bt_split.c revision 1.15 1 1.15 joerg /* $NetBSD: bt_split.c,v 1.15 2008/08/26 21:18:38 joerg Exp $ */
2 1.5 cgd
3 1.1 cgd /*-
4 1.4 cgd * Copyright (c) 1990, 1993, 1994
5 1.1 cgd * The Regents of the University of California. All rights reserved.
6 1.1 cgd *
7 1.1 cgd * This code is derived from software contributed to Berkeley by
8 1.1 cgd * Mike Olson.
9 1.1 cgd *
10 1.1 cgd * Redistribution and use in source and binary forms, with or without
11 1.1 cgd * modification, are permitted provided that the following conditions
12 1.1 cgd * are met:
13 1.1 cgd * 1. Redistributions of source code must retain the above copyright
14 1.1 cgd * notice, this list of conditions and the following disclaimer.
15 1.1 cgd * 2. Redistributions in binary form must reproduce the above copyright
16 1.1 cgd * notice, this list of conditions and the following disclaimer in the
17 1.1 cgd * documentation and/or other materials provided with the distribution.
18 1.13 agc * 3. Neither the name of the University nor the names of its contributors
19 1.1 cgd * may be used to endorse or promote products derived from this software
20 1.1 cgd * without specific prior written permission.
21 1.1 cgd *
22 1.1 cgd * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 1.1 cgd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 1.1 cgd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 1.1 cgd * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 1.1 cgd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 1.1 cgd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 1.1 cgd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 1.1 cgd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 1.1 cgd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 1.1 cgd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 1.1 cgd * SUCH DAMAGE.
33 1.1 cgd */
34 1.1 cgd
35 1.7 christos #include <sys/cdefs.h>
36 1.1 cgd #if defined(LIBC_SCCS) && !defined(lint)
37 1.5 cgd #if 0
38 1.6 cgd static char sccsid[] = "@(#)bt_split.c 8.9 (Berkeley) 7/26/94";
39 1.5 cgd #else
40 1.15 joerg __RCSID("$NetBSD: bt_split.c,v 1.15 2008/08/26 21:18:38 joerg Exp $");
41 1.5 cgd #endif
42 1.1 cgd #endif /* LIBC_SCCS and not lint */
43 1.1 cgd
44 1.8 jtc #include "namespace.h"
45 1.1 cgd #include <sys/types.h>
46 1.1 cgd
47 1.14 christos #include <assert.h>
48 1.1 cgd #include <limits.h>
49 1.1 cgd #include <stdio.h>
50 1.1 cgd #include <stdlib.h>
51 1.1 cgd #include <string.h>
52 1.1 cgd
53 1.1 cgd #include <db.h>
54 1.1 cgd #include "btree.h"
55 1.1 cgd
56 1.14 christos static int bt_broot(BTREE *, PAGE *, PAGE *, PAGE *);
57 1.14 christos static PAGE *bt_page(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
58 1.14 christos static int bt_preserve(BTREE *, pgno_t);
59 1.14 christos static PAGE *bt_psplit(BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t);
60 1.14 christos static PAGE *bt_root(BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t);
61 1.14 christos static int bt_rroot(BTREE *, PAGE *, PAGE *, PAGE *);
62 1.14 christos static recno_t rec_total(PAGE *);
63 1.1 cgd
64 1.1 cgd #ifdef STATISTICS
65 1.15 joerg unsigned long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved;
66 1.1 cgd #endif
67 1.1 cgd
68 1.1 cgd /*
69 1.1 cgd * __BT_SPLIT -- Split the tree.
70 1.1 cgd *
71 1.1 cgd * Parameters:
72 1.1 cgd * t: tree
73 1.1 cgd * sp: page to split
74 1.1 cgd * key: key to insert
75 1.1 cgd * data: data to insert
76 1.1 cgd * flags: BIGKEY/BIGDATA flags
77 1.1 cgd * ilen: insert length
78 1.1 cgd * skip: index to leave open
79 1.1 cgd *
80 1.1 cgd * Returns:
81 1.1 cgd * RET_ERROR, RET_SUCCESS
82 1.1 cgd */
83 1.1 cgd int
84 1.14 christos __bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
85 1.15 joerg size_t ilen, uint32_t argskip)
86 1.1 cgd {
87 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
88 1.7 christos BLEAF *bl = NULL, *tbl; /* pacify gcc */
89 1.1 cgd DBT a, b;
90 1.1 cgd EPGNO *parent;
91 1.1 cgd PAGE *h, *l, *r, *lchild, *rchild;
92 1.1 cgd indx_t nxtindex;
93 1.15 joerg uint16_t skip;
94 1.15 joerg uint32_t n, nbytes, nksize = 0; /* pacify gcc */
95 1.1 cgd int parentsplit;
96 1.1 cgd char *dest;
97 1.1 cgd
98 1.1 cgd /*
99 1.1 cgd * Split the page into two pages, l and r. The split routines return
100 1.1 cgd * a pointer to the page into which the key should be inserted and with
101 1.1 cgd * skip set to the offset which should be used. Additionally, l and r
102 1.1 cgd * are pinned.
