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 1.24 2008/01/25 05:49:08 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 * and key_beg.delete_tid 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 delete_tid set to the asof TID. Due to the
374 * way records are laid out, a forwards iteration may be required if
375 * ENOENT is returned to locate the historical record. Here's the
378 * delete_tid: 10 15 20
382 * Lets say we want to do a lookup AS-OF timestamp 12. We will traverse
383 * LEAF1 but the only record in LEAF1 has a termination (delete_tid) of 11,
384 * thus causing ENOENT to be returned. We really need to check record 18
385 * in LEAF2. If it also fails then the search fails (e.g. it might represent
386 * the range 14-18 and thus still not match our AS-OF timestamp of 12).
388 * btree_search() will set HAMMER_CURSOR_DELETE_CHECK and the
389 * cursor->delete_check TID if an iteration might be needed. In the above
390 * example delete_check would be set to 15.
393 hammer_btree_lookup(hammer_cursor_t cursor)
397 if (cursor->flags & HAMMER_CURSOR_ASOF) {
398 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
399 cursor->key_beg.delete_tid = cursor->asof;
401 cursor->flags &= ~HAMMER_CURSOR_DELETE_CHECK;
402 error = btree_search(cursor, 0);
403 if (error != ENOENT ||
404 (cursor->flags & HAMMER_CURSOR_DELETE_CHECK) == 0) {
407 * Stop if error other then ENOENT.
408 * Stop if ENOENT and not special case.
412 cursor->key_beg.delete_tid = cursor->delete_check;
416 error = btree_search(cursor, 0);
418 if (error == 0 && cursor->flags)
419 error = hammer_btree_extract(cursor, cursor->flags);
424 * Execute the logic required to start an iteration. The first record
425 * located within the specified range is returned and iteration control
426 * flags are adjusted for successive hammer_btree_iterate() calls.
429 hammer_btree_first(hammer_cursor_t cursor)
433 error = hammer_btree_lookup(cursor);
434 if (error == ENOENT) {
435 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
436 error = hammer_btree_iterate(cursor);
438 cursor->flags |= HAMMER_CURSOR_ATEDISK;
443 * Extract the record and/or data associated with the cursor's current
444 * position. Any prior record or data stored in the cursor is replaced.
445 * The cursor must be positioned at a leaf node.
447 * NOTE: Most extractions occur at the leaf of the B-Tree. The only
448 * extraction allowed at an internal element is at a cluster-push.
449 * Cluster-push elements have records but no data.
452 hammer_btree_extract(hammer_cursor_t cursor, int flags)
454 hammer_node_ondisk_t node;
455 hammer_btree_elm_t elm;
456 hammer_cluster_t cluster;
463 * A cluster record type has no data reference, the information
464 * is stored directly in the record and B-Tree element.
466 * The case where the data reference resolves to the same buffer
467 * as the record reference must be handled.
469 node = cursor->node->ondisk;
470 elm = &node->elms[cursor->index];
471 cluster = cursor->node->cluster;
472 cursor->flags &= ~HAMMER_CURSOR_DATA_EMBEDDED;
476 * There is nothing to extract for an internal element.
478 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
481 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
486 if ((flags & HAMMER_CURSOR_GET_RECORD)) {
487 cloff = elm->leaf.rec_offset;
488 cursor->record = hammer_bread(cluster, cloff,
489 HAMMER_FSBUF_RECORDS, &error,
490 &cursor->record_buffer);
495 if ((flags & HAMMER_CURSOR_GET_DATA) && error == 0) {
496 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD) {
498 * Only records have data references. Spike elements
502 } else if ((cloff ^ elm->leaf.data_offset) & ~HAMMER_BUFMASK) {
504 * The data is not in the same buffer as the last
505 * record we cached, but it could still be embedded
506 * in a record. Note that we may not have loaded the
507 * record's buffer above, depending on flags.
509 if ((elm->leaf.rec_offset ^ elm->leaf.data_offset) &
511 if (elm->leaf.data_len & HAMMER_BUFMASK)
512 buf_type = HAMMER_FSBUF_DATA;
514 buf_type = 0; /* pure data buffer */
516 buf_type = HAMMER_FSBUF_RECORDS;
518 cursor->data = hammer_bread(cluster,
519 elm->leaf.data_offset,
521 &cursor->data_buffer);
524 * Data in same buffer as record. Note that we
525 * leave any existing data_buffer intact, even
526 * though we don't use it in this case, in case
527 * other records extracted during an iteration
530 * The data must be embedded in the record for this
533 * Just assume the buffer type is correct.
535 cursor->data = (void *)
536 ((char *)cursor->record_buffer->ondisk +
537 (elm->leaf.data_offset & HAMMER_BUFMASK));
538 roff = (char *)cursor->data - (char *)cursor->record;
539 KKASSERT (roff >= 0 && roff < HAMMER_RECORD_SIZE);
540 cursor->flags |= HAMMER_CURSOR_DATA_EMBEDDED;
548 * Insert a leaf element into the B-Tree at the current cursor position.
549 * The cursor is positioned such that the element at and beyond the cursor
550 * are shifted to make room for the new record.
552 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
553 * flag set and that call must return ENOENT before this function can be
556 * ENOSPC is returned if there is no room to insert a new record.
559 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_elm_t elm)
561 hammer_node_ondisk_t node;
565 if ((error = hammer_cursor_upgrade(cursor)) != 0)
569 * Insert the element at the leaf node and update the count in the
570 * parent. It is possible for parent to be NULL, indicating that
571 * the root of the B-Tree in the cluster is a leaf. It is also
572 * possible for the leaf to be empty.
574 * Remember that the right-hand boundary is not included in the
577 hammer_modify_node(cursor->node);
578 node = cursor->node->ondisk;
580 KKASSERT(elm->base.btype != 0);
581 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
582 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
583 if (i != node->count) {
584 bcopy(&node->elms[i], &node->elms[i+1],
585 (node->count - i) * sizeof(*elm));
587 node->elms[i] = *elm;
591 * Debugging sanity checks. Note that the element to the left
592 * can match the element we are inserting if it is a SPIKE_END,
593 * because spike-end's represent a non-inclusive end to a range.
595 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
596 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
598 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->leaf.base) < 0);
600 if (i != node->count - 1)
601 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->leaf.base) > 0);
607 * Insert a cluster spike into the B-Tree at the current cursor position.
608 * The caller pre-positions the insertion cursor at ncluster's
609 * left bound in the originating cluster. Both the originating cluster
610 * and the target cluster must be serialized, EDEADLK is fatal.
612 * Basically we have to lay down the two spike elements and assert that
613 * the leaf's right bound does not bisect the ending element. The ending
614 * spike element is non-inclusive, just like a boundary. The target cluster's
615 * clu_btree_parent_offset may have to adjusted.
617 * NOTE: Serialization is usually accoplished by virtue of being the
618 * initial accessor of a cluster.
