db/btree/btree.h

   1 /*
   2  * Copyright (c) 2003 Apple Computer, Inc. All rights reserved.
   3  *
   4  * @APPLE_LICENSE_HEADER_START@
   5  *
   6  * Copyright (c) 1999-2003 Apple Computer, Inc.  All Rights Reserved.
   7  *
   8  * This file contains Original Code and/or Modifications of Original Code
   9  * as defined in and that are subject to the Apple Public Source License
  10  * Version 2.0 (the 'License'). You may not use this file except in
  11  * compliance with the License. Please obtain a copy of the License at
  12  * http://www.opensource.apple.com/apsl/ and read it before using this
  13  * file.
  14  *
  15  * The Original Code and all software distributed under the License are
  16  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  17  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  18  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  20  * Please see the License for the specific language governing rights and
  21  * limitations under the License.
  22  *
  23  * @APPLE_LICENSE_HEADER_END@
  24  */
  25 /*-
  26  * Copyright (c) 1991, 1993, 1994
  27  *      The Regents of the University of California.  All rights reserved.
  28  *
  29  * This code is derived from software contributed to Berkeley by
  30  * Mike Olson.
  31  *
  32  * Redistribution and use in source and binary forms, with or without
  33  * modification, are permitted provided that the following conditions
  34  * are met:
  35  * 1. Redistributions of source code must retain the above copyright
  36  *    notice, this list of conditions and the following disclaimer.
  37  * 2. Redistributions in binary form must reproduce the above copyright
  38  *    notice, this list of conditions and the following disclaimer in the
  39  *    documentation and/or other materials provided with the distribution.
  40  * 3. All advertising materials mentioning features or use of this software
  41  *    must display the following acknowledgement:
  42  *      This product includes software developed by the University of
  43  *      California, Berkeley and its contributors.
  44  * 4. Neither the name of the University nor the names of its contributors
  45  *    may be used to endorse or promote products derived from this software
  46  *    without specific prior written permission.
  47  *
  48  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  49  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  50  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  51  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  52  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  53  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  54  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  55  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  56  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  57  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  58  * SUCH DAMAGE.
  59  *
  60  *      @(#)btree.h     8.11 (Berkeley) 8/17/94
  61  * $FreeBSD: src/lib/libc/db/btree/btree.h,v 1.3 2002/03/22 23:41:40 obrien Exp $
  62  */
  63
  64 /* Macros to set/clear/test flags. */
  65 #define F_SET(p, f)     (p)->flags |= (f)
  66 #define F_CLR(p, f)     (p)->flags &= ~(f)
  67 #define F_ISSET(p, f)   ((p)->flags & (f))
  68
  69 #include <mpool.h>
  70
  71 #define DEFMINKEYPAGE   (2)             /* Minimum keys per page */
  72 #define MINCACHE        (5)             /* Minimum cached pages */
  73 #define MINPSIZE        (512)           /* Minimum page size */
  74
  75 /*
  76  * Page 0 of a btree file contains a copy of the meta-data.  This page is also
  77  * used as an out-of-band page, i.e. page pointers that point to nowhere point
  78  * to page 0.  Page 1 is the root of the btree.
  79  */
  80 #define P_INVALID        0              /* Invalid tree page number. */
  81 #define P_META           0              /* Tree metadata page number. */
  82 #define P_ROOT           1              /* Tree root page number. */
  83
  84 /*
  85  * There are five page layouts in the btree: btree internal pages (BINTERNAL),
  86  * btree leaf pages (BLEAF), recno internal pages (RINTERNAL), recno leaf pages
  87  * (RLEAF) and overflow pages.  All five page types have a page header (PAGE).
  88  * This implementation requires that values within structures NOT be padded.
  89  * (ANSI C permits random padding.)  If your compiler pads randomly you'll have
  90  * to do some work to get this package to run.
