146134Smao /*- 246134Smao * Copyright (c) 1990 The Regents of the University of California. 346134Smao * All rights reserved. 446134Smao * 546134Smao * This code is derived from software contributed to Berkeley by 646134Smao * Mike Olson. 746134Smao * 846134Smao * %sccs.include.redist.c% 946134Smao */ 1046134Smao 1146134Smao #if defined(LIBC_SCCS) && !defined(lint) 12*60234Sbostic static char sccsid[] = "@(#)bt_split.c 5.17 (Berkeley) 05/22/93"; 1346134Smao #endif /* LIBC_SCCS and not lint */ 1446134Smao 1546134Smao #include <sys/types.h> 1656741Sbostic 1750989Sbostic #define __DBINTERFACE_PRIVATE 1850989Sbostic #include <limits.h> 1950989Sbostic #include <stdio.h> 2046561Sbostic #include <stdlib.h> 2146561Sbostic #include <string.h> 2256741Sbostic 2357932Sbostic #include <db.h> 2446134Smao #include "btree.h" 2546134Smao 2657440Sbostic static int bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *)); 2757451Sbostic static PAGE *bt_page 2858078Sbostic __P((BTREE *, PAGE *, PAGE **, PAGE **, u_int *, size_t)); 2950989Sbostic static int bt_preserve __P((BTREE *, pgno_t)); 3057451Sbostic static PAGE *bt_psplit 3158078Sbostic __P((BTREE *, PAGE *, PAGE *, PAGE *, u_int *, size_t)); 3257451Sbostic static PAGE *bt_root 3358078Sbostic __P((BTREE *, PAGE *, PAGE **, PAGE **, u_int *, size_t)); 3450989Sbostic static int bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *)); 3550989Sbostic static recno_t rec_total __P((PAGE *)); 3650989Sbostic 3750989Sbostic #ifdef STATISTICS 3850989Sbostic u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved; 3950989Sbostic #endif 4050989Sbostic 4146134Smao /* 4250989Sbostic * __BT_SPLIT -- Split the tree. 4346134Smao * 4450989Sbostic * Parameters: 4550989Sbostic * t: tree 4657440Sbostic * sp: page to split 4750989Sbostic * key: key to insert 4850989Sbostic * data: data to insert 4950989Sbostic * flags: BIGKEY/BIGDATA flags 5057451Sbostic * ilen: insert length 5150989Sbostic * skip: index to leave open 5246134Smao * 5350989Sbostic * Returns: 5450989Sbostic * RET_ERROR, RET_SUCCESS 5546134Smao */ 5646134Smao int 5757451Sbostic __bt_split(t, sp, key, data, flags, ilen, skip) 5850989Sbostic BTREE *t; 5957440Sbostic PAGE *sp; 6050989Sbostic const DBT *key, *data; 6150989Sbostic u_long flags; 6257451Sbostic size_t ilen; 6358067Sbostic u_int skip; 6446134Smao { 6550989Sbostic BINTERNAL *bi; 6657451Sbostic BLEAF *bl, *tbl; 6750989Sbostic DBT a, b; 6850989Sbostic EPGNO *parent; 6957440Sbostic PAGE *h, *l, *r, *lchild, *rchild; 7057989Sbostic indx_t nxtindex; 7157451Sbostic size_t n, nbytes, nksize; 72*60234Sbostic int parentsplit; 7350989Sbostic char *dest; 7446134Smao 7546134Smao /* 7650989Sbostic * Split the page into two pages, l and r. The split routines return 7757440Sbostic * a pointer to the page into which the key should be inserted and with 7857440Sbostic * skip set to the offset which should be used. Additionally, l and r 7957440Sbostic * are pinned. 8046134Smao */ 8157440Sbostic h = sp->pgno == P_ROOT ? 8257451Sbostic bt_root(t, sp, &l, &r, &skip, ilen) : 8357451Sbostic bt_page(t, sp, &l, &r, &skip, ilen); 8450989Sbostic if (h == NULL) 8546134Smao return (RET_ERROR); 8646134Smao 8750989Sbostic /* 8857440Sbostic * Insert the new key/data pair into the leaf page. (Key inserts 8957440Sbostic * always cause a leaf page to split first.) 9050989Sbostic */ 9157451Sbostic h->linp[skip] = h->upper -= ilen; 9250989Sbostic dest = (char *)h + h->upper; 9360048Sbostic if (ISSET(t, R_RECNO)) 9450989Sbostic WR_RLEAF(dest, data, flags) 9551106Sbostic else 9650989Sbostic WR_BLEAF(dest, key, data, flags) 9746134Smao 9857440Sbostic /* If the root page was split, make it look right. */ 9957440Sbostic if (sp->pgno == P_ROOT && 10060048Sbostic (ISSET(t, R_RECNO) ? 