xref: /csrg-svn/lib/libc/db/btree/bt_split.c (revision 56402)
146134Smao /*-
246134Smao  * Copyright (c) 1990 The Regents of the University of California.
346134Smao  * All rights reserved.
446134Smao  *
546134Smao  * This code is derived from software contributed to Berkeley by
646134Smao  * Mike Olson.
746134Smao  *
846134Smao  * %sccs.include.redist.c%
946134Smao  */
1046134Smao 
1146134Smao #if defined(LIBC_SCCS) && !defined(lint)
12*56402Sbostic static char sccsid[] = "@(#)bt_split.c	5.6 (Berkeley) 10/03/92";
1346134Smao #endif /* LIBC_SCCS and not lint */
1446134Smao 
1546134Smao #include <sys/types.h>
1650989Sbostic #define	__DBINTERFACE_PRIVATE
1746134Smao #include <db.h>
1850989Sbostic #include <limits.h>
1950989Sbostic #include <stdio.h>
2046561Sbostic #include <stdlib.h>
2146561Sbostic #include <string.h>
2246134Smao #include "btree.h"
2346134Smao 
2450989Sbostic static int	 bt_preserve __P((BTREE *, pgno_t));
2550989Sbostic static PAGE	*bt_psplit __P((BTREE *, PAGE *, PAGE *, PAGE *, int *));
2650989Sbostic static PAGE	*bt_page __P((BTREE *, PAGE *, PAGE **, PAGE **, int *));
2750989Sbostic static PAGE	*bt_root __P((BTREE *, PAGE *, PAGE **, PAGE **, int *));
2850989Sbostic static int	 bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
2950989Sbostic static int	 bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *));
3050989Sbostic static recno_t	 rec_total __P((PAGE *));
3150989Sbostic 
3250989Sbostic #ifdef STATISTICS
3350989Sbostic u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
3450989Sbostic #endif
3550989Sbostic 
3646134Smao /*
3750989Sbostic  * __BT_SPLIT -- Split the tree.
3846134Smao  *
3950989Sbostic  * Parameters:
4050989Sbostic  *	t:	tree
4150989Sbostic  *	h:	page to split
4250989Sbostic  *	key:	key to insert
4350989Sbostic  *	data:	data to insert
4450989Sbostic  *	flags:	BIGKEY/BIGDATA flags
4550989Sbostic  *	nbytes:	length of insertion
4650989Sbostic  *	skip:	index to leave open
4746134Smao  *
4850989Sbostic  * Returns:
4950989Sbostic  *	RET_ERROR, RET_SUCCESS
5046134Smao  */
5146134Smao int
5250989Sbostic __bt_split(t, h, key, data, flags, nbytes, skip)
5350989Sbostic 	BTREE *t;
5450989Sbostic 	PAGE *h;
5550989Sbostic 	const DBT *key, *data;
5650989Sbostic 	u_long flags;
5750989Sbostic 	size_t nbytes;
5850989Sbostic 	int skip;
5946134Smao {
6050989Sbostic 	BINTERNAL *bi;
6150989Sbostic 	BLEAF *bl;
6250989Sbostic 	DBT a, b;
6350989Sbostic 	EPGNO *parent;
6450989Sbostic 	PAGE *l, *r, *lchild, *rchild;
6550989Sbostic 	index_t nxtindex;
6650989Sbostic 	size_t nksize;
6750989Sbostic 	int nosplit;
6850989Sbostic 	char *dest;
6946134Smao 
7046134Smao 	/*
7150989Sbostic 	 * Split the page into two pages, l and r.  The split routines return
7250989Sbostic 	 * a pointer to the page into which the key should be inserted and skip
7350989Sbostic 	 * set to the offset which should be used.  Additionally, l and r are
7450989Sbostic 	 * pinned.
7546134Smao 	 */
7650989Sbostic 	h = h->pgno == P_ROOT ?
7750989Sbostic 	    bt_root(t, h, &l, &r, &skip) : bt_page(t, h, &l, &r, &skip);
7850989Sbostic 	if (h == NULL)
7946134Smao 		return (RET_ERROR);
8046134Smao 
8150989Sbostic 	/*
8250989Sbostic 	 * Grab the space and insert the [rb]leaf structure.  Always a [rb]leaf
8350989Sbostic 	 * structure since key inserts always cause a leaf page to split first.
8450989Sbostic 	 */
8550989Sbostic 	h->linp[skip] = h->upper -= nbytes;
8650989Sbostic 	dest = (char *)h + h->upper;
8751106Sbostic 	if (ISSET(t, BTF_RECNO))
8850989Sbostic 		WR_RLEAF(dest, data, flags)
8951106Sbostic 	else
9050989Sbostic 		WR_BLEAF(dest, key, data, flags)
9146134Smao 
9246134Smao 	/*
9350989Sbostic 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
9450989Sbostic 	 * were traversed when we searched for the page that split.  Each stack
9550989Sbostic 	 * entry is a page number and a page index offset.  The offset is for
9650989Sbostic 	 * the page traversed on the search.  We've just split a page, so we
9750989Sbostic 	 * have to insert a new key into the parent page.
