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 1.71 2008/07/13 09:32:48 dillon Exp $
40 * HAMMER implements a modified B+Tree. In documentation this will
41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
42 * looks like a B+Tree (A B-Tree which stores its records only at the leafs
43 * of the tree), but adds two additional boundary elements which describe
44 * the left-most and right-most element a node is able to represent. In
45 * otherwords, we have boundary elements at the two ends of a B-Tree node
46 * instead of sub-tree pointers.
48 * A B-Tree internal node looks like this:
50 * B N N N N N N B <-- boundary and internal elements
51 * S S S S S S S <-- subtree pointers
53 * A B-Tree leaf node basically looks like this:
55 * L L L L L L L L <-- leaf elemenets
57 * The radix for an internal node is 1 less then a leaf but we get a
58 * number of significant benefits for our troubles.
60 * The big benefit to using a B-Tree containing boundary information
61 * is that it is possible to cache pointers into the middle of the tree
62 * and not have to start searches, insertions, OR deletions at the root
63 * node. In particular, searches are able to progress in a definitive
64 * direction from any point in the tree without revisting nodes. This
65 * greatly improves the efficiency of many operations, most especially
68 * B-Trees also make the stacking of trees fairly straightforward.
70 * INSERTIONS: A search performed with the intention of doing
71 * an insert will guarantee that the terminal leaf node is not full by
72 * splitting full nodes. Splits occur top-down during the dive down the
75 * DELETIONS: A deletion makes no attempt to proactively balance the
76 * tree and will recursively remove nodes that become empty. If a
77 * deadlock occurs a deletion may not be able to remove an empty leaf.
78 * Deletions never allow internal nodes to become empty (that would blow
85 static int btree_search(hammer_cursor_t cursor
, int flags
);
86 static int btree_split_internal(hammer_cursor_t cursor
);
87 static int btree_split_leaf(hammer_cursor_t cursor
);
88 static int btree_remove(hammer_cursor_t cursor
);
89 static int btree_node_is_full(hammer_node_ondisk_t node
);
90 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor
,
91 hammer_tid_t mirror_tid
);
92 static void hammer_make_separator(hammer_base_elm_t key1
,
93 hammer_base_elm_t key2
, hammer_base_elm_t dest
);
94 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor
);
97 * Iterate records after a search. The cursor is iterated forwards past
98 * the current record until a record matching the key-range requirements
99 * is found. ENOENT is returned if the iteration goes past the ending
102 * The iteration is inclusive of key_beg and can be inclusive or exclusive
103 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
105 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
106 * may be modified by B-Tree functions.
108 * cursor->key_beg may or may not be modified by this function during
109 * the iteration. XXX future - in case of an inverted lock we may have
110 * to reinitiate the lookup and set key_beg to properly pick up where we
113 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
116 hammer_btree_iterate(hammer_cursor_t cursor
)
118 hammer_node_ondisk_t node
;
119 hammer_btree_elm_t elm
;
125 * Skip past the current record
127 node
= cursor
->node
->ondisk
;
130 if (cursor
->index
< node
->count
&&
131 (cursor
->flags
& HAMMER_CURSOR_ATEDISK
)) {
136 * Loop until an element is found or we are done.
140 * We iterate up the tree and then index over one element
141 * while we are at the last element in the current node.
143 * If we are at the root of the filesystem, cursor_up
146 * XXX this could be optimized by storing the information in
147 * the parent reference.
149 * XXX we can lose the node lock temporarily, this could mess
152 ++hammer_stats_btree_iterations
;
153 hammer_flusher_clean_loose_ios(cursor
->trans
->hmp
);
155 if (cursor
->index
== node
->count
) {
156 if (hammer_debug_btree
) {
157 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
158 cursor
->node
->node_offset
,
160 (cursor
->parent
? cursor
->parent
->node_offset
: -1),
161 cursor
->parent_index
,
164 KKASSERT(cursor
->parent
== NULL
|| cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.subtree_offset
== cursor
->node
->node_offset
);
165 error
= hammer_cursor_up(cursor
);
168 /* reload stale pointer */
169 node
= cursor
->node
->ondisk
;
170 KKASSERT(cursor
->index
!= node
->count
);
173 * If we are reblocking we want to return internal
176 if (cursor
->flags
& HAMMER_CURSOR_REBLOCKING
) {
177 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
185 * Check internal or leaf element. Determine if the record
186 * at the cursor has gone beyond the end of our range.
188 * We recurse down through internal nodes.
190 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
191 elm
= &node
->elms
[cursor
->index
];
193 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
194 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
195 if (hammer_debug_btree
) {
196 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
197 cursor
->node
->node_offset
,
199 elm
[0].internal
.base
.obj_id
,
200 elm
[0].internal
.base
.rec_type
,
201 elm
[0].internal
.base
.key
,
202 elm
[0].internal
.base
.localization
,
206 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
207 cursor
->node
->node_offset
,
209 elm
[1].internal
.base
.obj_id
,
210 elm
[1].internal
.base
.rec_type
,
211 elm
[1].internal
.base
.key
,
212 elm
[1].internal
.base
.localization
,
221 if (r
== 0 && (cursor
->flags
&
222 HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
231 KKASSERT(elm
->internal
.subtree_offset
!= 0);
234 * If running the mirror filter see if we can skip
235 * one or more entire sub-trees. If we can we
236 * return the internal mode and the caller processes
237 * the skipped range (see mirror_read)
239 if (cursor
->flags
& HAMMER_CURSOR_MIRROR_FILTERED
) {
240 if (elm
->internal
.mirror_tid
<
241 cursor
->cmirror
->mirror_tid
) {
242 hammer_cursor_mirror_filter(cursor
);
247 error
= hammer_cursor_down(cursor
);
250 KKASSERT(cursor
->index
== 0);
251 /* reload stale pointer */
252 node
= cursor
->node
->ondisk
;
255 elm
= &node
->elms
[cursor
->index
];
256 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
->base
);
257 if (hammer_debug_btree
) {
258 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
259 cursor
->node
->node_offset
,
261 (elm
[0].leaf
.base
.btype
?
262 elm
[0].leaf
.base
.btype
: '?'),
263 elm
[0].leaf
.base
.obj_id
,
264 elm
[0].leaf
.base
.rec_type
,
265 elm
[0].leaf
.base
.key
,
266 elm
[0].leaf
.base
.localization
,
276 * We support both end-inclusive and
277 * end-exclusive searches.
280 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
285 switch(elm
->leaf
.base
.btype
) {
286 case HAMMER_BTREE_TYPE_RECORD
:
287 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
288 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
301 * node pointer invalid after loop
307 if (hammer_debug_btree
) {
308 int i
= cursor
->index
;
309 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
310 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
312 elm
->internal
.base
.obj_id
,
313 elm
->internal
.base
.rec_type
,
314 elm
->internal
.base
.key
,
315 elm
->internal
.base
.localization
324 * We hit an internal element that we could skip as part of a mirroring
325 * scan. Calculate the entire range being skipped.
327 * It is important to include any gaps between the parent's left_bound
328 * and the node's left_bound, and same goes for the right side.
331 hammer_cursor_mirror_filter(hammer_cursor_t cursor
)
333 struct hammer_cmirror
*cmirror
;
334 hammer_node_ondisk_t ondisk
;
335 hammer_btree_elm_t elm
;
337 ondisk
= cursor
->node
->ondisk
;
338 cmirror
= cursor
->cmirror
;
341 * Calculate the skipped range
343 elm
= &ondisk
->elms
[cursor
->index
];
344 if (cursor
->index
== 0)
345 cmirror
->skip_beg
= *cursor
->left_bound
;
347 cmirror
->skip_beg
= elm
->internal
.base
;
348 while (cursor
->index
< ondisk
->count
) {
349 if (elm
->internal
.mirror_tid
>= cmirror
->mirror_tid
)
354 if (cursor
->index
== ondisk
->count
)
355 cmirror
->skip_end
= *cursor
->right_bound
;
357 cmirror
->skip_end
= elm
->internal
.base
;
360 * clip the returned result.
362 if (hammer_btree_cmp(&cmirror
->skip_beg
, &cursor
->key_beg
) < 0)
363 cmirror
->skip_beg
= cursor
->key_beg
;
364 if (hammer_btree_cmp(&cmirror
->skip_end
, &cursor
->key_end
) > 0)
365 cmirror
->skip_end
= cursor
->key_end
;
369 * Iterate in the reverse direction. This is used by the pruning code to
370 * avoid overlapping records.
373 hammer_btree_iterate_reverse(hammer_cursor_t cursor
)
375 hammer_node_ondisk_t node
;
376 hammer_btree_elm_t elm
;
381 /* mirror filtering not supported for reverse iteration */
382 KKASSERT ((cursor
->flags
& HAMMER_CURSOR_MIRROR_FILTERED
) == 0);
385 * Skip past the current record. For various reasons the cursor
386 * may end up set to -1 or set to point at the end of the current
387 * node. These cases must be addressed.
389 node
= cursor
->node
->ondisk
;
392 if (cursor
->index
!= -1 &&
393 (cursor
->flags
& HAMMER_CURSOR_ATEDISK
)) {
396 if (cursor
->index
== cursor
->node
->ondisk
->count
)
400 * Loop until an element is found or we are done.
