1 | /*- |
2 | * Copyright (c) 1990, 1993, 1994 |
3 | * The Regents of the University of California. All rights reserved. |
4 | * |
5 | * This code is derived from software contributed to Berkeley by |
6 | * Mike Olson. |
7 | * |
8 | * Redistribution and use in source and binary forms, with or without |
9 | * modification, are permitted provided that the following conditions |
10 | * are met: |
11 | * 1. Redistributions of source code must retain the above copyright |
12 | * notice, this list of conditions and the following disclaimer. |
13 | * 2. Redistributions in binary form must reproduce the above copyright |
14 | * notice, this list of conditions and the following disclaimer in the |
15 | * documentation and/or other materials provided with the distribution. |
16 | * 3. All advertising materials mentioning features or use of this software |
17 | * must display the following acknowledgement: |
18 | * This product includes software developed by the University of |
19 | * California, Berkeley and its contributors. |
20 | * 4. Neither the name of the University nor the names of its contributors |
21 | * may be used to endorse or promote products derived from this software |
22 | * without specific prior written permission. |
23 | * |
24 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND |
25 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
26 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
27 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE |
28 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
29 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
30 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
31 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
32 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
33 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
34 | * SUCH DAMAGE. |
35 | */ |
36 | |
37 | #if defined(LIBC_SCCS) && !defined(lint) |
38 | static char sccsid[] = "@(#)bt_split.c 8.9 (Berkeley) 7/26/94" ; |
39 | #endif /* LIBC_SCCS and not lint */ |
40 | |
41 | #include <sys/types.h> |
42 | |
43 | #include <limits.h> |
44 | #include <stdio.h> |
45 | #include <stdlib.h> |
46 | #include <string.h> |
47 | |
48 | #include <db.h> |
49 | #include "btree.h" |
50 | |
51 | static int bt_broot __P((BTREE *, PAGE *, PAGE *, PAGE *)); |
52 | static PAGE *bt_page |
53 | __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t)); |
54 | static int bt_preserve __P((BTREE *, pgno_t)); |
55 | static PAGE *bt_psplit |
56 | __P((BTREE *, PAGE *, PAGE *, PAGE *, indx_t *, size_t)); |
57 | static PAGE *bt_root |
58 | __P((BTREE *, PAGE *, PAGE **, PAGE **, indx_t *, size_t)); |
59 | static int bt_rroot __P((BTREE *, PAGE *, PAGE *, PAGE *)); |
60 | static recno_t rec_total __P((PAGE *)); |
61 | |
62 | #ifdef STATISTICS |
63 | u_long bt_rootsplit, bt_split, bt_sortsplit, bt_pfxsaved; |
64 | #endif |
65 | |
66 | /* |
67 | * __BT_SPLIT -- Split the tree. |
68 | * |
69 | * Parameters: |
70 | * t: tree |
71 | * sp: page to split |
72 | * key: key to insert |
73 | * data: data to insert |
74 | * flags: BIGKEY/BIGDATA flags |
75 | * ilen: insert length |
76 | * skip: index to leave open |
77 | * |
78 | * Returns: |
79 | * RET_ERROR, RET_SUCCESS |
80 | */ |
81 | int |
82 | __bt_split(t, sp, key, data, flags, ilen, argskip) |
83 | BTREE *t; |
84 | PAGE *sp; |
85 | const DBT *key, *data; |
86 | int flags; |
87 | size_t ilen; |
88 | u_int32_t argskip; |
