2 * Copyright (c) 1990, 1993, 1994
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * @(#)bt_split.c 8.9 (Berkeley) 7/26/94
33 * $DragonFly: src/lib/libc/db/btree/bt_split.c,v 1.8 2005/11/19 20:46:32 swildner Exp $
36 #include <sys/types.h>
46 static int bt_broot (BTREE *, PAGE *, PAGE *, PAGE *);
48 (BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
49 static int bt_preserve (BTREE *, pgno_t);
50 static PAGE *bt_psplit
51 (BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
53 (BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
54 static int bt_rroot (BTREE *, PAGE *, PAGE *, PAGE *);
55 static recno_t rec_total (PAGE *);
58 u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
62 * __BT_SPLIT -- Split the tree.
68 * data: data to insert
69 * flags: BIGKEY/BIGDATA flags
71 * skip: index to leave open
74 * RET_ERROR, RET_SUCCESS
77 __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
78 size_t ilen, u_int32_t argskip)
84 PAGE *h, *l, *r, *lchild, *rchild;
87 u_int32_t n, nbytes, nksize;
95 * Split the page into two pages, l and r. The split routines return
96 * a pointer to the page into which the key should be inserted and with
97 * skip set to the offset which should be used. Additionally, l and r
101 h = sp->pgno == P_ROOT ?
102 bt_root(t, sp, &l, &r, &skip, ilen) :
103 bt_page(t, sp, &l, &r, &skip, ilen);
108 * Insert the new key/data pair into the leaf page. (Key inserts
109 * always cause a leaf page to split first.)
111 h->linp[skip] = h->upper -= ilen;
112 dest = (char *)h + h->upper;
113 if (F_ISSET(t, R_RECNO))
114 WR_RLEAF(dest, data, flags)
116 WR_BLEAF(dest, key, data, flags)
118 /* If the root page was split, make it look right. */
119 if (sp->pgno == P_ROOT &&
120 (F_ISSET(t, R_RECNO) ?
121 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
125 * Now we walk the parent page stack -- a LIFO stack of the pages that
126 * were traversed when we searched for the page that split. Each stack
127 * entry is a page number and a page index offset. The offset is for
128 * the page traversed on the search. We've just split a page, so we
129 * have to insert a new key into the parent page.
131 * If the insert into the parent page causes it to split, may have to
132 * continue splitting all the way up the tree. We stop if the root
133 * splits or the page inserted into didn't have to split to hold the
134 * new key. Some algorithms replace the key for the old page as well
135 * as the new page. We don't, as there's no reason to believe that the
136 * first key on the old page is any better than the key we have, and,
137 * in the case of a key being placed at index 0 causing the split, the
138 * key is unavailable.
140 * There are a maximum of 5 pages pinned at any time. We keep the left
141 * and right pages pinned while working on the parent. The 5 are the
142 * two children, left parent and right parent (when the parent splits)
143 * and the root page or the overflow key page when calling bt_preserve.
144 * This code must make sure that all pins are released other than the
145 * root page or overflow page which is unlocked elsewhere.
147 while ((parent = BT_POP(t)) != NULL) {
151 /* Get the parent page. */
152 if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
156 * The new key goes ONE AFTER the index, because the split
159 skip = parent->index + 1;
162 * Calculate the space needed on the parent page.
164 * Prefix trees: space hack when inserting into BINTERNAL
165 * pages. Retain only what's needed to distinguish between
166 * the new entry and the LAST entry on the page to its left.
167 * If the keys compare equal, retain the entire key. Note,
168 * we don't touch overflow keys, and the entire key must be
169 * retained for the next-to-left most key on the leftmost
170 * page of each level, or the search will fail. Applicable
171 * ONLY to internal pages that have leaf pages as children.
172 * Further reduction of the key between pairs of internal
173 * pages loses too much information.
175 switch (rchild->flags & P_TYPE) {
177 bi = GETBINTERNAL(rchild, 0);
178 nbytes = NBINTERNAL(bi->ksize);
181 bl = GETBLEAF(rchild, 0);
182 nbytes = NBINTERNAL(bl->ksize);
183 if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
184 (h->prevpg != P_INVALID || skip > 1)) {
185 tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
190 nksize = t->bt_pfx(&a, &b);
191 n = NBINTERNAL(nksize);
194 bt_pfxsaved += nbytes - n;
210 /* Split the parent page if necessary or shift the indices. */
211 if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
213 h = h->pgno == P_ROOT ?
