xref: /freebsd/crypto/krb5/src/plugins/kdb/db2/libdb2/btree/bt_split.c (revision 7f2fe78b9dd5f51c821d771b63d2e096f6fd49e9)
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 
37 #if defined(LIBC_SCCS) && !defined(lint)
38 static char sccsid[] = "@(#)bt_split.c	8.10 (Berkeley) 1/9/95";
39 #endif /* LIBC_SCCS and not lint */
40 
41 #include <sys/types.h>
42 
43 #include <limits.h>
44 #include <stdio.h>
45 #include <stdlib.h>
46 #include <string.h>
47 
48 #include "db-int.h"
49 #include "btree.h"
50 
51 static int	 bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *));
52 static PAGE	*bt_page
53 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
54 static int	 bt_preserve __P((BTREE *, db_pgno_t));
55 static PAGE	*bt_psplit
56 		    __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t));
57 static PAGE	*bt_root
58 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
59 static int	 bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
60 static recno_t	 rec_total __P((PAGE *));
61 
62 #ifdef STATISTICS
63 u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
64 #endif
65 
66 /*
67  * __BT_SPLIT -- Split the tree.
68  *
69  * Parameters:
70  *	t:	tree
71  *	sp:	page to split
72  *	key:	key to insert
73  *	data:	data to insert
74  *	flags:	BIGKEY/BIGDATA flags
75  *	ilen:	insert length
76  *	skip:	index to leave open
77  *
78  * Returns:
79  *	RET_ERROR, RET_SUCCESS
80  */
81 int
__bt_split(t,sp,key,data,flags,ilen,argskip)82 __bt_split(t, sp, key, data, flags, ilen, argskip)
83 	BTREE *t;
84 	PAGE *sp;
85 	const DBT *key, *data;
86 	int flags;
87 	size_t ilen;
88 	u_int32_t argskip;
89 {
90 	BINTERNAL *bi = NULL;
91 	BLEAF *bl = NULL, *tbl;
92 	DBT a, b;
93 	EPGNO *parent;
94 	PAGE *h, *l, *r, *lchild, *rchild;
95 	indx_t nxtindex;
96 	u_int16_t skip;
97 	u_int32_t n, nbytes, nksize = 0;
98 	int parentsplit;
99 	char *dest;
100 
101 	/*
102 	 * Split the page into two pages, l and r.  The split routines return
103 	 * a pointer to the page into which the key should be inserted and with
104 	 * skip set to the offset which should be used.  Additionally, l and r
105 	 * are pinned.
106 	 */
107 	skip = argskip;
108 	h = sp->pgno == P_ROOT ?
109 	    bt_root(t, sp, &l, &r, &skip, ilen) :
110 	    bt_page(t, sp, &l, &r, &skip, ilen);
111 	if (h == NULL)
112 		return (RET_ERROR);
113 
114 	/*
115 	 * Insert the new key/data pair into the leaf page.  (Key inserts
116 	 * always cause a leaf page to split first.)
117 	 */
118 	h->linp[skip] = h->upper -= ilen;
119 	dest = (char *)h + h->upper;
120 	if (F_ISSET(t, R_RECNO))
121 		WR_RLEAF(dest, data, flags)
122 	else
123 		WR_BLEAF(dest, key, data, flags)
124 
125 	/* If the root page was split, make it look right. */
126 	if (sp->pgno == P_ROOT &&
127 	    (F_ISSET(t, R_RECNO) ?
128 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
129 		goto err2;
130 
131 	/*
132 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
133 	 * were traversed when we searched for the page that split.  Each stack
134 	 * entry is a page number and a page index offset.  The offset is for
135 	 * the page traversed on the search.  We've just split a page, so we
136 	 * have to insert a new key into the parent page.