103 1.1 cgd */
104 1.4 cgd skip = argskip;
105 1.1 cgd h = sp->pgno == P_ROOT ?
106 1.1 cgd bt_root(t, sp, &l, &r, &skip, ilen) :
107 1.1 cgd bt_page(t, sp, &l, &r, &skip, ilen);
108 1.1 cgd if (h == NULL)
109 1.1 cgd return (RET_ERROR);
110 1.1 cgd
111 1.1 cgd /*
112 1.1 cgd * Insert the new key/data pair into the leaf page. (Key inserts
113 1.1 cgd * always cause a leaf page to split first.)
114 1.1 cgd */
115 1.14 christos _DBFIT(ilen, indx_t);
116 1.14 christos h->upper -= (indx_t)ilen;
117 1.14 christos h->linp[skip] = h->upper;
118 1.10 christos dest = (char *)(void *)h + h->upper;
119 1.6 cgd if (F_ISSET(t, R_RECNO))
120 1.14 christos WR_RLEAF(dest, data, flags);
121 1.1 cgd else
122 1.14 christos WR_BLEAF(dest, key, data, flags);
123 1.1 cgd
124 1.1 cgd /* If the root page was split, make it look right. */
125 1.1 cgd if (sp->pgno == P_ROOT &&
126 1.6 cgd (F_ISSET(t, R_RECNO) ?
127 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
128 1.1 cgd goto err2;
129 1.1 cgd
130 1.1 cgd /*
131 1.1 cgd * Now we walk the parent page stack -- a LIFO stack of the pages that
132 1.1 cgd * were traversed when we searched for the page that split. Each stack
133 1.1 cgd * entry is a page number and a page index offset. The offset is for
134 1.1 cgd * the page traversed on the search. We've just split a page, so we
135 1.1 cgd * have to insert a new key into the parent page.
136 1.1 cgd *
137 1.1 cgd * If the insert into the parent page causes it to split, may have to
138 1.1 cgd * continue splitting all the way up the tree. We stop if the root
139 1.1 cgd * splits or the page inserted into didn't have to split to hold the
140 1.1 cgd * new key. Some algorithms replace the key for the old page as well
141 1.1 cgd * as the new page. We don't, as there's no reason to believe that the
142 1.1 cgd * first key on the old page is any better than the key we have, and,
143 1.1 cgd * in the case of a key being placed at index 0 causing the split, the
144 1.1 cgd * key is unavailable.
145 1.1 cgd *
146 1.1 cgd * There are a maximum of 5 pages pinned at any time. We keep the left
147 1.1 cgd * and right pages pinned while working on the parent. The 5 are the
148 1.1 cgd * two children, left parent and right parent (when the parent splits)
149 1.1 cgd * and the root page or the overflow key page when calling bt_preserve.
150 1.1 cgd * This code must make sure that all pins are released other than the
151 1.1 cgd * root page or overflow page which is unlocked elsewhere.
152 1.1 cgd */
153 1.1 cgd while ((parent = BT_POP(t)) != NULL) {
154 1.1 cgd lchild = l;
155 1.1 cgd rchild = r;
156 1.1 cgd
157 1.1 cgd /* Get the parent page. */
158 1.1 cgd if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
159 1.1 cgd goto err2;
160 1.1 cgd
161 1.1 cgd /*
162 1.1 cgd * The new key goes ONE AFTER the index, because the split
163 1.1 cgd * was to the right.
164 1.1 cgd */
165 1.1 cgd skip = parent->index + 1;
166 1.1 cgd
167 1.1 cgd /*
168 1.1 cgd * Calculate the space needed on the parent page.
169 1.1 cgd *
170 1.1 cgd * Prefix trees: space hack when inserting into BINTERNAL
171 1.1 cgd * pages. Retain only what's needed to distinguish between
172 1.1 cgd * the new entry and the LAST entry on the page to its left.
173 1.1 cgd * If the keys compare equal, retain the entire key. Note,
174 1.1 cgd * we don't touch overflow keys, and the entire key must be
175 1.1 cgd * retained for the next-to-left most key on the leftmost
176 1.1 cgd * page of each level, or the search will fail. Applicable
177 1.1 cgd * ONLY to internal pages that have leaf pages as children.
178 1.1 cgd * Further reduction of the key between pairs of internal
179 1.1 cgd * pages loses too much information.
