames; scan < bottom; ++scan)
979 iam_unlock_bh(scan->bh);
980 DX_DEVAL(iam_lock_stats.dls_bh_full_again += !!result);
988 * Performs path lookup and returns with found leaf (if any) locked by htree
991 static int iam_lookup_lock(struct iam_path *path,
992 struct dynlock_handle **dl, enum dynlock_type lt)
996 while ((result = __iam_path_lookup(path)) == 0) {
998 *dl = iam_lock_htree(path->ip_container, path->ip_frame->leaf,
1001 iam_path_fini(path);
1005 do_corr(schedule());
1007 * while locking leaf we just found may get split so we need
1008 * to check this -bzzz
1010 if (iam_check_full_path(path, 1) == 0)
1012 iam_unlock_htree(path->ip_container, *dl);
1014 iam_path_fini(path);
1019 * Performs tree top-to-bottom traversal starting from root, and loads leaf
1022 static int iam_path_lookup(struct iam_path *path, int index)
1024 struct iam_leaf *leaf;
1027 leaf = &path->ip_leaf;
1028 result = iam_lookup_lock(path, &leaf->il_lock, DLT_WRITE);
1029 assert_inv(iam_path_check(path));
1030 do_corr(schedule());
1032 result = iam_leaf_load(path);
1034 do_corr(schedule());
1036 result = iam_leaf_ops(leaf)->
1037 ilookup(leaf, path->ip_ikey_target);
1039 result = iam_leaf_ops(leaf)->
1040 lookup(leaf, path->ip_key_target);
1041 do_corr(schedule());
1044 iam_leaf_unlock(leaf);
1050 * Common part of iam_it_{i,}get().
1052 static int __iam_it_get(struct iam_iterator *it, int index)
1055 assert_corr(it_state(it) == IAM_IT_DETACHED);
1057 result = iam_path_lookup(&it->ii_path, index);
1061 collision = result & IAM_LOOKUP_LAST;
1062 switch (result & ~IAM_LOOKUP_LAST) {
1063 case IAM_LOOKUP_EXACT:
1065 it->ii_state = IAM_IT_ATTACHED;
1069 it->ii_state = IAM_IT_ATTACHED;
1071 case IAM_LOOKUP_BEFORE:
1072 case IAM_LOOKUP_EMPTY:
1074 it->ii_state = IAM_IT_SKEWED;
1079 result |= collision;
1082 * See iam_it_get_exact() for explanation.
1084 assert_corr(result != -ENOENT);
1089 * Correct hash, but not the same key was found, iterate through hash
1090 * collision chain, looking for correct record.
1092 static int iam_it_collision(struct iam_iterator *it)
1096 assert(ergo(it_at_rec(it), !it_keyeq(it, it->ii_path.ip_key_target)));
1098 while ((result = iam_it_next(it)) == 0) {
1099 do_corr(schedule());
1100 if (it_ikeycmp(it, it->ii_path.ip_ikey_target) != 0)
1102 if (it_keyeq(it, it->ii_path.ip_key_target))
1109 * Attach iterator. After successful completion, @it points to record with
1110 * least key not larger than @k.
1112 * Return value: 0: positioned on existing record,
1113 * +ve: exact position found,
1116 * precondition: it_state(it) == IAM_IT_DETACHED
1117 * postcondition: ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1118 * it_keycmp(it, k) <= 0)
1120 int iam_it_get(struct iam_iterator *it, const struct iam_key *k)
1123 assert_corr(it_state(it) == IAM_IT_DETACHED);
1125 it->ii_path.ip_ikey_target = NULL;
1126 it->ii_path.ip_key_target = k;
1128 result = __iam_it_get(it, 0);
1130 if (result == IAM_LOOKUP_LAST) {
1131 result = iam_it_collision(it);
1135 result = __iam_it_get(it, 0);
1140 result &= ~IAM_LOOKUP_LAST;
1142 assert_corr(ergo(result > 0, it_keycmp(it, k) == 0));
1143 assert_corr(ergo(result == 0 && it_state(it) == IAM_IT_ATTACHED,
1144 it_keycmp(it, k) <= 0));
1149 * Attach iterator by index key.
1151 static int iam_it_iget(struct iam_iterator *it, const struct iam_ikey *k)
1153 assert_corr(it_state(it) == IAM_IT_DETACHED);
1155 it->ii_path.ip_ikey_target = k;
1156 return __iam_it_get(it, 1) & ~IAM_LOOKUP_LAST;
1160 * Attach iterator, and assure it points to the record (not skewed).
1162 * Return value: 0: positioned on existing record,
1163 * +ve: exact position found,
1166 * precondition: it_state(it) == IAM_IT_DETACHED &&
1167 * !(it->ii_flags&IAM_IT_WRITE)
1168 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED)
1170 int iam_it_get_at(struct iam_iterator *it, const struct iam_key *k)
1173 assert_corr(it_state(it) == IAM_IT_DETACHED &&
1174 !(it->ii_flags&IAM_IT_WRITE));
1175 result = iam_it_get(it, k);
1177 if (it_state(it) != IAM_IT_ATTACHED) {
1178 assert_corr(it_state(it) == IAM_IT_SKEWED);
1179 result = iam_it_next(it);
1182 assert_corr(ergo(result >= 0, it_state(it) == IAM_IT_ATTACHED));
1187 * Duplicates iterator.
1189 * postcondition: it_state(dst) == it_state(src) &&
1190 * iam_it_container(dst) == iam_it_container(src) &&
1191 * dst->ii_flags = src->ii_flags &&
1192 * ergo(it_state(src) == IAM_IT_ATTACHED,
1193 * iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1194 * iam_it_key_get(dst) == iam_it_key_get(src))
1196 void iam_it_dup(struct iam_iterator *dst, const struct iam_iterator *src)
1198 dst->ii_flags = src->ii_flags;
1199 dst->ii_state = src->ii_state;
1200 /* XXX not yet. iam_path_dup(&dst->ii_path, &src->ii_path); */
1202 * XXX: duplicate lock.