89 | { |
90 | BINTERNAL *bi = NULL; |
91 | BLEAF *bl = NULL, *tbl; |
92 | DBT a, b; |
93 | EPGNO *parent; |
94 | PAGE *h, *l, *r, *lchild, *rchild; |
95 | indx_t nxtindex; |
96 | u_int16_t skip; |
97 | u_int32_t n, nbytes, nksize = 0; |
98 | int parentsplit; |
99 | char *dest; |
100 | pgno_t pg_tmp; |
101 | |
102 | /* |
103 | * Split the page into two pages, l and r. The split routines return |
104 | * a pointer to the page into which the key should be inserted and with |
105 | * skip set to the offset which should be used. Additionally, l and r |
106 | * are pinned. |
107 | */ |
108 | skip = argskip; |
109 | h = sp->pgno == P_ROOT ? |
110 | bt_root(t, sp, &l, &r, &skip, ilen) : |
111 | bt_page(t, sp, &l, &r, &skip, ilen); |
112 | if (h == NULL) |
113 | return (RET_ERROR); |
114 | |
115 | /* |
116 | * Insert the new key/data pair into the leaf page. (Key inserts |
117 | * always cause a leaf page to split first.) |
118 | */ |
119 | h->linp[skip] = h->upper -= ilen; |
120 | dest = (char *)h + h->upper; |
121 | if (F_ISSET(t, R_RECNO)) |
122 | WR_RLEAF(dest, data, flags) |
123 | else |
124 | WR_BLEAF(dest, key, data, flags) |
125 | |
126 | /* If the root page was split, make it look right. */ |
127 | if (sp->pgno == P_ROOT && |
128 | (F_ISSET(t, R_RECNO) ? |
129 | bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) |
130 | goto err2; |
131 | |
132 | /* |
133 | * Now we walk the parent page stack -- a LIFO stack of the pages that |
134 | * were traversed when we searched for the page that split. Each stack |
135 | * entry is a page number and a page index offset. The offset is for |
136 | * the page traversed on the search. We've just split a page, so we |
137 | * have to insert a new key into the parent page. |
138 | * |
139 | * If the insert into the parent page causes it to split, may have to |
140 | * continue splitting all the way up the tree. We stop if the root |
141 | * splits or the page inserted into didn't have to split to hold the |
142 | * new key. Some algorithms replace the key for the old page as well |
143 | * as the new page. We don't, as there's no reason to believe that the |
144 | * first key on the old page is any better than the key we have, and, |
145 | * in the case of a key being placed at index 0 causing the split, the |
146 | * key is unavailable. |
147 | * |
148 | * There are a maximum of 5 pages pinned at any time. We keep the left |
149 | * and right pages pinned while working on the parent. The 5 are the |
150 | * two children, left parent and right parent (when the parent splits) |
151 | * and the root page or the overflow key page when calling bt_preserve. |
152 | * This code must make sure that all pins are released other than the |
153 | * root page or overflow page which is unlocked elsewhere. |
154 | */ |
155 | while ((parent = BT_POP(t)) != NULL) { |
156 | lchild = l; |
157 | rchild = r; |
158 | |
159 | /* Get the parent page. */ |
160 | if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL) |
161 | goto err2; |
162 | |
163 | /* |
164 | * The new key goes ONE AFTER the index, because the split |
165 | * was to the right. |
166 | */ |
167 | skip = parent->index + 1; |
168 | |
169 | /* |
170 | * Calculate the space needed on the parent page. |
171 | * |
172 | * Prefix trees: space hack when inserting into BINTERNAL |
173 | * pages. Retain only what's needed to distinguish between |
174 | * the new entry and the LAST entry on the page to its left. |
