2 * Copyright (c) 2007 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.27 2008/02/04 08:33:17 dillon Exp $
40 * HAMMER implements a modified B+Tree. In documentation this will
41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
42 * looks like a B+Tree (A B-Tree which stores its records only at the leafs
43 * of the tree), but adds two additional boundary elements which describe
44 * the left-most and right-most element a node is able to represent. In
45 * otherwords, we have boundary elements at the two ends of a B-Tree node
46 * instead of sub-tree pointers.
48 * A B-Tree internal node looks like this:
50 * B N N N N N N B <-- boundary and internal elements
51 * S S S S S S S <-- subtree pointers
53 * A B-Tree leaf node basically looks like this:
55 * L L L L L L L L <-- leaf elemenets
57 * The radix for an internal node is 1 less then a leaf but we get a
58 * number of significant benefits for our troubles.
60 * The big benefit to using a B-Tree containing boundary information
61 * is that it is possible to cache pointers into the middle of the tree
62 * and not have to start searches, insertions, OR deletions at the root
63 * node. In particular, searches are able to progress in a definitive
64 * direction from any point in the tree without revisting nodes. This
65 * greatly improves the efficiency of many operations, most especially
68 * B-Trees also make the stacking of trees fairly straightforward.
70 * SPIKES: Two leaf elements denoting a sub-range of keys may represent
71 * a spike, or a recursion into another cluster. Most standard B-Tree
72 * searches traverse spikes. The ending spike element is range-inclusive
73 * and does not operate quite like a right-bound.
75 * INSERTIONS: A search performed with the intention of doing
76 * an insert will guarantee that the terminal leaf node is not full by
77 * splitting full nodes. Splits occur top-down during the dive down the
80 * DELETIONS: A deletion makes no attempt to proactively balance the
81 * tree and will recursively remove nodes that become empty. Empty
82 * nodes are not allowed and a deletion may recurse upwards from the leaf.
83 * Rather then allow a deadlock a deletion may terminate early by setting
84 * an internal node's element's subtree_offset to 0. The deletion will
85 * then be resumed the next time a search encounters the element.
91 static int btree_search(hammer_cursor_t cursor, int flags);
92 static int btree_split_internal(hammer_cursor_t cursor);
93 static int btree_split_leaf(hammer_cursor_t cursor);
94 static int btree_remove(hammer_cursor_t cursor);
95 static int btree_remove_deleted_element(hammer_cursor_t cursor);
96 static int btree_set_parent(hammer_node_t node, hammer_btree_elm_t elm);
97 static int btree_node_is_almost_full(hammer_node_ondisk_t node);
98 static int btree_node_is_full(hammer_node_ondisk_t node);
99 static void hammer_make_separator(hammer_base_elm_t key1,
100 hammer_base_elm_t key2, hammer_base_elm_t dest);
103 * Iterate records after a search. The cursor is iterated forwards past
104 * the current record until a record matching the key-range requirements
105 * is found. ENOENT is returned if the iteration goes past the ending
108 * The iteration is inclusive of key_beg and can be inclusive or exclusive
109 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
111 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
112 * may be modified by B-Tree functions.
114 * cursor->key_beg may or may not be modified by this function during
115 * the iteration. XXX future - in case of an inverted lock we may have
116 * to reinitiate the lookup and set key_beg to properly pick up where we
119 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
122 hammer_btree_iterate(hammer_cursor_t cursor)
124 hammer_node_ondisk_t node;
125 hammer_btree_elm_t elm;
131 * Skip past the current record
133 node = cursor->node->ondisk;
136 if (cursor->index < node->count &&
137 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
142 * Loop until an element is found or we are done.
146 * We iterate up the tree and then index over one element
147 * while we are at the last element in the current node.
149 * NOTE: This can pop us up to another cluster.
151 * If we are at the root of the root cluster, cursor_up
154 * NOTE: hammer_cursor_up() will adjust cursor->key_beg
155 * when told to re-search for the cluster tag.
157 * XXX this could be optimized by storing the information in
158 * the parent reference.
160 * XXX we can lose the node lock temporarily, this could mess
163 if (cursor->index == node->count) {
164 error = hammer_cursor_up(cursor);
167 /* reload stale pointer */
168 node = cursor->node->ondisk;
169 KKASSERT(cursor->index != node->count);
175 * Check internal or leaf element. Determine if the record
176 * at the cursor has gone beyond the end of our range.
178 * Generally we recurse down through internal nodes. An
179 * internal node can only be returned if INCLUSTER is set
180 * and the node represents a cluster-push record.
182 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
183 elm = &node->elms[cursor->index];
184 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
185 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
186 if (hammer_debug_btree) {
187 kprintf("BRACKETL %d:%d:%08x[%d] %016llx %02x %016llx %d\n",
188 cursor->node->cluster->volume->vol_no,
189 cursor->node->cluster->clu_no,
190 cursor->node->node_offset,
192 elm[0].internal.base.obj_id,
193 elm[0].internal.base.rec_type,
194 elm[0].internal.base.key,
197 kprintf("BRACKETR %d:%d:%08x[%d] %016llx %02x %016llx %d\n",
198 cursor->node->cluster->volume->vol_no,
199 cursor->node->cluster->clu_no,
200 cursor->node->node_offset,
202 elm[1].internal.base.obj_id,
203 elm[1].internal.base.rec_type,
204 elm[1].internal.base.key,
213 if (r == 0 && (cursor->flags &
214 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
221 * When iterating try to clean up any deleted
222 * internal elements left over from btree_remove()
223 * deadlocks, but it is ok if we can't.
225 if (elm->internal.subtree_offset == 0) {
226 btree_remove_deleted_element(cursor);
227 /* note: elm also invalid */
228 } else if (elm->internal.subtree_offset != 0) {
229 error = hammer_cursor_down(cursor);
232 KKASSERT(cursor->index == 0);
234 /* reload stale pointer */
235 node = cursor->node->ondisk;
238 elm = &node->elms[cursor->index];
239 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
240 if (hammer_debug_btree) {
241 kprintf("ELEMENT %d:%d:%08x:%d %c %016llx %02x %016llx %d\n",
242 cursor->node->cluster->volume->vol_no,
243 cursor->node->cluster->clu_no,
244 cursor->node->node_offset,
246 (elm[0].leaf.base.btype ?
247 elm[0].leaf.base.btype : '?'),
248 elm[0].leaf.base.obj_id,
249 elm[0].leaf.base.rec_type,
250 elm[0].leaf.base.key,
260 * We support both end-inclusive and
261 * end-exclusive searches.
264 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
269 switch(elm->leaf.base.btype) {
270 case HAMMER_BTREE_TYPE_RECORD:
271 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
272 hammer_btree_chkts(cursor->asof, &elm->base)) {
277 case HAMMER_BTREE_TYPE_SPIKE_BEG:
279 * NOTE: This code assumes that the spike
280 * ending element immediately follows the
281 * spike beginning element.
284 * We must cursor-down via the SPIKE_END
285 * element, otherwise cursor->parent will
286 * not be set correctly for deletions.
288 * fall-through to avoid an improper
289 * termination from the conditional above.
291 KKASSERT(cursor->index + 1 < node->count);
293 KKASSERT(elm->leaf.base.btype ==
294 HAMMER_BTREE_TYPE_SPIKE_END);
297 case HAMMER_BTREE_TYPE_SPIKE_END:
299 * The SPIKE_END element is inclusive, NOT
300 * like a boundary, so be careful with the
303 * This code assumes that a preceding SPIKE_BEG
304 * has already been checked.
