xref: /dragonfly/lib/libc/db/btree/bt_split.c (revision 1de703da)
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  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by the University of
19  *	California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  * @(#)bt_split.c	8.9 (Berkeley) 7/26/94
37  */
38 
39 #include <sys/types.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 __P((BTREE *, PAGE *, PAGE *, PAGE *));
50 static PAGE	*bt_page
51 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
52 static int	 bt_preserve __P((BTREE *, pgno_t));
53 static PAGE	*bt_psplit
54 		    __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t));
55 static PAGE	*bt_root
56 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
57 static int	 bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
58 static recno_t	 rec_total __P((PAGE *));
59 
60 #ifdef STATISTICS
61 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
62 #endif
63 
64 /*
65  * __BT_SPLIT -- Split the tree.
66  *
67  * Parameters:
68  *	t:	tree
69  *	sp:	page to split
70  *	key:	key to insert
71  *	data:	data to insert
72  *	flags:	BIGKEY/BIGDATA flags
73  *	ilen:	insert length
74  *	skip:	index to leave open
75  *
76  * Returns:
77  *	RET_ERROR, RET_SUCCESS
78  */
79 int
80 __bt_split(t, sp, key, data, flags, ilen, argskip)
81 	BTREE *t;
82 	PAGE *sp;
83 	const DBT *key, *data;
84 	int flags;
85 	size_t ilen;
86 	u_int32_t argskip;
87 {
88 	BINTERNAL *bi;
89 	BLEAF *bl, *tbl;
90 	DBT a, b;
91 	EPGNO *parent;
92 	PAGE *h, *l, *r, *lchild, *rchild;
93 	indx_t nxtindex;
94 	u_int16_t skip;
95 	u_int32_t n, nbytes, nksize;
96 	int parentsplit;
97 	char *dest;
98 
99 	/*
100 	 * Split the page into two pages, l and r.  The split routines return
101 	 * a pointer to the page into which the key should be inserted and with
102 	 * skip set to the offset which should be used.  Additionally, l and r
103 	 * are pinned.
104 	 */
105 	skip = argskip;
106 	h = sp->pgno == P_ROOT ?
107 	    bt_root(t, sp, &l, &r, &skip, ilen) :
108 	    bt_page(t, sp, &l, &r, &skip, ilen);
109 	if (h == NULL)
110 		return (RET_ERROR);
111 
112 	/*
113 	 * Insert the new key/data pair into the leaf page.  (Key inserts
114 	 * always cause a leaf page to split first.)
115 	 */
116 	h->linp[skip] = h->upper -= ilen;
117 	dest = (char *)h + h->upper;
118 	if (F_ISSET(t, R_RECNO))
119 		WR_RLEAF(dest, data, flags)
120 	else
121 		WR_BLEAF(dest, key, data, flags)
122 
123 	/* If the root page was split, make it look right. */
124 	if (sp->pgno == P_ROOT &&
125 	    (F_ISSET(t, R_RECNO) ?
126 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
127 		goto err2;
128 
129 	/*
130 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
131 	 * were traversed when we searched for the page that split.  Each stack
132 	 * entry is a page number and a page index offset.  The offset is for
133 	 * the page traversed on the search.  We've just split a page, so we
134 	 * have to insert a new key into the parent page.
135 	 *
136 	 * If the insert into the parent page causes it to split, may have to
137 	 * continue splitting all the way up the tree.  We stop if the root
138 	 * splits or the page inserted into didn't have to split to hold the
139 	 * new key.  Some algorithms replace the key for the old page as well
140 	 * as the new page.  We don't, as there's no reason to believe that the
141 	 * first key on the old page is any better than the key we have, and,
142 	 * in the case of a key being placed at index 0 causing the split, the
143 	 * key is unavailable.
144 	 *
145 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
146 	 * and right pages pinned while working on the parent.   The 5 are the
147 	 * two children, left parent and right parent (when the parent splits)
148 	 * and the root page or the overflow key page when calling bt_preserve.
149 	 * This code must make sure that all pins are released other than the
150 	 * root page or overflow page which is unlocked elsewhere.