  91  */
  92 typedef struct _page {
  93         pgno_t  pgno;                   /* this page's page number */
  94         pgno_t  prevpg;                 /* left sibling */
  95         pgno_t  nextpg;                 /* right sibling */
  96
  97 #define P_BINTERNAL     0x01            /* btree internal page */
  98 #define P_BLEAF         0x02            /* leaf page */
  99 #define P_OVERFLOW      0x04            /* overflow page */
 100 #define P_RINTERNAL     0x08            /* recno internal page */
 101 #define P_RLEAF         0x10            /* leaf page */
 102 #define P_TYPE          0x1f            /* type mask */
 103 #define P_PRESERVE      0x20            /* never delete this chain of pages */
 104         u_int32_t flags;
 105
 106         indx_t  lower;                  /* lower bound of free space on page */
 107         indx_t  upper;                  /* upper bound of free space on page */
 108         indx_t  linp[1];                /* indx_t-aligned VAR. LENGTH DATA */
 109 } PAGE;
 110
 111 /* First and next index. */
 112 #define BTDATAOFF                                                       \
 113         (sizeof(pgno_t) + sizeof(pgno_t) + sizeof(pgno_t) +             \
 114             sizeof(u_int32_t) + sizeof(indx_t) + sizeof(indx_t))
 115 #define NEXTINDEX(p)    (((p)->lower - BTDATAOFF) / sizeof(indx_t))
 116
 117 /*
 118  * For pages other than overflow pages, there is an array of offsets into the
 119  * rest of the page immediately following the page header.  Each offset is to
 120  * an item which is unique to the type of page.  The h_lower offset is just
 121  * past the last filled-in index.  The h_upper offset is the first item on the
 122  * page.  Offsets are from the beginning of the page.
 123  *
 124  * If an item is too big to store on a single page, a flag is set and the item
 125  * is a { page, size } pair such that the page is the first page of an overflow
 126  * chain with size bytes of item.  Overflow pages are simply bytes without any
 127  * external structure.
 128  *
 129  * The page number and size fields in the items are pgno_t-aligned so they can
 130  * be manipulated without copying.  (This presumes that 32 bit items can be
 131  * manipulated on this system.)
 132  */
 133 #define LALIGN(n)       (((n) + sizeof(pgno_t) - 1) & ~(sizeof(pgno_t) - 1))
 134 #define NOVFLSIZE       (sizeof(pgno_t) + sizeof(u_int32_t))
 135
 136 /*
 137  * For the btree internal pages, the item is a key.  BINTERNALs are {key, pgno}
 138  * pairs, such that the key compares less than or equal to all of the records
 139  * on that page.  For a tree without duplicate keys, an internal page with two
 140  * consecutive keys, a and b, will have all records greater than or equal to a
 141  * and less than b stored on the page associated with a.  Duplicate keys are
 142  * somewhat special and can cause duplicate internal and leaf page records and
 143  * some minor modifications of the above rule.
 144  */
 145 typedef struct _binternal {
 146         u_int32_t ksize;                /* key size */
 147         pgno_t  pgno;                   /* page number stored on */
 148 #define P_BIGDATA       0x01            /* overflow data */
 149 #define P_BIGKEY        0x02            /* overflow key */
 150         u_char  flags;
 151         char    bytes[1];               /* data */
 152 } BINTERNAL;
 153
 154 /* Get the page's BINTERNAL structure at index indx. */
 155 #define GETBINTERNAL(pg, indx)                                          \
 156         ((BINTERNAL *)((char *)(pg) + (pg)->linp[indx]))
 157
 158 /* Get the number of bytes in the entry. */
 159 #define NBINTERNAL(len)                                                 \
 160         LALIGN(sizeof(u_int32_t) + sizeof(pgno_t) + sizeof(u_char) + (len))
 161
 162 /* Copy a BINTERNAL entry to the page. */
 163 #define WR_BINTERNAL(p, size, pgno, flags) {                            \
 164         *(u_int32_t *)p = size;                                         \
 165         p += sizeof(u_int32_t);                                         \
 166         *(pgno_t *)p = pgno;                                            \
 167         p += sizeof(pgno_t);                                            \
 168         *(u_char *)p = flags;                                           \
 169         p += sizeof(u_char);                                            \
 170 }
 171
 172 /*
 173  * For the recno internal pages, the item is a page number with the number of
 174  * keys found on that page and below.
