HAMMER: MFC to 2.0
[dragonfly.git] / sys / vfs / hammer / hammer_btree.c
blob98fe2544c0fda9d7657360c7888f4c6a58712fd6
1 /*
2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
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
16 * distribution.
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
32 * SUCH DAMAGE.
34 * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.71.2.4 2008/08/02 21:24:27 dillon Exp $
38 * HAMMER B-Tree index
40 * HAMMER implements a modified B+Tree. In documentation this will
41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
42 * looks like a B+Tree (A B-Tree which stores its records only at the leafs
43 * of the tree), but adds two additional boundary elements which describe
44 * the left-most and right-most element a node is able to represent. In
45 * otherwords, we have boundary elements at the two ends of a B-Tree node
46 * instead of sub-tree pointers.
48 * A B-Tree internal node looks like this:
50 * B N N N N N N B <-- boundary and internal elements
51 * S S S S S S S <-- subtree pointers
53 * A B-Tree leaf node basically looks like this:
55 * L L L L L L L L <-- leaf elemenets
57 * The radix for an internal node is 1 less then a leaf but we get a
58 * number of significant benefits for our troubles.
60 * The big benefit to using a B-Tree containing boundary information
61 * is that it is possible to cache pointers into the middle of the tree
62 * and not have to start searches, insertions, OR deletions at the root
63 * node. In particular, searches are able to progress in a definitive
64 * direction from any point in the tree without revisting nodes. This
65 * greatly improves the efficiency of many operations, most especially
66 * record appends.
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
73 * B-Tree.
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
79 * up the boundaries).
81 #include "hammer.h"
82 #include <sys/buf.h>
83 #include <sys/buf2.h>
85 static int btree_search(hammer_cursor_t cursor, int flags);
86 static int btree_split_internal(hammer_cursor_t cursor);
87 static int btree_split_leaf(hammer_cursor_t cursor);
88 static int btree_remove(hammer_cursor_t cursor);
89 static int btree_node_is_full(hammer_node_ondisk_t node);
90 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor,
91 hammer_tid_t mirror_tid);
92 static void hammer_make_separator(hammer_base_elm_t key1,
93 hammer_base_elm_t key2, hammer_base_elm_t dest);
94 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor);
97 * Iterate records after a search. The cursor is iterated forwards past
98 * the current record until a record matching the key-range requirements
99 * is found. ENOENT is returned if the iteration goes past the ending
100 * key.
102 * The iteration is inclusive of key_beg and can be inclusive or exclusive
103 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
105 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
106 * may be modified by B-Tree functions.
108 * cursor->key_beg may or may not be modified by this function during
109 * the iteration. XXX future - in case of an inverted lock we may have
110 * to reinitiate the lookup and set key_beg to properly pick up where we
111 * left off.
113 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
116 hammer_btree_iterate(hammer_cursor_t cursor)
118 hammer_node_ondisk_t node;
119 hammer_btree_elm_t elm;
120 int error = 0;
121 int r;
122 int s;
125 * Skip past the current record
127 node = cursor->node->ondisk;
128 if (node == NULL)
129 return(ENOENT);
130 if (cursor->index < node->count &&
131 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
132 ++cursor->index;
136 * Loop until an element is found or we are done.
138 for (;;) {
140 * We iterate up the tree and then index over one element
141 * while we are at the last element in the current node.
143 * If we are at the root of the filesystem, cursor_up
144 * returns ENOENT.
146 * XXX this could be optimized by storing the information in
147 * the parent reference.
149 * XXX we can lose the node lock temporarily, this could mess
150 * up our scan.
152 ++hammer_stats_btree_iterations;
153 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
155 if (cursor->index == node->count) {
156 if (hammer_debug_btree) {
157 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
158 cursor->node->node_offset,
159 cursor->index,
160 (cursor->parent ? cursor->parent->node_offset : -1),
161 cursor->parent_index,
162 curthread);
164 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
165 error = hammer_cursor_up(cursor);
166 if (error)
167 break;
168 /* reload stale pointer */
169 node = cursor->node->ondisk;
170 KKASSERT(cursor->index != node->count);
173 * If we are reblocking we want to return internal
174 * nodes.
176 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
177 cursor->flags |= HAMMER_CURSOR_ATEDISK;
178 return(0);
180 ++cursor->index;
181 continue;
185 * Check internal or leaf element. Determine if the record
186 * at the cursor has gone beyond the end of our range.
188 * We recurse down through internal nodes.
190 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
191 elm = &node->elms[cursor->index];
193 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
194 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
195 if (hammer_debug_btree) {
196 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
197 cursor->node->node_offset,
198 cursor->index,
199 elm[0].internal.base.obj_id,
200 elm[0].internal.base.rec_type,
201 elm[0].internal.base.key,
202 elm[0].internal.base.localization,
204 curthread
206 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
207 cursor->node->node_offset,
208 cursor->index + 1,
209 elm[1].internal.base.obj_id,
210 elm[1].internal.base.rec_type,
211 elm[1].internal.base.key,
212 elm[1].internal.base.localization,
217 if (r < 0) {
218 error = ENOENT;
219 break;
221 if (r == 0 && (cursor->flags &
222 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
223 error = ENOENT;
224 break;
226 KKASSERT(s <= 0);
229 * Better not be zero
231 KKASSERT(elm->internal.subtree_offset != 0);
234 * If running the mirror filter see if we can skip
235 * one or more entire sub-trees. If we can we
236 * return the internal mode and the caller processes
237 * the skipped range (see mirror_read)
239 if (cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) {
240 if (elm->internal.mirror_tid <
241 cursor->cmirror->mirror_tid) {
242 hammer_cursor_mirror_filter(cursor);
243 return(0);
247 error = hammer_cursor_down(cursor);
248 if (error)
249 break;
250 KKASSERT(cursor->index == 0);
251 /* reload stale pointer */
252 node = cursor->node->ondisk;
253 continue;
254 } else {
255 elm = &node->elms[cursor->index];
256 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
257 if (hammer_debug_btree) {
258 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
259 cursor->node->node_offset,
260 cursor->index,
261 (elm[0].leaf.base.btype ?
262 elm[0].leaf.base.btype : '?'),
263 elm[0].leaf.base.obj_id,
264 elm[0].leaf.base.rec_type,
265 elm[0].leaf.base.key,
266 elm[0].leaf.base.localization,
270 if (r < 0) {
271 error = ENOENT;
272 break;
276 * We support both end-inclusive and
277 * end-exclusive searches.
279 if (r == 0 &&
280 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
281 error = ENOENT;
282 break;
285 switch(elm->leaf.base.btype) {
286 case HAMMER_BTREE_TYPE_RECORD:
287 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
288 hammer_btree_chkts(cursor->asof, &elm->base)) {
289 ++cursor->index;
290 continue;
292 error = 0;
293 break;
294 default:
295 error = EINVAL;
296 break;
298 if (error)
299 break;
302 * node pointer invalid after loop
306 * Return entry
308 if (hammer_debug_btree) {
309 int i = cursor->index;
310 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
311 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
312 cursor->node, i,
313 elm->internal.base.obj_id,
314 elm->internal.base.rec_type,
315 elm->internal.base.key,
316 elm->internal.base.localization
319 return(0);
321 return(error);
325 * We hit an internal element that we could skip as part of a mirroring
326 * scan. Calculate the entire range being skipped.
328 * It is important to include any gaps between the parent's left_bound
329 * and the node's left_bound, and same goes for the right side.
331 static void
332 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
334 struct hammer_cmirror *cmirror;
335 hammer_node_ondisk_t ondisk;
336 hammer_btree_elm_t elm;
338 ondisk = cursor->node->ondisk;
339 cmirror = cursor->cmirror;
342 * Calculate the skipped range
344 elm = &ondisk->elms[cursor->index];
345 if (cursor->index == 0)
346 cmirror->skip_beg = *cursor->left_bound;
347 else
348 cmirror->skip_beg = elm->internal.base;
349 while (cursor->index < ondisk->count) {
350 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
351 break;
352 ++cursor->index;
353 ++elm;
355 if (cursor->index == ondisk->count)
356 cmirror->skip_end = *cursor->right_bound;
357 else
358 cmirror->skip_end = elm->internal.base;
361 * clip the returned result.
363 if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
364 cmirror->skip_beg = cursor->key_beg;
365 if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
366 cmirror->skip_end = cursor->key_end;
370 * Iterate in the reverse direction. This is used by the pruning code to
371 * avoid overlapping records.
374 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
376 hammer_node_ondisk_t node;
377 hammer_btree_elm_t elm;
378 int error = 0;
379 int r;
380 int s;
382 /* mirror filtering not supported for reverse iteration */
383 KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
386 * Skip past the current record. For various reasons the cursor
387 * may end up set to -1 or set to point at the end of the current
388 * node. These cases must be addressed.
390 node = cursor->node->ondisk;
391 if (node == NULL)
392 return(ENOENT);
393 if (cursor->index != -1 &&
394 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
395 --cursor->index;
397 if (cursor->index == cursor->node->ondisk->count)
398 --cursor->index;
401 * Loop until an element is found or we are done.
403 for (;;) {
404 ++hammer_stats_btree_iterations;
405 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
408 * We iterate up the tree and then index over one element
409 * while we are at the last element in the current node.
