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 1.76 2008/08/06 15:38:58 dillon Exp $
40 * HAMMER implements a modified B+Tree. In documentation this will
41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
42 * looks like a B+Tree (A B-Tree which stores its records only at the leafs
43 * of the tree), but adds two additional boundary elements which describe
44 * the left-most and right-most element a node is able to represent. In
45 * otherwords, we have boundary elements at the two ends of a B-Tree node
46 * instead of sub-tree pointers.
48 * A B-Tree internal node looks like this:
50 * B N N N N N N B <-- boundary and internal elements
51 * S S S S S S S <-- subtree pointers
53 * A B-Tree leaf node basically looks like this:
55 * L L L L L L L L <-- leaf elemenets
57 * The radix for an internal node is 1 less then a leaf but we get a
58 * number of significant benefits for our troubles.
60 * The big benefit to using a B-Tree containing boundary information
61 * is that it is possible to cache pointers into the middle of the tree
62 * and not have to start searches, insertions, OR deletions at the root
63 * node. In particular, searches are able to progress in a definitive
64 * direction from any point in the tree without revisting nodes. This
65 * greatly improves the efficiency of many operations, most especially
68 * B-Trees also make the stacking of trees fairly straightforward.
70 * INSERTIONS: A search performed with the intention of doing
71 * an insert will guarantee that the terminal leaf node is not full by
72 * splitting full nodes. Splits occur top-down during the dive down the
75 * DELETIONS: A deletion makes no attempt to proactively balance the
76 * tree and will recursively remove nodes that become empty. If a
77 * deadlock occurs a deletion may not be able to remove an empty leaf.
78 * Deletions never allow internal nodes to become empty (that would blow
85 static int btree_search(hammer_cursor_t cursor, int flags);
86 static int btree_split_internal(hammer_cursor_t cursor);
87 static int btree_split_leaf(hammer_cursor_t cursor);
88 static int btree_remove(hammer_cursor_t cursor);
89 static int btree_node_is_full(hammer_node_ondisk_t node);
90 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor,
91 hammer_tid_t mirror_tid);
92 static void hammer_make_separator(hammer_base_elm_t key1,
93 hammer_base_elm_t key2, hammer_base_elm_t dest);
94 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor);
97 * Iterate records after a search. The cursor is iterated forwards past
98 * the current record until a record matching the key-range requirements
99 * is found. ENOENT is returned if the iteration goes past the ending
102 * The iteration is inclusive of key_beg and can be inclusive or exclusive
103 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
105 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
106 * may be modified by B-Tree functions.
108 * cursor->key_beg may or may not be modified by this function during
109 * the iteration. XXX future - in case of an inverted lock we may have
110 * to reinitiate the lookup and set key_beg to properly pick up where we
113 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
116 hammer_btree_iterate(hammer_cursor_t cursor)
118 hammer_node_ondisk_t node;
119 hammer_btree_elm_t elm;
125 * Skip past the current record
127 node = cursor->node->ondisk;
130 if (cursor->index < node->count &&
131 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
136 * Loop until an element is found or we are done.
140 * We iterate up the tree and then index over one element
141 * while we are at the last element in the current node.
143 * If we are at the root of the filesystem, cursor_up
146 * XXX this could be optimized by storing the information in
147 * the parent reference.
149 * XXX we can lose the node lock temporarily, this could mess
152 ++hammer_stats_btree_iterations;
153 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
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);
173 * If we are reblocking we want to return internal
174 * nodes. Note that the internal node will be
175 * returned multiple times, on each upward recursion
176 * from its children. The caller selects which
177 * revisit it cares about (usually first or last only).
179 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
180 cursor->flags |= HAMMER_CURSOR_ATEDISK;
188 * Check internal or leaf element. Determine if the record
189 * at the cursor has gone beyond the end of our range.
191 * We recurse down through internal nodes.
193 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
194 elm = &node->elms[cursor->index];
196 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
197 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
198 if (hammer_debug_btree) {
199 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
200 cursor->node->node_offset,
202 elm[0].internal.base.obj_id,
203 elm[0].internal.base.rec_type,
204 elm[0].internal.base.key,
205 elm[0].internal.base.localization,
209 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
210 cursor->node->node_offset,
212 elm[1].internal.base.obj_id,
213 elm[1].internal.base.rec_type,
214 elm[1].internal.base.key,
215 elm[1].internal.base.localization,
224 if (r == 0 && (cursor->flags &
225 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
234 KKASSERT(elm->internal.subtree_offset != 0);
237 * If running the mirror filter see if we can skip
238 * one or more entire sub-trees. If we can we
239 * return the internal mode and the caller processes
240 * the skipped range (see mirror_read)
242 if (cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) {
243 if (elm->internal.mirror_tid <
244 cursor->cmirror->mirror_tid) {
245 hammer_cursor_mirror_filter(cursor);
250 error = hammer_cursor_down(cursor);
253 KKASSERT(cursor->index == 0);
254 /* reload stale pointer */
255 node = cursor->node->ondisk;
258 elm = &node->elms[cursor->index];
259 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
260 if (hammer_debug_btree) {
261 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
262 cursor->node->node_offset,
264 (elm[0].leaf.base.btype ?
265 elm[0].leaf.base.btype : '?'),
266 elm[0].leaf.base.obj_id,
267 elm[0].leaf.base.rec_type,
268 elm[0].leaf.base.key,
269 elm[0].leaf.base.localization,
279 * We support both end-inclusive and
280 * end-exclusive searches.
283 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
288 switch(elm->leaf.base.btype) {
289 case HAMMER_BTREE_TYPE_RECORD:
290 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
291 hammer_btree_chkts(cursor->asof, &elm->base)) {
305 * node pointer invalid after loop
311 if (hammer_debug_btree) {
312 int i = cursor->index;
313 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
314 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
316 elm->internal.base.obj_id,
317 elm->internal.base.rec_type,
318 elm->internal.base.key,
319 elm->internal.base.localization
328 * We hit an internal element that we could skip as part of a mirroring
329 * scan. Calculate the entire range being skipped.
331 * It is important to include any gaps between the parent's left_bound
332 * and the node's left_bound, and same goes for the right side.
335 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
337 struct hammer_cmirror *cmirror;
338 hammer_node_ondisk_t ondisk;
339 hammer_btree_elm_t elm;
341 ondisk = cursor->node->ondisk;
342 cmirror = cursor->cmirror;
345 * Calculate the skipped range
347 elm = &ondisk->elms[cursor->index];
348 if (cursor->index == 0)
349 cmirror->skip_beg = *cursor->left_bound;
351 cmirror->skip_beg = elm->internal.base;
352 while (cursor->index < ondisk->count) {
353 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
358 if (cursor->index == ondisk->count)
359 cmirror->skip_end = *cursor->right_bound;
361 cmirror->skip_end = elm->internal.base;
364 * clip the returned result.
366 if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
367 cmirror->skip_beg = cursor->key_beg;
368 if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
369 cmirror->skip_end = cursor->key_end;
373 * Iterate in the reverse direction. This is used by the pruning code to
374 * avoid overlapping records.
377 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
379 hammer_node_ondisk_t node;
380 hammer_btree_elm_t elm;
385 /* mirror filtering not supported for reverse iteration */
386 KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
389 * Skip past the current record. For various reasons the cursor
390 * may end up set to -1 or set to point at the end of the current
391 * node. These cases must be addressed.
393 node = cursor->node->ondisk;
396 if (cursor->index != -1 &&
397 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
400 if (cursor->index == cursor->node->ondisk->count)
404 * Loop until an element is found or we are done.
407 ++hammer_stats_btree_iterations;
408 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
411 * We iterate up the tree and then index over one element
412 * while we are at the last element in the current node.
414 if (cursor->index == -1) {
415 error = hammer_cursor_up(cursor);
417 cursor->index = 0; /* sanity */
420 /* reload stale pointer */
421 node = cursor->node->ondisk;
422 KKASSERT(cursor->index != node->count);
428 * Check internal or leaf element. Determine if the record
429 * at the cursor has gone beyond the end of our range.
431 * We recurse down through internal nodes.