10157440Sbostic bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 10257440Sbostic goto err2; 10357440Sbostic 10446134Smao /* 10550989Sbostic * Now we walk the parent page stack -- a LIFO stack of the pages that 10650989Sbostic * were traversed when we searched for the page that split. Each stack 10750989Sbostic * entry is a page number and a page index offset. The offset is for 10850989Sbostic * the page traversed on the search. We've just split a page, so we 10950989Sbostic * have to insert a new key into the parent page. 11050989Sbostic * 11150989Sbostic * If the insert into the parent page causes it to split, may have to 11250989Sbostic * continue splitting all the way up the tree. We stop if the root 11350989Sbostic * splits or the page inserted into didn't have to split to hold the 11450989Sbostic * new key. Some algorithms replace the key for the old page as well 11550989Sbostic * as the new page. We don't, as there's no reason to believe that the 11650989Sbostic * first key on the old page is any better than the key we have, and, 11750989Sbostic * in the case of a key being placed at index 0 causing the split, the 11850989Sbostic * key is unavailable. 11950989Sbostic * 12050989Sbostic * There are a maximum of 5 pages pinned at any time. We keep the left 12150989Sbostic * and right pages pinned while working on the parent. The 5 are the 12250989Sbostic * two children, left parent and right parent (when the parent splits) 12350989Sbostic * and the root page or the overflow key page when calling bt_preserve. 12450989Sbostic * This code must make sure that all pins are released other than the 12550989Sbostic * root page or overflow page which is unlocked elsewhere. 12646134Smao */ 127*60234Sbostic while ((parent = BT_POP(t)) != NULL) { 12850989Sbostic lchild = l; 12950989Sbostic rchild = r; 13046134Smao 13150989Sbostic /* Get the parent page. */ 13250989Sbostic if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL) 13350989Sbostic goto err2; 13446134Smao 13557451Sbostic /* 13657451Sbostic * The new key goes ONE AFTER the index, because the split 13757451Sbostic * was to the right. 13857451Sbostic */ 13950989Sbostic skip = parent->index + 1; 14046134Smao 14150989Sbostic /* 14250989Sbostic * Calculate the space needed on the parent page. 14350989Sbostic * 14457451Sbostic * Prefix trees: space hack when inserting into BINTERNAL 14557451Sbostic * pages. Retain only what's needed to distinguish between 14657451Sbostic * the new entry and the LAST entry on the page to its left. 14757451Sbostic * If the keys compare equal, retain the entire key. Note, 14857451Sbostic * we don't touch overflow keys, and the entire key must be 14957451Sbostic * retained for the next-to-left most key on the leftmost 15057451Sbostic * page of each level, or the search will fail. Applicable 15157451Sbostic * ONLY to internal pages that have leaf pages as children. 15257451Sbostic * Further reduction of the key between pairs of internal 15357451Sbostic * pages loses too much information. 15450989Sbostic */ 15550989Sbostic switch (rchild->flags & P_TYPE) { 15650989Sbostic case P_BINTERNAL: 15750989Sbostic bi = GETBINTERNAL(rchild, 0); 15850989Sbostic nbytes = NBINTERNAL(bi->ksize); 15950989Sbostic break; 16050989Sbostic case P_BLEAF: 16150989Sbostic bl = GETBLEAF(rchild, 0); 16256401Sbostic nbytes = NBINTERNAL(bl->ksize); 16357451Sbostic if (t->bt_pfx && !(bl->flags & P_BIGKEY) && 16457451Sbostic (h->prevpg != P_INVALID || skip > 1)) { 16550989Sbostic tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1); 16650989Sbostic a.size = tbl->ksize; 16750989Sbostic a.data = tbl->bytes; 16850989Sbostic b.size = bl->ksize; 16950989Sbostic b.data = bl->bytes; 17057451Sbostic nksize = t->bt_pfx(&a, &b); 17150989Sbostic n = NBINTERNAL(nksize); 17250989Sbostic if (n < nbytes) { 17350989Sbostic #ifdef STATISTICS 17450989Sbostic bt_pfxsaved += nbytes - n; 17550989Sbostic #endif 17650989Sbostic nbytes = n; 17750989Sbostic } else 17850989Sbostic nksize = 0; 17950989Sbostic } else 18050989Sbostic nksize = 0; 18150989Sbostic break; 18250989Sbostic case P_RINTERNAL: 18350989Sbostic case P_RLEAF: 18450989Sbostic nbytes = NRINTERNAL; 18550989Sbostic break; 18656741Sbostic default: 18756741Sbostic abort(); 18850989Sbostic } 18950989Sbostic 19050989Sbostic /* Split the parent page if necessary or shift the indices. */ 19157989Sbostic if (h->upper - h->lower < nbytes + sizeof(indx_t)) { 19257440Sbostic sp = h; 19350989Sbostic h = h->pgno == P_ROOT ? 