Home | History | Annotate | Line # | Download | only in btree
bt_split.c revision 1.9
      1 /*	$NetBSD: bt_split.c,v 1.9 1997/10/10 21:08:55 is Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1990, 1993, 1994
      5  *	The Regents of the University of California.  All rights reserved.
      6  *
      7  * This code is derived from software contributed to Berkeley by
      8  * Mike Olson.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  * 3. All advertising materials mentioning features or use of this software
     19  *    must display the following acknowledgement:
     20  *	This product includes software developed by the University of
     21  *	California, Berkeley and its contributors.
     22  * 4. Neither the name of the University nor the names of its contributors
     23  *    may be used to endorse or promote products derived from this software
     24  *    without specific prior written permission.
     25  *
     26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     36  * SUCH DAMAGE.
     37  */
     38 
     39 #include <sys/cdefs.h>
     40 #if defined(LIBC_SCCS) && !defined(lint)
     41 #if 0
     42 static char sccsid[] = "@(#)bt_split.c	8.9 (Berkeley) 7/26/94";
     43 #else
     44 __RCSID("$NetBSD: bt_split.c,v 1.9 1997/10/10 21:08:55 is Exp $");
     45 #endif
     46 #endif /* LIBC_SCCS and not lint */
     47 
     48 #include "namespace.h"
     49 #include <sys/types.h>
     50 
     51 #include <limits.h>
     52 #include <stdio.h>
     53 #include <stdlib.h>
     54 #include <string.h>
     55 
     56 #include <db.h>
     57 #include "btree.h"
     58 
     59 static int	 bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *));
     60 static PAGE	*bt_page
     61 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
     62 static int	 bt_preserve __P((BTREE *, pgno_t));
     63 static PAGE	*bt_psplit
     64 		    __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t));
     65 static PAGE	*bt_root
     66 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
     67 static int	 bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
     68 static recno_t	 rec_total __P((PAGE *));
     69 
     70 #ifdef STATISTICS
     71 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
     72 #endif
     73 
     74 /*
     75  * __BT_SPLIT -- Split the tree.
     76  *
     77  * Parameters:
     78  *	t:	tree
     79  *	sp:	page to split
     80  *	key:	key to insert
     81  *	data:	data to insert
     82  *	flags:	BIGKEY/BIGDATA flags
     83  *	ilen:	insert length
     84  *	skip:	index to leave open
     85  *
     86  * Returns:
     87  *	RET_ERROR, RET_SUCCESS
     88  */
     89 int
     90 __bt_split(t, sp, key, data, flags, ilen, argskip)
     91 	BTREE *t;
     92 	PAGE *sp;
     93 	const DBT *key, *data;
     94 	int flags;
     95 	size_t ilen;
     96 	u_int32_t argskip;
     97 {
     98 	BINTERNAL *bi = NULL;	/* pacify gcc */
     99 	BLEAF *bl = NULL, *tbl;	/* pacify gcc */
    100 	DBT a, b;
    101 	EPGNO *parent;
    102 	PAGE *h, *l, *r, *lchild, *rchild;
    103 	indx_t nxtindex;
    104 	u_int16_t skip;
    105 	u_int32_t n, nbytes, nksize = 0; /* pacify gcc */
    106 	int parentsplit;
    107 	char *dest;
    108 
    109 	/*
    110 	 * Split the page into two pages, l and r.  The split routines return
    111 	 * a pointer to the page into which the key should be inserted and with
    112 	 * skip set to the offset which should be used.  Additionally, l and r
    113 	 * are pinned.
    114 	 */
    115 	skip = argskip;
    116 	h = sp->pgno == P_ROOT ?
    117 	    bt_root(t, sp, &l, &r, &skip, ilen) :
    118 	    bt_page(t, sp, &l, &r, &skip, ilen);
    119 	if (h == NULL)
    120 		return (RET_ERROR);
    121 
    122 	/*
    123 	 * Insert the new key/data pair into the leaf page.  (Key inserts
    124 	 * always cause a leaf page to split first.)