151 	 */
152 	while ((parent = BT_POP(t)) != NULL) {
153 		lchild = l;
154 		rchild = r;
155 
156 		/* Get the parent page. */
157 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
158 			goto err2;
159 
160 	 	/*
161 		 * The new key goes ONE AFTER the index, because the split
162 		 * was to the right.
163 		 */
164 		skip = parent->index + 1;
165 
166 		/*
167 		 * Calculate the space needed on the parent page.
168 		 *
169 		 * Prefix trees: space hack when inserting into BINTERNAL
170 		 * pages.  Retain only what's needed to distinguish between
171 		 * the new entry and the LAST entry on the page to its left.
172 		 * If the keys compare equal, retain the entire key.  Note,
173 		 * we don't touch overflow keys, and the entire key must be
174 		 * retained for the next-to-left most key on the leftmost
175 		 * page of each level, or the search will fail.  Applicable
176 		 * ONLY to internal pages that have leaf pages as children.
177 		 * Further reduction of the key between pairs of internal
178 		 * pages loses too much information.
179 		 */
180 		switch (rchild->flags & P_TYPE) {
181 		case P_BINTERNAL:
182 			bi = GETBINTERNAL(rchild, 0);
183 			nbytes = NBINTERNAL(bi->ksize);
184 			break;
185 		case P_BLEAF:
186 			bl = GETBLEAF(rchild, 0);
187 			nbytes = NBINTERNAL(bl->ksize);
188 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
189 			    (h->prevpg != P_INVALID || skip > 1)) {
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 				nksize = t->bt_pfx(&a, &b);
196 				n = NBINTERNAL(nksize);
197 				if (n < nbytes) {
198 #ifdef STATISTICS
199 					bt_pfxsaved += nbytes - n;
200 #endif
201 					nbytes = n;
202 				} else
203 					nksize = 0;
204 			} else
205 				nksize = 0;
206 			break;
207 		case P_RINTERNAL:
208 		case P_RLEAF:
209 			nbytes = NRINTERNAL;
210 			break;
211 		default:
212 			abort();
213 		}
214 
215 		/* Split the parent page if necessary or shift the indices. */
216 		if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
217 			sp = h;
218 			h = h->pgno == P_ROOT ?
219 			    bt_root(t, h, &l, &r, &skip, nbytes) :
220 			    bt_page(t, h, &l, &r, &skip, nbytes);
221 			if (h == NULL)
222 				goto err1;
223 			parentsplit = 1;
224 		} else {
225 			if (skip < (nxtindex = NEXTINDEX(h)))
226 				memmove(h->linp + skip + 1, h->linp + skip,
227 				    (nxtindex - skip) * sizeof(indx_t));
228 			h->lower += sizeof(indx_t);
229 			parentsplit = 0;
230 		}
231 
232 		/* Insert the key into the parent page. */
233 		switch (rchild->flags & P_TYPE) {
234 		case P_BINTERNAL:
235 			h->linp[skip] = h->upper -= nbytes;
236 			dest = (char *)h + h->linp[skip];
237 			memmove(dest, bi, nbytes);
238 			((BINTERNAL *)dest)->pgno = rchild->pgno;
239 			break;
240 		case P_BLEAF:
241 			h->linp[skip] = h->upper -= nbytes;
242 			dest = (char *)h + h->linp[skip];
243 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
244 			    rchild->pgno, bl->flags & P_BIGKEY);
245 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
246 			if (bl->flags & P_BIGKEY &&
247 			    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
248 				goto err1;
249 			break;
250 		case P_RINTERNAL:
251 			/*
252 			 * Update the left page count.  If split
253 			 * added at index 0, fix the correct page.
254 			 */
255 			if (skip > 0)
256 				dest = (char *)h + h->linp[skip - 1];
257 			else
258 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
259 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
260 			((RINTERNAL *)dest)->pgno = lchild->pgno;
261 
262 			/* Update the right page count. */
263 			h->linp[skip] = h->upper -= nbytes;
264 			dest = (char *)h + h->linp[skip];
265 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
266 			((RINTERNAL *)dest)->pgno = rchild->pgno;
267 			break;
268 		case P_RLEAF:
269 			/*
270 			 * Update the left page count.  If split
271 			 * added at index 0, fix the correct page.
