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 1.43 2008/05/05 20:34:47 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. Empty
77 * nodes are not allowed and a deletion may recurse upwards from the leaf.
78 * Rather then allow a deadlock a deletion may terminate early by setting
79 * an internal node's element's subtree_offset to 0. The deletion will
80 * then be resumed the next time a search encounters the element.
86 static int btree_search(hammer_cursor_t cursor, int flags);
87 static int btree_split_internal(hammer_cursor_t cursor);
88 static int btree_split_leaf(hammer_cursor_t cursor);
89 static int btree_remove(hammer_cursor_t cursor);
90 static int btree_remove_deleted_element(hammer_cursor_t cursor);
91 static int btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
92 hammer_btree_elm_t elm);
93 static int btree_node_is_full(hammer_node_ondisk_t node);
94 static void hammer_make_separator(hammer_base_elm_t key1,
95 hammer_base_elm_t key2, hammer_base_elm_t dest);
96 static void hammer_btree_unlock_children(
97 struct hammer_node_locklist **locklistp);
100 * Iterate records after a search. The cursor is iterated forwards past
101 * the current record until a record matching the key-range requirements
102 * is found. ENOENT is returned if the iteration goes past the ending
105 * The iteration is inclusive of key_beg and can be inclusive or exclusive
106 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
108 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
109 * may be modified by B-Tree functions.
111 * cursor->key_beg may or may not be modified by this function during
112 * the iteration. XXX future - in case of an inverted lock we may have
113 * to reinitiate the lookup and set key_beg to properly pick up where we
116 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
119 hammer_btree_iterate(hammer_cursor_t cursor)
121 hammer_node_ondisk_t node;
122 hammer_btree_elm_t elm;
128 * Skip past the current record
130 node = cursor->node->ondisk;
133 if (cursor->index < node->count &&
134 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
139 * Loop until an element is found or we are done.
143 * We iterate up the tree and then index over one element
144 * while we are at the last element in the current node.
146 * If we are at the root of the filesystem, cursor_up
149 * XXX this could be optimized by storing the information in
150 * the parent reference.
152 * XXX we can lose the node lock temporarily, this could mess
155 if (cursor->index == node->count) {
156 if (hammer_debug_btree) {
157 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
158 cursor->node->node_offset,
160 (cursor->parent ? cursor->parent->node_offset : -1),
161 cursor->parent_index,
164 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
165 error = hammer_cursor_up(cursor);
168 /* reload stale pointer */
169 node = cursor->node->ondisk;
170 KKASSERT(cursor->index != node->count);
176 * Check internal or leaf element. Determine if the record
177 * at the cursor has gone beyond the end of our range.
179 * We recurse down through internal nodes.
181 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
182 elm = &node->elms[cursor->index];
183 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
184 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
185 if (hammer_debug_btree) {
186 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx %d (td=%p)\n",
187 cursor->node->node_offset,
189 elm[0].internal.base.obj_id,
190 elm[0].internal.base.rec_type,
191 elm[0].internal.base.key,
195 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx %d\n",
196 cursor->node->node_offset,
198 elm[1].internal.base.obj_id,
199 elm[1].internal.base.rec_type,
200 elm[1].internal.base.key,
209 if (r == 0 && (cursor->flags &
210 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
217 * When iterating try to clean up any deleted
218 * internal elements left over from btree_remove()
219 * deadlocks, but it is ok if we can't.
221 if (elm->internal.subtree_offset == 0) {
222 kprintf("REMOVE DELETED ELEMENT\n");
223 btree_remove_deleted_element(cursor);
224 /* note: elm also invalid */
225 } else if (elm->internal.subtree_offset != 0) {
226 error = hammer_cursor_down(cursor);
229 KKASSERT(cursor->index == 0);
231 /* reload stale pointer */
232 node = cursor->node->ondisk;
235 elm = &node->elms[cursor->index];
236 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
237 if (hammer_debug_btree) {
238 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx %d\n",
239 cursor->node->node_offset,
241 (elm[0].leaf.base.btype ?
242 elm[0].leaf.base.btype : '?'),
243 elm[0].leaf.base.obj_id,
244 elm[0].leaf.base.rec_type,
245 elm[0].leaf.base.key,
255 * We support both end-inclusive and
256 * end-exclusive searches.
259 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
264 switch(elm->leaf.base.btype) {
265 case HAMMER_BTREE_TYPE_RECORD:
266 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
267 hammer_btree_chkts(cursor->asof, &elm->base)) {
280 * node pointer invalid after loop
286 if (hammer_debug_btree) {
287 int i = cursor->index;
288 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
289 kprintf("ITERATE %p:%d %016llx %02x %016llx\n",
291 elm->internal.base.obj_id,
292 elm->internal.base.rec_type,
293 elm->internal.base.key
302 * Iterate in the reverse direction. This is used by the pruning code to
303 * avoid overlapping records.
306 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
308 hammer_node_ondisk_t node;
309 hammer_btree_elm_t elm;
315 * Skip past the current record. For various reasons the cursor
316 * may end up set to -1 or set to point at the end of the current
317 * node. These cases must be addressed.
319 node = cursor->node->ondisk;
322 if (cursor->index != -1 &&
323 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
326 if (cursor->index == cursor->node->ondisk->count)
330 * Loop until an element is found or we are done.
334 * We iterate up the tree and then index over one element
335 * while we are at the last element in the current node.
337 if (cursor->index == -1) {
338 error = hammer_cursor_up(cursor);
340 cursor->index = 0; /* sanity */
343 /* reload stale pointer */
344 node = cursor->node->ondisk;
345 KKASSERT(cursor->index != node->count);
351 * Check internal or leaf element. Determine if the record
352 * at the cursor has gone beyond the end of our range.
354 * We recurse down through internal nodes.
356 KKASSERT(cursor->index != node->count);
357 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
358 elm = &node->elms[cursor->index];
359 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
360 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
361 if (hammer_debug_btree) {
362 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx %d\n",
363 cursor->node->node_offset,
365 elm[0].internal.base.obj_id,
366 elm[0].internal.base.rec_type,
367 elm[0].internal.base.key,
370 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx %d\n",
371 cursor->node->node_offset,
373 elm[1].internal.base.obj_id,
374 elm[1].internal.base.rec_type,
375 elm[1].internal.base.key,
387 * When iterating try to clean up any deleted
388 * internal elements left over from btree_remove()
389 * deadlocks, but it is ok if we can't.
391 if (elm->internal.subtree_offset == 0) {
392 btree_remove_deleted_element(cursor);
393 /* note: elm also invalid */
394 } else if (elm->internal.subtree_offset != 0) {
395 error = hammer_cursor_down(cursor);
398 KKASSERT(cursor->index == 0);
399 cursor->index = cursor->node->ondisk->count - 1;
401 /* reload stale pointer */
402 node = cursor->node->ondisk;
405 elm = &node->elms[cursor->index];
406 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
407 if (hammer_debug_btree) {
408 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx %d\n",
409 cursor->node->node_offset,
411 (elm[0].leaf.base.btype ?