    125 	 */
    126 	h->linp[skip] = h->upper -= ilen;
    127 	dest = (char *)h + h->upper;
    128 	if (F_ISSET(t, R_RECNO))
    129 		WR_RLEAF(dest, data, flags)
    130 	else
    131 		WR_BLEAF(dest, key, data, flags)
    132 
    133 	/* If the root page was split, make it look right. */
    134 	if (sp->pgno == P_ROOT &&
    135 	    (F_ISSET(t, R_RECNO) ?
    136 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    137 		goto err2;
    138 
    139 	/*
    140 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
    141 	 * were traversed when we searched for the page that split.  Each stack
    142 	 * entry is a page number and a page index offset.  The offset is for
    143 	 * the page traversed on the search.  We've just split a page, so we
    144 	 * have to insert a new key into the parent page.
    145 	 *
    146 	 * If the insert into the parent page causes it to split, may have to
    147 	 * continue splitting all the way up the tree.  We stop if the root
    148 	 * splits or the page inserted into didn't have to split to hold the
    149 	 * new key.  Some algorithms replace the key for the old page as well
    150 	 * as the new page.  We don't, as there's no reason to believe that the
    151 	 * first key on the old page is any better than the key we have, and,
    152 	 * in the case of a key being placed at index 0 causing the split, the
    153 	 * key is unavailable.
    154 	 *
    155 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
    156 	 * and right pages pinned while working on the parent.   The 5 are the
    157 	 * two children, left parent and right parent (when the parent splits)
    158 	 * and the root page or the overflow key page when calling bt_preserve.
    159 	 * This code must make sure that all pins are released other than the
    160 	 * root page or overflow page which is unlocked elsewhere.
    161 	 */
    162 	while ((parent = BT_POP(t)) != NULL) {
    163 		lchild = l;
    164 		rchild = r;
    165 
    166 		/* Get the parent page. */
    167 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
    168 			goto err2;
    169 
    170 	 	/*
    171 		 * The new key goes ONE AFTER the index, because the split
    172 		 * was to the right.
    173 		 */
    174 		skip = parent->index + 1;
    175 
    176 		/*
    177 		 * Calculate the space needed on the parent page.
    178 		 *
    179 		 * Prefix trees: space hack when inserting into BINTERNAL
    180 		 * pages.  Retain only what's needed to distinguish between
    181 		 * the new entry and the LAST entry on the page to its left.
    182 		 * If the keys compare equal, retain the entire key.  Note,
    183 		 * we don't touch overflow keys, and the entire key must be
    184 		 * retained for the next-to-left most key on the leftmost
    185 		 * page of each level, or the search will fail.  Applicable
    186 		 * ONLY to internal pages that have leaf pages as children.
    187 		 * Further reduction of the key between pairs of internal
    188 		 * pages loses too much information.
    189 		 */
    190 		switch (rchild->flags & P_TYPE) {
    191 		case P_BINTERNAL:
    192 			bi = GETBINTERNAL(rchild, 0);
    193 			nbytes = NBINTERNAL(bi->ksize);
    194 			break;
    195 		case P_BLEAF:
    196 			bl = GETBLEAF(rchild, 0);
    197 			nbytes = NBINTERNAL(bl->ksize);
    198 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
    199 			    (h->prevpg != P_INVALID || skip > 1)) {
    200 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
    201 				a.size = tbl->ksize;
    202 				a.data = tbl->bytes;
    203 				b.size = bl->ksize;
    204 				b.data = bl->bytes;
    205 				nksize = t->bt_pfx(&a, &b);
    206 				n = NBINTERNAL(nksize);
    207 				if (n < nbytes) {
    208 #ifdef STATISTICS
    209 					bt_pfxsaved += nbytes - n;
    210 #endif
    211 					nbytes = n;
    212 				} else
    213 					nksize = 0;
    214 			} else
    215 				nksize = 0;
    216 			break;
    217 		case P_RINTERNAL:
    218 		case P_RLEAF:
    219 			nbytes = NRINTERNAL;
    220 			break;
    221 		default:
    222 			abort();
    223 		}
    224 
    225 		/* Split the parent page if necessary or shift the indices. */
    226 		if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
    227 			sp = h;
    228 			h = h->pgno == P_ROOT ?