9850989Sbostic 	 *
9950989Sbostic 	 * If the insert into the parent page causes it to split, may have to
10050989Sbostic 	 * continue splitting all the way up the tree.  We stop if the root
10150989Sbostic 	 * splits or the page inserted into didn't have to split to hold the
10250989Sbostic 	 * new key.  Some algorithms replace the key for the old page as well
10350989Sbostic 	 * as the new page.  We don't, as there's no reason to believe that the
10450989Sbostic 	 * first key on the old page is any better than the key we have, and,
10550989Sbostic 	 * in the case of a key being placed at index 0 causing the split, the
10650989Sbostic 	 * key is unavailable.
10750989Sbostic 	 *
10850989Sbostic 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
10950989Sbostic 	 * and right pages pinned while working on the parent.   The 5 are the
11050989Sbostic 	 * two children, left parent and right parent (when the parent splits)
11150989Sbostic 	 * and the root page or the overflow key page when calling bt_preserve.
11250989Sbostic 	 * This code must make sure that all pins are released other than the
11350989Sbostic 	 * root page or overflow page which is unlocked elsewhere.
11446134Smao 	 */
11550989Sbostic 	for (nosplit = 0; (parent = BT_POP(t)) != NULL;) {
11650989Sbostic 		lchild = l;
11750989Sbostic 		rchild = r;
11846134Smao 
11950989Sbostic 		/* Get the parent page. */
12050989Sbostic 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
12150989Sbostic 			goto err2;
12246134Smao 
12350989Sbostic 	 	/* The new key goes ONE AFTER the index. */
12450989Sbostic 		skip = parent->index + 1;
12546134Smao 
12650989Sbostic 		/*
12750989Sbostic 		 * Calculate the space needed on the parent page.
12850989Sbostic 		 *
12951106Sbostic 		 * Space hack when inserting into BINTERNAL pages.  Only need to
13050989Sbostic 		 * retain the number of bytes that will distinguish between the
13151106Sbostic 		 * new entry and the LAST entry on the page to its left.  If the
13251106Sbostic 		 * keys compare equal, retain the entire key.  Note, we don't
13351106Sbostic 		 * touch overflow keys and the entire key must be retained for
13451106Sbostic 		 * the next-to-leftmost key on the leftmost page of each level,
13551106Sbostic 		 * or the search will fail.
13650989Sbostic 		 */
13750989Sbostic 		switch (rchild->flags & P_TYPE) {
13850989Sbostic 		case P_BINTERNAL:
13950989Sbostic 			bi = GETBINTERNAL(rchild, 0);
14050989Sbostic 			nbytes = NBINTERNAL(bi->ksize);
14150989Sbostic 			if (t->bt_pfx && (h->prevpg != P_INVALID || skip > 1) &&
14250989Sbostic 			    !(bi->flags & P_BIGKEY)) {
14350989Sbostic 				BINTERNAL *tbi;
14450989Sbostic 				tbi =
14550989Sbostic 				    GETBINTERNAL(lchild, NEXTINDEX(lchild) - 1);
14650989Sbostic 				a.size = tbi->ksize;
14750989Sbostic 				a.data = tbi->bytes;
14850989Sbostic 				b.size = bi->ksize;
14950989Sbostic 				b.data = bi->bytes;
15050989Sbostic 				goto prefix;
15146134Smao 			}
15250989Sbostic 			break;
15350989Sbostic 		case P_BLEAF:
15450989Sbostic 			bl = GETBLEAF(rchild, 0);
15556401Sbostic 			nbytes = NBINTERNAL(bl->ksize);
15650989Sbostic 			if (t->bt_pfx && (h->prevpg != P_INVALID || skip > 1) &&
15750989Sbostic 			    !(bl->flags & P_BIGKEY)) {
15850989Sbostic 				BLEAF *tbl;
15950989Sbostic 				size_t n;
16050989Sbostic 
16150989Sbostic 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
16250989Sbostic 				a.size = tbl->ksize;
16350989Sbostic 				a.data = tbl->bytes;
16450989Sbostic 				b.size = bl->ksize;
16550989Sbostic 				b.data = bl->bytes;
16650989Sbostic prefix:				nksize = t->bt_pfx(&a, &b);
16750989Sbostic 				n = NBINTERNAL(nksize);
16850989Sbostic 				if (n < nbytes) {
16950989Sbostic #ifdef STATISTICS
17050989Sbostic 					bt_pfxsaved += nbytes - n;
17150989Sbostic #endif
17250989Sbostic 					nbytes = n;
17350989Sbostic 				} else
17450989Sbostic 					nksize = 0;
17550989Sbostic 			} else
17650989Sbostic 				nksize = 0;
17750989Sbostic 			break;
17850989Sbostic 		case P_RINTERNAL:
17950989Sbostic 		case P_RLEAF:
18050989Sbostic 			nbytes = NRINTERNAL;
18150989Sbostic 			break;
18250989Sbostic 		}
18350989Sbostic 
18450989Sbostic 		/* Split the parent page if necessary or shift the indices. */
18550989Sbostic 		if (h->upper - h->lower < nbytes + sizeof(index_t)) {
18650989Sbostic 			h = h->pgno == P_ROOT ?