403 ++hammer_stats_btree_iterations
;
404 hammer_flusher_clean_loose_ios(cursor
->trans
->hmp
);
407 * We iterate up the tree and then index over one element
408 * while we are at the last element in the current node.
410 if (cursor
->index
== -1) {
411 error
= hammer_cursor_up(cursor
);
413 cursor
->index
= 0; /* sanity */
416 /* reload stale pointer */
417 node
= cursor
->node
->ondisk
;
418 KKASSERT(cursor
->index
!= node
->count
);
424 * Check internal or leaf element. Determine if the record
425 * at the cursor has gone beyond the end of our range.
427 * We recurse down through internal nodes.
429 KKASSERT(cursor
->index
!= node
->count
);
430 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
431 elm
= &node
->elms
[cursor
->index
];
432 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
433 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
434 if (hammer_debug_btree
) {
435 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
436 cursor
->node
->node_offset
,
438 elm
[0].internal
.base
.obj_id
,
439 elm
[0].internal
.base
.rec_type
,
440 elm
[0].internal
.base
.key
,
441 elm
[0].internal
.base
.localization
,
444 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
445 cursor
->node
->node_offset
,
447 elm
[1].internal
.base
.obj_id
,
448 elm
[1].internal
.base
.rec_type
,
449 elm
[1].internal
.base
.key
,
450 elm
[1].internal
.base
.localization
,
464 KKASSERT(elm
->internal
.subtree_offset
!= 0);
466 error
= hammer_cursor_down(cursor
);
469 KKASSERT(cursor
->index
== 0);
470 /* reload stale pointer */
471 node
= cursor
->node
->ondisk
;
473 /* this can assign -1 if the leaf was empty */
474 cursor
->index
= node
->count
- 1;
477 elm
= &node
->elms
[cursor
->index
];
478 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
479 if (hammer_debug_btree
) {
480 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
481 cursor
->node
->node_offset
,
483 (elm
[0].leaf
.base
.btype
?
484 elm
[0].leaf
.base
.btype
: '?'),
485 elm
[0].leaf
.base
.obj_id
,
486 elm
[0].leaf
.base
.rec_type
,
487 elm
[0].leaf
.base
.key
,
488 elm
[0].leaf
.base
.localization
,
497 switch(elm
->leaf
.base
.btype
) {
498 case HAMMER_BTREE_TYPE_RECORD
:
499 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
500 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
513 * node pointer invalid after loop
519 if (hammer_debug_btree
) {
520 int i
= cursor
->index
;
521 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
522 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
524 elm
->internal
.base
.obj_id
,
525 elm
->internal
.base
.rec_type
,
526 elm
->internal
.base
.key
,
527 elm
->internal
.base
.localization
536 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
537 * could not be found, EDEADLK if inserting and a retry is needed, and a
538 * fatal error otherwise. When retrying, the caller must terminate the
539 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
541 * The cursor is suitably positioned for a deletion on success, and suitably
542 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
545 * The cursor may begin anywhere, the search will traverse the tree in
546 * either direction to locate the requested element.
548 * Most of the logic implementing historical searches is handled here. We
549 * do an initial lookup with create_tid set to the asof TID. Due to the
550 * way records are laid out, a backwards iteration may be required if
551 * ENOENT is returned to locate the historical record. Here's the
554 * create_tid: 10 15 20
558 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
559 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
560 * not visible and thus causes ENOENT to be returned. We really need
561 * to check record 11 in LEAF1. If it also fails then the search fails
562 * (e.g. it might represent the range 11-16 and thus still not match our
563 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
564 * further iterations.
566 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
567 * and the cursor->create_check TID if an iteration might be needed.
568 * In the above example create_check would be set to 14.
571 hammer_btree_lookup(hammer_cursor_t cursor
)
575 KKASSERT ((cursor
->flags
& HAMMER_CURSOR_INSERT
) == 0 ||
576 cursor
->trans
->sync_lock_refs
> 0);
577 ++hammer_stats_btree_lookups
;
578 if (cursor
->flags
& HAMMER_CURSOR_ASOF
) {
579 KKASSERT((cursor
->flags
& HAMMER_CURSOR_INSERT
) == 0);
580 cursor
->key_beg
.create_tid
= cursor
->asof
;
582 cursor
->flags
&= ~HAMMER_CURSOR_CREATE_CHECK
;
583 error
= btree_search(cursor
, 0);
584 if (error
!= ENOENT
||
585 (cursor
->flags
& HAMMER_CURSOR_CREATE_CHECK
) == 0) {
588 * Stop if error other then ENOENT.
589 * Stop if ENOENT and not special case.
593 if (hammer_debug_btree
) {
594 kprintf("CREATE_CHECK %016llx\n",
595 cursor
->create_check
);
597 cursor
->key_beg
.create_tid
= cursor
->create_check
;
601 error
= btree_search(cursor
, 0);
604 error
= hammer_btree_extract(cursor
, cursor
->flags
);
609 * Execute the logic required to start an iteration. The first record
610 * located within the specified range is returned and iteration control
611 * flags are adjusted for successive hammer_btree_iterate() calls.
614 hammer_btree_first(hammer_cursor_t cursor
)
618 error
= hammer_btree_lookup(cursor
);
619 if (error
== ENOENT
) {
620 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
621 error
= hammer_btree_iterate(cursor
);
623 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
628 * Similarly but for an iteration in the reverse direction.
630 * Set ATEDISK when iterating backwards to skip the current entry,
631 * which after an ENOENT lookup will be pointing beyond our end point.
634 hammer_btree_last(hammer_cursor_t cursor
)
636 struct hammer_base_elm save
;
639 save
= cursor
->key_beg
;
640 cursor
->key_beg
= cursor
->key_end
;
641 error
= hammer_btree_lookup(cursor
);
642 cursor
->key_beg
= save
;
643 if (error
== ENOENT
||
644 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
645 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
646 error
= hammer_btree_iterate_reverse(cursor
);
648 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
653 * Extract the record and/or data associated with the cursor's current
654 * position. Any prior record or data stored in the cursor is replaced.
655 * The cursor must be positioned at a leaf node.
657 * NOTE: All extractions occur at the leaf of the B-Tree.
660 hammer_btree_extract(hammer_cursor_t cursor
, int flags
)
663 hammer_node_ondisk_t node
;
664 hammer_btree_elm_t elm
;
665 hammer_off_t data_off
;
670 * The case where the data reference resolves to the same buffer
671 * as the record reference must be handled.
673 node
= cursor
->node
->ondisk
;
674 elm
= &node
->elms
[cursor
->index
];
676 hmp
= cursor
->node
->hmp
;
679 * There is nothing to extract for an internal element.
681 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
)
685 * Only record types have data.
687 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
688 cursor
->leaf
= &elm
->leaf
;
690 if ((flags
& HAMMER_CURSOR_GET_DATA
) == 0)
692 if (elm
->leaf
.base
.btype
!= HAMMER_BTREE_TYPE_RECORD
)
694 data_off
= elm
->leaf
.data_offset
;
695 data_len
= elm
->leaf
.data_len
;
702 KKASSERT(data_len
>= 0 && data_len
<= HAMMER_XBUFSIZE
);
703 cursor
->data
= hammer_bread_ext(hmp
, data_off
, data_len
,
704 &error
, &cursor
->data_buffer
);
705 if (hammer_crc_test_leaf(cursor
->data
, &elm
->leaf
) == 0)
706 Debugger("CRC FAILED: DATA");
712 * Insert a leaf element into the B-Tree at the current cursor position.
713 * The cursor is positioned such that the element at and beyond the cursor
714 * are shifted to make room for the new record.
716 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
717 * flag set and that call must return ENOENT before this function can be
720 * The caller may depend on the cursor's exclusive lock after return to
721 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
723 * ENOSPC is returned if there is no room to insert a new record.
726 hammer_btree_insert(hammer_cursor_t cursor
, hammer_btree_leaf_elm_t elm
,
729 hammer_node_ondisk_t node
;
734 if ((error
= hammer_cursor_upgrade_node(cursor
)) != 0)
736 ++hammer_stats_btree_inserts
;
739 * Insert the element at the leaf node and update the count in the
740 * parent. It is possible for parent to be NULL, indicating that
741 * the filesystem's ROOT B-Tree node is a leaf itself, which is
742 * possible. The root inode can never be deleted so the leaf should
745 * Remember that the right-hand boundary is not included in the
748 hammer_modify_node_all(cursor
->trans
, cursor
->node
);
749 node
= cursor
->node
->ondisk
;
751 KKASSERT(elm
->base
.btype
!= 0);
752 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
753 KKASSERT(node
->count
< HAMMER_BTREE_LEAF_ELMS
);
754 if (i
!= node
->count
) {
755 bcopy(&node
->elms
[i
], &node
->elms
[i
+1],
756 (node
->count
- i
) * sizeof(*elm
));
758 node
->elms
[i
].leaf
= *elm
;
760 hammer_cursor_inserted_element(cursor
->node
, i
);
763 * Update the leaf node's aggregate mirror_tid for mirroring
766 if (node
->mirror_tid
< elm
->base
.delete_tid
) {
767 node
->mirror_tid
= elm
->base
.delete_tid
;
770 if (node
->mirror_tid
< elm
->base
.create_tid
) {
771 node
->mirror_tid
= elm
->base
.create_tid
;
774 hammer_modify_node_done(cursor
->node
);
777 * Debugging sanity checks.