412 elm[0].leaf.base.btype : '?'),
413 elm[0].leaf.base.obj_id,
414 elm[0].leaf.base.rec_type,
415 elm[0].leaf.base.key,
424 switch(elm->leaf.base.btype) {
425 case HAMMER_BTREE_TYPE_RECORD:
426 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
427 hammer_btree_chkts(cursor->asof, &elm->base)) {
440 * node pointer invalid after loop
446 if (hammer_debug_btree) {
447 int i = cursor->index;
448 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
449 kprintf("ITERATE %p:%d %016llx %02x %016llx\n",
451 elm->internal.base.obj_id,
452 elm->internal.base.rec_type,
453 elm->internal.base.key
462 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
463 * could not be found, EDEADLK if inserting and a retry is needed, and a
464 * fatal error otherwise. When retrying, the caller must terminate the
465 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
467 * The cursor is suitably positioned for a deletion on success, and suitably
468 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
471 * The cursor may begin anywhere, the search will traverse the tree in
472 * either direction to locate the requested element.
474 * Most of the logic implementing historical searches is handled here. We
475 * do an initial lookup with create_tid set to the asof TID. Due to the
476 * way records are laid out, a backwards iteration may be required if
477 * ENOENT is returned to locate the historical record. Here's the
480 * create_tid: 10 15 20
484 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
485 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
486 * not visible and thus causes ENOENT to be returned. We really need
487 * to check record 11 in LEAF1. If it also fails then the search fails
488 * (e.g. it might represent the range 11-16 and thus still not match our
489 * AS-OF timestamp of 17).
491 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
492 * and the cursor->create_check TID if an iteration might be needed.
493 * In the above example create_check would be set to 14.
496 hammer_btree_lookup(hammer_cursor_t cursor)
500 if (cursor->flags & HAMMER_CURSOR_ASOF) {
501 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
502 cursor->key_beg.create_tid = cursor->asof;
504 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
505 error = btree_search(cursor, 0);
506 if (error != ENOENT ||
507 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
510 * Stop if error other then ENOENT.
511 * Stop if ENOENT and not special case.
515 if (hammer_debug_btree) {
516 kprintf("CREATE_CHECK %016llx\n",
517 cursor->create_check);
519 cursor->key_beg.create_tid = cursor->create_check;
523 error = btree_search(cursor, 0);
525 if (error == 0 && cursor->flags)
526 error = hammer_btree_extract(cursor, cursor->flags);
531 * Execute the logic required to start an iteration. The first record
532 * located within the specified range is returned and iteration control
533 * flags are adjusted for successive hammer_btree_iterate() calls.
536 hammer_btree_first(hammer_cursor_t cursor)
540 error = hammer_btree_lookup(cursor);
541 if (error == ENOENT) {
542 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
543 error = hammer_btree_iterate(cursor);
545 cursor->flags |= HAMMER_CURSOR_ATEDISK;
550 * Similarly but for an iteration in the reverse direction.
553 hammer_btree_last(hammer_cursor_t cursor)
555 struct hammer_base_elm save;
558 save = cursor->key_beg;
559 cursor->key_beg = cursor->key_end;
560 error = hammer_btree_lookup(cursor);
561 cursor->key_beg = save;
562 if (error == ENOENT ||
563 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
564 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
565 error = hammer_btree_iterate_reverse(cursor);
567 cursor->flags |= HAMMER_CURSOR_ATEDISK;
572 * Extract the record and/or data associated with the cursor's current
573 * position. Any prior record or data stored in the cursor is replaced.
574 * The cursor must be positioned at a leaf node.
576 * NOTE: All extractions occur at the leaf of the B-Tree.
579 hammer_btree_extract(hammer_cursor_t cursor, int flags)
582 hammer_node_ondisk_t node;
583 hammer_btree_elm_t elm;
584 hammer_off_t rec_off;
585 hammer_off_t data_off;
590 * The case where the data reference resolves to the same buffer
591 * as the record reference must be handled.
593 node = cursor->node->ondisk;
594 elm = &node->elms[cursor->index];
596 hmp = cursor->node->hmp;
597 flags |= cursor->flags & HAMMER_CURSOR_DATAEXTOK;
600 * There is nothing to extract for an internal element.
602 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
606 * Only record types have data.
608 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
609 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
610 flags &= ~HAMMER_CURSOR_GET_DATA;
611 data_off = elm->leaf.data_offset;
612 data_len = elm->leaf.data_len;
614 flags &= ~HAMMER_CURSOR_GET_DATA;
615 rec_off = elm->leaf.rec_offset;
618 * Extract the record if the record was requested or the data
619 * resides in the record buf.
621 if ((flags & HAMMER_CURSOR_GET_RECORD) ||
622 ((flags & HAMMER_CURSOR_GET_DATA) &&
623 ((rec_off ^ data_off) & ~HAMMER_BUFMASK64) == 0)) {
624 cursor->record = hammer_bread(hmp, rec_off, &error,
625 &cursor->record_buffer);
626 if (hammer_crc_test_record(cursor->record) == 0) {
627 Debugger("CRC FAILED: RECORD");
633 if ((flags & HAMMER_CURSOR_GET_DATA) && error == 0) {
634 if ((rec_off ^ data_off) & ~HAMMER_BUFMASK64) {
636 * Data and record are in different buffers.
638 cursor->data = hammer_bread(hmp, data_off, &error,
639 &cursor->data_buffer);
642 * Data resides in same buffer as record.
644 cursor->data = (void *)
645 ((char *)cursor->record_buffer->ondisk +
646 ((int32_t)data_off & HAMMER_BUFMASK));
648 KKASSERT(data_len >= 0 && data_len <= HAMMER_BUFSIZE);
650 crc32(cursor->data, data_len) != elm->leaf.data_crc) {
651 Debugger("CRC FAILED: DATA");
659 * Insert a leaf element into the B-Tree at the current cursor position.
660 * The cursor is positioned such that the element at and beyond the cursor
661 * are shifted to make room for the new record.
663 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
664 * flag set and that call must return ENOENT before this function can be
667 * The caller may depend on the cursor's exclusive lock after return to
668 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
670 * ENOSPC is returned if there is no room to insert a new record.
673 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_elm_t elm)
675 hammer_node_ondisk_t node;
679 if ((error = hammer_cursor_upgrade(cursor)) != 0)
683 * Insert the element at the leaf node and update the count in the
684 * parent. It is possible for parent to be NULL, indicating that
685 * the filesystem's ROOT B-Tree node is a leaf itself, which is
686 * possible. The root inode can never be deleted so the leaf should
689 * Remember that the right-hand boundary is not included in the
692 hammer_modify_node_all(cursor->trans, cursor->node);
693 node = cursor->node->ondisk;
695 KKASSERT(elm->base.btype != 0);
696 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
697 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
698 if (i != node->count) {
699 bcopy(&node->elms[i], &node->elms[i+1],
700 (node->count - i) * sizeof(*elm));
702 node->elms[i] = *elm;
704 hammer_modify_node_done(cursor->node);
707 * Debugging sanity checks.