    229 			    bt_root(t, h, &l, &r, &skip, nbytes) :
    230 			    bt_page(t, h, &l, &r, &skip, nbytes);
    231 			if (h == NULL)
    232 				goto err1;
    233 			parentsplit = 1;
    234 		} else {
    235 			if (skip < (nxtindex = NEXTINDEX(h)))
    236 				memmove(h->linp + skip + 1, h->linp + skip,
    237 				    (nxtindex - skip) * sizeof(indx_t));
    238 			h->lower += sizeof(indx_t);
    239 			parentsplit = 0;
    240 		}
    241 
    242 		/* Insert the key into the parent page. */
    243 		switch (rchild->flags & P_TYPE) {
    244 		case P_BINTERNAL:
    245 			h->linp[skip] = h->upper -= nbytes;
    246 			dest = (char *)h + h->linp[skip];
    247 			memmove(dest, bi, nbytes);
    248 			((BINTERNAL *)dest)->pgno = rchild->pgno;
    249 			break;
    250 		case P_BLEAF:
    251 			h->linp[skip] = h->upper -= nbytes;
    252 			dest = (char *)h + h->linp[skip];
    253 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
    254 			    rchild->pgno, bl->flags & P_BIGKEY);
    255 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
    256 			if (bl->flags & P_BIGKEY &&
    257 			    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
    258 				goto err1;
    259 			break;
    260 		case P_RINTERNAL:
    261 			/*
    262 			 * Update the left page count.  If split
    263 			 * added at index 0, fix the correct page.
    264 			 */
    265 			if (skip > 0)
    266 				dest = (char *)h + h->linp[skip - 1];
    267 			else
    268 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
    269 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
    270 			((RINTERNAL *)dest)->pgno = lchild->pgno;
    271 
    272 			/* Update the right page count. */
    273 			h->linp[skip] = h->upper -= nbytes;
    274 			dest = (char *)h + h->linp[skip];
    275 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
    276 			((RINTERNAL *)dest)->pgno = rchild->pgno;
    277 			break;
    278 		case P_RLEAF:
    279 			/*
    280 			 * Update the left page count.  If split
    281 			 * added at index 0, fix the correct page.
    282 			 */
    283 			if (skip > 0)
    284 				dest = (char *)h + h->linp[skip - 1];
    285 			else
    286 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
    287 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
    288 			((RINTERNAL *)dest)->pgno = lchild->pgno;
    289 
    290 			/* Update the right page count. */
    291 			h->linp[skip] = h->upper -= nbytes;
    292 			dest = (char *)h + h->linp[skip];
    293 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
    294 			((RINTERNAL *)dest)->pgno = rchild->pgno;
    295 			break;
    296 		default:
    297 			abort();
    298 		}
    299 
    300 		/* Unpin the held pages. */
    301 		if (!parentsplit) {
    302 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    303 			break;
    304 		}
    305 
    306 		/* If the root page was split, make it look right. */
    307 		if (sp->pgno == P_ROOT &&
    308 		    (F_ISSET(t, R_RECNO) ?
    309 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    310 			goto err1;
    311 
    312 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    313 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    314 	}
    315 
    316 	/* Unpin the held pages. */
    317 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
    318 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
    319 
    320 	/* Clear any pages left on the stack. */
    321 	return (RET_SUCCESS);
    322 
    323 	/*
    324 	 * If something fails in the above loop we were already walking back
    325 	 * up the tree and the tree is now inconsistent.  Nothing much we can
    326 	 * do about it but release any memory we're holding.
    327 	 */
    328 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    329 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    330 
    331 err2:	mpool_put(t->bt_mp, l, 0);
    332 	mpool_put(t->bt_mp, r, 0);
    333 	__dbpanic(t->bt_dbp);
    334 	return (RET_ERROR);
    335 }
    336 
    337 /*
    338  * BT_PAGE -- Split a non-root page of a btree.
    339  *
    340  * Parameters:
    341  *	t:	tree
    342  *	h:	root page
    343  *	lp:	pointer to left page pointer
    344  *	rp:	pointer to right page pointer
    345  *	skip:	pointer to index to leave open
    346  *	ilen:	insert length
    347  *
    348  * Returns:
    349  *	Pointer to page in which to insert or NULL on error.
