Home | History | Annotate | Line # | Download | only in btree
bt_split.c revision 1.19
      1 /*	$NetBSD: bt_split.c,v 1.19 2009/04/22 18:44:06 christos 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. Neither the name of the University nor the names of its contributors
     19  *    may be used to endorse or promote products derived from this software
     20  *    without specific prior written permission.
     21  *
     22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     32  * SUCH DAMAGE.
     33  */
     34 
     35 #if HAVE_NBTOOL_CONFIG_H
     36 #include "nbtool_config.h"
     37 #endif
     38 
     39 #include <sys/cdefs.h>
     40 __RCSID("$NetBSD: bt_split.c,v 1.19 2009/04/22 18:44:06 christos Exp $");
     41 
     42 #include "namespace.h"
     43 #include <sys/types.h>
     44 
     45 #include <assert.h>
     46 #include <limits.h>
     47 #include <stdio.h>
     48 #include <stdlib.h>
     49 #include <string.h>
     50 
     51 #include <db.h>
     52 #include "btree.h"
     53 
     54 static int	 bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
     55 static PAGE	*bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
     56 static int	 bt_preserve(BTREE *, pgno_t);
     57 static PAGE	*bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
     58 static PAGE	*bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
     59 static int	 bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
     60 static recno_t	 rec_total(PAGE *);
     61 
     62 #ifdef STATISTICS
     63 unsigned long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
     64 #endif
     65 
     66 /*
     67  * __BT_SPLIT -- Split the tree.
     68  *
     69  * Parameters:
     70  *	t:	tree
     71  *	sp:	page to split
     72  *	key:	key to insert
     73  *	data:	data to insert
     74  *	flags:	BIGKEY/BIGDATA flags
     75  *	ilen:	insert length
     76  *	skip:	index to leave open
     77  *
     78  * Returns:
     79  *	RET_ERROR, RET_SUCCESS
     80  */
     81 int
     82 __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
     83     size_t ilen, uint32_t argskip)
     84 {
     85 	BINTERNAL *bi = NULL;	/* pacify gcc */
     86 	BLEAF *bl = NULL, *tbl;	/* pacify gcc */
     87 	DBT a, b;
     88 	EPGNO *parent;
     89 	PAGE *h, *l, *r, *lchild, *rchild;
     90 	indx_t nxtindex;
     91 	uint16_t skip;
     92 	uint32_t n, nbytes, nksize = 0; /* pacify gcc */
     93 	int parentsplit;
     94 	char *dest;
     95 
     96 	/*
     97 	 * Split the page into two pages, l and r.  The split routines return
     98 	 * a pointer to the page into which the key should be inserted and with
     99 	 * skip set to the offset which should be used.  Additionally, l and r
    100 	 * are pinned.
    101 	 */
    102 	skip = argskip;
    103 	h = sp->pgno == P_ROOT ?
    104 	    bt_root(t, sp, &l, &r, &skip, ilen) :
    105 	    bt_page(t, sp, &l, &r, &skip, ilen);
    106 	if (h == NULL)
    107 		return (RET_ERROR);
    108 
    109 	/*
    110 	 * Insert the new key/data pair into the leaf page.  (Key inserts
    111 	 * always cause a leaf page to split first.)
    112 	 */
    113 	_DBFIT(ilen, indx_t);
    114 	h->upper -= (indx_t)ilen;
    115 	h->linp[skip] = h->upper;
    116 	dest = (char *)(void *)h + h->upper;
    117 	if (F_ISSET(t, R_RECNO))
    118 		WR_RLEAF(dest, data, flags);
    119 	else
    120 		WR_BLEAF(dest, key, data, flags);
    121 
    122 	/* If the root page was split, make it look right. */
    123 	if (sp->pgno == P_ROOT &&
    124 	    (F_ISSET(t, R_RECNO) ?
    125 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    126 		goto err2;
    127 
    128 	/*
    129 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
    130 	 * were traversed when we searched for the page that split.  Each stack
    131 	 * entry is a page number and a page index offset.  The offset is for
    132 	 * the page traversed on the search.  We've just split a page, so we
    133 	 * have to insert a new key into the parent page.