779 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->base
) <= 0);
780 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->base
) > 0);
782 KKASSERT(hammer_btree_cmp(&node
->elms
[i
-1].leaf
.base
, &elm
->base
) < 0);
784 if (i
!= node
->count
- 1)
785 KKASSERT(hammer_btree_cmp(&node
->elms
[i
+1].leaf
.base
, &elm
->base
) > 0);
791 * Delete a record from the B-Tree at the current cursor position.
792 * The cursor is positioned such that the current element is the one
795 * On return the cursor will be positioned after the deleted element and
796 * MAY point to an internal node. It will be suitable for the continuation
797 * of an iteration but not for an insertion or deletion.
799 * Deletions will attempt to partially rebalance the B-Tree in an upward
800 * direction, but will terminate rather then deadlock. Empty internal nodes
801 * are never allowed by a deletion which deadlocks may end up giving us an
802 * empty leaf. The pruner will clean up and rebalance the tree.
804 * This function can return EDEADLK, requiring the caller to retry the
805 * operation after clearing the deadlock.
808 hammer_btree_delete(hammer_cursor_t cursor
)
810 hammer_node_ondisk_t ondisk
;
812 hammer_node_t parent
;
816 KKASSERT (cursor
->trans
->sync_lock_refs
> 0);
817 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
819 ++hammer_stats_btree_deletes
;
822 * Delete the element from the leaf node.
824 * Remember that leaf nodes do not have boundaries.
827 ondisk
= node
->ondisk
;
830 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_LEAF
);
831 KKASSERT(i
>= 0 && i
< ondisk
->count
);
832 hammer_modify_node_all(cursor
->trans
, node
);
833 if (i
+ 1 != ondisk
->count
) {
834 bcopy(&ondisk
->elms
[i
+1], &ondisk
->elms
[i
],
835 (ondisk
->count
- i
- 1) * sizeof(ondisk
->elms
[0]));
838 hammer_modify_node_done(node
);
839 hammer_cursor_deleted_element(node
, i
);
842 * Validate local parent
844 if (ondisk
->parent
) {
845 parent
= cursor
->parent
;
847 KKASSERT(parent
!= NULL
);
848 KKASSERT(parent
->node_offset
== ondisk
->parent
);
852 * If the leaf becomes empty it must be detached from the parent,
853 * potentially recursing through to the filesystem root.
855 * This may reposition the cursor at one of the parent's of the
858 * Ignore deadlock errors, that simply means that btree_remove
859 * was unable to recurse and had to leave us with an empty leaf.
861 KKASSERT(cursor
->index
<= ondisk
->count
);
862 if (ondisk
->count
== 0) {
863 error
= btree_remove(cursor
);
864 if (error
== EDEADLK
)
869 KKASSERT(cursor
->parent
== NULL
||
870 cursor
->parent_index
< cursor
->parent
->ondisk
->count
);
875 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
877 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
879 * The search can begin ANYWHERE in the B-Tree. As a first step the search
880 * iterates up the tree as necessary to properly position itself prior to
881 * actually doing the sarch.
883 * INSERTIONS: The search will split full nodes and leaves on its way down
884 * and guarentee that the leaf it ends up on is not full. If we run out
885 * of space the search continues to the leaf (to position the cursor for
886 * the spike), but ENOSPC is returned.
888 * The search is only guarenteed to end up on a leaf if an error code of 0
889 * is returned, or if inserting and an error code of ENOENT is returned.
890 * Otherwise it can stop at an internal node. On success a search returns
893 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
894 * filesystem, and it is not simple code. Please note the following facts:
896 * - Internal node recursions have a boundary on the left AND right. The
897 * right boundary is non-inclusive. The create_tid is a generic part
898 * of the key for internal nodes.
900 * - Leaf nodes contain terminal elements only now.
902 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
903 * historical search. ASOF and INSERT are mutually exclusive. When
904 * doing an as-of lookup btree_search() checks for a right-edge boundary
905 * case. If while recursing down the left-edge differs from the key
906 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
907 * with cursor->create_check. This is used by btree_lookup() to iterate.
908 * The iteration backwards because as-of searches can wind up going
909 * down the wrong branch of the B-Tree.
913 btree_search(hammer_cursor_t cursor
, int flags
)
915 hammer_node_ondisk_t node
;
916 hammer_btree_elm_t elm
;
923 flags
|= cursor
->flags
;
924 ++hammer_stats_btree_searches
;
926 if (hammer_debug_btree
) {
927 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
928 cursor
->node
->node_offset
,
930 cursor
->key_beg
.obj_id
,
931 cursor
->key_beg
.rec_type
,
933 cursor
->key_beg
.create_tid
,
934 cursor
->key_beg
.localization
,
938 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
939 cursor
->parent
->node_offset
, cursor
->parent_index
,
940 cursor
->left_bound
->obj_id
,
941 cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.base
.obj_id
,
942 cursor
->right_bound
->obj_id
,
943 cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1].internal
.base
.obj_id
,
945 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
],
947 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1]
952 * Move our cursor up the tree until we find a node whos range covers
953 * the key we are trying to locate.
955 * The left bound is inclusive, the right bound is non-inclusive.
956 * It is ok to cursor up too far.
959 r
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->left_bound
);
960 s
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->right_bound
);
963 KKASSERT(cursor
->parent
);
964 ++hammer_stats_btree_iterations
;
965 error
= hammer_cursor_up(cursor
);
971 * The delete-checks below are based on node, not parent. Set the
972 * initial delete-check based on the parent.
975 KKASSERT(cursor
->left_bound
->create_tid
!= 1);
976 cursor
->create_check
= cursor
->left_bound
->create_tid
- 1;
977 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
981 * We better have ended up with a node somewhere.
983 KKASSERT(cursor
->node
!= NULL
);
986 * If we are inserting we can't start at a full node if the parent
987 * is also full (because there is no way to split the node),
988 * continue running up the tree until the requirement is satisfied
989 * or we hit the root of the filesystem.
991 * (If inserting we aren't doing an as-of search so we don't have
992 * to worry about create_check).
994 while ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
995 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
996 if (btree_node_is_full(cursor
->node
->ondisk
) == 0)
999 if (btree_node_is_full(cursor
->node
->ondisk
) ==0)
1002 if (cursor
->node
->ondisk
->parent
== 0 ||
1003 cursor
->parent
->ondisk
->count
!= HAMMER_BTREE_INT_ELMS
) {
1006 ++hammer_stats_btree_iterations
;
1007 error
= hammer_cursor_up(cursor
);
1008 /* node may have become stale */
1014 * Push down through internal nodes to locate the requested key.
1016 node
= cursor
->node
->ondisk
;
1017 while (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1019 * Scan the node to find the subtree index to push down into.
1020 * We go one-past, then back-up.
1022 * We must proactively remove deleted elements which may
1023 * have been left over from a deadlocked btree_remove().
1025 * The left and right boundaries are included in the loop
1026 * in order to detect edge cases.
1028 * If the separator only differs by create_tid (r == 1)
1029 * and we are doing an as-of search, we may end up going
1030 * down a branch to the left of the one containing the
1031 * desired key. This requires numerous special cases.
1033 ++hammer_stats_btree_iterations
;
1034 if (hammer_debug_btree
) {
1035 kprintf("SEARCH-I %016llx count=%d\n",
1036 cursor
->node
->node_offset
,
1041 * Try to shortcut the search before dropping into the
1042 * linear loop. Locate the first node where r <= 1.
1044 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
1045 while (i
<= node
->count
) {
1046 ++hammer_stats_btree_elements
;
1047 elm
= &node
->elms
[i
];
1048 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
1049 if (hammer_debug_btree
> 2) {
1050 kprintf(" IELM %p %d r=%d\n",
1051 &node
->elms
[i
], i
, r
);
1056 KKASSERT(elm
->base
.create_tid
!= 1);
1057 cursor
->create_check
= elm
->base
.create_tid
- 1;
1058 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
1062 if (hammer_debug_btree
) {
1063 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1068 * These cases occur when the parent's idea of the boundary
1069 * is wider then the child's idea of the boundary, and
1070 * require special handling. If not inserting we can
1071 * terminate the search early for these cases but the
1072 * child's boundaries cannot be unconditionally modified.
1076 * If i == 0 the search terminated to the LEFT of the
1077 * left_boundary but to the RIGHT of the parent's left
1082 elm
= &node
->elms
[0];
1085 * If we aren't inserting we can stop here.
1087 if ((flags
& (HAMMER_CURSOR_INSERT
|
1088 HAMMER_CURSOR_PRUNING
)) == 0) {
1094 * Correct a left-hand boundary mismatch.
1096 * We can only do this if we can upgrade the lock,
1097 * and synchronized as a background cursor (i.e.
1098 * inserting or pruning).
1100 * WARNING: We can only do this if inserting, i.e.
1101 * we are running on the backend.
1103 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1105 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1106 hammer_modify_node_field(cursor
->trans
, cursor
->node
,
1108 save
= node
->elms
[0].base
.btype
;
1109 node
->elms
[0].base
= *cursor
->left_bound
;
1110 node
->elms
[0].base
.btype
= save
;
1111 hammer_modify_node_done(cursor
->node
);
1112 } else if (i
== node
->count
+ 1) {
1114 * If i == node->count + 1 the search terminated to
1115 * the RIGHT of the right boundary but to the LEFT
1116 * of the parent's right boundary. If we aren't
1117 * inserting we can stop here.
