xref: /freebsd/lib/libc/db/btree/bt_split.c (revision 273c26a3c3bea87a241d6879abd4f991db180bf0)
1 /*-
2  * Copyright (c) 1990, 1993, 1994
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * Mike Olson.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #if defined(LIBC_SCCS) && !defined(lint)
34 static char sccsid[] = "@(#)bt_split.c	8.10 (Berkeley) 1/9/95";
35 #endif /* LIBC_SCCS and not lint */
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include <sys/param.h>
40 
41 #include <limits.h>
42 #include <stdio.h>
43 #include <stdlib.h>
44 #include <string.h>
45 
46 #include <db.h>
47 #include "btree.h"
48 
49 static int	 bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
50 static PAGE	*bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
51 static int	 bt_preserve(BTREE *, pgno_t);
52 static PAGE	*bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
53 static PAGE	*bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
54 static int	 bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
55 static recno_t	 rec_total(PAGE *);
56 
57 #ifdef STATISTICS
58 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
59 #endif
60 
61 /*
62  * __BT_SPLIT -- Split the tree.
63  *
64  * Parameters:
65  *	t:	tree
66  *	sp:	page to split
67  *	key:	key to insert
68  *	data:	data to insert
69  *	flags:	BIGKEY/BIGDATA flags
70  *	ilen:	insert length
71  *	skip:	index to leave open
72  *
73  * Returns:
74  *	RET_ERROR, RET_SUCCESS
75  */
76 int
77 __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
78     size_t ilen, u_int32_t argskip)
79 {
80 	BINTERNAL *bi;
81 	BLEAF *bl, *tbl;
82 	DBT a, b;
83 	EPGNO *parent;
84 	PAGE *h, *l, *r, *lchild, *rchild;
85 	indx_t nxtindex;
86 	u_int16_t skip;
87 	u_int32_t n, nbytes, nksize;
88 	int parentsplit;
89 	char *dest;
90 
91 	/*
92 	 * Split the page into two pages, l and r.  The split routines return
93 	 * a pointer to the page into which the key should be inserted and with
94 	 * skip set to the offset which should be used.  Additionally, l and r
95 	 * are pinned.
96 	 */
97 	skip = argskip;
98 	h = sp->pgno == P_ROOT ?
99 	    bt_root(t, sp, &l, &r, &skip, ilen) :
100 	    bt_page(t, sp, &l, &r, &skip, ilen);
101 	if (h == NULL)
102 		return (RET_ERROR);
103 
104 	/*
105 	 * Insert the new key/data pair into the leaf page.  (Key inserts
106 	 * always cause a leaf page to split first.)
107 	 */
108 	h->linp[skip] = h->upper -= ilen;
109 	dest = (char *)h + h->upper;
110 	if (F_ISSET(t, R_RECNO))
111 		WR_RLEAF(dest, data, flags)
112 	else
113 		WR_BLEAF(dest, key, data, flags)
114 
115 	/* If the root page was split, make it look right. */
116 	if (sp->pgno == P_ROOT &&
117 	    (F_ISSET(t, R_RECNO) ?
118 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
119 		goto err2;
120 
121 	/*
122 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
123 	 * were traversed when we searched for the page that split.  Each stack
124 	 * entry is a page number and a page index offset.  The offset is for
125 	 * the page traversed on the search.  We've just split a page, so we
126 	 * have to insert a new key into the parent page.
127 	 *
128 	 * If the insert into the parent page causes it to split, may have to
129 	 * continue splitting all the way up the tree.  We stop if the root
130 	 * splits or the page inserted into didn't have to split to hold the
131 	 * new key.  Some algorithms replace the key for the old page as well
132 	 * as the new page.  We don't, as there's no reason to believe that the
133 	 * first key on the old page is any better than the key we have, and,
134 	 * in the case of a key being placed at index 0 causing the split, the
135 	 * key is unavailable.
136 	 *
137 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
138 	 * and right pages pinned while working on the parent.   The 5 are the
139 	 * two children, left parent and right parent (when the parent splits)
140 	 * and the root page or the overflow key page when calling bt_preserve.
141 	 * This code must make sure that all pins are released other than the
142 	 * root page or overflow page which is unlocked elsewhere.
143 	 */
144 	while ((parent = BT_POP(t)) != NULL) {
145 		lchild = l;
146 		rchild = r;
147 
148 		/* Get the parent page. */
149 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
150 			goto err2;
151 
152 		/*
153 		 * The new key goes ONE AFTER the index, because the split
154 		 * was to the right.