180 1.1 cgd */
181 1.1 cgd switch (rchild->flags & P_TYPE) {
182 1.1 cgd case P_BINTERNAL:
183 1.1 cgd bi = GETBINTERNAL(rchild, 0);
184 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
185 1.1 cgd break;
186 1.1 cgd case P_BLEAF:
187 1.1 cgd bl = GETBLEAF(rchild, 0);
188 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
189 1.1 cgd if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
190 1.1 cgd (h->prevpg != P_INVALID || skip > 1)) {
191 1.14 christos size_t temp;
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 1.14 christos temp = t->bt_pfx(&a, &b);
198 1.15 joerg _DBFIT(temp, uint32_t);
199 1.15 joerg nksize = (uint32_t)temp;
200 1.1 cgd n = NBINTERNAL(nksize);
201 1.1 cgd if (n < nbytes) {
202 1.1 cgd #ifdef STATISTICS
203 1.1 cgd bt_pfxsaved += nbytes - n;
204 1.1 cgd #endif
205 1.1 cgd nbytes = n;
206 1.1 cgd } else
207 1.1 cgd nksize = 0;
208 1.1 cgd } else
209 1.1 cgd nksize = 0;
210 1.1 cgd break;
211 1.1 cgd case P_RINTERNAL:
212 1.1 cgd case P_RLEAF:
213 1.1 cgd nbytes = NRINTERNAL;
214 1.1 cgd break;
215 1.1 cgd default:
216 1.1 cgd abort();
217 1.1 cgd }
218 1.1 cgd
219 1.1 cgd /* Split the parent page if necessary or shift the indices. */
220 1.1 cgd if (h->upper - h->lower < nbytes + sizeof(indx_t)) {
221 1.1 cgd sp = h;
222 1.1 cgd h = h->pgno == P_ROOT ?
223 1.1 cgd bt_root(t, h, &l, &r, &skip, nbytes) :
224 1.1 cgd bt_page(t, h, &l, &r, &skip, nbytes);
225 1.1 cgd if (h == NULL)
226 1.1 cgd goto err1;
227 1.1 cgd parentsplit = 1;
228 1.1 cgd } else {
229 1.1 cgd if (skip < (nxtindex = NEXTINDEX(h)))
230 1.1 cgd memmove(h->linp + skip + 1, h->linp + skip,
231 1.1 cgd (nxtindex - skip) * sizeof(indx_t));
232 1.1 cgd h->lower += sizeof(indx_t);
233 1.1 cgd parentsplit = 0;
234 1.1 cgd }
235 1.1 cgd
236 1.1 cgd /* Insert the key into the parent page. */
237 1.6 cgd switch (rchild->flags & P_TYPE) {
238 1.1 cgd case P_BINTERNAL:
239 1.1 cgd h->linp[skip] = h->upper -= nbytes;
240 1.10 christos dest = (char *)(void *)h + h->linp[skip];
241 1.1 cgd memmove(dest, bi, nbytes);
242 1.10 christos ((BINTERNAL *)(void *)dest)->pgno = rchild->pgno;
243 1.1 cgd break;
244 1.1 cgd case P_BLEAF:
245 1.1 cgd h->linp[skip] = h->upper -= nbytes;
246 1.10 christos dest = (char *)(void *)h + h->linp[skip];
247 1.1 cgd WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
248 1.1 cgd rchild->pgno, bl->flags & P_BIGKEY);
249 1.1 cgd memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
250 1.1 cgd if (bl->flags & P_BIGKEY &&
251 1.10 christos bt_preserve(t, *(pgno_t *)(void *)bl->bytes) ==
252 1.10 christos RET_ERROR)
253 1.1 cgd goto err1;
254 1.1 cgd break;
255 1.1 cgd case P_RINTERNAL:
256 1.1 cgd /*
257 1.1 cgd * Update the left page count. If split
258 1.1 cgd * added at index 0, fix the correct page.
259 1.1 cgd */
260 1.1 cgd if (skip > 0)
261 1.10 christos dest = (char *)(void *)h + h->linp[skip - 1];
262 1.1 cgd else
263 1.10 christos dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
264 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = rec_total(lchild);
265 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
266 1.1 cgd
267 1.1 cgd /* Update the right page count. */
268 1.1 cgd h->linp[skip] = h->upper -= nbytes;
269 1.10 christos dest = (char *)(void *)h + h->linp[skip];
270 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = rec_total(rchild);
271 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
272 1.1 cgd break;
273 1.1 cgd case P_RLEAF:
274 1.1 cgd /*
275 1.1 cgd * Update the left page count. If split
276 1.1 cgd * added at index 0, fix the correct page.
277 1.1 cgd */
278 1.1 cgd if (skip > 0)
279 1.10 christos dest = (char *)(void *)h + h->linp[skip - 1];
280 1.1 cgd else
281 1.10 christos dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
282 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(lchild);
283 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
284 1.1 cgd
285 1.1 cgd /* Update the right page count. */
286 1.1 cgd h->linp[skip] = h->upper -= nbytes;
287 1.10 christos dest = (char *)(void *)h + h->linp[skip];
288 1.10 christos ((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(rchild);
289 1.10 christos ((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
290 1.1 cgd break;
291 1.1 cgd default:
292 1.1 cgd abort();
293 1.1 cgd }
294 1.1 cgd
295 1.1 cgd /* Unpin the held pages. */
296 1.1 cgd if (!parentsplit) {
297 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
298 1.1 cgd break;
299 1.1 cgd }
300 1.1 cgd
301 1.1 cgd /* If the root page was split, make it look right. */
302 1.1 cgd if (sp->pgno == P_ROOT &&
303 1.6 cgd (F_ISSET(t, R_RECNO) ?
