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 1.58 2008/06/26 04:06:22 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 if (cursor
->index
== node
->count
) {
151 if (hammer_debug_btree
) {
152 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
153 cursor
->node
->node_offset
,
155 (cursor
->parent
? cursor
->parent
->node_offset
: -1),
156 cursor
->parent_index
,
159 KKASSERT(cursor
->parent
== NULL
|| cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.subtree_offset
== cursor
->node
->node_offset
);
160 error
= hammer_cursor_up(cursor
);
163 /* reload stale pointer */
164 node
= cursor
->node
->ondisk
;
165 KKASSERT(cursor
->index
!= node
->count
);
168 * If we are reblocking we want to return internal
171 if (cursor
->flags
& HAMMER_CURSOR_REBLOCKING
) {
172 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
180 * Check internal or leaf element. Determine if the record
181 * at the cursor has gone beyond the end of our range.
183 * We recurse down through internal nodes.
185 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
186 elm
= &node
->elms
[cursor
->index
];
188 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
189 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
190 if (hammer_debug_btree
) {
191 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
192 cursor
->node
->node_offset
,
194 elm
[0].internal
.base
.obj_id
,
195 elm
[0].internal
.base
.rec_type
,
196 elm
[0].internal
.base
.key
,
197 elm
[0].internal
.base
.localization
,
201 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
202 cursor
->node
->node_offset
,
204 elm
[1].internal
.base
.obj_id
,
205 elm
[1].internal
.base
.rec_type
,
206 elm
[1].internal
.base
.key
,
207 elm
[1].internal
.base
.localization
,
216 if (r
== 0 && (cursor
->flags
&
217 HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
226 KKASSERT(elm
->internal
.subtree_offset
!= 0);
229 * If running the mirror filter see if we can skip
230 * the entire sub-tree.
232 if (cursor
->flags
& HAMMER_CURSOR_MIRROR_FILTERED
) {
233 if (elm
->internal
.mirror_tid
<
234 cursor
->mirror_tid
) {
240 error
= hammer_cursor_down(cursor
);
243 KKASSERT(cursor
->index
== 0);
244 /* reload stale pointer */
245 node
= cursor
->node
->ondisk
;
248 elm
= &node
->elms
[cursor
->index
];
249 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
->base
);
250 if (hammer_debug_btree
) {
251 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
252 cursor
->node
->node_offset
,
254 (elm
[0].leaf
.base
.btype
?
255 elm
[0].leaf
.base
.btype
: '?'),
256 elm
[0].leaf
.base
.obj_id
,
257 elm
[0].leaf
.base
.rec_type
,
258 elm
[0].leaf
.base
.key
,
259 elm
[0].leaf
.base
.localization
,
269 * We support both end-inclusive and
270 * end-exclusive searches.
273 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
278 switch(elm
->leaf
.base
.btype
) {
279 case HAMMER_BTREE_TYPE_RECORD
:
280 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
281 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
294 * node pointer invalid after loop
300 if (hammer_debug_btree
) {
301 int i
= cursor
->index
;
302 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
303 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
305 elm
->internal
.base
.obj_id
,
306 elm
->internal
.base
.rec_type
,
307 elm
->internal
.base
.key
,
308 elm
->internal
.base
.localization
317 * Iterate in the reverse direction. This is used by the pruning code to
318 * avoid overlapping records.
321 hammer_btree_iterate_reverse(hammer_cursor_t cursor
)
323 hammer_node_ondisk_t node
;
324 hammer_btree_elm_t elm
;
330 * Skip past the current record. For various reasons the cursor
331 * may end up set to -1 or set to point at the end of the current
332 * node. These cases must be addressed.
334 node
= cursor
->node
->ondisk
;
337 if (cursor
->index
!= -1 &&
338 (cursor
->flags
& HAMMER_CURSOR_ATEDISK
)) {
341 if (cursor
->index
== cursor
->node
->ondisk
->count
)
345 * Loop until an element is found or we are done.
349 * We iterate up the tree and then index over one element
350 * while we are at the last element in the current node.
352 if (cursor
->index
== -1) {
353 error
= hammer_cursor_up(cursor
);
355 cursor
->index
= 0; /* sanity */
358 /* reload stale pointer */
359 node
= cursor
->node
->ondisk
;
360 KKASSERT(cursor
->index
!= node
->count
);
366 * Check internal or leaf element. Determine if the record
367 * at the cursor has gone beyond the end of our range.
369 * We recurse down through internal nodes.
371 KKASSERT(cursor
->index
!= node
->count
);
372 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
373 elm
= &node
->elms
[cursor
->index
];
374 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
375 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
376 if (hammer_debug_btree
) {
377 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
378 cursor
->node
->node_offset
,
380 elm
[0].internal
.base
.obj_id
,
381 elm
[0].internal
.base
.rec_type
,
382 elm
[0].internal
.base
.key
,
383 elm
[0].internal
.base
.localization
,
386 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
387 cursor
->node
->node_offset
,
389 elm
[1].internal
.base
.obj_id
,
390 elm
[1].internal
.base
.rec_type
,
391 elm
[1].internal
.base
.key
,
392 elm
[1].internal
.base
.localization
,
406 KKASSERT(elm
->internal
.subtree_offset
!= 0);
408 error
= hammer_cursor_down(cursor
);
411 KKASSERT(cursor
->index
== 0);
412 /* reload stale pointer */
413 node
= cursor
->node
->ondisk
;
415 /* this can assign -1 if the leaf was empty */
416 cursor
->index
= node
->count
- 1;
419 elm
= &node
->elms
[cursor
->index
];
420 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
421 if (hammer_debug_btree
) {
422 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
423 cursor
->node
->node_offset
,
425 (elm
[0].leaf
.base
.btype
?
426 elm
[0].leaf
.base
.btype
: '?'),
427 elm
[0].leaf
.base
.obj_id
,
428 elm
[0].leaf
.base
.rec_type
,
429 elm
[0].leaf
.base
.key
,
430 elm
[0].leaf
.base
.localization
,
439 switch(elm
->leaf
.base
.btype
) {
440 case HAMMER_BTREE_TYPE_RECORD
:
441 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
442 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
455 * node pointer invalid after loop
461 if (hammer_debug_btree
) {
462 int i
= cursor
->index
;
463 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
464 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
466 elm
->internal
.base
.obj_id
,
467 elm
->internal
.base
.rec_type
,
468 elm
->internal
.base
.key
,
469 elm
->internal
.base
.localization
478 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
479 * could not be found, EDEADLK if inserting and a retry is needed, and a
480 * fatal error otherwise. When retrying, the caller must terminate the
481 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
483 * The cursor is suitably positioned for a deletion on success, and suitably
484 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
487 * The cursor may begin anywhere, the search will traverse the tree in
488 * either direction to locate the requested element.
490 * Most of the logic implementing historical searches is handled here. We
491 * do an initial lookup with create_tid set to the asof TID. Due to the
492 * way records are laid out, a backwards iteration may be required if
493 * ENOENT is returned to locate the historical record. Here's the
496 * create_tid: 10 15 20
500 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
501 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
502 * not visible and thus causes ENOENT to be returned. We really need
503 * to check record 11 in LEAF1. If it also fails then the search fails
504 * (e.g. it might represent the range 11-16 and thus still not match our
505 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
506 * further iterations.
508 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
509 * and the cursor->create_check TID if an iteration might be needed.
510 * In the above example create_check would be set to 14.
513 hammer_btree_lookup(hammer_cursor_t cursor
)
517 ++hammer_stats_btree_lookups
;
518 if (cursor
->flags
& HAMMER_CURSOR_ASOF
) {
519 KKASSERT((cursor
->flags
& HAMMER_CURSOR_INSERT
) == 0);
520 cursor
->key_beg
.create_tid
= cursor
->asof
;
522 cursor
->flags
&= ~HAMMER_CURSOR_CREATE_CHECK
;
523 error
= btree_search(cursor
, 0);
524 if (error
!= ENOENT
||
525 (cursor
->flags
& HAMMER_CURSOR_CREATE_CHECK
) == 0) {
528 * Stop if error other then ENOENT.
529 * Stop if ENOENT and not special case.