621 hammer_btree_insert_cluster(hammer_cursor_t cursor, hammer_cluster_t ncluster,
624 hammer_node_ondisk_t node;
625 hammer_btree_elm_t elm;
626 hammer_cluster_t ocluster;
627 const int esize = sizeof(*elm);
632 if ((error = hammer_cursor_upgrade(cursor)) != 0)
634 hammer_modify_node(cursor->node);
635 node = cursor->node->ondisk;
636 node_offset = cursor->node->node_offset;
639 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
640 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS - 2);
641 KKASSERT(i >= 0 && i <= node->count);
644 * Make sure the spike is legal or the B-Tree code will get really
647 * XXX the right bound my bisect the two spike elements. We
648 * need code here to 'fix' the right bound going up the tree
649 * instead of an assertion.
651 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_beg,
652 cursor->left_bound) >= 0);
653 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
654 cursor->right_bound) <= 0);
655 if (i != node->count) {
656 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
657 &node->elms[i].leaf.base) <= 0);
660 elm = &node->elms[i];
661 bcopy(elm, elm + 2, (node->count - i) * esize);
662 bzero(elm, 2 * esize);
665 elm[0].leaf.base = ncluster->ondisk->clu_btree_beg;
666 elm[0].leaf.base.btype = HAMMER_BTREE_TYPE_SPIKE_BEG;
667 elm[0].leaf.rec_offset = rec_offset;
668 elm[0].leaf.spike_clu_no = ncluster->clu_no;
669 elm[0].leaf.spike_vol_no = ncluster->volume->vol_no;
671 elm[1].leaf.base = ncluster->ondisk->clu_btree_end;
672 elm[1].leaf.base.btype = HAMMER_BTREE_TYPE_SPIKE_END;
673 elm[1].leaf.rec_offset = rec_offset;
674 elm[1].leaf.spike_clu_no = ncluster->clu_no;
675 elm[1].leaf.spike_vol_no = ncluster->volume->vol_no;
678 * SPIKE_END must be inclusive, not exclusive.
680 KKASSERT(elm[1].leaf.base.delete_tid != 1);
681 --elm[1].leaf.base.delete_tid;
684 * The target cluster's parent offset may have to be updated.
686 * NOTE: Modifying a cluster header does not mark it open, and
687 * flushing it will only clear an existing open flag if the cluster
688 * has been validated.
690 kprintf("INSERT CLUSTER %d:%d -> %d:%d ",
691 ncluster->ondisk->clu_btree_parent_vol_no,
692 ncluster->ondisk->clu_btree_parent_clu_no,
693 ncluster->volume->vol_no,
696 ocluster = cursor->node->cluster;
697 if (ncluster->ondisk->clu_btree_parent_offset != node_offset ||
698 ncluster->ondisk->clu_btree_parent_clu_no != ocluster->clu_no ||
699 ncluster->ondisk->clu_btree_parent_vol_no != ocluster->volume->vol_no) {
700 hammer_modify_cluster(ncluster);
701 ncluster->ondisk->clu_btree_parent_offset = node_offset;
702 ncluster->ondisk->clu_btree_parent_clu_no = ocluster->clu_no;
703 ncluster->ondisk->clu_btree_parent_vol_no = ocluster->volume->vol_no;
704 kprintf("(offset fixup)\n");
706 kprintf("(offset unchanged)\n");
713 * Delete a record from the B-Tree at the current cursor position.
714 * The cursor is positioned such that the current element is the one
717 * On return the cursor will be positioned after the deleted element and
718 * MAY point to an internal node. It will be suitable for the continuation
719 * of an iteration but not for an insertion or deletion.
721 * Deletions will attempt to partially rebalance the B-Tree in an upward
722 * direction, but will terminate rather then deadlock. Empty leaves are
723 * not allowed except at the root node of a cluster. An early termination
724 * will leave an internal node with an element whos subtree_offset is 0,
725 * a case detected and handled by btree_search().
727 * This function can return EDEADLK, requiring the caller to retry the
728 * operation after clearing the deadlock.
731 hammer_btree_delete(hammer_cursor_t cursor)
733 hammer_node_ondisk_t ondisk;
735 hammer_node_t parent;
739 if ((error = hammer_cursor_upgrade(cursor)) != 0)
743 * Delete the element from the leaf node.
745 * Remember that leaf nodes do not have boundaries.
748 ondisk = node->ondisk;
751 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
752 KKASSERT(i >= 0 && i < ondisk->count);
753 hammer_modify_node(node);
754 if (i + 1 != ondisk->count) {
755 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
756 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
761 * Validate local parent
763 if (ondisk->parent) {
764 parent = cursor->parent;
766 KKASSERT(parent != NULL);
767 KKASSERT(parent->node_offset == ondisk->parent);
768 KKASSERT(parent->cluster == node->cluster);
772 * If the leaf becomes empty it must be detached from the parent,
773 * potentially recursing through to the cluster root.
775 * This may reposition the cursor at one of the parent's of the
778 * Ignore deadlock errors, that simply means that btree_remove
779 * was unable to recurse and had to leave the subtree_offset
780 * in the parent set to 0.
782 KKASSERT(cursor->index <= ondisk->count);
783 if (ondisk->count == 0) {
785 error = btree_remove(cursor);
786 } while (error == EAGAIN);
787 if (error == EDEADLK)
792 KKASSERT(cursor->parent == NULL ||
793 cursor->parent_index < cursor->parent->ondisk->count);
798 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
800 * Search a cluster's B-Tree for cursor->key_beg, return the matching node.
802 * The search can begin ANYWHERE in the B-Tree. As a first step the search
803 * iterates up the tree as necessary to properly position itself prior to
804 * actually doing the sarch.
806 * INSERTIONS: The search will split full nodes and leaves on its way down
807 * and guarentee that the leaf it ends up on is not full. If we run out
808 * of space the search continues to the leaf (to position the cursor for
809 * the spike), but ENOSPC is returned.
811 * The search is only guarenteed to end up on a leaf if an error code of 0
812 * is returned, or if inserting and an error code of ENOENT is returned.
813 * Otherwise it can stop at an internal node. On success a search returns
814 * a leaf node unless INCLUSTER is set and the search located a cluster push
815 * node (which is an internal node).
817 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
818 * filesystem, and it is not simple code. Please note the following facts:
820 * - Internal node recursions have a boundary on the left AND right. The
821 * right boundary is non-inclusive. The delete_tid is a generic part
822 * of the key for internal nodes.
824 * - Leaf nodes contain terminal elements AND spikes. A spike recurses into
825 * another cluster and contains two leaf elements.. a beginning and an
826 * ending element. The SPIKE_END element is RANGE-EXCLUSIVE, just like a
827 * boundary. This means that it is possible to have two elements
828 * (a spike ending element and a record) side by side with the same key.
830 * - Because the SPIKE_END element is range inclusive, it cannot match the
831 * right boundary of the parent node. SPIKE_BEG and SPIKE_END elements
832 * always come in pairs, and always exist side by side in the same leaf.
834 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
835 * historical search. ASOF and INSERT are mutually exclusive. When
836 * doing an as-of lookup btree_search() checks for a right-edge boundary
837 * case. If while recursing down the right-edge differs from the key
838 * by ONLY its delete_tid, HAMMER_CURSOR_DELETE_CHECK is set along
839 * with cursor->delete_check. This is used by btree_lookup() to iterate.
840 * The iteration is necessary because as-of searches can wind up going
841 * down the wrong branch of the B-Tree.
