xref: /freebsd/lib/libc/db/btree/bt_split.c (revision e6dbc99d47ddb254d75822817592bb82b5ce4d97)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
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 defined(LIBC_SCCS) && !defined(lint)
36 static char sccsid[] = "@(#)bt_split.c	8.10 (Berkeley) 1/9/95";
37 #endif /* LIBC_SCCS and not lint */
38 #include <sys/param.h>
39 
40 #include <limits.h>
41 #include <stdio.h>
42 #include <stdlib.h>
43 #include <string.h>
44 
45 #include <db.h>
46 #include "btree.h"
47 
48 static int	 bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
49 static PAGE	*bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
50 static int	 bt_preserve(BTREE *, pgno_t);
51 static PAGE	*bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
52 static PAGE	*bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
53 static int	 bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
54 static recno_t	 rec_total(PAGE *);
55 
56 #ifdef STATISTICS
57 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
58 #endif
59 
60 /*
61  * __BT_SPLIT -- Split the tree.
62  *
63  * Parameters:
64  *	t:	tree
65  *	sp:	page to split
66  *	key:	key to insert
67  *	data:	data to insert
68  *	flags:	BIGKEY/BIGDATA flags
69  *	ilen:	insert length
70  *	skip:	index to leave open
71  *
72  * Returns:
73  *	RET_ERROR, RET_SUCCESS
74  */
75 int
76 __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
77     size_t ilen, u_int32_t argskip)
78 {
79 	BINTERNAL *bi;
80 	BLEAF *bl, *tbl;
81 	DBT a, b;
82 	EPGNO *parent;
83 	PAGE *h, *l, *r, *lchild, *rchild;
84 	indx_t nxtindex;
85 	u_int16_t skip;
86 	u_int32_t n, nbytes, nksize;
87 	int parentsplit;
88 	char *dest;
89 
90 	/*
91 	 * Split the page into two pages, l and r.  The split routines return
92 	 * a pointer to the page into which the key should be inserted and with
93 	 * skip set to the offset which should be used.  Additionally, l and r
94 	 * are pinned.
95 	 */
96 	skip = argskip;
97 	h = sp->pgno == P_ROOT ?
98 	    bt_root(t, sp, &l, &r, &skip, ilen) :
99 	    bt_page(t, sp, &l, &r, &skip, ilen);
100 	if (h == NULL)
101 		return (RET_ERROR);
102 
103 	/*
104 	 * Insert the new key/data pair into the leaf page.  (Key inserts
105 	 * always cause a leaf page to split first.)
106 	 */
107 	h->linp[skip] = h->upper -= ilen;
108 	dest = (char *)h + h->upper;
109 	if (F_ISSET(t, R_RECNO))
110 		WR_RLEAF(dest, data, flags)
111 	else
112 		WR_BLEAF(dest, key, data, flags)
113 
114 	/* If the root page was split, make it look right. */
115 	if (sp->pgno == P_ROOT &&
116 	    (F_ISSET(t, R_RECNO) ?
117 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
118 		goto err2;
119 
120 	/*
121 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
122 	 * were traversed when we searched for the page that split.  Each stack
123 	 * entry is a page number and a page index offset.  The offset is for
124 	 * the page traversed on the search.  We've just split a page, so we
125 	 * have to insert a new key into the parent page.
126 	 *
127 	 * If the insert into the parent page causes it to split, may have to
128 	 * continue splitting all the way up the tree.  We stop if the root
129 	 * splits or the page inserted into didn't have to split to hold the
130 	 * new key.  Some algorithms replace the key for the old page as well
131 	 * as the new page.  We don't, as there's no reason to believe that the
132 	 * first key on the old page is any better than the key we have, and,
133 	 * in the case of a key being placed at index 0 causing the split, the
134 	 * key is unavailable.
135 	 *
136 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
137 	 * and right pages pinned while working on the parent.   The 5 are the
138 	 * two children, left parent and right parent (when the parent splits)
139 	 * and the root page or the overflow key page when calling bt_preserve.
140 	 * This code must make sure that all pins are released other than the
141 	 * root page or overflow page which is unlocked elsewhere.
142 	 */
143 	while ((parent = BT_POP(t)) != NULL) {
144 		lchild = l;
145 		rchild = r;
146 
147 		/* Get the parent page. */
148 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
149 			goto err2;
150 
151 		/*
152 		 * The new key goes ONE AFTER the index, because the split
153 		 * was to the right.
