2 * Copyright (c) 2007-2008 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 1.48 2008/05/13 20:46:54 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 * INSERTIONS: A search performed with the intention of doing
71 * an insert will guarantee that the terminal leaf node is not full by
72 * splitting full nodes. Splits occur top-down during the dive down the
75 * DELETIONS: A deletion makes no attempt to proactively balance the
76 * tree and will recursively remove nodes that become empty. If a
77 * deadlock occurs a deletion may not be able to remove an empty leaf.
78 * Deletions never allow internal nodes to become empty (that would blow
85 static int btree_search(hammer_cursor_t cursor, int flags);
86 static int btree_split_internal(hammer_cursor_t cursor);
87 static int btree_split_leaf(hammer_cursor_t cursor);
88 static int btree_remove(hammer_cursor_t cursor);
89 static int btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
90 hammer_btree_elm_t elm);
91 static int btree_node_is_full(hammer_node_ondisk_t node);
92 static void hammer_make_separator(hammer_base_elm_t key1,
93 hammer_base_elm_t key2, hammer_base_elm_t dest);
94 static void hammer_btree_unlock_children(
95 struct hammer_node_locklist **locklistp);
98 * Iterate records after a search. The cursor is iterated forwards past
99 * the current record until a record matching the key-range requirements
100 * is found. ENOENT is returned if the iteration goes past the ending
103 * The iteration is inclusive of key_beg and can be inclusive or exclusive
104 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
106 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
107 * may be modified by B-Tree functions.
109 * cursor->key_beg may or may not be modified by this function during
110 * the iteration. XXX future - in case of an inverted lock we may have
111 * to reinitiate the lookup and set key_beg to properly pick up where we
114 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
117 hammer_btree_iterate(hammer_cursor_t cursor)
119 hammer_node_ondisk_t node;
120 hammer_btree_elm_t elm;
126 * Skip past the current record
128 node = cursor->node->ondisk;
131 if (cursor->index < node->count &&
132 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
137 * Loop until an element is found or we are done.
141 * We iterate up the tree and then index over one element
142 * while we are at the last element in the current node.
144 * If we are at the root of the filesystem, cursor_up
147 * XXX this could be optimized by storing the information in
148 * the parent reference.
150 * XXX we can lose the node lock temporarily, this could mess
153 if (cursor->index == node->count) {
154 if (hammer_debug_btree) {
155 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
156 cursor->node->node_offset,
158 (cursor->parent ? cursor->parent->node_offset : -1),
159 cursor->parent_index,
162 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
163 error = hammer_cursor_up(cursor);
166 /* reload stale pointer */
167 node = cursor->node->ondisk;
168 KKASSERT(cursor->index != node->count);
174 * Check internal or leaf element. Determine if the record
175 * at the cursor has gone beyond the end of our range.
177 * We recurse down through internal nodes.
179 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
180 elm = &node->elms[cursor->index];
181 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
182 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
183 if (hammer_debug_btree) {
184 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx %d (td=%p)\n",
185 cursor->node->node_offset,
187 elm[0].internal.base.obj_id,
188 elm[0].internal.base.rec_type,
189 elm[0].internal.base.key,
193 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx %d\n",
194 cursor->node->node_offset,
196 elm[1].internal.base.obj_id,
197 elm[1].internal.base.rec_type,
198 elm[1].internal.base.key,
207 if (r == 0 && (cursor->flags &
208 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
217 KKASSERT(elm->internal.subtree_offset != 0);
219 error = hammer_cursor_down(cursor);
222 KKASSERT(cursor->index == 0);
223 /* reload stale pointer */
224 node = cursor->node->ondisk;
227 elm = &node->elms[cursor->index];
228 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
229 if (hammer_debug_btree) {
230 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx %d\n",
231 cursor->node->node_offset,
233 (elm[0].leaf.base.btype ?
234 elm[0].leaf.base.btype : '?'),
235 elm[0].leaf.base.obj_id,
236 elm[0].leaf.base.rec_type,
237 elm[0].leaf.base.key,
247 * We support both end-inclusive and
248 * end-exclusive searches.
251 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
256 switch(elm->leaf.base.btype) {
257 case HAMMER_BTREE_TYPE_RECORD:
258 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
259 hammer_btree_chkts(cursor->asof, &elm->base)) {
272 * node pointer invalid after loop
278 if (hammer_debug_btree) {
279 int i = cursor->index;
280 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
281 kprintf("ITERATE %p:%d %016llx %02x %016llx\n",
283 elm->internal.base.obj_id,
284 elm->internal.base.rec_type,
285 elm->internal.base.key
294 * Iterate in the reverse direction. This is used by the pruning code to
295 * avoid overlapping records.
298 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
300 hammer_node_ondisk_t node;
301 hammer_btree_elm_t elm;
307 * Skip past the current record. For various reasons the cursor
308 * may end up set to -1 or set to point at the end of the current
309 * node. These cases must be addressed.
311 node = cursor->node->ondisk;
314 if (cursor->index != -1 &&
315 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
318 if (cursor->index == cursor->node->ondisk->count)
322 * Loop until an element is found or we are done.
326 * We iterate up the tree and then index over one element
327 * while we are at the last element in the current node.
329 if (cursor->index == -1) {
330 error = hammer_cursor_up(cursor);
332 cursor->index = 0; /* sanity */
335 /* reload stale pointer */
336 node = cursor->node->ondisk;
337 KKASSERT(cursor->index != node->count);
343 * Check internal or leaf element. Determine if the record
344 * at the cursor has gone beyond the end of our range.
346 * We recurse down through internal nodes.
348 KKASSERT(cursor->index != node->count);
349 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
350 elm = &node->elms[cursor->index];
351 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
352 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
353 if (hammer_debug_btree) {
354 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx %d\n",
355 cursor->node->node_offset,
357 elm[0].internal.base.obj_id,
358 elm[0].internal.base.rec_type,
359 elm[0].internal.base.key,
362 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx %d\n",
363 cursor->node->node_offset,
365 elm[1].internal.base.obj_id,
366 elm[1].internal.base.rec_type,
367 elm[1].internal.base.key,
381 KKASSERT(elm->internal.subtree_offset != 0);
383 error = hammer_cursor_down(cursor);
386 KKASSERT(cursor->index == 0);
387 /* reload stale pointer */
388 node = cursor->node->ondisk;
390 /* this can assign -1 if the leaf was empty */
391 cursor->index = node->count - 1;
394 elm = &node->elms[cursor->index];
395 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
396 if (hammer_debug_btree) {
397 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx %d\n",
398 cursor->node->node_offset,
400 (elm[0].leaf.base.btype ?