1204 assert_corr(it_state(dst) == it_state(src));
1205 assert_corr(iam_it_container(dst) == iam_it_container(src));
1206 assert_corr(dst->ii_flags = src->ii_flags);
1207 assert_corr(ergo(it_state(src) == IAM_IT_ATTACHED,
1208 iam_it_rec_get(dst) == iam_it_rec_get(src) &&
1209 iam_it_key_get(dst) == iam_it_key_get(src)));
1214 * Detach iterator. Does nothing it detached state.
1216 * postcondition: it_state(it) == IAM_IT_DETACHED
1218 void iam_it_put(struct iam_iterator *it)
1220 if (it->ii_state != IAM_IT_DETACHED) {
1221 it->ii_state = IAM_IT_DETACHED;
1222 iam_leaf_fini(&it->ii_path.ip_leaf);
1226 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1227 struct iam_ikey *ikey);
1231 * This function increments the frame pointer to search the next leaf
1232 * block, and reads in the necessary intervening nodes if the search
1233 * should be necessary. Whether or not the search is necessary is
1234 * controlled by the hash parameter. If the hash value is even, then
1235 * the search is only continued if the next block starts with that
1236 * hash value. This is used if we are searching for a specific file.
1238 * If the hash value is HASH_NB_ALWAYS, then always go to the next block.
1240 * This function returns 1 if the caller should continue to search,
1241 * or 0 if it should not. If there is an error reading one of the
1242 * index blocks, it will a negative error code.
1244 * If start_hash is non-null, it will be filled in with the starting
1245 * hash of the next page.
1247 static int iam_htree_advance(struct inode *dir, __u32 hash,
1248 struct iam_path *path, __u32 *start_hash,
1251 struct iam_frame *p;
1252 struct buffer_head *bh;
1253 int err, num_frames = 0;
1258 * Find the next leaf page by incrementing the frame pointer.
1259 * If we run out of entries in the interior node, loop around and
1260 * increment pointer in the parent node. When we break out of
1261 * this loop, num_frames indicates the number of interior
1262 * nodes need to be read.
1265 do_corr(schedule());
1270 p->at = iam_entry_shift(path, p->at, +1);
1271 if (p->at < iam_entry_shift(path, p->entries,
1272 dx_get_count(p->entries))) {
1273 p->leaf = dx_get_block(path, p->at);
1274 iam_unlock_bh(p->bh);
1277 iam_unlock_bh(p->bh);
1278 if (p == path->ip_frames)
1289 * If the hash is 1, then continue only if the next page has a
1290 * continuation hash of any value. This is used for readdir
1291 * handling. Otherwise, check to see if the hash matches the
1292 * desired contiuation hash. If it doesn't, return since
1293 * there's no point to read in the successive index pages.
1295 dx_get_ikey(path, p->at, (struct iam_ikey *)&bhash);
1297 *start_hash = bhash;
1298 if ((hash & 1) == 0) {
1299 if ((bhash & ~1) != hash)
1304 * If the hash is HASH_NB_ALWAYS, we always go to the next
1305 * block so no check is necessary
1307 while (num_frames--) {
1310 do_corr(schedule());
1312 idx = p->leaf = dx_get_block(path, p->at);
1313 iam_unlock_bh(p->bh);
1314 err = iam_path_descr(path)->id_ops->
1315 id_node_read(path->ip_container, idx, NULL, &bh);
1317 return err; /* Failure */
1320 assert_corr(p->bh != bh);
1322 p->entries = dx_node_get_entries(path, p);
1323 p->at = iam_entry_shift(path, p->entries, !compat);
1324 assert_corr(p->curidx != idx);
1327 assert_corr(p->leaf != dx_get_block(path, p->at));
1328 p->leaf = dx_get_block(path, p->at);
1329 iam_unlock_bh(p->bh);
1330 assert_inv(dx_node_check(path, p));
1336 static inline int iam_index_advance(struct iam_path *path)
1338 return iam_htree_advance(iam_path_obj(path), 0, path, NULL, 0);
1341 static void iam_unlock_array(struct iam_container *ic,
1342 struct dynlock_handle **lh)
1346 for (i = 0; i < DX_MAX_TREE_HEIGHT; ++i, ++lh) {
1348 iam_unlock_htree(ic, *lh);
1354 * Advance index part of @path to point to the next leaf. Returns 1 on
1355 * success, 0, when end of container was reached. Leaf node is locked.
1357 int iam_index_next(struct iam_container *c, struct iam_path *path)
1360 struct dynlock_handle *lh[DX_MAX_TREE_HEIGHT] = { NULL, };
1364 * Locking for iam_index_next()... is to be described.
1367 cursor = path->ip_frame->leaf;
1370 result = iam_index_lock(path, lh);
1371 do_corr(schedule());
1375 result = iam_check_full_path(path, 0);
1376 if (result == 0 && cursor == path->ip_frame->leaf) {
1377 result = iam_index_advance(path);
1379 assert_corr(result == 0 ||
1380 cursor != path->ip_frame->leaf);
1384 iam_unlock_array(c, lh);
1386 iam_path_release(path);
1387 do_corr(schedule());
1389 result = __iam_path_lookup(path);
1393 while (path->ip_frame->leaf != cursor) {
1394 do_corr(schedule());
1396 result = iam_index_lock(path, lh);
1397 do_corr(schedule());
1401 result = iam_check_full_path(path, 0);
1405 result = iam_index_advance(path);
1407 CERROR("cannot find cursor : %u\n",
1413 result = iam_check_full_path(path, 0);
1416 iam_unlock_array(c, lh);
1418 } while (result == -EAGAIN);
1422 iam_unlock_array(c, lh);
1427 * Move iterator one record right.