433 KKASSERT(cursor->index != node->count);
434 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
435 elm = &node->elms[cursor->index];
436 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
437 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
438 if (hammer_debug_btree) {
439 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
440 cursor->node->node_offset,
442 elm[0].internal.base.obj_id,
443 elm[0].internal.base.rec_type,
444 elm[0].internal.base.key,
445 elm[0].internal.base.localization,
448 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
449 cursor->node->node_offset,
451 elm[1].internal.base.obj_id,
452 elm[1].internal.base.rec_type,
453 elm[1].internal.base.key,
454 elm[1].internal.base.localization,
468 KKASSERT(elm->internal.subtree_offset != 0);
470 error = hammer_cursor_down(cursor);
473 KKASSERT(cursor->index == 0);
474 /* reload stale pointer */
475 node = cursor->node->ondisk;
477 /* this can assign -1 if the leaf was empty */
478 cursor->index = node->count - 1;
481 elm = &node->elms[cursor->index];
482 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
483 if (hammer_debug_btree) {
484 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
485 cursor->node->node_offset,
487 (elm[0].leaf.base.btype ?
488 elm[0].leaf.base.btype : '?'),
489 elm[0].leaf.base.obj_id,
490 elm[0].leaf.base.rec_type,
491 elm[0].leaf.base.key,
492 elm[0].leaf.base.localization,
501 switch(elm->leaf.base.btype) {
502 case HAMMER_BTREE_TYPE_RECORD:
503 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
504 hammer_btree_chkts(cursor->asof, &elm->base)) {
518 * node pointer invalid after loop
524 if (hammer_debug_btree) {
525 int i = cursor->index;
526 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
527 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
529 elm->internal.base.obj_id,
530 elm->internal.base.rec_type,
531 elm->internal.base.key,
532 elm->internal.base.localization
541 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
542 * could not be found, EDEADLK if inserting and a retry is needed, and a
543 * fatal error otherwise. When retrying, the caller must terminate the
544 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
546 * The cursor is suitably positioned for a deletion on success, and suitably
547 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
550 * The cursor may begin anywhere, the search will traverse the tree in
551 * either direction to locate the requested element.
553 * Most of the logic implementing historical searches is handled here. We
554 * do an initial lookup with create_tid set to the asof TID. Due to the
555 * way records are laid out, a backwards iteration may be required if
556 * ENOENT is returned to locate the historical record. Here's the
559 * create_tid: 10 15 20
563 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
564 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
565 * not visible and thus causes ENOENT to be returned. We really need
566 * to check record 11 in LEAF1. If it also fails then the search fails
567 * (e.g. it might represent the range 11-16 and thus still not match our
568 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
569 * further iterations.
571 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
572 * and the cursor->create_check TID if an iteration might be needed.
573 * In the above example create_check would be set to 14.
576 hammer_btree_lookup(hammer_cursor_t cursor)
580 KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
581 cursor->trans->sync_lock_refs > 0);
582 ++hammer_stats_btree_lookups;
583 if (cursor->flags & HAMMER_CURSOR_ASOF) {
584 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
585 cursor->key_beg.create_tid = cursor->asof;
587 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
588 error = btree_search(cursor, 0);
589 if (error != ENOENT ||
590 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
593 * Stop if error other then ENOENT.
594 * Stop if ENOENT and not special case.
598 if (hammer_debug_btree) {
599 kprintf("CREATE_CHECK %016llx\n",
600 cursor->create_check);
602 cursor->key_beg.create_tid = cursor->create_check;
606 error = btree_search(cursor, 0);
609 error = hammer_btree_extract(cursor, cursor->flags);
614 * Execute the logic required to start an iteration. The first record
615 * located within the specified range is returned and iteration control
616 * flags are adjusted for successive hammer_btree_iterate() calls.
619 hammer_btree_first(hammer_cursor_t cursor)
623 error = hammer_btree_lookup(cursor);
624 if (error == ENOENT) {
625 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
626 error = hammer_btree_iterate(cursor);
628 cursor->flags |= HAMMER_CURSOR_ATEDISK;
633 * Similarly but for an iteration in the reverse direction.
635 * Set ATEDISK when iterating backwards to skip the current entry,
636 * which after an ENOENT lookup will be pointing beyond our end point.
639 hammer_btree_last(hammer_cursor_t cursor)
641 struct hammer_base_elm save;
644 save = cursor->key_beg;
645 cursor->key_beg = cursor->key_end;
646 error = hammer_btree_lookup(cursor);
647 cursor->key_beg = save;
648 if (error == ENOENT ||
649 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
650 cursor->flags |= HAMMER_CURSOR_ATEDISK;
651 error = hammer_btree_iterate_reverse(cursor);
653 cursor->flags |= HAMMER_CURSOR_ATEDISK;
658 * Extract the record and/or data associated with the cursor's current
659 * position. Any prior record or data stored in the cursor is replaced.
660 * The cursor must be positioned at a leaf node.
662 * NOTE: All extractions occur at the leaf of the B-Tree.
665 hammer_btree_extract(hammer_cursor_t cursor, int flags)
667 hammer_node_ondisk_t node;
668 hammer_btree_elm_t elm;
669 hammer_off_t data_off;
675 * The case where the data reference resolves to the same buffer
676 * as the record reference must be handled.
678 node = cursor->node->ondisk;
679 elm = &node->elms[cursor->index];
681 hmp = cursor->node->hmp;
684 * There is nothing to extract for an internal element.
686 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
690 * Only record types have data.
692 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
693 cursor->leaf = &elm->leaf;
695 if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
697 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
699 data_off = elm->leaf.data_offset;
700 data_len = elm->leaf.data_len;
707 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
708 cursor->data = hammer_bread_ext(hmp, data_off, data_len,
709 &error, &cursor->data_buffer);
710 if (hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
711 kprintf("CRC DATA @ %016llx/%d FAILED\n",
712 elm->leaf.data_offset, elm->leaf.data_len);
713 Debugger("CRC FAILED: DATA");
720 * Insert a leaf element into the B-Tree at the current cursor position.
721 * The cursor is positioned such that the element at and beyond the cursor
722 * are shifted to make room for the new record.
724 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
725 * flag set and that call must return ENOENT before this function can be
728 * The caller may depend on the cursor's exclusive lock after return to
729 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
731 * ENOSPC is returned if there is no room to insert a new record.
734 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
737 hammer_node_ondisk_t node;
742 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
744 ++hammer_stats_btree_inserts;
747 * Insert the element at the leaf node and update the count in the
748 * parent. It is possible for parent to be NULL, indicating that
749 * the filesystem's ROOT B-Tree node is a leaf itself, which is
750 * possible. The root inode can never be deleted so the leaf should
753 * Remember that the right-hand boundary is not included in the
756 hammer_modify_node_all(cursor->trans, cursor->node);
757 node = cursor->node->ondisk;
759 KKASSERT(elm->base.btype != 0);
760 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
761 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
762 if (i != node->count) {
763 bcopy(&node->elms[i], &node->elms[i+1],
764 (node->count - i) * sizeof(*elm));
766 node->elms[i].leaf = *elm;
768 hammer_cursor_inserted_element(cursor->node, i);
771 * Update the leaf node's aggregate mirror_tid for mirroring
774 if (node->mirror_tid < elm->base.delete_tid) {
775 node->mirror_tid = elm->base.delete_tid;
778 if (node->mirror_tid < elm->base.create_tid) {
779 node->mirror_tid = elm->base.create_tid;
782 hammer_modify_node_done(cursor->node);
785 * Debugging sanity checks.
787 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
788 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
790 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
792 if (i != node->count - 1)
793 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
799 * Delete a record from the B-Tree at the current cursor position.
800 * The cursor is positioned such that the current element is the one
803 * On return the cursor will be positioned after the deleted element and
804 * MAY point to an internal node. It will be suitable for the continuation
805 * of an iteration but not for an insertion or deletion.
807 * Deletions will attempt to partially rebalance the B-Tree in an upward
808 * direction, but will terminate rather then deadlock. Empty internal nodes
809 * are never allowed by a deletion which deadlocks may end up giving us an
810 * empty leaf. The pruner will clean up and rebalance the tree.
812 * This function can return EDEADLK, requiring the caller to retry the
813 * operation after clearing the deadlock.
816 hammer_btree_delete(hammer_cursor_t cursor)
818 hammer_node_ondisk_t ondisk;
820 hammer_node_t parent;
824 KKASSERT (cursor->trans->sync_lock_refs > 0);
825 if ((error = hammer_cursor_upgrade(cursor)) != 0)
827 ++hammer_stats_btree_deletes;
830 * Delete the element from the leaf node.
