Adjust for the DragonFly ELF ABI
[dragonfly.git] / sys / vfs / hammer / hammer_btree.c
CommitLineData
427e5fc6
MD
1/*
2 * Copyright (c) 2007 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:
10 *
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.
20 *
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.
33 *
9944ae54 34 * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.18 2008/01/15 06:02:57 dillon Exp $
427e5fc6
MD
35 */
36
37/*
8cd0a023 38 * HAMMER B-Tree index
427e5fc6
MD
39 *
40 * HAMMER implements a modified B+Tree. In documentation this will
9944ae54 41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
427e5fc6
MD
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
8cd0a023 45 * otherwords, we have boundary elements at the two ends of a B-Tree node
427e5fc6
MD
46 * instead of sub-tree pointers.
47 *
8cd0a023 48 * A B-Tree internal node looks like this:
427e5fc6
MD
49 *
50 * B N N N N N N B <-- boundary and internal elements
51 * S S S S S S S <-- subtree pointers
52 *
8cd0a023 53 * A B-Tree leaf node basically looks like this:
427e5fc6
MD
54 *
55 * L L L L L L L L <-- leaf elemenets
56 *
8cd0a023
MD
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.
427e5fc6 59 *
8cd0a023
MD
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.
427e5fc6 67 *
8cd0a023
MD
68 * B-Trees also make the stacking of trees fairly straightforward.
69 *
70 * INTER-CLUSTER ELEMENTS: An element of an internal node may reference
71 * the root of another cluster rather then a node in the current cluster.
72 * This is known as an inter-cluster references. Only B-Tree searches
73 * will cross cluster boundaries. The rebalancing and collapse code does
74 * not attempt to move children between clusters. A major effect of this
75 * is that we have to relax minimum element count requirements and allow
76 * trees to become somewhat unabalanced.
77 *
78 * INSERTIONS AND DELETIONS: When inserting we split full nodes on our
79 * way down as an optimization. I originally experimented with rebalancing
80 * nodes on the way down for deletions but it created a huge mess due to
81 * the way inter-cluster linkages work. Instead, now I simply allow
82 * the tree to become unbalanced and allow leaf nodes to become empty.
83 * The delete code will try to clean things up from the bottom-up but
84 * will stop if related elements are not in-core or if it cannot get a node
85 * lock.
86 */
87#include "hammer.h"
88#include <sys/buf.h>
89#include <sys/buf2.h>
66325755 90
9944ae54
MD
91typedef enum btree_search_edge {
92 SEARCH_NONE,
93 SEARCH_LEFT_EDGE,
94 SEARCH_RIGHT_EDGE
95} btree_search_edge_t;
96
8cd0a023 97static int btree_search(hammer_cursor_t cursor, int flags);
9944ae54
MD
98static int btree_edge_internal(hammer_cursor_t cursor,
99 btree_search_edge_t edge);
8cd0a023
MD
100static int btree_split_internal(hammer_cursor_t cursor);
101static int btree_split_leaf(hammer_cursor_t cursor);
195c19a1 102static int btree_remove(hammer_cursor_t cursor);
7f7c1f84 103static int btree_set_parent(hammer_node_t node, hammer_btree_elm_t elm);
8cd0a023
MD
104#if 0
105static int btree_rebalance(hammer_cursor_t cursor);
106static int btree_collapse(hammer_cursor_t cursor);
107#endif
9944ae54 108static int btree_node_is_almost_full(hammer_node_ondisk_t node);
8cd0a023
MD
109static void hammer_make_separator(hammer_base_elm_t key1,
110 hammer_base_elm_t key2, hammer_base_elm_t dest);
66325755
MD
111
112/*
8cd0a023
MD
113 * Iterate records after a search. The cursor is iterated forwards past
114 * the current record until a record matching the key-range requirements
115 * is found. ENOENT is returned if the iteration goes past the ending
116 * key.
66325755 117 *
d26d0ae9
MD
118 * The iteration is inclusive of key_beg and can be inclusive or exclusive
119 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
66325755 120 *
8cd0a023 121 * cursor->key_beg may or may not be modified by this function during
d26d0ae9
MD
122 * the iteration. XXX future - in case of an inverted lock we may have
123 * to reinitiate the lookup and set key_beg to properly pick up where we
124 * left off.
66325755
MD
125 */
126int
8cd0a023 127hammer_btree_iterate(hammer_cursor_t cursor)
66325755 128{
8cd0a023
MD
129 hammer_node_ondisk_t node;
130 hammer_btree_elm_t elm;
66325755
MD
131 int error;
132 int r;
133 int s;
134
135 /*
8cd0a023 136 * Skip past the current record
66325755 137 */
8cd0a023 138 node = cursor->node->ondisk;
a89aec1b
MD
139 if (node == NULL)
140 return(ENOENT);
c0ade690
MD
141 if (cursor->index < node->count &&
142 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
66325755 143 ++cursor->index;
c0ade690 144 }
66325755 145
8cd0a023
MD
146 /*
147 * Loop until an element is found or we are done.
148 */
66325755
MD
149 for (;;) {
150 /*
8cd0a023
MD
151 * We iterate up the tree and then index over one element
152 * while we are at the last element in the current node.
153 *
154 * NOTE: This can pop us up to another cluster.
66325755 155 *
8cd0a023
MD
156 * If we are at the root of the root cluster, cursor_up
157 * returns ENOENT.
158 *
159 * NOTE: hammer_cursor_up() will adjust cursor->key_beg
160 * when told to re-search for the cluster tag.
66325755
MD
161 *
162 * XXX this could be optimized by storing the information in
163 * the parent reference.
195c19a1
MD
164 *
165 * XXX we can lose the node lock temporarily, this could mess
166 * up our scan.
66325755 167 */
8cd0a023 168 if (cursor->index == node->count) {
195c19a1 169 error = hammer_cursor_up(cursor, 0);
8cd0a023
MD
170 if (error)
171 break;
172 node = cursor->node->ondisk;
173 KKASSERT(cursor->index != node->count);
174 ++cursor->index;
175 continue;
66325755
MD
176 }
177
178 /*
d26d0ae9
MD
179 * Check internal or leaf element. Determine if the record
180 * at the cursor has gone beyond the end of our range.
66325755 181 *
d26d0ae9
MD
182 * Generally we recurse down through internal nodes. An
183 * internal node can only be returned if INCLUSTER is set
184 * and the node represents a cluster-push record. Internal
185 * elements do not contain create_tid/delete_tid information.
66325755 186 */
8cd0a023
MD
187 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
188 elm = &node->elms[cursor->index];
d26d0ae9
MD
189 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
190 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
b3deaf57
MD
191 if (hammer_debug_btree) {
192 kprintf("BRACKETL %p:%d %016llx %02x %016llx %d\n",
193 cursor->node, cursor->index,
194 elm[0].internal.base.obj_id,
195 elm[0].internal.base.rec_type,
196 elm[0].internal.base.key,
197 r
198 );
199 kprintf("BRACKETR %p:%d %016llx %02x %016llx %d\n",
200 cursor->node, cursor->index + 1,
201 elm[1].internal.base.obj_id,
202 elm[1].internal.base.rec_type,
203 elm[1].internal.base.key,
204 s
205 );
206 }
207
d26d0ae9
MD
208 if (r < 0) {
209 error = ENOENT;
210 break;
66325755 211 }
d26d0ae9
MD
212 if (r == 0 && (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
213 error = ENOENT;
8cd0a023 214 break;
d26d0ae9
MD
215 }
216 KKASSERT(s <= 0);
217 if ((cursor->flags & HAMMER_CURSOR_INCLUSTER) == 0 ||
218 elm->internal.rec_offset == 0) {
219 error = hammer_cursor_down(cursor);
220 if (error)
221 break;
222 KKASSERT(cursor->index == 0);
223 node = cursor->node->ondisk;
224 continue;
225 }
226 } else {
227 elm = &node->elms[cursor->index];
228 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
b3deaf57
MD
229 if (hammer_debug_btree) {
230 kprintf("ELEMENT %p:%d %016llx %02x %016llx %d\n",
231 cursor->node, cursor->index,
232 elm[0].leaf.base.obj_id,
233 elm[0].leaf.base.rec_type,
234 elm[0].leaf.base.key,
235 r
236 );
237 }
d26d0ae9
MD
238 if (r < 0) {
239 error = ENOENT;
240 break;
241 }
242 if (r == 0 && (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
243 error = ENOENT;
244 break;
245 }
246 if ((cursor->flags & HAMMER_CURSOR_ALLHISTORY) == 0 &&
247 hammer_btree_chkts(cursor->key_beg.create_tid,
248 &elm->base) != 0) {
249 ++cursor->index;
250 continue;
251 }
66325755
MD
252 }
253
254 /*
d26d0ae9 255 * Return entry
66325755 256 */
b3deaf57
MD
257 if (hammer_debug_btree) {
258 int i = cursor->index;
259 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
260 kprintf("ITERATE %p:%d %016llx %02x %016llx\n",
261 cursor->node, i,
262 elm->internal.base.obj_id,
263 elm->internal.base.rec_type,
264 elm->internal.base.key
265 );
266 }
d26d0ae9 267 return(0);
427e5fc6 268 }
66325755 269 return(error);
427e5fc6
MD
270}
271
272/*
8cd0a023
MD
273 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
274 * could not be found, and a fatal error otherwise.
275 *
276 * The cursor is suitably positioned for a deletion on success, and suitably
277 * positioned for an insertion on ENOENT.
427e5fc6 278 *
8cd0a023
MD
279 * The cursor may begin anywhere, the search will traverse clusters in
280 * either direction to locate the requested element.
427e5fc6
MD
281 */
282int
8cd0a023 283hammer_btree_lookup(hammer_cursor_t cursor)
427e5fc6 284{
66325755
MD
285 int error;
286
8cd0a023
MD
287 error = btree_search(cursor, 0);
288 if (error == 0 && cursor->flags)
289 error = hammer_btree_extract(cursor, cursor->flags);
66325755
MD
290 return(error);
291}
292
d26d0ae9
MD
293/*
294 * Execute the logic required to start an iteration. The first record
295 * located within the specified range is returned and iteration control
296 * flags are adjusted for successive hammer_btree_iterate() calls.
297 */
298int
299hammer_btree_first(hammer_cursor_t cursor)
300{
301 int error;
302
303 error = hammer_btree_lookup(cursor);
304 if (error == ENOENT) {
305 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
306 error = hammer_btree_iterate(cursor);
307 }
308 cursor->flags |= HAMMER_CURSOR_ATEDISK;
309 return(error);
310}
311
8cd0a023
MD
312/*
313 * Extract the record and/or data associated with the cursor's current
314 * position. Any prior record or data stored in the cursor is replaced.
315 * The cursor must be positioned at a leaf node.