533 if (hammer_debug_btree
) {
534 kprintf("CREATE_CHECK %016llx\n",
535 cursor
->create_check
);
537 cursor
->key_beg
.create_tid
= cursor
->create_check
;
541 error
= btree_search(cursor
, 0);
544 error
= hammer_btree_extract(cursor
, cursor
->flags
);
549 * Execute the logic required to start an iteration. The first record
550 * located within the specified range is returned and iteration control
551 * flags are adjusted for successive hammer_btree_iterate() calls.
554 hammer_btree_first(hammer_cursor_t cursor
)
558 error
= hammer_btree_lookup(cursor
);
559 if (error
== ENOENT
) {
560 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
561 error
= hammer_btree_iterate(cursor
);
563 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
568 * Similarly but for an iteration in the reverse direction.
570 * Set ATEDISK when iterating backwards to skip the current entry,
571 * which after an ENOENT lookup will be pointing beyond our end point.
574 hammer_btree_last(hammer_cursor_t cursor
)
576 struct hammer_base_elm save
;
579 save
= cursor
->key_beg
;
580 cursor
->key_beg
= cursor
->key_end
;
581 error
= hammer_btree_lookup(cursor
);
582 cursor
->key_beg
= save
;
583 if (error
== ENOENT
||
584 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
585 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
586 error
= hammer_btree_iterate_reverse(cursor
);
588 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
593 * Extract the record and/or data associated with the cursor's current
594 * position. Any prior record or data stored in the cursor is replaced.
595 * The cursor must be positioned at a leaf node.
597 * NOTE: All extractions occur at the leaf of the B-Tree.
600 hammer_btree_extract(hammer_cursor_t cursor
, int flags
)
603 hammer_node_ondisk_t node
;
604 hammer_btree_elm_t elm
;
605 hammer_off_t data_off
;
610 * The case where the data reference resolves to the same buffer
611 * as the record reference must be handled.
613 node
= cursor
->node
->ondisk
;
614 elm
= &node
->elms
[cursor
->index
];
616 hmp
= cursor
->node
->hmp
;
619 * There is nothing to extract for an internal element.
621 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
)
625 * Only record types have data.
627 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
628 cursor
->leaf
= &elm
->leaf
;
630 if ((flags
& HAMMER_CURSOR_GET_DATA
) == 0)
632 if (elm
->leaf
.base
.btype
!= HAMMER_BTREE_TYPE_RECORD
)
634 data_off
= elm
->leaf
.data_offset
;
635 data_len
= elm
->leaf
.data_len
;
642 KKASSERT(data_len
>= 0 && data_len
<= HAMMER_XBUFSIZE
);
643 cursor
->data
= hammer_bread_ext(hmp
, data_off
, data_len
,
644 &error
, &cursor
->data_buffer
);
645 if (hammer_crc_test_leaf(cursor
->data
, &elm
->leaf
) == 0)
646 Debugger("CRC FAILED: DATA");
652 * Insert a leaf element into the B-Tree at the current cursor position.
653 * The cursor is positioned such that the element at and beyond the cursor
654 * are shifted to make room for the new record.
656 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
657 * flag set and that call must return ENOENT before this function can be
660 * The caller may depend on the cursor's exclusive lock after return to
661 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
663 * ENOSPC is returned if there is no room to insert a new record.
666 hammer_btree_insert(hammer_cursor_t cursor
, hammer_btree_leaf_elm_t elm
)
668 hammer_node_ondisk_t node
;
672 if ((error
= hammer_cursor_upgrade_node(cursor
)) != 0)
674 ++hammer_stats_btree_inserts
;
677 * Insert the element at the leaf node and update the count in the
678 * parent. It is possible for parent to be NULL, indicating that
679 * the filesystem's ROOT B-Tree node is a leaf itself, which is
680 * possible. The root inode can never be deleted so the leaf should
683 * Remember that the right-hand boundary is not included in the
686 hammer_modify_node_all(cursor
->trans
, cursor
->node
);
687 node
= cursor
->node
->ondisk
;
689 KKASSERT(elm
->base
.btype
!= 0);
690 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
691 KKASSERT(node
->count
< HAMMER_BTREE_LEAF_ELMS
);
692 if (i
!= node
->count
) {
693 bcopy(&node
->elms
[i
], &node
->elms
[i
+1],
694 (node
->count
- i
) * sizeof(*elm
));
696 node
->elms
[i
].leaf
= *elm
;
700 * Update the leaf node's aggregate mirror_tid for mirroring
703 if (node
->mirror_tid
< elm
->base
.delete_tid
)
704 node
->mirror_tid
= elm
->base
.delete_tid
;
705 if (node
->mirror_tid
< elm
->base
.create_tid
)
706 node
->mirror_tid
= elm
->base
.create_tid
;
707 hammer_modify_node_done(cursor
->node
);
710 * What we really want to do is propogate mirror_tid all the way
711 * up the parent chain to the B-Tree root. That would be
712 * ultra-expensive, though.
714 if (cursor
->parent
&&
715 (cursor
->trans
->hmp
->hflags
& (HMNT_MASTERID
|HMNT_SLAVE
))) {
716 hammer_btree_mirror_propagate(cursor
->trans
, cursor
->parent
,
717 cursor
->parent_index
,
722 * Debugging sanity checks.
724 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->base
) <= 0);
725 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->base
) > 0);
727 KKASSERT(hammer_btree_cmp(&node
->elms
[i
-1].leaf
.base
, &elm
->base
) < 0);
729 if (i
!= node
->count
- 1)
730 KKASSERT(hammer_btree_cmp(&node
->elms
[i
+1].leaf
.base
, &elm
->base
) > 0);
736 * Delete a record from the B-Tree at the current cursor position.
737 * The cursor is positioned such that the current element is the one
740 * On return the cursor will be positioned after the deleted element and
741 * MAY point to an internal node. It will be suitable for the continuation
742 * of an iteration but not for an insertion or deletion.
744 * Deletions will attempt to partially rebalance the B-Tree in an upward
745 * direction, but will terminate rather then deadlock. Empty internal nodes
746 * are never allowed by a deletion which deadlocks may end up giving us an
747 * empty leaf. The pruner will clean up and rebalance the tree.
749 * This function can return EDEADLK, requiring the caller to retry the
750 * operation after clearing the deadlock.
753 hammer_btree_delete(hammer_cursor_t cursor
)
755 hammer_node_ondisk_t ondisk
;
757 hammer_node_t parent
;
761 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
763 ++hammer_stats_btree_deletes
;
766 * Delete the element from the leaf node.
768 * Remember that leaf nodes do not have boundaries.
771 ondisk
= node
->ondisk
;
774 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_LEAF
);
775 KKASSERT(i
>= 0 && i
< ondisk
->count
);
776 hammer_modify_node_all(cursor
->trans
, node
);
777 if (i
+ 1 != ondisk
->count
) {
778 bcopy(&ondisk
->elms
[i
+1], &ondisk
->elms
[i
],
779 (ondisk
->count
- i
- 1) * sizeof(ondisk
->elms
[0]));
782 hammer_modify_node_done(node
);
785 * Validate local parent
787 if (ondisk
->parent
) {
788 parent
= cursor
->parent
;
790 KKASSERT(parent
!= NULL
);
791 KKASSERT(parent
->node_offset
== ondisk
->parent
);
795 * If the leaf becomes empty it must be detached from the parent,
796 * potentially recursing through to the filesystem root.
798 * This may reposition the cursor at one of the parent's of the
801 * Ignore deadlock errors, that simply means that btree_remove
802 * was unable to recurse and had to leave us with an empty leaf.
804 KKASSERT(cursor
->index
<= ondisk
->count
);
805 if (ondisk
->count
== 0) {
806 error
= btree_remove(cursor
);
807 if (error
== EDEADLK
)
812 KKASSERT(cursor
->parent
== NULL
||
813 cursor
->parent_index
< cursor
->parent
->ondisk
->count
);
818 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
820 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
822 * The search can begin ANYWHERE in the B-Tree. As a first step the search
823 * iterates up the tree as necessary to properly position itself prior to
824 * actually doing the sarch.
826 * INSERTIONS: The search will split full nodes and leaves on its way down
827 * and guarentee that the leaf it ends up on is not full. If we run out
828 * of space the search continues to the leaf (to position the cursor for
829 * the spike), but ENOSPC is returned.
