Home | History | Annotate | Line # | Download | only in btree
bt_split.c revision 1.4
      1  1.1  cgd /*-
      2  1.4  cgd  * Copyright (c) 1990, 1993, 1994
      3  1.1  cgd  *	The Regents of the University of California.  All rights reserved.
      4  1.1  cgd  *
      5  1.1  cgd  * This code is derived from software contributed to Berkeley by
      6  1.1  cgd  * Mike Olson.
      7  1.1  cgd  *
      8  1.1  cgd  * Redistribution and use in source and binary forms, with or without
      9  1.1  cgd  * modification, are permitted provided that the following conditions
     10  1.1  cgd  * are met:
     11  1.1  cgd  * 1. Redistributions of source code must retain the above copyright
     12  1.1  cgd  *    notice, this list of conditions and the following disclaimer.
     13  1.1  cgd  * 2. Redistributions in binary form must reproduce the above copyright
     14  1.1  cgd  *    notice, this list of conditions and the following disclaimer in the
     15  1.1  cgd  *    documentation and/or other materials provided with the distribution.
     16  1.1  cgd  * 3. All advertising materials mentioning features or use of this software
     17  1.1  cgd  *    must display the following acknowledgement:
     18  1.1  cgd  *	This product includes software developed by the University of
     19  1.1  cgd  *	California, Berkeley and its contributors.
     20  1.1  cgd  * 4. Neither the name of the University nor the names of its contributors
     21  1.1  cgd  *    may be used to endorse or promote products derived from this software
     22  1.1  cgd  *    without specific prior written permission.
     23  1.1  cgd  *
     24  1.1  cgd  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     25  1.1  cgd  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     26  1.1  cgd  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     27  1.1  cgd  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     28  1.1  cgd  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     29  1.1  cgd  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     30  1.1  cgd  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     31  1.1  cgd  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     32  1.1  cgd  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     33  1.1  cgd  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     34  1.1  cgd  * SUCH DAMAGE.
     35  1.1  cgd  */
     36  1.1  cgd 
     37  1.1  cgd #if defined(LIBC_SCCS) && !defined(lint)
     38  1.4  cgd static char sccsid[] = "@(#)bt_split.c	8.6 (Berkeley) 6/16/94";
     39  1.1  cgd #endif /* LIBC_SCCS and not lint */
     40  1.1  cgd 
     41  1.1  cgd #include <sys/types.h>
     42  1.1  cgd 
     43  1.1  cgd #include <limits.h>
     44  1.1  cgd #include <stdio.h>
     45  1.1  cgd #include <stdlib.h>
     46  1.1  cgd #include <string.h>
     47  1.1  cgd 
     48  1.1  cgd #include <db.h>
     49  1.1  cgd #include "btree.h"
     50  1.1  cgd 
     51  1.1  cgd static int	 bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *));
     52  1.1  cgd static PAGE	*bt_page
     53  1.4  cgd 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
     54  1.1  cgd static int	 bt_preserve __P((BTREE *, pgno_t));
     55  1.1  cgd static PAGE	*bt_psplit
     56  1.4  cgd 		    __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t));
     57  1.1  cgd static PAGE	*bt_root
     58  1.4  cgd 		    __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t));
     59  1.1  cgd static int	 bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *));
     60  1.1  cgd static recno_t	 rec_total __P((PAGE *));
     61  1.1  cgd 
     62  1.1  cgd #ifdef STATISTICS
     63  1.1  cgd u_long	bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
     64  1.1  cgd #endif
     65  1.1  cgd 
     66  1.1  cgd /*
     67  1.1  cgd  * __BT_SPLIT -- Split the tree.
     68  1.1  cgd  *
     69  1.1  cgd  * Parameters:
     70  1.1  cgd  *	t:	tree
     71  1.1  cgd  *	sp:	page to split
     72  1.1  cgd  *	key:	key to insert
     73  1.1  cgd  *	data:	data to insert
     74  1.1  cgd  *	flags:	BIGKEY/BIGDATA flags
     75  1.1  cgd  *	ilen:	insert length
     76  1.1  cgd  *	skip:	index to leave open
     77  1.1  cgd  *
     78  1.1  cgd  * Returns:
     79  1.1  cgd  *	RET_ERROR, RET_SUCCESS
     80  1.1  cgd  */
     81  1.1  cgd int
     82  1.4  cgd __bt_split(t, sp, key, data, flags, ilen, argskip)
     83  1.1  cgd 	BTREE *t;
     84  1.1  cgd 	PAGE *sp;
     85  1.1  cgd 	const DBT *key, *data;
     86  1.4  cgd 	int flags;
     87  1.1  cgd 	size_t ilen;
     88  1.4  cgd 	u_int32_t argskip;
     89  1.1  cgd {
     90  1.1  cgd 	BINTERNAL *bi;
     91  1.1  cgd 	BLEAF *bl, *tbl;
     92  1.1  cgd 	DBT a, b;
     93  1.1  cgd 	EPGNO *parent;
     94  1.1  cgd 	PAGE *h, *l, *r, *lchild, *rchild;
     95  1.1  cgd 	indx_t nxtindex;
     96  1.4  cgd 	u_int16_t skip;
     97  1.4  cgd 	u_int32_t n, nbytes, nksize;
     98  1.1  cgd 	int parentsplit;
     99  1.1  cgd 	char *dest;
    100  1.1  cgd 
    101  1.1  cgd 	/*
    102  1.1  cgd 	 * Split the page into two pages, l and r.  The split routines return
    103  1.1  cgd 	 * a pointer to the page into which the key should be inserted and with
    104  1.1  cgd 	 * skip set to the offset which should be used.  Additionally, l and r
    105  1.1  cgd 	 * are pinned.
    106  1.1  cgd 	 */
    107  1.4  cgd 	skip = argskip;
    108  1.1  cgd 	h = sp->pgno == P_ROOT ?