845 btree_search(hammer_cursor_t cursor, int flags)
847 hammer_node_ondisk_t node;
848 hammer_cluster_t cluster;
849 hammer_btree_elm_t elm;
856 flags |= cursor->flags;
858 if (hammer_debug_btree) {
859 kprintf("SEARCH %d:%d:%08x[%d] %016llx %02x key=%016llx did=%016llx\n",
860 cursor->node->cluster->volume->vol_no,
861 cursor->node->cluster->clu_no,
862 cursor->node->node_offset,
864 cursor->key_beg.obj_id,
865 cursor->key_beg.rec_type,
867 cursor->key_beg.delete_tid
872 * Move our cursor up the tree until we find a node whos range covers
873 * the key we are trying to locate. This may move us between
876 * The left bound is inclusive, the right bound is non-inclusive.
877 * It is ok to cursor up too far so when cursoring across a cluster
880 * First see if we can skip the whole cluster. hammer_cursor_up()
881 * handles both cases but this way we don't check the cluster
882 * bounds when going up the tree within a cluster.
884 * NOTE: If INCLUSTER is set and we are at the root of the cluster,
885 * hammer_cursor_up() will return ENOENT.
887 cluster = cursor->node->cluster;
889 r = hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_beg);
890 s = hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_end);
894 error = hammer_cursor_toroot(cursor);
897 KKASSERT(cursor->parent);
898 error = hammer_cursor_up(cursor);
901 cluster = cursor->node->cluster;
904 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
905 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
908 KKASSERT(cursor->parent);
909 error = hammer_cursor_up(cursor);
915 * The delete-checks below are based on node, not parent. Set the
916 * initial delete-check based on the parent.
919 cursor->delete_check = cursor->right_bound->delete_tid;
920 cursor->flags |= HAMMER_CURSOR_DELETE_CHECK;
924 * We better have ended up with a node somewhere, and our second
925 * while loop had better not have traversed up a cluster.
927 KKASSERT(cursor->node != NULL && cursor->node->cluster == cluster);
930 * If we are inserting we can't start at a full node if the parent
931 * is also full (because there is no way to split the node),
932 * continue running up the tree until the requirement is satisfied
933 * or we hit the root of the current cluster.
935 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
936 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
937 if (btree_node_is_full(cursor->node->ondisk) == 0)
940 if (btree_node_is_almost_full(cursor->node->ondisk) ==0)
943 if (cursor->node->ondisk->parent == 0 ||
944 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
947 error = hammer_cursor_up(cursor);
948 /* cluster and node are now may become stale */
952 /* cluster = cursor->node->cluster; not needed until next cluster = */
956 * Push down through internal nodes to locate the requested key.
958 cluster = cursor->node->cluster;
959 node = cursor->node->ondisk;
960 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
962 * Scan the node to find the subtree index to push down into.
963 * We go one-past, then back-up.
965 * We must proactively remove deleted elements which may
966 * have been left over from a deadlocked btree_remove().
968 * The left and right boundaries are included in the loop
969 * in order to detect edge cases.
971 * If the separator only differs by delete_tid (r == -1)
972 * and we are doing an as-of search, we may end up going
973 * down a branch to the left of the one containing the
974 * desired key. This requires numerous special cases.
976 if (hammer_debug_btree) {
977 kprintf("SEARCH-I %d:%d:%08x count=%d\n",
978 cursor->node->cluster->volume->vol_no,
979 cursor->node->cluster->clu_no,
980 cursor->node->node_offset,
983 for (i = 0; i <= node->count; ++i) {
984 elm = &node->elms[i];
985 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
986 if (hammer_debug_btree > 2) {
987 kprintf(" IELM %p %d r=%d\n",
988 &node->elms[i], i, r);
992 cursor->delete_check =
993 elm->base.delete_tid;
995 HAMMER_CURSOR_DELETE_CHECK;
1000 if (hammer_debug_btree) {
1001 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1006 * These cases occur when the parent's idea of the boundary
1007 * is wider then the child's idea of the boundary, and
1008 * require special handling. If not inserting we can
1009 * terminate the search early for these cases but the
1010 * child's boundaries cannot be unconditionally modified.
1014 * If i == 0 the search terminated to the LEFT of the
1015 * left_boundary but to the RIGHT of the parent's left
1020 elm = &node->elms[0];
1023 * If we aren't inserting we can stop here.
1025 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1031 * Correct a left-hand boundary mismatch.
1033 * We can only do this if we can upgrade the lock.
1035 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1037 hammer_modify_node(cursor->node);
1038 save = node->elms[0].base.btype;
1039 node->elms[0].base = *cursor->left_bound;
1040 node->elms[0].base.btype = save;
1041 } else if (i == node->count + 1) {
1043 * If i == node->count + 1 the search terminated to
1044 * the RIGHT of the right boundary but to the LEFT
1045 * of the parent's right boundary. If we aren't
1046 * inserting we can stop here.
1048 * Note that the last element in this case is
1049 * elms[i-2] prior to adjustments to 'i'.
1052 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1058 * Correct a right-hand boundary mismatch.
1059 * (actual push-down record is i-2 prior to
1060 * adjustments to i).
1062 * We can only do this if we can upgrade the lock.
1064 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1066 elm = &node->elms[i];
1067 hammer_modify_node(cursor->node);
1068 elm->base = *cursor->right_bound;
1072 * The push-down index is now i - 1. If we had
1073 * terminated on the right boundary this will point
1074 * us at the last element.
1079 elm = &node->elms[i];
1081 if (hammer_debug_btree) {
1082 kprintf("RESULT-I %d:%d:%08x[%d] %016llx %02x "
1083 "key=%016llx did=%016llx\n",
1084 cursor->node->cluster->volume->vol_no,
1085 cursor->node->cluster->clu_no,
1086 cursor->node->node_offset,
1088 elm->internal.base.obj_id,
1089 elm->internal.base.rec_type,
1090 elm->internal.base.key,
1091 elm->internal.base.delete_tid
1096 * When searching try to clean up any deleted
1097 * internal elements left over from btree_remove()
1100 * If we fail and we are doing an insertion lookup,
1101 * we have to return EDEADLK, because an insertion lookup
1102 * must terminate at a leaf.
1104 if (elm->internal.subtree_offset == 0) {
1105 error = btree_remove_deleted_element(cursor);
1108 if (error == EDEADLK &&
1109 (flags & HAMMER_CURSOR_INSERT) == 0) {
1117 * Handle insertion and deletion requirements.
1119 * If inserting split full nodes. The split code will
1120 * adjust cursor->node and cursor->index if the current
1121 * index winds up in the new node.
1123 * If inserting and a left or right edge case was detected,
1124 * we cannot correct the left or right boundary and must
1125 * prepend and append an empty leaf node in order to make
1126 * the boundary correction.
1128 * If we run out of space we set enospc and continue on
1129 * to a leaf to provide the spike code with a good point
1130 * of entry. Enospc is reset if we cross a cluster boundary.
1132 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1133 if (btree_node_is_full(node)) {
1134 error = btree_split_internal(cursor);
1136 if (error != ENOSPC)
1141 * reload stale pointers
1144 node = cursor->node->ondisk;
1149 * Push down (push into new node, existing node becomes
1150 * the parent) and continue the search.