1119 * Note that the last element in this case is
1120 * elms[i-2] prior to adjustments to 'i'.
1123 if ((flags
& (HAMMER_CURSOR_INSERT
|
1124 HAMMER_CURSOR_PRUNING
)) == 0) {
1130 * Correct a right-hand boundary mismatch.
1131 * (actual push-down record is i-2 prior to
1132 * adjustments to i).
1134 * We can only do this if we can upgrade the lock,
1135 * and synchronized as a background cursor (i.e.
1136 * inserting or pruning).
1138 * WARNING: We can only do this if inserting, i.e.
1139 * we are running on the backend.
1141 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1143 elm
= &node
->elms
[i
];
1144 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1145 hammer_modify_node(cursor
->trans
, cursor
->node
,
1146 &elm
->base
, sizeof(elm
->base
));
1147 elm
->base
= *cursor
->right_bound
;
1148 hammer_modify_node_done(cursor
->node
);
1152 * The push-down index is now i - 1. If we had
1153 * terminated on the right boundary this will point
1154 * us at the last element.
1159 elm
= &node
->elms
[i
];
1161 if (hammer_debug_btree
) {
1162 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1163 "key=%016llx cre=%016llx lo=%02x\n",
1164 cursor
->node
->node_offset
,
1166 elm
->internal
.base
.obj_id
,
1167 elm
->internal
.base
.rec_type
,
1168 elm
->internal
.base
.key
,
1169 elm
->internal
.base
.create_tid
,
1170 elm
->internal
.base
.localization
1175 * We better have a valid subtree offset.
1177 KKASSERT(elm
->internal
.subtree_offset
!= 0);
1180 * Handle insertion and deletion requirements.
1182 * If inserting split full nodes. The split code will
1183 * adjust cursor->node and cursor->index if the current
1184 * index winds up in the new node.
1186 * If inserting and a left or right edge case was detected,
1187 * we cannot correct the left or right boundary and must
1188 * prepend and append an empty leaf node in order to make
1189 * the boundary correction.
1191 * If we run out of space we set enospc and continue on
1192 * to a leaf to provide the spike code with a good point
1195 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
1196 if (btree_node_is_full(node
)) {
1197 error
= btree_split_internal(cursor
);
1199 if (error
!= ENOSPC
)
1204 * reload stale pointers
1207 node
= cursor
->node
->ondisk
;
1212 * Push down (push into new node, existing node becomes
1213 * the parent) and continue the search.
1215 error
= hammer_cursor_down(cursor
);
1216 /* node may have become stale */
1219 node
= cursor
->node
->ondisk
;
1223 * We are at a leaf, do a linear search of the key array.
1225 * On success the index is set to the matching element and 0
1228 * On failure the index is set to the insertion point and ENOENT
1231 * Boundaries are not stored in leaf nodes, so the index can wind
1232 * up to the left of element 0 (index == 0) or past the end of
1233 * the array (index == node->count). It is also possible that the
1234 * leaf might be empty.
1236 ++hammer_stats_btree_iterations
;
1237 KKASSERT (node
->type
== HAMMER_BTREE_TYPE_LEAF
);
1238 KKASSERT(node
->count
<= HAMMER_BTREE_LEAF_ELMS
);
1239 if (hammer_debug_btree
) {
1240 kprintf("SEARCH-L %016llx count=%d\n",
1241 cursor
->node
->node_offset
,
1246 * Try to shortcut the search before dropping into the
1247 * linear loop. Locate the first node where r <= 1.
1249 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
1250 while (i
< node
->count
) {
1251 ++hammer_stats_btree_elements
;
1252 elm
= &node
->elms
[i
];
1254 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->leaf
.base
);
1256 if (hammer_debug_btree
> 1)
1257 kprintf(" ELM %p %d r=%d\n", &node
->elms
[i
], i
, r
);
1260 * We are at a record element. Stop if we've flipped past
1261 * key_beg, not counting the create_tid test. Allow the
1262 * r == 1 case (key_beg > element but differs only by its
1263 * create_tid) to fall through to the AS-OF check.
1265 KKASSERT (elm
->leaf
.base
.btype
== HAMMER_BTREE_TYPE_RECORD
);
1275 * Check our as-of timestamp against the element.
1277 if (flags
& HAMMER_CURSOR_ASOF
) {
1278 if (hammer_btree_chkts(cursor
->asof
,
1279 &node
->elms
[i
].base
) != 0) {
1285 if (r
> 0) { /* can only be +1 */
1293 if (hammer_debug_btree
) {
1294 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1295 cursor
->node
->node_offset
, i
);
1301 * The search of the leaf node failed. i is the insertion point.
1304 if (hammer_debug_btree
) {
1305 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1306 cursor
->node
->node_offset
, i
);
1310 * No exact match was found, i is now at the insertion point.
1312 * If inserting split a full leaf before returning. This
1313 * may have the side effect of adjusting cursor->node and
1317 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0 &&
1318 btree_node_is_full(node
)) {
1319 error
= btree_split_leaf(cursor
);
1321 if (error
!= ENOSPC
)
1326 * reload stale pointers
1330 node = &cursor->node->internal;
1335 * We reached a leaf but did not find the key we were looking for.
1336 * If this is an insert we will be properly positioned for an insert
1337 * (ENOENT) or spike (ENOSPC) operation.
1339 error
= enospc
? ENOSPC
: ENOENT
;
1345 * Heuristical search for the first element whos comparison is <= 1. May
1346 * return an index whos compare result is > 1 but may only return an index
1347 * whos compare result is <= 1 if it is the first element with that result.
1350 hammer_btree_search_node(hammer_base_elm_t elm
, hammer_node_ondisk_t node
)
1358 * Don't bother if the node does not have very many elements
1363 i
= b
+ (s
- b
) / 2;
1364 ++hammer_stats_btree_elements
;
1365 r
= hammer_btree_cmp(elm
, &node
->elms
[i
].leaf
.base
);
1376 /************************************************************************
1377 * SPLITTING AND MERGING *
1378 ************************************************************************
1380 * These routines do all the dirty work required to split and merge nodes.
1384 * Split an internal node into two nodes and move the separator at the split
1385 * point to the parent.
1387 * (cursor->node, cursor->index) indicates the element the caller intends
1388 * to push into. We will adjust node and index if that element winds
1389 * up in the split node.
1391 * If we are at the root of the filesystem a new root must be created with
1392 * two elements, one pointing to the original root and one pointing to the
1393 * newly allocated split node.
1397 btree_split_internal(hammer_cursor_t cursor
)
1399 hammer_node_ondisk_t ondisk
;
1401 hammer_node_t parent
;
1402 hammer_node_t new_node
;
1403 hammer_btree_elm_t elm
;
1404 hammer_btree_elm_t parent_elm
;
1405 hammer_node_locklist_t locklist
= NULL
;
1406 hammer_mount_t hmp
= cursor
->trans
->hmp
;
1412 const int esize
= sizeof(*elm
);
1414 error
= hammer_btree_lock_children(cursor
, &locklist
);
1417 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1419 ++hammer_stats_btree_splits
;
1422 * We are splitting but elms[split] will be promoted to the parent,
1423 * leaving the right hand node with one less element. If the
1424 * insertion point will be on the left-hand side adjust the split
1425 * point to give the right hand side one additional node.
1427 node
= cursor
->node
;
1428 ondisk
= node
->ondisk
;
1429 split
= (ondisk
->count
+ 1) / 2;
1430 if (cursor
->index
<= split
)
1434 * If we are at the root of the filesystem, create a new root node
1435 * with 1 element and split normally. Avoid making major
1436 * modifications until we know the whole operation will work.
1438 if (ondisk
->parent
== 0) {
1439 parent
= hammer_alloc_btree(cursor
->trans
, &error
);
1442 hammer_lock_ex(&parent
->lock
);
1443 hammer_modify_node_noundo(cursor
->trans
, parent
);
1444 ondisk
= parent
->ondisk
;
1447 ondisk
->mirror_tid
= node
->ondisk
->mirror_tid
;
1448 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1449 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1450 ondisk
->elms
[0].base
.btype
= node
->ondisk
->type
;
1451 ondisk
->elms
[0].internal
.subtree_offset
= node
->node_offset
;
1452 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1453 hammer_modify_node_done(parent
);
1454 /* ondisk->elms[1].base.btype - not used */
1456 parent_index
= 0; /* index of current node in parent */
1459 parent
= cursor
->parent
;
1460 parent_index
= cursor
->parent_index
;
1464 * Split node into new_node at the split point.
1466 * B O O O P N N B <-- P = node->elms[split]
1467 * 0 1 2 3 4 5 6 <-- subtree indices
1472 * B O O O B B N N B <--- inner boundary points are 'P'
1476 new_node
= hammer_alloc_btree(cursor
->trans
, &error
);
1477 if (new_node
== NULL
) {
1479 hammer_unlock(&parent
->lock
);
1480 hammer_delete_node(cursor
->trans
, parent
);
1481 hammer_rel_node(parent
);
1485 hammer_lock_ex(&new_node
->lock
);
1488 * Create the new node. P becomes the left-hand boundary in the
1489 * new node. Copy the right-hand boundary as well.
