xref: /openbsd-src/lib/libc/db/btree/bt_split.c (revision db3296cf5c1dd9058ceecc3a29fe4aaa0bd26000)
1 /*	$OpenBSD: bt_split.c,v 1.11 2003/06/02 20:18:33 millert Exp $	*/
2 
3 /*-
4  * Copyright (c) 1990, 1993, 1994
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Mike Olson.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 
35 #if defined(LIBC_SCCS) && !defined(lint)
36 #if 0
37 static char sccsid[] = "@(#)bt_split.c	8.10 (Berkeley) 1/9/95";
38 #else
39 static const char rcsid[] = "$OpenBSD: bt_split.c,v 1.11 2003/06/02 20:18:33 millert Exp $";
40 #endif
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.h>
51 #include "btree.h"
52 
53 static int	 bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
54 static PAGE	*bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
55 static int	 bt_preserve(BTREE *, pgno_t);
56 static PAGE	*bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
57 static PAGE	*bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
58 static int	 bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
59 static recno_t	 rec_total(PAGE *);
60 
61 #ifdef STATISTICS
62 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
63 #endif
64 
65 /*
66  * __BT_SPLIT -- Split the tree.
67  *
68  * Parameters:
69  *	t:	tree
70  *	sp:	page to split
71  *	key:	key to insert
72  *	data:	data to insert
73  *	flags:	BIGKEY/BIGDATA flags
74  *	ilen:	insert length
75  *	skip:	index to leave open
76  *
77  * Returns:
78  *	RET_ERROR, RET_SUCCESS
79  */
80 int
81 __bt_split(t, sp, key, data, flags, ilen, argskip)
82 	BTREE *t;
83 	PAGE *sp;
84 	const DBT *key, *data;
85 	int flags;
86 	size_t ilen;
87 	u_int32_t argskip;
88 {
89 	BINTERNAL *bi;
90 	BLEAF *bl, *tbl;
91 	DBT a, b;
92 	EPGNO *parent;
93 	PAGE *h, *l, *r, *lchild, *rchild;
94 	indx_t nxtindex;
95 	u_int16_t skip;
96 	u_int32_t n, nbytes, nksize;
97 	int parentsplit;
98 	char *dest;
99 
100 	/*
101 	 * Split the page into two pages, l and r.  The split routines return
102 	 * a pointer to the page into which the key should be inserted and with
103 	 * skip set to the offset which should be used.  Additionally, l and r
104 	 * are pinned.
105 	 */
106 	skip = argskip;
107 	h = sp->pgno == P_ROOT ?
108 	    bt_root(t, sp, &l, &r, &skip, ilen) :
109 	    bt_page(t, sp, &l, &r, &skip, ilen);
110 	if (h == NULL)
111 		return (RET_ERROR);
112 
113 	/*
114 	 * Insert the new key/data pair into the leaf page.  (Key inserts
115 	 * always cause a leaf page to split first.)
116 	 */
117 	h->linp[skip] = h->upper -= ilen;
118 	dest = (char *)h + h->upper;
119 	if (F_ISSET(t, R_RECNO))
120 		WR_RLEAF(dest, data, flags)
121 	else
122 		WR_BLEAF(dest, key, data, flags)
123 
124 	/* If the root page was split, make it look right. */
125 	if (sp->pgno == P_ROOT &&
126 	    (F_ISSET(t, R_RECNO) ?
127 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
128 		goto err2;
129 
130 	/*
131 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
132 	 * were traversed when we searched for the page that split.  Each stack
133 	 * entry is a page number and a page index offset.  The offset is for
134 	 * the page traversed on the search.  We've just split a page, so we
135 	 * have to insert a new key into the parent page.
136 	 *
137 	 * If the insert into the parent page causes it to split, may have to
138 	 * continue splitting all the way up the tree.  We stop if the root
139 	 * splits or the page inserted into didn't have to split to hold the
140 	 * new key.  Some algorithms replace the key for the old page as well
141 	 * as the new page.  We don't, as there's no reason to believe that the
142 	 * first key on the old page is any better than the key we have, and,
143 	 * in the case of a key being placed at index 0 causing the split, the
144 	 * key is unavailable.
145 	 *
146 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
147 	 * and right pages pinned while working on the parent.   The 5 are the
148 	 * two children, left parent and right parent (when the parent splits)
149 	 * and the root page or the overflow key page when calling bt_preserve.
150 	 * This code must make sure that all pins are released other than the
151 	 * root page or overflow page which is unlocked elsewhere.