18750989Sbostic 			    bt_root(t, h, &l, &r, &skip) :
18850989Sbostic 			    bt_page(t, h, &l, &r, &skip);
18950989Sbostic 			if (h == NULL)
19050989Sbostic 				goto err1;
19146134Smao 		} else {
19250989Sbostic 			if (skip < (nxtindex = NEXTINDEX(h)))
19350989Sbostic 				bcopy(h->linp + skip, h->linp + skip + 1,
19450989Sbostic 				    (nxtindex - skip) * sizeof(index_t));
19550989Sbostic 			h->lower += sizeof(index_t);
19650989Sbostic 			nosplit = 1;
19746134Smao 		}
19846134Smao 
19950989Sbostic 		/* Insert the key into the parent page. */
20050989Sbostic 		switch(rchild->flags & P_TYPE) {
20150989Sbostic 		case P_BINTERNAL:
20250989Sbostic 			h->linp[skip] = h->upper -= nbytes;
20350989Sbostic 			dest = (char *)h + h->linp[skip];
20450989Sbostic 			bcopy(bi, dest, nbytes);
20550989Sbostic 			if (nksize)
20650989Sbostic 				((BINTERNAL *)dest)->ksize = nksize;
20750989Sbostic 			((BINTERNAL *)dest)->pgno = rchild->pgno;
20850989Sbostic 			break;
20950989Sbostic 		case P_BLEAF:
21050989Sbostic 			h->linp[skip] = h->upper -= nbytes;
21150989Sbostic 			dest = (char *)h + h->linp[skip];
21250989Sbostic 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
21351106Sbostic 			    rchild->pgno, bl->flags & P_BIGKEY);
21450989Sbostic 			bcopy(bl->bytes, dest, nksize ? nksize : bl->ksize);
21550989Sbostic 			if (bl->flags & P_BIGKEY &&
21650989Sbostic 			    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
21750989Sbostic 				goto err1;
21850989Sbostic 			break;
21950989Sbostic 		case P_RINTERNAL:
22050989Sbostic 			/* Update both left and right page counts. */
22150989Sbostic 			h->linp[skip] = h->upper -= nbytes;
22250989Sbostic 			dest = (char *)h + h->linp[skip];
22350989Sbostic 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
22450989Sbostic 			((RINTERNAL *)dest)->pgno = rchild->pgno;
22550989Sbostic 			dest = (char *)h + h->linp[skip - 1];
22650989Sbostic 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
22750989Sbostic 			((RINTERNAL *)dest)->pgno = lchild->pgno;
22850989Sbostic 			break;
22950989Sbostic 		case P_RLEAF:
23050989Sbostic 			/* Update both left and right page counts. */
23150989Sbostic 			h->linp[skip] = h->upper -= nbytes;
23250989Sbostic 			dest = (char *)h + h->linp[skip];
23350989Sbostic 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
23450989Sbostic 			((RINTERNAL *)dest)->pgno = rchild->pgno;
23550989Sbostic 			dest = (char *)h + h->linp[skip - 1];
23650989Sbostic 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
23750989Sbostic 			((RINTERNAL *)dest)->pgno = lchild->pgno;
23850989Sbostic 			break;
23950989Sbostic 		}
24046134Smao 
24150989Sbostic 		/* Unpin the held pages. */
24250989Sbostic 		if (nosplit) {
24350989Sbostic 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
24450989Sbostic 			break;
24550989Sbostic 		}
24650989Sbostic 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
24750989Sbostic 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
24850989Sbostic 	}
24946134Smao 
25050989Sbostic 	/* Unpin the held pages. */
25150989Sbostic 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
25250989Sbostic 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
25350989Sbostic 
25451106Sbostic 	/* Clear any pages left on the stack. */
25551106Sbostic 	BT_CLR(t);
25650989Sbostic 	return (RET_SUCCESS);
25746134Smao 
25846134Smao 	/*
25950989Sbostic 	 * If something fails in the above loop we were already walking back
26050989Sbostic 	 * up the tree and the tree is now inconsistent.  Nothing much we can
26150989Sbostic 	 * do about it but release any memory we're holding.
