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 1.73 2008/07/15 22:13:16 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
176 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
177 cursor->flags |= HAMMER_CURSOR_ATEDISK;
185 * Check internal or leaf element. Determine if the record
186 * at the cursor has gone beyond the end of our range.
188 * We recurse down through internal nodes.
190 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
191 elm = &node->elms[cursor->index];
193 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
194 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
195 if (hammer_debug_btree) {
196 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
197 cursor->node->node_offset,
199 elm[0].internal.base.obj_id,
200 elm[0].internal.base.rec_type,
201 elm[0].internal.base.key,
202 elm[0].internal.base.localization,
206 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
207 cursor->node->node_offset,
209 elm[1].internal.base.obj_id,
210 elm[1].internal.base.rec_type,
211 elm[1].internal.base.key,
212 elm[1].internal.base.localization,
221 if (r == 0 && (cursor->flags &
222 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
231 KKASSERT(elm->internal.subtree_offset != 0);
234 * If running the mirror filter see if we can skip
235 * one or more entire sub-trees. If we can we
236 * return the internal mode and the caller processes
237 * the skipped range (see mirror_read)
239 if (cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) {
240 if (elm->internal.mirror_tid <
241 cursor->cmirror->mirror_tid) {
242 hammer_cursor_mirror_filter(cursor);
247 error = hammer_cursor_down(cursor);
250 KKASSERT(cursor->index == 0);
251 /* reload stale pointer */
252 node = cursor->node->ondisk;
255 elm = &node->elms[cursor->index];
256 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
257 if (hammer_debug_btree) {
258 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
259 cursor->node->node_offset,
261 (elm[0].leaf.base.btype ?
262 elm[0].leaf.base.btype : '?'),
263 elm[0].leaf.base.obj_id,
264 elm[0].leaf.base.rec_type,
265 elm[0].leaf.base.key,
266 elm[0].leaf.base.localization,
276 * We support both end-inclusive and
277 * end-exclusive searches.
280 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
285 switch(elm->leaf.base.btype) {
286 case HAMMER_BTREE_TYPE_RECORD:
287 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
288 hammer_btree_chkts(cursor->asof, &elm->base)) {
302 * node pointer invalid after loop
308 if (hammer_debug_btree) {
309 int i = cursor->index;
310 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
311 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
313 elm->internal.base.obj_id,
314 elm->internal.base.rec_type,
315 elm->internal.base.key,
316 elm->internal.base.localization
325 * We hit an internal element that we could skip as part of a mirroring
326 * scan. Calculate the entire range being skipped.
328 * It is important to include any gaps between the parent's left_bound
329 * and the node's left_bound, and same goes for the right side.
332 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
334 struct hammer_cmirror *cmirror;
335 hammer_node_ondisk_t ondisk;
336 hammer_btree_elm_t elm;
338 ondisk = cursor->node->ondisk;
339 cmirror = cursor->cmirror;
342 * Calculate the skipped range
344 elm = &ondisk->elms[cursor->index];
345 if (cursor->index == 0)
346 cmirror->skip_beg = *cursor->left_bound;
348 cmirror->skip_beg = elm->internal.base;
349 while (cursor->index < ondisk->count) {
350 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
355 if (cursor->index == ondisk->count)
356 cmirror->skip_end = *cursor->right_bound;
358 cmirror->skip_end = elm->internal.base;
361 * clip the returned result.
363 if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
364 cmirror->skip_beg = cursor->key_beg;
365 if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
366 cmirror->skip_end = cursor->key_end;
370 * Iterate in the reverse direction. This is used by the pruning code to
371 * avoid overlapping records.
374 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
376 hammer_node_ondisk_t node;
377 hammer_btree_elm_t elm;
382 /* mirror filtering not supported for reverse iteration */
383 KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
386 * Skip past the current record. For various reasons the cursor
387 * may end up set to -1 or set to point at the end of the current
388 * node. These cases must be addressed.
390 node = cursor->node->ondisk;
393 if (cursor->index != -1 &&
394 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
397 if (cursor->index == cursor->node->ondisk->count)
401 * Loop until an element is found or we are done.
404 ++hammer_stats_btree_iterations;
405 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
408 * We iterate up the tree and then index over one element
409 * while we are at the last element in the current node.
411 if (cursor->index == -1) {
412 error = hammer_cursor_up(cursor);
414 cursor->index = 0; /* sanity */
417 /* reload stale pointer */
418 node = cursor->node->ondisk;
419 KKASSERT(cursor->index != node->count);
425 * Check internal or leaf element. Determine if the record
426 * at the cursor has gone beyond the end of our range.
428 * We recurse down through internal nodes.
430 KKASSERT(cursor->index != node->count);
431 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
432 elm = &node->elms[cursor->index];
433 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
434 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
435 if (hammer_debug_btree) {
436 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
437 cursor->node->node_offset,
439 elm[0].internal.base.obj_id,
440 elm[0].internal.base.rec_type,
441 elm[0].internal.base.key,
442 elm[0].internal.base.localization,
445 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
446 cursor->node->node_offset,
448 elm[1].internal.base.obj_id,
449 elm[1].internal.base.rec_type,
450 elm[1].internal.base.key,
451 elm[1].internal.base.localization,
465 KKASSERT(elm->internal.subtree_offset != 0);
467 error = hammer_cursor_down(cursor);
470 KKASSERT(cursor->index == 0);
471 /* reload stale pointer */
472 node = cursor->node->ondisk;
474 /* this can assign -1 if the leaf was empty */
475 cursor->index = node->count - 1;
478 elm = &node->elms[cursor->index];
479 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
480 if (hammer_debug_btree) {
481 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
482 cursor->node->node_offset,
484 (elm[0].leaf.base.btype ?
485 elm[0].leaf.base.btype : '?'),
486 elm[0].leaf.base.obj_id,
487 elm[0].leaf.base.rec_type,
488 elm[0].leaf.base.key,
489 elm[0].leaf.base.localization,
498 switch(elm->leaf.base.btype) {
499 case HAMMER_BTREE_TYPE_RECORD:
500 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
501 hammer_btree_chkts(cursor->asof, &elm->base)) {
515 * node pointer invalid after loop
521 if (hammer_debug_btree) {
522 int i = cursor->index;
523 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
524 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
526 elm->internal.base.obj_id,
527 elm->internal.base.rec_type,
528 elm->internal.base.key,
529 elm->internal.base.localization
538 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
539 * could not be found, EDEADLK if inserting and a retry is needed, and a
540 * fatal error otherwise. When retrying, the caller must terminate the
541 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
543 * The cursor is suitably positioned for a deletion on success, and suitably
544 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
547 * The cursor may begin anywhere, the search will traverse the tree in
548 * either direction to locate the requested element.
550 * Most of the logic implementing historical searches is handled here. We
551 * do an initial lookup with create_tid set to the asof TID. Due to the
552 * way records are laid out, a backwards iteration may be required if
553 * ENOENT is returned to locate the historical record. Here's the
556 * create_tid: 10 15 20
560 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
561 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
562 * not visible and thus causes ENOENT to be returned. We really need
563 * to check record 11 in LEAF1. If it also fails then the search fails
564 * (e.g. it might represent the range 11-16 and thus still not match our
565 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
566 * further iterations.
568 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
569 * and the cursor->create_check TID if an iteration might be needed.
570 * In the above example create_check would be set to 14.