154 		 */
155 		skip = parent->index + 1;
156 
157 		/*
158 		 * Calculate the space needed on the parent page.
159 		 *
160 		 * Prefix trees: space hack when inserting into BINTERNAL
161 		 * pages.  Retain only what's needed to distinguish between
162 		 * the new entry and the LAST entry on the page to its left.
163 		 * If the keys compare equal, retain the entire key.  Note,
164 		 * we don't touch overflow keys, and the entire key must be
165 		 * retained for the next-to-left most key on the leftmost
166 		 * page of each level, or the search will fail.  Applicable
167 		 * ONLY to internal pages that have leaf pages as children.
168 		 * Further reduction of the key between pairs of internal
169 		 * pages loses too much information.
170 		 */
171 		switch (rchild->flags & P_TYPE) {
172 		case P_BINTERNAL:
173 			bi = GETBINTERNAL(rchild, 0);
174 			nbytes = NBINTERNAL(bi->ksize);
175 			break;
176 		case P_BLEAF:
177 			bl = GETBLEAF(rchild, 0);
178 			nbytes = NBINTERNAL(bl->ksize);
179 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
180 			    (h->prevpg != P_INVALID || skip > 1)) {
181 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
182 				a.size = tbl->ksize;
183 				a.data = tbl->bytes;
184 				b.size = bl->ksize;
185 				b.data = bl->bytes;
186 				nksize = t->bt_pfx(&a, &b);
187 				n = NBINTERNAL(nksize);
188 				if (n < nbytes) {
189 #ifdef STATISTICS
190 					bt_pfxsaved += nbytes - n;
191 #endif
192 					nbytes = n;
193 				} else
194 					nksize = 0;
195 			} else
196 				nksize = 0;
197 			break;
198 		case P_RINTERNAL:
199 		case P_RLEAF:
200 			nbytes = NRINTERNAL;
201 			break;
202 		default:
203 			abort();
204 		}
205 
206 		/* Split the parent page if necessary or shift the indices. */
207 		if ((u_int32_t)(h->upper - h->lower) < nbytes + sizeof(indx_t)) {
208 			sp = h;
209 			h = h->pgno == P_ROOT ?
210 			    bt_root(t, h, &l, &r, &skip, nbytes) :
211 			    bt_page(t, h, &l, &r, &skip, nbytes);
212 			if (h == NULL)
213 				goto err1;
214 			parentsplit = 1;
215 		} else {
216 			if (skip < (nxtindex = NEXTINDEX(h)))
217 				memmove(h->linp + skip + 1, h->linp + skip,
218 				    (nxtindex - skip) * sizeof(indx_t));
219 			h->lower += sizeof(indx_t);
220 			parentsplit = 0;
221 		}
222 
223 		/* Insert the key into the parent page. */
224 		switch (rchild->flags & P_TYPE) {
225 		case P_BINTERNAL:
226 			h->linp[skip] = h->upper -= nbytes;
227 			dest = (char *)h + h->linp[skip];
228 			memmove(dest, bi, nbytes);
229 			((BINTERNAL *)dest)->pgno = rchild->pgno;
230 			break;
231 		case P_BLEAF:
232 			h->linp[skip] = h->upper -= nbytes;
233 			dest = (char *)h + h->linp[skip];
234 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
235 			    rchild->pgno, bl->flags & P_BIGKEY);
236 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
237 			if (bl->flags & P_BIGKEY) {
238 				pgno_t pgno;
239 				memcpy(&pgno, bl->bytes, sizeof(pgno));
240 				if (bt_preserve(t, pgno) == RET_ERROR)
241 					goto err1;
242 			}
243 			break;
244 		case P_RINTERNAL:
245 			/*
246 			 * Update the left page count.  If split
247 			 * added at index 0, fix the correct page.
248 			 */
249 			if (skip > 0)
250 				dest = (char *)h + h->linp[skip - 1];
251 			else
252 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
253 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
254 			((RINTERNAL *)dest)->pgno = lchild->pgno;
255 
256 			/* Update the right page count. */
257 			h->linp[skip] = h->upper -= nbytes;
258 			dest = (char *)h + h->linp[skip];
259 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
260 			((RINTERNAL *)dest)->pgno = rchild->pgno;
261 			break;
262 		case P_RLEAF:
263 			/*
264 			 * Update the left page count.  If split
265 			 * added at index 0, fix the correct page.