709 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
710 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
712 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->leaf.base) < 0);
714 if (i != node->count - 1)
715 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->leaf.base) > 0);
721 * Delete a record from the B-Tree at the current cursor position.
722 * The cursor is positioned such that the current element is the one
725 * On return the cursor will be positioned after the deleted element and
726 * MAY point to an internal node. It will be suitable for the continuation
727 * of an iteration but not for an insertion or deletion.
729 * Deletions will attempt to partially rebalance the B-Tree in an upward
730 * direction, but will terminate rather then deadlock. Empty leaves are
731 * not allowed. An early termination will leave an internal node with an
732 * element whos subtree_offset is 0, a case detected and handled by
735 * This function can return EDEADLK, requiring the caller to retry the
736 * operation after clearing the deadlock.
739 hammer_btree_delete(hammer_cursor_t cursor)
741 hammer_node_ondisk_t ondisk;
743 hammer_node_t parent;
747 if ((error = hammer_cursor_upgrade(cursor)) != 0)
751 * Delete the element from the leaf node.
753 * Remember that leaf nodes do not have boundaries.
756 ondisk = node->ondisk;
759 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
760 KKASSERT(i >= 0 && i < ondisk->count);
761 hammer_modify_node_all(cursor->trans, node);
762 if (i + 1 != ondisk->count) {
763 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
764 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
767 hammer_modify_node_done(node);
770 * Validate local parent
772 if (ondisk->parent) {
773 parent = cursor->parent;
775 KKASSERT(parent != NULL);
776 KKASSERT(parent->node_offset == ondisk->parent);
780 * If the leaf becomes empty it must be detached from the parent,
781 * potentially recursing through to the filesystem root.
783 * This may reposition the cursor at one of the parent's of the
786 * Ignore deadlock errors, that simply means that btree_remove
787 * was unable to recurse and had to leave the subtree_offset
788 * in the parent set to 0.
790 KKASSERT(cursor->index <= ondisk->count);
791 if (ondisk->count == 0) {
793 error = btree_remove(cursor);
794 } while (error == EAGAIN);
795 if (error == EDEADLK)
800 KKASSERT(cursor->parent == NULL ||
801 cursor->parent_index < cursor->parent->ondisk->count);
806 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
808 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
810 * The search can begin ANYWHERE in the B-Tree. As a first step the search
811 * iterates up the tree as necessary to properly position itself prior to
812 * actually doing the sarch.
814 * INSERTIONS: The search will split full nodes and leaves on its way down
815 * and guarentee that the leaf it ends up on is not full. If we run out
816 * of space the search continues to the leaf (to position the cursor for
817 * the spike), but ENOSPC is returned.
819 * The search is only guarenteed to end up on a leaf if an error code of 0
820 * is returned, or if inserting and an error code of ENOENT is returned.
821 * Otherwise it can stop at an internal node. On success a search returns
824 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
825 * filesystem, and it is not simple code. Please note the following facts:
827 * - Internal node recursions have a boundary on the left AND right. The
828 * right boundary is non-inclusive. The create_tid is a generic part
829 * of the key for internal nodes.
831 * - Leaf nodes contain terminal elements only now.
833 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
834 * historical search. ASOF and INSERT are mutually exclusive. When
835 * doing an as-of lookup btree_search() checks for a right-edge boundary
836 * case. If while recursing down the left-edge differs from the key
837 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
838 * with cursor->create_check. This is used by btree_lookup() to iterate.
839 * The iteration backwards because as-of searches can wind up going
840 * down the wrong branch of the B-Tree.
844 btree_search(hammer_cursor_t cursor, int flags)
846 hammer_node_ondisk_t node;
847 hammer_btree_elm_t elm;
854 flags |= cursor->flags;
856 if (hammer_debug_btree) {
857 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx (td = %p)\n",
858 cursor->node->node_offset,
860 cursor->key_beg.obj_id,
861 cursor->key_beg.rec_type,
863 cursor->key_beg.create_tid,
867 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
868 cursor->parent->node_offset, cursor->parent_index,
869 cursor->left_bound->obj_id,
870 cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
871 cursor->right_bound->obj_id,
872 cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
874 &cursor->parent->ondisk->elms[cursor->parent_index],
876 &cursor->parent->ondisk->elms[cursor->parent_index+1]
881 * Move our cursor up the tree until we find a node whos range covers
882 * the key we are trying to locate.
884 * The left bound is inclusive, the right bound is non-inclusive.
885 * It is ok to cursor up too far.
888 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
889 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
892 KKASSERT(cursor->parent);
893 error = hammer_cursor_up(cursor);
899 * The delete-checks below are based on node, not parent. Set the
900 * initial delete-check based on the parent.
903 KKASSERT(cursor->left_bound->create_tid != 1);
904 cursor->create_check = cursor->left_bound->create_tid - 1;
905 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
909 * We better have ended up with a node somewhere.
911 KKASSERT(cursor->node != NULL);
914 * If we are inserting we can't start at a full node if the parent
915 * is also full (because there is no way to split the node),
916 * continue running up the tree until the requirement is satisfied
917 * or we hit the root of the filesystem.
919 * (If inserting we aren't doing an as-of search so we don't have
920 * to worry about create_check).
922 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
923 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
924 if (btree_node_is_full(cursor->node->ondisk) == 0)
927 if (btree_node_is_full(cursor->node->ondisk) ==0)
930 if (cursor->node->ondisk->parent == 0 ||
931 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
934 error = hammer_cursor_up(cursor);
935 /* node may have become stale */
942 * Push down through internal nodes to locate the requested key.
944 node = cursor->node->ondisk;
945 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
947 * Scan the node to find the subtree index to push down into.
948 * We go one-past, then back-up.
950 * We must proactively remove deleted elements which may
951 * have been left over from a deadlocked btree_remove().
953 * The left and right boundaries are included in the loop
954 * in order to detect edge cases.
956 * If the separator only differs by create_tid (r == 1)
957 * and we are doing an as-of search, we may end up going
958 * down a branch to the left of the one containing the
959 * desired key. This requires numerous special cases.
