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);
729 if (hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
730 kprintf("CRC DATA @ %016llx/%d FAILED\n",
731 (long long)elm->leaf.data_offset, elm->leaf.data_len);
732 if (hammer_debug_debug & 0x0001)
733 Debugger("CRC FAILED: DATA");
734 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM)
735 error = EDOM; /* less critical (mirroring) */
737 error = EIO; /* critical */
744 * Insert a leaf element into the B-Tree at the current cursor position.
745 * The cursor is positioned such that the element at and beyond the cursor
746 * are shifted to make room for the new record.
748 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
749 * flag set and that call must return ENOENT before this function can be
752 * The caller may depend on the cursor's exclusive lock after return to
753 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
755 * ENOSPC is returned if there is no room to insert a new record.
758 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
761 hammer_node_ondisk_t node;
766 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
768 ++hammer_stats_btree_inserts;
771 * Insert the element at the leaf node and update the count in the
772 * parent. It is possible for parent to be NULL, indicating that
773 * the filesystem's ROOT B-Tree node is a leaf itself, which is
774 * possible. The root inode can never be deleted so the leaf should
777 * Remember that the right-hand boundary is not included in the
780 hammer_modify_node_all(cursor->trans, cursor->node);
781 node = cursor->node->ondisk;
783 KKASSERT(elm->base.btype != 0);
784 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
785 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
786 if (i != node->count) {
787 bcopy(&node->elms[i], &node->elms[i+1],
788 (node->count - i) * sizeof(*elm));
790 node->elms[i].leaf = *elm;
792 hammer_cursor_inserted_element(cursor->node, i);
795 * Update the leaf node's aggregate mirror_tid for mirroring
798 if (node->mirror_tid < elm->base.delete_tid) {
799 node->mirror_tid = elm->base.delete_tid;
802 if (node->mirror_tid < elm->base.create_tid) {
803 node->mirror_tid = elm->base.create_tid;
806 hammer_modify_node_done(cursor->node);
809 * Debugging sanity checks.
811 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
812 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
814 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
816 if (i != node->count - 1)
817 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
823 * Delete a record from the B-Tree at the current cursor position.
824 * The cursor is positioned such that the current element is the one
827 * On return the cursor will be positioned after the deleted element and
828 * MAY point to an internal node. It will be suitable for the continuation
829 * of an iteration but not for an insertion or deletion.
831 * Deletions will attempt to partially rebalance the B-Tree in an upward
832 * direction, but will terminate rather then deadlock. Empty internal nodes
833 * are never allowed by a deletion which deadlocks may end up giving us an
834 * empty leaf. The pruner will clean up and rebalance the tree.
836 * This function can return EDEADLK, requiring the caller to retry the
837 * operation after clearing the deadlock.
840 hammer_btree_delete(hammer_cursor_t cursor)
842 hammer_node_ondisk_t ondisk;
844 hammer_node_t parent;
848 KKASSERT (cursor->trans->sync_lock_refs > 0);
849 if ((error = hammer_cursor_upgrade(cursor)) != 0)
851 ++hammer_stats_btree_deletes;
854 * Delete the element from the leaf node.
856 * Remember that leaf nodes do not have boundaries.
859 ondisk = node->ondisk;
862 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
863 KKASSERT(i >= 0 && i < ondisk->count);
864 hammer_modify_node_all(cursor->trans, node);
865 if (i + 1 != ondisk->count) {
866 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
867 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
870 hammer_modify_node_done(node);
871 hammer_cursor_deleted_element(node, i);
874 * Validate local parent
876 if (ondisk->parent) {
877 parent = cursor->parent;
879 KKASSERT(parent != NULL);
880 KKASSERT(parent->node_offset == ondisk->parent);
884 * If the leaf becomes empty it must be detached from the parent,
885 * potentially recursing through to the filesystem root.
887 * This may reposition the cursor at one of the parent's of the
890 * Ignore deadlock errors, that simply means that btree_remove
891 * was unable to recurse and had to leave us with an empty leaf.
893 KKASSERT(cursor->index <= ondisk->count);
894 if (ondisk->count == 0) {
895 error = btree_remove(cursor);
896 if (error == EDEADLK)
901 KKASSERT(cursor->parent == NULL ||
902 cursor->parent_index < cursor->parent->ondisk->count);
907 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
909 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
911 * The search can begin ANYWHERE in the B-Tree. As a first step the search
912 * iterates up the tree as necessary to properly position itself prior to
913 * actually doing the sarch.
915 * INSERTIONS: The search will split full nodes and leaves on its way down
916 * and guarentee that the leaf it ends up on is not full. If we run out
917 * of space the search continues to the leaf (to position the cursor for
918 * the spike), but ENOSPC is returned.
920 * The search is only guarenteed to end up on a leaf if an error code of 0
921 * is returned, or if inserting and an error code of ENOENT is returned.
922 * Otherwise it can stop at an internal node. On success a search returns
925 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
926 * filesystem, and it is not simple code. Please note the following facts:
928 * - Internal node recursions have a boundary on the left AND right. The
929 * right boundary is non-inclusive. The create_tid is a generic part
930 * of the key for internal nodes.
932 * - Leaf nodes contain terminal elements only now.
934 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
935 * historical search. ASOF and INSERT are mutually exclusive. When
936 * doing an as-of lookup btree_search() checks for a right-edge boundary
937 * case. If while recursing down the left-edge differs from the key
938 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
939 * with cursor->create_check. This is used by btree_lookup() to iterate.
940 * The iteration backwards because as-of searches can wind up going
941 * down the wrong branch of the B-Tree.
945 btree_search(hammer_cursor_t cursor, int flags)
947 hammer_node_ondisk_t node;
948 hammer_btree_elm_t elm;
955 flags |= cursor->flags;
956 ++hammer_stats_btree_searches;
958 if (hammer_debug_btree) {
959 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
960 (long long)cursor->node->node_offset,
962 (long long)cursor->key_beg.obj_id,
963 cursor->key_beg.rec_type,
964 (long long)cursor->key_beg.key,
965 (long long)cursor->key_beg.create_tid,
966 cursor->key_beg.localization,
970 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
971 (long long)cursor->parent->node_offset,
972 cursor->parent_index,
973 (long long)cursor->left_bound->obj_id,
974 (long long)cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
975 (long long)cursor->right_bound->obj_id,
976 (long long)cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
978 &cursor->parent->ondisk->elms[cursor->parent_index],
980 &cursor->parent->ondisk->elms[cursor->parent_index+1]
985 * Move our cursor up the tree until we find a node whos range covers
986 * the key we are trying to locate.
988 * The left bound is inclusive, the right bound is non-inclusive.
989 * It is ok to cursor up too far.
992 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
993 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
996 KKASSERT(cursor->parent);
997 ++hammer_stats_btree_iterations;
998 error = hammer_cursor_up(cursor);
1004 * The delete-checks below are based on node, not parent. Set the
1005 * initial delete-check based on the parent.