304 1.1 cgd bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
305 1.1 cgd goto err1;
306 1.1 cgd
307 1.1 cgd mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
308 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
309 1.1 cgd }
310 1.1 cgd
311 1.1 cgd /* Unpin the held pages. */
312 1.1 cgd mpool_put(t->bt_mp, l, MPOOL_DIRTY);
313 1.1 cgd mpool_put(t->bt_mp, r, MPOOL_DIRTY);
314 1.1 cgd
315 1.1 cgd /* Clear any pages left on the stack. */
316 1.1 cgd return (RET_SUCCESS);
317 1.1 cgd
318 1.1 cgd /*
319 1.1 cgd * If something fails in the above loop we were already walking back
320 1.1 cgd * up the tree and the tree is now inconsistent. Nothing much we can
321 1.1 cgd * do about it but release any memory we're holding.
322 1.1 cgd */
323 1.1 cgd err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
324 1.1 cgd mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
325 1.1 cgd
326 1.1 cgd err2: mpool_put(t->bt_mp, l, 0);
327 1.1 cgd mpool_put(t->bt_mp, r, 0);
328 1.1 cgd __dbpanic(t->bt_dbp);
329 1.1 cgd return (RET_ERROR);
330 1.1 cgd }
331 1.1 cgd
332 1.1 cgd /*
333 1.1 cgd * BT_PAGE -- Split a non-root page of a btree.
334 1.1 cgd *
335 1.1 cgd * Parameters:
336 1.1 cgd * t: tree
337 1.1 cgd * h: root page
338 1.1 cgd * lp: pointer to left page pointer
339 1.1 cgd * rp: pointer to right page pointer
340 1.1 cgd * skip: pointer to index to leave open
341 1.1 cgd * ilen: insert length
342 1.1 cgd *
343 1.1 cgd * Returns:
344 1.1 cgd * Pointer to page in which to insert or NULL on error.
345 1.1 cgd */
346 1.1 cgd static PAGE *
347 1.14 christos bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
348 1.1 cgd {
349 1.1 cgd PAGE *l, *r, *tp;
350 1.1 cgd pgno_t npg;
351 1.1 cgd
352 1.1 cgd #ifdef STATISTICS
353 1.1 cgd ++bt_split;
354 1.1 cgd #endif
355 1.1 cgd /* Put the new right page for the split into place. */
356 1.1 cgd if ((r = __bt_new(t, &npg)) == NULL)
357 1.1 cgd return (NULL);
358 1.1 cgd r->pgno = npg;
359 1.1 cgd r->lower = BTDATAOFF;
360 1.1 cgd r->upper = t->bt_psize;
361 1.1 cgd r->nextpg = h->nextpg;
362 1.1 cgd r->prevpg = h->pgno;
363 1.1 cgd r->flags = h->flags & P_TYPE;
364 1.1 cgd
365 1.1 cgd /*
366 1.1 cgd * If we're splitting the last page on a level because we're appending
367 1.1 cgd * a key to it (skip is NEXTINDEX()), it's likely that the data is
368 1.1 cgd * sorted. Adding an empty page on the side of the level is less work
369 1.1 cgd * and can push the fill factor much higher than normal. If we're
370 1.1 cgd * wrong it's no big deal, we'll just do the split the right way next
371 1.1 cgd * time. It may look like it's equally easy to do a similar hack for
372 1.1 cgd * reverse sorted data, that is, split the tree left, but it's not.
373 1.1 cgd * Don't even try.
374 1.1 cgd */
375 1.1 cgd if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
376 1.1 cgd #ifdef STATISTICS
377 1.1 cgd ++bt_sortsplit;
378 1.1 cgd #endif
379 1.1 cgd h->nextpg = r->pgno;
380 1.1 cgd r->lower = BTDATAOFF + sizeof(indx_t);
381 1.1 cgd *skip = 0;
382 1.1 cgd *lp = h;
383 1.1 cgd *rp = r;
384 1.1 cgd return (r);
385 1.1 cgd }
386 1.1 cgd
387 1.1 cgd /* Put the new left page for the split into place. */
388 1.4 cgd if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) {
389 1.1 cgd mpool_put(t->bt_mp, r, 0);
390 1.1 cgd return (NULL);
391 1.1 cgd }
392 1.6 cgd #ifdef PURIFY
393 1.6 cgd memset(l, 0xff, t->bt_psize);
394 1.6 cgd #endif
395 1.1 cgd l->pgno = h->pgno;
396 1.1 cgd l->nextpg = r->pgno;
397 1.1 cgd l->prevpg = h->prevpg;
398 1.1 cgd l->lower = BTDATAOFF;
399 1.1 cgd l->upper = t->bt_psize;
400 1.1 cgd l->flags = h->flags & P_TYPE;
401 1.1 cgd
402 1.1 cgd /* Fix up the previous pointer of the page after the split page. */
403 1.1 cgd if (h->nextpg != P_INVALID) {
404 1.1 cgd if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
405 1.1 cgd free(l);
406 1.1 cgd /* XXX mpool_free(t->bt_mp, r->pgno); */
407 1.1 cgd return (NULL);
408 1.1 cgd }
409 1.1 cgd tp->prevpg = r->pgno;
410 1.6 cgd mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
411 1.1 cgd }
412 1.1 cgd
413 1.1 cgd /*
414 1.1 cgd * Split right. The key/data pairs aren't sorted in the btree page so
415 1.1 cgd * it's simpler to copy the data from the split page onto two new pages
416 1.1 cgd * instead of copying half the data to the right page and compacting
417 1.1 cgd * the left page in place. Since the left page can't change, we have
418 1.1 cgd * to swap the original and the allocated left page after the split.
