2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.76 2008/08/06 15:38:58 dillon Exp $
40 * HAMMER implements a modified B+Tree. In documentation this will
41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
42 * looks like a B+Tree (A B-Tree which stores its records only at the leafs
43 * of the tree), but adds two additional boundary elements which describe
44 * the left-most and right-most element a node is able to represent. In
45 * otherwords, we have boundary elements at the two ends of a B-Tree node
46 * instead of sub-tree pointers.
48 * A B-Tree internal node looks like this:
50 * B N N N N N N B <-- boundary and internal elements
51 * S S S S S S S <-- subtree pointers
53 * A B-Tree leaf node basically looks like this:
55 * L L L L L L L L <-- leaf elemenets
57 * The radix for an internal node is 1 less then a leaf but we get a
58 * number of significant benefits for our troubles.
60 * The big benefit to using a B-Tree containing boundary information
61 * is that it is possible to cache pointers into the middle of the tree
62 * and not have to start searches, insertions, OR deletions at the root
63 * node. In particular, searches are able to progress in a definitive
64 * direction from any point in the tree without revisting nodes. This
65 * greatly improves the efficiency of many operations, most especially
68 * B-Trees also make the stacking of trees fairly straightforward.
70 * INSERTIONS: A search performed with the intention of doing
71 * an insert will guarantee that the terminal leaf node is not full by
72 * splitting full nodes. Splits occur top-down during the dive down the
75 * DELETIONS: A deletion makes no attempt to proactively balance the
76 * tree and will recursively remove nodes that become empty. If a
77 * deadlock occurs a deletion may not be able to remove an empty leaf.
78 * Deletions never allow internal nodes to become empty (that would blow
85 static int btree_search(hammer_cursor_t cursor
, int flags
);
86 static int btree_split_internal(hammer_cursor_t cursor
);
87 static int btree_split_leaf(hammer_cursor_t cursor
);
88 static int btree_remove(hammer_cursor_t cursor
);
89 static int btree_node_is_full(hammer_node_ondisk_t node
);
90 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor
,
91 hammer_tid_t mirror_tid
);
92 static void hammer_make_separator(hammer_base_elm_t key1
,
93 hammer_base_elm_t key2
, hammer_base_elm_t dest
);
94 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor
);
97 * Iterate records after a search. The cursor is iterated forwards past
98 * the current record until a record matching the key-range requirements
99 * is found. ENOENT is returned if the iteration goes past the ending
102 * The iteration is inclusive of key_beg and can be inclusive or exclusive
103 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
105 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
106 * may be modified by B-Tree functions.
108 * cursor->key_beg may or may not be modified by this function during
109 * the iteration. XXX future - in case of an inverted lock we may have
110 * to reinitiate the lookup and set key_beg to properly pick up where we
113 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
116 hammer_btree_iterate(hammer_cursor_t cursor
)
118 hammer_node_ondisk_t node
;
119 hammer_btree_elm_t elm
;
126 * Skip past the current record
128 hmp
= cursor
->trans
->hmp
;
129 node
= cursor
->node
->ondisk
;
132 if (cursor
->index
< node
->count
&&
133 (cursor
->flags
& HAMMER_CURSOR_ATEDISK
)) {
138 * HAMMER can wind up being cpu-bound.
140 if (++hmp
->check_yield
> hammer_yield_check
) {
141 hmp
->check_yield
= 0;
147 * Loop until an element is found or we are done.
151 * We iterate up the tree and then index over one element
152 * while we are at the last element in the current node.
154 * If we are at the root of the filesystem, cursor_up
157 * XXX this could be optimized by storing the information in
158 * the parent reference.
160 * XXX we can lose the node lock temporarily, this could mess
163 ++hammer_stats_btree_iterations
;
164 hammer_flusher_clean_loose_ios(hmp
);
166 if (cursor
->index
== node
->count
) {
167 if (hammer_debug_btree
) {
168 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
169 (long long)cursor
->node
->node_offset
,
171 (long long)(cursor
->parent
? cursor
->parent
->node_offset
: -1),
172 cursor
->parent_index
,
175 KKASSERT(cursor
->parent
== NULL
|| cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.subtree_offset
== cursor
->node
->node_offset
);
176 error
= hammer_cursor_up(cursor
);
179 /* reload stale pointer */
180 node
= cursor
->node
->ondisk
;
181 KKASSERT(cursor
->index
!= node
->count
);
184 * If we are reblocking we want to return internal
185 * nodes. Note that the internal node will be
186 * returned multiple times, on each upward recursion
187 * from its children. The caller selects which
188 * revisit it cares about (usually first or last only).
190 if (cursor
->flags
& HAMMER_CURSOR_REBLOCKING
) {
191 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
199 * Check internal or leaf element. Determine if the record
200 * at the cursor has gone beyond the end of our range.
202 * We recurse down through internal nodes.
204 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
205 elm
= &node
->elms
[cursor
->index
];
207 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
208 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
209 if (hammer_debug_btree
) {
210 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
211 (long long)cursor
->node
->node_offset
,
213 (long long)elm
[0].internal
.base
.obj_id
,
214 elm
[0].internal
.base
.rec_type
,
215 (long long)elm
[0].internal
.base
.key
,
216 elm
[0].internal
.base
.localization
,
220 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
221 (long long)cursor
->node
->node_offset
,
223 (long long)elm
[1].internal
.base
.obj_id
,
224 elm
[1].internal
.base
.rec_type
,
225 (long long)elm
[1].internal
.base
.key
,
226 elm
[1].internal
.base
.localization
,
235 if (r
== 0 && (cursor
->flags
&
236 HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
245 KKASSERT(elm
->internal
.subtree_offset
!= 0);
248 * If running the mirror filter see if we can skip
249 * one or more entire sub-trees. If we can we
250 * return the internal mode and the caller processes
251 * the skipped range (see mirror_read)
253 if (cursor
->flags
& HAMMER_CURSOR_MIRROR_FILTERED
) {
254 if (elm
->internal
.mirror_tid
<
255 cursor
->cmirror
->mirror_tid
) {
256 hammer_cursor_mirror_filter(cursor
);
261 error
= hammer_cursor_down(cursor
);
264 KKASSERT(cursor
->index
== 0);
265 /* reload stale pointer */
266 node
= cursor
->node
->ondisk
;
269 elm
= &node
->elms
[cursor
->index
];
270 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
->base
);
271 if (hammer_debug_btree
) {
272 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
273 (long long)cursor
->node
->node_offset
,
275 (elm
[0].leaf
.base
.btype
?
276 elm
[0].leaf
.base
.btype
: '?'),
277 (long long)elm
[0].leaf
.base
.obj_id
,
278 elm
[0].leaf
.base
.rec_type
,
279 (long long)elm
[0].leaf
.base
.key
,
280 elm
[0].leaf
.base
.localization
,
290 * We support both end-inclusive and
291 * end-exclusive searches.
294 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
299 switch(elm
->leaf
.base
.btype
) {
300 case HAMMER_BTREE_TYPE_RECORD
:
301 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
302 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
316 * node pointer invalid after loop
322 if (hammer_debug_btree
) {
323 int i
= cursor
->index
;
324 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
325 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
327 (long long)elm
->internal
.base
.obj_id
,
328 elm
->internal
.base
.rec_type
,
329 (long long)elm
->internal
.base
.key
,
330 elm
->internal
.base
.localization
339 * We hit an internal element that we could skip as part of a mirroring
340 * scan. Calculate the entire range being skipped.
342 * It is important to include any gaps between the parent's left_bound
343 * and the node's left_bound, and same goes for the right side.
346 hammer_cursor_mirror_filter(hammer_cursor_t cursor
)
348 struct hammer_cmirror
*cmirror
;
349 hammer_node_ondisk_t ondisk
;
350 hammer_btree_elm_t elm
;
352 ondisk
= cursor
->node
->ondisk
;
353 cmirror
= cursor
->cmirror
;
356 * Calculate the skipped range
358 elm
= &ondisk
->elms
[cursor
->index
];
359 if (cursor
->index
== 0)
360 cmirror
->skip_beg
= *cursor
->left_bound
;
362 cmirror
->skip_beg
= elm
->internal
.base
;
363 while (cursor
->index
< ondisk
->count
) {
364 if (elm
->internal
.mirror_tid
>= cmirror
->mirror_tid
)
369 if (cursor
->index
== ondisk
->count
)
370 cmirror
->skip_end
= *cursor
->right_bound
;
372 cmirror
->skip_end
= elm
->internal
.base
;
375 * clip the returned result.
377 if (hammer_btree_cmp(&cmirror
->skip_beg
, &cursor
->key_beg
) < 0)
378 cmirror
->skip_beg
= cursor
->key_beg
;
379 if (hammer_btree_cmp(&cmirror
->skip_end
, &cursor
->key_end
) > 0)
380 cmirror
->skip_end
= cursor
->key_end
;
384 * Iterate in the reverse direction. This is used by the pruning code to
385 * avoid overlapping records.
388 hammer_btree_iterate_reverse(hammer_cursor_t cursor
)
390 hammer_node_ondisk_t node
;
391 hammer_btree_elm_t elm
;
396 /* mirror filtering not supported for reverse iteration */
397 KKASSERT ((cursor
->flags
& HAMMER_CURSOR_MIRROR_FILTERED
) == 0);
400 * Skip past the current record. For various reasons the cursor
401 * may end up set to -1 or set to point at the end of the current
402 * node. These cases must be addressed.
404 node
= cursor
->node
->ondisk
;
407 if (cursor
->index
!= -1 &&
408 (cursor
->flags
& HAMMER_CURSOR_ATEDISK
)) {
411 if (cursor
->index
== cursor
->node
->ondisk
->count
)
415 * Loop until an element is found or we are done.
418 ++hammer_stats_btree_iterations
;
419 hammer_flusher_clean_loose_ios(cursor
->trans
->hmp
);
422 * We iterate up the tree and then index over one element
423 * while we are at the last element in the current node.
425 if (cursor
->index
== -1) {
426 error
= hammer_cursor_up(cursor
);
428 cursor
->index
= 0; /* sanity */
431 /* reload stale pointer */
432 node
= cursor
->node
->ondisk
;
433 KKASSERT(cursor
->index
!= node
->count
);
439 * Check internal or leaf element. Determine if the record
440 * at the cursor has gone beyond the end of our range.
442 * We recurse down through internal nodes.
