xref: /dragonfly/lib/libc/db/btree/bt_split.c (revision 99dd49c5)
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. 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  * @(#)bt_split.c	8.9 (Berkeley) 7/26/94
33  * $DragonFly: src/lib/libc/db/btree/bt_split.c,v 1.8 2005/11/19 20:46:32 swildner Exp $
34  */
35 
36 #include <sys/types.h>
37 
38 #include <limits.h>
39 #include <stdio.h>
40 #include <stdlib.h>
41 #include <string.h>
42 
43 #include <db.h>
44 #include "btree.h"
45 
46 static int	 bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
47 static PAGE	*bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
48 static int	 bt_preserve(BTREE *, pgno_t);
49 static PAGE	*bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
50 static PAGE	*bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
51 static int	 bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
52 static recno_t	 rec_total(PAGE *);
53 
54 #ifdef STATISTICS
55 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
56 #endif
57 
58 /*
59  * __BT_SPLIT -- Split the tree.
60  *
61  * Parameters:
62  *	t:	tree
63  *	sp:	page to split
64  *	key:	key to insert
65  *	data:	data to insert
66  *	flags:	BIGKEY/BIGDATA flags
67  *	ilen:	insert length
68  *	skip:	index to leave open
69  *
70  * Returns:
71  *	RET_ERROR, RET_SUCCESS
72  */
73 int
74 __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
75 	   size_t ilen, u_int32_t argskip)
76 {
77 	BINTERNAL *bi;
78 	BLEAF *bl, *tbl;
79 	DBT a, b;
80 	EPGNO *parent;
81 	PAGE *h, *l, *r, *lchild, *rchild;
82 	indx_t nxtindex;
83 	u_int16_t skip;
84 	u_int32_t n, nbytes, nksize;
85 	int parentsplit;
86 	char *dest;
87 
88 	bi = NULL;
89 	bl = NULL;
90 	nksize = 0;
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 (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 			    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
240 				goto err1;
241 			break;
242 		case P_RINTERNAL:
243 			/*
244 			 * Update the left page count.  If split
245 			 * added at index 0, fix the correct page.
246 			 */
247 			if (skip > 0)
248 				dest = (char *)h + h->linp[skip - 1];
249 			else
250 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
251 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
252 			((RINTERNAL *)dest)->pgno = lchild->pgno;
253 
254 			/* Update the right page count. */
255 			h->linp[skip] = h->upper -= nbytes;
256 			dest = (char *)h + h->linp[skip];
257 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
258 			((RINTERNAL *)dest)->pgno = rchild->pgno;
259 			break;
260 		case P_RLEAF:
261 			/*
262 			 * Update the left page count.  If split
263 			 * added at index 0, fix the correct page.
264 			 */
265 			if (skip > 0)
266 				dest = (char *)h + h->linp[skip - 1];
267 			else
268 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
269 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
270 			((RINTERNAL *)dest)->pgno = lchild->pgno;
271 
272 			/* Update the right page count. */
273 			h->linp[skip] = h->upper -= nbytes;
274 			dest = (char *)h + h->linp[skip];
275 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
276 			((RINTERNAL *)dest)->pgno = rchild->pgno;
277 			break;
278 		default:
279 			abort();
280 		}
281 
282 		/* Unpin the held pages. */
283 		if (!parentsplit) {
284 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
285 			break;
286 		}
287 
288 		/* If the root page was split, make it look right. */
289 		if (sp->pgno == P_ROOT &&
290 		    (F_ISSET(t, R_RECNO) ?
291 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
292 			goto err1;
293 
294 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
295 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
296 	}
297 
298 	/* Unpin the held pages. */
299 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
300 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
301 
302 	/* Clear any pages left on the stack. */
303 	return (RET_SUCCESS);
304 
305 	/*
306 	 * If something fails in the above loop we were already walking back
307 	 * up the tree and the tree is now inconsistent.  Nothing much we can
308 	 * do about it but release any memory we're holding.
309 	 */
310 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
311 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
312 
313 err2:	mpool_put(t->bt_mp, l, 0);
314 	mpool_put(t->bt_mp, r, 0);
315 	__dbpanic(t->bt_dbp);
316 	return (RET_ERROR);
317 }
318 
319 /*
320  * BT_PAGE -- Split a non-root page of a btree.
321  *
322  * Parameters:
323  *	t:	tree
324  *	h:	root page
325  *	lp:	pointer to left page pointer
326  *	rp:	pointer to right page pointer
327  *	skip:	pointer to index to leave open
328  *	ilen:	insert length
329  *
330  * Returns:
331  *	Pointer to page in which to insert or NULL on error.