175 | * If the keys compare equal, retain the entire key. Note, |
176 | * we don't touch overflow keys, and the entire key must be |
177 | * retained for the next-to-left most key on the leftmost |
178 | * page of each level, or the search will fail. Applicable |
179 | * ONLY to internal pages that have leaf pages as children. |
180 | * Further reduction of the key between pairs of internal |
181 | * pages loses too much information. |
182 | */ |
183 | switch (rchild->flags & P_TYPE) { |
184 | case P_BINTERNAL: |
185 | bi = GETBINTERNAL(rchild, 0); |
186 | nbytes = NBINTERNAL(bi->ksize); |
187 | break; |
188 | case P_BLEAF: |
189 | bl = GETBLEAF(rchild, 0); |
190 | nbytes = NBINTERNAL(bl->ksize); |
191 | if (t->bt_pfx && !(bl->flags & P_BIGKEY) && |
192 | (h->prevpg != P_INVALID || skip > 1)) { |
193 | tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1); |
194 | a.size = tbl->ksize; |
195 | a.data = tbl->bytes; |
196 | b.size = bl->ksize; |
197 | b.data = bl->bytes; |
198 | nksize = t->bt_pfx(&a, &b); |
199 | n = NBINTERNAL(nksize); |
200 | if (n < nbytes) { |
201 | #ifdef STATISTICS |
202 | bt_pfxsaved += nbytes - n; |
203 | #endif |
204 | nbytes = n; |
205 | } else |
206 | nksize = 0; |
207 | } else |
208 | nksize = 0; |
209 | break; |
210 | case P_RINTERNAL: |
211 | case P_RLEAF: |
212 | nbytes = NRINTERNAL; |
213 | break; |
214 | default: |
215 | abort(); |
216 | } |
217 | |
218 | /* Split the parent page if necessary or shift the indices. */ |
219 | if (h->upper - h->lower < nbytes + sizeof(indx_t)) { |
220 | sp = h; |
221 | h = h->pgno == P_ROOT ? |
222 | bt_root(t, h, &l, &r, &skip, nbytes) : |
223 | bt_page(t, h, &l, &r, &skip, nbytes); |
224 | if (h == NULL) |
225 | goto err1; |
226 | parentsplit = 1; |
227 | } else { |
228 | if (skip < (nxtindex = NEXTINDEX(h))) |
229 | memmove(h->linp + skip + 1, h->linp + skip, |
230 | (nxtindex - skip) * sizeof(indx_t)); |
231 | h->lower += sizeof(indx_t); |
232 | parentsplit = 0; |
233 | } |
234 | |
235 | /* Insert the key into the parent page. */ |
236 | switch (rchild->flags & P_TYPE) { |
237 | case P_BINTERNAL: |
238 | h->linp[skip] = h->upper -= nbytes; |
239 | dest = (char *)h + h->linp[skip]; |
240 | memmove(dest, bi, nbytes); |
241 | ((BINTERNAL *)dest)->pgno = rchild->pgno; |
242 | break; |
243 | case P_BLEAF: |
244 | h->linp[skip] = h->upper -= nbytes; |
245 | dest = (char *)h + h->linp[skip]; |
246 | WR_BINTERNAL(dest, nksize ? nksize : bl->ksize, |
247 | rchild->pgno, bl->flags & P_BIGKEY); |
248 | memmove(dest, bl->bytes, nksize ? nksize : bl->ksize); |
249 | /* Avoid alignment violation */ |
250 | memcpy(&pg_tmp, bl->bytes, sizeof(pgno_t)); |
251 | if (bl->flags & P_BIGKEY && |
252 | bt_preserve(t, pg_tmp) == RET_ERROR) |
253 | goto err1; |
254 | break; |
255 | case P_RINTERNAL: |
256 | /* |
257 | * Update the left page count. If split |
258 | * added at index 0, fix the correct page. |
259 | */ |
260 | if (skip > 0) |
261 | dest = (char *)h + h->linp[skip - 1]; |
262 | else |
263 | dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; |
264 | ((RINTERNAL *)dest)->nrecs = rec_total(lchild); |
265 | ((RINTERNAL *)dest)->pgno = lchild->pgno; |
266 | |
267 | /* Update the right page count. */ |
268 | h->linp[skip] = h->upper -= nbytes; |
269 | dest = (char *)h + h->linp[skip]; |
270 | ((RINTERNAL *)dest)->nrecs = rec_total(rchild); |
271 | ((RINTERNAL *)dest)->pgno = rchild->pgno; |
