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 1.60 2008/07/01 16:48:51 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 void hammer_make_separator(hammer_base_elm_t key1
,
91 hammer_base_elm_t key2
, hammer_base_elm_t dest
);
94 * Iterate records after a search. The cursor is iterated forwards past
95 * the current record until a record matching the key-range requirements
96 * is found. ENOENT is returned if the iteration goes past the ending
99 * The iteration is inclusive of key_beg and can be inclusive or exclusive
100 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
102 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
103 * may be modified by B-Tree functions.
105 * cursor->key_beg may or may not be modified by this function during
106 * the iteration. XXX future - in case of an inverted lock we may have
107 * to reinitiate the lookup and set key_beg to properly pick up where we
110 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
113 hammer_btree_iterate(hammer_cursor_t cursor
)
115 hammer_node_ondisk_t node
;
116 hammer_btree_elm_t elm
;
122 * Skip past the current record
124 node
= cursor
->node
->ondisk
;
127 if (cursor
->index
< node
->count
&&
128 (cursor
->flags
& HAMMER_CURSOR_ATEDISK
)) {
133 * Loop until an element is found or we are done.
137 * We iterate up the tree and then index over one element
138 * while we are at the last element in the current node.
140 * If we are at the root of the filesystem, cursor_up
143 * XXX this could be optimized by storing the information in
144 * the parent reference.
146 * XXX we can lose the node lock temporarily, this could mess
149 ++hammer_stats_btree_iterations
;
150 hammer_flusher_clean_loose_ios(cursor
->trans
->hmp
);
152 if (cursor
->index
== node
->count
) {
153 if (hammer_debug_btree
) {
154 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
155 cursor
->node
->node_offset
,
157 (cursor
->parent
? cursor
->parent
->node_offset
: -1),
158 cursor
->parent_index
,
161 KKASSERT(cursor
->parent
== NULL
|| cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.subtree_offset
== cursor
->node
->node_offset
);
162 error
= hammer_cursor_up(cursor
);
165 /* reload stale pointer */
166 node
= cursor
->node
->ondisk
;
167 KKASSERT(cursor
->index
!= node
->count
);
170 * If we are reblocking we want to return internal
173 if (cursor
->flags
& HAMMER_CURSOR_REBLOCKING
) {
174 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
182 * Check internal or leaf element. Determine if the record
183 * at the cursor has gone beyond the end of our range.
185 * We recurse down through internal nodes.
187 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
188 elm
= &node
->elms
[cursor
->index
];
190 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
191 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
192 if (hammer_debug_btree
) {
193 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
194 cursor
->node
->node_offset
,
196 elm
[0].internal
.base
.obj_id
,
197 elm
[0].internal
.base
.rec_type
,
198 elm
[0].internal
.base
.key
,
199 elm
[0].internal
.base
.localization
,
203 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
204 cursor
->node
->node_offset
,
206 elm
[1].internal
.base
.obj_id
,
207 elm
[1].internal
.base
.rec_type
,
208 elm
[1].internal
.base
.key
,
209 elm
[1].internal
.base
.localization
,
218 if (r
== 0 && (cursor
->flags
&
219 HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
228 KKASSERT(elm
->internal
.subtree_offset
!= 0);
231 * If running the mirror filter see if we can skip
232 * the entire sub-tree.
234 if (cursor
->flags
& HAMMER_CURSOR_MIRROR_FILTERED
) {
235 if (elm
->internal
.mirror_tid
<
236 cursor
->mirror_tid
) {
242 error
= hammer_cursor_down(cursor
);
245 KKASSERT(cursor
->index
== 0);
246 /* reload stale pointer */
247 node
= cursor
->node
->ondisk
;
250 elm
= &node
->elms
[cursor
->index
];
251 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
->base
);
252 if (hammer_debug_btree
) {
253 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
254 cursor
->node
->node_offset
,
256 (elm
[0].leaf
.base
.btype
?
257 elm
[0].leaf
.base
.btype
: '?'),
258 elm
[0].leaf
.base
.obj_id
,
259 elm
[0].leaf
.base
.rec_type
,
260 elm
[0].leaf
.base
.key
,
261 elm
[0].leaf
.base
.localization
,
271 * We support both end-inclusive and
272 * end-exclusive searches.
275 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
280 switch(elm
->leaf
.base
.btype
) {
281 case HAMMER_BTREE_TYPE_RECORD
:
282 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
283 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
296 * node pointer invalid after loop
302 if (hammer_debug_btree
) {
303 int i
= cursor
->index
;
304 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
305 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
307 elm
->internal
.base
.obj_id
,
308 elm
->internal
.base
.rec_type
,
309 elm
->internal
.base
.key
,
310 elm
->internal
.base
.localization
319 * Iterate in the reverse direction. This is used by the pruning code to
320 * avoid overlapping records.
323 hammer_btree_iterate_reverse(hammer_cursor_t cursor
)
325 hammer_node_ondisk_t node
;
326 hammer_btree_elm_t elm
;
332 * Skip past the current record. For various reasons the cursor
333 * may end up set to -1 or set to point at the end of the current
334 * node. These cases must be addressed.
336 node
= cursor
->node
->ondisk
;
339 if (cursor
->index
!= -1 &&
340 (cursor
->flags
& HAMMER_CURSOR_ATEDISK
)) {
343 if (cursor
->index
== cursor
->node
->ondisk
->count
)
347 * Loop until an element is found or we are done.
350 ++hammer_stats_btree_iterations
;
351 hammer_flusher_clean_loose_ios(cursor
->trans
->hmp
);
354 * We iterate up the tree and then index over one element
355 * while we are at the last element in the current node.
357 if (cursor
->index
== -1) {
358 error
= hammer_cursor_up(cursor
);
360 cursor
->index
= 0; /* sanity */
363 /* reload stale pointer */
364 node
= cursor
->node
->ondisk
;
365 KKASSERT(cursor
->index
!= node
->count
);
371 * Check internal or leaf element. Determine if the record
372 * at the cursor has gone beyond the end of our range.
374 * We recurse down through internal nodes.
376 KKASSERT(cursor
->index
!= node
->count
);
377 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
378 elm
= &node
->elms
[cursor
->index
];
379 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
380 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
381 if (hammer_debug_btree
) {
382 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
383 cursor
->node
->node_offset
,
385 elm
[0].internal
.base
.obj_id
,
386 elm
[0].internal
.base
.rec_type
,
387 elm
[0].internal
.base
.key
,
388 elm
[0].internal
.base
.localization
,
391 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
392 cursor
->node
->node_offset
,
394 elm
[1].internal
.base
.obj_id
,
395 elm
[1].internal
.base
.rec_type
,
396 elm
[1].internal
.base
.key
,
397 elm
[1].internal
.base
.localization
,
411 KKASSERT(elm
->internal
.subtree_offset
!= 0);
413 error
= hammer_cursor_down(cursor
);
416 KKASSERT(cursor
->index
== 0);
417 /* reload stale pointer */
418 node
= cursor
->node
->ondisk
;
420 /* this can assign -1 if the leaf was empty */
421 cursor
->index
= node
->count
- 1;
424 elm
= &node
->elms
[cursor
->index
];
425 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
426 if (hammer_debug_btree
) {
427 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
428 cursor
->node
->node_offset
,
430 (elm
[0].leaf
.base
.btype
?
431 elm
[0].leaf
.base
.btype
: '?'),
432 elm
[0].leaf
.base
.obj_id
,
433 elm
[0].leaf
.base
.rec_type
,
434 elm
[0].leaf
.base
.key
,
435 elm
[0].leaf
.base
.localization
,
444 switch(elm
->leaf
.base
.btype
) {
445 case HAMMER_BTREE_TYPE_RECORD
:
446 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
447 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
460 * node pointer invalid after loop
466 if (hammer_debug_btree
) {
467 int i
= cursor
->index
;
468 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
469 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
471 elm
->internal
.base
.obj_id
,
472 elm
->internal
.base
.rec_type
,
473 elm
->internal
.base
.key
,
474 elm
->internal
.base
.localization
483 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
484 * could not be found, EDEADLK if inserting and a retry is needed, and a
485 * fatal error otherwise. When retrying, the caller must terminate the
486 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
488 * The cursor is suitably positioned for a deletion on success, and suitably
489 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
492 * The cursor may begin anywhere, the search will traverse the tree in
493 * either direction to locate the requested element.