401 elm[0].leaf.base.btype : '?'),
402 elm[0].leaf.base.obj_id,
403 elm[0].leaf.base.rec_type,
404 elm[0].leaf.base.key,
413 switch(elm->leaf.base.btype) {
414 case HAMMER_BTREE_TYPE_RECORD:
415 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
416 hammer_btree_chkts(cursor->asof, &elm->base)) {
429 * node pointer invalid after loop
435 if (hammer_debug_btree) {
436 int i = cursor->index;
437 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
438 kprintf("ITERATE %p:%d %016llx %02x %016llx\n",
440 elm->internal.base.obj_id,
441 elm->internal.base.rec_type,
442 elm->internal.base.key
451 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
452 * could not be found, EDEADLK if inserting and a retry is needed, and a
453 * fatal error otherwise. When retrying, the caller must terminate the
454 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
456 * The cursor is suitably positioned for a deletion on success, and suitably
457 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
460 * The cursor may begin anywhere, the search will traverse the tree in
461 * either direction to locate the requested element.
463 * Most of the logic implementing historical searches is handled here. We
464 * do an initial lookup with create_tid set to the asof TID. Due to the
465 * way records are laid out, a backwards iteration may be required if
466 * ENOENT is returned to locate the historical record. Here's the
469 * create_tid: 10 15 20
473 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
474 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
475 * not visible and thus causes ENOENT to be returned. We really need
476 * to check record 11 in LEAF1. If it also fails then the search fails
477 * (e.g. it might represent the range 11-16 and thus still not match our
478 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
479 * further iterations.
481 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
482 * and the cursor->create_check TID if an iteration might be needed.
483 * In the above example create_check would be set to 14.
486 hammer_btree_lookup(hammer_cursor_t cursor)
490 if (cursor->flags & HAMMER_CURSOR_ASOF) {
491 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
492 cursor->key_beg.create_tid = cursor->asof;
494 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
495 error = btree_search(cursor, 0);
496 if (error != ENOENT ||
497 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
500 * Stop if error other then ENOENT.
501 * Stop if ENOENT and not special case.
505 if (hammer_debug_btree) {
506 kprintf("CREATE_CHECK %016llx\n",
507 cursor->create_check);
509 cursor->key_beg.create_tid = cursor->create_check;
513 error = btree_search(cursor, 0);
515 if (error == 0 && cursor->flags)
516 error = hammer_btree_extract(cursor, cursor->flags);
521 * Execute the logic required to start an iteration. The first record
522 * located within the specified range is returned and iteration control
523 * flags are adjusted for successive hammer_btree_iterate() calls.
526 hammer_btree_first(hammer_cursor_t cursor)
530 error = hammer_btree_lookup(cursor);
531 if (error == ENOENT) {
532 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
533 error = hammer_btree_iterate(cursor);
535 cursor->flags |= HAMMER_CURSOR_ATEDISK;
540 * Similarly but for an iteration in the reverse direction.
542 * Set ATEDISK when iterating backwards to skip the current entry,
543 * which after an ENOENT lookup will be pointing beyond our end point.
546 hammer_btree_last(hammer_cursor_t cursor)
548 struct hammer_base_elm save;
551 save = cursor->key_beg;
552 cursor->key_beg = cursor->key_end;
553 error = hammer_btree_lookup(cursor);
554 cursor->key_beg = save;
555 if (error == ENOENT ||
556 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
557 cursor->flags |= HAMMER_CURSOR_ATEDISK;
558 error = hammer_btree_iterate_reverse(cursor);
560 cursor->flags |= HAMMER_CURSOR_ATEDISK;
565 * Extract the record and/or data associated with the cursor's current
566 * position. Any prior record or data stored in the cursor is replaced.
567 * The cursor must be positioned at a leaf node.
569 * NOTE: All extractions occur at the leaf of the B-Tree.
572 hammer_btree_extract(hammer_cursor_t cursor, int flags)
575 hammer_node_ondisk_t node;
576 hammer_btree_elm_t elm;
577 hammer_off_t data_off;
582 * The case where the data reference resolves to the same buffer
583 * as the record reference must be handled.
585 node = cursor->node->ondisk;
586 elm = &node->elms[cursor->index];
588 hmp = cursor->node->hmp;
589 flags |= cursor->flags & HAMMER_CURSOR_DATAEXTOK;
592 * There is nothing to extract for an internal element.
594 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
598 * Only record types have data.
600 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
601 cursor->leaf = &elm->leaf;
602 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
603 flags &= ~HAMMER_CURSOR_GET_DATA;
604 data_off = elm->leaf.data_offset;
605 data_len = elm->leaf.data_len;
607 flags &= ~HAMMER_CURSOR_GET_DATA;
610 if ((flags & HAMMER_CURSOR_GET_DATA)) {
612 * Data and record are in different buffers.
614 cursor->data = hammer_bread(hmp, data_off, &error,
615 &cursor->data_buffer);
616 KKASSERT(data_len >= 0 && data_len <= HAMMER_BUFSIZE);
618 crc32(cursor->data, data_len) != elm->leaf.data_crc) {
619 Debugger("CRC FAILED: DATA");
627 * Insert a leaf element into the B-Tree at the current cursor position.
628 * The cursor is positioned such that the element at and beyond the cursor
629 * are shifted to make room for the new record.
631 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
632 * flag set and that call must return ENOENT before this function can be
635 * The caller may depend on the cursor's exclusive lock after return to
636 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
638 * ENOSPC is returned if there is no room to insert a new record.
641 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm)
643 hammer_node_ondisk_t node;
647 if ((error = hammer_cursor_upgrade(cursor)) != 0)
651 * Insert the element at the leaf node and update the count in the
652 * parent. It is possible for parent to be NULL, indicating that
653 * the filesystem's ROOT B-Tree node is a leaf itself, which is
654 * possible. The root inode can never be deleted so the leaf should
657 * Remember that the right-hand boundary is not included in the
660 hammer_modify_node_all(cursor->trans, cursor->node);
661 node = cursor->node->ondisk;
663 KKASSERT(elm->base.btype != 0);
664 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
665 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
666 if (i != node->count) {
667 bcopy(&node->elms[i], &node->elms[i+1],
668 (node->count - i) * sizeof(*elm));
670 node->elms[i].leaf = *elm;
672 hammer_modify_node_done(cursor->node);
675 * Debugging sanity checks.
677 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
678 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
680 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
682 if (i != node->count - 1)
683 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
689 * Delete a record from the B-Tree at the current cursor position.
690 * The cursor is positioned such that the current element is the one
693 * On return the cursor will be positioned after the deleted element and
694 * MAY point to an internal node. It will be suitable for the continuation
695 * of an iteration but not for an insertion or deletion.