1008 KKASSERT(cursor->left_bound->create_tid != 1);
1009 cursor->create_check = cursor->left_bound->create_tid - 1;
1010 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1014 * We better have ended up with a node somewhere.
1016 KKASSERT(cursor->node != NULL);
1019 * If we are inserting we can't start at a full node if the parent
1020 * is also full (because there is no way to split the node),
1021 * continue running up the tree until the requirement is satisfied
1022 * or we hit the root of the filesystem.
1024 * (If inserting we aren't doing an as-of search so we don't have
1025 * to worry about create_check).
1027 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1028 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1029 if (btree_node_is_full(cursor->node->ondisk) == 0)
1032 if (btree_node_is_full(cursor->node->ondisk) ==0)
1035 if (cursor->node->ondisk->parent == 0 ||
1036 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1039 ++hammer_stats_btree_iterations;
1040 error = hammer_cursor_up(cursor);
1041 /* node may have become stale */
1047 * Push down through internal nodes to locate the requested key.
1049 node = cursor->node->ondisk;
1050 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1052 * Scan the node to find the subtree index to push down into.
1053 * We go one-past, then back-up.
1055 * We must proactively remove deleted elements which may
1056 * have been left over from a deadlocked btree_remove().
1058 * The left and right boundaries are included in the loop
1059 * in order to detect edge cases.
1061 * If the separator only differs by create_tid (r == 1)
1062 * and we are doing an as-of search, we may end up going
1063 * down a branch to the left of the one containing the
1064 * desired key. This requires numerous special cases.
1066 ++hammer_stats_btree_iterations;
1067 if (hammer_debug_btree) {
1068 kprintf("SEARCH-I %016llx count=%d\n",
1069 (long long)cursor->node->node_offset,
1074 * Try to shortcut the search before dropping into the
1075 * linear loop. Locate the first node where r <= 1.
1077 i = hammer_btree_search_node(&cursor->key_beg, node);
1078 while (i <= node->count) {
1079 ++hammer_stats_btree_elements;
1080 elm = &node->elms[i];
1081 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1082 if (hammer_debug_btree > 2) {
1083 kprintf(" IELM %p %d r=%d\n",
1084 &node->elms[i], i, r);
1089 KKASSERT(elm->base.create_tid != 1);
1090 cursor->create_check = elm->base.create_tid - 1;
1091 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1095 if (hammer_debug_btree) {
1096 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1101 * These cases occur when the parent's idea of the boundary
1102 * is wider then the child's idea of the boundary, and
1103 * require special handling. If not inserting we can
1104 * terminate the search early for these cases but the
1105 * child's boundaries cannot be unconditionally modified.
1109 * If i == 0 the search terminated to the LEFT of the
1110 * left_boundary but to the RIGHT of the parent's left
1115 elm = &node->elms[0];
1118 * If we aren't inserting we can stop here.
1120 if ((flags & (HAMMER_CURSOR_INSERT |
1121 HAMMER_CURSOR_PRUNING)) == 0) {
1127 * Correct a left-hand boundary mismatch.
1129 * We can only do this if we can upgrade the lock,
1130 * and synchronized as a background cursor (i.e.
1131 * inserting or pruning).
1133 * WARNING: We can only do this if inserting, i.e.
1134 * we are running on the backend.
1136 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1138 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1139 hammer_modify_node_field(cursor->trans, cursor->node,
1141 save = node->elms[0].base.btype;
1142 node->elms[0].base = *cursor->left_bound;
1143 node->elms[0].base.btype = save;
1144 hammer_modify_node_done(cursor->node);
1145 } else if (i == node->count + 1) {
1147 * If i == node->count + 1 the search terminated to
1148 * the RIGHT of the right boundary but to the LEFT
1149 * of the parent's right boundary. If we aren't
1150 * inserting we can stop here.
1152 * Note that the last element in this case is
1153 * elms[i-2] prior to adjustments to 'i'.
1156 if ((flags & (HAMMER_CURSOR_INSERT |
1157 HAMMER_CURSOR_PRUNING)) == 0) {
1163 * Correct a right-hand boundary mismatch.
1164 * (actual push-down record is i-2 prior to
1165 * adjustments to i).
1167 * We can only do this if we can upgrade the lock,
1168 * and synchronized as a background cursor (i.e.
1169 * inserting or pruning).
1171 * WARNING: We can only do this if inserting, i.e.
1172 * we are running on the backend.
1174 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1176 elm = &node->elms[i];
1177 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1178 hammer_modify_node(cursor->trans, cursor->node,
1179 &elm->base, sizeof(elm->base));
1180 elm->base = *cursor->right_bound;
1181 hammer_modify_node_done(cursor->node);
1185 * The push-down index is now i - 1. If we had
1186 * terminated on the right boundary this will point
1187 * us at the last element.
1192 elm = &node->elms[i];
1194 if (hammer_debug_btree) {
1195 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1196 "key=%016llx cre=%016llx lo=%02x\n",
1197 (long long)cursor->node->node_offset,
1199 (long long)elm->internal.base.obj_id,
1200 elm->internal.base.rec_type,
1201 (long long)elm->internal.base.key,
1202 (long long)elm->internal.base.create_tid,
1203 elm->internal.base.localization
1208 * We better have a valid subtree offset.
1210 KKASSERT(elm->internal.subtree_offset != 0);
1213 * Handle insertion and deletion requirements.
1215 * If inserting split full nodes. The split code will
1216 * adjust cursor->node and cursor->index if the current
1217 * index winds up in the new node.
1219 * If inserting and a left or right edge case was detected,
1220 * we cannot correct the left or right boundary and must
1221 * prepend and append an empty leaf node in order to make
1222 * the boundary correction.
1224 * If we run out of space we set enospc and continue on
1225 * to a leaf to provide the spike code with a good point
1228 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1229 if (btree_node_is_full(node)) {
1230 error = btree_split_internal(cursor);
1232 if (error != ENOSPC)
1237 * reload stale pointers
1240 node = cursor->node->ondisk;
1245 * Push down (push into new node, existing node becomes
1246 * the parent) and continue the search.
1248 error = hammer_cursor_down(cursor);
1249 /* node may have become stale */
1252 node = cursor->node->ondisk;
1256 * We are at a leaf, do a linear search of the key array.
1258 * On success the index is set to the matching element and 0
1261 * On failure the index is set to the insertion point and ENOENT
1264 * Boundaries are not stored in leaf nodes, so the index can wind
1265 * up to the left of element 0 (index == 0) or past the end of
1266 * the array (index == node->count). It is also possible that the
1267 * leaf might be empty.
1269 ++hammer_stats_btree_iterations;
1270 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1271 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1272 if (hammer_debug_btree) {
1273 kprintf("SEARCH-L %016llx count=%d\n",
1274 (long long)cursor->node->node_offset,
1279 * Try to shortcut the search before dropping into the
1280 * linear loop. Locate the first node where r <= 1.