832 * Remember that leaf nodes do not have boundaries.
835 ondisk = node->ondisk;
838 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
839 KKASSERT(i >= 0 && i < ondisk->count);
840 hammer_modify_node_all(cursor->trans, node);
841 if (i + 1 != ondisk->count) {
842 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
843 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
846 hammer_modify_node_done(node);
847 hammer_cursor_deleted_element(node, i);
850 * Validate local parent
852 if (ondisk->parent) {
853 parent = cursor->parent;
855 KKASSERT(parent != NULL);
856 KKASSERT(parent->node_offset == ondisk->parent);
860 * If the leaf becomes empty it must be detached from the parent,
861 * potentially recursing through to the filesystem root.
863 * This may reposition the cursor at one of the parent's of the
866 * Ignore deadlock errors, that simply means that btree_remove
867 * was unable to recurse and had to leave us with an empty leaf.
869 KKASSERT(cursor->index <= ondisk->count);
870 if (ondisk->count == 0) {
871 error = btree_remove(cursor);
872 if (error == EDEADLK)
877 KKASSERT(cursor->parent == NULL ||
878 cursor->parent_index < cursor->parent->ondisk->count);
883 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
885 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
887 * The search can begin ANYWHERE in the B-Tree. As a first step the search
888 * iterates up the tree as necessary to properly position itself prior to
889 * actually doing the sarch.
891 * INSERTIONS: The search will split full nodes and leaves on its way down
892 * and guarentee that the leaf it ends up on is not full. If we run out
893 * of space the search continues to the leaf (to position the cursor for
894 * the spike), but ENOSPC is returned.
896 * The search is only guarenteed to end up on a leaf if an error code of 0
897 * is returned, or if inserting and an error code of ENOENT is returned.
898 * Otherwise it can stop at an internal node. On success a search returns
901 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
902 * filesystem, and it is not simple code. Please note the following facts:
904 * - Internal node recursions have a boundary on the left AND right. The
905 * right boundary is non-inclusive. The create_tid is a generic part
906 * of the key for internal nodes.
908 * - Leaf nodes contain terminal elements only now.
910 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
911 * historical search. ASOF and INSERT are mutually exclusive. When
912 * doing an as-of lookup btree_search() checks for a right-edge boundary
913 * case. If while recursing down the left-edge differs from the key
914 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
915 * with cursor->create_check. This is used by btree_lookup() to iterate.
916 * The iteration backwards because as-of searches can wind up going
917 * down the wrong branch of the B-Tree.
921 btree_search(hammer_cursor_t cursor, int flags)
923 hammer_node_ondisk_t node;
924 hammer_btree_elm_t elm;
931 flags |= cursor->flags;
932 ++hammer_stats_btree_searches;
934 if (hammer_debug_btree) {
935 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
936 cursor->node->node_offset,
938 cursor->key_beg.obj_id,
939 cursor->key_beg.rec_type,
941 cursor->key_beg.create_tid,
942 cursor->key_beg.localization,
946 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
947 cursor->parent->node_offset, cursor->parent_index,
948 cursor->left_bound->obj_id,
949 cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
950 cursor->right_bound->obj_id,
951 cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
953 &cursor->parent->ondisk->elms[cursor->parent_index],
955 &cursor->parent->ondisk->elms[cursor->parent_index+1]
960 * Move our cursor up the tree until we find a node whos range covers
961 * the key we are trying to locate.
963 * The left bound is inclusive, the right bound is non-inclusive.
964 * It is ok to cursor up too far.
967 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
968 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
971 KKASSERT(cursor->parent);
972 ++hammer_stats_btree_iterations;
973 error = hammer_cursor_up(cursor);
979 * The delete-checks below are based on node, not parent. Set the
980 * initial delete-check based on the parent.
983 KKASSERT(cursor->left_bound->create_tid != 1);
984 cursor->create_check = cursor->left_bound->create_tid - 1;
985 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
989 * We better have ended up with a node somewhere.
991 KKASSERT(cursor->node != NULL);
994 * If we are inserting we can't start at a full node if the parent
995 * is also full (because there is no way to split the node),
996 * continue running up the tree until the requirement is satisfied
997 * or we hit the root of the filesystem.
999 * (If inserting we aren't doing an as-of search so we don't have
1000 * to worry about create_check).
1002 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1003 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1004 if (btree_node_is_full(cursor->node->ondisk) == 0)
1007 if (btree_node_is_full(cursor->node->ondisk) ==0)
1010 if (cursor->node->ondisk->parent == 0 ||
1011 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1014 ++hammer_stats_btree_iterations;
1015 error = hammer_cursor_up(cursor);
1016 /* node may have become stale */
1022 * Push down through internal nodes to locate the requested key.
1024 node = cursor->node->ondisk;
1025 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1027 * Scan the node to find the subtree index to push down into.
1028 * We go one-past, then back-up.
1030 * We must proactively remove deleted elements which may
1031 * have been left over from a deadlocked btree_remove().
1033 * The left and right boundaries are included in the loop
1034 * in order to detect edge cases.
1036 * If the separator only differs by create_tid (r == 1)
1037 * and we are doing an as-of search, we may end up going
1038 * down a branch to the left of the one containing the
1039 * desired key. This requires numerous special cases.
1041 ++hammer_stats_btree_iterations;
1042 if (hammer_debug_btree) {
1043 kprintf("SEARCH-I %016llx count=%d\n",
1044 cursor->node->node_offset,
1049 * Try to shortcut the search before dropping into the
1050 * linear loop. Locate the first node where r <= 1.
1052 i = hammer_btree_search_node(&cursor->key_beg, node);
1053 while (i <= node->count) {
1054 ++hammer_stats_btree_elements;
1055 elm = &node->elms[i];
1056 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1057 if (hammer_debug_btree > 2) {
1058 kprintf(" IELM %p %d r=%d\n",
1059 &node->elms[i], i, r);
1064 KKASSERT(elm->base.create_tid != 1);
1065 cursor->create_check = elm->base.create_tid - 1;
1066 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1070 if (hammer_debug_btree) {
1071 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1076 * These cases occur when the parent's idea of the boundary
1077 * is wider then the child's idea of the boundary, and
1078 * require special handling. If not inserting we can
1079 * terminate the search early for these cases but the
1080 * child's boundaries cannot be unconditionally modified.
1084 * If i == 0 the search terminated to the LEFT of the
1085 * left_boundary but to the RIGHT of the parent's left
1090 elm = &node->elms[0];
1093 * If we aren't inserting we can stop here.
1095 if ((flags & (HAMMER_CURSOR_INSERT |
1096 HAMMER_CURSOR_PRUNING)) == 0) {
1102 * Correct a left-hand boundary mismatch.
1104 * We can only do this if we can upgrade the lock,
1105 * and synchronized as a background cursor (i.e.
1106 * inserting or pruning).
1108 * WARNING: We can only do this if inserting, i.e.
1109 * we are running on the backend.
1111 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1113 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1114 hammer_modify_node_field(cursor->trans, cursor->node,
1116 save = node->elms[0].base.btype;
1117 node->elms[0].base = *cursor->left_bound;
1118 node->elms[0].base.btype = save;
1119 hammer_modify_node_done(cursor->node);
1120 } else if (i == node->count + 1) {
1122 * If i == node->count + 1 the search terminated to
1123 * the RIGHT of the right boundary but to the LEFT
1124 * of the parent's right boundary. If we aren't
1125 * inserting we can stop here.
1127 * Note that the last element in this case is
1128 * elms[i-2] prior to adjustments to 'i'.
1131 if ((flags & (HAMMER_CURSOR_INSERT |
1132 HAMMER_CURSOR_PRUNING)) == 0) {
1138 * Correct a right-hand boundary mismatch.
1139 * (actual push-down record is i-2 prior to
1140 * adjustments to i).
1142 * We can only do this if we can upgrade the lock,
1143 * and synchronized as a background cursor (i.e.
1144 * inserting or pruning).
1146 * WARNING: We can only do this if inserting, i.e.
1147 * we are running on the backend.