831 * The search is only guarenteed to end up on a leaf if an error code of 0
832 * is returned, or if inserting and an error code of ENOENT is returned.
833 * Otherwise it can stop at an internal node. On success a search returns
836 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
837 * filesystem, and it is not simple code. Please note the following facts:
839 * - Internal node recursions have a boundary on the left AND right. The
840 * right boundary is non-inclusive. The create_tid is a generic part
841 * of the key for internal nodes.
843 * - Leaf nodes contain terminal elements only now.
845 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
846 * historical search. ASOF and INSERT are mutually exclusive. When
847 * doing an as-of lookup btree_search() checks for a right-edge boundary
848 * case. If while recursing down the left-edge differs from the key
849 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
850 * with cursor->create_check. This is used by btree_lookup() to iterate.
851 * The iteration backwards because as-of searches can wind up going
852 * down the wrong branch of the B-Tree.
856 btree_search(hammer_cursor_t cursor
, int flags
)
858 hammer_node_ondisk_t node
;
859 hammer_btree_elm_t elm
;
866 flags
|= cursor
->flags
;
867 ++hammer_stats_btree_searches
;
869 if (hammer_debug_btree
) {
870 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
871 cursor
->node
->node_offset
,
873 cursor
->key_beg
.obj_id
,
874 cursor
->key_beg
.rec_type
,
876 cursor
->key_beg
.create_tid
,
877 cursor
->key_beg
.localization
,
881 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
882 cursor
->parent
->node_offset
, cursor
->parent_index
,
883 cursor
->left_bound
->obj_id
,
884 cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.base
.obj_id
,
885 cursor
->right_bound
->obj_id
,
886 cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1].internal
.base
.obj_id
,
888 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
],
890 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1]
895 * Move our cursor up the tree until we find a node whos range covers
896 * the key we are trying to locate.
898 * The left bound is inclusive, the right bound is non-inclusive.
899 * It is ok to cursor up too far.
902 r
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->left_bound
);
903 s
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->right_bound
);
906 KKASSERT(cursor
->parent
);
907 ++hammer_stats_btree_iterations
;
908 error
= hammer_cursor_up(cursor
);
914 * The delete-checks below are based on node, not parent. Set the
915 * initial delete-check based on the parent.
918 KKASSERT(cursor
->left_bound
->create_tid
!= 1);
919 cursor
->create_check
= cursor
->left_bound
->create_tid
- 1;
920 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
924 * We better have ended up with a node somewhere.
926 KKASSERT(cursor
->node
!= NULL
);
929 * If we are inserting we can't start at a full node if the parent
930 * is also full (because there is no way to split the node),
931 * continue running up the tree until the requirement is satisfied
932 * or we hit the root of the filesystem.
934 * (If inserting we aren't doing an as-of search so we don't have
935 * to worry about create_check).
937 while ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
938 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
939 if (btree_node_is_full(cursor
->node
->ondisk
) == 0)
942 if (btree_node_is_full(cursor
->node
->ondisk
) ==0)
945 if (cursor
->node
->ondisk
->parent
== 0 ||
946 cursor
->parent
->ondisk
->count
!= HAMMER_BTREE_INT_ELMS
) {
949 ++hammer_stats_btree_iterations
;
950 error
= hammer_cursor_up(cursor
);
951 /* node may have become stale */
957 * Push down through internal nodes to locate the requested key.
959 node
= cursor
->node
->ondisk
;
960 while (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
962 * Scan the node to find the subtree index to push down into.
963 * We go one-past, then back-up.
965 * We must proactively remove deleted elements which may
966 * have been left over from a deadlocked btree_remove().
968 * The left and right boundaries are included in the loop
969 * in order to detect edge cases.
971 * If the separator only differs by create_tid (r == 1)
972 * and we are doing an as-of search, we may end up going
973 * down a branch to the left of the one containing the
974 * desired key. This requires numerous special cases.
976 ++hammer_stats_btree_iterations
;
977 if (hammer_debug_btree
) {
978 kprintf("SEARCH-I %016llx count=%d\n",
979 cursor
->node
->node_offset
,
984 * Try to shortcut the search before dropping into the
985 * linear loop. Locate the first node where r <= 1.
987 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
988 while (i
<= node
->count
) {
989 ++hammer_stats_btree_elements
;
990 elm
= &node
->elms
[i
];
991 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
992 if (hammer_debug_btree
> 2) {
993 kprintf(" IELM %p %d r=%d\n",
994 &node
->elms
[i
], i
, r
);
999 KKASSERT(elm
->base
.create_tid
!= 1);
1000 cursor
->create_check
= elm
->base
.create_tid
- 1;
1001 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
1005 if (hammer_debug_btree
) {
1006 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1011 * These cases occur when the parent's idea of the boundary
1012 * is wider then the child's idea of the boundary, and
1013 * require special handling. If not inserting we can
1014 * terminate the search early for these cases but the
1015 * child's boundaries cannot be unconditionally modified.
1019 * If i == 0 the search terminated to the LEFT of the
1020 * left_boundary but to the RIGHT of the parent's left
1025 elm
= &node
->elms
[0];
1028 * If we aren't inserting we can stop here.
1030 if ((flags
& (HAMMER_CURSOR_INSERT
|
1031 HAMMER_CURSOR_PRUNING
)) == 0) {
1037 * Correct a left-hand boundary mismatch.
1039 * We can only do this if we can upgrade the lock,
1040 * and synchronized as a background cursor (i.e.
1041 * inserting or pruning).
1043 * WARNING: We can only do this if inserting, i.e.
1044 * we are running on the backend.
1046 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1048 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1049 hammer_modify_node_field(cursor
->trans
, cursor
->node
,
1051 save
= node
->elms
[0].base
.btype
;
1052 node
->elms
[0].base
= *cursor
->left_bound
;
1053 node
->elms
[0].base
.btype
= save
;
1054 hammer_modify_node_done(cursor
->node
);
1055 } else if (i
== node
->count
+ 1) {
1057 * If i == node->count + 1 the search terminated to
1058 * the RIGHT of the right boundary but to the LEFT
1059 * of the parent's right boundary. If we aren't
1060 * inserting we can stop here.
1062 * Note that the last element in this case is
1063 * elms[i-2] prior to adjustments to 'i'.
1066 if ((flags
& (HAMMER_CURSOR_INSERT
|
1067 HAMMER_CURSOR_PRUNING
)) == 0) {
1073 * Correct a right-hand boundary mismatch.
1074 * (actual push-down record is i-2 prior to
1075 * adjustments to i).
1077 * We can only do this if we can upgrade the lock,
1078 * and synchronized as a background cursor (i.e.
1079 * inserting or pruning).
1081 * WARNING: We can only do this if inserting, i.e.
1082 * we are running on the backend.
1084 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1086 elm
= &node
->elms
[i
];
1087 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1088 hammer_modify_node(cursor
->trans
, cursor
->node
,
1089 &elm
->base
, sizeof(elm
->base
));
1090 elm
->base
= *cursor
->right_bound
;
1091 hammer_modify_node_done(cursor
->node
);
1095 * The push-down index is now i - 1. If we had
1096 * terminated on the right boundary this will point
1097 * us at the last element.
1102 elm
= &node
->elms
[i
];
1104 if (hammer_debug_btree
) {
1105 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1106 "key=%016llx cre=%016llx lo=%02x\n",
1107 cursor
->node
->node_offset
,
1109 elm
->internal
.base
.obj_id
,
1110 elm
->internal
.base
.rec_type
,
1111 elm
->internal
.base
.key
,
1112 elm
->internal
.base
.create_tid
,
1113 elm
->internal
.base
.localization
1118 * We better have a valid subtree offset.
1120 KKASSERT(elm
->internal
.subtree_offset
!= 0);
1123 * Handle insertion and deletion requirements.
1125 * If inserting split full nodes. The split code will
1126 * adjust cursor->node and cursor->index if the current
1127 * index winds up in the new node.
1129 * If inserting and a left or right edge case was detected,
1130 * we cannot correct the left or right boundary and must
1131 * prepend and append an empty leaf node in order to make
1132 * the boundary correction.