332  */
333 static PAGE *
334 bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
335 {
336 	PAGE *l, *r, *tp;
337 	pgno_t npg;
338 
339 #ifdef STATISTICS
340 	++bt_split;
341 #endif
342 	/* Put the new right page for the split into place. */
343 	if ((r = __bt_new(t, &npg)) == NULL)
344 		return (NULL);
345 	r->pgno = npg;
346 	r->lower = BTDATAOFF;
347 	r->upper = t->bt_psize;
348 	r->nextpg = h->nextpg;
349 	r->prevpg = h->pgno;
350 	r->flags = h->flags & P_TYPE;
351 
352 	/*
353 	 * If we're splitting the last page on a level because we're appending
354 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
355 	 * sorted.  Adding an empty page on the side of the level is less work
356 	 * and can push the fill factor much higher than normal.  If we're
357 	 * wrong it's no big deal, we'll just do the split the right way next
358 	 * time.  It may look like it's equally easy to do a similar hack for
359 	 * reverse sorted data, that is, split the tree left, but it's not.
360 	 * Don't even try.
361 	 */
362 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
363 #ifdef STATISTICS
364 		++bt_sortsplit;
365 #endif
366 		h->nextpg = r->pgno;
367 		r->lower = BTDATAOFF + sizeof(indx_t);
368 		*skip = 0;
369 		*lp = h;
370 		*rp = r;
371 		return (r);
372 	}
373 
374 	/* Put the new left page for the split into place. */
375 	if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
376 		mpool_put(t->bt_mp, r, 0);
377 		return (NULL);
378 	}
379 #ifdef PURIFY
380 	memset(l, 0xff, t->bt_psize);
381 #endif
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 	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 		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 		    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
551 			return (RET_ERROR);
552 		break;
553 	case P_BINTERNAL:
554 		bi = GETBINTERNAL(r, 0);
555 		nbytes = NBINTERNAL(bi->ksize);
556 		h->linp[1] = h->upper -= nbytes;
557 		dest = (char *)h + h->upper;
558 		memmove(dest, bi, nbytes);
559 		((BINTERNAL *)dest)->pgno = r->pgno;
560 		break;
561 	default:
562 		abort();
563 	}
564 
565 	/* There are two keys on the page. */
566 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
567 
568 	/* Unpin the root page, set to btree internal page. */
569 	h->flags &= ~P_TYPE;
570 	h->flags |= P_BINTERNAL;
571 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
572 
573 	return (RET_SUCCESS);
574 }
575 
576 /*
577  * BT_PSPLIT -- Do the real work of splitting the page.
578  *
579  * Parameters:
580  *	t:	tree
581  *	h:	page to be split
582  *	l:	page to put lower half of data
583  *	r:	page to put upper half of data
584  *	pskip:	pointer to index to leave open
585  *	ilen:	insert length
586  *
587  * Returns:
588  *	Pointer to page in which to insert.
589  */
590 static PAGE *
591 bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
592 {
593 	BINTERNAL *bi;
594 	BLEAF *bl;
595 	CURSOR *c;
596 	RLEAF *rl;
597 	PAGE *rval;
598 	void *src;
599 	indx_t full, half, nxt, off, skip, top, used;
600 	u_int32_t nbytes;
601 	int bigkeycnt, isbigkey;
602 
603 	src = NULL;
604 	/*
605 	 * Split the data to the left and right pages.  Leave the skip index
606 	 * open.  Additionally, make some effort not to split on an overflow
607 	 * key.  This makes internal page processing faster and can save
608 	 * space as overflow keys used by internal pages are never deleted.
609 	 */
610 	bigkeycnt = 0;
611 	skip = *pskip;
612 	full = t->bt_psize - BTDATAOFF;
613 	half = full / 2;
614 	used = 0;
615 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
616 		if (skip == off) {
617 			nbytes = ilen;
618 			isbigkey = 0;		/* XXX: not really known. */
619 		} else
620 			switch (h->flags & P_TYPE) {
621 			case P_BINTERNAL:
622 				src = bi = GETBINTERNAL(h, nxt);
623 				nbytes = NBINTERNAL(bi->ksize);
624 				isbigkey = bi->flags & P_BIGKEY;
625 				break;
626 			case P_BLEAF:
627 				src = bl = GETBLEAF(h, nxt);
628 				nbytes = NBLEAF(bl);
629 				isbigkey = bl->flags & P_BIGKEY;
630 				break;
631 			case P_RINTERNAL:
632 				src = GETRINTERNAL(h, nxt);
633 				nbytes = NRINTERNAL;
634 				isbigkey = 0;
635 				break;
636 			case P_RLEAF:
637 				src = rl = GETRLEAF(h, nxt);
638 				nbytes = NRLEAF(rl);
639 				isbigkey = 0;
640 				break;
641 			default:
642 				abort();
643 			}
644 
645 		/*
646 		 * If the key/data pairs are substantial fractions of the max
647 		 * possible size for the page, it's possible to get situations
648 		 * where we decide to try and copy too much onto the left page.