697 * Deletions will attempt to partially rebalance the B-Tree in an upward
698 * direction, but will terminate rather then deadlock. Empty internal nodes
699 * are never allowed by a deletion which deadlocks may end up giving us an
700 * empty leaf. The pruner will clean up and rebalance the tree.
702 * This function can return EDEADLK, requiring the caller to retry the
703 * operation after clearing the deadlock.
706 hammer_btree_delete(hammer_cursor_t cursor)
708 hammer_node_ondisk_t ondisk;
710 hammer_node_t parent;
714 if ((error = hammer_cursor_upgrade(cursor)) != 0)
718 * Delete the element from the leaf node.
720 * Remember that leaf nodes do not have boundaries.
723 ondisk = node->ondisk;
726 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
727 KKASSERT(i >= 0 && i < ondisk->count);
728 hammer_modify_node_all(cursor->trans, node);
729 if (i + 1 != ondisk->count) {
730 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
731 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
734 hammer_modify_node_done(node);
737 * Validate local parent
739 if (ondisk->parent) {
740 parent = cursor->parent;
742 KKASSERT(parent != NULL);
743 KKASSERT(parent->node_offset == ondisk->parent);
747 * If the leaf becomes empty it must be detached from the parent,
748 * potentially recursing through to the filesystem root.
750 * This may reposition the cursor at one of the parent's of the
753 * Ignore deadlock errors, that simply means that btree_remove
754 * was unable to recurse and had to leave us with an empty leaf.
756 KKASSERT(cursor->index <= ondisk->count);
757 if (ondisk->count == 0) {
758 error = btree_remove(cursor);
759 if (error == EDEADLK)
764 KKASSERT(cursor->parent == NULL ||
765 cursor->parent_index < cursor->parent->ondisk->count);
770 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
772 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
774 * The search can begin ANYWHERE in the B-Tree. As a first step the search
775 * iterates up the tree as necessary to properly position itself prior to
776 * actually doing the sarch.
778 * INSERTIONS: The search will split full nodes and leaves on its way down
779 * and guarentee that the leaf it ends up on is not full. If we run out
780 * of space the search continues to the leaf (to position the cursor for
781 * the spike), but ENOSPC is returned.
783 * The search is only guarenteed to end up on a leaf if an error code of 0
784 * is returned, or if inserting and an error code of ENOENT is returned.
785 * Otherwise it can stop at an internal node. On success a search returns
788 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
789 * filesystem, and it is not simple code. Please note the following facts:
791 * - Internal node recursions have a boundary on the left AND right. The
792 * right boundary is non-inclusive. The create_tid is a generic part
793 * of the key for internal nodes.
795 * - Leaf nodes contain terminal elements only now.
797 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
798 * historical search. ASOF and INSERT are mutually exclusive. When
799 * doing an as-of lookup btree_search() checks for a right-edge boundary
800 * case. If while recursing down the left-edge differs from the key
801 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
802 * with cursor->create_check. This is used by btree_lookup() to iterate.
803 * The iteration backwards because as-of searches can wind up going
804 * down the wrong branch of the B-Tree.
808 btree_search(hammer_cursor_t cursor, int flags)
810 hammer_node_ondisk_t node;
811 hammer_btree_elm_t elm;
818 flags |= cursor->flags;
820 if (hammer_debug_btree) {
821 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx (td = %p)\n",
822 cursor->node->node_offset,
824 cursor->key_beg.obj_id,
825 cursor->key_beg.rec_type,
827 cursor->key_beg.create_tid,
831 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
832 cursor->parent->node_offset, cursor->parent_index,
833 cursor->left_bound->obj_id,
834 cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
835 cursor->right_bound->obj_id,
836 cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
838 &cursor->parent->ondisk->elms[cursor->parent_index],
840 &cursor->parent->ondisk->elms[cursor->parent_index+1]
845 * Move our cursor up the tree until we find a node whos range covers
846 * the key we are trying to locate.
848 * The left bound is inclusive, the right bound is non-inclusive.
849 * It is ok to cursor up too far.
852 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
853 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
856 KKASSERT(cursor->parent);
857 error = hammer_cursor_up(cursor);
863 * The delete-checks below are based on node, not parent. Set the
864 * initial delete-check based on the parent.
867 KKASSERT(cursor->left_bound->create_tid != 1);
868 cursor->create_check = cursor->left_bound->create_tid - 1;
869 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
873 * We better have ended up with a node somewhere.
875 KKASSERT(cursor->node != NULL);
878 * If we are inserting we can't start at a full node if the parent
879 * is also full (because there is no way to split the node),
880 * continue running up the tree until the requirement is satisfied
881 * or we hit the root of the filesystem.
883 * (If inserting we aren't doing an as-of search so we don't have
884 * to worry about create_check).
886 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
887 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
888 if (btree_node_is_full(cursor->node->ondisk) == 0)
891 if (btree_node_is_full(cursor->node->ondisk) ==0)
894 if (cursor->node->ondisk->parent == 0 ||
895 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
898 error = hammer_cursor_up(cursor);
899 /* node may have become stale */
905 * Push down through internal nodes to locate the requested key.
907 node = cursor->node->ondisk;
908 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
910 * Scan the node to find the subtree index to push down into.
911 * We go one-past, then back-up.
913 * We must proactively remove deleted elements which may
914 * have been left over from a deadlocked btree_remove().
916 * The left and right boundaries are included in the loop
917 * in order to detect edge cases.
919 * If the separator only differs by create_tid (r == 1)
920 * and we are doing an as-of search, we may end up going
921 * down a branch to the left of the one containing the
922 * desired key. This requires numerous special cases.
924 if (hammer_debug_btree) {
925 kprintf("SEARCH-I %016llx count=%d\n",
926 cursor->node->node_offset,
929 for (i = 0; i <= node->count; ++i) {
930 elm = &node->elms[i];
931 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
932 if (hammer_debug_btree > 2) {
933 kprintf(" IELM %p %d r=%d\n",
934 &node->elms[i], i, r);
939 KKASSERT(elm->base.create_tid != 1);
940 cursor->create_check = elm->base.create_tid - 1;
941 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
944 if (hammer_debug_btree) {
945 kprintf("SEARCH-I preI=%d/%d r=%d\n",
950 * These cases occur when the parent's idea of the boundary
951 * is wider then the child's idea of the boundary, and
952 * require special handling. If not inserting we can
953 * terminate the search early for these cases but the
954 * child's boundaries cannot be unconditionally modified.
958 * If i == 0 the search terminated to the LEFT of the
959 * left_boundary but to the RIGHT of the parent's left
964 elm = &node->elms[0];
967 * If we aren't inserting we can stop here.
969 if ((flags & (HAMMER_CURSOR_INSERT |
970 HAMMER_CURSOR_PRUNING)) == 0) {
976 * Correct a left-hand boundary mismatch.