573 hammer_btree_lookup(hammer_cursor_t cursor)
577 KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
578 cursor->trans->sync_lock_refs > 0);
579 ++hammer_stats_btree_lookups;
580 if (cursor->flags & HAMMER_CURSOR_ASOF) {
581 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
582 cursor->key_beg.create_tid = cursor->asof;
584 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
585 error = btree_search(cursor, 0);
586 if (error != ENOENT ||
587 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
590 * Stop if error other then ENOENT.
591 * Stop if ENOENT and not special case.
595 if (hammer_debug_btree) {
596 kprintf("CREATE_CHECK %016llx\n",
597 cursor->create_check);
599 cursor->key_beg.create_tid = cursor->create_check;
603 error = btree_search(cursor, 0);
606 error = hammer_btree_extract(cursor, cursor->flags);
611 * Execute the logic required to start an iteration. The first record
612 * located within the specified range is returned and iteration control
613 * flags are adjusted for successive hammer_btree_iterate() calls.
616 hammer_btree_first(hammer_cursor_t cursor)
620 error = hammer_btree_lookup(cursor);
621 if (error == ENOENT) {
622 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
623 error = hammer_btree_iterate(cursor);
625 cursor->flags |= HAMMER_CURSOR_ATEDISK;
630 * Similarly but for an iteration in the reverse direction.
632 * Set ATEDISK when iterating backwards to skip the current entry,
633 * which after an ENOENT lookup will be pointing beyond our end point.
636 hammer_btree_last(hammer_cursor_t cursor)
638 struct hammer_base_elm save;
641 save = cursor->key_beg;
642 cursor->key_beg = cursor->key_end;
643 error = hammer_btree_lookup(cursor);
644 cursor->key_beg = save;
645 if (error == ENOENT ||
646 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
647 cursor->flags |= HAMMER_CURSOR_ATEDISK;
648 error = hammer_btree_iterate_reverse(cursor);
650 cursor->flags |= HAMMER_CURSOR_ATEDISK;
655 * Extract the record and/or data associated with the cursor's current
656 * position. Any prior record or data stored in the cursor is replaced.
657 * The cursor must be positioned at a leaf node.
659 * NOTE: All extractions occur at the leaf of the B-Tree.
662 hammer_btree_extract(hammer_cursor_t cursor, int flags)
665 hammer_node_ondisk_t node;
666 hammer_btree_elm_t elm;
667 hammer_off_t data_off;
672 * The case where the data reference resolves to the same buffer
673 * as the record reference must be handled.
675 node = cursor->node->ondisk;
676 elm = &node->elms[cursor->index];
678 hmp = cursor->node->hmp;
681 * There is nothing to extract for an internal element.
683 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
687 * Only record types have data.
689 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
690 cursor->leaf = &elm->leaf;
692 if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
694 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
696 data_off = elm->leaf.data_offset;
697 data_len = elm->leaf.data_len;
704 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
705 cursor->data = hammer_bread_ext(hmp, data_off, data_len,
706 &error, &cursor->data_buffer);
707 if (hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0)
708 Debugger("CRC FAILED: DATA");
714 * Insert a leaf element into the B-Tree at the current cursor position.
715 * The cursor is positioned such that the element at and beyond the cursor
716 * are shifted to make room for the new record.
718 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
719 * flag set and that call must return ENOENT before this function can be
722 * The caller may depend on the cursor's exclusive lock after return to
723 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
725 * ENOSPC is returned if there is no room to insert a new record.
728 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
731 hammer_node_ondisk_t node;
736 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
738 ++hammer_stats_btree_inserts;
741 * Insert the element at the leaf node and update the count in the
742 * parent. It is possible for parent to be NULL, indicating that
743 * the filesystem's ROOT B-Tree node is a leaf itself, which is
744 * possible. The root inode can never be deleted so the leaf should
747 * Remember that the right-hand boundary is not included in the
750 hammer_modify_node_all(cursor->trans, cursor->node);
751 node = cursor->node->ondisk;
753 KKASSERT(elm->base.btype != 0);
754 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
755 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
756 if (i != node->count) {
757 bcopy(&node->elms[i], &node->elms[i+1],
758 (node->count - i) * sizeof(*elm));
760 node->elms[i].leaf = *elm;
762 hammer_cursor_inserted_element(cursor->node, i);
765 * Update the leaf node's aggregate mirror_tid for mirroring
768 if (node->mirror_tid < elm->base.delete_tid) {
769 node->mirror_tid = elm->base.delete_tid;
772 if (node->mirror_tid < elm->base.create_tid) {
773 node->mirror_tid = elm->base.create_tid;
776 hammer_modify_node_done(cursor->node);
779 * Debugging sanity checks.
781 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
782 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
784 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
786 if (i != node->count - 1)
787 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
793 * Delete a record from the B-Tree at the current cursor position.
794 * The cursor is positioned such that the current element is the one
797 * On return the cursor will be positioned after the deleted element and
798 * MAY point to an internal node. It will be suitable for the continuation
799 * of an iteration but not for an insertion or deletion.
801 * Deletions will attempt to partially rebalance the B-Tree in an upward
802 * direction, but will terminate rather then deadlock. Empty internal nodes
803 * are never allowed by a deletion which deadlocks may end up giving us an
804 * empty leaf. The pruner will clean up and rebalance the tree.
806 * This function can return EDEADLK, requiring the caller to retry the
807 * operation after clearing the deadlock.
810 hammer_btree_delete(hammer_cursor_t cursor)
812 hammer_node_ondisk_t ondisk;
814 hammer_node_t parent;
818 KKASSERT (cursor->trans->sync_lock_refs > 0);
819 if ((error = hammer_cursor_upgrade(cursor)) != 0)
821 ++hammer_stats_btree_deletes;
824 * Delete the element from the leaf node.
826 * Remember that leaf nodes do not have boundaries.
829 ondisk = node->ondisk;
832 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
833 KKASSERT(i >= 0 && i < ondisk->count);
834 hammer_modify_node_all(cursor->trans, node);
835 if (i + 1 != ondisk->count) {
836 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
837 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
840 hammer_modify_node_done(node);
841 hammer_cursor_deleted_element(node, i);
844 * Validate local parent
846 if (ondisk->parent) {
847 parent = cursor->parent;
849 KKASSERT(parent != NULL);
850 KKASSERT(parent->node_offset == ondisk->parent);
854 * If the leaf becomes empty it must be detached from the parent,
855 * potentially recursing through to the filesystem root.
857 * This may reposition the cursor at one of the parent's of the
860 * Ignore deadlock errors, that simply means that btree_remove
861 * was unable to recurse and had to leave us with an empty leaf.
863 KKASSERT(cursor->index <= ondisk->count);
864 if (ondisk->count == 0) {
865 error = btree_remove(cursor);
866 if (error == EDEADLK)
871 KKASSERT(cursor->parent == NULL ||
872 cursor->parent_index < cursor->parent->ondisk->count);
877 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
879 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
881 * The search can begin ANYWHERE in the B-Tree. As a first step the search
882 * iterates up the tree as necessary to properly position itself prior to
883 * actually doing the sarch.
885 * INSERTIONS: The search will split full nodes and leaves on its way down
886 * and guarentee that the leaf it ends up on is not full. If we run out
887 * of space the search continues to the leaf (to position the cursor for
888 * the spike), but ENOSPC is returned.
890 * The search is only guarenteed to end up on a leaf if an error code of 0
891 * is returned, or if inserting and an error code of ENOENT is returned.
892 * Otherwise it can stop at an internal node. On success a search returns
895 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
896 * filesystem, and it is not simple code. Please note the following facts:
898 * - Internal node recursions have a boundary on the left AND right. The
899 * right boundary is non-inclusive. The create_tid is a generic part
900 * of the key for internal nodes.
