Home | History | Annotate | Download | only in btree
      1 #pragma ident	"%Z%%M%	%I%	%E% SMI"
      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 #if defined(LIBC_SCCS) && !defined(lint)
     40 static char sccsid[] = "@(#)bt_split.c	8.10 (Berkeley) 1/9/95";
     41 #endif /* LIBC_SCCS and not lint */
     42 
     43 #include <sys/types.h>
     44 
     45 #include <limits.h>
     46 #include <stdio.h>
     47 #include <stdlib.h>
     48 #include <string.h>
     49 
     50 #include "db-int.h"
     51 #include "btree.h"
     52 
     53 static int	 bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *));
     54 static PAGE	*bt_page
     55 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
     56 static int	 bt_preserve __P((BTREE *, db_pgno_t));
     57 static PAGE	*bt_psplit
     58 		    __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t));
     59 static PAGE	*bt_root
     60 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
     61 static int	 bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
     62 static recno_t	 rec_total __P((PAGE *));
     63 
     64 #ifdef STATISTICS
     65 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
     66 #endif
     67 
     68 /*
     69  * __BT_SPLIT -- Split the tree.
     70  *
     71  * Parameters:
     72  *	t:	tree
     73  *	sp:	page to split
     74  *	key:	key to insert
     75  *	data:	data to insert
     76  *	flags:	BIGKEY/BIGDATA flags
     77  *	ilen:	insert length
     78  *	skip:	index to leave open
     79  *
     80  * Returns:
     81  *	RET_ERROR, RET_SUCCESS
     82  */
     83 int
     84 __bt_split(t, sp, key, data, flags, ilen, argskip)
     85 	BTREE *t;
     86 	PAGE *sp;
     87 	const DBT *key, *data;
     88 	int flags;
     89 	size_t ilen;
     90 	u_int32_t argskip;
     91 {
     92 	BINTERNAL *bi;
     93 	BLEAF *bl, *tbl;
     94 	DBT a, b;
     95 	EPGNO *parent;
     96 	PAGE *h, *l, *r, *lchild, *rchild;
     97 	indx_t nxtindex;
     98 	u_int16_t skip;
     99 	u_int32_t n, nbytes, nksize;
    100 	int parentsplit;
    101 	char *dest;
    102 
    103 	/*
    104 	 * Split the page into two pages, l and r.  The split routines return
    105 	 * a pointer to the page into which the key should be inserted and with
    106 	 * skip set to the offset which should be used.  Additionally, l and r
    107 	 * are pinned.
    108 	 */
    109 	skip = argskip;
    110 	h = sp->pgno == P_ROOT ?
    111 	    bt_root(t, sp, &l, &r, &skip, ilen) :
    112 	    bt_page(t, sp, &l, &r, &skip, ilen);
    113 	if (h == NULL)
    114 		return (RET_ERROR);
    115 
    116 	/*
    117 	 * Insert the new key/data pair into the leaf page.  (Key inserts
    118 	 * always cause a leaf page to split first.)
    119 	 */
    120 	h->linp[skip] = h->upper -= ilen;
    121 	dest = (char *)h + h->upper;
    122 	if (F_ISSET(t, R_RECNO))
    123 		WR_RLEAF(dest, data, flags)
    124 	else
    125 		WR_BLEAF(dest, key, data, flags)
    126 
    127 	/* If the root page was split, make it look right. */
    128 	if (sp->pgno == P_ROOT &&
    129 	    (F_ISSET(t, R_RECNO) ?
    130 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    131 		goto err2;
    132 
    133 	/*
    134 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
    135 	 * were traversed when we searched for the page that split.  Each stack
    136 	 * entry is a page number and a page index offset.  The offset is for
    137 	 * the page traversed on the search.  We've just split a page, so we
    138 	 * have to insert a new key into the parent page.
    139 	 *
    140 	 * If the insert into the parent page causes it to split, may have to
    141 	 * continue splitting all the way up the tree.  We stop if the root
    142 	 * splits or the page inserted into didn't have to split to hold the
    143 	 * new key.  Some algorithms replace the key for the old page as well
    144 	 * as the new page.  We don't, as there's no reason to believe that the
    145 	 * first key on the old page is any better than the key we have, and,
    146 	 * in the case of a key being placed at index 0 causing the split, the
    147 	 * key is unavailable.