978 * We can only do this if we can upgrade the lock,
979 * and synchronized as a background cursor (i.e.
980 * inserting or pruning).
982 * WARNING: We can only do this if inserting, i.e.
983 * we are running on the backend.
985 if ((error = hammer_cursor_upgrade(cursor)) != 0)
987 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
988 hammer_modify_node_field(cursor->trans, cursor->node,
990 save = node->elms[0].base.btype;
991 node->elms[0].base = *cursor->left_bound;
992 node->elms[0].base.btype = save;
993 hammer_modify_node_done(cursor->node);
994 } else if (i == node->count + 1) {
996 * If i == node->count + 1 the search terminated to
997 * the RIGHT of the right boundary but to the LEFT
998 * of the parent's right boundary. If we aren't
999 * inserting we can stop here.
1001 * Note that the last element in this case is
1002 * elms[i-2] prior to adjustments to 'i'.
1005 if ((flags & (HAMMER_CURSOR_INSERT |
1006 HAMMER_CURSOR_PRUNING)) == 0) {
1012 * Correct a right-hand boundary mismatch.
1013 * (actual push-down record is i-2 prior to
1014 * adjustments to i).
1016 * We can only do this if we can upgrade the lock,
1017 * and synchronized as a background cursor (i.e.
1018 * inserting or pruning).
1020 * WARNING: We can only do this if inserting, i.e.
1021 * we are running on the backend.
1023 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1025 elm = &node->elms[i];
1026 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1027 hammer_modify_node(cursor->trans, cursor->node,
1028 &elm->base, sizeof(elm->base));
1029 elm->base = *cursor->right_bound;
1030 hammer_modify_node_done(cursor->node);
1034 * The push-down index is now i - 1. If we had
1035 * terminated on the right boundary this will point
1036 * us at the last element.
1041 elm = &node->elms[i];
1043 if (hammer_debug_btree) {
1044 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1045 "key=%016llx cre=%016llx\n",
1046 cursor->node->node_offset,
1048 elm->internal.base.obj_id,
1049 elm->internal.base.rec_type,
1050 elm->internal.base.key,
1051 elm->internal.base.create_tid
1056 * We better have a valid subtree offset.
1058 KKASSERT(elm->internal.subtree_offset != 0);
1061 * Handle insertion and deletion requirements.
1063 * If inserting split full nodes. The split code will
1064 * adjust cursor->node and cursor->index if the current
1065 * index winds up in the new node.
1067 * If inserting and a left or right edge case was detected,
1068 * we cannot correct the left or right boundary and must
1069 * prepend and append an empty leaf node in order to make
1070 * the boundary correction.
1072 * If we run out of space we set enospc and continue on
1073 * to a leaf to provide the spike code with a good point
1076 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1077 if (btree_node_is_full(node)) {
1078 error = btree_split_internal(cursor);
1080 if (error != ENOSPC)
1085 * reload stale pointers
1088 node = cursor->node->ondisk;
1093 * Push down (push into new node, existing node becomes
1094 * the parent) and continue the search.
1096 error = hammer_cursor_down(cursor);
1097 /* node may have become stale */
1100 node = cursor->node->ondisk;
1104 * We are at a leaf, do a linear search of the key array.
1106 * On success the index is set to the matching element and 0
1109 * On failure the index is set to the insertion point and ENOENT
1112 * Boundaries are not stored in leaf nodes, so the index can wind
1113 * up to the left of element 0 (index == 0) or past the end of
1114 * the array (index == node->count). It is also possible that the
1115 * leaf might be empty.
1117 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1118 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1119 if (hammer_debug_btree) {
1120 kprintf("SEARCH-L %016llx count=%d\n",
1121 cursor->node->node_offset,
1125 for (i = 0; i < node->count; ++i) {
1126 elm = &node->elms[i];
1128 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1130 if (hammer_debug_btree > 1)
1131 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1134 * We are at a record element. Stop if we've flipped past
1135 * key_beg, not counting the create_tid test. Allow the
1136 * r == 1 case (key_beg > element but differs only by its
1137 * create_tid) to fall through to the AS-OF check.
1139 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1147 * Check our as-of timestamp against the element.
1149 if (flags & HAMMER_CURSOR_ASOF) {
1150 if (hammer_btree_chkts(cursor->asof,
1151 &node->elms[i].base) != 0) {
1156 if (r > 0) /* can only be +1 */
1162 if (hammer_debug_btree) {
1163 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1164 cursor->node->node_offset, i);
1170 * The search of the leaf node failed. i is the insertion point.
1173 if (hammer_debug_btree) {
1174 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1175 cursor->node->node_offset, i);
1179 * No exact match was found, i is now at the insertion point.
1181 * If inserting split a full leaf before returning. This
1182 * may have the side effect of adjusting cursor->node and
1186 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1187 btree_node_is_full(node)) {
1188 error = btree_split_leaf(cursor);
1190 if (error != ENOSPC)
1195 * reload stale pointers
1199 node = &cursor->node->internal;
1204 * We reached a leaf but did not find the key we were looking for.
1205 * If this is an insert we will be properly positioned for an insert
1206 * (ENOENT) or spike (ENOSPC) operation.
1208 error = enospc ? ENOSPC : ENOENT;
1214 /************************************************************************
1215 * SPLITTING AND MERGING *
1216 ************************************************************************
1218 * These routines do all the dirty work required to split and merge nodes.
1222 * Split an internal node into two nodes and move the separator at the split
1223 * point to the parent.
1225 * (cursor->node, cursor->index) indicates the element the caller intends
1226 * to push into. We will adjust node and index if that element winds
1227 * up in the split node.
1229 * If we are at the root of the filesystem a new root must be created with
1230 * two elements, one pointing to the original root and one pointing to the
1231 * newly allocated split node.
1235 btree_split_internal(hammer_cursor_t cursor)
1237 hammer_node_ondisk_t ondisk;
1239 hammer_node_t parent;
1240 hammer_node_t new_node;
1241 hammer_btree_elm_t elm;
1242 hammer_btree_elm_t parent_elm;
1243 hammer_node_locklist_t locklist = NULL;
1244 hammer_mount_t hmp = cursor->trans->hmp;
1250 const int esize = sizeof(*elm);
1252 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1254 error = hammer_btree_lock_children(cursor, &locklist);
1259 * We are splitting but elms[split] will be promoted to the parent,
1260 * leaving the right hand node with one less element. If the
1261 * insertion point will be on the left-hand side adjust the split
1262 * point to give the right hand side one additional node.
1264 node = cursor->node;
1265 ondisk = node->ondisk;
1266 split = (ondisk->count + 1) / 2;
1267 if (cursor->index <= split)
1271 * If we are at the root of the filesystem, create a new root node
1272 * with 1 element and split normally. Avoid making major
1273 * modifications until we know the whole operation will work.
