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
38 * HAMMER implements a modified B+Tree. In documentation this will
39 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
40 * looks like a B+Tree (A B-Tree which stores its records only at the leafs
41 * of the tree), but adds two additional boundary elements which describe
42 * the left-most and right-most element a node is able to represent. In
43 * otherwords, we have boundary elements at the two ends of a B-Tree node
44 * instead of sub-tree pointers.
46 * A B-Tree internal node looks like this:
48 * B N N N N N N B <-- boundary and internal elements
49 * S S S S S S S <-- subtree pointers
51 * A B-Tree leaf node basically looks like this:
53 * L L L L L L L L <-- leaf elemenets
55 * The radix for an internal node is 1 less then a leaf but we get a
56 * number of significant benefits for our troubles.
58 * The big benefit to using a B-Tree containing boundary information
59 * is that it is possible to cache pointers into the middle of the tree
60 * and not have to start searches, insertions, OR deletions at the root
61 * node. In particular, searches are able to progress in a definitive
62 * direction from any point in the tree without revisting nodes. This
63 * greatly improves the efficiency of many operations, most especially
66 * B-Trees also make the stacking of trees fairly straightforward.
68 * INSERTIONS: A search performed with the intention of doing
69 * an insert will guarantee that the terminal leaf node is not full by
70 * splitting full nodes. Splits occur top-down during the dive down the
73 * DELETIONS: A deletion makes no attempt to proactively balance the
74 * tree and will recursively remove nodes that become empty. If a
75 * deadlock occurs a deletion may not be able to remove an empty leaf.
76 * Deletions never allow internal nodes to become empty (that would blow
83 static int btree_search(hammer_cursor_t cursor, int flags);
84 static int btree_split_internal(hammer_cursor_t cursor);
85 static int btree_split_leaf(hammer_cursor_t cursor);
86 static int btree_remove(hammer_cursor_t cursor);
87 static int btree_node_is_full(hammer_node_ondisk_t node);
88 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor,
89 hammer_tid_t mirror_tid);
90 static void hammer_make_separator(hammer_base_elm_t key1,
91 hammer_base_elm_t key2, hammer_base_elm_t dest);
92 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor);
95 * Iterate records after a search. The cursor is iterated forwards past
96 * the current record until a record matching the key-range requirements
97 * is found. ENOENT is returned if the iteration goes past the ending
100 * The iteration is inclusive of key_beg and can be inclusive or exclusive
101 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
103 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
104 * may be modified by B-Tree functions.
106 * cursor->key_beg may or may not be modified by this function during
107 * the iteration. XXX future - in case of an inverted lock we may have
108 * to reinitiate the lookup and set key_beg to properly pick up where we
111 * If HAMMER_CURSOR_ITERATE_CHECK is set it is possible that the cursor
112 * was reverse indexed due to being moved to a parent while unlocked,
113 * and something else might have inserted an element outside the iteration
114 * range. When this case occurs the iterator just keeps iterating until
115 * it gets back into the iteration range (instead of asserting).
117 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
120 hammer_btree_iterate(hammer_cursor_t cursor)
122 hammer_node_ondisk_t node;
123 hammer_btree_elm_t elm;
130 * Skip past the current record
132 hmp = cursor->trans->hmp;
133 node = cursor->node->ondisk;
136 if (cursor->index < node->count &&
137 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
142 * HAMMER can wind up being cpu-bound.
144 if (++hmp->check_yield > hammer_yield_check) {
145 hmp->check_yield = 0;
151 * Loop until an element is found or we are done.
155 * We iterate up the tree and then index over one element
156 * while we are at the last element in the current node.
158 * If we are at the root of the filesystem, cursor_up
161 * XXX this could be optimized by storing the information in
162 * the parent reference.
164 * XXX we can lose the node lock temporarily, this could mess
167 ++hammer_stats_btree_iterations;
168 hammer_flusher_clean_loose_ios(hmp);
170 if (cursor->index == node->count) {
171 if (hammer_debug_btree) {
172 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
173 (long long)cursor->node->node_offset,
175 (long long)(cursor->parent ? cursor->parent->node_offset : -1),
176 cursor->parent_index,
179 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
180 error = hammer_cursor_up(cursor);
183 /* reload stale pointer */
184 node = cursor->node->ondisk;
185 KKASSERT(cursor->index != node->count);
188 * If we are reblocking we want to return internal
189 * nodes. Note that the internal node will be
190 * returned multiple times, on each upward recursion
191 * from its children. The caller selects which
192 * revisit it cares about (usually first or last only).
194 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
195 cursor->flags |= HAMMER_CURSOR_ATEDISK;
203 * Check internal or leaf element. Determine if the record
204 * at the cursor has gone beyond the end of our range.
206 * We recurse down through internal nodes.
208 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
209 elm = &node->elms[cursor->index];
211 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
212 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
213 if (hammer_debug_btree) {
214 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
215 (long long)cursor->node->node_offset,
217 (long long)elm[0].internal.base.obj_id,
218 elm[0].internal.base.rec_type,
219 (long long)elm[0].internal.base.key,
220 elm[0].internal.base.localization,
224 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
225 (long long)cursor->node->node_offset,
227 (long long)elm[1].internal.base.obj_id,
228 elm[1].internal.base.rec_type,
229 (long long)elm[1].internal.base.key,
230 elm[1].internal.base.localization,
239 if (r == 0 && (cursor->flags &
240 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
248 KKASSERT(elm->internal.subtree_offset != 0);
252 * If running the mirror filter see if we
253 * can skip one or more entire sub-trees.
254 * If we can we return the internal node
255 * and the caller processes the skipped
256 * range (see mirror_read).
259 HAMMER_CURSOR_MIRROR_FILTERED) {
260 if (elm->internal.mirror_tid <
261 cursor->cmirror->mirror_tid) {
262 hammer_cursor_mirror_filter(cursor);
268 * Normally it would be impossible for the
269 * cursor to have gotten back-indexed,
270 * but it can happen if a node is deleted
271 * and the cursor is moved to its parent
272 * internal node. ITERATE_CHECK will be set.
274 KKASSERT(cursor->flags &
275 HAMMER_CURSOR_ITERATE_CHECK);
276 kprintf("hammer_btree_iterate: "
277 "DEBUG: Caught parent seek "
278 "in internal iteration\n");
281 error = hammer_cursor_down(cursor);
284 KKASSERT(cursor->index == 0);
285 /* reload stale pointer */
286 node = cursor->node->ondisk;
289 elm = &node->elms[cursor->index];
290 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
291 if (hammer_debug_btree) {
292 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
293 (long long)cursor->node->node_offset,
295 (elm[0].leaf.base.btype ?
296 elm[0].leaf.base.btype : '?'),
297 (long long)elm[0].leaf.base.obj_id,
298 elm[0].leaf.base.rec_type,
299 (long long)elm[0].leaf.base.key,
300 elm[0].leaf.base.localization,
310 * We support both end-inclusive and
311 * end-exclusive searches.
314 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
320 * If ITERATE_CHECK is set an unlocked cursor may
321 * have been moved to a parent and the iterate can
322 * happen upon elements that are not in the requested
325 if (cursor->flags & HAMMER_CURSOR_ITERATE_CHECK) {
326 s = hammer_btree_cmp(&cursor->key_beg,
329 kprintf("hammer_btree_iterate: "
330 "DEBUG: Caught parent seek "
331 "in leaf iteration\n");
336 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
341 switch(elm->leaf.base.btype) {
342 case HAMMER_BTREE_TYPE_RECORD:
343 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
344 hammer_btree_chkts(cursor->asof, &elm->base)) {
358 * node pointer invalid after loop
364 if (hammer_debug_btree) {
365 int i = cursor->index;
366 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
367 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
369 (long long)elm->internal.base.obj_id,
370 elm->internal.base.rec_type,
371 (long long)elm->internal.base.key,
372 elm->internal.base.localization
381 * We hit an internal element that we could skip as part of a mirroring
382 * scan. Calculate the entire range being skipped.
384 * It is important to include any gaps between the parent's left_bound
385 * and the node's left_bound, and same goes for the right side.
388 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
390 struct hammer_cmirror *cmirror;
391 hammer_node_ondisk_t ondisk;
392 hammer_btree_elm_t elm;
394 ondisk = cursor->node->ondisk;
395 cmirror = cursor->cmirror;
398 * Calculate the skipped range
400 elm = &ondisk->elms[cursor->index];
401 if (cursor->index == 0)
402 cmirror->skip_beg = *cursor->left_bound;
404 cmirror->skip_beg = elm->internal.base;
405 while (cursor->index < ondisk->count) {
406 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
411 if (cursor->index == ondisk->count)
412 cmirror->skip_end = *cursor->right_bound;
414 cmirror->skip_end = elm->internal.base;
417 * clip the returned result.
419 if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
420 cmirror->skip_beg = cursor->key_beg;
421 if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
422 cmirror->skip_end = cursor->key_end;
426 * Iterate in the reverse direction. This is used by the pruning code to
427 * avoid overlapping records.
430 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
432 hammer_node_ondisk_t node;
433 hammer_btree_elm_t elm;
439 /* mirror filtering not supported for reverse iteration */
440 KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
443 * Skip past the current record. For various reasons the cursor
444 * may end up set to -1 or set to point at the end of the current
445 * node. These cases must be addressed.
447 node = cursor->node->ondisk;
450 if (cursor->index != -1 &&
451 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
454 if (cursor->index == cursor->node->ondisk->count)
458 * HAMMER can wind up being cpu-bound.
460 hmp = cursor->trans->hmp;
461 if (++hmp->check_yield > hammer_yield_check) {
462 hmp->check_yield = 0;
467 * Loop until an element is found or we are done.