152 	 */
153 	while ((parent = BT_POP(t)) != NULL) {
154 		lchild = l;
155 		rchild = r;
156 
157 		/* Get the parent page. */
158 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
159 			goto err2;
160 
161 	 	/*
162 		 * The new key goes ONE AFTER the index, because the split
163 		 * was to the right.
164 		 */
165 		skip = parent->index + 1;
166 
167 		/*
168 		 * Calculate the space needed on the parent page.
169 		 *
170 		 * Prefix trees: space hack when inserting into BINTERNAL
171 		 * pages.  Retain only what's needed to distinguish between
172 		 * the new entry and the LAST entry on the page to its left.
173 		 * If the keys compare equal, retain the entire key.  Note,
174 		 * we don't touch overflow keys, and the entire key must be
175 		 * retained for the next-to-left most key on the leftmost
176 		 * page of each level, or the search will fail.  Applicable
177 		 * ONLY to internal pages that have leaf pages as children.
178 		 * Further reduction of the key between pairs of internal
179 		 * pages loses too much information.
180 		 */
181 		switch (rchild->flags & P_TYPE) {
182 		case P_BINTERNAL:
183 			bi = GETBINTERNAL(rchild, 0);
184 			nbytes = NBINTERNAL(bi->ksize);
185 			break;
186 		case P_BLEAF:
187 			bl = GETBLEAF(rchild, 0);
188 			nbytes = NBINTERNAL(bl->ksize);
189 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
190 			    (h->prevpg != P_INVALID || skip > 1)) {
191 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
192 				a.size = tbl->ksize;
193 				a.data = tbl->bytes;
194 				b.size = bl->ksize;
195 				b.data = bl->bytes;
196 				nksize = t->bt_pfx(&a, &b);
197 				n = NBINTERNAL(nksize);
198 				if (n < nbytes) {
199 #ifdef STATISTICS
200 					bt_pfxsaved += nbytes - n;
201 #endif
202 					nbytes = n;
203 				} else
204 					nksize = 0;
205 			} else
206 				nksize = 0;
207 			break;
208 		case P_RINTERNAL:
209 		case P_RLEAF:
210 			nbytes = NRINTERNAL;
211 			break;
212 		default:
213 			abort();
214 		}
215 
216 		/* Split the parent page if necessary or shift the indices. */
217 		if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
218 			sp = h;
219 			h = h->pgno == P_ROOT ?
220 			    bt_root(t, h, &l, &r, &skip, nbytes) :
221 			    bt_page(t, h, &l, &r, &skip, nbytes);
222 			if (h == NULL)
223 				goto err1;
224 			parentsplit = 1;
225 		} else {
226 			if (skip < (nxtindex = NEXTINDEX(h)))
227 				memmove(h->linp + skip + 1, h->linp + skip,
228 				    (nxtindex - skip) * sizeof(indx_t));
229 			h->lower += sizeof(indx_t);
230 			parentsplit = 0;
231 		}
232 
233 		/* Insert the key into the parent page. */
234 		switch (rchild->flags & P_TYPE) {
235 		case P_BINTERNAL:
236 			h->linp[skip] = h->upper -= nbytes;
237 			dest = (char *)h + h->linp[skip];
238 			memmove(dest, bi, nbytes);
239 			((BINTERNAL *)dest)->pgno = rchild->pgno;
240 			break;
241 		case P_BLEAF:
242 			h->linp[skip] = h->upper -= nbytes;
243 			dest = (char *)h + h->linp[skip];
244 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
245 			    rchild->pgno, bl->flags & P_BIGKEY);
246 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
247 			if (bl->flags & P_BIGKEY &&
248 			    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
249 				goto err1;
250 			break;
251 		case P_RINTERNAL:
252 			/*
253 			 * Update the left page count.  If split
254 			 * added at index 0, fix the correct page.
255 			 */
256 			if (skip > 0)
257 				dest = (char *)h + h->linp[skip - 1];
258 			else
259 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
260 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
261 			((RINTERNAL *)dest)->pgno = lchild->pgno;
262 
263 			/* Update the right page count. */
264 			h->linp[skip] = h->upper -= nbytes;
265 			dest = (char *)h + h->linp[skip];
266 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
267 			((RINTERNAL *)dest)->pgno = rchild->pgno;
268 			break;
269 		case P_RLEAF:
270 			/*
271 			 * Update the left page count.  If split
272 			 * added at index 0, fix the correct page.