495 * Most of the logic implementing historical searches is handled here. We
496 * do an initial lookup with create_tid set to the asof TID. Due to the
497 * way records are laid out, a backwards iteration may be required if
498 * ENOENT is returned to locate the historical record. Here's the
501 * create_tid: 10 15 20
505 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
506 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
507 * not visible and thus causes ENOENT to be returned. We really need
508 * to check record 11 in LEAF1. If it also fails then the search fails
509 * (e.g. it might represent the range 11-16 and thus still not match our
510 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
511 * further iterations.
513 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
514 * and the cursor->create_check TID if an iteration might be needed.
515 * In the above example create_check would be set to 14.
518 hammer_btree_lookup(hammer_cursor_t cursor
)
522 ++hammer_stats_btree_lookups
;
523 if (cursor
->flags
& HAMMER_CURSOR_ASOF
) {
524 KKASSERT((cursor
->flags
& HAMMER_CURSOR_INSERT
) == 0);
525 cursor
->key_beg
.create_tid
= cursor
->asof
;
527 cursor
->flags
&= ~HAMMER_CURSOR_CREATE_CHECK
;
528 error
= btree_search(cursor
, 0);
529 if (error
!= ENOENT
||
530 (cursor
->flags
& HAMMER_CURSOR_CREATE_CHECK
) == 0) {
533 * Stop if error other then ENOENT.
534 * Stop if ENOENT and not special case.
538 if (hammer_debug_btree
) {
539 kprintf("CREATE_CHECK %016llx\n",
540 cursor
->create_check
);
542 cursor
->key_beg
.create_tid
= cursor
->create_check
;
546 error
= btree_search(cursor
, 0);
549 error
= hammer_btree_extract(cursor
, cursor
->flags
);
554 * Execute the logic required to start an iteration. The first record
555 * located within the specified range is returned and iteration control
556 * flags are adjusted for successive hammer_btree_iterate() calls.
559 hammer_btree_first(hammer_cursor_t cursor
)
563 error
= hammer_btree_lookup(cursor
);
564 if (error
== ENOENT
) {
565 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
566 error
= hammer_btree_iterate(cursor
);
568 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
573 * Similarly but for an iteration in the reverse direction.
575 * Set ATEDISK when iterating backwards to skip the current entry,
576 * which after an ENOENT lookup will be pointing beyond our end point.
579 hammer_btree_last(hammer_cursor_t cursor
)
581 struct hammer_base_elm save
;
584 save
= cursor
->key_beg
;
585 cursor
->key_beg
= cursor
->key_end
;
586 error
= hammer_btree_lookup(cursor
);
587 cursor
->key_beg
= save
;
588 if (error
== ENOENT
||
589 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
590 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
591 error
= hammer_btree_iterate_reverse(cursor
);
593 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
598 * Extract the record and/or data associated with the cursor's current
599 * position. Any prior record or data stored in the cursor is replaced.
600 * The cursor must be positioned at a leaf node.
602 * NOTE: All extractions occur at the leaf of the B-Tree.
605 hammer_btree_extract(hammer_cursor_t cursor
, int flags
)
608 hammer_node_ondisk_t node
;
609 hammer_btree_elm_t elm
;
610 hammer_off_t data_off
;
615 * The case where the data reference resolves to the same buffer
616 * as the record reference must be handled.
618 node
= cursor
->node
->ondisk
;
619 elm
= &node
->elms
[cursor
->index
];
621 hmp
= cursor
->node
->hmp
;
624 * There is nothing to extract for an internal element.
626 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
)
630 * Only record types have data.
632 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
633 cursor
->leaf
= &elm
->leaf
;
635 if ((flags
& HAMMER_CURSOR_GET_DATA
) == 0)
637 if (elm
->leaf
.base
.btype
!= HAMMER_BTREE_TYPE_RECORD
)
639 data_off
= elm
->leaf
.data_offset
;
640 data_len
= elm
->leaf
.data_len
;
647 KKASSERT(data_len
>= 0 && data_len
<= HAMMER_XBUFSIZE
);
648 cursor
->data
= hammer_bread_ext(hmp
, data_off
, data_len
,
649 &error
, &cursor
->data_buffer
);
650 if (hammer_crc_test_leaf(cursor
->data
, &elm
->leaf
) == 0)
651 Debugger("CRC FAILED: DATA");
657 * Insert a leaf element into the B-Tree at the current cursor position.
658 * The cursor is positioned such that the element at and beyond the cursor
659 * are shifted to make room for the new record.
661 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
662 * flag set and that call must return ENOENT before this function can be
665 * The caller may depend on the cursor's exclusive lock after return to
666 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
668 * ENOSPC is returned if there is no room to insert a new record.
671 hammer_btree_insert(hammer_cursor_t cursor
, hammer_btree_leaf_elm_t elm
)
673 hammer_node_ondisk_t node
;
677 if ((error
= hammer_cursor_upgrade_node(cursor
)) != 0)
679 ++hammer_stats_btree_inserts
;
682 * Insert the element at the leaf node and update the count in the
683 * parent. It is possible for parent to be NULL, indicating that
684 * the filesystem's ROOT B-Tree node is a leaf itself, which is
685 * possible. The root inode can never be deleted so the leaf should
688 * Remember that the right-hand boundary is not included in the
691 hammer_modify_node_all(cursor
->trans
, cursor
->node
);
692 node
= cursor
->node
->ondisk
;
694 KKASSERT(elm
->base
.btype
!= 0);
695 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
696 KKASSERT(node
->count
< HAMMER_BTREE_LEAF_ELMS
);
697 if (i
!= node
->count
) {
698 bcopy(&node
->elms
[i
], &node
->elms
[i
+1],
699 (node
->count
- i
) * sizeof(*elm
));
701 node
->elms
[i
].leaf
= *elm
;
705 * Update the leaf node's aggregate mirror_tid for mirroring
708 if (node
->mirror_tid
< elm
->base
.delete_tid
)
709 node
->mirror_tid
= elm
->base
.delete_tid
;
710 if (node
->mirror_tid
< elm
->base
.create_tid
)
711 node
->mirror_tid
= elm
->base
.create_tid
;
712 hammer_modify_node_done(cursor
->node
);
715 * What we really want to do is propogate mirror_tid all the way
716 * up the parent chain to the B-Tree root. That would be
717 * ultra-expensive, though.
719 if (cursor
->parent
&&
720 (cursor
->trans
->hmp
->hflags
& (HMNT_MASTERID
|HMNT_SLAVE
))) {
721 hammer_btree_mirror_propagate(cursor
->trans
, cursor
->parent
,
722 cursor
->parent_index
,
727 * Debugging sanity checks.
729 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->base
) <= 0);
730 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->base
) > 0);
732 KKASSERT(hammer_btree_cmp(&node
->elms
[i
-1].leaf
.base
, &elm
->base
) < 0);
734 if (i
!= node
->count
- 1)
735 KKASSERT(hammer_btree_cmp(&node
->elms
[i
+1].leaf
.base
, &elm
->base
) > 0);
741 * Delete a record from the B-Tree at the current cursor position.
742 * The cursor is positioned such that the current element is the one
745 * On return the cursor will be positioned after the deleted element and
746 * MAY point to an internal node. It will be suitable for the continuation
747 * of an iteration but not for an insertion or deletion.
749 * Deletions will attempt to partially rebalance the B-Tree in an upward
750 * direction, but will terminate rather then deadlock. Empty internal nodes
751 * are never allowed by a deletion which deadlocks may end up giving us an
752 * empty leaf. The pruner will clean up and rebalance the tree.
754 * This function can return EDEADLK, requiring the caller to retry the
755 * operation after clearing the deadlock.
758 hammer_btree_delete(hammer_cursor_t cursor
)
760 hammer_node_ondisk_t ondisk
;
762 hammer_node_t parent
;
766 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
768 ++hammer_stats_btree_deletes
;
771 * Delete the element from the leaf node.
773 * Remember that leaf nodes do not have boundaries.
776 ondisk
= node
->ondisk
;
779 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_LEAF
);
780 KKASSERT(i
>= 0 && i
< ondisk
->count
);
781 hammer_modify_node_all(cursor
->trans
, node
);
782 if (i
+ 1 != ondisk
->count
) {
783 bcopy(&ondisk
->elms
[i
+1], &ondisk
->elms
[i
],
784 (ondisk
->count
- i
- 1) * sizeof(ondisk
->elms
[0]));
787 hammer_modify_node_done(node
);
790 * Validate local parent
792 if (ondisk
->parent
) {
793 parent
= cursor
->parent
;
795 KKASSERT(parent
!= NULL
);
796 KKASSERT(parent
->node_offset
== ondisk
->parent
);
800 * If the leaf becomes empty it must be detached from the parent,
801 * potentially recursing through to the filesystem root.