470 ++hammer_stats_btree_iterations;
471 hammer_flusher_clean_loose_ios(hmp);
474 * We iterate up the tree and then index over one element
475 * while we are at the last element in the current node.
477 if (cursor->index == -1) {
478 error = hammer_cursor_up(cursor);
480 cursor->index = 0; /* sanity */
483 /* reload stale pointer */
484 node = cursor->node->ondisk;
485 KKASSERT(cursor->index != node->count);
491 * Check internal or leaf element. Determine if the record
492 * at the cursor has gone beyond the end of our range.
494 * We recurse down through internal nodes.
496 KKASSERT(cursor->index != node->count);
497 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
498 elm = &node->elms[cursor->index];
499 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
500 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
501 if (hammer_debug_btree) {
502 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
503 (long long)cursor->node->node_offset,
505 (long long)elm[0].internal.base.obj_id,
506 elm[0].internal.base.rec_type,
507 (long long)elm[0].internal.base.key,
508 elm[0].internal.base.localization,
511 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
512 (long long)cursor->node->node_offset,
514 (long long)elm[1].internal.base.obj_id,
515 elm[1].internal.base.rec_type,
516 (long long)elm[1].internal.base.key,
517 elm[1].internal.base.localization,
528 * It shouldn't be possible to be seeked past key_end,
529 * even if the cursor got moved to a parent.
536 KKASSERT(elm->internal.subtree_offset != 0);
538 error = hammer_cursor_down(cursor);
541 KKASSERT(cursor->index == 0);
542 /* reload stale pointer */
543 node = cursor->node->ondisk;
545 /* this can assign -1 if the leaf was empty */
546 cursor->index = node->count - 1;
549 elm = &node->elms[cursor->index];
550 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
551 if (hammer_debug_btree) {
552 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
553 (long long)cursor->node->node_offset,
555 (elm[0].leaf.base.btype ?
556 elm[0].leaf.base.btype : '?'),
557 (long long)elm[0].leaf.base.obj_id,
558 elm[0].leaf.base.rec_type,
559 (long long)elm[0].leaf.base.key,
560 elm[0].leaf.base.localization,
570 * It shouldn't be possible to be seeked past key_end,
571 * even if the cursor got moved to a parent.
573 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
578 switch(elm->leaf.base.btype) {
579 case HAMMER_BTREE_TYPE_RECORD:
580 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
581 hammer_btree_chkts(cursor->asof, &elm->base)) {
595 * node pointer invalid after loop
601 if (hammer_debug_btree) {
602 int i = cursor->index;
603 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
604 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
606 (long long)elm->internal.base.obj_id,
607 elm->internal.base.rec_type,
608 (long long)elm->internal.base.key,
609 elm->internal.base.localization
618 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
619 * could not be found, EDEADLK if inserting and a retry is needed, and a
620 * fatal error otherwise. When retrying, the caller must terminate the
621 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
623 * The cursor is suitably positioned for a deletion on success, and suitably
624 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
627 * The cursor may begin anywhere, the search will traverse the tree in
628 * either direction to locate the requested element.
630 * Most of the logic implementing historical searches is handled here. We
631 * do an initial lookup with create_tid set to the asof TID. Due to the
632 * way records are laid out, a backwards iteration may be required if
633 * ENOENT is returned to locate the historical record. Here's the
636 * create_tid: 10 15 20
640 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
641 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
642 * not visible and thus causes ENOENT to be returned. We really need
643 * to check record 11 in LEAF1. If it also fails then the search fails
644 * (e.g. it might represent the range 11-16 and thus still not match our
645 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
646 * further iterations.
648 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
649 * and the cursor->create_check TID if an iteration might be needed.
650 * In the above example create_check would be set to 14.
653 hammer_btree_lookup(hammer_cursor_t cursor)
657 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
658 KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
659 cursor->trans->sync_lock_refs > 0);
660 ++hammer_stats_btree_lookups;
661 if (cursor->flags & HAMMER_CURSOR_ASOF) {
662 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
663 cursor->key_beg.create_tid = cursor->asof;
665 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
666 error = btree_search(cursor, 0);
667 if (error != ENOENT ||
668 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
671 * Stop if error other then ENOENT.
672 * Stop if ENOENT and not special case.
676 if (hammer_debug_btree) {
677 kprintf("CREATE_CHECK %016llx\n",
678 (long long)cursor->create_check);
680 cursor->key_beg.create_tid = cursor->create_check;
684 error = btree_search(cursor, 0);
687 error = hammer_btree_extract(cursor, cursor->flags);
692 * Execute the logic required to start an iteration. The first record
693 * located within the specified range is returned and iteration control
694 * flags are adjusted for successive hammer_btree_iterate() calls.
696 * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
697 * in a loop without worrying about it. Higher-level merged searches will
698 * adjust the flag appropriately.
701 hammer_btree_first(hammer_cursor_t cursor)
705 error = hammer_btree_lookup(cursor);
706 if (error == ENOENT) {
707 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
708 error = hammer_btree_iterate(cursor);
710 cursor->flags |= HAMMER_CURSOR_ATEDISK;
715 * Similarly but for an iteration in the reverse direction.
717 * Set ATEDISK when iterating backwards to skip the current entry,
718 * which after an ENOENT lookup will be pointing beyond our end point.
720 * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
721 * in a loop without worrying about it. Higher-level merged searches will
722 * adjust the flag appropriately.
725 hammer_btree_last(hammer_cursor_t cursor)
727 struct hammer_base_elm save;
730 save = cursor->key_beg;
731 cursor->key_beg = cursor->key_end;
732 error = hammer_btree_lookup(cursor);
733 cursor->key_beg = save;
734 if (error == ENOENT ||
735 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
736 cursor->flags |= HAMMER_CURSOR_ATEDISK;
737 error = hammer_btree_iterate_reverse(cursor);
739 cursor->flags |= HAMMER_CURSOR_ATEDISK;
744 * Extract the record and/or data associated with the cursor's current
745 * position. Any prior record or data stored in the cursor is replaced.
746 * The cursor must be positioned at a leaf node.
748 * NOTE: All extractions occur at the leaf of the B-Tree.
751 hammer_btree_extract(hammer_cursor_t cursor, int flags)
753 hammer_node_ondisk_t node;
754 hammer_btree_elm_t elm;
755 hammer_off_t data_off;
761 * The case where the data reference resolves to the same buffer
762 * as the record reference must be handled.
764 node = cursor->node->ondisk;
765 elm = &node->elms[cursor->index];
767 hmp = cursor->node->hmp;
770 * There is nothing to extract for an internal element.
772 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
776 * Only record types have data.
778 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
779 cursor->leaf = &elm->leaf;
781 if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
783 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
785 data_off = elm->leaf.data_offset;
786 data_len = elm->leaf.data_len;
793 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
794 cursor->data = hammer_bread_ext(hmp, data_off, data_len,
795 &error, &cursor->data_buffer);
798 * Mark the data buffer as not being meta-data if it isn't
799 * meta-data (sometimes bulk data is accessed via a volume
803 switch(elm->leaf.base.rec_type) {
804 case HAMMER_RECTYPE_DATA:
805 case HAMMER_RECTYPE_DB:
806 if ((data_off & HAMMER_ZONE_LARGE_DATA) == 0)
808 if (hammer_double_buffer == 0 ||
809 (cursor->flags & HAMMER_CURSOR_NOSWAPCACHE)) {
810 hammer_io_notmeta(cursor->data_buffer);
819 * Deal with CRC errors on the extracted data.
822 hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
823 kprintf("CRC DATA @ %016llx/%d FAILED\n",
824 (long long)elm->leaf.data_offset, elm->leaf.data_len);
825 if (hammer_debug_critical)
826 Debugger("CRC FAILED: DATA");
827 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM)
828 error = EDOM; /* less critical (mirroring) */
830 error = EIO; /* critical */
837 * Insert a leaf element into the B-Tree at the current cursor position.
838 * The cursor is positioned such that the element at and beyond the cursor
839 * are shifted to make room for the new record.
841 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
842 * flag set and that call must return ENOENT before this function can be
845 * The caller may depend on the cursor's exclusive lock after return to
846 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
848 * ENOSPC is returned if there is no room to insert a new record.
851 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
854 hammer_node_ondisk_t node;
859 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
861 ++hammer_stats_btree_inserts;
864 * Insert the element at the leaf node and update the count in the
865 * parent. It is possible for parent to be NULL, indicating that
866 * the filesystem's ROOT B-Tree node is a leaf itself, which is
867 * possible. The root inode can never be deleted so the leaf should
870 * Remember that the right-hand boundary is not included in the
873 hammer_modify_node_all(cursor->trans, cursor->node);
874 node = cursor->node->ondisk;
876 KKASSERT(elm->base.btype != 0);
877 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
878 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
879 if (i != node->count) {
880 bcopy(&node->elms[i], &node->elms[i+1],
881 (node->count - i) * sizeof(*elm));
883 node->elms[i].leaf = *elm;
885 hammer_cursor_inserted_element(cursor->node, i);
888 * Update the leaf node's aggregate mirror_tid for mirroring
891 if (node->mirror_tid < elm->base.delete_tid) {
892 node->mirror_tid = elm->base.delete_tid;
895 if (node->mirror_tid < elm->base.create_tid) {
896 node->mirror_tid = elm->base.create_tid;
899 hammer_modify_node_done(cursor->node);
902 * Debugging sanity checks.
904 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
905 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
907 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
909 if (i != node->count - 1)
910 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
916 * Delete a record from the B-Tree at the current cursor position.