1275 if (ondisk->parent == 0) {
1276 parent = hammer_alloc_btree(cursor->trans, &error);
1279 hammer_lock_ex(&parent->lock);
1280 hammer_modify_node_noundo(cursor->trans, parent);
1281 ondisk = parent->ondisk;
1284 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1285 ondisk->elms[0].base = hmp->root_btree_beg;
1286 ondisk->elms[0].base.btype = node->ondisk->type;
1287 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1288 ondisk->elms[1].base = hmp->root_btree_end;
1289 hammer_modify_node_done(parent);
1290 /* ondisk->elms[1].base.btype - not used */
1292 parent_index = 0; /* index of current node in parent */
1295 parent = cursor->parent;
1296 parent_index = cursor->parent_index;
1300 * Split node into new_node at the split point.
1302 * B O O O P N N B <-- P = node->elms[split]
1303 * 0 1 2 3 4 5 6 <-- subtree indices
1308 * B O O O B B N N B <--- inner boundary points are 'P'
1312 new_node = hammer_alloc_btree(cursor->trans, &error);
1313 if (new_node == NULL) {
1315 hammer_unlock(&parent->lock);
1316 hammer_delete_node(cursor->trans, parent);
1317 hammer_rel_node(parent);
1321 hammer_lock_ex(&new_node->lock);
1324 * Create the new node. P becomes the left-hand boundary in the
1325 * new node. Copy the right-hand boundary as well.
1327 * elm is the new separator.
1329 hammer_modify_node_noundo(cursor->trans, new_node);
1330 hammer_modify_node_all(cursor->trans, node);
1331 ondisk = node->ondisk;
1332 elm = &ondisk->elms[split];
1333 bcopy(elm, &new_node->ondisk->elms[0],
1334 (ondisk->count - split + 1) * esize);
1335 new_node->ondisk->count = ondisk->count - split;
1336 new_node->ondisk->parent = parent->node_offset;
1337 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1338 KKASSERT(ondisk->type == new_node->ondisk->type);
1341 * Cleanup the original node. Elm (P) becomes the new boundary,
1342 * its subtree_offset was moved to the new node. If we had created
1343 * a new root its parent pointer may have changed.
1345 elm->internal.subtree_offset = 0;
1346 ondisk->count = split;
1349 * Insert the separator into the parent, fixup the parent's
1350 * reference to the original node, and reference the new node.
1351 * The separator is P.
1353 * Remember that base.count does not include the right-hand boundary.
1355 hammer_modify_node_all(cursor->trans, parent);
1356 ondisk = parent->ondisk;
1357 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1358 parent_elm = &ondisk->elms[parent_index+1];
1359 bcopy(parent_elm, parent_elm + 1,
1360 (ondisk->count - parent_index) * esize);
1361 parent_elm->internal.base = elm->base; /* separator P */
1362 parent_elm->internal.base.btype = new_node->ondisk->type;
1363 parent_elm->internal.subtree_offset = new_node->node_offset;
1365 hammer_modify_node_done(parent);
1368 * The children of new_node need their parent pointer set to new_node.
1369 * The children have already been locked by
1370 * hammer_btree_lock_children().
1372 for (i = 0; i < new_node->ondisk->count; ++i) {
1373 elm = &new_node->ondisk->elms[i];
1374 error = btree_set_parent(cursor->trans, new_node, elm);
1376 panic("btree_split_internal: btree-fixup problem");
1379 hammer_modify_node_done(new_node);
1382 * The filesystem's root B-Tree pointer may have to be updated.
1385 hammer_volume_t volume;
1387 volume = hammer_get_root_volume(hmp, &error);
1388 KKASSERT(error == 0);
1390 hammer_modify_volume_field(cursor->trans, volume,
1392 volume->ondisk->vol0_btree_root = parent->node_offset;
1393 hammer_modify_volume_done(volume);
1394 node->ondisk->parent = parent->node_offset;
1395 if (cursor->parent) {
1396 hammer_unlock(&cursor->parent->lock);
1397 hammer_rel_node(cursor->parent);
1399 cursor->parent = parent; /* lock'd and ref'd */
1400 hammer_rel_volume(volume, 0);
1402 hammer_modify_node_done(node);
1406 * Ok, now adjust the cursor depending on which element the original
1407 * index was pointing at. If we are >= the split point the push node
1408 * is now in the new node.
1410 * NOTE: If we are at the split point itself we cannot stay with the
1411 * original node because the push index will point at the right-hand
1412 * boundary, which is illegal.
1414 * NOTE: The cursor's parent or parent_index must be adjusted for
1415 * the case where a new parent (new root) was created, and the case
1416 * where the cursor is now pointing at the split node.
1418 if (cursor->index >= split) {
1419 cursor->parent_index = parent_index + 1;
1420 cursor->index -= split;
1421 hammer_unlock(&cursor->node->lock);
1422 hammer_rel_node(cursor->node);
1423 cursor->node = new_node; /* locked and ref'd */
1425 cursor->parent_index = parent_index;
1426 hammer_unlock(&new_node->lock);
1427 hammer_rel_node(new_node);
1431 * Fixup left and right bounds
1433 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1434 cursor->left_bound = &parent_elm[0].internal.base;
1435 cursor->right_bound = &parent_elm[1].internal.base;
1436 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1437 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1438 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1439 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1442 hammer_btree_unlock_children(&locklist);
1443 hammer_cursor_downgrade(cursor);
1448 * Same as the above, but splits a full leaf node.
1454 btree_split_leaf(hammer_cursor_t cursor)
1456 hammer_node_ondisk_t ondisk;
1457 hammer_node_t parent;
1460 hammer_node_t new_leaf;
1461 hammer_btree_elm_t elm;
1462 hammer_btree_elm_t parent_elm;
1463 hammer_base_elm_t mid_boundary;
1468 const size_t esize = sizeof(*elm);
1470 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1473 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1474 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1475 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1476 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1479 * Calculate the split point. If the insertion point will be on
1480 * the left-hand side adjust the split point to give the right
1481 * hand side one additional node.
1483 * Spikes are made up of two leaf elements which cannot be
1486 leaf = cursor->node;
1487 ondisk = leaf->ondisk;
1488 split = (ondisk->count + 1) / 2;
1489 if (cursor->index <= split)
1494 elm = &ondisk->elms[split];
1496 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1497 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1498 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1499 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1502 * If we are at the root of the tree, create a new root node with
1503 * 1 element and split normally. Avoid making major modifications
1504 * until we know the whole operation will work.
