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 * with no valid sub-tree pointer for the right-most element.
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.
57 * The left-hand boundary (B in the left) is integrated into the first
58 * element so it doesn't require 2 elements to accomodate boundaries.
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, int *ndelete);
89 static __inline int btree_node_is_full(hammer_node_ondisk_t node);
90 static __inline int btree_max_elements(u_int8_t type);
91 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor,
92 hammer_tid_t mirror_tid);
93 static void hammer_make_separator(hammer_base_elm_t key1,
94 hammer_base_elm_t key2, hammer_base_elm_t dest);
95 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor);
98 * Iterate records after a search. The cursor is iterated forwards past
99 * the current record until a record matching the key-range requirements
100 * is found. ENOENT is returned if the iteration goes past the ending
103 * The iteration is inclusive of key_beg and can be inclusive or exclusive
104 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
106 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
107 * may be modified by B-Tree functions.
109 * cursor->key_beg may or may not be modified by this function during
110 * the iteration. XXX future - in case of an inverted lock we may have
111 * to reinitiate the lookup and set key_beg to properly pick up where we
114 * If HAMMER_CURSOR_ITERATE_CHECK is set it is possible that the cursor
115 * was reverse indexed due to being moved to a parent while unlocked,
116 * and something else might have inserted an element outside the iteration
117 * range. When this case occurs the iterator just keeps iterating until
118 * it gets back into the iteration range (instead of asserting).
120 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
123 hammer_btree_iterate(hammer_cursor_t cursor)
125 hammer_node_ondisk_t node;
126 hammer_btree_elm_t elm;
133 * Skip past the current record
135 hmp = cursor->trans->hmp;
136 node = cursor->node->ondisk;
139 if (cursor->index < node->count &&
140 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
145 * HAMMER can wind up being cpu-bound.
147 if (++hmp->check_yield > hammer_yield_check) {
148 hmp->check_yield = 0;
154 * Loop until an element is found or we are done.
158 * We iterate up the tree and then index over one element
159 * while we are at the last element in the current node.
161 * If we are at the root of the filesystem, cursor_up
164 * XXX this could be optimized by storing the information in
165 * the parent reference.
167 * XXX we can lose the node lock temporarily, this could mess
170 ++hammer_stats_btree_iterations;
171 hammer_flusher_clean_loose_ios(hmp);
173 if (cursor->index == node->count) {
174 if (hammer_debug_btree) {
175 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
176 (long long)cursor->node->node_offset,
178 (long long)(cursor->parent ? cursor->parent->node_offset : -1),
179 cursor->parent_index,
182 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
183 error = hammer_cursor_up(cursor);
186 /* reload stale pointer */
187 node = cursor->node->ondisk;
188 KKASSERT(cursor->index != node->count);
191 * If we are reblocking we want to return internal
192 * nodes. Note that the internal node will be
193 * returned multiple times, on each upward recursion
194 * from its children. The caller selects which
195 * revisit it cares about (usually first or last only).
197 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
198 cursor->flags |= HAMMER_CURSOR_ATEDISK;
206 * Check internal or leaf element. Determine if the record
207 * at the cursor has gone beyond the end of our range.
209 * We recurse down through internal nodes.
211 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
212 elm = &node->elms[cursor->index];
214 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
215 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
216 if (hammer_debug_btree) {
217 kprintf("BRACKETL %016llx[%d] %016llx %02x "
218 "key=%016llx lo=%02x %d (td=%p)\n",
219 (long long)cursor->node->node_offset,
221 (long long)elm[0].internal.base.obj_id,
222 elm[0].internal.base.rec_type,
223 (long long)elm[0].internal.base.key,
224 elm[0].internal.base.localization,
228 kprintf("BRACKETR %016llx[%d] %016llx %02x "
229 "key=%016llx lo=%02x %d\n",
230 (long long)cursor->node->node_offset,
232 (long long)elm[1].internal.base.obj_id,
233 elm[1].internal.base.rec_type,
234 (long long)elm[1].internal.base.key,
235 elm[1].internal.base.localization,
244 if (r == 0 && (cursor->flags &
245 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
253 KKASSERT(elm->internal.subtree_offset != 0);
257 * If running the mirror filter see if we
258 * can skip one or more entire sub-trees.
259 * If we can we return the internal node
260 * and the caller processes the skipped
261 * range (see mirror_read).
264 HAMMER_CURSOR_MIRROR_FILTERED) {
265 if (elm->internal.mirror_tid <
266 cursor->cmirror->mirror_tid) {
267 hammer_cursor_mirror_filter(cursor);
273 * Normally it would be impossible for the
274 * cursor to have gotten back-indexed,
275 * but it can happen if a node is deleted
276 * and the cursor is moved to its parent
277 * internal node. ITERATE_CHECK will be set.
279 KKASSERT(cursor->flags &
280 HAMMER_CURSOR_ITERATE_CHECK);
281 kprintf("hammer_btree_iterate: "
282 "DEBUG: Caught parent seek "
283 "in internal iteration\n");
286 error = hammer_cursor_down(cursor);
289 KKASSERT(cursor->index == 0);
290 /* reload stale pointer */
291 node = cursor->node->ondisk;
294 elm = &node->elms[cursor->index];
295 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
296 if (hammer_debug_btree) {
297 kprintf("ELEMENT %016llx:%d %c %016llx %02x "
298 "key=%016llx lo=%02x %d\n",
299 (long long)cursor->node->node_offset,
301 (elm[0].leaf.base.btype ?
302 elm[0].leaf.base.btype : '?'),
303 (long long)elm[0].leaf.base.obj_id,
304 elm[0].leaf.base.rec_type,
305 (long long)elm[0].leaf.base.key,
306 elm[0].leaf.base.localization,
316 * We support both end-inclusive and
317 * end-exclusive searches.
320 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
326 * If ITERATE_CHECK is set an unlocked cursor may
327 * have been moved to a parent and the iterate can
328 * happen upon elements that are not in the requested
331 if (cursor->flags & HAMMER_CURSOR_ITERATE_CHECK) {
332 s = hammer_btree_cmp(&cursor->key_beg,
335 kprintf("hammer_btree_iterate: "
336 "DEBUG: Caught parent seek "
337 "in leaf iteration\n");
342 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
347 switch(elm->leaf.base.btype) {
348 case HAMMER_BTREE_TYPE_RECORD:
349 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
350 hammer_btree_chkts(cursor->asof, &elm->base)) {
367 if (hammer_debug_btree) {
368 int i = cursor->index;
369 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
370 kprintf("ITERATE %p:%d %016llx %02x "
371 "key=%016llx lo=%02x\n",
373 (long long)elm->internal.base.obj_id,
374 elm->internal.base.rec_type,
375 (long long)elm->internal.base.key,
376 elm->internal.base.localization
385 * We hit an internal element that we could skip as part of a mirroring
386 * scan. Calculate the entire range being skipped.
388 * It is important to include any gaps between the parent's left_bound
389 * and the node's left_bound, and same goes for the right side.
392 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
394 struct hammer_cmirror *cmirror;
395 hammer_node_ondisk_t ondisk;
396 hammer_btree_elm_t elm;
398 ondisk = cursor->node->ondisk;
399 cmirror = cursor->cmirror;
402 * Calculate the skipped range
404 elm = &ondisk->elms[cursor->index];
405 if (cursor->index == 0)
406 cmirror->skip_beg = *cursor->left_bound;
408 cmirror->skip_beg = elm->internal.base;
409 while (cursor->index < ondisk->count) {
410 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
415 if (cursor->index == ondisk->count)
416 cmirror->skip_end = *cursor->right_bound;
418 cmirror->skip_end = elm->internal.base;
421 * clip the returned result.
423 if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
424 cmirror->skip_beg = cursor->key_beg;
425 if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
426 cmirror->skip_end = cursor->key_end;
430 * Iterate in the reverse direction. This is used by the pruning code to
431 * avoid overlapping records.
434 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
436 hammer_node_ondisk_t node;
437 hammer_btree_elm_t elm;
443 /* mirror filtering not supported for reverse iteration */
444 KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
447 * Skip past the current record. For various reasons the cursor
448 * may end up set to -1 or set to point at the end of the current
449 * node. These cases must be addressed.
451 node = cursor->node->ondisk;
454 if (cursor->index != -1 &&
455 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
458 if (cursor->index == cursor->node->ondisk->count)
462 * HAMMER can wind up being cpu-bound.
464 hmp = cursor->trans->hmp;
465 if (++hmp->check_yield > hammer_yield_check) {
466 hmp->check_yield = 0;
471 * Loop until an element is found or we are done.
474 ++hammer_stats_btree_iterations;
475 hammer_flusher_clean_loose_ios(hmp);
478 * We iterate up the tree and then index over one element
479 * while we are at the last element in the current node.
481 if (cursor->index == -1) {
482 error = hammer_cursor_up(cursor);
484 cursor->index = 0; /* sanity */
487 /* reload stale pointer */
488 node = cursor->node->ondisk;
489 KKASSERT(cursor->index != node->count);
495 * Check internal or leaf element. Determine if the record
496 * at the cursor has gone beyond the end of our range.
498 * We recurse down through internal nodes.