272 | break; |
273 | case P_RLEAF: |
274 | /* |
275 | * Update the left page count. If split |
276 | * added at index 0, fix the correct page. |
277 | */ |
278 | if (skip > 0) |
279 | dest = (char *)h + h->linp[skip - 1]; |
280 | else |
281 | dest = (char *)l + l->linp[NEXTINDEX(l) - 1]; |
282 | ((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild); |
283 | ((RINTERNAL *)dest)->pgno = lchild->pgno; |
284 | |
285 | /* Update the right page count. */ |
286 | h->linp[skip] = h->upper -= nbytes; |
287 | dest = (char *)h + h->linp[skip]; |
288 | ((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild); |
289 | ((RINTERNAL *)dest)->pgno = rchild->pgno; |
290 | break; |
291 | default: |
292 | abort(); |
293 | } |
294 | |
295 | /* Unpin the held pages. */ |
296 | if (!parentsplit) { |
297 | mpool_put(t->bt_mp, h, MPOOL_DIRTY); |
298 | break; |
299 | } |
300 | |
301 | /* If the root page was split, make it look right. */ |
302 | if (sp->pgno == P_ROOT && |
303 | (F_ISSET(t, R_RECNO) ? |
304 | bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR) |
305 | goto err1; |
306 | |
307 | mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); |
308 | mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); |
309 | } |
310 | |
311 | /* Unpin the held pages. */ |
312 | mpool_put(t->bt_mp, l, MPOOL_DIRTY); |
313 | mpool_put(t->bt_mp, r, MPOOL_DIRTY); |
314 | |
315 | /* Clear any pages left on the stack. */ |
316 | return (RET_SUCCESS); |
317 | |
318 | /* |
319 | * If something fails in the above loop we were already walking back |
320 | * up the tree and the tree is now inconsistent. Nothing much we can |
321 | * do about it but release any memory we're holding. |
322 | */ |
323 | err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY); |
324 | mpool_put(t->bt_mp, rchild, MPOOL_DIRTY); |
325 | |
326 | err2: mpool_put(t->bt_mp, l, 0); |
327 | mpool_put(t->bt_mp, r, 0); |
328 | __dbpanic(t->bt_dbp); |
329 | return (RET_ERROR); |
330 | } |
331 | |
332 | /* |
333 | * BT_PAGE -- Split a non-root page of a btree. |
334 | * |
335 | * Parameters: |
336 | * t: tree |
337 | * h: root page |
338 | * lp: pointer to left page pointer |
339 | * rp: pointer to right page pointer |
340 | * skip: pointer to index to leave open |
341 | * ilen: insert length |
342 | * |
343 | * Returns: |
344 | * Pointer to page in which to insert or NULL on error. |
345 | */ |
346 | static PAGE * |
347 | bt_page(t, h, lp, rp, skip, ilen) |
348 | BTREE *t; |
349 | PAGE *h, **lp, **rp; |
350 | indx_t *skip; |
351 | size_t ilen; |
352 | { |
353 | PAGE *l, *r, *tp; |
354 | pgno_t npg; |
355 | |
356 | #ifdef STATISTICS |
357 | ++bt_split; |
358 | #endif |
359 | /* Put the new right page for the split into place. */ |
360 | if ((r = __bt_new(t, &npg)) == NULL) |
361 | return (NULL); |
362 | r->pgno = npg; |
363 | r->lower = BTDATAOFF; |
364 | r->upper = t->bt_psize; |
365 | r->nextpg = h->nextpg; |
366 | r->prevpg = h->pgno; |
367 | r->flags = h->flags & P_TYPE; |
368 | |
369 | /* |
370 | * If we're splitting the last page on a level because we're appending |
371 | * a key to it (skip is NEXTINDEX()), it's likely that the data is |
372 | * sorted. Adding an empty page on the side of the level is less work |
373 | * and can push the fill factor much higher than normal. If we're |
374 | * wrong it's no big deal, we'll just do the split the right way next |
375 | * time. It may look like it's equally easy to do a similar hack for |