    148 	 *
    149 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
    150 	 * and right pages pinned while working on the parent.   The 5 are the
    151 	 * two children, left parent and right parent (when the parent splits)
    152 	 * and the root page or the overflow key page when calling bt_preserve.
    153 	 * This code must make sure that all pins are released other than the
    154 	 * root page or overflow page which is unlocked elsewhere.
    155 	 */
    156 	while ((parent = BT_POP(t)) != NULL) {
    157 		lchild = l;
    158 		rchild = r;
    159 
    160 		/* Get the parent page. */
    161 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
    162 			goto err2;
    163 
    164 	 	/*
    165 		 * The new key goes ONE AFTER the index, because the split
    166 		 * was to the right.
    167 		 */
    168 		skip = parent->index + 1;
    169 
    170 		/*
    171 		 * Calculate the space needed on the parent page.
    172 		 *
    173 		 * Prefix trees: space hack when inserting into BINTERNAL
    174 		 * pages.  Retain only what's needed to distinguish between
    175 		 * the new entry and the LAST entry on the page to its left.
    176 		 * If the keys compare equal, retain the entire key.  Note,
    177 		 * we don't touch overflow keys, and the entire key must be
    178 		 * retained for the next-to-left most key on the leftmost
    179 		 * page of each level, or the search will fail.  Applicable
    180 		 * ONLY to internal pages that have leaf pages as children.
    181 		 * Further reduction of the key between pairs of internal
    182 		 * pages loses too much information.
    183 		 */
    184 		switch (rchild->flags & P_TYPE) {
    185 		case P_BINTERNAL:
    186 			bi = GETBINTERNAL(rchild, 0);
    187 			nbytes = NBINTERNAL(bi->ksize);
    188 			break;
    189 		case P_BLEAF:
    190 			bl = GETBLEAF(rchild, 0);
    191 			nbytes = NBINTERNAL(bl->ksize);
    192 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
    193 			    (h->prevpg != P_INVALID || skip > 1)) {
    194 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
    195 				a.size = tbl->ksize;
    196 				a.data = tbl->bytes;
    197 				b.size = bl->ksize;
    198 				b.data = bl->bytes;
    199 				nksize = t->bt_pfx(&a, &b);
    200 				n = NBINTERNAL(nksize);
    201 				if (n < nbytes) {
    202 #ifdef STATISTICS
    203 					bt_pfxsaved += nbytes - n;
    204 #endif
    205 					nbytes = n;
    206 				} else
    207 					nksize = 0;
    208 			} else
    209 				nksize = 0;
    210 			break;
    211 		case P_RINTERNAL:
    212 		case P_RLEAF:
    213 			nbytes = NRINTERNAL;
    214 			break;
    215 		default:
    216 			abort();
    217 		}
    218 
    219 		/* Split the parent page if necessary or shift the indices. */
    220 		if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
    221 			sp = h;
    222 			h = h->pgno == P_ROOT ?
    223 			    bt_root(t, h, &l, &r, &skip, nbytes) :
    224 			    bt_page(t, h, &l, &r, &skip, nbytes);
    225 			if (h == NULL)
    226 				goto err1;
    227 			parentsplit = 1;
    228 		} else {
    229 			if (skip < (nxtindex = NEXTINDEX(h)))
    230 				memmove(h->linp + skip + 1, h->linp + skip,
    231 				    (nxtindex - skip) * sizeof(indx_t));
    232 			h->lower += sizeof(indx_t);
    233 			parentsplit = 0;
    234 		}
    235 
    236 		/* Insert the key into the parent page. */
    237 		switch (rchild->flags & P_TYPE) {
    238 		case P_BINTERNAL:
    239 			h->linp[skip] = h->upper -= nbytes;
    240 			dest = (char *)h + h->linp[skip];
    241 			memmove(dest, bi, nbytes);
    242 			((BINTERNAL *)dest)->pgno = rchild->pgno;
    243 			break;
    244 		case P_BLEAF:
    245 			h->linp[skip] = h->upper -= nbytes;
    246 			dest = (char *)h + h->linp[skip];
    247 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
    248 			    rchild->pgno, bl->flags & P_BIGKEY);
    249 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
    250 			if (bl->flags & P_BIGKEY &&
    251 			    bt_preserve(t, *(db_pgno_t *)bl->bytes) == RET_ERROR)
    252 				goto err1;
    253 			break;
    254 		case P_RINTERNAL:
    255 			/*
    256 			 * Update the left page count.  If split
    257 			 * added at index 0, fix the correct page.