19457451Sbostic bt_root(t, h, &l, &r, &skip, nbytes) : 19557451Sbostic bt_page(t, h, &l, &r, &skip, nbytes); 19650989Sbostic if (h == NULL) 19750989Sbostic goto err1; 198*60234Sbostic parentsplit = 1; 19946134Smao } else { 20050989Sbostic if (skip < (nxtindex = NEXTINDEX(h))) 20158017Sbostic memmove(h->linp + skip + 1, h->linp + skip, 20257989Sbostic (nxtindex - skip) * sizeof(indx_t)); 20357989Sbostic h->lower += sizeof(indx_t); 204*60234Sbostic parentsplit = 0; 20546134Smao } 20646134Smao 20750989Sbostic /* Insert the key into the parent page. */ 20850989Sbostic switch(rchild->flags & P_TYPE) { 20950989Sbostic case P_BINTERNAL: 21050989Sbostic h->linp[skip] = h->upper -= nbytes; 21150989Sbostic dest = (char *)h + h->linp[skip]; 21258017Sbostic memmove(dest, bi, nbytes); 21350989Sbostic ((BINTERNAL *)dest)->pgno = rchild->pgno; 21450989Sbostic break; 21550989Sbostic case P_BLEAF: 21650989Sbostic h->linp[skip] = h->upper -= nbytes; 21750989Sbostic dest = (char *)h + h->linp[skip]; 21850989Sbostic WR_BINTERNAL(dest, nksize ? nksize : bl->ksize, 21951106Sbostic rchild->pgno, bl->flags & P_BIGKEY); 22058017Sbostic memmove(dest, bl->bytes, nksize ? nksize : bl->ksize); 22150989Sbostic if (bl->flags & P_BIGKEY && 22250989Sbostic bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR) 22350989Sbostic goto err1; 22450989Sbostic break; 22550989Sbostic case P_RINTERNAL: 226*60234Sbostic /* 227*60234Sbostic * Update the left page count. If split 228*60234Sbostic * added at index 0, fix the correct page. 229*60234Sbostic */ 230*60234Sbostic if (skip > 0) 231*60234Sbostic dest = (char *)h + h->linp[skip - 1]; 232*60234Sbostic else 233*60234Sbostic dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; 234*60234Sbostic ((RINTERNAL *)dest)->nrecs = rec_total(lchild); 235*60234Sbostic ((RINTERNAL *)dest)->pgno = lchild->pgno; 236*60234Sbostic 237*60234Sbostic /* Update the right page count. */ 23850989Sbostic h->linp[skip] = h->upper -= nbytes; 23950989Sbostic dest = (char *)h + h->linp[skip]; 24050989Sbostic ((RINTERNAL *)dest)->nrecs = rec_total(rchild); 24150989Sbostic ((RINTERNAL *)dest)->pgno = rchild->pgno; 24250989Sbostic break; 24350989Sbostic case P_RLEAF: 244*60234Sbostic /* 245*60234Sbostic * Update the left page count. If split 246*60234Sbostic * added at index 0, fix the correct page. 247*60234Sbostic */ 248*60234Sbostic if (skip > 0) 249*60234Sbostic dest = (char *)h + h->linp[skip - 1]; 250*60234Sbostic else 251*60234Sbostic dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; 252*60234Sbostic ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild); 253*60234Sbostic ((RINTERNAL *)dest)->pgno = lchild->pgno; 254*60234Sbostic 255*60234Sbostic /* Update the right page count. */ 25650989Sbostic h->linp[skip] = h->upper -= nbytes; 25750989Sbostic dest = (char *)h + h->linp[skip]; 25850989Sbostic ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild); 25950989Sbostic ((RINTERNAL *)dest)->pgno = rchild->pgno; 26050989Sbostic break; 26156741Sbostic default: 26256741Sbostic abort(); 26350989Sbostic } 26446134Smao 26550989Sbostic /* Unpin the held pages. */ 266*60234Sbostic if (!parentsplit) { 26750989Sbostic mpool_put(t->bt_mp, h, MPOOL_DIRTY); 26850989Sbostic break; 26950989Sbostic } 27057440Sbostic 27157440Sbostic /* If the root page was split, make it look right. */ 27257440Sbostic if (sp->pgno == P_ROOT && 27360048Sbostic (ISSET(t, R_RECNO) ? 