 175  */
 176 typedef struct _rinternal {
 177         recno_t nrecs;                  /* number of records */
 178         pgno_t  pgno;                   /* page number stored below */
 179 } RINTERNAL;
 180
 181 /* Get the page's RINTERNAL structure at index indx. */
 182 #define GETRINTERNAL(pg, indx)                                          \
 183         ((RINTERNAL *)((char *)(pg) + (pg)->linp[indx]))
 184
 185 /* Get the number of bytes in the entry. */
 186 #define NRINTERNAL                                                      \
 187         LALIGN(sizeof(recno_t) + sizeof(pgno_t))
 188
 189 /* Copy a RINTERAL entry to the page. */
 190 #define WR_RINTERNAL(p, nrecs, pgno) {                                  \
 191         *(recno_t *)p = nrecs;                                          \
 192         p += sizeof(recno_t);                                           \
 193         *(pgno_t *)p = pgno;                                            \
 194 }
 195
 196 /* For the btree leaf pages, the item is a key and data pair. */
 197 typedef struct _bleaf {
 198         u_int32_t       ksize;          /* size of key */
 199         u_int32_t       dsize;          /* size of data */
 200         u_char  flags;                  /* P_BIGDATA, P_BIGKEY */
 201         char    bytes[1];               /* data */
 202 } BLEAF;
 203
 204 /* Get the page's BLEAF structure at index indx. */
 205 #define GETBLEAF(pg, indx)                                              \
 206         ((BLEAF *)((char *)(pg) + (pg)->linp[indx]))
 207
 208 /* Get the number of bytes in the entry. */
 209 #define NBLEAF(p)       NBLEAFDBT((p)->ksize, (p)->dsize)
 210
 211 /* Get the number of bytes in the user's key/data pair. */
 212 #define NBLEAFDBT(ksize, dsize)                                         \
 213         LALIGN(sizeof(u_int32_t) + sizeof(u_int32_t) + sizeof(u_char) + \
 214             (ksize) + (dsize))
 215
 216 /* Copy a BLEAF entry to the page. */
 217 #define WR_BLEAF(p, key, data, flags) {                                 \
 218         *(u_int32_t *)p = key->size;                                    \
 219         p += sizeof(u_int32_t);                                         \
 220         *(u_int32_t *)p = data->size;                                   \
 221         p += sizeof(u_int32_t);                                         \
 222         *(u_char *)p = flags;                                           \
 223         p += sizeof(u_char);                                            \
 224         memmove(p, key->data, key->size);                               \
 225         p += key->size;                                                 \
 226         memmove(p, data->data, data->size);                             \
 227 }
 228
 229 /* For the recno leaf pages, the item is a data entry. */
 230 typedef struct _rleaf {
 231         u_int32_t       dsize;          /* size of data */
 232         u_char  flags;                  /* P_BIGDATA */
 233         char    bytes[1];
 234 } RLEAF;
 235
 236 /* Get the page's RLEAF structure at index indx. */
 237 #define GETRLEAF(pg, indx)                                              \
 238         ((RLEAF *)((char *)(pg) + (pg)->linp[indx]))
 239
 240 /* Get the number of bytes in the entry. */
 241 #define NRLEAF(p)       NRLEAFDBT((p)->dsize)
 242
 243 /* Get the number of bytes from the user's data. */
 244 #define NRLEAFDBT(dsize)                                                \
 245         LALIGN(sizeof(u_int32_t) + sizeof(u_char) + (dsize))
 246
 247 /* Copy a RLEAF entry to the page. */
 248 #define WR_RLEAF(p, data, flags) {                                      \
 249         *(u_int32_t *)p = data->size;                                   \
 250         p += sizeof(u_int32_t);                                         \
 251         *(u_char *)p = flags;                                           \
 252         p += sizeof(u_char);                                            \
 253         memmove(p, data->data, data->size);                             \
 254 }
 255
 256 /*
 257  * A record in the tree is either a pointer to a page and an index in the page
 258  * or a page number and an index.  These structures are used as a cursor, stack
 259  * entry and search returns as well as to pass records to other routines.