    258 			 */
    259 			if (skip > 0)
    260 				dest = (char *)h + h->linp[skip - 1];
    261 			else
    262 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
    263 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
    264 			((RINTERNAL *)dest)->pgno = lchild->pgno;
    265 
    266 			/* Update the right page count. */
    267 			h->linp[skip] = h->upper -= nbytes;
    268 			dest = (char *)h + h->linp[skip];
    269 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
    270 			((RINTERNAL *)dest)->pgno = rchild->pgno;
    271 			break;
    272 		case P_RLEAF:
    273 			/*
    274 			 * Update the left page count.  If split
    275 			 * added at index 0, fix the correct page.
    276 			 */
    277 			if (skip > 0)
    278 				dest = (char *)h + h->linp[skip - 1];
    279 			else
    280 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
    281 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
    282 			((RINTERNAL *)dest)->pgno = lchild->pgno;
    283 
    284 			/* Update the right page count. */
    285 			h->linp[skip] = h->upper -= nbytes;
    286 			dest = (char *)h + h->linp[skip];
    287 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
    288 			((RINTERNAL *)dest)->pgno = rchild->pgno;
    289 			break;
    290 		default:
    291 			abort();
    292 		}
    293 
    294 		/* Unpin the held pages. */
    295 		if (!parentsplit) {
    296 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    297 			break;
    298 		}
    299 
    300 		/* If the root page was split, make it look right. */
    301 		if (sp->pgno == P_ROOT &&
    302 		    (F_ISSET(t, R_RECNO) ?
    303 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    304 			goto err1;
    305 
    306 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    307 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    308 	}
    309 
    310 	/* Unpin the held pages. */
    311 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
    312 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
    313 
    314 	/* Clear any pages left on the stack. */
    315 	return (RET_SUCCESS);
    316 
    317 	/*
    318 	 * If something fails in the above loop we were already walking back
    319 	 * up the tree and the tree is now inconsistent.  Nothing much we can
    320 	 * do about it but release any memory we're holding.
    321 	 */
    322 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    323 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    324 
    325 err2:	mpool_put(t->bt_mp, l, 0);
    326 	mpool_put(t->bt_mp, r, 0);
    327 	__dbpanic(t->bt_dbp);
    328 	return (RET_ERROR);
    329 }
    330 
    331 /*
    332  * BT_PAGE -- Split a non-root page of a btree.
    333  *
    334  * Parameters:
    335  *	t:	tree
    336  *	h:	root page
    337  *	lp:	pointer to left page pointer
    338  *	rp:	pointer to right page pointer
    339  *	skip:	pointer to index to leave open
    340  *	ilen:	insert length
    341  *
    342  * Returns:
    343  *	Pointer to page in which to insert or NULL on error.
    344  */
    345 static PAGE *
    346 bt_page(t, h, lp, rp, skip, ilen)
    347 	BTREE *t;
    348 	PAGE *h, **lp, **rp;
    349 	indx_t *skip;
    350 	size_t ilen;
    351 {
    352 	PAGE *l, *r, *tp;
    353 	db_pgno_t npg;
    354 
    355 #ifdef STATISTICS
    356 	++bt_split;
    357 #endif
    358 	/* Put the new right page for the split into place. */
    359 	if ((r = __bt_new(t, &npg)) == NULL)
    360 		return (NULL);
    361 	r->pgno = npg;
    362 	r->lower = BTDATAOFF;
    363 	r->upper = t->bt_psize;
    364 	r->nextpg = h->nextpg;
    365 	r->prevpg = h->pgno;
    366 	r->flags = h->flags & P_TYPE;
    367 
    368 	/*
    369 	 * If we're splitting the last page on a level because we're appending
    370 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
    371 	 * sorted.  Adding an empty page on the side of the level is less work
    372 	 * and can push the fill factor much higher than normal.  If we're
    373 	 * wrong it's no big deal, we'll just do the split the right way next
    374 	 * time.  It may look like it's equally easy to do a similar hack for
    375 	 * reverse sorted data, that is, split the tree left, but it's not.