1282 i = hammer_btree_search_node(&cursor->key_beg, node);
1283 while (i < node->count) {
1284 ++hammer_stats_btree_elements;
1285 elm = &node->elms[i];
1287 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1289 if (hammer_debug_btree > 1)
1290 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1293 * We are at a record element. Stop if we've flipped past
1294 * key_beg, not counting the create_tid test. Allow the
1295 * r == 1 case (key_beg > element but differs only by its
1296 * create_tid) to fall through to the AS-OF check.
1298 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1308 * Check our as-of timestamp against the element.
1310 if (flags & HAMMER_CURSOR_ASOF) {
1311 if (hammer_btree_chkts(cursor->asof,
1312 &node->elms[i].base) != 0) {
1318 if (r > 0) { /* can only be +1 */
1326 if (hammer_debug_btree) {
1327 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1328 (long long)cursor->node->node_offset, i);
1334 * The search of the leaf node failed. i is the insertion point.
1337 if (hammer_debug_btree) {
1338 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1339 (long long)cursor->node->node_offset, i);
1343 * No exact match was found, i is now at the insertion point.
1345 * If inserting split a full leaf before returning. This
1346 * may have the side effect of adjusting cursor->node and
1350 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1351 btree_node_is_full(node)) {
1352 error = btree_split_leaf(cursor);
1354 if (error != ENOSPC)
1359 * reload stale pointers
1363 node = &cursor->node->internal;
1368 * We reached a leaf but did not find the key we were looking for.
1369 * If this is an insert we will be properly positioned for an insert
1370 * (ENOENT) or spike (ENOSPC) operation.
1372 error = enospc ? ENOSPC : ENOENT;
1378 * Heuristical search for the first element whos comparison is <= 1. May
1379 * return an index whos compare result is > 1 but may only return an index
1380 * whos compare result is <= 1 if it is the first element with that result.
1383 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1391 * Don't bother if the node does not have very many elements
1396 i = b + (s - b) / 2;
1397 ++hammer_stats_btree_elements;
1398 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1409 /************************************************************************
1410 * SPLITTING AND MERGING *
1411 ************************************************************************
1413 * These routines do all the dirty work required to split and merge nodes.
1417 * Split an internal node into two nodes and move the separator at the split
1418 * point to the parent.
1420 * (cursor->node, cursor->index) indicates the element the caller intends
1421 * to push into. We will adjust node and index if that element winds
1422 * up in the split node.
1424 * If we are at the root of the filesystem a new root must be created with
1425 * two elements, one pointing to the original root and one pointing to the
1426 * newly allocated split node.
1430 btree_split_internal(hammer_cursor_t cursor)
1432 hammer_node_ondisk_t ondisk;
1434 hammer_node_t parent;
1435 hammer_node_t new_node;
1436 hammer_btree_elm_t elm;
1437 hammer_btree_elm_t parent_elm;
1438 struct hammer_node_lock lockroot;
1439 hammer_mount_t hmp = cursor->trans->hmp;
1446 const int esize = sizeof(*elm);
1448 hammer_node_lock_init(&lockroot, cursor->node);
1449 error = hammer_btree_lock_children(cursor, 1, &lockroot);
1452 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1454 ++hammer_stats_btree_splits;
1457 * We are splitting but elms[split] will be promoted to the parent,
1458 * leaving the right hand node with one less element. If the
1459 * insertion point will be on the left-hand side adjust the split
1460 * point to give the right hand side one additional node.
1462 node = cursor->node;
1463 ondisk = node->ondisk;
1464 split = (ondisk->count + 1) / 2;
1465 if (cursor->index <= split)
1469 * If we are at the root of the filesystem, create a new root node
1470 * with 1 element and split normally. Avoid making major
1471 * modifications until we know the whole operation will work.
1473 if (ondisk->parent == 0) {
1474 parent = hammer_alloc_btree(cursor->trans, node->node_offset,
1478 hammer_lock_ex(&parent->lock);
1479 hammer_modify_node_noundo(cursor->trans, parent);
1480 ondisk = parent->ondisk;
1483 ondisk->mirror_tid = node->ondisk->mirror_tid;
1484 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1485 ondisk->elms[0].base = hmp->root_btree_beg;
1486 ondisk->elms[0].base.btype = node->ondisk->type;
1487 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1488 ondisk->elms[1].base = hmp->root_btree_end;
1489 hammer_modify_node_done(parent);
1490 /* ondisk->elms[1].base.btype - not used */
1492 parent_index = 0; /* index of current node in parent */
1495 parent = cursor->parent;
1496 parent_index = cursor->parent_index;
1500 * Calculate a hint for the allocation of the new B-Tree node.
1501 * The most likely expansion is coming from the insertion point
1502 * at cursor->index, so try to localize the allocation of our
1503 * new node to accomodate that sub-tree.
1505 * Use the right-most sub-tree when expandinging on the right edge.
1506 * This is a very common case when copying a directory tree.
1508 if (cursor->index == ondisk->count)
1509 hint = ondisk->elms[cursor->index - 1].internal.subtree_offset;
1511 hint = ondisk->elms[cursor->index].internal.subtree_offset;
1514 * Split node into new_node at the split point.
1516 * B O O O P N N B <-- P = node->elms[split] (index 4)
1517 * 0 1 2 3 4 5 6 <-- subtree indices
1522 * B O O O B B N N B <--- inner boundary points are 'P'
1525 new_node = hammer_alloc_btree(cursor->trans, hint, &error);
1526 if (new_node == NULL) {
1528 hammer_unlock(&parent->lock);
1529 hammer_delete_node(cursor->trans, parent);
1530 hammer_rel_node(parent);
1534 hammer_lock_ex(&new_node->lock);
1537 * Create the new node. P becomes the left-hand boundary in the
1538 * new node. Copy the right-hand boundary as well.
1540 * elm is the new separator.
1542 hammer_modify_node_noundo(cursor->trans, new_node);
1543 hammer_modify_node_all(cursor->trans, node);
1544 ondisk = node->ondisk;
1545 elm = &ondisk->elms[split];
1546 bcopy(elm, &new_node->ondisk->elms[0],
1547 (ondisk->count - split + 1) * esize);
1548 new_node->ondisk->count = ondisk->count - split;
1549 new_node->ondisk->parent = parent->node_offset;
1550 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1551 new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1552 KKASSERT(ondisk->type == new_node->ondisk->type);
1553 hammer_cursor_split_node(node, new_node, split);
1556 * Cleanup the original node. Elm (P) becomes the new boundary,
1557 * its subtree_offset was moved to the new node. If we had created
1558 * a new root its parent pointer may have changed.