 260  *
 261  * One comment about searches.  Internal page searches must find the largest
 262  * record less than key in the tree so that descents work.  Leaf page searches
 263  * must find the smallest record greater than key so that the returned index
 264  * is the record's correct position for insertion.
 265  */
 266 typedef struct _epgno {
 267         pgno_t  pgno;                   /* the page number */
 268         indx_t  index;                  /* the index on the page */
 269 } EPGNO;
 270
 271 typedef struct _epg {
 272         PAGE    *page;                  /* the (pinned) page */
 273         indx_t   index;                 /* the index on the page */
 274 } EPG;
 275
 276 /*
 277  * About cursors.  The cursor (and the page that contained the key/data pair
 278  * that it referenced) can be deleted, which makes things a bit tricky.  If
 279  * there are no duplicates of the cursor key in the tree (i.e. B_NODUPS is set
 280  * or there simply aren't any duplicates of the key) we copy the key that it
 281  * referenced when it's deleted, and reacquire a new cursor key if the cursor
 282  * is used again.  If there are duplicates keys, we move to the next/previous
 283  * key, and set a flag so that we know what happened.  NOTE: if duplicate (to
 284  * the cursor) keys are added to the tree during this process, it is undefined
 285  * if they will be returned or not in a cursor scan.
 286  *
 287  * The flags determine the possible states of the cursor:
 288  *
 289  * CURS_INIT    The cursor references *something*.
 290  * CURS_ACQUIRE The cursor was deleted, and a key has been saved so that
 291  *              we can reacquire the right position in the tree.
 292  * CURS_AFTER, CURS_BEFORE
 293  *              The cursor was deleted, and now references a key/data pair
 294  *              that has not yet been returned, either before or after the
 295  *              deleted key/data pair.
 296  * XXX
 297  * This structure is broken out so that we can eventually offer multiple
 298  * cursors as part of the DB interface.
 299  */
 300 typedef struct _cursor {
 301         EPGNO    pg;                    /* B: Saved tree reference. */
 302         DBT      key;                   /* B: Saved key, or key.data == NULL. */
 303         recno_t  rcursor;               /* R: recno cursor (1-based) */
 304
 305 #define CURS_ACQUIRE    0x01            /*  B: Cursor needs to be reacquired. */
 306 #define CURS_AFTER      0x02            /*  B: Unreturned cursor after key. */
 307 #define CURS_BEFORE     0x04            /*  B: Unreturned cursor before key. */
 308 #define CURS_INIT       0x08            /* RB: Cursor initialized. */
 309         u_int8_t flags;
 310 } CURSOR;
 311
 312 /*
 313  * The metadata of the tree.  The nrecs field is used only by the RECNO code.
 314  * This is because the btree doesn't really need it and it requires that every
 315  * put or delete call modify the metadata.