214 bt_root(t, h, &l, &r, &skip, nbytes) :
215 bt_page(t, h, &l, &r, &skip, nbytes);
220 if (skip < (nxtindex = NEXTINDEX(h)))
221 memmove(h->linp + skip + 1, h->linp + skip,
222 (nxtindex - skip) * sizeof(indx_t));
223 h->lower += sizeof(indx_t);
227 /* Insert the key into the parent page. */
228 switch (rchild->flags & P_TYPE) {
230 h->linp[skip] = h->upper -= nbytes;
231 dest = (char *)h + h->linp[skip];
232 memmove(dest, bi, nbytes);
233 ((BINTERNAL *)dest)->pgno = rchild->pgno;
236 h->linp[skip] = h->upper -= nbytes;
237 dest = (char *)h + h->linp[skip];
238 WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
239 rchild->pgno, bl->flags & P_BIGKEY);
240 memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
241 if (bl->flags & P_BIGKEY &&
242 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
247 * Update the left page count. If split
248 * added at index 0, fix the correct page.
251 dest = (char *)h + h->linp[skip - 1];
253 dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
254 ((RINTERNAL *)dest)->nrecs = rec_total(lchild);
255 ((RINTERNAL *)dest)->pgno = lchild->pgno;
257 /* Update the right page count. */
258 h->linp[skip] = h->upper -= nbytes;
259 dest = (char *)h + h->linp[skip];
260 ((RINTERNAL *)dest)->nrecs = rec_total(rchild);
261 ((RINTERNAL *)dest)->pgno = rchild->pgno;
265 * Update the left page count. If split
266 * added at index 0, fix the correct page.
269 dest = (char *)h + h->linp[skip - 1];
271 dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
272 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
273 ((RINTERNAL *)dest)->pgno = lchild->pgno;
275 /* Update the right page count. */
276 h->linp[skip] = h->upper -= nbytes;
277 dest = (char *)h + h->linp[skip];
278 ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
279 ((RINTERNAL *)dest)->pgno = rchild->pgno;
285 /* Unpin the held pages. */
287 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
291 /* If the root page was split, make it look right. */
292 if (sp->pgno == P_ROOT &&
293 (F_ISSET(t, R_RECNO) ?
294 bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
297 mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
298 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
301 /* Unpin the held pages. */
302 mpool_put(t->bt_mp, l, MPOOL_DIRTY);
303 mpool_put(t->bt_mp, r, MPOOL_DIRTY);
305 /* Clear any pages left on the stack. */
306 return (RET_SUCCESS);
309 * If something fails in the above loop we were already walking back
310 * up the tree and the tree is now inconsistent. Nothing much we can
311 * do about it but release any memory we're holding.
313 err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
314 mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
316 err2: mpool_put(t->bt_mp, l, 0);
317 mpool_put(t->bt_mp, r, 0);
318 __dbpanic(t->bt_dbp);
323 * BT_PAGE -- Split a non-root page of a btree.
328 * lp: pointer to left page pointer
329 * rp: pointer to right page pointer
330 * skip: pointer to index to leave open
331 * ilen: insert length
334 * Pointer to page in which to insert or NULL on error.
337 bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
345 /* Put the new right page for the split into place. */
346 if ((r = __bt_new(t, &npg)) == NULL)
349 r->lower = BTDATAOFF;
350 r->upper = t->bt_psize;
351 r->nextpg = h->nextpg;
353 r->flags = h->flags & P_TYPE;
356 * If we're splitting the last page on a level because we're appending
357 * a key to it (skip is NEXTINDEX()), it's likely that the data is
358 * sorted. Adding an empty page on the side of the level is less work
359 * and can push the fill factor much higher than normal. If we're
360 * wrong it's no big deal, we'll just do the split the right way next
361 * time. It may look like it's equally easy to do a similar hack for
362 * reverse sorted data, that is, split the tree left, but it's not.