1560 elm->internal.subtree_offset = 0;
1561 ondisk->count = split;
1564 * Insert the separator into the parent, fixup the parent's
1565 * reference to the original node, and reference the new node.
1566 * The separator is P.
1568 * Remember that base.count does not include the right-hand boundary.
1570 hammer_modify_node_all(cursor->trans, parent);
1571 ondisk = parent->ondisk;
1572 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1573 parent_elm = &ondisk->elms[parent_index+1];
1574 bcopy(parent_elm, parent_elm + 1,
1575 (ondisk->count - parent_index) * esize);
1576 parent_elm->internal.base = elm->base; /* separator P */
1577 parent_elm->internal.base.btype = new_node->ondisk->type;
1578 parent_elm->internal.subtree_offset = new_node->node_offset;
1579 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1581 hammer_modify_node_done(parent);
1582 hammer_cursor_inserted_element(parent, parent_index + 1);
1585 * The children of new_node need their parent pointer set to new_node.
1586 * The children have already been locked by
1587 * hammer_btree_lock_children().
1589 for (i = 0; i < new_node->ondisk->count; ++i) {
1590 elm = &new_node->ondisk->elms[i];
1591 error = btree_set_parent(cursor->trans, new_node, elm);
1593 panic("btree_split_internal: btree-fixup problem");
1596 hammer_modify_node_done(new_node);
1599 * The filesystem's root B-Tree pointer may have to be updated.
1602 hammer_volume_t volume;
1604 volume = hammer_get_root_volume(hmp, &error);
1605 KKASSERT(error == 0);
1607 hammer_modify_volume_field(cursor->trans, volume,
1609 volume->ondisk->vol0_btree_root = parent->node_offset;
1610 hammer_modify_volume_done(volume);
1611 node->ondisk->parent = parent->node_offset;
1612 if (cursor->parent) {
1613 hammer_unlock(&cursor->parent->lock);
1614 hammer_rel_node(cursor->parent);
1616 cursor->parent = parent; /* lock'd and ref'd */
1617 hammer_rel_volume(volume, 0);
1619 hammer_modify_node_done(node);
1622 * Ok, now adjust the cursor depending on which element the original
1623 * index was pointing at. If we are >= the split point the push node
1624 * is now in the new node.
1626 * NOTE: If we are at the split point itself we cannot stay with the
1627 * original node because the push index will point at the right-hand
1628 * boundary, which is illegal.
1630 * NOTE: The cursor's parent or parent_index must be adjusted for
1631 * the case where a new parent (new root) was created, and the case
1632 * where the cursor is now pointing at the split node.
1634 if (cursor->index >= split) {
1635 cursor->parent_index = parent_index + 1;
1636 cursor->index -= split;
1637 hammer_unlock(&cursor->node->lock);
1638 hammer_rel_node(cursor->node);
1639 cursor->node = new_node; /* locked and ref'd */
1641 cursor->parent_index = parent_index;
1642 hammer_unlock(&new_node->lock);
1643 hammer_rel_node(new_node);
1647 * Fixup left and right bounds
1649 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1650 cursor->left_bound = &parent_elm[0].internal.base;
1651 cursor->right_bound = &parent_elm[1].internal.base;
1652 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1653 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1654 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1655 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1658 hammer_btree_unlock_children(cursor, &lockroot);
1659 hammer_cursor_downgrade(cursor);
1664 * Same as the above, but splits a full leaf node.
1670 btree_split_leaf(hammer_cursor_t cursor)
1672 hammer_node_ondisk_t ondisk;
1673 hammer_node_t parent;
1676 hammer_node_t new_leaf;
1677 hammer_btree_elm_t elm;
1678 hammer_btree_elm_t parent_elm;
1679 hammer_base_elm_t mid_boundary;
1685 const size_t esize = sizeof(*elm);
1687 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1689 ++hammer_stats_btree_splits;
1691 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1692 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1693 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1694 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1697 * Calculate the split point. If the insertion point will be on
1698 * the left-hand side adjust the split point to give the right
1699 * hand side one additional node.
1701 * Spikes are made up of two leaf elements which cannot be
1704 leaf = cursor->node;
1705 ondisk = leaf->ondisk;
1706 KKASSERT(ondisk->count > 2);
1707 split = (ondisk->count + 1) / 2;
1708 if (cursor->index <= split)
1713 elm = &ondisk->elms[split];
1715 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1716 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1717 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1718 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1721 * If we are at the root of the tree, create a new root node with
1722 * 1 element and split normally. Avoid making major modifications
1723 * until we know the whole operation will work.
1725 if (ondisk->parent == 0) {
1726 parent = hammer_alloc_btree(cursor->trans, leaf->node_offset,
1730 hammer_lock_ex(&parent->lock);
1731 hammer_modify_node_noundo(cursor->trans, parent);
1732 ondisk = parent->ondisk;
1735 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1736 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1737 ondisk->elms[0].base = hmp->root_btree_beg;
1738 ondisk->elms[0].base.btype = leaf->ondisk->type;
1739 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1740 ondisk->elms[1].base = hmp->root_btree_end;
1741 /* ondisk->elms[1].base.btype = not used */
1742 hammer_modify_node_done(parent);
1744 parent_index = 0; /* insertion point in parent */
1747 parent = cursor->parent;
1748 parent_index = cursor->parent_index;
1752 * Calculate a hint for the allocation of the new B-Tree leaf node.
1753 * For now just try to localize it within the same bigblock as
1756 * If the insertion point is at the end of the leaf we recognize a
1757 * likely append sequence of some sort (data, meta-data, inodes,
1758 * whatever). Set the hint to zero to allocate out of linear space
1759 * instead of trying to completely fill previously hinted space.
1761 * This also sets the stage for recursive splits to localize using
1764 ondisk = leaf->ondisk;
1765 if (cursor->index == ondisk->count)
1768 hint = leaf->node_offset;
1771 * Split leaf into new_leaf at the split point. Select a separator
1772 * value in-between the two leafs but with a bent towards the right
1773 * leaf since comparisons use an 'elm >= separator' inequality.
1782 new_leaf = hammer_alloc_btree(cursor->trans, hint, &error);
1783 if (new_leaf == NULL) {
1785 hammer_unlock(&parent->lock);
1786 hammer_delete_node(cursor->trans, parent);
1787 hammer_rel_node(parent);
1791 hammer_lock_ex(&new_leaf->lock);
1794 * Create the new node and copy the leaf elements from the split
1795 * point on to the new node.
1797 hammer_modify_node_all(cursor->trans, leaf);
1798 hammer_modify_node_noundo(cursor->trans, new_leaf);
1799 ondisk = leaf->ondisk;
1800 elm = &ondisk->elms[split];
1801 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1802 new_leaf->ondisk->count = ondisk->count - split;
1803 new_leaf->ondisk->parent = parent->node_offset;
1804 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1805 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1806 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1807 hammer_modify_node_done(new_leaf);
1808 hammer_cursor_split_node(leaf, new_leaf, split);
1811 * Cleanup the original node. Because this is a leaf node and
1812 * leaf nodes do not have a right-hand boundary, there
1813 * aren't any special edge cases to clean up. We just fixup the
1816 ondisk->count = split;
1819 * Insert the separator into the parent, fixup the parent's
1820 * reference to the original node, and reference the new node.