961 if (hammer_debug_btree) {
962 kprintf("SEARCH-I %016llx count=%d\n",
963 cursor->node->node_offset,
966 for (i = 0; i <= node->count; ++i) {
967 elm = &node->elms[i];
968 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
969 if (hammer_debug_btree > 2) {
970 kprintf(" IELM %p %d r=%d\n",
971 &node->elms[i], i, r);
976 KKASSERT(elm->base.create_tid != 1);
977 cursor->create_check = elm->base.create_tid - 1;
978 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
981 if (hammer_debug_btree) {
982 kprintf("SEARCH-I preI=%d/%d r=%d\n",
987 * These cases occur when the parent's idea of the boundary
988 * is wider then the child's idea of the boundary, and
989 * require special handling. If not inserting we can
990 * terminate the search early for these cases but the
991 * child's boundaries cannot be unconditionally modified.
995 * If i == 0 the search terminated to the LEFT of the
996 * left_boundary but to the RIGHT of the parent's left
1001 elm = &node->elms[0];
1004 * If we aren't inserting we can stop here.
1006 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1012 * Correct a left-hand boundary mismatch.
1014 * We can only do this if we can upgrade the lock.
1016 * WARNING: We can only do this if inserting, i.e.
1017 * we are running on the backend.
1019 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1021 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1022 hammer_modify_node_field(cursor->trans, cursor->node,
1024 save = node->elms[0].base.btype;
1025 node->elms[0].base = *cursor->left_bound;
1026 node->elms[0].base.btype = save;
1027 hammer_modify_node_done(cursor->node);
1028 } else if (i == node->count + 1) {
1030 * If i == node->count + 1 the search terminated to
1031 * the RIGHT of the right boundary but to the LEFT
1032 * of the parent's right boundary. If we aren't
1033 * inserting we can stop here.
1035 * Note that the last element in this case is
1036 * elms[i-2] prior to adjustments to 'i'.
1039 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
1045 * Correct a right-hand boundary mismatch.
1046 * (actual push-down record is i-2 prior to
1047 * adjustments to i).
1049 * We can only do this if we can upgrade the lock.
1051 * WARNING: We can only do this if inserting, i.e.
1052 * we are running on the backend.
1054 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1056 elm = &node->elms[i];
1057 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1058 hammer_modify_node(cursor->trans, cursor->node,
1059 &elm->base, sizeof(elm->base));
1060 elm->base = *cursor->right_bound;
1061 hammer_modify_node_done(cursor->node);
1065 * The push-down index is now i - 1. If we had
1066 * terminated on the right boundary this will point
1067 * us at the last element.
1072 elm = &node->elms[i];
1074 if (hammer_debug_btree) {
1075 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1076 "key=%016llx cre=%016llx\n",
1077 cursor->node->node_offset,
1079 elm->internal.base.obj_id,
1080 elm->internal.base.rec_type,
1081 elm->internal.base.key,
1082 elm->internal.base.create_tid
1087 * When searching try to clean up any deleted
1088 * internal elements left over from btree_remove()
1091 * If we fail and we are doing an insertion lookup,
1092 * we have to return EDEADLK, because an insertion lookup
1093 * must terminate at a leaf.
1095 if (elm->internal.subtree_offset == 0) {
1096 error = btree_remove_deleted_element(cursor);
1099 if (error == EDEADLK &&
1100 (flags & HAMMER_CURSOR_INSERT) == 0) {
1108 * Handle insertion and deletion requirements.
1110 * If inserting split full nodes. The split code will
1111 * adjust cursor->node and cursor->index if the current
1112 * index winds up in the new node.
1114 * If inserting and a left or right edge case was detected,
1115 * we cannot correct the left or right boundary and must
1116 * prepend and append an empty leaf node in order to make
1117 * the boundary correction.
1119 * If we run out of space we set enospc and continue on
1120 * to a leaf to provide the spike code with a good point
1123 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1124 if (btree_node_is_full(node)) {
1125 error = btree_split_internal(cursor);
1127 if (error != ENOSPC)
1132 * reload stale pointers
1135 node = cursor->node->ondisk;
1140 * Push down (push into new node, existing node becomes
1141 * the parent) and continue the search.
1143 error = hammer_cursor_down(cursor);
1144 /* node may have become stale */
1147 node = cursor->node->ondisk;
1151 * We are at a leaf, do a linear search of the key array.
1153 * If we encounter a spike element type within the necessary
1154 * range we push into it.
1156 * On success the index is set to the matching element and 0
1159 * On failure the index is set to the insertion point and ENOENT
1162 * Boundaries are not stored in leaf nodes, so the index can wind
1163 * up to the left of element 0 (index == 0) or past the end of
1164 * the array (index == node->count).
1166 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1167 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1168 if (hammer_debug_btree) {
1169 kprintf("SEARCH-L %016llx count=%d\n",
1170 cursor->node->node_offset,
1174 for (i = 0; i < node->count; ++i) {
1175 elm = &node->elms[i];
1177 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1179 if (hammer_debug_btree > 1)
1180 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1183 * We are at a record element. Stop if we've flipped past
1184 * key_beg, not counting the create_tid test. Allow the
1185 * r == 1 case (key_beg > element but differs only by its
1186 * create_tid) to fall through to the AS-OF check.
1188 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1196 * Check our as-of timestamp against the element.
1198 if (flags & HAMMER_CURSOR_ASOF) {
1199 if (hammer_btree_chkts(cursor->asof,
1200 &node->elms[i].base) != 0) {
1205 if (r > 0) /* can only be +1 */
1211 if (hammer_debug_btree) {
1212 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1213 cursor->node->node_offset, i);
1219 * The search of the leaf node failed. i is the insertion point.
1222 if (hammer_debug_btree) {
1223 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1224 cursor->node->node_offset, i);
1228 * No exact match was found, i is now at the insertion point.
1230 * If inserting split a full leaf before returning. This
1231 * may have the side effect of adjusting cursor->node and
1235 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1236 btree_node_is_full(node)) {
1237 error = btree_split_leaf(cursor);
1239 if (error != ENOSPC)
1244 * reload stale pointers
1248 node = &cursor->node->internal;
1253 * We reached a leaf but did not find the key we were looking for.
1254 * If this is an insert we will be properly positioned for an insert
1255 * (ENOENT) or spike (ENOSPC) operation.
1257 error = enospc ? ENOSPC : ENOENT;
1263 /************************************************************************
1264 * SPLITTING AND MERGING *
1265 ************************************************************************
1267 * These routines do all the dirty work required to split and merge nodes.
1271 * Split an internal node into two nodes and move the separator at the split
1272 * point to the parent.
1274 * (cursor->node, cursor->index) indicates the element the caller intends
1275 * to push into. We will adjust node and index if that element winds
1276 * up in the split node.
1278 * If we are at the root of the filesystem a new root must be created with
1279 * two elements, one pointing to the original root and one pointing to the
1280 * newly allocated split node.