26246134Smao 	 */
26350989Sbostic err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
26450989Sbostic 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
26546134Smao 
26650989Sbostic err2:	mpool_put(t->bt_mp, l, 0);
26750989Sbostic 	mpool_put(t->bt_mp, r, 0);
26850989Sbostic 	__dbpanic(t->bt_dbp);
26950989Sbostic 	return (RET_ERROR);
27046134Smao }
27146134Smao 
27246134Smao /*
27350989Sbostic  * BT_PAGE -- Split a non-root page of a btree.
27446134Smao  *
27550989Sbostic  * Parameters:
27650989Sbostic  *	t:	tree
27750989Sbostic  *	h:	root page
27850989Sbostic  *	lp:	pointer to left page pointer
27950989Sbostic  *	rp:	pointer to right page pointer
28050989Sbostic  *	skip:	pointer to index to leave open
28146134Smao  *
28250989Sbostic  * Returns:
28350989Sbostic  *	Pointer to page in which to insert or NULL on error.
28446134Smao  */
28550989Sbostic static PAGE *
28650989Sbostic bt_page(t, h, lp, rp, skip)
28750989Sbostic 	BTREE *t;
28850989Sbostic 	PAGE *h, **lp, **rp;
28950989Sbostic 	int *skip;
29046134Smao {
29150989Sbostic 	PAGE *l, *r, *tp;
29250989Sbostic 	pgno_t npg;
29346134Smao 
29450989Sbostic #ifdef STATISTICS
29550989Sbostic 	++bt_split;
29650989Sbostic #endif
29750989Sbostic 	/* Put the new right page for the split into place. */
29850989Sbostic 	if ((r = mpool_new(t->bt_mp, &npg)) == NULL)
29950989Sbostic 		return (NULL);
30050989Sbostic 	r->pgno = npg;
30150989Sbostic 	r->lower = BTDATAOFF;
30250989Sbostic 	r->upper = t->bt_psize;
30350989Sbostic 	r->nextpg = h->nextpg;
30450989Sbostic 	r->prevpg = h->pgno;
30550989Sbostic 	r->flags = h->flags & P_TYPE;
30646134Smao 
30750989Sbostic 	/*
30850989Sbostic 	 * If we're splitting the last page on a level because we're appending
30950989Sbostic 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
31050989Sbostic 	 * sorted.  Adding an empty page on the side of the level is less work
31150989Sbostic 	 * and can push the fill factor much higher than normal.  If we're
31250989Sbostic 	 * wrong it's no big deal, we'll just do the split the right way next
31350989Sbostic 	 * time.  It may look like it's equally easy to do a similar hack for
31450989Sbostic 	 * reverse sorted data, that is, split the tree left, but it's not.
31550989Sbostic 	 * Don't even try.
31650989Sbostic 	 */
31750989Sbostic 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
31850989Sbostic #ifdef STATISTICS
31950989Sbostic 		++bt_sortsplit;
32050989Sbostic #endif
32150989Sbostic 		h->nextpg = r->pgno;
32250989Sbostic 		r->lower = BTDATAOFF + sizeof(index_t);
32350989Sbostic 		*skip = 0;
32450989Sbostic 		*lp = h;
32550989Sbostic 		*rp = r;
32650989Sbostic 		return (r);
32750989Sbostic 	}
32846134Smao 
32950989Sbostic 	/* Put the new left page for the split into place. */
33050989Sbostic 	if ((l = malloc(t->bt_psize)) == NULL) {
33150989Sbostic 		mpool_put(t->bt_mp, r, 0);
33250989Sbostic 		return (NULL);
33350989Sbostic 	}
33450989Sbostic 	l->pgno = h->pgno;
33550989Sbostic 	l->nextpg = r->pgno;
33650989Sbostic 	l->prevpg = h->prevpg;
33750989Sbostic 	l->lower = BTDATAOFF;
33850989Sbostic 	l->upper = t->bt_psize;
33950989Sbostic 	l->flags = h->flags & P_TYPE;
34046134Smao 
34150989Sbostic 	/* Fix up the previous pointer of the page after the split page. */
34250989Sbostic 	if (h->nextpg != P_INVALID) {
34350989Sbostic 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
34450989Sbostic 			free(l);
34550989Sbostic 			/* XXX mpool_free(t->bt_mp, r->pgno); */
34650989Sbostic 			return (NULL);
34746134Smao 		}
34850989Sbostic 		tp->prevpg = r->pgno;
34950989Sbostic 		mpool_put(t->bt_mp, tp, 0);
35050989Sbostic 	}
35146134Smao 
35250989Sbostic 	/*
35350989Sbostic 	 * Split right.  The key/data pairs aren't sorted in the btree page so
35450989Sbostic 	 * it's simpler to copy the data from the split page onto two new pages
35550989Sbostic 	 * instead of copying half the data to the right page and compacting
35650989Sbostic 	 * the left page in place.  Since the left page can't change, we have
35750989Sbostic 	 * to swap the original and the allocated left page after the split.