419 1.1 cgd */
420 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
421 1.1 cgd
422 1.1 cgd /* Move the new left page onto the old left page. */
423 1.1 cgd memmove(h, l, t->bt_psize);
424 1.1 cgd if (tp == l)
425 1.1 cgd tp = h;
426 1.1 cgd free(l);
427 1.1 cgd
428 1.1 cgd *lp = h;
429 1.1 cgd *rp = r;
430 1.1 cgd return (tp);
431 1.1 cgd }
432 1.1 cgd
433 1.1 cgd /*
434 1.1 cgd * BT_ROOT -- Split the root page of a btree.
435 1.1 cgd *
436 1.1 cgd * Parameters:
437 1.1 cgd * t: tree
438 1.1 cgd * h: root page
439 1.1 cgd * lp: pointer to left page pointer
440 1.1 cgd * rp: pointer to right page pointer
441 1.1 cgd * skip: pointer to index to leave open
442 1.1 cgd * ilen: insert length
443 1.1 cgd *
444 1.1 cgd * Returns:
445 1.1 cgd * Pointer to page in which to insert or NULL on error.
446 1.1 cgd */
447 1.1 cgd static PAGE *
448 1.14 christos bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
449 1.1 cgd {
450 1.1 cgd PAGE *l, *r, *tp;
451 1.1 cgd pgno_t lnpg, rnpg;
452 1.1 cgd
453 1.1 cgd #ifdef STATISTICS
454 1.1 cgd ++bt_split;
455 1.1 cgd ++bt_rootsplit;
456 1.1 cgd #endif
457 1.1 cgd /* Put the new left and right pages for the split into place. */
458 1.1 cgd if ((l = __bt_new(t, &lnpg)) == NULL ||
459 1.1 cgd (r = __bt_new(t, &rnpg)) == NULL)
460 1.1 cgd return (NULL);
461 1.1 cgd l->pgno = lnpg;
462 1.1 cgd r->pgno = rnpg;
463 1.1 cgd l->nextpg = r->pgno;
464 1.1 cgd r->prevpg = l->pgno;
465 1.1 cgd l->prevpg = r->nextpg = P_INVALID;
466 1.1 cgd l->lower = r->lower = BTDATAOFF;
467 1.1 cgd l->upper = r->upper = t->bt_psize;
468 1.1 cgd l->flags = r->flags = h->flags & P_TYPE;
469 1.1 cgd
470 1.1 cgd /* Split the root page. */
471 1.1 cgd tp = bt_psplit(t, h, l, r, skip, ilen);
472 1.1 cgd
473 1.1 cgd *lp = l;
474 1.1 cgd *rp = r;
475 1.1 cgd return (tp);
476 1.1 cgd }
477 1.1 cgd
478 1.1 cgd /*
479 1.1 cgd * BT_RROOT -- Fix up the recno root page after it has been split.
480 1.1 cgd *
481 1.1 cgd * Parameters:
482 1.1 cgd * t: tree
483 1.1 cgd * h: root page
484 1.1 cgd * l: left page
485 1.1 cgd * r: right page
486 1.1 cgd *
487 1.1 cgd * Returns:
488 1.1 cgd * RET_ERROR, RET_SUCCESS
489 1.1 cgd */
490 1.1 cgd static int
491 1.14 christos bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
492 1.1 cgd {
493 1.1 cgd char *dest;
494 1.15 joerg uint32_t sz;
495 1.14 christos size_t temp;
496 1.14 christos
497 1.14 christos temp = t->bt_psize - NRINTERNAL;
498 1.15 joerg _DBFIT(temp, uint32_t);
499 1.15 joerg sz = (uint32_t)temp;
500 1.1 cgd
501 1.1 cgd /* Insert the left and right keys, set the header information. */
502 1.14 christos _DBFIT(sz, indx_t);
503 1.14 christos h->linp[0] = h->upper = (indx_t)sz;
504 1.10 christos dest = (char *)(void *)h + h->upper;
505 1.1 cgd WR_RINTERNAL(dest,
506 1.1 cgd l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
507 1.1 cgd
508 1.1 cgd h->linp[1] = h->upper -= NRINTERNAL;
509 1.10 christos dest = (char *)(void *)h + h->upper;
510 1.1 cgd WR_RINTERNAL(dest,
511 1.1 cgd r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
512 1.1 cgd
513 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
514 1.1 cgd
515 1.1 cgd /* Unpin the root page, set to recno internal page. */
516 1.1 cgd h->flags &= ~P_TYPE;
517 1.1 cgd h->flags |= P_RINTERNAL;
518 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
519 1.1 cgd
520 1.1 cgd return (RET_SUCCESS);
521 1.1 cgd }
522 1.1 cgd
523 1.1 cgd /*
524 1.1 cgd * BT_BROOT -- Fix up the btree root page after it has been split.