1491 * elm is the new separator.
1493 hammer_modify_node_noundo(cursor
->trans
, new_node
);
1494 hammer_modify_node_all(cursor
->trans
, node
);
1495 ondisk
= node
->ondisk
;
1496 elm
= &ondisk
->elms
[split
];
1497 bcopy(elm
, &new_node
->ondisk
->elms
[0],
1498 (ondisk
->count
- split
+ 1) * esize
);
1499 new_node
->ondisk
->count
= ondisk
->count
- split
;
1500 new_node
->ondisk
->parent
= parent
->node_offset
;
1501 new_node
->ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1502 new_node
->ondisk
->mirror_tid
= ondisk
->mirror_tid
;
1503 KKASSERT(ondisk
->type
== new_node
->ondisk
->type
);
1504 hammer_cursor_split_node(node
, new_node
, split
);
1507 * Cleanup the original node. Elm (P) becomes the new boundary,
1508 * its subtree_offset was moved to the new node. If we had created
1509 * a new root its parent pointer may have changed.
1511 elm
->internal
.subtree_offset
= 0;
1512 ondisk
->count
= split
;
1515 * Insert the separator into the parent, fixup the parent's
1516 * reference to the original node, and reference the new node.
1517 * The separator is P.
1519 * Remember that base.count does not include the right-hand boundary.
1521 hammer_modify_node_all(cursor
->trans
, parent
);
1522 ondisk
= parent
->ondisk
;
1523 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1524 parent_elm
= &ondisk
->elms
[parent_index
+1];
1525 bcopy(parent_elm
, parent_elm
+ 1,
1526 (ondisk
->count
- parent_index
) * esize
);
1527 parent_elm
->internal
.base
= elm
->base
; /* separator P */
1528 parent_elm
->internal
.base
.btype
= new_node
->ondisk
->type
;
1529 parent_elm
->internal
.subtree_offset
= new_node
->node_offset
;
1530 parent_elm
->internal
.mirror_tid
= new_node
->ondisk
->mirror_tid
;
1532 hammer_modify_node_done(parent
);
1533 hammer_cursor_inserted_element(parent
, parent_index
+ 1);
1536 * The children of new_node need their parent pointer set to new_node.
1537 * The children have already been locked by
1538 * hammer_btree_lock_children().
1540 for (i
= 0; i
< new_node
->ondisk
->count
; ++i
) {
1541 elm
= &new_node
->ondisk
->elms
[i
];
1542 error
= btree_set_parent(cursor
->trans
, new_node
, elm
);
1544 panic("btree_split_internal: btree-fixup problem");
1547 hammer_modify_node_done(new_node
);
1550 * The filesystem's root B-Tree pointer may have to be updated.
1553 hammer_volume_t volume
;
1555 volume
= hammer_get_root_volume(hmp
, &error
);
1556 KKASSERT(error
== 0);
1558 hammer_modify_volume_field(cursor
->trans
, volume
,
1560 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1561 hammer_modify_volume_done(volume
);
1562 node
->ondisk
->parent
= parent
->node_offset
;
1563 if (cursor
->parent
) {
1564 hammer_unlock(&cursor
->parent
->lock
);
1565 hammer_rel_node(cursor
->parent
);
1567 cursor
->parent
= parent
; /* lock'd and ref'd */
1568 hammer_rel_volume(volume
, 0);
1570 hammer_modify_node_done(node
);
1573 * Ok, now adjust the cursor depending on which element the original
1574 * index was pointing at. If we are >= the split point the push node
1575 * is now in the new node.
1577 * NOTE: If we are at the split point itself we cannot stay with the
1578 * original node because the push index will point at the right-hand
1579 * boundary, which is illegal.
1581 * NOTE: The cursor's parent or parent_index must be adjusted for
1582 * the case where a new parent (new root) was created, and the case
1583 * where the cursor is now pointing at the split node.
1585 if (cursor
->index
>= split
) {
1586 cursor
->parent_index
= parent_index
+ 1;
1587 cursor
->index
-= split
;
1588 hammer_unlock(&cursor
->node
->lock
);
1589 hammer_rel_node(cursor
->node
);
1590 cursor
->node
= new_node
; /* locked and ref'd */
1592 cursor
->parent_index
= parent_index
;
1593 hammer_unlock(&new_node
->lock
);
1594 hammer_rel_node(new_node
);
1598 * Fixup left and right bounds
1600 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1601 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1602 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1603 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1604 &cursor
->node
->ondisk
->elms
[0].internal
.base
) <= 0);
1605 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1606 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
].internal
.base
) >= 0);
1609 hammer_btree_unlock_children(&locklist
);
1610 hammer_cursor_downgrade(cursor
);
1615 * Same as the above, but splits a full leaf node.
1621 btree_split_leaf(hammer_cursor_t cursor
)
1623 hammer_node_ondisk_t ondisk
;
1624 hammer_node_t parent
;
1627 hammer_node_t new_leaf
;
1628 hammer_btree_elm_t elm
;
1629 hammer_btree_elm_t parent_elm
;
1630 hammer_base_elm_t mid_boundary
;
1635 const size_t esize
= sizeof(*elm
);
1637 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1639 ++hammer_stats_btree_splits
;
1641 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1642 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1643 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1644 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1647 * Calculate the split point. If the insertion point will be on
1648 * the left-hand side adjust the split point to give the right
1649 * hand side one additional node.
1651 * Spikes are made up of two leaf elements which cannot be
1654 leaf
= cursor
->node
;
1655 ondisk
= leaf
->ondisk
;
1656 split
= (ondisk
->count
+ 1) / 2;
1657 if (cursor
->index
<= split
)
1662 elm
= &ondisk
->elms
[split
];
1664 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
[-1].leaf
.base
) <= 0);
1665 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->leaf
.base
) <= 0);
1666 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->leaf
.base
) > 0);
1667 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
[1].leaf
.base
) > 0);
1670 * If we are at the root of the tree, create a new root node with
1671 * 1 element and split normally. Avoid making major modifications
1672 * until we know the whole operation will work.
1674 if (ondisk
->parent
== 0) {
1675 parent
= hammer_alloc_btree(cursor
->trans
, &error
);
1678 hammer_lock_ex(&parent
->lock
);
1679 hammer_modify_node_noundo(cursor
->trans
, parent
);
1680 ondisk
= parent
->ondisk
;
1683 ondisk
->mirror_tid
= leaf
->ondisk
->mirror_tid
;
1684 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1685 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1686 ondisk
->elms
[0].base
.btype
= leaf
->ondisk
->type
;
1687 ondisk
->elms
[0].internal
.subtree_offset
= leaf
->node_offset
;
1688 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1689 /* ondisk->elms[1].base.btype = not used */
1690 hammer_modify_node_done(parent
);
1692 parent_index
= 0; /* insertion point in parent */
1695 parent
= cursor
->parent
;
1696 parent_index
= cursor
->parent_index
;
1700 * Split leaf into new_leaf at the split point. Select a separator
1701 * value in-between the two leafs but with a bent towards the right
1702 * leaf since comparisons use an 'elm >= separator' inequality.
1711 new_leaf
= hammer_alloc_btree(cursor
->trans
, &error
);
1712 if (new_leaf
== NULL
) {
1714 hammer_unlock(&parent
->lock
);
1715 hammer_delete_node(cursor
->trans
, parent
);
1716 hammer_rel_node(parent
);
1720 hammer_lock_ex(&new_leaf
->lock
);
1723 * Create the new node and copy the leaf elements from the split
1724 * point on to the new node.
1726 hammer_modify_node_all(cursor
->trans
, leaf
);
1727 hammer_modify_node_noundo(cursor
->trans
, new_leaf
);
1728 ondisk
= leaf
->ondisk
;
1729 elm
= &ondisk
->elms
[split
];
1730 bcopy(elm
, &new_leaf
->ondisk
->elms
[0], (ondisk
->count
- split
) * esize
);
1731 new_leaf
->ondisk
->count
= ondisk
->count
- split
;
1732 new_leaf
->ondisk
->parent
= parent
->node_offset
;
1733 new_leaf
->ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
1734 new_leaf
->ondisk
->mirror_tid
= ondisk
->mirror_tid
;
1735 KKASSERT(ondisk
->type
== new_leaf
->ondisk
->type
);
1736 hammer_modify_node_done(new_leaf
);
1737 hammer_cursor_split_node(leaf
, new_leaf
, split
);
1740 * Cleanup the original node. Because this is a leaf node and
1741 * leaf nodes do not have a right-hand boundary, there
1742 * aren't any special edge cases to clean up. We just fixup the
1745 ondisk
->count
= split
;
1748 * Insert the separator into the parent, fixup the parent's
1749 * reference to the original node, and reference the new node.
1750 * The separator is P.
1752 * Remember that base.count does not include the right-hand boundary.