1506 if (ondisk->parent == 0) {
1507 parent = hammer_alloc_btree(cursor->trans, &error);
1510 hammer_lock_ex(&parent->lock);
1511 hammer_modify_node_noundo(cursor->trans, parent);
1512 ondisk = parent->ondisk;
1515 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1516 ondisk->elms[0].base = hmp->root_btree_beg;
1517 ondisk->elms[0].base.btype = leaf->ondisk->type;
1518 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1519 ondisk->elms[1].base = hmp->root_btree_end;
1520 /* ondisk->elms[1].base.btype = not used */
1521 hammer_modify_node_done(parent);
1523 parent_index = 0; /* insertion point in parent */
1526 parent = cursor->parent;
1527 parent_index = cursor->parent_index;
1531 * Split leaf into new_leaf at the split point. Select a separator
1532 * value in-between the two leafs but with a bent towards the right
1533 * leaf since comparisons use an 'elm >= separator' inequality.
1542 new_leaf = hammer_alloc_btree(cursor->trans, &error);
1543 if (new_leaf == NULL) {
1545 hammer_unlock(&parent->lock);
1546 hammer_delete_node(cursor->trans, parent);
1547 hammer_rel_node(parent);
1551 hammer_lock_ex(&new_leaf->lock);
1554 * Create the new node and copy the leaf elements from the split
1555 * point on to the new node.
1557 hammer_modify_node_all(cursor->trans, leaf);
1558 hammer_modify_node_noundo(cursor->trans, new_leaf);
1559 ondisk = leaf->ondisk;
1560 elm = &ondisk->elms[split];
1561 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1562 new_leaf->ondisk->count = ondisk->count - split;
1563 new_leaf->ondisk->parent = parent->node_offset;
1564 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1565 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1566 hammer_modify_node_done(new_leaf);
1569 * Cleanup the original node. Because this is a leaf node and
1570 * leaf nodes do not have a right-hand boundary, there
1571 * aren't any special edge cases to clean up. We just fixup the
1574 ondisk->count = split;
1577 * Insert the separator into the parent, fixup the parent's
1578 * reference to the original node, and reference the new node.
1579 * The separator is P.
1581 * Remember that base.count does not include the right-hand boundary.
1582 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1584 hammer_modify_node_all(cursor->trans, parent);
1585 ondisk = parent->ondisk;
1586 KKASSERT(split != 0);
1587 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1588 parent_elm = &ondisk->elms[parent_index+1];
1589 bcopy(parent_elm, parent_elm + 1,
1590 (ondisk->count - parent_index) * esize);
1592 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1593 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1594 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1595 mid_boundary = &parent_elm->base;
1597 hammer_modify_node_done(parent);
1600 * The filesystem's root B-Tree pointer may have to be updated.
1603 hammer_volume_t volume;
1605 volume = hammer_get_root_volume(hmp, &error);
1606 KKASSERT(error == 0);
1608 hammer_modify_volume_field(cursor->trans, volume,
1610 volume->ondisk->vol0_btree_root = parent->node_offset;
1611 hammer_modify_volume_done(volume);
1612 leaf->ondisk->parent = parent->node_offset;
1613 if (cursor->parent) {
1614 hammer_unlock(&cursor->parent->lock);
1615 hammer_rel_node(cursor->parent);
1617 cursor->parent = parent; /* lock'd and ref'd */
1618 hammer_rel_volume(volume, 0);
1620 hammer_modify_node_done(leaf);
1623 * Ok, now adjust the cursor depending on which element the original
1624 * index was pointing at. If we are >= the split point the push node
1625 * is now in the new node.
1627 * NOTE: If we are at the split point itself we need to select the
1628 * old or new node based on where key_beg's insertion point will be.
1629 * If we pick the wrong side the inserted element will wind up in
1630 * the wrong leaf node and outside that node's bounds.
1632 if (cursor->index > split ||
1633 (cursor->index == split &&
1634 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1635 cursor->parent_index = parent_index + 1;
1636 cursor->index -= split;
1637 hammer_unlock(&cursor->node->lock);
1638 hammer_rel_node(cursor->node);
1639 cursor->node = new_leaf;
1641 cursor->parent_index = parent_index;
1642 hammer_unlock(&new_leaf->lock);
1643 hammer_rel_node(new_leaf);
1647 * Fixup left and right bounds
1649 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1650 cursor->left_bound = &parent_elm[0].internal.base;
1651 cursor->right_bound = &parent_elm[1].internal.base;
1654 * Assert that the bounds are correct.
1656 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1657 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1658 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1659 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1660 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1661 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1664 hammer_cursor_downgrade(cursor);
1669 * Recursively correct the right-hand boundary's create_tid to (tid) as
1670 * long as the rest of the key matches. We have to recurse upward in
1671 * the tree as well as down the left side of each parent's right node.
1673 * Return EDEADLK if we were only partially successful, forcing the caller
1674 * to try again. The original cursor is not modified. This routine can
1675 * also fail with EDEADLK if it is forced to throw away a portion of its
1678 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1681 TAILQ_ENTRY(hammer_rhb) entry;
1686 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1689 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1691 struct hammer_rhb_list rhb_list;
1692 hammer_base_elm_t elm;
1693 hammer_node_t orig_node;
1694 struct hammer_rhb *rhb;
1698 TAILQ_INIT(&rhb_list);
1701 * Save our position so we can restore it on return. This also
1702 * gives us a stable 'elm'.
1704 orig_node = cursor->node;
1705 hammer_ref_node(orig_node);
1706 hammer_lock_sh(&orig_node->lock);
1707 orig_index = cursor->index;
1708 elm = &orig_node->ondisk->elms[orig_index].base;
1711 * Now build a list of parents going up, allocating a rhb
1712 * structure for each one.
1714 while (cursor->parent) {
1716 * Stop if we no longer have any right-bounds to fix up
1718 if (elm->obj_id != cursor->right_bound->obj_id ||
1719 elm->rec_type != cursor->right_bound->rec_type ||
1720 elm->key != cursor->right_bound->key) {
1725 * Stop if the right-hand bound's create_tid does not
1726 * need to be corrected.
1728 if (cursor->right_bound->create_tid >= tid)
1731 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1732 rhb->node = cursor->parent;
1733 rhb->index = cursor->parent_index;
1734 hammer_ref_node(rhb->node);
1735 hammer_lock_sh(&rhb->node->lock);
1736 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1738 hammer_cursor_up(cursor);
1742 * now safely adjust the right hand bound for each rhb. This may
1743 * also require taking the right side of the tree and iterating down
1747 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1748 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1749 hkprintf("CORRECT RHB %016llx index %d type=%c\n",
1750 rhb->node->node_offset,
1751 rhb->index, cursor->node->ondisk->type);
1754 TAILQ_REMOVE(&rhb_list, rhb, entry);
1755 hammer_unlock(&rhb->node->lock);
1756 hammer_rel_node(rhb->node);
1757 kfree(rhb, M_HAMMER);
1759 switch (cursor->node->ondisk->type) {
1760 case HAMMER_BTREE_TYPE_INTERNAL:
1762 * Right-boundary for parent at internal node
1763 * is one element to the right of the element whos
1764 * right boundary needs adjusting. We must then
1765 * traverse down the left side correcting any left
1766 * bounds (which may now be too far to the left).