1149 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1151 elm = &node->elms[i];
1152 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1153 hammer_modify_node(cursor->trans, cursor->node,
1154 &elm->base, sizeof(elm->base));
1155 elm->base = *cursor->right_bound;
1156 hammer_modify_node_done(cursor->node);
1160 * The push-down index is now i - 1. If we had
1161 * terminated on the right boundary this will point
1162 * us at the last element.
1167 elm = &node->elms[i];
1169 if (hammer_debug_btree) {
1170 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1171 "key=%016llx cre=%016llx lo=%02x\n",
1172 cursor->node->node_offset,
1174 elm->internal.base.obj_id,
1175 elm->internal.base.rec_type,
1176 elm->internal.base.key,
1177 elm->internal.base.create_tid,
1178 elm->internal.base.localization
1183 * We better have a valid subtree offset.
1185 KKASSERT(elm->internal.subtree_offset != 0);
1188 * Handle insertion and deletion requirements.
1190 * If inserting split full nodes. The split code will
1191 * adjust cursor->node and cursor->index if the current
1192 * index winds up in the new node.
1194 * If inserting and a left or right edge case was detected,
1195 * we cannot correct the left or right boundary and must
1196 * prepend and append an empty leaf node in order to make
1197 * the boundary correction.
1199 * If we run out of space we set enospc and continue on
1200 * to a leaf to provide the spike code with a good point
1203 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1204 if (btree_node_is_full(node)) {
1205 error = btree_split_internal(cursor);
1207 if (error != ENOSPC)
1212 * reload stale pointers
1215 node = cursor->node->ondisk;
1220 * Push down (push into new node, existing node becomes
1221 * the parent) and continue the search.
1223 error = hammer_cursor_down(cursor);
1224 /* node may have become stale */
1227 node = cursor->node->ondisk;
1231 * We are at a leaf, do a linear search of the key array.
1233 * On success the index is set to the matching element and 0
1236 * On failure the index is set to the insertion point and ENOENT
1239 * Boundaries are not stored in leaf nodes, so the index can wind
1240 * up to the left of element 0 (index == 0) or past the end of
1241 * the array (index == node->count). It is also possible that the
1242 * leaf might be empty.
1244 ++hammer_stats_btree_iterations;
1245 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1246 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1247 if (hammer_debug_btree) {
1248 kprintf("SEARCH-L %016llx count=%d\n",
1249 cursor->node->node_offset,
1254 * Try to shortcut the search before dropping into the
1255 * linear loop. Locate the first node where r <= 1.
1257 i = hammer_btree_search_node(&cursor->key_beg, node);
1258 while (i < node->count) {
1259 ++hammer_stats_btree_elements;
1260 elm = &node->elms[i];
1262 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1264 if (hammer_debug_btree > 1)
1265 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1268 * We are at a record element. Stop if we've flipped past
1269 * key_beg, not counting the create_tid test. Allow the
1270 * r == 1 case (key_beg > element but differs only by its
1271 * create_tid) to fall through to the AS-OF check.
1273 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1283 * Check our as-of timestamp against the element.
1285 if (flags & HAMMER_CURSOR_ASOF) {
1286 if (hammer_btree_chkts(cursor->asof,
1287 &node->elms[i].base) != 0) {
1293 if (r > 0) { /* can only be +1 */
1301 if (hammer_debug_btree) {
1302 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1303 cursor->node->node_offset, i);
1309 * The search of the leaf node failed. i is the insertion point.
1312 if (hammer_debug_btree) {
1313 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1314 cursor->node->node_offset, i);
1318 * No exact match was found, i is now at the insertion point.
1320 * If inserting split a full leaf before returning. This
1321 * may have the side effect of adjusting cursor->node and
1325 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1326 btree_node_is_full(node)) {
1327 error = btree_split_leaf(cursor);
1329 if (error != ENOSPC)
1334 * reload stale pointers
1338 node = &cursor->node->internal;
1343 * We reached a leaf but did not find the key we were looking for.
1344 * If this is an insert we will be properly positioned for an insert
1345 * (ENOENT) or spike (ENOSPC) operation.
1347 error = enospc ? ENOSPC : ENOENT;
1353 * Heuristical search for the first element whos comparison is <= 1. May
1354 * return an index whos compare result is > 1 but may only return an index
1355 * whos compare result is <= 1 if it is the first element with that result.
1358 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1366 * Don't bother if the node does not have very many elements
1371 i = b + (s - b) / 2;
1372 ++hammer_stats_btree_elements;
1373 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1384 /************************************************************************
1385 * SPLITTING AND MERGING *
1386 ************************************************************************
1388 * These routines do all the dirty work required to split and merge nodes.
1392 * Split an internal node into two nodes and move the separator at the split
1393 * point to the parent.
1395 * (cursor->node, cursor->index) indicates the element the caller intends
1396 * to push into. We will adjust node and index if that element winds
1397 * up in the split node.
1399 * If we are at the root of the filesystem a new root must be created with
1400 * two elements, one pointing to the original root and one pointing to the
1401 * newly allocated split node.
1405 btree_split_internal(hammer_cursor_t cursor)
1407 hammer_node_ondisk_t ondisk;
1409 hammer_node_t parent;
1410 hammer_node_t new_node;
1411 hammer_btree_elm_t elm;
1412 hammer_btree_elm_t parent_elm;
1413 struct hammer_node_lock lockroot;
1414 hammer_mount_t hmp = cursor->trans->hmp;
1420 const int esize = sizeof(*elm);
1422 hammer_node_lock_init(&lockroot, cursor->node);
1423 error = hammer_btree_lock_children(cursor, 1, &lockroot);
1426 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1428 ++hammer_stats_btree_splits;
1431 * We are splitting but elms[split] will be promoted to the parent,
1432 * leaving the right hand node with one less element. If the
1433 * insertion point will be on the left-hand side adjust the split
1434 * point to give the right hand side one additional node.
1436 node = cursor->node;
1437 ondisk = node->ondisk;
1438 split = (ondisk->count + 1) / 2;
1439 if (cursor->index <= split)
1443 * If we are at the root of the filesystem, create a new root node
1444 * with 1 element and split normally. Avoid making major
1445 * modifications until we know the whole operation will work.
1447 if (ondisk->parent == 0) {
1448 parent = hammer_alloc_btree(cursor->trans, &error);
1451 hammer_lock_ex(&parent->lock);
1452 hammer_modify_node_noundo(cursor->trans, parent);
1453 ondisk = parent->ondisk;
1456 ondisk->mirror_tid = node->ondisk->mirror_tid;
1457 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1458 ondisk->elms[0].base = hmp->root_btree_beg;
1459 ondisk->elms[0].base.btype = node->ondisk->type;
1460 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1461 ondisk->elms[1].base = hmp->root_btree_end;
1462 hammer_modify_node_done(parent);
1463 /* ondisk->elms[1].base.btype - not used */
1465 parent_index = 0; /* index of current node in parent */
1468 parent = cursor->parent;
1469 parent_index = cursor->parent_index;
1473 * Split node into new_node at the split point.
1475 * B O O O P N N B <-- P = node->elms[split]
1476 * 0 1 2 3 4 5 6 <-- subtree indices
1481 * B O O O B B N N B <--- inner boundary points are 'P'
1485 new_node = hammer_alloc_btree(cursor->trans, &error);
1486 if (new_node == NULL) {
1488 hammer_unlock(&parent->lock);
1489 hammer_delete_node(cursor->trans, parent);
1490 hammer_rel_node(parent);
1494 hammer_lock_ex(&new_node->lock);
1497 * Create the new node. P becomes the left-hand boundary in the
1498 * new node. Copy the right-hand boundary as well.
1500 * elm is the new separator.
1502 hammer_modify_node_noundo(cursor->trans, new_node);
1503 hammer_modify_node_all(cursor->trans, node);
1504 ondisk = node->ondisk;
1505 elm = &ondisk->elms[split];
1506 bcopy(elm, &new_node->ondisk->elms[0],
1507 (ondisk->count - split + 1) * esize);
1508 new_node->ondisk->count = ondisk->count - split;
1509 new_node->ondisk->parent = parent->node_offset;
1510 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1511 new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1512 KKASSERT(ondisk->type == new_node->ondisk->type);
1513 hammer_cursor_split_node(node, new_node, split);
1516 * Cleanup the original node. Elm (P) becomes the new boundary,
1517 * its subtree_offset was moved to the new node. If we had created
1518 * a new root its parent pointer may have changed.