1753 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1755 hammer_modify_node_all(cursor
->trans
, parent
);
1756 ondisk
= parent
->ondisk
;
1757 KKASSERT(split
!= 0);
1758 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1759 parent_elm
= &ondisk
->elms
[parent_index
+1];
1760 bcopy(parent_elm
, parent_elm
+ 1,
1761 (ondisk
->count
- parent_index
) * esize
);
1763 hammer_make_separator(&elm
[-1].base
, &elm
[0].base
, &parent_elm
->base
);
1764 parent_elm
->internal
.base
.btype
= new_leaf
->ondisk
->type
;
1765 parent_elm
->internal
.subtree_offset
= new_leaf
->node_offset
;
1766 parent_elm
->internal
.mirror_tid
= new_leaf
->ondisk
->mirror_tid
;
1767 mid_boundary
= &parent_elm
->base
;
1769 hammer_modify_node_done(parent
);
1770 hammer_cursor_inserted_element(parent
, parent_index
+ 1);
1773 * The filesystem's root B-Tree pointer may have to be updated.
1776 hammer_volume_t volume
;
1778 volume
= hammer_get_root_volume(hmp
, &error
);
1779 KKASSERT(error
== 0);
1781 hammer_modify_volume_field(cursor
->trans
, volume
,
1783 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1784 hammer_modify_volume_done(volume
);
1785 leaf
->ondisk
->parent
= parent
->node_offset
;
1786 if (cursor
->parent
) {
1787 hammer_unlock(&cursor
->parent
->lock
);
1788 hammer_rel_node(cursor
->parent
);
1790 cursor
->parent
= parent
; /* lock'd and ref'd */
1791 hammer_rel_volume(volume
, 0);
1793 hammer_modify_node_done(leaf
);
1796 * Ok, now adjust the cursor depending on which element the original
1797 * index was pointing at. If we are >= the split point the push node
1798 * is now in the new node.
1800 * NOTE: If we are at the split point itself we need to select the
1801 * old or new node based on where key_beg's insertion point will be.
1802 * If we pick the wrong side the inserted element will wind up in
1803 * the wrong leaf node and outside that node's bounds.
1805 if (cursor
->index
> split
||
1806 (cursor
->index
== split
&&
1807 hammer_btree_cmp(&cursor
->key_beg
, mid_boundary
) >= 0)) {
1808 cursor
->parent_index
= parent_index
+ 1;
1809 cursor
->index
-= split
;
1810 hammer_unlock(&cursor
->node
->lock
);
1811 hammer_rel_node(cursor
->node
);
1812 cursor
->node
= new_leaf
;
1814 cursor
->parent_index
= parent_index
;
1815 hammer_unlock(&new_leaf
->lock
);
1816 hammer_rel_node(new_leaf
);
1820 * Fixup left and right bounds
1822 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1823 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1824 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1827 * Assert that the bounds are correct.
1829 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1830 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1831 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1832 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1833 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &cursor
->key_beg
) <= 0);
1834 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &cursor
->key_beg
) > 0);
1837 hammer_cursor_downgrade(cursor
);
1844 * Recursively correct the right-hand boundary's create_tid to (tid) as
1845 * long as the rest of the key matches. We have to recurse upward in
1846 * the tree as well as down the left side of each parent's right node.
1848 * Return EDEADLK if we were only partially successful, forcing the caller
1849 * to try again. The original cursor is not modified. This routine can
1850 * also fail with EDEADLK if it is forced to throw away a portion of its
1853 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1856 TAILQ_ENTRY(hammer_rhb
) entry
;
1861 TAILQ_HEAD(hammer_rhb_list
, hammer_rhb
);
1864 hammer_btree_correct_rhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
1866 struct hammer_rhb_list rhb_list
;
1867 hammer_base_elm_t elm
;
1868 hammer_node_t orig_node
;
1869 struct hammer_rhb
*rhb
;
1873 TAILQ_INIT(&rhb_list
);
1876 * Save our position so we can restore it on return. This also
1877 * gives us a stable 'elm'.
1879 orig_node
= cursor
->node
;
1880 hammer_ref_node(orig_node
);
1881 hammer_lock_sh(&orig_node
->lock
);
1882 orig_index
= cursor
->index
;
1883 elm
= &orig_node
->ondisk
->elms
[orig_index
].base
;
1886 * Now build a list of parents going up, allocating a rhb
1887 * structure for each one.
1889 while (cursor
->parent
) {
1891 * Stop if we no longer have any right-bounds to fix up
1893 if (elm
->obj_id
!= cursor
->right_bound
->obj_id
||
1894 elm
->rec_type
!= cursor
->right_bound
->rec_type
||
1895 elm
->key
!= cursor
->right_bound
->key
) {
1900 * Stop if the right-hand bound's create_tid does not
1901 * need to be corrected.
1903 if (cursor
->right_bound
->create_tid
>= tid
)
1906 rhb
= kmalloc(sizeof(*rhb
), M_HAMMER
, M_WAITOK
|M_ZERO
);
1907 rhb
->node
= cursor
->parent
;
1908 rhb
->index
= cursor
->parent_index
;
1909 hammer_ref_node(rhb
->node
);
1910 hammer_lock_sh(&rhb
->node
->lock
);
1911 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
1913 hammer_cursor_up(cursor
);
1917 * now safely adjust the right hand bound for each rhb. This may
1918 * also require taking the right side of the tree and iterating down
1922 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1923 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
1926 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1927 hammer_unlock(&rhb
->node
->lock
);
1928 hammer_rel_node(rhb
->node
);
1929 kfree(rhb
, M_HAMMER
);
1931 switch (cursor
->node
->ondisk
->type
) {
1932 case HAMMER_BTREE_TYPE_INTERNAL
:
1934 * Right-boundary for parent at internal node
1935 * is one element to the right of the element whos
1936 * right boundary needs adjusting. We must then
1937 * traverse down the left side correcting any left
1938 * bounds (which may now be too far to the left).
1941 error
= hammer_btree_correct_lhb(cursor
, tid
);
1944 panic("hammer_btree_correct_rhb(): Bad node type");
1953 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
1954 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
1955 hammer_unlock(&rhb
->node
->lock
);
1956 hammer_rel_node(rhb
->node
);
1957 kfree(rhb
, M_HAMMER
);
1959 error
= hammer_cursor_seek(cursor
, orig_node
, orig_index
);
1960 hammer_unlock(&orig_node
->lock
);
1961 hammer_rel_node(orig_node
);
1966 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1967 * bound going downward starting at the current cursor position.
1969 * This function does not restore the cursor after use.
1972 hammer_btree_correct_lhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
1974 struct hammer_rhb_list rhb_list
;
1975 hammer_base_elm_t elm
;
1976 hammer_base_elm_t cmp
;
1977 struct hammer_rhb
*rhb
;
1980 TAILQ_INIT(&rhb_list
);
1982 cmp
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
1985 * Record the node and traverse down the left-hand side for all
1986 * matching records needing a boundary correction.
1990 rhb
= kmalloc(sizeof(*rhb
), M_HAMMER
, M_WAITOK
|M_ZERO
);
1991 rhb
->node
= cursor
->node
;
1992 rhb
->index
= cursor
->index
;
1993 hammer_ref_node(rhb
->node
);
1994 hammer_lock_sh(&rhb
->node
->lock
);
1995 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
1997 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1999 * Nothing to traverse down if we are at the right
2000 * boundary of an internal node.
2002 if (cursor
->index
== cursor
->node
->ondisk
->count
)
2005 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
2006 if (elm
->btype
== HAMMER_BTREE_TYPE_RECORD
)
2008 panic("Illegal leaf record type %02x", elm
->btype
);
2010 error
= hammer_cursor_down(cursor
);
2014 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
2015 if (elm
->obj_id
!= cmp
->obj_id
||
2016 elm
->rec_type
!= cmp
->rec_type
||
2017 elm
->key
!= cmp
->key
) {
2020 if (elm
->create_tid
>= tid
)
2026 * Now we can safely adjust the left-hand boundary from the bottom-up.
2027 * The last element we remove from the list is the caller's right hand
2028 * boundary, which must also be adjusted.
2030 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
2031 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
2034 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
2035 hammer_unlock(&rhb
->node
->lock
);
2036 hammer_rel_node(rhb
->node
);
2037 kfree(rhb
, M_HAMMER
);
2039 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
2040 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
2041 hammer_modify_node(cursor
->trans
, cursor
->node
,
2043 sizeof(elm
->create_tid
));
2044 elm
->create_tid
= tid
;
2045 hammer_modify_node_done(cursor
->node
);
2047 panic("hammer_btree_correct_lhb(): Bad element type");
2054 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
2055 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
2056 hammer_unlock(&rhb
->node
->lock
);
2057 hammer_rel_node(rhb
->node
);
2058 kfree(rhb
, M_HAMMER
);
2066 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2067 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2068 * the operation due to a deadlock, or some other error.
2070 * This routine is always called with an empty, locked leaf but may recurse
2071 * into want-to-be-empty parents as part of its operation.
2073 * It should also be noted that when removing empty leaves we must be sure
2074 * to test and update mirror_tid because another thread may have deadlocked
2075 * against us (or someone) trying to propagate it up and cannot retry once
2076 * the node has been deleted.
2078 * On return the cursor may end up pointing to an internal node, suitable
2079 * for further iteration but not for an immediate insertion or deletion.
2082 btree_remove(hammer_cursor_t cursor
)
2084 hammer_node_ondisk_t ondisk
;
2085 hammer_btree_elm_t elm
;
2087 hammer_node_t parent
;
2088 const int esize
= sizeof(*elm
);
2091 node
= cursor
->node
;
2094 * When deleting the root of the filesystem convert it to
2095 * an empty leaf node. Internal nodes cannot be empty.