525 1.1 cgd *
526 1.1 cgd * Parameters:
527 1.1 cgd * t: tree
528 1.1 cgd * h: root page
529 1.1 cgd * l: left page
530 1.1 cgd * r: right page
531 1.1 cgd *
532 1.1 cgd * Returns:
533 1.1 cgd * RET_ERROR, RET_SUCCESS
534 1.1 cgd */
535 1.1 cgd static int
536 1.14 christos bt_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
537 1.1 cgd {
538 1.7 christos BINTERNAL *bi = NULL; /* pacify gcc */
539 1.1 cgd BLEAF *bl;
540 1.15 joerg uint32_t nbytes;
541 1.1 cgd char *dest;
542 1.1 cgd
543 1.1 cgd /*
544 1.1 cgd * If the root page was a leaf page, change it into an internal page.
545 1.1 cgd * We copy the key we split on (but not the key's data, in the case of
546 1.1 cgd * a leaf page) to the new root page.
547 1.1 cgd *
548 1.1 cgd * The btree comparison code guarantees that the left-most key on any
549 1.1 cgd * level of the tree is never used, so it doesn't need to be filled in.
550 1.1 cgd */
551 1.1 cgd nbytes = NBINTERNAL(0);
552 1.1 cgd h->linp[0] = h->upper = t->bt_psize - nbytes;
553 1.10 christos dest = (char *)(void *)h + h->upper;
554 1.1 cgd WR_BINTERNAL(dest, 0, l->pgno, 0);
555 1.1 cgd
556 1.6 cgd switch (h->flags & P_TYPE) {
557 1.1 cgd case P_BLEAF:
558 1.1 cgd bl = GETBLEAF(r, 0);
559 1.1 cgd nbytes = NBINTERNAL(bl->ksize);
560 1.1 cgd h->linp[1] = h->upper -= nbytes;
561 1.10 christos dest = (char *)(void *)h + h->upper;
562 1.1 cgd WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
563 1.1 cgd memmove(dest, bl->bytes, bl->ksize);
564 1.1 cgd
565 1.1 cgd /*
566 1.1 cgd * If the key is on an overflow page, mark the overflow chain
567 1.1 cgd * so it isn't deleted when the leaf copy of the key is deleted.
568 1.1 cgd */
569 1.1 cgd if (bl->flags & P_BIGKEY &&
570 1.10 christos bt_preserve(t, *(pgno_t *)(void *)bl->bytes) == RET_ERROR)
571 1.1 cgd return (RET_ERROR);
572 1.1 cgd break;
573 1.1 cgd case P_BINTERNAL:
574 1.1 cgd bi = GETBINTERNAL(r, 0);
575 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
576 1.1 cgd h->linp[1] = h->upper -= nbytes;
577 1.10 christos dest = (char *)(void *)h + h->upper;
578 1.1 cgd memmove(dest, bi, nbytes);
579 1.10 christos ((BINTERNAL *)(void *)dest)->pgno = r->pgno;
580 1.1 cgd break;
581 1.1 cgd default:
582 1.1 cgd abort();
583 1.1 cgd }
584 1.1 cgd
585 1.1 cgd /* There are two keys on the page. */
586 1.1 cgd h->lower = BTDATAOFF + 2 * sizeof(indx_t);
587 1.1 cgd
588 1.1 cgd /* Unpin the root page, set to btree internal page. */
589 1.1 cgd h->flags &= ~P_TYPE;
590 1.1 cgd h->flags |= P_BINTERNAL;
591 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
592 1.1 cgd
593 1.1 cgd return (RET_SUCCESS);
594 1.1 cgd }
595 1.1 cgd
596 1.1 cgd /*
597 1.1 cgd * BT_PSPLIT -- Do the real work of splitting the page.
598 1.1 cgd *
599 1.1 cgd * Parameters:
600 1.1 cgd * t: tree
601 1.1 cgd * h: page to be split
602 1.1 cgd * l: page to put lower half of data
603 1.1 cgd * r: page to put upper half of data
604 1.1 cgd * pskip: pointer to index to leave open
605 1.1 cgd * ilen: insert length
606 1.1 cgd *
607 1.1 cgd * Returns:
608 1.1 cgd * Pointer to page in which to insert.