273 			 */
274 			if (skip > 0)
275 				dest = (char *)h + h->linp[skip - 1];
276 			else
277 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
278 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
279 			((RINTERNAL *)dest)->pgno = lchild->pgno;
280 
281 			/* Update the right page count. */
282 			h->linp[skip] = h->upper -= nbytes;
283 			dest = (char *)h + h->linp[skip];
284 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
285 			((RINTERNAL *)dest)->pgno = rchild->pgno;
286 			break;
287 		default:
288 			abort();
289 		}
290 
291 		/* Unpin the held pages. */
292 		if (!parentsplit) {
293 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
294 			break;
295 		}
296 
297 		/* If the root page was split, make it look right. */
298 		if (sp->pgno == P_ROOT &&
299 		    (F_ISSET(t, R_RECNO) ?
300 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
301 			goto err1;
302 
303 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
304 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
305 	}
306 
307 	/* Unpin the held pages. */
308 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
309 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
310 
311 	/* Clear any pages left on the stack. */
312 	return (RET_SUCCESS);
313 
314 	/*
315 	 * If something fails in the above loop we were already walking back
316 	 * up the tree and the tree is now inconsistent.  Nothing much we can
317 	 * do about it but release any memory we're holding.
318 	 */
319 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
320 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
321 
322 err2:	mpool_put(t->bt_mp, l, 0);
323 	mpool_put(t->bt_mp, r, 0);
324 	__dbpanic(t->bt_dbp);
325 	return (RET_ERROR);
326 }
327 
328 /*
329  * BT_PAGE -- Split a non-root page of a btree.
330  *
331  * Parameters:
332  *	t:	tree
333  *	h:	root page
334  *	lp:	pointer to left page pointer
335  *	rp:	pointer to right page pointer
336  *	skip:	pointer to index to leave open
337  *	ilen:	insert length
338  *
339  * Returns:
340  *	Pointer to page in which to insert or NULL on error.
341  */
342 static PAGE *
343 bt_page(t, h, lp, rp, skip, ilen)
344 	BTREE *t;
345 	PAGE *h, **lp, **rp;
346 	indx_t *skip;
347 	size_t ilen;
348 {
349 	PAGE *l, *r, *tp;
350 	pgno_t npg;
351 
352 #ifdef STATISTICS
353 	++bt_split;
354 #endif
355 	/* Put the new right page for the split into place. */
356 	if ((r = __bt_new(t, &npg)) == NULL)
357 		return (NULL);
358 	r->pgno = npg;
359 	r->lower = BTDATAOFF;
360 	r->upper = t->bt_psize;
361 	r->nextpg = h->nextpg;
362 	r->prevpg = h->pgno;
363 	r->flags = h->flags & P_TYPE;
364 
365 	/*
366 	 * If we're splitting the last page on a level because we're appending
367 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
368 	 * sorted.  Adding an empty page on the side of the level is less work
369 	 * and can push the fill factor much higher than normal.  If we're
370 	 * wrong it's no big deal, we'll just do the split the right way next
371 	 * time.  It may look like it's equally easy to do a similar hack for
372 	 * reverse sorted data, that is, split the tree left, but it's not.
373 	 * Don't even try.
374 	 */
375 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
376 #ifdef STATISTICS
377 		++bt_sortsplit;
378 #endif
379 		h->nextpg = r->pgno;
380 		r->lower = BTDATAOFF + sizeof(indx_t);
381 		*skip = 0;
382 		*lp = h;
383 		*rp = r;
384 		return (r);
385 	}
386 
387 	/* Put the new left page for the split into place. */
388 	if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
389 		mpool_put(t->bt_mp, r, 0);
390 		return (NULL);
391 	}
392 	memset(l, 0xff, t->bt_psize);
393 	l->pgno = h->pgno;
394 	l->nextpg = r->pgno;
395 	l->prevpg = h->prevpg;
396 	l->lower = BTDATAOFF;
397 	l->upper = t->bt_psize;
398 	l->flags = h->flags & P_TYPE;
399 
400 	/* Fix up the previous pointer of the page after the split page. */
401 	if (h->nextpg != P_INVALID) {
402 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
403 			free(l);
404 			/* XXX mpool_free(t->bt_mp, r->pgno); */
405 			return (NULL);
406 		}
407 		tp->prevpg = r->pgno;
408 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
409 	}
410 
411 	/*
412 	 * Split right.  The key/data pairs aren't sorted in the btree page so
413 	 * it's simpler to copy the data from the split page onto two new pages
414 	 * instead of copying half the data to the right page and compacting
415 	 * the left page in place.  Since the left page can't change, we have
416 	 * to swap the original and the allocated left page after the split.