    376 	 * Don't even try.
    377 	 */
    378 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
    379 #ifdef STATISTICS
    380 		++bt_sortsplit;
    381 #endif
    382 		h->nextpg = r->pgno;
    383 		r->lower = BTDATAOFF + sizeof(indx_t);
    384 		*skip = 0;
    385 		*lp = h;
    386 		*rp = r;
    387 		return (r);
    388 	}
    389 
    390 	/* Put the new left page for the split into place. */
    391 	if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
    392 		mpool_put(t->bt_mp, r, 0);
    393 		return (NULL);
    394 	}
    395 #ifdef PURIFY
    396 	memset(l, 0xff, t->bt_psize);
    397 #endif
    398 	l->pgno = h->pgno;
    399 	l->nextpg = r->pgno;
    400 	l->prevpg = h->prevpg;
    401 	l->lower = BTDATAOFF;
    402 	l->upper = t->bt_psize;
    403 	l->flags = h->flags & P_TYPE;
    404 
    405 	/* Fix up the previous pointer of the page after the split page. */
    406 	if (h->nextpg != P_INVALID) {
    407 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
    408 			free(l);
    409 			/* XXX mpool_free(t->bt_mp, r->pgno); */
    410 			return (NULL);
    411 		}
    412 		tp->prevpg = r->pgno;
    413 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
    414 	}
    415 
    416 	/*
    417 	 * Split right.  The key/data pairs aren't sorted in the btree page so
    418 	 * it's simpler to copy the data from the split page onto two new pages
    419 	 * instead of copying half the data to the right page and compacting
    420 	 * the left page in place.  Since the left page can't change, we have
    421 	 * to swap the original and the allocated left page after the split.
    422 	 */
    423 	tp = bt_psplit(t, h, l, r, skip, ilen);
    424 
    425 	/* Move the new left page onto the old left page. */
    426 	memmove(h, l, t->bt_psize);
    427 	if (tp == l)
    428 		tp = h;
    429 	free(l);
    430 
    431 	*lp = h;
    432 	*rp = r;
    433 	return (tp);
    434 }
    435 
    436 /*
    437  * BT_ROOT -- Split the root page of a btree.
    438  *
    439  * Parameters:
    440  *	t:	tree
    441  *	h:	root page
    442  *	lp:	pointer to left page pointer
    443  *	rp:	pointer to right page pointer
    444  *	skip:	pointer to index to leave open
    445  *	ilen:	insert length
    446  *
    447  * Returns:
    448  *	Pointer to page in which to insert or NULL on error.
    449  */
    450 static PAGE *
    451 bt_root(t, h, lp, rp, skip, ilen)
    452 	BTREE *t;
    453 	PAGE *h, **lp, **rp;
    454 	indx_t *skip;
    455 	size_t ilen;
    456 {
    457 	PAGE *l, *r, *tp;
    458 	db_pgno_t lnpg, rnpg;
    459 
    460 #ifdef STATISTICS
    461 	++bt_split;
    462 	++bt_rootsplit;
    463 #endif
    464 	/* Put the new left and right pages for the split into place. */
    465 	if ((l = __bt_new(t, &lnpg)) == NULL ||
    466 	    (r = __bt_new(t, &rnpg)) == NULL)
    467 		return (NULL);
    468 	l->pgno = lnpg;
    469 	r->pgno = rnpg;
    470 	l->nextpg = r->pgno;
    471 	r->prevpg = l->pgno;
    472 	l->prevpg = r->nextpg = P_INVALID;
    473 	l->lower = r->lower = BTDATAOFF;
    474 	l->upper = r->upper = t->bt_psize;
    475 	l->flags = r->flags = h->flags & P_TYPE;
    476 
    477 	/* Split the root page. */
    478 	tp = bt_psplit(t, h, l, r, skip, ilen);
    479 
    480 	*lp = l;
    481 	*rp = r;
    482 	return (tp);
    483 }
    484 
    485 /*
    486  * BT_RROOT -- Fix up the recno root page after it has been split.