902 * - Leaf nodes contain terminal elements only now.
904 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
905 * historical search. ASOF and INSERT are mutually exclusive. When
906 * doing an as-of lookup btree_search() checks for a right-edge boundary
907 * case. If while recursing down the left-edge differs from the key
908 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
909 * with cursor->create_check. This is used by btree_lookup() to iterate.
910 * The iteration backwards because as-of searches can wind up going
911 * down the wrong branch of the B-Tree.
915 btree_search(hammer_cursor_t cursor, int flags)
917 hammer_node_ondisk_t node;
918 hammer_btree_elm_t elm;
925 flags |= cursor->flags;
926 ++hammer_stats_btree_searches;
928 if (hammer_debug_btree) {
929 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
930 cursor->node->node_offset,
932 cursor->key_beg.obj_id,
933 cursor->key_beg.rec_type,
935 cursor->key_beg.create_tid,
936 cursor->key_beg.localization,
940 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
941 cursor->parent->node_offset, cursor->parent_index,
942 cursor->left_bound->obj_id,
943 cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
944 cursor->right_bound->obj_id,
945 cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
947 &cursor->parent->ondisk->elms[cursor->parent_index],
949 &cursor->parent->ondisk->elms[cursor->parent_index+1]
954 * Move our cursor up the tree until we find a node whos range covers
955 * the key we are trying to locate.
957 * The left bound is inclusive, the right bound is non-inclusive.
958 * It is ok to cursor up too far.
961 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
962 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
965 KKASSERT(cursor->parent);
966 ++hammer_stats_btree_iterations;
967 error = hammer_cursor_up(cursor);
973 * The delete-checks below are based on node, not parent. Set the
974 * initial delete-check based on the parent.
977 KKASSERT(cursor->left_bound->create_tid != 1);
978 cursor->create_check = cursor->left_bound->create_tid - 1;
979 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
983 * We better have ended up with a node somewhere.
985 KKASSERT(cursor->node != NULL);
988 * If we are inserting we can't start at a full node if the parent
989 * is also full (because there is no way to split the node),
990 * continue running up the tree until the requirement is satisfied
991 * or we hit the root of the filesystem.
993 * (If inserting we aren't doing an as-of search so we don't have
994 * to worry about create_check).
996 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
997 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
998 if (btree_node_is_full(cursor->node->ondisk) == 0)
1001 if (btree_node_is_full(cursor->node->ondisk) ==0)
1004 if (cursor->node->ondisk->parent == 0 ||
1005 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1008 ++hammer_stats_btree_iterations;
1009 error = hammer_cursor_up(cursor);
1010 /* node may have become stale */
1016 * Push down through internal nodes to locate the requested key.
1018 node = cursor->node->ondisk;
1019 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1021 * Scan the node to find the subtree index to push down into.
1022 * We go one-past, then back-up.
1024 * We must proactively remove deleted elements which may
1025 * have been left over from a deadlocked btree_remove().
1027 * The left and right boundaries are included in the loop
1028 * in order to detect edge cases.
1030 * If the separator only differs by create_tid (r == 1)
1031 * and we are doing an as-of search, we may end up going
1032 * down a branch to the left of the one containing the
1033 * desired key. This requires numerous special cases.
1035 ++hammer_stats_btree_iterations;
1036 if (hammer_debug_btree) {
1037 kprintf("SEARCH-I %016llx count=%d\n",
1038 cursor->node->node_offset,
1043 * Try to shortcut the search before dropping into the
1044 * linear loop. Locate the first node where r <= 1.
1046 i = hammer_btree_search_node(&cursor->key_beg, node);
1047 while (i <= node->count) {
1048 ++hammer_stats_btree_elements;
1049 elm = &node->elms[i];
1050 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1051 if (hammer_debug_btree > 2) {
1052 kprintf(" IELM %p %d r=%d\n",
1053 &node->elms[i], i, r);
1058 KKASSERT(elm->base.create_tid != 1);
1059 cursor->create_check = elm->base.create_tid - 1;
1060 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1064 if (hammer_debug_btree) {
1065 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1070 * These cases occur when the parent's idea of the boundary
1071 * is wider then the child's idea of the boundary, and
1072 * require special handling. If not inserting we can
1073 * terminate the search early for these cases but the
1074 * child's boundaries cannot be unconditionally modified.
1078 * If i == 0 the search terminated to the LEFT of the
1079 * left_boundary but to the RIGHT of the parent's left
1084 elm = &node->elms[0];
1087 * If we aren't inserting we can stop here.
1089 if ((flags & (HAMMER_CURSOR_INSERT |
1090 HAMMER_CURSOR_PRUNING)) == 0) {
1096 * Correct a left-hand boundary mismatch.
1098 * We can only do this if we can upgrade the lock,
1099 * and synchronized as a background cursor (i.e.
1100 * inserting or pruning).
1102 * WARNING: We can only do this if inserting, i.e.
1103 * we are running on the backend.
1105 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1107 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1108 hammer_modify_node_field(cursor->trans, cursor->node,
1110 save = node->elms[0].base.btype;
1111 node->elms[0].base = *cursor->left_bound;
1112 node->elms[0].base.btype = save;
1113 hammer_modify_node_done(cursor->node);
1114 } else if (i == node->count + 1) {
1116 * If i == node->count + 1 the search terminated to
1117 * the RIGHT of the right boundary but to the LEFT
1118 * of the parent's right boundary. If we aren't
1119 * inserting we can stop here.
1121 * Note that the last element in this case is
1122 * elms[i-2] prior to adjustments to 'i'.
1125 if ((flags & (HAMMER_CURSOR_INSERT |
1126 HAMMER_CURSOR_PRUNING)) == 0) {
1132 * Correct a right-hand boundary mismatch.
1133 * (actual push-down record is i-2 prior to
1134 * adjustments to i).
1136 * We can only do this if we can upgrade the lock,
1137 * and synchronized as a background cursor (i.e.
1138 * inserting or pruning).
1140 * WARNING: We can only do this if inserting, i.e.
1141 * we are running on the backend.
1143 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1145 elm = &node->elms[i];
1146 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1147 hammer_modify_node(cursor->trans, cursor->node,
1148 &elm->base, sizeof(elm->base));
1149 elm->base = *cursor->right_bound;
1150 hammer_modify_node_done(cursor->node);
1154 * The push-down index is now i - 1. If we had
1155 * terminated on the right boundary this will point
1156 * us at the last element.
1161 elm = &node->elms[i];
1163 if (hammer_debug_btree) {
1164 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1165 "key=%016llx cre=%016llx lo=%02x\n",
1166 cursor->node->node_offset,
1168 elm->internal.base.obj_id,
1169 elm->internal.base.rec_type,
1170 elm->internal.base.key,
1171 elm->internal.base.create_tid,
1172 elm->internal.base.localization
1177 * We better have a valid subtree offset.
1179 KKASSERT(elm->internal.subtree_offset != 0);
1182 * Handle insertion and deletion requirements.
1184 * If inserting split full nodes. The split code will
1185 * adjust cursor->node and cursor->index if the current
1186 * index winds up in the new node.
1188 * If inserting and a left or right edge case was detected,
1189 * we cannot correct the left or right boundary and must
1190 * prepend and append an empty leaf node in order to make
1191 * the boundary correction.