27457440Sbostic bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) 27557440Sbostic goto err1; 27657440Sbostic 27750989Sbostic mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 27850989Sbostic mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 27950989Sbostic } 28046134Smao 28150989Sbostic /* Unpin the held pages. */ 28250989Sbostic mpool_put(t->bt_mp, l, MPOOL_DIRTY); 28350989Sbostic mpool_put(t->bt_mp, r, MPOOL_DIRTY); 28450989Sbostic 28551106Sbostic /* Clear any pages left on the stack. */ 28650989Sbostic return (RET_SUCCESS); 28746134Smao 28846134Smao /* 28950989Sbostic * If something fails in the above loop we were already walking back 29050989Sbostic * up the tree and the tree is now inconsistent. Nothing much we can 29150989Sbostic * do about it but release any memory we're holding. 29246134Smao */ 29350989Sbostic err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); 29450989Sbostic mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); 29546134Smao 29650989Sbostic err2: mpool_put(t->bt_mp, l, 0); 29750989Sbostic mpool_put(t->bt_mp, r, 0); 29850989Sbostic __dbpanic(t->bt_dbp); 29950989Sbostic return (RET_ERROR); 30046134Smao } 30146134Smao 30246134Smao /* 30350989Sbostic * BT_PAGE -- Split a non-root page of a btree. 30446134Smao * 30550989Sbostic * Parameters: 30650989Sbostic * t: tree 30750989Sbostic * h: root page 30850989Sbostic * lp: pointer to left page pointer 30950989Sbostic * rp: pointer to right page pointer 31050989Sbostic * skip: pointer to index to leave open 31157451Sbostic * ilen: insert length 31246134Smao * 31350989Sbostic * Returns: 31450989Sbostic * Pointer to page in which to insert or NULL on error. 31546134Smao */ 31650989Sbostic static PAGE * 31757451Sbostic bt_page(t, h, lp, rp, skip, ilen) 31850989Sbostic BTREE *t; 31950989Sbostic PAGE *h, **lp, **rp; 32058078Sbostic u_int *skip; 32157451Sbostic size_t ilen; 32246134Smao { 32350989Sbostic PAGE *l, *r, *tp; 32450989Sbostic pgno_t npg; 32546134Smao 32650989Sbostic #ifdef STATISTICS 32750989Sbostic ++bt_split; 32850989Sbostic #endif 32950989Sbostic /* Put the new right page for the split into place. */ 33056492Sbostic if ((r = __bt_new(t, &npg)) == NULL) 33150989Sbostic return (NULL); 33250989Sbostic r->pgno = npg; 33350989Sbostic r->lower = BTDATAOFF; 33450989Sbostic r->upper = t->bt_psize; 33550989Sbostic r->nextpg = h->nextpg; 33650989Sbostic r->prevpg = h->pgno; 33750989Sbostic r->flags = h->flags & P_TYPE; 33846134Smao 33950989Sbostic /* 34050989Sbostic * If we're splitting the last page on a level because we're appending 34150989Sbostic * a key to it (skip is NEXTINDEX()), it's likely that the data is 34250989Sbostic * sorted. Adding an empty page on the side of the level is less work 34350989Sbostic * and can push the fill factor much higher than normal. If we're 34450989Sbostic * wrong it's no big deal, we'll just do the split the right way next 34550989Sbostic * time. It may look like it's equally easy to do a similar hack for 34650989Sbostic * reverse sorted data, that is, split the tree left, but it's not. 34750989Sbostic * Don't even try. 34850989Sbostic */ 34950989Sbostic if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) { 35050989Sbostic #ifdef STATISTICS 35150989Sbostic ++bt_sortsplit; 35250989Sbostic #endif 35350989Sbostic h->nextpg = r->pgno; 35457989Sbostic r->lower = BTDATAOFF + sizeof(indx_t); 35550989Sbostic *skip = 0; 35650989Sbostic *lp = h; 35750989Sbostic *rp = r; 35850989Sbostic return (r); 35950989Sbostic } 36046134Smao 36150989Sbostic /* Put the new left page for the split into place. */ 36250989Sbostic if ((l = malloc(t->bt_psize)) == NULL) { 36350989Sbostic mpool_put(t->bt_mp, r, 0); 36450989Sbostic return (NULL); 36550989Sbostic } 36650989Sbostic l->pgno = h->pgno; 36750989Sbostic l->nextpg = r->pgno; 36850989Sbostic l->prevpg = h->prevpg; 36950989Sbostic l->lower = BTDATAOFF; 37050989Sbostic l->upper = t->bt_psize; 37150989Sbostic l->flags = h->flags & P_TYPE; 37246134Smao 37350989Sbostic /* Fix up the previous pointer of the page after the split page. */ 37450989Sbostic if (h->nextpg != P_INVALID) { 37550989Sbostic