    134 	 *
    135 	 * If the insert into the parent page causes it to split, may have to
    136 	 * continue splitting all the way up the tree.  We stop if the root
    137 	 * splits or the page inserted into didn't have to split to hold the
    138 	 * new key.  Some algorithms replace the key for the old page as well
    139 	 * as the new page.  We don't, as there's no reason to believe that the
    140 	 * first key on the old page is any better than the key we have, and,
    141 	 * in the case of a key being placed at index 0 causing the split, the
    142 	 * key is unavailable.
    143 	 *
    144 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
    145 	 * and right pages pinned while working on the parent.   The 5 are the
    146 	 * two children, left parent and right parent (when the parent splits)
    147 	 * and the root page or the overflow key page when calling bt_preserve.
    148 	 * This code must make sure that all pins are released other than the
    149 	 * root page or overflow page which is unlocked elsewhere.
    150 	 */
    151 	while ((parent = BT_POP(t)) != NULL) {
    152 		lchild = l;
    153 		rchild = r;
    154 
    155 		/* Get the parent page. */
    156 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
    157 			goto err2;
    158 
    159 	 	/*
    160 		 * The new key goes ONE AFTER the index, because the split
    161 		 * was to the right.
    162 		 */
    163 		skip = parent->index + 1;
    164 
    165 		/*
    166 		 * Calculate the space needed on the parent page.
    167 		 *
    168 		 * Prefix trees: space hack when inserting into BINTERNAL
    169 		 * pages.  Retain only what's needed to distinguish between
    170 		 * the new entry and the LAST entry on the page to its left.
    171 		 * If the keys compare equal, retain the entire key.  Note,
    172 		 * we don't touch overflow keys, and the entire key must be
    173 		 * retained for the next-to-left most key on the leftmost
    174 		 * page of each level, or the search will fail.  Applicable
    175 		 * ONLY to internal pages that have leaf pages as children.
    176 		 * Further reduction of the key between pairs of internal
    177 		 * pages loses too much information.
    178 		 */
    179 		switch (rchild->flags & P_TYPE) {
    180 		case P_BINTERNAL:
    181 			bi = GETBINTERNAL(rchild, 0);
    182 			nbytes = NBINTERNAL(bi->ksize);
    183 			break;
    184 		case P_BLEAF:
    185 			bl = GETBLEAF(rchild, 0);
    186 			nbytes = NBINTERNAL(bl->ksize);
    187 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
    188 			    (h->prevpg != P_INVALID || skip > 1)) {
    189 				size_t temp;
    190 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
    191 				a.size = tbl->ksize;
    192 				a.data = tbl->bytes;
    193 				b.size = bl->ksize;
    194 				b.data = bl->bytes;
    195 				temp = t->bt_pfx(&a, &b);
    196 				_DBFIT(temp, uint32_t);
    197 				nksize = (uint32_t)temp;
    198 				n = NBINTERNAL(nksize);
    199 				if (n < nbytes) {
    200 #ifdef STATISTICS
    201 					bt_pfxsaved += nbytes - n;
    202 #endif
    203 					nbytes = n;
    204 				} else
    205 					nksize = 0;
    206 			} else
    207 				nksize = 0;
    208 			break;
    209 		case P_RINTERNAL:
    210 		case P_RLEAF:
    211 			nbytes = NRINTERNAL;
    212 			break;
    213 		default:
    214 			abort();
    215 		}
    216 
    217 		/* Split the parent page if necessary or shift the indices. */
    218 		if ((uint32_t)h->upper - (uint32_t)h->lower < nbytes + sizeof(indx_t)) {
    219 			sp = h;
    220 			h = h->pgno == P_ROOT ?