316 *
d26d0ae9
MD
317 * NOTE: Most extractions occur at the leaf of the B-Tree. The only
318 * extraction allowed at an internal element is at a cluster-push.
319 * Cluster-push elements have records but no data.
8cd0a023 320 */
66325755 321int
8cd0a023 322hammer_btree_extract(hammer_cursor_t cursor, int flags)
66325755 323{
8cd0a023
MD
324 hammer_node_ondisk_t node;
325 hammer_btree_elm_t elm;
326 hammer_cluster_t cluster;
c0ade690 327 u_int64_t buf_type;
427e5fc6 328 int32_t cloff;
d26d0ae9 329 int32_t roff;
427e5fc6 330 int error;
427e5fc6 331
8cd0a023 332 /*
427e5fc6 333 * A cluster record type has no data reference, the information
8cd0a023 334 * is stored directly in the record and B-Tree element.
427e5fc6
MD
335 *
336 * The case where the data reference resolves to the same buffer
337 * as the record reference must be handled.
338 */
8cd0a023 339 node = cursor->node->ondisk;
8cd0a023
MD
340 elm = &node->elms[cursor->index];
341 cluster = cursor->node->cluster;
d26d0ae9
MD
342 cursor->flags &= ~HAMMER_CURSOR_DATA_EMBEDDED;
343 cursor->data = NULL;
66325755
MD
344 error = 0;
345
d26d0ae9
MD
346 /*
347 * Internal elements can only be cluster pushes. A cluster push has
348 * no data reference.
349 */
350 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
9944ae54 351 cloff = elm->internal.rec_offset;
d26d0ae9
MD
352 KKASSERT(cloff != 0);
353 cursor->record = hammer_bread(cluster, cloff,
354 HAMMER_FSBUF_RECORDS, &error,
355 &cursor->record_buffer);
356 return(error);
357 }
358
359 /*
360 * Leaf element.
361 */
a89aec1b 362 if ((flags & HAMMER_CURSOR_GET_RECORD) && error == 0) {
8cd0a023
MD
363 cloff = elm->leaf.rec_offset;
364 cursor->record = hammer_bread(cluster, cloff,
365 HAMMER_FSBUF_RECORDS, &error,
366 &cursor->record_buffer);
427e5fc6
MD
367 } else {
368 cloff = 0;
369 }
a89aec1b 370 if ((flags & HAMMER_CURSOR_GET_DATA) && error == 0) {
8cd0a023
MD
371 if ((cloff ^ elm->leaf.data_offset) & ~HAMMER_BUFMASK) {
372 /*
c0ade690
MD
373 * The data is not in the same buffer as the last
374 * record we cached, but it could still be embedded
375 * in a record. Note that we may not have loaded the
376 * record's buffer above, depending on flags.
8cd0a023 377 */
c0ade690
MD
378 if ((elm->leaf.rec_offset ^ elm->leaf.data_offset) &
379 ~HAMMER_BUFMASK) {
380 if (elm->leaf.data_len & HAMMER_BUFMASK)
381 buf_type = HAMMER_FSBUF_DATA;
382 else
383 buf_type = 0; /* pure data buffer */
384 } else {
385 buf_type = HAMMER_FSBUF_RECORDS;
386 }
8cd0a023
MD
387 cursor->data = hammer_bread(cluster,
388 elm->leaf.data_offset,
c0ade690 389 buf_type, &error,
8cd0a023 390 &cursor->data_buffer);
427e5fc6 391 } else {
8cd0a023
MD
392 /*
393 * Data in same buffer as record. Note that we
394 * leave any existing data_buffer intact, even
395 * though we don't use it in this case, in case
396 * other records extracted during an iteration
397 * go back to it.
c0ade690 398 *
d26d0ae9
MD
399 * The data must be embedded in the record for this
400 * case to be hit.
401 *
c0ade690 402 * Just assume the buffer type is correct.
8cd0a023
MD
403 */
404 cursor->data = (void *)
405 ((char *)cursor->record_buffer->ondisk +
406 (elm->leaf.data_offset & HAMMER_BUFMASK));
d26d0ae9
MD
407 roff = (char *)cursor->data - (char *)cursor->record;
408 KKASSERT (roff >= 0 && roff < HAMMER_RECORD_SIZE);
409 cursor->flags |= HAMMER_CURSOR_DATA_EMBEDDED;
427e5fc6
MD
410 }
411 }
412 return(error);
413}
414
415
416/*
8cd0a023
MD
417 * Insert a leaf element into the B-Tree at the current cursor position.
418 * The cursor is positioned such that the element at and beyond the cursor
419 * are shifted to make room for the new record.
420 *
a89aec1b 421 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
8cd0a023
MD
422 * flag set and that call must return ENOENT before this function can be
423 * called.
424 *
425 * ENOSPC is returned if there is no room to insert a new record.
427e5fc6
MD
426 */
427int
8cd0a023 428hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_elm_t elm)
427e5fc6 429{
8cd0a023
MD
430 hammer_node_ondisk_t parent;
431 hammer_node_ondisk_t node;
427e5fc6
MD
432 int i;
433
427e5fc6
MD
434 /*
435 * Insert the element at the leaf node and update the count in the
436 * parent. It is possible for parent to be NULL, indicating that
8cd0a023 437 * the root of the B-Tree in the cluster is a leaf. It is also
427e5fc6
MD
438 * possible for the leaf to be empty.
439 *
440 * Remember that the right-hand boundary is not included in the
441 * count.
442 */
0b075555 443 hammer_modify_node(cursor->node);
8cd0a023 444 node = cursor->node->ondisk;
427e5fc6 445 i = cursor->index;
8cd0a023
MD
446 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
447 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
448 if (i != node->count) {
449 bcopy(&node->elms[i], &node->elms[i+1],
450 (node->count - i) * sizeof(*elm));
451 }
452 node->elms[i] = *elm;
453 ++node->count;
427e5fc6 454
b3deaf57
MD
455 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
456 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
457 if (i)
458 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->leaf.base) < 0);
459 if (i != node->count - 1)
460 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->leaf.base) > 0);
461
c0ade690
MD
462 /*
463 * Adjust the sub-tree count in the parent. note that the parent
464 * may be in a different cluster.
465 */
466 if (cursor->parent) {
0b075555 467 hammer_modify_node(cursor->parent);
c0ade690 468 parent = cursor->parent->ondisk;
427e5fc6 469 i = cursor->parent_index;
8cd0a023
MD
470 ++parent->elms[i].internal.subtree_count;
471 KKASSERT(parent->elms[i].internal.subtree_count <= node->count);
427e5fc6 472 }
427e5fc6
MD
473 return(0);
474}
475
9944ae54
MD
476/*
477 * Insert a cluster push into the B-Tree at the current cursor position.
478 * The cursor is positioned at a leaf after a failed btree_lookup.
479 *
480 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
481 * flag set and that call must return ENOENT before this function can be
482 * called.
483 *
484 * This routine is used ONLY during a recovery pass while the originating
485 * cluster is serialized. The leaf is broken up into up to three pieces,
486 * causing up to an additional internal elements to be added to the parent.
487 *
488 * ENOSPC is returned if there is no room to insert a new record.
489 */
490int
491hammer_btree_insert_cluster(hammer_cursor_t cursor, hammer_cluster_t ncluster,
492 int32_t rec_offset)
493{
494 hammer_cluster_t ocluster;
495 hammer_node_ondisk_t parent;
496 hammer_node_ondisk_t node;
497 hammer_node_ondisk_t xnode; /* additional leaf node */
498 hammer_node_t new_node;
499 hammer_btree_elm_t elm;
500 const int esize = sizeof(*elm);
501 u_int8_t save;
502 int error = 0;
503 int pi, i;
504
505 kprintf("cursor %p ncluster %p\n", cursor, ncluster);
506 hammer_modify_node(cursor->node);
507 node = cursor->node->ondisk;
508 i = cursor->index;
509 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
510 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
511
512 /*
513 * Make sure the spike is legal or the B-Tree code will get really
514 * confused.
515 */
516 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_beg,
517 cursor->left_bound) >= 0);
518 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
519 cursor->right_bound) <= 0);
520 if (i != node->count) {
521 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
522 &node->elms[i].leaf.base) <= 0);
523 }
524
525 /*
526 * If we are at the local root of the cluster a new root node
527 * must be created, because we need an internal node. The
528 * caller has already marked the source cluster as undergoing
529 * modification.
530 */
531 ocluster = cursor->node->cluster;
532 if (cursor->parent == NULL) {
533 cursor->parent = hammer_alloc_btree(ocluster, &error);
534 if (error)
535 return(error);
536 hammer_lock_ex(&cursor->parent->lock);
537 hammer_modify_node(cursor->parent);
538 parent = cursor->parent->ondisk;
539 parent->count = 1;
540 parent->parent = 0;
541 parent->type = HAMMER_BTREE_TYPE_INTERNAL;
542 parent->elms[0].base = ocluster->clu_btree_beg;
543 parent->elms[0].base.subtree_type = node->type;
544 parent->elms[0].internal.subtree_offset = cursor->node->node_offset;
545 parent->elms[0].internal.subtree_count = node->count;
546 parent->elms[1].base = ocluster->clu_btree_end;
547 cursor->parent_index = 0;
548 cursor->left_bound = &parent->elms[0].base;
549 cursor->right_bound = &parent->elms[1].base;
550 node->parent = cursor->parent->node_offset;
551 ocluster->ondisk->clu_btree_root = cursor->parent->node_offset;
552 kprintf("no parent\n");
553 } else {
554 kprintf("has parent\n");
555 }
556
557
558 KKASSERT(cursor->parent->ondisk->count <= HAMMER_BTREE_INT_ELMS - 2);
559
560 hammer_modify_node(cursor->parent);
561 parent = cursor->parent->ondisk;
562 pi = cursor->parent_index;
563
564 kprintf("%d node %d/%d (%c) offset=%d parent=%d\n",
565 cursor->node->cluster->clu_no,
566 i, node->count, node->type, cursor->node->node_offset, node->parent);
567
568 /*
569 * If the insertion point bisects the node we will need to allocate
570 * a second leaf node to copy the right hand side into.
571 */
572 if (i != 0 && i != node->count) {
573 new_node = hammer_alloc_btree(cursor->node->cluster, &error);
574 if (error)
575 return(error);
576 xnode = new_node->ondisk;
577 bcopy(&node->elms[i], &xnode->elms[0],
578 (node->count - i) * esize);
579 xnode->count = node->count - i;
580 xnode->parent = cursor->parent->node_offset;
581 xnode->type = HAMMER_BTREE_TYPE_LEAF;
582 node->count = i;
583 parent->elms[pi].internal.subtree_count = node->count;
584 } else {
585 new_node = NULL;
586 xnode = NULL;
587 }
588
589 /*
590 * Adjust the parent and set pi to point at the internal element
591 * which we intended to hold the spike.