1769 error = hammer_btree_correct_lhb(cursor, tid);
1772 panic("hammer_btree_correct_rhb(): Bad node type");
1781 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1782 TAILQ_REMOVE(&rhb_list, rhb, entry);
1783 hammer_unlock(&rhb->node->lock);
1784 hammer_rel_node(rhb->node);
1785 kfree(rhb, M_HAMMER);
1787 error = hammer_cursor_seek(cursor, orig_node, orig_index);
1788 hammer_unlock(&orig_node->lock);
1789 hammer_rel_node(orig_node);
1794 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1795 * bound going downward starting at the current cursor position.
1797 * This function does not restore the cursor after use.
1800 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1802 struct hammer_rhb_list rhb_list;
1803 hammer_base_elm_t elm;
1804 hammer_base_elm_t cmp;
1805 struct hammer_rhb *rhb;
1808 TAILQ_INIT(&rhb_list);
1810 cmp = &cursor->node->ondisk->elms[cursor->index].base;
1813 * Record the node and traverse down the left-hand side for all
1814 * matching records needing a boundary correction.
1818 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1819 rhb->node = cursor->node;
1820 rhb->index = cursor->index;
1821 hammer_ref_node(rhb->node);
1822 hammer_lock_sh(&rhb->node->lock);
1823 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1825 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1827 * Nothing to traverse down if we are at the right
1828 * boundary of an internal node.
1830 if (cursor->index == cursor->node->ondisk->count)
1833 elm = &cursor->node->ondisk->elms[cursor->index].base;
1834 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
1836 panic("Illegal leaf record type %02x", elm->btype);
1838 error = hammer_cursor_down(cursor);
1842 elm = &cursor->node->ondisk->elms[cursor->index].base;
1843 if (elm->obj_id != cmp->obj_id ||
1844 elm->rec_type != cmp->rec_type ||
1845 elm->key != cmp->key) {
1848 if (elm->create_tid >= tid)
1854 * Now we can safely adjust the left-hand boundary from the bottom-up.
1855 * The last element we remove from the list is the caller's right hand
1856 * boundary, which must also be adjusted.
1858 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1859 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1862 TAILQ_REMOVE(&rhb_list, rhb, entry);
1863 hammer_unlock(&rhb->node->lock);
1864 hammer_rel_node(rhb->node);
1865 kfree(rhb, M_HAMMER);
1867 elm = &cursor->node->ondisk->elms[cursor->index].base;
1868 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1869 hkprintf("hammer_btree_correct_lhb-I @%016llx[%d]\n",
1870 cursor->node->node_offset, cursor->index);
1871 hammer_modify_node(cursor->trans, cursor->node,
1873 sizeof(elm->create_tid));
1874 elm->create_tid = tid;
1875 hammer_modify_node_done(cursor->node);
1877 panic("hammer_btree_correct_lhb(): Bad element type");
1884 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1885 TAILQ_REMOVE(&rhb_list, rhb, entry);
1886 hammer_unlock(&rhb->node->lock);
1887 hammer_rel_node(rhb->node);
1888 kfree(rhb, M_HAMMER);
1894 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
1895 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
1896 * the operation due to a deadlock, or some other error.
1898 * This routine is always called with an empty, locked leaf but may recurse
1899 * into want-to-be-empty parents as part of its operation.
1901 * On return the cursor may end up pointing to an internal node, suitable
1902 * for further iteration but not for an immediate insertion or deletion.
1905 btree_remove(hammer_cursor_t cursor)
1907 hammer_node_ondisk_t ondisk;
1908 hammer_btree_elm_t elm;
1910 hammer_node_t parent;
1911 const int esize = sizeof(*elm);
1914 node = cursor->node;
1917 * When deleting the root of the filesystem convert it to
1918 * an empty leaf node. Internal nodes cannot be empty.
1920 if (node->ondisk->parent == 0) {
1921 KKASSERT(cursor->parent == NULL);
1922 hammer_modify_node_all(cursor->trans, node);
1923 ondisk = node->ondisk;
1924 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1926 hammer_modify_node_done(node);
1932 * Attempt to remove the parent's reference to the child. If the
1933 * parent would become empty we have to recurse. If we fail we
1934 * leave the parent pointing to an empty leaf node.
1936 parent = cursor->parent;
1938 if (parent->ondisk->count == 1) {
1940 * This special cursor_up_locked() call leaves the original
1941 * node exclusively locked and referenced, leaves the
1942 * original parent locked (as the new node), and locks the
1943 * new parent. It can return EDEADLK.
1945 error = hammer_cursor_up_locked(cursor);
1947 error = btree_remove(cursor);
1949 hammer_modify_node_all(cursor->trans, node);
1950 ondisk = node->ondisk;
1951 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
1953 hammer_modify_node_done(node);
1954 hammer_flush_node(node);
1955 hammer_delete_node(cursor->trans, node);
1957 kprintf("Warning: BTREE_REMOVE: Defering "
1958 "parent removal1 @ %016llx, skipping\n",
1961 hammer_unlock(&node->lock);
1962 hammer_rel_node(node);
1964 kprintf("Warning: BTREE_REMOVE: Defering parent "
1965 "removal2 @ %016llx, skipping\n",
1969 KKASSERT(parent->ondisk->count > 1);
1972 * Delete the subtree reference in the parent
1974 hammer_modify_node_all(cursor->trans, parent);
1975 ondisk = parent->ondisk;
1976 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
1977 elm = &ondisk->elms[cursor->parent_index];
1978 KKASSERT(elm->internal.subtree_offset == node->node_offset);
1979 KKASSERT(ondisk->count > 0);
1980 bcopy(&elm[1], &elm[0],
1981 (ondisk->count - cursor->parent_index) * esize);
1983 hammer_modify_node_done(parent);
1984 hammer_flush_node(node);
1985 hammer_delete_node(cursor->trans, node);
1988 * cursor->node is invalid, cursor up to make the cursor
1991 error = hammer_cursor_up(cursor);
1997 * The element (elm) has been moved to a new internal node (node).
1999 * If the element represents a pointer to an internal node that node's
2000 * parent must be adjusted to the element's new location.