376 | * reverse sorted data, that is, split the tree left, but it's not. |
377 | * Don't even try. |
378 | */ |
379 | if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) { |
380 | #ifdef STATISTICS |
381 | ++bt_sortsplit; |
382 | #endif |
383 | h->nextpg = r->pgno; |
384 | r->lower = BTDATAOFF + sizeof(indx_t); |
385 | *skip = 0; |
386 | *lp = h; |
387 | *rp = r; |
388 | return (r); |
389 | } |
390 | |
391 | /* Put the new left page for the split into place. */ |
392 | if ((l = (PAGE *)malloc(t->bt_psize)) == NULL) { |
393 | mpool_put(t->bt_mp, r, 0); |
394 | return (NULL); |
395 | } |
396 | #ifdef PURIFY |
397 | memset(l, 0xff, t->bt_psize); |
398 | #endif |
399 | l->pgno = h->pgno; |
400 | l->nextpg = r->pgno; |
401 | l->prevpg = h->prevpg; |
402 | l->lower = BTDATAOFF; |
403 | l->upper = t->bt_psize; |
404 | l->flags = h->flags & P_TYPE; |
405 | |
406 | /* Fix up the previous pointer of the page after the split page. */ |
407 | if (h->nextpg != P_INVALID) { |
408 | if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) { |
409 | free(l); |
410 | /* XXX mpool_free(t->bt_mp, r->pgno); */ |
411 | return (NULL); |
412 | } |
413 | tp->prevpg = r->pgno; |
414 | mpool_put(t->bt_mp, tp, MPOOL_DIRTY); |
415 | } |
416 | |
417 | /* |
418 | * Split right. The key/data pairs aren't sorted in the btree page so |
419 | * it's simpler to copy the data from the split page onto two new pages |
420 | * instead of copying half the data to the right page and compacting |
421 | * the left page in place. Since the left page can't change, we have |
422 | * to swap the original and the allocated left page after the split. |
423 | */ |
424 | tp = bt_psplit(t, h, l, r, skip, ilen); |
425 | |
426 | /* Move the new left page onto the old left page. */ |
427 | memmove(h, l, t->bt_psize); |
428 | if (tp == l) |
429 | tp = h; |
430 | free(l); |
431 | |
432 | *lp = h; |
433 | *rp = r; |
434 | return (tp); |
435 | } |
436 | |
437 | /* |
438 | * BT_ROOT -- Split the root page of a btree. |
439 | * |
440 | * Parameters: |
441 | * t: tree |
442 | * h: root page |
443 | * lp: pointer to left page pointer |
444 | * rp: pointer to right page pointer |
445 | * skip: pointer to index to leave open |
446 | * ilen: insert length |
447 | * |
448 | * Returns: |
449 | * Pointer to page in which to insert or NULL on error. |
450 | */ |
451 | static PAGE * |
452 | bt_root(t, h, lp, rp, skip, ilen) |
453 | BTREE *t; |
454 | PAGE *h, **lp, **rp; |
455 | indx_t *skip; |
456 | size_t ilen; |
457 | { |
458 | PAGE *l, *r, *tp; |
459 | pgno_t lnpg, rnpg; |
460 | |
461 | #ifdef STATISTICS |
462 | ++bt_split; |
463 | ++bt_rootsplit; |
464 | #endif |
465 | /* Put the new left and right pages for the split into place. */ |
466 | if ((l = __bt_new(t, &lnpg)) == NULL || |
467 | (r = __bt_new(t, &rnpg)) == NULL) |
468 | return (NULL); |
469 | l->pgno = lnpg; |
470 | r->pgno = rnpg; |
471 | l->nextpg = r->pgno; |
472 | r->prevpg = l->pgno; |
473 | l->prevpg = r->nextpg = P_INVALID; |
474 | l->lower = r->lower = BTDATAOFF; |
475 | l->upper = r->upper = t->bt_psize; |
476 | l->flags = r->flags = h->flags & P_TYPE; |
477 | |
478 | /* Split the root page. */ |
479 | tp = bt_psplit(t, h, l, r, skip, ilen); |
480 | |
481 | *lp = l; |
482 | *rp = r; |
483 | return (tp); |
484 | } |
485 | |
486 | /* |
487 | * BT_RROOT -- Fix up the recno root page after it has been split. |
488 | * |
489 | * Parameters: |