137 	 *
138 	 * If the insert into the parent page causes it to split, may have to
139 	 * continue splitting all the way up the tree.  We stop if the root
140 	 * splits or the page inserted into didn't have to split to hold the
141 	 * new key.  Some algorithms replace the key for the old page as well
142 	 * as the new page.  We don't, as there's no reason to believe that the
143 	 * first key on the old page is any better than the key we have, and,
144 	 * in the case of a key being placed at index 0 causing the split, the
145 	 * key is unavailable.
146 	 *
147 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
148 	 * and right pages pinned while working on the parent.   The 5 are the
149 	 * two children, left parent and right parent (when the parent splits)
150 	 * and the root page or the overflow key page when calling bt_preserve.
151 	 * This code must make sure that all pins are released other than the
152 	 * root page or overflow page which is unlocked elsewhere.
153 	 */
154 	while ((parent = BT_POP(t)) != NULL) {
155 		lchild = l;
156 		rchild = r;
157 
158 		/* Get the parent page. */
159 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
160 			goto err2;
161 
162 	 	/*
163 		 * The new key goes ONE AFTER the index, because the split
164 		 * was to the right.
165 		 */
166 		skip = parent->index + 1;
167 
168 		/*
169 		 * Calculate the space needed on the parent page.
170 		 *
171 		 * Prefix trees: space hack when inserting into BINTERNAL
172 		 * pages.  Retain only what's needed to distinguish between
173 		 * the new entry and the LAST entry on the page to its left.
174 		 * If the keys compare equal, retain the entire key.  Note,
175 		 * we don't touch overflow keys, and the entire key must be
176 		 * retained for the next-to-left most key on the leftmost
177 		 * page of each level, or the search will fail.  Applicable
178 		 * ONLY to internal pages that have leaf pages as children.
179 		 * Further reduction of the key between pairs of internal
180 		 * pages loses too much information.
181 		 */
182 		switch (rchild->flags & P_TYPE) {
183 		case P_BINTERNAL:
184 			bi = GETBINTERNAL(rchild, 0);
185 			nbytes = NBINTERNAL(bi->ksize);
186 			break;
187 		case P_BLEAF:
188 			bl = GETBLEAF(rchild, 0);
189 			nbytes = NBINTERNAL(bl->ksize);
190 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
191 			    (h->prevpg != P_INVALID || skip > 1)) {
192 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
193 				a.size = tbl->ksize;
194 				a.data = tbl->bytes;
195 				b.size = bl->ksize;
196 				b.data = bl->bytes;
197 				nksize = t->bt_pfx(&a, &b);
198 				n = NBINTERNAL(nksize);
199 				if (n < nbytes) {
200 #ifdef STATISTICS
201 					bt_pfxsaved += nbytes - n;
202 #endif
203 					nbytes = n;
204 				} else
205 					nksize = 0;
206 			} else
207 				nksize = 0;
208 			break;
209 		case P_RINTERNAL:
210 		case P_RLEAF:
211 			nbytes = NRINTERNAL;
212 			break;
213 		default:
214 			abort();
215 		}
216 
217 		/* Split the parent page if necessary or shift the indices. */
218 		if ((u_int32_t)h->upper - (u_int32_t)h->lower
219 		    < nbytes + sizeof(indx_t)) {
220 			sp = h;
221 			h = h->pgno == P_ROOT ?
222 			    bt_root(t, h, &l, &r, &skip, nbytes) :
223 			    bt_page(t, h, &l, &r, &skip, nbytes);
224 			if (h == NULL)
225 				goto err1;
226 			parentsplit = 1;
227 		} else {
228 			if (skip < (nxtindex = NEXTINDEX(h)))
229 				memmove(h->linp + skip + 1, h->linp + skip,
230 				    (nxtindex - skip) * sizeof(indx_t));
231 			h->lower += sizeof(indx_t);
232 			parentsplit = 0;
233 		}
234 
235 		/* Insert the key into the parent page. */
236 		switch (rchild->flags & P_TYPE) {
237 		case P_BINTERNAL:
238 			h->linp[skip] = h->upper -= nbytes;
239 			dest = (char *)h + h->linp[skip];
240 			memmove(dest, bi, nbytes);
241 			((BINTERNAL *)(void *)dest)->pgno = rchild->pgno;
242 			break;
243 		case P_BLEAF:
244 			h->linp[skip] = h->upper -= nbytes;
245 			dest = (char *)h + h->linp[skip];
246 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
247 			    rchild->pgno, bl->flags & P_BIGKEY);
248 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
249 			if (bl->flags & P_BIGKEY) {
250 				db_pgno_t pgno;
251 				memcpy(&pgno, bl->bytes, sizeof(pgno));
252 				if (bt_preserve(t, pgno) == RET_ERROR)
253 					goto err1;
254 			}
255 			break;
256 		case P_RINTERNAL:
257 			/*
258 			 * Update the left page count.  If split
259 			 * added at index 0, fix the correct page.