272 			 */
273 			if (skip > 0)
274 				dest = (char *)h + h->linp[skip - 1];
275 			else
276 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
277 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
278 			((RINTERNAL *)dest)->pgno = lchild->pgno;
279 
280 			/* Update the right page count. */
281 			h->linp[skip] = h->upper -= nbytes;
282 			dest = (char *)h + h->linp[skip];
283 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
284 			((RINTERNAL *)dest)->pgno = rchild->pgno;
285 			break;
286 		default:
287 			abort();
288 		}
289 
290 		/* Unpin the held pages. */
291 		if (!parentsplit) {
292 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
293 			break;
294 		}
295 
296 		/* If the root page was split, make it look right. */
297 		if (sp->pgno == P_ROOT &&
298 		    (F_ISSET(t, R_RECNO) ?
299 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
300 			goto err1;
301 
302 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
303 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
304 	}
305 
306 	/* Unpin the held pages. */
307 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
308 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
309 
310 	/* Clear any pages left on the stack. */
311 	return (RET_SUCCESS);
312 
313 	/*
314 	 * If something fails in the above loop we were already walking back
315 	 * up the tree and the tree is now inconsistent.  Nothing much we can
316 	 * do about it but release any memory we're holding.
317 	 */
318 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
319 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
320 
321 err2:	mpool_put(t->bt_mp, l, 0);
322 	mpool_put(t->bt_mp, r, 0);
323 	__dbpanic(t->bt_dbp);
324 	return (RET_ERROR);
325 }
326 
327 /*
328  * BT_PAGE -- Split a non-root page of a btree.
329  *
330  * Parameters:
331  *	t:	tree
332  *	h:	root page
333  *	lp:	pointer to left page pointer
334  *	rp:	pointer to right page pointer
335  *	skip:	pointer to index to leave open
336  *	ilen:	insert length
337  *
338  * Returns:
339  *	Pointer to page in which to insert or NULL on error.
340  */
341 static PAGE *
342 bt_page(t, h, lp, rp, skip, ilen)
343 	BTREE *t;
344 	PAGE *h, **lp, **rp;
345 	indx_t *skip;
346 	size_t ilen;
347 {
348 	PAGE *l, *r, *tp;
349 	pgno_t npg;
350 
351 #ifdef STATISTICS
352 	++bt_split;
353 #endif
354 	/* Put the new right page for the split into place. */
355 	if ((r = __bt_new(t, &npg)) == NULL)
356 		return (NULL);
357 	r->pgno = npg;
358 	r->lower = BTDATAOFF;
359 	r->upper = t->bt_psize;
360 	r->nextpg = h->nextpg;
361 	r->prevpg = h->pgno;
362 	r->flags = h->flags & P_TYPE;
363 
364 	/*
365 	 * If we're splitting the last page on a level because we're appending
366 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
367 	 * sorted.  Adding an empty page on the side of the level is less work
368 	 * and can push the fill factor much higher than normal.  If we're
369 	 * wrong it's no big deal, we'll just do the split the right way next
370 	 * time.  It may look like it's equally easy to do a similar hack for
371 	 * reverse sorted data, that is, split the tree left, but it's not.
372 	 * Don't even try.
373 	 */
374 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
375 #ifdef STATISTICS
376 		++bt_sortsplit;
377 #endif
378 		h->nextpg = r->pgno;
379 		r->lower = BTDATAOFF + sizeof(indx_t);
380 		*skip = 0;
381 		*lp = h;
382 		*rp = r;
383 		return (r);
384 	}
385 
386 	/* Put the new left page for the split into place. */
387 	if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
388 		mpool_put(t->bt_mp, r, 0);
389 		return (NULL);
390 	}
391 #ifdef PURIFY
392 	memset(l, 0xff, t->bt_psize);
393 #endif
394 	l->pgno = h->pgno;
395 	l->nextpg = r->pgno;
396 	l->prevpg = h->prevpg;
397 	l->lower = BTDATAOFF;
398 	l->upper = t->bt_psize;
399 	l->flags = h->flags & P_TYPE;
400 
401 	/* Fix up the previous pointer of the page after the split page. */
402 	if (h->nextpg != P_INVALID) {
403 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
404 			free(l);
405 			/* XXX mpool_free(t->bt_mp, r->pgno); */
406 			return (NULL);
407 		}
408 		tp->prevpg = r->pgno;
409 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
410 	}
411 
412 	/*
413 	 * Split right.  The key/data pairs aren't sorted in the btree page so
414 	 * it's simpler to copy the data from the split page onto two new pages
415 	 * instead of copying half the data to the right page and compacting
416 	 * the left page in place.  Since the left page can't change, we have
417 	 * to swap the original and the allocated left page after the split.