592 */
593 if (new_node) {
594 /*
595 * Insert spike after parent index. Spike is at pi + 1.
596 * Also include room after the spike for new_node
597 */
598 ++pi;
599 bcopy(&parent->elms[pi], &parent->elms[pi+2],
600 (parent->count - pi + 1) * esize);
601 parent->count += 2;
602 } else if (i == 0) {
603 /*
604 * Insert spike before parent index. Spike is at pi.
605 *
606 * cursor->node's index in the parent (cursor->parent_index)
607 * has now shifted over by one.
608 */
609 bcopy(&parent->elms[pi], &parent->elms[pi+1],
610 (parent->count - pi + 1) * esize);
611 ++parent->count;
612 ++cursor->parent_index;
613 } else {
614 /*
615 * Insert spike after parent index. Spike is at pi + 1.
616 */
617 ++pi;
618 bcopy(&parent->elms[pi], &parent->elms[pi+1],
619 (parent->count - pi + 1) * esize);
620 ++parent->count;
621 }
622
623 /*
624 * Load the spike into the parent at (pi).
625 *
626 * WARNING: subtree_type is actually overloaded within base.
627 * WARNING: subtree_clu_no is overloaded on subtree_offset
628 */
629 elm = &parent->elms[pi];
630 elm[0].internal.base = ncluster->ondisk->clu_btree_beg;
631 elm[0].internal.base.subtree_type = HAMMER_BTREE_TYPE_CLUSTER;
632 elm[0].internal.rec_offset = rec_offset;
633 elm[0].internal.subtree_clu_no = ncluster->clu_no;
634 elm[0].internal.subtree_vol_no = ncluster->volume->vol_no;
635 elm[0].internal.subtree_count = 0; /* XXX */
636
637 /*
638 * Load the new node into parent at (pi+1) if non-NULL, and also
639 * set the right-hand boundary for the spike.
640 *
641 * Because new_node is a leaf its elements do not point to any
642 * nodes so we don't have to scan it to adjust parent pointers.
643 *
644 * WARNING: subtree_type is actually overloaded within base.
645 * WARNING: subtree_clu_no is overloaded on subtree_offset
646 *
647 * XXX right-boundary may not match clu_btree_end if spike is
648 * at the end of the internal node. For now the cursor search
649 * insertion code will deal with it.
650 */
651 if (new_node) {
652 elm[1].internal.base = ncluster->ondisk->clu_btree_end;
653 elm[1].internal.base.subtree_type = HAMMER_BTREE_TYPE_LEAF;
654 elm[1].internal.subtree_offset = new_node->node_offset;
655 elm[1].internal.subtree_count = xnode->count;
656 elm[1].internal.subtree_vol_no = -1;
657 elm[1].internal.rec_offset = 0;
658 } else {
659 /*
660 * The right boundary is only the base part of elm[1].
661 * The rest belongs to elm[1]'s recursion. Note however
662 * that subtree_type is overloaded within base so we
663 * have to retain it as well.
664 */
665 save = elm[1].internal.base.subtree_type;
666 elm[1].internal.base = ncluster->ondisk->clu_btree_end;
667 elm[1].internal.base.subtree_type = save;
668 }
669
670 /*
671 * The boundaries stored in the cursor for node are probably all
672 * messed up now, fix them.
673 */
674 cursor->left_bound = &parent->elms[cursor->parent_index].base;
675 cursor->right_bound = &parent->elms[cursor->parent_index+1].base;
676
677 KKASSERT(hammer_btree_cmp(&ncluster->ondisk->clu_btree_end,
678 &elm[1].internal.base) <= 0);
679
680
681 /*
682 * Adjust the target cluster's parent offset
683 */
684 hammer_modify_cluster(ncluster);
685 ncluster->ondisk->clu_btree_parent_offset = cursor->parent->node_offset;
686
687 if (new_node)
688 hammer_rel_node(new_node);
689
690 return(0);
691}
692
427e5fc6 693/*
8cd0a023
MD
694 * Delete a record from the B-Tree's at the current cursor position.
695 * The cursor is positioned such that the current element is the one
696 * to be deleted.
697 *
195c19a1
MD
698 * On return the cursor will be positioned after the deleted element and
699 * MAY point to an internal node. It will be suitable for the continuation
700 * of an iteration but not for an insertion or deletion.
8cd0a023 701 *
195c19a1
MD
702 * Deletions will attempt to partially rebalance the B-Tree in an upward
703 * direction. It is possible to end up with empty leafs. An empty internal
704 * node is impossible (worst case: it has one element pointing to an empty
705 * leaf).
427e5fc6
MD
706 */
707int
8cd0a023 708hammer_btree_delete(hammer_cursor_t cursor)
427e5fc6 709{
8cd0a023
MD
710 hammer_node_ondisk_t ondisk;
711 hammer_node_t node;
712 hammer_node_t parent;
713 hammer_btree_elm_t elm;
714 int error;
427e5fc6
MD
715 int i;
716
427e5fc6 717 /*
8cd0a023 718 * Delete the element from the leaf node.
427e5fc6 719 *
8cd0a023 720 * Remember that leaf nodes do not have boundaries.
427e5fc6 721 */
8cd0a023
MD
722 node = cursor->node;
723 ondisk = node->ondisk;
427e5fc6
MD
724 i = cursor->index;
725
8cd0a023 726 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
0b075555 727 hammer_modify_node(node);
8cd0a023
MD
728 if (i + 1 != ondisk->count) {
729 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
730 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
731 }
732 --ondisk->count;
733 if (cursor->parent != NULL) {
427e5fc6
MD
734 /*
735 * Adjust parent's notion of the leaf's count. subtree_count
8cd0a023 736 * is only approximate, it is allowed to be too small but
427e5fc6
MD
737 * never allowed to be too large. Make sure we don't drop
738 * the count below 0.
739 */
8cd0a023 740 parent = cursor->parent;
0b075555 741 hammer_modify_node(parent);
8cd0a023
MD
742 elm = &parent->ondisk->elms[cursor->parent_index];
743 if (elm->internal.subtree_count)
744 --elm->internal.subtree_count;
745 KKASSERT(elm->internal.subtree_count <= ondisk->count);
427e5fc6 746 }
427e5fc6 747
8cd0a023 748 /*
195c19a1
MD
749 * It is possible, but not desireable, to stop here. If the element
750 * count drops to 0 (which is allowed for a leaf), try recursively
751 * remove the B-Tree node.
8cd0a023 752 *
195c19a1
MD
753 * XXX rebalancing calls would go here too.
754 *
755 * This may reposition the cursor at one of the parent's of the
756 * current node.
8cd0a023 757 */
b3deaf57 758 KKASSERT(cursor->index <= ondisk->count);
8cd0a023 759 if (ondisk->count == 0) {
195c19a1
MD
760 error = btree_remove(cursor);
761 if (error == EAGAIN)
8cd0a023 762 error = 0;
8cd0a023 763 } else {
8cd0a023
MD
764 error = 0;
765 }
766 return(error);
767}
427e5fc6
MD
768
769/*
8cd0a023
MD
770 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
771 *
772 * Search a cluster's B-Tree for cursor->key_beg, return the matching node.
773 *
d26d0ae9
MD
774 * The search can begin ANYWHERE in the B-Tree. As a first step the search
775 * iterates up the tree as necessary to properly position itself prior to
776 * actually doing the sarch.
777 *
8cd0a023 778 * INSERTIONS: The search will split full nodes and leaves on its way down
d26d0ae9
MD
779 * and guarentee that the leaf it ends up on is not full. If we run out
780 * of space the search continues to the leaf (to position the cursor for
781 * the spike), but ENOSPC is returned.
427e5fc6 782 *
61aeeb33
MD
783 * XXX this isn't optimal - we really need to just locate the end point and
784 * insert space going up, and if we get a deadlock just release and retry
785 * the operation. Or something like that. The insertion code can transit
786 * multiple clusters and run splits in unnecessary clusters.
787 *
d26d0ae9 788 * DELETIONS: The search will rebalance the tree on its way down. XXX
fbc6e32a
MD
789 *
790 * The search is only guarenteed to end up on a leaf if an error code of 0
791 * is returned, or if inserting and an error code of ENOENT is returned.
d26d0ae9
MD
792 * Otherwise it can stop at an internal node. On success a search returns
793 * a leaf node unless INCLUSTER is set and the search located a cluster push
794 * node (which is an internal node).
427e5fc6 795 */
8cd0a023 796static
427e5fc6 797int
8cd0a023 798btree_search(hammer_cursor_t cursor, int flags)
427e5fc6 799{
8cd0a023
MD
800 hammer_node_ondisk_t node;
801 hammer_cluster_t cluster;
61aeeb33 802 hammer_btree_elm_t elm;
9944ae54 803 btree_search_edge_t edge;
8cd0a023 804 int error;
d26d0ae9 805 int enospc = 0;
8cd0a023
MD
806 int i;
807 int r;
808
809 flags |= cursor->flags;
810
b3deaf57 811 if (hammer_debug_btree) {
d113fda1 812 kprintf("SEARCH %p:%d %016llx %02x key=%016llx tid=%016llx\n",
b3deaf57
MD
813 cursor->node, cursor->index,
814 cursor->key_beg.obj_id,
815 cursor->key_beg.rec_type,
d113fda1
MD
816 cursor->key_beg.key,
817 cursor->key_beg.create_tid
b3deaf57
MD
818 );
819 }
820
8cd0a023
MD
821 /*
822 * Move our cursor up the tree until we find a node whos range covers
823 * the key we are trying to locate. This may move us between
824 * clusters.
825 *
826 * The left bound is inclusive, the right bound is non-inclusive.
827 * It is ok to cursor up too far so when cursoring across a cluster
828 * boundary.
829 *
830 * First see if we can skip the whole cluster. hammer_cursor_up()
831 * handles both cases but this way we don't check the cluster
832 * bounds when going up the tree within a cluster.
d26d0ae9
MD
833 *
834 * NOTE: If INCLUSTER is set and we are at the root of the cluster,
835 * hammer_cursor_up() will return ENOENT.
8cd0a023
MD
836 */
837 cluster = cursor->node->cluster;
838 while (
839 hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_beg) < 0 ||
840 hammer_btree_cmp(&cursor->key_beg, &cluster->clu_btree_end) >= 0) {
841 error = hammer_cursor_toroot(cursor);
842 if (error)
843 goto done;
9944ae54 844 KKASSERT(cursor->parent);
195c19a1 845 error = hammer_cursor_up(cursor, 0);
8cd0a023
MD
846 if (error)
847 goto done;
848 cluster = cursor->node->cluster;
427e5fc6
MD
849 }
850
851 /*
8cd0a023
MD
852 * Deal with normal cursoring within a cluster. The right bound
853 * is non-inclusive. That is, the bounds form a separator.