    350  */
    351 static PAGE *
    352 bt_page(t, h, lp, rp, skip, ilen)
    353 	BTREE *t;
    354 	PAGE *h, **lp, **rp;
    355 	indx_t *skip;
    356 	size_t ilen;
    357 {
    358 	PAGE *l, *r, *tp;
    359 	pgno_t npg;
    360 
    361 #ifdef STATISTICS
    362 	++bt_split;
    363 #endif
    364 	/* Put the new right page for the split into place. */
    365 	if ((r = __bt_new(t, &npg)) == NULL)
    366 		return (NULL);
    367 	r->pgno = npg;
    368 	r->lower = BTDATAOFF;
    369 	r->upper = t->bt_psize;
    370 	r->nextpg = h->nextpg;
    371 	r->prevpg = h->pgno;
    372 	r->flags = h->flags & P_TYPE;
    373 
    374 	/*
    375 	 * If we're splitting the last page on a level because we're appending
    376 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
    377 	 * sorted.  Adding an empty page on the side of the level is less work
    378 	 * and can push the fill factor much higher than normal.  If we're
    379 	 * wrong it's no big deal, we'll just do the split the right way next
    380 	 * time.  It may look like it's equally easy to do a similar hack for
    381 	 * reverse sorted data, that is, split the tree left, but it's not.
    382 	 * Don't even try.
    383 	 */
    384 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
    385 #ifdef STATISTICS
    386 		++bt_sortsplit;
    387 #endif
    388 		h->nextpg = r->pgno;
    389 		r->lower = BTDATAOFF + sizeof(indx_t);
    390 		*skip = 0;
    391 		*lp = h;
    392 		*rp = r;
    393 		return (r);
    394 	}
    395 
    396 	/* Put the new left page for the split into place. */
    397 	if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
    398 		mpool_put(t->bt_mp, r, 0);
    399 		return (NULL);
    400 	}
    401 #ifdef PURIFY
    402 	memset(l, 0xff, t->bt_psize);
    403 #endif
    404 	l->pgno = h->pgno;
    405 	l->nextpg = r->pgno;
    406 	l->prevpg = h->prevpg;
    407 	l->lower = BTDATAOFF;
    408 	l->upper = t->bt_psize;
    409 	l->flags = h->flags & P_TYPE;
    410 
    411 	/* Fix up the previous pointer of the page after the split page. */
    412 	if (h->nextpg != P_INVALID) {
    413 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
    414 			free(l);
    415 			/* XXX mpool_free(t->bt_mp, r->pgno); */
    416 			return (NULL);
    417 		}
    418 		tp->prevpg = r->pgno;
    419 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
    420 	}
    421 
    422 	/*
    423 	 * Split right.  The key/data pairs aren't sorted in the btree page so
    424 	 * it's simpler to copy the data from the split page onto two new pages
    425 	 * instead of copying half the data to the right page and compacting
    426 	 * the left page in place.  Since the left page can't change, we have
    427 	 * to swap the original and the allocated left page after the split.
    428 	 */
    429 	tp = bt_psplit(t, h, l, r, skip, ilen);
    430 
    431 	/* Move the new left page onto the old left page. */
    432 	memmove(h, l, t->bt_psize);
    433 	if (tp == l)
    434 		tp = h;
    435 	free(l);
    436 
    437 	*lp = h;
    438 	*rp = r;
    439 	return (tp);
    440 }
    441 
    442 /*
    443  * BT_ROOT -- Split the root page of a btree.
    444  *
    445  * Parameters:
    446  *	t:	tree
    447  *	h:	root page
    448  *	lp:	pointer to left page pointer
    449  *	rp:	pointer to right page pointer
    450  *	skip:	pointer to index to leave open
    451  *	ilen:	insert length
    452  *
    453  * Returns:
    454  *	Pointer to page in which to insert or NULL on error.