1821 * The separator is P.
1823 * Remember that base.count does not include the right-hand boundary.
1824 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1826 hammer_modify_node_all(cursor->trans, parent);
1827 ondisk = parent->ondisk;
1828 KKASSERT(split != 0);
1829 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1830 parent_elm = &ondisk->elms[parent_index+1];
1831 bcopy(parent_elm, parent_elm + 1,
1832 (ondisk->count - parent_index) * esize);
1834 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1835 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1836 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1837 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1838 mid_boundary = &parent_elm->base;
1840 hammer_modify_node_done(parent);
1841 hammer_cursor_inserted_element(parent, parent_index + 1);
1844 * The filesystem's root B-Tree pointer may have to be updated.
1847 hammer_volume_t volume;
1849 volume = hammer_get_root_volume(hmp, &error);
1850 KKASSERT(error == 0);
1852 hammer_modify_volume_field(cursor->trans, volume,
1854 volume->ondisk->vol0_btree_root = parent->node_offset;
1855 hammer_modify_volume_done(volume);
1856 leaf->ondisk->parent = parent->node_offset;
1857 if (cursor->parent) {
1858 hammer_unlock(&cursor->parent->lock);
1859 hammer_rel_node(cursor->parent);
1861 cursor->parent = parent; /* lock'd and ref'd */
1862 hammer_rel_volume(volume, 0);
1864 hammer_modify_node_done(leaf);
1867 * Ok, now adjust the cursor depending on which element the original
1868 * index was pointing at. If we are >= the split point the push node
1869 * is now in the new node.
1871 * NOTE: If we are at the split point itself we need to select the
1872 * old or new node based on where key_beg's insertion point will be.
1873 * If we pick the wrong side the inserted element will wind up in
1874 * the wrong leaf node and outside that node's bounds.
1876 if (cursor->index > split ||
1877 (cursor->index == split &&
1878 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1879 cursor->parent_index = parent_index + 1;
1880 cursor->index -= split;
1881 hammer_unlock(&cursor->node->lock);
1882 hammer_rel_node(cursor->node);
1883 cursor->node = new_leaf;
1885 cursor->parent_index = parent_index;
1886 hammer_unlock(&new_leaf->lock);
1887 hammer_rel_node(new_leaf);
1891 * Fixup left and right bounds
1893 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1894 cursor->left_bound = &parent_elm[0].internal.base;
1895 cursor->right_bound = &parent_elm[1].internal.base;
1898 * Assert that the bounds are correct.
1900 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1901 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1902 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1903 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1904 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1905 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1908 hammer_cursor_downgrade(cursor);
1915 * Recursively correct the right-hand boundary's create_tid to (tid) as
1916 * long as the rest of the key matches. We have to recurse upward in
1917 * the tree as well as down the left side of each parent's right node.
1919 * Return EDEADLK if we were only partially successful, forcing the caller
1920 * to try again. The original cursor is not modified. This routine can
1921 * also fail with EDEADLK if it is forced to throw away a portion of its
1924 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1927 TAILQ_ENTRY(hammer_rhb) entry;
1932 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1935 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1937 struct hammer_mount *hmp;
1938 struct hammer_rhb_list rhb_list;
1939 hammer_base_elm_t elm;
1940 hammer_node_t orig_node;
1941 struct hammer_rhb *rhb;
1945 TAILQ_INIT(&rhb_list);
1946 hmp = cursor->trans->hmp;
1949 * Save our position so we can restore it on return. This also
1950 * gives us a stable 'elm'.
1952 orig_node = cursor->node;
1953 hammer_ref_node(orig_node);
1954 hammer_lock_sh(&orig_node->lock);
1955 orig_index = cursor->index;
1956 elm = &orig_node->ondisk->elms[orig_index].base;
1959 * Now build a list of parents going up, allocating a rhb
1960 * structure for each one.
1962 while (cursor->parent) {
1964 * Stop if we no longer have any right-bounds to fix up
1966 if (elm->obj_id != cursor->right_bound->obj_id ||
1967 elm->rec_type != cursor->right_bound->rec_type ||
1968 elm->key != cursor->right_bound->key) {
1973 * Stop if the right-hand bound's create_tid does not
1974 * need to be corrected.
1976 if (cursor->right_bound->create_tid >= tid)
1979 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
1980 rhb->node = cursor->parent;
1981 rhb->index = cursor->parent_index;
1982 hammer_ref_node(rhb->node);
1983 hammer_lock_sh(&rhb->node->lock);
1984 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1986 hammer_cursor_up(cursor);
1990 * now safely adjust the right hand bound for each rhb. This may
1991 * also require taking the right side of the tree and iterating down
1995 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1996 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1999 TAILQ_REMOVE(&rhb_list, rhb, entry);
2000 hammer_unlock(&rhb->node->lock);
2001 hammer_rel_node(rhb->node);
2002 kfree(rhb, hmp->m_misc);
2004 switch (cursor->node->ondisk->type) {
2005 case HAMMER_BTREE_TYPE_INTERNAL:
2007 * Right-boundary for parent at internal node
2008 * is one element to the right of the element whos
2009 * right boundary needs adjusting. We must then
2010 * traverse down the left side correcting any left
2011 * bounds (which may now be too far to the left).
2014 error = hammer_btree_correct_lhb(cursor, tid);
2017 panic("hammer_btree_correct_rhb(): Bad node type");
2026 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2027 TAILQ_REMOVE(&rhb_list, rhb, entry);
2028 hammer_unlock(&rhb->node->lock);
2029 hammer_rel_node(rhb->node);
2030 kfree(rhb, hmp->m_misc);
2032 error = hammer_cursor_seek(cursor, orig_node, orig_index);
2033 hammer_unlock(&orig_node->lock);
2034 hammer_rel_node(orig_node);
2039 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2040 * bound going downward starting at the current cursor position.
2042 * This function does not restore the cursor after use.
2045 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
2047 struct hammer_rhb_list rhb_list;
2048 hammer_base_elm_t elm;
2049 hammer_base_elm_t cmp;
2050 struct hammer_rhb *rhb;
2051 struct hammer_mount *hmp;
2054 TAILQ_INIT(&rhb_list);
2055 hmp = cursor->trans->hmp;
2057 cmp = &cursor->node->ondisk->elms[cursor->index].base;
2060 * Record the node and traverse down the left-hand side for all
2061 * matching records needing a boundary correction.