1429 * Return value: 0: success,
1430 * +1: end of container reached
1433 * precondition: (it_state(it) == IAM_IT_ATTACHED ||
1434 * it_state(it) == IAM_IT_SKEWED) && it->ii_flags&IAM_IT_MOVE
1435 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED) &&
1436 * ergo(result > 0, it_state(it) == IAM_IT_DETACHED)
1438 int iam_it_next(struct iam_iterator *it)
1441 struct iam_path *path;
1442 struct iam_leaf *leaf;
1443 do_corr(struct iam_ikey *ik_orig);
1445 /* assert_corr(it->ii_flags&IAM_IT_MOVE); */
1446 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1447 it_state(it) == IAM_IT_SKEWED);
1449 path = &it->ii_path;
1450 leaf = &path->ip_leaf;
1452 assert_corr(iam_leaf_is_locked(leaf));
1455 do_corr(ik_orig = it_at_rec(it) ?
1456 iam_it_ikey_get(it, iam_path_ikey(path, 2)) : NULL);
1457 if (it_before(it)) {
1458 assert_corr(!iam_leaf_at_end(leaf));
1459 it->ii_state = IAM_IT_ATTACHED;
1461 if (!iam_leaf_at_end(leaf))
1462 /* advance within leaf node */
1463 iam_leaf_next(leaf);
1465 * multiple iterations may be necessary due to empty leaves.
1467 while (result == 0 && iam_leaf_at_end(leaf)) {
1468 do_corr(schedule());
1469 /* advance index portion of the path */
1470 result = iam_index_next(iam_it_container(it), path);
1471 assert_corr(iam_leaf_is_locked(leaf));
1473 struct dynlock_handle *lh;
1474 lh = iam_lock_htree(iam_it_container(it),
1475 path->ip_frame->leaf,
1478 iam_leaf_fini(leaf);
1480 result = iam_leaf_load(path);
1482 iam_leaf_start(leaf);
1485 } else if (result == 0)
1486 /* end of container reached */
1492 it->ii_state = IAM_IT_ATTACHED;
1494 assert_corr(ergo(result == 0, it_state(it) == IAM_IT_ATTACHED));
1495 assert_corr(ergo(result > 0, it_state(it) == IAM_IT_DETACHED));
1496 assert_corr(ergo(result == 0 && ik_orig != NULL,
1497 it_ikeycmp(it, ik_orig) >= 0));
1502 * Return pointer to the record under iterator.
1504 * precondition: it_state(it) == IAM_IT_ATTACHED && it_at_rec(it)
1505 * postcondition: it_state(it) == IAM_IT_ATTACHED
1507 struct iam_rec *iam_it_rec_get(const struct iam_iterator *it)
1509 assert_corr(it_state(it) == IAM_IT_ATTACHED);
1510 assert_corr(it_at_rec(it));
1511 return iam_leaf_rec(&it->ii_path.ip_leaf);
1514 static void iam_it_reccpy(struct iam_iterator *it, const struct iam_rec *r)
1516 struct iam_leaf *folio;
1518 folio = &it->ii_path.ip_leaf;
1519 iam_leaf_ops(folio)->rec_set(folio, r);
1523 * Replace contents of record under iterator.
1525 * precondition: it_state(it) == IAM_IT_ATTACHED &&
1526 * it->ii_flags&IAM_IT_WRITE
1527 * postcondition: it_state(it) == IAM_IT_ATTACHED &&
1528 * ergo(result == 0, !memcmp(iam_it_rec_get(it), r, ...))
1530 int iam_it_rec_set(handle_t *h,
1531 struct iam_iterator *it, const struct iam_rec *r)
1534 struct iam_path *path;
1535 struct buffer_head *bh;
1537 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
1538 it->ii_flags&IAM_IT_WRITE);
1539 assert_corr(it_at_rec(it));
1541 path = &it->ii_path;
1542 bh = path->ip_leaf.il_bh;
1543 result = iam_txn_add(h, path, bh);
1545 iam_it_reccpy(it, r);
1546 result = iam_txn_dirty(h, path, bh);
1552 * Return pointer to the index key under iterator.
1554 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1555 * it_state(it) == IAM_IT_SKEWED
1557 static struct iam_ikey *iam_it_ikey_get(const struct iam_iterator *it,
1558 struct iam_ikey *ikey)
1560 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1561 it_state(it) == IAM_IT_SKEWED);
1562 assert_corr(it_at_rec(it));
1563 return iam_leaf_ikey(&it->ii_path.ip_leaf, ikey);
1567 * Return pointer to the key under iterator.