if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) { 37650989Sbostic free(l); 37750989Sbostic /* XXX mpool_free(t->bt_mp, r->pgno); */ 37850989Sbostic return (NULL); 37946134Smao } 38050989Sbostic tp->prevpg = r->pgno; 38150989Sbostic mpool_put(t->bt_mp, tp, 0); 38250989Sbostic } 38346134Smao 38450989Sbostic /* 38550989Sbostic * Split right. The key/data pairs aren't sorted in the btree page so 38650989Sbostic * it's simpler to copy the data from the split page onto two new pages 38750989Sbostic * instead of copying half the data to the right page and compacting 38850989Sbostic * the left page in place. Since the left page can't change, we have 38950989Sbostic * to swap the original and the allocated left page after the split. 39050989Sbostic */ 39157451Sbostic tp = bt_psplit(t, h, l, r, skip, ilen); 39246134Smao 39350989Sbostic /* Move the new left page onto the old left page. */ 39458017Sbostic memmove(h, l, t->bt_psize); 39550989Sbostic if (tp == l) 39650989Sbostic tp = h; 39750989Sbostic free(l); 39850989Sbostic 39950989Sbostic *lp = h; 40050989Sbostic *rp = r; 40150989Sbostic return (tp); 40246134Smao } 40346134Smao 40446134Smao /* 40551106Sbostic * BT_ROOT -- Split the root page of a btree. 40646134Smao * 40750989Sbostic * Parameters: 40850989Sbostic * t: tree 40950989Sbostic * h: root page 41050989Sbostic * lp: pointer to left page pointer 41150989Sbostic * rp: pointer to right page pointer 41250989Sbostic * skip: pointer to index to leave open 41357451Sbostic * ilen: insert length 41446134Smao * 41550989Sbostic * Returns: 41650989Sbostic * Pointer to page in which to insert or NULL on error. 41746134Smao */ 41850989Sbostic static PAGE * 41957451Sbostic bt_root(t, h, lp, rp, skip, ilen) 42050989Sbostic BTREE *t; 42150989Sbostic PAGE *h, **lp, **rp; 42258078Sbostic u_int *skip; 42357451Sbostic size_t ilen; 42446134Smao { 42550989Sbostic PAGE *l, *r, *tp; 42650989Sbostic pgno_t lnpg, rnpg; 42746134Smao 42850989Sbostic #ifdef STATISTICS 42950989Sbostic ++bt_split; 43050989Sbostic ++bt_rootsplit; 43150989Sbostic #endif 43250989Sbostic /* Put the new left and right pages for the split into place. */ 43356492Sbostic if ((l = __bt_new(t, &lnpg)) == NULL || 43456492Sbostic (r = __bt_new(t, &rnpg)) == NULL) 43550989Sbostic return (NULL); 43650989Sbostic l->pgno = lnpg; 43750989Sbostic r->pgno = rnpg; 43850989Sbostic l->nextpg = r->pgno; 43950989Sbostic r->prevpg = l->pgno; 44050989Sbostic l->prevpg = r->nextpg = P_INVALID; 44150989Sbostic l->lower = r->lower = BTDATAOFF; 44250989Sbostic l->upper = r->upper = t->bt_psize; 44350989Sbostic l->flags = r->flags = h->flags & P_TYPE; 44450989Sbostic 44550989Sbostic /* Split the root page. */ 44657451Sbostic tp = bt_psplit(t, h, l, r, skip, ilen); 44750989Sbostic 44850989Sbostic *lp = l; 44950989Sbostic *rp = r; 45050989Sbostic return (tp); 45146134Smao } 45246134Smao 45346134Smao /* 45457440Sbostic * BT_RROOT -- Fix up the recno root page after it has been split. 45546134Smao * 45650989Sbostic * Parameters: 45750989Sbostic * t: tree 45850989Sbostic * h: root page 45957440Sbostic * l: left page 46057440Sbostic * r: right page 46146134Smao * 46250989Sbostic * Returns: 46350989Sbostic * RET_ERROR, RET_SUCCESS 46446134Smao */ 46550989Sbostic static int 46650989Sbostic bt_rroot(t, h, l, r) 46750989Sbostic BTREE *t; 46850989Sbostic PAGE *h, *l, *r; 46946134Smao { 47050989Sbostic char *dest; 47146134Smao 47250989Sbostic /* Insert the left and right keys, set the header information. */ 47350989Sbostic h->linp[0] = h->upper = t->bt_psize - NRINTERNAL; 47450989Sbostic dest = (char *)h + h->upper; 47550989Sbostic WR_RINTERNAL(dest, 47650989Sbostic l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno); 47746134Smao 47850989Sbostic h->linp[1] = h->upper -= NRINTERNAL; 47950989Sbostic dest = (char *)h + h->upper; 48050989Sbostic WR_RINTERNAL(dest, 48150989Sbostic r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno); 48246134Smao 48357989Sbostic h->lower = BTDATAOFF + 2 * sizeof(indx_t); 48446134Smao 48550989Sbostic /* Unpin the root page, set to recno internal page. */ 48650989Sbostic h->flags &= ~P_TYPE; 48750989Sbostic h->flags |= P_RINTERNAL; 48850989Sbostic mpool_put(t->bt_mp, h, MPOOL_DIRTY); 48946134Smao 49050989Sbostic return (RET_SUCCESS); 49150989Sbostic } 49246134Smao 49350989Sbostic /* 49457440Sbostic * BT_BROOT -- Fix up the btree root page after it has been split. 