306 if (cursor->flags & HAMMER_CURSOR_INCLUSTER)
308 error = hammer_cursor_down(cursor);
311 KKASSERT(cursor->index == 0);
312 /* reload stale pointer */
313 node = cursor->node->ondisk;
316 * If the cluster root is empty it and its
317 * related spike can be deleted. Ignore
320 if (node->count == 0) {
321 error = hammer_cursor_upgrade(cursor);
323 error = btree_remove(cursor);
324 hammer_cursor_downgrade(cursor);
326 /* reload stale pointer */
327 node = cursor->node->ondisk;
338 * node pointer invalid after loop
344 if (hammer_debug_btree) {
345 int i = cursor->index;
346 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
347 kprintf("ITERATE %p:%d %016llx %02x %016llx\n",
349 elm->internal.base.obj_id,
350 elm->internal.base.rec_type,
351 elm->internal.base.key
360 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
361 * could not be found, EDEADLK if inserting and a retry is needed, and a
362 * fatal error otherwise. When retrying, the caller must terminate the
363 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
365 * The cursor is suitably positioned for a deletion on success, and suitably
366 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
369 * The cursor may begin anywhere, the search will traverse clusters in
370 * either direction to locate the requested element.
372 * Most of the logic implementing historical searches is handled here. We
373 * do an initial lookup with create_tid set to the asof TID. Due to the
374 * way records are laid out, a backwards iteration may be required if
375 * ENOENT is returned to locate the historical record. Here's the
378 * create_tid: 10 15 20
382 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
383 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
384 * not visible and thus causes ENOENT to be returned. We really need
385 * to check record 11 in LEAF1. If it also fails then the search fails
386 * (e.g. it might represent the range 11-16 and thus still not match our
387 * AS-OF timestamp of 17).
389 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
390 * and the cursor->create_check TID if an iteration might be needed.
391 * In the above example create_check would be set to 14.
394 hammer_btree_lookup(hammer_cursor_t cursor)
398 if (cursor->flags & HAMMER_CURSOR_ASOF) {
399 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
400 cursor->key_beg.create_tid = cursor->asof;
402 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
403 error = btree_search(cursor, 0);
404 if (error != ENOENT ||
405 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
408 * Stop if error other then ENOENT.
409 * Stop if ENOENT and not special case.
413 cursor->key_beg.create_tid = cursor->create_check;
417 error = btree_search(cursor, 0);
419 if (error == 0 && cursor->flags)
420 error = hammer_btree_extract(cursor, cursor->flags);
425 * Execute the logic required to start an iteration. The first record
426 * located within the specified range is returned and iteration control
427 * flags are adjusted for successive hammer_btree_iterate() calls.
430 hammer_btree_first(hammer_cursor_t cursor)
434 error = hammer_btree_lookup(cursor);
435 if (error == ENOENT) {
436 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
437 error = hammer_btree_iterate(cursor);
439 cursor->flags |= HAMMER_CURSOR_ATEDISK;
444 * Extract the record and/or data associated with the cursor's current
445 * position. Any prior record or data stored in the cursor is replaced.
446 * The cursor must be positioned at a leaf node.
448 * NOTE: Most extractions occur at the leaf of the B-Tree. The only
449 * extraction allowed at an internal element is at a cluster-push.
450 * Cluster-push elements have records but no data.
453 hammer_btree_extract(hammer_cursor_t cursor, int flags)
455 hammer_node_ondisk_t node;
456 hammer_btree_elm_t elm;
457 hammer_cluster_t cluster;
464 * A cluster record type has no data reference, the information
465 * is stored directly in the record and B-Tree element.
467 * The case where the data reference resolves to the same buffer
468 * as the record reference must be handled.
470 node = cursor->node->ondisk;
471 elm = &node->elms[cursor->index];
472 cluster = cursor->node->cluster;
473 cursor->flags &= ~HAMMER_CURSOR_DATA_EMBEDDED;
477 * There is nothing to extract for an internal element.
479 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
482 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
487 if ((flags & HAMMER_CURSOR_GET_RECORD)) {
488 cloff = elm->leaf.rec_offset;
489 cursor->record = hammer_bread(cluster, cloff,
490 HAMMER_FSBUF_RECORDS, &error,
491 &cursor->record_buffer);
496 if ((flags & HAMMER_CURSOR_GET_DATA) && error == 0) {
497 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD) {
499 * Only records have data references. Spike elements
503 } else if ((cloff ^ elm->leaf.data_offset) & ~HAMMER_BUFMASK) {
505 * The data is not in the same buffer as the last
506 * record we cached, but it could still be embedded
507 * in a record. Note that we may not have loaded the
508 * record's buffer above, depending on flags.
510 if ((elm->leaf.rec_offset ^ elm->leaf.data_offset) &
512 if (elm->leaf.data_len & HAMMER_BUFMASK)
513 buf_type = HAMMER_FSBUF_DATA;
515 buf_type = 0; /* pure data buffer */
517 buf_type = HAMMER_FSBUF_RECORDS;
519 cursor->data = hammer_bread(cluster,
520 elm->leaf.data_offset,
522 &cursor->data_buffer);
525 * Data in same buffer as record. Note that we
526 * leave any existing data_buffer intact, even
527 * though we don't use it in this case, in case
528 * other records extracted during an iteration
531 * The data must be embedded in the record for this
534 * Just assume the buffer type is correct.
536 cursor->data = (void *)
537 ((char *)cursor->record_buffer->ondisk +
538 (elm->leaf.data_offset & HAMMER_BUFMASK));
539 roff = (char *)cursor->data - (char *)cursor->record;
540 KKASSERT (roff >= 0 && roff < HAMMER_RECORD_SIZE);
541 cursor->flags |= HAMMER_CURSOR_DATA_EMBEDDED;
549 * Insert a leaf element into the B-Tree at the current cursor position.
550 * The cursor is positioned such that the element at and beyond the cursor
551 * are shifted to make room for the new record.
553 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
554 * flag set and that call must return ENOENT before this function can be
557 * ENOSPC is returned if there is no room to insert a new record.
560 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_elm_t elm)
562 hammer_node_ondisk_t node;
566 if ((error = hammer_cursor_upgrade(cursor)) != 0)
570 * Insert the element at the leaf node and update the count in the
571 * parent. It is possible for parent to be NULL, indicating that
572 * the root of the B-Tree in the cluster is a leaf. It is also
573 * possible for the leaf to be empty.
575 * Remember that the right-hand boundary is not included in the
578 hammer_modify_node(cursor->node);
579 node = cursor->node->ondisk;
581 KKASSERT(elm->base.btype != 0);
582 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
583 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
584 if (i != node->count) {
585 bcopy(&node->elms[i], &node->elms[i+1],
586 (node->count - i) * sizeof(*elm));
588 node->elms[i] = *elm;
592 * Debugging sanity checks. Note that the element to the left
593 * can match the element we are inserting if it is a SPIKE_END,
594 * because spike-end's represent a non-inclusive end to a range.
596 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
597 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
599 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->leaf.base) < 0);
601 if (i != node->count - 1)
602 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->leaf.base) > 0);
608 * Insert a cluster spike into the B-Tree at the current cursor position.
609 * The caller pre-positions the insertion cursor at ncluster's
610 * left bound in the originating cluster. Both the originating cluster
611 * and the target cluster must be serialized, EDEADLK is fatal.
613 * Basically we have to lay down the two spike elements and assert that
614 * the leaf's right bound does not bisect the ending element. The ending
615 * spike element is non-inclusive, just like a boundary. The target cluster's
616 * clu_btree_parent_offset may have to adjusted.
618 * NOTE: Serialization is usually accoplished by virtue of being the
619 * initial accessor of a cluster.
622 hammer_btree_insert_cluster(hammer_cursor_t cursor, hammer_cluster_t ncluster,
625 hammer_node_ondisk_t node;
626 hammer_btree_elm_t elm;
627 hammer_cluster_t ocluster;
628 const int esize = sizeof(*elm);
633 if ((error = hammer_cursor_upgrade(cursor)) != 0)
635 hammer_modify_node(cursor->node);
636 node = cursor->node->ondisk;
637 node_offset = cursor->node->node_offset;
640 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
641 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS - 2);
642 KKASSERT(i >= 0 && i <= node->count);
645 * Make sure the spike is legal or the B-Tree code will get really
648 * XXX the right bound my bisect the two spike elements. We
649 * need code here to 'fix' the right bound going up the tree
650 * instead of an assertion.