1134 * If we run out of space we set enospc and continue on
1135 * to a leaf to provide the spike code with a good point
1138 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
1139 if (btree_node_is_full(node
)) {
1140 error
= btree_split_internal(cursor
);
1142 if (error
!= ENOSPC
)
1147 * reload stale pointers
1150 node
= cursor
->node
->ondisk
;
1155 * Push down (push into new node, existing node becomes
1156 * the parent) and continue the search.
1158 error
= hammer_cursor_down(cursor
);
1159 /* node may have become stale */
1162 node
= cursor
->node
->ondisk
;
1166 * We are at a leaf, do a linear search of the key array.
1168 * On success the index is set to the matching element and 0
1171 * On failure the index is set to the insertion point and ENOENT
1174 * Boundaries are not stored in leaf nodes, so the index can wind
1175 * up to the left of element 0 (index == 0) or past the end of
1176 * the array (index == node->count). It is also possible that the
1177 * leaf might be empty.
1179 ++hammer_stats_btree_iterations
;
1180 KKASSERT (node
->type
== HAMMER_BTREE_TYPE_LEAF
);
1181 KKASSERT(node
->count
<= HAMMER_BTREE_LEAF_ELMS
);
1182 if (hammer_debug_btree
) {
1183 kprintf("SEARCH-L %016llx count=%d\n",
1184 cursor
->node
->node_offset
,
1189 * Try to shortcut the search before dropping into the
1190 * linear loop. Locate the first node where r <= 1.
1192 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
1193 while (i
< node
->count
) {
1194 ++hammer_stats_btree_elements
;
1195 elm
= &node
->elms
[i
];
1197 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->leaf
.base
);
1199 if (hammer_debug_btree
> 1)
1200 kprintf(" ELM %p %d r=%d\n", &node
->elms
[i
], i
, r
);
1203 * We are at a record element. Stop if we've flipped past
1204 * key_beg, not counting the create_tid test. Allow the
1205 * r == 1 case (key_beg > element but differs only by its
1206 * create_tid) to fall through to the AS-OF check.
1208 KKASSERT (elm
->leaf
.base
.btype
== HAMMER_BTREE_TYPE_RECORD
);
1218 * Check our as-of timestamp against the element.
1220 if (flags
& HAMMER_CURSOR_ASOF
) {
1221 if (hammer_btree_chkts(cursor
->asof
,
1222 &node
->elms
[i
].base
) != 0) {
1228 if (r
> 0) { /* can only be +1 */
1236 if (hammer_debug_btree
) {
1237 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1238 cursor
->node
->node_offset
, i
);
1244 * The search of the leaf node failed. i is the insertion point.
1247 if (hammer_debug_btree
) {
1248 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1249 cursor
->node
->node_offset
, i
);
1253 * No exact match was found, i is now at the insertion point.
1255 * If inserting split a full leaf before returning. This
1256 * may have the side effect of adjusting cursor->node and
1260 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0 &&
1261 btree_node_is_full(node
)) {
1262 error
= btree_split_leaf(cursor
);
1264 if (error
!= ENOSPC
)
1269 * reload stale pointers
1273 node = &cursor->node->internal;
1278 * We reached a leaf but did not find the key we were looking for.
1279 * If this is an insert we will be properly positioned for an insert
1280 * (ENOENT) or spike (ENOSPC) operation.
1282 error
= enospc
? ENOSPC
: ENOENT
;
1288 * Heuristical search for the first element whos comparison is <= 1. May
1289 * return an index whos compare result is > 1 but may only return an index
1290 * whos compare result is <= 1 if it is the first element with that result.
1293 hammer_btree_search_node(hammer_base_elm_t elm
, hammer_node_ondisk_t node
)
1301 * Don't bother if the node does not have very many elements
1306 i
= b
+ (s
- b
) / 2;
1307 ++hammer_stats_btree_elements
;
1308 r
= hammer_btree_cmp(elm
, &node
->elms
[i
].leaf
.base
);
1319 /************************************************************************
1320 * SPLITTING AND MERGING *
1321 ************************************************************************
1323 * These routines do all the dirty work required to split and merge nodes.
1327 * Split an internal node into two nodes and move the separator at the split
1328 * point to the parent.
1330 * (cursor->node, cursor->index) indicates the element the caller intends
1331 * to push into. We will adjust node and index if that element winds
1332 * up in the split node.
1334 * If we are at the root of the filesystem a new root must be created with
1335 * two elements, one pointing to the original root and one pointing to the
1336 * newly allocated split node.
1340 btree_split_internal(hammer_cursor_t cursor
)
1342 hammer_node_ondisk_t ondisk
;
1344 hammer_node_t parent
;
1345 hammer_node_t new_node
;
1346 hammer_btree_elm_t elm
;
1347 hammer_btree_elm_t parent_elm
;
1348 hammer_node_locklist_t locklist
= NULL
;
1349 hammer_mount_t hmp
= cursor
->trans
->hmp
;
1355 const int esize
= sizeof(*elm
);
1357 error
= hammer_btree_lock_children(cursor
, &locklist
);
1360 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1362 ++hammer_stats_btree_splits
;
1365 * We are splitting but elms[split] will be promoted to the parent,
1366 * leaving the right hand node with one less element. If the
1367 * insertion point will be on the left-hand side adjust the split
1368 * point to give the right hand side one additional node.
1370 node
= cursor
->node
;
1371 ondisk
= node
->ondisk
;
1372 split
= (ondisk
->count
+ 1) / 2;
1373 if (cursor
->index
<= split
)
1377 * If we are at the root of the filesystem, create a new root node
1378 * with 1 element and split normally. Avoid making major
1379 * modifications until we know the whole operation will work.
1381 if (ondisk
->parent
== 0) {
1382 parent
= hammer_alloc_btree(cursor
->trans
, &error
);
1385 hammer_lock_ex(&parent
->lock
);
1386 hammer_modify_node_noundo(cursor
->trans
, parent
);
1387 ondisk
= parent
->ondisk
;
1390 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1391 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1392 ondisk
->elms
[0].base
.btype
= node
->ondisk
->type
;
1393 ondisk
->elms
[0].internal
.subtree_offset
= node
->node_offset
;
1394 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1395 hammer_modify_node_done(parent
);
1396 /* ondisk->elms[1].base.btype - not used */
1398 parent_index
= 0; /* index of current node in parent */
1401 parent
= cursor
->parent
;
1402 parent_index
= cursor
->parent_index
;
1406 * Split node into new_node at the split point.
1408 * B O O O P N N B <-- P = node->elms[split]
1409 * 0 1 2 3 4 5 6 <-- subtree indices
1414 * B O O O B B N N B <--- inner boundary points are 'P'
1418 new_node
= hammer_alloc_btree(cursor
->trans
, &error
);
1419 if (new_node
== NULL
) {
1421 hammer_unlock(&parent
->lock
);
1422 hammer_delete_node(cursor
->trans
, parent
);
1423 hammer_rel_node(parent
);
1427 hammer_lock_ex(&new_node
->lock
);
1430 * Create the new node. P becomes the left-hand boundary in the
1431 * new node. Copy the right-hand boundary as well.
1433 * elm is the new separator.
1435 hammer_modify_node_noundo(cursor
->trans
, new_node
);
1436 hammer_modify_node_all(cursor
->trans
, node
);
1437 ondisk
= node
->ondisk
;
1438 elm
= &ondisk
->elms
[split
];
1439 bcopy(elm
, &new_node
->ondisk
->elms
[0],
1440 (ondisk
->count
- split
+ 1) * esize
);
1441 new_node
->ondisk
->count
= ondisk
->count
- split
;
1442 new_node
->ondisk
->parent
= parent
->node_offset
;
1443 new_node
->ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1444 KKASSERT(ondisk
->type
== new_node
->ondisk
->type
);
1447 * Cleanup the original node. Elm (P) becomes the new boundary,
1448 * its subtree_offset was moved to the new node. If we had created
1449 * a new root its parent pointer may have changed.
1451 elm
->internal
.subtree_offset
= 0;
1452 ondisk
->count
= split
;
1455 * Insert the separator into the parent, fixup the parent's
1456 * reference to the original node, and reference the new node.