    109  1.1  cgd 	    bt_root(t, sp, &l, &r, &skip, ilen) :
    110  1.1  cgd 	    bt_page(t, sp, &l, &r, &skip, ilen);
    111  1.1  cgd 	if (h == NULL)
    112  1.1  cgd 		return (RET_ERROR);
    113  1.1  cgd 
    114  1.1  cgd 	/*
    115  1.1  cgd 	 * Insert the new key/data pair into the leaf page.  (Key inserts
    116  1.1  cgd 	 * always cause a leaf page to split first.)
    117  1.1  cgd 	 */
    118  1.1  cgd 	h->linp[skip] = h->upper -= ilen;
    119  1.1  cgd 	dest = (char *)h + h->upper;
    120  1.1  cgd 	if (ISSET(t, R_RECNO))
    121  1.1  cgd 		WR_RLEAF(dest, data, flags)
    122  1.1  cgd 	else
    123  1.1  cgd 		WR_BLEAF(dest, key, data, flags)
    124  1.1  cgd 
    125  1.1  cgd 	/* If the root page was split, make it look right. */
    126  1.1  cgd 	if (sp->pgno == P_ROOT &&
    127  1.1  cgd 	    (ISSET(t, R_RECNO) ?
    128  1.1  cgd 	    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    129  1.1  cgd 		goto err2;
    130  1.1  cgd 
    131  1.1  cgd 	/*
    132  1.1  cgd 	 * Now we walk the parent page stack -- a LIFO stack of the pages that
    133  1.1  cgd 	 * were traversed when we searched for the page that split.  Each stack
    134  1.1  cgd 	 * entry is a page number and a page index offset.  The offset is for
    135  1.1  cgd 	 * the page traversed on the search.  We've just split a page, so we
    136  1.1  cgd 	 * have to insert a new key into the parent page.
    137  1.1  cgd 	 *
    138  1.1  cgd 	 * If the insert into the parent page causes it to split, may have to
    139  1.1  cgd 	 * continue splitting all the way up the tree.  We stop if the root
    140  1.1  cgd 	 * splits or the page inserted into didn't have to split to hold the
    141  1.1  cgd 	 * new key.  Some algorithms replace the key for the old page as well
    142  1.1  cgd 	 * as the new page.  We don't, as there's no reason to believe that the
    143  1.1  cgd 	 * first key on the old page is any better than the key we have, and,
    144  1.1  cgd 	 * in the case of a key being placed at index 0 causing the split, the
    145  1.1  cgd 	 * key is unavailable.
    146  1.1  cgd 	 *
    147  1.1  cgd 	 * There are a maximum of 5 pages pinned at any time.  We keep the left
    148  1.1  cgd 	 * and right pages pinned while working on the parent.   The 5 are the
    149  1.1  cgd 	 * two children, left parent and right parent (when the parent splits)
    150  1.1  cgd 	 * and the root page or the overflow key page when calling bt_preserve.
    151  1.1  cgd 	 * This code must make sure that all pins are released other than the
    152  1.1  cgd 	 * root page or overflow page which is unlocked elsewhere.
    153  1.1  cgd 	 */
    154  1.1  cgd 	while ((parent = BT_POP(t)) != NULL) {
    155  1.1  cgd 		lchild = l;
    156  1.1  cgd 		rchild = r;
    157  1.1  cgd 
    158  1.1  cgd 		/* Get the parent page. */
    159  1.1  cgd 		if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
    160  1.1  cgd 			goto err2;
    161  1.1  cgd 
    162  1.1  cgd 	 	/*
    163  1.1  cgd 		 * The new key goes ONE AFTER the index, because the split
    164  1.1  cgd 		 * was to the right.
    165  1.1  cgd 		 */
    166  1.1  cgd 		skip = parent->index + 1;
    167  1.1  cgd 
    168  1.1  cgd 		/*
    169  1.1  cgd 		 * Calculate the space needed on the parent page.
    170  1.1  cgd 		 *
    171  1.1  cgd 		 * Prefix trees: space hack when inserting into BINTERNAL
    172  1.1  cgd 		 * pages.  Retain only what's needed to distinguish between
    173  1.1  cgd 		 * the new entry and the LAST entry on the page to its left.
    174  1.1  cgd 		 * If the keys compare equal, retain the entire key.  Note,
    175  1.1  cgd 		 * we don't touch overflow keys, and the entire key must be
    176  1.1  cgd 		 * retained for the next-to-left most key on the leftmost
    177  1.1  cgd 		 * page of each level, or the search will fail.  Applicable
    178  1.1  cgd 		 * ONLY to internal pages that have leaf pages as children.
    179  1.1  cgd 		 * Further reduction of the key between pairs of internal
    180  1.1  cgd 		 * pages loses too much information.