652 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_beg,
653 cursor->left_bound) >= 0);
654 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
655 cursor->right_bound) <= 0);
656 if (i != node->count) {
657 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
658 &node->elms[i].leaf.base) <= 0);
661 elm = &node->elms[i];
662 bcopy(elm, elm + 2, (node->count - i) * esize);
663 bzero(elm, 2 * esize);
666 elm[0].leaf.base = ncluster->ondisk->clu_btree_beg;
667 elm[0].leaf.base.btype = HAMMER_BTREE_TYPE_SPIKE_BEG;
668 elm[0].leaf.rec_offset = rec_offset;
669 elm[0].leaf.spike_clu_no = ncluster->clu_no;
670 elm[0].leaf.spike_vol_no = ncluster->volume->vol_no;
672 elm[1].leaf.base = ncluster->ondisk->clu_btree_end;
673 elm[1].leaf.base.btype = HAMMER_BTREE_TYPE_SPIKE_END;
674 elm[1].leaf.rec_offset = rec_offset;
675 elm[1].leaf.spike_clu_no = ncluster->clu_no;
676 elm[1].leaf.spike_vol_no = ncluster->volume->vol_no;
679 * SPIKE_END must be inclusive, not exclusive.
681 KKASSERT(elm[1].leaf.base.create_tid != 1);
682 --elm[1].leaf.base.create_tid;
685 * The target cluster's parent offset may have to be updated.
687 * NOTE: Modifying a cluster header does not mark it open, and
688 * flushing it will only clear an existing open flag if the cluster
689 * has been validated.
691 if (hammer_debug_general & 0x40) {
692 kprintf("INSERT CLUSTER %d:%d -> %d:%d ",
693 ncluster->ondisk->clu_btree_parent_vol_no,
694 ncluster->ondisk->clu_btree_parent_clu_no,
695 ncluster->volume->vol_no,
699 ocluster = cursor->node->cluster;
700 if (ncluster->ondisk->clu_btree_parent_offset != node_offset ||
701 ncluster->ondisk->clu_btree_parent_clu_no != ocluster->clu_no ||
702 ncluster->ondisk->clu_btree_parent_vol_no != ocluster->volume->vol_no) {
703 hammer_modify_cluster(ncluster);
704 ncluster->ondisk->clu_btree_parent_offset = node_offset;
705 ncluster->ondisk->clu_btree_parent_clu_no = ocluster->clu_no;
706 ncluster->ondisk->clu_btree_parent_vol_no = ocluster->volume->vol_no;
707 if (hammer_debug_general & 0x40)
708 kprintf("(offset fixup)\n");
710 if (hammer_debug_general & 0x40)
711 kprintf("(offset unchanged)\n");
718 * Delete a record from the B-Tree at the current cursor position.
719 * The cursor is positioned such that the current element is the one
722 * On return the cursor will be positioned after the deleted element and
723 * MAY point to an internal node. It will be suitable for the continuation
724 * of an iteration but not for an insertion or deletion.
726 * Deletions will attempt to partially rebalance the B-Tree in an upward
727 * direction, but will terminate rather then deadlock. Empty leaves are
728 * not allowed except at the root node of a cluster. An early termination
729 * will leave an internal node with an element whos subtree_offset is 0,
730 * a case detected and handled by btree_search().
732 * This function can return EDEADLK, requiring the caller to retry the
733 * operation after clearing the deadlock.
736 hammer_btree_delete(hammer_cursor_t cursor)
738 hammer_node_ondisk_t ondisk;
740 hammer_node_t parent;
744 if ((error = hammer_cursor_upgrade(cursor)) != 0)
748 * Delete the element from the leaf node.
750 * Remember that leaf nodes do not have boundaries.
753 ondisk = node->ondisk;
756 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
757 KKASSERT(i >= 0 && i < ondisk->count);
758 hammer_modify_node(node);
759 if (i + 1 != ondisk->count) {
760 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
761 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
766 * Validate local parent
768 if (ondisk->parent) {
769 parent = cursor->parent;
771 KKASSERT(parent != NULL);
772 KKASSERT(parent->node_offset == ondisk->parent);
773 KKASSERT(parent->cluster == node->cluster);
777 * If the leaf becomes empty it must be detached from the parent,
778 * potentially recursing through to the cluster root.
780 * This may reposition the cursor at one of the parent's of the
783 * Ignore deadlock errors, that simply means that btree_remove
784 * was unable to recurse and had to leave the subtree_offset
785 * in the parent set to 0.
787 KKASSERT(cursor->index <= ondisk->count);
788 if (ondisk->count == 0) {
790 error = btree_remove(cursor);
791 } while (error == EAGAIN);
792 if (error == EDEADLK)
797 KKASSERT(cursor->parent == NULL ||
798 cursor->parent_index < cursor->parent->ondisk->count);
803 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
805 * Search a cluster's B-Tree for cursor->key_beg, return the matching node.
807 * The search can begin ANYWHERE in the B-Tree. As a first step the search
808 * iterates up the tree as necessary to properly position itself prior to
809 * actually doing the sarch.
811 * INSERTIONS: The search will split full nodes and leaves on its way down
812 * and guarentee that the leaf it ends up on is not full. If we run out
813 * of space the search continues to the leaf (to position the cursor for
814 * the spike), but ENOSPC is returned.
816 * The search is only guarenteed to end up on a leaf if an error code of 0
817 * is returned, or if inserting and an error code of ENOENT is returned.
818 * Otherwise it can stop at an internal node. On success a search returns
819 * a leaf node unless INCLUSTER is set and the search located a cluster push
820 * node (which is an internal node).
822 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
823 * filesystem, and it is not simple code. Please note the following facts:
825 * - Internal node recursions have a boundary on the left AND right. The
826 * right boundary is non-inclusive. The create_tid is a generic part
827 * of the key for internal nodes.
829 * - Leaf nodes contain terminal elements AND spikes. A spike recurses into
830 * another cluster and contains two leaf elements.. a beginning and an
831 * ending element. The SPIKE_END element is RANGE-EXCLUSIVE, just like a
832 * boundary. This means that it is possible to have two elements
833 * (a spike ending element and a record) side by side with the same key.
835 * - Because the SPIKE_END element is range inclusive, it cannot match the
836 * right boundary of the parent node. SPIKE_BEG and SPIKE_END elements
837 * always come in pairs, and always exist side by side in the same leaf.
839 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
840 * historical search. ASOF and INSERT are mutually exclusive. When
841 * doing an as-of lookup btree_search() checks for a right-edge boundary
842 * case. If while recursing down the left-edge differs from the key
843 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
844 * with cursor->create_check. This is used by btree_lookup() to iterate.
845 * The iteration backwards because as-of searches can wind up going
846 * down the wrong branch of the B-Tree.
850 btree_search(hammer_cursor_t cursor, int flags)
852 hammer_node_ondisk_t node;
853 hammer_cluster_t cluster;
854 hammer_btree_elm_t elm;
861 flags |= cursor->flags;
863 if (hammer_debug_btree) {
864 kprintf("SEARCH %d:%d:%08x[%d] %016llx %02x key=%016llx cre=%016llx\n",
865 cursor->node->cluster->volume->vol_no,
866 cursor->node->cluster->clu_no,
867 cursor->node->node_offset,
869 cursor->key_beg.obj_id,
870 cursor->key_beg.rec_type,
872 cursor->key_beg.create_tid
877 * Move our cursor up the tree until we find a node whos range covers
878 * the key we are trying to locate. This may move us between
881 * The left bound is inclusive, the right bound is non-inclusive.
882 * It is ok to cursor up too far so when cursoring across a cluster
885 * First see if we can skip the whole cluster. hammer_cursor_up()
886 * handles both cases but this way we don't check the cluster
887 * bounds when going up the tree within a cluster.
889 * NOTE: If INCLUSTER is set and we are at the root of the cluster,
890 * hammer_cursor_up() will return ENOENT.
892 cluster = cursor->node->cluster;
894 r = hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_beg);
895 s = hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_end);
899 error = hammer_cursor_toroot(cursor);
902 KKASSERT(cursor->parent);
903 error = hammer_cursor_up(cursor);
906 cluster = cursor->node->cluster;
909 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
910 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
913 KKASSERT(cursor->parent);
914 error = hammer_cursor_up(cursor);
920 * The delete-checks below are based on node, not parent. Set the
921 * initial delete-check based on the parent.