609 1.1 cgd */
610 1.1 cgd static PAGE *
611 1.14 christos bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
612 1.1 cgd {
613 1.1 cgd BINTERNAL *bi;
614 1.1 cgd BLEAF *bl;
615 1.6 cgd CURSOR *c;
616 1.1 cgd RLEAF *rl;
617 1.1 cgd PAGE *rval;
618 1.7 christos void *src = NULL; /* pacify gcc */
619 1.1 cgd indx_t full, half, nxt, off, skip, top, used;
620 1.15 joerg uint32_t nbytes;
621 1.14 christos size_t temp;
622 1.1 cgd int bigkeycnt, isbigkey;
623 1.1 cgd
624 1.1 cgd /*
625 1.1 cgd * Split the data to the left and right pages. Leave the skip index
626 1.1 cgd * open. Additionally, make some effort not to split on an overflow
627 1.1 cgd * key. This makes internal page processing faster and can save
628 1.1 cgd * space as overflow keys used by internal pages are never deleted.
629 1.1 cgd */
630 1.1 cgd bigkeycnt = 0;
631 1.1 cgd skip = *pskip;
632 1.14 christos temp = t->bt_psize - BTDATAOFF;
633 1.14 christos _DBFIT(temp, indx_t);
634 1.14 christos full = (indx_t)temp;
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 1.15 joerg _DBFIT(ilen, uint32_t);
640 1.15 joerg nbytes = (uint32_t)ilen;
641 1.1 cgd isbigkey = 0; /* XXX: not really known. */
642 1.1 cgd } else
643 1.1 cgd switch (h->flags & P_TYPE) {
644 1.1 cgd case P_BINTERNAL:
645 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
646 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
647 1.1 cgd isbigkey = bi->flags & P_BIGKEY;
648 1.1 cgd break;
649 1.1 cgd case P_BLEAF:
650 1.1 cgd src = bl = GETBLEAF(h, nxt);
651 1.1 cgd nbytes = NBLEAF(bl);
652 1.1 cgd isbigkey = bl->flags & P_BIGKEY;
653 1.1 cgd break;
654 1.1 cgd case P_RINTERNAL:
655 1.1 cgd src = GETRINTERNAL(h, nxt);
656 1.1 cgd nbytes = NRINTERNAL;
657 1.1 cgd isbigkey = 0;
658 1.1 cgd break;
659 1.1 cgd case P_RLEAF:
660 1.1 cgd src = rl = GETRLEAF(h, nxt);
661 1.1 cgd nbytes = NRLEAF(rl);
662 1.1 cgd isbigkey = 0;
663 1.1 cgd break;
664 1.1 cgd default:
665 1.1 cgd abort();
666 1.1 cgd }
667 1.1 cgd
668 1.1 cgd /*
669 1.1 cgd * If the key/data pairs are substantial fractions of the max
670 1.1 cgd * possible size for the page, it's possible to get situations
671 1.1 cgd * where we decide to try and copy too much onto the left page.
672 1.1 cgd * Make sure that doesn't happen.
673 1.1 cgd */
674 1.9 is if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
675 1.9 is nxt == top - 1) {
676 1.1 cgd --off;
677 1.1 cgd break;
678 1.1 cgd }
679 1.1 cgd
680 1.1 cgd /* Copy the key/data pair, if not the skipped index. */
681 1.1 cgd if (skip != off) {
682 1.1 cgd ++nxt;
683 1.1 cgd
684 1.1 cgd l->linp[off] = l->upper -= nbytes;
685 1.10 christos memmove((char *)(void *)l + l->upper, src, nbytes);
686 1.1 cgd }
687 1.1 cgd
688 1.14 christos temp = nbytes + sizeof(indx_t);
689 1.14 christos _DBFIT(temp, indx_t);
690 1.14 christos used += (indx_t)temp;
691 1.1 cgd if (used >= half) {
692 1.1 cgd if (!isbigkey || bigkeycnt == 3)
693 1.1 cgd break;
694 1.1 cgd else
695 1.1 cgd ++bigkeycnt;
696 1.1 cgd }
697 1.1 cgd }
698 1.1 cgd
699 1.1 cgd /*
700 1.1 cgd * Off is the last offset that's valid for the left page.
701 1.1 cgd * Nxt is the first offset to be placed on the right page.
702 1.1 cgd */
703 1.14 christos temp = (off + 1) * sizeof(indx_t);
704 1.14 christos _DBFIT(temp, indx_t);
705 1.14 christos l->lower += (indx_t)temp;
706 1.1 cgd
707 1.1 cgd /*
708 1.1 cgd * If splitting the page that the cursor was on, the cursor has to be
709 1.1 cgd * adjusted to point to the same record as before the split. If the
710 1.1 cgd * cursor is at or past the skipped slot, the cursor is incremented by
711 1.1 cgd * one. If the cursor is on the right page, it is decremented by the
712 1.1 cgd * number of records split to the left page.
713 1.1 cgd */
714 1.6 cgd c = &t->bt_cursor;
715 1.6 cgd if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
716 1.6 cgd if (c->pg.index >= skip)
717 1.6 cgd ++c->pg.index;
718 1.6 cgd if (c->pg.index < nxt) /* Left page. */
719 1.6 cgd c->pg.pgno = l->pgno;
720 1.1 cgd else { /* Right page. */
721 1.6 cgd c->pg.pgno = r->pgno;
722 1.6 cgd c->pg.index -= nxt;
723 1.1 cgd }
724 1.1 cgd }
725 1.1 cgd
726 1.1 cgd /*
727 1.1 cgd * If the skipped index was on the left page, just return that page.