1152 error = hammer_cursor_down(cursor);
1153 /* node and cluster become stale */
1156 node = cursor->node->ondisk;
1157 cluster = cursor->node->cluster;
1161 * We are at a leaf, do a linear search of the key array.
1163 * If we encounter a spike element type within the necessary
1164 * range we push into it. Note that SPIKE_END is non-inclusive
1165 * of the spike range.
1167 * On success the index is set to the matching element and 0
1170 * On failure the index is set to the insertion point and ENOENT
1173 * Boundaries are not stored in leaf nodes, so the index can wind
1174 * up to the left of element 0 (index == 0) or past the end of
1175 * the array (index == node->count).
1177 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1178 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1179 if (hammer_debug_btree) {
1180 kprintf("SEARCH-L %d:%d:%08x count=%d\n",
1181 cursor->node->cluster->volume->vol_no,
1182 cursor->node->cluster->clu_no,
1183 cursor->node->node_offset,
1187 for (i = 0; i < node->count; ++i) {
1188 elm = &node->elms[i];
1190 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1192 if (hammer_debug_btree > 1)
1193 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1195 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_BEG) {
1197 * SPIKE_BEG. Stop if we are to the left of the
1198 * spike begin element.
1200 * If we are not the last element in the leaf continue
1201 * the loop looking for the SPIKE_END. If we are
1202 * the last element, however, then push into the
1205 * If doing an as-of search a Spike demark on a
1206 * delete_tid boundary must be pushed into and an
1207 * iteration will be forced if it turned out to be
1210 * If not doing an as-of search exact comparisons
1213 * enospc must be reset because we have crossed a
1218 * Set the delete check if the stop element
1219 * only differs by its delete_tid.
1222 cursor->delete_check =
1223 elm->base.delete_tid;
1225 HAMMER_CURSOR_DELETE_CHECK;
1229 if (i != node->count - 1)
1231 panic("btree_search: illegal spike, no SPIKE_END "
1232 "in leaf node! %p\n", cursor->node);
1235 * XXX This is not currently legal, you can only
1236 * cursor_down() from a SPIKE_END element, otherwise
1237 * the cursor parent is pointing at the wrong element
1240 if (flags & HAMMER_CURSOR_INCLUSTER)
1243 error = hammer_cursor_down(cursor);
1250 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_END) {
1252 * SPIKE_END. We can only hit this case if we are
1253 * greater or equal to SPIKE_BEG.
1255 * If we are <= SPIKE_END we must push into
1256 * it, otherwise continue the search. The SPIKE_END
1257 * element is range-inclusive.
1259 * enospc must be reset because we have crossed a
1266 * We're trying to recurse, set the delete check
1267 * if the right boundary only differs by its
1268 * delete_tid and delete_tid is not 0. Remember
1269 * the spike_end is inclusive, so we have to set
1270 * delete_check one past. A delete_tid of 0
1271 * represents infinity and cannot be incremented.
1273 * We also have to set it for an exact match,
1274 * because the SPIKE_END is still an (inclusive)
1275 * boundary, not a match.
1277 if ((r == 0 || r == -1) && elm->base.delete_tid != 0) {
1278 cursor->delete_check = elm->base.delete_tid + 1;
1279 cursor->flags |= HAMMER_CURSOR_DELETE_CHECK;
1281 if (flags & HAMMER_CURSOR_INCLUSTER)
1284 error = hammer_cursor_down(cursor);
1292 * We are at a record element. Stop if we've flipped past
1293 * key_beg, not counting the delete_tid test.
1295 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1303 * Check our as-of timestamp against the element. A
1304 * delete_tid that matches exactly in an as-of search
1305 * is actually a no-match (the record was deleted as-of
1306 * our timestamp so it isn't visible).
1308 if (flags & HAMMER_CURSOR_ASOF) {
1309 if (hammer_btree_chkts(cursor->asof,
1310 &node->elms[i].base) != 0) {
1315 if (r < 0) /* can only be -1 */
1322 if (hammer_debug_btree) {
1323 kprintf("RESULT-L %d:%d:%08x[%d] (SUCCESS)\n",
1324 cursor->node->cluster->volume->vol_no,
1325 cursor->node->cluster->clu_no,
1326 cursor->node->node_offset,
1333 * The search of the leaf node failed. i is the insertion point.
1336 if (hammer_debug_btree) {
1337 kprintf("RESULT-L %d:%d:%08x[%d] (FAILED)\n",
1338 cursor->node->cluster->volume->vol_no,
1339 cursor->node->cluster->clu_no,
1340 cursor->node->node_offset,
1345 * No exact match was found, i is now at the insertion point.
1347 * If inserting split a full leaf before returning. This
1348 * may have the side effect of adjusting cursor->node and
1351 * For now the leaf must have at least 2 free elements to accomodate
1352 * the insertion of a spike during recovery. See the
1353 * hammer_btree_insert_cluster() function.
1356 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1357 btree_node_is_almost_full(node)) {
1358 error = btree_split_leaf(cursor);
1360 if (error != ENOSPC)
1365 * reload stale pointers
1369 node = &cursor->node->internal;
1374 * We reached a leaf but did not find the key we were looking for.
1375 * If this is an insert we will be properly positioned for an insert
1376 * (ENOENT) or spike (ENOSPC) operation.
1378 error = enospc ? ENOSPC : ENOENT;
1384 /************************************************************************
1385 * SPLITTING AND MERGING *
1386 ************************************************************************
1388 * These routines do all the dirty work required to split and merge nodes.
1392 * Split an internal node into two nodes and move the separator at the split
1393 * point to the parent.
1395 * (cursor->node, cursor->index) indicates the element the caller intends
1396 * to push into. We will adjust node and index if that element winds
1397 * up in the split node.
1399 * If we are at the root of a cluster a new root must be created with two
1400 * elements, one pointing to the original root and one pointing to the
1401 * newly allocated split node.
1403 * NOTE! Being at the root of a cluster is different from being at the
1404 * root of the root cluster. cursor->parent will not be NULL and
1405 * cursor->node->ondisk.parent must be tested against 0. Theoretically
1406 * we could propogate the algorithm into the parent and deal with multiple
1407 * 'roots' in the cluster header, but it's easier not to.
1411 btree_split_internal(hammer_cursor_t cursor)
1413 hammer_node_ondisk_t ondisk;
1415 hammer_node_t parent;
1416 hammer_node_t new_node;
1417 hammer_btree_elm_t elm;
1418 hammer_btree_elm_t parent_elm;
1419 hammer_node_locklist_t locklist = NULL;
1425 const int esize = sizeof(*elm);
1427 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1429 if ((cursor->flags & HAMMER_CURSOR_RECOVER) == 0) {
1430 error = hammer_btree_lock_children(cursor, &locklist);
1436 * We are splitting but elms[split] will be promoted to the parent,
1437 * leaving the right hand node with one less element. If the
1438 * insertion point will be on the left-hand side adjust the split
1439 * point to give the right hand side one additional node.
1441 node = cursor->node;
1442 ondisk = node->ondisk;
1443 split = (ondisk->count + 1) / 2;
1444 if (cursor->index <= split)
1448 * If we are at the root of the cluster, create a new root node with
1449 * 1 element and split normally. Avoid making major modifications
1450 * until we know the whole operation will work.