444 KKASSERT(cursor
->index
!= node
->count
);
445 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
446 elm
= &node
->elms
[cursor
->index
];
447 r
= hammer_btree_cmp(&cursor
->key_end
, &elm
[0].base
);
448 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
[1].base
);
449 if (hammer_debug_btree
) {
450 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
451 (long long)cursor
->node
->node_offset
,
453 (long long)elm
[0].internal
.base
.obj_id
,
454 elm
[0].internal
.base
.rec_type
,
455 (long long)elm
[0].internal
.base
.key
,
456 elm
[0].internal
.base
.localization
,
459 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
460 (long long)cursor
->node
->node_offset
,
462 (long long)elm
[1].internal
.base
.obj_id
,
463 elm
[1].internal
.base
.rec_type
,
464 (long long)elm
[1].internal
.base
.key
,
465 elm
[1].internal
.base
.localization
,
479 KKASSERT(elm
->internal
.subtree_offset
!= 0);
481 error
= hammer_cursor_down(cursor
);
484 KKASSERT(cursor
->index
== 0);
485 /* reload stale pointer */
486 node
= cursor
->node
->ondisk
;
488 /* this can assign -1 if the leaf was empty */
489 cursor
->index
= node
->count
- 1;
492 elm
= &node
->elms
[cursor
->index
];
493 s
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
494 if (hammer_debug_btree
) {
495 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
496 (long long)cursor
->node
->node_offset
,
498 (elm
[0].leaf
.base
.btype
?
499 elm
[0].leaf
.base
.btype
: '?'),
500 (long long)elm
[0].leaf
.base
.obj_id
,
501 elm
[0].leaf
.base
.rec_type
,
502 (long long)elm
[0].leaf
.base
.key
,
503 elm
[0].leaf
.base
.localization
,
512 switch(elm
->leaf
.base
.btype
) {
513 case HAMMER_BTREE_TYPE_RECORD
:
514 if ((cursor
->flags
& HAMMER_CURSOR_ASOF
) &&
515 hammer_btree_chkts(cursor
->asof
, &elm
->base
)) {
529 * node pointer invalid after loop
535 if (hammer_debug_btree
) {
536 int i
= cursor
->index
;
537 hammer_btree_elm_t elm
= &cursor
->node
->ondisk
->elms
[i
];
538 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
540 (long long)elm
->internal
.base
.obj_id
,
541 elm
->internal
.base
.rec_type
,
542 (long long)elm
->internal
.base
.key
,
543 elm
->internal
.base
.localization
552 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
553 * could not be found, EDEADLK if inserting and a retry is needed, and a
554 * fatal error otherwise. When retrying, the caller must terminate the
555 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
557 * The cursor is suitably positioned for a deletion on success, and suitably
558 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
561 * The cursor may begin anywhere, the search will traverse the tree in
562 * either direction to locate the requested element.
564 * Most of the logic implementing historical searches is handled here. We
565 * do an initial lookup with create_tid set to the asof TID. Due to the
566 * way records are laid out, a backwards iteration may be required if
567 * ENOENT is returned to locate the historical record. Here's the
570 * create_tid: 10 15 20
574 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
575 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
576 * not visible and thus causes ENOENT to be returned. We really need
577 * to check record 11 in LEAF1. If it also fails then the search fails
578 * (e.g. it might represent the range 11-16 and thus still not match our
579 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
580 * further iterations.
582 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
583 * and the cursor->create_check TID if an iteration might be needed.
584 * In the above example create_check would be set to 14.
587 hammer_btree_lookup(hammer_cursor_t cursor
)
591 KKASSERT ((cursor
->flags
& HAMMER_CURSOR_INSERT
) == 0 ||
592 cursor
->trans
->sync_lock_refs
> 0);
593 ++hammer_stats_btree_lookups
;
594 if (cursor
->flags
& HAMMER_CURSOR_ASOF
) {
595 KKASSERT((cursor
->flags
& HAMMER_CURSOR_INSERT
) == 0);
596 cursor
->key_beg
.create_tid
= cursor
->asof
;
598 cursor
->flags
&= ~HAMMER_CURSOR_CREATE_CHECK
;
599 error
= btree_search(cursor
, 0);
600 if (error
!= ENOENT
||
601 (cursor
->flags
& HAMMER_CURSOR_CREATE_CHECK
) == 0) {
604 * Stop if error other then ENOENT.
605 * Stop if ENOENT and not special case.
609 if (hammer_debug_btree
) {
610 kprintf("CREATE_CHECK %016llx\n",
611 (long long)cursor
->create_check
);
613 cursor
->key_beg
.create_tid
= cursor
->create_check
;
617 error
= btree_search(cursor
, 0);
620 error
= hammer_btree_extract(cursor
, cursor
->flags
);
625 * Execute the logic required to start an iteration. The first record
626 * located within the specified range is returned and iteration control
627 * flags are adjusted for successive hammer_btree_iterate() calls.
629 * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
630 * in a loop without worrying about it. Higher-level merged searches will
631 * adjust the flag appropriately.
634 hammer_btree_first(hammer_cursor_t cursor
)
638 error
= hammer_btree_lookup(cursor
);
639 if (error
== ENOENT
) {
640 cursor
->flags
&= ~HAMMER_CURSOR_ATEDISK
;
641 error
= hammer_btree_iterate(cursor
);
643 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
648 * Similarly but for an iteration in the reverse direction.
650 * Set ATEDISK when iterating backwards to skip the current entry,
651 * which after an ENOENT lookup will be pointing beyond our end point.
653 * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
654 * in a loop without worrying about it. Higher-level merged searches will
655 * adjust the flag appropriately.
658 hammer_btree_last(hammer_cursor_t cursor
)
660 struct hammer_base_elm save
;
663 save
= cursor
->key_beg
;
664 cursor
->key_beg
= cursor
->key_end
;
665 error
= hammer_btree_lookup(cursor
);
666 cursor
->key_beg
= save
;
667 if (error
== ENOENT
||
668 (cursor
->flags
& HAMMER_CURSOR_END_INCLUSIVE
) == 0) {
669 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
670 error
= hammer_btree_iterate_reverse(cursor
);
672 cursor
->flags
|= HAMMER_CURSOR_ATEDISK
;
677 * Extract the record and/or data associated with the cursor's current
678 * position. Any prior record or data stored in the cursor is replaced.
679 * The cursor must be positioned at a leaf node.
681 * NOTE: All extractions occur at the leaf of the B-Tree.
684 hammer_btree_extract(hammer_cursor_t cursor
, int flags
)
686 hammer_node_ondisk_t node
;
687 hammer_btree_elm_t elm
;
688 hammer_off_t data_off
;
694 * The case where the data reference resolves to the same buffer
695 * as the record reference must be handled.
697 node
= cursor
->node
->ondisk
;
698 elm
= &node
->elms
[cursor
->index
];
700 hmp
= cursor
->node
->hmp
;
703 * There is nothing to extract for an internal element.
705 if (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
)
709 * Only record types have data.
711 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
712 cursor
->leaf
= &elm
->leaf
;
714 if ((flags
& HAMMER_CURSOR_GET_DATA
) == 0)
716 if (elm
->leaf
.base
.btype
!= HAMMER_BTREE_TYPE_RECORD
)
718 data_off
= elm
->leaf
.data_offset
;
719 data_len
= elm
->leaf
.data_len
;
726 KKASSERT(data_len
>= 0 && data_len
<= HAMMER_XBUFSIZE
);
727 cursor
->data
= hammer_bread_ext(hmp
, data_off
, data_len
,
728 &error
, &cursor
->data_buffer
);
731 * Mark the data buffer as not being meta-data if it isn't
732 * meta-data (sometimes bulk data is accessed via a volume
736 switch(elm
->leaf
.base
.rec_type
) {
737 case HAMMER_RECTYPE_DATA
:
738 case HAMMER_RECTYPE_DB
:
739 hammer_io_notmeta(cursor
->data_buffer
);
747 * Deal with CRC errors on the extracted data.
750 hammer_crc_test_leaf(cursor
->data
, &elm
->leaf
) == 0) {
751 kprintf("CRC DATA @ %016llx/%d FAILED\n",
752 (long long)elm
->leaf
.data_offset
, elm
->leaf
.data_len
);
753 if (hammer_debug_critical
)
754 Debugger("CRC FAILED: DATA");
755 if (cursor
->trans
->flags
& HAMMER_TRANSF_CRCDOM
)
756 error
= EDOM
; /* less critical (mirroring) */
758 error
= EIO
; /* critical */
765 * Insert a leaf element into the B-Tree at the current cursor position.
766 * The cursor is positioned such that the element at and beyond the cursor
767 * are shifted to make room for the new record.
769 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
770 * flag set and that call must return ENOENT before this function can be
773 * The caller may depend on the cursor's exclusive lock after return to
774 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
776 * ENOSPC is returned if there is no room to insert a new record.
779 hammer_btree_insert(hammer_cursor_t cursor
, hammer_btree_leaf_elm_t elm
,
782 hammer_node_ondisk_t node
;
787 if ((error
= hammer_cursor_upgrade_node(cursor
)) != 0)
789 ++hammer_stats_btree_inserts
;
792 * Insert the element at the leaf node and update the count in the
793 * parent. It is possible for parent to be NULL, indicating that
794 * the filesystem's ROOT B-Tree node is a leaf itself, which is
795 * possible. The root inode can never be deleted so the leaf should
798 * Remember that the right-hand boundary is not included in the
801 hammer_modify_node_all(cursor
->trans
, cursor
->node
);
802 node
= cursor
->node
->ondisk
;
804 KKASSERT(elm
->base
.btype
!= 0);
805 KKASSERT(node
->type
== HAMMER_BTREE_TYPE_LEAF
);
806 KKASSERT(node
->count
< HAMMER_BTREE_LEAF_ELMS
);
807 if (i
!= node
->count
) {
808 bcopy(&node
->elms
[i
], &node
->elms
[i
+1],
809 (node
->count
- i
) * sizeof(*elm
));
811 node
->elms
[i
].leaf
= *elm
;
813 hammer_cursor_inserted_element(cursor
->node
, i
);
816 * Update the leaf node's aggregate mirror_tid for mirroring
819 if (node
->mirror_tid
< elm
->base
.delete_tid
) {
820 node
->mirror_tid
= elm
->base
.delete_tid
;
823 if (node
->mirror_tid
< elm
->base
.create_tid
) {
824 node
->mirror_tid
= elm
->base
.create_tid
;
827 hammer_modify_node_done(cursor
->node
);
830 * Debugging sanity checks.
832 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->base
) <= 0);
833 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->base
) > 0);
835 KKASSERT(hammer_btree_cmp(&node
->elms
[i
-1].leaf
.base
, &elm
->base
) < 0);
837 if (i
!= node
->count
- 1)
838 KKASSERT(hammer_btree_cmp(&node
->elms
[i
+1].leaf
.base
, &elm
->base
) > 0);
844 * Delete a record from the B-Tree at the current cursor position.
845 * The cursor is positioned such that the current element is the one
848 * On return the cursor will be positioned after the deleted element and
849 * MAY point to an internal node. It will be suitable for the continuation
850 * of an iteration but not for an insertion or deletion.
852 * Deletions will attempt to partially rebalance the B-Tree in an upward
853 * direction, but will terminate rather then deadlock. Empty internal nodes
854 * are never allowed by a deletion which deadlocks may end up giving us an
855 * empty leaf. The pruner will clean up and rebalance the tree.