    181  1.1  cgd 		 */
    182  1.1  cgd 		switch (rchild->flags & P_TYPE) {
    183  1.1  cgd 		case P_BINTERNAL:
    184  1.1  cgd 			bi = GETBINTERNAL(rchild, 0);
    185  1.1  cgd 			nbytes = NBINTERNAL(bi->ksize);
    186  1.1  cgd 			break;
    187  1.1  cgd 		case P_BLEAF:
    188  1.1  cgd 			bl = GETBLEAF(rchild, 0);
    189  1.1  cgd 			nbytes = NBINTERNAL(bl->ksize);
    190  1.1  cgd 			if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
    191  1.1  cgd 			    (h->prevpg != P_INVALID || skip > 1)) {
    192  1.1  cgd 				tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
    193  1.1  cgd 				a.size = tbl->ksize;
    194  1.1  cgd 				a.data = tbl->bytes;
    195  1.1  cgd 				b.size = bl->ksize;
    196  1.1  cgd 				b.data = bl->bytes;
    197  1.1  cgd 				nksize = t->bt_pfx(&a, &b);
    198  1.1  cgd 				n = NBINTERNAL(nksize);
    199  1.1  cgd 				if (n < nbytes) {
    200  1.1  cgd #ifdef STATISTICS
    201  1.1  cgd 					bt_pfxsaved += nbytes - n;
    202  1.1  cgd #endif
    203  1.1  cgd 					nbytes = n;
    204  1.1  cgd 				} else
    205  1.1  cgd 					nksize = 0;
    206  1.1  cgd 			} else
    207  1.1  cgd 				nksize = 0;
    208  1.1  cgd 			break;
    209  1.1  cgd 		case P_RINTERNAL:
    210  1.1  cgd 		case P_RLEAF:
    211  1.1  cgd 			nbytes = NRINTERNAL;
    212  1.1  cgd 			break;
    213  1.1  cgd 		default:
    214  1.1  cgd 			abort();
    215  1.1  cgd 		}
    216  1.1  cgd 
    217  1.1  cgd 		/* Split the parent page if necessary or shift the indices. */
    218  1.1  cgd 		if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
    219  1.1  cgd 			sp = h;
    220  1.1  cgd 			h = h->pgno == P_ROOT ?
    221  1.1  cgd 			    bt_root(t, h, &l, &r, &skip, nbytes) :
    222  1.1  cgd 			    bt_page(t, h, &l, &r, &skip, nbytes);
    223  1.1  cgd 			if (h == NULL)
    224  1.1  cgd 				goto err1;
    225  1.1  cgd 			parentsplit = 1;
    226  1.1  cgd 		} else {
    227  1.1  cgd 			if (skip < (nxtindex = NEXTINDEX(h)))
    228  1.1  cgd 				memmove(h->linp + skip + 1, h->linp + skip,
    229  1.1  cgd 				    (nxtindex - skip) * sizeof(indx_t));
    230  1.1  cgd 			h->lower += sizeof(indx_t);
    231  1.1  cgd 			parentsplit = 0;
    232  1.1  cgd 		}
    233  1.1  cgd 
    234  1.1  cgd 		/* Insert the key into the parent page. */
    235  1.1  cgd 		switch(rchild->flags & P_TYPE) {
    236  1.1  cgd 		case P_BINTERNAL:
    237  1.1  cgd 			h->linp[skip] = h->upper -= nbytes;
    238  1.1  cgd 			dest = (char *)h + h->linp[skip];
    239  1.1  cgd 			memmove(dest, bi, nbytes);
    240  1.1  cgd 			((BINTERNAL *)dest)->pgno = rchild->pgno;
    241  1.1  cgd 			break;
    242  1.1  cgd 		case P_BLEAF:
    243  1.1  cgd 			h->linp[skip] = h->upper -= nbytes;
    244  1.1  cgd 			dest = (char *)h + h->linp[skip];
    245  1.1  cgd 			WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
    246  1.1  cgd 			    rchild->pgno, bl->flags & P_BIGKEY);
    247  1.1  cgd 			memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
    248  1.1  cgd 			if (bl->flags & P_BIGKEY &&
    249  1.1  cgd 			    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
    250  1.1  cgd 				goto err1;
    251  1.1  cgd 			break;
    252  1.1  cgd 		case P_RINTERNAL:
    253  1.1  cgd 			/*
    254  1.1  cgd 			 * Update the left page count.  If split
    255  1.1  cgd 			 * added at index 0, fix the correct page.
    256  1.1  cgd 			 */
    257  1.1  cgd 			if (skip > 0)
    258  1.1  cgd 				dest = (char *)h + h->linp[skip - 1];
    259  1.1  cgd 			else
    260  1.1  cgd 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
    261  1.1  cgd 			((RINTERNAL *)dest)->nrecs = rec_total(lchild);
    262  1.1  cgd 			((RINTERNAL *)dest)->pgno = lchild->pgno;
    263  1.1  cgd 
    264  1.1  cgd 			/* Update the right page count. */
    265  1.1  cgd 			h->linp[skip] = h->upper -= nbytes;
    266  1.1  cgd 			dest = (char *)h + h->linp[skip];
    267  1.1  cgd 			((RINTERNAL *)dest)->nrecs = rec_total(rchild);
    268  1.1  cgd 			((RINTERNAL *)dest)->pgno = rchild->pgno;
    269  1.1  cgd 			break;
    270  1.1  cgd 		case P_RLEAF:
    271  1.1  cgd 			/*
    272  1.1  cgd 			 * Update the left page count.  If split
    273  1.1  cgd 			 * added at index 0, fix the correct page.
    274  1.1  cgd 			 */
    275  1.1  cgd 			if (skip > 0)
    276  1.1  cgd 				dest = (char *)h + h->linp[skip - 1];
    277  1.1  cgd 			else
    278  1.1  cgd 				dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
    279  1.1  cgd 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
    280  1.1  cgd 			((RINTERNAL *)dest)->pgno = lchild->pgno;
    281  1.1  cgd 
    282  1.1  cgd 			/* Update the right page count. */
    283  1.1  cgd 			h->linp[skip] = h->upper -= nbytes;
    284  1.1  cgd 			dest = (char *)h + h->linp[skip];
    285  1.1  cgd 			((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
    286  1.1  cgd 			((RINTERNAL *)dest)->pgno = rchild->pgno;
    287  1.1  cgd 			break;
    288  1.1  cgd 		default:
    289  1.1  cgd 			abort();
    290  1.1  cgd 		}
    291  1.1  cgd 
    292  1.1  cgd 		/* Unpin the held pages. */
    293  1.1  cgd 		if (!parentsplit) {
    294  1.1  cgd 			mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    295  1.1  cgd 			break;
    296  1.1  cgd 		}
    297  1.1  cgd 
    298  1.1  cgd 		/* If the root page was split, make it look right. */
    299  1.1  cgd 		if (sp->pgno == P_ROOT &&
    300  1.1  cgd 		    (ISSET(t, R_RECNO) ?