803 * This may reposition the cursor at one of the parent's of the
806 * Ignore deadlock errors, that simply means that btree_remove
807 * was unable to recurse and had to leave us with an empty leaf.
809 KKASSERT(cursor
->index
<= ondisk
->count
);
810 if (ondisk
->count
== 0) {
811 error
= btree_remove(cursor
);
812 if (error
== EDEADLK
)
817 KKASSERT(cursor
->parent
== NULL
||
818 cursor
->parent_index
< cursor
->parent
->ondisk
->count
);
823 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
825 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
827 * The search can begin ANYWHERE in the B-Tree. As a first step the search
828 * iterates up the tree as necessary to properly position itself prior to
829 * actually doing the sarch.
831 * INSERTIONS: The search will split full nodes and leaves on its way down
832 * and guarentee that the leaf it ends up on is not full. If we run out
833 * of space the search continues to the leaf (to position the cursor for
834 * the spike), but ENOSPC is returned.
836 * The search is only guarenteed to end up on a leaf if an error code of 0
837 * is returned, or if inserting and an error code of ENOENT is returned.
838 * Otherwise it can stop at an internal node. On success a search returns
841 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
842 * filesystem, and it is not simple code. Please note the following facts:
844 * - Internal node recursions have a boundary on the left AND right. The
845 * right boundary is non-inclusive. The create_tid is a generic part
846 * of the key for internal nodes.
848 * - Leaf nodes contain terminal elements only now.
850 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
851 * historical search. ASOF and INSERT are mutually exclusive. When
852 * doing an as-of lookup btree_search() checks for a right-edge boundary
853 * case. If while recursing down the left-edge differs from the key
854 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
855 * with cursor->create_check. This is used by btree_lookup() to iterate.
856 * The iteration backwards because as-of searches can wind up going
857 * down the wrong branch of the B-Tree.
861 btree_search(hammer_cursor_t cursor
, int flags
)
863 hammer_node_ondisk_t node
;
864 hammer_btree_elm_t elm
;
871 flags
|= cursor
->flags
;
872 ++hammer_stats_btree_searches
;
874 if (hammer_debug_btree
) {
875 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
876 cursor
->node
->node_offset
,
878 cursor
->key_beg
.obj_id
,
879 cursor
->key_beg
.rec_type
,
881 cursor
->key_beg
.create_tid
,
882 cursor
->key_beg
.localization
,
886 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
887 cursor
->parent
->node_offset
, cursor
->parent_index
,
888 cursor
->left_bound
->obj_id
,
889 cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.base
.obj_id
,
890 cursor
->right_bound
->obj_id
,
891 cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1].internal
.base
.obj_id
,
893 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
],
895 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1]
900 * Move our cursor up the tree until we find a node whos range covers
901 * the key we are trying to locate.
903 * The left bound is inclusive, the right bound is non-inclusive.
904 * It is ok to cursor up too far.
907 r
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->left_bound
);
908 s
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->right_bound
);
911 KKASSERT(cursor
->parent
);
912 ++hammer_stats_btree_iterations
;
913 error
= hammer_cursor_up(cursor
);
919 * The delete-checks below are based on node, not parent. Set the
920 * initial delete-check based on the parent.
923 KKASSERT(cursor
->left_bound
->create_tid
!= 1);
924 cursor
->create_check
= cursor
->left_bound
->create_tid
- 1;
925 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
929 * We better have ended up with a node somewhere.
931 KKASSERT(cursor
->node
!= NULL
);
934 * If we are inserting we can't start at a full node if the parent
935 * is also full (because there is no way to split the node),
936 * continue running up the tree until the requirement is satisfied
937 * or we hit the root of the filesystem.
939 * (If inserting we aren't doing an as-of search so we don't have
940 * to worry about create_check).
942 while ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
943 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
944 if (btree_node_is_full(cursor
->node
->ondisk
) == 0)
947 if (btree_node_is_full(cursor
->node
->ondisk
) ==0)
950 if (cursor
->node
->ondisk
->parent
== 0 ||
951 cursor
->parent
->ondisk
->count
!= HAMMER_BTREE_INT_ELMS
) {
954 ++hammer_stats_btree_iterations
;
955 error
= hammer_cursor_up(cursor
);
956 /* node may have become stale */
962 * Push down through internal nodes to locate the requested key.
964 node
= cursor
->node
->ondisk
;
965 while (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
967 * Scan the node to find the subtree index to push down into.
968 * We go one-past, then back-up.
970 * We must proactively remove deleted elements which may
971 * have been left over from a deadlocked btree_remove().
973 * The left and right boundaries are included in the loop
974 * in order to detect edge cases.
976 * If the separator only differs by create_tid (r == 1)
977 * and we are doing an as-of search, we may end up going
978 * down a branch to the left of the one containing the
979 * desired key. This requires numerous special cases.
981 ++hammer_stats_btree_iterations
;
982 if (hammer_debug_btree
) {
983 kprintf("SEARCH-I %016llx count=%d\n",
984 cursor
->node
->node_offset
,
989 * Try to shortcut the search before dropping into the
990 * linear loop. Locate the first node where r <= 1.
992 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
993 while (i
<= node
->count
) {
994 ++hammer_stats_btree_elements
;
995 elm
= &node
->elms
[i
];
996 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
997 if (hammer_debug_btree
> 2) {
998 kprintf(" IELM %p %d r=%d\n",
999 &node
->elms
[i
], i
, r
);
1004 KKASSERT(elm
->base
.create_tid
!= 1);
1005 cursor
->create_check
= elm
->base
.create_tid
- 1;
1006 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
1010 if (hammer_debug_btree
) {
1011 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1016 * These cases occur when the parent's idea of the boundary
1017 * is wider then the child's idea of the boundary, and
1018 * require special handling. If not inserting we can
1019 * terminate the search early for these cases but the
1020 * child's boundaries cannot be unconditionally modified.
1024 * If i == 0 the search terminated to the LEFT of the
1025 * left_boundary but to the RIGHT of the parent's left
1030 elm
= &node
->elms
[0];
1033 * If we aren't inserting we can stop here.
1035 if ((flags
& (HAMMER_CURSOR_INSERT
|
1036 HAMMER_CURSOR_PRUNING
)) == 0) {
1042 * Correct a left-hand boundary mismatch.
1044 * We can only do this if we can upgrade the lock,
1045 * and synchronized as a background cursor (i.e.
1046 * inserting or pruning).
1048 * WARNING: We can only do this if inserting, i.e.
1049 * we are running on the backend.
1051 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1053 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1054 hammer_modify_node_field(cursor
->trans
, cursor
->node
,
1056 save
= node
->elms
[0].base
.btype
;
1057 node
->elms
[0].base
= *cursor
->left_bound
;
1058 node
->elms
[0].base
.btype
= save
;
1059 hammer_modify_node_done(cursor
->node
);
1060 } else if (i
== node
->count
+ 1) {
1062 * If i == node->count + 1 the search terminated to
1063 * the RIGHT of the right boundary but to the LEFT
1064 * of the parent's right boundary. If we aren't
1065 * inserting we can stop here.
1067 * Note that the last element in this case is
1068 * elms[i-2] prior to adjustments to 'i'.
1071 if ((flags
& (HAMMER_CURSOR_INSERT
|
1072 HAMMER_CURSOR_PRUNING
)) == 0) {
1078 * Correct a right-hand boundary mismatch.
1079 * (actual push-down record is i-2 prior to
1080 * adjustments to i).
1082 * We can only do this if we can upgrade the lock,
1083 * and synchronized as a background cursor (i.e.
1084 * inserting or pruning).
1086 * WARNING: We can only do this if inserting, i.e.
1087 * we are running on the backend.
1089 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1091 elm
= &node
->elms
[i
];
1092 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1093 hammer_modify_node(cursor
->trans
, cursor
->node
,
1094 &elm
->base
, sizeof(elm
->base
));
1095 elm
->base
= *cursor
->right_bound
;
1096 hammer_modify_node_done(cursor
->node
);
1100 * The push-down index is now i - 1. If we had
1101 * terminated on the right boundary this will point
1102 * us at the last element.