427e5fc6 854 */
8cd0a023
MD
855 while (hammer_btree_cmp(&cursor->key_beg, cursor->left_bound) < 0 ||
856 hammer_btree_cmp(&cursor->key_beg, cursor->right_bound) >= 0) {
9944ae54 857 KKASSERT(cursor->parent);
195c19a1 858 error = hammer_cursor_up(cursor, 0);
8cd0a023
MD
859 if (error)
860 goto done;
427e5fc6 861 }
427e5fc6 862
8cd0a023
MD
863 /*
864 * We better have ended up with a node somewhere, and our second
865 * while loop had better not have traversed up a cluster.
866 */
867 KKASSERT(cursor->node != NULL && cursor->node->cluster == cluster);
868
869 /*
870 * If we are inserting we can't start at a full node if the parent
871 * is also full (because there is no way to split the node),
872 * continue running up the tree until we hit the root of the
873 * root cluster or until the requirement is satisfied.
874 *
875 * NOTE: These cursor-up's CAN continue to cross cluster boundaries.
876 *
9944ae54
MD
877 * NOTE: We must guarantee at least two open spots in the parent
878 * to deal with hammer_btree_insert_cluster().
879 *
8cd0a023
MD
880 * XXX as an optimization it should be possible to unbalance the tree
881 * and stop at the root of the current cluster.
882 */
61aeeb33 883 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
9944ae54 884 if (btree_node_is_almost_full(cursor->node->ondisk) == 0)
8cd0a023
MD
885 break;
886 if (cursor->parent == NULL)
887 break;
888 if (cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS)
889 break;
195c19a1 890 error = hammer_cursor_up(cursor, 0);
8cd0a023
MD
891 /* cluster and node are now may become stale */
892 if (error)
893 goto done;
427e5fc6 894 }
8cd0a023 895 /* cluster = cursor->node->cluster; not needed until next cluster = */
427e5fc6 896
8cd0a023
MD
897#if 0
898 /*
899 * If we are deleting we can't start at an internal node with only
900 * one element unless it is root, because all of our code assumes
901 * that internal nodes will never be empty. Just do this generally
902 * for both leaf and internal nodes to get better balance.
903 *
904 * This handles the case where the cursor is sitting at a leaf and
905 * either the leaf or parent contain an insufficient number of
906 * elements.
907 *
908 * NOTE: These cursor-up's CAN continue to cross cluster boundaries.
a89aec1b
MD
909 *
910 * XXX NOTE: Iterations may not set this flag anyway.
8cd0a023 911 */
a89aec1b 912 while (flags & HAMMER_CURSOR_DELETE) {
8cd0a023
MD
913 if (cursor->node->ondisk->count > 1)
914 break;
915 if (cursor->parent == NULL)
916 break;
917 KKASSERT(cursor->node->ondisk->count != 0);
195c19a1 918 error = hammer_cursor_up(cursor, 0);
8cd0a023
MD
919 /* cluster and node are now may become stale */
920 if (error)
921 goto done;
922 }
923#endif
427e5fc6 924
8cd0a023
MD
925/*new_cluster:*/
926 /*
927 * Push down through internal nodes to locate the requested key.
928 */
929 cluster = cursor->node->cluster;
930 node = cursor->node->ondisk;
931 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
932#if 0
427e5fc6 933 /*
8cd0a023
MD
934 * If we are a the root node and deleting, try to collapse
935 * all of the root's children into the root. This is the
936 * only point where tree depth is reduced.
a89aec1b
MD
937 *
938 * XXX NOTE: Iterations may not set this flag anyway.
427e5fc6 939 */
a89aec1b 940 if ((flags & HAMMER_CURSOR_DELETE) && cursor->parent == NULL) {
8cd0a023
MD
941 error = btree_collapse(cursor);
942 /* node becomes stale after call */
d26d0ae9 943 /* XXX ENOSPC */
8cd0a023
MD
944 if (error)
945 goto done;
66325755 946 }
8cd0a023
MD
947 node = cursor->node->ondisk;
948#endif
8cd0a023
MD
949 /*
950 * Scan the node to find the subtree index to push down into.
fbc6e32a 951 * We go one-past, then back-up.
d113fda1
MD
952 *
953 * We have a serious issue with the midpoints for internal
954 * nodes when the midpoint bisects two historical records
955 * (where only create_tid is different). Short of iterating
956 * through the record's entire history the only solution is
957 * to calculate a midpoint that isn't a midpoint in that
958 * case. Please see hammer_make_separator() for more
959 * information.
9944ae54
MD
960 *
961 * The right boundary is included in the search.
8cd0a023 962 */
9944ae54 963 for (i = 0; i <= node->count; ++i) {
61aeeb33
MD
964 elm = &node->elms[i];
965 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
8cd0a023
MD
966 if (r < 0)
967 break;
968 }
8cd0a023
MD
969
970 /*
9944ae54
MD
971 * These cases occur when the parent's idea of the boundary
972 * is wider then the child's idea of the boundary, and
973 * require special handling. If not inserting we can
974 * terminate the search early for these cases but the
975 * child's boundaries cannot be unconditionally modified.
8cd0a023 976 */
9944ae54 977 edge = SEARCH_NONE;
fbc6e32a 978 if (i == 0) {
9944ae54
MD
979 /*
980 * If i == 0 the search terminated to the LEFT of the
981 * left_boundary but to the RIGHT of the parent's left
982 * boundary.
983 */
fbc6e32a 984 u_int8_t save;
d26d0ae9 985
fbc6e32a
MD
986 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
987 cursor->index = 0;
988 return(ENOENT);
989 }
9944ae54
MD
990 elm = &node->elms[0];
991
992 if (elm->base.subtree_type ==
993 HAMMER_BTREE_TYPE_CLUSTER) {
994 edge = SEARCH_LEFT_EDGE;
995 } else {
996 /*
997 * Correct a left-hand boundary mismatch.
998 */
999 hammer_modify_node(cursor->node);
1000 save = node->elms[0].base.subtree_type;
1001 node->elms[0].base = *cursor->left_bound;
1002 node->elms[0].base.subtree_type = save;
1003 }
1004 } else if (i == node->count + 1) {
d26d0ae9 1005 /*
9944ae54
MD
1006 * If i == node->count + 1 the search terminated to
1007 * the RIGHT of the right boundary but to the LEFT
1008 * of the parent's right boundary.
d113fda1 1009 *
9944ae54
MD
1010 * Note that the last element in this case is
1011 * elms[i-2] prior to adjustments to 'i'.
d26d0ae9 1012 */
9944ae54 1013 --i;
d113fda1 1014 if ((flags & HAMMER_CURSOR_INSERT) == 0) {
9944ae54
MD
1015 cursor->index = i;
1016 return(ENOENT);
d26d0ae9
MD
1017 }
1018
9944ae54
MD
1019 elm = &node->elms[i];
1020 if (elm[-1].base.subtree_type ==
1021 HAMMER_BTREE_TYPE_CLUSTER) {
1022 edge = SEARCH_RIGHT_EDGE;
1023 } else {
d26d0ae9 1024 hammer_modify_node(cursor->node);
61aeeb33 1025 elm->base = *cursor->right_bound;
d26d0ae9 1026 }
fbc6e32a
MD
1027 } else {
1028 /*
9944ae54
MD
1029 * The push-down index is now i - 1. If we had
1030 * terminated on the right boundary this will point
1031 * us at the last element.
fbc6e32a
MD
1032 */
1033 --i;
1034 }
8cd0a023
MD
1035 cursor->index = i;
1036
b3deaf57 1037 if (hammer_debug_btree) {
61aeeb33 1038 elm = &node->elms[i];
d113fda1 1039 kprintf("SEARCH-I %p:%d %016llx %02x key=%016llx tid=%016llx\n",
b3deaf57
MD
1040 cursor->node, i,
1041 elm->internal.base.obj_id,
1042 elm->internal.base.rec_type,
d113fda1
MD
1043 elm->internal.base.key,
1044 elm->internal.base.create_tid
b3deaf57
MD
1045 );
1046 }
1047
8cd0a023
MD
1048 /*
1049 * Handle insertion and deletion requirements.
1050 *
1051 * If inserting split full nodes. The split code will
1052 * adjust cursor->node and cursor->index if the current
1053 * index winds up in the new node.
61aeeb33 1054 *
9944ae54
MD
1055 * If inserting and a left or right edge case was detected,
1056 * we cannot correct the left or right boundary and must
1057 * prepend and append an empty leaf node in order to make
1058 * the boundary correction.
1059 *
61aeeb33
MD
1060 * If we run out of space we set enospc and continue on
1061 * to a leaf to provide the spike code with a good point
1062 * of entry. Enospc is reset if we cross a cluster boundary.
8cd0a023 1063 */
61aeeb33 1064 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
9944ae54 1065 if (btree_node_is_almost_full(node)) {
8cd0a023 1066 error = btree_split_internal(cursor);
d26d0ae9
MD
1067 if (error) {
1068 if (error != ENOSPC)
1069 goto done;
1070 enospc = 1;
d26d0ae9 1071 }
8cd0a023
MD
1072 /*
1073 * reload stale pointers
1074 */
1075 i = cursor->index;
1076 node = cursor->node->ondisk;
1077 }
9944ae54
MD
1078 if (edge != SEARCH_NONE && enospc == 0) {
1079 error = btree_edge_internal(cursor, edge);
1080 if (error) {
1081 if (error != ENOSPC)
1082 goto done;
1083 enospc = 1;
1084 }
1085 /*
1086 * reload stale pointers
1087 */
1088 i = cursor->index;
1089 node = cursor->node->ondisk;
1090 }
8cd0a023
MD
1091 }
1092
1093#if 0
427e5fc6 1094 /*
8cd0a023
MD
1095 * If deleting rebalance - do not allow the child to have
1096 * just one element or we will not be able to delete it.
1097 *
1098 * Neither internal or leaf nodes (except a root-leaf) are
1099 * allowed to drop to 0 elements. (XXX - well, leaf nodes
1100 * can at the moment).
1101 *
1102 * Our separators may have been reorganized after rebalancing,
1103 * so we have to pop back up and rescan.
1104 *
1105 * XXX test for subtree_count < maxelms / 2, minus 1 or 2
1106 * for hysteresis?
a89aec1b
MD
1107 *
1108 * XXX NOTE: Iterations may not set this flag anyway.
427e5fc6 1109 */
a89aec1b 1110 if (flags & HAMMER_CURSOR_DELETE) {
8cd0a023
MD
1111 if (node->elms[i].internal.subtree_count <= 1) {
1112 error = btree_rebalance(cursor);
1113 if (error)
1114 goto done;
1115 /* cursor->index is invalid after call */
1116 goto new_cluster;
1117 }
1118 }
1119#endif
d26d0ae9 1120 /*
61aeeb33
MD
1121 * A non-zero rec_offset specifies a cluster push.