728 1.1 cgd * Otherwise, adjust the skip index to reflect the new position on
729 1.1 cgd * the right page.
730 1.1 cgd */
731 1.1 cgd if (skip <= off) {
732 1.12 mycroft skip = MAX_PAGE_OFFSET;
733 1.1 cgd rval = l;
734 1.1 cgd } else {
735 1.1 cgd rval = r;
736 1.1 cgd *pskip -= nxt;
737 1.1 cgd }
738 1.1 cgd
739 1.1 cgd for (off = 0; nxt < top; ++off) {
740 1.1 cgd if (skip == nxt) {
741 1.1 cgd ++off;
742 1.12 mycroft skip = MAX_PAGE_OFFSET;
743 1.1 cgd }
744 1.1 cgd switch (h->flags & P_TYPE) {
745 1.1 cgd case P_BINTERNAL:
746 1.1 cgd src = bi = GETBINTERNAL(h, nxt);
747 1.1 cgd nbytes = NBINTERNAL(bi->ksize);
748 1.1 cgd break;
749 1.1 cgd case P_BLEAF:
750 1.1 cgd src = bl = GETBLEAF(h, nxt);
751 1.1 cgd nbytes = NBLEAF(bl);
752 1.1 cgd break;
753 1.1 cgd case P_RINTERNAL:
754 1.1 cgd src = GETRINTERNAL(h, nxt);
755 1.1 cgd nbytes = NRINTERNAL;
756 1.1 cgd break;
757 1.1 cgd case P_RLEAF:
758 1.1 cgd src = rl = GETRLEAF(h, nxt);
759 1.1 cgd nbytes = NRLEAF(rl);
760 1.1 cgd break;
761 1.1 cgd default:
762 1.1 cgd abort();
763 1.1 cgd }
764 1.1 cgd ++nxt;
765 1.1 cgd r->linp[off] = r->upper -= nbytes;
766 1.10 christos memmove((char *)(void *)r + r->upper, src, nbytes);
767 1.1 cgd }
768 1.14 christos temp = off * sizeof(indx_t);
769 1.14 christos _DBFIT(temp, indx_t);
770 1.14 christos r->lower += (indx_t)temp;
771 1.1 cgd
772 1.1 cgd /* If the key is being appended to the page, adjust the index. */
773 1.1 cgd if (skip == top)
774 1.1 cgd r->lower += sizeof(indx_t);
775 1.1 cgd
776 1.1 cgd return (rval);
777 1.1 cgd }
778 1.1 cgd
779 1.1 cgd /*
780 1.1 cgd * BT_PRESERVE -- Mark a chain of pages as used by an internal node.
781 1.1 cgd *
782 1.1 cgd * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the
783 1.1 cgd * record that references them gets deleted. Chains pointed to by internal
784 1.1 cgd * pages never get deleted. This routine marks a chain as pointed to by an
785 1.1 cgd * internal page.
786 1.1 cgd *
787 1.1 cgd * Parameters:
788 1.1 cgd * t: tree
789 1.1 cgd * pg: page number of first page in the chain.
790 1.1 cgd *
791 1.1 cgd * Returns:
792 1.1 cgd * RET_SUCCESS, RET_ERROR.
793 1.1 cgd */
794 1.1 cgd static int
795 1.14 christos bt_preserve(BTREE *t, pgno_t pg)
796 1.1 cgd {
797 1.1 cgd PAGE *h;
798 1.1 cgd
799 1.1 cgd if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
800 1.1 cgd return (RET_ERROR);
801 1.1 cgd h->flags |= P_PRESERVE;
802 1.1 cgd mpool_put(t->bt_mp, h, MPOOL_DIRTY);
803 1.1 cgd return (RET_SUCCESS);
804 1.1 cgd }
805 1.1 cgd
806 1.1 cgd /*
807 1.1 cgd * REC_TOTAL -- Return the number of recno entries below a page.
808 1.1 cgd *
809 1.1 cgd * Parameters:
810 1.1 cgd * h: page
811 1.1 cgd *
812 1.1 cgd * Returns:
813 1.1 cgd * The number of recno entries below a page.
814 1.1 cgd *
815 1.1 cgd * XXX
816 1.1 cgd * These values could be set by the bt_psplit routine. The problem is that the
817 1.1 cgd * entry has to be popped off of the stack etc. or the values have to be passed
818 1.1 cgd * all the way back to bt_split/bt_rroot and it's not very clean.
819 1.1 cgd */
820 1.1 cgd static recno_t
821 1.14 christos rec_total(PAGE *h)
822 1.1 cgd {
823 1.1 cgd recno_t recs;
824 1.1 cgd indx_t nxt, top;
825 1.1 cgd
826 1.1 cgd for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt)
827 1.1 cgd recs += GETRINTERNAL(h, nxt)->nrecs;
828 1.1 cgd return (recs);
829 1.1 cgd }
830