1107 elm
= &node
->elms
[i
];
1109 if (hammer_debug_btree
) {
1110 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1111 "key=%016llx cre=%016llx lo=%02x\n",
1112 cursor
->node
->node_offset
,
1114 elm
->internal
.base
.obj_id
,
1115 elm
->internal
.base
.rec_type
,
1116 elm
->internal
.base
.key
,
1117 elm
->internal
.base
.create_tid
,
1118 elm
->internal
.base
.localization
1123 * We better have a valid subtree offset.
1125 KKASSERT(elm
->internal
.subtree_offset
!= 0);
1128 * Handle insertion and deletion requirements.
1130 * If inserting split full nodes. The split code will
1131 * adjust cursor->node and cursor->index if the current
1132 * index winds up in the new node.
1134 * If inserting and a left or right edge case was detected,
1135 * we cannot correct the left or right boundary and must
1136 * prepend and append an empty leaf node in order to make
1137 * the boundary correction.
1139 * If we run out of space we set enospc and continue on
1140 * to a leaf to provide the spike code with a good point
1143 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
1144 if (btree_node_is_full(node
)) {
1145 error
= btree_split_internal(cursor
);
1147 if (error
!= ENOSPC
)
1152 * reload stale pointers
1155 node
= cursor
->node
->ondisk
;
1160 * Push down (push into new node, existing node becomes
1161 * the parent) and continue the search.
1163 error
= hammer_cursor_down(cursor
);
1164 /* node may have become stale */
1167 node
= cursor
->node
->ondisk
;
1171 * We are at a leaf, do a linear search of the key array.
1173 * On success the index is set to the matching element and 0
1176 * On failure the index is set to the insertion point and ENOENT
1179 * Boundaries are not stored in leaf nodes, so the index can wind
1180 * up to the left of element 0 (index == 0) or past the end of
1181 * the array (index == node->count). It is also possible that the
1182 * leaf might be empty.
1184 ++hammer_stats_btree_iterations
;
1185 KKASSERT (node
->type
== HAMMER_BTREE_TYPE_LEAF
);
1186 KKASSERT(node
->count
<= HAMMER_BTREE_LEAF_ELMS
);
1187 if (hammer_debug_btree
) {
1188 kprintf("SEARCH-L %016llx count=%d\n",
1189 cursor
->node
->node_offset
,
1194 * Try to shortcut the search before dropping into the
1195 * linear loop. Locate the first node where r <= 1.
1197 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
1198 while (i
< node
->count
) {
1199 ++hammer_stats_btree_elements
;
1200 elm
= &node
->elms
[i
];
1202 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->leaf
.base
);
1204 if (hammer_debug_btree
> 1)
1205 kprintf(" ELM %p %d r=%d\n", &node
->elms
[i
], i
, r
);
1208 * We are at a record element. Stop if we've flipped past
1209 * key_beg, not counting the create_tid test. Allow the
1210 * r == 1 case (key_beg > element but differs only by its
1211 * create_tid) to fall through to the AS-OF check.
1213 KKASSERT (elm
->leaf
.base
.btype
== HAMMER_BTREE_TYPE_RECORD
);
1223 * Check our as-of timestamp against the element.
1225 if (flags
& HAMMER_CURSOR_ASOF
) {
1226 if (hammer_btree_chkts(cursor
->asof
,
1227 &node
->elms
[i
].base
) != 0) {
1233 if (r
> 0) { /* can only be +1 */
1241 if (hammer_debug_btree
) {
1242 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1243 cursor
->node
->node_offset
, i
);
1249 * The search of the leaf node failed. i is the insertion point.
1252 if (hammer_debug_btree
) {
1253 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1254 cursor
->node
->node_offset
, i
);
1258 * No exact match was found, i is now at the insertion point.
1260 * If inserting split a full leaf before returning. This
1261 * may have the side effect of adjusting cursor->node and
1265 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0 &&
1266 btree_node_is_full(node
)) {
1267 error
= btree_split_leaf(cursor
);
1269 if (error
!= ENOSPC
)
1274 * reload stale pointers
1278 node = &cursor->node->internal;
1283 * We reached a leaf but did not find the key we were looking for.
1284 * If this is an insert we will be properly positioned for an insert
1285 * (ENOENT) or spike (ENOSPC) operation.
1287 error
= enospc
? ENOSPC
: ENOENT
;
1293 * Heuristical search for the first element whos comparison is <= 1. May
1294 * return an index whos compare result is > 1 but may only return an index
1295 * whos compare result is <= 1 if it is the first element with that result.
1298 hammer_btree_search_node(hammer_base_elm_t elm
, hammer_node_ondisk_t node
)
1306 * Don't bother if the node does not have very many elements
1311 i
= b
+ (s
- b
) / 2;
1312 ++hammer_stats_btree_elements
;
1313 r
= hammer_btree_cmp(elm
, &node
->elms
[i
].leaf
.base
);
1324 /************************************************************************
1325 * SPLITTING AND MERGING *
1326 ************************************************************************
1328 * These routines do all the dirty work required to split and merge nodes.
1332 * Split an internal node into two nodes and move the separator at the split
1333 * point to the parent.
1335 * (cursor->node, cursor->index) indicates the element the caller intends
1336 * to push into. We will adjust node and index if that element winds
1337 * up in the split node.
1339 * If we are at the root of the filesystem a new root must be created with
1340 * two elements, one pointing to the original root and one pointing to the
1341 * newly allocated split node.
1345 btree_split_internal(hammer_cursor_t cursor
)
1347 hammer_node_ondisk_t ondisk
;
1349 hammer_node_t parent
;
1350 hammer_node_t new_node
;
1351 hammer_btree_elm_t elm
;
1352 hammer_btree_elm_t parent_elm
;
1353 hammer_node_locklist_t locklist
= NULL
;
1354 hammer_mount_t hmp
= cursor
->trans
->hmp
;
1360 const int esize
= sizeof(*elm
);
1362 error
= hammer_btree_lock_children(cursor
, &locklist
);
1365 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1367 ++hammer_stats_btree_splits
;
1370 * We are splitting but elms[split] will be promoted to the parent,
1371 * leaving the right hand node with one less element. If the
1372 * insertion point will be on the left-hand side adjust the split
1373 * point to give the right hand side one additional node.
1375 node
= cursor
->node
;
1376 ondisk
= node
->ondisk
;
1377 split
= (ondisk
->count
+ 1) / 2;
1378 if (cursor
->index
<= split
)
1382 * If we are at the root of the filesystem, create a new root node
1383 * with 1 element and split normally. Avoid making major
1384 * modifications until we know the whole operation will work.
1386 if (ondisk
->parent
== 0) {
1387 parent
= hammer_alloc_btree(cursor
->trans
, &error
);
1390 hammer_lock_ex(&parent
->lock
);
1391 hammer_modify_node_noundo(cursor
->trans
, parent
);
1392 ondisk
= parent
->ondisk
;
1395 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1396 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1397 ondisk
->elms
[0].base
.btype
= node
->ondisk
->type
;
1398 ondisk
->elms
[0].internal
.subtree_offset
= node
->node_offset
;
1399 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1400 hammer_modify_node_done(parent
);
1401 /* ondisk->elms[1].base.btype - not used */
1403 parent_index
= 0; /* index of current node in parent */
1406 parent
= cursor
->parent
;
1407 parent_index
= cursor
->parent_index
;
1411 * Split node into new_node at the split point.
1413 * B O O O P N N B <-- P = node->elms[split]
1414 * 0 1 2 3 4 5 6 <-- subtree indices
1419 * B O O O B B N N B <--- inner boundary points are 'P'
1423 new_node
= hammer_alloc_btree(cursor
->trans
, &error
);
1424 if (new_node
== NULL
) {
1426 hammer_unlock(&parent
->lock
);
1427 hammer_delete_node(cursor
->trans
, parent
);
1428 hammer_rel_node(parent
);
1432 hammer_lock_ex(&new_node
->lock
);
1435 * Create the new node. P becomes the left-hand boundary in the
1436 * new node. Copy the right-hand boundary as well.
1438 * elm is the new separator.