260 			 */
261 			if (skip > 0)
262 				dest = (char *)h + h->linp[skip - 1];
263 			else
264 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
265 			((RINTERNAL *)(void *)dest)->nrecs = rec_total(lchild);
266 			((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
267 
268 			/* Update the right page count. */
269 			h->linp[skip] = h->upper -= nbytes;
270 			dest = (char *)h + h->linp[skip];
271 			((RINTERNAL *)(void *)dest)->nrecs = rec_total(rchild);
272 			((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
273 			break;
274 		case P_RLEAF:
275 			/*
276 			 * Update the left page count.  If split
277 			 * added at index 0, fix the correct page.
278 			 */
279 			if (skip > 0)
280 				dest = (char *)h + h->linp[skip - 1];
281 			else
282 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
283 			((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(lchild);
284 			((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
285 
286 			/* Update the right page count. */
287 			h->linp[skip] = h->upper -= nbytes;
288 			dest = (char *)h + h->linp[skip];
289 			((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(rchild);
290 			((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
291 			break;
292 		default:
293 			abort();
294 		}
295 
296 		/* Unpin the held pages. */
297 		if (!parentsplit) {
298 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
299 			break;
300 		}
301 
302 		/* If the root page was split, make it look right. */
303 		if (sp->pgno == P_ROOT &&
304 		    (F_ISSET(t, R_RECNO) ?
305 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
306 			goto err1;
307 
308 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
309 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
310 	}
311 
312 	/* Unpin the held pages. */
313 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
314 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
315 
316 	/* Clear any pages left on the stack. */
317 	return (RET_SUCCESS);
318 
319 	/*
320 	 * If something fails in the above loop we were already walking back
321 	 * up the tree and the tree is now inconsistent.  Nothing much we can
322 	 * do about it but release any memory we're holding.
323 	 */
324 err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
325 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
326 
327 err2:	mpool_put(t->bt_mp, l, 0);
328 	mpool_put(t->bt_mp, r, 0);
329 	__dbpanic(t->bt_dbp);
330 	return (RET_ERROR);
331 }
332 
333 /*
334  * BT_PAGE -- Split a non-root page of a btree.
335  *
336  * Parameters:
337  *	t:	tree
338  *	h:	root page
339  *	lp:	pointer to left page pointer
340  *	rp:	pointer to right page pointer
341  *	skip:	pointer to index to leave open
342  *	ilen:	insert length
343  *
344  * Returns:
345  *	Pointer to page in which to insert or NULL on error.
346  */
347 static PAGE *
bt_page(t,h,lp,rp,skip,ilen)348 bt_page(t, h, lp, rp, skip, ilen)
349 	BTREE *t;
350 	PAGE *h, **lp, **rp;
351 	indx_t *skip;
352 	size_t ilen;
353 {
354 	PAGE *l, *r, *tp;
355 	db_pgno_t npg;
356 
357 #ifdef STATISTICS
358 	++bt_split;
359 #endif
360 	/* Put the new right page for the split into place. */
361 	if ((r = __bt_new(t, &npg)) == NULL)
362 		return (NULL);
363 	r->pgno = npg;
364 	r->lower = BTDATAOFF;
365 	r->upper = t->bt_psize;
366 	r->nextpg = h->nextpg;
367 	r->prevpg = h->pgno;
368 	r->flags = h->flags & P_TYPE;
369 
370 	/*
371 	 * If we're splitting the last page on a level because we're appending
372 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
373 	 * sorted.  Adding an empty page on the side of the level is less work
374 	 * and can push the fill factor much higher than normal.  If we're
375 	 * wrong it's no big deal, we'll just do the split the right way next
376 	 * time.  It may look like it's equally easy to do a similar hack for
377 	 * reverse sorted data, that is, split the tree left, but it's not.