2097 ondisk
= node
->ondisk
;
2098 if (ondisk
->parent
== 0) {
2099 KKASSERT(cursor
->parent
== NULL
);
2100 hammer_modify_node_all(cursor
->trans
, node
);
2101 KKASSERT(ondisk
== node
->ondisk
);
2102 ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
2104 hammer_modify_node_done(node
);
2109 parent
= cursor
->parent
;
2110 hammer_cursor_removed_node(node
, parent
, cursor
->parent_index
);
2113 * Attempt to remove the parent's reference to the child. If the
2114 * parent would become empty we have to recurse. If we fail we
2115 * leave the parent pointing to an empty leaf node.
2117 if (parent
->ondisk
->count
== 1) {
2119 * This special cursor_up_locked() call leaves the original
2120 * node exclusively locked and referenced, leaves the
2121 * original parent locked (as the new node), and locks the
2122 * new parent. It can return EDEADLK.
2124 error
= hammer_cursor_up_locked(cursor
);
2126 error
= btree_remove(cursor
);
2128 hammer_modify_node_all(cursor
->trans
, node
);
2129 ondisk
= node
->ondisk
;
2130 ondisk
->type
= HAMMER_BTREE_TYPE_DELETED
;
2132 hammer_modify_node_done(node
);
2133 hammer_flush_node(node
);
2134 hammer_delete_node(cursor
->trans
, node
);
2136 kprintf("Warning: BTREE_REMOVE: Defering "
2137 "parent removal1 @ %016llx, skipping\n",
2140 hammer_unlock(&node
->lock
);
2141 hammer_rel_node(node
);
2143 kprintf("Warning: BTREE_REMOVE: Defering parent "
2144 "removal2 @ %016llx, skipping\n",
2148 KKASSERT(parent
->ondisk
->count
> 1);
2150 hammer_modify_node_all(cursor
->trans
, parent
);
2151 ondisk
= parent
->ondisk
;
2152 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
);
2154 elm
= &ondisk
->elms
[cursor
->parent_index
];
2155 KKASSERT(elm
->internal
.subtree_offset
== node
->node_offset
);
2156 KKASSERT(ondisk
->count
> 0);
2159 * We must retain the highest mirror_tid. The deleted
2160 * range is now encompassed by the element to the left.
2161 * If we are already at the left edge the new left edge
2162 * inherits mirror_tid.
2164 * Note that bounds of the parent to our parent may create
2165 * a gap to the left of our left-most node or to the right
2166 * of our right-most node. The gap is silently included
2167 * in the mirror_tid's area of effect from the point of view
2170 if (cursor
->parent_index
) {
2171 if (elm
[-1].internal
.mirror_tid
<
2172 elm
[0].internal
.mirror_tid
) {
2173 elm
[-1].internal
.mirror_tid
=
2174 elm
[0].internal
.mirror_tid
;
2177 if (elm
[1].internal
.mirror_tid
<
2178 elm
[0].internal
.mirror_tid
) {
2179 elm
[1].internal
.mirror_tid
=
2180 elm
[0].internal
.mirror_tid
;
2185 * Delete the subtree reference in the parent
2187 bcopy(&elm
[1], &elm
[0],
2188 (ondisk
->count
- cursor
->parent_index
) * esize
);
2190 hammer_modify_node_done(parent
);
2191 hammer_cursor_deleted_element(parent
, cursor
->parent_index
);
2192 hammer_flush_node(node
);
2193 hammer_delete_node(cursor
->trans
, node
);
2196 * cursor->node is invalid, cursor up to make the cursor
2199 error
= hammer_cursor_up(cursor
);
2205 * Propagate cursor->trans->tid up the B-Tree starting at the current
2206 * cursor position using pseudofs info gleaned from the passed inode.
2208 * The passed inode has no relationship to the cursor position other
2209 * then being in the same pseudofs as the insertion or deletion we
2210 * are propagating the mirror_tid for.
2213 hammer_btree_do_propagation(hammer_cursor_t cursor
,
2214 hammer_pseudofs_inmem_t pfsm
,
2215 hammer_btree_leaf_elm_t leaf
)
2217 hammer_cursor_t ncursor
;
2218 hammer_tid_t mirror_tid
;
2222 * We only propagate the mirror_tid up if we are in master or slave
2223 * mode. We do not bother if we are in no-mirror mode.
2225 * If pfsm is NULL we propagate (from mirror_write).
2228 pfsm
->pfsd
.master_id
< 0 &&
2229 (pfsm
->pfsd
.mirror_flags
& HAMMER_PFSD_SLAVE
) == 0) {
2234 * This is a bit of a hack because we cannot deadlock or return
2235 * EDEADLK here. The related operation has already completed and
2236 * we must propagate the mirror_tid now regardless.
2238 * Generate a new cursor which inherits the original's locks and
2239 * unlock the original. Use the new cursor to propagate the
2240 * mirror_tid. Then clean up the new cursor and reacquire locks
2243 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2244 * original's locks and the original is tracked and must be
2247 mirror_tid
= cursor
->node
->ondisk
->mirror_tid
;
2248 KKASSERT(mirror_tid
!= 0);
2249 ncursor
= hammer_push_cursor(cursor
);
2250 error
= hammer_btree_mirror_propagate(ncursor
, mirror_tid
);
2251 KKASSERT(error
== 0);
2252 hammer_pop_cursor(cursor
, ncursor
);
2257 * Propagate a mirror TID update upwards through the B-Tree to the root.
2259 * A locked internal node must be passed in. The node will remain locked
2262 * This function syncs mirror_tid at the specified internal node's element,
2263 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2266 hammer_btree_mirror_propagate(hammer_cursor_t cursor
, hammer_tid_t mirror_tid
)
2268 hammer_btree_internal_elm_t elm
;
2273 error
= hammer_cursor_up(cursor
);
2275 error
= hammer_cursor_upgrade(cursor
);
2276 while (error
== EDEADLK
) {
2277 hammer_recover_cursor(cursor
);
2278 error
= hammer_cursor_upgrade(cursor
);
2282 node
= cursor
->node
;
2283 KKASSERT (node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
);
2286 * Adjust the node's element
2288 elm
= &node
->ondisk
->elms
[cursor
->index
].internal
;
2289 if (elm
->mirror_tid
>= mirror_tid
)
2291 hammer_modify_node(cursor
->trans
, node
, &elm
->mirror_tid
,
2292 sizeof(elm
->mirror_tid
));
2293 elm
->mirror_tid
= mirror_tid
;
2294 hammer_modify_node_done(node
);
2295 if (hammer_debug_general
& 0x0002) {
2296 kprintf("mirror_propagate: propagate "
2297 "%016llx @%016llx:%d\n",
2298 mirror_tid
, node
->node_offset
, cursor
->index
);
2303 * Adjust the node's mirror_tid aggregator
2305 if (node
->ondisk
->mirror_tid
>= mirror_tid
)
2307 hammer_modify_node_field(cursor
->trans
, node
, mirror_tid
);
2308 node
->ondisk
->mirror_tid
= mirror_tid
;
2309 hammer_modify_node_done(node
);
2310 if (hammer_debug_general
& 0x0002) {
2311 kprintf("mirror_propagate: propagate "
2312 "%016llx @%016llx\n",
2313 mirror_tid
, node
->node_offset
);
2316 if (error
== ENOENT
)
2322 hammer_btree_get_parent(hammer_node_t node
, int *parent_indexp
, int *errorp
,
2325 hammer_node_t parent
;
2326 hammer_btree_elm_t elm
;
2332 parent
= hammer_get_node(node
->hmp
, node
->ondisk
->parent
, 0, errorp
);
2334 KKASSERT(parent
== NULL
);
2337 KKASSERT ((parent
->flags
& HAMMER_NODE_DELETED
) == 0);
2342 if (try_exclusive
) {
2343 if (hammer_lock_ex_try(&parent
->lock
)) {
2344 hammer_rel_node(parent
);
2349 hammer_lock_sh(&parent
->lock
);
2353 * Figure out which element in the parent is pointing to the
2356 if (node
->ondisk
->count
) {
2357 i
= hammer_btree_search_node(&node
->ondisk
->elms
[0].base
,
2362 while (i
< parent
->ondisk
->count
) {
2363 elm
= &parent
->ondisk
->elms
[i
];
2364 if (elm
->internal
.subtree_offset
== node
->node_offset
)
2368 if (i
== parent
->ondisk
->count
) {
2369 hammer_unlock(&parent
->lock
);
2370 panic("Bad B-Tree link: parent %p node %p\n", parent
, node
);
2373 KKASSERT(*errorp
== 0);
2378 * The element (elm) has been moved to a new internal node (node).
2380 * If the element represents a pointer to an internal node that node's
2381 * parent must be adjusted to the element's new location.
2383 * XXX deadlock potential here with our exclusive locks
2386 btree_set_parent(hammer_transaction_t trans
, hammer_node_t node
,
2387 hammer_btree_elm_t elm
)
2389 hammer_node_t child
;
2394 switch(elm
->base
.btype
) {
2395 case HAMMER_BTREE_TYPE_INTERNAL
:
2396 case HAMMER_BTREE_TYPE_LEAF
:
2397 child
= hammer_get_node(node
->hmp
, elm
->internal
.subtree_offset
,
2400 hammer_modify_node_field(trans
, child
, parent
);
2401 child
->ondisk
->parent
= node
->node_offset
;
2402 hammer_modify_node_done(child
);
2403 hammer_rel_node(child
);
2413 * Exclusively lock all the children of node. This is used by the split
2414 * code to prevent anyone from accessing the children of a cursor node
2415 * while we fix-up its parent offset.