411 if (cursor->index == -1) {
412 error = hammer_cursor_up(cursor);
413 if (error) {
414 cursor->index = 0; /* sanity */
415 break;
417 /* reload stale pointer */
418 node = cursor->node->ondisk;
419 KKASSERT(cursor->index != node->count);
420 --cursor->index;
421 continue;
425 * Check internal or leaf element. Determine if the record
426 * at the cursor has gone beyond the end of our range.
428 * We recurse down through internal nodes.
430 KKASSERT(cursor->index != node->count);
431 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
432 elm = &node->elms[cursor->index];
433 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
434 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
435 if (hammer_debug_btree) {
436 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
437 cursor->node->node_offset,
438 cursor->index,
439 elm[0].internal.base.obj_id,
440 elm[0].internal.base.rec_type,
441 elm[0].internal.base.key,
442 elm[0].internal.base.localization,
445 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
446 cursor->node->node_offset,
447 cursor->index + 1,
448 elm[1].internal.base.obj_id,
449 elm[1].internal.base.rec_type,
450 elm[1].internal.base.key,
451 elm[1].internal.base.localization,
456 if (s >= 0) {
457 error = ENOENT;
458 break;
460 KKASSERT(r >= 0);
463 * Better not be zero
465 KKASSERT(elm->internal.subtree_offset != 0);
467 error = hammer_cursor_down(cursor);
468 if (error)
469 break;
470 KKASSERT(cursor->index == 0);
471 /* reload stale pointer */
472 node = cursor->node->ondisk;
474 /* this can assign -1 if the leaf was empty */
475 cursor->index = node->count - 1;
476 continue;
477 } else {
478 elm = &node->elms[cursor->index];
479 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
480 if (hammer_debug_btree) {
481 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
482 cursor->node->node_offset,
483 cursor->index,
484 (elm[0].leaf.base.btype ?
485 elm[0].leaf.base.btype : '?'),
486 elm[0].leaf.base.obj_id,
487 elm[0].leaf.base.rec_type,
488 elm[0].leaf.base.key,
489 elm[0].leaf.base.localization,
493 if (s > 0) {
494 error = ENOENT;
495 break;
498 switch(elm->leaf.base.btype) {
499 case HAMMER_BTREE_TYPE_RECORD:
500 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
501 hammer_btree_chkts(cursor->asof, &elm->base)) {
502 --cursor->index;
503 continue;
505 error = 0;
506 break;
507 default:
508 error = EINVAL;
509 break;
511 if (error)
512 break;
515 * node pointer invalid after loop
519 * Return entry
521 if (hammer_debug_btree) {
522 int i = cursor->index;
523 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
524 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
525 cursor->node, i,
526 elm->internal.base.obj_id,
527 elm->internal.base.rec_type,
528 elm->internal.base.key,
529 elm->internal.base.localization
532 return(0);
534 return(error);
538 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
539 * could not be found, EDEADLK if inserting and a retry is needed, and a
540 * fatal error otherwise. When retrying, the caller must terminate the
541 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
543 * The cursor is suitably positioned for a deletion on success, and suitably
544 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
545 * specified.
547 * The cursor may begin anywhere, the search will traverse the tree in
548 * either direction to locate the requested element.
550 * Most of the logic implementing historical searches is handled here. We
551 * do an initial lookup with create_tid set to the asof TID. Due to the
552 * way records are laid out, a backwards iteration may be required if
553 * ENOENT is returned to locate the historical record. Here's the
554 * problem:
556 * create_tid: 10 15 20
557 * LEAF1 LEAF2
558 * records: (11) (18)
560 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
561 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
562 * not visible and thus causes ENOENT to be returned. We really need
563 * to check record 11 in LEAF1. If it also fails then the search fails
564 * (e.g. it might represent the range 11-16 and thus still not match our
565 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
566 * further iterations.
568 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
569 * and the cursor->create_check TID if an iteration might be needed.
570 * In the above example create_check would be set to 14.
573 hammer_btree_lookup(hammer_cursor_t cursor)
575 int error;
577 KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
578 cursor->trans->sync_lock_refs > 0);
579 ++hammer_stats_btree_lookups;
580 if (cursor->flags & HAMMER_CURSOR_ASOF) {
581 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
582 cursor->key_beg.create_tid = cursor->asof;
583 for (;;) {
584 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
585 error = btree_search(cursor, 0);
586 if (error != ENOENT ||
587 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
589 * Stop if no error.
590 * Stop if error other then ENOENT.
591 * Stop if ENOENT and not special case.
593 break;
595 if (hammer_debug_btree) {
596 kprintf("CREATE_CHECK %016llx\n",
597 cursor->create_check);
599 cursor->key_beg.create_tid = cursor->create_check;
600 /* loop */
602 } else {
603 error = btree_search(cursor, 0);
605 if (error == 0)
606 error = hammer_btree_extract(cursor, cursor->flags);
607 return(error);
611 * Execute the logic required to start an iteration. The first record
612 * located within the specified range is returned and iteration control
613 * flags are adjusted for successive hammer_btree_iterate() calls.
616 hammer_btree_first(hammer_cursor_t cursor)
618 int error;
620 error = hammer_btree_lookup(cursor);
621 if (error == ENOENT) {
622 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
623 error = hammer_btree_iterate(cursor);
625 cursor->flags |= HAMMER_CURSOR_ATEDISK;
626 return(error);
630 * Similarly but for an iteration in the reverse direction.
632 * Set ATEDISK when iterating backwards to skip the current entry,
633 * which after an ENOENT lookup will be pointing beyond our end point.
636 hammer_btree_last(hammer_cursor_t cursor)
638 struct hammer_base_elm save;
639 int error;
641 save = cursor->key_beg;
642 cursor->key_beg = cursor->key_end;
643 error = hammer_btree_lookup(cursor);
644 cursor->key_beg = save;
645 if (error == ENOENT ||
646 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
647 cursor->flags |= HAMMER_CURSOR_ATEDISK;
648 error = hammer_btree_iterate_reverse(cursor);
650 cursor->flags |= HAMMER_CURSOR_ATEDISK;
651 return(error);
655 * Extract the record and/or data associated with the cursor's current
656 * position. Any prior record or data stored in the cursor is replaced.
657 * The cursor must be positioned at a leaf node.
659 * NOTE: All extractions occur at the leaf of the B-Tree.
662 hammer_btree_extract(hammer_cursor_t cursor, int flags)
664 hammer_mount_t hmp;
665 hammer_node_ondisk_t node;
666 hammer_btree_elm_t elm;
667 hammer_off_t data_off;
668 int32_t data_len;
669 int error;
672 * The case where the data reference resolves to the same buffer
673 * as the record reference must be handled.
675 node = cursor->node->ondisk;
676 elm = &node->elms[cursor->index];
677 cursor->data = NULL;
678 hmp = cursor->node->hmp;
681 * There is nothing to extract for an internal element.
683 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
684 return(EINVAL);
687 * Only record types have data.
689 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
690 cursor->leaf = &elm->leaf;
692 if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
693 return(0);
694 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
695 return(0);
696 data_off = elm->leaf.data_offset;
697 data_len = elm->leaf.data_len;
698 if (data_off == 0)
699 return(0);
702 * Load the data
704 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
705 cursor->data = hammer_bread_ext(hmp, data_off, data_len,
706 &error, &cursor->data_buffer);
707 if (hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
708 kprintf("CRC DATA @ %016llx/%d FAILED\n",
709 elm->leaf.data_offset, elm->leaf.data_len);
710 Debugger("CRC FAILED: DATA");
712 return(error);
717 * Insert a leaf element into the B-Tree at the current cursor position.
718 * The cursor is positioned such that the element at and beyond the cursor
719 * are shifted to make room for the new record.
721 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
722 * flag set and that call must return ENOENT before this function can be
723 * called.
725 * The caller may depend on the cursor's exclusive lock after return to
726 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
728 * ENOSPC is returned if there is no room to insert a new record.
731 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
732 int *doprop)
734 hammer_node_ondisk_t node;
735 int i;
736 int error;
738 *doprop = 0;
739 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
740 return(error);
741 ++hammer_stats_btree_inserts;
744 * Insert the element at the leaf node and update the count in the
745 * parent. It is possible for parent to be NULL, indicating that
746 * the filesystem's ROOT B-Tree node is a leaf itself, which is
747 * possible. The root inode can never be deleted so the leaf should
748 * never be empty.
750 * Remember that the right-hand boundary is not included in the
751 * count.
753 hammer_modify_node_all(cursor->trans, cursor->node);
754 node = cursor->node->ondisk;
755 i = cursor->index;
756 KKASSERT(elm->base.btype != 0);
757 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
758 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
759 if (i != node->count) {
760 bcopy(&node->elms[i], &node->elms[i+1],
761 (node->count - i) * sizeof(*elm));
763 node->elms[i].leaf = *elm;
764 ++node->count;
765 hammer_cursor_inserted_element(cursor->node, i);
768 * Update the leaf node's aggregate mirror_tid for mirroring
769 * support.
771 if (node->mirror_tid < elm->base.delete_tid) {
772 node->mirror_tid = elm->base.delete_tid;
773 *doprop = 1;
775 if (node->mirror_tid < elm->base.create_tid) {
776 node->mirror_tid = elm->base.create_tid;
777 *doprop = 1;
779 hammer_modify_node_done(cursor->node);
782 * Debugging sanity checks.