1193 * If we run out of space we set enospc and continue on
1194 * to a leaf to provide the spike code with a good point
1197 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1198 if (btree_node_is_full(node)) {
1199 error = btree_split_internal(cursor);
1201 if (error != ENOSPC)
1206 * reload stale pointers
1209 node = cursor->node->ondisk;
1214 * Push down (push into new node, existing node becomes
1215 * the parent) and continue the search.
1217 error = hammer_cursor_down(cursor);
1218 /* node may have become stale */
1221 node = cursor->node->ondisk;
1225 * We are at a leaf, do a linear search of the key array.
1227 * On success the index is set to the matching element and 0
1230 * On failure the index is set to the insertion point and ENOENT
1233 * Boundaries are not stored in leaf nodes, so the index can wind
1234 * up to the left of element 0 (index == 0) or past the end of
1235 * the array (index == node->count). It is also possible that the
1236 * leaf might be empty.
1238 ++hammer_stats_btree_iterations;
1239 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1240 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1241 if (hammer_debug_btree) {
1242 kprintf("SEARCH-L %016llx count=%d\n",
1243 cursor->node->node_offset,
1248 * Try to shortcut the search before dropping into the
1249 * linear loop. Locate the first node where r <= 1.
1251 i = hammer_btree_search_node(&cursor->key_beg, node);
1252 while (i < node->count) {
1253 ++hammer_stats_btree_elements;
1254 elm = &node->elms[i];
1256 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1258 if (hammer_debug_btree > 1)
1259 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1262 * We are at a record element. Stop if we've flipped past
1263 * key_beg, not counting the create_tid test. Allow the
1264 * r == 1 case (key_beg > element but differs only by its
1265 * create_tid) to fall through to the AS-OF check.
1267 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1277 * Check our as-of timestamp against the element.
1279 if (flags & HAMMER_CURSOR_ASOF) {
1280 if (hammer_btree_chkts(cursor->asof,
1281 &node->elms[i].base) != 0) {
1287 if (r > 0) { /* can only be +1 */
1295 if (hammer_debug_btree) {
1296 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1297 cursor->node->node_offset, i);
1303 * The search of the leaf node failed. i is the insertion point.
1306 if (hammer_debug_btree) {
1307 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1308 cursor->node->node_offset, i);
1312 * No exact match was found, i is now at the insertion point.
1314 * If inserting split a full leaf before returning. This
1315 * may have the side effect of adjusting cursor->node and
1319 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1320 btree_node_is_full(node)) {
1321 error = btree_split_leaf(cursor);
1323 if (error != ENOSPC)
1328 * reload stale pointers
1332 node = &cursor->node->internal;
1337 * We reached a leaf but did not find the key we were looking for.
1338 * If this is an insert we will be properly positioned for an insert
1339 * (ENOENT) or spike (ENOSPC) operation.
1341 error = enospc ? ENOSPC : ENOENT;
1347 * Heuristical search for the first element whos comparison is <= 1. May
1348 * return an index whos compare result is > 1 but may only return an index
1349 * whos compare result is <= 1 if it is the first element with that result.
1352 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1360 * Don't bother if the node does not have very many elements
1365 i = b + (s - b) / 2;
1366 ++hammer_stats_btree_elements;
1367 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1378 /************************************************************************
1379 * SPLITTING AND MERGING *
1380 ************************************************************************
1382 * These routines do all the dirty work required to split and merge nodes.
1386 * Split an internal node into two nodes and move the separator at the split
1387 * point to the parent.
1389 * (cursor->node, cursor->index) indicates the element the caller intends
1390 * to push into. We will adjust node and index if that element winds
1391 * up in the split node.
1393 * If we are at the root of the filesystem a new root must be created with
1394 * two elements, one pointing to the original root and one pointing to the
1395 * newly allocated split node.
1399 btree_split_internal(hammer_cursor_t cursor)
1401 hammer_node_ondisk_t ondisk;
1403 hammer_node_t parent;
1404 hammer_node_t new_node;
1405 hammer_btree_elm_t elm;
1406 hammer_btree_elm_t parent_elm;
1407 hammer_node_locklist_t locklist = NULL;
1408 hammer_mount_t hmp = cursor->trans->hmp;
1414 const int esize = sizeof(*elm);
1416 error = hammer_btree_lock_children(cursor, &locklist);
1419 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1421 ++hammer_stats_btree_splits;
1424 * We are splitting but elms[split] will be promoted to the parent,
1425 * leaving the right hand node with one less element. If the
1426 * insertion point will be on the left-hand side adjust the split
1427 * point to give the right hand side one additional node.
1429 node = cursor->node;
1430 ondisk = node->ondisk;
1431 split = (ondisk->count + 1) / 2;
1432 if (cursor->index <= split)
1436 * If we are at the root of the filesystem, create a new root node
1437 * with 1 element and split normally. Avoid making major
1438 * modifications until we know the whole operation will work.
1440 if (ondisk->parent == 0) {
1441 parent = hammer_alloc_btree(cursor->trans, &error);
1444 hammer_lock_ex(&parent->lock);
1445 hammer_modify_node_noundo(cursor->trans, parent);
1446 ondisk = parent->ondisk;
1449 ondisk->mirror_tid = node->ondisk->mirror_tid;
1450 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1451 ondisk->elms[0].base = hmp->root_btree_beg;
1452 ondisk->elms[0].base.btype = node->ondisk->type;
1453 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1454 ondisk->elms[1].base = hmp->root_btree_end;
1455 hammer_modify_node_done(parent);
1456 /* ondisk->elms[1].base.btype - not used */
1458 parent_index = 0; /* index of current node in parent */
1461 parent = cursor->parent;
1462 parent_index = cursor->parent_index;
1466 * Split node into new_node at the split point.
1468 * B O O O P N N B <-- P = node->elms[split]
1469 * 0 1 2 3 4 5 6 <-- subtree indices
1474 * B O O O B B N N B <--- inner boundary points are 'P'
1478 new_node = hammer_alloc_btree(cursor->trans, &error);
1479 if (new_node == NULL) {
1481 hammer_unlock(&parent->lock);
1482 hammer_delete_node(cursor->trans, parent);
1483 hammer_rel_node(parent);
1487 hammer_lock_ex(&new_node->lock);
1490 * Create the new node. P becomes the left-hand boundary in the
1491 * new node. Copy the right-hand boundary as well.
1493 * elm is the new separator.
1495 hammer_modify_node_noundo(cursor->trans, new_node);
1496 hammer_modify_node_all(cursor->trans, node);
1497 ondisk = node->ondisk;
1498 elm = &ondisk->elms[split];
1499 bcopy(elm, &new_node->ondisk->elms[0],
1500 (ondisk->count - split + 1) * esize);
1501 new_node->ondisk->count = ondisk->count - split;
1502 new_node->ondisk->parent = parent->node_offset;
1503 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1504 new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1505 KKASSERT(ondisk->type == new_node->ondisk->type);
1506 hammer_cursor_split_node(node, new_node, split);
1509 * Cleanup the original node. Elm (P) becomes the new boundary,
1510 * its subtree_offset was moved to the new node. If we had created
1511 * a new root its parent pointer may have changed.
1513 elm->internal.subtree_offset = 0;
1514 ondisk->count = split;
1517 * Insert the separator into the parent, fixup the parent's
1518 * reference to the original node, and reference the new node.
1519 * The separator is P.
1521 * Remember that base.count does not include the right-hand boundary.