49550989Sbostic * 49650989Sbostic * Parameters: 49750989Sbostic * t: tree 49850989Sbostic * h: root page 49957440Sbostic * l: left page 50057440Sbostic * r: right page 50150989Sbostic * 50250989Sbostic * Returns: 50350989Sbostic * RET_ERROR, RET_SUCCESS 50450989Sbostic */ 50550989Sbostic static int 50650989Sbostic bt_broot(t, h, l, r) 50750989Sbostic BTREE *t; 50850989Sbostic PAGE *h, *l, *r; 50950989Sbostic { 51050989Sbostic BINTERNAL *bi; 51150989Sbostic BLEAF *bl; 51250989Sbostic size_t nbytes; 51350989Sbostic char *dest; 51446134Smao 51546134Smao /* 51650989Sbostic * If the root page was a leaf page, change it into an internal page. 51750989Sbostic * We copy the key we split on (but not the key's data, in the case of 51856401Sbostic * a leaf page) to the new root page. 51950989Sbostic * 52050989Sbostic * The btree comparison code guarantees that the left-most key on any 52157440Sbostic * level of the tree is never used, so it doesn't need to be filled in. 52246134Smao */ 52356401Sbostic nbytes = NBINTERNAL(0); 52450989Sbostic h->linp[0] = h->upper = t->bt_psize - nbytes; 52550989Sbostic dest = (char *)h + h->upper; 52650989Sbostic WR_BINTERNAL(dest, 0, l->pgno, 0); 52746134Smao 52850989Sbostic switch(h->flags & P_TYPE) { 52950989Sbostic case P_BLEAF: 53050989Sbostic bl = GETBLEAF(r, 0); 53150989Sbostic nbytes = NBINTERNAL(bl->ksize); 53250989Sbostic h->linp[1] = h->upper -= nbytes; 53350989Sbostic dest = (char *)h + h->upper; 53450989Sbostic WR_BINTERNAL(dest, bl->ksize, r->pgno, 0); 53558017Sbostic memmove(dest, bl->bytes, bl->ksize); 53650989Sbostic 53756401Sbostic /* 53856401Sbostic * If the key is on an overflow page, mark the overflow chain 53956401Sbostic * so it isn't deleted when the leaf copy of the key is deleted. 54056401Sbostic */ 54150989Sbostic if (bl->flags & P_BIGKEY && 54250989Sbostic bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR) 54350989Sbostic return (RET_ERROR); 54450989Sbostic break; 54550989Sbostic case P_BINTERNAL: 54650989Sbostic bi = GETBINTERNAL(r, 0); 54750989Sbostic nbytes = NBINTERNAL(bi->ksize); 54850989Sbostic h->linp[1] = h->upper -= nbytes; 54950989Sbostic dest = (char *)h + h->upper; 55058017Sbostic memmove(dest, bi, nbytes); 55150989Sbostic ((BINTERNAL *)dest)->pgno = r->pgno; 55250989Sbostic break; 55356741Sbostic default: 55456741Sbostic abort(); 55546134Smao } 55657440Sbostic 55757440Sbostic /* There are two keys on the page. */ 55857989Sbostic h->lower = BTDATAOFF + 2 * sizeof(indx_t); 55946134Smao 56050989Sbostic /* Unpin the root page, set to btree internal page. */ 56150989Sbostic h->flags &= ~P_TYPE; 56250989Sbostic h->flags |= P_BINTERNAL; 56350989Sbostic mpool_put(t->bt_mp, h, MPOOL_DIRTY); 56446134Smao 56546134Smao return (RET_SUCCESS); 56646134Smao } 56746134Smao 56846134Smao /* 56950989Sbostic * BT_PSPLIT -- Do the real work of splitting the page. 57046134Smao * 57150989Sbostic * Parameters: 57250989Sbostic * t: tree 57350989Sbostic * h: page to be split 57450989Sbostic * l: page to put lower half of data 57550989Sbostic * r: page to put upper half of data 57657440Sbostic * pskip: pointer to index to leave open 57757451Sbostic * ilen: insert length 57846134Smao * 57950989Sbostic * Returns: 58050989Sbostic * Pointer to page in which to insert. 