784 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
785 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
786 if (i) {
787 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
789 if (i != node->count - 1)
790 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
792 return(0);
796 * Delete a record from the B-Tree at the current cursor position.
797 * The cursor is positioned such that the current element is the one
798 * to be deleted.
800 * On return the cursor will be positioned after the deleted element and
801 * MAY point to an internal node. It will be suitable for the continuation
802 * of an iteration but not for an insertion or deletion.
804 * Deletions will attempt to partially rebalance the B-Tree in an upward
805 * direction, but will terminate rather then deadlock. Empty internal nodes
806 * are never allowed by a deletion which deadlocks may end up giving us an
807 * empty leaf. The pruner will clean up and rebalance the tree.
809 * This function can return EDEADLK, requiring the caller to retry the
810 * operation after clearing the deadlock.
813 hammer_btree_delete(hammer_cursor_t cursor)
815 hammer_node_ondisk_t ondisk;
816 hammer_node_t node;
817 hammer_node_t parent;
818 int error;
819 int i;
821 KKASSERT (cursor->trans->sync_lock_refs > 0);
822 if ((error = hammer_cursor_upgrade(cursor)) != 0)
823 return(error);
824 ++hammer_stats_btree_deletes;
827 * Delete the element from the leaf node.
829 * Remember that leaf nodes do not have boundaries.
831 node = cursor->node;
832 ondisk = node->ondisk;
833 i = cursor->index;
835 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
836 KKASSERT(i >= 0 && i < ondisk->count);
837 hammer_modify_node_all(cursor->trans, node);
838 if (i + 1 != ondisk->count) {
839 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
840 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
842 --ondisk->count;
843 hammer_modify_node_done(node);
844 hammer_cursor_deleted_element(node, i);
847 * Validate local parent
849 if (ondisk->parent) {
850 parent = cursor->parent;
852 KKASSERT(parent != NULL);
853 KKASSERT(parent->node_offset == ondisk->parent);
857 * If the leaf becomes empty it must be detached from the parent,
858 * potentially recursing through to the filesystem root.
860 * This may reposition the cursor at one of the parent's of the
861 * current node.
863 * Ignore deadlock errors, that simply means that btree_remove
864 * was unable to recurse and had to leave us with an empty leaf.
866 KKASSERT(cursor->index <= ondisk->count);
867 if (ondisk->count == 0) {
868 error = btree_remove(cursor);
869 if (error == EDEADLK)
870 error = 0;
871 } else {
872 error = 0;
874 KKASSERT(cursor->parent == NULL ||
875 cursor->parent_index < cursor->parent->ondisk->count);
876 return(error);
880 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
882 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
884 * The search can begin ANYWHERE in the B-Tree. As a first step the search
885 * iterates up the tree as necessary to properly position itself prior to
886 * actually doing the sarch.
888 * INSERTIONS: The search will split full nodes and leaves on its way down
889 * and guarentee that the leaf it ends up on is not full. If we run out
890 * of space the search continues to the leaf (to position the cursor for
891 * the spike), but ENOSPC is returned.
893 * The search is only guarenteed to end up on a leaf if an error code of 0
894 * is returned, or if inserting and an error code of ENOENT is returned.
895 * Otherwise it can stop at an internal node. On success a search returns
896 * a leaf node.
898 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
899 * filesystem, and it is not simple code. Please note the following facts:
901 * - Internal node recursions have a boundary on the left AND right. The
902 * right boundary is non-inclusive. The create_tid is a generic part
903 * of the key for internal nodes.
905 * - Leaf nodes contain terminal elements only now.
907 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
908 * historical search. ASOF and INSERT are mutually exclusive. When
909 * doing an as-of lookup btree_search() checks for a right-edge boundary
910 * case. If while recursing down the left-edge differs from the key
911 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
912 * with cursor->create_check. This is used by btree_lookup() to iterate.
913 * The iteration backwards because as-of searches can wind up going
914 * down the wrong branch of the B-Tree.
916 static
918 btree_search(hammer_cursor_t cursor, int flags)
920 hammer_node_ondisk_t node;
921 hammer_btree_elm_t elm;
922 int error;
923 int enospc = 0;
924 int i;
925 int r;
926 int s;
928 flags |= cursor->flags;
929 ++hammer_stats_btree_searches;
931 if (hammer_debug_btree) {
932 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
933 cursor->node->node_offset,
934 cursor->index,
935 cursor->key_beg.obj_id,
936 cursor->key_beg.rec_type,
937 cursor->key_beg.key,
938 cursor->key_beg.create_tid,
939 cursor->key_beg.localization,
940 curthread
942 if (cursor->parent)
943 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
944 cursor->parent->node_offset, cursor->parent_index,
945 cursor->left_bound->obj_id,
946 cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
947 cursor->right_bound->obj_id,
948 cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
949 cursor->left_bound,
950 &cursor->parent->ondisk->elms[cursor->parent_index],
951 cursor->right_bound,
952 &cursor->parent->ondisk->elms[cursor->parent_index+1]
957 * Move our cursor up the tree until we find a node whos range covers
958 * the key we are trying to locate.
960 * The left bound is inclusive, the right bound is non-inclusive.
961 * It is ok to cursor up too far.
963 for (;;) {
964 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
965 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
966 if (r >= 0 && s < 0)
967 break;
968 KKASSERT(cursor->parent);
969 ++hammer_stats_btree_iterations;
970 error = hammer_cursor_up(cursor);
971 if (error)
972 goto done;
976 * The delete-checks below are based on node, not parent. Set the
977 * initial delete-check based on the parent.
979 if (r == 1) {
980 KKASSERT(cursor->left_bound->create_tid != 1);
981 cursor->create_check = cursor->left_bound->create_tid - 1;
982 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
986 * We better have ended up with a node somewhere.
988 KKASSERT(cursor->node != NULL);
991 * If we are inserting we can't start at a full node if the parent
992 * is also full (because there is no way to split the node),
993 * continue running up the tree until the requirement is satisfied
994 * or we hit the root of the filesystem.
996 * (If inserting we aren't doing an as-of search so we don't have
997 * to worry about create_check).
999 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1000 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1001 if (btree_node_is_full(cursor->node->ondisk) == 0)
1002 break;
1003 } else {
1004 if (btree_node_is_full(cursor->node->ondisk) ==0)
1005 break;
1007 if (cursor->node->ondisk->parent == 0 ||
1008 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1009 break;
1011 ++hammer_stats_btree_iterations;
1012 error = hammer_cursor_up(cursor);
1013 /* node may have become stale */
1014 if (error)
1015 goto done;
1019 * Push down through internal nodes to locate the requested key.
1021 node = cursor->node->ondisk;
1022 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1024 * Scan the node to find the subtree index to push down into.
1025 * We go one-past, then back-up.
1027 * We must proactively remove deleted elements which may
1028 * have been left over from a deadlocked btree_remove().
1030 * The left and right boundaries are included in the loop
1031 * in order to detect edge cases.
1033 * If the separator only differs by create_tid (r == 1)
1034 * and we are doing an as-of search, we may end up going
1035 * down a branch to the left of the one containing the
1036 * desired key. This requires numerous special cases.
1038 ++hammer_stats_btree_iterations;
1039 if (hammer_debug_btree) {
1040 kprintf("SEARCH-I %016llx count=%d\n",
1041 cursor->node->node_offset,
1042 node->count);
1046 * Try to shortcut the search before dropping into the
1047 * linear loop. Locate the first node where r <= 1.
1049 i = hammer_btree_search_node(&cursor->key_beg, node);
1050 while (i <= node->count) {
1051 ++hammer_stats_btree_elements;
1052 elm = &node->elms[i];
1053 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1054 if (hammer_debug_btree > 2) {
1055 kprintf(" IELM %p %d r=%d\n",
1056 &node->elms[i], i, r);
1058 if (r < 0)
1059 break;
1060 if (r == 1) {
1061 KKASSERT(elm->base.create_tid != 1);
1062 cursor->create_check = elm->base.create_tid - 1;
1063 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1065 ++i;
1067 if (hammer_debug_btree) {
1068 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1069 i, node->count, r);
1073 * These cases occur when the parent's idea of the boundary
1074 * is wider then the child's idea of the boundary, and
1075 * require special handling. If not inserting we can
1076 * terminate the search early for these cases but the
1077 * child's boundaries cannot be unconditionally modified.
1079 if (i == 0) {
1081 * If i == 0 the search terminated to the LEFT of the
1082 * left_boundary but to the RIGHT of the parent's left
1083 * boundary.
1085 u_int8_t save;
1087 elm = &node->elms[0];
1090 * If we aren't inserting we can stop here.
1092 if ((flags & (HAMMER_CURSOR_INSERT |
1093 HAMMER_CURSOR_PRUNING)) == 0) {
1094 cursor->index = 0;
1095 return(ENOENT);
1099 * Correct a left-hand boundary mismatch.
1101 * We can only do this if we can upgrade the lock,
1102 * and synchronized as a background cursor (i.e.
1103 * inserting or pruning).
1105 * WARNING: We can only do this if inserting, i.e.
1106 * we are running on the backend.