500 KKASSERT(cursor->index != node->count);
501 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
502 elm = &node->elms[cursor->index];
504 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
505 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
506 if (hammer_debug_btree) {
507 kprintf("BRACKETL %016llx[%d] %016llx %02x "
508 "key=%016llx lo=%02x %d (td=%p)\n",
509 (long long)cursor->node->node_offset,
511 (long long)elm[0].internal.base.obj_id,
512 elm[0].internal.base.rec_type,
513 (long long)elm[0].internal.base.key,
514 elm[0].internal.base.localization,
518 kprintf("BRACKETR %016llx[%d] %016llx %02x "
519 "key=%016llx lo=%02x %d\n",
520 (long long)cursor->node->node_offset,
522 (long long)elm[1].internal.base.obj_id,
523 elm[1].internal.base.rec_type,
524 (long long)elm[1].internal.base.key,
525 elm[1].internal.base.localization,
536 * It shouldn't be possible to be seeked past key_end,
537 * even if the cursor got moved to a parent.
544 KKASSERT(elm->internal.subtree_offset != 0);
546 error = hammer_cursor_down(cursor);
549 KKASSERT(cursor->index == 0);
550 /* reload stale pointer */
551 node = cursor->node->ondisk;
553 /* this can assign -1 if the leaf was empty */
554 cursor->index = node->count - 1;
557 elm = &node->elms[cursor->index];
558 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
559 if (hammer_debug_btree) {
560 kprintf("ELEMENTR %016llx:%d %c %016llx %02x "
561 "key=%016llx lo=%02x %d\n",
562 (long long)cursor->node->node_offset,
564 (elm[0].leaf.base.btype ?
565 elm[0].leaf.base.btype : '?'),
566 (long long)elm[0].leaf.base.obj_id,
567 elm[0].leaf.base.rec_type,
568 (long long)elm[0].leaf.base.key,
569 elm[0].leaf.base.localization,
579 * It shouldn't be possible to be seeked past key_end,
580 * even if the cursor got moved to a parent.
582 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
587 switch(elm->leaf.base.btype) {
588 case HAMMER_BTREE_TYPE_RECORD:
589 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
590 hammer_btree_chkts(cursor->asof, &elm->base)) {
607 if (hammer_debug_btree) {
608 int i = cursor->index;
609 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
610 kprintf("ITERATER %p:%d %016llx %02x "
611 "key=%016llx lo=%02x\n",
613 (long long)elm->internal.base.obj_id,
614 elm->internal.base.rec_type,
615 (long long)elm->internal.base.key,
616 elm->internal.base.localization
625 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
626 * could not be found, EDEADLK if inserting and a retry is needed, and a
627 * fatal error otherwise. When retrying, the caller must terminate the
628 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
630 * The cursor is suitably positioned for a deletion on success, and suitably
631 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
634 * The cursor may begin anywhere, the search will traverse the tree in
635 * either direction to locate the requested element.
637 * Most of the logic implementing historical searches is handled here. We
638 * do an initial lookup with create_tid set to the asof TID. Due to the
639 * way records are laid out, a backwards iteration may be required if
640 * ENOENT is returned to locate the historical record. Here's the
643 * create_tid: 10 15 20
647 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
648 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
649 * not visible and thus causes ENOENT to be returned. We really need
650 * to check record 11 in LEAF1. If it also fails then the search fails
651 * (e.g. it might represent the range 11-16 and thus still not match our
652 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
653 * further iterations.
655 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
656 * and the cursor->create_check TID if an iteration might be needed.
657 * In the above example create_check would be set to 14.
660 hammer_btree_lookup(hammer_cursor_t cursor)
664 cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
665 KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
666 cursor->trans->sync_lock_refs > 0);
667 ++hammer_stats_btree_lookups;
668 if (cursor->flags & HAMMER_CURSOR_ASOF) {
669 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
670 cursor->key_beg.create_tid = cursor->asof;
672 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
673 error = btree_search(cursor, 0);
674 if (error != ENOENT ||
675 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
678 * Stop if error other then ENOENT.
679 * Stop if ENOENT and not special case.
683 if (hammer_debug_btree) {
684 kprintf("CREATE_CHECK %016llx\n",
685 (long long)cursor->create_check);
687 cursor->key_beg.create_tid = cursor->create_check;
691 error = btree_search(cursor, 0);
694 error = hammer_btree_extract(cursor, cursor->flags);
699 * Execute the logic required to start an iteration. The first record
700 * located within the specified range is returned and iteration control
701 * flags are adjusted for successive hammer_btree_iterate() calls.
703 * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
704 * in a loop without worrying about it. Higher-level merged searches will
705 * adjust the flag appropriately.
708 hammer_btree_first(hammer_cursor_t cursor)
712 error = hammer_btree_lookup(cursor);
713 if (error == ENOENT) {
714 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
715 error = hammer_btree_iterate(cursor);
717 cursor->flags |= HAMMER_CURSOR_ATEDISK;
722 * Similarly but for an iteration in the reverse direction.
724 * Set ATEDISK when iterating backwards to skip the current entry,
725 * which after an ENOENT lookup will be pointing beyond our end point.
727 * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
728 * in a loop without worrying about it. Higher-level merged searches will
729 * adjust the flag appropriately.
732 hammer_btree_last(hammer_cursor_t cursor)
734 struct hammer_base_elm save;
737 save = cursor->key_beg;
738 cursor->key_beg = cursor->key_end;
739 error = hammer_btree_lookup(cursor);
740 cursor->key_beg = save;
741 if (error == ENOENT ||
742 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
743 cursor->flags |= HAMMER_CURSOR_ATEDISK;
744 error = hammer_btree_iterate_reverse(cursor);
746 cursor->flags |= HAMMER_CURSOR_ATEDISK;
751 * Extract the record and/or data associated with the cursor's current
752 * position. Any prior record or data stored in the cursor is replaced.
754 * NOTE: All extractions occur at the leaf of the B-Tree.
757 hammer_btree_extract(hammer_cursor_t cursor, int flags)
759 hammer_node_ondisk_t node;
760 hammer_btree_elm_t elm;
761 hammer_off_t data_off;
767 * Certain types of corruption can result in a NULL node pointer.
769 if (cursor->node == NULL) {
770 kprintf("hammer: NULL cursor->node, filesystem might "
771 "have gotten corrupted\n");
776 * The case where the data reference resolves to the same buffer
777 * as the record reference must be handled.
779 node = cursor->node->ondisk;
780 elm = &node->elms[cursor->index];
782 hmp = cursor->node->hmp;
785 * There is nothing to extract for an internal element.
787 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
791 * Only record types have data.
793 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
794 cursor->leaf = &elm->leaf;
796 if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
798 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
800 data_off = elm->leaf.data_offset;
801 data_len = elm->leaf.data_len;
808 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
809 cursor->data = hammer_bread_ext(hmp, data_off, data_len,
810 &error, &cursor->data_buffer);
813 * Mark the data buffer as not being meta-data if it isn't
814 * meta-data (sometimes bulk data is accessed via a volume
818 switch(elm->leaf.base.rec_type) {
819 case HAMMER_RECTYPE_DATA:
820 case HAMMER_RECTYPE_DB:
821 if ((data_off & HAMMER_ZONE_LARGE_DATA) == 0)
823 if (hammer_double_buffer == 0 ||
824 (cursor->flags & HAMMER_CURSOR_NOSWAPCACHE)) {
825 hammer_io_notmeta(cursor->data_buffer);
834 * Deal with CRC errors on the extracted data.
837 hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
838 kprintf("CRC DATA @ %016llx/%d FAILED\n",
839 (long long)elm->leaf.data_offset, elm->leaf.data_len);
840 if (hammer_debug_critical)
841 Debugger("CRC FAILED: DATA");
842 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM)
843 error = EDOM; /* less critical (mirroring) */
845 error = EIO; /* critical */
852 * Insert a leaf element into the B-Tree at the current cursor position.
853 * The cursor is positioned such that the element at and beyond the cursor
854 * are shifted to make room for the new record.
856 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
857 * flag set and that call must return ENOENT before this function can be
858 * called. ENOSPC is returned if there is no room to insert a new record.
860 * The caller may depend on the cursor's exclusive lock after return to
861 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
864 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
867 hammer_node_ondisk_t node;
872 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
874 ++hammer_stats_btree_inserts;
877 * Insert the element at the leaf node and update the count in the
878 * parent. It is possible for parent to be NULL, indicating that
879 * the filesystem's ROOT B-Tree node is a leaf itself, which is
880 * possible. The root inode can never be deleted so the leaf should
883 * Remember that leaf nodes do not have boundaries.
885 hammer_modify_node_all(cursor->trans, cursor->node);
886 node = cursor->node->ondisk;
888 KKASSERT(elm->base.btype != 0);
889 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
890 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
891 if (i != node->count) {
892 bcopy(&node->elms[i], &node->elms[i+1],
893 (node->count - i) * sizeof(*elm));
895 node->elms[i].leaf = *elm;
897 hammer_cursor_inserted_element(cursor->node, i);
900 * Update the leaf node's aggregate mirror_tid for mirroring
903 if (node->mirror_tid < elm->base.delete_tid) {
904 node->mirror_tid = elm->base.delete_tid;
907 if (node->mirror_tid < elm->base.create_tid) {
908 node->mirror_tid = elm->base.create_tid;
911 hammer_modify_node_done(cursor->node);
914 * Debugging sanity checks.
916 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
917 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
919 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
921 if (i != node->count - 1)
922 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
928 * Delete a record from the B-Tree at the current cursor position.