35850989Sbostic 	 */
35950989Sbostic 	tp = bt_psplit(t, h, l, r, skip);
36046134Smao 
36150989Sbostic 	/* Move the new left page onto the old left page. */
36250989Sbostic 	bcopy(l, h, t->bt_psize);
36350989Sbostic 	if (tp == l)
36450989Sbostic 		tp = h;
36550989Sbostic 	free(l);
36650989Sbostic 
36750989Sbostic 	*lp = h;
36850989Sbostic 	*rp = r;
36950989Sbostic 	return (tp);
37046134Smao }
37146134Smao 
37246134Smao /*
37351106Sbostic  * BT_ROOT -- Split the root page of a btree.
37446134Smao  *
37550989Sbostic  * Parameters:
37650989Sbostic  *	t:	tree
37750989Sbostic  *	h:	root page
37850989Sbostic  *	lp:	pointer to left page pointer
37950989Sbostic  *	rp:	pointer to right page pointer
38050989Sbostic  *	skip:	pointer to index to leave open
38146134Smao  *
38250989Sbostic  * Returns:
38350989Sbostic  *	Pointer to page in which to insert or NULL on error.
38446134Smao  */
38550989Sbostic static PAGE *
38650989Sbostic bt_root(t, h, lp, rp, skip)
38750989Sbostic 	BTREE *t;
38850989Sbostic 	PAGE *h, **lp, **rp;
38950989Sbostic 	int *skip;
39046134Smao {
39150989Sbostic 	PAGE *l, *r, *tp;
39250989Sbostic 	pgno_t lnpg, rnpg;
39346134Smao 
39450989Sbostic #ifdef STATISTICS
39550989Sbostic 	++bt_split;
39650989Sbostic 	++bt_rootsplit;
39750989Sbostic #endif
39850989Sbostic 	/* Put the new left and right pages for the split into place. */
39950989Sbostic 	if ((l = mpool_new(t->bt_mp, &lnpg)) == NULL ||
40050989Sbostic 	    (r = mpool_new(t->bt_mp, &rnpg)) == NULL)
40150989Sbostic 		return (NULL);
40250989Sbostic 	l->pgno = lnpg;
40350989Sbostic 	r->pgno = rnpg;
40450989Sbostic 	l->nextpg = r->pgno;
40550989Sbostic 	r->prevpg = l->pgno;
40650989Sbostic 	l->prevpg = r->nextpg = P_INVALID;
40750989Sbostic 	l->lower = r->lower = BTDATAOFF;
40850989Sbostic 	l->upper = r->upper = t->bt_psize;
40950989Sbostic 	l->flags = r->flags = h->flags & P_TYPE;
41050989Sbostic 
41150989Sbostic 	/* Split the root page. */
41250989Sbostic 	tp = bt_psplit(t, h, l, r, skip);
41350989Sbostic 
41450989Sbostic 	/* Make the root page look right. */
41550989Sbostic 	if ((ISSET(t, BTF_RECNO) ?
41650989Sbostic 	    bt_rroot(t, h, l, r) : bt_broot(t, h, l, r)) == RET_ERROR)
41750989Sbostic 		return (NULL);
41850989Sbostic 
41950989Sbostic 	*lp = l;
42050989Sbostic 	*rp = r;
42150989Sbostic 	return (tp);
42246134Smao }
42346134Smao 
42446134Smao /*
42550989Sbostic  * BT_RROOT -- Fix up the recno root page after the split.
42646134Smao  *
42750989Sbostic  * Parameters:
42850989Sbostic  *	t:	tree
42950989Sbostic  *	h:	root page
43046134Smao  *
43150989Sbostic  * Returns:
43250989Sbostic  *	RET_ERROR, RET_SUCCESS
43346134Smao  */
43450989Sbostic static int
43550989Sbostic bt_rroot(t, h, l, r)
43650989Sbostic 	BTREE *t;
43750989Sbostic 	PAGE *h, *l, *r;
43846134Smao {
43950989Sbostic 	char *dest;
44046134Smao 
44150989Sbostic 	/* Insert the left and right keys, set the header information. */
44250989Sbostic 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
44350989Sbostic 	dest = (char *)h + h->upper;
44450989Sbostic 	WR_RINTERNAL(dest,
44550989Sbostic 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
44646134Smao 
44750989Sbostic 	h->linp[1] = h->upper -= NRINTERNAL;
44850989Sbostic 	dest = (char *)h + h->upper;
44950989Sbostic 	WR_RINTERNAL(dest,
45050989Sbostic 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
45146134Smao 
45250989Sbostic 	h->lower = BTDATAOFF + 2 * sizeof(index_t);
45346134Smao 
45450989Sbostic 	/* Unpin the root page, set to recno internal page. */
45550989Sbostic 	h->flags &= ~P_TYPE;
45650989Sbostic 	h->flags |= P_RINTERNAL;
45750989Sbostic 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
45846134Smao 
45950989Sbostic 	return (RET_SUCCESS);
46050989Sbostic }
46146134Smao 
46250989Sbostic /*
46350989Sbostic  * BT_BROOT -- Fix up the btree root page after the split.