857 * This function can return EDEADLK, requiring the caller to retry the
858 * operation after clearing the deadlock.
861 hammer_btree_delete(hammer_cursor_t cursor
)
863 hammer_node_ondisk_t ondisk
;
865 hammer_node_t parent
;
869 KKASSERT (cursor
->trans
->sync_lock_refs
> 0);
870 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
872 ++hammer_stats_btree_deletes
;
875 * Delete the element from the leaf node.
877 * Remember that leaf nodes do not have boundaries.
880 ondisk
= node
->ondisk
;
883 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_LEAF
);
884 KKASSERT(i
>= 0 && i
< ondisk
->count
);
885 hammer_modify_node_all(cursor
->trans
, node
);
886 if (i
+ 1 != ondisk
->count
) {
887 bcopy(&ondisk
->elms
[i
+1], &ondisk
->elms
[i
],
888 (ondisk
->count
- i
- 1) * sizeof(ondisk
->elms
[0]));
891 hammer_modify_node_done(node
);
892 hammer_cursor_deleted_element(node
, i
);
895 * Validate local parent
897 if (ondisk
->parent
) {
898 parent
= cursor
->parent
;
900 KKASSERT(parent
!= NULL
);
901 KKASSERT(parent
->node_offset
== ondisk
->parent
);
905 * If the leaf becomes empty it must be detached from the parent,
906 * potentially recursing through to the filesystem root.
908 * This may reposition the cursor at one of the parent's of the
911 * Ignore deadlock errors, that simply means that btree_remove
912 * was unable to recurse and had to leave us with an empty leaf.
914 KKASSERT(cursor
->index
<= ondisk
->count
);
915 if (ondisk
->count
== 0) {
916 error
= btree_remove(cursor
);
917 if (error
== EDEADLK
)
922 KKASSERT(cursor
->parent
== NULL
||
923 cursor
->parent_index
< cursor
->parent
->ondisk
->count
);
928 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
930 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
932 * The search can begin ANYWHERE in the B-Tree. As a first step the search
933 * iterates up the tree as necessary to properly position itself prior to
934 * actually doing the sarch.
936 * INSERTIONS: The search will split full nodes and leaves on its way down
937 * and guarentee that the leaf it ends up on is not full. If we run out
938 * of space the search continues to the leaf (to position the cursor for
939 * the spike), but ENOSPC is returned.
941 * The search is only guarenteed to end up on a leaf if an error code of 0
942 * is returned, or if inserting and an error code of ENOENT is returned.
943 * Otherwise it can stop at an internal node. On success a search returns
946 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
947 * filesystem, and it is not simple code. Please note the following facts:
949 * - Internal node recursions have a boundary on the left AND right. The
950 * right boundary is non-inclusive. The create_tid is a generic part
951 * of the key for internal nodes.
953 * - Leaf nodes contain terminal elements only now.
955 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
956 * historical search. ASOF and INSERT are mutually exclusive. When
957 * doing an as-of lookup btree_search() checks for a right-edge boundary
958 * case. If while recursing down the left-edge differs from the key
959 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
960 * with cursor->create_check. This is used by btree_lookup() to iterate.
961 * The iteration backwards because as-of searches can wind up going
962 * down the wrong branch of the B-Tree.
966 btree_search(hammer_cursor_t cursor
, int flags
)
968 hammer_node_ondisk_t node
;
969 hammer_btree_elm_t elm
;
976 flags
|= cursor
->flags
;
977 ++hammer_stats_btree_searches
;
979 if (hammer_debug_btree
) {
980 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
981 (long long)cursor
->node
->node_offset
,
983 (long long)cursor
->key_beg
.obj_id
,
984 cursor
->key_beg
.rec_type
,
985 (long long)cursor
->key_beg
.key
,
986 (long long)cursor
->key_beg
.create_tid
,
987 cursor
->key_beg
.localization
,
991 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
992 (long long)cursor
->parent
->node_offset
,
993 cursor
->parent_index
,
994 (long long)cursor
->left_bound
->obj_id
,
995 (long long)cursor
->parent
->ondisk
->elms
[cursor
->parent_index
].internal
.base
.obj_id
,
996 (long long)cursor
->right_bound
->obj_id
,
997 (long long)cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1].internal
.base
.obj_id
,
999 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
],
1000 cursor
->right_bound
,
1001 &cursor
->parent
->ondisk
->elms
[cursor
->parent_index
+1]
1006 * Move our cursor up the tree until we find a node whos range covers
1007 * the key we are trying to locate.
1009 * The left bound is inclusive, the right bound is non-inclusive.
1010 * It is ok to cursor up too far.
1013 r
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->left_bound
);
1014 s
= hammer_btree_cmp(&cursor
->key_beg
, cursor
->right_bound
);
1015 if (r
>= 0 && s
< 0)
1017 KKASSERT(cursor
->parent
);
1018 ++hammer_stats_btree_iterations
;
1019 error
= hammer_cursor_up(cursor
);
1025 * The delete-checks below are based on node, not parent. Set the
1026 * initial delete-check based on the parent.
1029 KKASSERT(cursor
->left_bound
->create_tid
!= 1);
1030 cursor
->create_check
= cursor
->left_bound
->create_tid
- 1;
1031 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
1035 * We better have ended up with a node somewhere.
1037 KKASSERT(cursor
->node
!= NULL
);
1040 * If we are inserting we can't start at a full node if the parent
1041 * is also full (because there is no way to split the node),
1042 * continue running up the tree until the requirement is satisfied
1043 * or we hit the root of the filesystem.
1045 * (If inserting we aren't doing an as-of search so we don't have
1046 * to worry about create_check).
1048 while ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
1049 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1050 if (btree_node_is_full(cursor
->node
->ondisk
) == 0)
1053 if (btree_node_is_full(cursor
->node
->ondisk
) ==0)
1056 if (cursor
->node
->ondisk
->parent
== 0 ||
1057 cursor
->parent
->ondisk
->count
!= HAMMER_BTREE_INT_ELMS
) {
1060 ++hammer_stats_btree_iterations
;
1061 error
= hammer_cursor_up(cursor
);
1062 /* node may have become stale */
1068 * Push down through internal nodes to locate the requested key.
1070 node
= cursor
->node
->ondisk
;
1071 while (node
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
1073 * Scan the node to find the subtree index to push down into.
1074 * We go one-past, then back-up.
1076 * We must proactively remove deleted elements which may
1077 * have been left over from a deadlocked btree_remove().
1079 * The left and right boundaries are included in the loop
1080 * in order to detect edge cases.
1082 * If the separator only differs by create_tid (r == 1)
1083 * and we are doing an as-of search, we may end up going
1084 * down a branch to the left of the one containing the
1085 * desired key. This requires numerous special cases.
1087 ++hammer_stats_btree_iterations
;
1088 if (hammer_debug_btree
) {
1089 kprintf("SEARCH-I %016llx count=%d\n",
1090 (long long)cursor
->node
->node_offset
,
1095 * Try to shortcut the search before dropping into the
1096 * linear loop. Locate the first node where r <= 1.
1098 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
1099 while (i
<= node
->count
) {
1100 ++hammer_stats_btree_elements
;
1101 elm
= &node
->elms
[i
];
1102 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->base
);
1103 if (hammer_debug_btree
> 2) {
1104 kprintf(" IELM %p %d r=%d\n",
1105 &node
->elms
[i
], i
, r
);
1110 KKASSERT(elm
->base
.create_tid
!= 1);
1111 cursor
->create_check
= elm
->base
.create_tid
- 1;
1112 cursor
->flags
|= HAMMER_CURSOR_CREATE_CHECK
;
1116 if (hammer_debug_btree
) {
1117 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1122 * These cases occur when the parent's idea of the boundary
1123 * is wider then the child's idea of the boundary, and
1124 * require special handling. If not inserting we can
1125 * terminate the search early for these cases but the
1126 * child's boundaries cannot be unconditionally modified.
1130 * If i == 0 the search terminated to the LEFT of the
1131 * left_boundary but to the RIGHT of the parent's left
1136 elm
= &node
->elms
[0];
1139 * If we aren't inserting we can stop here.
1141 if ((flags
& (HAMMER_CURSOR_INSERT
|
1142 HAMMER_CURSOR_PRUNING
)) == 0) {
1148 * Correct a left-hand boundary mismatch.
1150 * We can only do this if we can upgrade the lock,
1151 * and synchronized as a background cursor (i.e.
1152 * inserting or pruning).
1154 * WARNING: We can only do this if inserting, i.e.
1155 * we are running on the backend.
1157 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1159 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1160 hammer_modify_node_field(cursor
->trans
, cursor
->node
,
1162 save
= node
->elms
[0].base
.btype
;
1163 node
->elms
[0].base
= *cursor
->left_bound
;
1164 node
->elms
[0].base
.btype
= save
;
1165 hammer_modify_node_done(cursor
->node
);
1166 } else if (i
== node
->count
+ 1) {
1168 * If i == node->count + 1 the search terminated to
1169 * the RIGHT of the right boundary but to the LEFT
1170 * of the parent's right boundary. If we aren't
1171 * inserting we can stop here.
1173 * Note that the last element in this case is
1174 * elms[i-2] prior to adjustments to 'i'.
1177 if ((flags
& (HAMMER_CURSOR_INSERT
|
1178 HAMMER_CURSOR_PRUNING
)) == 0) {
1184 * Correct a right-hand boundary mismatch.
1185 * (actual push-down record is i-2 prior to
1186 * adjustments to i).
1188 * We can only do this if we can upgrade the lock,
1189 * and synchronized as a background cursor (i.e.
1190 * inserting or pruning).
1192 * WARNING: We can only do this if inserting, i.e.
1193 * we are running on the backend.
1195 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1197 elm
= &node
->elms
[i
];
1198 KKASSERT(cursor
->flags
& HAMMER_CURSOR_BACKEND
);
1199 hammer_modify_node(cursor
->trans
, cursor
->node
,
1200 &elm
->base
, sizeof(elm
->base
));
1201 elm
->base
= *cursor
->right_bound
;
1202 hammer_modify_node_done(cursor
->node
);
1206 * The push-down index is now i - 1. If we had
1207 * terminated on the right boundary this will point
1208 * us at the last element.
1213 elm
= &node
->elms
[i
];
1215 if (hammer_debug_btree
) {
1216 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1217 "key=%016llx cre=%016llx lo=%02x\n",
1218 (long long)cursor
->node
->node_offset
,
1220 (long long)elm
->internal
.base
.obj_id
,
1221 elm
->internal
.base
.rec_type
,
1222 (long long)elm
->internal
.base
.key
,
1223 (long long)elm
->internal
.base
.create_tid
,
1224 elm
->internal
.base
.localization
1229 * We better have a valid subtree offset.