378 	 * Don't even try.
379 	 */
380 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
381 #ifdef STATISTICS
382 		++bt_sortsplit;
383 #endif
384 		h->nextpg = r->pgno;
385 		r->lower = BTDATAOFF + sizeof(indx_t);
386 		*skip = 0;
387 		*lp = h;
388 		*rp = r;
389 		return (r);
390 	}
391 
392 	/* Put the new left page for the split into place. */
393 	if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
394 		mpool_put(t->bt_mp, r, 0);
395 		return (NULL);
396 	}
397 #if 1 /* def PURIFY */
398 	memset(l, 0xff, t->bt_psize);
399 #endif
400 	l->pgno = h->pgno;
401 	l->nextpg = r->pgno;
402 	l->prevpg = h->prevpg;
403 	l->lower = BTDATAOFF;
404 	l->upper = t->bt_psize;
405 	l->flags = h->flags & P_TYPE;
406 
407 	/* Fix up the previous pointer of the page after the split page. */
408 	if (h->nextpg != P_INVALID) {
409 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
410 			free(l);
411 			/* XXX mpool_free(t->bt_mp, r->pgno); */
412 			return (NULL);
413 		}
414 		tp->prevpg = r->pgno;
415 		mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
416 	}
417 
418 	/*
419 	 * Split right.  The key/data pairs aren't sorted in the btree page so
420 	 * it's simpler to copy the data from the split page onto two new pages
421 	 * instead of copying half the data to the right page and compacting
422 	 * the left page in place.  Since the left page can't change, we have
423 	 * to swap the original and the allocated left page after the split.
424 	 */
425 	tp = bt_psplit(t, h, l, r, skip, ilen);
426 
427 	/* Move the new left page onto the old left page. */
428 	memmove(h, l, t->bt_psize);
429 	if (tp == l)
430 		tp = h;
431 	free(l);
432 
433 	*lp = h;
434 	*rp = r;
435 	return (tp);
436 }
437 
438 /*
439  * BT_ROOT -- Split the root page of a btree.
440  *
441  * Parameters:
442  *	t:	tree
443  *	h:	root page
444  *	lp:	pointer to left page pointer
445  *	rp:	pointer to right page pointer
446  *	skip:	pointer to index to leave open
447  *	ilen:	insert length
448  *
449  * Returns:
450  *	Pointer to page in which to insert or NULL on error.
451  */
452 static PAGE *
bt_root(t,h,lp,rp,skip,ilen)453 bt_root(t, h, lp, rp, skip, ilen)
454 	BTREE *t;
455 	PAGE *h, **lp, **rp;
456 	indx_t *skip;
457 	size_t ilen;
458 {
459 	PAGE *l, *r, *tp;
460 	db_pgno_t lnpg, rnpg;
461 
462 #ifdef STATISTICS
463 	++bt_split;
464 	++bt_rootsplit;
465 #endif
466 	/* Put the new left and right pages for the split into place. */
467 	if ((l = __bt_new(t, &lnpg)) == NULL ||
468 	    (r = __bt_new(t, &rnpg)) == NULL)
469 		return (NULL);
470 	l->pgno = lnpg;
471 	r->pgno = rnpg;
472 	l->nextpg = r->pgno;
473 	r->prevpg = l->pgno;
474 	l->prevpg = r->nextpg = P_INVALID;
475 	l->lower = r->lower = BTDATAOFF;
476 	l->upper = r->upper = t->bt_psize;
477 	l->flags = r->flags = h->flags & P_TYPE;
478 
479 	/* Split the root page. */
480 	tp = bt_psplit(t, h, l, r, skip, ilen);
481 
482 	*lp = l;
483 	*rp = r;
484 	return (tp);
485 }
486 
487 /*
488  * BT_RROOT -- Fix up the recno root page after it has been split.