490 | * t: tree |
491 | * h: root page |
492 | * l: left page |
493 | * r: right page |
494 | * |
495 | * Returns: |
496 | * RET_ERROR, RET_SUCCESS |
497 | */ |
498 | static int |
499 | bt_rroot(t, h, l, r) |
500 | BTREE *t; |
501 | PAGE *h, *l, *r; |
502 | { |
503 | char *dest; |
504 | |
505 | /* Insert the left and right keys, set the header information. */ |
506 | h->linp[0] = h->upper = t->bt_psize - NRINTERNAL; |
507 | dest = (char *)h + h->upper; |
508 | WR_RINTERNAL(dest, |
509 | l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno); |
510 | |
511 | h->linp[1] = h->upper -= NRINTERNAL; |
512 | dest = (char *)h + h->upper; |
513 | WR_RINTERNAL(dest, |
514 | r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno); |
515 | |
516 | h->lower = BTDATAOFF + 2 * sizeof(indx_t); |
517 | |
518 | /* Unpin the root page, set to recno internal page. */ |
519 | h->flags &= ~P_TYPE; |
520 | h->flags |= P_RINTERNAL; |
521 | mpool_put(t->bt_mp, h, MPOOL_DIRTY); |
522 | |
523 | return (RET_SUCCESS); |
524 | } |
525 | |
526 | /* |
527 | * BT_BROOT -- Fix up the btree root page after it has been split. |
528 | * |
529 | * Parameters: |
530 | * t: tree |
531 | * h: root page |
532 | * l: left page |
533 | * r: right page |
534 | * |
535 | * Returns: |
536 | * RET_ERROR, RET_SUCCESS |
537 | */ |
538 | static int |
539 | bt_broot(t, h, l, r) |
540 | BTREE *t; |
541 | PAGE *h, *l, *r; |
542 | { |
543 | BINTERNAL *bi; |
544 | BLEAF *bl; |
545 | u_int32_t nbytes; |
546 | char *dest; |
547 | pgno_t pg_tmp; |
548 | |
549 | /* |
550 | * If the root page was a leaf page, change it into an internal page. |
551 | * We copy the key we split on (but not the key's data, in the case of |
552 | * a leaf page) to the new root page. |
553 | * |
554 | * The btree comparison code guarantees that the left-most key on any |
555 | * level of the tree is never used, so it doesn't need to be filled in. |
556 | */ |
557 | nbytes = NBINTERNAL(0); |
558 | h->linp[0] = h->upper = t->bt_psize - nbytes; |
559 | dest = (char *)h + h->upper; |
560 | WR_BINTERNAL(dest, 0, l->pgno, 0); |
561 | |
562 | switch (h->flags & P_TYPE) { |
563 | case P_BLEAF: |
564 | bl = GETBLEAF(r, 0); |
565 | nbytes = NBINTERNAL(bl->ksize); |
566 | h->linp[1] = h->upper -= nbytes; |
567 | dest = (char *)h + h->upper; |
568 | WR_BINTERNAL(dest, bl->ksize, r->pgno, 0); |
569 | memmove(dest, bl->bytes, bl->ksize); |
570 | |
571 | /* |
572 | * If the key is on an overflow page, mark the overflow chain |
573 | * so it isn't deleted when the leaf copy of the key is deleted. |
574 | */ |
575 | /* Avoid alignment violation */ |
576 | memcpy(&pg_tmp, bl->bytes, sizeof(pgno_t)); |
577 | if (bl->flags & P_BIGKEY && |
578 | bt_preserve(t, pg_tmp) == RET_ERROR) |
579 | return (RET_ERROR); |
580 | break; |
581 | case P_BINTERNAL: |
582 | bi = GETBINTERNAL(r, 0); |
583 | nbytes = NBINTERNAL(bi->ksize); |
584 | h->linp[1] = h->upper -= nbytes; |
585 | dest = (char *)h + h->upper; |
586 | memmove(dest, bi, nbytes); |
587 | ((BINTERNAL *)dest)->pgno = r->pgno; |
588 | break; |
589 | default: |
590 | abort(); |
591 | } |
592 | |
593 | /* There are two keys on the page. */ |
594 | h->lower = BTDATAOFF + 2 * sizeof(indx_t); |
595 | |
596 | /* Unpin the root page, set to btree internal page. */ |
597 | h->flags &= ~P_TYPE; |
598 | h->flags |= P_BINTERNAL; |
599 | mpool_put(t->bt_mp, h, MPOOL_DIRTY); |
600 | |
601 | return (RET_SUCCESS); |
