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 1.53 2008/06/14 01:42:13 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 hammer_btree_search_node(hammer_base_elm_t elm,
87 hammer_node_ondisk_t node);
88 static int btree_split_internal(hammer_cursor_t cursor);
89 static int btree_split_leaf(hammer_cursor_t cursor);
90 static int btree_remove(hammer_cursor_t cursor);
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);
96 * Iterate records after a search. The cursor is iterated forwards past
97 * the current record until a record matching the key-range requirements
98 * is found. ENOENT is returned if the iteration goes past the ending
101 * The iteration is inclusive of key_beg and can be inclusive or exclusive
102 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
104 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
105 * may be modified by B-Tree functions.
107 * cursor->key_beg may or may not be modified by this function during
108 * the iteration. XXX future - in case of an inverted lock we may have
109 * to reinitiate the lookup and set key_beg to properly pick up where we
112 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
115 hammer_btree_iterate(hammer_cursor_t cursor)
117 hammer_node_ondisk_t node;
118 hammer_btree_elm_t elm;
124 * Skip past the current record
126 node = cursor->node->ondisk;
129 if (cursor->index < node->count &&
130 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
135 * Loop until an element is found or we are done.
139 * We iterate up the tree and then index over one element
140 * while we are at the last element in the current node.
142 * If we are at the root of the filesystem, cursor_up
145 * XXX this could be optimized by storing the information in
146 * the parent reference.
148 * XXX we can lose the node lock temporarily, this could mess
151 ++hammer_stats_btree_iterations;
152 if (cursor->index == node->count) {
153 if (hammer_debug_btree) {
154 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
155 cursor->node->node_offset,
157 (cursor->parent ? cursor->parent->node_offset : -1),
158 cursor->parent_index,
161 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
162 error = hammer_cursor_up(cursor);
165 /* reload stale pointer */
166 node = cursor->node->ondisk;
167 KKASSERT(cursor->index != node->count);
170 * If we are reblocking we want to return internal
173 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
174 cursor->flags |= HAMMER_CURSOR_ATEDISK;
182 * Check internal or leaf element. Determine if the record
183 * at the cursor has gone beyond the end of our range.
185 * We recurse down through internal nodes.
187 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
188 elm = &node->elms[cursor->index];
189 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
190 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
191 if (hammer_debug_btree) {
192 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
193 cursor->node->node_offset,
195 elm[0].internal.base.obj_id,
196 elm[0].internal.base.rec_type,
197 elm[0].internal.base.key,
198 elm[0].internal.base.localization,
202 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
203 cursor->node->node_offset,
205 elm[1].internal.base.obj_id,
206 elm[1].internal.base.rec_type,
207 elm[1].internal.base.key,
208 elm[1].internal.base.localization,
217 if (r == 0 && (cursor->flags &
218 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
227 KKASSERT(elm->internal.subtree_offset != 0);
229 error = hammer_cursor_down(cursor);
232 KKASSERT(cursor->index == 0);
233 /* reload stale pointer */
234 node = cursor->node->ondisk;
237 elm = &node->elms[cursor->index];
238 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
239 if (hammer_debug_btree) {
240 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
241 cursor->node->node_offset,
243 (elm[0].leaf.base.btype ?
244 elm[0].leaf.base.btype : '?'),
245 elm[0].leaf.base.obj_id,
246 elm[0].leaf.base.rec_type,
247 elm[0].leaf.base.key,
248 elm[0].leaf.base.localization,
258 * We support both end-inclusive and
259 * end-exclusive searches.
262 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
267 switch(elm->leaf.base.btype) {
268 case HAMMER_BTREE_TYPE_RECORD:
269 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
270 hammer_btree_chkts(cursor->asof, &elm->base)) {
283 * node pointer invalid after loop
289 if (hammer_debug_btree) {
290 int i = cursor->index;
291 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
292 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
294 elm->internal.base.obj_id,
295 elm->internal.base.rec_type,
296 elm->internal.base.key,
297 elm->internal.base.localization
306 * Iterate in the reverse direction. This is used by the pruning code to
307 * avoid overlapping records.
310 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
312 hammer_node_ondisk_t node;
313 hammer_btree_elm_t elm;
319 * Skip past the current record. For various reasons the cursor
320 * may end up set to -1 or set to point at the end of the current
321 * node. These cases must be addressed.
323 node = cursor->node->ondisk;
326 if (cursor->index != -1 &&
327 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
330 if (cursor->index == cursor->node->ondisk->count)
334 * Loop until an element is found or we are done.
338 * We iterate up the tree and then index over one element
339 * while we are at the last element in the current node.
341 if (cursor->index == -1) {
342 error = hammer_cursor_up(cursor);
344 cursor->index = 0; /* sanity */
347 /* reload stale pointer */
348 node = cursor->node->ondisk;
349 KKASSERT(cursor->index != node->count);
355 * Check internal or leaf element. Determine if the record
356 * at the cursor has gone beyond the end of our range.
358 * We recurse down through internal nodes.
360 KKASSERT(cursor->index != node->count);
361 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
362 elm = &node->elms[cursor->index];
363 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
364 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
365 if (hammer_debug_btree) {
366 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
367 cursor->node->node_offset,
369 elm[0].internal.base.obj_id,
370 elm[0].internal.base.rec_type,
371 elm[0].internal.base.key,
372 elm[0].internal.base.localization,
375 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
376 cursor->node->node_offset,
378 elm[1].internal.base.obj_id,
379 elm[1].internal.base.rec_type,
380 elm[1].internal.base.key,
381 elm[1].internal.base.localization,
395 KKASSERT(elm->internal.subtree_offset != 0);
397 error = hammer_cursor_down(cursor);
400 KKASSERT(cursor->index == 0);
401 /* reload stale pointer */
402 node = cursor->node->ondisk;
404 /* this can assign -1 if the leaf was empty */
405 cursor->index = node->count - 1;
408 elm = &node->elms[cursor->index];
409 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
410 if (hammer_debug_btree) {
411 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
412 cursor->node->node_offset,
414 (elm[0].leaf.base.btype ?
415 elm[0].leaf.base.btype : '?'),
416 elm[0].leaf.base.obj_id,
417 elm[0].leaf.base.rec_type,
418 elm[0].leaf.base.key,
419 elm[0].leaf.base.localization,
428 switch(elm->leaf.base.btype) {
429 case HAMMER_BTREE_TYPE_RECORD:
430 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
431 hammer_btree_chkts(cursor->asof, &elm->base)) {
444 * node pointer invalid after loop
450 if (hammer_debug_btree) {
451 int i = cursor->index;
452 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
453 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
455 elm->internal.base.obj_id,
456 elm->internal.base.rec_type,
457 elm->internal.base.key,
458 elm->internal.base.localization
467 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
468 * could not be found, EDEADLK if inserting and a retry is needed, and a
469 * fatal error otherwise. When retrying, the caller must terminate the
470 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
472 * The cursor is suitably positioned for a deletion on success, and suitably
473 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
476 * The cursor may begin anywhere, the search will traverse the tree in
477 * either direction to locate the requested element.