58146134Smao */ 58250989Sbostic static PAGE * 58357451Sbostic bt_psplit(t, h, l, r, pskip, ilen) 58450989Sbostic BTREE *t; 58550989Sbostic PAGE *h, *l, *r; 58658078Sbostic u_int *pskip; 58757451Sbostic size_t ilen; 58846134Smao { 58950989Sbostic BINTERNAL *bi; 59050989Sbostic BLEAF *bl; 59150989Sbostic RLEAF *rl; 59250989Sbostic EPGNO *c; 59350989Sbostic PAGE *rval; 59458067Sbostic void *src; 59558067Sbostic indx_t full, half, nxt, off, skip, top, used; 59650989Sbostic size_t nbytes; 59758067Sbostic int bigkeycnt, isbigkey; 59846134Smao 59946134Smao /* 60057451Sbostic * Split the data to the left and right pages. Leave the skip index 60157451Sbostic * open. Additionally, make some effort not to split on an overflow 60257451Sbostic * key. This makes internal page processing faster and can save 60357451Sbostic * space as overflow keys used by internal pages are never deleted. 60446134Smao */ 60550989Sbostic bigkeycnt = 0; 60651106Sbostic skip = *pskip; 60757451Sbostic full = t->bt_psize - BTDATAOFF; 60857451Sbostic half = full / 2; 60957451Sbostic used = 0; 61050989Sbostic for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) { 61157451Sbostic if (skip == off) { 61257451Sbostic nbytes = ilen; 61357451Sbostic isbigkey = 0; /* XXX: not really known. */ 61457451Sbostic } else 61557451Sbostic switch (h->flags & P_TYPE) { 61657451Sbostic case P_BINTERNAL: 61757451Sbostic src = bi = GETBINTERNAL(h, nxt); 61857451Sbostic nbytes = NBINTERNAL(bi->ksize); 61957451Sbostic isbigkey = bi->flags & P_BIGKEY; 62057451Sbostic break; 62157451Sbostic case P_BLEAF: 62257451Sbostic src = bl = GETBLEAF(h, nxt); 62357451Sbostic nbytes = NBLEAF(bl); 62457451Sbostic isbigkey = bl->flags & P_BIGKEY; 62557451Sbostic break; 62657451Sbostic case P_RINTERNAL: 62757451Sbostic src = GETRINTERNAL(h, nxt); 62857451Sbostic nbytes = NRINTERNAL; 62957451Sbostic isbigkey = 0; 63057451Sbostic break; 63157451Sbostic case P_RLEAF: 63257451Sbostic src = rl = GETRLEAF(h, nxt); 63357451Sbostic nbytes = NRLEAF(rl); 63457451Sbostic isbigkey = 0; 63557451Sbostic break; 63657451Sbostic default: 63757451Sbostic abort(); 63857451Sbostic } 63957451Sbostic 64057451Sbostic /* 64157451Sbostic * If the key/data pairs are substantial fractions of the max 64257451Sbostic * possible size for the page, it's possible to get situations 64357451Sbostic * where we decide to try and copy too much onto the left page. 64457451Sbostic * Make sure that doesn't happen. 64557451Sbostic */ 64657451Sbostic if (skip <= off && used + nbytes >= full) { 64757451Sbostic --off; 64850989Sbostic break; 64950989Sbostic } 65046134Smao 65157451Sbostic /* Copy the key/data pair, if not the skipped index. */ 65257451Sbostic if (skip != off) { 65357451Sbostic ++nxt; 65457451Sbostic 65557451Sbostic l->linp[off] = l->upper -= nbytes; 65658017Sbostic memmove((char *)l + l->upper, src, nbytes); 65757451Sbostic } 65857451Sbostic 65957451Sbostic used += nbytes; 66057451Sbostic if (used >= half) { 66157440Sbostic if (!isbigkey || bigkeycnt == 3) 66257440Sbostic break; 66357440Sbostic else 66457440Sbostic ++bigkeycnt; 66557451Sbostic } 66650989Sbostic } 66757451Sbostic 66857451Sbostic /* 66957451Sbostic * Off is the last offset that's valid for the left page. 67057451Sbostic * Nxt is the first offset to be placed on the right page. 67157451Sbostic */ 67257989Sbostic l->lower += (off + 1) * sizeof(indx_t); 67346134Smao 67450989Sbostic /* 67551106Sbostic * If splitting the page that the cursor was on, the cursor has to be 67651106Sbostic * adjusted to point to the same record as before the split. If the 67757451Sbostic * cursor is at or past the skipped slot, the cursor is incremented by 67857451Sbostic * one. If the cursor is on the right page, it is decremented by the 67957451Sbostic * number of records split to the left page. 68051106Sbostic * 68160048Sbostic * Don't bother checking for the B_SEQINIT flag, the page number will 68251106Sbostic * be P_INVALID. 