2417 * If we don't lock the children we can really mess up cursors which block
2418 * trying to cursor-up into our node.
2420 * On failure EDEADLK (or some other error) is returned. If a deadlock
2421 * error is returned the cursor is adjusted to block on termination.
2424 hammer_btree_lock_children(hammer_cursor_t cursor
,
2425 struct hammer_node_locklist
**locklistp
)
2428 hammer_node_locklist_t item
;
2429 hammer_node_ondisk_t ondisk
;
2430 hammer_btree_elm_t elm
;
2431 hammer_node_t child
;
2435 node
= cursor
->node
;
2436 ondisk
= node
->ondisk
;
2440 * We really do not want to block on I/O with exclusive locks held,
2441 * pre-get the children before trying to lock the mess.
2443 for (i
= 0; i
< ondisk
->count
; ++i
) {
2444 ++hammer_stats_btree_elements
;
2445 elm
= &ondisk
->elms
[i
];
2446 if (elm
->base
.btype
!= HAMMER_BTREE_TYPE_LEAF
&&
2447 elm
->base
.btype
!= HAMMER_BTREE_TYPE_INTERNAL
) {
2450 child
= hammer_get_node(node
->hmp
,
2451 elm
->internal
.subtree_offset
,
2454 hammer_rel_node(child
);
2460 for (i
= 0; error
== 0 && i
< ondisk
->count
; ++i
) {
2461 ++hammer_stats_btree_elements
;
2462 elm
= &ondisk
->elms
[i
];
2464 switch(elm
->base
.btype
) {
2465 case HAMMER_BTREE_TYPE_INTERNAL
:
2466 case HAMMER_BTREE_TYPE_LEAF
:
2467 KKASSERT(elm
->internal
.subtree_offset
!= 0);
2468 child
= hammer_get_node(node
->hmp
,
2469 elm
->internal
.subtree_offset
,
2477 if (hammer_lock_ex_try(&child
->lock
) != 0) {
2478 if (cursor
->deadlk_node
== NULL
) {
2479 cursor
->deadlk_node
= child
;
2480 hammer_ref_node(cursor
->deadlk_node
);
2483 hammer_rel_node(child
);
2485 item
= kmalloc(sizeof(*item
),
2486 M_HAMMER
, M_WAITOK
);
2487 item
->next
= *locklistp
;
2494 hammer_btree_unlock_children(locklistp
);
2500 * Release previously obtained node locks.
2503 hammer_btree_unlock_children(struct hammer_node_locklist
**locklistp
)
2505 hammer_node_locklist_t item
;
2507 while ((item
= *locklistp
) != NULL
) {
2508 *locklistp
= item
->next
;
2509 hammer_unlock(&item
->node
->lock
);
2510 hammer_rel_node(item
->node
);
2511 kfree(item
, M_HAMMER
);
2515 /************************************************************************
2516 * MISCELLANIOUS SUPPORT *
2517 ************************************************************************/
2520 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2522 * Note that for this particular function a return value of -1, 0, or +1
2523 * can denote a match if create_tid is otherwise discounted. A create_tid
2524 * of zero is considered to be 'infinity' in comparisons.
2526 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2529 hammer_btree_cmp(hammer_base_elm_t key1
, hammer_base_elm_t key2
)
2531 if (key1
->localization
< key2
->localization
)
2533 if (key1
->localization
> key2
->localization
)
2536 if (key1
->obj_id
< key2
->obj_id
)
2538 if (key1
->obj_id
> key2
->obj_id
)
2541 if (key1
->rec_type
< key2
->rec_type
)
2543 if (key1
->rec_type
> key2
->rec_type
)
2546 if (key1
->key
< key2
->key
)
2548 if (key1
->key
> key2
->key
)
2552 * A create_tid of zero indicates a record which is undeletable
2553 * and must be considered to have a value of positive infinity.
2555 if (key1
->create_tid
== 0) {
2556 if (key2
->create_tid
== 0)
2560 if (key2
->create_tid
== 0)
2562 if (key1
->create_tid
< key2
->create_tid
)
2564 if (key1
->create_tid
> key2
->create_tid
)
2570 * Test a timestamp against an element to determine whether the
2571 * element is visible. A timestamp of 0 means 'infinity'.
2574 hammer_btree_chkts(hammer_tid_t asof
, hammer_base_elm_t base
)
2577 if (base
->delete_tid
)
2581 if (asof
< base
->create_tid
)
2583 if (base
->delete_tid
&& asof
>= base
->delete_tid
)
2589 * Create a separator half way inbetween key1 and key2. For fields just
2590 * one unit apart, the separator will match key2. key1 is on the left-hand
2591 * side and key2 is on the right-hand side.
2593 * key2 must be >= the separator. It is ok for the separator to match key2.
2595 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2598 * NOTE: It might be beneficial to just scrap this whole mess and just
2599 * set the separator to key2.
2601 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2602 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2605 hammer_make_separator(hammer_base_elm_t key1
, hammer_base_elm_t key2
,
2606 hammer_base_elm_t dest
)
2608 bzero(dest
, sizeof(*dest
));
2610 dest
->rec_type
= key2
->rec_type
;
2611 dest
->key
= key2
->key
;
2612 dest
->obj_id
= key2
->obj_id
;
2613 dest
->create_tid
= key2
->create_tid
;
2615 MAKE_SEPARATOR(key1
, key2
, dest
, localization
);
2616 if (key1
->localization
== key2
->localization
) {
2617 MAKE_SEPARATOR(key1
, key2
, dest
, obj_id
);
2618 if (key1
->obj_id
== key2
->obj_id
) {
2619 MAKE_SEPARATOR(key1
, key2
, dest
, rec_type
);
2620 if (key1
->rec_type
== key2
->rec_type
) {
2621 MAKE_SEPARATOR(key1
, key2
, dest
, key
);
2623 * Don't bother creating a separator for
2624 * create_tid, which also conveniently avoids
2625 * having to handle the create_tid == 0
2626 * (infinity) case. Just leave create_tid
2629 * Worst case, dest matches key2 exactly,
2630 * which is acceptable.
2637 #undef MAKE_SEPARATOR
2640 * Return whether a generic internal or leaf node is full
2643 btree_node_is_full(hammer_node_ondisk_t node
)
2645 switch(node
->type
) {
2646 case HAMMER_BTREE_TYPE_INTERNAL
:
2647 if (node
->count
== HAMMER_BTREE_INT_ELMS
)
2650 case HAMMER_BTREE_TYPE_LEAF
:
2651 if (node
->count
== HAMMER_BTREE_LEAF_ELMS
)
2655 panic("illegal btree subtype");
2662 btree_max_elements(u_int8_t type
)
2664 if (type
== HAMMER_BTREE_TYPE_LEAF
)
2665 return(HAMMER_BTREE_LEAF_ELMS
);
2666 if (type
== HAMMER_BTREE_TYPE_INTERNAL
)
2667 return(HAMMER_BTREE_INT_ELMS
);
2668 panic("btree_max_elements: bad type %d\n", type
);
2673 hammer_print_btree_node(hammer_node_ondisk_t ondisk
)
2675 hammer_btree_elm_t elm
;
2678 kprintf("node %p count=%d parent=%016llx type=%c\n",
2679 ondisk
, ondisk
->count
, ondisk
->parent
, ondisk
->type
);
2682 * Dump both boundary elements if an internal node
2684 if (ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
2685 for (i
= 0; i
<= ondisk
->count
; ++i
) {
2686 elm
= &ondisk
->elms
[i
];
2687 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2690 for (i
= 0; i
< ondisk
->count
; ++i
) {
2691 elm
= &ondisk
->elms
[i
];
2692 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2698 hammer_print_btree_elm(hammer_btree_elm_t elm
, u_int8_t type
, int i
)
2701 kprintf("\tobj_id = %016llx\n", elm
->base
.obj_id
);
2702 kprintf("\tkey = %016llx\n", elm
->base
.key
);
2703 kprintf("\tcreate_tid = %016llx\n", elm
->base
.create_tid
);
2704 kprintf("\tdelete_tid = %016llx\n", elm
->base
.delete_tid
);
2705 kprintf("\trec_type = %04x\n", elm
->base
.rec_type
);
2706 kprintf("\tobj_type = %02x\n", elm
->base
.obj_type
);
2707 kprintf("\tbtype = %02x (%c)\n",
2709 (elm
->base
.btype
? elm
->base
.btype
: '?'));
2710 kprintf("\tlocalization = %02x\n", elm
->base
.localization
);
2713 case HAMMER_BTREE_TYPE_INTERNAL
:
2714 kprintf("\tsubtree_off = %016llx\n",
2715 elm
->internal
.subtree_offset
);
2717 case HAMMER_BTREE_TYPE_RECORD
:
2718 kprintf("\tdata_offset = %016llx\n", elm
->leaf
.data_offset
);
2719 kprintf("\tdata_len = %08x\n", elm
->leaf
.data_len
);
2720 kprintf("\tdata_crc = %08x\n", elm
->leaf
.data_crc
);