1231 KKASSERT(elm
->internal
.subtree_offset
!= 0);
1234 * Handle insertion and deletion requirements.
1236 * If inserting split full nodes. The split code will
1237 * adjust cursor->node and cursor->index if the current
1238 * index winds up in the new node.
1240 * If inserting and a left or right edge case was detected,
1241 * we cannot correct the left or right boundary and must
1242 * prepend and append an empty leaf node in order to make
1243 * the boundary correction.
1245 * If we run out of space we set enospc and continue on
1246 * to a leaf to provide the spike code with a good point
1249 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0) {
1250 if (btree_node_is_full(node
)) {
1251 error
= btree_split_internal(cursor
);
1253 if (error
!= ENOSPC
)
1258 * reload stale pointers
1261 node
= cursor
->node
->ondisk
;
1266 * Push down (push into new node, existing node becomes
1267 * the parent) and continue the search.
1269 error
= hammer_cursor_down(cursor
);
1270 /* node may have become stale */
1273 node
= cursor
->node
->ondisk
;
1277 * We are at a leaf, do a linear search of the key array.
1279 * On success the index is set to the matching element and 0
1282 * On failure the index is set to the insertion point and ENOENT
1285 * Boundaries are not stored in leaf nodes, so the index can wind
1286 * up to the left of element 0 (index == 0) or past the end of
1287 * the array (index == node->count). It is also possible that the
1288 * leaf might be empty.
1290 ++hammer_stats_btree_iterations
;
1291 KKASSERT (node
->type
== HAMMER_BTREE_TYPE_LEAF
);
1292 KKASSERT(node
->count
<= HAMMER_BTREE_LEAF_ELMS
);
1293 if (hammer_debug_btree
) {
1294 kprintf("SEARCH-L %016llx count=%d\n",
1295 (long long)cursor
->node
->node_offset
,
1300 * Try to shortcut the search before dropping into the
1301 * linear loop. Locate the first node where r <= 1.
1303 i
= hammer_btree_search_node(&cursor
->key_beg
, node
);
1304 while (i
< node
->count
) {
1305 ++hammer_stats_btree_elements
;
1306 elm
= &node
->elms
[i
];
1308 r
= hammer_btree_cmp(&cursor
->key_beg
, &elm
->leaf
.base
);
1310 if (hammer_debug_btree
> 1)
1311 kprintf(" ELM %p %d r=%d\n", &node
->elms
[i
], i
, r
);
1314 * We are at a record element. Stop if we've flipped past
1315 * key_beg, not counting the create_tid test. Allow the
1316 * r == 1 case (key_beg > element but differs only by its
1317 * create_tid) to fall through to the AS-OF check.
1319 KKASSERT (elm
->leaf
.base
.btype
== HAMMER_BTREE_TYPE_RECORD
);
1329 * Check our as-of timestamp against the element.
1331 if (flags
& HAMMER_CURSOR_ASOF
) {
1332 if (hammer_btree_chkts(cursor
->asof
,
1333 &node
->elms
[i
].base
) != 0) {
1339 if (r
> 0) { /* can only be +1 */
1347 if (hammer_debug_btree
) {
1348 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1349 (long long)cursor
->node
->node_offset
, i
);
1355 * The search of the leaf node failed. i is the insertion point.
1358 if (hammer_debug_btree
) {
1359 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1360 (long long)cursor
->node
->node_offset
, i
);
1364 * No exact match was found, i is now at the insertion point.
1366 * If inserting split a full leaf before returning. This
1367 * may have the side effect of adjusting cursor->node and
1371 if ((flags
& HAMMER_CURSOR_INSERT
) && enospc
== 0 &&
1372 btree_node_is_full(node
)) {
1373 error
= btree_split_leaf(cursor
);
1375 if (error
!= ENOSPC
)
1380 * reload stale pointers
1384 node = &cursor->node->internal;
1389 * We reached a leaf but did not find the key we were looking for.
1390 * If this is an insert we will be properly positioned for an insert
1391 * (ENOENT) or spike (ENOSPC) operation.
1393 error
= enospc
? ENOSPC
: ENOENT
;
1399 * Heuristical search for the first element whos comparison is <= 1. May
1400 * return an index whos compare result is > 1 but may only return an index
1401 * whos compare result is <= 1 if it is the first element with that result.
1404 hammer_btree_search_node(hammer_base_elm_t elm
, hammer_node_ondisk_t node
)
1412 * Don't bother if the node does not have very many elements
1417 i
= b
+ (s
- b
) / 2;
1418 ++hammer_stats_btree_elements
;
1419 r
= hammer_btree_cmp(elm
, &node
->elms
[i
].leaf
.base
);
1430 /************************************************************************
1431 * SPLITTING AND MERGING *
1432 ************************************************************************
1434 * These routines do all the dirty work required to split and merge nodes.
1438 * Split an internal node into two nodes and move the separator at the split
1439 * point to the parent.
1441 * (cursor->node, cursor->index) indicates the element the caller intends
1442 * to push into. We will adjust node and index if that element winds
1443 * up in the split node.
1445 * If we are at the root of the filesystem a new root must be created with
1446 * two elements, one pointing to the original root and one pointing to the
1447 * newly allocated split node.
1451 btree_split_internal(hammer_cursor_t cursor
)
1453 hammer_node_ondisk_t ondisk
;
1455 hammer_node_t parent
;
1456 hammer_node_t new_node
;
1457 hammer_btree_elm_t elm
;
1458 hammer_btree_elm_t parent_elm
;
1459 struct hammer_node_lock lockroot
;
1460 hammer_mount_t hmp
= cursor
->trans
->hmp
;
1467 const int esize
= sizeof(*elm
);
1469 hammer_node_lock_init(&lockroot
, cursor
->node
);
1470 error
= hammer_btree_lock_children(cursor
, 1, &lockroot
);
1473 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1475 ++hammer_stats_btree_splits
;
1478 * Calculate the split point. If the insertion point is at the
1479 * end of the leaf we adjust the split point significantly to the
1480 * right to try to optimize node fill and flag it. If we hit
1481 * that same leaf again our heuristic failed and we don't try
1482 * to optimize node fill (it could lead to a degenerate case).
1484 node
= cursor
->node
;
1485 ondisk
= node
->ondisk
;
1486 KKASSERT(ondisk
->count
> 4);
1487 if (cursor
->index
== ondisk
->count
&&
1488 (node
->flags
& HAMMER_NODE_NONLINEAR
) == 0) {
1489 split
= (ondisk
->count
+ 1) * 3 / 4;
1490 node
->flags
|= HAMMER_NODE_NONLINEAR
;
1493 * We are splitting but elms[split] will be promoted to
1494 * the parent, leaving the right hand node with one less
1495 * element. If the insertion point will be on the
1496 * left-hand side adjust the split point to give the
1497 * right hand side one additional node.
1499 split
= (ondisk
->count
+ 1) / 2;
1500 if (cursor
->index
<= split
)
1505 * If we are at the root of the filesystem, create a new root node
1506 * with 1 element and split normally. Avoid making major
1507 * modifications until we know the whole operation will work.
1509 if (ondisk
->parent
== 0) {
1510 parent
= hammer_alloc_btree(cursor
->trans
, node
->node_offset
,
1514 hammer_lock_ex(&parent
->lock
);
1515 hammer_modify_node_noundo(cursor
->trans
, parent
);
1516 ondisk
= parent
->ondisk
;
1519 ondisk
->mirror_tid
= node
->ondisk
->mirror_tid
;
1520 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1521 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1522 ondisk
->elms
[0].base
.btype
= node
->ondisk
->type
;
1523 ondisk
->elms
[0].internal
.subtree_offset
= node
->node_offset
;
1524 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1525 hammer_modify_node_done(parent
);
1526 /* ondisk->elms[1].base.btype - not used */
1528 parent_index
= 0; /* index of current node in parent */
1531 parent
= cursor
->parent
;
1532 parent_index
= cursor
->parent_index
;
1536 * Calculate a hint for the allocation of the new B-Tree node.
1537 * The most likely expansion is coming from the insertion point
1538 * at cursor->index, so try to localize the allocation of our
1539 * new node to accomodate that sub-tree.
1541 * Use the right-most sub-tree when expandinging on the right edge.
1542 * This is a very common case when copying a directory tree.
1544 if (cursor
->index
== ondisk
->count
)
1545 hint
= ondisk
->elms
[cursor
->index
- 1].internal
.subtree_offset
;
1547 hint
= ondisk
->elms
[cursor
->index
].internal
.subtree_offset
;
1550 * Split node into new_node at the split point.
1552 * B O O O P N N B <-- P = node->elms[split] (index 4)
1553 * 0 1 2 3 4 5 6 <-- subtree indices
1558 * B O O O B B N N B <--- inner boundary points are 'P'
1561 new_node
= hammer_alloc_btree(cursor
->trans
, hint
, &error
);
1562 if (new_node
== NULL
) {
1564 hammer_unlock(&parent
->lock
);
1565 hammer_delete_node(cursor
->trans
, parent
);
1566 hammer_rel_node(parent
);
1570 hammer_lock_ex(&new_node
->lock
);
1573 * Create the new node. P becomes the left-hand boundary in the
1574 * new node. Copy the right-hand boundary as well.
1576 * elm is the new separator.
1578 hammer_modify_node_noundo(cursor
->trans
, new_node
);
1579 hammer_modify_node_all(cursor
->trans
, node
);
1580 ondisk
= node
->ondisk
;
1581 elm
= &ondisk
->elms
[split
];
1582 bcopy(elm
, &new_node
->ondisk
->elms
[0],
1583 (ondisk
->count
- split
+ 1) * esize
);
1584 new_node
->ondisk
->count
= ondisk
->count
- split
;
1585 new_node
->ondisk
->parent
= parent
->node_offset
;
1586 new_node
->ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1587 new_node
->ondisk
->mirror_tid
= ondisk
->mirror_tid
;
1588 KKASSERT(ondisk
->type
== new_node
->ondisk
->type
);
1589 hammer_cursor_split_node(node
, new_node
, split
);
1592 * Cleanup the original node. Elm (P) becomes the new boundary,
1593 * its subtree_offset was moved to the new node. If we had created
1594 * a new root its parent pointer may have changed.
1596 elm
->internal
.subtree_offset
= 0;
1597 ondisk
->count
= split
;
1600 * Insert the separator into the parent, fixup the parent's
1601 * reference to the original node, and reference the new node.
1602 * The separator is P.
1604 * Remember that base.count does not include the right-hand boundary.