2065 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2066 rhb->node = cursor->node;
2067 rhb->index = cursor->index;
2068 hammer_ref_node(rhb->node);
2069 hammer_lock_sh(&rhb->node->lock);
2070 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2072 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2074 * Nothing to traverse down if we are at the right
2075 * boundary of an internal node.
2077 if (cursor->index == cursor->node->ondisk->count)
2080 elm = &cursor->node->ondisk->elms[cursor->index].base;
2081 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2083 panic("Illegal leaf record type %02x", elm->btype);
2085 error = hammer_cursor_down(cursor);
2089 elm = &cursor->node->ondisk->elms[cursor->index].base;
2090 if (elm->obj_id != cmp->obj_id ||
2091 elm->rec_type != cmp->rec_type ||
2092 elm->key != cmp->key) {
2095 if (elm->create_tid >= tid)
2101 * Now we can safely adjust the left-hand boundary from the bottom-up.
2102 * The last element we remove from the list is the caller's right hand
2103 * boundary, which must also be adjusted.
2105 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2106 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2109 TAILQ_REMOVE(&rhb_list, rhb, entry);
2110 hammer_unlock(&rhb->node->lock);
2111 hammer_rel_node(rhb->node);
2112 kfree(rhb, hmp->m_misc);
2114 elm = &cursor->node->ondisk->elms[cursor->index].base;
2115 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2116 hammer_modify_node(cursor->trans, cursor->node,
2118 sizeof(elm->create_tid));
2119 elm->create_tid = tid;
2120 hammer_modify_node_done(cursor->node);
2122 panic("hammer_btree_correct_lhb(): Bad element type");
2129 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2130 TAILQ_REMOVE(&rhb_list, rhb, entry);
2131 hammer_unlock(&rhb->node->lock);
2132 hammer_rel_node(rhb->node);
2133 kfree(rhb, hmp->m_misc);
2141 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2142 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2143 * the operation due to a deadlock, or some other error.
2145 * This routine is initially called with an empty leaf and may be
2146 * recursively called with single-element internal nodes.
2148 * It should also be noted that when removing empty leaves we must be sure
2149 * to test and update mirror_tid because another thread may have deadlocked
2150 * against us (or someone) trying to propagate it up and cannot retry once
2151 * the node has been deleted.
2153 * On return the cursor may end up pointing to an internal node, suitable
2154 * for further iteration but not for an immediate insertion or deletion.
2157 btree_remove(hammer_cursor_t cursor)
2159 hammer_node_ondisk_t ondisk;
2160 hammer_btree_elm_t elm;
2162 hammer_node_t parent;
2163 const int esize = sizeof(*elm);
2166 node = cursor->node;
2169 * When deleting the root of the filesystem convert it to
2170 * an empty leaf node. Internal nodes cannot be empty.
2172 ondisk = node->ondisk;
2173 if (ondisk->parent == 0) {
2174 KKASSERT(cursor->parent == NULL);
2175 hammer_modify_node_all(cursor->trans, node);
2176 KKASSERT(ondisk == node->ondisk);
2177 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2179 hammer_modify_node_done(node);
2184 parent = cursor->parent;
2185 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2188 * Attempt to remove the parent's reference to the child. If the
2189 * parent would become empty we have to recurse. If we fail we
2190 * leave the parent pointing to an empty leaf node.
2192 * We have to recurse successfully before we can delete the internal
2193 * node as it is illegal to have empty internal nodes. Even though
2194 * the operation may be aborted we must still fixup any unlocked
2195 * cursors as if we had deleted the element prior to recursing
2196 * (by calling hammer_cursor_deleted_element()) so those cursors
2197 * are properly forced up the chain by the recursion.
2199 if (parent->ondisk->count == 1) {
2201 * This special cursor_up_locked() call leaves the original
2202 * node exclusively locked and referenced, leaves the
2203 * original parent locked (as the new node), and locks the
2204 * new parent. It can return EDEADLK.
2206 error = hammer_cursor_up_locked(cursor);
2208 hammer_cursor_deleted_element(cursor->node, 0);
2209 error = btree_remove(cursor);
2211 hammer_modify_node_all(cursor->trans, node);
2212 ondisk = node->ondisk;
2213 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2215 hammer_modify_node_done(node);
2216 hammer_flush_node(node);
2217 hammer_delete_node(cursor->trans, node);
2220 * Defer parent removal because we could not
2221 * get the lock, just let the leaf remain
2226 hammer_unlock(&node->lock);
2227 hammer_rel_node(node);
2230 * Defer parent removal because we could not
2231 * get the lock, just let the leaf remain
2237 KKASSERT(parent->ondisk->count > 1);
2239 hammer_modify_node_all(cursor->trans, parent);
2240 ondisk = parent->ondisk;
2241 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2243 elm = &ondisk->elms[cursor->parent_index];
2244 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2245 KKASSERT(ondisk->count > 0);
2248 * We must retain the highest mirror_tid. The deleted
2249 * range is now encompassed by the element to the left.
2250 * If we are already at the left edge the new left edge
2251 * inherits mirror_tid.
2253 * Note that bounds of the parent to our parent may create
2254 * a gap to the left of our left-most node or to the right
2255 * of our right-most node. The gap is silently included
2256 * in the mirror_tid's area of effect from the point of view
2259 if (cursor->parent_index) {
2260 if (elm[-1].internal.mirror_tid <
2261 elm[0].internal.mirror_tid) {
2262 elm[-1].internal.mirror_tid =
2263 elm[0].internal.mirror_tid;
2266 if (elm[1].internal.mirror_tid <
2267 elm[0].internal.mirror_tid) {
2268 elm[1].internal.mirror_tid =
2269 elm[0].internal.mirror_tid;
2274 * Delete the subtree reference in the parent
2276 bcopy(&elm[1], &elm[0],
2277 (ondisk->count - cursor->parent_index) * esize);
2279 hammer_modify_node_done(parent);
2280 hammer_cursor_deleted_element(parent, cursor->parent_index);
2281 hammer_flush_node(node);
2282 hammer_delete_node(cursor->trans, node);
2285 * cursor->node is invalid, cursor up to make the cursor
2288 error = hammer_cursor_up(cursor);
2294 * Propagate cursor->trans->tid up the B-Tree starting at the current
2295 * cursor position using pseudofs info gleaned from the passed inode.
2297 * The passed inode has no relationship to the cursor position other
2298 * then being in the same pseudofs as the insertion or deletion we
2299 * are propagating the mirror_tid for.
2301 * WARNING! Because we push and pop the passed cursor, it may be
2302 * modified by other B-Tree operations while it is unlocked
2303 * and things like the node & leaf pointers, and indexes might
2307 hammer_btree_do_propagation(hammer_cursor_t cursor,
2308 hammer_pseudofs_inmem_t pfsm,
2309 hammer_btree_leaf_elm_t leaf)
2311 hammer_cursor_t ncursor;
2312 hammer_tid_t mirror_tid;
2316 * We do not propagate a mirror_tid if the filesystem was mounted
2317 * in no-mirror mode.