1452 * The root of the cluster is different from the root of the root
1453 * cluster. Use the node's on-disk structure's parent offset to
1456 if (ondisk->parent == 0) {
1457 parent = hammer_alloc_btree(node->cluster, &error);
1460 hammer_lock_ex(&parent->lock);
1461 hammer_modify_node(parent);
1462 ondisk = parent->ondisk;
1465 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1466 ondisk->elms[0].base = node->cluster->clu_btree_beg;
1467 ondisk->elms[0].base.btype = node->ondisk->type;
1468 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1469 ondisk->elms[1].base = node->cluster->clu_btree_end;
1470 /* ondisk->elms[1].base.btype - not used */
1472 parent_index = 0; /* index of current node in parent */
1475 parent = cursor->parent;
1476 parent_index = cursor->parent_index;
1477 KKASSERT(parent->cluster == node->cluster);
1481 * Split node into new_node at the split point.
1483 * B O O O P N N B <-- P = node->elms[split]
1484 * 0 1 2 3 4 5 6 <-- subtree indices
1489 * B O O O B B N N B <--- inner boundary points are 'P'
1493 new_node = hammer_alloc_btree(node->cluster, &error);
1494 if (new_node == NULL) {
1496 hammer_unlock(&parent->lock);
1497 parent->flags |= HAMMER_NODE_DELETED;
1498 hammer_rel_node(parent);
1502 hammer_lock_ex(&new_node->lock);
1505 * Create the new node. P becomes the left-hand boundary in the
1506 * new node. Copy the right-hand boundary as well.
1508 * elm is the new separator.
1510 hammer_modify_node(new_node);
1511 hammer_modify_node(node);
1512 ondisk = node->ondisk;
1513 elm = &ondisk->elms[split];
1514 bcopy(elm, &new_node->ondisk->elms[0],
1515 (ondisk->count - split + 1) * esize);
1516 new_node->ondisk->count = ondisk->count - split;
1517 new_node->ondisk->parent = parent->node_offset;
1518 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1519 KKASSERT(ondisk->type == new_node->ondisk->type);
1522 * Cleanup the original node. Elm (P) becomes the new boundary,
1523 * its subtree_offset was moved to the new node. If we had created
1524 * a new root its parent pointer may have changed.
1526 elm->internal.subtree_offset = 0;
1527 ondisk->count = split;
1530 * Insert the separator into the parent, fixup the parent's
1531 * reference to the original node, and reference the new node.
1532 * The separator is P.
1534 * Remember that base.count does not include the right-hand boundary.
1536 hammer_modify_node(parent);
1537 ondisk = parent->ondisk;
1538 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1539 parent_elm = &ondisk->elms[parent_index+1];
1540 bcopy(parent_elm, parent_elm + 1,
1541 (ondisk->count - parent_index) * esize);
1542 parent_elm->internal.base = elm->base; /* separator P */
1543 parent_elm->internal.base.btype = new_node->ondisk->type;
1544 parent_elm->internal.subtree_offset = new_node->node_offset;
1548 * The children of new_node need their parent pointer set to new_node.
1549 * The children have already been locked by
1550 * hammer_btree_lock_children().
1552 for (i = 0; i < new_node->ondisk->count; ++i) {
1553 elm = &new_node->ondisk->elms[i];
1554 error = btree_set_parent(new_node, elm);
1556 panic("btree_split_internal: btree-fixup problem");
1561 * The cluster's root pointer may have to be updated.
1564 hammer_modify_cluster(node->cluster);
1565 node->cluster->ondisk->clu_btree_root = parent->node_offset;
1566 node->ondisk->parent = parent->node_offset;
1567 if (cursor->parent) {
1568 hammer_unlock(&cursor->parent->lock);
1569 hammer_rel_node(cursor->parent);
1571 cursor->parent = parent; /* lock'd and ref'd */
1576 * Ok, now adjust the cursor depending on which element the original
1577 * index was pointing at. If we are >= the split point the push node
1578 * is now in the new node.
1580 * NOTE: If we are at the split point itself we cannot stay with the
1581 * original node because the push index will point at the right-hand
1582 * boundary, which is illegal.
1584 * NOTE: The cursor's parent or parent_index must be adjusted for
1585 * the case where a new parent (new root) was created, and the case
1586 * where the cursor is now pointing at the split node.
1588 if (cursor->index >= split) {
1589 cursor->parent_index = parent_index + 1;
1590 cursor->index -= split;
1591 hammer_unlock(&cursor->node->lock);
1592 hammer_rel_node(cursor->node);
1593 cursor->node = new_node; /* locked and ref'd */
1595 cursor->parent_index = parent_index;
1596 hammer_unlock(&new_node->lock);
1597 hammer_rel_node(new_node);
1601 * Fixup left and right bounds
1603 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1604 cursor->left_bound = &parent_elm[0].internal.base;
1605 cursor->right_bound = &parent_elm[1].internal.base;
1606 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1607 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1608 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1609 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1612 hammer_btree_unlock_children(&locklist);
1613 hammer_cursor_downgrade(cursor);
1618 * Same as the above, but splits a full leaf node.
1624 btree_split_leaf(hammer_cursor_t cursor)
1626 hammer_node_ondisk_t ondisk;
1627 hammer_node_t parent;
1629 hammer_node_t new_leaf;
1630 hammer_btree_elm_t elm;
1631 hammer_btree_elm_t parent_elm;
1632 hammer_base_elm_t mid_boundary;
1633 hammer_node_locklist_t locklist = NULL;
1639 const size_t esize = sizeof(*elm);
1641 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1643 if ((cursor->flags & HAMMER_CURSOR_RECOVER) == 0) {
1644 error = hammer_btree_lock_children(cursor, &locklist);
1650 * Calculate the split point. If the insertion point will be on
1651 * the left-hand side adjust the split point to give the right
1652 * hand side one additional node.
1654 * Spikes are made up of two leaf elements which cannot be
1657 leaf = cursor->node;
1658 ondisk = leaf->ondisk;
1659 split = (ondisk->count + 1) / 2;
1660 if (cursor->index <= split)
1664 elm = &ondisk->elms[split];
1665 if (elm->leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_END) {
1667 elm[-1].leaf.base.btype == HAMMER_BTREE_TYPE_SPIKE_BEG);
1672 * If we are at the root of the tree, create a new root node with
1673 * 1 element and split normally. Avoid making major modifications
1674 * until we know the whole operation will work.
1676 if (ondisk->parent == 0) {
1677 parent = hammer_alloc_btree(leaf->cluster, &error);
1680 hammer_lock_ex(&parent->lock);
1681 hammer_modify_node(parent);
1682 ondisk = parent->ondisk;
1685 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1686 ondisk->elms[0].base = leaf->cluster->clu_btree_beg;
1687 ondisk->elms[0].base.btype = leaf->ondisk->type;
1688 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1689 ondisk->elms[1].base = leaf->cluster->clu_btree_end;
1690 /* ondisk->elms[1].base.btype = not used */
1692 parent_index = 0; /* insertion point in parent */
1695 parent = cursor->parent;
1696 parent_index = cursor->parent_index;
1697 KKASSERT(parent->cluster == leaf->cluster);
1701 * Split leaf into new_leaf at the split point. Select a separator
1702 * value in-between the two leafs but with a bent towards the right
1703 * leaf since comparisons use an 'elm >= separator' inequality.