649 		 * Make sure that doesn't happen.
650 		 */
651 		if ((skip <= off &&
652 		    used + nbytes + sizeof(indx_t) >= full) || nxt == top - 1) {
653 			--off;
654 			break;
655 		}
656 
657 		/* Copy the key/data pair, if not the skipped index. */
658 		if (skip != off) {
659 			++nxt;
660 
661 			l->linp[off] = l->upper -= nbytes;
662 			memmove((char *)l + l->upper, src, nbytes);
663 		}
664 
665 		used += nbytes + sizeof(indx_t);
666 		if (used >= half) {
667 			if (!isbigkey || bigkeycnt == 3)
668 				break;
669 			else
670 				++bigkeycnt;
671 		}
672 	}
673 
674 	/*
675 	 * Off is the last offset that's valid for the left page.
676 	 * Nxt is the first offset to be placed on the right page.
677 	 */
678 	l->lower += (off + 1) * sizeof(indx_t);
679 
680 	/*
681 	 * If splitting the page that the cursor was on, the cursor has to be
682 	 * adjusted to point to the same record as before the split.  If the
683 	 * cursor is at or past the skipped slot, the cursor is incremented by
684 	 * one.  If the cursor is on the right page, it is decremented by the
685 	 * number of records split to the left page.
686 	 */
687 	c = &t->bt_cursor;
688 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
689 		if (c->pg.index >= skip)
690 			++c->pg.index;
691 		if (c->pg.index < nxt)			/* Left page. */
692 			c->pg.pgno = l->pgno;
693 		else {					/* Right page. */
694 			c->pg.pgno = r->pgno;
695 			c->pg.index -= nxt;
696 		}
697 	}
698 
699 	/*
700 	 * If the skipped index was on the left page, just return that page.
701 	 * Otherwise, adjust the skip index to reflect the new position on
702 	 * the right page.
703 	 */
704 	if (skip <= off) {
705 		skip = MAX_PAGE_OFFSET;
706 		rval = l;
707 	} else {
708 		rval = r;
709 		*pskip -= nxt;
710 	}
711 
712 	for (off = 0; nxt < top; ++off) {
713 		if (skip == nxt) {
714 			++off;
715 			skip = MAX_PAGE_OFFSET;
716 		}
717 		switch (h->flags & P_TYPE) {
718 		case P_BINTERNAL:
719 			src = bi = GETBINTERNAL(h, nxt);
720 			nbytes = NBINTERNAL(bi->ksize);
721 			break;
722 		case P_BLEAF:
723 			src = bl = GETBLEAF(h, nxt);
724 			nbytes = NBLEAF(bl);
725 			break;
726 		case P_RINTERNAL:
727 			src = GETRINTERNAL(h, nxt);
728 			nbytes = NRINTERNAL;
729 			break;
730 		case P_RLEAF:
731 			src = rl = GETRLEAF(h, nxt);
732 			nbytes = NRLEAF(rl);
733 			break;
734 		default:
735 			abort();
736 		}
737 		++nxt;
738 		r->linp[off] = r->upper -= nbytes;
739 		memmove((char *)r + r->upper, src, nbytes);
740 	}
741 	r->lower += off * sizeof(indx_t);
742 
743 	/* If the key is being appended to the page, adjust the index. */
744 	if (skip == top)
745 		r->lower += sizeof(indx_t);
746 
747 	return (rval);
748 }
749 
750 /*
751  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
752  *
753  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
754  * record that references them gets deleted.  Chains pointed to by internal
755  * pages never get deleted.  This routine marks a chain as pointed to by an
756  * internal page.
757  *
758  * Parameters:
759  *	t:	tree
760  *	pg:	page number of first page in the chain.
761  *
762  * Returns:
763  *	RET_SUCCESS, RET_ERROR.
764  */
765 static int
766 bt_preserve(BTREE *t, pgno_t pg)
767 {
768 	PAGE *h;
769 
770 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
771 		return (RET_ERROR);
772 	h->flags |= P_PRESERVE;
773 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
774 	return (RET_SUCCESS);
775 }
776 
777 /*
778  * REC_TOTAL -- Return the number of recno entries below a page.
779  *
780  * Parameters:
781  *	h:	page
782  *
783  * Returns:
784  *	The number of recno entries below a page.
785  *
786  * XXX
787  * These values could be set by the bt_psplit routine.  The problem is that the
788  * entry has to be popped off of the stack etc. or the values have to be passed
789  * all the way back to bt_split/bt_rroot and it's not very clean.
790  */
791 static recno_t
792 rec_total(PAGE *h)
793 {
794 	recno_t recs;
795 	indx_t nxt, top;
796 
797 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
798 		recs += GETRINTERNAL(h, nxt)->nrecs;
799 	return (recs);
800 }
801