46450989Sbostic  *
46550989Sbostic  * Parameters:
46650989Sbostic  *	t:	tree
46750989Sbostic  *	h:	root page
46850989Sbostic  *
46950989Sbostic  * Returns:
47050989Sbostic  *	RET_ERROR, RET_SUCCESS
47150989Sbostic  */
47250989Sbostic static int
47350989Sbostic bt_broot(t, h, l, r)
47450989Sbostic 	BTREE *t;
47550989Sbostic 	PAGE *h, *l, *r;
47650989Sbostic {
47750989Sbostic 	BINTERNAL *bi;
47850989Sbostic 	BLEAF *bl;
47950989Sbostic 	size_t nbytes;
48050989Sbostic 	char *dest;
48146134Smao 
48246134Smao 	/*
48350989Sbostic 	 * If the root page was a leaf page, change it into an internal page.
48450989Sbostic 	 * We copy the key we split on (but not the key's data, in the case of
48556401Sbostic 	 * a leaf page) to the new root page.
48650989Sbostic 	 *
48750989Sbostic 	 * The btree comparison code guarantees that the left-most key on any
48850989Sbostic 	 * level of the tree is never used, so it doesn't need to be filled
48950989Sbostic 	 * in.  (This is not just convenience -- if the insert index is 0, we
49050989Sbostic 	 * don't *have* a key to fill in.)  The right key is available because
49150989Sbostic 	 * the split code guarantees not to split on the skipped index.
49246134Smao 	 */
49356401Sbostic 	nbytes = NBINTERNAL(0);
49450989Sbostic 	h->linp[0] = h->upper = t->bt_psize - nbytes;
49550989Sbostic 	dest = (char *)h + h->upper;
49650989Sbostic 	WR_BINTERNAL(dest, 0, l->pgno, 0);
49746134Smao 
49850989Sbostic 	switch(h->flags & P_TYPE) {
49950989Sbostic 	case P_BLEAF:
50050989Sbostic 		bl = GETBLEAF(r, 0);
50150989Sbostic 		nbytes = NBINTERNAL(bl->ksize);
50250989Sbostic 		h->linp[1] = h->upper -= nbytes;
50350989Sbostic 		dest = (char *)h + h->upper;
50450989Sbostic 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
50550989Sbostic 		bcopy(bl->bytes, dest, bl->ksize);
50650989Sbostic 
50756401Sbostic 		/*
50856401Sbostic 		 * If the key is on an overflow page, mark the overflow chain
50956401Sbostic 		 * so it isn't deleted when the leaf copy of the key is deleted.
51056401Sbostic 		 */
51150989Sbostic 		if (bl->flags & P_BIGKEY &&
51250989Sbostic 		    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
51350989Sbostic 			return (RET_ERROR);
51450989Sbostic 		break;
51550989Sbostic 	case P_BINTERNAL:
51650989Sbostic 		bi = GETBINTERNAL(r, 0);
51750989Sbostic 		nbytes = NBINTERNAL(bi->ksize);
51850989Sbostic 		h->linp[1] = h->upper -= nbytes;
51950989Sbostic 		dest = (char *)h + h->upper;
52050989Sbostic 		bcopy(bi, dest, nbytes);
52150989Sbostic 		((BINTERNAL *)dest)->pgno = r->pgno;
52250989Sbostic 		break;
52346134Smao 	}
52450989Sbostic 	h->lower = BTDATAOFF + 2 * sizeof(index_t);
52546134Smao 
52650989Sbostic 	/* Unpin the root page, set to btree internal page. */
52750989Sbostic 	h->flags &= ~P_TYPE;
52850989Sbostic 	h->flags |= P_BINTERNAL;
52950989Sbostic 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
53046134Smao 
53146134Smao 	return (RET_SUCCESS);
53246134Smao }
53346134Smao 
53446134Smao /*
53550989Sbostic  * BT_PSPLIT -- Do the real work of splitting the page.
53646134Smao  *
53750989Sbostic  * Parameters:
53850989Sbostic  *	t:	tree
53950989Sbostic  *	h:	page to be split
54050989Sbostic  *	l:	page to put lower half of data
54150989Sbostic  *	r:	page to put upper half of data
54250989Sbostic  *	skip:	pointer to index to leave open
54346134Smao  *
54450989Sbostic  * Returns:
54550989Sbostic  *	Pointer to page in which to insert.