1520 elm->internal.subtree_offset = 0;
1521 ondisk->count = split;
1524 * Insert the separator into the parent, fixup the parent's
1525 * reference to the original node, and reference the new node.
1526 * The separator is P.
1528 * Remember that base.count does not include the right-hand boundary.
1530 hammer_modify_node_all(cursor->trans, parent);
1531 ondisk = parent->ondisk;
1532 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1533 parent_elm = &ondisk->elms[parent_index+1];
1534 bcopy(parent_elm, parent_elm + 1,
1535 (ondisk->count - parent_index) * esize);
1536 parent_elm->internal.base = elm->base; /* separator P */
1537 parent_elm->internal.base.btype = new_node->ondisk->type;
1538 parent_elm->internal.subtree_offset = new_node->node_offset;
1539 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1541 hammer_modify_node_done(parent);
1542 hammer_cursor_inserted_element(parent, parent_index + 1);
1545 * The children of new_node need their parent pointer set to new_node.
1546 * The children have already been locked by
1547 * hammer_btree_lock_children().
1549 for (i = 0; i < new_node->ondisk->count; ++i) {
1550 elm = &new_node->ondisk->elms[i];
1551 error = btree_set_parent(cursor->trans, new_node, elm);
1553 panic("btree_split_internal: btree-fixup problem");
1556 hammer_modify_node_done(new_node);
1559 * The filesystem's root B-Tree pointer may have to be updated.
1562 hammer_volume_t volume;
1564 volume = hammer_get_root_volume(hmp, &error);
1565 KKASSERT(error == 0);
1567 hammer_modify_volume_field(cursor->trans, volume,
1569 volume->ondisk->vol0_btree_root = parent->node_offset;
1570 hammer_modify_volume_done(volume);
1571 node->ondisk->parent = parent->node_offset;
1572 if (cursor->parent) {
1573 hammer_unlock(&cursor->parent->lock);
1574 hammer_rel_node(cursor->parent);
1576 cursor->parent = parent; /* lock'd and ref'd */
1577 hammer_rel_volume(volume, 0);
1579 hammer_modify_node_done(node);
1582 * Ok, now adjust the cursor depending on which element the original
1583 * index was pointing at. If we are >= the split point the push node
1584 * is now in the new node.
1586 * NOTE: If we are at the split point itself we cannot stay with the
1587 * original node because the push index will point at the right-hand
1588 * boundary, which is illegal.
1590 * NOTE: The cursor's parent or parent_index must be adjusted for
1591 * the case where a new parent (new root) was created, and the case
1592 * where the cursor is now pointing at the split node.
1594 if (cursor->index >= split) {
1595 cursor->parent_index = parent_index + 1;
1596 cursor->index -= split;
1597 hammer_unlock(&cursor->node->lock);
1598 hammer_rel_node(cursor->node);
1599 cursor->node = new_node; /* locked and ref'd */
1601 cursor->parent_index = parent_index;
1602 hammer_unlock(&new_node->lock);
1603 hammer_rel_node(new_node);
1607 * Fixup left and right bounds
1609 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1610 cursor->left_bound = &parent_elm[0].internal.base;
1611 cursor->right_bound = &parent_elm[1].internal.base;
1612 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1613 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1614 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1615 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1618 hammer_btree_unlock_children(cursor, &lockroot);
1619 hammer_cursor_downgrade(cursor);
1624 * Same as the above, but splits a full leaf node.
1630 btree_split_leaf(hammer_cursor_t cursor)
1632 hammer_node_ondisk_t ondisk;
1633 hammer_node_t parent;
1636 hammer_node_t new_leaf;
1637 hammer_btree_elm_t elm;
1638 hammer_btree_elm_t parent_elm;
1639 hammer_base_elm_t mid_boundary;
1644 const size_t esize = sizeof(*elm);
1646 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1648 ++hammer_stats_btree_splits;
1650 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1651 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1652 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1653 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1656 * Calculate the split point. If the insertion point will be on
1657 * the left-hand side adjust the split point to give the right
1658 * hand side one additional node.
1660 * Spikes are made up of two leaf elements which cannot be
1663 leaf = cursor->node;
1664 ondisk = leaf->ondisk;
1665 split = (ondisk->count + 1) / 2;
1666 if (cursor->index <= split)
1671 elm = &ondisk->elms[split];
1673 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1674 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1675 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1676 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1679 * If we are at the root of the tree, create a new root node with
1680 * 1 element and split normally. Avoid making major modifications
1681 * until we know the whole operation will work.
1683 if (ondisk->parent == 0) {
1684 parent = hammer_alloc_btree(cursor->trans, &error);
1687 hammer_lock_ex(&parent->lock);
1688 hammer_modify_node_noundo(cursor->trans, parent);
1689 ondisk = parent->ondisk;
1692 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1693 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1694 ondisk->elms[0].base = hmp->root_btree_beg;
1695 ondisk->elms[0].base.btype = leaf->ondisk->type;
1696 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1697 ondisk->elms[1].base = hmp->root_btree_end;
1698 /* ondisk->elms[1].base.btype = not used */
1699 hammer_modify_node_done(parent);
1701 parent_index = 0; /* insertion point in parent */
1704 parent = cursor->parent;
1705 parent_index = cursor->parent_index;
1709 * Split leaf into new_leaf at the split point. Select a separator
1710 * value in-between the two leafs but with a bent towards the right
1711 * leaf since comparisons use an 'elm >= separator' inequality.
1720 new_leaf = hammer_alloc_btree(cursor->trans, &error);
1721 if (new_leaf == NULL) {
1723 hammer_unlock(&parent->lock);
1724 hammer_delete_node(cursor->trans, parent);
1725 hammer_rel_node(parent);
1729 hammer_lock_ex(&new_leaf->lock);
1732 * Create the new node and copy the leaf elements from the split
1733 * point on to the new node.
1735 hammer_modify_node_all(cursor->trans, leaf);
1736 hammer_modify_node_noundo(cursor->trans, new_leaf);
1737 ondisk = leaf->ondisk;
1738 elm = &ondisk->elms[split];
1739 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1740 new_leaf->ondisk->count = ondisk->count - split;
1741 new_leaf->ondisk->parent = parent->node_offset;
1742 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1743 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1744 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1745 hammer_modify_node_done(new_leaf);
1746 hammer_cursor_split_node(leaf, new_leaf, split);
1749 * Cleanup the original node. Because this is a leaf node and
1750 * leaf nodes do not have a right-hand boundary, there
1751 * aren't any special edge cases to clean up. We just fixup the
1754 ondisk->count = split;
1757 * Insert the separator into the parent, fixup the parent's
1758 * reference to the original node, and reference the new node.
1759 * The separator is P.
1761 * Remember that base.count does not include the right-hand boundary.
1762 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1764 hammer_modify_node_all(cursor->trans, parent);
1765 ondisk = parent->ondisk;
1766 KKASSERT(split != 0);
1767 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1768 parent_elm = &ondisk->elms[parent_index+1];
1769 bcopy(parent_elm, parent_elm + 1,
1770 (ondisk->count - parent_index) * esize);
1772 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1773 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1774 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1775 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1776 mid_boundary = &parent_elm->base;
1778 hammer_modify_node_done(parent);
1779 hammer_cursor_inserted_element(parent, parent_index + 1);
1782 * The filesystem's root B-Tree pointer may have to be updated.
1785 hammer_volume_t volume;
1787 volume = hammer_get_root_volume(hmp, &error);
1788 KKASSERT(error == 0);
1790 hammer_modify_volume_field(cursor->trans, volume,
1792 volume->ondisk->vol0_btree_root = parent->node_offset;
1793 hammer_modify_volume_done(volume);
1794 leaf->ondisk->parent = parent->node_offset;
1795 if (cursor->parent) {
1796 hammer_unlock(&cursor->parent->lock);
1797 hammer_rel_node(cursor->parent);
1799 cursor->parent = parent; /* lock'd and ref'd */
1800 hammer_rel_volume(volume, 0);
1802 hammer_modify_node_done(leaf);
1805 * Ok, now adjust the cursor depending on which element the original
1806 * index was pointing at. If we are >= the split point the push node
1807 * is now in the new node.