1284 btree_split_internal(hammer_cursor_t cursor)
1286 hammer_node_ondisk_t ondisk;
1288 hammer_node_t parent;
1289 hammer_node_t new_node;
1290 hammer_btree_elm_t elm;
1291 hammer_btree_elm_t parent_elm;
1292 hammer_node_locklist_t locklist = NULL;
1293 hammer_mount_t hmp = cursor->trans->hmp;
1299 const int esize = sizeof(*elm);
1301 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1303 error = hammer_btree_lock_children(cursor, &locklist);
1308 * We are splitting but elms[split] will be promoted to the parent,
1309 * leaving the right hand node with one less element. If the
1310 * insertion point will be on the left-hand side adjust the split
1311 * point to give the right hand side one additional node.
1313 node = cursor->node;
1314 ondisk = node->ondisk;
1315 split = (ondisk->count + 1) / 2;
1316 if (cursor->index <= split)
1320 * If we are at the root of the filesystem, create a new root node
1321 * with 1 element and split normally. Avoid making major
1322 * modifications until we know the whole operation will work.
1324 if (ondisk->parent == 0) {
1325 parent = hammer_alloc_btree(cursor->trans, &error);
1328 hammer_lock_ex(&parent->lock);
1329 hammer_modify_node_noundo(cursor->trans, parent);
1330 ondisk = parent->ondisk;
1333 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1334 ondisk->elms[0].base = hmp->root_btree_beg;
1335 ondisk->elms[0].base.btype = node->ondisk->type;
1336 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1337 ondisk->elms[1].base = hmp->root_btree_end;
1338 hammer_modify_node_done(parent);
1339 /* ondisk->elms[1].base.btype - not used */
1341 parent_index = 0; /* index of current node in parent */
1344 parent = cursor->parent;
1345 parent_index = cursor->parent_index;
1349 * Split node into new_node at the split point.
1351 * B O O O P N N B <-- P = node->elms[split]
1352 * 0 1 2 3 4 5 6 <-- subtree indices
1357 * B O O O B B N N B <--- inner boundary points are 'P'
1361 new_node = hammer_alloc_btree(cursor->trans, &error);
1362 if (new_node == NULL) {
1364 hammer_unlock(&parent->lock);
1365 hammer_delete_node(cursor->trans, parent);
1366 hammer_rel_node(parent);
1370 hammer_lock_ex(&new_node->lock);
1373 * Create the new node. P becomes the left-hand boundary in the
1374 * new node. Copy the right-hand boundary as well.
1376 * elm is the new separator.
1378 hammer_modify_node_noundo(cursor->trans, new_node);
1379 hammer_modify_node_all(cursor->trans, node);
1380 ondisk = node->ondisk;
1381 elm = &ondisk->elms[split];
1382 bcopy(elm, &new_node->ondisk->elms[0],
1383 (ondisk->count - split + 1) * esize);
1384 new_node->ondisk->count = ondisk->count - split;
1385 new_node->ondisk->parent = parent->node_offset;
1386 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1387 KKASSERT(ondisk->type == new_node->ondisk->type);
1390 * Cleanup the original node. Elm (P) becomes the new boundary,
1391 * its subtree_offset was moved to the new node. If we had created
1392 * a new root its parent pointer may have changed.
1394 elm->internal.subtree_offset = 0;
1395 ondisk->count = split;
1398 * Insert the separator into the parent, fixup the parent's
1399 * reference to the original node, and reference the new node.
1400 * The separator is P.
1402 * Remember that base.count does not include the right-hand boundary.
1404 hammer_modify_node_all(cursor->trans, parent);
1405 ondisk = parent->ondisk;
1406 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1407 parent_elm = &ondisk->elms[parent_index+1];
1408 bcopy(parent_elm, parent_elm + 1,
1409 (ondisk->count - parent_index) * esize);
1410 parent_elm->internal.base = elm->base; /* separator P */
1411 parent_elm->internal.base.btype = new_node->ondisk->type;
1412 parent_elm->internal.subtree_offset = new_node->node_offset;
1414 hammer_modify_node_done(parent);
1417 * The children of new_node need their parent pointer set to new_node.
1418 * The children have already been locked by
1419 * hammer_btree_lock_children().
1421 for (i = 0; i < new_node->ondisk->count; ++i) {
1422 elm = &new_node->ondisk->elms[i];
1423 error = btree_set_parent(cursor->trans, new_node, elm);
1425 panic("btree_split_internal: btree-fixup problem");
1428 hammer_modify_node_done(new_node);
1431 * The filesystem's root B-Tree pointer may have to be updated.
1434 hammer_volume_t volume;
1436 volume = hammer_get_root_volume(hmp, &error);
1437 KKASSERT(error == 0);
1439 hammer_modify_volume_field(cursor->trans, volume,
1441 volume->ondisk->vol0_btree_root = parent->node_offset;
1442 hammer_modify_volume_done(volume);
1443 node->ondisk->parent = parent->node_offset;
1444 if (cursor->parent) {
1445 hammer_unlock(&cursor->parent->lock);
1446 hammer_rel_node(cursor->parent);
1448 cursor->parent = parent; /* lock'd and ref'd */
1449 hammer_rel_volume(volume, 0);
1451 hammer_modify_node_done(node);
1455 * Ok, now adjust the cursor depending on which element the original
1456 * index was pointing at. If we are >= the split point the push node
1457 * is now in the new node.
1459 * NOTE: If we are at the split point itself we cannot stay with the
1460 * original node because the push index will point at the right-hand
1461 * boundary, which is illegal.
1463 * NOTE: The cursor's parent or parent_index must be adjusted for
1464 * the case where a new parent (new root) was created, and the case
1465 * where the cursor is now pointing at the split node.
1467 if (cursor->index >= split) {
1468 cursor->parent_index = parent_index + 1;
1469 cursor->index -= split;
1470 hammer_unlock(&cursor->node->lock);
1471 hammer_rel_node(cursor->node);
1472 cursor->node = new_node; /* locked and ref'd */
1474 cursor->parent_index = parent_index;
1475 hammer_unlock(&new_node->lock);
1476 hammer_rel_node(new_node);
1480 * Fixup left and right bounds
1482 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1483 cursor->left_bound = &parent_elm[0].internal.base;
1484 cursor->right_bound = &parent_elm[1].internal.base;
1485 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1486 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1487 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1488 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1491 hammer_btree_unlock_children(&locklist);
1492 hammer_cursor_downgrade(cursor);
1497 * Same as the above, but splits a full leaf node.
1503 btree_split_leaf(hammer_cursor_t cursor)
1505 hammer_node_ondisk_t ondisk;
1506 hammer_node_t parent;
1509 hammer_node_t new_leaf;
1510 hammer_btree_elm_t elm;
1511 hammer_btree_elm_t parent_elm;
1512 hammer_base_elm_t mid_boundary;
1517 const size_t esize = sizeof(*elm);
1519 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1522 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1523 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1524 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1525 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1528 * Calculate the split point. If the insertion point will be on
1529 * the left-hand side adjust the split point to give the right
1530 * hand side one additional node.