    487  *
    488  * Parameters:
    489  *	t:	tree
    490  *	h:	root page
    491  *	l:	left page
    492  *	r:	right page
    493  *
    494  * Returns:
    495  *	RET_ERROR, RET_SUCCESS
    496  */
    497 static int
    498 bt_rroot(t, h, l, r)
    499 	BTREE *t;
    500 	PAGE *h, *l, *r;
    501 {
    502 	char *dest;
    503 
    504 	/* Insert the left and right keys, set the header information. */
    505 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
    506 	dest = (char *)h + h->upper;
    507 	WR_RINTERNAL(dest,
    508 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
    509 
    510 	h->linp[1] = h->upper -= NRINTERNAL;
    511 	dest = (char *)h + h->upper;
    512 	WR_RINTERNAL(dest,
    513 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
    514 
    515 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    516 
    517 	/* Unpin the root page, set to recno internal page. */
    518 	h->flags &= ~P_TYPE;
    519 	h->flags |= P_RINTERNAL;
    520 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    521 
    522 	return (RET_SUCCESS);
    523 }
    524 
    525 /*
    526  * BT_BROOT -- Fix up the btree root page after it has been split.
    527  *
    528  * Parameters:
    529  *	t:	tree
    530  *	h:	root page
    531  *	l:	left page
    532  *	r:	right page
    533  *
    534  * Returns:
    535  *	RET_ERROR, RET_SUCCESS
    536  */
    537 static int
    538 bt_broot(t, h, l, r)
    539 	BTREE *t;
    540 	PAGE *h, *l, *r;
    541 {
    542 	BINTERNAL *bi;
    543 	BLEAF *bl;
    544 	u_int32_t nbytes;
    545 	char *dest;
    546 
    547 	/*
    548 	 * If the root page was a leaf page, change it into an internal page.
    549 	 * We copy the key we split on (but not the key's data, in the case of
    550 	 * a leaf page) to the new root page.
    551 	 *
    552 	 * The btree comparison code guarantees that the left-most key on any
    553 	 * level of the tree is never used, so it doesn't need to be filled in.
    554 	 */
    555 	nbytes = NBINTERNAL(0);
    556 	h->linp[0] = h->upper = t->bt_psize - nbytes;
    557 	dest = (char *)h + h->upper;
    558 	WR_BINTERNAL(dest, 0, l->pgno, 0);
    559 
    560 	switch (h->flags & P_TYPE) {
    561 	case P_BLEAF:
    562 		bl = GETBLEAF(r, 0);
    563 		nbytes = NBINTERNAL(bl->ksize);
    564 		h->linp[1] = h->upper -= nbytes;
    565 		dest = (char *)h + h->upper;
    566 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
    567 		memmove(dest, bl->bytes, bl->ksize);
    568 
    569 		/*
    570 		 * If the key is on an overflow page, mark the overflow chain
    571 		 * so it isn't deleted when the leaf copy of the key is deleted.
    572 		 */
    573 		if (bl->flags & P_BIGKEY &&
    574 		    bt_preserve(t, *(db_pgno_t *)bl->bytes) == RET_ERROR)
    575 			return (RET_ERROR);
    576 		break;
    577 	case P_BINTERNAL:
    578 		bi = GETBINTERNAL(r, 0);
    579 		nbytes = NBINTERNAL(bi->ksize);
    580 		h->linp[1] = h->upper -= nbytes;
    581 		dest = (char *)h + h->upper;
    582 		memmove(dest, bi, nbytes);
    583 		((BINTERNAL *)dest)->pgno = r->pgno;
    584 		break;
    585 	default:
    586 		abort();
    587 	}
    588 
    589 	/* There are two keys on the page. */
    590 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    591 
    592 	/* Unpin the root page, set to btree internal page. */
    593 	h->flags &= ~P_TYPE;
    594 	h->flags |= P_BINTERNAL;
    595 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    596 
    597 	return (RET_SUCCESS);
    598 }
    599 
    600 /*
    601  * BT_PSPLIT -- Do the real work of splitting the page.