602 | } |
603 | |
604 | /* |
605 | * BT_PSPLIT -- Do the real work of splitting the page. |
606 | * |
607 | * Parameters: |
608 | * t: tree |
609 | * h: page to be split |
610 | * l: page to put lower half of data |
611 | * r: page to put upper half of data |
612 | * pskip: pointer to index to leave open |
613 | * ilen: insert length |
614 | * |
615 | * Returns: |
616 | * Pointer to page in which to insert. |
617 | */ |
618 | static PAGE * |
619 | bt_psplit(t, h, l, r, pskip, ilen) |
620 | BTREE *t; |
621 | PAGE *h, *l, *r; |
622 | indx_t *pskip; |
623 | size_t ilen; |
624 | { |
625 | BINTERNAL *bi; |
626 | BLEAF *bl; |
627 | CURSOR *c; |
628 | RLEAF *rl; |
629 | PAGE *rval; |
630 | void *src = NULL; |
631 | indx_t full, half, nxt, off, skip, top, used; |
632 | u_int32_t nbytes; |
633 | int bigkeycnt, isbigkey; |
634 | |
635 | /* |
636 | * Split the data to the left and right pages. Leave the skip index |
637 | * open. Additionally, make some effort not to split on an overflow |
638 | * key. This makes internal page processing faster and can save |
639 | * space as overflow keys used by internal pages are never deleted. |
640 | */ |
641 | bigkeycnt = 0; |
642 | skip = *pskip; |
643 | full = t->bt_psize - BTDATAOFF; |
644 | half = full / 2; |
645 | used = 0; |
646 | for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) { |
647 | if (skip == off) { |
648 | nbytes = ilen; |
649 | isbigkey = 0; /* XXX: not really known. */ |
650 | } else |
651 | switch (h->flags & P_TYPE) { |
652 | case P_BINTERNAL: |
653 | src = bi = GETBINTERNAL(h, nxt); |
654 | nbytes = NBINTERNAL(bi->ksize); |
655 | isbigkey = bi->flags & P_BIGKEY; |
656 | break; |
657 | case P_BLEAF: |
658 | src = bl = GETBLEAF(h, nxt); |
659 | nbytes = NBLEAF(bl); |
660 | isbigkey = bl->flags & P_BIGKEY; |
661 | break; |
662 | case P_RINTERNAL: |
663 | src = GETRINTERNAL(h, nxt); |
664 | nbytes = NRINTERNAL; |
665 | isbigkey = 0; |
666 | break; |
667 | case P_RLEAF: |
668 | src = rl = GETRLEAF(h, nxt); |
669 | nbytes = NRLEAF(rl); |
670 | isbigkey = 0; |
671 | break; |
672 | default: |
673 | abort(); |
674 | } |
675 | |
676 | /* |
677 | * If the key/data pairs are substantial fractions of the max |
678 | * possible size for the page, it's possible to get situations |
679 | * where we decide to try and copy too much onto the left page. |
680 | * Make sure that doesn't happen. |
681 | */ |
682 | if ((skip <= off && |
683 | used + nbytes + sizeof(indx_t) >= full) || nxt == top - 1) { |
684 | --off; |
685 | break; |
686 | } |
687 | |
688 | /* Copy the key/data pair, if not the skipped index. */ |
689 | if (skip != off) { |
690 | ++nxt; |
691 | |
692 | l->linp[off] = l->upper -= nbytes; |
693 | memmove((char *)l + l->upper, src, nbytes); |
694 | } |
695 | |
696 | used += nbytes + sizeof(indx_t); |
697 | if (used >= half) { |
698 | if (!isbigkey || bigkeycnt == 3) |
699 | break; |
700 | else |
701 | ++bigkeycnt; |
702 | } |
703 | } |
704 | |
705 | /* |
706 | * Off is the last offset that's valid for the left page. |
707 | * Nxt is the first offset to be placed on the right page. |
708 | */ |
709 | l->lower += (off + 1) * sizeof(indx_t); |
710 | |
711 | /* |
712 | * If splitting the page that the cursor was on, the cursor has to be |
713 | * adjusted to point to the same record as before the split. If the |
714 | * cursor is at or past the skipped slot, the cursor is incremented by |