266 			 */
267 			if (skip > 0)
268 				dest = (char *)h + h->linp[skip - 1];
269 			else
270 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
271 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
272 			((RINTERNAL *)dest)->pgno = lchild->pgno;
273 
274 			/* Update the right page count. */
275 			h->linp[skip] = h->upper -= nbytes;
276 			dest = (char *)h + h->linp[skip];
277 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
278 			((RINTERNAL *)dest)->pgno = rchild->pgno;
279 			break;
280 		default:
281 			abort();
282 		}
283 
284 		/* Unpin the held pages. */
285 		if (!parentsplit) {
286 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
287 			break;
288 		}
289 
290 		/* If the root page was split, make it look right. */
291 		if (sp->pgno == P_ROOT &&
292 		    (F_ISSET(t, R_RECNO) ?
293 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
294 			goto err1;
295 
296 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
297 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
298 	}
299 
300 	/* Unpin the held pages. */
301 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
302 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
303 
304 	/* Clear any pages left on the stack. */
305 	return (RET_SUCCESS);
306 
307 	/*
308 	 * If something fails in the above loop we were already walking back
309 	 * up the tree and the tree is now inconsistent.  Nothing much we can
310 	 * do about it but release any memory we're holding.
311 	 */
312 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
313 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
314 
315 err2:	mpool_put(t->bt_mp, l, 0);
316 	mpool_put(t->bt_mp, r, 0);
317 	__dbpanic(t->bt_dbp);
318 	return (RET_ERROR);
319 }
320 
321 /*
322  * BT_PAGE -- Split a non-root page of a btree.
323  *
324  * Parameters:
325  *	t:	tree
326  *	h:	root page
327  *	lp:	pointer to left page pointer
328  *	rp:	pointer to right page pointer
329  *	skip:	pointer to index to leave open
330  *	ilen:	insert length
331  *
332  * Returns:
333  *	Pointer to page in which to insert or NULL on error.
334  */
335 static PAGE *
336 bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
337 {
338 	PAGE *l, *r, *tp;
339 	pgno_t npg;
340 
341 #ifdef STATISTICS
342 	++bt_split;
343 #endif
344 	/* Put the new right page for the split into place. */
345 	if ((r = __bt_new(t, &npg)) == NULL)
346 		return (NULL);
347 	r->pgno = npg;
348 	r->lower = BTDATAOFF;
349 	r->upper = t->bt_psize;
350 	r->nextpg = h->nextpg;
351 	r->prevpg = h->pgno;
352 	r->flags = h->flags & P_TYPE;
353 
354 	/*
355 	 * If we're splitting the last page on a level because we're appending
356 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
357 	 * sorted.  Adding an empty page on the side of the level is less work
358 	 * and can push the fill factor much higher than normal.  If we're
359 	 * wrong it's no big deal, we'll just do the split the right way next
360 	 * time.  It may look like it's equally easy to do a similar hack for
361 	 * reverse sorted data, that is, split the tree left, but it's not.
362 	 * Don't even try.
363 	 */
364 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
365 #ifdef STATISTICS
366 		++bt_sortsplit;
367 #endif
368 		h->nextpg = r->pgno;
369 		r->lower = BTDATAOFF + sizeof(indx_t);
370 		*skip = 0;
371 		*lp = h;
372 		*rp = r;
373 		return (r);
374 	}
375 
376 	/* Put the new left page for the split into place. */
377 	if ((l = (PAGE *)calloc(1, t->bt_psize)) == NULL) {
378 		mpool_put(t->bt_mp, r, 0);
379 		return (NULL);
380 	}
381 	l->pgno = h->pgno;
382 	l->nextpg = r->pgno;
383 	l->prevpg = h->prevpg;
384 	l->lower = BTDATAOFF;
385 	l->upper = t->bt_psize;
386 	l->flags = h->flags & P_TYPE;
387 
388 	/* Fix up the previous pointer of the page after the split page. */
389 	if (h->nextpg != P_INVALID) {
390 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
391 			free(l);
392 			/* XXX mpool_free(t->bt_mp, r->pgno); */
393 			return (NULL);
394 		}
395 		tp->prevpg = r->pgno;
396 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
397 	}
398 
399 	/*
400 	 * Split right.  The key/data pairs aren't sorted in the btree page so
401 	 * it's simpler to copy the data from the split page onto two new pages
402 	 * instead of copying half the data to the right page and compacting
403 	 * the left page in place.  Since the left page can't change, we have
404 	 * to swap the original and the allocated left page after the split.