1809 * NOTE: If we are at the split point itself we need to select the
1810 * old or new node based on where key_beg's insertion point will be.
1811 * If we pick the wrong side the inserted element will wind up in
1812 * the wrong leaf node and outside that node's bounds.
1814 if (cursor->index > split ||
1815 (cursor->index == split &&
1816 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1817 cursor->parent_index = parent_index + 1;
1818 cursor->index -= split;
1819 hammer_unlock(&cursor->node->lock);
1820 hammer_rel_node(cursor->node);
1821 cursor->node = new_leaf;
1823 cursor->parent_index = parent_index;
1824 hammer_unlock(&new_leaf->lock);
1825 hammer_rel_node(new_leaf);
1829 * Fixup left and right bounds
1831 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1832 cursor->left_bound = &parent_elm[0].internal.base;
1833 cursor->right_bound = &parent_elm[1].internal.base;
1836 * Assert that the bounds are correct.
1838 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1839 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1840 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1841 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1842 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1843 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1846 hammer_cursor_downgrade(cursor);
1853 * Recursively correct the right-hand boundary's create_tid to (tid) as
1854 * long as the rest of the key matches. We have to recurse upward in
1855 * the tree as well as down the left side of each parent's right node.
1857 * Return EDEADLK if we were only partially successful, forcing the caller
1858 * to try again. The original cursor is not modified. This routine can
1859 * also fail with EDEADLK if it is forced to throw away a portion of its
1862 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1865 TAILQ_ENTRY(hammer_rhb) entry;
1870 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1873 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1875 struct hammer_mount *hmp;
1876 struct hammer_rhb_list rhb_list;
1877 hammer_base_elm_t elm;
1878 hammer_node_t orig_node;
1879 struct hammer_rhb *rhb;
1883 TAILQ_INIT(&rhb_list);
1884 hmp = cursor->trans->hmp;
1887 * Save our position so we can restore it on return. This also
1888 * gives us a stable 'elm'.
1890 orig_node = cursor->node;
1891 hammer_ref_node(orig_node);
1892 hammer_lock_sh(&orig_node->lock);
1893 orig_index = cursor->index;
1894 elm = &orig_node->ondisk->elms[orig_index].base;
1897 * Now build a list of parents going up, allocating a rhb
1898 * structure for each one.
1900 while (cursor->parent) {
1902 * Stop if we no longer have any right-bounds to fix up
1904 if (elm->obj_id != cursor->right_bound->obj_id ||
1905 elm->rec_type != cursor->right_bound->rec_type ||
1906 elm->key != cursor->right_bound->key) {
1911 * Stop if the right-hand bound's create_tid does not
1912 * need to be corrected.
1914 if (cursor->right_bound->create_tid >= tid)
1917 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
1918 rhb->node = cursor->parent;
1919 rhb->index = cursor->parent_index;
1920 hammer_ref_node(rhb->node);
1921 hammer_lock_sh(&rhb->node->lock);
1922 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1924 hammer_cursor_up(cursor);
1928 * now safely adjust the right hand bound for each rhb. This may
1929 * also require taking the right side of the tree and iterating down
1933 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1934 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1937 TAILQ_REMOVE(&rhb_list, rhb, entry);
1938 hammer_unlock(&rhb->node->lock);
1939 hammer_rel_node(rhb->node);
1940 kfree(rhb, hmp->m_misc);
1942 switch (cursor->node->ondisk->type) {
1943 case HAMMER_BTREE_TYPE_INTERNAL:
1945 * Right-boundary for parent at internal node
1946 * is one element to the right of the element whos
1947 * right boundary needs adjusting. We must then
1948 * traverse down the left side correcting any left
1949 * bounds (which may now be too far to the left).
1952 error = hammer_btree_correct_lhb(cursor, tid);
1955 panic("hammer_btree_correct_rhb(): Bad node type");
1964 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1965 TAILQ_REMOVE(&rhb_list, rhb, entry);
1966 hammer_unlock(&rhb->node->lock);
1967 hammer_rel_node(rhb->node);
1968 kfree(rhb, hmp->m_misc);
1970 error = hammer_cursor_seek(cursor, orig_node, orig_index);
1971 hammer_unlock(&orig_node->lock);
1972 hammer_rel_node(orig_node);
1977 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1978 * bound going downward starting at the current cursor position.
1980 * This function does not restore the cursor after use.
1983 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1985 struct hammer_rhb_list rhb_list;
1986 hammer_base_elm_t elm;
1987 hammer_base_elm_t cmp;
1988 struct hammer_rhb *rhb;
1989 struct hammer_mount *hmp;
1992 TAILQ_INIT(&rhb_list);
1993 hmp = cursor->trans->hmp;
1995 cmp = &cursor->node->ondisk->elms[cursor->index].base;
1998 * Record the node and traverse down the left-hand side for all
1999 * matching records needing a boundary correction.
2003 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2004 rhb->node = cursor->node;
2005 rhb->index = cursor->index;
2006 hammer_ref_node(rhb->node);
2007 hammer_lock_sh(&rhb->node->lock);
2008 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2010 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2012 * Nothing to traverse down if we are at the right
2013 * boundary of an internal node.
2015 if (cursor->index == cursor->node->ondisk->count)
2018 elm = &cursor->node->ondisk->elms[cursor->index].base;
2019 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2021 panic("Illegal leaf record type %02x", elm->btype);
2023 error = hammer_cursor_down(cursor);
2027 elm = &cursor->node->ondisk->elms[cursor->index].base;
2028 if (elm->obj_id != cmp->obj_id ||
2029 elm->rec_type != cmp->rec_type ||
2030 elm->key != cmp->key) {
2033 if (elm->create_tid >= tid)
2039 * Now we can safely adjust the left-hand boundary from the bottom-up.
2040 * The last element we remove from the list is the caller's right hand
2041 * boundary, which must also be adjusted.
2043 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2044 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2047 TAILQ_REMOVE(&rhb_list, rhb, entry);
2048 hammer_unlock(&rhb->node->lock);
2049 hammer_rel_node(rhb->node);
2050 kfree(rhb, hmp->m_misc);
2052 elm = &cursor->node->ondisk->elms[cursor->index].base;
2053 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2054 hammer_modify_node(cursor->trans, cursor->node,
2056 sizeof(elm->create_tid));
2057 elm->create_tid = tid;
2058 hammer_modify_node_done(cursor->node);
2060 panic("hammer_btree_correct_lhb(): Bad element type");
2067 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2068 TAILQ_REMOVE(&rhb_list, rhb, entry);
2069 hammer_unlock(&rhb->node->lock);
2070 hammer_rel_node(rhb->node);
2071 kfree(rhb, hmp->m_misc);
2079 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2080 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2081 * the operation due to a deadlock, or some other error.
2083 * This routine is initially called with an empty leaf and may be
2084 * recursively called with single-element internal nodes.
2086 * It should also be noted that when removing empty leaves we must be sure
2087 * to test and update mirror_tid because another thread may have deadlocked
2088 * against us (or someone) trying to propagate it up and cannot retry once
2089 * the node has been deleted.
2091 * On return the cursor may end up pointing to an internal node, suitable
2092 * for further iteration but not for an immediate insertion or deletion.
2095 btree_remove(hammer_cursor_t cursor)
2097 hammer_node_ondisk_t ondisk;
2098 hammer_btree_elm_t elm;
2100 hammer_node_t parent;
2101 const int esize = sizeof(*elm);
2104 node = cursor->node;
2107 * When deleting the root of the filesystem convert it to
2108 * an empty leaf node. Internal nodes cannot be empty.
2110 ondisk = node->ondisk;
2111 if (ondisk->parent == 0) {
2112 KKASSERT(cursor->parent == NULL);
2113 hammer_modify_node_all(cursor->trans, node);
2114 KKASSERT(ondisk == node->ondisk);
2115 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2117 hammer_modify_node_done(node);
2122 parent = cursor->parent;
2123 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2126 * Attempt to remove the parent's reference to the child. If the
2127 * parent would become empty we have to recurse. If we fail we
2128 * leave the parent pointing to an empty leaf node.