1108 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1109 return(error);
1110 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1111 hammer_modify_node_field(cursor->trans, cursor->node,
1112 elms[0]);
1113 save = node->elms[0].base.btype;
1114 node->elms[0].base = *cursor->left_bound;
1115 node->elms[0].base.btype = save;
1116 hammer_modify_node_done(cursor->node);
1117 } else if (i == node->count + 1) {
1119 * If i == node->count + 1 the search terminated to
1120 * the RIGHT of the right boundary but to the LEFT
1121 * of the parent's right boundary. If we aren't
1122 * inserting we can stop here.
1124 * Note that the last element in this case is
1125 * elms[i-2] prior to adjustments to 'i'.
1127 --i;
1128 if ((flags & (HAMMER_CURSOR_INSERT |
1129 HAMMER_CURSOR_PRUNING)) == 0) {
1130 cursor->index = i;
1131 return (ENOENT);
1135 * Correct a right-hand boundary mismatch.
1136 * (actual push-down record is i-2 prior to
1137 * adjustments to i).
1139 * We can only do this if we can upgrade the lock,
1140 * and synchronized as a background cursor (i.e.
1141 * inserting or pruning).
1143 * WARNING: We can only do this if inserting, i.e.
1144 * we are running on the backend.
1146 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1147 return(error);
1148 elm = &node->elms[i];
1149 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1150 hammer_modify_node(cursor->trans, cursor->node,
1151 &elm->base, sizeof(elm->base));
1152 elm->base = *cursor->right_bound;
1153 hammer_modify_node_done(cursor->node);
1154 --i;
1155 } else {
1157 * The push-down index is now i - 1. If we had
1158 * terminated on the right boundary this will point
1159 * us at the last element.
1161 --i;
1163 cursor->index = i;
1164 elm = &node->elms[i];
1166 if (hammer_debug_btree) {
1167 kprintf("RESULT-I %016llx[%d] %016llx %02x "
1168 "key=%016llx cre=%016llx lo=%02x\n",
1169 cursor->node->node_offset,
1171 elm->internal.base.obj_id,
1172 elm->internal.base.rec_type,
1173 elm->internal.base.key,
1174 elm->internal.base.create_tid,
1175 elm->internal.base.localization
1180 * We better have a valid subtree offset.
1182 KKASSERT(elm->internal.subtree_offset != 0);
1185 * Handle insertion and deletion requirements.
1187 * If inserting split full nodes. The split code will
1188 * adjust cursor->node and cursor->index if the current
1189 * index winds up in the new node.
1191 * If inserting and a left or right edge case was detected,
1192 * we cannot correct the left or right boundary and must
1193 * prepend and append an empty leaf node in order to make
1194 * the boundary correction.
1196 * If we run out of space we set enospc and continue on
1197 * to a leaf to provide the spike code with a good point
1198 * of entry.
1200 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1201 if (btree_node_is_full(node)) {
1202 error = btree_split_internal(cursor);
1203 if (error) {
1204 if (error != ENOSPC)
1205 goto done;
1206 enospc = 1;
1209 * reload stale pointers
1211 i = cursor->index;
1212 node = cursor->node->ondisk;
1217 * Push down (push into new node, existing node becomes
1218 * the parent) and continue the search.
1220 error = hammer_cursor_down(cursor);
1221 /* node may have become stale */
1222 if (error)
1223 goto done;
1224 node = cursor->node->ondisk;
1228 * We are at a leaf, do a linear search of the key array.
1230 * On success the index is set to the matching element and 0
1231 * is returned.
1233 * On failure the index is set to the insertion point and ENOENT
1234 * is returned.
1236 * Boundaries are not stored in leaf nodes, so the index can wind
1237 * up to the left of element 0 (index == 0) or past the end of
1238 * the array (index == node->count). It is also possible that the
1239 * leaf might be empty.
1241 ++hammer_stats_btree_iterations;
1242 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1243 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1244 if (hammer_debug_btree) {
1245 kprintf("SEARCH-L %016llx count=%d\n",
1246 cursor->node->node_offset,
1247 node->count);
1251 * Try to shortcut the search before dropping into the
1252 * linear loop. Locate the first node where r <= 1.
1254 i = hammer_btree_search_node(&cursor->key_beg, node);
1255 while (i < node->count) {
1256 ++hammer_stats_btree_elements;
1257 elm = &node->elms[i];
1259 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1261 if (hammer_debug_btree > 1)
1262 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1265 * We are at a record element. Stop if we've flipped past
1266 * key_beg, not counting the create_tid test. Allow the
1267 * r == 1 case (key_beg > element but differs only by its
1268 * create_tid) to fall through to the AS-OF check.
1270 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1272 if (r < 0)
1273 goto failed;
1274 if (r > 1) {
1275 ++i;
1276 continue;
1280 * Check our as-of timestamp against the element.
1282 if (flags & HAMMER_CURSOR_ASOF) {
1283 if (hammer_btree_chkts(cursor->asof,
1284 &node->elms[i].base) != 0) {
1285 ++i;
1286 continue;
1288 /* success */
1289 } else {
1290 if (r > 0) { /* can only be +1 */
1291 ++i;
1292 continue;
1294 /* success */
1296 cursor->index = i;
1297 error = 0;
1298 if (hammer_debug_btree) {
1299 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1300 cursor->node->node_offset, i);
1302 goto done;
1306 * The search of the leaf node failed. i is the insertion point.
1308 failed:
1309 if (hammer_debug_btree) {
1310 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1311 cursor->node->node_offset, i);
1315 * No exact match was found, i is now at the insertion point.
1317 * If inserting split a full leaf before returning. This
1318 * may have the side effect of adjusting cursor->node and
1319 * cursor->index.
1321 cursor->index = i;
1322 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1323 btree_node_is_full(node)) {
1324 error = btree_split_leaf(cursor);
1325 if (error) {
1326 if (error != ENOSPC)
1327 goto done;
1328 enospc = 1;
1331 * reload stale pointers
1333 /* NOT USED
1334 i = cursor->index;
1335 node = &cursor->node->internal;
1340 * We reached a leaf but did not find the key we were looking for.
1341 * If this is an insert we will be properly positioned for an insert
1342 * (ENOENT) or spike (ENOSPC) operation.
1344 error = enospc ? ENOSPC : ENOENT;
1345 done:
1346 return(error);
1350 * Heuristical search for the first element whos comparison is <= 1. May
1351 * return an index whos compare result is > 1 but may only return an index
1352 * whos compare result is <= 1 if it is the first element with that result.
1355 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1357 int b;
1358 int s;
1359 int i;
1360 int r;
1363 * Don't bother if the node does not have very many elements
1365 b = 0;
1366 s = node->count;
1367 while (s - b > 4) {
1368 i = b + (s - b) / 2;
1369 ++hammer_stats_btree_elements;
1370 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1371 if (r <= 1) {
1372 s = i;
1373 } else {
1374 b = i;
1377 return(b);
1381 /************************************************************************
1382 * SPLITTING AND MERGING *
1383 ************************************************************************
1385 * These routines do all the dirty work required to split and merge nodes.
1389 * Split an internal node into two nodes and move the separator at the split
1390 * point to the parent.
1392 * (cursor->node, cursor->index) indicates the element the caller intends
1393 * to push into. We will adjust node and index if that element winds
1394 * up in the split node.
1396 * If we are at the root of the filesystem a new root must be created with
1397 * two elements, one pointing to the original root and one pointing to the
1398 * newly allocated split node.
1400 static
1402 btree_split_internal(hammer_cursor_t cursor)
1404 hammer_node_ondisk_t ondisk;
1405 hammer_node_t node;
1406 hammer_node_t parent;
1407 hammer_node_t new_node;
1408 hammer_btree_elm_t elm;
1409 hammer_btree_elm_t parent_elm;
1410 hammer_node_locklist_t locklist = NULL;
1411 hammer_mount_t hmp = cursor->trans->hmp;
1412 int parent_index;
1413 int made_root;
1414 int split;
1415 int error;
1416 int i;
1417 const int esize = sizeof(*elm);
1419 error = hammer_btree_lock_children(cursor, &locklist);
1420 if (error)
1421 goto done;
1422 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1423 goto done;
1424 ++hammer_stats_btree_splits;
1427 * We are splitting but elms[split] will be promoted to the parent,
1428 * leaving the right hand node with one less element. If the
1429 * insertion point will be on the left-hand side adjust the split
1430 * point to give the right hand side one additional node.
1432 node = cursor->node;
1433 ondisk = node->ondisk;
1434 split = (ondisk->count + 1) / 2;
1435 if (cursor->index <= split)
1436 --split;
1439 * If we are at the root of the filesystem, create a new root node
1440 * with 1 element and split normally. Avoid making major
1441 * modifications until we know the whole operation will work.
1443 if (ondisk->parent == 0) {
1444 parent = hammer_alloc_btree(cursor->trans, &error);
1445 if (parent == NULL)
1446 goto done;
1447 hammer_lock_ex(&parent->lock);
1448 hammer_modify_node_noundo(cursor->trans, parent);
1449 ondisk = parent->ondisk;
1450 ondisk->count = 1;
1451 ondisk->parent = 0;
1452 ondisk->mirror_tid = node->ondisk->mirror_tid;
1453 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1454 ondisk->elms[0].base = hmp->root_btree_beg;
1455 ondisk->elms[0].base.btype = node->ondisk->type;
1456 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1457 ondisk->elms[1].base = hmp->root_btree_end;
1458 hammer_modify_node_done(parent);
1459 /* ondisk->elms[1].base.btype - not used */
1460 made_root = 1;
1461 parent_index = 0; /* index of current node in parent */
1462 } else {
1463 made_root = 0;
1464 parent = cursor->parent;
1465 parent_index = cursor->parent_index;
1469 * Split node into new_node at the split point.