1569 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1570 * it_state(it) == IAM_IT_SKEWED
1572 struct iam_key *iam_it_key_get(const struct iam_iterator *it)
1574 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1575 it_state(it) == IAM_IT_SKEWED);
1576 assert_corr(it_at_rec(it));
1577 return iam_leaf_key(&it->ii_path.ip_leaf);
1581 * Return size of key under iterator (in bytes)
1583 * precondition: it_state(it) == IAM_IT_ATTACHED ||
1584 * it_state(it) == IAM_IT_SKEWED
1586 int iam_it_key_size(const struct iam_iterator *it)
1588 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
1589 it_state(it) == IAM_IT_SKEWED);
1590 assert_corr(it_at_rec(it));
1591 return iam_leaf_key_size(&it->ii_path.ip_leaf);
1594 static struct buffer_head *
1595 iam_new_node(handle_t *h, struct iam_container *c, iam_ptr_t *b, int *e)
1597 struct inode *inode = c->ic_object;
1598 struct buffer_head *bh = NULL;
1599 struct iam_idle_head *head;
1600 struct buffer_head *idle;
1604 if (c->ic_idle_bh == NULL)
1607 mutex_lock(&c->ic_idle_mutex);
1608 if (unlikely(c->ic_idle_bh == NULL)) {
1609 mutex_unlock(&c->ic_idle_mutex);
1613 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
1614 count = le16_to_cpu(head->iih_count);
1616 *e = ldiskfs_journal_get_write_access(h, c->ic_idle_bh);
1621 *b = le32_to_cpu(head->iih_blks[count]);
1622 head->iih_count = cpu_to_le16(count);
1623 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_idle_bh);
1627 mutex_unlock(&c->ic_idle_mutex);
1628 bh = __ldiskfs_bread(NULL, inode, *b, 0);
1629 if (IS_ERR_OR_NULL(bh)) {
1639 /* The block itself which contains the iam_idle_head is
1640 * also an idle block, and can be used as the new node. */
1641 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
1642 c->ic_descr->id_root_gap +
1643 sizeof(struct dx_countlimit));
1644 *e = ldiskfs_journal_get_write_access(h, c->ic_root_bh);
1648 *b = le32_to_cpu(*idle_blocks);
1649 iam_lock_bh(c->ic_root_bh);
1650 *idle_blocks = head->iih_next;
1651 iam_unlock_bh(c->ic_root_bh);
1652 *e = ldiskfs_handle_dirty_metadata(h, inode, c->ic_root_bh);
1654 iam_lock_bh(c->ic_root_bh);
1655 *idle_blocks = cpu_to_le32(*b);
1656 iam_unlock_bh(c->ic_root_bh);
1661 idle = iam_load_idle_blocks(c, le32_to_cpu(*idle_blocks));
1662 if (idle != NULL && IS_ERR(idle)) {
1664 c->ic_idle_bh = NULL;
1669 c->ic_idle_bh = idle;
1670 mutex_unlock(&c->ic_idle_mutex);
1673 /* get write access for the found buffer head */
1674 *e = ldiskfs_journal_get_write_access(h, bh);
1678 ldiskfs_std_error(inode->i_sb, *e);
1680 /* Clear the reused node as new node does. */
1681 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1682 set_buffer_uptodate(bh);
1687 bh = osd_ldiskfs_append(h, inode, b);
1696 mutex_unlock(&c->ic_idle_mutex);
1697 ldiskfs_std_error(inode->i_sb, *e);
1702 * Insertion of new record. Interaction with jbd during non-trivial case (when
1703 * split happens) is as following:
1705 * - new leaf node is involved into transaction by iam_new_node();
1707 * - old leaf node is involved into transaction by iam_add_rec();
1709 * - leaf where insertion point ends in, is marked dirty by iam_add_rec();
1711 * - leaf without insertion point is marked dirty (as @new_leaf) by
1714 * - split index nodes are involved into transaction and marked dirty by
1715 * split_index_node().
1717 * - "safe" index node, which is no split, but where new pointer is inserted
1718 * is involved into transaction and marked dirty by split_index_node().
1720 * - index node where pointer to new leaf is inserted is involved into
1721 * transaction by split_index_node() and marked dirty by iam_add_rec().
1723 * - inode is marked dirty by iam_add_rec().
1727 static int iam_new_leaf(handle_t *handle, struct iam_leaf *leaf)
1731 struct buffer_head *new_leaf;
1732 struct buffer_head *old_leaf;
1733 struct iam_container *c;
1735 struct iam_path *path;
1737 c = iam_leaf_container(leaf);
1738 path = leaf->il_path;
1741 new_leaf = iam_new_node(handle, c, &blknr, &err);
1742 do_corr(schedule());
1743 if (new_leaf != NULL) {
1744 struct dynlock_handle *lh;
1746 lh = iam_lock_htree(c, blknr, DLT_WRITE);
1747 do_corr(schedule());
1749 iam_leaf_ops(leaf)->init_new(c, new_leaf);
1750 do_corr(schedule());
1751 old_leaf = leaf->il_bh;
1752 iam_leaf_split(leaf, &new_leaf, blknr);
1753 if (old_leaf != leaf->il_bh) {
1755 * Switched to the new leaf.
1757 iam_leaf_unlock(leaf);
1759 path->ip_frame->leaf = blknr;
1761 iam_unlock_htree(path->ip_container, lh);
1762 do_corr(schedule());
1763 err = iam_txn_dirty(handle, path, new_leaf);
1765 err = ldiskfs_mark_inode_dirty(handle, obj);
1766 do_corr(schedule());
1771 assert_inv(iam_path_check(iam_leaf_path(leaf)));
1775 static inline void dx_set_limit(struct iam_entry *entries, unsigned value)
1777 ((struct dx_countlimit *) entries)->limit = cpu_to_le16(value);
1780 static int iam_shift_entries(struct iam_path *path,
1781 struct iam_frame *frame, unsigned count,
1782 struct iam_entry *entries, struct iam_entry *entries2,
1789 struct iam_frame *parent = frame - 1;
1790 struct iam_ikey *pivot = iam_path_ikey(path, 3);
1792 delta = dx_index_is_compat(path) ? 0 : +1;
1794 count1 = count/2 + delta;
1795 count2 = count - count1;
1796 dx_get_ikey(path, iam_entry_shift(path, entries, count1), pivot);
1798 dxtrace(printk("Split index %d/%d\n", count1, count2));
1800 memcpy((char *) iam_entry_shift(path, entries2, delta),
1801 (char *) iam_entry_shift(path, entries, count1),
1802 count2 * iam_entry_size(path));
1804 dx_set_count(entries2, count2 + delta);
1805 dx_set_limit(entries2, dx_node_limit(path));
1808 * NOTE: very subtle piece of code competing dx_probe() may find 2nd
1809 * level index in root index, then we insert new index here and set
1810 * new count in that 2nd level index. so, dx_probe() may see 2nd level
1811 * index w/o hash it looks for. the solution is to check root index
1812 * after we locked just founded 2nd level index -bzzz
1814 iam_insert_key_lock(path, parent, pivot, newblock);
1817 * now old and new 2nd level index blocks contain all pointers, so
1818 * dx_probe() may find it in the both. it's OK -bzzz
1820 iam_lock_bh(frame->bh);
1821 dx_set_count(entries, count1);
1822 iam_unlock_bh(frame->bh);
1825 * now old 2nd level index block points to first half of leafs. it's
1826 * importand that dx_probe() must check root index block for changes
1827 * under dx_lock_bh(frame->bh) -bzzz
1834 int split_index_node(handle_t *handle, struct iam_path *path,
1835 struct dynlock_handle **lh)
1838 struct iam_entry *entries; /* old block contents */
1839 struct iam_entry *entries2; /* new block contents */
1840 struct iam_frame *frame, *safe;
1841 struct buffer_head *bh_new[DX_MAX_TREE_HEIGHT] = {NULL};
1842 u32 newblock[DX_MAX_TREE_HEIGHT] = {0};
1843 struct dynlock_handle *lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1844 struct dynlock_handle *new_lock[DX_MAX_TREE_HEIGHT] = {NULL,};
1845 struct inode *dir = iam_path_obj(path);
1846 struct iam_descr *descr;
1850 descr = iam_path_descr(path);
1852 * Algorithm below depends on this.