418 	 */
419 	tp = bt_psplit(t, h, l, r, skip, ilen);
420 
421 	/* Move the new left page onto the old left page. */
422 	memmove(h, l, t->bt_psize);
423 	if (tp == l)
424 		tp = h;
425 	free(l);
426 
427 	*lp = h;
428 	*rp = r;
429 	return (tp);
430 }
431 
432 /*
433  * BT_ROOT -- Split the root page of a btree.
434  *
435  * Parameters:
436  *	t:	tree
437  *	h:	root page
438  *	lp:	pointer to left page pointer
439  *	rp:	pointer to right page pointer
440  *	skip:	pointer to index to leave open
441  *	ilen:	insert length
442  *
443  * Returns:
444  *	Pointer to page in which to insert or NULL on error.
445  */
446 static PAGE *
447 bt_root(t, h, lp, rp, skip, ilen)
448 	BTREE *t;
449 	PAGE *h, **lp, **rp;
450 	indx_t *skip;
451 	size_t ilen;
452 {
453 	PAGE *l, *r, *tp;
454 	pgno_t lnpg, rnpg;
455 
456 #ifdef STATISTICS
457 	++bt_split;
458 	++bt_rootsplit;
459 #endif
460 	/* Put the new left and right pages for the split into place. */
461 	if ((l = __bt_new(t, &lnpg)) == NULL ||
462 	    (r = __bt_new(t, &rnpg)) == NULL)
463 		return (NULL);
464 	l->pgno = lnpg;
465 	r->pgno = rnpg;
466 	l->nextpg = r->pgno;
467 	r->prevpg = l->pgno;
468 	l->prevpg = r->nextpg = P_INVALID;
469 	l->lower = r->lower = BTDATAOFF;
470 	l->upper = r->upper = t->bt_psize;
471 	l->flags = r->flags = h->flags & P_TYPE;
472 
473 	/* Split the root page. */
474 	tp = bt_psplit(t, h, l, r, skip, ilen);
475 
476 	*lp = l;
477 	*rp = r;
478 	return (tp);
479 }
480 
481 /*
482  * BT_RROOT -- Fix up the recno root page after it has been split.
483  *
484  * Parameters:
485  *	t:	tree
486  *	h:	root page
487  *	l:	left page
488  *	r:	right page
489  *
490  * Returns:
491  *	RET_ERROR, RET_SUCCESS
492  */
493 static int
494 bt_rroot(t, h, l, r)
495 	BTREE *t;
496 	PAGE *h, *l, *r;
497 {
498 	char *dest;
499 
500 	/* Insert the left and right keys, set the header information. */
501 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
502 	dest = (char *)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 *)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(t, h, l, r)
535 	BTREE *t;
536 	PAGE *h, *l, *r;
537 {
538 	BINTERNAL *bi;
539 	BLEAF *bl;
540 	u_int32_t nbytes;
541 	char *dest;
542 
543 	/*
544 	 * If the root page was a leaf page, change it into an internal page.
545 	 * We copy the key we split on (but not the key's data, in the case of
546 	 * a leaf page) to the new root page.
547 	 *
548 	 * The btree comparison code guarantees that the left-most key on any
549 	 * level of the tree is never used, so it doesn't need to be filled in.
550 	 */
551 	nbytes = NBINTERNAL(0);
552 	h->linp[0] = h->upper = t->bt_psize - nbytes;
553 	dest = (char *)h + h->upper;
554 	WR_BINTERNAL(dest, 0, l->pgno, 0);
555 
556 	switch (h->flags & P_TYPE) {
557 	case P_BLEAF:
558 		bl = GETBLEAF(r, 0);
559 		nbytes = NBINTERNAL(bl->ksize);
560 		h->linp[1] = h->upper -= nbytes;
561 		dest = (char *)h + h->upper;
562 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
563 		memmove(dest, bl->bytes, bl->ksize);
564 
565 		/*
566 		 * If the key is on an overflow page, mark the overflow chain
567 		 * so it isn't deleted when the leaf copy of the key is deleted.