1712 new_leaf = hammer_alloc_btree(leaf->cluster, &error);
1713 if (new_leaf == NULL) {
1715 hammer_unlock(&parent->lock);
1716 parent->flags |= HAMMER_NODE_DELETED;
1717 hammer_rel_node(parent);
1721 hammer_lock_ex(&new_leaf->lock);
1724 * Create the new node. P (elm) become the left-hand boundary in the
1725 * new node. Copy the right-hand boundary as well.
1727 hammer_modify_node(leaf);
1728 hammer_modify_node(new_leaf);
1729 ondisk = leaf->ondisk;
1730 elm = &ondisk->elms[split];
1731 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1732 new_leaf->ondisk->count = ondisk->count - split;
1733 new_leaf->ondisk->parent = parent->node_offset;
1734 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1735 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1738 * Cleanup the original node. Because this is a leaf node and
1739 * leaf nodes do not have a right-hand boundary, there
1740 * aren't any special edge cases to clean up. We just fixup the
1743 ondisk->count = split;
1746 * Insert the separator into the parent, fixup the parent's
1747 * reference to the original node, and reference the new node.
1748 * The separator is P.
1750 * Remember that base.count does not include the right-hand boundary.
1751 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1753 hammer_modify_node(parent);
1754 ondisk = parent->ondisk;
1755 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1756 parent_elm = &ondisk->elms[parent_index+1];
1757 bcopy(parent_elm, parent_elm + 1,
1758 (ondisk->count - parent_index) * esize);
1761 * Create the separator. XXX At the moment use exactly the
1762 * right-hand element if this is a recovery operation in order
1763 * to guarantee that it does not bisect the spike elements in a
1764 * later call to hammer_btree_insert_cluster().
1766 if (cursor->flags & HAMMER_CURSOR_RECOVER) {
1767 parent_elm->base = elm[0].base;
1769 hammer_make_separator(&elm[-1].base, &elm[0].base,
1772 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1773 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1774 mid_boundary = &parent_elm->base;
1778 * The children of new_leaf need their parent pointer set to new_leaf.
1779 * The children have already been locked by btree_lock_children().
1781 * The leaf's elements are either TYPE_RECORD or TYPE_SPIKE_*. Only
1782 * elements of BTREE_TYPE_SPIKE_END really requires any action.
1784 for (i = 0; i < new_leaf->ondisk->count; ++i) {
1785 elm = &new_leaf->ondisk->elms[i];
1786 error = btree_set_parent(new_leaf, elm);
1788 panic("btree_split_internal: btree-fixup problem");
1793 * The cluster's root pointer may have to be updated.
1796 hammer_modify_cluster(leaf->cluster);
1797 leaf->cluster->ondisk->clu_btree_root = parent->node_offset;
1798 leaf->ondisk->parent = parent->node_offset;
1799 if (cursor->parent) {
1800 hammer_unlock(&cursor->parent->lock);
1801 hammer_rel_node(cursor->parent);
1803 cursor->parent = parent; /* lock'd and ref'd */
1807 * Ok, now adjust the cursor depending on which element the original
1808 * index was pointing at. If we are >= the split point the push node
1809 * is now in the new node.
1811 * NOTE: If we are at the split point itself we need to select the
1812 * old or new node based on where key_beg's insertion point will be.
1813 * If we pick the wrong side the inserted element will wind up in
1814 * the wrong leaf node and outside that node's bounds.
1816 if (cursor->index > split ||
1817 (cursor->index == split &&
1818 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1819 cursor->parent_index = parent_index + 1;
1820 cursor->index -= split;
1821 hammer_unlock(&cursor->node->lock);
1822 hammer_rel_node(cursor->node);
1823 cursor->node = new_leaf;
1825 cursor->parent_index = parent_index;
1826 hammer_unlock(&new_leaf->lock);
1827 hammer_rel_node(new_leaf);
1831 * Fixup left and right bounds
1833 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1834 cursor->left_bound = &parent_elm[0].internal.base;
1835 cursor->right_bound = &parent_elm[1].internal.base;
1838 * Note: The right assertion is typically > 0, but if the last element
1839 * is a SPIKE_END it can be == 0 because the spike-end is non-inclusive
1840 * of the range being spiked.
1842 * This may seem a bit odd but it works.
1844 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1845 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1846 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1847 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) >= 0);
1850 hammer_btree_unlock_children(&locklist);
1851 hammer_cursor_downgrade(cursor);
1856 * Attempt to remove the empty B-Tree node at (cursor->node). Returns 0
1857 * on success, EAGAIN if we could not acquire the necessary locks, or some
1858 * other error. This node can be a leaf node or an internal node.
1860 * On return the cursor may end up pointing at an internal node, suitable
1861 * for further iteration but not for an immediate insertion or deletion.
1863 * cursor->node may be an internal node or a leaf node.
1865 * NOTE: If cursor->node has one element it is the parent trying to delete
1866 * that element, make sure cursor->index is properly adjusted on success.
1869 btree_remove(hammer_cursor_t cursor)
1871 hammer_node_ondisk_t ondisk;
1872 hammer_btree_elm_t elm;
1875 hammer_node_t parent;
1876 const int esize = sizeof(*elm);
1880 * If we are at the root of the cluster we must be able to
1881 * successfully delete the HAMMER_BTREE_SPIKE_* leaf elements in
1882 * the parent in order to be able to destroy the cluster.
1884 node = cursor->node;
1886 if (node->ondisk->parent == 0) {
1887 hammer_modify_node(node);
1888 ondisk = node->ondisk;
1889 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1895 * When trying to delete a cluster we need to exclusively
1896 * lock the cluster root, its parent (leaf in parent cluster),
1897 * AND the parent of that leaf if it's going to be empty,
1898 * because we can't leave around an empty leaf.
1900 * XXX this is messy due to potentially recursive locks.
1901 * downgrade the cursor, get a second shared lock on the
1902 * node that cannot deadlock because we only own shared locks
1903 * then, cursor-up, and re-upgrade everything. If the
1904 * upgrades EDEADLK then don't try to remove the cluster
1907 if ((parent = cursor->parent) != NULL) {
1908 hammer_cursor_downgrade(cursor);
1910 hammer_ref_node(save);
1911 hammer_lock_sh(&save->lock);
1914 * After the cursor up save has the empty root node
1915 * of the target cluster to be deleted, cursor->node
1916 * is at the leaf containing the spikes, and
1917 * cursor->parent is the parent of that leaf.
1919 * cursor->node and cursor->parent are both in the
1920 * parent cluster of the cluster being deleted.
1922 error = hammer_cursor_up(cursor);
1925 error = hammer_cursor_upgrade(cursor);
1927 error = hammer_lock_upgrade(&save->lock);
1930 /* may be EDEADLK */
1931 kprintf("BTREE_REMOVE: Cannot delete cluster\n");
1932 Debugger("BTREE_REMOVE");
1935 * cursor->node is now the leaf in the parent
1936 * cluster containing the spike elements.