1606 hammer_modify_node_all(cursor
->trans
, parent
);
1607 ondisk
= parent
->ondisk
;
1608 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1609 parent_elm
= &ondisk
->elms
[parent_index
+1];
1610 bcopy(parent_elm
, parent_elm
+ 1,
1611 (ondisk
->count
- parent_index
) * esize
);
1612 parent_elm
->internal
.base
= elm
->base
; /* separator P */
1613 parent_elm
->internal
.base
.btype
= new_node
->ondisk
->type
;
1614 parent_elm
->internal
.subtree_offset
= new_node
->node_offset
;
1615 parent_elm
->internal
.mirror_tid
= new_node
->ondisk
->mirror_tid
;
1617 hammer_modify_node_done(parent
);
1618 hammer_cursor_inserted_element(parent
, parent_index
+ 1);
1621 * The children of new_node need their parent pointer set to new_node.
1622 * The children have already been locked by
1623 * hammer_btree_lock_children().
1625 for (i
= 0; i
< new_node
->ondisk
->count
; ++i
) {
1626 elm
= &new_node
->ondisk
->elms
[i
];
1627 error
= btree_set_parent(cursor
->trans
, new_node
, elm
);
1629 panic("btree_split_internal: btree-fixup problem");
1632 hammer_modify_node_done(new_node
);
1635 * The filesystem's root B-Tree pointer may have to be updated.
1638 hammer_volume_t volume
;
1640 volume
= hammer_get_root_volume(hmp
, &error
);
1641 KKASSERT(error
== 0);
1643 hammer_modify_volume_field(cursor
->trans
, volume
,
1645 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1646 hammer_modify_volume_done(volume
);
1647 node
->ondisk
->parent
= parent
->node_offset
;
1648 if (cursor
->parent
) {
1649 hammer_unlock(&cursor
->parent
->lock
);
1650 hammer_rel_node(cursor
->parent
);
1652 cursor
->parent
= parent
; /* lock'd and ref'd */
1653 hammer_rel_volume(volume
, 0);
1655 hammer_modify_node_done(node
);
1658 * Ok, now adjust the cursor depending on which element the original
1659 * index was pointing at. If we are >= the split point the push node
1660 * is now in the new node.
1662 * NOTE: If we are at the split point itself we cannot stay with the
1663 * original node because the push index will point at the right-hand
1664 * boundary, which is illegal.
1666 * NOTE: The cursor's parent or parent_index must be adjusted for
1667 * the case where a new parent (new root) was created, and the case
1668 * where the cursor is now pointing at the split node.
1670 if (cursor
->index
>= split
) {
1671 cursor
->parent_index
= parent_index
+ 1;
1672 cursor
->index
-= split
;
1673 hammer_unlock(&cursor
->node
->lock
);
1674 hammer_rel_node(cursor
->node
);
1675 cursor
->node
= new_node
; /* locked and ref'd */
1677 cursor
->parent_index
= parent_index
;
1678 hammer_unlock(&new_node
->lock
);
1679 hammer_rel_node(new_node
);
1683 * Fixup left and right bounds
1685 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1686 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1687 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1688 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1689 &cursor
->node
->ondisk
->elms
[0].internal
.base
) <= 0);
1690 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1691 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
].internal
.base
) >= 0);
1694 hammer_btree_unlock_children(cursor
, &lockroot
);
1695 hammer_cursor_downgrade(cursor
);
1700 * Same as the above, but splits a full leaf node.
1706 btree_split_leaf(hammer_cursor_t cursor
)
1708 hammer_node_ondisk_t ondisk
;
1709 hammer_node_t parent
;
1712 hammer_node_t new_leaf
;
1713 hammer_btree_elm_t elm
;
1714 hammer_btree_elm_t parent_elm
;
1715 hammer_base_elm_t mid_boundary
;
1721 const size_t esize
= sizeof(*elm
);
1723 if ((error
= hammer_cursor_upgrade(cursor
)) != 0)
1725 ++hammer_stats_btree_splits
;
1727 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1728 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1729 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1730 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1733 * Calculate the split point. If the insertion point is at the
1734 * end of the leaf we adjust the split point significantly to the
1735 * right to try to optimize node fill and flag it. If we hit
1736 * that same leaf again our heuristic failed and we don't try
1737 * to optimize node fill (it could lead to a degenerate case).
1739 * Spikes are made up of two leaf elements which cannot be
1742 leaf
= cursor
->node
;
1743 ondisk
= leaf
->ondisk
;
1744 KKASSERT(ondisk
->count
> 4);
1745 if (cursor
->index
== ondisk
->count
&&
1746 (leaf
->flags
& HAMMER_NODE_NONLINEAR
) == 0) {
1747 split
= (ondisk
->count
+ 1) * 3 / 4;
1748 leaf
->flags
|= HAMMER_NODE_NONLINEAR
;
1750 split
= (ondisk
->count
+ 1) / 2;
1755 * If the insertion point is at the split point shift the
1756 * split point left so we don't have to worry about
1758 if (cursor
->index
== split
)
1761 KKASSERT(split
> 0 && split
< ondisk
->count
);
1766 elm
= &ondisk
->elms
[split
];
1768 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
[-1].leaf
.base
) <= 0);
1769 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &elm
->leaf
.base
) <= 0);
1770 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
->leaf
.base
) > 0);
1771 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &elm
[1].leaf
.base
) > 0);
1774 * If we are at the root of the tree, create a new root node with
1775 * 1 element and split normally. Avoid making major modifications
1776 * until we know the whole operation will work.
1778 if (ondisk
->parent
== 0) {
1779 parent
= hammer_alloc_btree(cursor
->trans
, leaf
->node_offset
,
1783 hammer_lock_ex(&parent
->lock
);
1784 hammer_modify_node_noundo(cursor
->trans
, parent
);
1785 ondisk
= parent
->ondisk
;
1788 ondisk
->mirror_tid
= leaf
->ondisk
->mirror_tid
;
1789 ondisk
->type
= HAMMER_BTREE_TYPE_INTERNAL
;
1790 ondisk
->elms
[0].base
= hmp
->root_btree_beg
;
1791 ondisk
->elms
[0].base
.btype
= leaf
->ondisk
->type
;
1792 ondisk
->elms
[0].internal
.subtree_offset
= leaf
->node_offset
;
1793 ondisk
->elms
[1].base
= hmp
->root_btree_end
;
1794 /* ondisk->elms[1].base.btype = not used */
1795 hammer_modify_node_done(parent
);
1797 parent_index
= 0; /* insertion point in parent */
1800 parent
= cursor
->parent
;
1801 parent_index
= cursor
->parent_index
;
1805 * Calculate a hint for the allocation of the new B-Tree leaf node.
1806 * For now just try to localize it within the same bigblock as
1809 * If the insertion point is at the end of the leaf we recognize a
1810 * likely append sequence of some sort (data, meta-data, inodes,
1811 * whatever). Set the hint to zero to allocate out of linear space
1812 * instead of trying to completely fill previously hinted space.
1814 * This also sets the stage for recursive splits to localize using
1817 ondisk
= leaf
->ondisk
;
1818 if (cursor
->index
== ondisk
->count
)
1821 hint
= leaf
->node_offset
;
1824 * Split leaf into new_leaf at the split point. Select a separator
1825 * value in-between the two leafs but with a bent towards the right
1826 * leaf since comparisons use an 'elm >= separator' inequality.
1835 new_leaf
= hammer_alloc_btree(cursor
->trans
, hint
, &error
);
1836 if (new_leaf
== NULL
) {
1838 hammer_unlock(&parent
->lock
);
1839 hammer_delete_node(cursor
->trans
, parent
);
1840 hammer_rel_node(parent
);
1844 hammer_lock_ex(&new_leaf
->lock
);
1847 * Create the new node and copy the leaf elements from the split
1848 * point on to the new node.
1850 hammer_modify_node_all(cursor
->trans
, leaf
);
1851 hammer_modify_node_noundo(cursor
->trans
, new_leaf
);
1852 ondisk
= leaf
->ondisk
;
1853 elm
= &ondisk
->elms
[split
];
1854 bcopy(elm
, &new_leaf
->ondisk
->elms
[0], (ondisk
->count
- split
) * esize
);
1855 new_leaf
->ondisk
->count
= ondisk
->count
- split
;
1856 new_leaf
->ondisk
->parent
= parent
->node_offset
;
1857 new_leaf
->ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
1858 new_leaf
->ondisk
->mirror_tid
= ondisk
->mirror_tid
;
1859 KKASSERT(ondisk
->type
== new_leaf
->ondisk
->type
);
1860 hammer_modify_node_done(new_leaf
);
1861 hammer_cursor_split_node(leaf
, new_leaf
, split
);
1864 * Cleanup the original node. Because this is a leaf node and
1865 * leaf nodes do not have a right-hand boundary, there
1866 * aren't any special edge cases to clean up. We just fixup the
1869 ondisk
->count
= split
;
1872 * Insert the separator into the parent, fixup the parent's
1873 * reference to the original node, and reference the new node.
1874 * The separator is P.
1876 * Remember that base.count does not include the right-hand boundary.
1877 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1879 hammer_modify_node_all(cursor
->trans
, parent
);
1880 ondisk
= parent
->ondisk
;
1881 KKASSERT(split
!= 0);
1882 KKASSERT(ondisk
->count
!= HAMMER_BTREE_INT_ELMS
);
1883 parent_elm
= &ondisk
->elms
[parent_index
+1];
1884 bcopy(parent_elm
, parent_elm
+ 1,
1885 (ondisk
->count
- parent_index
) * esize
);
1887 hammer_make_separator(&elm
[-1].base
, &elm
[0].base
, &parent_elm
->base
);
1888 parent_elm
->internal
.base
.btype
= new_leaf
->ondisk
->type
;
1889 parent_elm
->internal
.subtree_offset
= new_leaf
->node_offset
;
1890 parent_elm
->internal
.mirror_tid
= new_leaf
->ondisk
->mirror_tid
;
1891 mid_boundary
= &parent_elm
->base
;
1893 hammer_modify_node_done(parent
);
1894 hammer_cursor_inserted_element(parent
, parent_index
+ 1);
1897 * The filesystem's root B-Tree pointer may have to be updated.
1900 hammer_volume_t volume
;
1902 volume
= hammer_get_root_volume(hmp
, &error
);
1903 KKASSERT(error
== 0);
1905 hammer_modify_volume_field(cursor
->trans
, volume
,
1907 volume
->ondisk
->vol0_btree_root
= parent
->node_offset
;
1908 hammer_modify_volume_done(volume
);
1909 leaf
->ondisk
->parent
= parent
->node_offset
;
1910 if (cursor
->parent
) {
1911 hammer_unlock(&cursor
->parent
->lock
);
1912 hammer_rel_node(cursor
->parent
);
1914 cursor
->parent
= parent
; /* lock'd and ref'd */
1915 hammer_rel_volume(volume
, 0);
1917 hammer_modify_node_done(leaf
);
1920 * Ok, now adjust the cursor depending on which element the original
1921 * index was pointing at. If we are >= the split point the push node
1922 * is now in the new node.
1924 * NOTE: If we are at the split point itself we need to select the
1925 * old or new node based on where key_beg's insertion point will be.