479 * Most of the logic implementing historical searches is handled here. We
480 * do an initial lookup with create_tid set to the asof TID. Due to the
481 * way records are laid out, a backwards iteration may be required if
482 * ENOENT is returned to locate the historical record. Here's the
485 * create_tid: 10 15 20
489 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
490 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
491 * not visible and thus causes ENOENT to be returned. We really need
492 * to check record 11 in LEAF1. If it also fails then the search fails
493 * (e.g. it might represent the range 11-16 and thus still not match our
494 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
495 * further iterations.
497 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
498 * and the cursor->create_check TID if an iteration might be needed.
499 * In the above example create_check would be set to 14.
502 hammer_btree_lookup(hammer_cursor_t cursor)
506 if (cursor->flags & HAMMER_CURSOR_ASOF) {
507 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
508 cursor->key_beg.create_tid = cursor->asof;
510 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
511 error = btree_search(cursor, 0);
512 if (error != ENOENT ||
513 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
516 * Stop if error other then ENOENT.
517 * Stop if ENOENT and not special case.
521 if (hammer_debug_btree) {
522 kprintf("CREATE_CHECK %016llx\n",
523 cursor->create_check);
525 cursor->key_beg.create_tid = cursor->create_check;
529 error = btree_search(cursor, 0);
532 error = hammer_btree_extract(cursor, cursor->flags);
537 * Execute the logic required to start an iteration. The first record
538 * located within the specified range is returned and iteration control
539 * flags are adjusted for successive hammer_btree_iterate() calls.
542 hammer_btree_first(hammer_cursor_t cursor)
546 error = hammer_btree_lookup(cursor);
547 if (error == ENOENT) {
548 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
549 error = hammer_btree_iterate(cursor);
551 cursor->flags |= HAMMER_CURSOR_ATEDISK;
556 * Similarly but for an iteration in the reverse direction.
558 * Set ATEDISK when iterating backwards to skip the current entry,
559 * which after an ENOENT lookup will be pointing beyond our end point.
562 hammer_btree_last(hammer_cursor_t cursor)
564 struct hammer_base_elm save;
567 save = cursor->key_beg;
568 cursor->key_beg = cursor->key_end;
569 error = hammer_btree_lookup(cursor);
570 cursor->key_beg = save;
571 if (error == ENOENT ||
572 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
573 cursor->flags |= HAMMER_CURSOR_ATEDISK;
574 error = hammer_btree_iterate_reverse(cursor);
576 cursor->flags |= HAMMER_CURSOR_ATEDISK;
581 * Extract the record and/or data associated with the cursor's current
582 * position. Any prior record or data stored in the cursor is replaced.
583 * The cursor must be positioned at a leaf node.
585 * NOTE: All extractions occur at the leaf of the B-Tree.
588 hammer_btree_extract(hammer_cursor_t cursor, int flags)
591 hammer_node_ondisk_t node;
592 hammer_btree_elm_t elm;
593 hammer_off_t data_off;
598 * The case where the data reference resolves to the same buffer
599 * as the record reference must be handled.
601 node = cursor->node->ondisk;
602 elm = &node->elms[cursor->index];
604 hmp = cursor->node->hmp;
607 * There is nothing to extract for an internal element.
609 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
613 * Only record types have data.
615 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
616 cursor->leaf = &elm->leaf;
617 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
618 flags &= ~HAMMER_CURSOR_GET_DATA;
619 data_off = elm->leaf.data_offset;
620 data_len = elm->leaf.data_len;
622 flags &= ~HAMMER_CURSOR_GET_DATA;
625 if ((flags & HAMMER_CURSOR_GET_DATA)) {
627 * Data and record are in different buffers.
629 cursor->data = hammer_bread(hmp, data_off, &error,
630 &cursor->data_buffer);
631 KKASSERT(data_len >= 0 && data_len <= HAMMER_BUFSIZE);
633 crc32(cursor->data, data_len) != elm->leaf.data_crc) {
634 Debugger("CRC FAILED: DATA");
642 * Insert a leaf element into the B-Tree at the current cursor position.
643 * The cursor is positioned such that the element at and beyond the cursor
644 * are shifted to make room for the new record.
646 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
647 * flag set and that call must return ENOENT before this function can be
650 * The caller may depend on the cursor's exclusive lock after return to
651 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
653 * ENOSPC is returned if there is no room to insert a new record.
656 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm)
658 hammer_node_ondisk_t node;
662 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
666 * Insert the element at the leaf node and update the count in the
667 * parent. It is possible for parent to be NULL, indicating that
668 * the filesystem's ROOT B-Tree node is a leaf itself, which is
669 * possible. The root inode can never be deleted so the leaf should
672 * Remember that the right-hand boundary is not included in the
675 hammer_modify_node_all(cursor->trans, cursor->node);
676 node = cursor->node->ondisk;
678 KKASSERT(elm->base.btype != 0);
679 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
680 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
681 if (i != node->count) {
682 bcopy(&node->elms[i], &node->elms[i+1],
683 (node->count - i) * sizeof(*elm));
685 node->elms[i].leaf = *elm;
687 hammer_modify_node_done(cursor->node);
690 * Debugging sanity checks.
692 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
693 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
695 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
697 if (i != node->count - 1)
698 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
704 * Delete a record from the B-Tree at the current cursor position.
705 * The cursor is positioned such that the current element is the one
708 * On return the cursor will be positioned after the deleted element and
709 * MAY point to an internal node. It will be suitable for the continuation
710 * of an iteration but not for an insertion or deletion.
712 * Deletions will attempt to partially rebalance the B-Tree in an upward
713 * direction, but will terminate rather then deadlock. Empty internal nodes
714 * are never allowed by a deletion which deadlocks may end up giving us an
715 * empty leaf. The pruner will clean up and rebalance the tree.
717 * This function can return EDEADLK, requiring the caller to retry the
718 * operation after clearing the deadlock.
721 hammer_btree_delete(hammer_cursor_t cursor)
723 hammer_node_ondisk_t ondisk;
725 hammer_node_t parent;
729 if ((error = hammer_cursor_upgrade(cursor)) != 0)
733 * Delete the element from the leaf node.
735 * Remember that leaf nodes do not have boundaries.