924 KKASSERT(cursor->left_bound->create_tid != 1);
925 cursor->create_check = cursor->left_bound->create_tid - 1;
926 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
930 * We better have ended up with a node somewhere, and our second
931 * while loop had better not have traversed up a cluster.
933 KKASSERT(cursor->node != NULL && cursor->node->cluster == cluster);
936 * If we are inserting we can't start at a full node if the parent
937 * is also full (because there is no way to split the node),
938 * continue running up the tree until the requirement is satisfied
939 * or we hit the root of the current cluster.
941 * (If inserting we aren't doing an as-of search so we don't have
942 * to worry about create_check).
944 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
945 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
946 if (btree_node_is_full(cursor->node->ondisk) == 0)
949 if (btree_node_is_almost_full(cursor->node->ondisk) ==0)
952 if (cursor->node->ondisk->parent == 0 ||
953 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
956 error = hammer_cursor_up(cursor);
957 /* cluster and node are now may become stale */
961 /* cluster = cursor->node->cluster; not needed until next cluster = */
965 * Push down through internal nodes to locate the requested key.
967 cluster = cursor->node->cluster;
968 node = cursor->node->ondisk;
969 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
971 * Scan the node to find the subtree index to push down into.
972 * We go one-past, then back-up.
974 * We must proactively remove deleted elements which may
975 * have been left over from a deadlocked btree_remove().
977 * The left and right boundaries are included in the loop
978 * in order to detect edge cases.
980 * If the separator only differs by create_tid (r == 1)
981 * and we are doing an as-of search, we may end up going
982 * down a branch to the left of the one containing the
983 * desired key. This requires numerous special cases.
985 if (hammer_debug_btree) {
986 kprintf("SEARCH-I %d:%d:%08x count=%d\n",
987 cursor->node->cluster->volume->vol_no,
988 cursor->node->cluster->clu_no,
989 cursor->node->node_offset,
992 for (i = 0; i <= node->count; ++i) {
993 elm = &node->elms[i];
994 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
995 if (hammer_debug_btree > 2) {
996 kprintf(" IELM %p %d r=%d\n",
997 &node->elms[i], i, r);
1002 KKASSERT(elm->base.create_tid != 1);
1003 cursor->create_check = elm->base.create_tid - 1;
1004 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1007 if (hammer_debug_btree) {
1008 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1013 * These cases occur when the parent's idea of the boundary
1014 * is wider then the child's idea of the boundary, and
1015 * require special handling. If not inserting we can
1016 * terminate the search early for these cases but the
1017 * child's boundaries cannot be unconditionally modified.
1021 * If i == 0 the search terminated to the LEFT of the
1022 * left_boundary but to the RIGHT of the parent's left
1027 elm = &node->elms[0];
1030 * If we aren't inserting we can stop here.
1032 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1038 * Correct a left-hand boundary mismatch.
1040 * We can only do this if we can upgrade the lock.
1042 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1044 hammer_modify_node(cursor->node);
1045 save = node->elms[0].base.btype;
1046 node->elms[0].base = *cursor->left_bound;
1047 node->elms[0].base.btype = save;
1048 } else if (i == node->count + 1) {
1050 * If i == node->count + 1 the search terminated to
1051 * the RIGHT of the right boundary but to the LEFT
1052 * of the parent's right boundary. If we aren't
1053 * inserting we can stop here.
1055 * Note that the last element in this case is
1056 * elms[i-2] prior to adjustments to 'i'.
1059 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1065 * Correct a right-hand boundary mismatch.
1066 * (actual push-down record is i-2 prior to
1067 * adjustments to i).
1069 * We can only do this if we can upgrade the lock.
1071 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1073 elm = &node->elms[i];
1074 hammer_modify_node(cursor->node);
1075 elm->base = *cursor->right_bound;
1079 * The push-down index is now i - 1. If we had
1080 * terminated on the right boundary this will point
1081 * us at the last element.
1086 elm = &node->elms[i];
1088 if (hammer_debug_btree) {
1089 kprintf("RESULT-I %d:%d:%08x[%d] %016llx %02x "
1090 "key=%016llx cre=%016llx\n",
1091 cursor->node->cluster->volume->vol_no,
1092 cursor->node->cluster->clu_no,
1093 cursor->node->node_offset,
1095 elm->internal.base.obj_id,
1096 elm->internal.base.rec_type,
1097 elm->internal.base.key,
1098 elm->internal.base.create_tid
1103 * When searching try to clean up any deleted
1104 * internal elements left over from btree_remove()
1107 * If we fail and we are doing an insertion lookup,
1108 * we have to return EDEADLK, because an insertion lookup
1109 * must terminate at a leaf.
1111 if (elm->internal.subtree_offset == 0) {
1112 error = btree_remove_deleted_element(cursor);
1115 if (error == EDEADLK &&
1116 (flags & HAMMER_CURSOR_INSERT) == 0) {
1124 * Handle insertion and deletion requirements.
1126 * If inserting split full nodes. The split code will
1127 * adjust cursor->node and cursor->index if the current
1128 * index winds up in the new node.
1130 * If inserting and a left or right edge case was detected,
1131 * we cannot correct the left or right boundary and must
1132 * prepend and append an empty leaf node in order to make
1133 * the boundary correction.
1135 * If we run out of space we set enospc and continue on
1136 * to a leaf to provide the spike code with a good point
1137 * of entry. Enospc is reset if we cross a cluster boundary.
1139 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1140 if (btree_node_is_full(node)) {
1141 error = btree_split_internal(cursor);
1143 if (error != ENOSPC)
1148 * reload stale pointers
1151 node = cursor->node->ondisk;
1156 * Push down (push into new node, existing node becomes
1157 * the parent) and continue the search.
1159 error = hammer_cursor_down(cursor);
1160 /* node and cluster become stale */
1163 node = cursor->node->ondisk;
1164 cluster = cursor->node->cluster;
1168 * We are at a leaf, do a linear search of the key array.
1170 * If we encounter a spike element type within the necessary
1171 * range we push into it. Note that SPIKE_END is non-inclusive
1172 * of the spike range.
1174 * On success the index is set to the matching element and 0
1177 * On failure the index is set to the insertion point and ENOENT
1180 * Boundaries are not stored in leaf nodes, so the index can wind
1181 * up to the left of element 0 (index == 0) or past the end of
1182 * the array (index == node->count).
1184 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1185 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1186 if (hammer_debug_btree) {
1187 kprintf("SEARCH-L %d:%d:%08x count=%d\n",
1188 cursor->node->cluster->volume->vol_no,
1189 cursor->node->cluster->clu_no,
1190 cursor->node->node_offset,
1194 for (i = 0; i < node->count; ++i) {
1195 elm = &node->elms[i];
1197 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1199 if (hammer_debug_btree > 1)
1200 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1202 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_BEG) {
1204 * SPIKE_BEG. Stop if we are to the left of the
1205 * spike begin element.
1207 * If we are not the last element in the leaf continue
1208 * the loop looking for the SPIKE_END. If we are
1209 * the last element, however, then push into the
1212 * If doing an as-of search a Spike demark on a
1213 * create_tid boundary must be pushed into and an
1214 * iteration will be forced if it turned out to be
1217 * If not doing an as-of search exact comparisons
1220 * enospc must be reset because we have crossed a
1227 * Set the create_check if the spike element
1228 * only differs by its create_tid.
1231 cursor->create_check = elm->base.create_tid;
1232 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1234 if (i != node->count - 1)
1236 panic("btree_search: illegal spike, no SPIKE_END "
1237 "in leaf node! %p\n", cursor->node);
1239 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_END) {
1241 * SPIKE_END. We can only hit this case if we are
1242 * greater or equal to SPIKE_BEG.
1244 * If we are <= SPIKE_END we must push into
1245 * it, otherwise continue the search. The SPIKE_END
1246 * element is range-inclusive.
1248 * enospc must be reset because we have crossed a
1253 * Continue the search but check for a
1254 * create_tid boundary. Because the
1255 * SPIKE_END is inclusive we do not have
1256 * to subtract 1 to force an iteration to
1257 * go down the spike.