1471 * B O O O P N N B <-- P = node->elms[split]
1472 * 0 1 2 3 4 5 6 <-- subtree indices
1474 * x x P x x
1475 * s S S s
1476 * / \
1477 * B O O O B B N N B <--- inner boundary points are 'P'
1478 * 0 1 2 3 4 5 6
1481 new_node = hammer_alloc_btree(cursor->trans, &error);
1482 if (new_node == NULL) {
1483 if (made_root) {
1484 hammer_unlock(&parent->lock);
1485 hammer_delete_node(cursor->trans, parent);
1486 hammer_rel_node(parent);
1488 goto done;
1490 hammer_lock_ex(&new_node->lock);
1493 * Create the new node. P becomes the left-hand boundary in the
1494 * new node. Copy the right-hand boundary as well.
1496 * elm is the new separator.
1498 hammer_modify_node_noundo(cursor->trans, new_node);
1499 hammer_modify_node_all(cursor->trans, node);
1500 ondisk = node->ondisk;
1501 elm = &ondisk->elms[split];
1502 bcopy(elm, &new_node->ondisk->elms[0],
1503 (ondisk->count - split + 1) * esize);
1504 new_node->ondisk->count = ondisk->count - split;
1505 new_node->ondisk->parent = parent->node_offset;
1506 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1507 new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1508 KKASSERT(ondisk->type == new_node->ondisk->type);
1509 hammer_cursor_split_node(node, new_node, split);
1512 * Cleanup the original node. Elm (P) becomes the new boundary,
1513 * its subtree_offset was moved to the new node. If we had created
1514 * a new root its parent pointer may have changed.
1516 elm->internal.subtree_offset = 0;
1517 ondisk->count = split;
1520 * Insert the separator into the parent, fixup the parent's
1521 * reference to the original node, and reference the new node.
1522 * The separator is P.
1524 * Remember that base.count does not include the right-hand boundary.
1526 hammer_modify_node_all(cursor->trans, parent);
1527 ondisk = parent->ondisk;
1528 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1529 parent_elm = &ondisk->elms[parent_index+1];
1530 bcopy(parent_elm, parent_elm + 1,
1531 (ondisk->count - parent_index) * esize);
1532 parent_elm->internal.base = elm->base; /* separator P */
1533 parent_elm->internal.base.btype = new_node->ondisk->type;
1534 parent_elm->internal.subtree_offset = new_node->node_offset;
1535 parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1536 ++ondisk->count;
1537 hammer_modify_node_done(parent);
1538 hammer_cursor_inserted_element(parent, parent_index + 1);
1541 * The children of new_node need their parent pointer set to new_node.
1542 * The children have already been locked by
1543 * hammer_btree_lock_children().
1545 for (i = 0; i < new_node->ondisk->count; ++i) {
1546 elm = &new_node->ondisk->elms[i];
1547 error = btree_set_parent(cursor->trans, new_node, elm);
1548 if (error) {
1549 panic("btree_split_internal: btree-fixup problem");
1552 hammer_modify_node_done(new_node);
1555 * The filesystem's root B-Tree pointer may have to be updated.
1557 if (made_root) {
1558 hammer_volume_t volume;
1560 volume = hammer_get_root_volume(hmp, &error);
1561 KKASSERT(error == 0);
1563 hammer_modify_volume_field(cursor->trans, volume,
1564 vol0_btree_root);
1565 volume->ondisk->vol0_btree_root = parent->node_offset;
1566 hammer_modify_volume_done(volume);
1567 node->ondisk->parent = parent->node_offset;
1568 if (cursor->parent) {
1569 hammer_unlock(&cursor->parent->lock);
1570 hammer_rel_node(cursor->parent);
1572 cursor->parent = parent; /* lock'd and ref'd */
1573 hammer_rel_volume(volume, 0);
1575 hammer_modify_node_done(node);
1578 * Ok, now adjust the cursor depending on which element the original
1579 * index was pointing at. If we are >= the split point the push node
1580 * is now in the new node.
1582 * NOTE: If we are at the split point itself we cannot stay with the
1583 * original node because the push index will point at the right-hand
1584 * boundary, which is illegal.
1586 * NOTE: The cursor's parent or parent_index must be adjusted for
1587 * the case where a new parent (new root) was created, and the case
1588 * where the cursor is now pointing at the split node.
1590 if (cursor->index >= split) {
1591 cursor->parent_index = parent_index + 1;
1592 cursor->index -= split;
1593 hammer_unlock(&cursor->node->lock);
1594 hammer_rel_node(cursor->node);
1595 cursor->node = new_node; /* locked and ref'd */
1596 } else {
1597 cursor->parent_index = parent_index;
1598 hammer_unlock(&new_node->lock);
1599 hammer_rel_node(new_node);
1603 * Fixup left and right bounds
1605 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1606 cursor->left_bound = &parent_elm[0].internal.base;
1607 cursor->right_bound = &parent_elm[1].internal.base;
1608 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1609 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1610 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1611 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1613 done:
1614 hammer_btree_unlock_children(&locklist);
1615 hammer_cursor_downgrade(cursor);
1616 return (error);
1620 * Same as the above, but splits a full leaf node.
1622 * This function
1624 static
1626 btree_split_leaf(hammer_cursor_t cursor)
1628 hammer_node_ondisk_t ondisk;
1629 hammer_node_t parent;
1630 hammer_node_t leaf;
1631 hammer_mount_t hmp;
1632 hammer_node_t new_leaf;
1633 hammer_btree_elm_t elm;
1634 hammer_btree_elm_t parent_elm;
1635 hammer_base_elm_t mid_boundary;
1636 int parent_index;
1637 int made_root;
1638 int split;
1639 int error;
1640 const size_t esize = sizeof(*elm);
1642 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1643 return(error);
1644 ++hammer_stats_btree_splits;
1646 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1647 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1648 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1649 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1652 * Calculate the split point. If the insertion point will be on
1653 * the left-hand side adjust the split point to give the right
1654 * hand side one additional node.
1656 * Spikes are made up of two leaf elements which cannot be
1657 * safely split.
1659 leaf = cursor->node;
1660 ondisk = leaf->ondisk;
1661 split = (ondisk->count + 1) / 2;
1662 if (cursor->index <= split)
1663 --split;
1664 error = 0;
1665 hmp = leaf->hmp;
1667 elm = &ondisk->elms[split];
1669 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1670 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1671 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1672 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1675 * If we are at the root of the tree, create a new root node with
1676 * 1 element and split normally. Avoid making major modifications
1677 * until we know the whole operation will work.
1679 if (ondisk->parent == 0) {
1680 parent = hammer_alloc_btree(cursor->trans, &error);
1681 if (parent == NULL)
1682 goto done;
1683 hammer_lock_ex(&parent->lock);
1684 hammer_modify_node_noundo(cursor->trans, parent);
1685 ondisk = parent->ondisk;
1686 ondisk->count = 1;
1687 ondisk->parent = 0;
1688 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1689 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1690 ondisk->elms[0].base = hmp->root_btree_beg;
1691 ondisk->elms[0].base.btype = leaf->ondisk->type;
1692 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1693 ondisk->elms[1].base = hmp->root_btree_end;
1694 /* ondisk->elms[1].base.btype = not used */
1695 hammer_modify_node_done(parent);
1696 made_root = 1;
1697 parent_index = 0; /* insertion point in parent */
1698 } else {
1699 made_root = 0;
1700 parent = cursor->parent;
1701 parent_index = cursor->parent_index;
1705 * Split leaf into new_leaf at the split point. Select a separator
1706 * value in-between the two leafs but with a bent towards the right
1707 * leaf since comparisons use an 'elm >= separator' inequality.
1709 * L L L L L L L L
1711 * x x P x x
1712 * s S S s
1713 * / \
1714 * L L L L L L L L
1716 new_leaf = hammer_alloc_btree(cursor->trans, &error);
1717 if (new_leaf == NULL) {
1718 if (made_root) {
1719 hammer_unlock(&parent->lock);
1720 hammer_delete_node(cursor->trans, parent);
1721 hammer_rel_node(parent);
1723 goto done;
1725 hammer_lock_ex(&new_leaf->lock);
1728 * Create the new node and copy the leaf elements from the split
1729 * point on to the new node.
1731 hammer_modify_node_all(cursor->trans, leaf);
1732 hammer_modify_node_noundo(cursor->trans, new_leaf);
1733 ondisk = leaf->ondisk;
1734 elm = &ondisk->elms[split];
1735 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1736 new_leaf->ondisk->count = ondisk->count - split;
1737 new_leaf->ondisk->parent = parent->node_offset;
1738 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1739 new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1740 KKASSERT(ondisk->type == new_leaf->ondisk->type);
1741 hammer_modify_node_done(new_leaf);
1742 hammer_cursor_split_node(leaf, new_leaf, split);
1745 * Cleanup the original node. Because this is a leaf node and
1746 * leaf nodes do not have a right-hand boundary, there
1747 * aren't any special edge cases to clean up. We just fixup the
1748 * count.