1854 assert_corr(dx_root_limit(path) < dx_node_limit(path));
1856 frame = path->ip_frame;
1857 entries = frame->entries;
1860 * Tall-tree handling: we might have to split multiple index blocks
1861 * all the way up to tree root. Tricky point here is error handling:
1862 * to avoid complicated undo/rollback we
1864 * - first allocate all necessary blocks
1866 * - insert pointers into them atomically.
1870 * Locking: leaf is already locked. htree-locks are acquired on all
1871 * index nodes that require split bottom-to-top, on the "safe" node,
1872 * and on all new nodes
1875 dxtrace(printk("using %u of %u node entries\n",
1876 dx_get_count(entries), dx_get_limit(entries)));
1878 /* What levels need split? */
1879 for (nr_splet = 0; frame >= path->ip_frames &&
1880 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1881 --frame, ++nr_splet) {
1882 do_corr(schedule());
1883 if (nr_splet == DX_MAX_TREE_HEIGHT) {
1885 CWARN(dir->i_sb, __FUNCTION__,
1886 "Directory index full!\n");
1896 * Lock all nodes, bottom to top.
1898 for (frame = path->ip_frame, i = nr_splet; i >= 0; --i, --frame) {
1899 do_corr(schedule());
1900 lock[i] = iam_lock_htree(path->ip_container, frame->curidx,
1902 if (lock[i] == NULL) {
1909 * Check for concurrent index modification.
1911 err = iam_check_full_path(path, 1);
1915 * And check that the same number of nodes is to be split.
1917 for (i = 0, frame = path->ip_frame; frame >= path->ip_frames &&
1918 dx_get_count(frame->entries) == dx_get_limit(frame->entries);
1922 if (i != nr_splet) {
1927 /* Go back down, allocating blocks, locking them, and adding into
1929 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
1930 bh_new[i] = iam_new_node(handle, path->ip_container,
1931 &newblock[i], &err);
1932 do_corr(schedule());
1934 descr->id_ops->id_node_init(path->ip_container,
1937 new_lock[i] = iam_lock_htree(path->ip_container, newblock[i],
1939 if (new_lock[i] == NULL) {
1943 do_corr(schedule());
1944 BUFFER_TRACE(frame->bh, "get_write_access");
1945 err = ldiskfs_journal_get_write_access(handle, frame->bh);
1949 /* Add "safe" node to transaction too */
1950 if (safe + 1 != path->ip_frames) {
1951 do_corr(schedule());
1952 err = ldiskfs_journal_get_write_access(handle, safe->bh);
1957 /* Go through nodes once more, inserting pointers */
1958 for (frame = safe + 1, i = 0; i < nr_splet; ++i, ++frame) {
1961 struct buffer_head *bh2;
1962 struct buffer_head *bh;
1964 entries = frame->entries;
1965 count = dx_get_count(entries);
1966 idx = iam_entry_diff(path, frame->at, entries);
1969 entries2 = dx_get_entries(path, bh2->b_data, 0);
1972 if (frame == path->ip_frames) {
1973 /* splitting root node. Tricky point:
1975 * In the "normal" B-tree we'd split root *and* add
1976 * new root to the tree with pointers to the old root
1977 * and its sibling (thus introducing two new nodes).
1979 * In htree it's enough to add one node, because
1980 * capacity of the root node is smaller than that of
1983 struct iam_frame *frames;
1984 struct iam_entry *next;
1986 assert_corr(i == 0);
1988 do_corr(schedule());
1990 frames = path->ip_frames;
1991 memcpy((char *) entries2, (char *) entries,
1992 count * iam_entry_size(path));
1993 dx_set_limit(entries2, dx_node_limit(path));
1996 iam_lock_bh(frame->bh);
1997 next = descr->id_ops->id_root_inc(path->ip_container,
1999 dx_set_block(path, next, newblock[0]);
2000 iam_unlock_bh(frame->bh);
2002 do_corr(schedule());
2003 /* Shift frames in the path */
2004 memmove(frames + 2, frames + 1,
2005 (sizeof path->ip_frames) - 2 * sizeof frames[0]);
2006 /* Add new access path frame */
2007 frames[1].at = iam_entry_shift(path, entries2, idx);
2008 frames[1].entries = entries = entries2;
2010 assert_inv(dx_node_check(path, frame));
2013 assert_inv(dx_node_check(path, frame));
2014 bh_new[0] = NULL; /* buffer head is "consumed" */
2015 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2018 do_corr(schedule());
2020 /* splitting non-root index node. */
2021 struct iam_frame *parent = frame - 1;
2023 do_corr(schedule());
2024 count = iam_shift_entries(path, frame, count,
2025 entries, entries2, newblock[i]);
2026 /* Which index block gets the new entry? */
2028 int d = dx_index_is_compat(path) ? 0 : +1;
2030 frame->at = iam_entry_shift(path, entries2,
2032 frame->entries = entries = entries2;
2033 frame->curidx = newblock[i];
2034 swap(frame->bh, bh2);
2035 assert_corr(lock[i + 1] != NULL);
2036 assert_corr(new_lock[i] != NULL);
2037 swap(lock[i + 1], new_lock[i]);
2039 parent->at = iam_entry_shift(path,
2042 assert_inv(dx_node_check(path, frame));
2043 assert_inv(dx_node_check(path, parent));
2044 dxtrace(dx_show_index ("node", frame->entries));
2045 dxtrace(dx_show_index ("node",
2046 ((struct dx_node *) bh2->b_data)->entries));
2047 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh2);
2050 do_corr(schedule());
2051 err = ldiskfs_handle_dirty_metadata(handle, NULL,
2056 do_corr(schedule());
2057 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2062 * This function was called to make insertion of new leaf
2063 * possible. Check that it fulfilled its obligations.