738 ondisk = node->ondisk;
741 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
742 KKASSERT(i >= 0 && i < ondisk->count);
743 hammer_modify_node_all(cursor->trans, node);
744 if (i + 1 != ondisk->count) {
745 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
746 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
749 hammer_modify_node_done(node);
752 * Validate local parent
754 if (ondisk->parent) {
755 parent = cursor->parent;
757 KKASSERT(parent != NULL);
758 KKASSERT(parent->node_offset == ondisk->parent);
762 * If the leaf becomes empty it must be detached from the parent,
763 * potentially recursing through to the filesystem root.
765 * This may reposition the cursor at one of the parent's of the
768 * Ignore deadlock errors, that simply means that btree_remove
769 * was unable to recurse and had to leave us with an empty leaf.
771 KKASSERT(cursor->index <= ondisk->count);
772 if (ondisk->count == 0) {
773 error = btree_remove(cursor);
774 if (error == EDEADLK)
779 KKASSERT(cursor->parent == NULL ||
780 cursor->parent_index < cursor->parent->ondisk->count);
785 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
787 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
789 * The search can begin ANYWHERE in the B-Tree. As a first step the search
790 * iterates up the tree as necessary to properly position itself prior to
791 * actually doing the sarch.
793 * INSERTIONS: The search will split full nodes and leaves on its way down
794 * and guarentee that the leaf it ends up on is not full. If we run out
795 * of space the search continues to the leaf (to position the cursor for
796 * the spike), but ENOSPC is returned.
798 * The search is only guarenteed to end up on a leaf if an error code of 0
799 * is returned, or if inserting and an error code of ENOENT is returned.
800 * Otherwise it can stop at an internal node. On success a search returns
803 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
804 * filesystem, and it is not simple code. Please note the following facts:
806 * - Internal node recursions have a boundary on the left AND right. The
807 * right boundary is non-inclusive. The create_tid is a generic part
808 * of the key for internal nodes.
810 * - Leaf nodes contain terminal elements only now.
812 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
813 * historical search. ASOF and INSERT are mutually exclusive. When
814 * doing an as-of lookup btree_search() checks for a right-edge boundary
815 * case. If while recursing down the left-edge differs from the key
816 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
817 * with cursor->create_check. This is used by btree_lookup() to iterate.
818 * The iteration backwards because as-of searches can wind up going
819 * down the wrong branch of the B-Tree.
823 btree_search(hammer_cursor_t cursor, int flags)
825 hammer_node_ondisk_t node;
826 hammer_btree_elm_t elm;
833 flags |= cursor->flags;
835 if (hammer_debug_btree) {
836 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
837 cursor->node->node_offset,
839 cursor->key_beg.obj_id,
840 cursor->key_beg.rec_type,
842 cursor->key_beg.create_tid,
843 cursor->key_beg.localization,
847 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
848 cursor->parent->node_offset, cursor->parent_index,
849 cursor->left_bound->obj_id,
850 cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
851 cursor->right_bound->obj_id,
852 cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
854 &cursor->parent->ondisk->elms[cursor->parent_index],
856 &cursor->parent->ondisk->elms[cursor->parent_index+1]
861 * Move our cursor up the tree until we find a node whos range covers
862 * the key we are trying to locate.
864 * The left bound is inclusive, the right bound is non-inclusive.
865 * It is ok to cursor up too far.
868 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
869 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
872 KKASSERT(cursor->parent);
873 error = hammer_cursor_up(cursor);
879 * The delete-checks below are based on node, not parent. Set the
880 * initial delete-check based on the parent.
883 KKASSERT(cursor->left_bound->create_tid != 1);
884 cursor->create_check = cursor->left_bound->create_tid - 1;
885 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
889 * We better have ended up with a node somewhere.
891 KKASSERT(cursor->node != NULL);
894 * If we are inserting we can't start at a full node if the parent
895 * is also full (because there is no way to split the node),
896 * continue running up the tree until the requirement is satisfied
897 * or we hit the root of the filesystem.
899 * (If inserting we aren't doing an as-of search so we don't have
900 * to worry about create_check).
902 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
903 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
904 if (btree_node_is_full(cursor->node->ondisk) == 0)
907 if (btree_node_is_full(cursor->node->ondisk) ==0)
910 if (cursor->node->ondisk->parent == 0 ||
911 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
914 error = hammer_cursor_up(cursor);
915 /* node may have become stale */
921 * Push down through internal nodes to locate the requested key.
923 node = cursor->node->ondisk;
924 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
926 * Scan the node to find the subtree index to push down into.
927 * We go one-past, then back-up.
929 * We must proactively remove deleted elements which may
930 * have been left over from a deadlocked btree_remove().
932 * The left and right boundaries are included in the loop
933 * in order to detect edge cases.
935 * If the separator only differs by create_tid (r == 1)
936 * and we are doing an as-of search, we may end up going
937 * down a branch to the left of the one containing the
938 * desired key. This requires numerous special cases.
940 ++hammer_stats_btree_iterations;
941 if (hammer_debug_btree) {
942 kprintf("SEARCH-I %016llx count=%d\n",
943 cursor->node->node_offset,
948 * Try to shortcut the search before dropping into the
949 * linear loop. Locate the first node where r <= 1.
951 i = hammer_btree_search_node(&cursor->key_beg, node);
952 while (i <= node->count) {
953 elm = &node->elms[i];
954 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
955 if (hammer_debug_btree > 2) {
956 kprintf(" IELM %p %d r=%d\n",
957 &node->elms[i], i, r);
962 KKASSERT(elm->base.create_tid != 1);
963 cursor->create_check = elm->base.create_tid - 1;
964 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
968 if (hammer_debug_btree) {
969 kprintf("SEARCH-I preI=%d/%d r=%d\n",
974 * These cases occur when the parent's idea of the boundary
975 * is wider then the child's idea of the boundary, and
976 * require special handling. If not inserting we can
977 * terminate the search early for these cases but the
978 * child's boundaries cannot be unconditionally modified.
982 * If i == 0 the search terminated to the LEFT of the
983 * left_boundary but to the RIGHT of the parent's left
988 elm = &node->elms[0];
991 * If we aren't inserting we can stop here.
993 if ((flags & (HAMMER_CURSOR_INSERT |
994 HAMMER_CURSOR_PRUNING)) == 0) {
1000 * Correct a left-hand boundary mismatch.
1002 * We can only do this if we can upgrade the lock,
1003 * and synchronized as a background cursor (i.e.
1004 * inserting or pruning).
1006 * WARNING: We can only do this if inserting, i.e.
1007 * we are running on the backend.
1009 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1011 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1012 hammer_modify_node_field(cursor->trans, cursor->node,
1014 save = node->elms[0].base.btype;
1015 node->elms[0].base = *cursor->left_bound;
1016 node->elms[0].base.btype = save;
1017 hammer_modify_node_done(cursor->node);
1018 } else if (i == node->count + 1) {
1020 * If i == node->count + 1 the search terminated to
1021 * the RIGHT of the right boundary but to the LEFT
1022 * of the parent's right boundary. If we aren't
1023 * inserting we can stop here.