1750 ondisk->count = split;
1753 * Insert the separator into the parent, fixup the parent's
1754 * reference to the original node, and reference the new node.
1755 * The separator is P.
1757 * Remember that base.count does not include the right-hand boundary.
1758 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1760 hammer_modify_node_all(cursor->trans, parent);
1761 ondisk = parent->ondisk;
1762 KKASSERT(split != 0);
1763 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1764 parent_elm = &ondisk->elms[parent_index+1];
1765 bcopy(parent_elm, parent_elm + 1,
1766 (ondisk->count - parent_index) * esize);
1768 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1769 parent_elm->internal.base.btype = new_leaf->ondisk->type;
1770 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1771 parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1772 mid_boundary = &parent_elm->base;
1773 ++ondisk->count;
1774 hammer_modify_node_done(parent);
1775 hammer_cursor_inserted_element(parent, parent_index + 1);
1778 * The filesystem's root B-Tree pointer may have to be updated.
1780 if (made_root) {
1781 hammer_volume_t volume;
1783 volume = hammer_get_root_volume(hmp, &error);
1784 KKASSERT(error == 0);
1786 hammer_modify_volume_field(cursor->trans, volume,
1787 vol0_btree_root);
1788 volume->ondisk->vol0_btree_root = parent->node_offset;
1789 hammer_modify_volume_done(volume);
1790 leaf->ondisk->parent = parent->node_offset;
1791 if (cursor->parent) {
1792 hammer_unlock(&cursor->parent->lock);
1793 hammer_rel_node(cursor->parent);
1795 cursor->parent = parent; /* lock'd and ref'd */
1796 hammer_rel_volume(volume, 0);
1798 hammer_modify_node_done(leaf);
1801 * Ok, now adjust the cursor depending on which element the original
1802 * index was pointing at. If we are >= the split point the push node
1803 * is now in the new node.
1805 * NOTE: If we are at the split point itself we need to select the
1806 * old or new node based on where key_beg's insertion point will be.
1807 * If we pick the wrong side the inserted element will wind up in
1808 * the wrong leaf node and outside that node's bounds.
1810 if (cursor->index > split ||
1811 (cursor->index == split &&
1812 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
1813 cursor->parent_index = parent_index + 1;
1814 cursor->index -= split;
1815 hammer_unlock(&cursor->node->lock);
1816 hammer_rel_node(cursor->node);
1817 cursor->node = new_leaf;
1818 } else {
1819 cursor->parent_index = parent_index;
1820 hammer_unlock(&new_leaf->lock);
1821 hammer_rel_node(new_leaf);
1825 * Fixup left and right bounds
1827 parent_elm = &parent->ondisk->elms[cursor->parent_index];
1828 cursor->left_bound = &parent_elm[0].internal.base;
1829 cursor->right_bound = &parent_elm[1].internal.base;
1832 * Assert that the bounds are correct.
1834 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1835 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1836 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1837 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1838 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1839 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
1841 done:
1842 hammer_cursor_downgrade(cursor);
1843 return (error);
1846 #if 0
1849 * Recursively correct the right-hand boundary's create_tid to (tid) as
1850 * long as the rest of the key matches. We have to recurse upward in
1851 * the tree as well as down the left side of each parent's right node.
1853 * Return EDEADLK if we were only partially successful, forcing the caller
1854 * to try again. The original cursor is not modified. This routine can
1855 * also fail with EDEADLK if it is forced to throw away a portion of its
1856 * record history.
1858 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1860 struct hammer_rhb {
1861 TAILQ_ENTRY(hammer_rhb) entry;
1862 hammer_node_t node;
1863 int index;
1866 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1869 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1871 struct hammer_rhb_list rhb_list;
1872 hammer_base_elm_t elm;
1873 hammer_node_t orig_node;
1874 struct hammer_rhb *rhb;
1875 int orig_index;
1876 int error;
1878 TAILQ_INIT(&rhb_list);
1881 * Save our position so we can restore it on return. This also
1882 * gives us a stable 'elm'.
1884 orig_node = cursor->node;
1885 hammer_ref_node(orig_node);
1886 hammer_lock_sh(&orig_node->lock);
1887 orig_index = cursor->index;
1888 elm = &orig_node->ondisk->elms[orig_index].base;
1891 * Now build a list of parents going up, allocating a rhb
1892 * structure for each one.
1894 while (cursor->parent) {
1896 * Stop if we no longer have any right-bounds to fix up
1898 if (elm->obj_id != cursor->right_bound->obj_id ||
1899 elm->rec_type != cursor->right_bound->rec_type ||
1900 elm->key != cursor->right_bound->key) {
1901 break;
1905 * Stop if the right-hand bound's create_tid does not
1906 * need to be corrected.
1908 if (cursor->right_bound->create_tid >= tid)
1909 break;
1911 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1912 rhb->node = cursor->parent;
1913 rhb->index = cursor->parent_index;
1914 hammer_ref_node(rhb->node);
1915 hammer_lock_sh(&rhb->node->lock);
1916 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1918 hammer_cursor_up(cursor);
1922 * now safely adjust the right hand bound for each rhb. This may
1923 * also require taking the right side of the tree and iterating down
1924 * ITS left side.
1926 error = 0;
1927 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1928 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1929 if (error)
1930 break;
1931 TAILQ_REMOVE(&rhb_list, rhb, entry);
1932 hammer_unlock(&rhb->node->lock);
1933 hammer_rel_node(rhb->node);
1934 kfree(rhb, M_HAMMER);
1936 switch (cursor->node->ondisk->type) {
1937 case HAMMER_BTREE_TYPE_INTERNAL:
1939 * Right-boundary for parent at internal node
1940 * is one element to the right of the element whos
1941 * right boundary needs adjusting. We must then
1942 * traverse down the left side correcting any left
1943 * bounds (which may now be too far to the left).
1945 ++cursor->index;
1946 error = hammer_btree_correct_lhb(cursor, tid);
1947 break;
1948 default:
1949 panic("hammer_btree_correct_rhb(): Bad node type");
1950 error = EINVAL;
1951 break;
1956 * Cleanup
1958 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1959 TAILQ_REMOVE(&rhb_list, rhb, entry);
1960 hammer_unlock(&rhb->node->lock);
1961 hammer_rel_node(rhb->node);
1962 kfree(rhb, M_HAMMER);
1964 error = hammer_cursor_seek(cursor, orig_node, orig_index);
1965 hammer_unlock(&orig_node->lock);
1966 hammer_rel_node(orig_node);
1967 return (error);
1971 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1972 * bound going downward starting at the current cursor position.
1974 * This function does not restore the cursor after use.
1977 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1979 struct hammer_rhb_list rhb_list;
1980 hammer_base_elm_t elm;
1981 hammer_base_elm_t cmp;
1982 struct hammer_rhb *rhb;
1983 int error;
1985 TAILQ_INIT(&rhb_list);
1987 cmp = &cursor->node->ondisk->elms[cursor->index].base;
1990 * Record the node and traverse down the left-hand side for all
1991 * matching records needing a boundary correction.
1993 error = 0;
1994 for (;;) {
1995 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1996 rhb->node = cursor->node;
1997 rhb->index = cursor->index;
1998 hammer_ref_node(rhb->node);
1999 hammer_lock_sh(&rhb->node->lock);
2000 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2002 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2004 * Nothing to traverse down if we are at the right
2005 * boundary of an internal node.
2007 if (cursor->index == cursor->node->ondisk->count)
2008 break;
2009 } else {
2010 elm = &cursor->node->ondisk->elms[cursor->index].base;
2011 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2012 break;
2013 panic("Illegal leaf record type %02x", elm->btype);
2015 error = hammer_cursor_down(cursor);
2016 if (error)
2017 break;
2019 elm = &cursor->node->ondisk->elms[cursor->index].base;
2020 if (elm->obj_id != cmp->obj_id ||
2021 elm->rec_type != cmp->rec_type ||
2022 elm->key != cmp->key) {
2023 break;
2025 if (elm->create_tid >= tid)
2026 break;
2031 * Now we can safely adjust the left-hand boundary from the bottom-up.
2032 * The last element we remove from the list is the caller's right hand
2033 * boundary, which must also be adjusted.
2035 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2036 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2037 if (error)
2038 break;
2039 TAILQ_REMOVE(&rhb_list, rhb, entry);
2040 hammer_unlock(&rhb->node->lock);
2041 hammer_rel_node(rhb->node);
2042 kfree(rhb, M_HAMMER);
2044 elm = &cursor->node->ondisk->elms[cursor->index].base;
2045 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2046 hammer_modify_node(cursor->trans, cursor->node,
2047 &elm->create_tid,
2048 sizeof(elm->create_tid));
2049 elm->create_tid = tid;
2050 hammer_modify_node_done(cursor->node);
2051 } else {
2052 panic("hammer_btree_correct_lhb(): Bad element type");
2057 * Cleanup
2059 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2060 TAILQ_REMOVE(&rhb_list, rhb, entry);
2061 hammer_unlock(&rhb->node->lock);
2062 hammer_rel_node(rhb->node);
2063 kfree(rhb, M_HAMMER);
2065 return (error);
2068 #endif
2071 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2072 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2073 * the operation due to a deadlock, or some other error.