    455  */
    456 static PAGE *
    457 bt_root(t, h, lp, rp, skip, ilen)
    458 	BTREE *t;
    459 	PAGE *h, **lp, **rp;
    460 	indx_t *skip;
    461 	size_t ilen;
    462 {
    463 	PAGE *l, *r, *tp;
    464 	pgno_t lnpg, rnpg;
    465 
    466 #ifdef STATISTICS
    467 	++bt_split;
    468 	++bt_rootsplit;
    469 #endif
    470 	/* Put the new left and right pages for the split into place. */
    471 	if ((l = __bt_new(t, &lnpg)) == NULL ||
    472 	    (r = __bt_new(t, &rnpg)) == NULL)
    473 		return (NULL);
    474 	l->pgno = lnpg;
    475 	r->pgno = rnpg;
    476 	l->nextpg = r->pgno;
    477 	r->prevpg = l->pgno;
    478 	l->prevpg = r->nextpg = P_INVALID;
    479 	l->lower = r->lower = BTDATAOFF;
    480 	l->upper = r->upper = t->bt_psize;
    481 	l->flags = r->flags = h->flags & P_TYPE;
    482 
    483 	/* Split the root page. */
    484 	tp = bt_psplit(t, h, l, r, skip, ilen);
    485 
    486 	*lp = l;
    487 	*rp = r;
    488 	return (tp);
    489 }
    490 
    491 /*
    492  * BT_RROOT -- Fix up the recno root page after it has been split.
    493  *
    494  * Parameters:
    495  *	t:	tree
    496  *	h:	root page
    497  *	l:	left page
    498  *	r:	right page
    499  *
    500  * Returns:
    501  *	RET_ERROR, RET_SUCCESS
    502  */
    503 static int
    504 bt_rroot(t, h, l, r)
    505 	BTREE *t;
    506 	PAGE *h, *l, *r;
    507 {
    508 	char *dest;
    509 
    510 	/* Insert the left and right keys, set the header information. */
    511 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
    512 	dest = (char *)h + h->upper;
    513 	WR_RINTERNAL(dest,
    514 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
    515 
    516 	h->linp[1] = h->upper -= NRINTERNAL;
    517 	dest = (char *)h + h->upper;
    518 	WR_RINTERNAL(dest,
    519 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
    520 
    521 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    522 
    523 	/* Unpin the root page, set to recno internal page. */
    524 	h->flags &= ~P_TYPE;
    525 	h->flags |= P_RINTERNAL;
    526 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    527 
    528 	return (RET_SUCCESS);
    529 }
    530 
    531 /*
    532  * BT_BROOT -- Fix up the btree root page after it has been split.
    533  *
    534  * Parameters:
    535  *	t:	tree
    536  *	h:	root page
    537  *	l:	left page
    538  *	r:	right page
    539  *
    540  * Returns:
    541  *	RET_ERROR, RET_SUCCESS
    542  */
    543 static int
    544 bt_broot(t, h, l, r)
    545 	BTREE *t;
    546 	PAGE *h, *l, *r;
    547 {
    548 	BINTERNAL *bi = NULL;	/* pacify gcc */
    549 	BLEAF *bl;
    550 	u_int32_t nbytes;
    551 	char *dest;
    552 
    553 	/*
    554 	 * If the root page was a leaf page, change it into an internal page.
    555 	 * We copy the key we split on (but not the key's data, in the case of
    556 	 * a leaf page) to the new root page.
    557 	 *
    558 	 * The btree comparison code guarantees that the left-most key on any
    559 	 * level of the tree is never used, so it doesn't need to be filled in.
    560 	 */
    561 	nbytes = NBINTERNAL(0);
    562 	h->linp[0] = h->upper = t->bt_psize - nbytes;
    563 	dest = (char *)h + h->upper;
    564 	WR_BINTERNAL(dest, 0, l->pgno, 0);
    565 
    566 	switch (h->flags & P_TYPE) {
    567 	case P_BLEAF:
    568 		bl = GETBLEAF(r, 0);
    569 		nbytes = NBINTERNAL(bl->ksize);
    570 		h->linp[1] = h->upper -= nbytes;
    571 		dest = (char *)h + h->upper;
    572 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
    573 		memmove(dest, bl->bytes, bl->ksize);
    574 
    575 		/*
    576 		 * If the key is on an overflow page, mark the overflow chain
    577 		 * so it isn't deleted when the leaf copy of the key is deleted.