2130 * We have to recurse successfully before we can delete the internal
2131 * node as it is illegal to have empty internal nodes. Even though
2132 * the operation may be aborted we must still fixup any unlocked
2133 * cursors as if we had deleted the element prior to recursing
2134 * (by calling hammer_cursor_deleted_element()) so those cursors
2135 * are properly forced up the chain by the recursion.
2137 if (parent->ondisk->count == 1) {
2139 * This special cursor_up_locked() call leaves the original
2140 * node exclusively locked and referenced, leaves the
2141 * original parent locked (as the new node), and locks the
2142 * new parent. It can return EDEADLK.
2144 error = hammer_cursor_up_locked(cursor);
2146 hammer_cursor_deleted_element(cursor->node, 0);
2147 error = btree_remove(cursor);
2149 hammer_modify_node_all(cursor->trans, node);
2150 ondisk = node->ondisk;
2151 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2153 hammer_modify_node_done(node);
2154 hammer_flush_node(node);
2155 hammer_delete_node(cursor->trans, node);
2157 kprintf("Warning: BTREE_REMOVE: Defering "
2158 "parent removal1 @ %016llx, skipping\n",
2161 hammer_unlock(&node->lock);
2162 hammer_rel_node(node);
2164 kprintf("Warning: BTREE_REMOVE: Defering parent "
2165 "removal2 @ %016llx, skipping\n",
2169 KKASSERT(parent->ondisk->count > 1);
2171 hammer_modify_node_all(cursor->trans, parent);
2172 ondisk = parent->ondisk;
2173 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2175 elm = &ondisk->elms[cursor->parent_index];
2176 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2177 KKASSERT(ondisk->count > 0);
2180 * We must retain the highest mirror_tid. The deleted
2181 * range is now encompassed by the element to the left.
2182 * If we are already at the left edge the new left edge
2183 * inherits mirror_tid.
2185 * Note that bounds of the parent to our parent may create
2186 * a gap to the left of our left-most node or to the right
2187 * of our right-most node. The gap is silently included
2188 * in the mirror_tid's area of effect from the point of view
2191 if (cursor->parent_index) {
2192 if (elm[-1].internal.mirror_tid <
2193 elm[0].internal.mirror_tid) {
2194 elm[-1].internal.mirror_tid =
2195 elm[0].internal.mirror_tid;
2198 if (elm[1].internal.mirror_tid <
2199 elm[0].internal.mirror_tid) {
2200 elm[1].internal.mirror_tid =
2201 elm[0].internal.mirror_tid;
2206 * Delete the subtree reference in the parent
2208 bcopy(&elm[1], &elm[0],
2209 (ondisk->count - cursor->parent_index) * esize);
2211 hammer_modify_node_done(parent);
2212 hammer_cursor_deleted_element(parent, cursor->parent_index);
2213 hammer_flush_node(node);
2214 hammer_delete_node(cursor->trans, node);
2217 * cursor->node is invalid, cursor up to make the cursor
2220 error = hammer_cursor_up(cursor);
2226 * Propagate cursor->trans->tid up the B-Tree starting at the current
2227 * cursor position using pseudofs info gleaned from the passed inode.
2229 * The passed inode has no relationship to the cursor position other
2230 * then being in the same pseudofs as the insertion or deletion we
2231 * are propagating the mirror_tid for.
2234 hammer_btree_do_propagation(hammer_cursor_t cursor,
2235 hammer_pseudofs_inmem_t pfsm,
2236 hammer_btree_leaf_elm_t leaf)
2238 hammer_cursor_t ncursor;
2239 hammer_tid_t mirror_tid;
2243 * We do not propagate a mirror_tid if the filesystem was mounted
2244 * in no-mirror mode.
2246 if (cursor->trans->hmp->master_id < 0)
2250 * This is a bit of a hack because we cannot deadlock or return
2251 * EDEADLK here. The related operation has already completed and
2252 * we must propagate the mirror_tid now regardless.
2254 * Generate a new cursor which inherits the original's locks and
2255 * unlock the original. Use the new cursor to propagate the
2256 * mirror_tid. Then clean up the new cursor and reacquire locks
2259 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2260 * original's locks and the original is tracked and must be
2263 mirror_tid = cursor->node->ondisk->mirror_tid;
2264 KKASSERT(mirror_tid != 0);
2265 ncursor = hammer_push_cursor(cursor);
2266 error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2267 KKASSERT(error == 0);
2268 hammer_pop_cursor(cursor, ncursor);
2273 * Propagate a mirror TID update upwards through the B-Tree to the root.
2275 * A locked internal node must be passed in. The node will remain locked
2278 * This function syncs mirror_tid at the specified internal node's element,
2279 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2282 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2284 hammer_btree_internal_elm_t elm;
2289 error = hammer_cursor_up(cursor);
2291 error = hammer_cursor_upgrade(cursor);
2292 while (error == EDEADLK) {
2293 hammer_recover_cursor(cursor);
2294 error = hammer_cursor_upgrade(cursor);
2298 node = cursor->node;
2299 KKASSERT (node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2302 * Adjust the node's element
2304 elm = &node->ondisk->elms[cursor->index].internal;
2305 if (elm->mirror_tid >= mirror_tid)
2307 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2308 sizeof(elm->mirror_tid));
2309 elm->mirror_tid = mirror_tid;
2310 hammer_modify_node_done(node);
2311 if (hammer_debug_general & 0x0002) {
2312 kprintf("mirror_propagate: propagate "
2313 "%016llx @%016llx:%d\n",
2314 mirror_tid, node->node_offset, cursor->index);
2319 * Adjust the node's mirror_tid aggregator
2321 if (node->ondisk->mirror_tid >= mirror_tid)
2323 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2324 node->ondisk->mirror_tid = mirror_tid;
2325 hammer_modify_node_done(node);
2326 if (hammer_debug_general & 0x0002) {
2327 kprintf("mirror_propagate: propagate "
2328 "%016llx @%016llx\n",
2329 mirror_tid, node->node_offset);
2332 if (error == ENOENT)
2338 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2339 int *parent_indexp, int *errorp, int try_exclusive)
2341 hammer_node_t parent;
2342 hammer_btree_elm_t elm;
2348 parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2350 KKASSERT(parent == NULL);
2353 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2358 if (try_exclusive) {
2359 if (hammer_lock_ex_try(&parent->lock)) {
2360 hammer_rel_node(parent);
2365 hammer_lock_sh(&parent->lock);
2369 * Figure out which element in the parent is pointing to the
2372 if (node->ondisk->count) {
2373 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2378 while (i < parent->ondisk->count) {
2379 elm = &parent->ondisk->elms[i];
2380 if (elm->internal.subtree_offset == node->node_offset)
2384 if (i == parent->ondisk->count) {
2385 hammer_unlock(&parent->lock);
2386 panic("Bad B-Tree link: parent %p node %p\n", parent, node);
2389 KKASSERT(*errorp == 0);
2394 * The element (elm) has been moved to a new internal node (node).
2396 * If the element represents a pointer to an internal node that node's
2397 * parent must be adjusted to the element's new location.
2399 * XXX deadlock potential here with our exclusive locks
2402 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2403 hammer_btree_elm_t elm)
2405 hammer_node_t child;
2410 switch(elm->base.btype) {
2411 case HAMMER_BTREE_TYPE_INTERNAL:
2412 case HAMMER_BTREE_TYPE_LEAF:
2413 child = hammer_get_node(trans, elm->internal.subtree_offset,
2416 hammer_modify_node_field(trans, child, parent);
2417 child->ondisk->parent = node->node_offset;
2418 hammer_modify_node_done(child);
2419 hammer_rel_node(child);
2429 * Initialize the root of a recursive B-Tree node lock list structure.
2432 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2434 TAILQ_INIT(&parent->list);
2435 parent->parent = NULL;
2436 parent->node = node;
2438 parent->count = node->ondisk->count;
2439 parent->copy = NULL;
2444 * Exclusively lock all the children of node. This is used by the split
2445 * code to prevent anyone from accessing the children of a cursor node
2446 * while we fix-up its parent offset.
2448 * If we don't lock the children we can really mess up cursors which block
2449 * trying to cursor-up into our node.
2451 * On failure EDEADLK (or some other error) is returned. If a deadlock
2452 * error is returned the cursor is adjusted to block on termination.
2454 * The caller is responsible for managing parent->node, the root's node
2455 * is usually aliased from a cursor.