489  *
490  * Parameters:
491  *	t:	tree
492  *	h:	root page
493  *	l:	left page
494  *	r:	right page
495  *
496  * Returns:
497  *	RET_ERROR, RET_SUCCESS
498  */
499 static int
bt_rroot(t,h,l,r)500 bt_rroot(t, h, l, r)
501 	BTREE *t;
502 	PAGE *h, *l, *r;
503 {
504 	char *dest;
505 
506 	/* Insert the left and right keys, set the header information. */
507 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
508 	dest = (char *)h + h->upper;
509 	WR_RINTERNAL(dest,
510 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
511 
512 	h->linp[1] = h->upper -= NRINTERNAL;
513 	dest = (char *)h + h->upper;
514 	WR_RINTERNAL(dest,
515 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
516 
517 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
518 
519 	/* Unpin the root page, set to recno internal page. */
520 	h->flags &= ~P_TYPE;
521 	h->flags |= P_RINTERNAL;
522 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
523 
524 	return (RET_SUCCESS);
525 }
526 
527 /*
528  * BT_BROOT -- Fix up the btree root page after it has been split.
529  *
530  * Parameters:
531  *	t:	tree
532  *	h:	root page
533  *	l:	left page
534  *	r:	right page
535  *
536  * Returns:
537  *	RET_ERROR, RET_SUCCESS
538  */
539 static int
bt_broot(t,h,l,r)540 bt_broot(t, h, l, r)
541 	BTREE *t;
542 	PAGE *h, *l, *r;
543 {
544 	BINTERNAL *bi;
545 	BLEAF *bl;
546 	u_int32_t nbytes;
547 	char *dest;
548 
549 	/*
550 	 * If the root page was a leaf page, change it into an internal page.
551 	 * We copy the key we split on (but not the key's data, in the case of
552 	 * a leaf page) to the new root page.
553 	 *
554 	 * The btree comparison code guarantees that the left-most key on any
555 	 * level of the tree is never used, so it doesn't need to be filled in.
556 	 */
557 	nbytes = NBINTERNAL(0);
558 	h->linp[0] = h->upper = t->bt_psize - nbytes;
559 	dest = (char *)h + h->upper;
560 	WR_BINTERNAL(dest, 0, l->pgno, 0);
561 
562 	switch (h->flags & P_TYPE) {
563 	case P_BLEAF:
564 		bl = GETBLEAF(r, 0);
565 		nbytes = NBINTERNAL(bl->ksize);
566 		h->linp[1] = h->upper -= nbytes;
567 		dest = (char *)h + h->upper;
568 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
569 		memmove(dest, bl->bytes, bl->ksize);
570 
571 		/*
572 		 * If the key is on an overflow page, mark the overflow chain
573 		 * so it isn't deleted when the leaf copy of the key is deleted.
574 		 */
575 		if (bl->flags & P_BIGKEY) {
576 			db_pgno_t pgno;
577 			memcpy(&pgno, bl->bytes, sizeof(pgno));
578 			if (bt_preserve(t, pgno) == RET_ERROR)
579 				return (RET_ERROR);
580 		}
581 		break;
582 	case P_BINTERNAL:
583 		bi = GETBINTERNAL(r, 0);
584 		nbytes = NBINTERNAL(bi->ksize);
585 		h->linp[1] = h->upper -= nbytes;
586 		dest = (char *)h + h->upper;
587 		memmove(dest, bi, nbytes);
588 		((BINTERNAL *)(void *)dest)->pgno = r->pgno;
589 		break;
590 	default:
591 		abort();
592 	}
593 
594 	/* There are two keys on the page. */
595 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
596 
597 	/* Unpin the root page, set to btree internal page. */
598 	h->flags &= ~P_TYPE;
599 	h->flags |= P_BINTERNAL;
600 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
601 
602 	return (RET_SUCCESS);
603 }
604 
605 /*
606  * BT_PSPLIT -- Do the real work of splitting the page.