    301  1.1  cgd 		    bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
    302  1.1  cgd 			goto err1;
    303  1.1  cgd 
    304  1.1  cgd 		mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    305  1.1  cgd 		mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    306  1.1  cgd 	}
    307  1.1  cgd 
    308  1.1  cgd 	/* Unpin the held pages. */
    309  1.1  cgd 	mpool_put(t->bt_mp, l, MPOOL_DIRTY);
    310  1.1  cgd 	mpool_put(t->bt_mp, r, MPOOL_DIRTY);
    311  1.1  cgd 
    312  1.1  cgd 	/* Clear any pages left on the stack. */
    313  1.1  cgd 	return (RET_SUCCESS);
    314  1.1  cgd 
    315  1.1  cgd 	/*
    316  1.1  cgd 	 * If something fails in the above loop we were already walking back
    317  1.1  cgd 	 * up the tree and the tree is now inconsistent.  Nothing much we can
    318  1.1  cgd 	 * do about it but release any memory we're holding.
    319  1.1  cgd 	 */
    320  1.1  cgd err1:	mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
    321  1.1  cgd 	mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
    322  1.1  cgd 
    323  1.1  cgd err2:	mpool_put(t->bt_mp, l, 0);
    324  1.1  cgd 	mpool_put(t->bt_mp, r, 0);
    325  1.1  cgd 	__dbpanic(t->bt_dbp);
    326  1.1  cgd 	return (RET_ERROR);
    327  1.1  cgd }
    328  1.1  cgd 
    329  1.1  cgd /*
    330  1.1  cgd  * BT_PAGE -- Split a non-root page of a btree.
    331  1.1  cgd  *
    332  1.1  cgd  * Parameters:
    333  1.1  cgd  *	t:	tree
    334  1.1  cgd  *	h:	root page
    335  1.1  cgd  *	lp:	pointer to left page pointer
    336  1.1  cgd  *	rp:	pointer to right page pointer
    337  1.1  cgd  *	skip:	pointer to index to leave open
    338  1.1  cgd  *	ilen:	insert length
    339  1.1  cgd  *
    340  1.1  cgd  * Returns:
    341  1.1  cgd  *	Pointer to page in which to insert or NULL on error.
    342  1.1  cgd  */
    343  1.1  cgd static PAGE *
    344  1.1  cgd bt_page(t, h, lp, rp, skip, ilen)
    345  1.1  cgd 	BTREE *t;
    346  1.1  cgd 	PAGE *h, **lp, **rp;
    347  1.4  cgd 	indx_t *skip;
    348  1.1  cgd 	size_t ilen;
    349  1.1  cgd {
    350  1.1  cgd 	PAGE *l, *r, *tp;
    351  1.1  cgd 	pgno_t npg;
    352  1.1  cgd 
    353  1.1  cgd #ifdef STATISTICS
    354  1.1  cgd 	++bt_split;
    355  1.1  cgd #endif
    356  1.1  cgd 	/* Put the new right page for the split into place. */
    357  1.1  cgd 	if ((r = __bt_new(t, &npg)) == NULL)
    358  1.1  cgd 		return (NULL);
    359  1.1  cgd 	r->pgno = npg;
    360  1.1  cgd 	r->lower = BTDATAOFF;
    361  1.1  cgd 	r->upper = t->bt_psize;
    362  1.1  cgd 	r->nextpg = h->nextpg;
    363  1.1  cgd 	r->prevpg = h->pgno;
    364  1.1  cgd 	r->flags = h->flags & P_TYPE;
    365  1.1  cgd 
    366  1.1  cgd 	/*
    367  1.1  cgd 	 * If we're splitting the last page on a level because we're appending
    368  1.1  cgd 	 * a key to it (skip is NEXTINDEX()), it's likely that the data is
    369  1.1  cgd 	 * sorted.  Adding an empty page on the side of the level is less work
    370  1.1  cgd 	 * and can push the fill factor much higher than normal.  If we're
    371  1.1  cgd 	 * wrong it's no big deal, we'll just do the split the right way next
    372  1.1  cgd 	 * time.  It may look like it's equally easy to do a similar hack for
    373  1.1  cgd 	 * reverse sorted data, that is, split the tree left, but it's not.
    374  1.1  cgd 	 * Don't even try.
    375  1.1  cgd 	 */
    376  1.1  cgd 	if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
    377  1.1  cgd #ifdef STATISTICS
    378  1.1  cgd 		++bt_sortsplit;
    379  1.1  cgd #endif
    380  1.1  cgd 		h->nextpg = r->pgno;
    381  1.1  cgd 		r->lower = BTDATAOFF + sizeof(indx_t);
    382  1.1  cgd 		*skip = 0;
    383  1.1  cgd 		*lp = h;
    384  1.1  cgd 		*rp = r;
    385  1.1  cgd 		return (r);
    386  1.1  cgd 	}
    387  1.1  cgd 
    388  1.1  cgd 	/* Put the new left page for the split into place. */
    389  1.4  cgd 	if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
    390  1.1  cgd 		mpool_put(t->bt_mp, r, 0);
    391  1.1  cgd 		return (NULL);
    392  1.1  cgd 	}
    393  1.1  cgd 	l->pgno = h->pgno;
    394  1.1  cgd 	l->nextpg = r->pgno;
    395  1.1  cgd 	l->prevpg = h->prevpg;
    396  1.1  cgd 	l->lower = BTDATAOFF;
    397  1.1  cgd 	l->upper = t->bt_psize;
    398  1.1  cgd 	l->flags = h->flags & P_TYPE;
    399  1.1  cgd 
    400  1.1  cgd 	/* Fix up the previous pointer of the page after the split page. */
    401  1.1  cgd 	if (h->nextpg != P_INVALID) {
    402  1.1  cgd 		if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
    403  1.1  cgd 			free(l);
    404  1.1  cgd 			/* XXX mpool_free(t->bt_mp, r->pgno); */
    405  1.1  cgd 			return (NULL);
    406  1.1  cgd 		}
    407  1.1  cgd 		tp->prevpg = r->pgno;
    408  1.1  cgd 		mpool_put(t->bt_mp, tp, 0);
    409  1.1  cgd 	}
    410  1.1  cgd 
    411  1.1  cgd 	/*
    412  1.1  cgd 	 * Split right.  The key/data pairs aren't sorted in the btree page so
    413  1.1  cgd 	 * it's simpler to copy the data from the split page onto two new pages
    414  1.1  cgd 	 * instead of copying half the data to the right page and compacting
    415  1.1  cgd 	 * the left page in place.  Since the left page can't change, we have
    416  1.1  cgd 	 * to swap the original and the allocated left page after the split.