2458 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2459 hammer_node_lock_t parent)
2462 hammer_node_lock_t item;
2463 hammer_node_ondisk_t ondisk;
2464 hammer_btree_elm_t elm;
2465 hammer_node_t child;
2466 struct hammer_mount *hmp;
2470 node = parent->node;
2471 ondisk = node->ondisk;
2473 hmp = cursor->trans->hmp;
2476 * We really do not want to block on I/O with exclusive locks held,
2477 * pre-get the children before trying to lock the mess. This is
2478 * only done one-level deep for now.
2480 for (i = 0; i < ondisk->count; ++i) {
2481 ++hammer_stats_btree_elements;
2482 elm = &ondisk->elms[i];
2483 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2484 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2487 child = hammer_get_node(cursor->trans,
2488 elm->internal.subtree_offset,
2491 hammer_rel_node(child);
2497 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2498 ++hammer_stats_btree_elements;
2499 elm = &ondisk->elms[i];
2501 switch(elm->base.btype) {
2502 case HAMMER_BTREE_TYPE_INTERNAL:
2503 case HAMMER_BTREE_TYPE_LEAF:
2504 KKASSERT(elm->internal.subtree_offset != 0);
2505 child = hammer_get_node(cursor->trans,
2506 elm->internal.subtree_offset,
2514 if (hammer_lock_ex_try(&child->lock) != 0) {
2515 if (cursor->deadlk_node == NULL) {
2516 cursor->deadlk_node = child;
2517 hammer_ref_node(cursor->deadlk_node);
2520 hammer_rel_node(child);
2522 item = kmalloc(sizeof(*item), hmp->m_misc,
2524 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2525 TAILQ_INIT(&item->list);
2526 item->parent = parent;
2529 item->count = child->ondisk->count;
2532 * Recurse (used by the rebalancing code)
2534 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2535 error = hammer_btree_lock_children(
2544 hammer_btree_unlock_children(cursor, parent);
2549 * Create an in-memory copy of all B-Tree nodes listed, recursively,
2550 * including the parent.
2553 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2555 hammer_mount_t hmp = cursor->trans->hmp;
2556 hammer_node_lock_t item;
2558 if (parent->copy == NULL) {
2559 parent->copy = kmalloc(sizeof(*parent->copy), hmp->m_misc,
2561 *parent->copy = *parent->node->ondisk;
2563 TAILQ_FOREACH(item, &parent->list, entry) {
2564 hammer_btree_lock_copy(cursor, item);
2569 * Recursively sync modified copies to the media.
2572 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2574 hammer_node_lock_t item;
2576 if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2577 hammer_modify_node_all(cursor->trans, parent->node);
2578 *parent->node->ondisk = *parent->copy;
2579 hammer_modify_node_done(parent->node);
2580 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2581 hammer_flush_node(parent->node);
2582 hammer_delete_node(cursor->trans, parent->node);
2585 TAILQ_FOREACH(item, &parent->list, entry) {
2586 hammer_btree_sync_copy(cursor, item);
2591 * Release previously obtained node locks. The caller is responsible for
2592 * cleaning up parent->node itself (its usually just aliased from a cursor),
2593 * but this function will take care of the copies.
2596 hammer_btree_unlock_children(hammer_cursor_t cursor, hammer_node_lock_t parent)
2598 hammer_node_lock_t item;
2601 kfree(parent->copy, cursor->trans->hmp->m_misc);
2602 parent->copy = NULL; /* safety */
2604 while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2605 TAILQ_REMOVE(&parent->list, item, entry);
2606 hammer_btree_unlock_children(cursor, item);
2607 hammer_unlock(&item->node->lock);
2608 hammer_rel_node(item->node);
2609 kfree(item, cursor->trans->hmp->m_misc);
2613 /************************************************************************
2614 * MISCELLANIOUS SUPPORT *
2615 ************************************************************************/
2618 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2620 * Note that for this particular function a return value of -1, 0, or +1
2621 * can denote a match if create_tid is otherwise discounted. A create_tid
2622 * of zero is considered to be 'infinity' in comparisons.
2624 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2627 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2629 if (key1->localization < key2->localization)
2631 if (key1->localization > key2->localization)
2634 if (key1->obj_id < key2->obj_id)
2636 if (key1->obj_id > key2->obj_id)
2639 if (key1->rec_type < key2->rec_type)
2641 if (key1->rec_type > key2->rec_type)
2644 if (key1->key < key2->key)
2646 if (key1->key > key2->key)
2650 * A create_tid of zero indicates a record which is undeletable
2651 * and must be considered to have a value of positive infinity.
2653 if (key1->create_tid == 0) {
2654 if (key2->create_tid == 0)
2658 if (key2->create_tid == 0)
2660 if (key1->create_tid < key2->create_tid)
2662 if (key1->create_tid > key2->create_tid)
2668 * Test a timestamp against an element to determine whether the
2669 * element is visible. A timestamp of 0 means 'infinity'.
2672 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2675 if (base->delete_tid)
2679 if (asof < base->create_tid)
2681 if (base->delete_tid && asof >= base->delete_tid)
2687 * Create a separator half way inbetween key1 and key2. For fields just
2688 * one unit apart, the separator will match key2. key1 is on the left-hand
2689 * side and key2 is on the right-hand side.
2691 * key2 must be >= the separator. It is ok for the separator to match key2.
2693 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2696 * NOTE: It might be beneficial to just scrap this whole mess and just
2697 * set the separator to key2.
2699 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2700 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2703 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2704 hammer_base_elm_t dest)
2706 bzero(dest, sizeof(*dest));
2708 dest->rec_type = key2->rec_type;
2709 dest->key = key2->key;
2710 dest->obj_id = key2->obj_id;
2711 dest->create_tid = key2->create_tid;
2713 MAKE_SEPARATOR(key1, key2, dest, localization);
2714 if (key1->localization == key2->localization) {
2715 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2716 if (key1->obj_id == key2->obj_id) {
2717 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2718 if (key1->rec_type == key2->rec_type) {
2719 MAKE_SEPARATOR(key1, key2, dest, key);
2721 * Don't bother creating a separator for
2722 * create_tid, which also conveniently avoids
2723 * having to handle the create_tid == 0
2724 * (infinity) case. Just leave create_tid
2727 * Worst case, dest matches key2 exactly,
2728 * which is acceptable.
2735 #undef MAKE_SEPARATOR
2738 * Return whether a generic internal or leaf node is full
2741 btree_node_is_full(hammer_node_ondisk_t node)
2743 switch(node->type) {
2744 case HAMMER_BTREE_TYPE_INTERNAL:
2745 if (node->count == HAMMER_BTREE_INT_ELMS)
2748 case HAMMER_BTREE_TYPE_LEAF:
2749 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2753 panic("illegal btree subtype");
2760 btree_max_elements(u_int8_t type)
2762 if (type == HAMMER_BTREE_TYPE_LEAF)
2763 return(HAMMER_BTREE_LEAF_ELMS);
2764 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2765 return(HAMMER_BTREE_INT_ELMS);
2766 panic("btree_max_elements: bad type %d\n", type);
2771 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2773 hammer_btree_elm_t elm;
2776 kprintf("node %p count=%d parent=%016llx type=%c\n",
2777 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2780 * Dump both boundary elements if an internal node
2782 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2783 for (i = 0; i <= ondisk->count; ++i) {
2784 elm = &ondisk->elms[i];
2785 hammer_print_btree_elm(elm, ondisk->type, i);
2788 for (i = 0; i < ondisk->count; ++i) {
2789 elm = &ondisk->elms[i];
2790 hammer_print_btree_elm(elm, ondisk->type, i);
2796 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2799 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
2800 kprintf("\tkey = %016llx\n", elm->base.key);
2801 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2802 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2803 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2804 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2805 kprintf("\tbtype = %02x (%c)\n",
2807 (elm->base.btype ? elm->base.btype : '?'));
2808 kprintf("\tlocalization = %02x\n", elm->base.localization);
2811 case HAMMER_BTREE_TYPE_INTERNAL:
2812 kprintf("\tsubtree_off = %016llx\n",
2813 elm->internal.subtree_offset);
2815 case HAMMER_BTREE_TYPE_RECORD:
2816 kprintf("\tdata_offset = %016llx\n", elm->leaf.data_offset);
2817 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2818 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);