917 * The cursor is positioned such that the current element is the one
920 * On return the cursor will be positioned after the deleted element and
921 * MAY point to an internal node. It will be suitable for the continuation
922 * of an iteration but not for an insertion or deletion.
924 * Deletions will attempt to partially rebalance the B-Tree in an upward
925 * direction, but will terminate rather then deadlock. Empty internal nodes
926 * are never allowed by a deletion which deadlocks may end up giving us an
927 * empty leaf. The pruner will clean up and rebalance the tree.
929 * This function can return EDEADLK, requiring the caller to retry the
930 * operation after clearing the deadlock.
933 hammer_btree_delete(hammer_cursor_t cursor)
935 hammer_node_ondisk_t ondisk;
937 hammer_node_t parent;
941 KKASSERT (cursor->trans->sync_lock_refs > 0);
942 if ((error = hammer_cursor_upgrade(cursor)) != 0)
944 ++hammer_stats_btree_deletes;
947 * Delete the element from the leaf node.
949 * Remember that leaf nodes do not have boundaries.
952 ondisk = node->ondisk;
955 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
956 KKASSERT(i >= 0 && i < ondisk->count);
957 hammer_modify_node_all(cursor->trans, node);
958 if (i + 1 != ondisk->count) {
959 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
960 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
963 hammer_modify_node_done(node);
964 hammer_cursor_deleted_element(node, i);
967 * Validate local parent
969 if (ondisk->parent) {
970 parent = cursor->parent;
972 KKASSERT(parent != NULL);
973 KKASSERT(parent->node_offset == ondisk->parent);
977 * If the leaf becomes empty it must be detached from the parent,
978 * potentially recursing through to the filesystem root.
980 * This may reposition the cursor at one of the parent's of the
983 * Ignore deadlock errors, that simply means that btree_remove
984 * was unable to recurse and had to leave us with an empty leaf.
986 KKASSERT(cursor->index <= ondisk->count);
987 if (ondisk->count == 0) {
988 error = btree_remove(cursor);
989 if (error == EDEADLK)
994 KKASSERT(cursor->parent == NULL ||
995 cursor->parent_index < cursor->parent->ondisk->count);
1000 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
1002 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
1004 * The search can begin ANYWHERE in the B-Tree. As a first step the search
1005 * iterates up the tree as necessary to properly position itself prior to
1006 * actually doing the sarch.
1008 * INSERTIONS: The search will split full nodes and leaves on its way down
1009 * and guarentee that the leaf it ends up on is not full. If we run out
1010 * of space the search continues to the leaf (to position the cursor for
1011 * the spike), but ENOSPC is returned.
1013 * The search is only guarenteed to end up on a leaf if an error code of 0
1014 * is returned, or if inserting and an error code of ENOENT is returned.
1015 * Otherwise it can stop at an internal node. On success a search returns
1018 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
1019 * filesystem, and it is not simple code. Please note the following facts:
1021 * - Internal node recursions have a boundary on the left AND right. The
1022 * right boundary is non-inclusive. The create_tid is a generic part
1023 * of the key for internal nodes.
1025 * - Leaf nodes contain terminal elements only now.
1027 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
1028 * historical search. ASOF and INSERT are mutually exclusive. When
1029 * doing an as-of lookup btree_search() checks for a right-edge boundary
1030 * case. If while recursing down the left-edge differs from the key
1031 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
1032 * with cursor->create_check. This is used by btree_lookup() to iterate.
1033 * The iteration backwards because as-of searches can wind up going
1034 * down the wrong branch of the B-Tree.
1038 btree_search(hammer_cursor_t cursor, int flags)
1040 hammer_node_ondisk_t node;
1041 hammer_btree_elm_t elm;
1048 flags |= cursor->flags;
1049 ++hammer_stats_btree_searches;
1051 if (hammer_debug_btree) {
1052 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
1053 (long long)cursor->node->node_offset,
1055 (long long)cursor->key_beg.obj_id,
1056 cursor->key_beg.rec_type,
1057 (long long)cursor->key_beg.key,
1058 (long long)cursor->key_beg.create_tid,
1059 cursor->key_beg.localization,
1063 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
1064 (long long)cursor->parent->node_offset,
1065 cursor->parent_index,
1066 (long long)cursor->left_bound->obj_id,
1067 (long long)cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
1068 (long long)cursor->right_bound->obj_id,
1069 (long long)cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
1071 &cursor->parent->ondisk->elms[cursor->parent_index],
1072 cursor->right_bound,
1073 &cursor->parent->ondisk->elms[cursor->parent_index+1]
1078 * Move our cursor up the tree until we find a node whos range covers
1079 * the key we are trying to locate.
1081 * The left bound is inclusive, the right bound is non-inclusive.
1082 * It is ok to cursor up too far.
1085 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
1086 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
1087 if (r >= 0 && s < 0)
1089 KKASSERT(cursor->parent);
1090 ++hammer_stats_btree_iterations;
1091 error = hammer_cursor_up(cursor);
1097 * The delete-checks below are based on node, not parent. Set the
1098 * initial delete-check based on the parent.
1101 KKASSERT(cursor->left_bound->create_tid != 1);
1102 cursor->create_check = cursor->left_bound->create_tid - 1;
1103 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1107 * We better have ended up with a node somewhere.
1109 KKASSERT(cursor->node != NULL);
1112 * If we are inserting we can't start at a full node if the parent
1113 * is also full (because there is no way to split the node),
1114 * continue running up the tree until the requirement is satisfied
1115 * or we hit the root of the filesystem.
1117 * (If inserting we aren't doing an as-of search so we don't have
1118 * to worry about create_check).
1120 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1121 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1122 if (btree_node_is_full(cursor->node->ondisk) == 0)
1125 if (btree_node_is_full(cursor->node->ondisk) ==0)
1128 if (cursor->node->ondisk->parent == 0 ||
1129 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1132 ++hammer_stats_btree_iterations;
1133 error = hammer_cursor_up(cursor);
1134 /* node may have become stale */
1140 * Push down through internal nodes to locate the requested key.
1142 node = cursor->node->ondisk;
1143 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1145 * Scan the node to find the subtree index to push down into.
1146 * We go one-past, then back-up.
1148 * We must proactively remove deleted elements which may
1149 * have been left over from a deadlocked btree_remove().
1151 * The left and right boundaries are included in the loop
1152 * in order to detect edge cases.
1154 * If the separator only differs by create_tid (r == 1)
1155 * and we are doing an as-of search, we may end up going
1156 * down a branch to the left of the one containing the
1157 * desired key. This requires numerous special cases.
1159 ++hammer_stats_btree_iterations;
1160 if (hammer_debug_btree) {
1161 kprintf("SEARCH-I %016llx count=%d\n",
1162 (long long)cursor->node->node_offset,
1167 * Try to shortcut the search before dropping into the
1168 * linear loop. Locate the first node where r <= 1.
1170 i = hammer_btree_search_node(&cursor->key_beg, node);
1171 while (i <= node->count) {
1172 ++hammer_stats_btree_elements;
1173 elm = &node->elms[i];
1174 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1175 if (hammer_debug_btree > 2) {
1176 kprintf(" IELM %p %d r=%d\n",
1177 &node->elms[i], i, r);
1182 KKASSERT(elm->base.create_tid != 1);
1183 cursor->create_check = elm->base.create_tid - 1;
1184 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1188 if (hammer_debug_btree) {
1189 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1194 * These cases occur when the parent's idea of the boundary
1195 * is wider then the child's idea of the boundary, and
1196 * require special handling. If not inserting we can
1197 * terminate the search early for these cases but the
1198 * child's boundaries cannot be unconditionally modified.
1202 * If i == 0 the search terminated to the LEFT of the
1203 * left_boundary but to the RIGHT of the parent's left
1208 elm = &node->elms[0];
1211 * If we aren't inserting we can stop here.
1213 if ((flags & (HAMMER_CURSOR_INSERT |
1214 HAMMER_CURSOR_PRUNING)) == 0) {
1220 * Correct a left-hand boundary mismatch.
1222 * We can only do this if we can upgrade the lock,
1223 * and synchronized as a background cursor (i.e.
1224 * inserting or pruning).
1226 * WARNING: We can only do this if inserting, i.e.
1227 * we are running on the backend.
1229 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1231 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1232 hammer_modify_node_field(cursor->trans, cursor->node,
1234 save = node->elms[0].base.btype;
1235 node->elms[0].base = *cursor->left_bound;
1236 node->elms[0].base.btype = save;
1237 hammer_modify_node_done(cursor->node);
1238 } else if (i == node->count + 1) {
1240 * If i == node->count + 1 the search terminated to
1241 * the RIGHT of the right boundary but to the LEFT
1242 * of the parent's right boundary. If we aren't
1243 * inserting we can stop here.
1245 * Note that the last element in this case is
1246 * elms[i-2] prior to adjustments to 'i'.
1249 if ((flags & (HAMMER_CURSOR_INSERT |
1250 HAMMER_CURSOR_PRUNING)) == 0) {
1256 * Correct a right-hand boundary mismatch.
1257 * (actual push-down record is i-2 prior to
1258 * adjustments to i).
1260 * We can only do this if we can upgrade the lock,
1261 * and synchronized as a background cursor (i.e.
1262 * inserting or pruning).
1264 * WARNING: We can only do this if inserting, i.e.
1265 * we are running on the backend.
1267 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1269 elm = &node->elms[i];
1270 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1271 hammer_modify_node(cursor->trans, cursor->node,
1272 &elm->base, sizeof(elm->base));
1273 elm->base = *cursor->right_bound;
1274 hammer_modify_node_done(cursor->node);
1278 * The push-down index is now i - 1. If we had
1279 * terminated on the right boundary this will point
1280 * us at the last element.