1260 cursor->create_check =
1261 elm->base.create_tid;
1263 HAMMER_CURSOR_CREATE_CHECK;
1267 if (flags & HAMMER_CURSOR_INCLUSTER)
1270 error = hammer_cursor_down(cursor);
1278 * We are at a record element. Stop if we've flipped past
1279 * key_beg, not counting the create_tid test. Allow the
1280 * r == 1 case (key_beg > element but differs only by its
1281 * create_tid) to fall through to the AS-OF check.
1283 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1291 * Check our as-of timestamp against the element.
1293 if (flags & HAMMER_CURSOR_ASOF) {
1294 if (hammer_btree_chkts(cursor->asof,
1295 &node->elms[i].base) != 0) {
1300 if (r > 0) /* can only be +1 */
1307 if (hammer_debug_btree) {
1308 kprintf("RESULT-L %d:%d:%08x[%d] (SUCCESS)\n",
1309 cursor->node->cluster->volume->vol_no,
1310 cursor->node->cluster->clu_no,
1311 cursor->node->node_offset,
1318 * The search of the leaf node failed. i is the insertion point.
1321 if (hammer_debug_btree) {
1322 kprintf("RESULT-L %d:%d:%08x[%d] (FAILED)\n",
1323 cursor->node->cluster->volume->vol_no,
1324 cursor->node->cluster->clu_no,
1325 cursor->node->node_offset,
1330 * No exact match was found, i is now at the insertion point.
1332 * If inserting split a full leaf before returning. This
1333 * may have the side effect of adjusting cursor->node and
1336 * For now the leaf must have at least 2 free elements to accomodate
1337 * the insertion of a spike during recovery. See the
1338 * hammer_btree_insert_cluster() function.
1341 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1342 btree_node_is_almost_full(node)) {
1343 error = btree_split_leaf(cursor);
1345 if (error != ENOSPC)
1350 * reload stale pointers
1354 node = &cursor->node->internal;
1359 * We reached a leaf but did not find the key we were looking for.
1360 * If this is an insert we will be properly positioned for an insert
1361 * (ENOENT) or spike (ENOSPC) operation.
1363 error = enospc ? ENOSPC : ENOENT;
1369 /************************************************************************
1370 * SPLITTING AND MERGING *
1371 ************************************************************************
1373 * These routines do all the dirty work required to split and merge nodes.
1377 * Split an internal node into two nodes and move the separator at the split
1378 * point to the parent.
1380 * (cursor->node, cursor->index) indicates the element the caller intends
1381 * to push into. We will adjust node and index if that element winds
1382 * up in the split node.
1384 * If we are at the root of a cluster a new root must be created with two
1385 * elements, one pointing to the original root and one pointing to the
1386 * newly allocated split node.
1388 * NOTE! Being at the root of a cluster is different from being at the
1389 * root of the root cluster. cursor->parent will not be NULL and
1390 * cursor->node->ondisk.parent must be tested against 0. Theoretically
1391 * we could propogate the algorithm into the parent and deal with multiple
1392 * 'roots' in the cluster header, but it's easier not to.
1396 btree_split_internal(hammer_cursor_t cursor)
1398 hammer_node_ondisk_t ondisk;
1400 hammer_node_t parent;
1401 hammer_node_t new_node;
1402 hammer_btree_elm_t elm;
1403 hammer_btree_elm_t parent_elm;
1404 hammer_node_locklist_t locklist = NULL;
1410 const int esize = sizeof(*elm);
1412 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1414 if ((cursor->flags & HAMMER_CURSOR_RECOVER) == 0) {
1415 error = hammer_btree_lock_children(cursor, &locklist);
1421 * We are splitting but elms[split] will be promoted to the parent,
1422 * leaving the right hand node with one less element. If the
1423 * insertion point will be on the left-hand side adjust the split
1424 * point to give the right hand side one additional node.
1426 node = cursor->node;
1427 ondisk = node->ondisk;
1428 split = (ondisk->count + 1) / 2;
1429 if (cursor->index <= split)
1433 * If we are at the root of the cluster, create a new root node with
1434 * 1 element and split normally. Avoid making major modifications
1435 * until we know the whole operation will work.
1437 * The root of the cluster is different from the root of the root
1438 * cluster. Use the node's on-disk structure's parent offset to
1441 if (ondisk->parent == 0) {
1442 parent = hammer_alloc_btree(node->cluster, &error);
1445 hammer_lock_ex(&parent->lock);
1446 hammer_modify_node(parent);
1447 ondisk = parent->ondisk;
1450 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1451 ondisk->elms[0].base = node->cluster->clu_btree_beg;
1452 ondisk->elms[0].base.btype = node->ondisk->type;
1453 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1454 ondisk->elms[1].base = node->cluster->clu_btree_end;
1455 /* ondisk->elms[1].base.btype - not used */
1457 parent_index = 0; /* index of current node in parent */
1460 parent = cursor->parent;
1461 parent_index = cursor->parent_index;
1462 KKASSERT(parent->cluster == node->cluster);
1466 * Split node into new_node at the split point.
1468 * B O O O P N N B <-- P = node->elms[split]
1469 * 0 1 2 3 4 5 6 <-- subtree indices
1474 * B O O O B B N N B <--- inner boundary points are 'P'
1478 new_node = hammer_alloc_btree(node->cluster, &error);
1479 if (new_node == NULL) {
1481 hammer_unlock(&parent->lock);
1482 parent->flags |= HAMMER_NODE_DELETED;
1483 hammer_rel_node(parent);
1487 hammer_lock_ex(&new_node->lock);
1490 * Create the new node. P becomes the left-hand boundary in the
1491 * new node. Copy the right-hand boundary as well.
1493 * elm is the new separator.
1495 hammer_modify_node(new_node);
1496 hammer_modify_node(node);
1497 ondisk = node->ondisk;
1498 elm = &ondisk->elms[split];
1499 bcopy(elm, &new_node->ondisk->elms[0],
1500 (ondisk->count - split + 1) * esize);
1501 new_node->ondisk->count = ondisk->count - split;
1502 new_node->ondisk->parent = parent->node_offset;
1503 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1504 KKASSERT(ondisk->type == new_node->ondisk->type);
1507 * Cleanup the original node. Elm (P) becomes the new boundary,
1508 * its subtree_offset was moved to the new node. If we had created
1509 * a new root its parent pointer may have changed.
1511 elm->internal.subtree_offset = 0;
1512 ondisk->count = split;
1515 * Insert the separator into the parent, fixup the parent's
1516 * reference to the original node, and reference the new node.
1517 * The separator is P.
1519 * Remember that base.count does not include the right-hand boundary.
1521 hammer_modify_node(parent);
1522 ondisk = parent->ondisk;
1523 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1524 parent_elm = &ondisk->elms[parent_index+1];
1525 bcopy(parent_elm, parent_elm + 1,
1526 (ondisk->count - parent_index) * esize);
1527 parent_elm->internal.base = elm->base; /* separator P */
1528 parent_elm->internal.base.btype = new_node->ondisk->type;
1529 parent_elm->internal.subtree_offset = new_node->node_offset;
1533 * The children of new_node need their parent pointer set to new_node.
1534 * The children have already been locked by
1535 * hammer_btree_lock_children().
1537 for (i = 0; i < new_node->ondisk->count; ++i) {
1538 elm = &new_node->ondisk->elms[i];
1539 error = btree_set_parent(new_node, elm);
1541 panic("btree_split_internal: btree-fixup problem");
1546 * The cluster's root pointer may have to be updated.
1549 hammer_modify_cluster(node->cluster);
1550 node->cluster->ondisk->clu_btree_root = parent->node_offset;
1551 node->ondisk->parent = parent->node_offset;
1552 if (cursor->parent) {
1553 hammer_unlock(&cursor->parent->lock);
1554 hammer_rel_node(cursor->parent);
1556 cursor->parent = parent; /* lock'd and ref'd */
1561 * Ok, now adjust the cursor depending on which element the original
1562 * index was pointing at. If we are >= the split point the push node
1563 * is now in the new node.