 316  */
 317 typedef struct _btmeta {
 318         u_int32_t       magic;          /* magic number */
 319         u_int32_t       version;        /* version */
 320         u_int32_t       psize;          /* page size */
 321         u_int32_t       free;           /* page number of first free page */
 322         u_int32_t       nrecs;          /* R: number of records */
 323
 324 #define SAVEMETA        (B_NODUPS | R_RECNO)
 325         u_int32_t       flags;          /* bt_flags & SAVEMETA */
 326 } BTMETA;
 327
 328 /* The in-memory btree/recno data structure. */
 329 typedef struct _btree {
 330         MPOOL    *bt_mp;                /* memory pool cookie */
 331
 332         DB       *bt_dbp;               /* pointer to enclosing DB */
 333
 334         EPG       bt_cur;               /* current (pinned) page */
 335         PAGE     *bt_pinned;            /* page pinned across calls */
 336
 337         CURSOR    bt_cursor;            /* cursor */
 338
 339 #define BT_PUSH(t, p, i) {                                              \
 340         t->bt_sp->pgno = p;                                             \
 341         t->bt_sp->index = i;                                            \
 342         ++t->bt_sp;                                                     \
 343 }
 344 #define BT_POP(t)       (t->bt_sp == t->bt_stack ? NULL : --t->bt_sp)
 345 #define BT_CLR(t)       (t->bt_sp = t->bt_stack)
 346         EPGNO     bt_stack[50];         /* stack of parent pages */
 347         EPGNO    *bt_sp;                /* current stack pointer */
 348
 349         DBT       bt_rkey;              /* returned key */
 350         DBT       bt_rdata;             /* returned data */
 351
 352         int       bt_fd;                /* tree file descriptor */
 353
 354         pgno_t    bt_free;              /* next free page */
 355         u_int32_t bt_psize;             /* page size */
 356         indx_t    bt_ovflsize;          /* cut-off for key/data overflow */
 357         int       bt_lorder;            /* byte order */
 358                                         /* sorted order */
 359         enum { NOT, BACK, FORWARD } bt_order;
 360         EPGNO     bt_last;              /* last insert */
 361
 362                                         /* B: key comparison function */
 363         int     (*bt_cmp)(const DBT *, const DBT *);
 364                                         /* B: prefix comparison function */
 365         size_t  (*bt_pfx)(const DBT *, const DBT *);
 366                                         /* R: recno input function */
 367         int     (*bt_irec)(struct _btree *, recno_t);
 368
 369         FILE     *bt_rfp;               /* R: record FILE pointer */
 370         int       bt_rfd;               /* R: record file descriptor */
 371
 372         caddr_t   bt_cmap;              /* R: current point in mapped space */
 373         caddr_t   bt_smap;              /* R: start of mapped space */
 374         caddr_t   bt_emap;              /* R: end of mapped space */
 375         size_t    bt_msize;             /* R: size of mapped region. */
 376
 377         recno_t   bt_nrecs;             /* R: number of records */
 378         size_t    bt_reclen;            /* R: fixed record length */
 379         u_char    bt_bval;              /* R: delimiting byte/pad character */
 380
 381 /*
 382  * NB:
 383  * B_NODUPS and R_RECNO are stored on disk, and may not be changed.
 384  */
 385 #define B_INMEM         0x00001         /* in-memory tree */
 386 #define B_METADIRTY     0x00002         /* need to write metadata */
 387 #define B_MODIFIED      0x00004         /* tree modified */
 388 #define B_NEEDSWAP      0x00008         /* if byte order requires swapping */
 389 #define B_RDONLY        0x00010         /* read-only tree */
 390
 391 #define B_NODUPS        0x00020         /* no duplicate keys permitted */
 392 #define R_RECNO         0x00080         /* record oriented tree */
 393
 394 #define R_CLOSEFP       0x00040         /* opened a file pointer */
 395 #define R_EOF           0x00100         /* end of input file reached. */
 396 #define R_FIXLEN        0x00200         /* fixed length records */
 397 #define R_MEMMAPPED     0x00400         /* memory mapped file. */
 398 #define R_INMEM         0x00800         /* in-memory file */
 399 #define R_MODIFIED      0x01000         /* modified file */
 400 #define R_RDONLY        0x02000         /* read-only file */
 401
 402 #define B_DB_LOCK       0x04000         /* DB_LOCK specified. */
 403 #define B_DB_SHMEM      0x08000         /* DB_SHMEM specified. */
 404 #define B_DB_TXN        0x10000         /* DB_TXN specified. */
 405         u_int32_t flags;
 406 } BTREE;
 407
 408 #include "extern.h"