1457 * The separator is P.
1459 * Remember that base.count does not include the right-hand boundary.
1461 hammer_modify_node_all(cursor
->trans
, parent
);
1462 ondisk
= parent
->ondisk
;
1463 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1464 parent_elm
= &ondisk
->elms
[parent_index
+1];
1465 bcopy(parent_elm
, parent_elm
+ 1,
1466 (ondisk
->count
- parent_index
) * esize
);
1467 parent_elm
->internal
.base
= elm
->base
; /* separator P */
1468 parent_elm
->internal
.base
.btype
= new_node
->ondisk
->type
;
1469 parent_elm
->internal
.subtree_offset
= new_node
->node_offset
;
1471 hammer_modify_node_done(parent
);
1474 * The children of new_node need their parent pointer set to new_node.
1475 * The children have already been locked by
1476 * hammer_btree_lock_children().
1478 for (i
= 0; i
< new_node
->ondisk
->count
; ++i
) {
1479 elm
= &new_node
->ondisk
->elms
[i
];
1480 error
= btree_set_parent(cursor
->trans
, new_node
, elm
);
1482 panic("btree_split_internal: btree-fixup problem");
1485 hammer_modify_node_done(new_node
);
1488 * The filesystem's root B-Tree pointer may have to be updated.
1491 hammer_volume_t volume
;
1493 volume
= hammer_get_root_volume(hmp
, &error
);
1494 KKASSERT(error
== 0);
1496 hammer_modify_volume_field(cursor
->trans
, volume
,
1498 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1499 hammer_modify_volume_done(volume
);
1500 node
->ondisk
->parent
= parent
->node_offset
;
1501 if (cursor
->parent
) {
1502 hammer_unlock(&cursor
->parent
->lock
);
1503 hammer_rel_node(cursor
->parent
);
1505 cursor
->parent
= parent
; /* lock'd and ref'd */
1506 hammer_rel_volume(volume
, 0);
1508 hammer_modify_node_done(node
);
1512 * Ok, now adjust the cursor depending on which element the original
1513 * index was pointing at. If we are >= the split point the push node
1514 * is now in the new node.
1516 * NOTE: If we are at the split point itself we cannot stay with the
1517 * original node because the push index will point at the right-hand
1518 * boundary, which is illegal.
1520 * NOTE: The cursor's parent or parent_index must be adjusted for
1521 * the case where a new parent (new root) was created, and the case
1522 * where the cursor is now pointing at the split node.
1524 if (cursor
->index
>= split
) {
1525 cursor
->parent_index
= parent_index
+ 1;
1526 cursor
->index
-= split
;
1527 hammer_unlock(&cursor
->node
->lock
);
1528 hammer_rel_node(cursor
->node
);
1529 cursor
->node
= new_node
; /* locked and ref'd */
1531 cursor
->parent_index
= parent_index
;
1532 hammer_unlock(&new_node
->lock
);
1533 hammer_rel_node(new_node
);
1537 * Fixup left and right bounds
1539 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1540 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1541 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1542 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1543 &cursor
->node
->ondisk
->elms
[0].internal
.base
) <= 0);
1544 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1545 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
].internal
.base
) >= 0);
1548 hammer_btree_unlock_children(&locklist
);
1549 hammer_cursor_downgrade(cursor
);
1554 * Same as the above, but splits a full leaf node.
1560 btree_split_leaf(hammer_cursor_t cursor
)
1562 hammer_node_ondisk_t ondisk
;
1563 hammer_node_t parent
;
1566 hammer_node_t new_leaf
;
1567 hammer_btree_elm_t elm
;
1568 hammer_btree_elm_t parent_elm
;
1569 hammer_base_elm_t mid_boundary
;
1574 const size_t esize
= sizeof(*elm
);
1576 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1578 ++hammer_stats_btree_splits
;
1580 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1581 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1582 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1583 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1586 * Calculate the split point. If the insertion point will be on
1587 * the left-hand side adjust the split point to give the right
1588 * hand side one additional node.
1590 * Spikes are made up of two leaf elements which cannot be
1593 leaf
= cursor
->node
;
1594 ondisk
= leaf
->ondisk
;
1595 split
= (ondisk
->count
+ 1) / 2;
1596 if (cursor
->index
<= split
)
1601 elm
= &ondisk
->elms
[split
];
1603 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
[-1].leaf
.base
) <= 0);
1604 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->leaf
.base
) <= 0);
1605 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->leaf
.base
) > 0);
1606 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
[1].leaf
.base
) > 0);
1609 * If we are at the root of the tree, create a new root node with
1610 * 1 element and split normally. Avoid making major modifications
1611 * until we know the whole operation will work.
1613 if (ondisk
->parent
== 0) {
1614 parent
= hammer_alloc_btree(cursor
->trans
, &error
);
1617 hammer_lock_ex(&parent
->lock
);
1618 hammer_modify_node_noundo(cursor
->trans
, parent
);
1619 ondisk
= parent
->ondisk
;
1622 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1623 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1624 ondisk
->elms
[0].base
.btype
= leaf
->ondisk
->type
;
1625 ondisk
->elms
[0].internal
.subtree_offset
= leaf
->node_offset
;
1626 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1627 /* ondisk->elms[1].base.btype = not used */
1628 hammer_modify_node_done(parent
);
1630 parent_index
= 0; /* insertion point in parent */
1633 parent
= cursor
->parent
;
1634 parent_index
= cursor
->parent_index
;
1638 * Split leaf into new_leaf at the split point. Select a separator
1639 * value in-between the two leafs but with a bent towards the right
1640 * leaf since comparisons use an 'elm >= separator' inequality.
1649 new_leaf
= hammer_alloc_btree(cursor
->trans
, &error
);
1650 if (new_leaf
== NULL
) {
1652 hammer_unlock(&parent
->lock
);
1653 hammer_delete_node(cursor
->trans
, parent
);
1654 hammer_rel_node(parent
);
1658 hammer_lock_ex(&new_leaf
->lock
);
1661 * Create the new node and copy the leaf elements from the split
1662 * point on to the new node.
1664 hammer_modify_node_all(cursor
->trans
, leaf
);
1665 hammer_modify_node_noundo(cursor
->trans
, new_leaf
);
1666 ondisk
= leaf
->ondisk
;
1667 elm
= &ondisk
->elms
[split
];
1668 bcopy(elm
, &new_leaf
->ondisk
->elms
[0], (ondisk
->count
- split
) * esize
);
1669 new_leaf
->ondisk
->count
= ondisk
->count
- split
;
1670 new_leaf
->ondisk
->parent
= parent
->node_offset
;
1671 new_leaf
->ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
1672 KKASSERT(ondisk
->type
== new_leaf
->ondisk
->type
);
1673 hammer_modify_node_done(new_leaf
);
1676 * Cleanup the original node. Because this is a leaf node and
1677 * leaf nodes do not have a right-hand boundary, there
1678 * aren't any special edge cases to clean up. We just fixup the
1681 ondisk
->count
= split
;
1684 * Insert the separator into the parent, fixup the parent's
1685 * reference to the original node, and reference the new node.
1686 * The separator is P.
1688 * Remember that base.count does not include the right-hand boundary.
1689 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1691 hammer_modify_node_all(cursor
->trans
, parent
);
1692 ondisk
= parent
->ondisk
;
1693 KKASSERT(split
!= 0);
1694 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1695 parent_elm
= &ondisk
->elms
[parent_index
+1];
1696 bcopy(parent_elm
, parent_elm
+ 1,
1697 (ondisk
->count
- parent_index
) * esize
);
1699 hammer_make_separator(&elm
[-1].base
, &elm
[0].base
, &parent_elm
->base
);
1700 parent_elm
->internal
.base
.btype
= new_leaf
->ondisk
->type
;
1701 parent_elm
->internal
.subtree_offset
= new_leaf
->node_offset
;
1702 mid_boundary
= &parent_elm
->base
;
1704 hammer_modify_node_done(parent
);
1707 * The filesystem's root B-Tree pointer may have to be updated.