1523 hammer_modify_node_all(cursor->trans, parent);
1524 ondisk = parent->ondisk;
1525 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1526 parent_elm = &ondisk->elms[parent_index+1];
1527 bcopy(parent_elm, parent_elm + 1,
1528 (ondisk->count - parent_index) * esize);
1529 parent_elm->internal.base = elm->base; /* separator P */
1530 parent_elm->internal.base.btype = new_node->ondisk->type;
1531 parent_elm->internal.subtree_offset = new_node->node_offset;
1532 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1534 hammer_modify_node_done(parent);
1535 hammer_cursor_inserted_element(parent, parent_index + 1);
1538 * The children of new_node need their parent pointer set to new_node.
1539 * The children have already been locked by
1540 * hammer_btree_lock_children().
1542 for (i = 0; i < new_node->ondisk->count; ++i) {
1543 elm = &new_node->ondisk->elms[i];
1544 error = btree_set_parent(cursor->trans, new_node, elm);
1546 panic("btree_split_internal: btree-fixup problem");
1549 hammer_modify_node_done(new_node);
1552 * The filesystem's root B-Tree pointer may have to be updated.
1555 hammer_volume_t volume;
1557 volume = hammer_get_root_volume(hmp, &error);
1558 KKASSERT(error == 0);
1560 hammer_modify_volume_field(cursor->trans, volume,
1562 volume->ondisk->vol0_btree_root = parent->node_offset;
1563 hammer_modify_volume_done(volume);
1564 node->ondisk->parent = parent->node_offset;
1565 if (cursor->parent) {
1566 hammer_unlock(&cursor->parent->lock);
1567 hammer_rel_node(cursor->parent);
1569 cursor->parent = parent; /* lock'd and ref'd */
1570 hammer_rel_volume(volume, 0);
1572 hammer_modify_node_done(node);
1575 * Ok, now adjust the cursor depending on which element the original
1576 * index was pointing at. If we are >= the split point the push node
1577 * is now in the new node.
1579 * NOTE: If we are at the split point itself we cannot stay with the
1580 * original node because the push index will point at the right-hand
1581 * boundary, which is illegal.
1583 * NOTE: The cursor's parent or parent_index must be adjusted for
1584 * the case where a new parent (new root) was created, and the case
1585 * where the cursor is now pointing at the split node.
1587 if (cursor->index >= split) {
1588 cursor->parent_index = parent_index + 1;
1589 cursor->index -= split;
1590 hammer_unlock(&cursor->node->lock);
1591 hammer_rel_node(cursor->node);
1592 cursor->node = new_node; /* locked and ref'd */
1594 cursor->parent_index = parent_index;
1595 hammer_unlock(&new_node->lock);
1596 hammer_rel_node(new_node);
1600 * Fixup left and right bounds
1602 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1603 cursor->left_bound = &parent_elm[0].internal.base;
1604 cursor->right_bound = &parent_elm[1].internal.base;
1605 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1606 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1607 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1608 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1611 hammer_btree_unlock_children(&locklist);
1612 hammer_cursor_downgrade(cursor);
1617 * Same as the above, but splits a full leaf node.
1623 btree_split_leaf(hammer_cursor_t cursor)
1625 hammer_node_ondisk_t ondisk;
1626 hammer_node_t parent;
1629 hammer_node_t new_leaf;
1630 hammer_btree_elm_t elm;
1631 hammer_btree_elm_t parent_elm;
1632 hammer_base_elm_t mid_boundary;
1637 const size_t esize = sizeof(*elm);
1639 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1641 ++hammer_stats_btree_splits;
1643 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1644 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1645 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1646 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1649 * Calculate the split point. If the insertion point will be on
1650 * the left-hand side adjust the split point to give the right
1651 * hand side one additional node.
1653 * Spikes are made up of two leaf elements which cannot be
1656 leaf = cursor->node;
1657 ondisk = leaf->ondisk;
1658 split = (ondisk->count + 1) / 2;
1659 if (cursor->index <= split)
1664 elm = &ondisk->elms[split];
1666 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1667 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1668 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1669 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1672 * If we are at the root of the tree, create a new root node with
1673 * 1 element and split normally. Avoid making major modifications
1674 * until we know the whole operation will work.
1676 if (ondisk->parent == 0) {
1677 parent = hammer_alloc_btree(cursor->trans, &error);
1680 hammer_lock_ex(&parent->lock);
1681 hammer_modify_node_noundo(cursor->trans, parent);
1682 ondisk = parent->ondisk;
1685 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1686 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1687 ondisk->elms[0].base = hmp->root_btree_beg;
1688 ondisk->elms[0].base.btype = leaf->ondisk->type;
1689 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1690 ondisk->elms[1].base = hmp->root_btree_end;
1691 /* ondisk->elms[1].base.btype = not used */
1692 hammer_modify_node_done(parent);
1694 parent_index = 0; /* insertion point in parent */
1697 parent = cursor->parent;
1698 parent_index = cursor->parent_index;
1702 * Split leaf into new_leaf at the split point. Select a separator
1703 * value in-between the two leafs but with a bent towards the right
1704 * leaf since comparisons use an 'elm >= separator' inequality.
1713 new_leaf = hammer_alloc_btree(cursor->trans, &error);
1714 if (new_leaf == NULL) {
1716 hammer_unlock(&parent->lock);
1717 hammer_delete_node(cursor->trans, parent);
1718 hammer_rel_node(parent);
1722 hammer_lock_ex(&new_leaf->lock);
1725 * Create the new node and copy the leaf elements from the split
1726 * point on to the new node.
1728 hammer_modify_node_all(cursor->trans, leaf);
1729 hammer_modify_node_noundo(cursor->trans, new_leaf);
1730 ondisk = leaf->ondisk;
1731 elm = &ondisk->elms[split];
1732 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1733 new_leaf->ondisk->count = ondisk->count - split;
1734 new_leaf->ondisk->parent = parent->node_offset;
1735 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1736 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1737 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1738 hammer_modify_node_done(new_leaf);
1739 hammer_cursor_split_node(leaf, new_leaf, split);
1742 * Cleanup the original node. Because this is a leaf node and
1743 * leaf nodes do not have a right-hand boundary, there
1744 * aren't any special edge cases to clean up. We just fixup the
1747 ondisk->count = split;
1750 * Insert the separator into the parent, fixup the parent's
1751 * reference to the original node, and reference the new node.
1752 * The separator is P.
1754 * Remember that base.count does not include the right-hand boundary.
1755 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1757 hammer_modify_node_all(cursor->trans, parent);
1758 ondisk = parent->ondisk;
1759 KKASSERT(split != 0);
1760 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1761 parent_elm = &ondisk->elms[parent_index+1];
1762 bcopy(parent_elm, parent_elm + 1,
1763 (ondisk->count - parent_index) * esize);
1765 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1766 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1767 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1768 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1769 mid_boundary = &parent_elm->base;
1771 hammer_modify_node_done(parent);
1772 hammer_cursor_inserted_element(parent, parent_index + 1);
1775 * The filesystem's root B-Tree pointer may have to be updated.
1778 hammer_volume_t volume;
1780 volume = hammer_get_root_volume(hmp, &error);
1781 KKASSERT(error == 0);
1783 hammer_modify_volume_field(cursor->trans, volume,
1785 volume->ondisk->vol0_btree_root = parent->node_offset;
1786 hammer_modify_volume_done(volume);
1787 leaf->ondisk->parent = parent->node_offset;
1788 if (cursor->parent) {
1789 hammer_unlock(&cursor->parent->lock);
1790 hammer_rel_node(cursor->parent);
1792 cursor->parent = parent; /* lock'd and ref'd */
1793 hammer_rel_volume(volume, 0);
1795 hammer_modify_node_done(leaf);
1798 * Ok, now adjust the cursor depending on which element the original
1799 * index was pointing at. If we are >= the split point the push node
1800 * is now in the new node.