929 * The cursor is positioned such that the current element is the one
932 * On return the cursor will be positioned after the deleted element and
933 * MAY point to an internal node. It will be suitable for the continuation
934 * of an iteration but not for an insertion or deletion.
936 * Deletions will attempt to partially rebalance the B-Tree in an upward
937 * direction, but will terminate rather then deadlock. Empty internal nodes
938 * are never allowed by a deletion which deadlocks may end up giving us an
939 * empty leaf. The pruner will clean up and rebalance the tree.
941 * This function can return EDEADLK, requiring the caller to retry the
942 * operation after clearing the deadlock.
944 * This function will store the number of deleted btree nodes in *ndelete
945 * if ndelete is not NULL.
948 hammer_btree_delete(hammer_cursor_t cursor, int *ndelete)
950 hammer_node_ondisk_t ondisk;
952 hammer_node_t parent __debugvar;
956 KKASSERT (cursor->trans->sync_lock_refs > 0);
959 if ((error = hammer_cursor_upgrade(cursor)) != 0)
961 ++hammer_stats_btree_deletes;
964 * Delete the element from the leaf node.
966 * Remember that leaf nodes do not have boundaries.
969 ondisk = node->ondisk;
972 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
973 KKASSERT(i >= 0 && i < ondisk->count);
974 hammer_modify_node_all(cursor->trans, node);
975 if (i + 1 != ondisk->count) {
976 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
977 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
980 hammer_modify_node_done(node);
981 hammer_cursor_deleted_element(node, i);
984 * Validate local parent
986 if (ondisk->parent) {
987 parent = cursor->parent;
989 KKASSERT(parent != NULL);
990 KKASSERT(parent->node_offset == ondisk->parent);
994 * If the leaf becomes empty it must be detached from the parent,
995 * potentially recursing through to the filesystem root.
997 * This may reposition the cursor at one of the parent's of the
1000 * Ignore deadlock errors, that simply means that btree_remove
1001 * was unable to recurse and had to leave us with an empty leaf.
1003 KKASSERT(cursor->index <= ondisk->count);
1004 if (ondisk->count == 0) {
1005 error = btree_remove(cursor, ndelete);
1006 if (error == EDEADLK)
1011 KKASSERT(cursor->parent == NULL ||
1012 cursor->parent_index < cursor->parent->ondisk->count);
1017 * PRIMARY B-TREE SEARCH SUPPORT PROCEDURE
1019 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
1021 * The search can begin ANYWHERE in the B-Tree. As a first step the search
1022 * iterates up the tree as necessary to properly position itself prior to
1023 * actually doing the sarch.
1025 * INSERTIONS: The search will split full nodes and leaves on its way down
1026 * and guarentee that the leaf it ends up on is not full. If we run out
1027 * of space the search continues to the leaf, but ENOSPC is returned.
1029 * The search is only guarenteed to end up on a leaf if an error code of 0
1030 * is returned, or if inserting and an error code of ENOENT is returned.
1031 * Otherwise it can stop at an internal node. On success a search returns
1034 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
1035 * filesystem, and it is not simple code. Please note the following facts:
1037 * - Internal node recursions have a boundary on the left AND right. The
1038 * right boundary is non-inclusive. The create_tid is a generic part
1039 * of the key for internal nodes.
1041 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
1042 * historical search. ASOF and INSERT are mutually exclusive. When
1043 * doing an as-of lookup btree_search() checks for a right-edge boundary
1044 * case. If while recursing down the left-edge differs from the key
1045 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
1046 * with cursor->create_check. This is used by btree_lookup() to iterate.
1047 * The iteration backwards because as-of searches can wind up going
1048 * down the wrong branch of the B-Tree.
1052 btree_search(hammer_cursor_t cursor, int flags)
1054 hammer_node_ondisk_t node;
1055 hammer_btree_elm_t elm;
1062 flags |= cursor->flags;
1063 ++hammer_stats_btree_searches;
1065 if (hammer_debug_btree) {
1066 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td=%p)\n",
1067 (long long)cursor->node->node_offset,
1069 (long long)cursor->key_beg.obj_id,
1070 cursor->key_beg.rec_type,
1071 (long long)cursor->key_beg.key,
1072 (long long)cursor->key_beg.create_tid,
1073 cursor->key_beg.localization,
1077 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
1078 (long long)cursor->parent->node_offset,
1079 cursor->parent_index,
1080 (long long)cursor->left_bound->obj_id,
1081 (long long)cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
1082 (long long)cursor->right_bound->obj_id,
1083 (long long)cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
1085 &cursor->parent->ondisk->elms[cursor->parent_index],
1086 cursor->right_bound,
1087 &cursor->parent->ondisk->elms[cursor->parent_index+1]
1092 * Move our cursor up the tree until we find a node whos range covers
1093 * the key we are trying to locate.
1095 * The left bound is inclusive, the right bound is non-inclusive.
1096 * It is ok to cursor up too far.
1099 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
1100 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
1101 if (r >= 0 && s < 0)
1103 KKASSERT(cursor->parent);
1104 ++hammer_stats_btree_iterations;
1105 error = hammer_cursor_up(cursor);
1111 * The delete-checks below are based on node, not parent. Set the
1112 * initial delete-check based on the parent.
1115 KKASSERT(cursor->left_bound->create_tid != 1);
1116 cursor->create_check = cursor->left_bound->create_tid - 1;
1117 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1121 * We better have ended up with a node somewhere.
1123 KKASSERT(cursor->node != NULL);
1126 * If we are inserting we can't start at a full node if the parent
1127 * is also full (because there is no way to split the node),
1128 * continue running up the tree until the requirement is satisfied
1129 * or we hit the root of the filesystem.
1131 * (If inserting we aren't doing an as-of search so we don't have
1132 * to worry about create_check).
1134 while (flags & HAMMER_CURSOR_INSERT) {
1135 if (btree_node_is_full(cursor->node->ondisk) == 0)
1137 if (cursor->node->ondisk->parent == 0 ||
1138 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1141 ++hammer_stats_btree_iterations;
1142 error = hammer_cursor_up(cursor);
1143 /* node may have become stale */
1149 * Push down through internal nodes to locate the requested key.
1151 node = cursor->node->ondisk;
1152 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1154 * Scan the node to find the subtree index to push down into.
1155 * We go one-past, then back-up.
1157 * We must proactively remove deleted elements which may
1158 * have been left over from a deadlocked btree_remove().
1160 * The left and right boundaries are included in the loop
1161 * in order to detect edge cases.
1163 * If the separator only differs by create_tid (r == 1)
1164 * and we are doing an as-of search, we may end up going
1165 * down a branch to the left of the one containing the
1166 * desired key. This requires numerous special cases.
1168 ++hammer_stats_btree_iterations;
1169 if (hammer_debug_btree) {
1170 kprintf("SEARCH-I %016llx count=%d\n",
1171 (long long)cursor->node->node_offset,
1176 * Try to shortcut the search before dropping into the
1177 * linear loop. Locate the first node where r <= 1.
1179 i = hammer_btree_search_node(&cursor->key_beg, node);
1180 while (i <= node->count) {
1181 ++hammer_stats_btree_elements;
1182 elm = &node->elms[i];
1183 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1184 if (hammer_debug_btree > 2) {
1185 kprintf(" IELM %p %d r=%d\n",
1186 &node->elms[i], i, r);
1191 KKASSERT(elm->base.create_tid != 1);
1192 cursor->create_check = elm->base.create_tid - 1;
1193 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1197 if (hammer_debug_btree) {
1198 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1203 * These cases occur when the parent's idea of the boundary
1204 * is wider then the child's idea of the boundary, and
1205 * require special handling. If not inserting we can
1206 * terminate the search early for these cases but the
1207 * child's boundaries cannot be unconditionally modified.
1211 * If i == 0 the search terminated to the LEFT of the
1212 * left_boundary but to the RIGHT of the parent's left
1217 elm = &node->elms[0];
1220 * If we aren't inserting we can stop here.
1222 if ((flags & (HAMMER_CURSOR_INSERT |
1223 HAMMER_CURSOR_PRUNING)) == 0) {
1229 * Correct a left-hand boundary mismatch.
1231 * We can only do this if we can upgrade the lock,
1232 * and synchronized as a background cursor (i.e.
1233 * inserting or pruning).
1235 * WARNING: We can only do this if inserting, i.e.
1236 * we are running on the backend.
1238 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1240 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1241 hammer_modify_node_field(cursor->trans, cursor->node,
1243 save = node->elms[0].base.btype;
1244 node->elms[0].base = *cursor->left_bound;
1245 node->elms[0].base.btype = save;
1246 hammer_modify_node_done(cursor->node);
1247 } else if (i == node->count + 1) {
1249 * If i == node->count + 1 the search terminated to
1250 * the RIGHT of the right boundary but to the LEFT
1251 * of the parent's right boundary. If we aren't
1252 * inserting we can stop here.
1254 * Note that the last element in this case is
1255 * elms[i-2] prior to adjustments to 'i'.
1258 if ((flags & (HAMMER_CURSOR_INSERT |
1259 HAMMER_CURSOR_PRUNING)) == 0) {
1265 * Correct a right-hand boundary mismatch.
1266 * (actual push-down record is i-2 prior to
1267 * adjustments to i).
1269 * We can only do this if we can upgrade the lock,
1270 * and synchronized as a background cursor (i.e.
1271 * inserting or pruning).