365 if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
370 r->lower = BTDATAOFF + sizeof(indx_t);
377 /* Put the new left page for the split into place. */
378 if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
379 mpool_put(t->bt_mp, r, 0);
383 memset(l, 0xff, t->bt_psize);
387 l->prevpg = h->prevpg;
388 l->lower = BTDATAOFF;
389 l->upper = t->bt_psize;
390 l->flags = h->flags & P_TYPE;
392 /* Fix up the previous pointer of the page after the split page. */
393 if (h->nextpg != P_INVALID) {
394 if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
396 /* XXX mpool_free(t->bt_mp, r->pgno); */
399 tp->prevpg = r->pgno;
400 mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
404 * Split right. The key/data pairs aren't sorted in the btree page so
405 * it's simpler to copy the data from the split page onto two new pages
406 * instead of copying half the data to the right page and compacting
407 * the left page in place. Since the left page can't change, we have
408 * to swap the original and the allocated left page after the split.
410 tp = bt_psplit(t, h, l, r, skip, ilen);
412 /* Move the new left page onto the old left page. */
413 memmove(h, l, t->bt_psize);
424 * BT_ROOT -- Split the root page of a btree.
429 * lp: pointer to left page pointer
430 * rp: pointer to right page pointer
431 * skip: pointer to index to leave open
432 * ilen: insert length
435 * Pointer to page in which to insert or NULL on error.
438 bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
447 /* Put the new left and right pages for the split into place. */
448 if ((l = __bt_new(t, &lnpg)) == NULL ||
449 (r = __bt_new(t, &rnpg)) == NULL)
455 l->prevpg = r->nextpg = P_INVALID;
456 l->lower = r->lower = BTDATAOFF;
457 l->upper = r->upper = t->bt_psize;
458 l->flags = r->flags = h->flags & P_TYPE;
460 /* Split the root page. */
461 tp = bt_psplit(t, h, l, r, skip, ilen);
469 * BT_RROOT -- Fix up the recno root page after it has been split.
478 * RET_ERROR, RET_SUCCESS
481 bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
485 /* Insert the left and right keys, set the header information. */
486 h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
487 dest = (char *)h + h->upper;
489 l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
491 h->linp[1] = h->upper -= NRINTERNAL;
492 dest = (char *)h + h->upper;
494 r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
496 h->lower = BTDATAOFF + 2 * sizeof(indx_t);
498 /* Unpin the root page, set to recno internal page. */
500 h->flags |= P_RINTERNAL;
501 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
503 return (RET_SUCCESS);
507 * BT_BROOT -- Fix up the btree root page after it has been split.
516 * RET_ERROR, RET_SUCCESS
519 bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
527 * If the root page was a leaf page, change it into an internal page.
528 * We copy the key we split on (but not the key's data, in the case of
529 * a leaf page) to the new root page.
531 * The btree comparison code guarantees that the left-most key on any
532 * level of the tree is never used, so it doesn't need to be filled in.
534 nbytes = NBINTERNAL(0);
535 h->linp[0] = h->upper = t->bt_psize - nbytes;
536 dest = (char *)h + h->upper;
537 WR_BINTERNAL(dest, 0, l->pgno, 0);
539 switch (h->flags & P_TYPE) {
542 nbytes = NBINTERNAL(bl->ksize);
543 h->linp[1] = h->upper -= nbytes;
544 dest = (char *)h + h->upper;
545 WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
546 memmove(dest, bl->bytes, bl->ksize);
549 * If the key is on an overflow page, mark the overflow chain
550 * so it isn't deleted when the leaf copy of the key is deleted.
552 if (bl->flags & P_BIGKEY &&
553 bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
557 bi = GETBINTERNAL(r, 0);
558 nbytes = NBINTERNAL(bi->ksize);
559 h->linp[1] = h->upper -= nbytes;
560 dest = (char *)h + h->upper;
561 memmove(dest, bi, nbytes);
562 ((BINTERNAL *)dest)->pgno = r->pgno;
568 /* There are two keys on the page. */
569 h->lower = BTDATAOFF + 2 * sizeof(indx_t);
571 /* Unpin the root page, set to btree internal page. */
573 h->flags |= P_BINTERNAL;
574 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
576 return (RET_SUCCESS);
580 * BT_PSPLIT -- Do the real work of splitting the page.