1025 * Note that the last element in this case is
1026 * elms[i-2] prior to adjustments to 'i'.
1029 if ((flags & (HAMMER_CURSOR_INSERT |
1030 HAMMER_CURSOR_PRUNING)) == 0) {
1036 * Correct a right-hand boundary mismatch.
1037 * (actual push-down record is i-2 prior to
1038 * adjustments to i).
1040 * We can only do this if we can upgrade the lock,
1041 * and synchronized as a background cursor (i.e.
1042 * inserting or pruning).
1044 * WARNING: We can only do this if inserting, i.e.
1045 * we are running on the backend.
1047 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1049 elm = &node->elms[i];
1050 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1051 hammer_modify_node(cursor->trans, cursor->node,
1052 &elm->base, sizeof(elm->base));
1053 elm->base = *cursor->right_bound;
1054 hammer_modify_node_done(cursor->node);
1058 * The push-down index is now i - 1. If we had
1059 * terminated on the right boundary this will point
1060 * us at the last element.
1065 elm = &node->elms[i];
1067 if (hammer_debug_btree) {
1068 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1069 "key=%016llx cre=%016llx lo=%02x\n",
1070 cursor->node->node_offset,
1072 elm->internal.base.obj_id,
1073 elm->internal.base.rec_type,
1074 elm->internal.base.key,
1075 elm->internal.base.create_tid,
1076 elm->internal.base.localization
1081 * We better have a valid subtree offset.
1083 KKASSERT(elm->internal.subtree_offset != 0);
1086 * Handle insertion and deletion requirements.
1088 * If inserting split full nodes. The split code will
1089 * adjust cursor->node and cursor->index if the current
1090 * index winds up in the new node.
1092 * If inserting and a left or right edge case was detected,
1093 * we cannot correct the left or right boundary and must
1094 * prepend and append an empty leaf node in order to make
1095 * the boundary correction.
1097 * If we run out of space we set enospc and continue on
1098 * to a leaf to provide the spike code with a good point
1101 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1102 if (btree_node_is_full(node)) {
1103 error = btree_split_internal(cursor);
1105 if (error != ENOSPC)
1110 * reload stale pointers
1113 node = cursor->node->ondisk;
1118 * Push down (push into new node, existing node becomes
1119 * the parent) and continue the search.
1121 error = hammer_cursor_down(cursor);
1122 /* node may have become stale */
1125 node = cursor->node->ondisk;
1129 * We are at a leaf, do a linear search of the key array.
1131 * On success the index is set to the matching element and 0
1134 * On failure the index is set to the insertion point and ENOENT
1137 * Boundaries are not stored in leaf nodes, so the index can wind
1138 * up to the left of element 0 (index == 0) or past the end of
1139 * the array (index == node->count). It is also possible that the
1140 * leaf might be empty.
1142 ++hammer_stats_btree_iterations;
1143 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1144 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1145 if (hammer_debug_btree) {
1146 kprintf("SEARCH-L %016llx count=%d\n",
1147 cursor->node->node_offset,
1152 * Try to shortcut the search before dropping into the
1153 * linear loop. Locate the first node where r <= 1.
1155 i = hammer_btree_search_node(&cursor->key_beg, node);
1156 while (i < node->count) {
1157 elm = &node->elms[i];
1159 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1161 if (hammer_debug_btree > 1)
1162 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1165 * We are at a record element. Stop if we've flipped past
1166 * key_beg, not counting the create_tid test. Allow the
1167 * r == 1 case (key_beg > element but differs only by its
1168 * create_tid) to fall through to the AS-OF check.
1170 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1180 * Check our as-of timestamp against the element.
1182 if (flags & HAMMER_CURSOR_ASOF) {
1183 if (hammer_btree_chkts(cursor->asof,
1184 &node->elms[i].base) != 0) {
1190 if (r > 0) { /* can only be +1 */
1198 if (hammer_debug_btree) {
1199 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1200 cursor->node->node_offset, i);
1206 * The search of the leaf node failed. i is the insertion point.
1209 if (hammer_debug_btree) {
1210 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1211 cursor->node->node_offset, i);
1215 * No exact match was found, i is now at the insertion point.
1217 * If inserting split a full leaf before returning. This
1218 * may have the side effect of adjusting cursor->node and
1222 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1223 btree_node_is_full(node)) {
1224 error = btree_split_leaf(cursor);
1226 if (error != ENOSPC)
1231 * reload stale pointers
1235 node = &cursor->node->internal;
1240 * We reached a leaf but did not find the key we were looking for.
1241 * If this is an insert we will be properly positioned for an insert
1242 * (ENOENT) or spike (ENOSPC) operation.
1244 error = enospc ? ENOSPC : ENOENT;
1250 * Heuristical search for the first element whos comparison is <= 1. May
1251 * return an index whos compare result is > 1 but may only return an index
1252 * whos compare result is <= 1 if it is the first element with that result.
1255 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1263 * Don't bother if the node does not have very many elements
1268 i = b + (s - b) / 2;
1269 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1280 /************************************************************************
1281 * SPLITTING AND MERGING *
1282 ************************************************************************
1284 * These routines do all the dirty work required to split and merge nodes.
1288 * Split an internal node into two nodes and move the separator at the split
1289 * point to the parent.
1291 * (cursor->node, cursor->index) indicates the element the caller intends
1292 * to push into. We will adjust node and index if that element winds
1293 * up in the split node.
1295 * If we are at the root of the filesystem a new root must be created with
1296 * two elements, one pointing to the original root and one pointing to the
1297 * newly allocated split node.
1301 btree_split_internal(hammer_cursor_t cursor)
1303 hammer_node_ondisk_t ondisk;
1305 hammer_node_t parent;
1306 hammer_node_t new_node;
1307 hammer_btree_elm_t elm;
1308 hammer_btree_elm_t parent_elm;
1309 hammer_node_locklist_t locklist = NULL;
1310 hammer_mount_t hmp = cursor->trans->hmp;
1316 const int esize = sizeof(*elm);
1318 error = hammer_btree_lock_children(cursor, &locklist);
1321 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1325 * We are splitting but elms[split] will be promoted to the parent,
1326 * leaving the right hand node with one less element. If the
1327 * insertion point will be on the left-hand side adjust the split
1328 * point to give the right hand side one additional node.
1330 node = cursor->node;
1331 ondisk = node->ondisk;
1332 split = (ondisk->count + 1) / 2;
1333 if (cursor->index <= split)
1337 * If we are at the root of the filesystem, create a new root node
1338 * with 1 element and split normally. Avoid making major
1339 * modifications until we know the whole operation will work.