2075 * This routine is always called with an empty, locked leaf but may recurse
2076 * into want-to-be-empty parents as part of its operation.
2078 * It should also be noted that when removing empty leaves we must be sure
2079 * to test and update mirror_tid because another thread may have deadlocked
2080 * against us (or someone) trying to propagate it up and cannot retry once
2081 * the node has been deleted.
2083 * On return the cursor may end up pointing to an internal node, suitable
2084 * for further iteration but not for an immediate insertion or deletion.
2086 static int
2087 btree_remove(hammer_cursor_t cursor)
2089 hammer_node_ondisk_t ondisk;
2090 hammer_btree_elm_t elm;
2091 hammer_node_t node;
2092 hammer_node_t parent;
2093 const int esize = sizeof(*elm);
2094 int error;
2096 node = cursor->node;
2099 * When deleting the root of the filesystem convert it to
2100 * an empty leaf node. Internal nodes cannot be empty.
2102 ondisk = node->ondisk;
2103 if (ondisk->parent == 0) {
2104 KKASSERT(cursor->parent == NULL);
2105 hammer_modify_node_all(cursor->trans, node);
2106 KKASSERT(ondisk == node->ondisk);
2107 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2108 ondisk->count = 0;
2109 hammer_modify_node_done(node);
2110 cursor->index = 0;
2111 return(0);
2114 parent = cursor->parent;
2115 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2118 * Attempt to remove the parent's reference to the child. If the
2119 * parent would become empty we have to recurse. If we fail we
2120 * leave the parent pointing to an empty leaf node.
2122 if (parent->ondisk->count == 1) {
2124 * This special cursor_up_locked() call leaves the original
2125 * node exclusively locked and referenced, leaves the
2126 * original parent locked (as the new node), and locks the
2127 * new parent. It can return EDEADLK.
2129 error = hammer_cursor_up_locked(cursor);
2130 if (error == 0) {
2131 error = btree_remove(cursor);
2132 if (error == 0) {
2133 hammer_modify_node_all(cursor->trans, node);
2134 ondisk = node->ondisk;
2135 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2136 ondisk->count = 0;
2137 hammer_modify_node_done(node);
2138 hammer_flush_node(node);
2139 hammer_delete_node(cursor->trans, node);
2140 } else {
2141 kprintf("Warning: BTREE_REMOVE: Defering "
2142 "parent removal1 @ %016llx, skipping\n",
2143 node->node_offset);
2145 hammer_unlock(&node->lock);
2146 hammer_rel_node(node);
2147 } else {
2148 kprintf("Warning: BTREE_REMOVE: Defering parent "
2149 "removal2 @ %016llx, skipping\n",
2150 node->node_offset);
2152 } else {
2153 KKASSERT(parent->ondisk->count > 1);
2155 hammer_modify_node_all(cursor->trans, parent);
2156 ondisk = parent->ondisk;
2157 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2159 elm = &ondisk->elms[cursor->parent_index];
2160 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2161 KKASSERT(ondisk->count > 0);
2164 * We must retain the highest mirror_tid. The deleted
2165 * range is now encompassed by the element to the left.
2166 * If we are already at the left edge the new left edge
2167 * inherits mirror_tid.
2169 * Note that bounds of the parent to our parent may create
2170 * a gap to the left of our left-most node or to the right
2171 * of our right-most node. The gap is silently included
2172 * in the mirror_tid's area of effect from the point of view
2173 * of the scan.
2175 if (cursor->parent_index) {
2176 if (elm[-1].internal.mirror_tid <
2177 elm[0].internal.mirror_tid) {
2178 elm[-1].internal.mirror_tid =
2179 elm[0].internal.mirror_tid;
2181 } else {
2182 if (elm[1].internal.mirror_tid <
2183 elm[0].internal.mirror_tid) {
2184 elm[1].internal.mirror_tid =
2185 elm[0].internal.mirror_tid;
2190 * Delete the subtree reference in the parent
2192 bcopy(&elm[1], &elm[0],
2193 (ondisk->count - cursor->parent_index) * esize);
2194 --ondisk->count;
2195 hammer_modify_node_done(parent);
2196 hammer_cursor_deleted_element(parent, cursor->parent_index);
2197 hammer_flush_node(node);
2198 hammer_delete_node(cursor->trans, node);
2201 * cursor->node is invalid, cursor up to make the cursor
2202 * valid again.
2204 error = hammer_cursor_up(cursor);
2206 return (error);
2210 * Propagate cursor->trans->tid up the B-Tree starting at the current
2211 * cursor position using pseudofs info gleaned from the passed inode.
2213 * The passed inode has no relationship to the cursor position other
2214 * then being in the same pseudofs as the insertion or deletion we
2215 * are propagating the mirror_tid for.
2217 void
2218 hammer_btree_do_propagation(hammer_cursor_t cursor,
2219 hammer_pseudofs_inmem_t pfsm,
2220 hammer_btree_leaf_elm_t leaf)
2222 hammer_cursor_t ncursor;
2223 hammer_tid_t mirror_tid;
2224 int error;
2227 * We do not propagate a mirror_tid if the filesystem was mounted
2228 * in no-mirror mode.
2230 if (cursor->trans->hmp->master_id < 0)
2231 return;
2234 * This is a bit of a hack because we cannot deadlock or return
2235 * EDEADLK here. The related operation has already completed and
2236 * we must propagate the mirror_tid now regardless.
2238 * Generate a new cursor which inherits the original's locks and
2239 * unlock the original. Use the new cursor to propagate the
2240 * mirror_tid. Then clean up the new cursor and reacquire locks
2241 * on the original.
2243 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2244 * original's locks and the original is tracked and must be
2245 * re-locked.
2247 mirror_tid = cursor->node->ondisk->mirror_tid;
2248 KKASSERT(mirror_tid != 0);
2249 ncursor = hammer_push_cursor(cursor);
2250 error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2251 KKASSERT(error == 0);
2252 hammer_pop_cursor(cursor, ncursor);
2257 * Propagate a mirror TID update upwards through the B-Tree to the root.
2259 * A locked internal node must be passed in. The node will remain locked
2260 * on return.
2262 * This function syncs mirror_tid at the specified internal node's element,
2263 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2265 static int
2266 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2268 hammer_btree_internal_elm_t elm;
2269 hammer_node_t node;
2270 int error;
2272 for (;;) {
2273 error = hammer_cursor_up(cursor);
2274 if (error == 0)
2275 error = hammer_cursor_upgrade(cursor);
2276 while (error == EDEADLK) {
2277 hammer_recover_cursor(cursor);
2278 error = hammer_cursor_upgrade(cursor);
2280 if (error)
2281 break;
2282 node = cursor->node;
2283 KKASSERT (node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2286 * Adjust the node's element
2288 elm = &node->ondisk->elms[cursor->index].internal;
2289 if (elm->mirror_tid >= mirror_tid)
2290 break;
2291 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2292 sizeof(elm->mirror_tid));
2293 elm->mirror_tid = mirror_tid;
2294 hammer_modify_node_done(node);
2295 if (hammer_debug_general & 0x0002) {
2296 kprintf("mirror_propagate: propagate "
2297 "%016llx @%016llx:%d\n",
2298 mirror_tid, node->node_offset, cursor->index);
2303 * Adjust the node's mirror_tid aggregator
2305 if (node->ondisk->mirror_tid >= mirror_tid)
2306 return(0);
2307 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2308 node->ondisk->mirror_tid = mirror_tid;
2309 hammer_modify_node_done(node);
2310 if (hammer_debug_general & 0x0002) {
2311 kprintf("mirror_propagate: propagate "
2312 "%016llx @%016llx\n",
2313 mirror_tid, node->node_offset);
2316 if (error == ENOENT)
2317 error = 0;
2318 return(error);
2321 hammer_node_t
2322 hammer_btree_get_parent(hammer_node_t node, int *parent_indexp, int *errorp,
2323 int try_exclusive)
2325 hammer_node_t parent;
2326 hammer_btree_elm_t elm;
2327 int i;
2330 * Get the node
2332 parent = hammer_get_node(node->hmp, node->ondisk->parent, 0, errorp);
2333 if (*errorp) {
2334 KKASSERT(parent == NULL);
2335 return(NULL);
2337 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2340 * Lock the node
2342 if (try_exclusive) {
2343 if (hammer_lock_ex_try(&parent->lock)) {
2344 hammer_rel_node(parent);
2345 *errorp = EDEADLK;
2346 return(NULL);
2348 } else {
2349 hammer_lock_sh(&parent->lock);
2353 * Figure out which element in the parent is pointing to the
2354 * child.
2356 if (node->ondisk->count) {
2357 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2358 parent->ondisk);
2359 } else {
2360 i = 0;
2362 while (i < parent->ondisk->count) {
2363 elm = &parent->ondisk->elms[i];
2364 if (elm->internal.subtree_offset == node->node_offset)
2365 break;
2366 ++i;
2368 if (i == parent->ondisk->count) {
2369 hammer_unlock(&parent->lock);
2370 panic("Bad B-Tree link: parent %p node %p\n", parent, node);
2372 *parent_indexp = i;
2373 KKASSERT(*errorp == 0);
2374 return(parent);
2378 * The element (elm) has been moved to a new internal node (node).