417 	 */
418 	tp = bt_psplit(t, h, l, r, skip, ilen);
419 
420 	/* Move the new left page onto the old left page. */
421 	memmove(h, l, t->bt_psize);
422 	if (tp == l)
423 		tp = h;
424 	free(l);
425 
426 	*lp = h;
427 	*rp = r;
428 	return (tp);
429 }
430 
431 /*
432  * BT_ROOT -- Split the root page of a btree.
433  *
434  * Parameters:
435  *	t:	tree
436  *	h:	root page
437  *	lp:	pointer to left page pointer
438  *	rp:	pointer to right page pointer
439  *	skip:	pointer to index to leave open
440  *	ilen:	insert length
441  *
442  * Returns:
443  *	Pointer to page in which to insert or NULL on error.
444  */
445 static PAGE *
446 bt_root(t, h, lp, rp, skip, ilen)
447 	BTREE *t;
448 	PAGE *h, **lp, **rp;
449 	indx_t *skip;
450 	size_t ilen;
451 {
452 	PAGE *l, *r, *tp;
453 	pgno_t lnpg, rnpg;
454 
455 #ifdef STATISTICS
456 	++bt_split;
457 	++bt_rootsplit;
458 #endif
459 	/* Put the new left and right pages for the split into place. */
460 	if ((l = __bt_new(t, &lnpg)) == NULL ||
461 	    (r = __bt_new(t, &rnpg)) == NULL)
462 		return (NULL);
463 	l->pgno = lnpg;
464 	r->pgno = rnpg;
465 	l->nextpg = r->pgno;
466 	r->prevpg = l->pgno;
467 	l->prevpg = r->nextpg = P_INVALID;
468 	l->lower = r->lower = BTDATAOFF;
469 	l->upper = r->upper = t->bt_psize;
470 	l->flags = r->flags = h->flags & P_TYPE;
471 
472 	/* Split the root page. */
473 	tp = bt_psplit(t, h, l, r, skip, ilen);
474 
475 	*lp = l;
476 	*rp = r;
477 	return (tp);
478 }
479 
480 /*
481  * BT_RROOT -- Fix up the recno root page after it has been split.
482  *
483  * Parameters:
484  *	t:	tree
485  *	h:	root page
486  *	l:	left page
487  *	r:	right page
488  *
489  * Returns:
490  *	RET_ERROR, RET_SUCCESS
491  */
492 static int
493 bt_rroot(t, h, l, r)
494 	BTREE *t;
495 	PAGE *h, *l, *r;
496 {
497 	char *dest;
498 
499 	/* Insert the left and right keys, set the header information. */
500 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
501 	dest = (char *)h + h->upper;
502 	WR_RINTERNAL(dest,
503 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
504 
505 	h->linp[1] = h->upper -= NRINTERNAL;
506 	dest = (char *)h + h->upper;
507 	WR_RINTERNAL(dest,
508 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
509 
510 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
511 
512 	/* Unpin the root page, set to recno internal page. */
513 	h->flags &= ~P_TYPE;
514 	h->flags |= P_RINTERNAL;
515 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
516 
517 	return (RET_SUCCESS);
518 }
519 
520 /*
521  * BT_BROOT -- Fix up the btree root page after it has been split.
522  *
523  * Parameters:
524  *	t:	tree
525  *	h:	root page
526  *	l:	left page
527  *	r:	right page
528  *
529  * Returns:
530  *	RET_ERROR, RET_SUCCESS
531  */
532 static int
533 bt_broot(t, h, l, r)
534 	BTREE *t;
535 	PAGE *h, *l, *r;
536 {
537 	BINTERNAL *bi;
538 	BLEAF *bl;
539 	u_int32_t nbytes;
540 	char *dest;
541 
542 	/*
543 	 * If the root page was a leaf page, change it into an internal page.
544 	 * We copy the key we split on (but not the key's data, in the case of
545 	 * a leaf page) to the new root page.
546 	 *
547 	 * The btree comparison code guarantees that the left-most key on any
548 	 * level of the tree is never used, so it doesn't need to be filled in.