1285 elm = &node->elms[i];
1287 if (hammer_debug_btree) {
1288 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1289 "key=%016llx cre=%016llx lo=%02x\n",
1290 (long long)cursor->node->node_offset,
1292 (long long)elm->internal.base.obj_id,
1293 elm->internal.base.rec_type,
1294 (long long)elm->internal.base.key,
1295 (long long)elm->internal.base.create_tid,
1296 elm->internal.base.localization
1301 * We better have a valid subtree offset.
1303 KKASSERT(elm->internal.subtree_offset != 0);
1306 * Handle insertion and deletion requirements.
1308 * If inserting split full nodes. The split code will
1309 * adjust cursor->node and cursor->index if the current
1310 * index winds up in the new node.
1312 * If inserting and a left or right edge case was detected,
1313 * we cannot correct the left or right boundary and must
1314 * prepend and append an empty leaf node in order to make
1315 * the boundary correction.
1317 * If we run out of space we set enospc and continue on
1318 * to a leaf to provide the spike code with a good point
1321 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1322 if (btree_node_is_full(node)) {
1323 error = btree_split_internal(cursor);
1325 if (error != ENOSPC)
1330 * reload stale pointers
1333 node = cursor->node->ondisk;
1338 * Push down (push into new node, existing node becomes
1339 * the parent) and continue the search.
1341 error = hammer_cursor_down(cursor);
1342 /* node may have become stale */
1345 node = cursor->node->ondisk;
1349 * We are at a leaf, do a linear search of the key array.
1351 * On success the index is set to the matching element and 0
1354 * On failure the index is set to the insertion point and ENOENT
1357 * Boundaries are not stored in leaf nodes, so the index can wind
1358 * up to the left of element 0 (index == 0) or past the end of
1359 * the array (index == node->count). It is also possible that the
1360 * leaf might be empty.
1362 ++hammer_stats_btree_iterations;
1363 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1364 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1365 if (hammer_debug_btree) {
1366 kprintf("SEARCH-L %016llx count=%d\n",
1367 (long long)cursor->node->node_offset,
1372 * Try to shortcut the search before dropping into the
1373 * linear loop. Locate the first node where r <= 1.
1375 i = hammer_btree_search_node(&cursor->key_beg, node);
1376 while (i < node->count) {
1377 ++hammer_stats_btree_elements;
1378 elm = &node->elms[i];
1380 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1382 if (hammer_debug_btree > 1)
1383 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1386 * We are at a record element. Stop if we've flipped past
1387 * key_beg, not counting the create_tid test. Allow the
1388 * r == 1 case (key_beg > element but differs only by its
1389 * create_tid) to fall through to the AS-OF check.
1391 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1401 * Check our as-of timestamp against the element.
1403 if (flags & HAMMER_CURSOR_ASOF) {
1404 if (hammer_btree_chkts(cursor->asof,
1405 &node->elms[i].base) != 0) {
1411 if (r > 0) { /* can only be +1 */
1419 if (hammer_debug_btree) {
1420 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1421 (long long)cursor->node->node_offset, i);
1427 * The search of the leaf node failed. i is the insertion point.
1430 if (hammer_debug_btree) {
1431 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1432 (long long)cursor->node->node_offset, i);
1436 * No exact match was found, i is now at the insertion point.
1438 * If inserting split a full leaf before returning. This
1439 * may have the side effect of adjusting cursor->node and
1443 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1444 btree_node_is_full(node)) {
1445 error = btree_split_leaf(cursor);
1447 if (error != ENOSPC)
1452 * reload stale pointers
1456 node = &cursor->node->internal;
1461 * We reached a leaf but did not find the key we were looking for.
1462 * If this is an insert we will be properly positioned for an insert
1463 * (ENOENT) or spike (ENOSPC) operation.
1465 error = enospc ? ENOSPC : ENOENT;
1471 * Heuristical search for the first element whos comparison is <= 1. May
1472 * return an index whos compare result is > 1 but may only return an index
1473 * whos compare result is <= 1 if it is the first element with that result.
1476 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1484 * Don't bother if the node does not have very many elements
1489 i = b + (s - b) / 2;
1490 ++hammer_stats_btree_elements;
1491 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1502 /************************************************************************
1503 * SPLITTING AND MERGING *
1504 ************************************************************************
1506 * These routines do all the dirty work required to split and merge nodes.
1510 * Split an internal node into two nodes and move the separator at the split
1511 * point to the parent.
1513 * (cursor->node, cursor->index) indicates the element the caller intends
1514 * to push into. We will adjust node and index if that element winds
1515 * up in the split node.
1517 * If we are at the root of the filesystem a new root must be created with
1518 * two elements, one pointing to the original root and one pointing to the
1519 * newly allocated split node.
1523 btree_split_internal(hammer_cursor_t cursor)
1525 hammer_node_ondisk_t ondisk;
1527 hammer_node_t parent;
1528 hammer_node_t new_node;
1529 hammer_btree_elm_t elm;
1530 hammer_btree_elm_t parent_elm;
1531 struct hammer_node_lock lockroot;
1532 hammer_mount_t hmp = cursor->trans->hmp;
1539 const int esize = sizeof(*elm);
1541 hammer_node_lock_init(&lockroot, cursor->node);
1542 error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
1545 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1547 ++hammer_stats_btree_splits;
1550 * Calculate the split point. If the insertion point is at the
1551 * end of the leaf we adjust the split point significantly to the
1552 * right to try to optimize node fill and flag it. If we hit
1553 * that same leaf again our heuristic failed and we don't try
1554 * to optimize node fill (it could lead to a degenerate case).
1556 node = cursor->node;
1557 ondisk = node->ondisk;
1558 KKASSERT(ondisk->count > 4);
1559 if (cursor->index == ondisk->count &&
1560 (node->flags & HAMMER_NODE_NONLINEAR) == 0) {
1561 split = (ondisk->count + 1) * 3 / 4;
1562 node->flags |= HAMMER_NODE_NONLINEAR;
1565 * We are splitting but elms[split] will be promoted to
1566 * the parent, leaving the right hand node with one less
1567 * element. If the insertion point will be on the
1568 * left-hand side adjust the split point to give the
1569 * right hand side one additional node.
1571 split = (ondisk->count + 1) / 2;
1572 if (cursor->index <= split)
1577 * If we are at the root of the filesystem, create a new root node
1578 * with 1 element and split normally. Avoid making major
1579 * modifications until we know the whole operation will work.
1581 if (ondisk->parent == 0) {
1582 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1585 hammer_lock_ex(&parent->lock);
1586 hammer_modify_node_noundo(cursor->trans, parent);
1587 ondisk = parent->ondisk;
1590 ondisk->mirror_tid = node->ondisk->mirror_tid;
1591 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1592 ondisk->elms[0].base = hmp->root_btree_beg;
1593 ondisk->elms[0].base.btype = node->ondisk->type;
1594 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1595 ondisk->elms[1].base = hmp->root_btree_end;
1596 hammer_modify_node_done(parent);
1597 /* ondisk->elms[1].base.btype - not used */
1599 parent_index = 0; /* index of current node in parent */
1602 parent = cursor->parent;
1603 parent_index = cursor->parent_index;
1607 * Calculate a hint for the allocation of the new B-Tree node.
1608 * The most likely expansion is coming from the insertion point
1609 * at cursor->index, so try to localize the allocation of our
1610 * new node to accomodate that sub-tree.
1612 * Use the right-most sub-tree when expandinging on the right edge.
1613 * This is a very common case when copying a directory tree.
1615 if (cursor->index == ondisk->count)
1616 hint = ondisk->elms[cursor->index - 1].internal.subtree_offset;
1618 hint = ondisk->elms[cursor->index].internal.subtree_offset;
1621 * Split node into new_node at the split point.
1623 * B O O O P N N B <-- P = node->elms[split] (index 4)
1624 * 0 1 2 3 4 5 6 <-- subtree indices
1629 * B O O O B B N N B <--- inner boundary points are 'P'
1632 new_node = hammer_alloc_btree(cursor->trans, 0, &error);
1633 if (new_node == NULL) {
1635 hammer_unlock(&parent->lock);
1636 hammer_delete_node(cursor->trans, parent);
1637 hammer_rel_node(parent);
1641 hammer_lock_ex(&new_node->lock);
1644 * Create the new node. P becomes the left-hand boundary in the
1645 * new node. Copy the right-hand boundary as well.
1647 * elm is the new separator.
1649 hammer_modify_node_noundo(cursor->trans, new_node);
1650 hammer_modify_node_all(cursor->trans, node);
1651 ondisk = node->ondisk;
1652 elm = &ondisk->elms[split];
1653 bcopy(elm, &new_node->ondisk->elms[0],
1654 (ondisk->count - split + 1) * esize);
1655 new_node->ondisk->count = ondisk->count - split;
1656 new_node->ondisk->parent = parent->node_offset;
1657 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1658 new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1659 KKASSERT(ondisk->type == new_node->ondisk->type);
1660 hammer_cursor_split_node(node, new_node, split);
1663 * Cleanup the original node. Elm (P) becomes the new boundary,
1664 * its subtree_offset was moved to the new node. If we had created
1665 * a new root its parent pointer may have changed.
1667 elm->internal.subtree_offset = 0;
1668 ondisk->count = split;
1671 * Insert the separator into the parent, fixup the parent's
1672 * reference to the original node, and reference the new node.
1673 * The separator is P.
1675 * Remember that base.count does not include the right-hand boundary.