2380 * If the element represents a pointer to an internal node that node's
2381 * parent must be adjusted to the element's new location.
2383 * XXX deadlock potential here with our exclusive locks
2386 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2387 hammer_btree_elm_t elm)
2389 hammer_node_t child;
2390 int error;
2392 error = 0;
2394 switch(elm->base.btype) {
2395 case HAMMER_BTREE_TYPE_INTERNAL:
2396 case HAMMER_BTREE_TYPE_LEAF:
2397 child = hammer_get_node(node->hmp, elm->internal.subtree_offset,
2398 0, &error);
2399 if (error == 0) {
2400 hammer_modify_node_field(trans, child, parent);
2401 child->ondisk->parent = node->node_offset;
2402 hammer_modify_node_done(child);
2403 hammer_rel_node(child);
2405 break;
2406 default:
2407 break;
2409 return(error);
2413 * Exclusively lock all the children of node. This is used by the split
2414 * code to prevent anyone from accessing the children of a cursor node
2415 * while we fix-up its parent offset.
2417 * If we don't lock the children we can really mess up cursors which block
2418 * trying to cursor-up into our node.
2420 * On failure EDEADLK (or some other error) is returned. If a deadlock
2421 * error is returned the cursor is adjusted to block on termination.
2424 hammer_btree_lock_children(hammer_cursor_t cursor,
2425 struct hammer_node_locklist **locklistp)
2427 hammer_node_t node;
2428 hammer_node_locklist_t item;
2429 hammer_node_ondisk_t ondisk;
2430 hammer_btree_elm_t elm;
2431 hammer_node_t child;
2432 int error;
2433 int i;
2435 node = cursor->node;
2436 ondisk = node->ondisk;
2437 error = 0;
2440 * We really do not want to block on I/O with exclusive locks held,
2441 * pre-get the children before trying to lock the mess.
2443 for (i = 0; i < ondisk->count; ++i) {
2444 ++hammer_stats_btree_elements;
2445 elm = &ondisk->elms[i];
2446 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2447 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2448 continue;
2450 child = hammer_get_node(node->hmp,
2451 elm->internal.subtree_offset,
2452 0, &error);
2453 if (child)
2454 hammer_rel_node(child);
2458 * Do it for real
2460 for (i = 0; error == 0 && i < ondisk->count; ++i) {
2461 ++hammer_stats_btree_elements;
2462 elm = &ondisk->elms[i];
2464 switch(elm->base.btype) {
2465 case HAMMER_BTREE_TYPE_INTERNAL:
2466 case HAMMER_BTREE_TYPE_LEAF:
2467 KKASSERT(elm->internal.subtree_offset != 0);
2468 child = hammer_get_node(node->hmp,
2469 elm->internal.subtree_offset,
2470 0, &error);
2471 break;
2472 default:
2473 child = NULL;
2474 break;
2476 if (child) {
2477 if (hammer_lock_ex_try(&child->lock) != 0) {
2478 if (cursor->deadlk_node == NULL) {
2479 cursor->deadlk_node = child;
2480 hammer_ref_node(cursor->deadlk_node);
2482 error = EDEADLK;
2483 hammer_rel_node(child);
2484 } else {
2485 item = kmalloc(sizeof(*item),
2486 M_HAMMER, M_WAITOK);
2487 item->next = *locklistp;
2488 item->node = child;
2489 *locklistp = item;
2493 if (error)
2494 hammer_btree_unlock_children(locklistp);
2495 return(error);
2500 * Release previously obtained node locks.
2502 void
2503 hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2505 hammer_node_locklist_t item;
2507 while ((item = *locklistp) != NULL) {
2508 *locklistp = item->next;
2509 hammer_unlock(&item->node->lock);
2510 hammer_rel_node(item->node);
2511 kfree(item, M_HAMMER);
2515 /************************************************************************
2516 * MISCELLANIOUS SUPPORT *
2517 ************************************************************************/
2520 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2522 * Note that for this particular function a return value of -1, 0, or +1
2523 * can denote a match if create_tid is otherwise discounted. A create_tid
2524 * of zero is considered to be 'infinity' in comparisons.
2526 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2529 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2531 if (key1->localization < key2->localization)
2532 return(-5);
2533 if (key1->localization > key2->localization)
2534 return(5);
2536 if (key1->obj_id < key2->obj_id)
2537 return(-4);
2538 if (key1->obj_id > key2->obj_id)
2539 return(4);
2541 if (key1->rec_type < key2->rec_type)
2542 return(-3);
2543 if (key1->rec_type > key2->rec_type)
2544 return(3);
2546 if (key1->key < key2->key)
2547 return(-2);
2548 if (key1->key > key2->key)
2549 return(2);
2552 * A create_tid of zero indicates a record which is undeletable
2553 * and must be considered to have a value of positive infinity.
2555 if (key1->create_tid == 0) {
2556 if (key2->create_tid == 0)
2557 return(0);
2558 return(1);
2560 if (key2->create_tid == 0)
2561 return(-1);
2562 if (key1->create_tid < key2->create_tid)
2563 return(-1);
2564 if (key1->create_tid > key2->create_tid)
2565 return(1);
2566 return(0);
2570 * Test a timestamp against an element to determine whether the
2571 * element is visible. A timestamp of 0 means 'infinity'.
2574 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
2576 if (asof == 0) {
2577 if (base->delete_tid)
2578 return(1);
2579 return(0);
2581 if (asof < base->create_tid)
2582 return(-1);
2583 if (base->delete_tid && asof >= base->delete_tid)
2584 return(1);
2585 return(0);
2589 * Create a separator half way inbetween key1 and key2. For fields just
2590 * one unit apart, the separator will match key2. key1 is on the left-hand
2591 * side and key2 is on the right-hand side.
2593 * key2 must be >= the separator. It is ok for the separator to match key2.
2595 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2596 * key2.
2598 * NOTE: It might be beneficial to just scrap this whole mess and just
2599 * set the separator to key2.
2601 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2602 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2604 static void
2605 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2606 hammer_base_elm_t dest)
2608 bzero(dest, sizeof(*dest));
2610 dest->rec_type = key2->rec_type;
2611 dest->key = key2->key;
2612 dest->obj_id = key2->obj_id;
2613 dest->create_tid = key2->create_tid;
2615 MAKE_SEPARATOR(key1, key2, dest, localization);
2616 if (key1->localization == key2->localization) {
2617 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2618 if (key1->obj_id == key2->obj_id) {
2619 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2620 if (key1->rec_type == key2->rec_type) {
2621 MAKE_SEPARATOR(key1, key2, dest, key);
2623 * Don't bother creating a separator for
2624 * create_tid, which also conveniently avoids
2625 * having to handle the create_tid == 0
2626 * (infinity) case. Just leave create_tid
2627 * set to key2.
2629 * Worst case, dest matches key2 exactly,
2630 * which is acceptable.
2637 #undef MAKE_SEPARATOR
2640 * Return whether a generic internal or leaf node is full
2642 static int
2643 btree_node_is_full(hammer_node_ondisk_t node)
2645 switch(node->type) {
2646 case HAMMER_BTREE_TYPE_INTERNAL:
2647 if (node->count == HAMMER_BTREE_INT_ELMS)
2648 return(1);
2649 break;
2650 case HAMMER_BTREE_TYPE_LEAF:
2651 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2652 return(1);
2653 break;
2654 default:
2655 panic("illegal btree subtype");
2657 return(0);
2660 #if 0
2661 static int
2662 btree_max_elements(u_int8_t type)
2664 if (type == HAMMER_BTREE_TYPE_LEAF)
2665 return(HAMMER_BTREE_LEAF_ELMS);
2666 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2667 return(HAMMER_BTREE_INT_ELMS);
2668 panic("btree_max_elements: bad type %d\n", type);
2670 #endif
2672 void
2673 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2675 hammer_btree_elm_t elm;
2676 int i;
2678 kprintf("node %p count=%d parent=%016llx type=%c\n",
2679 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2682 * Dump both boundary elements if an internal node
2684 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2685 for (i = 0; i <= ondisk->count; ++i) {
2686 elm = &ondisk->elms[i];
2687 hammer_print_btree_elm(elm, ondisk->type, i);
2689 } else {
2690 for (i = 0; i < ondisk->count; ++i) {
2691 elm = &ondisk->elms[i];
2692 hammer_print_btree_elm(elm, ondisk->type, i);
2697 void
2698 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2700 kprintf(" %2d", i);
2701 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
2702 kprintf("\tkey = %016llx\n", elm->base.key);
2703 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2704 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2705 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2706 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
2707 kprintf("\tbtype = %02x (%c)\n",
2708 elm->base.btype,
2709 (elm->base.btype ? elm->base.btype : '?'));
2710 kprintf("\tlocalization = %02x\n", elm->base.localization);
2712 switch(type) {
2713 case HAMMER_BTREE_TYPE_INTERNAL:
2714 kprintf("\tsubtree_off = %016llx\n",
2715 elm->internal.subtree_offset);
2716 break;
2717 case HAMMER_BTREE_TYPE_RECORD:
2718 kprintf("\tdata_offset = %016llx\n", elm->leaf.data_offset);
2719 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2720 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);
2721 break;