1926 * If we pick the wrong side the inserted element will wind up in
1927 * the wrong leaf node and outside that node's bounds.
1929 if (cursor
->index
> split
||
1930 (cursor
->index
== split
&&
1931 hammer_btree_cmp(&cursor
->key_beg
, mid_boundary
) >= 0)) {
1932 cursor
->parent_index
= parent_index
+ 1;
1933 cursor
->index
-= split
;
1934 hammer_unlock(&cursor
->node
->lock
);
1935 hammer_rel_node(cursor
->node
);
1936 cursor
->node
= new_leaf
;
1938 cursor
->parent_index
= parent_index
;
1939 hammer_unlock(&new_leaf
->lock
);
1940 hammer_rel_node(new_leaf
);
1944 * Fixup left and right bounds
1946 parent_elm
= &parent
->ondisk
->elms
[cursor
->parent_index
];
1947 cursor
->left_bound
= &parent_elm
[0].internal
.base
;
1948 cursor
->right_bound
= &parent_elm
[1].internal
.base
;
1951 * Assert that the bounds are correct.
1953 KKASSERT(hammer_btree_cmp(cursor
->left_bound
,
1954 &cursor
->node
->ondisk
->elms
[0].leaf
.base
) <= 0);
1955 KKASSERT(hammer_btree_cmp(cursor
->right_bound
,
1956 &cursor
->node
->ondisk
->elms
[cursor
->node
->ondisk
->count
-1].leaf
.base
) > 0);
1957 KKASSERT(hammer_btree_cmp(cursor
->left_bound
, &cursor
->key_beg
) <= 0);
1958 KKASSERT(hammer_btree_cmp(cursor
->right_bound
, &cursor
->key_beg
) > 0);
1961 hammer_cursor_downgrade(cursor
);
1968 * Recursively correct the right-hand boundary's create_tid to (tid) as
1969 * long as the rest of the key matches. We have to recurse upward in
1970 * the tree as well as down the left side of each parent's right node.
1972 * Return EDEADLK if we were only partially successful, forcing the caller
1973 * to try again. The original cursor is not modified. This routine can
1974 * also fail with EDEADLK if it is forced to throw away a portion of its
1977 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1980 TAILQ_ENTRY(hammer_rhb
) entry
;
1985 TAILQ_HEAD(hammer_rhb_list
, hammer_rhb
);
1988 hammer_btree_correct_rhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
1990 struct hammer_mount
*hmp
;
1991 struct hammer_rhb_list rhb_list
;
1992 hammer_base_elm_t elm
;
1993 hammer_node_t orig_node
;
1994 struct hammer_rhb
*rhb
;
1998 TAILQ_INIT(&rhb_list
);
1999 hmp
= cursor
->trans
->hmp
;
2002 * Save our position so we can restore it on return. This also
2003 * gives us a stable 'elm'.
2005 orig_node
= cursor
->node
;
2006 hammer_ref_node(orig_node
);
2007 hammer_lock_sh(&orig_node
->lock
);
2008 orig_index
= cursor
->index
;
2009 elm
= &orig_node
->ondisk
->elms
[orig_index
].base
;
2012 * Now build a list of parents going up, allocating a rhb
2013 * structure for each one.
2015 while (cursor
->parent
) {
2017 * Stop if we no longer have any right-bounds to fix up
2019 if (elm
->obj_id
!= cursor
->right_bound
->obj_id
||
2020 elm
->rec_type
!= cursor
->right_bound
->rec_type
||
2021 elm
->key
!= cursor
->right_bound
->key
) {
2026 * Stop if the right-hand bound's create_tid does not
2027 * need to be corrected.
2029 if (cursor
->right_bound
->create_tid
>= tid
)
2032 rhb
= kmalloc(sizeof(*rhb
), hmp
->m_misc
, M_WAITOK
|M_ZERO
);
2033 rhb
->node
= cursor
->parent
;
2034 rhb
->index
= cursor
->parent_index
;
2035 hammer_ref_node(rhb
->node
);
2036 hammer_lock_sh(&rhb
->node
->lock
);
2037 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
2039 hammer_cursor_up(cursor
);
2043 * now safely adjust the right hand bound for each rhb. This may
2044 * also require taking the right side of the tree and iterating down
2048 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
2049 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
2052 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
2053 hammer_unlock(&rhb
->node
->lock
);
2054 hammer_rel_node(rhb
->node
);
2055 kfree(rhb
, hmp
->m_misc
);
2057 switch (cursor
->node
->ondisk
->type
) {
2058 case HAMMER_BTREE_TYPE_INTERNAL
:
2060 * Right-boundary for parent at internal node
2061 * is one element to the right of the element whos
2062 * right boundary needs adjusting. We must then
2063 * traverse down the left side correcting any left
2064 * bounds (which may now be too far to the left).
2067 error
= hammer_btree_correct_lhb(cursor
, tid
);
2070 panic("hammer_btree_correct_rhb(): Bad node type");
2079 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
2080 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
2081 hammer_unlock(&rhb
->node
->lock
);
2082 hammer_rel_node(rhb
->node
);
2083 kfree(rhb
, hmp
->m_misc
);
2085 error
= hammer_cursor_seek(cursor
, orig_node
, orig_index
);
2086 hammer_unlock(&orig_node
->lock
);
2087 hammer_rel_node(orig_node
);
2092 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2093 * bound going downward starting at the current cursor position.
2095 * This function does not restore the cursor after use.
2098 hammer_btree_correct_lhb(hammer_cursor_t cursor
, hammer_tid_t tid
)
2100 struct hammer_rhb_list rhb_list
;
2101 hammer_base_elm_t elm
;
2102 hammer_base_elm_t cmp
;
2103 struct hammer_rhb
*rhb
;
2104 struct hammer_mount
*hmp
;
2107 TAILQ_INIT(&rhb_list
);
2108 hmp
= cursor
->trans
->hmp
;
2110 cmp
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
2113 * Record the node and traverse down the left-hand side for all
2114 * matching records needing a boundary correction.
2118 rhb
= kmalloc(sizeof(*rhb
), hmp
->m_misc
, M_WAITOK
|M_ZERO
);
2119 rhb
->node
= cursor
->node
;
2120 rhb
->index
= cursor
->index
;
2121 hammer_ref_node(rhb
->node
);
2122 hammer_lock_sh(&rhb
->node
->lock
);
2123 TAILQ_INSERT_HEAD(&rhb_list
, rhb
, entry
);
2125 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
2127 * Nothing to traverse down if we are at the right
2128 * boundary of an internal node.
2130 if (cursor
->index
== cursor
->node
->ondisk
->count
)
2133 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
2134 if (elm
->btype
== HAMMER_BTREE_TYPE_RECORD
)
2136 panic("Illegal leaf record type %02x", elm
->btype
);
2138 error
= hammer_cursor_down(cursor
);
2142 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
2143 if (elm
->obj_id
!= cmp
->obj_id
||
2144 elm
->rec_type
!= cmp
->rec_type
||
2145 elm
->key
!= cmp
->key
) {
2148 if (elm
->create_tid
>= tid
)
2154 * Now we can safely adjust the left-hand boundary from the bottom-up.
2155 * The last element we remove from the list is the caller's right hand
2156 * boundary, which must also be adjusted.
2158 while (error
== 0 && (rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
2159 error
= hammer_cursor_seek(cursor
, rhb
->node
, rhb
->index
);
2162 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
2163 hammer_unlock(&rhb
->node
->lock
);
2164 hammer_rel_node(rhb
->node
);
2165 kfree(rhb
, hmp
->m_misc
);
2167 elm
= &cursor
->node
->ondisk
->elms
[cursor
->index
].base
;
2168 if (cursor
->node
->ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
2169 hammer_modify_node(cursor
->trans
, cursor
->node
,
2171 sizeof(elm
->create_tid
));
2172 elm
->create_tid
= tid
;
2173 hammer_modify_node_done(cursor
->node
);
2175 panic("hammer_btree_correct_lhb(): Bad element type");
2182 while ((rhb
= TAILQ_FIRST(&rhb_list
)) != NULL
) {
2183 TAILQ_REMOVE(&rhb_list
, rhb
, entry
);
2184 hammer_unlock(&rhb
->node
->lock
);
2185 hammer_rel_node(rhb
->node
);
2186 kfree(rhb
, hmp
->m_misc
);
2194 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2195 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2196 * the operation due to a deadlock, or some other error.
2198 * This routine is initially called with an empty leaf and may be
2199 * recursively called with single-element internal nodes.
2201 * It should also be noted that when removing empty leaves we must be sure
2202 * to test and update mirror_tid because another thread may have deadlocked
2203 * against us (or someone) trying to propagate it up and cannot retry once
2204 * the node has been deleted.
2206 * On return the cursor may end up pointing to an internal node, suitable
2207 * for further iteration but not for an immediate insertion or deletion.
2210 btree_remove(hammer_cursor_t cursor
)
2212 hammer_node_ondisk_t ondisk
;
2213 hammer_btree_elm_t elm
;
2215 hammer_node_t parent
;
2216 const int esize
= sizeof(*elm
);
2219 node
= cursor
->node
;
2222 * When deleting the root of the filesystem convert it to
2223 * an empty leaf node. Internal nodes cannot be empty.
2225 ondisk
= node
->ondisk
;
2226 if (ondisk
->parent
== 0) {
2227 KKASSERT(cursor
->parent
== NULL
);
2228 hammer_modify_node_all(cursor
->trans
, node
);
2229 KKASSERT(ondisk
== node
->ondisk
);
2230 ondisk
->type
= HAMMER_BTREE_TYPE_LEAF
;
2232 hammer_modify_node_done(node
);
2237 parent
= cursor
->parent
;
2240 * Attempt to remove the parent's reference to the child. If the
2241 * parent would become empty we have to recurse. If we fail we
2242 * leave the parent pointing to an empty leaf node.
2244 * We have to recurse successfully before we can delete the internal
2245 * node as it is illegal to have empty internal nodes. Even though
2246 * the operation may be aborted we must still fixup any unlocked
2247 * cursors as if we had deleted the element prior to recursing
2248 * (by calling hammer_cursor_deleted_element()) so those cursors
2249 * are properly forced up the chain by the recursion.
2251 if (parent
->ondisk
->count
== 1) {
2253 * This special cursor_up_locked() call leaves the original
2254 * node exclusively locked and referenced, leaves the
2255 * original parent locked (as the new node), and locks the
2256 * new parent. It can return EDEADLK.
2258 * We cannot call hammer_cursor_removed_node() until we are
2259 * actually able to remove the node. If we did then tracked
2260 * cursors in the middle of iterations could be repointed
2261 * to a parent node. If this occurs they could end up
2262 * scanning newly inserted records into the node (that could
2263 * not be deleted) when they push down again.