2002 * XXX deadlock potential here with our exclusive locks
2006 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2007 hammer_btree_elm_t elm)
2009 hammer_node_t child;
2014 switch(elm->base.btype) {
2015 case HAMMER_BTREE_TYPE_INTERNAL:
2016 case HAMMER_BTREE_TYPE_LEAF:
2017 child = hammer_get_node(node->hmp, elm->internal.subtree_offset,
2020 hammer_modify_node_field(trans, child, parent);
2021 child->ondisk->parent = node->node_offset;
2022 hammer_modify_node_done(child);
2023 hammer_rel_node(child);
2033 * Exclusively lock all the children of node. This is used by the split
2034 * code to prevent anyone from accessing the children of a cursor node
2035 * while we fix-up its parent offset.
2037 * If we don't lock the children we can really mess up cursors which block
2038 * trying to cursor-up into our node.
2040 * On failure EDEADLK (or some other error) is returned. If a deadlock
2041 * error is returned the cursor is adjusted to block on termination.
2044 hammer_btree_lock_children(hammer_cursor_t cursor,
2045 struct hammer_node_locklist **locklistp)
2048 hammer_node_locklist_t item;
2049 hammer_node_ondisk_t ondisk;
2050 hammer_btree_elm_t elm;
2051 hammer_node_t child;
2055 node = cursor->node;
2056 ondisk = node->ondisk;
2058 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2059 elm = &ondisk->elms[i];
2061 switch(elm->base.btype) {
2062 case HAMMER_BTREE_TYPE_INTERNAL:
2063 case HAMMER_BTREE_TYPE_LEAF:
2064 KKASSERT(elm->internal.subtree_offset != 0);
2065 child = hammer_get_node(node->hmp,
2066 elm->internal.subtree_offset,
2074 if (hammer_lock_ex_try(&child->lock) != 0) {
2075 if (cursor->deadlk_node == NULL) {
2076 cursor->deadlk_node = child;
2077 hammer_ref_node(cursor->deadlk_node);
2080 hammer_rel_node(child);
2082 item = kmalloc(sizeof(*item),
2083 M_HAMMER, M_WAITOK);
2084 item->next = *locklistp;
2091 hammer_btree_unlock_children(locklistp);
2097 * Release previously obtained node locks.
2100 hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2102 hammer_node_locklist_t item;
2104 while ((item = *locklistp) != NULL) {
2105 *locklistp = item->next;
2106 hammer_unlock(&item->node->lock);
2107 hammer_rel_node(item->node);
2108 kfree(item, M_HAMMER);
2112 /************************************************************************
2113 * MISCELLANIOUS SUPPORT *
2114 ************************************************************************/
2117 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2119 * Note that for this particular function a return value of -1, 0, or +1
2120 * can denote a match if create_tid is otherwise discounted. A create_tid
2121 * of zero is considered to be 'infinity' in comparisons.
2123 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2126 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2128 if (key1->obj_id < key2->obj_id)
2130 if (key1->obj_id > key2->obj_id)
2133 if (key1->rec_type < key2->rec_type)
2135 if (key1->rec_type > key2->rec_type)
2138 if (key1->key < key2->key)
2140 if (key1->key > key2->key)
2144 * A create_tid of zero indicates a record which is undeletable
2145 * and must be considered to have a value of positive infinity.
2147 if (key1->create_tid == 0) {
2148 if (key2->create_tid == 0)
2152 if (key2->create_tid == 0)
2154 if (key1->create_tid < key2->create_tid)
2156 if (key1->create_tid > key2->create_tid)
2162 * Test a timestamp against an element to determine whether the
2163 * element is visible. A timestamp of 0 means 'infinity'.
2166 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2169 if (base->delete_tid)
2173 if (asof < base->create_tid)
2175 if (base->delete_tid && asof >= base->delete_tid)
2181 * Create a separator half way inbetween key1 and key2. For fields just
2182 * one unit apart, the separator will match key2. key1 is on the left-hand
2183 * side and key2 is on the right-hand side.
2185 * key2 must be >= the separator. It is ok for the separator to match key2.
2187 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2190 * NOTE: It might be beneficial to just scrap this whole mess and just
2191 * set the separator to key2.
2193 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2194 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2197 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2198 hammer_base_elm_t dest)
2200 bzero(dest, sizeof(*dest));
2202 dest->rec_type = key2->rec_type;
2203 dest->key = key2->key;
2204 dest->create_tid = key2->create_tid;
2206 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2207 if (key1->obj_id == key2->obj_id) {
2208 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2209 if (key1->rec_type == key2->rec_type) {
2210 MAKE_SEPARATOR(key1, key2, dest, key);
2212 * Don't bother creating a separator for create_tid,
2213 * which also conveniently avoids having to handle
2214 * the create_tid == 0 (infinity) case. Just leave
2215 * create_tid set to key2.
2217 * Worst case, dest matches key2 exactly, which is
2224 #undef MAKE_SEPARATOR
2227 * Return whether a generic internal or leaf node is full
2230 btree_node_is_full(hammer_node_ondisk_t node)
2232 switch(node->type) {
2233 case HAMMER_BTREE_TYPE_INTERNAL:
2234 if (node->count == HAMMER_BTREE_INT_ELMS)
2237 case HAMMER_BTREE_TYPE_LEAF:
2238 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2242 panic("illegal btree subtype");
2249 btree_max_elements(u_int8_t type)
2251 if (type == HAMMER_BTREE_TYPE_LEAF)
2252 return(HAMMER_BTREE_LEAF_ELMS);
2253 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2254 return(HAMMER_BTREE_INT_ELMS);
2255 panic("btree_max_elements: bad type %d\n", type);
2260 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2262 hammer_btree_elm_t elm;
2265 kprintf("node %p count=%d parent=%016llx type=%c\n",
2266 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2269 * Dump both boundary elements if an internal node
2271 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2272 for (i = 0; i <= ondisk->count; ++i) {
2273 elm = &ondisk->elms[i];
2274 hammer_print_btree_elm(elm, ondisk->type, i);
2277 for (i = 0; i < ondisk->count; ++i) {
2278 elm = &ondisk->elms[i];
2279 hammer_print_btree_elm(elm, ondisk->type, i);
2285 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2288 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
2289 kprintf("\tkey = %016llx\n", elm->base.key);
2290 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2291 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2292 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2293 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2294 kprintf("\tbtype = %02x (%c)\n",
2296 (elm->base.btype ? elm->base.btype : '?'));
2299 case HAMMER_BTREE_TYPE_INTERNAL:
2300 kprintf("\tsubtree_off = %016llx\n",
2301 elm->internal.subtree_offset);
2303 case HAMMER_BTREE_TYPE_RECORD:
2304 kprintf("\tatime = %016llx\n", elm->leaf.atime);
2305 kprintf("\tdata_offset = %016llx\n", elm->leaf.data_offset);
2306 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2307 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);