1938 * The cursor should be pointing at the
1939 * SPIKE_END element.
1941 * Remove the spike elements and recurse
1942 * if the leaf becomes empty.
1944 node = cursor->node;
1945 hammer_modify_node(node);
1946 ondisk = node->ondisk;
1947 KKASSERT(cursor->index > 0);
1949 elm = &ondisk->elms[cursor->index];
1950 KKASSERT(elm[0].leaf.base.btype ==
1951 HAMMER_BTREE_TYPE_SPIKE_BEG);
1952 KKASSERT(elm[1].leaf.base.btype ==
1953 HAMMER_BTREE_TYPE_SPIKE_END);
1956 * Ok, remove it and the underlying record.
1958 hammer_free_record(node->cluster,
1959 elm->leaf.rec_offset,
1960 HAMMER_RECTYPE_CLUSTER);
1961 bcopy(elm + 2, elm, (ondisk->count -
1962 cursor->index - 2) * esize);
1964 save->flags |= HAMMER_NODE_DELETED;
1965 save->cluster->flags |= HAMMER_CLUSTER_DELETED;
1966 hammer_flush_node(save);
1967 hammer_unlock(&save->lock);
1968 hammer_rel_node(save);
1969 if (ondisk->count == 0)
1977 * Zero-out the parent's reference to the child and flag the
1978 * child for destruction. This ensures that the child is not
1979 * reused while other references to it exist.
1981 parent = cursor->parent;
1982 hammer_modify_node(parent);
1983 ondisk = parent->ondisk;
1984 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
1985 elm = &ondisk->elms[cursor->parent_index];
1986 KKASSERT(elm->internal.subtree_offset == node->node_offset);
1987 elm->internal.subtree_offset = 0;
1989 hammer_flush_node(node);
1990 node->flags |= HAMMER_NODE_DELETED;
1993 * If the parent would otherwise not become empty we can physically
1994 * remove the zero'd element. Note however that in order to
1995 * guarentee a valid cursor we still need to be able to cursor up
1996 * because we no longer have a node.
1998 * This collapse will change the parent's boundary elements, making
1999 * them wider. The new boundaries are recursively corrected in
2002 * XXX we can theoretically recalculate the midpoint but there isn't
2003 * much of a reason to do it.
2005 error = hammer_cursor_up(cursor);
2007 error = hammer_cursor_upgrade(cursor);
2010 kprintf("BTREE_REMOVE: Cannot lock parent, skipping\n");
2011 Debugger("BTREE_REMOVE");
2016 * Remove the internal element from the parent. The bcopy must
2017 * include the right boundary element.
2019 KKASSERT(parent == cursor->node && ondisk == parent->ondisk);
2022 /* ondisk is node's ondisk */
2023 /* elm is node's element */
2026 * Remove the internal element that we zero'd out. Tell the caller
2027 * to loop if it hits zero (to try to avoid eating up precious kernel
2030 KKASSERT(ondisk->count > 0);
2031 bcopy(&elm[1], &elm[0], (ondisk->count - cursor->index) * esize);
2033 if (ondisk->count == 0)
2039 * Attempt to remove the deleted internal element at the current cursor
2040 * position. If we are unable to remove the element we return EDEADLK.
2042 * If the current internal node becomes empty we delete it in the parent
2043 * and cursor up, looping until we finish or we deadlock.
2045 * On return, if successful, the cursor will be pointing at the next
2046 * iterative position in the B-Tree. If unsuccessful the cursor will be
2047 * pointing at the last deleted internal element that could not be
2052 btree_remove_deleted_element(hammer_cursor_t cursor)
2055 hammer_btree_elm_t elm;
2058 if ((error = hammer_cursor_upgrade(cursor)) != 0)
2060 node = cursor->node;
2061 elm = &node->ondisk->elms[cursor->index];
2062 if (elm->internal.subtree_offset == 0) {
2064 error = btree_remove(cursor);
2065 kprintf("BTREE REMOVE DELETED ELEMENT %d\n", error);
2066 } while (error == EAGAIN);
2072 * The element (elm) has been moved to a new internal node (node).
2074 * If the element represents a pointer to an internal node that node's
2075 * parent must be adjusted to the element's new location.
2077 * If the element represents a spike the target cluster's header must
2078 * be adjusted to point to the element's new location. This only
2079 * applies to HAMMER_SPIKE_END.
2081 * GET_CLUSTER_NORECOVER must be used to avoid a recovery recursion during
2082 * the rebuild of the recovery cluster's B-Tree, which can blow the kernel
2085 * XXX deadlock potential here with our exclusive locks
2089 btree_set_parent(hammer_node_t node, hammer_btree_elm_t elm)
2091 hammer_volume_t volume;
2092 hammer_cluster_t cluster;
2093 hammer_node_t child;
2098 switch(elm->base.btype) {
2099 case HAMMER_BTREE_TYPE_INTERNAL:
2100 case HAMMER_BTREE_TYPE_LEAF:
2101 child = hammer_get_node(node->cluster,
2102 elm->internal.subtree_offset, &error);
2104 hammer_modify_node(child);
2105 child->ondisk->parent = node->node_offset;
2106 hammer_rel_node(child);
2109 case HAMMER_BTREE_TYPE_SPIKE_END:
2110 volume = hammer_get_volume(node->cluster->volume->hmp,
2111 elm->leaf.spike_vol_no, &error);
2114 cluster = hammer_get_cluster(volume, elm->leaf.spike_clu_no,
2115 &error, GET_CLUSTER_NORECOVER);
2116 hammer_rel_volume(volume, 0);
2119 hammer_modify_cluster(cluster);
2120 cluster->ondisk->clu_btree_parent_offset = node->node_offset;
2121 KKASSERT(cluster->ondisk->clu_btree_parent_clu_no ==
2122 node->cluster->clu_no);
2123 KKASSERT(cluster->ondisk->clu_btree_parent_vol_no ==
2124 node->cluster->volume->vol_no);
2125 hammer_rel_cluster(cluster, 0);
2134 * Exclusively lock all the children of node. This is used by the split
2135 * code to prevent anyone from accessing the children of a cursor node
2136 * while we fix-up its parent offset.
2138 * If we don't lock the children we can really mess up cursors which block
2139 * trying to cursor-up into our node.
2141 * WARNING: Cannot be used when doing B-tree operations on a recovery
2142 * cluster because the target cluster may require recovery, resulting
2143 * in a deep recursion which blows the kernel stack.
2145 * On failure EDEADLK (or some other error) is returned. If a deadlock
2146 * error is returned the cursor is adjusted to block on termination.