155 		 */
156 		skip = parent->index + 1;
157 
158 		/*
159 		 * Calculate the space needed on the parent page.
160 		 *
161 		 * Prefix trees: space hack when inserting into BINTERNAL
162 		 * pages.  Retain only what's needed to distinguish between
163 		 * the new entry and the LAST entry on the page to its left.
164 		 * If the keys compare equal, retain the entire key.  Note,
165 		 * we don't touch overflow keys, and the entire key must be
166 		 * retained for the next-to-left most key on the leftmost
167 		 * page of each level, or the search will fail.  Applicable
168 		 * ONLY to internal pages that have leaf pages as children.
169 		 * Further reduction of the key between pairs of internal
170 		 * pages loses too much information.
171 		 */
172 		switch (rchild->flags & P_TYPE) {
173 		case P_BINTERNAL:
174 			bi = GETBINTERNAL(rchild, 0);
175 			nbytes = NBINTERNAL(bi->ksize);
176 			break;
177 		case P_BLEAF:
178 			bl = GETBLEAF(rchild, 0);
179 			nbytes = NBINTERNAL(bl->ksize);
180 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
181 			    (h->prevpg != P_INVALID || skip > 1)) {
182 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
183 				a.size = tbl->ksize;
184 				a.data = tbl->bytes;
185 				b.size = bl->ksize;
186 				b.data = bl->bytes;
187 				nksize = t->bt_pfx(&a, &b);
188 				n = NBINTERNAL(nksize);
189 				if (n < nbytes) {
190 #ifdef STATISTICS
191 					bt_pfxsaved += nbytes - n;
192 #endif
193 					nbytes = n;
194 				} else
195 					nksize = 0;
196 			} else
197 				nksize = 0;
198 			break;
199 		case P_RINTERNAL:
200 		case P_RLEAF:
201 			nbytes = NRINTERNAL;
202 			break;
203 		default:
204 			abort();
205 		}
206 
207 		/* Split the parent page if necessary or shift the indices. */
208 		if ((u_int32_t)(h->upper - h->lower) < nbytes + sizeof(indx_t)) {
209 			sp = h;
210 			h = h->pgno == P_ROOT ?
211 			    bt_root(t, h, &l, &r, &skip, nbytes) :
212 			    bt_page(t, h, &l, &r, &skip, nbytes);
213 			if (h == NULL)
214 				goto err1;
215 			parentsplit = 1;
216 		} else {
217 			if (skip < (nxtindex = NEXTINDEX(h)))
218 				memmove(h->linp + skip + 1, h->linp + skip,
219 				    (nxtindex - skip) * sizeof(indx_t));
220 			h->lower += sizeof(indx_t);
221 			parentsplit = 0;
222 		}
223 
224 		/* Insert the key into the parent page. */
225 		switch (rchild->flags & P_TYPE) {
226 		case P_BINTERNAL:
227 			h->linp[skip] = h->upper -= nbytes;
228 			dest = (char *)h + h->linp[skip];
229 			memmove(dest, bi, nbytes);
230 			((BINTERNAL *)dest)->pgno = rchild->pgno;
231 			break;
232 		case P_BLEAF:
233 			h->linp[skip] = h->upper -= nbytes;
234 			dest = (char *)h + h->linp[skip];
235 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
236 			    rchild->pgno, bl->flags & P_BIGKEY);
237 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
238 			if (bl->flags & P_BIGKEY) {
239 				pgno_t pgno;
240 				memcpy(&pgno, bl->bytes, sizeof(pgno));
241 				if (bt_preserve(t, pgno) == RET_ERROR)
242 					goto err1;
243 			}
244 			break;
245 		case P_RINTERNAL:
246 			/*
247 			 * Update the left page count.  If split
248 			 * added at index 0, fix the correct page.
249 			 */
250 			if (skip > 0)
251 				dest = (char *)h + h->linp[skip - 1];
252 			else
253 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
254 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
255 			((RINTERNAL *)dest)->pgno = lchild->pgno;
256 
257 			/* Update the right page count. */
258 			h->linp[skip] = h->upper -= nbytes;
259 			dest = (char *)h + h->linp[skip];
260 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
261 			((RINTERNAL *)dest)->pgno = rchild->pgno;
262 			break;
263 		case P_RLEAF:
264 			/*
265 			 * Update the left page count.  If split
266 			 * added at index 0, fix the correct page.