607  *
608  * Parameters:
609  *	t:	tree
610  *	h:	page to be split
611  *	l:	page to put lower half of data
612  *	r:	page to put upper half of data
613  *	pskip:	pointer to index to leave open
614  *	ilen:	insert length
615  *
616  * Returns:
617  *	Pointer to page in which to insert.
618  */
619 static PAGE *
bt_psplit(t,h,l,r,pskip,ilen)620 bt_psplit(t, h, l, r, pskip, ilen)
621 	BTREE *t;
622 	PAGE *h, *l, *r;
623 	indx_t *pskip;
624 	size_t ilen;
625 {
626 	BINTERNAL *bi;
627 	BLEAF *bl;
628 	CURSOR *c;
629 	RLEAF *rl;
630 	PAGE *rval;
631 	void *src;
632 	indx_t full, half, nxt, off, skip, top, used;
633 	u_int32_t nbytes;
634 	int bigkeycnt, isbigkey;
635 
636 	/*
637 	 * Split the data to the left and right pages.  Leave the skip index
638 	 * open.  Additionally, make some effort not to split on an overflow
639 	 * key.  This makes internal page processing faster and can save
640 	 * space as overflow keys used by internal pages are never deleted.
641 	 */
642 	bigkeycnt = 0;
643 	skip = *pskip;
644 	full = t->bt_psize - BTDATAOFF;
645 	half = full / 2;
646 	used = 0;
647 	src = NULL;	     /* silence gcc "uninitialized" warning */
648 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
649 		if (skip == off) {
650 			nbytes = ilen;
651 			isbigkey = 0;		/* XXX: not really known. */
652 		} else
653 			switch (h->flags & P_TYPE) {
654 			case P_BINTERNAL:
655 				src = bi = GETBINTERNAL(h, nxt);
656 				nbytes = NBINTERNAL(bi->ksize);
657 				isbigkey = bi->flags & P_BIGKEY;
658 				break;
659 			case P_BLEAF:
660 				src = bl = GETBLEAF(h, nxt);
661 				nbytes = NBLEAF(bl);
662 				isbigkey = bl->flags & P_BIGKEY;
663 				break;
664 			case P_RINTERNAL:
665 				src = GETRINTERNAL(h, nxt);
666 				nbytes = NRINTERNAL;
667 				isbigkey = 0;
668 				break;
669 			case P_RLEAF:
670 				src = rl = GETRLEAF(h, nxt);
671 				nbytes = NRLEAF(rl);
672 				isbigkey = 0;
673 				break;
674 			default:
675 				abort();
676 			}
677 
678 		/*
679 		 * If the key/data pairs are substantial fractions of the max
680 		 * possible size for the page, it's possible to get situations
681 		 * where we decide to try and copy too much onto the left page.
682 		 * Make sure that doesn't happen.
683 		 */
684 		if ((skip <= off && used + nbytes + sizeof(indx_t) >= full)
685 		    || nxt == top - 1) {
686 			--off;
687 			break;
688 		}
689 
690 		/* Copy the key/data pair, if not the skipped index. */
691 		if (skip != off) {
692 			++nxt;
693 
694 			l->linp[off] = l->upper -= nbytes;
695 			memmove((char *)l + l->upper, src, nbytes);
696 		}
697 
698 		used += nbytes + sizeof(indx_t);
699 		if (used >= half) {
700 			if (!isbigkey || bigkeycnt == 3)
701 				break;
702 			else
703 				++bigkeycnt;
704 		}
705 	}
706 
707 	/*
708 	 * Off is the last offset that's valid for the left page.
709 	 * Nxt is the first offset to be placed on the right page.