    221 			    bt_root(t, h, &l, &r, &skip, nbytes) :
    222 			    bt_page(t, h, &l, &r, &skip, nbytes);
    223 			if (h == NULL)
    224 				goto err1;
    225 			parentsplit = 1;
    226 		} else {
    227 			if (skip < (nxtindex = NEXTINDEX(h)))
    228 				memmove(h->linp + skip + 1, h->linp + skip,
    229 				    (nxtindex - skip) * sizeof(indx_t));
    230 			h->lower += sizeof(indx_t);
    231 			parentsplit = 0;
    232 		}
    233 
    234 		/* Insert the key into the parent page. */
    235 		switch (rchild->flags & P_TYPE) {
    236 		case P_BINTERNAL:
    237 			h->linp[skip] = h->upper -= nbytes;
    238 			dest = (char *)(void *)h + h->linp[skip];
    239 			memmove(dest, bi, nbytes);
    240 			((BINTERNAL *)(void *)dest)->pgno = rchild->pgno;
    241 			break;
    242 		case P_BLEAF:
    243 			h->linp[skip] = h->upper -= nbytes;
    244 			dest = (char *)(void *)h + h->linp[skip];
    245 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
    246 			    rchild->pgno, bl->flags & P_BIGKEY);
    247 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
    248 			if (bl->flags & P_BIGKEY &&
    249 			    bt_preserve(t, *(pgno_t *)(void *)bl->bytes) ==
    250 			    RET_ERROR)
    251 				goto err1;
    252 			break;
    253 		case P_RINTERNAL:
    254 			/*
    255 			 * Update the left page count.  If split
    256 			 * added at index 0, fix the correct page.
    257 			 */
    258 			if (skip > 0)
    259 				dest = (char *)(void *)h + h->linp[skip - 1];
    260 			else
    261 				dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
    262 			((RINTERNAL *)(void *)dest)->nrecs = rec_total(lchild);
    263 			((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
    264 
    265 			/* Update the right page count. */
    266 			h->linp[skip] = h->upper -= nbytes;
    267 			dest = (char *)(void *)h + h->linp[skip];
    268 			((RINTERNAL *)(void *)dest)->nrecs = rec_total(rchild);
    269 			((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
    270 			break;
    271 		case P_RLEAF:
    272 			/*
    273 			 * Update the left page count.  If split
    274 			 * added at index 0, fix the correct page.
    275 			 */
    276 			if (skip > 0)
    277 				dest = (char *)(void *)h + h->linp[skip - 1];
    278 			else
    279 				dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
    280 			((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(lchild);
    281 			((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
    282 
    283 			/* Update the right page count. */
    284 			h->linp[skip] = h->upper -= nbytes;
    285 			dest = (char *)(void *)h + h->linp[skip];
    286 			((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(rchild);
    287 			((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
    288 			break;
    289 		default:
    290 			abort();
    291 		}
    292 
    293 		/* Unpin the held pages. */
    294 		if (!parentsplit) {
    295 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    296 			break;
    297 		}
    298 
    299 		/* If the root page was split, make it look right. */
    300 		if (sp->pgno == P_ROOT &&
    301 		    (F_ISSET(t, R_RECNO) ?
    302 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    303 			goto err1;
    304 
    305 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    306 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    307 	}
    308 
    309 	/* Unpin the held pages. */
    310 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
    311 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
    312 
    313 	/* Clear any pages left on the stack. */
    314 	return (RET_SUCCESS);
    315 
    316 	/*
    317 	 * If something fails in the above loop we were already walking back
    318 	 * up the tree and the tree is now inconsistent.  Nothing much we can
    319 	 * do about it but release any memory we're holding.
    320 	 */
    321 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    322 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    323 
    324 err2:	mpool_put(t->bt_mp, l, 0);
    325 	mpool_put(t->bt_mp, r, 0);
    326 	__dbpanic(t->bt_dbp);
    327 	return (RET_ERROR);
    328 }
    329 
    330 /*
    331  * BT_PAGE -- Split a non-root page of a btree.
    332  *
    333  * Parameters:
    334  *	t:	tree
    335  *	h:	root page
    336  *	lp:	pointer to left page pointer
    337  *	rp:	pointer to right page pointer
    338  *	skip:	pointer to index to leave open
    339  *	ilen:	insert length
    340  *
    341  * Returns:
    342  *	Pointer to page in which to insert or NULL on error.