1677 hammer_modify_node_all(cursor->trans, parent);
1678 ondisk = parent->ondisk;
1679 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1680 parent_elm = &ondisk->elms[parent_index+1];
1681 bcopy(parent_elm, parent_elm + 1,
1682 (ondisk->count - parent_index) * esize);
1683 parent_elm->internal.base = elm->base; /* separator P */
1684 parent_elm->internal.base.btype = new_node->ondisk->type;
1685 parent_elm->internal.subtree_offset = new_node->node_offset;
1686 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1688 hammer_modify_node_done(parent);
1689 hammer_cursor_inserted_element(parent, parent_index + 1);
1692 * The children of new_node need their parent pointer set to new_node.
1693 * The children have already been locked by
1694 * hammer_btree_lock_children().
1696 for (i = 0; i < new_node->ondisk->count; ++i) {
1697 elm = &new_node->ondisk->elms[i];
1698 error = btree_set_parent(cursor->trans, new_node, elm);
1700 panic("btree_split_internal: btree-fixup problem");
1703 hammer_modify_node_done(new_node);
1706 * The filesystem's root B-Tree pointer may have to be updated.
1709 hammer_volume_t volume;
1711 volume = hammer_get_root_volume(hmp, &error);
1712 KKASSERT(error == 0);
1714 hammer_modify_volume_field(cursor->trans, volume,
1716 volume->ondisk->vol0_btree_root = parent->node_offset;
1717 hammer_modify_volume_done(volume);
1718 node->ondisk->parent = parent->node_offset;
1719 if (cursor->parent) {
1720 hammer_unlock(&cursor->parent->lock);
1721 hammer_rel_node(cursor->parent);
1723 cursor->parent = parent; /* lock'd and ref'd */
1724 hammer_rel_volume(volume, 0);
1726 hammer_modify_node_done(node);
1729 * Ok, now adjust the cursor depending on which element the original
1730 * index was pointing at. If we are >= the split point the push node
1731 * is now in the new node.
1733 * NOTE: If we are at the split point itself we cannot stay with the
1734 * original node because the push index will point at the right-hand
1735 * boundary, which is illegal.
1737 * NOTE: The cursor's parent or parent_index must be adjusted for
1738 * the case where a new parent (new root) was created, and the case
1739 * where the cursor is now pointing at the split node.
1741 if (cursor->index >= split) {
1742 cursor->parent_index = parent_index + 1;
1743 cursor->index -= split;
1744 hammer_unlock(&cursor->node->lock);
1745 hammer_rel_node(cursor->node);
1746 cursor->node = new_node; /* locked and ref'd */
1748 cursor->parent_index = parent_index;
1749 hammer_unlock(&new_node->lock);
1750 hammer_rel_node(new_node);
1754 * Fixup left and right bounds
1756 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1757 cursor->left_bound = &parent_elm[0].internal.base;
1758 cursor->right_bound = &parent_elm[1].internal.base;
1759 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1760 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1761 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1762 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1765 hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
1766 hammer_cursor_downgrade(cursor);
1771 * Same as the above, but splits a full leaf node.
1777 btree_split_leaf(hammer_cursor_t cursor)
1779 hammer_node_ondisk_t ondisk;
1780 hammer_node_t parent;
1783 hammer_node_t new_leaf;
1784 hammer_btree_elm_t elm;
1785 hammer_btree_elm_t parent_elm;
1786 hammer_base_elm_t mid_boundary;
1792 const size_t esize = sizeof(*elm);
1794 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1796 ++hammer_stats_btree_splits;
1798 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1799 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1800 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1801 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1804 * Calculate the split point. If the insertion point is at the
1805 * end of the leaf we adjust the split point significantly to the
1806 * right to try to optimize node fill and flag it. If we hit
1807 * that same leaf again our heuristic failed and we don't try
1808 * to optimize node fill (it could lead to a degenerate case).
1810 * Spikes are made up of two leaf elements which cannot be
1813 leaf = cursor->node;
1814 ondisk = leaf->ondisk;
1815 KKASSERT(ondisk->count > 4);
1816 if (cursor->index == ondisk->count &&
1817 (leaf->flags & HAMMER_NODE_NONLINEAR) == 0) {
1818 split = (ondisk->count + 1) * 3 / 4;
1819 leaf->flags |= HAMMER_NODE_NONLINEAR;
1821 split = (ondisk->count + 1) / 2;
1826 * If the insertion point is at the split point shift the
1827 * split point left so we don't have to worry about
1829 if (cursor->index == split)
1832 KKASSERT(split > 0 && split < ondisk->count);
1837 elm = &ondisk->elms[split];
1839 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1840 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1841 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1842 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1845 * If we are at the root of the tree, create a new root node with
1846 * 1 element and split normally. Avoid making major modifications
1847 * until we know the whole operation will work.
1849 if (ondisk->parent == 0) {
1850 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1853 hammer_lock_ex(&parent->lock);
1854 hammer_modify_node_noundo(cursor->trans, parent);
1855 ondisk = parent->ondisk;
1858 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1859 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1860 ondisk->elms[0].base = hmp->root_btree_beg;
1861 ondisk->elms[0].base.btype = leaf->ondisk->type;
1862 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1863 ondisk->elms[1].base = hmp->root_btree_end;
1864 /* ondisk->elms[1].base.btype = not used */
1865 hammer_modify_node_done(parent);
1867 parent_index = 0; /* insertion point in parent */
1870 parent = cursor->parent;
1871 parent_index = cursor->parent_index;
1875 * Calculate a hint for the allocation of the new B-Tree leaf node.
1876 * For now just try to localize it within the same bigblock as
1879 * If the insertion point is at the end of the leaf we recognize a
1880 * likely append sequence of some sort (data, meta-data, inodes,
1881 * whatever). Set the hint to zero to allocate out of linear space
1882 * instead of trying to completely fill previously hinted space.
1884 * This also sets the stage for recursive splits to localize using
1887 ondisk = leaf->ondisk;
1888 if (cursor->index == ondisk->count)
1891 hint = leaf->node_offset;
1894 * Split leaf into new_leaf at the split point. Select a separator
1895 * value in-between the two leafs but with a bent towards the right
1896 * leaf since comparisons use an 'elm >= separator' inequality.
1905 new_leaf = hammer_alloc_btree(cursor->trans, 0, &error);
1906 if (new_leaf == NULL) {
1908 hammer_unlock(&parent->lock);
1909 hammer_delete_node(cursor->trans, parent);
1910 hammer_rel_node(parent);
1914 hammer_lock_ex(&new_leaf->lock);
1917 * Create the new node and copy the leaf elements from the split
1918 * point on to the new node.
1920 hammer_modify_node_all(cursor->trans, leaf);
1921 hammer_modify_node_noundo(cursor->trans, new_leaf);
1922 ondisk = leaf->ondisk;
1923 elm = &ondisk->elms[split];
1924 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1925 new_leaf->ondisk->count = ondisk->count - split;
1926 new_leaf->ondisk->parent = parent->node_offset;
1927 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1928 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1929 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1930 hammer_modify_node_done(new_leaf);
1931 hammer_cursor_split_node(leaf, new_leaf, split);
1934 * Cleanup the original node. Because this is a leaf node and
1935 * leaf nodes do not have a right-hand boundary, there
1936 * aren't any special edge cases to clean up. We just fixup the
1939 ondisk->count = split;
1942 * Insert the separator into the parent, fixup the parent's
1943 * reference to the original node, and reference the new node.
1944 * The separator is P.
1946 * Remember that base.count does not include the right-hand boundary.
1947 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1949 hammer_modify_node_all(cursor->trans, parent);
1950 ondisk = parent->ondisk;
1951 KKASSERT(split != 0);
1952 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1953 parent_elm = &ondisk->elms[parent_index+1];
1954 bcopy(parent_elm, parent_elm + 1,
1955 (ondisk->count - parent_index) * esize);
1957 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1958 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1959 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1960 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1961 mid_boundary = &parent_elm->base;
1963 hammer_modify_node_done(parent);
1964 hammer_cursor_inserted_element(parent, parent_index + 1);
1967 * The filesystem's root B-Tree pointer may have to be updated.
1970 hammer_volume_t volume;
1972 volume = hammer_get_root_volume(hmp, &error);
1973 KKASSERT(error == 0);
1975 hammer_modify_volume_field(cursor->trans, volume,
1977 volume->ondisk->vol0_btree_root = parent->node_offset;
1978 hammer_modify_volume_done(volume);
1979 leaf->ondisk->parent = parent->node_offset;
1980 if (cursor->parent) {
1981 hammer_unlock(&cursor->parent->lock);
1982 hammer_rel_node(cursor->parent);
1984 cursor->parent = parent; /* lock'd and ref'd */
1985 hammer_rel_volume(volume, 0);
1987 hammer_modify_node_done(leaf);
1990 * Ok, now adjust the cursor depending on which element the original
1991 * index was pointing at. If we are >= the split point the push node
1992 * is now in the new node.
1994 * NOTE: If we are at the split point itself we need to select the
1995 * old or new node based on where key_beg's insertion point will be.
1996 * If we pick the wrong side the inserted element will wind up in
1997 * the wrong leaf node and outside that node's bounds.
1999 if (cursor->index > split ||
2000 (cursor->index == split &&
2001 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
2002 cursor->parent_index = parent_index + 1;
2003 cursor->index -= split;
2004 hammer_unlock(&cursor->node->lock);
2005 hammer_rel_node(cursor->node);
2006 cursor->node = new_leaf;
2008 cursor->parent_index = parent_index;
2009 hammer_unlock(&new_leaf->lock);
2010 hammer_rel_node(new_leaf);
2014 * Fixup left and right bounds
2016 parent_elm = &parent->ondisk->elms[cursor->parent_index];
2017 cursor->left_bound = &parent_elm[0].internal.base;
2018 cursor->right_bound = &parent_elm[1].internal.base;
2021 * Assert that the bounds are correct.