568 		 */
569 		if (bl->flags & P_BIGKEY &&
570 		    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
571 			return (RET_ERROR);
572 		break;
573 	case P_BINTERNAL:
574 		bi = GETBINTERNAL(r, 0);
575 		nbytes = NBINTERNAL(bi->ksize);
576 		h->linp[1] = h->upper -= nbytes;
577 		dest = (char *)h + h->upper;
578 		memmove(dest, bi, nbytes);
579 		((BINTERNAL *)dest)->pgno = r->pgno;
580 		break;
581 	default:
582 		abort();
583 	}
584 
585 	/* There are two keys on the page. */
586 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
587 
588 	/* Unpin the root page, set to btree internal page. */
589 	h->flags &= ~P_TYPE;
590 	h->flags |= P_BINTERNAL;
591 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
592 
593 	return (RET_SUCCESS);
594 }
595 
596 /*
597  * BT_PSPLIT -- Do the real work of splitting the page.
598  *
599  * Parameters:
600  *	t:	tree
601  *	h:	page to be split
602  *	l:	page to put lower half of data
603  *	r:	page to put upper half of data
604  *	pskip:	pointer to index to leave open
605  *	ilen:	insert length
606  *
607  * Returns:
608  *	Pointer to page in which to insert.
609  */
610 static PAGE *
611 bt_psplit(t, h, l, r, pskip, ilen)
612 	BTREE *t;
613 	PAGE *h, *l, *r;
614 	indx_t *pskip;
615 	size_t ilen;
616 {
617 	BINTERNAL *bi;
618 	BLEAF *bl;
619 	CURSOR *c;
620 	RLEAF *rl;
621 	PAGE *rval;
622 	void *src;
623 	indx_t full, half, nxt, off, skip, top, used;
624 	u_int32_t nbytes;
625 	int bigkeycnt, isbigkey;
626 
627 	/*
628 	 * Split the data to the left and right pages.  Leave the skip index
629 	 * open.  Additionally, make some effort not to split on an overflow
630 	 * key.  This makes internal page processing faster and can save
631 	 * space as overflow keys used by internal pages are never deleted.
632 	 */
633 	bigkeycnt = 0;
634 	skip = *pskip;
635 	full = t->bt_psize - BTDATAOFF;
636 	half = full / 2;
637 	used = 0;
638 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
639 		if (skip == off) {
640 			nbytes = ilen;
641 			isbigkey = 0;		/* XXX: not really known. */
642 		} else
643 			switch (h->flags & P_TYPE) {
644 			case P_BINTERNAL:
645 				src = bi = GETBINTERNAL(h, nxt);
646 				nbytes = NBINTERNAL(bi->ksize);
647 				isbigkey = bi->flags & P_BIGKEY;
648 				break;
649 			case P_BLEAF:
650 				src = bl = GETBLEAF(h, nxt);
651 				nbytes = NBLEAF(bl);
652 				isbigkey = bl->flags & P_BIGKEY;
653 				break;
654 			case P_RINTERNAL:
655 				src = GETRINTERNAL(h, nxt);
656 				nbytes = NRINTERNAL;
657 				isbigkey = 0;
658 				break;
659 			case P_RLEAF:
660 				src = rl = GETRLEAF(h, nxt);
661 				nbytes = NRLEAF(rl);
662 				isbigkey = 0;
663 				break;
664 			default:
665 				abort();
666 			}
667 
668 		/*
669 		 * If the key/data pairs are substantial fractions of the max
670 		 * possible size for the page, it's possible to get situations
671 		 * where we decide to try and copy too much onto the left page.
672 		 * Make sure that doesn't happen.