267 			 */
268 			if (skip > 0)
269 				dest = (char *)h + h->linp[skip - 1];
270 			else
271 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
272 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
273 			((RINTERNAL *)dest)->pgno = lchild->pgno;
274 
275 			/* Update the right page count. */
276 			h->linp[skip] = h->upper -= nbytes;
277 			dest = (char *)h + h->linp[skip];
278 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
279 			((RINTERNAL *)dest)->pgno = rchild->pgno;
280 			break;
281 		default:
282 			abort();
283 		}
284 
285 		/* Unpin the held pages. */
286 		if (!parentsplit) {
287 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
288 			break;
289 		}
290 
291 		/* If the root page was split, make it look right. */
292 		if (sp->pgno == P_ROOT &&
293 		    (F_ISSET(t, R_RECNO) ?
294 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
295 			goto err1;
296 
297 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
298 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
299 	}
300 
301 	/* Unpin the held pages. */
302 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
303 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
304 
305 	/* Clear any pages left on the stack. */
306 	return (RET_SUCCESS);
307 
308 	/*
309 	 * If something fails in the above loop we were already walking back
310 	 * up the tree and the tree is now inconsistent.  Nothing much we can
311 	 * do about it but release any memory we're holding.
312 	 */
313 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
314 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
315 
316 err2:	mpool_put(t->bt_mp, l, 0);
317 	mpool_put(t->bt_mp, r, 0);
318 	__dbpanic(t->bt_dbp);
319 	return (RET_ERROR);
320 }
321 
322 /*
323  * BT_PAGE -- Split a non-root page of a btree.
324  *
325  * Parameters:
326  *	t:	tree
327  *	h:	root page
328  *	lp:	pointer to left page pointer
329  *	rp:	pointer to right page pointer
330  *	skip:	pointer to index to leave open
331  *	ilen:	insert length
332  *
333  * Returns:
334  *	Pointer to page in which to insert or NULL on error.
335  */
336 static PAGE *
337 bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
338 {
339 	PAGE *l, *r, *tp;
340 	pgno_t npg;
341 
342 #ifdef STATISTICS
343 	++bt_split;
344 #endif
345 	/* Put the new right page for the split into place. */
346 	if ((r = __bt_new(t, &npg)) == NULL)
347 		return (NULL);
348 	r->pgno = npg;
349 	r->lower = BTDATAOFF;
350 	r->upper = t->bt_psize;
351 	r->nextpg = h->nextpg;
352 	r->prevpg = h->pgno;
353 	r->flags = h->flags & P_TYPE;
354 
355 	/*
356 	 * If we're splitting the last page on a level because we're appending
357 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
358 	 * sorted.  Adding an empty page on the side of the level is less work
359 	 * and can push the fill factor much higher than normal.  If we're
360 	 * wrong it's no big deal, we'll just do the split the right way next
361 	 * time.  It may look like it's equally easy to do a similar hack for
362 	 * reverse sorted data, that is, split the tree left, but it's not.
363 	 * Don't even try.
364 	 */
365 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
366 #ifdef STATISTICS
367 		++bt_sortsplit;
368 #endif
369 		h->nextpg = r->pgno;
370 		r->lower = BTDATAOFF + sizeof(indx_t);
371 		*skip = 0;
372 		*lp = h;
373 		*rp = r;
374 		return (r);
375 	}
376 
377 	/* Put the new left page for the split into place. */
378 	if ((l = (PAGE *)calloc(1, t->bt_psize)) == NULL) {
379 		mpool_put(t->bt_mp, r, 0);
380 		return (NULL);
381 	}
382 	l->pgno = h->pgno;
383 	l->nextpg = r->pgno;
384 	l->prevpg = h->prevpg;
385 	l->lower = BTDATAOFF;
386 	l->upper = t->bt_psize;
387 	l->flags = h->flags & P_TYPE;
388 
389 	/* Fix up the previous pointer of the page after the split page. */
390 	if (h->nextpg != P_INVALID) {
391 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
392 			free(l);
393 			/* XXX mpool_free(t->bt_mp, r->pgno); */
394 			return (NULL);
395 		}
396 		tp->prevpg = r->pgno;
397 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
398 	}
399 
400 	/*
401 	 * Split right.  The key/data pairs aren't sorted in the btree page so
402 	 * it's simpler to copy the data from the split page onto two new pages
403 	 * instead of copying half the data to the right page and compacting
404 	 * the left page in place.  Since the left page can't change, we have
405 	 * to swap the original and the allocated left page after the split.