710 	 */
711 	l->lower += (off + 1) * sizeof(indx_t);
712 
713 	/*
714 	 * If splitting the page that the cursor was on, the cursor has to be
715 	 * adjusted to point to the same record as before the split.  If the
716 	 * cursor is at or past the skipped slot, the cursor is incremented by
717 	 * one.  If the cursor is on the right page, it is decremented by the
718 	 * number of records split to the left page.
719 	 */
720 	c = &t->bt_cursor;
721 	if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
722 		if (c->pg.index >= skip)
723 			++c->pg.index;
724 		if (c->pg.index < nxt)			/* Left page. */
725 			c->pg.pgno = l->pgno;
726 		else {					/* Right page. */
727 			c->pg.pgno = r->pgno;
728 			c->pg.index -= nxt;
729 		}
730 	}
731 
732 	/*
733 	 * If the skipped index was on the left page, just return that page.
734 	 * Otherwise, adjust the skip index to reflect the new position on
735 	 * the right page.
736 	 */
737 	if (skip <= off) {
738 		skip = (indx_t)-1;
739 		rval = l;
740 	} else {
741 		rval = r;
742 		*pskip -= nxt;
743 	}
744 
745 	for (off = 0; nxt < top; ++off) {
746 		if (skip == nxt) {
747 			++off;
748 			skip = (indx_t)-1;
749 		}
750 		switch (h->flags & P_TYPE) {
751 		case P_BINTERNAL:
752 			src = bi = GETBINTERNAL(h, nxt);
753 			nbytes = NBINTERNAL(bi->ksize);
754 			break;
755 		case P_BLEAF:
756 			src = bl = GETBLEAF(h, nxt);
757 			nbytes = NBLEAF(bl);
758 			break;
759 		case P_RINTERNAL:
760 			src = GETRINTERNAL(h, nxt);
761 			nbytes = NRINTERNAL;
762 			break;
763 		case P_RLEAF:
764 			src = rl = GETRLEAF(h, nxt);
765 			nbytes = NRLEAF(rl);
766 			break;
767 		default:
768 			abort();
769 		}
770 		++nxt;
771 		r->linp[off] = r->upper -= nbytes;
772 		memmove((char *)r + r->upper, src, nbytes);
773 	}
774 	r->lower += off * sizeof(indx_t);
775 
776 	/* If the key is being appended to the page, adjust the index. */
777 	if (skip == top)
778 		r->lower += sizeof(indx_t);
779 
780 	return (rval);
781 }
782 
783 /*
784  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
785  *
786  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
787  * record that references them gets deleted.  Chains pointed to by internal
788  * pages never get deleted.  This routine marks a chain as pointed to by an
789  * internal page.
790  *
791  * Parameters:
792  *	t:	tree
793  *	pg:	page number of first page in the chain.
794  *
795  * Returns:
796  *	RET_SUCCESS, RET_ERROR.
797  */
798 static int
bt_preserve(t,pg)799 bt_preserve(t, pg)
800 	BTREE *t;
801 	db_pgno_t pg;
802 {
803 	PAGE *h;
804 
805 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
806 		return (RET_ERROR);
807 	h->flags |= P_PRESERVE;
808 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
809 	return (RET_SUCCESS);
810 }
811 
812 /*
813  * REC_TOTAL -- Return the number of recno entries below a page.
814  *
815  * Parameters:
816  *	h:	page
817  *
818  * Returns:
819  *	The number of recno entries below a page.
820  *
821  * XXX
822  * These values could be set by the bt_psplit routine.  The problem is that the
823  * entry has to be popped off of the stack etc. or the values have to be passed
824  * all the way back to bt_split/bt_rroot and it's not very clean.
825  */
826 static recno_t
rec_total(h)827 rec_total(h)
828 	PAGE *h;
829 {
830 	recno_t recs;
831 	indx_t nxt, top;
832 
833 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
834 		recs += GETRINTERNAL(h, nxt)->nrecs;
835 	return (recs);
836 }
837