1802 * NOTE: If we are at the split point itself we need to select the
1803 * old or new node based on where key_beg's insertion point will be.
1804 * If we pick the wrong side the inserted element will wind up in
1805 * the wrong leaf node and outside that node's bounds.
1807 if (cursor->index > split ||
1808 (cursor->index == split &&
1809 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1810 cursor->parent_index = parent_index + 1;
1811 cursor->index -= split;
1812 hammer_unlock(&cursor->node->lock);
1813 hammer_rel_node(cursor->node);
1814 cursor->node = new_leaf;
1816 cursor->parent_index = parent_index;
1817 hammer_unlock(&new_leaf->lock);
1818 hammer_rel_node(new_leaf);
1822 * Fixup left and right bounds
1824 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1825 cursor->left_bound = &parent_elm[0].internal.base;
1826 cursor->right_bound = &parent_elm[1].internal.base;
1829 * Assert that the bounds are correct.
1831 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1832 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1833 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1834 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1835 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1836 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1839 hammer_cursor_downgrade(cursor);
1846 * Recursively correct the right-hand boundary's create_tid to (tid) as
1847 * long as the rest of the key matches. We have to recurse upward in
1848 * the tree as well as down the left side of each parent's right node.
1850 * Return EDEADLK if we were only partially successful, forcing the caller
1851 * to try again. The original cursor is not modified. This routine can
1852 * also fail with EDEADLK if it is forced to throw away a portion of its
1855 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1858 TAILQ_ENTRY(hammer_rhb) entry;
1863 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1866 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1868 struct hammer_rhb_list rhb_list;
1869 hammer_base_elm_t elm;
1870 hammer_node_t orig_node;
1871 struct hammer_rhb *rhb;
1875 TAILQ_INIT(&rhb_list);
1878 * Save our position so we can restore it on return. This also
1879 * gives us a stable 'elm'.
1881 orig_node = cursor->node;
1882 hammer_ref_node(orig_node);
1883 hammer_lock_sh(&orig_node->lock);
1884 orig_index = cursor->index;
1885 elm = &orig_node->ondisk->elms[orig_index].base;
1888 * Now build a list of parents going up, allocating a rhb
1889 * structure for each one.
1891 while (cursor->parent) {
1893 * Stop if we no longer have any right-bounds to fix up
1895 if (elm->obj_id != cursor->right_bound->obj_id ||
1896 elm->rec_type != cursor->right_bound->rec_type ||
1897 elm->key != cursor->right_bound->key) {
1902 * Stop if the right-hand bound's create_tid does not
1903 * need to be corrected.
1905 if (cursor->right_bound->create_tid >= tid)
1908 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1909 rhb->node = cursor->parent;
1910 rhb->index = cursor->parent_index;
1911 hammer_ref_node(rhb->node);
1912 hammer_lock_sh(&rhb->node->lock);
1913 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1915 hammer_cursor_up(cursor);
1919 * now safely adjust the right hand bound for each rhb. This may
1920 * also require taking the right side of the tree and iterating down
1924 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1925 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1928 TAILQ_REMOVE(&rhb_list, rhb, entry);
1929 hammer_unlock(&rhb->node->lock);
1930 hammer_rel_node(rhb->node);
1931 kfree(rhb, M_HAMMER);
1933 switch (cursor->node->ondisk->type) {
1934 case HAMMER_BTREE_TYPE_INTERNAL:
1936 * Right-boundary for parent at internal node
1937 * is one element to the right of the element whos
1938 * right boundary needs adjusting. We must then
1939 * traverse down the left side correcting any left
1940 * bounds (which may now be too far to the left).
1943 error = hammer_btree_correct_lhb(cursor, tid);
1946 panic("hammer_btree_correct_rhb(): Bad node type");
1955 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1956 TAILQ_REMOVE(&rhb_list, rhb, entry);
1957 hammer_unlock(&rhb->node->lock);
1958 hammer_rel_node(rhb->node);
1959 kfree(rhb, M_HAMMER);
1961 error = hammer_cursor_seek(cursor, orig_node, orig_index);
1962 hammer_unlock(&orig_node->lock);
1963 hammer_rel_node(orig_node);
1968 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1969 * bound going downward starting at the current cursor position.
1971 * This function does not restore the cursor after use.
1974 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1976 struct hammer_rhb_list rhb_list;
1977 hammer_base_elm_t elm;
1978 hammer_base_elm_t cmp;
1979 struct hammer_rhb *rhb;
1982 TAILQ_INIT(&rhb_list);
1984 cmp = &cursor->node->ondisk->elms[cursor->index].base;
1987 * Record the node and traverse down the left-hand side for all
1988 * matching records needing a boundary correction.
1992 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1993 rhb->node = cursor->node;
1994 rhb->index = cursor->index;
1995 hammer_ref_node(rhb->node);
1996 hammer_lock_sh(&rhb->node->lock);
1997 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1999 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2001 * Nothing to traverse down if we are at the right
2002 * boundary of an internal node.
2004 if (cursor->index == cursor->node->ondisk->count)
2007 elm = &cursor->node->ondisk->elms[cursor->index].base;
2008 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2010 panic("Illegal leaf record type %02x", elm->btype);
2012 error = hammer_cursor_down(cursor);
2016 elm = &cursor->node->ondisk->elms[cursor->index].base;
2017 if (elm->obj_id != cmp->obj_id ||
2018 elm->rec_type != cmp->rec_type ||
2019 elm->key != cmp->key) {
2022 if (elm->create_tid >= tid)
2028 * Now we can safely adjust the left-hand boundary from the bottom-up.
2029 * The last element we remove from the list is the caller's right hand
2030 * boundary, which must also be adjusted.
2032 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2033 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2036 TAILQ_REMOVE(&rhb_list, rhb, entry);
2037 hammer_unlock(&rhb->node->lock);
2038 hammer_rel_node(rhb->node);
2039 kfree(rhb, M_HAMMER);
2041 elm = &cursor->node->ondisk->elms[cursor->index].base;
2042 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2043 hammer_modify_node(cursor->trans, cursor->node,
2045 sizeof(elm->create_tid));
2046 elm->create_tid = tid;
2047 hammer_modify_node_done(cursor->node);
2049 panic("hammer_btree_correct_lhb(): Bad element type");
2056 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2057 TAILQ_REMOVE(&rhb_list, rhb, entry);
2058 hammer_unlock(&rhb->node->lock);
2059 hammer_rel_node(rhb->node);
2060 kfree(rhb, M_HAMMER);
2068 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2069 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2070 * the operation due to a deadlock, or some other error.
2072 * This routine is always called with an empty, locked leaf but may recurse
2073 * into want-to-be-empty parents as part of its operation.
2075 * It should also be noted that when removing empty leaves we must be sure
2076 * to test and update mirror_tid because another thread may have deadlocked
2077 * against us (or someone) trying to propagate it up and cannot retry once
2078 * the node has been deleted.
2080 * On return the cursor may end up pointing to an internal node, suitable
2081 * for further iteration but not for an immediate insertion or deletion.
2084 btree_remove(hammer_cursor_t cursor)
2086 hammer_node_ondisk_t ondisk;
2087 hammer_btree_elm_t elm;
2089 hammer_node_t parent;
2090 const int esize = sizeof(*elm);
2093 node = cursor->node;
2096 * When deleting the root of the filesystem convert it to
2097 * an empty leaf node. Internal nodes cannot be empty.