2065 assert_corr(dx_get_count(path->ip_frame->entries) <
2066 dx_get_limit(path->ip_frame->entries));
2067 assert_corr(lock[nr_splet] != NULL);
2068 *lh = lock[nr_splet];
2069 lock[nr_splet] = NULL;
2072 * Log ->i_size modification.
2074 err = ldiskfs_mark_inode_dirty(handle, dir);
2080 ldiskfs_std_error(dir->i_sb, err);
2083 iam_unlock_array(path->ip_container, lock);
2084 iam_unlock_array(path->ip_container, new_lock);
2086 assert_corr(err || iam_frame_is_locked(path, path->ip_frame));
2088 do_corr(schedule());
2089 for (i = 0; i < ARRAY_SIZE(bh_new); ++i) {
2090 if (bh_new[i] != NULL)
2096 static int iam_add_rec(handle_t *handle, struct iam_iterator *it,
2097 struct iam_path *path,
2098 const struct iam_key *k, const struct iam_rec *r)
2101 struct iam_leaf *leaf;
2103 leaf = &path->ip_leaf;
2104 assert_inv(iam_path_check(path));
2105 err = iam_txn_add(handle, path, leaf->il_bh);
2107 do_corr(schedule());
2108 if (!iam_leaf_can_add(leaf, k, r)) {
2109 struct dynlock_handle *lh = NULL;
2112 assert_corr(lh == NULL);
2113 do_corr(schedule());
2114 err = split_index_node(handle, path, &lh);
2115 if (err == -EAGAIN) {
2116 assert_corr(lh == NULL);
2118 iam_path_fini(path);
2119 it->ii_state = IAM_IT_DETACHED;
2121 do_corr(schedule());
2122 err = iam_it_get_exact(it, k);
2124 err = +1; /* repeat split */
2129 assert_inv(iam_path_check(path));
2131 assert_corr(lh != NULL);
2132 do_corr(schedule());
2133 err = iam_new_leaf(handle, leaf);
2135 err = iam_txn_dirty(handle, path,
2136 path->ip_frame->bh);
2138 iam_unlock_htree(path->ip_container, lh);
2139 do_corr(schedule());
2142 iam_leaf_rec_add(leaf, k, r);
2143 err = iam_txn_dirty(handle, path, leaf->il_bh);
2146 assert_inv(iam_path_check(path));
2151 * Insert new record with key @k and contents from @r, shifting records to the
2152 * right. On success, iterator is positioned on the newly inserted record.
2154 * precondition: it->ii_flags&IAM_IT_WRITE &&
2155 * (it_state(it) == IAM_IT_ATTACHED ||
2156 * it_state(it) == IAM_IT_SKEWED) &&
2157 * ergo(it_state(it) == IAM_IT_ATTACHED,
2158 * it_keycmp(it, k) <= 0) &&
2159 * ergo(it_before(it), it_keycmp(it, k) > 0));
2160 * postcondition: ergo(result == 0,
2161 * it_state(it) == IAM_IT_ATTACHED &&
2162 * it_keycmp(it, k) == 0 &&
2163 * !memcmp(iam_it_rec_get(it), r, ...))
2165 int iam_it_rec_insert(handle_t *h, struct iam_iterator *it,
2166 const struct iam_key *k, const struct iam_rec *r)
2169 struct iam_path *path;
2171 path = &it->ii_path;
2173 assert_corr(it->ii_flags&IAM_IT_WRITE);
2174 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2175 it_state(it) == IAM_IT_SKEWED);
2176 assert_corr(ergo(it_state(it) == IAM_IT_ATTACHED,
2177 it_keycmp(it, k) <= 0));
2178 assert_corr(ergo(it_before(it), it_keycmp(it, k) > 0));
2179 result = iam_add_rec(h, it, path, k, r);
2181 it->ii_state = IAM_IT_ATTACHED;
2182 assert_corr(ergo(result == 0,
2183 it_state(it) == IAM_IT_ATTACHED &&
2184 it_keycmp(it, k) == 0));
2188 static inline int iam_idle_blocks_limit(struct inode *inode)
2190 return (inode->i_sb->s_blocksize - sizeof(struct iam_idle_head)) >> 2;
2194 * If the leaf cannnot be recycled, we will lose one block for reusing.
2195 * It is not a serious issue because it almost the same of non-recycle.
2197 static iam_ptr_t iam_index_shrink(handle_t *h, struct iam_path *p,
2198 struct iam_leaf *l, struct buffer_head **bh)
2200 struct iam_container *c = p->ip_container;
2201 struct inode *inode = c->ic_object;
2202 struct iam_frame *frame = p->ip_frame;
2203 struct iam_entry *entries;
2204 struct iam_entry *pos;
2205 struct dynlock_handle *lh;
2209 if (c->ic_idle_failed)
2212 if (unlikely(frame == NULL))
2215 if (!iam_leaf_empty(l))
2218 lh = iam_lock_htree(c, frame->curidx, DLT_WRITE);
2220 CWARN("%s: No memory to recycle idle blocks\n",
2221 osd_ino2name(inode));
2225 rc = iam_txn_add(h, p, frame->bh);
2227 iam_unlock_htree(c, lh);
2231 iam_lock_bh(frame->bh);
2232 entries = frame->entries;
2233 count = dx_get_count(entries);
2234 /* NOT shrink the last entry in the index node, which can be reused
2235 * directly by next new node. */
2237 iam_unlock_bh(frame->bh);
2238 iam_unlock_htree(c, lh);
2242 pos = iam_find_position(p, frame);
2243 /* There may be some new leaf nodes have been added or empty leaf nodes
2244 * have been shrinked during my delete operation.