    417  1.1  cgd 	 */
    418  1.1  cgd 	tp = bt_psplit(t, h, l, r, skip, ilen);
    419  1.1  cgd 
    420  1.1  cgd 	/* Move the new left page onto the old left page. */
    421  1.1  cgd 	memmove(h, l, t->bt_psize);
    422  1.1  cgd 	if (tp == l)
    423  1.1  cgd 		tp = h;
    424  1.1  cgd 	free(l);
    425  1.1  cgd 
    426  1.1  cgd 	*lp = h;
    427  1.1  cgd 	*rp = r;
    428  1.1  cgd 	return (tp);
    429  1.1  cgd }
    430  1.1  cgd 
    431  1.1  cgd /*
    432  1.1  cgd  * BT_ROOT -- Split the root page of a btree.
    433  1.1  cgd  *
    434  1.1  cgd  * Parameters:
    435  1.1  cgd  *	t:	tree
    436  1.1  cgd  *	h:	root page
    437  1.1  cgd  *	lp:	pointer to left page pointer
    438  1.1  cgd  *	rp:	pointer to right page pointer
    439  1.1  cgd  *	skip:	pointer to index to leave open
    440  1.1  cgd  *	ilen:	insert length
    441  1.1  cgd  *
    442  1.1  cgd  * Returns:
    443  1.1  cgd  *	Pointer to page in which to insert or NULL on error.
    444  1.1  cgd  */
    445  1.1  cgd static PAGE *
    446  1.1  cgd bt_root(t, h, lp, rp, skip, ilen)
    447  1.1  cgd 	BTREE *t;
    448  1.1  cgd 	PAGE *h, **lp, **rp;
    449  1.4  cgd 	indx_t *skip;
    450  1.1  cgd 	size_t ilen;
    451  1.1  cgd {
    452  1.1  cgd 	PAGE *l, *r, *tp;
    453  1.1  cgd 	pgno_t lnpg, rnpg;
    454  1.1  cgd 
    455  1.1  cgd #ifdef STATISTICS
    456  1.1  cgd 	++bt_split;
    457  1.1  cgd 	++bt_rootsplit;
    458  1.1  cgd #endif
    459  1.1  cgd 	/* Put the new left and right pages for the split into place. */
    460  1.1  cgd 	if ((l = __bt_new(t, &lnpg)) == NULL ||
    461  1.1  cgd 	    (r = __bt_new(t, &rnpg)) == NULL)
    462  1.1  cgd 		return (NULL);
    463  1.1  cgd 	l->pgno = lnpg;
    464  1.1  cgd 	r->pgno = rnpg;
    465  1.1  cgd 	l->nextpg = r->pgno;
    466  1.1  cgd 	r->prevpg = l->pgno;
    467  1.1  cgd 	l->prevpg = r->nextpg = P_INVALID;
    468  1.1  cgd 	l->lower = r->lower = BTDATAOFF;
    469  1.1  cgd 	l->upper = r->upper = t->bt_psize;
    470  1.1  cgd 	l->flags = r->flags = h->flags & P_TYPE;
    471  1.1  cgd 
    472  1.1  cgd 	/* Split the root page. */
    473  1.1  cgd 	tp = bt_psplit(t, h, l, r, skip, ilen);
    474  1.1  cgd 
    475  1.1  cgd 	*lp = l;
    476  1.1  cgd 	*rp = r;
    477  1.1  cgd 	return (tp);
    478  1.1  cgd }
    479  1.1  cgd 
    480  1.1  cgd /*
    481  1.1  cgd  * BT_RROOT -- Fix up the recno root page after it has been split.
    482  1.1  cgd  *
    483  1.1  cgd  * Parameters:
    484  1.1  cgd  *	t:	tree
    485  1.1  cgd  *	h:	root page
    486  1.1  cgd  *	l:	left page
    487  1.1  cgd  *	r:	right page
    488  1.1  cgd  *
    489  1.1  cgd  * Returns:
    490  1.1  cgd  *	RET_ERROR, RET_SUCCESS
    491  1.1  cgd  */
    492  1.1  cgd static int
    493  1.1  cgd bt_rroot(t, h, l, r)
    494  1.1  cgd 	BTREE *t;
    495  1.1  cgd 	PAGE *h, *l, *r;
    496  1.1  cgd {
    497  1.1  cgd 	char *dest;
    498  1.1  cgd 
    499  1.1  cgd 	/* Insert the left and right keys, set the header information. */
    500  1.1  cgd 	h->linp[0] = h->upper = t->bt_psize - NRINTERNAL;
    501  1.1  cgd 	dest = (char *)h + h->upper;
    502  1.1  cgd 	WR_RINTERNAL(dest,
    503  1.1  cgd 	    l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
    504  1.1  cgd 
    505  1.1  cgd 	h->linp[1] = h->upper -= NRINTERNAL;
    506  1.1  cgd 	dest = (char *)h + h->upper;
    507  1.1  cgd 	WR_RINTERNAL(dest,
    508  1.1  cgd 	    r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
    509  1.1  cgd 
    510  1.1  cgd 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    511  1.1  cgd 
    512  1.1  cgd 	/* Unpin the root page, set to recno internal page. */
    513  1.1  cgd 	h->flags &= ~P_TYPE;
    514  1.1  cgd 	h->flags |= P_RINTERNAL;
    515  1.1  cgd 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    516  1.1  cgd 
    517  1.1  cgd 	return (RET_SUCCESS);
    518  1.1  cgd }
    519  1.1  cgd 
    520  1.1  cgd /*
    521  1.1  cgd  * BT_BROOT -- Fix up the btree root page after it has been split.