    578 		 */
    579 		if (bl->flags & P_BIGKEY &&
    580 		    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
    581 			return (RET_ERROR);
    582 		break;
    583 	case P_BINTERNAL:
    584 		bi = GETBINTERNAL(r, 0);
    585 		nbytes = NBINTERNAL(bi->ksize);
    586 		h->linp[1] = h->upper -= nbytes;
    587 		dest = (char *)h + h->upper;
    588 		memmove(dest, bi, nbytes);
    589 		((BINTERNAL *)dest)->pgno = r->pgno;
    590 		break;
    591 	default:
    592 		abort();
    593 	}
    594 
    595 	/* There are two keys on the page. */
    596 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    597 
    598 	/* Unpin the root page, set to btree internal page. */
    599 	h->flags &= ~P_TYPE;
    600 	h->flags |= P_BINTERNAL;
    601 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    602 
    603 	return (RET_SUCCESS);
    604 }
    605 
    606 /*
    607  * BT_PSPLIT -- Do the real work of splitting the page.
    608  *
    609  * Parameters:
    610  *	t:	tree
    611  *	h:	page to be split
    612  *	l:	page to put lower half of data
    613  *	r:	page to put upper half of data
    614  *	pskip:	pointer to index to leave open
    615  *	ilen:	insert length
    616  *
    617  * Returns:
    618  *	Pointer to page in which to insert.
    619  */
    620 static PAGE *
    621 bt_psplit(t, h, l, r, pskip, ilen)
    622 	BTREE *t;
    623 	PAGE *h, *l, *r;
    624 	indx_t *pskip;
    625 	size_t ilen;
    626 {
    627 	BINTERNAL *bi;
    628 	BLEAF *bl;
    629 	CURSOR *c;
    630 	RLEAF *rl;
    631 	PAGE *rval;
    632 	void *src = NULL;	/* pacify gcc */
    633 	indx_t full, half, nxt, off, skip, top, used;
    634 	u_int32_t nbytes;
    635 	int bigkeycnt, isbigkey;
    636 
    637 	/*
    638 	 * Split the data to the left and right pages.  Leave the skip index
    639 	 * open.  Additionally, make some effort not to split on an overflow
    640 	 * key.  This makes internal page processing faster and can save
    641 	 * space as overflow keys used by internal pages are never deleted.
    642 	 */
    643 	bigkeycnt = 0;
    644 	skip = *pskip;
    645 	full = t->bt_psize - BTDATAOFF;
    646 	half = full / 2;
    647 	used = 0;
    648 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
    649 		if (skip == off) {
    650 			nbytes = ilen;
    651 			isbigkey = 0;		/* XXX: not really known. */
    652 		} else
    653 			switch (h->flags & P_TYPE) {
    654 			case P_BINTERNAL:
    655 				src = bi = GETBINTERNAL(h, nxt);
    656 				nbytes = NBINTERNAL(bi->ksize);
    657 				isbigkey = bi->flags & P_BIGKEY;
    658 				break;
    659 			case P_BLEAF:
    660 				src = bl = GETBLEAF(h, nxt);
    661 				nbytes = NBLEAF(bl);
    662 				isbigkey = bl->flags & P_BIGKEY;
    663 				break;
    664 			case P_RINTERNAL:
    665 				src = GETRINTERNAL(h, nxt);
    666 				nbytes = NRINTERNAL;
    667 				isbigkey = 0;
    668 				break;
    669 			case P_RLEAF:
    670 				src = rl = GETRLEAF(h, nxt);
    671 				nbytes = NRLEAF(rl);
    672 				isbigkey = 0;
    673 				break;
    674 			default:
    675 				abort();
    676 			}
    677 
    678 		/*
    679 		 * If the key/data pairs are substantial fractions of the max
    680 		 * possible size for the page, it's possible to get situations
    681 		 * where we decide to try and copy too much onto the left page.
    682 		 * Make sure that doesn't happen.
    683 		 */
    684 		if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
    685 		    nxt == top - 1) {
    686 			--off;
    687 			break;
    688 		}
    689 
    690 		/* Copy the key/data pair, if not the skipped index. */
    691 		if (skip != off) {
    692 			++nxt;
    693 
    694 			l->linp[off] = l->upper -= nbytes;
    695 			memmove((char *)l + l->upper, src, nbytes);
    696 		}
    697 
    698 		used += nbytes + sizeof(indx_t);
    699 		if (used >= half) {
    700 			if (!isbigkey || bigkeycnt == 3)
    701 				break;
    702 			else
    703 				++bigkeycnt;
    704 		}
    705 	}
    706 
    707 	/*
    708 	 * Off is the last offset that's valid for the left page.