2023 KKASSERT(hammer_btree_cmp(cursor->left_bound,
2024 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
2025 KKASSERT(hammer_btree_cmp(cursor->right_bound,
2026 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
2027 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
2028 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
2031 hammer_cursor_downgrade(cursor);
2038 * Recursively correct the right-hand boundary's create_tid to (tid) as
2039 * long as the rest of the key matches. We have to recurse upward in
2040 * the tree as well as down the left side of each parent's right node.
2042 * Return EDEADLK if we were only partially successful, forcing the caller
2043 * to try again. The original cursor is not modified. This routine can
2044 * also fail with EDEADLK if it is forced to throw away a portion of its
2047 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
2050 TAILQ_ENTRY(hammer_rhb) entry;
2055 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
2058 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
2060 struct hammer_mount *hmp;
2061 struct hammer_rhb_list rhb_list;
2062 hammer_base_elm_t elm;
2063 hammer_node_t orig_node;
2064 struct hammer_rhb *rhb;
2068 TAILQ_INIT(&rhb_list);
2069 hmp = cursor->trans->hmp;
2072 * Save our position so we can restore it on return. This also
2073 * gives us a stable 'elm'.
2075 orig_node = cursor->node;
2076 hammer_ref_node(orig_node);
2077 hammer_lock_sh(&orig_node->lock);
2078 orig_index = cursor->index;
2079 elm = &orig_node->ondisk->elms[orig_index].base;
2082 * Now build a list of parents going up, allocating a rhb
2083 * structure for each one.
2085 while (cursor->parent) {
2087 * Stop if we no longer have any right-bounds to fix up
2089 if (elm->obj_id != cursor->right_bound->obj_id ||
2090 elm->rec_type != cursor->right_bound->rec_type ||
2091 elm->key != cursor->right_bound->key) {
2096 * Stop if the right-hand bound's create_tid does not
2097 * need to be corrected.
2099 if (cursor->right_bound->create_tid >= tid)
2102 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2103 rhb->node = cursor->parent;
2104 rhb->index = cursor->parent_index;
2105 hammer_ref_node(rhb->node);
2106 hammer_lock_sh(&rhb->node->lock);
2107 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2109 hammer_cursor_up(cursor);
2113 * now safely adjust the right hand bound for each rhb. This may
2114 * also require taking the right side of the tree and iterating down
2118 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2119 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2122 TAILQ_REMOVE(&rhb_list, rhb, entry);
2123 hammer_unlock(&rhb->node->lock);
2124 hammer_rel_node(rhb->node);
2125 kfree(rhb, hmp->m_misc);
2127 switch (cursor->node->ondisk->type) {
2128 case HAMMER_BTREE_TYPE_INTERNAL:
2130 * Right-boundary for parent at internal node
2131 * is one element to the right of the element whos
2132 * right boundary needs adjusting. We must then
2133 * traverse down the left side correcting any left
2134 * bounds (which may now be too far to the left).
2137 error = hammer_btree_correct_lhb(cursor, tid);
2140 panic("hammer_btree_correct_rhb(): Bad node type");
2149 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2150 TAILQ_REMOVE(&rhb_list, rhb, entry);
2151 hammer_unlock(&rhb->node->lock);
2152 hammer_rel_node(rhb->node);
2153 kfree(rhb, hmp->m_misc);
2155 error = hammer_cursor_seek(cursor, orig_node, orig_index);
2156 hammer_unlock(&orig_node->lock);
2157 hammer_rel_node(orig_node);
2162 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2163 * bound going downward starting at the current cursor position.
2165 * This function does not restore the cursor after use.
2168 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
2170 struct hammer_rhb_list rhb_list;
2171 hammer_base_elm_t elm;
2172 hammer_base_elm_t cmp;
2173 struct hammer_rhb *rhb;
2174 struct hammer_mount *hmp;
2177 TAILQ_INIT(&rhb_list);
2178 hmp = cursor->trans->hmp;
2180 cmp = &cursor->node->ondisk->elms[cursor->index].base;
2183 * Record the node and traverse down the left-hand side for all
2184 * matching records needing a boundary correction.
2188 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2189 rhb->node = cursor->node;
2190 rhb->index = cursor->index;
2191 hammer_ref_node(rhb->node);
2192 hammer_lock_sh(&rhb->node->lock);
2193 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2195 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2197 * Nothing to traverse down if we are at the right
2198 * boundary of an internal node.
2200 if (cursor->index == cursor->node->ondisk->count)
2203 elm = &cursor->node->ondisk->elms[cursor->index].base;
2204 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2206 panic("Illegal leaf record type %02x", elm->btype);
2208 error = hammer_cursor_down(cursor);
2212 elm = &cursor->node->ondisk->elms[cursor->index].base;
2213 if (elm->obj_id != cmp->obj_id ||
2214 elm->rec_type != cmp->rec_type ||
2215 elm->key != cmp->key) {
2218 if (elm->create_tid >= tid)
2224 * Now we can safely adjust the left-hand boundary from the bottom-up.
2225 * The last element we remove from the list is the caller's right hand
2226 * boundary, which must also be adjusted.
2228 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2229 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2232 TAILQ_REMOVE(&rhb_list, rhb, entry);
2233 hammer_unlock(&rhb->node->lock);
2234 hammer_rel_node(rhb->node);
2235 kfree(rhb, hmp->m_misc);
2237 elm = &cursor->node->ondisk->elms[cursor->index].base;
2238 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2239 hammer_modify_node(cursor->trans, cursor->node,
2241 sizeof(elm->create_tid));
2242 elm->create_tid = tid;
2243 hammer_modify_node_done(cursor->node);
2245 panic("hammer_btree_correct_lhb(): Bad element type");
2252 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2253 TAILQ_REMOVE(&rhb_list, rhb, entry);
2254 hammer_unlock(&rhb->node->lock);
2255 hammer_rel_node(rhb->node);
2256 kfree(rhb, hmp->m_misc);
2264 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2265 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2266 * the operation due to a deadlock, or some other error.
2268 * This routine is initially called with an empty leaf and may be
2269 * recursively called with single-element internal nodes.
2271 * It should also be noted that when removing empty leaves we must be sure
2272 * to test and update mirror_tid because another thread may have deadlocked
2273 * against us (or someone) trying to propagate it up and cannot retry once
2274 * the node has been deleted.
2276 * On return the cursor may end up pointing to an internal node, suitable
2277 * for further iteration but not for an immediate insertion or deletion.
2280 btree_remove(hammer_cursor_t cursor)
2282 hammer_node_ondisk_t ondisk;
2283 hammer_btree_elm_t elm;
2285 hammer_node_t parent;
2286 const int esize = sizeof(*elm);
2289 node = cursor->node;
2292 * When deleting the root of the filesystem convert it to
2293 * an empty leaf node. Internal nodes cannot be empty.
2295 ondisk = node->ondisk;
2296 if (ondisk->parent == 0) {
2297 KKASSERT(cursor->parent == NULL);
2298 hammer_modify_node_all(cursor->trans, node);
2299 KKASSERT(ondisk == node->ondisk);
2300 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2302 hammer_modify_node_done(node);
2307 parent = cursor->parent;
2310 * Attempt to remove the parent's reference to the child. If the
2311 * parent would become empty we have to recurse. If we fail we
2312 * leave the parent pointing to an empty leaf node.
2314 * We have to recurse successfully before we can delete the internal
2315 * node as it is illegal to have empty internal nodes. Even though
2316 * the operation may be aborted we must still fixup any unlocked
2317 * cursors as if we had deleted the element prior to recursing
2318 * (by calling hammer_cursor_deleted_element()) so those cursors
2319 * are properly forced up the chain by the recursion.
2321 if (parent->ondisk->count == 1) {
2323 * This special cursor_up_locked() call leaves the original
2324 * node exclusively locked and referenced, leaves the
2325 * original parent locked (as the new node), and locks the
2326 * new parent. It can return EDEADLK.
2328 * We cannot call hammer_cursor_removed_node() until we are
2329 * actually able to remove the node. If we did then tracked
2330 * cursors in the middle of iterations could be repointed
2331 * to a parent node. If this occurs they could end up
2332 * scanning newly inserted records into the node (that could
2333 * not be deleted) when they push down again.
2335 * Due to the way the recursion works the final parent is left
2336 * in cursor->parent after the recursion returns. Each
2337 * layer on the way back up is thus able to call
2338 * hammer_cursor_removed_node() and 'jump' the node up to
2339 * the (same) final parent.
2341 * NOTE! The local variable 'parent' is invalid after we
2342 * call hammer_cursor_up_locked().
2344 error = hammer_cursor_up_locked(cursor);
2348 hammer_cursor_deleted_element(cursor->node, 0);
2349 error = btree_remove(cursor);
2351 KKASSERT(node != cursor->node);
2352 hammer_cursor_removed_node(
2355 hammer_modify_node_all(cursor->trans, node);
2356 ondisk = node->ondisk;
2357 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2359 hammer_modify_node_done(node);
2360 hammer_flush_node(node, 0);
2361 hammer_delete_node(cursor->trans, node);
2364 * Defer parent removal because we could not
2365 * get the lock, just let the leaf remain
2370 hammer_unlock(&node->lock);
2371 hammer_rel_node(node);
2374 * Defer parent removal because we could not
2375 * get the lock, just let the leaf remain
2381 KKASSERT(parent->ondisk->count > 1);
2383 hammer_modify_node_all(cursor->trans, parent);
2384 ondisk = parent->ondisk;
2385 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2387 elm = &ondisk->elms[cursor->parent_index];
2388 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2389 KKASSERT(ondisk->count > 0);
2392 * We must retain the highest mirror_tid. The deleted
2393 * range is now encompassed by the element to the left.