1710 hammer_volume_t volume
;
1712 volume
= hammer_get_root_volume(hmp
, &error
);
1713 KKASSERT(error
== 0);
1715 hammer_modify_volume_field(cursor
->trans
, volume
,
1717 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1718 hammer_modify_volume_done(volume
);
1719 leaf
->ondisk
->parent
= parent
->node_offset
;
1720 if (cursor
->parent
) {
1721 hammer_unlock(&cursor
->parent
->lock
);
1722 hammer_rel_node(cursor
->parent
);
1724 cursor
->parent
= parent
; /* lock'd and ref'd */
1725 hammer_rel_volume(volume
, 0);
1727 hammer_modify_node_done(leaf
);
1730 * Ok, now adjust the cursor depending on which element the original
1731 * index was pointing at. If we are >= the split point the push node
1732 * is now in the new node.
1734 * NOTE: If we are at the split point itself we need to select the
1735 * old or new node based on where key_beg's insertion point will be.
1736 * If we pick the wrong side the inserted element will wind up in
1737 * the wrong leaf node and outside that node's bounds.
1739 if (cursor
->index
> split
||
1740 (cursor
->index
== split
&&
1741 hammer_btree_cmp(&cursor
->key_beg
, mid_boundary
) >= 0)) {
1742 cursor
->parent_index
= parent_index
+ 1;
1743 cursor
->index
-= split
;
1744 hammer_unlock(&cursor
->node
->lock
);
1745 hammer_rel_node(cursor
->node
);
1746 cursor
->node
= new_leaf
;
1748 cursor
->parent_index
= parent_index
;
1749 hammer_unlock(&new_leaf
->lock
);
1750 hammer_rel_node(new_leaf
);
1754 * Fixup left and right bounds
1756 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1757 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1758 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1761 * Assert that the bounds are correct.
1763 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1764 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1765 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1766 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1767 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &cursor
->key_beg
) <= 0);
1768 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &cursor
->key_beg
) > 0);
1771 hammer_cursor_downgrade(cursor
);
1776 * Recursively correct the right-hand boundary's create_tid to (tid) as
1777 * long as the rest of the key matches. We have to recurse upward in
1778 * the tree as well as down the left side of each parent's right node.
1780 * Return EDEADLK if we were only partially successful, forcing the caller
1781 * to try again. The original cursor is not modified. This routine can
1782 * also fail with EDEADLK if it is forced to throw away a portion of its
1785 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1788 TAILQ_ENTRY(hammer_rhb
) entry
;
1793 TAILQ_HEAD(hammer_rhb_list
, hammer_rhb
);
1796 hammer_btree_correct_rhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
1798 struct hammer_rhb_list rhb_list
;
1799 hammer_base_elm_t elm
;
1800 hammer_node_t orig_node
;
1801 struct hammer_rhb
*rhb
;
1805 TAILQ_INIT(&rhb_list
);
1808 * Save our position so we can restore it on return. This also
1809 * gives us a stable 'elm'.
1811 orig_node
= cursor
->node
;
1812 hammer_ref_node(orig_node
);
1813 hammer_lock_sh(&orig_node
->lock
);
1814 orig_index
= cursor
->index
;
1815 elm
= &orig_node
->ondisk
->elms
[orig_index
].base
;
1818 * Now build a list of parents going up, allocating a rhb
1819 * structure for each one.
1821 while (cursor
->parent
) {
1823 * Stop if we no longer have any right-bounds to fix up
1825 if (elm
->obj_id
!= cursor
->right_bound
->obj_id
||
1826 elm
->rec_type
!= cursor
->right_bound
->rec_type
||
1827 elm
->key
!= cursor
->right_bound
->key
) {
1832 * Stop if the right-hand bound's create_tid does not
1833 * need to be corrected.
1835 if (cursor
->right_bound
->create_tid
>= tid
)
1838 rhb
= kmalloc(sizeof(*rhb
), M_HAMMER
, M_WAITOK
|M_ZERO
);
1839 rhb
->node
= cursor
->parent
;
1840 rhb
->index
= cursor
->parent_index
;
1841 hammer_ref_node(rhb
->node
);
1842 hammer_lock_sh(&rhb
->node
->lock
);
1843 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
1845 hammer_cursor_up(cursor
);
1849 * now safely adjust the right hand bound for each rhb. This may
1850 * also require taking the right side of the tree and iterating down
1854 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1855 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
1858 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1859 hammer_unlock(&rhb
->node
->lock
);
1860 hammer_rel_node(rhb
->node
);
1861 kfree(rhb
, M_HAMMER
);
1863 switch (cursor
->node
->ondisk
->type
) {
1864 case HAMMER_BTREE_TYPE_INTERNAL
:
1866 * Right-boundary for parent at internal node
1867 * is one element to the right of the element whos
1868 * right boundary needs adjusting. We must then
1869 * traverse down the left side correcting any left
1870 * bounds (which may now be too far to the left).
1873 error
= hammer_btree_correct_lhb(cursor
, tid
);
1876 panic("hammer_btree_correct_rhb(): Bad node type");
1885 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1886 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1887 hammer_unlock(&rhb
->node
->lock
);
1888 hammer_rel_node(rhb
->node
);
1889 kfree(rhb
, M_HAMMER
);
1891 error
= hammer_cursor_seek(cursor
, orig_node
, orig_index
);
1892 hammer_unlock(&orig_node
->lock
);
1893 hammer_rel_node(orig_node
);
1898 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1899 * bound going downward starting at the current cursor position.
1901 * This function does not restore the cursor after use.
1904 hammer_btree_correct_lhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
1906 struct hammer_rhb_list rhb_list
;
1907 hammer_base_elm_t elm
;
1908 hammer_base_elm_t cmp
;
1909 struct hammer_rhb
*rhb
;
1912 TAILQ_INIT(&rhb_list
);
1914 cmp
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1917 * Record the node and traverse down the left-hand side for all
1918 * matching records needing a boundary correction.
1922 rhb
= kmalloc(sizeof(*rhb
), M_HAMMER
, M_WAITOK
|M_ZERO
);
1923 rhb
->node
= cursor
->node
;
1924 rhb
->index
= cursor
->index
;
1925 hammer_ref_node(rhb
->node
);
1926 hammer_lock_sh(&rhb
->node
->lock
);
1927 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
1929 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1931 * Nothing to traverse down if we are at the right
1932 * boundary of an internal node.
1934 if (cursor
->index
== cursor
->node
->ondisk
->count
)
1937 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1938 if (elm
->btype
== HAMMER_BTREE_TYPE_RECORD
)
1940 panic("Illegal leaf record type %02x", elm
->btype
);
1942 error
= hammer_cursor_down(cursor
);
1946 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1947 if (elm
->obj_id
!= cmp
->obj_id
||
1948 elm
->rec_type
!= cmp
->rec_type
||
1949 elm
->key
!= cmp
->key
) {
1952 if (elm
->create_tid
>= tid
)
1958 * Now we can safely adjust the left-hand boundary from the bottom-up.
1959 * The last element we remove from the list is the caller's right hand
1960 * boundary, which must also be adjusted.
1962 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1963 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
1966 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1967 hammer_unlock(&rhb
->node
->lock
);
1968 hammer_rel_node(rhb
->node
);
1969 kfree(rhb
, M_HAMMER
);
1971 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1972 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1973 hammer_modify_node(cursor
->trans
, cursor
->node
,
1975 sizeof(elm
->create_tid
));
1976 elm
->create_tid
= tid
;
1977 hammer_modify_node_done(cursor
->node
);
1979 panic("hammer_btree_correct_lhb(): Bad element type");
1986 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1987 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1988 hammer_unlock(&rhb
->node
->lock
);
1989 hammer_rel_node(rhb
->node
);
1990 kfree(rhb
, M_HAMMER
);
1996 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
1997 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
1998 * the operation due to a deadlock, or some other error.
2000 * This routine is always called with an empty, locked leaf but may recurse
2001 * into want-to-be-empty parents as part of its operation.
2003 * It should also be noted that when removing empty leaves we must be sure
2004 * to test and update mirror_tid because another thread may have deadlocked
2005 * against us (or someone) trying to propogate it up and cannot retry once
2006 * the node has been deleted.
2008 * On return the cursor may end up pointing to an internal node, suitable
2009 * for further iteration but not for an immediate insertion or deletion.