    343  */
    344 static PAGE *
    345 bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
    346 {
    347 	PAGE *l, *r, *tp;
    348 	pgno_t npg;
    349 
    350 #ifdef STATISTICS
    351 	++bt_split;
    352 #endif
    353 	/* Put the new right page for the split into place. */
    354 	if ((r = __bt_new(t, &npg)) == NULL)
    355 		return (NULL);
    356 	r->pgno = npg;
    357 	r->lower = BTDATAOFF;
    358 	r->upper = t->bt_psize;
    359 	r->nextpg = h->nextpg;
    360 	r->prevpg = h->pgno;
    361 	r->flags = h->flags & P_TYPE;
    362 
    363 	/*
    364 	 * If we're splitting the last page on a level because we're appending
    365 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
    366 	 * sorted.  Adding an empty page on the side of the level is less work
    367 	 * and can push the fill factor much higher than normal.  If we're
    368 	 * wrong it's no big deal, we'll just do the split the right way next
    369 	 * time.  It may look like it's equally easy to do a similar hack for
    370 	 * reverse sorted data, that is, split the tree left, but it's not.
    371 	 * Don't even try.
    372 	 */
    373 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
    374 #ifdef STATISTICS
    375 		++bt_sortsplit;
    376 #endif
    377 		h->nextpg = r->pgno;
    378 		r->lower = BTDATAOFF + sizeof(indx_t);
    379 		*skip = 0;
    380 		*lp = h;
    381 		*rp = r;
    382 		return (r);
    383 	}
    384 
    385 	/* Put the new left page for the split into place. */
    386 	if ((l = calloc(1, t->bt_psize)) == NULL) {
    387 		mpool_put(t->bt_mp, r, 0);
    388 		return (NULL);
    389 	}
    390 #ifdef PURIFY
    391 	memset(l, 0xff, t->bt_psize);
    392 #endif
    393 	l->pgno = h->pgno;
    394 	l->nextpg = r->pgno;
    395 	l->prevpg = h->prevpg;
    396 	l->lower = BTDATAOFF;
    397 	l->upper = t->bt_psize;
    398 	l->flags = h->flags & P_TYPE;
    399 
    400 	/* Fix up the previous pointer of the page after the split page. */
    401 	if (h->nextpg != P_INVALID) {
    402 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
    403 			free(l);
    404 			/* XXX mpool_free(t->bt_mp, r->pgno); */
    405 			return (NULL);
    406 		}
    407 		tp->prevpg = r->pgno;
    408 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
    409 	}
    410 
    411 	/*
    412 	 * Split right.  The key/data pairs aren't sorted in the btree page so
    413 	 * it's simpler to copy the data from the split page onto two new pages
    414 	 * instead of copying half the data to the right page and compacting
    415 	 * the left page in place.  Since the left page can't change, we have
    416 	 * to swap the original and the allocated left page after the split.
    417 	 */
    418 	tp = bt_psplit(t, h, l, r, skip, ilen);
    419 
    420 	/* Move the new left page onto the old left page. */
    421 	memmove(h, l, t->bt_psize);
    422 	if (tp == l)
    423 		tp = h;
    424 	free(l);
    425 
    426 	*lp = h;
    427 	*rp = r;
    428 	return (tp);
    429 }
    430 
    431 /*
    432  * BT_ROOT -- Split the root page of a btree.
    433  *
    434  * Parameters:
    435  *	t:	tree
    436  *	h:	root page
    437  *	lp:	pointer to left page pointer
    438  *	rp:	pointer to right page pointer
    439  *	skip:	pointer to index to leave open
    440  *	ilen:	insert length
    441  *
    442  * Returns:
    443  *	Pointer to page in which to insert or NULL on error.
    444  */
    445 static PAGE *
    446 bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
    447 {
    448 	PAGE *l, *r, *tp;
    449 	pgno_t lnpg, rnpg;
    450 
    451 #ifdef STATISTICS
    452 	++bt_split;
    453 	++bt_rootsplit;
    454 #endif
    455 	/* Put the new left and right pages for the split into place. */
    456 	if ((l = __bt_new(t, &lnpg)) == NULL ||
    457 	    (r = __bt_new(t, &rnpg)) == NULL)
    458 		return (NULL);
    459 	l->pgno = lnpg;
    460 	r->pgno = rnpg;
    461 	l->nextpg = r->pgno;
    462 	r->prevpg = l->pgno;
    463 	l->prevpg = r->nextpg = P_INVALID;
    464 	l->lower = r->lower = BTDATAOFF;
    465 	l->upper = r->upper = t->bt_psize;
    466 	l->flags = r->flags = h->flags & P_TYPE;
    467 
    468 	/* Split the root page. */
    469 	tp = bt_psplit(t, h, l, r, skip, ilen);
    470 
    471 	*lp = l;
    472 	*rp = r;
    473 	return (tp);
    474 }
    475 
    476 /*
    477  * BT_RROOT -- Fix up the recno root page after it has been split.