1565 * NOTE: If we are at the split point itself we cannot stay with the
1566 * original node because the push index will point at the right-hand
1567 * boundary, which is illegal.
1569 * NOTE: The cursor's parent or parent_index must be adjusted for
1570 * the case where a new parent (new root) was created, and the case
1571 * where the cursor is now pointing at the split node.
1573 if (cursor->index >= split) {
1574 cursor->parent_index = parent_index + 1;
1575 cursor->index -= split;
1576 hammer_unlock(&cursor->node->lock);
1577 hammer_rel_node(cursor->node);
1578 cursor->node = new_node; /* locked and ref'd */
1580 cursor->parent_index = parent_index;
1581 hammer_unlock(&new_node->lock);
1582 hammer_rel_node(new_node);
1586 * Fixup left and right bounds
1588 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1589 cursor->left_bound = &parent_elm[0].internal.base;
1590 cursor->right_bound = &parent_elm[1].internal.base;
1591 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1592 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1593 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1594 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1597 hammer_btree_unlock_children(&locklist);
1598 hammer_cursor_downgrade(cursor);
1603 * Same as the above, but splits a full leaf node.
1609 btree_split_leaf(hammer_cursor_t cursor)
1611 hammer_node_ondisk_t ondisk;
1612 hammer_node_t parent;
1614 hammer_node_t new_leaf;
1615 hammer_btree_elm_t elm;
1616 hammer_btree_elm_t parent_elm;
1617 hammer_base_elm_t mid_boundary;
1618 hammer_node_locklist_t locklist = NULL;
1624 const size_t esize = sizeof(*elm);
1626 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1628 if ((cursor->flags & HAMMER_CURSOR_RECOVER) == 0) {
1629 error = hammer_btree_lock_children(cursor, &locklist);
1635 * Calculate the split point. If the insertion point will be on
1636 * the left-hand side adjust the split point to give the right
1637 * hand side one additional node.
1639 * Spikes are made up of two leaf elements which cannot be
1642 leaf = cursor->node;
1643 ondisk = leaf->ondisk;
1644 split = (ondisk->count + 1) / 2;
1645 if (cursor->index <= split)
1649 elm = &ondisk->elms[split];
1650 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_END) {
1652 elm[-1].leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_BEG);
1657 * If we are at the root of the tree, create a new root node with
1658 * 1 element and split normally. Avoid making major modifications
1659 * until we know the whole operation will work.
1661 if (ondisk->parent == 0) {
1662 parent = hammer_alloc_btree(leaf->cluster, &error);
1665 hammer_lock_ex(&parent->lock);
1666 hammer_modify_node(parent);
1667 ondisk = parent->ondisk;
1670 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1671 ondisk->elms[0].base = leaf->cluster->clu_btree_beg;
1672 ondisk->elms[0].base.btype = leaf->ondisk->type;
1673 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1674 ondisk->elms[1].base = leaf->cluster->clu_btree_end;
1675 /* ondisk->elms[1].base.btype = not used */
1677 parent_index = 0; /* insertion point in parent */
1680 parent = cursor->parent;
1681 parent_index = cursor->parent_index;
1682 KKASSERT(parent->cluster == leaf->cluster);
1686 * Split leaf into new_leaf at the split point. Select a separator
1687 * value in-between the two leafs but with a bent towards the right
1688 * leaf since comparisons use an 'elm >= separator' inequality.
1697 new_leaf = hammer_alloc_btree(leaf->cluster, &error);
1698 if (new_leaf == NULL) {
1700 hammer_unlock(&parent->lock);
1701 parent->flags |= HAMMER_NODE_DELETED;
1702 hammer_rel_node(parent);
1706 hammer_lock_ex(&new_leaf->lock);
1709 * Create the new node. P (elm) become the left-hand boundary in the
1710 * new node. Copy the right-hand boundary as well.
1712 hammer_modify_node(leaf);
1713 hammer_modify_node(new_leaf);
1714 ondisk = leaf->ondisk;
1715 elm = &ondisk->elms[split];
1716 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1717 new_leaf->ondisk->count = ondisk->count - split;
1718 new_leaf->ondisk->parent = parent->node_offset;
1719 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1720 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1723 * Cleanup the original node. Because this is a leaf node and
1724 * leaf nodes do not have a right-hand boundary, there
1725 * aren't any special edge cases to clean up. We just fixup the
1728 ondisk->count = split;
1731 * Insert the separator into the parent, fixup the parent's
1732 * reference to the original node, and reference the new node.
1733 * The separator is P.
1735 * Remember that base.count does not include the right-hand boundary.
1736 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1738 hammer_modify_node(parent);
1739 ondisk = parent->ondisk;
1740 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1741 parent_elm = &ondisk->elms[parent_index+1];
1742 bcopy(parent_elm, parent_elm + 1,
1743 (ondisk->count - parent_index) * esize);
1746 * Create the separator. XXX At the moment use exactly the
1747 * right-hand element if this is a recovery operation in order
1748 * to guarantee that it does not bisect the spike elements in a
1749 * later call to hammer_btree_insert_cluster().
1751 if (cursor->flags & HAMMER_CURSOR_RECOVER) {
1752 parent_elm->base = elm[0].base;
1754 hammer_make_separator(&elm[-1].base, &elm[0].base,
1757 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1758 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1759 mid_boundary = &parent_elm->base;
1763 * The children of new_leaf need their parent pointer set to new_leaf.
1764 * The children have already been locked by btree_lock_children().
1766 * The leaf's elements are either TYPE_RECORD or TYPE_SPIKE_*. Only
1767 * elements of BTREE_TYPE_SPIKE_END really requires any action.
1769 for (i = 0; i < new_leaf->ondisk->count; ++i) {
1770 elm = &new_leaf->ondisk->elms[i];
1771 error = btree_set_parent(new_leaf, elm);
1773 panic("btree_split_internal: btree-fixup problem");
1778 * The cluster's root pointer may have to be updated.
1781 hammer_modify_cluster(leaf->cluster);
1782 leaf->cluster->ondisk->clu_btree_root = parent->node_offset;
1783 leaf->ondisk->parent = parent->node_offset;
1784 if (cursor->parent) {
1785 hammer_unlock(&cursor->parent->lock);
1786 hammer_rel_node(cursor->parent);
1788 cursor->parent = parent; /* lock'd and ref'd */
1792 * Ok, now adjust the cursor depending on which element the original
1793 * index was pointing at. If we are >= the split point the push node
1794 * is now in the new node.
1796 * NOTE: If we are at the split point itself we need to select the
1797 * old or new node based on where key_beg's insertion point will be.
1798 * If we pick the wrong side the inserted element will wind up in
1799 * the wrong leaf node and outside that node's bounds.
1801 if (cursor->index > split ||
1802 (cursor->index == split &&
1803 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1804 cursor->parent_index = parent_index + 1;
1805 cursor->index -= split;
1806 hammer_unlock(&cursor->node->lock);
1807 hammer_rel_node(cursor->node);
1808 cursor->node = new_leaf;
1810 cursor->parent_index = parent_index;
1811 hammer_unlock(&new_leaf->lock);
1812 hammer_rel_node(new_leaf);
1816 * Fixup left and right bounds
1818 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1819 cursor->left_bound = &parent_elm[0].internal.base;
1820 cursor->right_bound = &parent_elm[1].internal.base;
1823 * Note: The right assertion is typically > 0, but if the last element
1824 * is a SPIKE_END it can be == 0 because the spike-end is non-inclusive
1825 * of the range being spiked.
1827 * This may seem a bit odd but it works.
1829 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1830 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1831 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1832 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) >= 0);
1835 hammer_btree_unlock_children(&locklist);
1836 hammer_cursor_downgrade(cursor);
1841 * Attempt to remove the empty B-Tree node at (cursor->node). Returns 0
1842 * on success, EAGAIN if we could not acquire the necessary locks, or some
1843 * other error. This node can be a leaf node or an internal node.
1845 * On return the cursor may end up pointing at an internal node, suitable
1846 * for further iteration but not for an immediate insertion or deletion.