1341 if (ondisk->parent == 0) {
1342 parent = hammer_alloc_btree(cursor->trans, &error);
1345 hammer_lock_ex(&parent->lock);
1346 hammer_modify_node_noundo(cursor->trans, parent);
1347 ondisk = parent->ondisk;
1350 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1351 ondisk->elms[0].base = hmp->root_btree_beg;
1352 ondisk->elms[0].base.btype = node->ondisk->type;
1353 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1354 ondisk->elms[1].base = hmp->root_btree_end;
1355 hammer_modify_node_done(parent);
1356 /* ondisk->elms[1].base.btype - not used */
1358 parent_index = 0; /* index of current node in parent */
1361 parent = cursor->parent;
1362 parent_index = cursor->parent_index;
1366 * Split node into new_node at the split point.
1368 * B O O O P N N B <-- P = node->elms[split]
1369 * 0 1 2 3 4 5 6 <-- subtree indices
1374 * B O O O B B N N B <--- inner boundary points are 'P'
1378 new_node = hammer_alloc_btree(cursor->trans, &error);
1379 if (new_node == NULL) {
1381 hammer_unlock(&parent->lock);
1382 hammer_delete_node(cursor->trans, parent);
1383 hammer_rel_node(parent);
1387 hammer_lock_ex(&new_node->lock);
1390 * Create the new node. P becomes the left-hand boundary in the
1391 * new node. Copy the right-hand boundary as well.
1393 * elm is the new separator.
1395 hammer_modify_node_noundo(cursor->trans, new_node);
1396 hammer_modify_node_all(cursor->trans, node);
1397 ondisk = node->ondisk;
1398 elm = &ondisk->elms[split];
1399 bcopy(elm, &new_node->ondisk->elms[0],
1400 (ondisk->count - split + 1) * esize);
1401 new_node->ondisk->count = ondisk->count - split;
1402 new_node->ondisk->parent = parent->node_offset;
1403 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1404 KKASSERT(ondisk->type == new_node->ondisk->type);
1407 * Cleanup the original node. Elm (P) becomes the new boundary,
1408 * its subtree_offset was moved to the new node. If we had created
1409 * a new root its parent pointer may have changed.
1411 elm->internal.subtree_offset = 0;
1412 ondisk->count = split;
1415 * Insert the separator into the parent, fixup the parent's
1416 * reference to the original node, and reference the new node.
1417 * The separator is P.
1419 * Remember that base.count does not include the right-hand boundary.
1421 hammer_modify_node_all(cursor->trans, parent);
1422 ondisk = parent->ondisk;
1423 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1424 parent_elm = &ondisk->elms[parent_index+1];
1425 bcopy(parent_elm, parent_elm + 1,
1426 (ondisk->count - parent_index) * esize);
1427 parent_elm->internal.base = elm->base; /* separator P */
1428 parent_elm->internal.base.btype = new_node->ondisk->type;
1429 parent_elm->internal.subtree_offset = new_node->node_offset;
1431 hammer_modify_node_done(parent);
1434 * The children of new_node need their parent pointer set to new_node.
1435 * The children have already been locked by
1436 * hammer_btree_lock_children().
1438 for (i = 0; i < new_node->ondisk->count; ++i) {
1439 elm = &new_node->ondisk->elms[i];
1440 error = btree_set_parent(cursor->trans, new_node, elm);
1442 panic("btree_split_internal: btree-fixup problem");
1445 hammer_modify_node_done(new_node);
1448 * The filesystem's root B-Tree pointer may have to be updated.
1451 hammer_volume_t volume;
1453 volume = hammer_get_root_volume(hmp, &error);
1454 KKASSERT(error == 0);
1456 hammer_modify_volume_field(cursor->trans, volume,
1458 volume->ondisk->vol0_btree_root = parent->node_offset;
1459 hammer_modify_volume_done(volume);
1460 node->ondisk->parent = parent->node_offset;
1461 if (cursor->parent) {
1462 hammer_unlock(&cursor->parent->lock);
1463 hammer_rel_node(cursor->parent);
1465 cursor->parent = parent; /* lock'd and ref'd */
1466 hammer_rel_volume(volume, 0);
1468 hammer_modify_node_done(node);
1472 * Ok, now adjust the cursor depending on which element the original
1473 * index was pointing at. If we are >= the split point the push node
1474 * is now in the new node.
1476 * NOTE: If we are at the split point itself we cannot stay with the
1477 * original node because the push index will point at the right-hand
1478 * boundary, which is illegal.
1480 * NOTE: The cursor's parent or parent_index must be adjusted for
1481 * the case where a new parent (new root) was created, and the case
1482 * where the cursor is now pointing at the split node.
1484 if (cursor->index >= split) {
1485 cursor->parent_index = parent_index + 1;
1486 cursor->index -= split;
1487 hammer_unlock(&cursor->node->lock);
1488 hammer_rel_node(cursor->node);
1489 cursor->node = new_node; /* locked and ref'd */
1491 cursor->parent_index = parent_index;
1492 hammer_unlock(&new_node->lock);
1493 hammer_rel_node(new_node);
1497 * Fixup left and right bounds
1499 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1500 cursor->left_bound = &parent_elm[0].internal.base;
1501 cursor->right_bound = &parent_elm[1].internal.base;
1502 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1503 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1504 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1505 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1508 hammer_btree_unlock_children(&locklist);
1509 hammer_cursor_downgrade(cursor);
1514 * Same as the above, but splits a full leaf node.
1520 btree_split_leaf(hammer_cursor_t cursor)
1522 hammer_node_ondisk_t ondisk;
1523 hammer_node_t parent;
1526 hammer_node_t new_leaf;
1527 hammer_btree_elm_t elm;
1528 hammer_btree_elm_t parent_elm;
1529 hammer_base_elm_t mid_boundary;
1534 const size_t esize = sizeof(*elm);
1536 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1539 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1540 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1541 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1542 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1545 * Calculate the split point. If the insertion point will be on
1546 * the left-hand side adjust the split point to give the right
1547 * hand side one additional node.
1549 * Spikes are made up of two leaf elements which cannot be
1552 leaf = cursor->node;
1553 ondisk = leaf->ondisk;
1554 split = (ondisk->count + 1) / 2;
1555 if (cursor->index <= split)
1560 elm = &ondisk->elms[split];
1562 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1563 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1564 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1565 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1568 * If we are at the root of the tree, create a new root node with
1569 * 1 element and split normally. Avoid making major modifications
1570 * until we know the whole operation will work.