    602  *
    603  * Parameters:
    604  *	t:	tree
    605  *	h:	page to be split
    606  *	l:	page to put lower half of data
    607  *	r:	page to put upper half of data
    608  *	pskip:	pointer to index to leave open
    609  *	ilen:	insert length
    610  *
    611  * Returns:
    612  *	Pointer to page in which to insert.
    613  */
    614 static PAGE *
    615 bt_psplit(t, h, l, r, pskip, ilen)
    616 	BTREE *t;
    617 	PAGE *h, *l, *r;
    618 	indx_t *pskip;
    619 	size_t ilen;
    620 {
    621 	BINTERNAL *bi;
    622 	BLEAF *bl;
    623 	CURSOR *c;
    624 	RLEAF *rl;
    625 	PAGE *rval;
    626 	void *src;
    627 	indx_t full, half, nxt, off, skip, top, used;
    628 	u_int32_t nbytes;
    629 	int bigkeycnt, isbigkey;
    630 
    631 	/*
    632 	 * Split the data to the left and right pages.  Leave the skip index
    633 	 * open.  Additionally, make some effort not to split on an overflow
    634 	 * key.  This makes internal page processing faster and can save
    635 	 * space as overflow keys used by internal pages are never deleted.
    636 	 */
    637 	bigkeycnt = 0;
    638 	skip = *pskip;
    639 	full = t->bt_psize - BTDATAOFF;
    640 	half = full / 2;
    641 	used = 0;
    642 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
    643 		if (skip == off) {
    644 			nbytes = ilen;
    645 			isbigkey = 0;		/* XXX: not really known. */
    646 		} else
    647 			switch (h->flags & P_TYPE) {
    648 			case P_BINTERNAL:
    649 				src = bi = GETBINTERNAL(h, nxt);
    650 				nbytes = NBINTERNAL(bi->ksize);
    651 				isbigkey = bi->flags & P_BIGKEY;
    652 				break;
    653 			case P_BLEAF:
    654 				src = bl = GETBLEAF(h, nxt);
    655 				nbytes = NBLEAF(bl);
    656 				isbigkey = bl->flags & P_BIGKEY;
    657 				break;
    658 			case P_RINTERNAL:
    659 				src = GETRINTERNAL(h, nxt);
    660 				nbytes = NRINTERNAL;
    661 				isbigkey = 0;
    662 				break;
    663 			case P_RLEAF:
    664 				src = rl = GETRLEAF(h, nxt);
    665 				nbytes = NRLEAF(rl);
    666 				isbigkey = 0;
    667 				break;
    668 			default:
    669 				abort();
    670 			}
    671 
    672 		/*
    673 		 * If the key/data pairs are substantial fractions of the max
    674 		 * possible size for the page, it's possible to get situations
    675 		 * where we decide to try and copy too much onto the left page.
    676 		 * Make sure that doesn't happen.
    677 		 */
    678 		if ((skip <= off && used + nbytes + sizeof(indx_t) >= full)
    679 		    || nxt == top - 1) {
    680 			--off;
    681 			break;
    682 		}
    683 
    684 		/* Copy the key/data pair, if not the skipped index. */
    685 		if (skip != off) {
    686 			++nxt;
    687 
    688 			l->linp[off] = l->upper -= nbytes;
    689 			memmove((char *)l + l->upper, src, nbytes);
    690 		}
    691 
    692 		used += nbytes + sizeof(indx_t);
    693 		if (used >= half) {
    694 			if (!isbigkey || bigkeycnt == 3)
    695 				break;
    696 			else
    697 				++bigkeycnt;
    698 		}
    699 	}
    700 
    701 	/*
    702 	 * Off is the last offset that's valid for the left page.
    703 	 * Nxt is the first offset to be placed on the right page.
    704 	 */
    705 	l->lower += (off + 1) * sizeof(indx_t);
    706 
    707 	/*
    708 	 * If splitting the page that the cursor was on, the cursor has to be
    709 	 * adjusted to point to the same record as before the split.  If the
    710 	 * cursor is at or past the skipped slot, the cursor is incremented by
    711 	 * one.  If the cursor is on the right page, it is decremented by the
    712 	 * number of records split to the left page.