2394 * If we are already at the left edge the new left edge
2395 * inherits mirror_tid.
2397 * Note that bounds of the parent to our parent may create
2398 * a gap to the left of our left-most node or to the right
2399 * of our right-most node. The gap is silently included
2400 * in the mirror_tid's area of effect from the point of view
2403 if (cursor->parent_index) {
2404 if (elm[-1].internal.mirror_tid <
2405 elm[0].internal.mirror_tid) {
2406 elm[-1].internal.mirror_tid =
2407 elm[0].internal.mirror_tid;
2410 if (elm[1].internal.mirror_tid <
2411 elm[0].internal.mirror_tid) {
2412 elm[1].internal.mirror_tid =
2413 elm[0].internal.mirror_tid;
2418 * Delete the subtree reference in the parent. Include
2419 * boundary element at end.
2421 bcopy(&elm[1], &elm[0],
2422 (ondisk->count - cursor->parent_index) * esize);
2424 hammer_modify_node_done(parent);
2425 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2426 hammer_cursor_deleted_element(parent, cursor->parent_index);
2427 hammer_flush_node(node, 0);
2428 hammer_delete_node(cursor->trans, node);
2431 * cursor->node is invalid, cursor up to make the cursor
2432 * valid again. We have to flag the condition in case
2433 * another thread wiggles an insertion in during an
2436 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK;
2437 error = hammer_cursor_up(cursor);
2443 * Propagate cursor->trans->tid up the B-Tree starting at the current
2444 * cursor position using pseudofs info gleaned from the passed inode.
2446 * The passed inode has no relationship to the cursor position other
2447 * then being in the same pseudofs as the insertion or deletion we
2448 * are propagating the mirror_tid for.
2450 * WARNING! Because we push and pop the passed cursor, it may be
2451 * modified by other B-Tree operations while it is unlocked
2452 * and things like the node & leaf pointers, and indexes might
2456 hammer_btree_do_propagation(hammer_cursor_t cursor,
2457 hammer_pseudofs_inmem_t pfsm,
2458 hammer_btree_leaf_elm_t leaf)
2460 hammer_cursor_t ncursor;
2461 hammer_tid_t mirror_tid;
2465 * We do not propagate a mirror_tid if the filesystem was mounted
2466 * in no-mirror mode.
2468 if (cursor->trans->hmp->master_id < 0)
2472 * This is a bit of a hack because we cannot deadlock or return
2473 * EDEADLK here. The related operation has already completed and
2474 * we must propagate the mirror_tid now regardless.
2476 * Generate a new cursor which inherits the original's locks and
2477 * unlock the original. Use the new cursor to propagate the
2478 * mirror_tid. Then clean up the new cursor and reacquire locks
2481 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2482 * original's locks and the original is tracked and must be
2485 mirror_tid = cursor->node->ondisk->mirror_tid;
2486 KKASSERT(mirror_tid != 0);
2487 ncursor = hammer_push_cursor(cursor);
2488 error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2489 KKASSERT(error == 0);
2490 hammer_pop_cursor(cursor, ncursor);
2491 /* WARNING: cursor's leaf pointer may change after pop */
2496 * Propagate a mirror TID update upwards through the B-Tree to the root.
2498 * A locked internal node must be passed in. The node will remain locked
2501 * This function syncs mirror_tid at the specified internal node's element,
2502 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2505 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2507 hammer_btree_internal_elm_t elm;
2512 error = hammer_cursor_up(cursor);
2514 error = hammer_cursor_upgrade(cursor);
2517 * We can ignore HAMMER_CURSOR_ITERATE_CHECK, the
2518 * cursor will still be properly positioned for
2519 * mirror propagation, just not for iterations.
2521 while (error == EDEADLK) {
2522 hammer_recover_cursor(cursor);
2523 error = hammer_cursor_upgrade(cursor);
2529 * If the cursor deadlocked it could end up at a leaf
2530 * after we lost the lock.
2532 node = cursor->node;
2533 if (node->ondisk->type != HAMMER_BTREE_TYPE_INTERNAL)
2537 * Adjust the node's element
2539 elm = &node->ondisk->elms[cursor->index].internal;
2540 if (elm->mirror_tid >= mirror_tid)
2542 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2543 sizeof(elm->mirror_tid));
2544 elm->mirror_tid = mirror_tid;
2545 hammer_modify_node_done(node);
2546 if (hammer_debug_general & 0x0002) {
2547 kprintf("mirror_propagate: propagate "
2548 "%016llx @%016llx:%d\n",
2549 (long long)mirror_tid,
2550 (long long)node->node_offset,
2556 * Adjust the node's mirror_tid aggregator
2558 if (node->ondisk->mirror_tid >= mirror_tid)
2560 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2561 node->ondisk->mirror_tid = mirror_tid;
2562 hammer_modify_node_done(node);
2563 if (hammer_debug_general & 0x0002) {
2564 kprintf("mirror_propagate: propagate "
2565 "%016llx @%016llx\n",
2566 (long long)mirror_tid,
2567 (long long)node->node_offset);
2570 if (error == ENOENT)
2576 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2577 int *parent_indexp, int *errorp, int try_exclusive)
2579 hammer_node_t parent;
2580 hammer_btree_elm_t elm;
2586 parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2588 KKASSERT(parent == NULL);
2591 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2596 if (try_exclusive) {
2597 if (hammer_lock_ex_try(&parent->lock)) {
2598 hammer_rel_node(parent);
2603 hammer_lock_sh(&parent->lock);
2607 * Figure out which element in the parent is pointing to the
2610 if (node->ondisk->count) {
2611 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2616 while (i < parent->ondisk->count) {
2617 elm = &parent->ondisk->elms[i];
2618 if (elm->internal.subtree_offset == node->node_offset)
2622 if (i == parent->ondisk->count) {
2623 hammer_unlock(&parent->lock);
2624 panic("Bad B-Tree link: parent %p node %p", parent, node);
2627 KKASSERT(*errorp == 0);
2632 * The element (elm) has been moved to a new internal node (node).
2634 * If the element represents a pointer to an internal node that node's
2635 * parent must be adjusted to the element's new location.
2637 * XXX deadlock potential here with our exclusive locks
2640 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2641 hammer_btree_elm_t elm)
2643 hammer_node_t child;
2648 switch(elm->base.btype) {
2649 case HAMMER_BTREE_TYPE_INTERNAL:
2650 case HAMMER_BTREE_TYPE_LEAF:
2651 child = hammer_get_node(trans, elm->internal.subtree_offset,
2654 hammer_modify_node_field(trans, child, parent);
2655 child->ondisk->parent = node->node_offset;
2656 hammer_modify_node_done(child);
2657 hammer_rel_node(child);
2667 * Initialize the root of a recursive B-Tree node lock list structure.
2670 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2672 TAILQ_INIT(&parent->list);
2673 parent->parent = NULL;
2674 parent->node = node;
2676 parent->count = node->ondisk->count;
2677 parent->copy = NULL;
2682 * Initialize a cache of hammer_node_lock's including space allocated
2685 * This is used by the rebalancing code to preallocate the copy space
2686 * for ~4096 B-Tree nodes (16MB of data) prior to acquiring any HAMMER
2687 * locks, otherwise we can blow out the pageout daemon's emergency
2688 * reserve and deadlock it.
2690 * NOTE: HAMMER_NODE_LOCK_LCACHE is not set on items cached in the lcache.
2691 * The flag is set when the item is pulled off the cache for use.
2694 hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
2697 hammer_node_lock_t item;
2700 for (count = 1; depth; --depth)
2701 count *= HAMMER_BTREE_LEAF_ELMS;
2702 bzero(lcache, sizeof(*lcache));
2703 TAILQ_INIT(&lcache->list);
2705 item = kmalloc(sizeof(*item), hmp->m_misc, M_WAITOK|M_ZERO);
2706 item->copy = kmalloc(sizeof(*item->copy),
2707 hmp->m_misc, M_WAITOK);
2708 TAILQ_INIT(&item->list);
2709 TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2715 hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache)
2717 hammer_node_lock_t item;
2719 while ((item = TAILQ_FIRST(&lcache->list)) != NULL) {
2720 TAILQ_REMOVE(&lcache->list, item, entry);
2721 KKASSERT(item->copy);
2722 KKASSERT(TAILQ_EMPTY(&item->list));
2723 kfree(item->copy, hmp->m_misc);
2724 kfree(item, hmp->m_misc);
2726 KKASSERT(lcache->copy == NULL);
2730 * Exclusively lock all the children of node. This is used by the split
2731 * code to prevent anyone from accessing the children of a cursor node
2732 * while we fix-up its parent offset.
2734 * If we don't lock the children we can really mess up cursors which block
2735 * trying to cursor-up into our node.
2737 * On failure EDEADLK (or some other error) is returned. If a deadlock
2738 * error is returned the cursor is adjusted to block on termination.
2740 * The caller is responsible for managing parent->node, the root's node
2741 * is usually aliased from a cursor.
2744 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2745 hammer_node_lock_t parent,
2746 hammer_node_lock_t lcache)
2749 hammer_node_lock_t item;
2750 hammer_node_ondisk_t ondisk;
2751 hammer_btree_elm_t elm;
2752 hammer_node_t child;
2753 struct hammer_mount *hmp;
2757 node = parent->node;
2758 ondisk = node->ondisk;
2760 hmp = cursor->trans->hmp;
2763 * We really do not want to block on I/O with exclusive locks held,
2764 * pre-get the children before trying to lock the mess. This is
2765 * only done one-level deep for now.