1440 hammer_modify_node_noundo(cursor
->trans
, new_node
);
1441 hammer_modify_node_all(cursor
->trans
, node
);
1442 ondisk
= node
->ondisk
;
1443 elm
= &ondisk
->elms
[split
];
1444 bcopy(elm
, &new_node
->ondisk
->elms
[0],
1445 (ondisk
->count
- split
+ 1) * esize
);
1446 new_node
->ondisk
->count
= ondisk
->count
- split
;
1447 new_node
->ondisk
->parent
= parent
->node_offset
;
1448 new_node
->ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1449 KKASSERT(ondisk
->type
== new_node
->ondisk
->type
);
1452 * Cleanup the original node. Elm (P) becomes the new boundary,
1453 * its subtree_offset was moved to the new node. If we had created
1454 * a new root its parent pointer may have changed.
1456 elm
->internal
.subtree_offset
= 0;
1457 ondisk
->count
= split
;
1460 * Insert the separator into the parent, fixup the parent's
1461 * reference to the original node, and reference the new node.
1462 * The separator is P.
1464 * Remember that base.count does not include the right-hand boundary.
1466 hammer_modify_node_all(cursor
->trans
, parent
);
1467 ondisk
= parent
->ondisk
;
1468 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1469 parent_elm
= &ondisk
->elms
[parent_index
+1];
1470 bcopy(parent_elm
, parent_elm
+ 1,
1471 (ondisk
->count
- parent_index
) * esize
);
1472 parent_elm
->internal
.base
= elm
->base
; /* separator P */
1473 parent_elm
->internal
.base
.btype
= new_node
->ondisk
->type
;
1474 parent_elm
->internal
.subtree_offset
= new_node
->node_offset
;
1476 hammer_modify_node_done(parent
);
1479 * The children of new_node need their parent pointer set to new_node.
1480 * The children have already been locked by
1481 * hammer_btree_lock_children().
1483 for (i
= 0; i
< new_node
->ondisk
->count
; ++i
) {
1484 elm
= &new_node
->ondisk
->elms
[i
];
1485 error
= btree_set_parent(cursor
->trans
, new_node
, elm
);
1487 panic("btree_split_internal: btree-fixup problem");
1490 hammer_modify_node_done(new_node
);
1493 * The filesystem's root B-Tree pointer may have to be updated.
1496 hammer_volume_t volume
;
1498 volume
= hammer_get_root_volume(hmp
, &error
);
1499 KKASSERT(error
== 0);
1501 hammer_modify_volume_field(cursor
->trans
, volume
,
1503 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1504 hammer_modify_volume_done(volume
);
1505 node
->ondisk
->parent
= parent
->node_offset
;
1506 if (cursor
->parent
) {
1507 hammer_unlock(&cursor
->parent
->lock
);
1508 hammer_rel_node(cursor
->parent
);
1510 cursor
->parent
= parent
; /* lock'd and ref'd */
1511 hammer_rel_volume(volume
, 0);
1513 hammer_modify_node_done(node
);
1517 * Ok, now adjust the cursor depending on which element the original
1518 * index was pointing at. If we are >= the split point the push node
1519 * is now in the new node.
1521 * NOTE: If we are at the split point itself we cannot stay with the
1522 * original node because the push index will point at the right-hand
1523 * boundary, which is illegal.
1525 * NOTE: The cursor's parent or parent_index must be adjusted for
1526 * the case where a new parent (new root) was created, and the case
1527 * where the cursor is now pointing at the split node.
1529 if (cursor
->index
>= split
) {
1530 cursor
->parent_index
= parent_index
+ 1;
1531 cursor
->index
-= split
;
1532 hammer_unlock(&cursor
->node
->lock
);
1533 hammer_rel_node(cursor
->node
);
1534 cursor
->node
= new_node
; /* locked and ref'd */
1536 cursor
->parent_index
= parent_index
;
1537 hammer_unlock(&new_node
->lock
);
1538 hammer_rel_node(new_node
);
1542 * Fixup left and right bounds
1544 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1545 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1546 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1547 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1548 &cursor
->node
->ondisk
->elms
[0].internal
.base
) <= 0);
1549 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1550 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
].internal
.base
) >= 0);
1553 hammer_btree_unlock_children(&locklist
);
1554 hammer_cursor_downgrade(cursor
);
1559 * Same as the above, but splits a full leaf node.
1565 btree_split_leaf(hammer_cursor_t cursor
)
1567 hammer_node_ondisk_t ondisk
;
1568 hammer_node_t parent
;
1571 hammer_node_t new_leaf
;
1572 hammer_btree_elm_t elm
;
1573 hammer_btree_elm_t parent_elm
;
1574 hammer_base_elm_t mid_boundary
;
1579 const size_t esize
= sizeof(*elm
);
1581 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1583 ++hammer_stats_btree_splits
;
1585 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1586 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1587 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1588 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1591 * Calculate the split point. If the insertion point will be on
1592 * the left-hand side adjust the split point to give the right
1593 * hand side one additional node.
1595 * Spikes are made up of two leaf elements which cannot be
1598 leaf
= cursor
->node
;
1599 ondisk
= leaf
->ondisk
;
1600 split
= (ondisk
->count
+ 1) / 2;
1601 if (cursor
->index
<= split
)
1606 elm
= &ondisk
->elms
[split
];
1608 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
[-1].leaf
.base
) <= 0);
1609 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->leaf
.base
) <= 0);
1610 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->leaf
.base
) > 0);
1611 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
[1].leaf
.base
) > 0);
1614 * If we are at the root of the tree, create a new root node with
1615 * 1 element and split normally. Avoid making major modifications
1616 * until we know the whole operation will work.
1618 if (ondisk
->parent
== 0) {
1619 parent
= hammer_alloc_btree(cursor
->trans
, &error
);
1622 hammer_lock_ex(&parent
->lock
);
1623 hammer_modify_node_noundo(cursor
->trans
, parent
);
1624 ondisk
= parent
->ondisk
;
1627 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1628 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1629 ondisk
->elms
[0].base
.btype
= leaf
->ondisk
->type
;
1630 ondisk
->elms
[0].internal
.subtree_offset
= leaf
->node_offset
;
1631 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1632 /* ondisk->elms[1].base.btype = not used */
1633 hammer_modify_node_done(parent
);
1635 parent_index
= 0; /* insertion point in parent */
1638 parent
= cursor
->parent
;
1639 parent_index
= cursor
->parent_index
;
1643 * Split leaf into new_leaf at the split point. Select a separator
1644 * value in-between the two leafs but with a bent towards the right
1645 * leaf since comparisons use an 'elm >= separator' inequality.
1654 new_leaf
= hammer_alloc_btree(cursor
->trans
, &error
);
1655 if (new_leaf
== NULL
) {
1657 hammer_unlock(&parent
->lock
);
1658 hammer_delete_node(cursor
->trans
, parent
);
1659 hammer_rel_node(parent
);
1663 hammer_lock_ex(&new_leaf
->lock
);
1666 * Create the new node and copy the leaf elements from the split
1667 * point on to the new node.
1669 hammer_modify_node_all(cursor
->trans
, leaf
);
1670 hammer_modify_node_noundo(cursor
->trans
, new_leaf
);
1671 ondisk
= leaf
->ondisk
;
1672 elm
= &ondisk
->elms
[split
];
1673 bcopy(elm
, &new_leaf
->ondisk
->elms
[0], (ondisk
->count
- split
) * esize
);
1674 new_leaf
->ondisk
->count
= ondisk
->count
- split
;
1675 new_leaf
->ondisk
->parent
= parent
->node_offset
;
1676 new_leaf
->ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
1677 KKASSERT(ondisk
->type
== new_leaf
->ondisk
->type
);
1678 hammer_modify_node_done(new_leaf
);
1681 * Cleanup the original node. Because this is a leaf node and
1682 * leaf nodes do not have a right-hand boundary, there
1683 * aren't any special edge cases to clean up. We just fixup the
1686 ondisk
->count
= split
;
1689 * Insert the separator into the parent, fixup the parent's
1690 * reference to the original node, and reference the new node.
1691 * The separator is P.
1693 * Remember that base.count does not include the right-hand boundary.
1694 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1696 hammer_modify_node_all(cursor
->trans
, parent
);
1697 ondisk
= parent
->ondisk
;
1698 KKASSERT(split
!= 0);
1699 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1700 parent_elm
= &ondisk
->elms
[parent_index
+1];
1701 bcopy(parent_elm
, parent_elm
+ 1,
1702 (ondisk
->count
- parent_index
) * esize
);
1704 hammer_make_separator(&elm
[-1].base
, &elm
[0].base
, &parent_elm
->base
);
1705 parent_elm
->internal
.base
.btype
= new_leaf
->ondisk
->type
;
1706 parent_elm
->internal
.subtree_offset
= new_leaf
->node_offset
;
1707 mid_boundary
= &parent_elm
->base
;
1709 hammer_modify_node_done(parent
);
1712 * The filesystem's root B-Tree pointer may have to be updated.