1122 * If this is a cluster push we reset the enospc flag,
1123 * which reenables the insertion code in the new cluster.
1124 * This also ensures that if a spike occurs both its node
1125 * and its parent will be in the same cluster.
d26d0ae9 1126 *
61aeeb33
MD
1127 * If INCLUSTER is set we terminate at the cluster boundary.
1128 * In this case we must determine whether key_beg is within
1129 * the cluster's boundary or not. XXX
d26d0ae9 1130 */
61aeeb33
MD
1131 elm = &node->elms[i];
1132 if (elm->internal.rec_offset) {
9944ae54 1133 KKASSERT(elm->base.subtree_type ==
61aeeb33
MD
1134 HAMMER_BTREE_TYPE_CLUSTER);
1135 enospc = 0;
1136 if (flags & HAMMER_CURSOR_INCLUSTER) {
1137 KKASSERT((flags & HAMMER_CURSOR_INSERT) == 0);
1138 r = hammer_btree_cmp(&cursor->key_beg,
1139 &elm[1].base);
1140 error = (r < 0) ? 0 : ENOENT;
1141 goto done;
1142 }
d26d0ae9 1143 }
427e5fc6
MD
1144
1145 /*
8cd0a023 1146 * Push down (push into new node, existing node becomes
d26d0ae9 1147 * the parent) and continue the search.
427e5fc6 1148 */
8cd0a023
MD
1149 error = hammer_cursor_down(cursor);
1150 /* node and cluster become stale */
1151 if (error)
1152 goto done;
1153 node = cursor->node->ondisk;
1154 cluster = cursor->node->cluster;
427e5fc6 1155 }
427e5fc6 1156
8cd0a023
MD
1157 /*
1158 * We are at a leaf, do a linear search of the key array.
d26d0ae9
MD
1159 *
1160 * On success the index is set to the matching element and 0
1161 * is returned.
1162 *
1163 * On failure the index is set to the insertion point and ENOENT
1164 * is returned.
8cd0a023
MD
1165 *
1166 * Boundaries are not stored in leaf nodes, so the index can wind
1167 * up to the left of element 0 (index == 0) or past the end of
1168 * the array (index == node->count).
1169 */
1170 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1171
1172 for (i = 0; i < node->count; ++i) {
1173 r = hammer_btree_cmp(&cursor->key_beg, &node->elms[i].base);
427e5fc6 1174
427e5fc6 1175 /*
d113fda1
MD
1176 * Stop if we've flipped past key_beg. This includes a
1177 * record whos create_tid is larger then our asof id.
427e5fc6 1178 */
8cd0a023
MD
1179 if (r < 0)
1180 break;
427e5fc6 1181
66325755 1182 /*
d113fda1
MD
1183 * Return an exact match. In this case we have to do special
1184 * checks if the only difference in the records is the
1185 * create_ts, in order to properly match against our as-of
1186 * query.
66325755 1187 */
d113fda1
MD
1188 if (r >= 0 && r <= 1) {
1189 if ((cursor->flags & HAMMER_CURSOR_ALLHISTORY) == 0 &&
1190 hammer_btree_chkts(cursor->key_beg.create_tid,
1191 &node->elms[i].base) != 0) {
1192 continue;
1193 }
8cd0a023
MD
1194 cursor->index = i;
1195 error = 0;
b3deaf57
MD
1196 if (hammer_debug_btree) {
1197 kprintf("SEARCH-L %p:%d (SUCCESS)\n",
1198 cursor->node, i);
1199 }
8cd0a023 1200 goto done;
66325755 1201 }
427e5fc6 1202 }
8cd0a023 1203
b3deaf57
MD
1204 if (hammer_debug_btree) {
1205 kprintf("SEARCH-L %p:%d (FAILED)\n",
1206 cursor->node, i);
1207 }
1208
8cd0a023
MD
1209 /*
1210 * No exact match was found, i is now at the insertion point.
1211 *
1212 * If inserting split a full leaf before returning. This
1213 * may have the side effect of adjusting cursor->node and
1214 * cursor->index.
1215 */
1216 cursor->index = i;
9944ae54 1217 if ((flags & HAMMER_CURSOR_INSERT) && btree_node_is_almost_full(node)) {
8cd0a023 1218 error = btree_split_leaf(cursor);
d26d0ae9
MD
1219 if (error) {
1220 if (error != ENOSPC)
1221 goto done;
1222 enospc = 1;
1223 flags &= ~HAMMER_CURSOR_INSERT;
1224 }
1225 /*
1226 * reload stale pointers
1227 */
8cd0a023
MD
1228 /* NOT USED
1229 i = cursor->index;
1230 node = &cursor->node->internal;
1231 */
8cd0a023 1232 }
d26d0ae9
MD
1233
1234 /*
1235 * We reached a leaf but did not find the key we were looking for.
1236 * If this is an insert we will be properly positioned for an insert
1237 * (ENOENT) or spike (ENOSPC) operation.
1238 */
1239 error = enospc ? ENOSPC : ENOENT;
8cd0a023 1240done:
427e5fc6
MD
1241 return(error);
1242}
1243
8cd0a023 1244
427e5fc6 1245/************************************************************************
8cd0a023 1246 * SPLITTING AND MERGING *
427e5fc6
MD
1247 ************************************************************************
1248 *
1249 * These routines do all the dirty work required to split and merge nodes.
1250 */
1251
9944ae54
MD
1252/*
1253 * This case occurs when we are trying to insert and have come across a
1254 * mismatched left or right boundary which could not be adjusted due to
1255 * being part of a spike. In order to be able to adjust the boundary
1256 * we have to prepend or append an empty leaf node.
1257 */
1258static
1259int
1260btree_edge_internal(hammer_cursor_t cursor, btree_search_edge_t edge)
1261{
1262 hammer_node_ondisk_t old_disk;
1263 hammer_node_ondisk_t new_disk;
1264 hammer_node_t new_node;
1265 hammer_btree_elm_t elm;
1266 int error;
1267 int n;
1268 const int esize = sizeof(*elm);
1269
1270 old_disk = cursor->node->ondisk;
1271 KKASSERT(old_disk->type == HAMMER_BTREE_TYPE_INTERNAL);
1272 KKASSERT(old_disk->count < HAMMER_BTREE_INT_ELMS);
1273
1274 /*
1275 * Allocate a new leaf node.
1276 */
1277 new_node = hammer_alloc_btree(cursor->node->cluster, &error);
1278 if (error)
1279 return(error);
1280
1281 hammer_lock_ex(&new_node->lock);
1282 hammer_modify_node(cursor->node);
1283 hammer_modify_node(new_node);
1284 new_disk = new_node->ondisk;
1285 n = old_disk->count;
1286
1287 /*
1288 * Prepend or append the leaf node and correct the boundary
1289 * mismatch.
1290 */
1291 switch(edge) {
1292 case SEARCH_LEFT_EDGE:
1293 KKASSERT(cursor->index == 0);
1294 elm = &old_disk->elms[0];
1295 bcopy(elm, elm + 1, (n + 1) * esize);
1296 elm->base = *cursor->left_bound;
1297 break;
1298 case SEARCH_RIGHT_EDGE:
1299 KKASSERT(cursor->index == old_disk->count);
1300 elm = &old_disk->elms[n];
1301 elm[1].base = *cursor->right_bound;
1302 break;
1303 default:
1304 panic("btree_edge_internal: bad edge");
1305 break;
1306 }
1307 ++old_disk->count;
1308 elm->base.subtree_type = HAMMER_BTREE_TYPE_LEAF;
1309 elm->internal.subtree_offset = new_node->node_offset;
1310 elm->internal.subtree_vol_no = -1;
1311 elm->internal.subtree_count = 0;
1312
1313 new_disk->count = 0;
1314 new_disk->parent = cursor->node->node_offset;
1315 new_disk->type = HAMMER_BTREE_TYPE_LEAF;
1316
1317 hammer_unlock(&new_node->lock);
1318 hammer_rel_node(new_node);
1319
1320 /*
1321 * Cursor->index remains unchanged. It now points to our new leaf
1322 * node and cursor->node's boundaries have been synchronized with
1323 * the parent.
1324 */
1325 return(0);
1326}
1327
427e5fc6 1328/*
8cd0a023 1329 * Split an internal node into two nodes and move the separator at the split
427e5fc6
MD
1330 * point to the parent. Note that the parent's parent's element pointing
1331 * to our parent will have an incorrect subtree_count (we don't update it).
1332 * It will be low, which is ok.
1333 *
8cd0a023
MD
1334 * (cursor->node, cursor->index) indicates the element the caller intends
1335 * to push into. We will adjust node and index if that element winds
427e5fc6 1336 * up in the split node.
8cd0a023
MD
1337 *
1338 * If we are at the root of a cluster a new root must be created with two
1339 * elements, one pointing to the original root and one pointing to the
1340 * newly allocated split node.
1341 *
1342 * NOTE! Being at the root of a cluster is different from being at the
1343 * root of the root cluster. cursor->parent will not be NULL and
1344 * cursor->node->ondisk.parent must be tested against 0. Theoretically
1345 * we could propogate the algorithm into the parent and deal with multiple
1346 * 'roots' in the cluster header, but it's easier not to.
427e5fc6
MD
1347 */
1348static
1349int
8cd0a023 1350btree_split_internal(hammer_cursor_t cursor)
427e5fc6 1351{
8cd0a023
MD
1352 hammer_node_ondisk_t ondisk;
1353 hammer_node_t node;
1354 hammer_node_t parent;
1355 hammer_node_t new_node;
1356 hammer_btree_elm_t elm;
1357 hammer_btree_elm_t parent_elm;
427e5fc6
MD
1358 int parent_index;
1359 int made_root;
1360 int split;
1361 int error;
7f7c1f84 1362 int i;
8cd0a023 1363 const int esize = sizeof(*elm);
427e5fc6
MD
1364
1365 /*
1366 * We are splitting but elms[split] will be promoted to the parent,
1367 * leaving the right hand node with one less element. If the
1368 * insertion point will be on the left-hand side adjust the split
1369 * point to give the right hand side one additional node.
1370 */
8cd0a023
MD
1371 node = cursor->node;
1372 ondisk = node->ondisk;
1373 split = (ondisk->count + 1) / 2;
427e5fc6
MD
1374 if (cursor->index <= split)
1375 --split;
1376 error = 0;
1377
1378 /*
8cd0a023 1379 * If we are at the root of the cluster, create a new root node with
427e5fc6
MD
1380 * 1 element and split normally. Avoid making major modifications
1381 * until we know the whole operation will work.