1848 * cursor->node may be an internal node or a leaf node.
1850 * NOTE: If cursor->node has one element it is the parent trying to delete
1851 * that element, make sure cursor->index is properly adjusted on success.
1854 btree_remove(hammer_cursor_t cursor)
1856 hammer_node_ondisk_t ondisk;
1857 hammer_btree_elm_t elm;
1860 hammer_node_t parent;
1861 const int esize = sizeof(*elm);
1865 * If we are at the root of the cluster we must be able to
1866 * successfully delete the HAMMER_BTREE_SPIKE_* leaf elements in
1867 * the parent in order to be able to destroy the cluster.
1869 node = cursor->node;
1871 if (node->ondisk->parent == 0) {
1872 hammer_modify_node(node);
1873 ondisk = node->ondisk;
1874 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1880 * When trying to delete a cluster we need to exclusively
1881 * lock the cluster root, its parent (leaf in parent cluster),
1882 * AND the parent of that leaf if it's going to be empty,
1883 * because we can't leave around an empty leaf.
1885 * XXX this is messy due to potentially recursive locks.
1886 * downgrade the cursor, get a second shared lock on the
1887 * node that cannot deadlock because we only own shared locks
1888 * then, cursor-up, and re-upgrade everything. If the
1889 * upgrades EDEADLK then don't try to remove the cluster
1892 if ((parent = cursor->parent) != NULL) {
1893 hammer_cursor_downgrade(cursor);
1895 hammer_ref_node(save);
1896 hammer_lock_sh(&save->lock);
1899 * After the cursor up save has the empty root node
1900 * of the target cluster to be deleted, cursor->node
1901 * is at the leaf containing the spikes, and
1902 * cursor->parent is the parent of that leaf.
1904 * cursor->node and cursor->parent are both in the
1905 * parent cluster of the cluster being deleted.
1907 error = hammer_cursor_up(cursor);
1910 error = hammer_cursor_upgrade(cursor);
1912 error = hammer_lock_upgrade(&save->lock);
1915 /* may be EDEADLK */
1916 kprintf("BTREE_REMOVE: Cannot delete cluster\n");
1917 Debugger("BTREE_REMOVE");
1920 * cursor->node is now the leaf in the parent
1921 * cluster containing the spike elements.
1923 * The cursor should be pointing at the
1924 * SPIKE_END element.
1926 * Remove the spike elements and recurse
1927 * if the leaf becomes empty.
1929 node = cursor->node;
1930 hammer_modify_node(node);
1931 ondisk = node->ondisk;
1932 KKASSERT(cursor->index > 0);
1934 elm = &ondisk->elms[cursor->index];
1935 KKASSERT(elm[0].leaf.base.btype ==
1936 HAMMER_BTREE_TYPE_SPIKE_BEG);
1937 KKASSERT(elm[1].leaf.base.btype ==
1938 HAMMER_BTREE_TYPE_SPIKE_END);
1941 * Ok, remove it and the underlying record.
1943 hammer_free_record(node->cluster,
1944 elm->leaf.rec_offset,
1945 HAMMER_RECTYPE_CLUSTER);
1946 bcopy(elm + 2, elm, (ondisk->count -
1947 cursor->index - 2) * esize);
1949 save->flags |= HAMMER_NODE_DELETED;
1950 save->cluster->flags |= HAMMER_CLUSTER_DELETED;
1951 hammer_flush_node(save);
1952 hammer_unlock(&save->lock);
1953 hammer_rel_node(save);
1954 if (ondisk->count == 0)
1962 * Zero-out the parent's reference to the child and flag the
1963 * child for destruction. This ensures that the child is not
1964 * reused while other references to it exist.
1966 parent = cursor->parent;
1967 hammer_modify_node(parent);
1968 ondisk = parent->ondisk;
1969 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
1970 elm = &ondisk->elms[cursor->parent_index];
1971 KKASSERT(elm->internal.subtree_offset == node->node_offset);
1972 elm->internal.subtree_offset = 0;
1974 hammer_flush_node(node);
1975 node->flags |= HAMMER_NODE_DELETED;
1978 * If the parent would otherwise not become empty we can physically
1979 * remove the zero'd element. Note however that in order to
1980 * guarentee a valid cursor we still need to be able to cursor up
1981 * because we no longer have a node.
1983 * This collapse will change the parent's boundary elements, making
1984 * them wider. The new boundaries are recursively corrected in
1987 * XXX we can theoretically recalculate the midpoint but there isn't
1988 * much of a reason to do it.
1990 error = hammer_cursor_up(cursor);
1992 error = hammer_cursor_upgrade(cursor);
1995 kprintf("BTREE_REMOVE: Cannot lock parent, skipping\n");
1996 Debugger("BTREE_REMOVE");
2001 * Remove the internal element from the parent. The bcopy must
2002 * include the right boundary element.
2004 KKASSERT(parent == cursor->node && ondisk == parent->ondisk);
2007 /* ondisk is node's ondisk */
2008 /* elm is node's element */
2011 * Remove the internal element that we zero'd out. Tell the caller
2012 * to loop if it hits zero (to try to avoid eating up precious kernel
2015 KKASSERT(ondisk->count > 0);
2016 bcopy(&elm[1], &elm[0], (ondisk->count - cursor->index) * esize);
2018 if (ondisk->count == 0)
2024 * Attempt to remove the deleted internal element at the current cursor
2025 * position. If we are unable to remove the element we return EDEADLK.
2027 * If the current internal node becomes empty we delete it in the parent
2028 * and cursor up, looping until we finish or we deadlock.
2030 * On return, if successful, the cursor will be pointing at the next
2031 * iterative position in the B-Tree. If unsuccessful the cursor will be
2032 * pointing at the last deleted internal element that could not be
2037 btree_remove_deleted_element(hammer_cursor_t cursor)
2040 hammer_btree_elm_t elm;
2043 if ((error = hammer_cursor_upgrade(cursor)) != 0)
2045 node = cursor->node;
2046 elm = &node->ondisk->elms[cursor->index];
2047 if (elm->internal.subtree_offset == 0) {
2049 error = btree_remove(cursor);
2050 kprintf("BTREE REMOVE DELETED ELEMENT %d\n", error);
2051 } while (error == EAGAIN);
2057 * The element (elm) has been moved to a new internal node (node).
2059 * If the element represents a pointer to an internal node that node's
2060 * parent must be adjusted to the element's new location.
2062 * If the element represents a spike the target cluster's header must
2063 * be adjusted to point to the element's new location. This only
2064 * applies to HAMMER_SPIKE_END.
2066 * GET_CLUSTER_NORECOVER must be used to avoid a recovery recursion during
2067 * the rebuild of the recovery cluster's B-Tree, which can blow the kernel
2070 * XXX deadlock potential here with our exclusive locks
2074 btree_set_parent(hammer_node_t node, hammer_btree_elm_t elm)
2076 hammer_volume_t volume;
2077 hammer_cluster_t cluster;
2078 hammer_node_t child;
2083 switch(elm->base.btype) {
2084 case HAMMER_BTREE_TYPE_INTERNAL:
2085 case HAMMER_BTREE_TYPE_LEAF:
2086 child = hammer_get_node(node->cluster,
2087 elm->internal.subtree_offset, &error);
2089 hammer_modify_node(child);
2090 child->ondisk->parent = node->node_offset;
2091 hammer_rel_node(child);
2094 case HAMMER_BTREE_TYPE_SPIKE_END:
2095 volume = hammer_get_volume(node->cluster->volume->hmp,
2096 elm->leaf.spike_vol_no, &error);
2099 cluster = hammer_get_cluster(volume, elm->leaf.spike_clu_no,
2100 &error, GET_CLUSTER_NORECOVER);
2101 hammer_rel_volume(volume, 0);
2104 hammer_modify_cluster(cluster);
2105 cluster->ondisk->clu_btree_parent_offset = node->node_offset;
2106 KKASSERT(cluster->ondisk->clu_btree_parent_clu_no ==
2107 node->cluster->clu_no);
2108 KKASSERT(cluster->ondisk->clu_btree_parent_vol_no ==
2109 node->cluster->volume->vol_no);
2110 hammer_rel_cluster(cluster, 0);
2119 * Exclusively lock all the children of node. This is used by the split
2120 * code to prevent anyone from accessing the children of a cursor node
2121 * while we fix-up its parent offset.