406 	 */
407 	tp = bt_psplit(t, h, l, r, skip, ilen);
408 
409 	/* Move the new left page onto the old left page. */
410 	memmove(h, l, t->bt_psize);
411 	if (tp == l)
412 		tp = h;
413 	free(l);
414 
415 	*lp = h;
416 	*rp = r;
417 	return (tp);
418 }
419 
420 /*
421  * BT_ROOT -- Split the root page of a btree.
422  *
423  * Parameters:
424  *	t:	tree
425  *	h:	root page
426  *	lp:	pointer to left page pointer
427  *	rp:	pointer to right page pointer
428  *	skip:	pointer to index to leave open
429  *	ilen:	insert length
430  *
431  * Returns:
432  *	Pointer to page in which to insert or NULL on error.
433  */
434 static PAGE *
435 bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
436 {
437 	PAGE *l, *r, *tp;
438 	pgno_t lnpg, rnpg;
439 
440 #ifdef STATISTICS
441 	++bt_split;
442 	++bt_rootsplit;
443 #endif
444 	/* Put the new left and right pages for the split into place. */
445 	if ((l = __bt_new(t, &lnpg)) == NULL ||
446 	    (r = __bt_new(t, &rnpg)) == NULL)
447 		return (NULL);
448 	l->pgno = lnpg;
449 	r->pgno = rnpg;
450 	l->nextpg = r->pgno;
451 	r->prevpg = l->pgno;
452 	l->prevpg = r->nextpg = P_INVALID;
453 	l->lower = r->lower = BTDATAOFF;
454 	l->upper = r->upper = t->bt_psize;
455 	l->flags = r->flags = h->flags & P_TYPE;
456 
457 	/* Split the root page. */
458 	tp = bt_psplit(t, h, l, r, skip, ilen);
459 
460 	*lp = l;
461 	*rp = r;
462 	return (tp);
463 }
464 
465 /*
466  * BT_RROOT -- Fix up the recno root page after it has been split.
467  *
468  * Parameters:
469  *	t:	tree
470  *	h:	root page
471  *	l:	left page
472  *	r:	right page
473  *
474  * Returns:
475  *	RET_ERROR, RET_SUCCESS
476  */
477 static int
478 bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
479 {
480 	char *dest;
481 
482 	/* Insert the left and right keys, set the header information. */
483 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
484 	dest = (char *)h + h->upper;
485 	WR_RINTERNAL(dest,
486 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
487 
488 	__PAST_END(h->linp, 1) = h->upper -= NRINTERNAL;
489 	dest = (char *)h + h->upper;
490 	WR_RINTERNAL(dest,
491 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
492 
493 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
494 
495 	/* Unpin the root page, set to recno internal page. */
496 	h->flags &= ~P_TYPE;
497 	h->flags |= P_RINTERNAL;
498 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
499 
500 	return (RET_SUCCESS);
501 }
502 
503 /*
504  * BT_BROOT -- Fix up the btree root page after it has been split.
505  *
506  * Parameters:
507  *	t:	tree
508  *	h:	root page
509  *	l:	left page
510  *	r:	right page
511  *
512  * Returns:
513  *	RET_ERROR, RET_SUCCESS
514  */
515 static int
516 bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
517 {
518 	BINTERNAL *bi;
519 	BLEAF *bl;
520 	u_int32_t nbytes;
521 	char *dest;
522 
523 	/*
524 	 * If the root page was a leaf page, change it into an internal page.
525 	 * We copy the key we split on (but not the key's data, in the case of
526 	 * a leaf page) to the new root page.
527 	 *
528 	 * The btree comparison code guarantees that the left-most key on any
529 	 * level of the tree is never used, so it doesn't need to be filled in.