2265 * Due to the way the recursion works the final parent is left
2266 * in cursor->parent after the recursion returns. Each
2267 * layer on the way back up is thus able to call
2268 * hammer_cursor_removed_node() and 'jump' the node up to
2269 * the (same) final parent.
2271 * NOTE! The local variable 'parent' is invalid after we
2272 * call hammer_cursor_up_locked().
2274 error
= hammer_cursor_up_locked(cursor
);
2278 hammer_cursor_deleted_element(cursor
->node
, 0);
2279 error
= btree_remove(cursor
);
2281 KKASSERT(node
!= cursor
->node
);
2282 hammer_cursor_removed_node(
2285 hammer_modify_node_all(cursor
->trans
, node
);
2286 ondisk
= node
->ondisk
;
2287 ondisk
->type
= HAMMER_BTREE_TYPE_DELETED
;
2289 hammer_modify_node_done(node
);
2290 hammer_flush_node(node
);
2291 hammer_delete_node(cursor
->trans
, node
);
2294 * Defer parent removal because we could not
2295 * get the lock, just let the leaf remain
2300 hammer_unlock(&node
->lock
);
2301 hammer_rel_node(node
);
2304 * Defer parent removal because we could not
2305 * get the lock, just let the leaf remain
2311 KKASSERT(parent
->ondisk
->count
> 1);
2313 hammer_modify_node_all(cursor
->trans
, parent
);
2314 ondisk
= parent
->ondisk
;
2315 KKASSERT(ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
);
2317 elm
= &ondisk
->elms
[cursor
->parent_index
];
2318 KKASSERT(elm
->internal
.subtree_offset
== node
->node_offset
);
2319 KKASSERT(ondisk
->count
> 0);
2322 * We must retain the highest mirror_tid. The deleted
2323 * range is now encompassed by the element to the left.
2324 * If we are already at the left edge the new left edge
2325 * inherits mirror_tid.
2327 * Note that bounds of the parent to our parent may create
2328 * a gap to the left of our left-most node or to the right
2329 * of our right-most node. The gap is silently included
2330 * in the mirror_tid's area of effect from the point of view
2333 if (cursor
->parent_index
) {
2334 if (elm
[-1].internal
.mirror_tid
<
2335 elm
[0].internal
.mirror_tid
) {
2336 elm
[-1].internal
.mirror_tid
=
2337 elm
[0].internal
.mirror_tid
;
2340 if (elm
[1].internal
.mirror_tid
<
2341 elm
[0].internal
.mirror_tid
) {
2342 elm
[1].internal
.mirror_tid
=
2343 elm
[0].internal
.mirror_tid
;
2348 * Delete the subtree reference in the parent. Include
2349 * boundary element at end.
2351 bcopy(&elm
[1], &elm
[0],
2352 (ondisk
->count
- cursor
->parent_index
) * esize
);
2354 hammer_modify_node_done(parent
);
2355 hammer_cursor_removed_node(node
, parent
, cursor
->parent_index
);
2356 hammer_cursor_deleted_element(parent
, cursor
->parent_index
);
2357 hammer_flush_node(node
);
2358 hammer_delete_node(cursor
->trans
, node
);
2361 * cursor->node is invalid, cursor up to make the cursor
2364 error
= hammer_cursor_up(cursor
);
2370 * Propagate cursor->trans->tid up the B-Tree starting at the current
2371 * cursor position using pseudofs info gleaned from the passed inode.
2373 * The passed inode has no relationship to the cursor position other
2374 * then being in the same pseudofs as the insertion or deletion we
2375 * are propagating the mirror_tid for.
2377 * WARNING! Because we push and pop the passed cursor, it may be
2378 * modified by other B-Tree operations while it is unlocked
2379 * and things like the node & leaf pointers, and indexes might
2383 hammer_btree_do_propagation(hammer_cursor_t cursor
,
2384 hammer_pseudofs_inmem_t pfsm
,
2385 hammer_btree_leaf_elm_t leaf
)
2387 hammer_cursor_t ncursor
;
2388 hammer_tid_t mirror_tid
;
2392 * We do not propagate a mirror_tid if the filesystem was mounted
2393 * in no-mirror mode.
2395 if (cursor
->trans
->hmp
->master_id
< 0)
2399 * This is a bit of a hack because we cannot deadlock or return
2400 * EDEADLK here. The related operation has already completed and
2401 * we must propagate the mirror_tid now regardless.
2403 * Generate a new cursor which inherits the original's locks and
2404 * unlock the original. Use the new cursor to propagate the
2405 * mirror_tid. Then clean up the new cursor and reacquire locks
2408 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2409 * original's locks and the original is tracked and must be
2412 mirror_tid
= cursor
->node
->ondisk
->mirror_tid
;
2413 KKASSERT(mirror_tid
!= 0);
2414 ncursor
= hammer_push_cursor(cursor
);
2415 error
= hammer_btree_mirror_propagate(ncursor
, mirror_tid
);
2416 KKASSERT(error
== 0);
2417 hammer_pop_cursor(cursor
, ncursor
);
2418 /* WARNING: cursor's leaf pointer may change after pop */
2423 * Propagate a mirror TID update upwards through the B-Tree to the root.
2425 * A locked internal node must be passed in. The node will remain locked
2428 * This function syncs mirror_tid at the specified internal node's element,
2429 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2432 hammer_btree_mirror_propagate(hammer_cursor_t cursor
, hammer_tid_t mirror_tid
)
2434 hammer_btree_internal_elm_t elm
;
2439 error
= hammer_cursor_up(cursor
);
2441 error
= hammer_cursor_upgrade(cursor
);
2442 while (error
== EDEADLK
) {
2443 hammer_recover_cursor(cursor
);
2444 error
= hammer_cursor_upgrade(cursor
);
2450 * If the cursor deadlocked it could end up at a leaf
2451 * after we lost the lock.
2453 node
= cursor
->node
;
2454 if (node
->ondisk
->type
!= HAMMER_BTREE_TYPE_INTERNAL
)
2458 * Adjust the node's element
2460 elm
= &node
->ondisk
->elms
[cursor
->index
].internal
;
2461 if (elm
->mirror_tid
>= mirror_tid
)
2463 hammer_modify_node(cursor
->trans
, node
, &elm
->mirror_tid
,
2464 sizeof(elm
->mirror_tid
));
2465 elm
->mirror_tid
= mirror_tid
;
2466 hammer_modify_node_done(node
);
2467 if (hammer_debug_general
& 0x0002) {
2468 kprintf("mirror_propagate: propagate "
2469 "%016llx @%016llx:%d\n",
2470 (long long)mirror_tid
,
2471 (long long)node
->node_offset
,
2477 * Adjust the node's mirror_tid aggregator
2479 if (node
->ondisk
->mirror_tid
>= mirror_tid
)
2481 hammer_modify_node_field(cursor
->trans
, node
, mirror_tid
);
2482 node
->ondisk
->mirror_tid
= mirror_tid
;
2483 hammer_modify_node_done(node
);
2484 if (hammer_debug_general
& 0x0002) {
2485 kprintf("mirror_propagate: propagate "
2486 "%016llx @%016llx\n",
2487 (long long)mirror_tid
,
2488 (long long)node
->node_offset
);
2491 if (error
== ENOENT
)
2497 hammer_btree_get_parent(hammer_transaction_t trans
, hammer_node_t node
,
2498 int *parent_indexp
, int *errorp
, int try_exclusive
)
2500 hammer_node_t parent
;
2501 hammer_btree_elm_t elm
;
2507 parent
= hammer_get_node(trans
, node
->ondisk
->parent
, 0, errorp
);
2509 KKASSERT(parent
== NULL
);
2512 KKASSERT ((parent
->flags
& HAMMER_NODE_DELETED
) == 0);
2517 if (try_exclusive
) {
2518 if (hammer_lock_ex_try(&parent
->lock
)) {
2519 hammer_rel_node(parent
);
2524 hammer_lock_sh(&parent
->lock
);
2528 * Figure out which element in the parent is pointing to the
2531 if (node
->ondisk
->count
) {
2532 i
= hammer_btree_search_node(&node
->ondisk
->elms
[0].base
,
2537 while (i
< parent
->ondisk
->count
) {
2538 elm
= &parent
->ondisk
->elms
[i
];
2539 if (elm
->internal
.subtree_offset
== node
->node_offset
)
2543 if (i
== parent
->ondisk
->count
) {
2544 hammer_unlock(&parent
->lock
);
2545 panic("Bad B-Tree link: parent %p node %p\n", parent
, node
);
2548 KKASSERT(*errorp
== 0);
2553 * The element (elm) has been moved to a new internal node (node).
2555 * If the element represents a pointer to an internal node that node's
2556 * parent must be adjusted to the element's new location.
2558 * XXX deadlock potential here with our exclusive locks
2561 btree_set_parent(hammer_transaction_t trans
, hammer_node_t node
,
2562 hammer_btree_elm_t elm
)
2564 hammer_node_t child
;
2569 switch(elm
->base
.btype
) {
2570 case HAMMER_BTREE_TYPE_INTERNAL
:
2571 case HAMMER_BTREE_TYPE_LEAF
:
2572 child
= hammer_get_node(trans
, elm
->internal
.subtree_offset
,
2575 hammer_modify_node_field(trans
, child
, parent
);
2576 child
->ondisk
->parent
= node
->node_offset
;
2577 hammer_modify_node_done(child
);
2578 hammer_rel_node(child
);
2588 * Initialize the root of a recursive B-Tree node lock list structure.
2591 hammer_node_lock_init(hammer_node_lock_t parent
, hammer_node_t node
)
2593 TAILQ_INIT(&parent
->list
);
2594 parent
->parent
= NULL
;
2595 parent
->node
= node
;
2597 parent
->count
= node
->ondisk
->count
;
2598 parent
->copy
= NULL
;
2603 * Exclusively lock all the children of node. This is used by the split
2604 * code to prevent anyone from accessing the children of a cursor node
2605 * while we fix-up its parent offset.
2607 * If we don't lock the children we can really mess up cursors which block
2608 * trying to cursor-up into our node.
2610 * On failure EDEADLK (or some other error) is returned. If a deadlock
2611 * error is returned the cursor is adjusted to block on termination.
2613 * The caller is responsible for managing parent->node, the root's node
2614 * is usually aliased from a cursor.
2617 hammer_btree_lock_children(hammer_cursor_t cursor
, int depth
,
2618 hammer_node_lock_t parent
)
2621 hammer_node_lock_t item
;
2622 hammer_node_ondisk_t ondisk
;
2623 hammer_btree_elm_t elm
;
2624 hammer_node_t child
;
2625 struct hammer_mount
*hmp
;
2629 node
= parent
->node
;
2630 ondisk
= node
->ondisk
;
2632 hmp
= cursor
->trans
->hmp
;
2635 * We really do not want to block on I/O with exclusive locks held,
2636 * pre-get the children before trying to lock the mess. This is
2637 * only done one-level deep for now.