1273 * WARNING: We can only do this if inserting, i.e.
1274 * we are running on the backend.
1276 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1278 elm = &node->elms[i];
1279 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1280 hammer_modify_node(cursor->trans, cursor->node,
1281 &elm->base, sizeof(elm->base));
1282 elm->base = *cursor->right_bound;
1283 hammer_modify_node_done(cursor->node);
1287 * The push-down index is now i - 1. If we had
1288 * terminated on the right boundary this will point
1289 * us at the last element.
1294 elm = &node->elms[i];
1296 if (hammer_debug_btree) {
1297 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1298 "key=%016llx cre=%016llx lo=%02x\n",
1299 (long long)cursor->node->node_offset,
1301 (long long)elm->internal.base.obj_id,
1302 elm->internal.base.rec_type,
1303 (long long)elm->internal.base.key,
1304 (long long)elm->internal.base.create_tid,
1305 elm->internal.base.localization
1310 * We better have a valid subtree offset.
1312 KKASSERT(elm->internal.subtree_offset != 0);
1315 * Handle insertion and deletion requirements.
1317 * If inserting split full nodes. The split code will
1318 * adjust cursor->node and cursor->index if the current
1319 * index winds up in the new node.
1321 * If inserting and a left or right edge case was detected,
1322 * we cannot correct the left or right boundary and must
1323 * prepend and append an empty leaf node in order to make
1324 * the boundary correction.
1326 * If we run out of space we set enospc but continue on
1329 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1330 if (btree_node_is_full(node)) {
1331 error = btree_split_internal(cursor);
1333 if (error != ENOSPC)
1338 * reload stale pointers
1341 node = cursor->node->ondisk;
1346 * Push down (push into new node, existing node becomes
1347 * the parent) and continue the search.
1349 error = hammer_cursor_down(cursor);
1350 /* node may have become stale */
1353 node = cursor->node->ondisk;
1357 * We are at a leaf, do a linear search of the key array.
1359 * On success the index is set to the matching element and 0
1362 * On failure the index is set to the insertion point and ENOENT
1365 * Boundaries are not stored in leaf nodes, so the index can wind
1366 * up to the left of element 0 (index == 0) or past the end of
1367 * the array (index == node->count). It is also possible that the
1368 * leaf might be empty.
1370 ++hammer_stats_btree_iterations;
1371 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1372 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1373 if (hammer_debug_btree) {
1374 kprintf("SEARCH-L %016llx count=%d\n",
1375 (long long)cursor->node->node_offset,
1380 * Try to shortcut the search before dropping into the
1381 * linear loop. Locate the first node where r <= 1.
1383 i = hammer_btree_search_node(&cursor->key_beg, node);
1384 while (i < node->count) {
1385 ++hammer_stats_btree_elements;
1386 elm = &node->elms[i];
1388 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1390 if (hammer_debug_btree > 1)
1391 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1394 * We are at a record element. Stop if we've flipped past
1395 * key_beg, not counting the create_tid test. Allow the
1396 * r == 1 case (key_beg > element but differs only by its
1397 * create_tid) to fall through to the AS-OF check.
1399 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1409 * Check our as-of timestamp against the element.
1411 if (flags & HAMMER_CURSOR_ASOF) {
1412 if (hammer_btree_chkts(cursor->asof,
1413 &node->elms[i].base) != 0) {
1419 if (r > 0) { /* can only be +1 */
1427 if (hammer_debug_btree) {
1428 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1429 (long long)cursor->node->node_offset, i);
1435 * The search of the leaf node failed. i is the insertion point.
1438 if (hammer_debug_btree) {
1439 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1440 (long long)cursor->node->node_offset, i);
1444 * No exact match was found, i is now at the insertion point.
1446 * If inserting split a full leaf before returning. This
1447 * may have the side effect of adjusting cursor->node and
1451 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1452 btree_node_is_full(node)) {
1453 error = btree_split_leaf(cursor);
1455 if (error != ENOSPC)
1460 * reload stale pointers
1464 node = &cursor->node->internal;
1469 * We reached a leaf but did not find the key we were looking for.
1470 * If this is an insert we will be properly positioned for an insert
1471 * (ENOENT) or unable to insert (ENOSPC).
1473 error = enospc ? ENOSPC : ENOENT;
1479 * Heuristical search for the first element whos comparison is <= 1. May
1480 * return an index whos compare result is > 1 but may only return an index
1481 * whos compare result is <= 1 if it is the first element with that result.
1484 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1492 * Don't bother if the node does not have very many elements
1497 i = b + (s - b) / 2;
1498 ++hammer_stats_btree_elements;
1499 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1510 /************************************************************************
1511 * SPLITTING AND MERGING *
1512 ************************************************************************
1514 * These routines do all the dirty work required to split and merge nodes.
1518 * Split an internal node into two nodes and move the separator at the split
1519 * point to the parent.
1521 * (cursor->node, cursor->index) indicates the element the caller intends
1522 * to push into. We will adjust node and index if that element winds
1523 * up in the split node.
1525 * If we are at the root of the filesystem a new root must be created with
1526 * two elements, one pointing to the original root and one pointing to the
1527 * newly allocated split node.
1531 btree_split_internal(hammer_cursor_t cursor)
1533 hammer_node_ondisk_t ondisk;
1535 hammer_node_t parent;
1536 hammer_node_t new_node;
1537 hammer_btree_elm_t elm;
1538 hammer_btree_elm_t parent_elm;
1539 struct hammer_node_lock lockroot;
1540 hammer_mount_t hmp = cursor->trans->hmp;
1546 const int esize = sizeof(*elm);
1548 hammer_node_lock_init(&lockroot, cursor->node);
1549 error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
1552 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1554 ++hammer_stats_btree_splits;
1557 * Calculate the split point. If the insertion point is at the
1558 * end of the leaf we adjust the split point significantly to the
1559 * right to try to optimize node fill and flag it. If we hit
1560 * that same leaf again our heuristic failed and we don't try
1561 * to optimize node fill (it could lead to a degenerate case).
1563 node = cursor->node;
1564 ondisk = node->ondisk;
1565 KKASSERT(ondisk->count > 4);
1566 if (cursor->index == ondisk->count &&
1567 (node->flags & HAMMER_NODE_NONLINEAR) == 0) {
1568 split = (ondisk->count + 1) * 3 / 4;
1569 node->flags |= HAMMER_NODE_NONLINEAR;
1572 * We are splitting but elms[split] will be promoted to
1573 * the parent, leaving the right hand node with one less
1574 * element. If the insertion point will be on the
1575 * left-hand side adjust the split point to give the
1576 * right hand side one additional node.
1578 split = (ondisk->count + 1) / 2;
1579 if (cursor->index <= split)
1584 * If we are at the root of the filesystem, create a new root node
1585 * with 1 element and split normally. Avoid making major
1586 * modifications until we know the whole operation will work.
1588 if (ondisk->parent == 0) {
1589 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1592 hammer_lock_ex(&parent->lock);
1593 hammer_modify_node_noundo(cursor->trans, parent);
1594 ondisk = parent->ondisk;
1597 ondisk->mirror_tid = node->ondisk->mirror_tid;
1598 ondisk->signature = node->ondisk->signature;
1599 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1600 ondisk->elms[0].base = hmp->root_btree_beg;
1601 ondisk->elms[0].base.btype = node->ondisk->type;
1602 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1603 ondisk->elms[0].internal.mirror_tid = ondisk->mirror_tid;
1604 ondisk->elms[1].base = hmp->root_btree_end;
1605 hammer_modify_node_done(parent);
1607 parent_index = 0; /* index of current node in parent */
1610 parent = cursor->parent;
1611 parent_index = cursor->parent_index;
1615 * Split node into new_node at the split point.
1617 * B O O O P N N B <-- P = node->elms[split] (index 4)
1618 * 0 1 2 3 4 5 6 <-- subtree indices
1623 * B O O O B B N N B <--- inner boundary points are 'P'
1626 new_node = hammer_alloc_btree(cursor->trans, 0, &error);
1627 if (new_node == NULL) {
1629 hammer_unlock(&parent->lock);
1630 hammer_delete_node(cursor->trans, parent);
1631 hammer_rel_node(parent);
1635 hammer_lock_ex(&new_node->lock);
1638 * Create the new node. P becomes the left-hand boundary in the
1639 * new node. Copy the right-hand boundary as well.
1641 * elm is the new separator.
1643 hammer_modify_node_noundo(cursor->trans, new_node);
1644 hammer_modify_node_all(cursor->trans, node);
1645 ondisk = node->ondisk;
1646 elm = &ondisk->elms[split];
1647 bcopy(elm, &new_node->ondisk->elms[0],
1648 (ondisk->count - split + 1) * esize);
1649 new_node->ondisk->count = ondisk->count - split;
1650 new_node->ondisk->parent = parent->node_offset;
1651 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1652 new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1653 KKASSERT(ondisk->type == new_node->ondisk->type);
1654 hammer_cursor_split_node(node, new_node, split);
1657 * Cleanup the original node. Elm (P) becomes the new boundary,
1658 * its subtree_offset was moved to the new node. If we had created
1659 * a new root its parent pointer may have changed.
1661 elm->internal.subtree_offset = 0;
1662 ondisk->count = split;
1665 * Insert the separator into the parent, fixup the parent's
1666 * reference to the original node, and reference the new node.
1667 * The separator is P.
1669 * Remember that base.count does not include the right-hand boundary.