2319 if (cursor->trans->hmp->master_id < 0)
2323 * This is a bit of a hack because we cannot deadlock or return
2324 * EDEADLK here. The related operation has already completed and
2325 * we must propagate the mirror_tid now regardless.
2327 * Generate a new cursor which inherits the original's locks and
2328 * unlock the original. Use the new cursor to propagate the
2329 * mirror_tid. Then clean up the new cursor and reacquire locks
2332 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2333 * original's locks and the original is tracked and must be
2336 mirror_tid = cursor->node->ondisk->mirror_tid;
2337 KKASSERT(mirror_tid != 0);
2338 ncursor = hammer_push_cursor(cursor);
2339 error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2340 KKASSERT(error == 0);
2341 hammer_pop_cursor(cursor, ncursor);
2342 /* WARNING: cursor's leaf pointer may change after pop */
2347 * Propagate a mirror TID update upwards through the B-Tree to the root.
2349 * A locked internal node must be passed in. The node will remain locked
2352 * This function syncs mirror_tid at the specified internal node's element,
2353 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2356 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2358 hammer_btree_internal_elm_t elm;
2363 error = hammer_cursor_up(cursor);
2365 error = hammer_cursor_upgrade(cursor);
2366 while (error == EDEADLK) {
2367 hammer_recover_cursor(cursor);
2368 error = hammer_cursor_upgrade(cursor);
2372 node = cursor->node;
2373 KKASSERT (node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2376 * Adjust the node's element
2378 elm = &node->ondisk->elms[cursor->index].internal;
2379 if (elm->mirror_tid >= mirror_tid)
2381 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2382 sizeof(elm->mirror_tid));
2383 elm->mirror_tid = mirror_tid;
2384 hammer_modify_node_done(node);
2385 if (hammer_debug_general & 0x0002) {
2386 kprintf("mirror_propagate: propagate "
2387 "%016llx @%016llx:%d\n",
2388 (long long)mirror_tid,
2389 (long long)node->node_offset,
2395 * Adjust the node's mirror_tid aggregator
2397 if (node->ondisk->mirror_tid >= mirror_tid)
2399 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2400 node->ondisk->mirror_tid = mirror_tid;
2401 hammer_modify_node_done(node);
2402 if (hammer_debug_general & 0x0002) {
2403 kprintf("mirror_propagate: propagate "
2404 "%016llx @%016llx\n",
2405 (long long)mirror_tid,
2406 (long long)node->node_offset);
2409 if (error == ENOENT)
2415 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2416 int *parent_indexp, int *errorp, int try_exclusive)
2418 hammer_node_t parent;
2419 hammer_btree_elm_t elm;
2425 parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2427 KKASSERT(parent == NULL);
2430 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2435 if (try_exclusive) {
2436 if (hammer_lock_ex_try(&parent->lock)) {
2437 hammer_rel_node(parent);
2442 hammer_lock_sh(&parent->lock);
2446 * Figure out which element in the parent is pointing to the
2449 if (node->ondisk->count) {
2450 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2455 while (i < parent->ondisk->count) {
2456 elm = &parent->ondisk->elms[i];
2457 if (elm->internal.subtree_offset == node->node_offset)
2461 if (i == parent->ondisk->count) {
2462 hammer_unlock(&parent->lock);
2463 panic("Bad B-Tree link: parent %p node %p\n", parent, node);
2466 KKASSERT(*errorp == 0);
2471 * The element (elm) has been moved to a new internal node (node).
2473 * If the element represents a pointer to an internal node that node's
2474 * parent must be adjusted to the element's new location.
2476 * XXX deadlock potential here with our exclusive locks
2479 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2480 hammer_btree_elm_t elm)
2482 hammer_node_t child;
2487 switch(elm->base.btype) {
2488 case HAMMER_BTREE_TYPE_INTERNAL:
2489 case HAMMER_BTREE_TYPE_LEAF:
2490 child = hammer_get_node(trans, elm->internal.subtree_offset,
2493 hammer_modify_node_field(trans, child, parent);
2494 child->ondisk->parent = node->node_offset;
2495 hammer_modify_node_done(child);
2496 hammer_rel_node(child);
2506 * Initialize the root of a recursive B-Tree node lock list structure.
2509 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2511 TAILQ_INIT(&parent->list);
2512 parent->parent = NULL;
2513 parent->node = node;
2515 parent->count = node->ondisk->count;
2516 parent->copy = NULL;
2521 * Exclusively lock all the children of node. This is used by the split
2522 * code to prevent anyone from accessing the children of a cursor node
2523 * while we fix-up its parent offset.
2525 * If we don't lock the children we can really mess up cursors which block
2526 * trying to cursor-up into our node.
2528 * On failure EDEADLK (or some other error) is returned. If a deadlock
2529 * error is returned the cursor is adjusted to block on termination.
2531 * The caller is responsible for managing parent->node, the root's node
2532 * is usually aliased from a cursor.
2535 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2536 hammer_node_lock_t parent)
2539 hammer_node_lock_t item;
2540 hammer_node_ondisk_t ondisk;
2541 hammer_btree_elm_t elm;
2542 hammer_node_t child;
2543 struct hammer_mount *hmp;
2547 node = parent->node;
2548 ondisk = node->ondisk;
2550 hmp = cursor->trans->hmp;
2553 * We really do not want to block on I/O with exclusive locks held,
2554 * pre-get the children before trying to lock the mess. This is
2555 * only done one-level deep for now.
2557 for (i = 0; i < ondisk->count; ++i) {
2558 ++hammer_stats_btree_elements;
2559 elm = &ondisk->elms[i];
2560 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2561 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2564 child = hammer_get_node(cursor->trans,
2565 elm->internal.subtree_offset,
2568 hammer_rel_node(child);
2574 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2575 ++hammer_stats_btree_elements;
2576 elm = &ondisk->elms[i];
2578 switch(elm->base.btype) {
2579 case HAMMER_BTREE_TYPE_INTERNAL:
2580 case HAMMER_BTREE_TYPE_LEAF:
2581 KKASSERT(elm->internal.subtree_offset != 0);
2582 child = hammer_get_node(cursor->trans,
2583 elm->internal.subtree_offset,
2591 if (hammer_lock_ex_try(&child->lock) != 0) {
2592 if (cursor->deadlk_node == NULL) {
2593 cursor->deadlk_node = child;
2594 hammer_ref_node(cursor->deadlk_node);
2597 hammer_rel_node(child);
2599 item = kmalloc(sizeof(*item), hmp->m_misc,
2601 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2602 TAILQ_INIT(&item->list);
2603 item->parent = parent;
2606 item->count = child->ondisk->count;
2609 * Recurse (used by the rebalancing code)
2611 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2612 error = hammer_btree_lock_children(
2621 hammer_btree_unlock_children(cursor, parent);
2626 * Create an in-memory copy of all B-Tree nodes listed, recursively,
2627 * including the parent.