530 	 */
531 	nbytes = NBINTERNAL(0);
532 	h->linp[0] = h->upper = t->bt_psize - nbytes;
533 	dest = (char *)h + h->upper;
534 	WR_BINTERNAL(dest, 0, l->pgno, 0);
535 
536 	switch (h->flags & P_TYPE) {
537 	case P_BLEAF:
538 		bl = GETBLEAF(r, 0);
539 		nbytes = NBINTERNAL(bl->ksize);
540 		__PAST_END(h->linp, 1) = h->upper -= nbytes;
541 		dest = (char *)h + h->upper;
542 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
543 		memmove(dest, bl->bytes, bl->ksize);
544 
545 		/*
546 		 * If the key is on an overflow page, mark the overflow chain
547 		 * so it isn't deleted when the leaf copy of the key is deleted.
548 		 */
549 	if (bl->flags & P_BIGKEY) {
550 			pgno_t pgno;
551 			memcpy(&pgno, bl->bytes, sizeof(pgno));
552 			if (bt_preserve(t, pgno) == RET_ERROR)
553 				return (RET_ERROR);
554 		}
555 		break;
556 	case P_BINTERNAL:
557 		bi = GETBINTERNAL(r, 0);
558 		nbytes = NBINTERNAL(bi->ksize);
559 		__PAST_END(h->linp, 1) = h->upper -= nbytes;
560 		dest = (char *)h + h->upper;
561 		memmove(dest, bi, nbytes);
562 		((BINTERNAL *)dest)->pgno = r->pgno;
563 		break;
564 	default:
565 		abort();
566 	}
567 
568 	/* There are two keys on the page. */
569 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
570 
571 	/* Unpin the root page, set to btree internal page. */
572 	h->flags &= ~P_TYPE;
573 	h->flags |= P_BINTERNAL;
574 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
575 
576 	return (RET_SUCCESS);
577 }
578 
579 /*
580  * BT_PSPLIT -- Do the real work of splitting the page.
581  *
582  * Parameters:
583  *	t:	tree
584  *	h:	page to be split
585  *	l:	page to put lower half of data
586  *	r:	page to put upper half of data
587  *	pskip:	pointer to index to leave open
588  *	ilen:	insert length
589  *
590  * Returns:
591  *	Pointer to page in which to insert.
592  */
593 static PAGE *
594 bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
595 {
596 	BINTERNAL *bi;
597 	BLEAF *bl;
598 	CURSOR *c;
599 	RLEAF *rl;
600 	PAGE *rval;
601 	void *src;
602 	indx_t full, half, nxt, off, skip, top, used;
603 	u_int32_t nbytes;
604 	int bigkeycnt, isbigkey;
605 
606 	/*
607 	 * Split the data to the left and right pages.  Leave the skip index
608 	 * open.  Additionally, make some effort not to split on an overflow
609 	 * key.  This makes internal page processing faster and can save
610 	 * space as overflow keys used by internal pages are never deleted.
611 	 */
612 	bigkeycnt = 0;
613 	skip = *pskip;
614 	full = t->bt_psize - BTDATAOFF;
615 	half = full / 2;
616 	used = 0;
617 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
618 		if (skip == off) {
619 			nbytes = ilen;
620 			isbigkey = 0;		/* XXX: not really known. */
621 		} else
622 			switch (h->flags & P_TYPE) {
623 			case P_BINTERNAL:
624 				src = bi = GETBINTERNAL(h, nxt);
625 				nbytes = NBINTERNAL(bi->ksize);
626 				isbigkey = bi->flags & P_BIGKEY;
627 				break;
628 			case P_BLEAF:
629 				src = bl = GETBLEAF(h, nxt);
630 				nbytes = NBLEAF(bl);
631 				isbigkey = bl->flags & P_BIGKEY;
632 				break;
633 			case P_RINTERNAL:
634 				src = GETRINTERNAL(h, nxt);
635 				nbytes = NRINTERNAL;
636 				isbigkey = 0;
637 				break;
638 			case P_RLEAF:
639 				src = rl = GETRLEAF(h, nxt);
640 				nbytes = NRLEAF(rl);
641 				isbigkey = 0;
642 				break;
643 			default:
644 				abort();
645 			}
646 
647 		/*
648 		 * If the key/data pairs are substantial fractions of the max
649 		 * possible size for the page, it's possible to get situations
650 		 * where we decide to try and copy too much onto the left page.
651 		 * Make sure that doesn't happen.