2639 for (i
= 0; i
< ondisk
->count
; ++i
) {
2640 ++hammer_stats_btree_elements
;
2641 elm
= &ondisk
->elms
[i
];
2642 if (elm
->base
.btype
!= HAMMER_BTREE_TYPE_LEAF
&&
2643 elm
->base
.btype
!= HAMMER_BTREE_TYPE_INTERNAL
) {
2646 child
= hammer_get_node(cursor
->trans
,
2647 elm
->internal
.subtree_offset
,
2650 hammer_rel_node(child
);
2656 for (i
= 0; error
== 0 && i
< ondisk
->count
; ++i
) {
2657 ++hammer_stats_btree_elements
;
2658 elm
= &ondisk
->elms
[i
];
2660 switch(elm
->base
.btype
) {
2661 case HAMMER_BTREE_TYPE_INTERNAL
:
2662 case HAMMER_BTREE_TYPE_LEAF
:
2663 KKASSERT(elm
->internal
.subtree_offset
!= 0);
2664 child
= hammer_get_node(cursor
->trans
,
2665 elm
->internal
.subtree_offset
,
2673 if (hammer_lock_ex_try(&child
->lock
) != 0) {
2674 if (cursor
->deadlk_node
== NULL
) {
2675 cursor
->deadlk_node
= child
;
2676 hammer_ref_node(cursor
->deadlk_node
);
2679 hammer_rel_node(child
);
2681 item
= kmalloc(sizeof(*item
), hmp
->m_misc
,
2683 TAILQ_INSERT_TAIL(&parent
->list
, item
, entry
);
2684 TAILQ_INIT(&item
->list
);
2685 item
->parent
= parent
;
2688 item
->count
= child
->ondisk
->count
;
2691 * Recurse (used by the rebalancing code)
2693 if (depth
> 1 && elm
->base
.btype
== HAMMER_BTREE_TYPE_INTERNAL
) {
2694 error
= hammer_btree_lock_children(
2703 hammer_btree_unlock_children(cursor
, parent
);
2708 * Create an in-memory copy of all B-Tree nodes listed, recursively,
2709 * including the parent.
2712 hammer_btree_lock_copy(hammer_cursor_t cursor
, hammer_node_lock_t parent
)
2714 hammer_mount_t hmp
= cursor
->trans
->hmp
;
2715 hammer_node_lock_t item
;
2717 if (parent
->copy
== NULL
) {
2718 parent
->copy
= kmalloc(sizeof(*parent
->copy
), hmp
->m_misc
,
2720 *parent
->copy
= *parent
->node
->ondisk
;
2722 TAILQ_FOREACH(item
, &parent
->list
, entry
) {
2723 hammer_btree_lock_copy(cursor
, item
);
2728 * Recursively sync modified copies to the media.
2731 hammer_btree_sync_copy(hammer_cursor_t cursor
, hammer_node_lock_t parent
)
2733 hammer_node_lock_t item
;
2736 if (parent
->flags
& HAMMER_NODE_LOCK_UPDATED
) {
2738 hammer_modify_node_all(cursor
->trans
, parent
->node
);
2739 *parent
->node
->ondisk
= *parent
->copy
;
2740 hammer_modify_node_done(parent
->node
);
2741 if (parent
->copy
->type
== HAMMER_BTREE_TYPE_DELETED
) {
2742 hammer_flush_node(parent
->node
);
2743 hammer_delete_node(cursor
->trans
, parent
->node
);
2746 TAILQ_FOREACH(item
, &parent
->list
, entry
) {
2747 count
+= hammer_btree_sync_copy(cursor
, item
);
2753 * Release previously obtained node locks. The caller is responsible for
2754 * cleaning up parent->node itself (its usually just aliased from a cursor),
2755 * but this function will take care of the copies.
2758 hammer_btree_unlock_children(hammer_cursor_t cursor
, hammer_node_lock_t parent
)
2760 hammer_node_lock_t item
;
2763 kfree(parent
->copy
, cursor
->trans
->hmp
->m_misc
);
2764 parent
->copy
= NULL
; /* safety */
2766 while ((item
= TAILQ_FIRST(&parent
->list
)) != NULL
) {
2767 TAILQ_REMOVE(&parent
->list
, item
, entry
);
2768 hammer_btree_unlock_children(cursor
, item
);
2769 hammer_unlock(&item
->node
->lock
);
2770 hammer_rel_node(item
->node
);
2771 kfree(item
, cursor
->trans
->hmp
->m_misc
);
2775 /************************************************************************
2776 * MISCELLANIOUS SUPPORT *
2777 ************************************************************************/
2780 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2782 * Note that for this particular function a return value of -1, 0, or +1
2783 * can denote a match if create_tid is otherwise discounted. A create_tid
2784 * of zero is considered to be 'infinity' in comparisons.
2786 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2789 hammer_btree_cmp(hammer_base_elm_t key1
, hammer_base_elm_t key2
)
2791 if (key1
->localization
< key2
->localization
)
2793 if (key1
->localization
> key2
->localization
)
2796 if (key1
->obj_id
< key2
->obj_id
)
2798 if (key1
->obj_id
> key2
->obj_id
)
2801 if (key1
->rec_type
< key2
->rec_type
)
2803 if (key1
->rec_type
> key2
->rec_type
)
2806 if (key1
->key
< key2
->key
)
2808 if (key1
->key
> key2
->key
)
2812 * A create_tid of zero indicates a record which is undeletable
2813 * and must be considered to have a value of positive infinity.
2815 if (key1
->create_tid
== 0) {
2816 if (key2
->create_tid
== 0)
2820 if (key2
->create_tid
== 0)
2822 if (key1
->create_tid
< key2
->create_tid
)
2824 if (key1
->create_tid
> key2
->create_tid
)
2830 * Test a timestamp against an element to determine whether the
2831 * element is visible. A timestamp of 0 means 'infinity'.
2834 hammer_btree_chkts(hammer_tid_t asof
, hammer_base_elm_t base
)
2837 if (base
->delete_tid
)
2841 if (asof
< base
->create_tid
)
2843 if (base
->delete_tid
&& asof
>= base
->delete_tid
)
2849 * Create a separator half way inbetween key1 and key2. For fields just
2850 * one unit apart, the separator will match key2. key1 is on the left-hand
2851 * side and key2 is on the right-hand side.
2853 * key2 must be >= the separator. It is ok for the separator to match key2.
2855 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2858 * NOTE: It might be beneficial to just scrap this whole mess and just
2859 * set the separator to key2.
2861 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2862 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2865 hammer_make_separator(hammer_base_elm_t key1
, hammer_base_elm_t key2
,
2866 hammer_base_elm_t dest
)
2868 bzero(dest
, sizeof(*dest
));
2870 dest
->rec_type
= key2
->rec_type
;
2871 dest
->key
= key2
->key
;
2872 dest
->obj_id
= key2
->obj_id
;
2873 dest
->create_tid
= key2
->create_tid
;
2875 MAKE_SEPARATOR(key1
, key2
, dest
, localization
);
2876 if (key1
->localization
== key2
->localization
) {
2877 MAKE_SEPARATOR(key1
, key2
, dest
, obj_id
);
2878 if (key1
->obj_id
== key2
->obj_id
) {
2879 MAKE_SEPARATOR(key1
, key2
, dest
, rec_type
);
2880 if (key1
->rec_type
== key2
->rec_type
) {
2881 MAKE_SEPARATOR(key1
, key2
, dest
, key
);
2883 * Don't bother creating a separator for
2884 * create_tid, which also conveniently avoids
2885 * having to handle the create_tid == 0
2886 * (infinity) case. Just leave create_tid
2889 * Worst case, dest matches key2 exactly,
2890 * which is acceptable.
2897 #undef MAKE_SEPARATOR
2900 * Return whether a generic internal or leaf node is full
2903 btree_node_is_full(hammer_node_ondisk_t node
)
2905 switch(node
->type
) {
2906 case HAMMER_BTREE_TYPE_INTERNAL
:
2907 if (node
->count
== HAMMER_BTREE_INT_ELMS
)
2910 case HAMMER_BTREE_TYPE_LEAF
:
2911 if (node
->count
== HAMMER_BTREE_LEAF_ELMS
)
2915 panic("illegal btree subtype");
2922 btree_max_elements(u_int8_t type
)
2924 if (type
== HAMMER_BTREE_TYPE_LEAF
)
2925 return(HAMMER_BTREE_LEAF_ELMS
);
2926 if (type
== HAMMER_BTREE_TYPE_INTERNAL
)
2927 return(HAMMER_BTREE_INT_ELMS
);
2928 panic("btree_max_elements: bad type %d\n", type
);
2933 hammer_print_btree_node(hammer_node_ondisk_t ondisk
)
2935 hammer_btree_elm_t elm
;
2938 kprintf("node %p count=%d parent=%016llx type=%c\n",
2939 ondisk
, ondisk
->count
,
2940 (long long)ondisk
->parent
, ondisk
->type
);
2943 * Dump both boundary elements if an internal node
2945 if (ondisk
->type
== HAMMER_BTREE_TYPE_INTERNAL
) {
2946 for (i
= 0; i
<= ondisk
->count
; ++i
) {
2947 elm
= &ondisk
->elms
[i
];
2948 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2951 for (i
= 0; i
< ondisk
->count
; ++i
) {
2952 elm
= &ondisk
->elms
[i
];
2953 hammer_print_btree_elm(elm
, ondisk
->type
, i
);
2959 hammer_print_btree_elm(hammer_btree_elm_t elm
, u_int8_t type
, int i
)
2962 kprintf("\tobj_id = %016llx\n", (long long)elm
->base
.obj_id
);
2963 kprintf("\tkey = %016llx\n", (long long)elm
->base
.key
);
2964 kprintf("\tcreate_tid = %016llx\n", (long long)elm
->base
.create_tid
);
2965 kprintf("\tdelete_tid = %016llx\n", (long long)elm
->base
.delete_tid
);
2966 kprintf("\trec_type = %04x\n", elm
->base
.rec_type
);
2967 kprintf("\tobj_type = %02x\n", elm
->base
.obj_type
);
2968 kprintf("\tbtype = %02x (%c)\n",
2970 (elm
->base
.btype
? elm
->base
.btype
: '?'));
2971 kprintf("\tlocalization = %02x\n", elm
->base
.localization
);
2974 case HAMMER_BTREE_TYPE_INTERNAL
:
2975 kprintf("\tsubtree_off = %016llx\n",
2976 (long long)elm
->internal
.subtree_offset
);
2978 case HAMMER_BTREE_TYPE_RECORD
:
2979 kprintf("\tdata_offset = %016llx\n",
2980 (long long)elm
->leaf
.data_offset
);
2981 kprintf("\tdata_len = %08x\n", elm
->leaf
.data_len
);
2982 kprintf("\tdata_crc = %08x\n", elm
->leaf
.data_crc
);