1671 hammer_modify_node_all(cursor->trans, parent);
1672 ondisk = parent->ondisk;
1673 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1674 parent_elm = &ondisk->elms[parent_index+1];
1675 bcopy(parent_elm, parent_elm + 1,
1676 (ondisk->count - parent_index) * esize);
1677 parent_elm->internal.base = elm->base; /* separator P */
1678 parent_elm->internal.base.btype = new_node->ondisk->type;
1679 parent_elm->internal.subtree_offset = new_node->node_offset;
1680 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1682 hammer_modify_node_done(parent);
1683 hammer_cursor_inserted_element(parent, parent_index + 1);
1686 * The children of new_node need their parent pointer set to new_node.
1687 * The children have already been locked by
1688 * hammer_btree_lock_children().
1690 for (i = 0; i < new_node->ondisk->count; ++i) {
1691 elm = &new_node->ondisk->elms[i];
1692 error = btree_set_parent(cursor->trans, new_node, elm);
1694 panic("btree_split_internal: btree-fixup problem");
1697 hammer_modify_node_done(new_node);
1700 * The filesystem's root B-Tree pointer may have to be updated.
1703 hammer_volume_t volume;
1705 volume = hammer_get_root_volume(hmp, &error);
1706 KKASSERT(error == 0);
1708 hammer_modify_volume_field(cursor->trans, volume,
1710 volume->ondisk->vol0_btree_root = parent->node_offset;
1711 hammer_modify_volume_done(volume);
1712 node->ondisk->parent = parent->node_offset;
1713 /* node->ondisk->signature = 0; */
1714 if (cursor->parent) {
1715 hammer_unlock(&cursor->parent->lock);
1716 hammer_rel_node(cursor->parent);
1718 cursor->parent = parent; /* lock'd and ref'd */
1719 hammer_rel_volume(volume, 0);
1721 hammer_modify_node_done(node);
1724 * Ok, now adjust the cursor depending on which element the original
1725 * index was pointing at. If we are >= the split point the push node
1726 * is now in the new node.
1728 * NOTE: If we are at the split point itself we cannot stay with the
1729 * original node because the push index will point at the right-hand
1730 * boundary, which is illegal.
1732 * NOTE: The cursor's parent or parent_index must be adjusted for
1733 * the case where a new parent (new root) was created, and the case
1734 * where the cursor is now pointing at the split node.
1736 if (cursor->index >= split) {
1737 cursor->parent_index = parent_index + 1;
1738 cursor->index -= split;
1739 hammer_unlock(&cursor->node->lock);
1740 hammer_rel_node(cursor->node);
1741 cursor->node = new_node; /* locked and ref'd */
1743 cursor->parent_index = parent_index;
1744 hammer_unlock(&new_node->lock);
1745 hammer_rel_node(new_node);
1749 * Fixup left and right bounds
1751 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1752 cursor->left_bound = &parent_elm[0].internal.base;
1753 cursor->right_bound = &parent_elm[1].internal.base;
1754 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1755 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1756 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1757 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1760 hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
1761 hammer_cursor_downgrade(cursor);
1766 * Same as the above, but splits a full leaf node.
1770 btree_split_leaf(hammer_cursor_t cursor)
1772 hammer_node_ondisk_t ondisk;
1773 hammer_node_t parent;
1776 hammer_node_t new_leaf;
1777 hammer_btree_elm_t elm;
1778 hammer_btree_elm_t parent_elm;
1779 hammer_base_elm_t mid_boundary;
1784 const size_t esize = sizeof(*elm);
1786 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1788 ++hammer_stats_btree_splits;
1790 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1791 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1792 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1793 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1796 * Calculate the split point. If the insertion point is at the
1797 * end of the leaf we adjust the split point significantly to the
1798 * right to try to optimize node fill and flag it. If we hit
1799 * that same leaf again our heuristic failed and we don't try
1800 * to optimize node fill (it could lead to a degenerate case).
1802 leaf = cursor->node;
1803 ondisk = leaf->ondisk;
1804 KKASSERT(ondisk->count > 4);
1805 if (cursor->index == ondisk->count &&
1806 (leaf->flags & HAMMER_NODE_NONLINEAR) == 0) {
1807 split = (ondisk->count + 1) * 3 / 4;
1808 leaf->flags |= HAMMER_NODE_NONLINEAR;
1810 split = (ondisk->count + 1) / 2;
1815 * If the insertion point is at the split point shift the
1816 * split point left so we don't have to worry about
1818 if (cursor->index == split)
1821 KKASSERT(split > 0 && split < ondisk->count);
1826 elm = &ondisk->elms[split];
1828 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1829 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1830 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1831 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1834 * If we are at the root of the tree, create a new root node with
1835 * 1 element and split normally. Avoid making major modifications
1836 * until we know the whole operation will work.
1838 if (ondisk->parent == 0) {
1839 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1842 hammer_lock_ex(&parent->lock);
1843 hammer_modify_node_noundo(cursor->trans, parent);
1844 ondisk = parent->ondisk;
1847 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1848 ondisk->signature = leaf->ondisk->signature;
1849 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1850 ondisk->elms[0].base = hmp->root_btree_beg;
1851 ondisk->elms[0].base.btype = leaf->ondisk->type;
1852 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1853 ondisk->elms[0].internal.mirror_tid = ondisk->mirror_tid;
1854 ondisk->elms[1].base = hmp->root_btree_end;
1855 hammer_modify_node_done(parent);
1857 parent_index = 0; /* insertion point in parent */
1860 parent = cursor->parent;
1861 parent_index = cursor->parent_index;
1865 * Split leaf into new_leaf at the split point. Select a separator
1866 * value in-between the two leafs but with a bent towards the right
1867 * leaf since comparisons use an 'elm >= separator' inequality.
1876 new_leaf = hammer_alloc_btree(cursor->trans, 0, &error);
1877 if (new_leaf == NULL) {
1879 hammer_unlock(&parent->lock);
1880 hammer_delete_node(cursor->trans, parent);
1881 hammer_rel_node(parent);
1885 hammer_lock_ex(&new_leaf->lock);
1888 * Create the new node and copy the leaf elements from the split
1889 * point on to the new node.
1891 hammer_modify_node_all(cursor->trans, leaf);
1892 hammer_modify_node_noundo(cursor->trans, new_leaf);
1893 ondisk = leaf->ondisk;
1894 elm = &ondisk->elms[split];
1895 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1896 new_leaf->ondisk->count = ondisk->count - split;
1897 new_leaf->ondisk->parent = parent->node_offset;
1898 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1899 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1900 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1901 hammer_modify_node_done(new_leaf);
1902 hammer_cursor_split_node(leaf, new_leaf, split);
1905 * Cleanup the original node. Because this is a leaf node and
1906 * leaf nodes do not have a right-hand boundary, there
1907 * aren't any special edge cases to clean up. We just fixup the
1910 ondisk->count = split;
1913 * Insert the separator into the parent, fixup the parent's
1914 * reference to the original node, and reference the new node.
1915 * The separator is P.
1917 * Remember that base.count does not include the right-hand boundary.
1918 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1920 hammer_modify_node_all(cursor->trans, parent);
1921 ondisk = parent->ondisk;
1922 KKASSERT(split != 0);
1923 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1924 parent_elm = &ondisk->elms[parent_index+1];
1925 bcopy(parent_elm, parent_elm + 1,
1926 (ondisk->count - parent_index) * esize);
1928 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1929 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1930 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1931 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1932 mid_boundary = &parent_elm->base;
1934 hammer_modify_node_done(parent);
1935 hammer_cursor_inserted_element(parent, parent_index + 1);
1938 * The filesystem's root B-Tree pointer may have to be updated.
1941 hammer_volume_t volume;
1943 volume = hammer_get_root_volume(hmp, &error);
1944 KKASSERT(error == 0);
1946 hammer_modify_volume_field(cursor->trans, volume,
1948 volume->ondisk->vol0_btree_root = parent->node_offset;
1949 hammer_modify_volume_done(volume);
1950 leaf->ondisk->parent = parent->node_offset;
1951 /* leaf->ondisk->signature = 0; */
1952 if (cursor->parent) {
1953 hammer_unlock(&cursor->parent->lock);
1954 hammer_rel_node(cursor->parent);
1956 cursor->parent = parent; /* lock'd and ref'd */
1957 hammer_rel_volume(volume, 0);
1959 hammer_modify_node_done(leaf);
1962 * Ok, now adjust the cursor depending on which element the original
1963 * index was pointing at. If we are >= the split point the push node
1964 * is now in the new node.
1966 * NOTE: If we are at the split point itself we need to select the
1967 * old or new node based on where key_beg's insertion point will be.
1968 * If we pick the wrong side the inserted element will wind up in
1969 * the wrong leaf node and outside that node's bounds.
1971 if (cursor->index > split ||
1972 (cursor->index == split &&
1973 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1974 cursor->parent_index = parent_index + 1;
1975 cursor->index -= split;
1976 hammer_unlock(&cursor->node->lock);
1977 hammer_rel_node(cursor->node);
1978 cursor->node = new_leaf;
1980 cursor->parent_index = parent_index;
1981 hammer_unlock(&new_leaf->lock);
1982 hammer_rel_node(new_leaf);
1986 * Fixup left and right bounds
1988 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1989 cursor->left_bound = &parent_elm[0].internal.base;
1990 cursor->right_bound = &parent_elm[1].internal.base;
1993 * Assert that the bounds are correct.