8cd0a023
MD
1382 *
1383 * The root of the cluster is different from the root of the root
1384 * cluster. Use the node's on-disk structure's parent offset to
1385 * detect the case.
427e5fc6 1386 */
8cd0a023
MD
1387 if (ondisk->parent == 0) {
1388 parent = hammer_alloc_btree(node->cluster, &error);
427e5fc6
MD
1389 if (parent == NULL)
1390 return(error);
8cd0a023 1391 hammer_lock_ex(&parent->lock);
0b075555 1392 hammer_modify_node(parent);
8cd0a023
MD
1393 ondisk = parent->ondisk;
1394 ondisk->count = 1;
1395 ondisk->parent = 0;
1396 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1397 ondisk->elms[0].base = node->cluster->clu_btree_beg;
9944ae54 1398 ondisk->elms[0].base.subtree_type = node->ondisk->type;
8cd0a023
MD
1399 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1400 ondisk->elms[1].base = node->cluster->clu_btree_end;
9944ae54 1401 /* ondisk->elms[1].base.subtree_Type - not used */
427e5fc6 1402 made_root = 1;
8cd0a023 1403 parent_index = 0; /* index of current node in parent */
427e5fc6
MD
1404 } else {
1405 made_root = 0;
8cd0a023
MD
1406 parent = cursor->parent;
1407 parent_index = cursor->parent_index;
195c19a1 1408 KKASSERT(parent->cluster == node->cluster);
427e5fc6 1409 }
427e5fc6
MD
1410
1411 /*
1412 * Split node into new_node at the split point.
1413 *
1414 * B O O O P N N B <-- P = node->elms[split]
1415 * 0 1 2 3 4 5 6 <-- subtree indices
1416 *
1417 * x x P x x
1418 * s S S s
1419 * / \
1420 * B O O O B B N N B <--- inner boundary points are 'P'
1421 * 0 1 2 3 4 5 6
1422 *
1423 */
8cd0a023 1424 new_node = hammer_alloc_btree(node->cluster, &error);
427e5fc6 1425 if (new_node == NULL) {
8cd0a023
MD
1426 if (made_root) {
1427 hammer_unlock(&parent->lock);
b3deaf57 1428 parent->flags |= HAMMER_NODE_DELETED;
8cd0a023
MD
1429 hammer_rel_node(parent);
1430 }
427e5fc6
MD
1431 return(error);
1432 }
8cd0a023 1433 hammer_lock_ex(&new_node->lock);
427e5fc6
MD
1434
1435 /*
8cd0a023 1436 * Create the new node. P becomes the left-hand boundary in the
427e5fc6
MD
1437 * new node. Copy the right-hand boundary as well.
1438 *
1439 * elm is the new separator.
1440 */
0b075555
MD
1441 hammer_modify_node(new_node);
1442 hammer_modify_node(node);
8cd0a023
MD
1443 ondisk = node->ondisk;
1444 elm = &ondisk->elms[split];
1445 bcopy(elm, &new_node->ondisk->elms[0],
1446 (ondisk->count - split + 1) * esize);
1447 new_node->ondisk->count = ondisk->count - split;
1448 new_node->ondisk->parent = parent->node_offset;
1449 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1450 KKASSERT(ondisk->type == new_node->ondisk->type);
427e5fc6
MD
1451
1452 /*
1453 * Cleanup the original node. P becomes the new boundary, its
1454 * subtree_offset was moved to the new node. If we had created
1455 * a new root its parent pointer may have changed.
1456 */
8cd0a023 1457 elm->internal.subtree_offset = 0;
0b075555 1458 elm->internal.rec_offset = 0;
c0ade690 1459 ondisk->count = split;
427e5fc6
MD
1460
1461 /*
1462 * Insert the separator into the parent, fixup the parent's
1463 * reference to the original node, and reference the new node.
1464 * The separator is P.
1465 *
1466 * Remember that base.count does not include the right-hand boundary.
1467 */
0b075555 1468 hammer_modify_node(parent);
8cd0a023 1469 ondisk = parent->ondisk;
d26d0ae9 1470 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
8cd0a023
MD
1471 ondisk->elms[parent_index].internal.subtree_count = split;
1472 parent_elm = &ondisk->elms[parent_index+1];
427e5fc6 1473 bcopy(parent_elm, parent_elm + 1,
8cd0a023
MD
1474 (ondisk->count - parent_index) * esize);
1475 parent_elm->internal.base = elm->base; /* separator P */
9944ae54 1476 parent_elm->internal.base.subtree_type = new_node->ondisk->type;
8cd0a023
MD
1477 parent_elm->internal.subtree_offset = new_node->node_offset;
1478 parent_elm->internal.subtree_count = new_node->ondisk->count;
0b075555
MD
1479 parent_elm->internal.subtree_vol_no = 0;
1480 parent_elm->internal.rec_offset = 0;
76376933 1481 ++ondisk->count;
427e5fc6 1482
7f7c1f84
MD
1483 /*
1484 * The children of new_node need their parent pointer set to new_node.
1485 */
1486 for (i = 0; i < new_node->ondisk->count; ++i) {
1487 elm = &new_node->ondisk->elms[i];
1488 error = btree_set_parent(new_node, elm);
1489 if (error) {
1490 panic("btree_split_internal: btree-fixup problem");
1491 }
1492 }
1493
427e5fc6
MD
1494 /*
1495 * The cluster's root pointer may have to be updated.
1496 */
1497 if (made_root) {
8cd0a023 1498 hammer_modify_cluster(node->cluster);
0b075555 1499 node->cluster->ondisk->clu_btree_root = parent->node_offset;
8cd0a023
MD
1500 node->ondisk->parent = parent->node_offset;
1501 if (cursor->parent) {
1502 hammer_unlock(&cursor->parent->lock);
1503 hammer_rel_node(cursor->parent);
1504 }
1505 cursor->parent = parent; /* lock'd and ref'd */
427e5fc6
MD
1506 }
1507
8cd0a023 1508
427e5fc6
MD
1509 /*
1510 * Ok, now adjust the cursor depending on which element the original
1511 * index was pointing at. If we are >= the split point the push node
1512 * is now in the new node.
1513 *
1514 * NOTE: If we are at the split point itself we cannot stay with the
1515 * original node because the push index will point at the right-hand
1516 * boundary, which is illegal.
8cd0a023
MD
1517 *
1518 * NOTE: The cursor's parent or parent_index must be adjusted for
1519 * the case where a new parent (new root) was created, and the case
1520 * where the cursor is now pointing at the split node.
427e5fc6
MD
1521 */
1522 if (cursor->index >= split) {
8cd0a023 1523 cursor->parent_index = parent_index + 1;
427e5fc6 1524 cursor->index -= split;
8cd0a023
MD
1525 hammer_unlock(&cursor->node->lock);
1526 hammer_rel_node(cursor->node);
1527 cursor->node = new_node; /* locked and ref'd */
1528 } else {
1529 cursor->parent_index = parent_index;
1530 hammer_unlock(&new_node->lock);
1531 hammer_rel_node(new_node);
427e5fc6 1532 }
76376933
MD
1533
1534 /*
1535 * Fixup left and right bounds
1536 */
1537 parent_elm = &parent->ondisk->elms[cursor->parent_index];
fbc6e32a
MD
1538 cursor->left_bound = &parent_elm[0].internal.base;
1539 cursor->right_bound = &parent_elm[1].internal.base;
b3deaf57
MD
1540 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1541 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1542 KKASSERT(hammer_btree_cmp(cursor->right_bound,
9944ae54 1543 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
76376933 1544
427e5fc6
MD
1545 return (0);
1546}
1547
1548/*
1549 * Same as the above, but splits a full leaf node.
1550 */
1551static
1552int
8cd0a023 1553btree_split_leaf(hammer_cursor_t cursor)
427e5fc6 1554{
8cd0a023
MD
1555 hammer_node_ondisk_t ondisk;
1556 hammer_node_t parent;
1557 hammer_node_t leaf;
1558 hammer_node_t new_leaf;
1559 hammer_btree_elm_t elm;
1560 hammer_btree_elm_t parent_elm;
b3deaf57 1561 hammer_base_elm_t mid_boundary;
427e5fc6
MD
1562 int parent_index;
1563 int made_root;
1564 int split;
1565 int error;
8cd0a023 1566 const size_t esize = sizeof(*elm);
427e5fc6
MD
1567
1568 /*
8cd0a023
MD
1569 * Calculate the split point. If the insertion point will be on
1570 * the left-hand side adjust the split point to give the right
1571 * hand side one additional node.
427e5fc6 1572 */
8cd0a023
MD
1573 leaf = cursor->node;
1574 ondisk = leaf->ondisk;
1575 split = (ondisk->count + 1) / 2;
427e5fc6
MD
1576 if (cursor->index <= split)
1577 --split;
1578 error = 0;
1579
1580 /*
1581 * If we are at the root of the tree, create a new root node with
1582 * 1 element and split normally. Avoid making major modifications
1583 * until we know the whole operation will work.
1584 */
8cd0a023
MD
1585 if (ondisk->parent == 0) {
1586 parent = hammer_alloc_btree(leaf->cluster, &error);
427e5fc6
MD
1587 if (parent == NULL)
1588 return(error);
8cd0a023 1589 hammer_lock_ex(&parent->lock);
0b075555 1590 hammer_modify_node(parent);
8cd0a023
MD
1591 ondisk = parent->ondisk;
1592 ondisk->count = 1;
1593 ondisk->parent = 0;
1594 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1595 ondisk->elms[0].base = leaf->cluster->clu_btree_beg;
9944ae54 1596 ondisk->elms[0].base.subtree_type = leaf->ondisk->type;
8cd0a023
MD
1597 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1598 ondisk->elms[1].base = leaf->cluster->clu_btree_end;
9944ae54 1599 /* ondisk->elms[1].base.subtree_type = not used */
427e5fc6 1600 made_root = 1;
8cd0a023 1601 parent_index = 0; /* insertion point in parent */
427e5fc6
MD
1602 } else {
1603 made_root = 0;
8cd0a023
MD
1604 parent = cursor->parent;
1605 parent_index = cursor->parent_index;
195c19a1 1606 KKASSERT(parent->cluster == leaf->cluster);
427e5fc6 1607 }
427e5fc6
MD
1608
1609 /*
1610 * Split leaf into new_leaf at the split point. Select a separator
1611 * value in-between the two leafs but with a bent towards the right
1612 * leaf since comparisons use an 'elm >= separator' inequality.