715 | * one. If the cursor is on the right page, it is decremented by the |
716 | * number of records split to the left page. |
717 | */ |
718 | c = &t->bt_cursor; |
719 | if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) { |
720 | if (c->pg.index >= skip) |
721 | ++c->pg.index; |
722 | if (c->pg.index < nxt) /* Left page. */ |
723 | c->pg.pgno = l->pgno; |
724 | else { /* Right page. */ |
725 | c->pg.pgno = r->pgno; |
726 | c->pg.index -= nxt; |
727 | } |
728 | } |
729 | |
730 | /* |
731 | * If the skipped index was on the left page, just return that page. |
732 | * Otherwise, adjust the skip index to reflect the new position on |
733 | * the right page. |
734 | */ |
735 | if (skip <= off) { |
736 | skip = 0; |
737 | rval = l; |
738 | } else { |
739 | rval = r; |
740 | *pskip -= nxt; |
741 | } |
742 | |
743 | for (off = 0; nxt < top; ++off) { |
744 | if (skip == nxt) { |
745 | ++off; |
746 | skip = 0; |
747 | } |
748 | switch (h->flags & P_TYPE) { |
749 | case P_BINTERNAL: |
750 | src = bi = GETBINTERNAL(h, nxt); |
751 | nbytes = NBINTERNAL(bi->ksize); |
752 | break; |
753 | case P_BLEAF: |
754 | src = bl = GETBLEAF(h, nxt); |
755 | nbytes = NBLEAF(bl); |
756 | break; |
757 | case P_RINTERNAL: |
758 | src = GETRINTERNAL(h, nxt); |
759 | nbytes = NRINTERNAL; |
760 | break; |
761 | case P_RLEAF: |
762 | src = rl = GETRLEAF(h, nxt); |
763 | nbytes = NRLEAF(rl); |
764 | break; |
765 | default: |
766 | abort(); |
767 | } |
768 | ++nxt; |
769 | r->linp[off] = r->upper -= nbytes; |
770 | memmove((char *)r + r->upper, src, nbytes); |
771 | } |
772 | r->lower += off * sizeof(indx_t); |
773 | |
774 | /* If the key is being appended to the page, adjust the index. */ |
775 | if (skip == top) |
776 | r->lower += sizeof(indx_t); |
777 | |
778 | return (rval); |
779 | } |
780 | |
781 | /* |
782 | * BT_PRESERVE -- Mark a chain of pages as used by an internal node. |
783 | * |
784 | * Chains of indirect blocks pointed to by leaf nodes get reclaimed when the |
785 | * record that references them gets deleted. Chains pointed to by internal |
786 | * pages never get deleted. This routine marks a chain as pointed to by an |
787 | * internal page. |
788 | * |
789 | * Parameters: |
790 | * t: tree |
791 | * pg: page number of first page in the chain. |
792 | * |
793 | * Returns: |
794 | * RET_SUCCESS, RET_ERROR. |
795 | */ |
796 | static int |
797 | bt_preserve(t, pg) |
798 | BTREE *t; |
799 | pgno_t pg; |
800 | { |
801 | PAGE *h; |
802 | |
803 | if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL) |
804 | return (RET_ERROR); |
805 | h->flags |= P_PRESERVE; |
806 | mpool_put(t->bt_mp, h, MPOOL_DIRTY); |
807 | return (RET_SUCCESS); |
808 | } |
809 | |
810 | /* |
811 | * REC_TOTAL -- Return the number of recno entries below a page. |
812 | * |
813 | * Parameters: |
814 | * h: page |
815 | * |
816 | * Returns: |
817 | * The number of recno entries below a page. |
818 | * |
819 | * XXX |
820 | * These values could be set by the bt_psplit routine. The problem is that the |
821 | * entry has to be popped off of the stack etc. or the values have to be passed |
822 | * all the way back to bt_split/bt_rroot and it's not very clean. |
823 | */ |
824 | static recno_t |
825 | rec_total(h) |
826 | PAGE *h; |
827 | { |
828 | recno_t recs; |
829 | indx_t nxt, top; |
830 | |
831 | for (recs = 0, nxt = 0, top = NEXTINDEX(h); nxt < top; ++nxt) |
832 | recs += GETRINTERNAL(h, nxt)->nrecs; |
833 | return (recs); |
834 | } |
835 | |