1995 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1996 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1997 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1998 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1999 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
2000 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
2003 hammer_cursor_downgrade(cursor);
2010 * Recursively correct the right-hand boundary's create_tid to (tid) as
2011 * long as the rest of the key matches. We have to recurse upward in
2012 * the tree as well as down the left side of each parent's right node.
2014 * Return EDEADLK if we were only partially successful, forcing the caller
2015 * to try again. The original cursor is not modified. This routine can
2016 * also fail with EDEADLK if it is forced to throw away a portion of its
2019 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
2022 TAILQ_ENTRY(hammer_rhb) entry;
2027 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
2030 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
2032 struct hammer_mount *hmp;
2033 struct hammer_rhb_list rhb_list;
2034 hammer_base_elm_t elm;
2035 hammer_node_t orig_node;
2036 struct hammer_rhb *rhb;
2040 TAILQ_INIT(&rhb_list);
2041 hmp = cursor->trans->hmp;
2044 * Save our position so we can restore it on return. This also
2045 * gives us a stable 'elm'.
2047 orig_node = cursor->node;
2048 hammer_ref_node(orig_node);
2049 hammer_lock_sh(&orig_node->lock);
2050 orig_index = cursor->index;
2051 elm = &orig_node->ondisk->elms[orig_index].base;
2054 * Now build a list of parents going up, allocating a rhb
2055 * structure for each one.
2057 while (cursor->parent) {
2059 * Stop if we no longer have any right-bounds to fix up
2061 if (elm->obj_id != cursor->right_bound->obj_id ||
2062 elm->rec_type != cursor->right_bound->rec_type ||
2063 elm->key != cursor->right_bound->key) {
2068 * Stop if the right-hand bound's create_tid does not
2069 * need to be corrected.
2071 if (cursor->right_bound->create_tid >= tid)
2074 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2075 rhb->node = cursor->parent;
2076 rhb->index = cursor->parent_index;
2077 hammer_ref_node(rhb->node);
2078 hammer_lock_sh(&rhb->node->lock);
2079 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2081 hammer_cursor_up(cursor);
2085 * now safely adjust the right hand bound for each rhb. This may
2086 * also require taking the right side of the tree and iterating down
2090 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2091 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2094 TAILQ_REMOVE(&rhb_list, rhb, entry);
2095 hammer_unlock(&rhb->node->lock);
2096 hammer_rel_node(rhb->node);
2097 kfree(rhb, hmp->m_misc);
2099 switch (cursor->node->ondisk->type) {
2100 case HAMMER_BTREE_TYPE_INTERNAL:
2102 * Right-boundary for parent at internal node
2103 * is one element to the right of the element whos
2104 * right boundary needs adjusting. We must then
2105 * traverse down the left side correcting any left
2106 * bounds (which may now be too far to the left).
2109 error = hammer_btree_correct_lhb(cursor, tid);
2112 panic("hammer_btree_correct_rhb(): Bad node type");
2121 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
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 error = hammer_cursor_seek(cursor, orig_node, orig_index);
2128 hammer_unlock(&orig_node->lock);
2129 hammer_rel_node(orig_node);
2134 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2135 * bound going downward starting at the current cursor position.
2137 * This function does not restore the cursor after use.
2140 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
2142 struct hammer_rhb_list rhb_list;
2143 hammer_base_elm_t elm;
2144 hammer_base_elm_t cmp;
2145 struct hammer_rhb *rhb;
2146 struct hammer_mount *hmp;
2149 TAILQ_INIT(&rhb_list);
2150 hmp = cursor->trans->hmp;
2152 cmp = &cursor->node->ondisk->elms[cursor->index].base;
2155 * Record the node and traverse down the left-hand side for all
2156 * matching records needing a boundary correction.
2160 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2161 rhb->node = cursor->node;
2162 rhb->index = cursor->index;
2163 hammer_ref_node(rhb->node);
2164 hammer_lock_sh(&rhb->node->lock);
2165 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2167 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2169 * Nothing to traverse down if we are at the right
2170 * boundary of an internal node.
2172 if (cursor->index == cursor->node->ondisk->count)
2175 elm = &cursor->node->ondisk->elms[cursor->index].base;
2176 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2178 panic("Illegal leaf record type %02x", elm->btype);
2180 error = hammer_cursor_down(cursor);
2184 elm = &cursor->node->ondisk->elms[cursor->index].base;
2185 if (elm->obj_id != cmp->obj_id ||
2186 elm->rec_type != cmp->rec_type ||
2187 elm->key != cmp->key) {
2190 if (elm->create_tid >= tid)
2196 * Now we can safely adjust the left-hand boundary from the bottom-up.
2197 * The last element we remove from the list is the caller's right hand
2198 * boundary, which must also be adjusted.
2200 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2201 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2204 TAILQ_REMOVE(&rhb_list, rhb, entry);
2205 hammer_unlock(&rhb->node->lock);
2206 hammer_rel_node(rhb->node);
2207 kfree(rhb, hmp->m_misc);
2209 elm = &cursor->node->ondisk->elms[cursor->index].base;
2210 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2211 hammer_modify_node(cursor->trans, cursor->node,
2213 sizeof(elm->create_tid));
2214 elm->create_tid = tid;
2215 hammer_modify_node_done(cursor->node);
2217 panic("hammer_btree_correct_lhb(): Bad element type");
2224 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2225 TAILQ_REMOVE(&rhb_list, rhb, entry);
2226 hammer_unlock(&rhb->node->lock);
2227 hammer_rel_node(rhb->node);
2228 kfree(rhb, hmp->m_misc);
2236 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2237 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2238 * the operation due to a deadlock, or some other error.
2240 * This routine is initially called with an empty leaf and may be
2241 * recursively called with single-element internal nodes.
2243 * It should also be noted that when removing empty leaves we must be sure
2244 * to test and update mirror_tid because another thread may have deadlocked
2245 * against us (or someone) trying to propagate it up and cannot retry once
2246 * the node has been deleted.
2248 * On return the cursor may end up pointing to an internal node, suitable
2249 * for further iteration but not for an immediate insertion or deletion.
2252 btree_remove(hammer_cursor_t cursor, int *ndelete)
2254 hammer_node_ondisk_t ondisk;
2255 hammer_btree_elm_t elm;
2257 hammer_node_t parent;
2258 const int esize = sizeof(*elm);
2261 node = cursor->node;
2264 * When deleting the root of the filesystem convert it to
2265 * an empty leaf node. Internal nodes cannot be empty.
2267 ondisk = node->ondisk;
2268 if (ondisk->parent == 0) {
2269 KKASSERT(cursor->parent == NULL);
2270 hammer_modify_node_all(cursor->trans, node);
2271 KKASSERT(ondisk == node->ondisk);
2272 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2274 hammer_modify_node_done(node);
2279 parent = cursor->parent;
2282 * Attempt to remove the parent's reference to the child. If the
2283 * parent would become empty we have to recurse. If we fail we
2284 * leave the parent pointing to an empty leaf node.
2286 * We have to recurse successfully before we can delete the internal
2287 * node as it is illegal to have empty internal nodes. Even though
2288 * the operation may be aborted we must still fixup any unlocked
2289 * cursors as if we had deleted the element prior to recursing
2290 * (by calling hammer_cursor_deleted_element()) so those cursors
2291 * are properly forced up the chain by the recursion.
2293 if (parent->ondisk->count == 1) {
2295 * This special cursor_up_locked() call leaves the original
2296 * node exclusively locked and referenced, leaves the
2297 * original parent locked (as the new node), and locks the
2298 * new parent. It can return EDEADLK.
2300 * We cannot call hammer_cursor_removed_node() until we are
2301 * actually able to remove the node. If we did then tracked
2302 * cursors in the middle of iterations could be repointed
2303 * to a parent node. If this occurs they could end up
2304 * scanning newly inserted records into the node (that could
2305 * not be deleted) when they push down again.
2307 * Due to the way the recursion works the final parent is left
2308 * in cursor->parent after the recursion returns. Each
2309 * layer on the way back up is thus able to call
2310 * hammer_cursor_removed_node() and 'jump' the node up to
2311 * the (same) final parent.
2313 * NOTE! The local variable 'parent' is invalid after we
2314 * call hammer_cursor_up_locked().
2316 error = hammer_cursor_up_locked(cursor);
2320 hammer_cursor_deleted_element(cursor->node, 0);
2321 error = btree_remove(cursor, ndelete);
2323 KKASSERT(node != cursor->node);
2324 hammer_cursor_removed_node(
2327 hammer_modify_node_all(cursor->trans, node);
2328 ondisk = node->ondisk;
2329 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2331 hammer_modify_node_done(node);
2332 hammer_flush_node(node, 0);
2333 hammer_delete_node(cursor->trans, node);
2338 * Defer parent removal because we could not
2339 * get the lock, just let the leaf remain
2344 hammer_unlock(&node->lock);
2345 hammer_rel_node(node);
2348 * Defer parent removal because we could not
2349 * get the lock, just let the leaf remain
2355 KKASSERT(parent->ondisk->count > 1);
2357 hammer_modify_node_all(cursor->trans, parent);
2358 ondisk = parent->ondisk;
2359 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2361 elm = &ondisk->elms[cursor->parent_index];
2362 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2363 KKASSERT(ondisk->count > 0);
2366 * We must retain the highest mirror_tid. The deleted
2367 * range is now encompassed by the element to the left.