1613 *
1614 * L L L L L L L L
1615 *
1616 * x x P x x
1617 * s S S s
1618 * / \
1619 * L L L L L L L L
1620 */
8cd0a023 1621 new_leaf = hammer_alloc_btree(leaf->cluster, &error);
427e5fc6 1622 if (new_leaf == NULL) {
8cd0a023
MD
1623 if (made_root) {
1624 hammer_unlock(&parent->lock);
b3deaf57 1625 parent->flags |= HAMMER_NODE_DELETED;
8cd0a023
MD
1626 hammer_rel_node(parent);
1627 }
427e5fc6
MD
1628 return(error);
1629 }
8cd0a023 1630 hammer_lock_ex(&new_leaf->lock);
427e5fc6
MD
1631
1632 /*
1633 * Create the new node. P become the left-hand boundary in the
1634 * new node. Copy the right-hand boundary as well.
1635 */
0b075555
MD
1636 hammer_modify_node(leaf);
1637 hammer_modify_node(new_leaf);
8cd0a023
MD
1638 ondisk = leaf->ondisk;
1639 elm = &ondisk->elms[split];
1640 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1641 new_leaf->ondisk->count = ondisk->count - split;
1642 new_leaf->ondisk->parent = parent->node_offset;
1643 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1644 KKASSERT(ondisk->type == new_leaf->ondisk->type);
427e5fc6
MD
1645
1646 /*
8cd0a023
MD
1647 * Cleanup the original node. Because this is a leaf node and
1648 * leaf nodes do not have a right-hand boundary, there
c0ade690
MD
1649 * aren't any special edge cases to clean up. We just fixup the
1650 * count.
427e5fc6 1651 */
c0ade690 1652 ondisk->count = split;
427e5fc6
MD
1653
1654 /*
1655 * Insert the separator into the parent, fixup the parent's
1656 * reference to the original node, and reference the new node.
1657 * The separator is P.
1658 *
1659 * Remember that base.count does not include the right-hand boundary.
1660 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1661 */
0b075555 1662 hammer_modify_node(parent);
8cd0a023 1663 ondisk = parent->ondisk;
d26d0ae9 1664 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
8cd0a023
MD
1665 ondisk->elms[parent_index].internal.subtree_count = split;
1666 parent_elm = &ondisk->elms[parent_index+1];
d26d0ae9
MD
1667 bcopy(parent_elm, parent_elm + 1,
1668 (ondisk->count - parent_index) * esize);
427e5fc6 1669 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
9944ae54 1670 parent_elm->internal.base.subtree_type = new_leaf->ondisk->type;
8cd0a023
MD
1671 parent_elm->internal.subtree_offset = new_leaf->node_offset;
1672 parent_elm->internal.subtree_count = new_leaf->ondisk->count;
0b075555
MD
1673 parent_elm->internal.subtree_vol_no = 0;
1674 parent_elm->internal.rec_offset = 0;
b3deaf57 1675 mid_boundary = &parent_elm->base;
76376933 1676 ++ondisk->count;
427e5fc6
MD
1677
1678 /*
1679 * The cluster's root pointer may have to be updated.
1680 */
1681 if (made_root) {
8cd0a023 1682 hammer_modify_cluster(leaf->cluster);
0b075555 1683 leaf->cluster->ondisk->clu_btree_root = parent->node_offset;
8cd0a023
MD
1684 leaf->ondisk->parent = parent->node_offset;
1685 if (cursor->parent) {
1686 hammer_unlock(&cursor->parent->lock);
1687 hammer_rel_node(cursor->parent);
1688 }
1689 cursor->parent = parent; /* lock'd and ref'd */
427e5fc6 1690 }
8cd0a023 1691
427e5fc6
MD
1692 /*
1693 * Ok, now adjust the cursor depending on which element the original
1694 * index was pointing at. If we are >= the split point the push node
1695 * is now in the new node.
1696 *
b3deaf57
MD
1697 * NOTE: If we are at the split point itself we need to select the
1698 * old or new node based on where key_beg's insertion point will be.
1699 * If we pick the wrong side the inserted element will wind up in
1700 * the wrong leaf node and outside that node's bounds.
427e5fc6 1701 */
b3deaf57
MD
1702 if (cursor->index > split ||
1703 (cursor->index == split &&
1704 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
8cd0a023 1705 cursor->parent_index = parent_index + 1;
427e5fc6 1706 cursor->index -= split;
8cd0a023
MD
1707 hammer_unlock(&cursor->node->lock);
1708 hammer_rel_node(cursor->node);
1709 cursor->node = new_leaf;
1710 } else {
1711 cursor->parent_index = parent_index;
1712 hammer_unlock(&new_leaf->lock);
1713 hammer_rel_node(new_leaf);
427e5fc6 1714 }
76376933
MD
1715
1716 /*
1717 * Fixup left and right bounds
1718 */
1719 parent_elm = &parent->ondisk->elms[cursor->parent_index];
fbc6e32a
MD
1720 cursor->left_bound = &parent_elm[0].internal.base;
1721 cursor->right_bound = &parent_elm[1].internal.base;
b3deaf57
MD
1722 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1723 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1724 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1725 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
76376933 1726
427e5fc6
MD
1727 return (0);
1728}
1729
1730/*
195c19a1
MD
1731 * Attempt to remove the empty B-Tree node at (cursor->node). Returns 0
1732 * on success, EAGAIN if we could not acquire the necessary locks, or some
1733 * other error.
8cd0a023 1734 *
195c19a1 1735 * On return the cursor may end up pointing at an internal node, suitable
b3deaf57 1736 * for further iteration but not for an immediate insertion or deletion.
8cd0a023 1737 *
195c19a1 1738 * cursor->node may be an internal node or a leaf node.
b3deaf57
MD
1739 *
1740 * NOTE: If cursor->node has one element it is the parent trying to delete
1741 * that element, make sure cursor->index is properly adjusted on success.
8cd0a023
MD
1742 */
1743int
195c19a1 1744btree_remove(hammer_cursor_t cursor)
8cd0a023
MD
1745{
1746 hammer_node_ondisk_t ondisk;
195c19a1
MD
1747 hammer_btree_elm_t elm;
1748 hammer_node_t save;
1749 hammer_node_t node;
8cd0a023
MD
1750 hammer_node_t parent;
1751 int error;
195c19a1 1752 int i;
8cd0a023
MD
1753
1754 /*
195c19a1
MD
1755 * If we are at the root of the root cluster there is nothing to
1756 * remove, but an internal node at the root of a cluster is not
1757 * allowed to be empty so convert it to a leaf node.
8cd0a023 1758 */
195c19a1 1759 if (cursor->parent == NULL) {
0b075555 1760 hammer_modify_node(cursor->node);
195c19a1
MD
1761 ondisk = cursor->node->ondisk;
1762 KKASSERT(ondisk->parent == 0);
1763 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1764 ondisk->count = 0;
b3deaf57 1765 cursor->index = 0;
195c19a1 1766 kprintf("EMPTY ROOT OF ROOT CLUSTER -> LEAF\n");
8cd0a023
MD
1767 return(0);
1768 }
1769
1770 /*
195c19a1
MD
1771 * Retain a reference to cursor->node, ex-lock again (2 locks now)
1772 * so we do not lose the lock when we cursor around.
8cd0a023 1773 */
195c19a1
MD
1774 save = cursor->node;
1775 hammer_ref_node(save);
1776 hammer_lock_ex(&save->lock);
8cd0a023
MD
1777
1778 /*
195c19a1
MD
1779 * We need to be able to lock the parent of the parent. Do this
1780 * non-blocking and return EAGAIN if the lock cannot be acquired.
1781 * non-blocking is required in order to avoid a deadlock.
1782 *
1783 * After we cursor up, parent is moved to node and the new parent
1784 * is the parent of the parent.
8cd0a023 1785 */
195c19a1
MD
1786 error = hammer_cursor_up(cursor, 1);
1787 if (error) {
b3deaf57
MD
1788 kprintf("BTREE_REMOVE: Cannot lock parent, skipping\n");
1789 goto failure;
8cd0a023 1790 }
195c19a1
MD
1791
1792 /*
1793 * At this point we want to remove the element at (node, index),
1794 * which is now the (original) parent pointing to the saved node.
1795 * Removing the element allows us to then free the node it was
1796 * pointing to.
1797 *
1798 * However, an internal node is not allowed to have 0 elements, so
1799 * if the count would drop to 0 we have to recurse. It is possible
1800 * for the recursion to fail.
1801 *
1802 * NOTE: The cursor is in an indeterminant position after recursing,
1803 * but will still be suitable for an iteration.
1804 */
1805 node = cursor->node;
1806 KKASSERT(node->ondisk->count > 0);
1807 if (node->ondisk->count == 1) {
1808 error = btree_remove(cursor);
1809 if (error == 0) {
b3deaf57
MD
1810 /*kprintf("BTREE_REMOVE: Successful!\n");*/
1811 goto success;
195c19a1
MD
1812 } else {
1813 kprintf("BTREE_REMOVE: Recursion failed %d\n", error);
b3deaf57 1814 goto failure;
8cd0a023
MD
1815 }
1816 }
195c19a1
MD
1817
1818 /*
1819 * Remove the element at (node, index) and adjust the parent's
1820 * subtree_count.
fbc6e32a 1821 *
d26d0ae9
MD
1822 * NOTE! If removing element 0 an internal node's left-hand boundary
1823 * will no longer match its parent. If removing a mid-element the
1824 * boundary will no longer match a child's left hand or right hand
1825 * boundary.
1826 *
1827 * BxBxBxB remove a (x[0]): internal node's left-hand
1828 * | | | boundary no longer matches
1829 * a b c parent.
1830 *
1831 * remove b (x[1]): a's right hand boundary no
1832 * longer matches parent.
fbc6e32a 1833 *
d26d0ae9
MD
1834 * remove c (x[2]): b's right hand boundary no
1835 * longer matches parent.
1836 *
1837 * These cases are corrected in btree_search().
195c19a1 1838 */
fbc6e32a 1839#if 0
195c19a1 1840 kprintf("BTREE_REMOVE: Removing element %d\n", cursor->index);
fbc6e32a 1841#endif
b3deaf57
MD
1842 KKASSERT(node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
1843 KKASSERT(cursor->index < node->ondisk->count);
0b075555 1844 hammer_modify_node(node);
195c19a1
MD
1845 ondisk = node->ondisk;
1846 i = cursor->index;
9944ae54
MD
1847
1848 /*
1849 * WARNING: For historical lookups to work properly we cannot
1850 * recalculate the mid-point or we might blow up historical searches
1851 * which depend on the mid-point matching the first right-hand element
1852 * XXX
1853 */
195c19a1
MD
1854 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
1855 (ondisk->count - i) * sizeof(ondisk->elms[0]));
1856 --ondisk->count;
195c19a1
MD
1857
1858 /*
1859 * Adjust the parent-parent's (now parent) reference to the parent
1860 * (now node).