652 		 */
653 		if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
654 		    nxt == top - 1) {
655 			--off;
656 			break;
657 		}
658 
659 		/* Copy the key/data pair, if not the skipped index. */
660 		if (skip != off) {
661 			++nxt;
662 
663 			l->linp[off] = l->upper -= nbytes;
664 			memmove((char *)l + l->upper, src, nbytes);
665 		}
666 
667 		used += nbytes + sizeof(indx_t);
668 		if (used >= half) {
669 			if (!isbigkey || bigkeycnt == 3)
670 				break;
671 			else
672 				++bigkeycnt;
673 		}
674 	}
675 
676 	/*
677 	 * Off is the last offset that's valid for the left page.
678 	 * Nxt is the first offset to be placed on the right page.
679 	 */
680 	l->lower += (off + 1) * sizeof(indx_t);
681 
682 	/*
683 	 * If splitting the page that the cursor was on, the cursor has to be
684 	 * adjusted to point to the same record as before the split.  If the
685 	 * cursor is at or past the skipped slot, the cursor is incremented by
686 	 * one.  If the cursor is on the right page, it is decremented by the
687 	 * number of records split to the left page.
688 	 */
689 	c = &t->bt_cursor;
690 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
691 		if (c->pg.index >= skip)
692 			++c->pg.index;
693 		if (c->pg.index < nxt)			/* Left page. */
694 			c->pg.pgno = l->pgno;
695 		else {					/* Right page. */
696 			c->pg.pgno = r->pgno;
697 			c->pg.index -= nxt;
698 		}
699 	}
700 
701 	/*
702 	 * If the skipped index was on the left page, just return that page.
703 	 * Otherwise, adjust the skip index to reflect the new position on
704 	 * the right page.
705 	 */
706 	if (skip <= off) {
707 		skip = MAX_PAGE_OFFSET;
708 		rval = l;
709 	} else {
710 		rval = r;
711 		*pskip -= nxt;
712 	}
713 
714 	for (off = 0; nxt < top; ++off) {
715 		if (skip == nxt) {
716 			++off;
717 			skip = MAX_PAGE_OFFSET;
718 		}
719 		switch (h->flags & P_TYPE) {
720 		case P_BINTERNAL:
721 			src = bi = GETBINTERNAL(h, nxt);
722 			nbytes = NBINTERNAL(bi->ksize);
723 			break;
724 		case P_BLEAF:
725 			src = bl = GETBLEAF(h, nxt);
726 			nbytes = NBLEAF(bl);
727 			break;
728 		case P_RINTERNAL:
729 			src = GETRINTERNAL(h, nxt);
730 			nbytes = NRINTERNAL;
731 			break;
732 		case P_RLEAF:
733 			src = rl = GETRLEAF(h, nxt);
734 			nbytes = NRLEAF(rl);
735 			break;
736 		default:
737 			abort();
738 		}
739 		++nxt;
740 		r->linp[off] = r->upper -= nbytes;
741 		memmove((char *)r + r->upper, src, nbytes);
742 	}
743 	r->lower += off * sizeof(indx_t);
744 
745 	/* If the key is being appended to the page, adjust the index. */
746 	if (skip == top)
747 		r->lower += sizeof(indx_t);
748 
749 	return (rval);
750 }
751 
752 /*
753  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
754  *
755  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
756  * record that references them gets deleted.  Chains pointed to by internal
757  * pages never get deleted.  This routine marks a chain as pointed to by an
758  * internal page.
759  *
760  * Parameters:
761  *	t:	tree
762  *	pg:	page number of first page in the chain.
763  *
764  * Returns:
765  *	RET_SUCCESS, RET_ERROR.
766  */
767 static int
768 bt_preserve(BTREE *t, pgno_t pg)
769 {
770 	PAGE *h;
771 
772 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
773 		return (RET_ERROR);
774 	h->flags |= P_PRESERVE;
775 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
776 	return (RET_SUCCESS);
777 }
778 
779 /*
780  * REC_TOTAL -- Return the number of recno entries below a page.
781  *
782  * Parameters:
783  *	h:	page
784  *
785  * Returns:
786  *	The number of recno entries below a page.
787  *
788  * XXX
789  * These values could be set by the bt_psplit routine.  The problem is that the
790  * entry has to be popped off of the stack etc. or the values have to be passed
791  * all the way back to bt_split/bt_rroot and it's not very clean.
792  */
793 static recno_t
794 rec_total(PAGE *h)
795 {
796 	recno_t recs;
797 	indx_t nxt, top;
798 
799 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
800 		recs += GETRINTERNAL(h, nxt)->nrecs;
801 	return (recs);
802 }
803