1715 hammer_volume_t volume
;
1717 volume
= hammer_get_root_volume(hmp
, &error
);
1718 KKASSERT(error
== 0);
1720 hammer_modify_volume_field(cursor
->trans
, volume
,
1722 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1723 hammer_modify_volume_done(volume
);
1724 leaf
->ondisk
->parent
= parent
->node_offset
;
1725 if (cursor
->parent
) {
1726 hammer_unlock(&cursor
->parent
->lock
);
1727 hammer_rel_node(cursor
->parent
);
1729 cursor
->parent
= parent
; /* lock'd and ref'd */
1730 hammer_rel_volume(volume
, 0);
1732 hammer_modify_node_done(leaf
);
1735 * Ok, now adjust the cursor depending on which element the original
1736 * index was pointing at. If we are >= the split point the push node
1737 * is now in the new node.
1739 * NOTE: If we are at the split point itself we need to select the
1740 * old or new node based on where key_beg's insertion point will be.
1741 * If we pick the wrong side the inserted element will wind up in
1742 * the wrong leaf node and outside that node's bounds.
1744 if (cursor
->index
> split
||
1745 (cursor
->index
== split
&&
1746 hammer_btree_cmp(&cursor
->key_beg
, mid_boundary
) >= 0)) {
1747 cursor
->parent_index
= parent_index
+ 1;
1748 cursor
->index
-= split
;
1749 hammer_unlock(&cursor
->node
->lock
);
1750 hammer_rel_node(cursor
->node
);
1751 cursor
->node
= new_leaf
;
1753 cursor
->parent_index
= parent_index
;
1754 hammer_unlock(&new_leaf
->lock
);
1755 hammer_rel_node(new_leaf
);
1759 * Fixup left and right bounds
1761 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1762 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1763 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1766 * Assert that the bounds are correct.
1768 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1769 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1770 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1771 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1772 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &cursor
->key_beg
) <= 0);
1773 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &cursor
->key_beg
) > 0);
1776 hammer_cursor_downgrade(cursor
);
1781 * Recursively correct the right-hand boundary's create_tid to (tid) as
1782 * long as the rest of the key matches. We have to recurse upward in
1783 * the tree as well as down the left side of each parent's right node.
1785 * Return EDEADLK if we were only partially successful, forcing the caller
1786 * to try again. The original cursor is not modified. This routine can
1787 * also fail with EDEADLK if it is forced to throw away a portion of its
1790 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1793 TAILQ_ENTRY(hammer_rhb
) entry
;
1798 TAILQ_HEAD(hammer_rhb_list
, hammer_rhb
);
1801 hammer_btree_correct_rhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
1803 struct hammer_rhb_list rhb_list
;
1804 hammer_base_elm_t elm
;
1805 hammer_node_t orig_node
;
1806 struct hammer_rhb
*rhb
;
1810 TAILQ_INIT(&rhb_list
);
1813 * Save our position so we can restore it on return. This also
1814 * gives us a stable 'elm'.
1816 orig_node
= cursor
->node
;
1817 hammer_ref_node(orig_node
);
1818 hammer_lock_sh(&orig_node
->lock
);
1819 orig_index
= cursor
->index
;
1820 elm
= &orig_node
->ondisk
->elms
[orig_index
].base
;
1823 * Now build a list of parents going up, allocating a rhb
1824 * structure for each one.
1826 while (cursor
->parent
) {
1828 * Stop if we no longer have any right-bounds to fix up
1830 if (elm
->obj_id
!= cursor
->right_bound
->obj_id
||
1831 elm
->rec_type
!= cursor
->right_bound
->rec_type
||
1832 elm
->key
!= cursor
->right_bound
->key
) {
1837 * Stop if the right-hand bound's create_tid does not
1838 * need to be corrected.
1840 if (cursor
->right_bound
->create_tid
>= tid
)
1843 rhb
= kmalloc(sizeof(*rhb
), M_HAMMER
, M_WAITOK
|M_ZERO
);
1844 rhb
->node
= cursor
->parent
;
1845 rhb
->index
= cursor
->parent_index
;
1846 hammer_ref_node(rhb
->node
);
1847 hammer_lock_sh(&rhb
->node
->lock
);
1848 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
1850 hammer_cursor_up(cursor
);
1854 * now safely adjust the right hand bound for each rhb. This may
1855 * also require taking the right side of the tree and iterating down
1859 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1860 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
1863 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1864 hammer_unlock(&rhb
->node
->lock
);
1865 hammer_rel_node(rhb
->node
);
1866 kfree(rhb
, M_HAMMER
);
1868 switch (cursor
->node
->ondisk
->type
) {
1869 case HAMMER_BTREE_TYPE_INTERNAL
:
1871 * Right-boundary for parent at internal node
1872 * is one element to the right of the element whos
1873 * right boundary needs adjusting. We must then
1874 * traverse down the left side correcting any left
1875 * bounds (which may now be too far to the left).
1878 error
= hammer_btree_correct_lhb(cursor
, tid
);
1881 panic("hammer_btree_correct_rhb(): Bad node type");
1890 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1891 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1892 hammer_unlock(&rhb
->node
->lock
);
1893 hammer_rel_node(rhb
->node
);
1894 kfree(rhb
, M_HAMMER
);
1896 error
= hammer_cursor_seek(cursor
, orig_node
, orig_index
);
1897 hammer_unlock(&orig_node
->lock
);
1898 hammer_rel_node(orig_node
);
1903 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1904 * bound going downward starting at the current cursor position.
1906 * This function does not restore the cursor after use.
1909 hammer_btree_correct_lhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
1911 struct hammer_rhb_list rhb_list
;
1912 hammer_base_elm_t elm
;
1913 hammer_base_elm_t cmp
;
1914 struct hammer_rhb
*rhb
;
1917 TAILQ_INIT(&rhb_list
);
1919 cmp
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1922 * Record the node and traverse down the left-hand side for all
1923 * matching records needing a boundary correction.
1927 rhb
= kmalloc(sizeof(*rhb
), M_HAMMER
, M_WAITOK
|M_ZERO
);
1928 rhb
->node
= cursor
->node
;
1929 rhb
->index
= cursor
->index
;
1930 hammer_ref_node(rhb
->node
);
1931 hammer_lock_sh(&rhb
->node
->lock
);
1932 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
1934 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1936 * Nothing to traverse down if we are at the right
1937 * boundary of an internal node.
1939 if (cursor
->index
== cursor
->node
->ondisk
->count
)
1942 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1943 if (elm
->btype
== HAMMER_BTREE_TYPE_RECORD
)
1945 panic("Illegal leaf record type %02x", elm
->btype
);
1947 error
= hammer_cursor_down(cursor
);
1951 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1952 if (elm
->obj_id
!= cmp
->obj_id
||
1953 elm
->rec_type
!= cmp
->rec_type
||
1954 elm
->key
!= cmp
->key
) {
1957 if (elm
->create_tid
>= tid
)
1963 * Now we can safely adjust the left-hand boundary from the bottom-up.
1964 * The last element we remove from the list is the caller's right hand
1965 * boundary, which must also be adjusted.
1967 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1968 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
1971 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1972 hammer_unlock(&rhb
->node
->lock
);
1973 hammer_rel_node(rhb
->node
);
1974 kfree(rhb
, M_HAMMER
);
1976 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1977 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1978 hammer_modify_node(cursor
->trans
, cursor
->node
,
1980 sizeof(elm
->create_tid
));
1981 elm
->create_tid
= tid
;
1982 hammer_modify_node_done(cursor
->node
);
1984 panic("hammer_btree_correct_lhb(): Bad element type");
1991 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1992 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1993 hammer_unlock(&rhb
->node
->lock
);
1994 hammer_rel_node(rhb
->node
);
1995 kfree(rhb
, M_HAMMER
);
2001 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2002 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2003 * the operation due to a deadlock, or some other error.
2005 * This routine is always called with an empty, locked leaf but may recurse
2006 * into want-to-be-empty parents as part of its operation.