1861 */
1862 if ((parent = cursor->parent) != NULL) {
1863 elm = &parent->ondisk->elms[cursor->parent_index];
1864 if (elm->internal.subtree_count != ondisk->count) {
195c19a1 1865 hammer_modify_node(parent);
0b075555 1866 elm->internal.subtree_count = ondisk->count;
195c19a1 1867 }
9944ae54
MD
1868 if (elm->base.subtree_type != HAMMER_BTREE_TYPE_CLUSTER &&
1869 elm->base.subtree_type != ondisk->type) {
195c19a1 1870 hammer_modify_node(parent);
9944ae54 1871 elm->base.subtree_type = ondisk->type;
8cd0a023
MD
1872 }
1873 }
b3deaf57
MD
1874
1875success:
195c19a1 1876 /*
b3deaf57
MD
1877 * Free the saved node. If the saved node was the root of a
1878 * cluster, free the entire cluster.
195c19a1
MD
1879 */
1880 hammer_flush_node(save);
b3deaf57
MD
1881 save->flags |= HAMMER_NODE_DELETED;
1882
1883 error = 0;
1884failure:
195c19a1
MD
1885 hammer_unlock(&save->lock);
1886 hammer_rel_node(save);
b3deaf57 1887 return(error);
8cd0a023
MD
1888}
1889
7f7c1f84
MD
1890/*
1891 * The child represented by the element in internal node node needs
1892 * to have its parent pointer adjusted.
1893 */
1894static
1895int
1896btree_set_parent(hammer_node_t node, hammer_btree_elm_t elm)
1897{
1898 hammer_volume_t volume;
1899 hammer_cluster_t cluster;
1900 hammer_node_t child;
1901 int error;
1902
1903 error = 0;
1904
9944ae54 1905 switch(elm->internal.base.subtree_type) {
7f7c1f84
MD
1906 case HAMMER_BTREE_TYPE_LEAF:
1907 case HAMMER_BTREE_TYPE_INTERNAL:
1908 child = hammer_get_node(node->cluster,
1909 elm->internal.subtree_offset, &error);
1910 if (error == 0) {
0b075555 1911 hammer_modify_node(child);
7f7c1f84
MD
1912 hammer_lock_ex(&child->lock);
1913 child->ondisk->parent = node->node_offset;
7f7c1f84
MD
1914 hammer_unlock(&child->lock);
1915 hammer_rel_node(child);
1916 }
1917 break;
1918 case HAMMER_BTREE_TYPE_CLUSTER:
1919 volume = hammer_get_volume(node->cluster->volume->hmp,
d26d0ae9 1920 elm->internal.subtree_vol_no, &error);
7f7c1f84
MD
1921 if (error)
1922 break;
1923 cluster = hammer_get_cluster(volume,
d26d0ae9 1924 elm->internal.subtree_clu_no,
7f7c1f84
MD
1925 &error, 0);
1926 hammer_rel_volume(volume, 0);
1927 if (error)
1928 break;
0b075555 1929 hammer_modify_cluster(cluster);
7f7c1f84
MD
1930 hammer_lock_ex(&cluster->io.lock);
1931 cluster->ondisk->clu_btree_parent_offset = node->node_offset;
1932 hammer_unlock(&cluster->io.lock);
1933 KKASSERT(cluster->ondisk->clu_btree_parent_clu_no ==
1934 node->cluster->clu_no);
1935 KKASSERT(cluster->ondisk->clu_btree_parent_vol_no ==
1936 node->cluster->volume->vol_no);
7f7c1f84
MD
1937 hammer_rel_cluster(cluster, 0);
1938 break;
1939 default:
1940 hammer_print_btree_elm(elm, HAMMER_BTREE_TYPE_INTERNAL, -1);
1941 panic("btree_set_parent: bad subtree_type");
1942 break; /* NOT REACHED */
1943 }
1944 return(error);
1945}
1946
8cd0a023
MD
1947/************************************************************************
1948 * MISCELLANIOUS SUPPORT *
1949 ************************************************************************/
1950
1951/*
d26d0ae9 1952 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
8cd0a023 1953 *
d113fda1
MD
1954 * Note that for this particular function a return value of -1, 0, or +1
1955 * can denote a match if create_tid is otherwise discounted.
1956 *
8cd0a023 1957 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
8cd0a023
MD
1958 */
1959int
1960hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
1961{
d26d0ae9
MD
1962 if (key1->obj_id < key2->obj_id)
1963 return(-4);
1964 if (key1->obj_id > key2->obj_id)
1965 return(4);
8cd0a023 1966
d26d0ae9
MD
1967 if (key1->rec_type < key2->rec_type)
1968 return(-3);
1969 if (key1->rec_type > key2->rec_type)
1970 return(3);
8cd0a023 1971
8cd0a023
MD
1972 if (key1->key < key2->key)
1973 return(-2);
1974 if (key1->key > key2->key)
1975 return(2);
d113fda1
MD
1976
1977 if (key1->create_tid < key2->create_tid)
1978 return(-1);
1979 if (key1->create_tid > key2->create_tid)
1980 return(1);
8cd0a023
MD
1981 return(0);
1982}
1983
c0ade690 1984/*
d26d0ae9
MD
1985 * Test a non-zero timestamp against an element to determine whether the
1986 * element is visible.
c0ade690
MD
1987 */
1988int
d26d0ae9 1989hammer_btree_chkts(hammer_tid_t create_tid, hammer_base_elm_t base)
c0ade690 1990{
d26d0ae9
MD
1991 if (create_tid < base->create_tid)
1992 return(-1);
1993 if (base->delete_tid && create_tid >= base->delete_tid)
1994 return(1);
c0ade690
MD
1995 return(0);
1996}
1997
8cd0a023
MD
1998/*
1999 * Create a separator half way inbetween key1 and key2. For fields just
2000 * one unit apart, the separator will match key2.
2001 *
d113fda1 2002 * At the moment require that the separator never match key2 exactly.
8cd0a023 2003 *
d113fda1
MD
2004 * We have to special case the separator between two historical keys,
2005 * where all elements except create_tid match. In this case our B-Tree
2006 * searches can't figure out which branch of an internal node to go down
2007 * unless the mid point's create_tid is exactly key2.
2008 * (see btree_search()'s scan code on HAMMER_BTREE_TYPE_INTERNAL).
8cd0a023
MD
2009 */
2010#define MAKE_SEPARATOR(key1, key2, dest, field) \
2011 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2012
2013static void
2014hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2015 hammer_base_elm_t dest)
2016{
2017 bzero(dest, sizeof(*dest));
2018 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2019 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2020 MAKE_SEPARATOR(key1, key2, dest, key);
d113fda1
MD
2021 if (key1->obj_id == key2->obj_id &&
2022 key1->rec_type == key2->rec_type &&
2023 key1->key == key2->key) {
2024 dest->create_tid = key2->create_tid;
2025 } else {
2026 dest->create_tid = 0;
2027 }
8cd0a023
MD
2028}
2029
2030#undef MAKE_SEPARATOR
2031
9944ae54 2032#if 0
8cd0a023
MD
2033/*
2034 * Return whether a generic internal or leaf node is full
2035 */
2036static int
2037btree_node_is_full(hammer_node_ondisk_t node)
2038{
2039 switch(node->type) {
2040 case HAMMER_BTREE_TYPE_INTERNAL:
2041 if (node->count == HAMMER_BTREE_INT_ELMS)
2042 return(1);
2043 break;
2044 case HAMMER_BTREE_TYPE_LEAF:
2045 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2046 return(1);
2047 break;
2048 default:
2049 panic("illegal btree subtype");
2050 }
2051 return(0);
2052}
9944ae54
MD
2053#endif
2054
2055/*
2056 * Return whether a generic internal or leaf node is almost full. This
2057 * routine is used as a helper for search insertions to guarentee at
2058 * least 2 available slots in the internal node(s) leading up to a leaf,
2059 * so hammer_btree_insert_cluster() will function properly.
2060 */
2061static int
2062btree_node_is_almost_full(hammer_node_ondisk_t node)
2063{
2064 switch(node->type) {
2065 case HAMMER_BTREE_TYPE_INTERNAL:
2066 if (node->count > HAMMER_BTREE_INT_ELMS - 2)
2067 return(1);
2068 break;
2069 case HAMMER_BTREE_TYPE_LEAF:
2070 if (node->count > HAMMER_BTREE_LEAF_ELMS - 2)
2071 return(1);
2072 break;
2073 default:
2074 panic("illegal btree subtype");
2075 }
2076 return(0);
2077}
8cd0a023
MD
2078
2079#if 0
2080static int
2081btree_max_elements(u_int8_t type)
2082{
2083 if (type == HAMMER_BTREE_TYPE_LEAF)
2084 return(HAMMER_BTREE_LEAF_ELMS);
2085 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2086 return(HAMMER_BTREE_INT_ELMS);
2087 panic("btree_max_elements: bad type %d\n", type);
2088}
2089#endif
2090
c0ade690
MD
2091void
2092hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2093{
2094 hammer_btree_elm_t elm;
2095 int i;
2096
2097 kprintf("node %p count=%d parent=%d type=%c\n",
2098 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2099
2100 /*
2101 * Dump both boundary elements if an internal node
2102 */
2103 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2104 for (i = 0; i <= ondisk->count; ++i) {
2105 elm = &ondisk->elms[i];
2106 hammer_print_btree_elm(elm, ondisk->type, i);
2107 }
2108 } else {
2109 for (i = 0; i < ondisk->count; ++i) {
2110 elm = &ondisk->elms[i];
2111 hammer_print_btree_elm(elm, ondisk->type, i);
2112 }
2113 }
2114}
2115
2116void
2117hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2118{
2119 kprintf(" %2d", i);
2120 kprintf("\tobjid = %016llx\n", elm->base.obj_id);
2121 kprintf("\tkey = %016llx\n", elm->base.key);
2122 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2123 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2124 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2125 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
9944ae54 2126 kprintf("\tsubtree_type = %02x\n", elm->base.subtree_type);
c0ade690
MD
2127
2128 if (type == HAMMER_BTREE_TYPE_INTERNAL) {
2129 if (elm->internal.rec_offset) {
2130 kprintf("\tcluster_rec = %08x\n",
2131 elm->internal.rec_offset);
2132 kprintf("\tcluster_id = %08x\n",
d26d0ae9 2133 elm->internal.subtree_clu_no);
c0ade690 2134 kprintf("\tvolno = %08x\n",
d26d0ae9 2135 elm->internal.subtree_vol_no);
c0ade690
MD
2136 } else {
2137 kprintf("\tsubtree_off = %08x\n",
2138 elm->internal.subtree_offset);
2139 }
2140 kprintf("\tsubtree_count= %d\n", elm->internal.subtree_count);
2141 } else {
2142 kprintf("\trec_offset = %08x\n", elm->leaf.rec_offset);
2143 kprintf("\tdata_offset = %08x\n", elm->leaf.data_offset);
2144 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2145 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);
2146 }
2147}