2767 for (i = 0; i < ondisk->count; ++i) {
2768 ++hammer_stats_btree_elements;
2769 elm = &ondisk->elms[i];
2770 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2771 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2774 child = hammer_get_node(cursor->trans,
2775 elm->internal.subtree_offset,
2778 hammer_rel_node(child);
2784 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2785 ++hammer_stats_btree_elements;
2786 elm = &ondisk->elms[i];
2788 switch(elm->base.btype) {
2789 case HAMMER_BTREE_TYPE_INTERNAL:
2790 case HAMMER_BTREE_TYPE_LEAF:
2791 KKASSERT(elm->internal.subtree_offset != 0);
2792 child = hammer_get_node(cursor->trans,
2793 elm->internal.subtree_offset,
2801 if (hammer_lock_ex_try(&child->lock) != 0) {
2802 if (cursor->deadlk_node == NULL) {
2803 cursor->deadlk_node = child;
2804 hammer_ref_node(cursor->deadlk_node);
2807 hammer_rel_node(child);
2810 item = TAILQ_FIRST(&lcache->list);
2811 KKASSERT(item != NULL);
2812 item->flags |= HAMMER_NODE_LOCK_LCACHE;
2813 TAILQ_REMOVE(&lcache->list,
2816 item = kmalloc(sizeof(*item),
2819 TAILQ_INIT(&item->list);
2822 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2823 item->parent = parent;
2826 item->count = child->ondisk->count;
2829 * Recurse (used by the rebalancing code)
2831 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2832 error = hammer_btree_lock_children(
2842 hammer_btree_unlock_children(hmp, parent, lcache);
2847 * Create an in-memory copy of all B-Tree nodes listed, recursively,
2848 * including the parent.
2851 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2853 hammer_mount_t hmp = cursor->trans->hmp;
2854 hammer_node_lock_t item;
2856 if (parent->copy == NULL) {
2857 KKASSERT((parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0);
2858 parent->copy = kmalloc(sizeof(*parent->copy),
2859 hmp->m_misc, M_WAITOK);
2861 KKASSERT((parent->flags & HAMMER_NODE_LOCK_UPDATED) == 0);
2862 *parent->copy = *parent->node->ondisk;
2863 TAILQ_FOREACH(item, &parent->list, entry) {
2864 hammer_btree_lock_copy(cursor, item);
2869 * Recursively sync modified copies to the media.
2872 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2874 hammer_node_lock_t item;
2877 if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2879 hammer_modify_node_all(cursor->trans, parent->node);
2880 *parent->node->ondisk = *parent->copy;
2881 hammer_modify_node_done(parent->node);
2882 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2883 hammer_flush_node(parent->node, 0);
2884 hammer_delete_node(cursor->trans, parent->node);
2887 TAILQ_FOREACH(item, &parent->list, entry) {
2888 count += hammer_btree_sync_copy(cursor, item);
2894 * Release previously obtained node locks. The caller is responsible for
2895 * cleaning up parent->node itself (its usually just aliased from a cursor),
2896 * but this function will take care of the copies.
2898 * NOTE: The root node is not placed in the lcache and node->copy is not
2899 * deallocated when lcache != NULL.
2902 hammer_btree_unlock_children(hammer_mount_t hmp, hammer_node_lock_t parent,
2903 hammer_node_lock_t lcache)
2905 hammer_node_lock_t item;
2906 hammer_node_ondisk_t copy;
2908 while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2909 TAILQ_REMOVE(&parent->list, item, entry);
2910 hammer_btree_unlock_children(hmp, item, lcache);
2911 hammer_unlock(&item->node->lock);
2912 hammer_rel_node(item->node);
2915 * NOTE: When placing the item back in the lcache
2916 * the flag is cleared by the bzero().
2917 * Remaining fields are cleared as a safety
2920 KKASSERT(item->flags & HAMMER_NODE_LOCK_LCACHE);
2921 KKASSERT(TAILQ_EMPTY(&item->list));
2923 bzero(item, sizeof(*item));
2924 TAILQ_INIT(&item->list);
2927 bzero(copy, sizeof(*copy));
2928 TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2930 kfree(item, hmp->m_misc);
2933 if (parent->copy && (parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0) {
2934 kfree(parent->copy, hmp->m_misc);
2935 parent->copy = NULL; /* safety */
2939 /************************************************************************
2940 * MISCELLANIOUS SUPPORT *
2941 ************************************************************************/
2944 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2946 * Note that for this particular function a return value of -1, 0, or +1
2947 * can denote a match if create_tid is otherwise discounted. A create_tid
2948 * of zero is considered to be 'infinity' in comparisons.
2950 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2953 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2955 if (key1->localization < key2->localization)
2957 if (key1->localization > key2->localization)
2960 if (key1->obj_id < key2->obj_id)
2962 if (key1->obj_id > key2->obj_id)
2965 if (key1->rec_type < key2->rec_type)
2967 if (key1->rec_type > key2->rec_type)
2970 if (key1->key < key2->key)
2972 if (key1->key > key2->key)
2976 * A create_tid of zero indicates a record which is undeletable
2977 * and must be considered to have a value of positive infinity.
2979 if (key1->create_tid == 0) {
2980 if (key2->create_tid == 0)
2984 if (key2->create_tid == 0)
2986 if (key1->create_tid < key2->create_tid)
2988 if (key1->create_tid > key2->create_tid)
2994 * Test a timestamp against an element to determine whether the
2995 * element is visible. A timestamp of 0 means 'infinity'.
2998 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
3001 if (base->delete_tid)
3005 if (asof < base->create_tid)
3007 if (base->delete_tid && asof >= base->delete_tid)
3013 * Create a separator half way inbetween key1 and key2. For fields just
3014 * one unit apart, the separator will match key2. key1 is on the left-hand
3015 * side and key2 is on the right-hand side.
3017 * key2 must be >= the separator. It is ok for the separator to match key2.
3019 * NOTE: Even if key1 does not match key2, the separator may wind up matching
3022 * NOTE: It might be beneficial to just scrap this whole mess and just
3023 * set the separator to key2.
3025 #define MAKE_SEPARATOR(key1, key2, dest, field) \
3026 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
3029 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
3030 hammer_base_elm_t dest)
3032 bzero(dest, sizeof(*dest));
3034 dest->rec_type = key2->rec_type;
3035 dest->key = key2->key;
3036 dest->obj_id = key2->obj_id;
3037 dest->create_tid = key2->create_tid;
3039 MAKE_SEPARATOR(key1, key2, dest, localization);
3040 if (key1->localization == key2->localization) {
3041 MAKE_SEPARATOR(key1, key2, dest, obj_id);
3042 if (key1->obj_id == key2->obj_id) {
3043 MAKE_SEPARATOR(key1, key2, dest, rec_type);
3044 if (key1->rec_type == key2->rec_type) {
3045 MAKE_SEPARATOR(key1, key2, dest, key);
3047 * Don't bother creating a separator for
3048 * create_tid, which also conveniently avoids
3049 * having to handle the create_tid == 0
3050 * (infinity) case. Just leave create_tid
3053 * Worst case, dest matches key2 exactly,
3054 * which is acceptable.
3061 #undef MAKE_SEPARATOR
3064 * Return whether a generic internal or leaf node is full
3067 btree_node_is_full(hammer_node_ondisk_t node)
3069 switch(node->type) {
3070 case HAMMER_BTREE_TYPE_INTERNAL:
3071 if (node->count == HAMMER_BTREE_INT_ELMS)
3074 case HAMMER_BTREE_TYPE_LEAF:
3075 if (node->count == HAMMER_BTREE_LEAF_ELMS)
3079 panic("illegal btree subtype");
3086 btree_max_elements(u_int8_t type)
3088 if (type == HAMMER_BTREE_TYPE_LEAF)
3089 return(HAMMER_BTREE_LEAF_ELMS);
3090 if (type == HAMMER_BTREE_TYPE_INTERNAL)
3091 return(HAMMER_BTREE_INT_ELMS);
3092 panic("btree_max_elements: bad type %d", type);
3097 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
3099 hammer_btree_elm_t elm;
3102 kprintf("node %p count=%d parent=%016llx type=%c\n",
3103 ondisk, ondisk->count,
3104 (long long)ondisk->parent, ondisk->type);
3107 * Dump both boundary elements if an internal node
3109 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
3110 for (i = 0; i <= ondisk->count; ++i) {
3111 elm = &ondisk->elms[i];
3112 hammer_print_btree_elm(elm, ondisk->type, i);
3115 for (i = 0; i < ondisk->count; ++i) {
3116 elm = &ondisk->elms[i];
3117 hammer_print_btree_elm(elm, ondisk->type, i);
3123 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
3126 kprintf("\tobj_id = %016llx\n", (long long)elm->base.obj_id);
3127 kprintf("\tkey = %016llx\n", (long long)elm->base.key);
3128 kprintf("\tcreate_tid = %016llx\n", (long long)elm->base.create_tid);
3129 kprintf("\tdelete_tid = %016llx\n", (long long)elm->base.delete_tid);
3130 kprintf("\trec_type = %04x\n", elm->base.rec_type);
3131 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
3132 kprintf("\tbtype = %02x (%c)\n",
3134 (elm->base.btype ? elm->base.btype : '?'));
3135 kprintf("\tlocalization = %02x\n", elm->base.localization);
3138 case HAMMER_BTREE_TYPE_INTERNAL:
3139 kprintf("\tsubtree_off = %016llx\n",
3140 (long long)elm->internal.subtree_offset);
3142 case HAMMER_BTREE_TYPE_RECORD:
3143 kprintf("\tdata_offset = %016llx\n",
3144 (long long)elm->leaf.data_offset);
3145 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
3146 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);