1532 * Spikes are made up of two leaf elements which cannot be
1535 leaf = cursor->node;
1536 ondisk = leaf->ondisk;
1537 split = (ondisk->count + 1) / 2;
1538 if (cursor->index <= split)
1543 elm = &ondisk->elms[split];
1545 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1546 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1547 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1548 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1551 * If we are at the root of the tree, create a new root node with
1552 * 1 element and split normally. Avoid making major modifications
1553 * until we know the whole operation will work.
1555 if (ondisk->parent == 0) {
1556 parent = hammer_alloc_btree(cursor->trans, &error);
1559 hammer_lock_ex(&parent->lock);
1560 hammer_modify_node_noundo(cursor->trans, parent);
1561 ondisk = parent->ondisk;
1564 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1565 ondisk->elms[0].base = hmp->root_btree_beg;
1566 ondisk->elms[0].base.btype = leaf->ondisk->type;
1567 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1568 ondisk->elms[1].base = hmp->root_btree_end;
1569 /* ondisk->elms[1].base.btype = not used */
1570 hammer_modify_node_done(parent);
1572 parent_index = 0; /* insertion point in parent */
1575 parent = cursor->parent;
1576 parent_index = cursor->parent_index;
1580 * Split leaf into new_leaf at the split point. Select a separator
1581 * value in-between the two leafs but with a bent towards the right
1582 * leaf since comparisons use an 'elm >= separator' inequality.
1591 new_leaf = hammer_alloc_btree(cursor->trans, &error);
1592 if (new_leaf == NULL) {
1594 hammer_unlock(&parent->lock);
1595 hammer_delete_node(cursor->trans, parent);
1596 hammer_rel_node(parent);
1600 hammer_lock_ex(&new_leaf->lock);
1603 * Create the new node and copy the leaf elements from the split
1604 * point on to the new node.
1606 hammer_modify_node_all(cursor->trans, leaf);
1607 hammer_modify_node_noundo(cursor->trans, new_leaf);
1608 ondisk = leaf->ondisk;
1609 elm = &ondisk->elms[split];
1610 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1611 new_leaf->ondisk->count = ondisk->count - split;
1612 new_leaf->ondisk->parent = parent->node_offset;
1613 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1614 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1615 hammer_modify_node_done(new_leaf);
1618 * Cleanup the original node. Because this is a leaf node and
1619 * leaf nodes do not have a right-hand boundary, there
1620 * aren't any special edge cases to clean up. We just fixup the
1623 ondisk->count = split;
1626 * Insert the separator into the parent, fixup the parent's
1627 * reference to the original node, and reference the new node.
1628 * The separator is P.
1630 * Remember that base.count does not include the right-hand boundary.
1631 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1633 hammer_modify_node_all(cursor->trans, parent);
1634 ondisk = parent->ondisk;
1635 KKASSERT(split != 0);
1636 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1637 parent_elm = &ondisk->elms[parent_index+1];
1638 bcopy(parent_elm, parent_elm + 1,
1639 (ondisk->count - parent_index) * esize);
1641 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1642 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1643 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1644 mid_boundary = &parent_elm->base;
1646 hammer_modify_node_done(parent);
1649 * The filesystem's root B-Tree pointer may have to be updated.
1652 hammer_volume_t volume;
1654 volume = hammer_get_root_volume(hmp, &error);
1655 KKASSERT(error == 0);
1657 hammer_modify_volume_field(cursor->trans, volume,
1659 volume->ondisk->vol0_btree_root = parent->node_offset;
1660 hammer_modify_volume_done(volume);
1661 leaf->ondisk->parent = parent->node_offset;
1662 if (cursor->parent) {
1663 hammer_unlock(&cursor->parent->lock);
1664 hammer_rel_node(cursor->parent);
1666 cursor->parent = parent; /* lock'd and ref'd */
1667 hammer_rel_volume(volume, 0);
1669 hammer_modify_node_done(leaf);
1672 * Ok, now adjust the cursor depending on which element the original
1673 * index was pointing at. If we are >= the split point the push node
1674 * is now in the new node.
1676 * NOTE: If we are at the split point itself we need to select the
1677 * old or new node based on where key_beg's insertion point will be.
1678 * If we pick the wrong side the inserted element will wind up in
1679 * the wrong leaf node and outside that node's bounds.
1681 if (cursor->index > split ||
1682 (cursor->index == split &&
1683 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1684 cursor->parent_index = parent_index + 1;
1685 cursor->index -= split;
1686 hammer_unlock(&cursor->node->lock);
1687 hammer_rel_node(cursor->node);
1688 cursor->node = new_leaf;
1690 cursor->parent_index = parent_index;
1691 hammer_unlock(&new_leaf->lock);
1692 hammer_rel_node(new_leaf);
1696 * Fixup left and right bounds
1698 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1699 cursor->left_bound = &parent_elm[0].internal.base;
1700 cursor->right_bound = &parent_elm[1].internal.base;
1703 * Assert that the bounds are correct.
1705 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1706 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1707 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1708 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1709 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1710 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1713 hammer_cursor_downgrade(cursor);
1718 * Recursively correct the right-hand boundary's create_tid to (tid) as
1719 * long as the rest of the key matches. We have to recurse upward in
1720 * the tree as well as down the left side of each parent's right node.
1722 * Return EDEADLK if we were only partially successful, forcing the caller
1723 * to try again. The original cursor is not modified. This routine can
1724 * also fail with EDEADLK if it is forced to throw away a portion of its
1727 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1730 TAILQ_ENTRY(hammer_rhb) entry;
1735 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1738 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1740 struct hammer_rhb_list rhb_list;
1741 hammer_base_elm_t elm;
1742 hammer_node_t orig_node;
1743 struct hammer_rhb *rhb;
1747 TAILQ_INIT(&rhb_list);
1750 * Save our position so we can restore it on return. This also
1751 * gives us a stable 'elm'.
1753 orig_node = cursor->node;
1754 hammer_ref_node(orig_node);
1755 hammer_lock_sh(&orig_node->lock);
1756 orig_index = cursor->index;
1757 elm = &orig_node->ondisk->elms[orig_index].base;
1760 * Now build a list of parents going up, allocating a rhb
1761 * structure for each one.
1763 while (cursor->parent) {
1765 * Stop if we no longer have any right-bounds to fix up
1767 if (elm->obj_id != cursor->right_bound->obj_id ||
1768 elm->rec_type != cursor->right_bound->rec_type ||
1769 elm->key != cursor->right_bound->key) {
1774 * Stop if the right-hand bound's create_tid does not
1775 * need to be corrected.