1572 if (ondisk->parent == 0) {
1573 parent = hammer_alloc_btree(cursor->trans, &error);
1576 hammer_lock_ex(&parent->lock);
1577 hammer_modify_node_noundo(cursor->trans, parent);
1578 ondisk = parent->ondisk;
1581 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1582 ondisk->elms[0].base = hmp->root_btree_beg;
1583 ondisk->elms[0].base.btype = leaf->ondisk->type;
1584 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1585 ondisk->elms[1].base = hmp->root_btree_end;
1586 /* ondisk->elms[1].base.btype = not used */
1587 hammer_modify_node_done(parent);
1589 parent_index = 0; /* insertion point in parent */
1592 parent = cursor->parent;
1593 parent_index = cursor->parent_index;
1597 * Split leaf into new_leaf at the split point. Select a separator
1598 * value in-between the two leafs but with a bent towards the right
1599 * leaf since comparisons use an 'elm >= separator' inequality.
1608 new_leaf = hammer_alloc_btree(cursor->trans, &error);
1609 if (new_leaf == NULL) {
1611 hammer_unlock(&parent->lock);
1612 hammer_delete_node(cursor->trans, parent);
1613 hammer_rel_node(parent);
1617 hammer_lock_ex(&new_leaf->lock);
1620 * Create the new node and copy the leaf elements from the split
1621 * point on to the new node.
1623 hammer_modify_node_all(cursor->trans, leaf);
1624 hammer_modify_node_noundo(cursor->trans, new_leaf);
1625 ondisk = leaf->ondisk;
1626 elm = &ondisk->elms[split];
1627 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1628 new_leaf->ondisk->count = ondisk->count - split;
1629 new_leaf->ondisk->parent = parent->node_offset;
1630 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1631 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1632 hammer_modify_node_done(new_leaf);
1635 * Cleanup the original node. Because this is a leaf node and
1636 * leaf nodes do not have a right-hand boundary, there
1637 * aren't any special edge cases to clean up. We just fixup the
1640 ondisk->count = split;
1643 * Insert the separator into the parent, fixup the parent's
1644 * reference to the original node, and reference the new node.
1645 * The separator is P.
1647 * Remember that base.count does not include the right-hand boundary.
1648 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1650 hammer_modify_node_all(cursor->trans, parent);
1651 ondisk = parent->ondisk;
1652 KKASSERT(split != 0);
1653 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1654 parent_elm = &ondisk->elms[parent_index+1];
1655 bcopy(parent_elm, parent_elm + 1,
1656 (ondisk->count - parent_index) * esize);
1658 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1659 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1660 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1661 mid_boundary = &parent_elm->base;
1663 hammer_modify_node_done(parent);
1666 * The filesystem's root B-Tree pointer may have to be updated.
1669 hammer_volume_t volume;
1671 volume = hammer_get_root_volume(hmp, &error);
1672 KKASSERT(error == 0);
1674 hammer_modify_volume_field(cursor->trans, volume,
1676 volume->ondisk->vol0_btree_root = parent->node_offset;
1677 hammer_modify_volume_done(volume);
1678 leaf->ondisk->parent = parent->node_offset;
1679 if (cursor->parent) {
1680 hammer_unlock(&cursor->parent->lock);
1681 hammer_rel_node(cursor->parent);
1683 cursor->parent = parent; /* lock'd and ref'd */
1684 hammer_rel_volume(volume, 0);
1686 hammer_modify_node_done(leaf);
1689 * Ok, now adjust the cursor depending on which element the original
1690 * index was pointing at. If we are >= the split point the push node
1691 * is now in the new node.
1693 * NOTE: If we are at the split point itself we need to select the
1694 * old or new node based on where key_beg's insertion point will be.
1695 * If we pick the wrong side the inserted element will wind up in
1696 * the wrong leaf node and outside that node's bounds.
1698 if (cursor->index > split ||
1699 (cursor->index == split &&
1700 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1701 cursor->parent_index = parent_index + 1;
1702 cursor->index -= split;
1703 hammer_unlock(&cursor->node->lock);
1704 hammer_rel_node(cursor->node);
1705 cursor->node = new_leaf;
1707 cursor->parent_index = parent_index;
1708 hammer_unlock(&new_leaf->lock);
1709 hammer_rel_node(new_leaf);
1713 * Fixup left and right bounds
1715 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1716 cursor->left_bound = &parent_elm[0].internal.base;
1717 cursor->right_bound = &parent_elm[1].internal.base;
1720 * Assert that the bounds are correct.
1722 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1723 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1724 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1725 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1726 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1727 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1730 hammer_cursor_downgrade(cursor);
1735 * Recursively correct the right-hand boundary's create_tid to (tid) as
1736 * long as the rest of the key matches. We have to recurse upward in
1737 * the tree as well as down the left side of each parent's right node.
1739 * Return EDEADLK if we were only partially successful, forcing the caller
1740 * to try again. The original cursor is not modified. This routine can
1741 * also fail with EDEADLK if it is forced to throw away a portion of its
1744 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1747 TAILQ_ENTRY(hammer_rhb) entry;
1752 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1755 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1757 struct hammer_rhb_list rhb_list;
1758 hammer_base_elm_t elm;
1759 hammer_node_t orig_node;
1760 struct hammer_rhb *rhb;
1764 TAILQ_INIT(&rhb_list);
1767 * Save our position so we can restore it on return. This also
1768 * gives us a stable 'elm'.
1770 orig_node = cursor->node;
1771 hammer_ref_node(orig_node);
1772 hammer_lock_sh(&orig_node->lock);
1773 orig_index = cursor->index;
1774 elm = &orig_node->ondisk->elms[orig_index].base;
1777 * Now build a list of parents going up, allocating a rhb
1778 * structure for each one.
1780 while (cursor->parent) {
1782 * Stop if we no longer have any right-bounds to fix up
1784 if (elm->obj_id != cursor->right_bound->obj_id ||
1785 elm->rec_type != cursor->right_bound->rec_type ||
1786 elm->key != cursor->right_bound->key) {
1791 * Stop if the right-hand bound's create_tid does not
1792 * need to be corrected.
1794 if (cursor->right_bound->create_tid >= tid)
1797 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1798 rhb->node = cursor->parent;
1799 rhb->index = cursor->parent_index;
1800 hammer_ref_node(rhb->node);
1801 hammer_lock_sh(&rhb->node->lock);
1802 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1804 hammer_cursor_up(cursor);
1808 * now safely adjust the right hand bound for each rhb. This may
1809 * also require taking the right side of the tree and iterating down
1813 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1814 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1817 TAILQ_REMOVE(&rhb_list, rhb, entry);
1818 hammer_unlock(&rhb->node->lock);
1819 hammer_rel_node(rhb->node);
1820 kfree(rhb, M_HAMMER);
1822 switch (cursor->node->ondisk->type) {
1823 case HAMMER_BTREE_TYPE_INTERNAL:
1825 * Right-boundary for parent at internal node
1826 * is one element to the right of the element whos
1827 * right boundary needs adjusting. We must then
1828 * traverse down the left side correcting any left
1829 * bounds (which may now be too far to the left).