68350989Sbostic */ 68450989Sbostic c = &t->bt_bcursor; 68557451Sbostic if (c->pgno == h->pgno) { 68657451Sbostic if (c->index >= skip) 68757451Sbostic ++c->index; 68857451Sbostic if (c->index < nxt) /* Left page. */ 68950989Sbostic c->pgno = l->pgno; 69057451Sbostic else { /* Right page. */ 69150989Sbostic c->pgno = r->pgno; 69257451Sbostic c->index -= nxt; 69346134Smao } 69457451Sbostic } 69546134Smao 69650989Sbostic /* 69757451Sbostic * If the skipped index was on the left page, just return that page. 69857451Sbostic * Otherwise, adjust the skip index to reflect the new position on 69957451Sbostic * the right page. 70050989Sbostic */ 70157451Sbostic if (skip <= off) { 70257451Sbostic skip = 0; 70357451Sbostic rval = l; 70457451Sbostic } else { 70550989Sbostic rval = r; 70657451Sbostic *pskip -= nxt; 70757451Sbostic } 70846134Smao 70950989Sbostic for (off = 0; nxt < top; ++off) { 71051106Sbostic if (skip == nxt) { 71157451Sbostic ++off; 71251106Sbostic skip = 0; 71346134Smao } 71450989Sbostic switch (h->flags & P_TYPE) { 71550989Sbostic case P_BINTERNAL: 71650989Sbostic src = bi = GETBINTERNAL(h, nxt); 71750989Sbostic nbytes = NBINTERNAL(bi->ksize); 71850989Sbostic break; 71950989Sbostic case P_BLEAF: 72050989Sbostic src = bl = GETBLEAF(h, nxt); 72150989Sbostic nbytes = NBLEAF(bl); 72250989Sbostic break; 72350989Sbostic case P_RINTERNAL: 72450989Sbostic src = GETRINTERNAL(h, nxt); 72550989Sbostic nbytes = NRINTERNAL; 72650989Sbostic break; 72750989Sbostic case P_RLEAF: 72850989Sbostic src = rl = GETRLEAF(h, nxt); 72950989Sbostic nbytes = NRLEAF(rl); 73050989Sbostic break; 73156741Sbostic default: 73256741Sbostic abort(); 73350989Sbostic } 73450989Sbostic ++nxt; 73550989Sbostic r->linp[off] = r->upper -= nbytes; 73658017Sbostic memmove((char *)r + r->upper, src, nbytes); 73750989Sbostic } 73857989Sbostic r->lower += off * sizeof(indx_t); 73946134Smao 74050989Sbostic /* If the key is being appended to the page, adjust the index. */ 74151106Sbostic if (skip == top) 74257989Sbostic r->lower += sizeof(indx_t); 74346134Smao 74450989Sbostic return (rval); 74550989Sbostic } 74646134Smao 74750989Sbostic /* 74850989Sbostic * BT_PRESERVE -- Mark a chain of pages as used by an internal node. 74950989Sbostic * 75050989Sbostic * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the 75150989Sbostic * record that references them gets deleted. Chains pointed to by internal 75250989Sbostic * pages never get deleted. This routine marks a chain as pointed to by an 75350989Sbostic * internal page. 75450989Sbostic * 75550989Sbostic * Parameters: 75650989Sbostic * t: tree 75750989Sbostic * pg: page number of first page in the chain. 75850989Sbostic * 75950989Sbostic * Returns: 76050989Sbostic * RET_SUCCESS, RET_ERROR. 76150989Sbostic */ 76250989Sbostic static int 76350989Sbostic bt_preserve(t, pg) 76450989Sbostic BTREE *t; 76550989Sbostic pgno_t pg; 76650989Sbostic { 76750989Sbostic PAGE *h; 76846134Smao 76950989Sbostic if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL) 77050989Sbostic return (RET_ERROR); 77150989Sbostic h->flags |= P_PRESERVE; 77250989Sbostic mpool_put(t->bt_mp, h, MPOOL_DIRTY); 77346134Smao return (RET_SUCCESS); 77446134Smao } 77550989Sbostic 77650989Sbostic /* 77750989Sbostic * REC_TOTAL -- Return the number of recno entries below a page. 77850989Sbostic * 77950989Sbostic * Parameters: 78050989Sbostic * h: page 78150989Sbostic * 78250989Sbostic * Returns: 78350989Sbostic * The number of recno entries below a page. 78450989Sbostic * 78550989Sbostic * XXX 78650989Sbostic * These values could be set by the bt_psplit routine. The problem is that the 78750989Sbostic * entry has to be popped off of the stack etc. or the values have to be passed 78850989Sbostic * all the way back to bt_split/bt_rroot and it's not very clean. 78950989Sbostic */ 79050989Sbostic static recno_t 79150989Sbostic rec_total(h) 79250989Sbostic PAGE *h; 79350989Sbostic { 79450989Sbostic recno_t recs; 79557989Sbostic indx_t nxt, top; 79650989Sbostic 79750989Sbostic for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt) 79850989Sbostic recs += GETRINTERNAL(h, nxt)->nrecs; 79950989Sbostic return (recs); 80050989Sbostic } 801