2099 ondisk = node->ondisk;
2100 if (ondisk->parent == 0) {
2101 KKASSERT(cursor->parent == NULL);
2102 hammer_modify_node_all(cursor->trans, node);
2103 KKASSERT(ondisk == node->ondisk);
2104 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2106 hammer_modify_node_done(node);
2111 parent = cursor->parent;
2112 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2115 * Attempt to remove the parent's reference to the child. If the
2116 * parent would become empty we have to recurse. If we fail we
2117 * leave the parent pointing to an empty leaf node.
2119 if (parent->ondisk->count == 1) {
2121 * This special cursor_up_locked() call leaves the original
2122 * node exclusively locked and referenced, leaves the
2123 * original parent locked (as the new node), and locks the
2124 * new parent. It can return EDEADLK.
2126 error = hammer_cursor_up_locked(cursor);
2128 error = btree_remove(cursor);
2130 hammer_modify_node_all(cursor->trans, node);
2131 ondisk = node->ondisk;
2132 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2134 hammer_modify_node_done(node);
2135 hammer_flush_node(node);
2136 hammer_delete_node(cursor->trans, node);
2138 kprintf("Warning: BTREE_REMOVE: Defering "
2139 "parent removal1 @ %016llx, skipping\n",
2142 hammer_unlock(&node->lock);
2143 hammer_rel_node(node);
2145 kprintf("Warning: BTREE_REMOVE: Defering parent "
2146 "removal2 @ %016llx, skipping\n",
2150 KKASSERT(parent->ondisk->count > 1);
2152 hammer_modify_node_all(cursor->trans, parent);
2153 ondisk = parent->ondisk;
2154 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2156 elm = &ondisk->elms[cursor->parent_index];
2157 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2158 KKASSERT(ondisk->count > 0);
2161 * We must retain the highest mirror_tid. The deleted
2162 * range is now encompassed by the element to the left.
2163 * If we are already at the left edge the new left edge
2164 * inherits mirror_tid.
2166 * Note that bounds of the parent to our parent may create
2167 * a gap to the left of our left-most node or to the right
2168 * of our right-most node. The gap is silently included
2169 * in the mirror_tid's area of effect from the point of view
2172 if (cursor->parent_index) {
2173 if (elm[-1].internal.mirror_tid <
2174 elm[0].internal.mirror_tid) {
2175 elm[-1].internal.mirror_tid =
2176 elm[0].internal.mirror_tid;
2179 if (elm[1].internal.mirror_tid <
2180 elm[0].internal.mirror_tid) {
2181 elm[1].internal.mirror_tid =
2182 elm[0].internal.mirror_tid;
2187 * Delete the subtree reference in the parent
2189 bcopy(&elm[1], &elm[0],
2190 (ondisk->count - cursor->parent_index) * esize);
2192 hammer_modify_node_done(parent);
2193 hammer_cursor_deleted_element(parent, cursor->parent_index);
2194 hammer_flush_node(node);
2195 hammer_delete_node(cursor->trans, node);
2198 * cursor->node is invalid, cursor up to make the cursor
2201 error = hammer_cursor_up(cursor);
2207 * Propagate cursor->trans->tid up the B-Tree starting at the current
2208 * cursor position using pseudofs info gleaned from the passed inode.
2210 * The passed inode has no relationship to the cursor position other
2211 * then being in the same pseudofs as the insertion or deletion we
2212 * are propagating the mirror_tid for.
2215 hammer_btree_do_propagation(hammer_cursor_t cursor,
2216 hammer_pseudofs_inmem_t pfsm,
2217 hammer_btree_leaf_elm_t leaf)
2219 hammer_cursor_t ncursor;
2220 hammer_tid_t mirror_tid;
2224 * We only propagate the mirror_tid up if we are in master or slave
2225 * mode. We do not bother if we are in no-mirror mode.
2227 * If pfsm is NULL we propagate (from mirror_write).
2230 pfsm->pfsd.master_id < 0 &&
2231 (pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE) == 0) {
2236 * This is a bit of a hack because we cannot deadlock or return
2237 * EDEADLK here. The related operation has already completed and
2238 * we must propagate the mirror_tid now regardless.
2240 * Generate a new cursor which inherits the original's locks and
2241 * unlock the original. Use the new cursor to propagate the
2242 * mirror_tid. Then clean up the new cursor and reacquire locks
2245 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2246 * original's locks and the original is tracked and must be
2249 mirror_tid = cursor->node->ondisk->mirror_tid;
2250 KKASSERT(mirror_tid != 0);
2251 ncursor = hammer_push_cursor(cursor);
2252 error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2253 KKASSERT(error == 0);
2254 hammer_pop_cursor(cursor, ncursor);
2259 * Propagate a mirror TID update upwards through the B-Tree to the root.
2261 * A locked internal node must be passed in. The node will remain locked
2264 * This function syncs mirror_tid at the specified internal node's element,
2265 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2268 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2270 hammer_btree_internal_elm_t elm;
2275 error = hammer_cursor_up(cursor);
2277 error = hammer_cursor_upgrade(cursor);
2278 while (error == EDEADLK) {
2279 hammer_recover_cursor(cursor);
2280 error = hammer_cursor_upgrade(cursor);
2284 node = cursor->node;
2285 KKASSERT (node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2288 * Adjust the node's element
2290 elm = &node->ondisk->elms[cursor->index].internal;
2291 if (elm->mirror_tid >= mirror_tid)
2293 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2294 sizeof(elm->mirror_tid));
2295 elm->mirror_tid = mirror_tid;
2296 hammer_modify_node_done(node);
2297 if (hammer_debug_general & 0x0002) {
2298 kprintf("mirror_propagate: propagate "
2299 "%016llx @%016llx:%d\n",
2300 mirror_tid, node->node_offset, cursor->index);
2305 * Adjust the node's mirror_tid aggregator
2307 if (node->ondisk->mirror_tid >= mirror_tid)
2309 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2310 node->ondisk->mirror_tid = mirror_tid;
2311 hammer_modify_node_done(node);
2312 if (hammer_debug_general & 0x0002) {
2313 kprintf("mirror_propagate: propagate "
2314 "%016llx @%016llx\n",
2315 mirror_tid, node->node_offset);
2318 if (error == ENOENT)
2324 hammer_btree_get_parent(hammer_node_t node, int *parent_indexp, int *errorp,
2327 hammer_node_t parent;
2328 hammer_btree_elm_t elm;
2334 parent = hammer_get_node(node->hmp, node->ondisk->parent, 0, errorp);
2336 KKASSERT(parent == NULL);
2339 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2344 if (try_exclusive) {
2345 if (hammer_lock_ex_try(&parent->lock)) {
2346 hammer_rel_node(parent);
2351 hammer_lock_sh(&parent->lock);
2355 * Figure out which element in the parent is pointing to the
2358 if (node->ondisk->count) {
2359 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2364 while (i < parent->ondisk->count) {
2365 elm = &parent->ondisk->elms[i];
2366 if (elm->internal.subtree_offset == node->node_offset)
2370 if (i == parent->ondisk->count) {
2371 hammer_unlock(&parent->lock);
2372 panic("Bad B-Tree link: parent %p node %p\n", parent, node);
2375 KKASSERT(*errorp == 0);
2380 * The element (elm) has been moved to a new internal node (node).
2382 * If the element represents a pointer to an internal node that node's
2383 * parent must be adjusted to the element's new location.