1832 error = hammer_btree_correct_lhb(cursor, tid);
1835 panic("hammer_btree_correct_rhb(): Bad node type");
1844 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1845 TAILQ_REMOVE(&rhb_list, rhb, entry);
1846 hammer_unlock(&rhb->node->lock);
1847 hammer_rel_node(rhb->node);
1848 kfree(rhb, M_HAMMER);
1850 error = hammer_cursor_seek(cursor, orig_node, orig_index);
1851 hammer_unlock(&orig_node->lock);
1852 hammer_rel_node(orig_node);
1857 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1858 * bound going downward starting at the current cursor position.
1860 * This function does not restore the cursor after use.
1863 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1865 struct hammer_rhb_list rhb_list;
1866 hammer_base_elm_t elm;
1867 hammer_base_elm_t cmp;
1868 struct hammer_rhb *rhb;
1871 TAILQ_INIT(&rhb_list);
1873 cmp = &cursor->node->ondisk->elms[cursor->index].base;
1876 * Record the node and traverse down the left-hand side for all
1877 * matching records needing a boundary correction.
1881 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1882 rhb->node = cursor->node;
1883 rhb->index = cursor->index;
1884 hammer_ref_node(rhb->node);
1885 hammer_lock_sh(&rhb->node->lock);
1886 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1888 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1890 * Nothing to traverse down if we are at the right
1891 * boundary of an internal node.
1893 if (cursor->index == cursor->node->ondisk->count)
1896 elm = &cursor->node->ondisk->elms[cursor->index].base;
1897 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
1899 panic("Illegal leaf record type %02x", elm->btype);
1901 error = hammer_cursor_down(cursor);
1905 elm = &cursor->node->ondisk->elms[cursor->index].base;
1906 if (elm->obj_id != cmp->obj_id ||
1907 elm->rec_type != cmp->rec_type ||
1908 elm->key != cmp->key) {
1911 if (elm->create_tid >= tid)
1917 * Now we can safely adjust the left-hand boundary from the bottom-up.
1918 * The last element we remove from the list is the caller's right hand
1919 * boundary, which must also be adjusted.
1921 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1922 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1925 TAILQ_REMOVE(&rhb_list, rhb, entry);
1926 hammer_unlock(&rhb->node->lock);
1927 hammer_rel_node(rhb->node);
1928 kfree(rhb, M_HAMMER);
1930 elm = &cursor->node->ondisk->elms[cursor->index].base;
1931 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1932 hammer_modify_node(cursor->trans, cursor->node,
1934 sizeof(elm->create_tid));
1935 elm->create_tid = tid;
1936 hammer_modify_node_done(cursor->node);
1938 panic("hammer_btree_correct_lhb(): Bad element type");
1945 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1946 TAILQ_REMOVE(&rhb_list, rhb, entry);
1947 hammer_unlock(&rhb->node->lock);
1948 hammer_rel_node(rhb->node);
1949 kfree(rhb, M_HAMMER);
1955 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
1956 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
1957 * the operation due to a deadlock, or some other error.
1959 * This routine is always called with an empty, locked leaf but may recurse
1960 * into want-to-be-empty parents as part of its operation.
1962 * On return the cursor may end up pointing to an internal node, suitable
1963 * for further iteration but not for an immediate insertion or deletion.
1966 btree_remove(hammer_cursor_t cursor)
1968 hammer_node_ondisk_t ondisk;
1969 hammer_btree_elm_t elm;
1971 hammer_node_t parent;
1972 const int esize = sizeof(*elm);
1975 node = cursor->node;
1978 * When deleting the root of the filesystem convert it to
1979 * an empty leaf node. Internal nodes cannot be empty.
1981 if (node->ondisk->parent == 0) {
1982 KKASSERT(cursor->parent == NULL);
1983 hammer_modify_node_all(cursor->trans, node);
1984 ondisk = node->ondisk;
1985 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1987 hammer_modify_node_done(node);
1993 * Attempt to remove the parent's reference to the child. If the
1994 * parent would become empty we have to recurse. If we fail we
1995 * leave the parent pointing to an empty leaf node.
1997 parent = cursor->parent;
1999 if (parent->ondisk->count == 1) {
2001 * This special cursor_up_locked() call leaves the original
2002 * node exclusively locked and referenced, leaves the
2003 * original parent locked (as the new node), and locks the
2004 * new parent. It can return EDEADLK.
2006 error = hammer_cursor_up_locked(cursor);
2008 error = btree_remove(cursor);
2010 hammer_modify_node_all(cursor->trans, node);
2011 ondisk = node->ondisk;
2012 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2014 hammer_modify_node_done(node);
2015 hammer_flush_node(node);
2016 hammer_delete_node(cursor->trans, node);
2018 kprintf("Warning: BTREE_REMOVE: Defering "
2019 "parent removal1 @ %016llx, skipping\n",
2022 hammer_unlock(&node->lock);
2023 hammer_rel_node(node);
2025 kprintf("Warning: BTREE_REMOVE: Defering parent "
2026 "removal2 @ %016llx, skipping\n",
2030 KKASSERT(parent->ondisk->count > 1);
2033 * Delete the subtree reference in the parent
2035 hammer_modify_node_all(cursor->trans, parent);
2036 ondisk = parent->ondisk;
2037 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2038 elm = &ondisk->elms[cursor->parent_index];
2039 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2040 KKASSERT(ondisk->count > 0);
2041 bcopy(&elm[1], &elm[0],
2042 (ondisk->count - cursor->parent_index) * esize);
2044 hammer_modify_node_done(parent);
2045 hammer_flush_node(node);
2046 hammer_delete_node(cursor->trans, node);
2049 * cursor->node is invalid, cursor up to make the cursor
2052 error = hammer_cursor_up(cursor);
2058 * The element (elm) has been moved to a new internal node (node).
2060 * If the element represents a pointer to an internal node that node's
2061 * parent must be adjusted to the element's new location.
2063 * XXX deadlock potential here with our exclusive locks
2066 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2067 hammer_btree_elm_t elm)
2069 hammer_node_t child;
2074 switch(elm->base.btype) {
2075 case HAMMER_BTREE_TYPE_INTERNAL:
2076 case HAMMER_BTREE_TYPE_LEAF:
2077 child = hammer_get_node(node->hmp, elm->internal.subtree_offset,
2080 hammer_modify_node_field(trans, child, parent);
2081 child->ondisk->parent = node->node_offset;
2082 hammer_modify_node_done(child);
2083 hammer_rel_node(child);
2093 * Exclusively lock all the children of node. This is used by the split
2094 * code to prevent anyone from accessing the children of a cursor node
2095 * while we fix-up its parent offset.
2097 * If we don't lock the children we can really mess up cursors which block
2098 * trying to cursor-up into our node.