2246 * If the empty leaf is not under current index node because the index
2247 * node has been split, then just skip the empty leaf, which is rare. */
2248 if (unlikely(frame->leaf != dx_get_block(p, pos))) {
2249 iam_unlock_bh(frame->bh);
2250 iam_unlock_htree(c, lh);
2255 if (frame->at < iam_entry_shift(p, entries, count - 1)) {
2256 struct iam_entry *n = iam_entry_shift(p, frame->at, 1);
2258 memmove(frame->at, n,
2259 (char *)iam_entry_shift(p, entries, count) - (char *)n);
2260 frame->at_shifted = 1;
2262 dx_set_count(entries, count - 1);
2263 iam_unlock_bh(frame->bh);
2264 rc = iam_txn_dirty(h, p, frame->bh);
2265 iam_unlock_htree(c, lh);
2275 iam_install_idle_blocks(handle_t *h, struct iam_path *p, struct buffer_head *bh,
2276 __u32 *idle_blocks, iam_ptr_t blk)
2278 struct iam_container *c = p->ip_container;
2279 struct buffer_head *old = c->ic_idle_bh;
2280 struct iam_idle_head *head;
2283 head = (struct iam_idle_head *)(bh->b_data);
2284 head->iih_magic = cpu_to_le16(IAM_IDLE_HEADER_MAGIC);
2285 head->iih_count = 0;
2286 head->iih_next = *idle_blocks;
2287 /* The bh already get_write_accessed. */
2288 rc = iam_txn_dirty(h, p, bh);
2292 rc = iam_txn_add(h, p, c->ic_root_bh);
2296 iam_lock_bh(c->ic_root_bh);
2297 *idle_blocks = cpu_to_le32(blk);
2298 iam_unlock_bh(c->ic_root_bh);
2299 rc = iam_txn_dirty(h, p, c->ic_root_bh);
2301 /* NOT release old before new assigned. */
2306 iam_lock_bh(c->ic_root_bh);
2307 *idle_blocks = head->iih_next;
2308 iam_unlock_bh(c->ic_root_bh);
2314 * If the leaf cannnot be recycled, we will lose one block for reusing.
2315 * It is not a serious issue because it almost the same of non-recycle.
2317 static void iam_recycle_leaf(handle_t *h, struct iam_path *p,
2318 struct buffer_head *bh, iam_ptr_t blk)
2320 struct iam_container *c = p->ip_container;
2321 struct inode *inode = c->ic_object;
2322 struct iam_idle_head *head;
2327 mutex_lock(&c->ic_idle_mutex);
2328 if (unlikely(c->ic_idle_failed)) {
2333 idle_blocks = (__u32 *)(c->ic_root_bh->b_data +
2334 c->ic_descr->id_root_gap +
2335 sizeof(struct dx_countlimit));
2336 /* It is the first idle block. */
2337 if (c->ic_idle_bh == NULL) {
2338 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2342 head = (struct iam_idle_head *)(c->ic_idle_bh->b_data);
2343 count = le16_to_cpu(head->iih_count);
2344 /* Current ic_idle_bh is full, to be replaced by the leaf. */
2345 if (count == iam_idle_blocks_limit(inode)) {
2346 rc = iam_install_idle_blocks(h, p, bh, idle_blocks, blk);
2350 /* Just add to ic_idle_bh. */
2351 rc = iam_txn_add(h, p, c->ic_idle_bh);
2355 head->iih_blks[count] = cpu_to_le32(blk);
2356 head->iih_count = cpu_to_le16(count + 1);
2357 rc = iam_txn_dirty(h, p, c->ic_idle_bh);
2360 mutex_unlock(&c->ic_idle_mutex);
2362 CWARN("%s: idle blocks failed, will lose the blk %u\n",
2363 osd_ino2name(inode), blk);
2367 * Delete record under iterator.
2369 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2370 * it->ii_flags&IAM_IT_WRITE &&
2372 * postcondition: it_state(it) == IAM_IT_ATTACHED ||
2373 * it_state(it) == IAM_IT_DETACHED
2375 int iam_it_rec_delete(handle_t *h, struct iam_iterator *it)
2378 struct iam_leaf *leaf;
2379 struct iam_path *path;
2381 assert_corr(it_state(it) == IAM_IT_ATTACHED &&
2382 it->ii_flags&IAM_IT_WRITE);
2383 assert_corr(it_at_rec(it));
2385 path = &it->ii_path;
2386 leaf = &path->ip_leaf;
2388 assert_inv(iam_path_check(path));
2390 result = iam_txn_add(h, path, leaf->il_bh);
2392 * no compaction for now.
2395 iam_rec_del(leaf, it->ii_flags&IAM_IT_MOVE);
2396 result = iam_txn_dirty(h, path, leaf->il_bh);
2397 if (result == 0 && iam_leaf_at_end(leaf)) {
2398 struct buffer_head *bh = NULL;
2401 blk = iam_index_shrink(h, path, leaf, &bh);
2402 if (it->ii_flags & IAM_IT_MOVE) {
2403 result = iam_it_next(it);
2409 iam_recycle_leaf(h, path, bh, blk);
2414 assert_inv(iam_path_check(path));
2415 assert_corr(it_state(it) == IAM_IT_ATTACHED ||
2416 it_state(it) == IAM_IT_DETACHED);
2421 * Convert iterator to cookie.