    522  1.1  cgd  *
    523  1.1  cgd  * Parameters:
    524  1.1  cgd  *	t:	tree
    525  1.1  cgd  *	h:	root page
    526  1.1  cgd  *	l:	left page
    527  1.1  cgd  *	r:	right page
    528  1.1  cgd  *
    529  1.1  cgd  * Returns:
    530  1.1  cgd  *	RET_ERROR, RET_SUCCESS
    531  1.1  cgd  */
    532  1.1  cgd static int
    533  1.1  cgd bt_broot(t, h, l, r)
    534  1.1  cgd 	BTREE *t;
    535  1.1  cgd 	PAGE *h, *l, *r;
    536  1.1  cgd {
    537  1.1  cgd 	BINTERNAL *bi;
    538  1.1  cgd 	BLEAF *bl;
    539  1.4  cgd 	u_int32_t nbytes;
    540  1.1  cgd 	char *dest;
    541  1.1  cgd 
    542  1.1  cgd 	/*
    543  1.1  cgd 	 * If the root page was a leaf page, change it into an internal page.
    544  1.1  cgd 	 * We copy the key we split on (but not the key's data, in the case of
    545  1.1  cgd 	 * a leaf page) to the new root page.
    546  1.1  cgd 	 *
    547  1.1  cgd 	 * The btree comparison code guarantees that the left-most key on any
    548  1.1  cgd 	 * level of the tree is never used, so it doesn't need to be filled in.
    549  1.1  cgd 	 */
    550  1.1  cgd 	nbytes = NBINTERNAL(0);
    551  1.1  cgd 	h->linp[0] = h->upper = t->bt_psize - nbytes;
    552  1.1  cgd 	dest = (char *)h + h->upper;
    553  1.1  cgd 	WR_BINTERNAL(dest, 0, l->pgno, 0);
    554  1.1  cgd 
    555  1.1  cgd 	switch(h->flags & P_TYPE) {
    556  1.1  cgd 	case P_BLEAF:
    557  1.1  cgd 		bl = GETBLEAF(r, 0);
    558  1.1  cgd 		nbytes = NBINTERNAL(bl->ksize);
    559  1.1  cgd 		h->linp[1] = h->upper -= nbytes;
    560  1.1  cgd 		dest = (char *)h + h->upper;
    561  1.1  cgd 		WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
    562  1.1  cgd 		memmove(dest, bl->bytes, bl->ksize);
    563  1.1  cgd 
    564  1.1  cgd 		/*
    565  1.1  cgd 		 * If the key is on an overflow page, mark the overflow chain
    566  1.1  cgd 		 * so it isn't deleted when the leaf copy of the key is deleted.
    567  1.1  cgd 		 */
    568  1.1  cgd 		if (bl->flags & P_BIGKEY &&
    569  1.1  cgd 		    bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
    570  1.1  cgd 			return (RET_ERROR);
    571  1.1  cgd 		break;
    572  1.1  cgd 	case P_BINTERNAL:
    573  1.1  cgd 		bi = GETBINTERNAL(r, 0);
    574  1.1  cgd 		nbytes = NBINTERNAL(bi->ksize);
    575  1.1  cgd 		h->linp[1] = h->upper -= nbytes;
    576  1.1  cgd 		dest = (char *)h + h->upper;
    577  1.1  cgd 		memmove(dest, bi, nbytes);
    578  1.1  cgd 		((BINTERNAL *)dest)->pgno = r->pgno;
    579  1.1  cgd 		break;
    580  1.1  cgd 	default:
    581  1.1  cgd 		abort();
    582  1.1  cgd 	}
    583  1.1  cgd 
    584  1.1  cgd 	/* There are two keys on the page. */
    585  1.1  cgd 	h->lower = BTDATAOFF + 2 * sizeof(indx_t);
    586  1.1  cgd 
    587  1.1  cgd 	/* Unpin the root page, set to btree internal page. */
    588  1.1  cgd 	h->flags &= ~P_TYPE;
    589  1.1  cgd 	h->flags |= P_BINTERNAL;
    590  1.1  cgd 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    591  1.1  cgd 
    592  1.1  cgd 	return (RET_SUCCESS);
    593  1.1  cgd }
    594  1.1  cgd 
    595  1.1  cgd /*
    596  1.1  cgd  * BT_PSPLIT -- Do the real work of splitting the page.
    597  1.1  cgd  *
    598  1.1  cgd  * Parameters:
    599  1.1  cgd  *	t:	tree
    600  1.1  cgd  *	h:	page to be split
    601  1.1  cgd  *	l:	page to put lower half of data
    602  1.1  cgd  *	r:	page to put upper half of data
    603  1.1  cgd  *	pskip:	pointer to index to leave open
    604  1.1  cgd  *	ilen:	insert length
    605  1.1  cgd  *
    606  1.1  cgd  * Returns:
    607  1.1  cgd  *	Pointer to page in which to insert.