    713 	 */
    714 	c = &t->bt_cursor;
    715 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
    716 		if (c->pg.index >= skip)
    717 			++c->pg.index;
    718 		if (c->pg.index < nxt)			/* Left page. */
    719 			c->pg.pgno = l->pgno;
    720 		else {					/* Right page. */
    721 			c->pg.pgno = r->pgno;
    722 			c->pg.index -= nxt;
    723 		}
    724 	}
    725 
    726 	/*
    727 	 * If the skipped index was on the left page, just return that page.
    728 	 * Otherwise, adjust the skip index to reflect the new position on
    729 	 * the right page.
    730 	 */
    731 	if (skip <= off) {
    732 		/*
    733 		 * If we get here then 'skip' is in the left page.  We do
    734 		 * not want to mix this with the right page, so we assign
    735 		 * an unrealistic value (-1).
    736 		 */
    737 		skip = (indx_t)-1;
    738 		rval = l;
    739 	} else {
    740 		rval = r;
    741 		*pskip -= nxt;
    742 	}
    743 
    744 	for (off = 0; nxt < top; ++off) {
    745 		if (skip == nxt) {
    746 			++off;
    747 			/*
    748 			 * Assign 'skip' an unrealistic value (-1) to ensure
    749 			 * it is not matched again.
    750 			 */
    751 			skip = (indx_t)-1;
    752 		}
    753 		switch (h->flags & P_TYPE) {
    754 		case P_BINTERNAL:
    755 			src = bi = GETBINTERNAL(h, nxt);
    756 			nbytes = NBINTERNAL(bi->ksize);
    757 			break;
    758 		case P_BLEAF:
    759 			src = bl = GETBLEAF(h, nxt);
    760 			nbytes = NBLEAF(bl);
    761 			break;
    762 		case P_RINTERNAL:
    763 			src = GETRINTERNAL(h, nxt);
    764 			nbytes = NRINTERNAL;
    765 			break;
    766 		case P_RLEAF:
    767 			src = rl = GETRLEAF(h, nxt);
    768 			nbytes = NRLEAF(rl);
    769 			break;
    770 		default:
    771 			abort();
    772 		}
    773 		++nxt;
    774 		r->linp[off] = r->upper -= nbytes;
    775 		memmove((char *)r + r->upper, src, nbytes);
    776 	}
    777 	r->lower += off * sizeof(indx_t);
    778 
    779 	/* If the key is being appended to the page, adjust the index. */
    780 	if (skip == top)
    781 		r->lower += sizeof(indx_t);
    782 
    783 	return (rval);
    784 }
    785 
    786 /*
    787  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
    788  *
    789  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
    790  * record that references them gets deleted.  Chains pointed to by internal
    791  * pages never get deleted.  This routine marks a chain as pointed to by an
    792  * internal page.
    793  *
    794  * Parameters:
    795  *	t:	tree
    796  *	pg:	page number of first page in the chain.
    797  *
    798  * Returns:
    799  *	RET_SUCCESS, RET_ERROR.
    800  */
    801 static int
    802 bt_preserve(t, pg)
    803 	BTREE *t;
    804 	db_pgno_t pg;
    805 {
    806 	PAGE *h;
    807 
    808 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
    809 		return (RET_ERROR);
    810 	h->flags |= P_PRESERVE;
    811 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    812 	return (RET_SUCCESS);
    813 }
    814 
    815 /*
    816  * REC_TOTAL -- Return the number of recno entries below a page.
    817  *
    818  * Parameters:
    819  *	h:	page
    820  *
    821  * Returns:
    822  *	The number of recno entries below a page.
    823  *
    824  * XXX
    825  * These values could be set by the bt_psplit routine.  The problem is that the
    826  * entry has to be popped off of the stack etc. or the values have to be passed
    827  * all the way back to bt_split/bt_rroot and it's not very clean.
    828  */
    829 static recno_t
    830 rec_total(h)
    831 	PAGE *h;
    832 {
    833 	recno_t recs;
    834 	indx_t nxt, top;
    835 
    836 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
    837 		recs += GETRINTERNAL(h, nxt)->nrecs;
    838 	return (recs);
    839 }
    840