2008 * It should also be noted that when removing empty leaves we must be sure
2009 * to test and update mirror_tid because another thread may have deadlocked
2010 * against us (or someone) trying to propogate it up and cannot retry once
2011 * the node has been deleted.
2013 * On return the cursor may end up pointing to an internal node, suitable
2014 * for further iteration but not for an immediate insertion or deletion.
2017 btree_remove(hammer_cursor_t cursor
)
2019 hammer_node_ondisk_t ondisk
;
2020 hammer_btree_elm_t elm
;
2022 hammer_node_t parent
;
2023 const int esize
= sizeof(*elm
);
2026 node
= cursor
->node
;
2029 * When deleting the root of the filesystem convert it to
2030 * an empty leaf node. Internal nodes cannot be empty.
2032 ondisk
= node
->ondisk
;
2033 if (ondisk
->parent
== 0) {
2034 KKASSERT(cursor
->parent
== NULL
);
2035 hammer_modify_node_all(cursor
->trans
, node
);
2036 KKASSERT(ondisk
== node
->ondisk
);
2037 ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
2039 hammer_modify_node_done(node
);
2044 parent
= cursor
->parent
;
2047 * If another thread deadlocked trying to propogate mirror_tid up
2048 * we have to finish the job before deleting node. XXX
2050 if (parent
->ondisk
->mirror_tid
< node
->ondisk
->mirror_tid
&&
2051 (cursor
->trans
->hmp
->hflags
& (HMNT_MASTERID
|HMNT_SLAVE
))) {
2052 hammer_btree_mirror_propagate(cursor
->trans
,
2054 cursor
->parent_index
,
2055 node
->ondisk
->mirror_tid
);
2060 * Attempt to remove the parent's reference to the child. If the
2061 * parent would become empty we have to recurse. If we fail we
2062 * leave the parent pointing to an empty leaf node.
2064 if (parent
->ondisk
->count
== 1) {
2066 * This special cursor_up_locked() call leaves the original
2067 * node exclusively locked and referenced, leaves the
2068 * original parent locked (as the new node), and locks the
2069 * new parent. It can return EDEADLK.
2071 error
= hammer_cursor_up_locked(cursor
);
2073 error
= btree_remove(cursor
);
2075 hammer_modify_node_all(cursor
->trans
, node
);
2076 ondisk
= node
->ondisk
;
2077 ondisk
->type
= HAMMER_BTREE_TYPE_DELETED
;
2079 hammer_modify_node_done(node
);
2080 hammer_flush_node(node
);
2081 hammer_delete_node(cursor
->trans
, node
);
2083 kprintf("Warning: BTREE_REMOVE: Defering "
2084 "parent removal1 @ %016llx, skipping\n",
2087 hammer_unlock(&node
->lock
);
2088 hammer_rel_node(node
);
2090 kprintf("Warning: BTREE_REMOVE: Defering parent "
2091 "removal2 @ %016llx, skipping\n",
2095 KKASSERT(parent
->ondisk
->count
> 1);
2098 * Delete the subtree reference in the parent
2100 hammer_modify_node_all(cursor
->trans
, parent
);
2101 ondisk
= parent
->ondisk
;
2102 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
);
2104 elm
= &ondisk
->elms
[cursor
->parent_index
];
2105 KKASSERT(elm
->internal
.subtree_offset
== node
->node_offset
);
2106 KKASSERT(ondisk
->count
> 0);
2107 bcopy(&elm
[1], &elm
[0],
2108 (ondisk
->count
- cursor
->parent_index
) * esize
);
2110 hammer_modify_node_done(parent
);
2111 hammer_flush_node(node
);
2112 hammer_delete_node(cursor
->trans
, node
);
2115 * cursor->node is invalid, cursor up to make the cursor
2118 error
= hammer_cursor_up(cursor
);
2124 * Propagate a mirror TID update upwards through the B-Tree to the root.
2126 * A locked internal node must be passed in. The node will remain locked
2129 * This function syncs mirror_tid at the specified internal node's element,
2130 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2133 hammer_btree_mirror_propagate(hammer_transaction_t trans
, hammer_node_t node
,
2134 int index
, hammer_tid_t mirror_tid
)
2136 hammer_btree_internal_elm_t elm
;
2137 hammer_node_t parent
;
2141 KKASSERT (node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
);
2144 * Adjust the node's element
2146 elm
= &node
->ondisk
->elms
[index
].internal
;
2147 if (elm
->mirror_tid
>= mirror_tid
)
2149 hammer_modify_node(trans
, node
, &elm
->mirror_tid
,
2150 sizeof(elm
->mirror_tid
));
2151 elm
->mirror_tid
= mirror_tid
;
2152 hammer_modify_node_done(node
);
2155 * Adjust the node's mirror_tid aggragator
2157 if (node
->ondisk
->mirror_tid
>= mirror_tid
)
2159 hammer_modify_node_field(trans
, node
, mirror_tid
);
2160 node
->ondisk
->mirror_tid
= mirror_tid
;
2161 hammer_modify_node_done(node
);
2165 if (node
->ondisk
->parent
&&
2166 (trans
->hmp
->hflags
& (HMNT_MASTERID
|HMNT_SLAVE
))) {
2167 parent
= hammer_btree_get_parent(node
, &parent_index
,
2170 hammer_btree_mirror_propagate(trans
, parent
,
2171 parent_index
, mirror_tid
);
2172 hammer_unlock(&parent
->lock
);
2173 hammer_rel_node(parent
);
2180 hammer_btree_get_parent(hammer_node_t node
, int *parent_indexp
, int *errorp
,
2183 hammer_node_t parent
;
2184 hammer_btree_elm_t elm
;
2190 parent
= hammer_get_node(node
->hmp
, node
->ondisk
->parent
, 0, errorp
);
2192 KKASSERT(parent
== NULL
);
2195 KKASSERT ((parent
->flags
& HAMMER_NODE_DELETED
) == 0);
2200 if (try_exclusive
) {
2201 if (hammer_lock_ex_try(&parent
->lock
)) {
2202 hammer_rel_node(parent
);
2207 hammer_lock_sh(&parent
->lock
);
2211 * Figure out which element in the parent is pointing to the
2214 if (node
->ondisk
->count
) {
2215 i
= hammer_btree_search_node(&node
->ondisk
->elms
[0].base
,
2220 while (i
< parent
->ondisk
->count
) {
2221 elm
= &parent
->ondisk
->elms
[i
];
2222 if (elm
->internal
.subtree_offset
== node
->node_offset
)
2226 if (i
== parent
->ondisk
->count
) {
2227 hammer_unlock(&parent
->lock
);
2228 panic("Bad B-Tree link: parent %p node %p\n", parent
, node
);
2231 KKASSERT(*errorp
== 0);
2236 * The element (elm) has been moved to a new internal node (node).
2238 * If the element represents a pointer to an internal node that node's
2239 * parent must be adjusted to the element's new location.
2241 * XXX deadlock potential here with our exclusive locks
2244 btree_set_parent(hammer_transaction_t trans
, hammer_node_t node
,
2245 hammer_btree_elm_t elm
)
2247 hammer_node_t child
;
2252 switch(elm
->base
.btype
) {
2253 case HAMMER_BTREE_TYPE_INTERNAL
:
2254 case HAMMER_BTREE_TYPE_LEAF
:
2255 child
= hammer_get_node(node
->hmp
, elm
->internal
.subtree_offset
,
2258 hammer_modify_node_field(trans
, child
, parent
);
2259 child
->ondisk
->parent
= node
->node_offset
;
2260 hammer_modify_node_done(child
);
2261 hammer_rel_node(child
);
2271 * Exclusively lock all the children of node. This is used by the split
2272 * code to prevent anyone from accessing the children of a cursor node
2273 * while we fix-up its parent offset.
2275 * If we don't lock the children we can really mess up cursors which block
2276 * trying to cursor-up into our node.
2278 * On failure EDEADLK (or some other error) is returned. If a deadlock
2279 * error is returned the cursor is adjusted to block on termination.
2282 hammer_btree_lock_children(hammer_cursor_t cursor
,
2283 struct hammer_node_locklist
**locklistp
)
2286 hammer_node_locklist_t item
;
2287 hammer_node_ondisk_t ondisk
;
2288 hammer_btree_elm_t elm
;
2289 hammer_node_t child
;
2293 node
= cursor
->node
;
2294 ondisk
= node
->ondisk
;
2298 * We really do not want to block on I/O with exclusive locks held,
2299 * pre-get the children before trying to lock the mess.