1777 if (cursor->right_bound->create_tid >= tid)
1780 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1781 rhb->node = cursor->parent;
1782 rhb->index = cursor->parent_index;
1783 hammer_ref_node(rhb->node);
1784 hammer_lock_sh(&rhb->node->lock);
1785 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1787 hammer_cursor_up(cursor);
1791 * now safely adjust the right hand bound for each rhb. This may
1792 * also require taking the right side of the tree and iterating down
1796 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1797 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1798 kprintf("CORRECT RHB %016llx index %d type=%c\n",
1799 rhb->node->node_offset,
1800 rhb->index, cursor->node->ondisk->type);
1803 TAILQ_REMOVE(&rhb_list, rhb, entry);
1804 hammer_unlock(&rhb->node->lock);
1805 hammer_rel_node(rhb->node);
1806 kfree(rhb, M_HAMMER);
1808 switch (cursor->node->ondisk->type) {
1809 case HAMMER_BTREE_TYPE_INTERNAL:
1811 * Right-boundary for parent at internal node
1812 * is one element to the right of the element whos
1813 * right boundary needs adjusting. We must then
1814 * traverse down the left side correcting any left
1815 * bounds (which may now be too far to the left).
1818 error = hammer_btree_correct_lhb(cursor, tid);
1821 panic("hammer_btree_correct_rhb(): Bad node type");
1830 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1831 TAILQ_REMOVE(&rhb_list, rhb, entry);
1832 hammer_unlock(&rhb->node->lock);
1833 hammer_rel_node(rhb->node);
1834 kfree(rhb, M_HAMMER);
1836 error = hammer_cursor_seek(cursor, orig_node, orig_index);
1837 hammer_unlock(&orig_node->lock);
1838 hammer_rel_node(orig_node);
1843 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1844 * bound going downward starting at the current cursor position.
1846 * This function does not restore the cursor after use.
1849 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1851 struct hammer_rhb_list rhb_list;
1852 hammer_base_elm_t elm;
1853 hammer_base_elm_t cmp;
1854 struct hammer_rhb *rhb;
1857 TAILQ_INIT(&rhb_list);
1859 cmp = &cursor->node->ondisk->elms[cursor->index].base;
1862 * Record the node and traverse down the left-hand side for all
1863 * matching records needing a boundary correction.
1867 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1868 rhb->node = cursor->node;
1869 rhb->index = cursor->index;
1870 hammer_ref_node(rhb->node);
1871 hammer_lock_sh(&rhb->node->lock);
1872 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1874 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1876 * Nothing to traverse down if we are at the right
1877 * boundary of an internal node.
1879 if (cursor->index == cursor->node->ondisk->count)
1882 elm = &cursor->node->ondisk->elms[cursor->index].base;
1883 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
1885 panic("Illegal leaf record type %02x", elm->btype);
1887 error = hammer_cursor_down(cursor);
1891 elm = &cursor->node->ondisk->elms[cursor->index].base;
1892 if (elm->obj_id != cmp->obj_id ||
1893 elm->rec_type != cmp->rec_type ||
1894 elm->key != cmp->key) {
1897 if (elm->create_tid >= tid)
1903 * Now we can safely adjust the left-hand boundary from the bottom-up.
1904 * The last element we remove from the list is the caller's right hand
1905 * boundary, which must also be adjusted.
1907 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1908 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1911 TAILQ_REMOVE(&rhb_list, rhb, entry);
1912 hammer_unlock(&rhb->node->lock);
1913 hammer_rel_node(rhb->node);
1914 kfree(rhb, M_HAMMER);
1916 elm = &cursor->node->ondisk->elms[cursor->index].base;
1917 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1918 kprintf("hammer_btree_correct_lhb-I @%016llx[%d]\n",
1919 cursor->node->node_offset, cursor->index);
1920 hammer_modify_node(cursor->trans, cursor->node,
1922 sizeof(elm->create_tid));
1923 elm->create_tid = tid;
1924 hammer_modify_node_done(cursor->node);
1926 panic("hammer_btree_correct_lhb(): Bad element type");
1933 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1934 TAILQ_REMOVE(&rhb_list, rhb, entry);
1935 hammer_unlock(&rhb->node->lock);
1936 hammer_rel_node(rhb->node);
1937 kfree(rhb, M_HAMMER);
1943 * Attempt to remove the empty B-Tree node at (cursor->node). Returns 0
1944 * on success, EAGAIN if we could not acquire the necessary locks, or some
1945 * other error. This node can be a leaf node or an internal node.
1947 * On return the cursor may end up pointing at an internal node, suitable
1948 * for further iteration but not for an immediate insertion or deletion.
1950 * cursor->node may be an internal node or a leaf node.
1952 * NOTE: If cursor->node has one element it is the parent trying to delete
1953 * that element, make sure cursor->index is properly adjusted on success.
1956 btree_remove(hammer_cursor_t cursor)
1958 hammer_node_ondisk_t ondisk;
1959 hammer_btree_elm_t elm;
1961 hammer_node_t parent;
1962 const int esize = sizeof(*elm);
1965 node = cursor->node;
1968 * When deleting the root of the filesystem convert it to
1969 * an empty leaf node. Internal nodes cannot be empty.
1971 if (node->ondisk->parent == 0) {
1972 hammer_modify_node_all(cursor->trans, node);
1973 ondisk = node->ondisk;
1974 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1976 hammer_modify_node_done(node);
1982 * Zero-out the parent's reference to the child and flag the
1983 * child for destruction. This ensures that the child is not
1984 * reused while other references to it exist.
1986 parent = cursor->parent;
1987 hammer_modify_node_all(cursor->trans, parent);
1988 ondisk = parent->ondisk;
1989 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
1990 elm = &ondisk->elms[cursor->parent_index];
1991 KKASSERT(elm->internal.subtree_offset == node->node_offset);
1992 elm->internal.subtree_offset = 0;
1994 hammer_flush_node(node);
1995 hammer_delete_node(cursor->trans, node);
1998 * If the parent would otherwise not become empty we can physically
1999 * remove the zero'd element. Note however that in order to
2000 * guarentee a valid cursor we still need to be able to cursor up
2001 * because we no longer have a node.
2003 * This collapse will change the parent's boundary elements, making
2004 * them wider. The new boundaries are recursively corrected in
2007 * XXX we can theoretically recalculate the midpoint but there isn't
2008 * much of a reason to do it.
2010 error = hammer_cursor_up(cursor);
2012 error = hammer_cursor_upgrade(cursor);
2015 kprintf("BTREE_REMOVE: Cannot lock parent, skipping\n");
2016 Debugger("BTREE_REMOVE");
2017 hammer_modify_node_done(parent);
2022 * Remove the internal element from the parent. The bcopy must
2023 * include the right boundary element.
2025 KKASSERT(parent == cursor->node && ondisk == parent->ondisk);
2028 /* ondisk is node's ondisk */
2029 /* elm is node's element */
2032 * Remove the internal element that we zero'd out. Tell the caller
2033 * to loop if it hits zero (to try to avoid eating up precious kernel
2036 KKASSERT(ondisk->count > 0);
2037 bcopy(&elm[1], &elm[0], (ondisk->count - cursor->index) * esize);
2039 if (ondisk->count == 0)
2041 hammer_modify_node_done(node);
2046 * Attempt to remove the deleted internal element at the current cursor
2047 * position. If we are unable to remove the element we return EDEADLK.