405 	 */
406 	tp = bt_psplit(t, h, l, r, skip, ilen);
407 
408 	/* Move the new left page onto the old left page. */
409 	memmove(h, l, t->bt_psize);
410 	if (tp == l)
411 		tp = h;
412 	free(l);
413 
414 	*lp = h;
415 	*rp = r;
416 	return (tp);
417 }
418 
419 /*
420  * BT_ROOT -- Split the root page of a btree.
421  *
422  * Parameters:
423  *	t:	tree
424  *	h:	root page
425  *	lp:	pointer to left page pointer
426  *	rp:	pointer to right page pointer
427  *	skip:	pointer to index to leave open
428  *	ilen:	insert length
429  *
430  * Returns:
431  *	Pointer to page in which to insert or NULL on error.
432  */
433 static PAGE *
434 bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
435 {
436 	PAGE *l, *r, *tp;
437 	pgno_t lnpg, rnpg;
438 
439 #ifdef STATISTICS
440 	++bt_split;
441 	++bt_rootsplit;
442 #endif
443 	/* Put the new left and right pages for the split into place. */
444 	if ((l = __bt_new(t, &lnpg)) == NULL ||
445 	    (r = __bt_new(t, &rnpg)) == NULL)
446 		return (NULL);
447 	l->pgno = lnpg;
448 	r->pgno = rnpg;
449 	l->nextpg = r->pgno;
450 	r->prevpg = l->pgno;
451 	l->prevpg = r->nextpg = P_INVALID;
452 	l->lower = r->lower = BTDATAOFF;
453 	l->upper = r->upper = t->bt_psize;
454 	l->flags = r->flags = h->flags & P_TYPE;
455 
456 	/* Split the root page. */
457 	tp = bt_psplit(t, h, l, r, skip, ilen);
458 
459 	*lp = l;
460 	*rp = r;
461 	return (tp);
462 }
463 
464 /*
465  * BT_RROOT -- Fix up the recno root page after it has been split.
466  *
467  * Parameters:
468  *	t:	tree
469  *	h:	root page
470  *	l:	left page
471  *	r:	right page
472  *
473  * Returns:
474  *	RET_ERROR, RET_SUCCESS
475  */
476 static int
477 bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
478 {
479 	char *dest;
480 
481 	/* Insert the left and right keys, set the header information. */
482 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
483 	dest = (char *)h + h->upper;
484 	WR_RINTERNAL(dest,
485 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
486 
487 	__PAST_END(h->linp, 1) = h->upper -= NRINTERNAL;
488 	dest = (char *)h + h->upper;
489 	WR_RINTERNAL(dest,
490 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
491 
492 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
493 
494 	/* Unpin the root page, set to recno internal page. */
495 	h->flags &= ~P_TYPE;
496 	h->flags |= P_RINTERNAL;
497 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
498 
499 	return (RET_SUCCESS);
500 }
501 
502 /*
503  * BT_BROOT -- Fix up the btree root page after it has been split.
504  *
505  * Parameters:
506  *	t:	tree
507  *	h:	root page
508  *	l:	left page
509  *	r:	right page
510  *
511  * Returns:
512  *	RET_ERROR, RET_SUCCESS
513  */
514 static int
515 bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
516 {
517 	BINTERNAL *bi;
518 	BLEAF *bl;
519 	u_int32_t nbytes;
520 	char *dest;
521 
522 	/*
523 	 * If the root page was a leaf page, change it into an internal page.
524 	 * We copy the key we split on (but not the key's data, in the case of
525 	 * a leaf page) to the new root page.
526 	 *
527 	 * The btree comparison code guarantees that the left-most key on any
528 	 * level of the tree is never used, so it doesn't need to be filled in.