549 	 */
550 	nbytes = NBINTERNAL(0);
551 	h->linp[0] = h->upper = t->bt_psize - nbytes;
552 	dest = (char *)h + h->upper;
553 	WR_BINTERNAL(dest, 0, l->pgno, 0);
554 
555 	switch (h->flags & P_TYPE) {
556 	case P_BLEAF:
557 		bl = GETBLEAF(r, 0);
558 		nbytes = NBINTERNAL(bl->ksize);
559 		h->linp[1] = h->upper -= nbytes;
560 		dest = (char *)h + h->upper;
561 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
562 		memmove(dest, bl->bytes, bl->ksize);
563 
564 		/*
565 		 * If the key is on an overflow page, mark the overflow chain
566 		 * so it isn't deleted when the leaf copy of the key is deleted.
567 		 */
568 		if (bl->flags & P_BIGKEY &&
569 		    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
570 			return (RET_ERROR);
571 		break;
572 	case P_BINTERNAL:
573 		bi = GETBINTERNAL(r, 0);
574 		nbytes = NBINTERNAL(bi->ksize);
575 		h->linp[1] = h->upper -= nbytes;
576 		dest = (char *)h + h->upper;
577 		memmove(dest, bi, nbytes);
578 		((BINTERNAL *)dest)->pgno = r->pgno;
579 		break;
580 	default:
581 		abort();
582 	}
583 
584 	/* There are two keys on the page. */
585 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
586 
587 	/* Unpin the root page, set to btree internal page. */
588 	h->flags &= ~P_TYPE;
589 	h->flags |= P_BINTERNAL;
590 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
591 
592 	return (RET_SUCCESS);
593 }
594 
595 /*
596  * BT_PSPLIT -- Do the real work of splitting the page.
597  *
598  * Parameters:
599  *	t:	tree
600  *	h:	page to be split
601  *	l:	page to put lower half of data
602  *	r:	page to put upper half of data
603  *	pskip:	pointer to index to leave open
604  *	ilen:	insert length
605  *
606  * Returns:
607  *	Pointer to page in which to insert.
608  */
609 static PAGE *
610 bt_psplit(t, h, l, r, pskip, ilen)
611 	BTREE *t;
612 	PAGE *h, *l, *r;
613 	indx_t *pskip;
614 	size_t ilen;
615 {
616 	BINTERNAL *bi;
617 	BLEAF *bl;
618 	CURSOR *c;
619 	RLEAF *rl;
620 	PAGE *rval;
621 	void *src;
622 	indx_t full, half, nxt, off, skip, top, used;
623 	u_int32_t nbytes;
624 	int bigkeycnt, isbigkey;
625 
626 	/*
627 	 * Split the data to the left and right pages.  Leave the skip index
628 	 * open.  Additionally, make some effort not to split on an overflow
629 	 * key.  This makes internal page processing faster and can save
630 	 * space as overflow keys used by internal pages are never deleted.
631 	 */
632 	bigkeycnt = 0;
633 	skip = *pskip;
634 	full = t->bt_psize - BTDATAOFF;
635 	half = full / 2;
636 	used = 0;
637 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
638 		if (skip == off) {
639 			nbytes = ilen;
640 			isbigkey = 0;		/* XXX: not really known. */
641 		} else
642 			switch (h->flags & P_TYPE) {
643 			case P_BINTERNAL:
644 				src = bi = GETBINTERNAL(h, nxt);
645 				nbytes = NBINTERNAL(bi->ksize);
646 				isbigkey = bi->flags & P_BIGKEY;
647 				break;
648 			case P_BLEAF:
649 				src = bl = GETBLEAF(h, nxt);
650 				nbytes = NBLEAF(bl);
651 				isbigkey = bl->flags & P_BIGKEY;
652 				break;
653 			case P_RINTERNAL:
654 				src = GETRINTERNAL(h, nxt);
655 				nbytes = NRINTERNAL;
656 				isbigkey = 0;
657 				break;
658 			case P_RLEAF:
659 				src = rl = GETRLEAF(h, nxt);
660 				nbytes = NRLEAF(rl);
661 				isbigkey = 0;
662 				break;
663 			default:
664 				abort();
665 			}
666 
667 		/*
668 		 * If the key/data pairs are substantial fractions of the max
669 		 * possible size for the page, it's possible to get situations
670 		 * where we decide to try and copy too much onto the left page.
671 		 * Make sure that doesn't happen.