54646134Smao  */
54750989Sbostic static PAGE *
54851106Sbostic bt_psplit(t, h, l, r, pskip)
54950989Sbostic 	BTREE *t;
55050989Sbostic 	PAGE *h, *l, *r;
55151106Sbostic 	int *pskip;
55246134Smao {
55350989Sbostic 	BINTERNAL *bi;
55450989Sbostic 	BLEAF *bl;
55550989Sbostic 	RLEAF *rl;
55650989Sbostic 	EPGNO *c;
55750989Sbostic 	PAGE *rval;
55851106Sbostic 	index_t half, skip;
55950989Sbostic 	size_t nbytes;
56050989Sbostic 	void *src;
56150989Sbostic 	int bigkeycnt, isbigkey, nxt, off, top;
56246134Smao 
56346134Smao 	/*
56450989Sbostic 	 * Split the data to the left and right pages. Leave the skip index
56550989Sbostic 	 * open and guarantee that the split doesn't happen on that index (the
56650989Sbostic 	 * right key must be available for the parent page).  Additionally,
56750989Sbostic 	 * make some effort not to split on an overflow key.  This makes it
56850989Sbostic 	 * faster to process internal pages and can save space since overflow
56950989Sbostic 	 * keys used by internal pages are never deleted.
57046134Smao 	 */
57150989Sbostic 	bigkeycnt = 0;
57251106Sbostic 	skip = *pskip;
57350989Sbostic 	half = (t->bt_psize - BTDATAOFF) / 2;
57450989Sbostic 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
57551106Sbostic 		if (skip == off)
57650989Sbostic 			continue;
57750989Sbostic 		switch (h->flags & P_TYPE) {
57850989Sbostic 		case P_BINTERNAL:
57950989Sbostic 			src = bi = GETBINTERNAL(h, nxt);
58050989Sbostic 			nbytes = NBINTERNAL(bi->ksize);
58150989Sbostic 			isbigkey = bi->flags & P_BIGKEY;
58250989Sbostic 			break;
58350989Sbostic 		case P_BLEAF:
58450989Sbostic 			src = bl = GETBLEAF(h, nxt);
58550989Sbostic 			nbytes = NBLEAF(bl);
58650989Sbostic 			isbigkey = bl->flags & P_BIGKEY;
58750989Sbostic 			break;
58850989Sbostic 		case P_RINTERNAL:
58950989Sbostic 			src = GETRINTERNAL(h, nxt);
59050989Sbostic 			nbytes = NRINTERNAL;
59150989Sbostic 			isbigkey = 0;
59250989Sbostic 			break;
59350989Sbostic 		case P_RLEAF:
59450989Sbostic 			src = rl = GETRLEAF(h, nxt);
59550989Sbostic 			nbytes = NRLEAF(rl);
59650989Sbostic 			isbigkey = 0;
59750989Sbostic 			break;
59850989Sbostic 		}
59950989Sbostic 		++nxt;
60050989Sbostic 		l->linp[off] = l->upper -= nbytes;
60150989Sbostic 		bcopy(src, (char *)l + l->upper, nbytes);
60246134Smao 
60350989Sbostic 		/* There's no empirical justification for the '3'. */
604*56402Sbostic 		if (half < nbytes) {
605*56402Sbostic 			if (skip != off + 1)
606*56402Sbostic 				if (!isbigkey || bigkeycnt == 3)
607*56402Sbostic 					break;
608*56402Sbostic 				else
609*56402Sbostic 					++bigkeycnt;
610*56402Sbostic 		} else
61150989Sbostic 			half -= nbytes;
61250989Sbostic 	}
61350989Sbostic 	l->lower += (off + 1) * sizeof(index_t);
61446134Smao 
61550989Sbostic 	/*
61651106Sbostic 	 * If splitting the page that the cursor was on, the cursor has to be
61751106Sbostic 	 * adjusted to point to the same record as before the split.  If the
61851106Sbostic 	 * skipped slot and the cursor are both on the left page and the cursor
61951106Sbostic 	 * is on or past the skipped slot, the cursor is incremented by one.
62051106Sbostic 	 * If the skipped slot and the cursor are both on the right page and
62151106Sbostic 	 * the cursor is on or past the skipped slot, the cursor is incremented
62251106Sbostic 	 * by one.  If the skipped slot and the cursor aren't on the same page,
62351106Sbostic 	 * the cursor isn't changed.  Regardless of the relationship of the
62451106Sbostic 	 * skipped slot and the cursor, if the cursor is on the right page it
62551106Sbostic 	 * is decremented by the number of records split to the left page.
62651106Sbostic 	 *
62751106Sbostic 	 * Don't bother checking for the BTF_SEQINIT flag, the page number will
62851106Sbostic 	 * be P_INVALID.
62950989Sbostic 	 */
63050989Sbostic 	c = &t->bt_bcursor;
63150989Sbostic 	if (c->pgno == h->pgno)
63251106Sbostic 		if (c->index < off) {			/* left page */
63350989Sbostic 			c->pgno = l->pgno;
63451106Sbostic 			if (c->index >= skip)
63551106Sbostic 				++c->index;
63651106Sbostic 		} else {				/* right page */
63750989Sbostic 			c->pgno = r->pgno;
63851106Sbostic 			if (c->index >= skip && skip > off)
63951106Sbostic 				++c->index;
64050989Sbostic 			c->index -= off;
64146134Smao 		}
64246134Smao 
64350989Sbostic 	/*
64450989Sbostic 	 * Decide which page to return, and adjust the skip index if the
64550989Sbostic 	 * to-be-inserted-upon page has changed.