2012 btree_remove(hammer_cursor_t cursor
)
2014 hammer_node_ondisk_t ondisk
;
2015 hammer_btree_elm_t elm
;
2017 hammer_node_t parent
;
2018 const int esize
= sizeof(*elm
);
2021 node
= cursor
->node
;
2024 * When deleting the root of the filesystem convert it to
2025 * an empty leaf node. Internal nodes cannot be empty.
2027 ondisk
= node
->ondisk
;
2028 if (ondisk
->parent
== 0) {
2029 KKASSERT(cursor
->parent
== NULL
);
2030 hammer_modify_node_all(cursor
->trans
, node
);
2031 KKASSERT(ondisk
== node
->ondisk
);
2032 ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
2034 hammer_modify_node_done(node
);
2039 parent
= cursor
->parent
;
2042 * If another thread deadlocked trying to propogate mirror_tid up
2043 * we have to finish the job before deleting node. XXX
2045 if (parent
->ondisk
->mirror_tid
< node
->ondisk
->mirror_tid
&&
2046 (cursor
->trans
->hmp
->hflags
& (HMNT_MASTERID
|HMNT_SLAVE
))) {
2047 hammer_btree_mirror_propagate(cursor
->trans
,
2049 cursor
->parent_index
,
2050 node
->ondisk
->mirror_tid
);
2055 * Attempt to remove the parent's reference to the child. If the
2056 * parent would become empty we have to recurse. If we fail we
2057 * leave the parent pointing to an empty leaf node.
2059 if (parent
->ondisk
->count
== 1) {
2061 * This special cursor_up_locked() call leaves the original
2062 * node exclusively locked and referenced, leaves the
2063 * original parent locked (as the new node), and locks the
2064 * new parent. It can return EDEADLK.
2066 error
= hammer_cursor_up_locked(cursor
);
2068 error
= btree_remove(cursor
);
2070 hammer_modify_node_all(cursor
->trans
, node
);
2071 ondisk
= node
->ondisk
;
2072 ondisk
->type
= HAMMER_BTREE_TYPE_DELETED
;
2074 hammer_modify_node_done(node
);
2075 hammer_flush_node(node
);
2076 hammer_delete_node(cursor
->trans
, node
);
2078 kprintf("Warning: BTREE_REMOVE: Defering "
2079 "parent removal1 @ %016llx, skipping\n",
2082 hammer_unlock(&node
->lock
);
2083 hammer_rel_node(node
);
2085 kprintf("Warning: BTREE_REMOVE: Defering parent "
2086 "removal2 @ %016llx, skipping\n",
2090 KKASSERT(parent
->ondisk
->count
> 1);
2093 * Delete the subtree reference in the parent
2095 hammer_modify_node_all(cursor
->trans
, parent
);
2096 ondisk
= parent
->ondisk
;
2097 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
);
2099 elm
= &ondisk
->elms
[cursor
->parent_index
];
2100 KKASSERT(elm
->internal
.subtree_offset
== node
->node_offset
);
2101 KKASSERT(ondisk
->count
> 0);
2102 bcopy(&elm
[1], &elm
[0],
2103 (ondisk
->count
- cursor
->parent_index
) * esize
);
2105 hammer_modify_node_done(parent
);
2106 hammer_flush_node(node
);
2107 hammer_delete_node(cursor
->trans
, node
);
2110 * cursor->node is invalid, cursor up to make the cursor
2113 error
= hammer_cursor_up(cursor
);
2119 * Propagate a mirror TID update upwards through the B-Tree to the root.
2121 * A locked internal node must be passed in. The node will remain locked
2124 * This function syncs mirror_tid at the specified internal node's element,
2125 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2128 hammer_btree_mirror_propagate(hammer_transaction_t trans
, hammer_node_t node
,
2129 int index
, hammer_tid_t mirror_tid
)
2131 hammer_btree_internal_elm_t elm
;
2132 hammer_node_t parent
;
2136 KKASSERT (node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
);
2139 * Adjust the node's element
2141 elm
= &node
->ondisk
->elms
[index
].internal
;
2142 if (elm
->mirror_tid
>= mirror_tid
)
2144 hammer_modify_node(trans
, node
, &elm
->mirror_tid
,
2145 sizeof(elm
->mirror_tid
));
2146 elm
->mirror_tid
= mirror_tid
;
2147 hammer_modify_node_done(node
);
2150 * Adjust the node's mirror_tid aggragator
2152 if (node
->ondisk
->mirror_tid
>= mirror_tid
)
2154 hammer_modify_node_field(trans
, node
, mirror_tid
);
2155 node
->ondisk
->mirror_tid
= mirror_tid
;
2156 hammer_modify_node_done(node
);
2160 if (node
->ondisk
->parent
&&
2161 (trans
->hmp
->hflags
& (HMNT_MASTERID
|HMNT_SLAVE
))) {
2162 parent
= hammer_btree_get_parent(node
, &parent_index
,
2165 hammer_btree_mirror_propagate(trans
, parent
,
2166 parent_index
, mirror_tid
);
2167 hammer_unlock(&parent
->lock
);
2168 hammer_rel_node(parent
);
2175 hammer_btree_get_parent(hammer_node_t node
, int *parent_indexp
, int *errorp
,
2178 hammer_node_t parent
;
2179 hammer_btree_elm_t elm
;
2185 parent
= hammer_get_node(node
->hmp
, node
->ondisk
->parent
, 0, errorp
);
2187 KKASSERT(parent
== NULL
);
2190 KKASSERT ((parent
->flags
& HAMMER_NODE_DELETED
) == 0);
2195 if (try_exclusive
) {
2196 if (hammer_lock_ex_try(&parent
->lock
)) {
2197 hammer_rel_node(parent
);
2202 hammer_lock_sh(&parent
->lock
);
2206 * Figure out which element in the parent is pointing to the
2209 if (node
->ondisk
->count
) {
2210 i
= hammer_btree_search_node(&node
->ondisk
->elms
[0].base
,
2215 while (i
< parent
->ondisk
->count
) {
2216 elm
= &parent
->ondisk
->elms
[i
];
2217 if (elm
->internal
.subtree_offset
== node
->node_offset
)
2221 if (i
== parent
->ondisk
->count
) {
2222 hammer_unlock(&parent
->lock
);
2223 panic("Bad B-Tree link: parent %p node %p\n", parent
, node
);
2226 KKASSERT(*errorp
== 0);
2231 * The element (elm) has been moved to a new internal node (node).
2233 * If the element represents a pointer to an internal node that node's
2234 * parent must be adjusted to the element's new location.
2236 * XXX deadlock potential here with our exclusive locks
2239 btree_set_parent(hammer_transaction_t trans
, hammer_node_t node
,
2240 hammer_btree_elm_t elm
)
2242 hammer_node_t child
;
2247 switch(elm
->base
.btype
) {
2248 case HAMMER_BTREE_TYPE_INTERNAL
:
2249 case HAMMER_BTREE_TYPE_LEAF
:
2250 child
= hammer_get_node(node
->hmp
, elm
->internal
.subtree_offset
,
2253 hammer_modify_node_field(trans
, child
, parent
);
2254 child
->ondisk
->parent
= node
->node_offset
;
2255 hammer_modify_node_done(child
);
2256 hammer_rel_node(child
);
2266 * Exclusively lock all the children of node. This is used by the split
2267 * code to prevent anyone from accessing the children of a cursor node
2268 * while we fix-up its parent offset.
2270 * If we don't lock the children we can really mess up cursors which block
2271 * trying to cursor-up into our node.
2273 * On failure EDEADLK (or some other error) is returned. If a deadlock
2274 * error is returned the cursor is adjusted to block on termination.
2277 hammer_btree_lock_children(hammer_cursor_t cursor
,
2278 struct hammer_node_locklist
**locklistp
)
2281 hammer_node_locklist_t item
;
2282 hammer_node_ondisk_t ondisk
;
2283 hammer_btree_elm_t elm
;
2284 hammer_node_t child
;
2288 node
= cursor
->node
;
2289 ondisk
= node
->ondisk
;
2293 * We really do not want to block on I/O with exclusive locks held,
2294 * pre-get the children before trying to lock the mess.