2368 * If we are already at the left edge the new left edge
2369 * inherits mirror_tid.
2371 * Note that bounds of the parent to our parent may create
2372 * a gap to the left of our left-most node or to the right
2373 * of our right-most node. The gap is silently included
2374 * in the mirror_tid's area of effect from the point of view
2377 if (cursor->parent_index) {
2378 if (elm[-1].internal.mirror_tid <
2379 elm[0].internal.mirror_tid) {
2380 elm[-1].internal.mirror_tid =
2381 elm[0].internal.mirror_tid;
2384 if (elm[1].internal.mirror_tid <
2385 elm[0].internal.mirror_tid) {
2386 elm[1].internal.mirror_tid =
2387 elm[0].internal.mirror_tid;
2392 * Delete the subtree reference in the parent. Include
2393 * boundary element at end.
2395 bcopy(&elm[1], &elm[0],
2396 (ondisk->count - cursor->parent_index) * esize);
2398 hammer_modify_node_done(parent);
2399 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2400 hammer_cursor_deleted_element(parent, cursor->parent_index);
2401 hammer_flush_node(node, 0);
2402 hammer_delete_node(cursor->trans, node);
2405 * cursor->node is invalid, cursor up to make the cursor
2406 * valid again. We have to flag the condition in case
2407 * another thread wiggles an insertion in during an
2410 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK;
2411 error = hammer_cursor_up(cursor);
2419 * Propagate cursor->trans->tid up the B-Tree starting at the current
2420 * cursor position using pseudofs info gleaned from the passed inode.
2422 * The passed inode has no relationship to the cursor position other
2423 * then being in the same pseudofs as the insertion or deletion we
2424 * are propagating the mirror_tid for.
2426 * WARNING! Because we push and pop the passed cursor, it may be
2427 * modified by other B-Tree operations while it is unlocked
2428 * and things like the node & leaf pointers, and indexes might
2432 hammer_btree_do_propagation(hammer_cursor_t cursor,
2433 hammer_pseudofs_inmem_t pfsm,
2434 hammer_btree_leaf_elm_t leaf)
2436 hammer_cursor_t ncursor;
2437 hammer_tid_t mirror_tid;
2438 int error __debugvar;
2441 * We do not propagate a mirror_tid if the filesystem was mounted
2442 * in no-mirror mode.
2444 if (cursor->trans->hmp->master_id < 0)
2448 * This is a bit of a hack because we cannot deadlock or return
2449 * EDEADLK here. The related operation has already completed and
2450 * we must propagate the mirror_tid now regardless.
2452 * Generate a new cursor which inherits the original's locks and
2453 * unlock the original. Use the new cursor to propagate the
2454 * mirror_tid. Then clean up the new cursor and reacquire locks
2457 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2458 * original's locks and the original is tracked and must be
2461 mirror_tid = cursor->node->ondisk->mirror_tid;
2462 KKASSERT(mirror_tid != 0);
2463 ncursor = hammer_push_cursor(cursor);
2464 error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2465 KKASSERT(error == 0);
2466 hammer_pop_cursor(cursor, ncursor);
2467 /* WARNING: cursor's leaf pointer may change after pop */
2472 * Propagate a mirror TID update upwards through the B-Tree to the root.
2474 * A locked internal node must be passed in. The node will remain locked
2477 * This function syncs mirror_tid at the specified internal node's element,
2478 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2481 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2483 hammer_btree_internal_elm_t elm;
2488 error = hammer_cursor_up(cursor);
2490 error = hammer_cursor_upgrade(cursor);
2493 * We can ignore HAMMER_CURSOR_ITERATE_CHECK, the
2494 * cursor will still be properly positioned for
2495 * mirror propagation, just not for iterations.
2497 while (error == EDEADLK) {
2498 hammer_recover_cursor(cursor);
2499 error = hammer_cursor_upgrade(cursor);
2505 * If the cursor deadlocked it could end up at a leaf
2506 * after we lost the lock.
2508 node = cursor->node;
2509 if (node->ondisk->type != HAMMER_BTREE_TYPE_INTERNAL)
2513 * Adjust the node's element
2515 elm = &node->ondisk->elms[cursor->index].internal;
2516 if (elm->mirror_tid >= mirror_tid)
2518 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2519 sizeof(elm->mirror_tid));
2520 elm->mirror_tid = mirror_tid;
2521 hammer_modify_node_done(node);
2522 if (hammer_debug_general & 0x0002) {
2523 kprintf("mirror_propagate: propagate "
2524 "%016llx @%016llx:%d\n",
2525 (long long)mirror_tid,
2526 (long long)node->node_offset,
2532 * Adjust the node's mirror_tid aggregator
2534 if (node->ondisk->mirror_tid >= mirror_tid)
2536 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2537 node->ondisk->mirror_tid = mirror_tid;
2538 hammer_modify_node_done(node);
2539 if (hammer_debug_general & 0x0002) {
2540 kprintf("mirror_propagate: propagate "
2541 "%016llx @%016llx\n",
2542 (long long)mirror_tid,
2543 (long long)node->node_offset);
2546 if (error == ENOENT)
2552 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2553 int *parent_indexp, int *errorp, int try_exclusive)
2555 hammer_node_t parent;
2556 hammer_btree_elm_t elm;
2562 parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2564 KKASSERT(parent == NULL);
2567 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2572 if (try_exclusive) {
2573 if (hammer_lock_ex_try(&parent->lock)) {
2574 hammer_rel_node(parent);
2579 hammer_lock_sh(&parent->lock);
2583 * Figure out which element in the parent is pointing to the
2586 if (node->ondisk->count) {
2587 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2592 while (i < parent->ondisk->count) {
2593 elm = &parent->ondisk->elms[i];
2594 if (elm->internal.subtree_offset == node->node_offset)
2598 if (i == parent->ondisk->count) {
2599 hammer_unlock(&parent->lock);
2600 panic("Bad B-Tree link: parent %p node %p", parent, node);
2603 KKASSERT(*errorp == 0);
2608 * The element (elm) has been moved to a new internal node (node).
2610 * If the element represents a pointer to an internal node that node's
2611 * parent must be adjusted to the element's new location.
2613 * XXX deadlock potential here with our exclusive locks
2616 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2617 hammer_btree_elm_t elm)
2619 hammer_node_t child;
2624 switch(elm->base.btype) {
2625 case HAMMER_BTREE_TYPE_INTERNAL:
2626 case HAMMER_BTREE_TYPE_LEAF:
2627 child = hammer_get_node(trans, elm->internal.subtree_offset,
2630 hammer_modify_node_field(trans, child, parent);
2631 child->ondisk->parent = node->node_offset;
2632 hammer_modify_node_done(child);
2633 hammer_rel_node(child);
2643 * Initialize the root of a recursive B-Tree node lock list structure.
2646 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2648 TAILQ_INIT(&parent->list);
2649 parent->parent = NULL;
2650 parent->node = node;
2652 parent->count = node->ondisk->count;
2653 parent->copy = NULL;
2658 * Initialize a cache of hammer_node_lock's including space allocated
2661 * This is used by the rebalancing code to preallocate the copy space
2662 * for ~4096 B-Tree nodes (16MB of data) prior to acquiring any HAMMER
2663 * locks, otherwise we can blow out the pageout daemon's emergency
2664 * reserve and deadlock it.
2666 * NOTE: HAMMER_NODE_LOCK_LCACHE is not set on items cached in the lcache.
2667 * The flag is set when the item is pulled off the cache for use.
2670 hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
2673 hammer_node_lock_t item;
2676 for (count = 1; depth; --depth)
2677 count *= HAMMER_BTREE_LEAF_ELMS;
2678 bzero(lcache, sizeof(*lcache));
2679 TAILQ_INIT(&lcache->list);
2681 item = kmalloc(sizeof(*item), hmp->m_misc, M_WAITOK|M_ZERO);
2682 item->copy = kmalloc(sizeof(*item->copy),
2683 hmp->m_misc, M_WAITOK);
2684 TAILQ_INIT(&item->list);
2685 TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2691 hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache)
2693 hammer_node_lock_t item;
2695 while ((item = TAILQ_FIRST(&lcache->list)) != NULL) {
2696 TAILQ_REMOVE(&lcache->list, item, entry);
2697 KKASSERT(item->copy);
2698 KKASSERT(TAILQ_EMPTY(&item->list));
2699 kfree(item->copy, hmp->m_misc);
2700 kfree(item, hmp->m_misc);
2702 KKASSERT(lcache->copy == NULL);
2706 * Exclusively lock all the children of node. This is used by the split
2707 * code to prevent anyone from accessing the children of a cursor node
2708 * while we fix-up its parent offset.
2710 * If we don't lock the children we can really mess up cursors which block
2711 * trying to cursor-up into our node.
2713 * On failure EDEADLK (or some other error) is returned. If a deadlock
2714 * error is returned the cursor is adjusted to block on termination.
2716 * The caller is responsible for managing parent->node, the root's node
2717 * is usually aliased from a cursor.
2720 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2721 hammer_node_lock_t parent,
2722 hammer_node_lock_t lcache)
2725 hammer_node_lock_t item;
2726 hammer_node_ondisk_t ondisk;
2727 hammer_btree_elm_t elm;
2728 hammer_node_t child;
2729 struct hammer_mount *hmp;
2733 node = parent->node;
2734 ondisk = node->ondisk;
2736 hmp = cursor->trans->hmp;
2739 * We really do not want to block on I/O with exclusive locks held,
2740 * pre-get the children before trying to lock the mess. This is
2741 * only done one-level deep for now.