2149 hammer_btree_lock_children(hammer_cursor_t cursor,
2150 struct hammer_node_locklist **locklistp)
2153 hammer_node_locklist_t item;
2154 hammer_node_ondisk_t ondisk;
2155 hammer_btree_elm_t elm;
2156 hammer_volume_t volume;
2157 hammer_cluster_t cluster;
2158 hammer_node_t child;
2162 node = cursor->node;
2163 ondisk = node->ondisk;
2165 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2166 elm = &ondisk->elms[i];
2169 switch(elm->base.btype) {
2170 case HAMMER_BTREE_TYPE_INTERNAL:
2171 case HAMMER_BTREE_TYPE_LEAF:
2172 child = hammer_get_node(node->cluster,
2173 elm->internal.subtree_offset,
2176 case HAMMER_BTREE_TYPE_SPIKE_END:
2177 volume = hammer_get_volume(node->cluster->volume->hmp,
2178 elm->leaf.spike_vol_no,
2182 cluster = hammer_get_cluster(volume,
2183 elm->leaf.spike_clu_no,
2186 hammer_rel_volume(volume, 0);
2189 KKASSERT(cluster->ondisk->clu_btree_root != 0);
2190 child = hammer_get_node(cluster,
2191 cluster->ondisk->clu_btree_root,
2193 hammer_rel_cluster(cluster, 0);
2199 if (hammer_lock_ex_try(&child->lock) != 0) {
2200 if (cursor->deadlk_node == NULL) {
2201 cursor->deadlk_node = node;
2202 hammer_ref_node(cursor->deadlk_node);
2206 item = kmalloc(sizeof(*item),
2207 M_HAMMER, M_WAITOK);
2208 item->next = *locklistp;
2215 hammer_btree_unlock_children(locklistp);
2221 * Release previously obtained node locks.
2224 hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2226 hammer_node_locklist_t item;
2228 while ((item = *locklistp) != NULL) {
2229 *locklistp = item->next;
2230 hammer_unlock(&item->node->lock);
2231 hammer_rel_node(item->node);
2232 kfree(item, M_HAMMER);
2236 /************************************************************************
2237 * MISCELLANIOUS SUPPORT *
2238 ************************************************************************/
2241 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2243 * Note that for this particular function a return value of -1, 0, or +1
2244 * can denote a match if delete_tid is otherwise discounted. A delete_tid
2245 * of zero is considered to be 'infinity' in comparisons.
2247 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2250 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2252 if (key1->obj_id < key2->obj_id)
2254 if (key1->obj_id > key2->obj_id)
2257 if (key1->rec_type < key2->rec_type)
2259 if (key1->rec_type > key2->rec_type)
2262 if (key1->key < key2->key)
2264 if (key1->key > key2->key)
2268 * A delete_tid of zero indicates a record which has not been
2269 * deleted yet and must be considered to have a value of positive
2272 if (key1->delete_tid == 0) {
2273 if (key2->delete_tid == 0)
2277 if (key2->delete_tid == 0)
2279 if (key1->delete_tid < key2->delete_tid)
2281 if (key1->delete_tid > key2->delete_tid)
2287 * Test a timestamp against an element to determine whether the
2288 * element is visible. A timestamp of 0 means 'infinity'.
2291 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2294 if (base->delete_tid)
2298 if (asof < base->create_tid)
2300 if (base->delete_tid && asof >= base->delete_tid)
2306 * Create a separator half way inbetween key1 and key2. For fields just
2307 * one unit apart, the separator will match key2. key1 is on the left-hand
2308 * side and key2 is on the right-hand side.
2310 * delete_tid has to be special cased because a value of 0 represents
2311 * infinity, and records with a delete_tid of 0 can be replaced with
2312 * a non-zero delete_tid when deleted and must maintain their proper
2313 * (as in the same) position in the B-Tree.
2315 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2316 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2319 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2320 hammer_base_elm_t dest)
2322 bzero(dest, sizeof(*dest));
2323 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2324 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2325 MAKE_SEPARATOR(key1, key2, dest, key);
2327 if (key1->obj_id == key2->obj_id &&
2328 key1->rec_type == key2->rec_type &&
2329 key1->key == key2->key) {
2330 if (key1->delete_tid == 0) {
2332 * Oops, a delete_tid of 0 means 'infinity', so
2333 * if everything matches this just isn't legal.
2335 panic("key1->delete_tid of 0 is impossible here");
2337 KKASSERT(key1->btype == HAMMER_BTREE_TYPE_SPIKE_END);
2338 dest->delete_tid = key1->delete_tid;
2340 } else if (key2->delete_tid == 0) {
2341 dest->delete_tid = key1->delete_tid + 1;
2343 MAKE_SEPARATOR(key1, key2, dest, delete_tid);
2346 dest->delete_tid = 0;
2350 #undef MAKE_SEPARATOR
2353 * Return whether a generic internal or leaf node is full
2356 btree_node_is_full(hammer_node_ondisk_t node)
2358 switch(node->type) {
2359 case HAMMER_BTREE_TYPE_INTERNAL:
2360 if (node->count == HAMMER_BTREE_INT_ELMS)
2363 case HAMMER_BTREE_TYPE_LEAF:
2364 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2368 panic("illegal btree subtype");
2374 * Return whether a generic internal or leaf node is almost full. This
2375 * routine is used as a helper for search insertions to guarentee at
2376 * least 2 available slots in the internal node(s) leading up to a leaf,
2377 * so hammer_btree_insert_cluster() will function properly.
2380 btree_node_is_almost_full(hammer_node_ondisk_t node)
2382 switch(node->type) {
2383 case HAMMER_BTREE_TYPE_INTERNAL:
2384 if (node->count > HAMMER_BTREE_INT_ELMS - 2)
2387 case HAMMER_BTREE_TYPE_LEAF:
2388 if (node->count > HAMMER_BTREE_LEAF_ELMS - 2)
2392 panic("illegal btree subtype");
2399 btree_max_elements(u_int8_t type)
2401 if (type == HAMMER_BTREE_TYPE_LEAF)
2402 return(HAMMER_BTREE_LEAF_ELMS);
2403 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2404 return(HAMMER_BTREE_INT_ELMS);
2405 panic("btree_max_elements: bad type %d\n", type);
2410 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2412 hammer_btree_elm_t elm;
2415 kprintf("node %p count=%d parent=%d type=%c\n",
2416 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2419 * Dump both boundary elements if an internal node
2421 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2422 for (i = 0; i <= ondisk->count; ++i) {
2423 elm = &ondisk->elms[i];
2424 hammer_print_btree_elm(elm, ondisk->type, i);
2427 for (i = 0; i < ondisk->count; ++i) {
2428 elm = &ondisk->elms[i];
2429 hammer_print_btree_elm(elm, ondisk->type, i);
2435 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2438 kprintf("\tobjid = %016llx\n", elm->base.obj_id);
2439 kprintf("\tkey = %016llx\n", elm->base.key);
2440 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2441 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2442 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2443 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2444 kprintf("\tbtype = %02x (%c)\n",
2446 (elm->base.btype ? elm->base.btype : '?'));
2449 case HAMMER_BTREE_TYPE_INTERNAL:
2450 kprintf("\tsubtree_off = %08x\n",
2451 elm->internal.subtree_offset);
2453 case HAMMER_BTREE_TYPE_SPIKE_BEG:
2454 case HAMMER_BTREE_TYPE_SPIKE_END:
2455 kprintf("\tspike_clu_no = %d\n", elm->leaf.spike_clu_no);
2456 kprintf("\tspike_vol_no = %d\n", elm->leaf.spike_vol_no);
2458 case HAMMER_BTREE_TYPE_RECORD:
2459 kprintf("\trec_offset = %08x\n", elm->leaf.rec_offset);
2460 kprintf("\tdata_offset = %08x\n", elm->leaf.data_offset);
2461 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2462 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);