2123 * If we don't lock the children we can really mess up cursors which block
2124 * trying to cursor-up into our node.
2126 * WARNING: Cannot be used when doing B-tree operations on a recovery
2127 * cluster because the target cluster may require recovery, resulting
2128 * in a deep recursion which blows the kernel stack.
2130 * On failure EDEADLK (or some other error) is returned. If a deadlock
2131 * error is returned the cursor is adjusted to block on termination.
2134 hammer_btree_lock_children(hammer_cursor_t cursor,
2135 struct hammer_node_locklist **locklistp)
2138 hammer_node_locklist_t item;
2139 hammer_node_ondisk_t ondisk;
2140 hammer_btree_elm_t elm;
2141 hammer_volume_t volume;
2142 hammer_cluster_t cluster;
2143 hammer_node_t child;
2147 node = cursor->node;
2148 ondisk = node->ondisk;
2150 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2151 elm = &ondisk->elms[i];
2154 switch(elm->base.btype) {
2155 case HAMMER_BTREE_TYPE_INTERNAL:
2156 case HAMMER_BTREE_TYPE_LEAF:
2157 child = hammer_get_node(node->cluster,
2158 elm->internal.subtree_offset,
2161 case HAMMER_BTREE_TYPE_SPIKE_END:
2162 volume = hammer_get_volume(node->cluster->volume->hmp,
2163 elm->leaf.spike_vol_no,
2167 cluster = hammer_get_cluster(volume,
2168 elm->leaf.spike_clu_no,
2171 hammer_rel_volume(volume, 0);
2174 KKASSERT(cluster->ondisk->clu_btree_root != 0);
2175 child = hammer_get_node(cluster,
2176 cluster->ondisk->clu_btree_root,
2178 hammer_rel_cluster(cluster, 0);
2184 if (hammer_lock_ex_try(&child->lock) != 0) {
2185 if (cursor->deadlk_node == NULL) {
2186 cursor->deadlk_node = node;
2187 hammer_ref_node(cursor->deadlk_node);
2191 item = kmalloc(sizeof(*item),
2192 M_HAMMER, M_WAITOK);
2193 item->next = *locklistp;
2200 hammer_btree_unlock_children(locklistp);
2206 * Release previously obtained node locks.
2209 hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2211 hammer_node_locklist_t item;
2213 while ((item = *locklistp) != NULL) {
2214 *locklistp = item->next;
2215 hammer_unlock(&item->node->lock);
2216 hammer_rel_node(item->node);
2217 kfree(item, M_HAMMER);
2221 /************************************************************************
2222 * MISCELLANIOUS SUPPORT *
2223 ************************************************************************/
2226 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2228 * Note that for this particular function a return value of -1, 0, or +1
2229 * can denote a match if create_tid is otherwise discounted. A create_tid
2230 * of zero is considered to be 'infinity' in comparisons.
2232 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2235 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2237 if (key1->obj_id < key2->obj_id)
2239 if (key1->obj_id > key2->obj_id)
2242 if (key1->rec_type < key2->rec_type)
2244 if (key1->rec_type > key2->rec_type)
2247 if (key1->key < key2->key)
2249 if (key1->key > key2->key)
2253 * A create_tid of zero indicates a record which is undeletable
2254 * and must be considered to have a value of positive infinity.
2256 if (key1->create_tid == 0) {
2257 if (key2->create_tid == 0)
2261 if (key2->create_tid == 0)
2263 if (key1->create_tid < key2->create_tid)
2265 if (key1->create_tid > key2->create_tid)
2271 * Test a timestamp against an element to determine whether the
2272 * element is visible. A timestamp of 0 means 'infinity'.
2275 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2278 if (base->delete_tid)
2282 if (asof < base->create_tid)
2284 if (base->delete_tid && asof >= base->delete_tid)
2290 * Create a separator half way inbetween key1 and key2. For fields just
2291 * one unit apart, the separator will match key2. key1 is on the left-hand
2292 * side and key2 is on the right-hand side.
2294 * create_tid has to be special cased because a value of 0 represents
2297 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2298 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2301 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2302 hammer_base_elm_t dest)
2304 bzero(dest, sizeof(*dest));
2305 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2306 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2307 MAKE_SEPARATOR(key1, key2, dest, key);
2309 if (key1->obj_id == key2->obj_id &&
2310 key1->rec_type == key2->rec_type &&
2311 key1->key == key2->key) {
2312 if (key1->create_tid == 0) {
2314 * Oops, a create_tid of 0 means 'infinity', so
2315 * if everything matches this just isn't legal.
2317 panic("key1->create_tid of 0 is impossible here");
2318 } else if (key2->create_tid == 0) {
2319 dest->create_tid = key1->create_tid + 1;
2321 MAKE_SEPARATOR(key1, key2, dest, create_tid);
2324 dest->create_tid = 0;
2328 #undef MAKE_SEPARATOR
2331 * Return whether a generic internal or leaf node is full
2334 btree_node_is_full(hammer_node_ondisk_t node)
2336 switch(node->type) {
2337 case HAMMER_BTREE_TYPE_INTERNAL:
2338 if (node->count == HAMMER_BTREE_INT_ELMS)
2341 case HAMMER_BTREE_TYPE_LEAF:
2342 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2346 panic("illegal btree subtype");
2352 * Return whether a generic internal or leaf node is almost full. This
2353 * routine is used as a helper for search insertions to guarentee at
2354 * least 2 available slots in the internal node(s) leading up to a leaf,
2355 * so hammer_btree_insert_cluster() will function properly.
2358 btree_node_is_almost_full(hammer_node_ondisk_t node)
2360 switch(node->type) {
2361 case HAMMER_BTREE_TYPE_INTERNAL:
2362 if (node->count > HAMMER_BTREE_INT_ELMS - 2)
2365 case HAMMER_BTREE_TYPE_LEAF:
2366 if (node->count > HAMMER_BTREE_LEAF_ELMS - 2)
2370 panic("illegal btree subtype");
2377 btree_max_elements(u_int8_t type)
2379 if (type == HAMMER_BTREE_TYPE_LEAF)
2380 return(HAMMER_BTREE_LEAF_ELMS);
2381 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2382 return(HAMMER_BTREE_INT_ELMS);
2383 panic("btree_max_elements: bad type %d\n", type);
2388 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2390 hammer_btree_elm_t elm;
2393 kprintf("node %p count=%d parent=%d type=%c\n",
2394 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2397 * Dump both boundary elements if an internal node
2399 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2400 for (i = 0; i <= ondisk->count; ++i) {
2401 elm = &ondisk->elms[i];
2402 hammer_print_btree_elm(elm, ondisk->type, i);
2405 for (i = 0; i < ondisk->count; ++i) {
2406 elm = &ondisk->elms[i];
2407 hammer_print_btree_elm(elm, ondisk->type, i);
2413 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2416 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
2417 kprintf("\tkey = %016llx\n", elm->base.key);
2418 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2419 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2420 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2421 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2422 kprintf("\tbtype = %02x (%c)\n",
2424 (elm->base.btype ? elm->base.btype : '?'));
2427 case HAMMER_BTREE_TYPE_INTERNAL:
2428 kprintf("\tsubtree_off = %08x\n",
2429 elm->internal.subtree_offset);
2431 case HAMMER_BTREE_TYPE_SPIKE_BEG:
2432 case HAMMER_BTREE_TYPE_SPIKE_END:
2433 kprintf("\tspike_clu_no = %d\n", elm->leaf.spike_clu_no);
2434 kprintf("\tspike_vol_no = %d\n", elm->leaf.spike_vol_no);
2436 case HAMMER_BTREE_TYPE_RECORD:
2437 kprintf("\trec_offset = %08x\n", elm->leaf.rec_offset);
2438 kprintf("\tdata_offset = %08x\n", elm->leaf.data_offset);
2439 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2440 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);