529 	 */
530 	nbytes = NBINTERNAL(0);
531 	h->linp[0] = h->upper = t->bt_psize - nbytes;
532 	dest = (char *)h + h->upper;
533 	WR_BINTERNAL(dest, 0, l->pgno, 0);
534 
535 	switch (h->flags & P_TYPE) {
536 	case P_BLEAF:
537 		bl = GETBLEAF(r, 0);
538 		nbytes = NBINTERNAL(bl->ksize);
539 		__PAST_END(h->linp, 1) = h->upper -= nbytes;
540 		dest = (char *)h + h->upper;
541 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
542 		memmove(dest, bl->bytes, bl->ksize);
543 
544 		/*
545 		 * If the key is on an overflow page, mark the overflow chain
546 		 * so it isn't deleted when the leaf copy of the key is deleted.
547 		 */
548 	if (bl->flags & P_BIGKEY) {
549 			pgno_t pgno;
550 			memcpy(&pgno, bl->bytes, sizeof(pgno));
551 			if (bt_preserve(t, pgno) == RET_ERROR)
552 				return (RET_ERROR);
553 		}
554 		break;
555 	case P_BINTERNAL:
556 		bi = GETBINTERNAL(r, 0);
557 		nbytes = NBINTERNAL(bi->ksize);
558 		__PAST_END(h->linp, 1) = h->upper -= nbytes;
559 		dest = (char *)h + h->upper;
560 		memmove(dest, bi, nbytes);
561 		((BINTERNAL *)dest)->pgno = r->pgno;
562 		break;
563 	default:
564 		abort();
565 	}
566 
567 	/* There are two keys on the page. */
568 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
569 
570 	/* Unpin the root page, set to btree internal page. */
571 	h->flags &= ~P_TYPE;
572 	h->flags |= P_BINTERNAL;
573 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
574 
575 	return (RET_SUCCESS);
576 }
577 
578 /*
579  * BT_PSPLIT -- Do the real work of splitting the page.
580  *
581  * Parameters:
582  *	t:	tree
583  *	h:	page to be split
584  *	l:	page to put lower half of data
585  *	r:	page to put upper half of data
586  *	pskip:	pointer to index to leave open
587  *	ilen:	insert length
588  *
589  * Returns:
590  *	Pointer to page in which to insert.
591  */
592 static PAGE *
593 bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
594 {
595 	BINTERNAL *bi;
596 	BLEAF *bl;
597 	CURSOR *c;
598 	RLEAF *rl;
599 	PAGE *rval;
600 	void *src;
601 	indx_t full, half, nxt, off, skip, top, used;
602 	u_int32_t nbytes;
603 	int bigkeycnt, isbigkey;
604 
605 	/*
606 	 * Split the data to the left and right pages.  Leave the skip index
607 	 * open.  Additionally, make some effort not to split on an overflow
608 	 * key.  This makes internal page processing faster and can save
609 	 * space as overflow keys used by internal pages are never deleted.
610 	 */
611 	bigkeycnt = 0;
612 	skip = *pskip;
613 	full = t->bt_psize - BTDATAOFF;
614 	half = full / 2;
615 	used = 0;
616 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
617 		if (skip == off) {
618 			nbytes = ilen;
619 			isbigkey = 0;		/* XXX: not really known. */
620 		} else
621 			switch (h->flags & P_TYPE) {
622 			case P_BINTERNAL:
623 				src = bi = GETBINTERNAL(h, nxt);
624 				nbytes = NBINTERNAL(bi->ksize);
625 				isbigkey = bi->flags & P_BIGKEY;
626 				break;
627 			case P_BLEAF:
628 				src = bl = GETBLEAF(h, nxt);
629 				nbytes = NBLEAF(bl);
630 				isbigkey = bl->flags & P_BIGKEY;
631 				break;
632 			case P_RINTERNAL:
633 				src = GETRINTERNAL(h, nxt);
634 				nbytes = NRINTERNAL;
635 				isbigkey = 0;
636 				break;
637 			case P_RLEAF:
638 				src = rl = GETRLEAF(h, nxt);
639 				nbytes = NRLEAF(rl);
640 				isbigkey = 0;
641 				break;
642 			default:
643 				abort();
644 			}
645 
646 		/*
647 		 * If the key/data pairs are substantial fractions of the max
648 		 * possible size for the page, it's possible to get situations
649 		 * where we decide to try and copy too much onto the left page.
650 		 * Make sure that doesn't happen.