64650989Sbostic 	 */
64751106Sbostic 	if (skip > off) {
64850989Sbostic 		rval = r;
64951106Sbostic 		*pskip -= off + 1;
65050989Sbostic 	} else
65150989Sbostic 		rval = l;
65246134Smao 
65350989Sbostic 	for (off = 0; nxt < top; ++off) {
65451106Sbostic 		if (skip == nxt) {
65551106Sbostic 			skip = 0;
65650989Sbostic 			continue;
65746134Smao 		}
65850989Sbostic 		switch (h->flags & P_TYPE) {
65950989Sbostic 		case P_BINTERNAL:
66050989Sbostic 			src = bi = GETBINTERNAL(h, nxt);
66150989Sbostic 			nbytes = NBINTERNAL(bi->ksize);
66250989Sbostic 			break;
66350989Sbostic 		case P_BLEAF:
66450989Sbostic 			src = bl = GETBLEAF(h, nxt);
66550989Sbostic 			nbytes = NBLEAF(bl);
66650989Sbostic 			break;
66750989Sbostic 		case P_RINTERNAL:
66850989Sbostic 			src = GETRINTERNAL(h, nxt);
66950989Sbostic 			nbytes = NRINTERNAL;
67050989Sbostic 			break;
67150989Sbostic 		case P_RLEAF:
67250989Sbostic 			src = rl = GETRLEAF(h, nxt);
67350989Sbostic 			nbytes = NRLEAF(rl);
67450989Sbostic 			break;
67550989Sbostic 		}
67650989Sbostic 		++nxt;
67750989Sbostic 		r->linp[off] = r->upper -= nbytes;
67850989Sbostic 		bcopy(src, (char *)r + r->upper, nbytes);
67950989Sbostic 	}
68050989Sbostic 	r->lower += off * sizeof(index_t);
68146134Smao 
68250989Sbostic 	/* If the key is being appended to the page, adjust the index. */
68351106Sbostic 	if (skip == top)
68450989Sbostic 		r->lower += sizeof(index_t);
68546134Smao 
68650989Sbostic 	return (rval);
68750989Sbostic }
68846134Smao 
68950989Sbostic /*
69050989Sbostic  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
69150989Sbostic  *
69250989Sbostic  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
69350989Sbostic  * record that references them gets deleted.  Chains pointed to by internal
69450989Sbostic  * pages never get deleted.  This routine marks a chain as pointed to by an
69550989Sbostic  * internal page.
69650989Sbostic  *
69750989Sbostic  * Parameters:
69850989Sbostic  *	t:	tree
69950989Sbostic  *	pg:	page number of first page in the chain.
70050989Sbostic  *
70150989Sbostic  * Returns:
70250989Sbostic  *	RET_SUCCESS, RET_ERROR.
70350989Sbostic  */
70450989Sbostic static int
70550989Sbostic bt_preserve(t, pg)
70650989Sbostic 	BTREE *t;
70750989Sbostic 	pgno_t pg;
70850989Sbostic {
70950989Sbostic 	PAGE *h;
71046134Smao 
71150989Sbostic 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
71250989Sbostic 		return (RET_ERROR);
71350989Sbostic 	h->flags |= P_PRESERVE;
71450989Sbostic 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
71546134Smao 	return (RET_SUCCESS);
71646134Smao }
71750989Sbostic 
71850989Sbostic /*
71950989Sbostic  * REC_TOTAL -- Return the number of recno entries below a page.
72050989Sbostic  *
72150989Sbostic  * Parameters:
72250989Sbostic  *	h:	page
72350989Sbostic  *
72450989Sbostic  * Returns:
72550989Sbostic  *	The number of recno entries below a page.
72650989Sbostic  *
72750989Sbostic  * XXX
72850989Sbostic  * These values could be set by the bt_psplit routine.  The problem is that the
72950989Sbostic  * entry has to be popped off of the stack etc. or the values have to be passed
73050989Sbostic  * all the way back to bt_split/bt_rroot and it's not very clean.
73150989Sbostic  */
73250989Sbostic static recno_t
73350989Sbostic rec_total(h)
73450989Sbostic 	PAGE *h;
73550989Sbostic {
73650989Sbostic 	recno_t recs;
73750989Sbostic 	index_t nxt, top;
73850989Sbostic 
73950989Sbostic 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
74050989Sbostic 		recs += GETRINTERNAL(h, nxt)->nrecs;
74150989Sbostic 	return (recs);
74250989Sbostic }
743