673 		 */
674 		if (skip <= off &&
675 		    used + nbytes + sizeof(indx_t) >= full || nxt == top - 1) {
676 			--off;
677 			break;
678 		}
679 
680 		/* Copy the key/data pair, if not the skipped index. */
681 		if (skip != off) {
682 			++nxt;
683 
684 			l->linp[off] = l->upper -= nbytes;
685 			memmove((char *)l + l->upper, src, nbytes);
686 		}
687 
688 		used += nbytes + sizeof(indx_t);
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 	l->lower += (off + 1) * sizeof(indx_t);
702 
703 	/*
704 	 * If splitting the page that the cursor was on, the cursor has to be
705 	 * adjusted to point to the same record as before the split.  If the
706 	 * cursor is at or past the skipped slot, the cursor is incremented by
707 	 * one.  If the cursor is on the right page, it is decremented by the
708 	 * number of records split to the left page.
709 	 */
710 	c = &t->bt_cursor;
711 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
712 		if (c->pg.index >= skip)
713 			++c->pg.index;
714 		if (c->pg.index < nxt)			/* Left page. */
715 			c->pg.pgno = l->pgno;
716 		else {					/* Right page. */
717 			c->pg.pgno = r->pgno;
718 			c->pg.index -= nxt;
719 		}
720 	}
721 
722 	/*
723 	 * If the skipped index was on the left page, just return that page.
724 	 * Otherwise, adjust the skip index to reflect the new position on
725 	 * the right page.
726 	 */
727 	if (skip <= off) {
728 		skip = 0;
729 		rval = l;
730 	} else {
731 		rval = r;
732 		*pskip -= nxt;
733 	}
734 
735 	for (off = 0; nxt < top; ++off) {
736 		if (skip == nxt) {
737 			++off;
738 			skip = 0;
739 		}
740 		switch (h->flags & P_TYPE) {
741 		case P_BINTERNAL:
742 			src = bi = GETBINTERNAL(h, nxt);
743 			nbytes = NBINTERNAL(bi->ksize);
744 			break;
745 		case P_BLEAF:
746 			src = bl = GETBLEAF(h, nxt);
747 			nbytes = NBLEAF(bl);
748 			break;
749 		case P_RINTERNAL:
750 			src = GETRINTERNAL(h, nxt);
751 			nbytes = NRINTERNAL;
752 			break;
753 		case P_RLEAF:
754 			src = rl = GETRLEAF(h, nxt);
755 			nbytes = NRLEAF(rl);
756 			break;
757 		default:
758 			abort();
759 		}
760 		++nxt;
761 		r->linp[off] = r->upper -= nbytes;
762 		memmove((char *)r + r->upper, src, nbytes);
763 	}
764 	r->lower += off * sizeof(indx_t);
765 
766 	/* If the key is being appended to the page, adjust the index. */
767 	if (skip == top)
768 		r->lower += sizeof(indx_t);
769 
770 	return (rval);
771 }
772 
773 /*
774  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
775  *
776  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
777  * record that references them gets deleted.  Chains pointed to by internal
778  * pages never get deleted.  This routine marks a chain as pointed to by an
779  * internal page.
780  *
781  * Parameters:
782  *	t:	tree
783  *	pg:	page number of first page in the chain.
784  *
785  * Returns:
786  *	RET_SUCCESS, RET_ERROR.
787  */
788 static int
789 bt_preserve(t, pg)
790 	BTREE *t;
791 	pgno_t pg;
792 {
793 	PAGE *h;
794 
795 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
796 		return (RET_ERROR);
797 	h->flags |= P_PRESERVE;
798 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
799 	return (RET_SUCCESS);
800 }
801 
802 /*
803  * REC_TOTAL -- Return the number of recno entries below a page.
804  *
805  * Parameters:
806  *	h:	page
807  *
808  * Returns:
809  *	The number of recno entries below a page.
810  *
811  * XXX
812  * These values could be set by the bt_psplit routine.  The problem is that the
813  * entry has to be popped off of the stack etc. or the values have to be passed
814  * all the way back to bt_split/bt_rroot and it's not very clean.
815  */
816 static recno_t
817 rec_total(h)
818 	PAGE *h;
819 {
820 	recno_t recs;
821 	indx_t nxt, top;
822 
823 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
824 		recs += GETRINTERNAL(h, nxt)->nrecs;
825 	return (recs);
826 }
827