    478  *
    479  * Parameters:
    480  *	t:	tree
    481  *	h:	root page
    482  *	l:	left page
    483  *	r:	right page
    484  *
    485  * Returns:
    486  *	RET_ERROR, RET_SUCCESS
    487  */
    488 static int
    489 bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
    490 {
    491 	char *dest;
    492 	uint32_t sz;
    493 	size_t temp;
    494 
    495 	temp = t->bt_psize - NRINTERNAL;
    496 	_DBFIT(temp, uint32_t);
    497 	sz = (uint32_t)temp;
    498 
    499 	/* Insert the left and right keys, set the header information. */
    500 	_DBFIT(sz, indx_t);
    501 	h->linp[0] = h->upper = (indx_t)sz;
    502 	dest = (char *)(void *)h + h->upper;
    503 	WR_RINTERNAL(dest,
    504 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
    505 
    506 	h->linp[1] = h->upper -= NRINTERNAL;
    507 	dest = (char *)(void *)h + h->upper;
    508 	WR_RINTERNAL(dest,
    509 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
    510 
    511 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    512 
    513 	/* Unpin the root page, set to recno internal page. */
    514 	h->flags &= ~P_TYPE;
    515 	h->flags |= P_RINTERNAL;
    516 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    517 
    518 	return (RET_SUCCESS);
    519 }
    520 
    521 /*
    522  * BT_BROOT -- Fix up the btree root page after it has been split.
    523  *
    524  * Parameters:
    525  *	t:	tree
    526  *	h:	root page
    527  *	l:	left page
    528  *	r:	right page
    529  *
    530  * Returns:
    531  *	RET_ERROR, RET_SUCCESS
    532  */
    533 static int
    534 bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
    535 {
    536 	BINTERNAL *bi = NULL;	/* pacify gcc */
    537 	BLEAF *bl;
    538 	uint32_t nbytes;
    539 	char *dest;
    540 
    541 	/*
    542 	 * If the root page was a leaf page, change it into an internal page.
    543 	 * We copy the key we split on (but not the key's data, in the case of
    544 	 * a leaf page) to the new root page.
    545 	 *
    546 	 * The btree comparison code guarantees that the left-most key on any
    547 	 * level of the tree is never used, so it doesn't need to be filled in.
    548 	 */
    549 	nbytes = NBINTERNAL(0);
    550 	h->linp[0] = h->upper = t->bt_psize - nbytes;
    551 	dest = (char *)(void *)h + h->upper;
    552 	WR_BINTERNAL(dest, 0, l->pgno, 0);
    553 
    554 	switch (h->flags & P_TYPE) {
    555 	case P_BLEAF:
    556 		bl = GETBLEAF(r, 0);
    557 		nbytes = NBINTERNAL(bl->ksize);
    558 		h->linp[1] = h->upper -= nbytes;
    559 		dest = (char *)(void *)h + h->upper;
    560 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
    561 		memmove(dest, bl->bytes, bl->ksize);
    562 
    563 		/*
    564 		 * If the key is on an overflow page, mark the overflow chain
    565 		 * so it isn't deleted when the leaf copy of the key is deleted.
    566 		 */
    567 		if (bl->flags & P_BIGKEY &&
    568 		    bt_preserve(t, *(pgno_t *)(void *)bl->bytes) == RET_ERROR)
    569 			return (RET_ERROR);
    570 		break;
    571 	case P_BINTERNAL:
    572 		bi = GETBINTERNAL(r, 0);
    573 		nbytes = NBINTERNAL(bi->ksize);
    574 		h->linp[1] = h->upper -= nbytes;
    575 		dest = (char *)(void *)h + h->upper;
    576 		memmove(dest, bi, nbytes);
    577 		((BINTERNAL *)(void *)dest)->pgno = r->pgno;
    578 		break;
    579 	default:
    580 		abort();
    581 	}
    582 
    583 	/* There are two keys on the page. */
    584 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    585 
    586 	/* Unpin the root page, set to btree internal page. */
    587 	h->flags &= ~P_TYPE;
    588 	h->flags |= P_BINTERNAL;
    589 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    590 
    591 	return (RET_SUCCESS);
    592 }
    593 
    594 /*
    595  * BT_PSPLIT -- Do the real work of splitting the page.