    608  1.1  cgd  */
    609  1.1  cgd static PAGE *
    610  1.1  cgd bt_psplit(t, h, l, r, pskip, ilen)
    611  1.1  cgd 	BTREE *t;
    612  1.1  cgd 	PAGE *h, *l, *r;
    613  1.4  cgd 	indx_t *pskip;
    614  1.1  cgd 	size_t ilen;
    615  1.1  cgd {
    616  1.1  cgd 	BINTERNAL *bi;
    617  1.1  cgd 	BLEAF *bl;
    618  1.1  cgd 	RLEAF *rl;
    619  1.1  cgd 	EPGNO *c;
    620  1.1  cgd 	PAGE *rval;
    621  1.1  cgd 	void *src;
    622  1.1  cgd 	indx_t full, half, nxt, off, skip, top, used;
    623  1.4  cgd 	u_int32_t nbytes;
    624  1.1  cgd 	int bigkeycnt, isbigkey;
    625  1.1  cgd 
    626  1.1  cgd 	/*
    627  1.1  cgd 	 * Split the data to the left and right pages.  Leave the skip index
    628  1.1  cgd 	 * open.  Additionally, make some effort not to split on an overflow
    629  1.1  cgd 	 * key.  This makes internal page processing faster and can save
    630  1.1  cgd 	 * space as overflow keys used by internal pages are never deleted.
    631  1.1  cgd 	 */
    632  1.1  cgd 	bigkeycnt = 0;
    633  1.1  cgd 	skip = *pskip;
    634  1.1  cgd 	full = t->bt_psize - BTDATAOFF;
    635  1.1  cgd 	half = full / 2;
    636  1.1  cgd 	used = 0;
    637  1.1  cgd 	for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
    638  1.1  cgd 		if (skip == off) {
    639  1.1  cgd 			nbytes = ilen;
    640  1.1  cgd 			isbigkey = 0;		/* XXX: not really known. */
    641  1.1  cgd 		} else
    642  1.1  cgd 			switch (h->flags & P_TYPE) {
    643  1.1  cgd 			case P_BINTERNAL:
    644  1.1  cgd 				src = bi = GETBINTERNAL(h, nxt);
    645  1.1  cgd 				nbytes = NBINTERNAL(bi->ksize);
    646  1.1  cgd 				isbigkey = bi->flags & P_BIGKEY;
    647  1.1  cgd 				break;
    648  1.1  cgd 			case P_BLEAF:
    649  1.1  cgd 				src = bl = GETBLEAF(h, nxt);
    650  1.1  cgd 				nbytes = NBLEAF(bl);
    651  1.1  cgd 				isbigkey = bl->flags & P_BIGKEY;
    652  1.1  cgd 				break;
    653  1.1  cgd 			case P_RINTERNAL:
    654  1.1  cgd 				src = GETRINTERNAL(h, nxt);
    655  1.1  cgd 				nbytes = NRINTERNAL;
    656  1.1  cgd 				isbigkey = 0;
    657  1.1  cgd 				break;
    658  1.1  cgd 			case P_RLEAF:
    659  1.1  cgd 				src = rl = GETRLEAF(h, nxt);
    660  1.1  cgd 				nbytes = NRLEAF(rl);
    661  1.1  cgd 				isbigkey = 0;
    662  1.1  cgd 				break;
    663  1.1  cgd 			default:
    664  1.1  cgd 				abort();
    665  1.1  cgd 			}
    666  1.1  cgd 
    667  1.1  cgd 		/*
    668  1.1  cgd 		 * If the key/data pairs are substantial fractions of the max
    669  1.1  cgd 		 * possible size for the page, it's possible to get situations
    670  1.1  cgd 		 * where we decide to try and copy too much onto the left page.
    671  1.1  cgd 		 * Make sure that doesn't happen.
    672  1.1  cgd 		 */
    673  1.1  cgd 		if (skip <= off && used + nbytes >= full) {
    674  1.1  cgd 			--off;
    675  1.1  cgd 			break;
    676  1.1  cgd 		}
    677  1.1  cgd 
    678  1.1  cgd 		/* Copy the key/data pair, if not the skipped index. */
    679  1.1  cgd 		if (skip != off) {
    680  1.1  cgd 			++nxt;
    681  1.1  cgd 
    682  1.1  cgd 			l->linp[off] = l->upper -= nbytes;
    683  1.1  cgd 			memmove((char *)l + l->upper, src, nbytes);
    684  1.1  cgd 		}
    685  1.1  cgd 
    686  1.1  cgd 		used += nbytes;
    687  1.1  cgd 		if (used >= half) {
    688  1.1  cgd 			if (!isbigkey || bigkeycnt == 3)
    689  1.1  cgd 				break;
    690  1.1  cgd 			else
    691  1.1  cgd 				++bigkeycnt;
    692  1.1  cgd 		}
    693  1.1  cgd 	}
    694  1.1  cgd 
    695  1.1  cgd 	/*
    696  1.1  cgd 	 * Off is the last offset that's valid for the left page.
    697  1.1  cgd 	 * Nxt is the first offset to be placed on the right page.
    698  1.1  cgd 	 */
    699  1.1  cgd 	l->lower += (off + 1) * sizeof(indx_t);
    700  1.1  cgd 
    701  1.1  cgd 	/*
    702  1.1  cgd 	 * If splitting the page that the cursor was on, the cursor has to be
    703  1.1  cgd 	 * adjusted to point to the same record as before the split.  If the
    704  1.1  cgd 	 * cursor is at or past the skipped slot, the cursor is incremented by
    705  1.1  cgd 	 * one.  If the cursor is on the right page, it is decremented by the
    706  1.1  cgd 	 * number of records split to the left page.
    707  1.1  cgd 	 *
    708  1.1  cgd 	 * Don't bother checking for the B_SEQINIT flag, the page number will
    709  1.1  cgd 	 * be P_INVALID.