2100 * On failure EDEADLK (or some other error) is returned. If a deadlock
2101 * error is returned the cursor is adjusted to block on termination.
2104 hammer_btree_lock_children(hammer_cursor_t cursor,
2105 struct hammer_node_locklist **locklistp)
2108 hammer_node_locklist_t item;
2109 hammer_node_ondisk_t ondisk;
2110 hammer_btree_elm_t elm;
2111 hammer_node_t child;
2115 node = cursor->node;
2116 ondisk = node->ondisk;
2120 * We really do not want to block on I/O with exclusive locks held,
2121 * pre-get the children before trying to lock the mess.
2123 for (i = 0; i < ondisk->count; ++i) {
2124 elm = &ondisk->elms[i];
2125 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2126 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2129 child = hammer_get_node(node->hmp,
2130 elm->internal.subtree_offset,
2133 hammer_rel_node(child);
2139 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2140 elm = &ondisk->elms[i];
2142 switch(elm->base.btype) {
2143 case HAMMER_BTREE_TYPE_INTERNAL:
2144 case HAMMER_BTREE_TYPE_LEAF:
2145 KKASSERT(elm->internal.subtree_offset != 0);
2146 child = hammer_get_node(node->hmp,
2147 elm->internal.subtree_offset,
2155 if (hammer_lock_ex_try(&child->lock) != 0) {
2156 if (cursor->deadlk_node == NULL) {
2157 cursor->deadlk_node = child;
2158 hammer_ref_node(cursor->deadlk_node);
2161 hammer_rel_node(child);
2163 item = kmalloc(sizeof(*item),
2164 M_HAMMER, M_WAITOK);
2165 item->next = *locklistp;
2172 hammer_btree_unlock_children(locklistp);
2178 * Release previously obtained node locks.
2181 hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2183 hammer_node_locklist_t item;
2185 while ((item = *locklistp) != NULL) {
2186 *locklistp = item->next;
2187 hammer_unlock(&item->node->lock);
2188 hammer_rel_node(item->node);
2189 kfree(item, M_HAMMER);
2193 /************************************************************************
2194 * MISCELLANIOUS SUPPORT *
2195 ************************************************************************/
2198 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2200 * Note that for this particular function a return value of -1, 0, or +1
2201 * can denote a match if create_tid is otherwise discounted. A create_tid
2202 * of zero is considered to be 'infinity' in comparisons.
2204 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2207 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2209 if (key1->localization < key2->localization)
2211 if (key1->localization > key2->localization)
2214 if (key1->obj_id < key2->obj_id)
2216 if (key1->obj_id > key2->obj_id)
2219 if (key1->rec_type < key2->rec_type)
2221 if (key1->rec_type > key2->rec_type)
2224 if (key1->key < key2->key)
2226 if (key1->key > key2->key)
2230 * A create_tid of zero indicates a record which is undeletable
2231 * and must be considered to have a value of positive infinity.
2233 if (key1->create_tid == 0) {
2234 if (key2->create_tid == 0)
2238 if (key2->create_tid == 0)
2240 if (key1->create_tid < key2->create_tid)
2242 if (key1->create_tid > key2->create_tid)
2248 * Test a timestamp against an element to determine whether the
2249 * element is visible. A timestamp of 0 means 'infinity'.
2252 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2255 if (base->delete_tid)
2259 if (asof < base->create_tid)
2261 if (base->delete_tid && asof >= base->delete_tid)
2267 * Create a separator half way inbetween key1 and key2. For fields just
2268 * one unit apart, the separator will match key2. key1 is on the left-hand
2269 * side and key2 is on the right-hand side.
2271 * key2 must be >= the separator. It is ok for the separator to match key2.
2273 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2276 * NOTE: It might be beneficial to just scrap this whole mess and just
2277 * set the separator to key2.
2279 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2280 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2283 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2284 hammer_base_elm_t dest)
2286 bzero(dest, sizeof(*dest));
2288 dest->rec_type = key2->rec_type;
2289 dest->key = key2->key;
2290 dest->obj_id = key2->obj_id;
2291 dest->create_tid = key2->create_tid;
2293 MAKE_SEPARATOR(key1, key2, dest, localization);
2294 if (key1->localization == key2->localization) {
2295 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2296 if (key1->obj_id == key2->obj_id) {
2297 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2298 if (key1->rec_type == key2->rec_type) {
2299 MAKE_SEPARATOR(key1, key2, dest, key);
2301 * Don't bother creating a separator for
2302 * create_tid, which also conveniently avoids
2303 * having to handle the create_tid == 0
2304 * (infinity) case. Just leave create_tid
2307 * Worst case, dest matches key2 exactly,
2308 * which is acceptable.
2315 #undef MAKE_SEPARATOR
2318 * Return whether a generic internal or leaf node is full
2321 btree_node_is_full(hammer_node_ondisk_t node)
2323 switch(node->type) {
2324 case HAMMER_BTREE_TYPE_INTERNAL:
2325 if (node->count == HAMMER_BTREE_INT_ELMS)
2328 case HAMMER_BTREE_TYPE_LEAF:
2329 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2333 panic("illegal btree subtype");
2340 btree_max_elements(u_int8_t type)
2342 if (type == HAMMER_BTREE_TYPE_LEAF)
2343 return(HAMMER_BTREE_LEAF_ELMS);
2344 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2345 return(HAMMER_BTREE_INT_ELMS);
2346 panic("btree_max_elements: bad type %d\n", type);
2351 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2353 hammer_btree_elm_t elm;
2356 kprintf("node %p count=%d parent=%016llx type=%c\n",
2357 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2360 * Dump both boundary elements if an internal node
2362 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2363 for (i = 0; i <= ondisk->count; ++i) {
2364 elm = &ondisk->elms[i];
2365 hammer_print_btree_elm(elm, ondisk->type, i);
2368 for (i = 0; i < ondisk->count; ++i) {
2369 elm = &ondisk->elms[i];
2370 hammer_print_btree_elm(elm, ondisk->type, i);
2376 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2379 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
2380 kprintf("\tkey = %016llx\n", elm->base.key);
2381 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2382 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2383 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2384 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2385 kprintf("\tbtype = %02x (%c)\n",
2387 (elm->base.btype ? elm->base.btype : '?'));
2388 kprintf("\tlocalization = %02x\n", elm->base.localization);
2391 case HAMMER_BTREE_TYPE_INTERNAL:
2392 kprintf("\tsubtree_off = %016llx\n",
2393 elm->internal.subtree_offset);
2395 case HAMMER_BTREE_TYPE_RECORD:
2396 kprintf("\tatime = %016llx\n", elm->leaf.atime);
2397 kprintf("\tdata_offset = %016llx\n", elm->leaf.data_offset);
2398 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2399 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);