2385 * XXX deadlock potential here with our exclusive locks
2388 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2389 hammer_btree_elm_t elm)
2391 hammer_node_t child;
2396 switch(elm->base.btype) {
2397 case HAMMER_BTREE_TYPE_INTERNAL:
2398 case HAMMER_BTREE_TYPE_LEAF:
2399 child = hammer_get_node(node->hmp, elm->internal.subtree_offset,
2402 hammer_modify_node_field(trans, child, parent);
2403 child->ondisk->parent = node->node_offset;
2404 hammer_modify_node_done(child);
2405 hammer_rel_node(child);
2415 * Exclusively lock all the children of node. This is used by the split
2416 * code to prevent anyone from accessing the children of a cursor node
2417 * while we fix-up its parent offset.
2419 * If we don't lock the children we can really mess up cursors which block
2420 * trying to cursor-up into our node.
2422 * On failure EDEADLK (or some other error) is returned. If a deadlock
2423 * error is returned the cursor is adjusted to block on termination.
2426 hammer_btree_lock_children(hammer_cursor_t cursor,
2427 struct hammer_node_locklist **locklistp)
2430 hammer_node_locklist_t item;
2431 hammer_node_ondisk_t ondisk;
2432 hammer_btree_elm_t elm;
2433 hammer_node_t child;
2437 node = cursor->node;
2438 ondisk = node->ondisk;
2442 * We really do not want to block on I/O with exclusive locks held,
2443 * pre-get the children before trying to lock the mess.
2445 for (i = 0; i < ondisk->count; ++i) {
2446 ++hammer_stats_btree_elements;
2447 elm = &ondisk->elms[i];
2448 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2449 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2452 child = hammer_get_node(node->hmp,
2453 elm->internal.subtree_offset,
2456 hammer_rel_node(child);
2462 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2463 ++hammer_stats_btree_elements;
2464 elm = &ondisk->elms[i];
2466 switch(elm->base.btype) {
2467 case HAMMER_BTREE_TYPE_INTERNAL:
2468 case HAMMER_BTREE_TYPE_LEAF:
2469 KKASSERT(elm->internal.subtree_offset != 0);
2470 child = hammer_get_node(node->hmp,
2471 elm->internal.subtree_offset,
2479 if (hammer_lock_ex_try(&child->lock) != 0) {
2480 if (cursor->deadlk_node == NULL) {
2481 cursor->deadlk_node = child;
2482 hammer_ref_node(cursor->deadlk_node);
2485 hammer_rel_node(child);
2487 item = kmalloc(sizeof(*item),
2488 M_HAMMER, M_WAITOK);
2489 item->next = *locklistp;
2496 hammer_btree_unlock_children(locklistp);
2502 * Release previously obtained node locks.
2505 hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2507 hammer_node_locklist_t item;
2509 while ((item = *locklistp) != NULL) {
2510 *locklistp = item->next;
2511 hammer_unlock(&item->node->lock);
2512 hammer_rel_node(item->node);
2513 kfree(item, M_HAMMER);
2517 /************************************************************************
2518 * MISCELLANIOUS SUPPORT *
2519 ************************************************************************/
2522 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2524 * Note that for this particular function a return value of -1, 0, or +1
2525 * can denote a match if create_tid is otherwise discounted. A create_tid
2526 * of zero is considered to be 'infinity' in comparisons.
2528 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2531 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2533 if (key1->localization < key2->localization)
2535 if (key1->localization > key2->localization)
2538 if (key1->obj_id < key2->obj_id)
2540 if (key1->obj_id > key2->obj_id)
2543 if (key1->rec_type < key2->rec_type)
2545 if (key1->rec_type > key2->rec_type)
2548 if (key1->key < key2->key)
2550 if (key1->key > key2->key)
2554 * A create_tid of zero indicates a record which is undeletable
2555 * and must be considered to have a value of positive infinity.
2557 if (key1->create_tid == 0) {
2558 if (key2->create_tid == 0)
2562 if (key2->create_tid == 0)
2564 if (key1->create_tid < key2->create_tid)
2566 if (key1->create_tid > key2->create_tid)
2572 * Test a timestamp against an element to determine whether the
2573 * element is visible. A timestamp of 0 means 'infinity'.
2576 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2579 if (base->delete_tid)
2583 if (asof < base->create_tid)
2585 if (base->delete_tid && asof >= base->delete_tid)
2591 * Create a separator half way inbetween key1 and key2. For fields just
2592 * one unit apart, the separator will match key2. key1 is on the left-hand
2593 * side and key2 is on the right-hand side.
2595 * key2 must be >= the separator. It is ok for the separator to match key2.
2597 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2600 * NOTE: It might be beneficial to just scrap this whole mess and just
2601 * set the separator to key2.
2603 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2604 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2607 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2608 hammer_base_elm_t dest)
2610 bzero(dest, sizeof(*dest));
2612 dest->rec_type = key2->rec_type;
2613 dest->key = key2->key;
2614 dest->obj_id = key2->obj_id;
2615 dest->create_tid = key2->create_tid;
2617 MAKE_SEPARATOR(key1, key2, dest, localization);
2618 if (key1->localization == key2->localization) {
2619 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2620 if (key1->obj_id == key2->obj_id) {
2621 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2622 if (key1->rec_type == key2->rec_type) {
2623 MAKE_SEPARATOR(key1, key2, dest, key);
2625 * Don't bother creating a separator for
2626 * create_tid, which also conveniently avoids
2627 * having to handle the create_tid == 0
2628 * (infinity) case. Just leave create_tid
2631 * Worst case, dest matches key2 exactly,
2632 * which is acceptable.
2639 #undef MAKE_SEPARATOR
2642 * Return whether a generic internal or leaf node is full
2645 btree_node_is_full(hammer_node_ondisk_t node)
2647 switch(node->type) {
2648 case HAMMER_BTREE_TYPE_INTERNAL:
2649 if (node->count == HAMMER_BTREE_INT_ELMS)
2652 case HAMMER_BTREE_TYPE_LEAF:
2653 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2657 panic("illegal btree subtype");
2664 btree_max_elements(u_int8_t type)
2666 if (type == HAMMER_BTREE_TYPE_LEAF)
2667 return(HAMMER_BTREE_LEAF_ELMS);
2668 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2669 return(HAMMER_BTREE_INT_ELMS);
2670 panic("btree_max_elements: bad type %d\n", type);
2675 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2677 hammer_btree_elm_t elm;
2680 kprintf("node %p count=%d parent=%016llx type=%c\n",
2681 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2684 * Dump both boundary elements if an internal node
2686 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2687 for (i = 0; i <= ondisk->count; ++i) {
2688 elm = &ondisk->elms[i];
2689 hammer_print_btree_elm(elm, ondisk->type, i);
2692 for (i = 0; i < ondisk->count; ++i) {
2693 elm = &ondisk->elms[i];
2694 hammer_print_btree_elm(elm, ondisk->type, i);
2700 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2703 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
2704 kprintf("\tkey = %016llx\n", elm->base.key);
2705 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2706 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2707 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2708 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2709 kprintf("\tbtype = %02x (%c)\n",
2711 (elm->base.btype ? elm->base.btype : '?'));
2712 kprintf("\tlocalization = %02x\n", elm->base.localization);
2715 case HAMMER_BTREE_TYPE_INTERNAL:
2716 kprintf("\tsubtree_off = %016llx\n",
2717 elm->internal.subtree_offset);
2719 case HAMMER_BTREE_TYPE_RECORD:
2720 kprintf("\tdata_offset = %016llx\n", elm->leaf.data_offset);
2721 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2722 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);