    710  1.1  cgd 	 */
    711  1.1  cgd 	c = &t->bt_bcursor;
    712  1.1  cgd 	if (c->pgno == h->pgno) {
    713  1.1  cgd 		if (c->index >= skip)
    714  1.1  cgd 			++c->index;
    715  1.1  cgd 		if (c->index < nxt)			/* Left page. */
    716  1.1  cgd 			c->pgno = l->pgno;
    717  1.1  cgd 		else {					/* Right page. */
    718  1.1  cgd 			c->pgno = r->pgno;
    719  1.1  cgd 			c->index -= nxt;
    720  1.1  cgd 		}
    721  1.1  cgd 	}
    722  1.1  cgd 
    723  1.1  cgd 	/*
    724  1.1  cgd 	 * If the skipped index was on the left page, just return that page.
    725  1.1  cgd 	 * Otherwise, adjust the skip index to reflect the new position on
    726  1.1  cgd 	 * the right page.
    727  1.1  cgd 	 */
    728  1.1  cgd 	if (skip <= off) {
    729  1.1  cgd 		skip = 0;
    730  1.1  cgd 		rval = l;
    731  1.1  cgd 	} else {
    732  1.1  cgd 		rval = r;
    733  1.1  cgd 		*pskip -= nxt;
    734  1.1  cgd 	}
    735  1.1  cgd 
    736  1.1  cgd 	for (off = 0; nxt < top; ++off) {
    737  1.1  cgd 		if (skip == nxt) {
    738  1.1  cgd 			++off;
    739  1.1  cgd 			skip = 0;
    740  1.1  cgd 		}
    741  1.1  cgd 		switch (h->flags & P_TYPE) {
    742  1.1  cgd 		case P_BINTERNAL:
    743  1.1  cgd 			src = bi = GETBINTERNAL(h, nxt);
    744  1.1  cgd 			nbytes = NBINTERNAL(bi->ksize);
    745  1.1  cgd 			break;
    746  1.1  cgd 		case P_BLEAF:
    747  1.1  cgd 			src = bl = GETBLEAF(h, nxt);
    748  1.1  cgd 			nbytes = NBLEAF(bl);
    749  1.1  cgd 			break;
    750  1.1  cgd 		case P_RINTERNAL:
    751  1.1  cgd 			src = GETRINTERNAL(h, nxt);
    752  1.1  cgd 			nbytes = NRINTERNAL;
    753  1.1  cgd 			break;
    754  1.1  cgd 		case P_RLEAF:
    755  1.1  cgd 			src = rl = GETRLEAF(h, nxt);
    756  1.1  cgd 			nbytes = NRLEAF(rl);
    757  1.1  cgd 			break;
    758  1.1  cgd 		default:
    759  1.1  cgd 			abort();
    760  1.1  cgd 		}
    761  1.1  cgd 		++nxt;
    762  1.1  cgd 		r->linp[off] = r->upper -= nbytes;
    763  1.1  cgd 		memmove((char *)r + r->upper, src, nbytes);
    764  1.1  cgd 	}
    765  1.1  cgd 	r->lower += off * sizeof(indx_t);
    766  1.1  cgd 
    767  1.1  cgd 	/* If the key is being appended to the page, adjust the index. */
    768  1.1  cgd 	if (skip == top)
    769  1.1  cgd 		r->lower += sizeof(indx_t);
    770  1.1  cgd 
    771  1.1  cgd 	return (rval);
    772  1.1  cgd }
    773  1.1  cgd 
    774  1.1  cgd /*
    775  1.1  cgd  * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
    776  1.1  cgd  *
    777  1.1  cgd  * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
    778  1.1  cgd  * record that references them gets deleted.  Chains pointed to by internal
    779  1.1  cgd  * pages never get deleted.  This routine marks a chain as pointed to by an
    780  1.1  cgd  * internal page.
    781  1.1  cgd  *
    782  1.1  cgd  * Parameters:
    783  1.1  cgd  *	t:	tree
    784  1.1  cgd  *	pg:	page number of first page in the chain.
    785  1.1  cgd  *
    786  1.1  cgd  * Returns:
    787  1.1  cgd  *	RET_SUCCESS, RET_ERROR.
    788  1.1  cgd  */
    789  1.1  cgd static int
    790  1.1  cgd bt_preserve(t, pg)
    791  1.1  cgd 	BTREE *t;
    792  1.1  cgd 	pgno_t pg;
    793  1.1  cgd {
    794  1.1  cgd 	PAGE *h;
    795  1.1  cgd 
    796  1.1  cgd 	if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
    797  1.1  cgd 		return (RET_ERROR);
    798  1.1  cgd 	h->flags |= P_PRESERVE;
    799  1.1  cgd 	mpool_put(t->bt_mp, h, MPOOL_DIRTY);
    800  1.1  cgd 	return (RET_SUCCESS);
    801  1.1  cgd }
    802  1.1  cgd 
    803  1.1  cgd /*
    804  1.1  cgd  * REC_TOTAL -- Return the number of recno entries below a page.
    805  1.1  cgd  *
    806  1.1  cgd  * Parameters:
    807  1.1  cgd  *	h:	page
    808  1.1  cgd  *
    809  1.1  cgd  * Returns:
    810  1.1  cgd  *	The number of recno entries below a page.
    811  1.1  cgd  *
    812  1.1  cgd  * XXX
    813  1.1  cgd  * These values could be set by the bt_psplit routine.  The problem is that the
    814  1.1  cgd  * entry has to be popped off of the stack etc. or the values have to be passed
    815  1.1  cgd  * all the way back to bt_split/bt_rroot and it's not very clean.
    816  1.1  cgd  */
    817  1.1  cgd static recno_t
    818  1.1  cgd rec_total(h)
    819  1.1  cgd 	PAGE *h;
    820  1.1  cgd {
    821  1.1  cgd 	recno_t recs;
    822  1.1  cgd 	indx_t nxt, top;
    823  1.1  cgd 
    824  1.1  cgd 	for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
    825  1.1  cgd 		recs += GETRINTERNAL(h, nxt)->nrecs;
    826  1.1  cgd 	return (recs);
    827  1.1  cgd }
    828