672 		 */
673 		if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
674 		    nxt == top - 1) {
675 			--off;
676 			break;
677 		}
678 
679 		/* Copy the key/data pair, if not the skipped index. */
680 		if (skip != off) {
681 			++nxt;
682 
683 			l->linp[off] = l->upper -= nbytes;
684 			memmove((char *)l + l->upper, src, nbytes);
685 		}
686 
687 		used += nbytes + sizeof(indx_t);
688 		if (used >= half) {
689 			if (!isbigkey || bigkeycnt == 3)
690 				break;
691 			else
692 				++bigkeycnt;
693 		}
694 	}
695 
696 	/*
697 	 * Off is the last offset that's valid for the left page.
698 	 * Nxt is the first offset to be placed on the right page.
699 	 */
700 	l->lower += (off + 1) * sizeof(indx_t);
701 
702 	/*
703 	 * If splitting the page that the cursor was on, the cursor has to be
704 	 * adjusted to point to the same record as before the split.  If the
705 	 * cursor is at or past the skipped slot, the cursor is incremented by
706 	 * one.  If the cursor is on the right page, it is decremented by the
707 	 * number of records split to the left page.
708 	 */
709 	c = &t->bt_cursor;
710 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
711 		if (c->pg.index >= skip)
712 			++c->pg.index;
713 		if (c->pg.index < nxt)			/* Left page. */
714 			c->pg.pgno = l->pgno;
715 		else {					/* Right page. */
716 			c->pg.pgno = r->pgno;
717 			c->pg.index -= nxt;
718 		}
719 	}
720 
721 	/*
722 	 * If the skipped index was on the left page, just return that page.
723 	 * Otherwise, adjust the skip index to reflect the new position on
724 	 * the right page.
725 	 */
726 	if (skip <= off) {
727 		skip = MAX_PAGE_OFFSET;
728 		rval = l;
729 	} else {
730 		rval = r;
731 		*pskip -= nxt;
732 	}
733 
734 	for (off = 0; nxt < top; ++off) {
735 		if (skip == nxt) {
736 			++off;
737 			skip = MAX_PAGE_OFFSET;
738 		}
739 		switch (h->flags & P_TYPE) {
740 		case P_BINTERNAL:
741 			src = bi = GETBINTERNAL(h, nxt);
742 			nbytes = NBINTERNAL(bi->ksize);
743 			break;
744 		case P_BLEAF:
745 			src = bl = GETBLEAF(h, nxt);
746 			nbytes = NBLEAF(bl);
747 			break;
748 		case P_RINTERNAL:
749 			src = GETRINTERNAL(h, nxt);
750 			nbytes = NRINTERNAL;
751 			break;
752 		case P_RLEAF:
753 			src = rl = GETRLEAF(h, nxt);
754 			nbytes = NRLEAF(rl);
755 			break;
756 		default:
757 			abort();
758 		}
759 		++nxt;
760 		r->linp[off] = r->upper -= nbytes;
761 		memmove((char *)r + r->upper, src, nbytes);
762 	}
763 	r->lower += off * sizeof(indx_t);
764 
765 	/* If the key is being appended to the page, adjust the index. */
766 	if (skip == top)
767 		r->lower += sizeof(indx_t);
768 
769 	return (rval);
770 }
771 
772 /*
773  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
774  *
775  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
776  * record that references them gets deleted.  Chains pointed to by internal
777  * pages never get deleted.  This routine marks a chain as pointed to by an
778  * internal page.
779  *
780  * Parameters:
781  *	t:	tree
782  *	pg:	page number of first page in the chain.
783  *
784  * Returns:
785  *	RET_SUCCESS, RET_ERROR.
786  */
787 static int
788 bt_preserve(t, pg)
789 	BTREE *t;
790 	pgno_t pg;
791 {
792 	PAGE *h;
793 
794 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
795 		return (RET_ERROR);
796 	h->flags |= P_PRESERVE;
797 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
798 	return (RET_SUCCESS);
799 }
800 
801 /*
802  * REC_TOTAL -- Return the number of recno entries below a page.
803  *
804  * Parameters:
805  *	h:	page
806  *
807  * Returns:
808  *	The number of recno entries below a page.
809  *
810  * XXX
811  * These values could be set by the bt_psplit routine.  The problem is that the
812  * entry has to be popped off of the stack etc. or the values have to be passed
813  * all the way back to bt_split/bt_rroot and it's not very clean.
814  */
815 static recno_t
816 rec_total(h)
817 	PAGE *h;
818 {
819 	recno_t recs;
820 	indx_t nxt, top;
821 
822 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
823 		recs += GETRINTERNAL(h, nxt)->nrecs;
824 	return (recs);
825 }
826