2049 * If the current internal node becomes empty we delete it in the parent
2050 * and cursor up, looping until we finish or we deadlock.
2052 * On return, if successful, the cursor will be pointing at the next
2053 * iterative position in the B-Tree. If unsuccessful the cursor will be
2054 * pointing at the last deleted internal element that could not be
2059 btree_remove_deleted_element(hammer_cursor_t cursor)
2062 hammer_btree_elm_t elm;
2065 if ((error = hammer_cursor_upgrade(cursor)) != 0)
2067 node = cursor->node;
2068 elm = &node->ondisk->elms[cursor->index];
2069 if (elm->internal.subtree_offset == 0) {
2071 error = btree_remove(cursor);
2072 kprintf("BTREE REMOVE DELETED ELEMENT %d\n", error);
2073 } while (error == EAGAIN);
2079 * The element (elm) has been moved to a new internal node (node).
2081 * If the element represents a pointer to an internal node that node's
2082 * parent must be adjusted to the element's new location.
2084 * XXX deadlock potential here with our exclusive locks
2088 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2089 hammer_btree_elm_t elm)
2091 hammer_node_t child;
2096 switch(elm->base.btype) {
2097 case HAMMER_BTREE_TYPE_INTERNAL:
2098 case HAMMER_BTREE_TYPE_LEAF:
2099 child = hammer_get_node(node->hmp, elm->internal.subtree_offset,
2102 hammer_modify_node_field(trans, child, parent);
2103 child->ondisk->parent = node->node_offset;
2104 hammer_modify_node_done(child);
2105 hammer_rel_node(child);
2115 * Exclusively lock all the children of node. This is used by the split
2116 * code to prevent anyone from accessing the children of a cursor node
2117 * while we fix-up its parent offset.
2119 * If we don't lock the children we can really mess up cursors which block
2120 * trying to cursor-up into our node.
2122 * On failure EDEADLK (or some other error) is returned. If a deadlock
2123 * error is returned the cursor is adjusted to block on termination.
2126 hammer_btree_lock_children(hammer_cursor_t cursor,
2127 struct hammer_node_locklist **locklistp)
2130 hammer_node_locklist_t item;
2131 hammer_node_ondisk_t ondisk;
2132 hammer_btree_elm_t elm;
2133 hammer_node_t child;
2137 node = cursor->node;
2138 ondisk = node->ondisk;
2140 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2141 elm = &ondisk->elms[i];
2143 switch(elm->base.btype) {
2144 case HAMMER_BTREE_TYPE_INTERNAL:
2145 case HAMMER_BTREE_TYPE_LEAF:
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->obj_id < key2->obj_id)
2211 if (key1->obj_id > key2->obj_id)
2214 if (key1->rec_type < key2->rec_type)
2216 if (key1->rec_type > key2->rec_type)
2219 if (key1->key < key2->key)
2221 if (key1->key > key2->key)
2225 * A create_tid of zero indicates a record which is undeletable
2226 * and must be considered to have a value of positive infinity.
2228 if (key1->create_tid == 0) {
2229 if (key2->create_tid == 0)
2233 if (key2->create_tid == 0)
2235 if (key1->create_tid < key2->create_tid)
2237 if (key1->create_tid > key2->create_tid)
2243 * Test a timestamp against an element to determine whether the
2244 * element is visible. A timestamp of 0 means 'infinity'.
2247 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2250 if (base->delete_tid)
2254 if (asof < base->create_tid)
2256 if (base->delete_tid && asof >= base->delete_tid)
2262 * Create a separator half way inbetween key1 and key2. For fields just
2263 * one unit apart, the separator will match key2. key1 is on the left-hand
2264 * side and key2 is on the right-hand side.
2266 * key2 must be >= the separator. It is ok for the separator to match key2.
2268 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2271 * NOTE: It might be beneficial to just scrap this whole mess and just
2272 * set the separator to key2.
2274 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2275 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2278 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2279 hammer_base_elm_t dest)
2281 bzero(dest, sizeof(*dest));
2283 dest->rec_type = key2->rec_type;
2284 dest->key = key2->key;
2285 dest->create_tid = key2->create_tid;
2287 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2288 if (key1->obj_id == key2->obj_id) {
2289 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2290 if (key1->rec_type == key2->rec_type) {
2291 MAKE_SEPARATOR(key1, key2, dest, key);
2293 * Don't bother creating a separator for create_tid,
2294 * which also conveniently avoids having to handle
2295 * the create_tid == 0 (infinity) case. Just leave
2296 * create_tid set to key2.
2298 * Worst case, dest matches key2 exactly, which is
2305 #undef MAKE_SEPARATOR
2308 * Return whether a generic internal or leaf node is full
2311 btree_node_is_full(hammer_node_ondisk_t node)
2313 switch(node->type) {
2314 case HAMMER_BTREE_TYPE_INTERNAL:
2315 if (node->count == HAMMER_BTREE_INT_ELMS)
2318 case HAMMER_BTREE_TYPE_LEAF:
2319 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2323 panic("illegal btree subtype");
2330 btree_max_elements(u_int8_t type)
2332 if (type == HAMMER_BTREE_TYPE_LEAF)
2333 return(HAMMER_BTREE_LEAF_ELMS);
2334 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2335 return(HAMMER_BTREE_INT_ELMS);
2336 panic("btree_max_elements: bad type %d\n", type);
2341 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2343 hammer_btree_elm_t elm;
2346 kprintf("node %p count=%d parent=%016llx type=%c\n",
2347 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2350 * Dump both boundary elements if an internal node
2352 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2353 for (i = 0; i <= ondisk->count; ++i) {
2354 elm = &ondisk->elms[i];
2355 hammer_print_btree_elm(elm, ondisk->type, i);
2358 for (i = 0; i < ondisk->count; ++i) {
2359 elm = &ondisk->elms[i];
2360 hammer_print_btree_elm(elm, ondisk->type, i);
2366 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2369 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
2370 kprintf("\tkey = %016llx\n", elm->base.key);
2371 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2372 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2373 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2374 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2375 kprintf("\tbtype = %02x (%c)\n",
2377 (elm->base.btype ? elm->base.btype : '?'));
2380 case HAMMER_BTREE_TYPE_INTERNAL:
2381 kprintf("\tsubtree_off = %016llx\n",
2382 elm->internal.subtree_offset);
2384 case HAMMER_BTREE_TYPE_RECORD:
2385 kprintf("\trec_offset = %016llx\n", elm->leaf.rec_offset);
2386 kprintf("\tdata_offset = %016llx\n", elm->leaf.data_offset);
2387 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2388 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);