2301 for (i
= 0; i
< ondisk
->count
; ++i
) {
2302 ++hammer_stats_btree_elements
;
2303 elm
= &ondisk
->elms
[i
];
2304 if (elm
->base
.btype
!= HAMMER_BTREE_TYPE_LEAF
&&
2305 elm
->base
.btype
!= HAMMER_BTREE_TYPE_INTERNAL
) {
2308 child
= hammer_get_node(node
->hmp
,
2309 elm
->internal
.subtree_offset
,
2312 hammer_rel_node(child
);
2318 for (i
= 0; error
== 0 && i
< ondisk
->count
; ++i
) {
2319 ++hammer_stats_btree_elements
;
2320 elm
= &ondisk
->elms
[i
];
2322 switch(elm
->base
.btype
) {
2323 case HAMMER_BTREE_TYPE_INTERNAL
:
2324 case HAMMER_BTREE_TYPE_LEAF
:
2325 KKASSERT(elm
->internal
.subtree_offset
!= 0);
2326 child
= hammer_get_node(node
->hmp
,
2327 elm
->internal
.subtree_offset
,
2335 if (hammer_lock_ex_try(&child
->lock
) != 0) {
2336 if (cursor
->deadlk_node
== NULL
) {
2337 cursor
->deadlk_node
= child
;
2338 hammer_ref_node(cursor
->deadlk_node
);
2341 hammer_rel_node(child
);
2343 item
= kmalloc(sizeof(*item
),
2344 M_HAMMER
, M_WAITOK
);
2345 item
->next
= *locklistp
;
2352 hammer_btree_unlock_children(locklistp
);
2358 * Release previously obtained node locks.
2361 hammer_btree_unlock_children(struct hammer_node_locklist
**locklistp
)
2363 hammer_node_locklist_t item
;
2365 while ((item
= *locklistp
) != NULL
) {
2366 *locklistp
= item
->next
;
2367 hammer_unlock(&item
->node
->lock
);
2368 hammer_rel_node(item
->node
);
2369 kfree(item
, M_HAMMER
);
2373 /************************************************************************
2374 * MISCELLANIOUS SUPPORT *
2375 ************************************************************************/
2378 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2380 * Note that for this particular function a return value of -1, 0, or +1
2381 * can denote a match if create_tid is otherwise discounted. A create_tid
2382 * of zero is considered to be 'infinity' in comparisons.
2384 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2387 hammer_btree_cmp(hammer_base_elm_t key1
, hammer_base_elm_t key2
)
2389 if (key1
->localization
< key2
->localization
)
2391 if (key1
->localization
> key2
->localization
)
2394 if (key1
->obj_id
< key2
->obj_id
)
2396 if (key1
->obj_id
> key2
->obj_id
)
2399 if (key1
->rec_type
< key2
->rec_type
)
2401 if (key1
->rec_type
> key2
->rec_type
)
2404 if (key1
->key
< key2
->key
)
2406 if (key1
->key
> key2
->key
)
2410 * A create_tid of zero indicates a record which is undeletable
2411 * and must be considered to have a value of positive infinity.
2413 if (key1
->create_tid
== 0) {
2414 if (key2
->create_tid
== 0)
2418 if (key2
->create_tid
== 0)
2420 if (key1
->create_tid
< key2
->create_tid
)
2422 if (key1
->create_tid
> key2
->create_tid
)
2428 * Test a timestamp against an element to determine whether the
2429 * element is visible. A timestamp of 0 means 'infinity'.
2432 hammer_btree_chkts(hammer_tid_t asof
, hammer_base_elm_t base
)
2435 if (base
->delete_tid
)
2439 if (asof
< base
->create_tid
)
2441 if (base
->delete_tid
&& asof
>= base
->delete_tid
)
2447 * Create a separator half way inbetween key1 and key2. For fields just
2448 * one unit apart, the separator will match key2. key1 is on the left-hand
2449 * side and key2 is on the right-hand side.
2451 * key2 must be >= the separator. It is ok for the separator to match key2.
2453 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2456 * NOTE: It might be beneficial to just scrap this whole mess and just
2457 * set the separator to key2.
2459 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2460 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2463 hammer_make_separator(hammer_base_elm_t key1
, hammer_base_elm_t key2
,
2464 hammer_base_elm_t dest
)
2466 bzero(dest
, sizeof(*dest
));
2468 dest
->rec_type
= key2
->rec_type
;
2469 dest
->key
= key2
->key
;
2470 dest
->obj_id
= key2
->obj_id
;
2471 dest
->create_tid
= key2
->create_tid
;
2473 MAKE_SEPARATOR(key1
, key2
, dest
, localization
);
2474 if (key1
->localization
== key2
->localization
) {
2475 MAKE_SEPARATOR(key1
, key2
, dest
, obj_id
);
2476 if (key1
->obj_id
== key2
->obj_id
) {
2477 MAKE_SEPARATOR(key1
, key2
, dest
, rec_type
);
2478 if (key1
->rec_type
== key2
->rec_type
) {
2479 MAKE_SEPARATOR(key1
, key2
, dest
, key
);
2481 * Don't bother creating a separator for
2482 * create_tid, which also conveniently avoids
2483 * having to handle the create_tid == 0
2484 * (infinity) case. Just leave create_tid
2487 * Worst case, dest matches key2 exactly,
2488 * which is acceptable.
2495 #undef MAKE_SEPARATOR
2498 * Return whether a generic internal or leaf node is full
2501 btree_node_is_full(hammer_node_ondisk_t node
)
2503 switch(node
->type
) {
2504 case HAMMER_BTREE_TYPE_INTERNAL
:
2505 if (node
->count
== HAMMER_BTREE_INT_ELMS
)
2508 case HAMMER_BTREE_TYPE_LEAF
:
2509 if (node
->count
== HAMMER_BTREE_LEAF_ELMS
)
2513 panic("illegal btree subtype");
2520 btree_max_elements(u_int8_t type
)
2522 if (type
== HAMMER_BTREE_TYPE_LEAF
)
2523 return(HAMMER_BTREE_LEAF_ELMS
);
2524 if (type
== HAMMER_BTREE_TYPE_INTERNAL
)
2525 return(HAMMER_BTREE_INT_ELMS
);
2526 panic("btree_max_elements: bad type %d\n", type
);
2531 hammer_print_btree_node(hammer_node_ondisk_t ondisk
)
2533 hammer_btree_elm_t elm
;
2536 kprintf("node %p count=%d parent=%016llx type=%c\n",
2537 ondisk
, ondisk
->count
, ondisk
->parent
, ondisk
->type
);
2540 * Dump both boundary elements if an internal node
2542 if (ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
2543 for (i
= 0; i
<= ondisk
->count
; ++i
) {
2544 elm
= &ondisk
->elms
[i
];
2545 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2548 for (i
= 0; i
< ondisk
->count
; ++i
) {
2549 elm
= &ondisk
->elms
[i
];
2550 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2556 hammer_print_btree_elm(hammer_btree_elm_t elm
, u_int8_t type
, int i
)
2559 kprintf("\tobj_id = %016llx\n", elm
->base
.obj_id
);
2560 kprintf("\tkey = %016llx\n", elm
->base
.key
);
2561 kprintf("\tcreate_tid = %016llx\n", elm
->base
.create_tid
);
2562 kprintf("\tdelete_tid = %016llx\n", elm
->base
.delete_tid
);
2563 kprintf("\trec_type = %04x\n", elm
->base
.rec_type
);
2564 kprintf("\tobj_type = %02x\n", elm
->base
.obj_type
);
2565 kprintf("\tbtype = %02x (%c)\n",
2567 (elm
->base
.btype
? elm
->base
.btype
: '?'));
2568 kprintf("\tlocalization = %02x\n", elm
->base
.localization
);
2571 case HAMMER_BTREE_TYPE_INTERNAL
:
2572 kprintf("\tsubtree_off = %016llx\n",
2573 elm
->internal
.subtree_offset
);
2575 case HAMMER_BTREE_TYPE_RECORD
:
2576 kprintf("\tdata_offset = %016llx\n", elm
->leaf
.data_offset
);
2577 kprintf("\tdata_len = %08x\n", elm
->leaf
.data_len
);
2578 kprintf("\tdata_crc = %08x\n", elm
->leaf
.data_crc
);