    596  *
    597  * Parameters:
    598  *	t:	tree
    599  *	h:	page to be split
    600  *	l:	page to put lower half of data
    601  *	r:	page to put upper half of data
    602  *	pskip:	pointer to index to leave open
    603  *	ilen:	insert length
    604  *
    605  * Returns:
    606  *	Pointer to page in which to insert.
    607  */
    608 static PAGE *
    609 bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
    610 {
    611 	BINTERNAL *bi;
    612 	BLEAF *bl;
    613 	CURSOR *c;
    614 	RLEAF *rl;
    615 	PAGE *rval;
    616 	void *src = NULL;	/* pacify gcc */
    617 	indx_t full, half, nxt, off, skip, top, used;
    618 	uint32_t nbytes;
    619 	size_t temp;
    620 	int bigkeycnt, isbigkey;
    621 
    622 	/*
    623 	 * Split the data to the left and right pages.  Leave the skip index
    624 	 * open.  Additionally, make some effort not to split on an overflow
    625 	 * key.  This makes internal page processing faster and can save
    626 	 * space as overflow keys used by internal pages are never deleted.
    627 	 */
    628 	bigkeycnt = 0;
    629 	skip = *pskip;
    630 	temp = t->bt_psize - BTDATAOFF;
    631 	_DBFIT(temp, indx_t);
    632 	full = (indx_t)temp;
    633 	half = full / 2;
    634 	used = 0;
    635 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
    636 		if (skip == off) {
    637 			_DBFIT(ilen, uint32_t);
    638 			nbytes = (uint32_t)ilen;
    639 			isbigkey = 0;		/* XXX: not really known. */
    640 		} else
    641 			switch (h->flags & P_TYPE) {
    642 			case P_BINTERNAL:
    643 				src = bi = GETBINTERNAL(h, nxt);
    644 				nbytes = NBINTERNAL(bi->ksize);
    645 				isbigkey = bi->flags & P_BIGKEY;
    646 				break;
    647 			case P_BLEAF:
    648 				src = bl = GETBLEAF(h, nxt);
    649 				nbytes = NBLEAF(bl);
    650 				isbigkey = bl->flags & P_BIGKEY;
    651 				break;
    652 			case P_RINTERNAL:
    653 				src = GETRINTERNAL(h, nxt);
    654 				nbytes = NRINTERNAL;
    655 				isbigkey = 0;
    656 				break;
    657 			case P_RLEAF:
    658 				src = rl = GETRLEAF(h, nxt);
    659 				nbytes = NRLEAF(rl);
    660 				isbigkey = 0;
    661 				break;
    662 			default:
    663 				abort();
    664 			}
    665 
    666 		/*
    667 		 * If the key/data pairs are substantial fractions of the max
    668 		 * possible size for the page, it's possible to get situations
    669 		 * where we decide to try and copy too much onto the left page.
    670 		 * Make sure that doesn't happen.
    671 		 */
    672 		if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
    673 		    nxt == top - 1) {
    674 			--off;
    675 			break;
    676 		}
    677 
    678 		/* Copy the key/data pair, if not the skipped index. */
    679 		if (skip != off) {
    680 			++nxt;
    681 
    682 			l->linp[off] = l->upper -= nbytes;
    683 			memmove((char *)(void *)l + l->upper, src, nbytes);
    684 		}
    685 
    686 		temp = nbytes + sizeof(indx_t);
    687 		_DBFIT(temp, indx_t);
    688 		used += (indx_t)temp;
    689 		if (used >= half) {
    690 			if (!isbigkey || bigkeycnt == 3)
    691 				break;
    692 			else
    693 				++bigkeycnt;
    694 		}
    695 	}
    696 
    697 	/*
    698 	 * Off is the last offset that's valid for the left page.
    699 	 * Nxt is the first offset to be placed on the right page.