2423 * precondition: it_state(it) == IAM_IT_ATTACHED &&
2424 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2425 * postcondition: it_state(it) == IAM_IT_ATTACHED
2427 iam_pos_t iam_it_store(const struct iam_iterator *it)
2431 assert_corr(it_state(it) == IAM_IT_ATTACHED);
2432 assert_corr(it_at_rec(it));
2433 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <=
2437 return *(iam_pos_t *)iam_it_ikey_get(it, (void *)&result);
2441 * Restore iterator from cookie.
2443 * precondition: it_state(it) == IAM_IT_DETACHED && it->ii_flags&IAM_IT_MOVE &&
2444 * iam_path_descr(it->ii_path)->id_key_size <= sizeof(iam_pos_t)
2445 * postcondition: ergo(result == 0, it_state(it) == IAM_IT_ATTACHED &&
2446 * iam_it_store(it) == pos)
2448 int iam_it_load(struct iam_iterator *it, iam_pos_t pos)
2450 assert_corr(it_state(it) == IAM_IT_DETACHED &&
2451 it->ii_flags&IAM_IT_MOVE);
2452 assert_corr(iam_it_container(it)->ic_descr->id_ikey_size <= sizeof pos);
2453 return iam_it_iget(it, (struct iam_ikey *)&pos);
2456 /***********************************************************************/
2458 /***********************************************************************/
2460 static inline int ptr_inside(void *base, size_t size, void *ptr)
2462 return (base <= ptr) && (ptr < base + size);
2465 static int iam_frame_invariant(struct iam_frame *f)
2469 f->bh->b_data != NULL &&
2470 ptr_inside(f->bh->b_data, f->bh->b_size, f->entries) &&
2471 ptr_inside(f->bh->b_data, f->bh->b_size, f->at) &&
2472 f->entries <= f->at);
2475 static int iam_leaf_invariant(struct iam_leaf *l)
2479 l->il_bh->b_data != NULL &&
2480 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_entries) &&
2481 ptr_inside(l->il_bh->b_data, l->il_bh->b_size, l->il_at) &&
2482 l->il_entries <= l->il_at;
2485 static int iam_path_invariant(struct iam_path *p)
2489 if (p->ip_container == NULL ||
2490 p->ip_indirect < 0 || p->ip_indirect > DX_MAX_TREE_HEIGHT - 1 ||
2491 p->ip_frame != p->ip_frames + p->ip_indirect ||
2492 !iam_leaf_invariant(&p->ip_leaf))
2494 for (i = 0; i < ARRAY_SIZE(p->ip_frames); ++i) {
2495 if (i <= p->ip_indirect) {
2496 if (!iam_frame_invariant(&p->ip_frames[i]))
2503 int iam_it_invariant(struct iam_iterator *it)
2506 (it->ii_state == IAM_IT_DETACHED ||
2507 it->ii_state == IAM_IT_ATTACHED ||
2508 it->ii_state == IAM_IT_SKEWED) &&
2509 !(it->ii_flags & ~(IAM_IT_MOVE | IAM_IT_WRITE)) &&
2510 ergo(it->ii_state == IAM_IT_ATTACHED ||
2511 it->ii_state == IAM_IT_SKEWED,
2512 iam_path_invariant(&it->ii_path) &&
2513 equi(it_at_rec(it), it->ii_state == IAM_IT_SKEWED));
2517 * Search container @c for record with key @k. If record is found, its data
2518 * are moved into @r.
2520 * Return values: 0: found, -ENOENT: not-found, -ve: error
2522 int iam_lookup(struct iam_container *c, const struct iam_key *k,
2523 struct iam_rec *r, struct iam_path_descr *pd)
2525 struct iam_iterator it;
2528 iam_it_init(&it, c, 0, pd);
2530 result = iam_it_get_exact(&it, k);
2533 * record with required key found, copy it into user buffer
2535 iam_reccpy(&it.ii_path.ip_leaf, r);
2542 * Insert new record @r with key @k into container @c (within context of
2545 * Return values: 0: success, -ve: error, including -EEXIST when record with
2546 * given key is already present.
2548 * postcondition: ergo(result == 0 || result == -EEXIST,
2549 * iam_lookup(c, k, r2) > 0;
2551 int iam_insert(handle_t *h, struct iam_container *c, const struct iam_key *k,
2552 const struct iam_rec *r, struct iam_path_descr *pd)
2554 struct iam_iterator it;
2557 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2559 result = iam_it_get_exact(&it, k);
2560 if (result == -ENOENT)
2561 result = iam_it_rec_insert(h, &it, k, r);
2562 else if (result == 0)
2570 * Update record with the key @k in container @c (within context of
2571 * transaction @h), new record is given by @r.
2573 * Return values: +1: skip because of the same rec value, 0: success,
2574 * -ve: error, including -ENOENT if no record with the given key found.
2576 int iam_update(handle_t *h, struct iam_container *c, const struct iam_key *k,
2577 const struct iam_rec *r, struct iam_path_descr *pd)
2579 struct iam_iterator it;
2580 struct iam_leaf *folio;
2583 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2585 result = iam_it_get_exact(&it, k);
2587 folio = &it.ii_path.ip_leaf;
2588 result = iam_leaf_ops(folio)->rec_eq(folio, r);
2590 iam_it_rec_set(h, &it, r);
2600 * Delete existing record with key @k.
2602 * Return values: 0: success, -ENOENT: not-found, -ve: other error.
2604 * postcondition: ergo(result == 0 || result == -ENOENT,
2605 * !iam_lookup(c, k, *));
2607 int iam_delete(handle_t *h, struct iam_container *c, const struct iam_key *k,
2608 struct iam_path_descr *pd)
2610 struct iam_iterator it;
2613 iam_it_init(&it, c, IAM_IT_WRITE, pd);
2615 result = iam_it_get_exact(&it, k);
2617 iam_it_rec_delete(h, &it);
2623 int iam_root_limit(int rootgap, int blocksize, int size)
2628 limit = (blocksize - rootgap) / size;
2629 nlimit = blocksize / size;
2630 if (limit == nlimit)