2296 for (i
= 0; i
< ondisk
->count
; ++i
) {
2297 ++hammer_stats_btree_elements
;
2298 elm
= &ondisk
->elms
[i
];
2299 if (elm
->base
.btype
!= HAMMER_BTREE_TYPE_LEAF
&&
2300 elm
->base
.btype
!= HAMMER_BTREE_TYPE_INTERNAL
) {
2303 child
= hammer_get_node(node
->hmp
,
2304 elm
->internal
.subtree_offset
,
2307 hammer_rel_node(child
);
2313 for (i
= 0; error
== 0 && i
< ondisk
->count
; ++i
) {
2314 ++hammer_stats_btree_elements
;
2315 elm
= &ondisk
->elms
[i
];
2317 switch(elm
->base
.btype
) {
2318 case HAMMER_BTREE_TYPE_INTERNAL
:
2319 case HAMMER_BTREE_TYPE_LEAF
:
2320 KKASSERT(elm
->internal
.subtree_offset
!= 0);
2321 child
= hammer_get_node(node
->hmp
,
2322 elm
->internal
.subtree_offset
,
2330 if (hammer_lock_ex_try(&child
->lock
) != 0) {
2331 if (cursor
->deadlk_node
== NULL
) {
2332 cursor
->deadlk_node
= child
;
2333 hammer_ref_node(cursor
->deadlk_node
);
2336 hammer_rel_node(child
);
2338 item
= kmalloc(sizeof(*item
),
2339 M_HAMMER
, M_WAITOK
);
2340 item
->next
= *locklistp
;
2347 hammer_btree_unlock_children(locklistp
);
2353 * Release previously obtained node locks.
2356 hammer_btree_unlock_children(struct hammer_node_locklist
**locklistp
)
2358 hammer_node_locklist_t item
;
2360 while ((item
= *locklistp
) != NULL
) {
2361 *locklistp
= item
->next
;
2362 hammer_unlock(&item
->node
->lock
);
2363 hammer_rel_node(item
->node
);
2364 kfree(item
, M_HAMMER
);
2368 /************************************************************************
2369 * MISCELLANIOUS SUPPORT *
2370 ************************************************************************/
2373 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2375 * Note that for this particular function a return value of -1, 0, or +1
2376 * can denote a match if create_tid is otherwise discounted. A create_tid
2377 * of zero is considered to be 'infinity' in comparisons.
2379 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2382 hammer_btree_cmp(hammer_base_elm_t key1
, hammer_base_elm_t key2
)
2384 if (key1
->localization
< key2
->localization
)
2386 if (key1
->localization
> key2
->localization
)
2389 if (key1
->obj_id
< key2
->obj_id
)
2391 if (key1
->obj_id
> key2
->obj_id
)
2394 if (key1
->rec_type
< key2
->rec_type
)
2396 if (key1
->rec_type
> key2
->rec_type
)
2399 if (key1
->key
< key2
->key
)
2401 if (key1
->key
> key2
->key
)
2405 * A create_tid of zero indicates a record which is undeletable
2406 * and must be considered to have a value of positive infinity.
2408 if (key1
->create_tid
== 0) {
2409 if (key2
->create_tid
== 0)
2413 if (key2
->create_tid
== 0)
2415 if (key1
->create_tid
< key2
->create_tid
)
2417 if (key1
->create_tid
> key2
->create_tid
)
2423 * Test a timestamp against an element to determine whether the
2424 * element is visible. A timestamp of 0 means 'infinity'.
2427 hammer_btree_chkts(hammer_tid_t asof
, hammer_base_elm_t base
)
2430 if (base
->delete_tid
)
2434 if (asof
< base
->create_tid
)
2436 if (base
->delete_tid
&& asof
>= base
->delete_tid
)
2442 * Create a separator half way inbetween key1 and key2. For fields just
2443 * one unit apart, the separator will match key2. key1 is on the left-hand
2444 * side and key2 is on the right-hand side.
2446 * key2 must be >= the separator. It is ok for the separator to match key2.
2448 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2451 * NOTE: It might be beneficial to just scrap this whole mess and just
2452 * set the separator to key2.
2454 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2455 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2458 hammer_make_separator(hammer_base_elm_t key1
, hammer_base_elm_t key2
,
2459 hammer_base_elm_t dest
)
2461 bzero(dest
, sizeof(*dest
));
2463 dest
->rec_type
= key2
->rec_type
;
2464 dest
->key
= key2
->key
;
2465 dest
->obj_id
= key2
->obj_id
;
2466 dest
->create_tid
= key2
->create_tid
;
2468 MAKE_SEPARATOR(key1
, key2
, dest
, localization
);
2469 if (key1
->localization
== key2
->localization
) {
2470 MAKE_SEPARATOR(key1
, key2
, dest
, obj_id
);
2471 if (key1
->obj_id
== key2
->obj_id
) {
2472 MAKE_SEPARATOR(key1
, key2
, dest
, rec_type
);
2473 if (key1
->rec_type
== key2
->rec_type
) {
2474 MAKE_SEPARATOR(key1
, key2
, dest
, key
);
2476 * Don't bother creating a separator for
2477 * create_tid, which also conveniently avoids
2478 * having to handle the create_tid == 0
2479 * (infinity) case. Just leave create_tid
2482 * Worst case, dest matches key2 exactly,
2483 * which is acceptable.
2490 #undef MAKE_SEPARATOR
2493 * Return whether a generic internal or leaf node is full
2496 btree_node_is_full(hammer_node_ondisk_t node
)
2498 switch(node
->type
) {
2499 case HAMMER_BTREE_TYPE_INTERNAL
:
2500 if (node
->count
== HAMMER_BTREE_INT_ELMS
)
2503 case HAMMER_BTREE_TYPE_LEAF
:
2504 if (node
->count
== HAMMER_BTREE_LEAF_ELMS
)
2508 panic("illegal btree subtype");
2515 btree_max_elements(u_int8_t type
)
2517 if (type
== HAMMER_BTREE_TYPE_LEAF
)
2518 return(HAMMER_BTREE_LEAF_ELMS
);
2519 if (type
== HAMMER_BTREE_TYPE_INTERNAL
)
2520 return(HAMMER_BTREE_INT_ELMS
);
2521 panic("btree_max_elements: bad type %d\n", type
);
2526 hammer_print_btree_node(hammer_node_ondisk_t ondisk
)
2528 hammer_btree_elm_t elm
;
2531 kprintf("node %p count=%d parent=%016llx type=%c\n",
2532 ondisk
, ondisk
->count
, ondisk
->parent
, ondisk
->type
);
2535 * Dump both boundary elements if an internal node
2537 if (ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
2538 for (i
= 0; i
<= ondisk
->count
; ++i
) {
2539 elm
= &ondisk
->elms
[i
];
2540 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2543 for (i
= 0; i
< ondisk
->count
; ++i
) {
2544 elm
= &ondisk
->elms
[i
];
2545 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2551 hammer_print_btree_elm(hammer_btree_elm_t elm
, u_int8_t type
, int i
)
2554 kprintf("\tobj_id = %016llx\n", elm
->base
.obj_id
);
2555 kprintf("\tkey = %016llx\n", elm
->base
.key
);
2556 kprintf("\tcreate_tid = %016llx\n", elm
->base
.create_tid
);
2557 kprintf("\tdelete_tid = %016llx\n", elm
->base
.delete_tid
);
2558 kprintf("\trec_type = %04x\n", elm
->base
.rec_type
);
2559 kprintf("\tobj_type = %02x\n", elm
->base
.obj_type
);
2560 kprintf("\tbtype = %02x (%c)\n",
2562 (elm
->base
.btype
? elm
->base
.btype
: '?'));
2563 kprintf("\tlocalization = %02x\n", elm
->base
.localization
);
2566 case HAMMER_BTREE_TYPE_INTERNAL
:
2567 kprintf("\tsubtree_off = %016llx\n",
2568 elm
->internal
.subtree_offset
);
2570 case HAMMER_BTREE_TYPE_RECORD
:
2571 kprintf("\tdata_offset = %016llx\n", elm
->leaf
.data_offset
);
2572 kprintf("\tdata_len = %08x\n", elm
->leaf
.data_len
);
2573 kprintf("\tdata_crc = %08x\n", elm
->leaf
.data_crc
);