    700 	 */
    701 	temp = (off + 1) * sizeof(indx_t);
    702 	_DBFIT(temp, indx_t);
    703 	l->lower += (indx_t)temp;
    704 
    705 	/*
    706 	 * If splitting the page that the cursor was on, the cursor has to be
    707 	 * adjusted to point to the same record as before the split.  If the
    708 	 * cursor is at or past the skipped slot, the cursor is incremented by
    709 	 * one.  If the cursor is on the right page, it is decremented by the
    710 	 * number of records split to the left page.
    711 	 */
    712 	c = &t->bt_cursor;
    713 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
    714 		if (c->pg.index >= skip)
    715 			++c->pg.index;
    716 		if (c->pg.index < nxt)			/* Left page. */
    717 			c->pg.pgno = l->pgno;
    718 		else {					/* Right page. */
    719 			c->pg.pgno = r->pgno;
    720 			c->pg.index -= nxt;
    721 		}
    722 	}
    723 
    724 	/*
    725 	 * If the skipped index was on the left page, just return that page.
    726 	 * Otherwise, adjust the skip index to reflect the new position on
    727 	 * the right page.
    728 	 */
    729 	if (skip <= off) {
    730 		skip = MAX_PAGE_OFFSET;
    731 		rval = l;
    732 	} else {
    733 		rval = r;
    734 		*pskip -= nxt;
    735 	}
    736 
    737 	for (off = 0; nxt < top; ++off) {
    738 		if (skip == nxt) {
    739 			++off;
    740 			skip = MAX_PAGE_OFFSET;
    741 		}
    742 		switch (h->flags & P_TYPE) {
    743 		case P_BINTERNAL:
    744 			src = bi = GETBINTERNAL(h, nxt);
    745 			nbytes = NBINTERNAL(bi->ksize);
    746 			break;
    747 		case P_BLEAF:
    748 			src = bl = GETBLEAF(h, nxt);
    749 			nbytes = NBLEAF(bl);
    750 			break;
    751 		case P_RINTERNAL:
    752 			src = GETRINTERNAL(h, nxt);
    753 			nbytes = NRINTERNAL;
    754 			break;
    755 		case P_RLEAF:
    756 			src = rl = GETRLEAF(h, nxt);
    757 			nbytes = NRLEAF(rl);
    758 			break;
    759 		default:
    760 			abort();
    761 		}
    762 		++nxt;
    763 		r->linp[off] = r->upper -= nbytes;
    764 		memmove((char *)(void *)r + r->upper, src, nbytes);
    765 	}
    766 	temp = off * sizeof(indx_t);
    767 	_DBFIT(temp, indx_t);
    768 	r->lower += (indx_t)temp;
    769 
    770 	/* If the key is being appended to the page, adjust the index. */
    771 	if (skip == top)
    772 		r->lower += sizeof(indx_t);
    773 
    774 	return (rval);
    775 }
    776 
    777 /*
    778  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
    779  *
    780  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
    781  * record that references them gets deleted.  Chains pointed to by internal
    782  * pages never get deleted.  This routine marks a chain as pointed to by an
    783  * internal page.
    784  *
    785  * Parameters:
    786  *	t:	tree
    787  *	pg:	page number of first page in the chain.
    788  *
    789  * Returns:
    790  *	RET_SUCCESS, RET_ERROR.
    791  */
    792 static int
    793 bt_preserve(BTREE *t, pgno_t pg)
    794 {
    795 	PAGE *h;
    796 
    797 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
    798 		return (RET_ERROR);
    799 	h->flags |= P_PRESERVE;
    800 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    801 	return (RET_SUCCESS);
    802 }
    803 
    804 /*
    805  * REC_TOTAL -- Return the number of recno entries below a page.
    806  *
    807  * Parameters:
    808  *	h:	page
    809  *
    810  * Returns:
    811  *	The number of recno entries below a page.
    812  *
    813  * XXX
    814  * These values could be set by the bt_psplit routine.  The problem is that the
    815  * entry has to be popped off of the stack etc. or the values have to be passed
    816  * all the way back to bt_split/bt_rroot and it's not very clean.
    817  */
    818 static recno_t
    819 rec_total(PAGE *h)
    820 {
    821 	recno_t recs;
    822 	indx_t nxt, top;
    823 
    824 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
    825 		recs += GETRINTERNAL(h, nxt)->nrecs;
    826 	return (recs);
    827 }
    828