2630 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2632 hammer_mount_t hmp = cursor->trans->hmp;
2633 hammer_node_lock_t item;
2635 if (parent->copy == NULL) {
2636 parent->copy = kmalloc(sizeof(*parent->copy), hmp->m_misc,
2638 *parent->copy = *parent->node->ondisk;
2640 TAILQ_FOREACH(item, &parent->list, entry) {
2641 hammer_btree_lock_copy(cursor, item);
2646 * Recursively sync modified copies to the media.
2649 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2651 hammer_node_lock_t item;
2654 if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2656 hammer_modify_node_all(cursor->trans, parent->node);
2657 *parent->node->ondisk = *parent->copy;
2658 hammer_modify_node_done(parent->node);
2659 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2660 hammer_flush_node(parent->node);
2661 hammer_delete_node(cursor->trans, parent->node);
2664 TAILQ_FOREACH(item, &parent->list, entry) {
2665 count += hammer_btree_sync_copy(cursor, item);
2671 * Release previously obtained node locks. The caller is responsible for
2672 * cleaning up parent->node itself (its usually just aliased from a cursor),
2673 * but this function will take care of the copies.
2676 hammer_btree_unlock_children(hammer_cursor_t cursor, hammer_node_lock_t parent)
2678 hammer_node_lock_t item;
2681 kfree(parent->copy, cursor->trans->hmp->m_misc);
2682 parent->copy = NULL; /* safety */
2684 while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2685 TAILQ_REMOVE(&parent->list, item, entry);
2686 hammer_btree_unlock_children(cursor, item);
2687 hammer_unlock(&item->node->lock);
2688 hammer_rel_node(item->node);
2689 kfree(item, cursor->trans->hmp->m_misc);
2693 /************************************************************************
2694 * MISCELLANIOUS SUPPORT *
2695 ************************************************************************/
2698 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2700 * Note that for this particular function a return value of -1, 0, or +1
2701 * can denote a match if create_tid is otherwise discounted. A create_tid
2702 * of zero is considered to be 'infinity' in comparisons.
2704 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2707 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2709 if (key1->localization < key2->localization)
2711 if (key1->localization > key2->localization)
2714 if (key1->obj_id < key2->obj_id)
2716 if (key1->obj_id > key2->obj_id)
2719 if (key1->rec_type < key2->rec_type)
2721 if (key1->rec_type > key2->rec_type)
2724 if (key1->key < key2->key)
2726 if (key1->key > key2->key)
2730 * A create_tid of zero indicates a record which is undeletable
2731 * and must be considered to have a value of positive infinity.
2733 if (key1->create_tid == 0) {
2734 if (key2->create_tid == 0)
2738 if (key2->create_tid == 0)
2740 if (key1->create_tid < key2->create_tid)
2742 if (key1->create_tid > key2->create_tid)
2748 * Test a timestamp against an element to determine whether the
2749 * element is visible. A timestamp of 0 means 'infinity'.
2752 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2755 if (base->delete_tid)
2759 if (asof < base->create_tid)
2761 if (base->delete_tid && asof >= base->delete_tid)
2767 * Create a separator half way inbetween key1 and key2. For fields just
2768 * one unit apart, the separator will match key2. key1 is on the left-hand
2769 * side and key2 is on the right-hand side.
2771 * key2 must be >= the separator. It is ok for the separator to match key2.
2773 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2776 * NOTE: It might be beneficial to just scrap this whole mess and just
2777 * set the separator to key2.
2779 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2780 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2783 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2784 hammer_base_elm_t dest)
2786 bzero(dest, sizeof(*dest));
2788 dest->rec_type = key2->rec_type;
2789 dest->key = key2->key;
2790 dest->obj_id = key2->obj_id;
2791 dest->create_tid = key2->create_tid;
2793 MAKE_SEPARATOR(key1, key2, dest, localization);
2794 if (key1->localization == key2->localization) {
2795 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2796 if (key1->obj_id == key2->obj_id) {
2797 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2798 if (key1->rec_type == key2->rec_type) {
2799 MAKE_SEPARATOR(key1, key2, dest, key);
2801 * Don't bother creating a separator for
2802 * create_tid, which also conveniently avoids
2803 * having to handle the create_tid == 0
2804 * (infinity) case. Just leave create_tid
2807 * Worst case, dest matches key2 exactly,
2808 * which is acceptable.
2815 #undef MAKE_SEPARATOR
2818 * Return whether a generic internal or leaf node is full
2821 btree_node_is_full(hammer_node_ondisk_t node)
2823 switch(node->type) {
2824 case HAMMER_BTREE_TYPE_INTERNAL:
2825 if (node->count == HAMMER_BTREE_INT_ELMS)
2828 case HAMMER_BTREE_TYPE_LEAF:
2829 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2833 panic("illegal btree subtype");
2840 btree_max_elements(u_int8_t type)
2842 if (type == HAMMER_BTREE_TYPE_LEAF)
2843 return(HAMMER_BTREE_LEAF_ELMS);
2844 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2845 return(HAMMER_BTREE_INT_ELMS);
2846 panic("btree_max_elements: bad type %d\n", type);
2851 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2853 hammer_btree_elm_t elm;
2856 kprintf("node %p count=%d parent=%016llx type=%c\n",
2857 ondisk, ondisk->count,
2858 (long long)ondisk->parent, ondisk->type);
2861 * Dump both boundary elements if an internal node
2863 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2864 for (i = 0; i <= ondisk->count; ++i) {
2865 elm = &ondisk->elms[i];
2866 hammer_print_btree_elm(elm, ondisk->type, i);
2869 for (i = 0; i < ondisk->count; ++i) {
2870 elm = &ondisk->elms[i];
2871 hammer_print_btree_elm(elm, ondisk->type, i);
2877 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2880 kprintf("\tobj_id = %016llx\n", (long long)elm->base.obj_id);
2881 kprintf("\tkey = %016llx\n", (long long)elm->base.key);
2882 kprintf("\tcreate_tid = %016llx\n", (long long)elm->base.create_tid);
2883 kprintf("\tdelete_tid = %016llx\n", (long long)elm->base.delete_tid);
2884 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2885 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2886 kprintf("\tbtype = %02x (%c)\n",
2888 (elm->base.btype ? elm->base.btype : '?'));
2889 kprintf("\tlocalization = %02x\n", elm->base.localization);
2892 case HAMMER_BTREE_TYPE_INTERNAL:
2893 kprintf("\tsubtree_off = %016llx\n",
2894 (long long)elm->internal.subtree_offset);
2896 case HAMMER_BTREE_TYPE_RECORD:
2897 kprintf("\tdata_offset = %016llx\n",
2898 (long long)elm->leaf.data_offset);
2899 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2900 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);