584 * h: page to be split
585 * l: page to put lower half of data
586 * r: page to put upper half of data
587 * pskip: pointer to index to leave open
588 * ilen: insert length
591 * Pointer to page in which to insert.
594 bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
602 indx_t full, half, nxt, off, skip, top, used;
604 int bigkeycnt, isbigkey;
608 * Split the data to the left and right pages. Leave the skip index
609 * open. Additionally, make some effort not to split on an overflow
610 * key. This makes internal page processing faster and can save
611 * space as overflow keys used by internal pages are never deleted.
615 full = t->bt_psize - BTDATAOFF;
618 for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
621 isbigkey = 0; /* XXX: not really known. */
623 switch (h->flags & P_TYPE) {
625 src = bi = GETBINTERNAL(h, nxt);
626 nbytes = NBINTERNAL(bi->ksize);
627 isbigkey = bi->flags & P_BIGKEY;
630 src = bl = GETBLEAF(h, nxt);
632 isbigkey = bl->flags & P_BIGKEY;
635 src = GETRINTERNAL(h, nxt);
640 src = rl = GETRLEAF(h, nxt);
649 * If the key/data pairs are substantial fractions of the max
650 * possible size for the page, it's possible to get situations
651 * where we decide to try and copy too much onto the left page.
652 * Make sure that doesn't happen.
655 used + nbytes + sizeof(indx_t) >= full) || nxt == top - 1) {
660 /* Copy the key/data pair, if not the skipped index. */
664 l->linp[off] = l->upper -= nbytes;
665 memmove((char *)l + l->upper, src, nbytes);
668 used += nbytes + sizeof(indx_t);
670 if (!isbigkey || bigkeycnt == 3)
678 * Off is the last offset that's valid for the left page.
679 * Nxt is the first offset to be placed on the right page.
681 l->lower += (off + 1) * sizeof(indx_t);
684 * If splitting the page that the cursor was on, the cursor has to be
685 * adjusted to point to the same record as before the split. If the
686 * cursor is at or past the skipped slot, the cursor is incremented by
687 * one. If the cursor is on the right page, it is decremented by the
688 * number of records split to the left page.
691 if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
692 if (c->pg.index >= skip)
694 if (c->pg.index < nxt) /* Left page. */
695 c->pg.pgno = l->pgno;
696 else { /* Right page. */
697 c->pg.pgno = r->pgno;
703 * If the skipped index was on the left page, just return that page.
704 * Otherwise, adjust the skip index to reflect the new position on
708 skip = MAX_PAGE_OFFSET;
715 for (off = 0; nxt < top; ++off) {
718 skip = MAX_PAGE_OFFSET;
720 switch (h->flags & P_TYPE) {
722 src = bi = GETBINTERNAL(h, nxt);
723 nbytes = NBINTERNAL(bi->ksize);
726 src = bl = GETBLEAF(h, nxt);
730 src = GETRINTERNAL(h, nxt);
734 src = rl = GETRLEAF(h, nxt);
741 r->linp[off] = r->upper -= nbytes;
742 memmove((char *)r + r->upper, src, nbytes);
744 r->lower += off * sizeof(indx_t);
746 /* If the key is being appended to the page, adjust the index. */
748 r->lower += sizeof(indx_t);
754 * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
756 * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
757 * record that references them gets deleted. Chains pointed to by internal
758 * pages never get deleted. This routine marks a chain as pointed to by an
763 * pg: page number of first page in the chain.
766 * RET_SUCCESS, RET_ERROR.
769 bt_preserve(BTREE *t, pgno_t pg)
773 if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
775 h->flags |= P_PRESERVE;
776 mpool_put(t->bt_mp, h, MPOOL_DIRTY);
777 return (RET_SUCCESS);
781 * REC_TOTAL -- Return the number of recno entries below a page.
787 * The number of recno entries below a page.
790 * These values could be set by the bt_psplit routine. The problem is that the
791 * entry has to be popped off of the stack etc. or the values have to be passed
792 * all the way back to bt_split/bt_rroot and it's not very clean.
800 for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
801 recs += GETRINTERNAL(h, nxt)->nrecs;