2743 for (i = 0; i < ondisk->count; ++i) {
2744 ++hammer_stats_btree_elements;
2745 elm = &ondisk->elms[i];
2746 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2747 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2750 child = hammer_get_node(cursor->trans,
2751 elm->internal.subtree_offset,
2754 hammer_rel_node(child);
2760 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2761 ++hammer_stats_btree_elements;
2762 elm = &ondisk->elms[i];
2764 switch(elm->base.btype) {
2765 case HAMMER_BTREE_TYPE_INTERNAL:
2766 case HAMMER_BTREE_TYPE_LEAF:
2767 KKASSERT(elm->internal.subtree_offset != 0);
2768 child = hammer_get_node(cursor->trans,
2769 elm->internal.subtree_offset,
2777 if (hammer_lock_ex_try(&child->lock) != 0) {
2778 if (cursor->deadlk_node == NULL) {
2779 cursor->deadlk_node = child;
2780 hammer_ref_node(cursor->deadlk_node);
2783 hammer_rel_node(child);
2786 item = TAILQ_FIRST(&lcache->list);
2787 KKASSERT(item != NULL);
2788 item->flags |= HAMMER_NODE_LOCK_LCACHE;
2789 TAILQ_REMOVE(&lcache->list,
2792 item = kmalloc(sizeof(*item),
2795 TAILQ_INIT(&item->list);
2798 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2799 item->parent = parent;
2802 item->count = child->ondisk->count;
2805 * Recurse (used by the rebalancing code)
2807 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2808 error = hammer_btree_lock_children(
2818 hammer_btree_unlock_children(hmp, parent, lcache);
2823 * Create an in-memory copy of all B-Tree nodes listed, recursively,
2824 * including the parent.
2827 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2829 hammer_mount_t hmp = cursor->trans->hmp;
2830 hammer_node_lock_t item;
2832 if (parent->copy == NULL) {
2833 KKASSERT((parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0);
2834 parent->copy = kmalloc(sizeof(*parent->copy),
2835 hmp->m_misc, M_WAITOK);
2837 KKASSERT((parent->flags & HAMMER_NODE_LOCK_UPDATED) == 0);
2838 *parent->copy = *parent->node->ondisk;
2839 TAILQ_FOREACH(item, &parent->list, entry) {
2840 hammer_btree_lock_copy(cursor, item);
2845 * Recursively sync modified copies to the media.
2848 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2850 hammer_node_lock_t item;
2853 if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2855 hammer_modify_node_all(cursor->trans, parent->node);
2856 *parent->node->ondisk = *parent->copy;
2857 hammer_modify_node_done(parent->node);
2858 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2859 hammer_flush_node(parent->node, 0);
2860 hammer_delete_node(cursor->trans, parent->node);
2863 TAILQ_FOREACH(item, &parent->list, entry) {
2864 count += hammer_btree_sync_copy(cursor, item);
2870 * Release previously obtained node locks. The caller is responsible for
2871 * cleaning up parent->node itself (its usually just aliased from a cursor),
2872 * but this function will take care of the copies.
2874 * NOTE: The root node is not placed in the lcache and node->copy is not
2875 * deallocated when lcache != NULL.
2878 hammer_btree_unlock_children(hammer_mount_t hmp, hammer_node_lock_t parent,
2879 hammer_node_lock_t lcache)
2881 hammer_node_lock_t item;
2882 hammer_node_ondisk_t copy;
2884 while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2885 TAILQ_REMOVE(&parent->list, item, entry);
2886 hammer_btree_unlock_children(hmp, item, lcache);
2887 hammer_unlock(&item->node->lock);
2888 hammer_rel_node(item->node);
2891 * NOTE: When placing the item back in the lcache
2892 * the flag is cleared by the bzero().
2893 * Remaining fields are cleared as a safety
2896 KKASSERT(item->flags & HAMMER_NODE_LOCK_LCACHE);
2897 KKASSERT(TAILQ_EMPTY(&item->list));
2899 bzero(item, sizeof(*item));
2900 TAILQ_INIT(&item->list);
2903 bzero(copy, sizeof(*copy));
2904 TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2906 kfree(item, hmp->m_misc);
2909 if (parent->copy && (parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0) {
2910 kfree(parent->copy, hmp->m_misc);
2911 parent->copy = NULL; /* safety */
2915 /************************************************************************
2916 * MISCELLANIOUS SUPPORT *
2917 ************************************************************************/
2920 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2922 * Note that for this particular function a return value of -1, 0, or +1
2923 * can denote a match if create_tid is otherwise discounted. A create_tid
2924 * of zero is considered to be 'infinity' in comparisons.
2926 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2929 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2931 if (key1->localization < key2->localization)
2933 if (key1->localization > key2->localization)
2936 if (key1->obj_id < key2->obj_id)
2938 if (key1->obj_id > key2->obj_id)
2941 if (key1->rec_type < key2->rec_type)
2943 if (key1->rec_type > key2->rec_type)
2946 if (key1->key < key2->key)
2948 if (key1->key > key2->key)
2952 * A create_tid of zero indicates a record which is undeletable
2953 * and must be considered to have a value of positive infinity.
2955 if (key1->create_tid == 0) {
2956 if (key2->create_tid == 0)
2960 if (key2->create_tid == 0)
2962 if (key1->create_tid < key2->create_tid)
2964 if (key1->create_tid > key2->create_tid)
2970 * Test a timestamp against an element to determine whether the
2971 * element is visible. A timestamp of 0 means 'infinity'.
2974 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2977 if (base->delete_tid)
2981 if (asof < base->create_tid)
2983 if (base->delete_tid && asof >= base->delete_tid)
2989 * Create a separator half way inbetween key1 and key2. For fields just
2990 * one unit apart, the separator will match key2. key1 is on the left-hand
2991 * side and key2 is on the right-hand side.
2993 * key2 must be >= the separator. It is ok for the separator to match key2.
2995 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2998 * NOTE: It might be beneficial to just scrap this whole mess and just
2999 * set the separator to key2.
3001 #define MAKE_SEPARATOR(key1, key2, dest, field) \
3002 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
3005 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
3006 hammer_base_elm_t dest)
3008 bzero(dest, sizeof(*dest));
3010 dest->rec_type = key2->rec_type;
3011 dest->key = key2->key;
3012 dest->obj_id = key2->obj_id;
3013 dest->create_tid = key2->create_tid;
3015 MAKE_SEPARATOR(key1, key2, dest, localization);
3016 if (key1->localization == key2->localization) {
3017 MAKE_SEPARATOR(key1, key2, dest, obj_id);
3018 if (key1->obj_id == key2->obj_id) {
3019 MAKE_SEPARATOR(key1, key2, dest, rec_type);
3020 if (key1->rec_type == key2->rec_type) {
3021 MAKE_SEPARATOR(key1, key2, dest, key);
3023 * Don't bother creating a separator for
3024 * create_tid, which also conveniently avoids
3025 * having to handle the create_tid == 0
3026 * (infinity) case. Just leave create_tid
3029 * Worst case, dest matches key2 exactly,
3030 * which is acceptable.
3037 #undef MAKE_SEPARATOR
3040 * Return whether a generic internal or leaf node is full
3044 btree_node_is_full(hammer_node_ondisk_t node)
3046 return(btree_max_elements(node->type) == node->count);
3051 btree_max_elements(u_int8_t type)
3055 n = hammer_node_max_elements(type);
3057 panic("btree_max_elements: bad type %d", type);
3062 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
3064 hammer_btree_elm_t elm;
3067 kprintf("node %p count=%d parent=%016llx type=%c\n",
3068 ondisk, ondisk->count,
3069 (long long)ondisk->parent, ondisk->type);
3072 * Dump both boundary elements if an internal node
3074 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
3075 for (i = 0; i <= ondisk->count; ++i) {
3076 elm = &ondisk->elms[i];
3077 hammer_print_btree_elm(elm, ondisk->type, i);
3080 for (i = 0; i < ondisk->count; ++i) {
3081 elm = &ondisk->elms[i];
3082 hammer_print_btree_elm(elm, ondisk->type, i);
3088 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
3091 kprintf("\tobj_id = %016llx\n", (long long)elm->base.obj_id);
3092 kprintf("\tkey = %016llx\n", (long long)elm->base.key);
3093 kprintf("\tcreate_tid = %016llx\n", (long long)elm->base.create_tid);
3094 kprintf("\tdelete_tid = %016llx\n", (long long)elm->base.delete_tid);
3095 kprintf("\trec_type = %04x\n", elm->base.rec_type);
3096 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
3097 kprintf("\tbtype = %02x (%c)\n",
3099 (elm->base.btype ? elm->base.btype : '?'));
3100 kprintf("\tlocalization = %02x\n", elm->base.localization);
3103 case HAMMER_BTREE_TYPE_INTERNAL:
3104 kprintf("\tsubtree_off = %016llx\n",
3105 (long long)elm->internal.subtree_offset);
3107 case HAMMER_BTREE_TYPE_RECORD:
3108 kprintf("\tdata_offset = %016llx\n",
3109 (long long)elm->leaf.data_offset);
3110 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
3111 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);