    709 	 * Nxt is the first offset to be placed on the right page.
    710 	 */
    711 	l->lower += (off + 1) * sizeof(indx_t);
    712 
    713 	/*
    714 	 * If splitting the page that the cursor was on, the cursor has to be
    715 	 * adjusted to point to the same record as before the split.  If the
    716 	 * cursor is at or past the skipped slot, the cursor is incremented by
    717 	 * one.  If the cursor is on the right page, it is decremented by the
    718 	 * number of records split to the left page.
    719 	 */
    720 	c = &t->bt_cursor;
    721 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
    722 		if (c->pg.index >= skip)
    723 			++c->pg.index;
    724 		if (c->pg.index < nxt)			/* Left page. */
    725 			c->pg.pgno = l->pgno;
    726 		else {					/* Right page. */
    727 			c->pg.pgno = r->pgno;
    728 			c->pg.index -= nxt;
    729 		}
    730 	}
    731 
    732 	/*
    733 	 * If the skipped index was on the left page, just return that page.
    734 	 * Otherwise, adjust the skip index to reflect the new position on
    735 	 * the right page.
    736 	 */
    737 	if (skip <= off) {
    738 		skip = 0;
    739 		rval = l;
    740 	} else {
    741 		rval = r;
    742 		*pskip -= nxt;
    743 	}
    744 
    745 	for (off = 0; nxt < top; ++off) {
    746 		if (skip == nxt) {
    747 			++off;
    748 			skip = 0;
    749 		}
    750 		switch (h->flags & P_TYPE) {
    751 		case P_BINTERNAL:
    752 			src = bi = GETBINTERNAL(h, nxt);
    753 			nbytes = NBINTERNAL(bi->ksize);
    754 			break;
    755 		case P_BLEAF:
    756 			src = bl = GETBLEAF(h, nxt);
    757 			nbytes = NBLEAF(bl);
    758 			break;
    759 		case P_RINTERNAL:
    760 			src = GETRINTERNAL(h, nxt);
    761 			nbytes = NRINTERNAL;
    762 			break;
    763 		case P_RLEAF:
    764 			src = rl = GETRLEAF(h, nxt);
    765 			nbytes = NRLEAF(rl);
    766 			break;
    767 		default:
    768 			abort();
    769 		}
    770 		++nxt;
    771 		r->linp[off] = r->upper -= nbytes;
    772 		memmove((char *)r + r->upper, src, nbytes);
    773 	}
    774 	r->lower += off * sizeof(indx_t);
    775 
    776 	/* If the key is being appended to the page, adjust the index. */
    777 	if (skip == top)
    778 		r->lower += sizeof(indx_t);
    779 
    780 	return (rval);
    781 }
    782 
    783 /*
    784  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
    785  *
    786  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
    787  * record that references them gets deleted.  Chains pointed to by internal
    788  * pages never get deleted.  This routine marks a chain as pointed to by an
    789  * internal page.
    790  *
    791  * Parameters:
    792  *	t:	tree
    793  *	pg:	page number of first page in the chain.
    794  *
    795  * Returns:
    796  *	RET_SUCCESS, RET_ERROR.
    797  */
    798 static int
    799 bt_preserve(t, pg)
    800 	BTREE *t;
    801 	pgno_t pg;
    802 {
    803 	PAGE *h;
    804 
    805 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
    806 		return (RET_ERROR);
    807 	h->flags |= P_PRESERVE;
    808 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    809 	return (RET_SUCCESS);
    810 }
    811 
    812 /*
    813  * REC_TOTAL -- Return the number of recno entries below a page.
    814  *
    815  * Parameters:
    816  *	h:	page
    817  *
    818  * Returns:
    819  *	The number of recno entries below a page.
    820  *
    821  * XXX
    822  * These values could be set by the bt_psplit routine.  The problem is that the
    823  * entry has to be popped off of the stack etc. or the values have to be passed
    824  * all the way back to bt_split/bt_rroot and it's not very clean.
    825  */
    826 static recno_t
    827 rec_total(h)
    828 	PAGE *h;
    829 {
    830 	recno_t recs;
    831 	indx_t nxt, top;
    832 
    833 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
    834 		recs += GETRINTERNAL(h, nxt)->nrecs;
    835 	return (recs);
    836 }
    837