651 		 */
652 		if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
653 		    nxt == top - 1) {
654 			--off;
655 			break;
656 		}
657 
658 		/* Copy the key/data pair, if not the skipped index. */
659 		if (skip != off) {
660 			++nxt;
661 
662 			l->linp[off] = l->upper -= nbytes;
663 			memmove((char *)l + l->upper, src, nbytes);
664 		}
665 
666 		used += nbytes + sizeof(indx_t);
667 		if (used >= half) {
668 			if (!isbigkey || bigkeycnt == 3)
669 				break;
670 			else
671 				++bigkeycnt;
672 		}
673 	}
674 
675 	/*
676 	 * Off is the last offset that's valid for the left page.
677 	 * Nxt is the first offset to be placed on the right page.
678 	 */
679 	l->lower += (off + 1) * sizeof(indx_t);
680 
681 	/*
682 	 * If splitting the page that the cursor was on, the cursor has to be
683 	 * adjusted to point to the same record as before the split.  If the
684 	 * cursor is at or past the skipped slot, the cursor is incremented by
685 	 * one.  If the cursor is on the right page, it is decremented by the
686 	 * number of records split to the left page.
687 	 */
688 	c = &t->bt_cursor;
689 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
690 		if (c->pg.index >= skip)
691 			++c->pg.index;
692 		if (c->pg.index < nxt)			/* Left page. */
693 			c->pg.pgno = l->pgno;
694 		else {					/* Right page. */
695 			c->pg.pgno = r->pgno;
696 			c->pg.index -= nxt;
697 		}
698 	}
699 
700 	/*
701 	 * If the skipped index was on the left page, just return that page.
702 	 * Otherwise, adjust the skip index to reflect the new position on
703 	 * the right page.
704 	 */
705 	if (skip <= off) {
706 		skip = MAX_PAGE_OFFSET;
707 		rval = l;
708 	} else {
709 		rval = r;
710 		*pskip -= nxt;
711 	}
712 
713 	for (off = 0; nxt < top; ++off) {
714 		if (skip == nxt) {
715 			++off;
716 			skip = MAX_PAGE_OFFSET;
717 		}
718 		switch (h->flags & P_TYPE) {
719 		case P_BINTERNAL:
720 			src = bi = GETBINTERNAL(h, nxt);
721 			nbytes = NBINTERNAL(bi->ksize);
722 			break;
723 		case P_BLEAF:
724 			src = bl = GETBLEAF(h, nxt);
725 			nbytes = NBLEAF(bl);
726 			break;
727 		case P_RINTERNAL:
728 			src = GETRINTERNAL(h, nxt);
729 			nbytes = NRINTERNAL;
730 			break;
731 		case P_RLEAF:
732 			src = rl = GETRLEAF(h, nxt);
733 			nbytes = NRLEAF(rl);
734 			break;
735 		default:
736 			abort();
737 		}
738 		++nxt;
739 		r->linp[off] = r->upper -= nbytes;
740 		memmove((char *)r + r->upper, src, nbytes);
741 	}
742 	r->lower += off * sizeof(indx_t);
743 
744 	/* If the key is being appended to the page, adjust the index. */
745 	if (skip == top)
746 		r->lower += sizeof(indx_t);
747 
748 	return (rval);
749 }
750 
751 /*
752  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
753  *
754  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
755  * record that references them gets deleted.  Chains pointed to by internal
756  * pages never get deleted.  This routine marks a chain as pointed to by an
757  * internal page.
758  *
759  * Parameters:
760  *	t:	tree
761  *	pg:	page number of first page in the chain.
762  *
763  * Returns:
764  *	RET_SUCCESS, RET_ERROR.
765  */
766 static int
767 bt_preserve(BTREE *t, pgno_t pg)
768 {
769 	PAGE *h;
770 
771 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
772 		return (RET_ERROR);
773 	h->flags |= P_PRESERVE;
774 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
775 	return (RET_SUCCESS);
776 }
777 
778 /*
779  * REC_TOTAL -- Return the number of recno entries below a page.
780  *
781  * Parameters:
782  *	h:	page
783  *
784  * Returns:
785  *	The number of recno entries below a page.
786  *
787  * XXX
788  * These values could be set by the bt_psplit routine.  The problem is that the
789  * entry has to be popped off of the stack etc. or the values have to be passed
790  * all the way back to bt_split/bt_rroot and it's not very clean.
791  */
792 static recno_t
793 rec_total(PAGE *h)
794 {
795 	recno_t recs;
796 	indx_t nxt, top;
797 
798 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
799 		recs += GETRINTERNAL(h, nxt)->nrecs;
800 	return (recs);
801 }
802