1 /*- 2 * Copyright (c) 1990, 1993 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 * %sccs.include.redist.c% 9 */ 10 11 #if defined(LIBC_SCCS) && !defined(lint) 12 static char sccsid[] = "@(#)bt_open.c 8.5 (Berkeley) 02/21/94"; 13 #endif /* LIBC_SCCS and not lint */ 14 15 /* 16 * Implementation of btree access method for 4.4BSD. 17 * 18 * The design here was originally based on that of the btree access method 19 * used in the Postgres database system at UC Berkeley. This implementation 20 * is wholly independent of the Postgres code. 21 */ 22 23 #include <sys/param.h> 24 #include <sys/stat.h> 25 26 #include <errno.h> 27 #include <fcntl.h> 28 #include <limits.h> 29 #include <signal.h> 30 #include <stdio.h> 31 #include <stdlib.h> 32 #include <string.h> 33 #include <unistd.h> 34 35 #include <db.h> 36 #include "btree.h" 37 38 static int byteorder __P((void)); 39 static int nroot __P((BTREE *)); 40 static int tmp __P((void)); 41 42 /* 43 * __BT_OPEN -- Open a btree. 44 * 45 * Creates and fills a DB struct, and calls the routine that actually 46 * opens the btree. 47 * 48 * Parameters: 49 * fname: filename (NULL for in-memory trees) 50 * flags: open flag bits 51 * mode: open permission bits 52 * b: BTREEINFO pointer 53 * 54 * Returns: 55 * NULL on failure, pointer to DB on success. 56 * 57 */ 58 DB * 59 __bt_open(fname, flags, mode, openinfo, dflags) 60 const char *fname; 61 int flags, mode, dflags; 62 const BTREEINFO *openinfo; 63 { 64 struct stat sb; 65 BTMETA m; 66 BTREE *t; 67 BTREEINFO b; 68 DB *dbp; 69 pgno_t ncache; 70 ssize_t nr; 71 int machine_lorder; 72 73 t = NULL; 74 75 /* 76 * Intention is to make sure all of the user's selections are okay 77 * here and then use them without checking. Can't be complete, since 78 * we don't know the right page size, lorder or flags until the backing 79 * file is opened. Also, the file's page size can cause the cachesize 80 * to change. 81 */ 82 machine_lorder = byteorder(); 83 if (openinfo) { 84 b = *openinfo; 85 86 /* Flags: R_DUP. */ 87 if (b.flags & ~(R_DUP)) 88 goto einval; 89 90 /* 91 * Page size must be indx_t aligned and >= MINPSIZE. Default 92 * page size is set farther on, based on the underlying file 93 * transfer size. 94 */ 95 if (b.psize && 96 (b.psize < MINPSIZE || b.psize > MAX_PAGE_OFFSET + 1 || 97 b.psize & sizeof(indx_t) - 1)) 98 goto einval; 99 100 /* Minimum number of keys per page; absolute minimum is 2. */ 101 if (b.minkeypage) { 102 if (b.minkeypage < 2) 103 goto einval; 104 } else 105 b.minkeypage = DEFMINKEYPAGE; 106 107 /* If no comparison, use default comparison and prefix. */ 108 if (b.compare == NULL) { 109 b.compare = __bt_defcmp; 110 if (b.prefix == NULL) 111 b.prefix = __bt_defpfx; 112 } 113 114 if (b.lorder == 0) 115 b.lorder = machine_lorder; 116 } else { 117 b.compare = __bt_defcmp; 118 b.cachesize = 0; 119 b.flags = 0; 120 b.lorder = machine_lorder; 121 b.minkeypage = DEFMINKEYPAGE; 122 b.prefix = __bt_defpfx; 123 b.psize = 0; 124 } 125 126 /* Check for the ubiquitous PDP-11. */ 127 if (b.lorder != BIG_ENDIAN && b.lorder != LITTLE_ENDIAN) 128 goto einval; 129 130 /* Allocate and initialize DB and BTREE structures. */ 131 if ((t = (BTREE *)malloc(sizeof(BTREE))) == NULL) 132 goto err; 133 memset(t, 0, sizeof(BTREE)); 134 t->bt_bcursor.pgno = P_INVALID; 135 t->bt_fd = -1; /* Don't close unopened fd on error. */ 136 t->bt_lorder = b.lorder; 137 t->bt_order = NOT; 138 t->bt_cmp = b.compare; 139 t->bt_pfx = b.prefix; 140 t->bt_rfd = -1; 141 142 if ((t->bt_dbp = dbp = (DB *)malloc(sizeof(DB))) == NULL) 143 goto err; 144 t->bt_flags = 0; 145 if (t->bt_lorder != machine_lorder) 146 SET(t, B_NEEDSWAP); 147 148 dbp->type = DB_BTREE; 149 dbp->internal = t; 150 dbp->close = __bt_close; 151 dbp->del = __bt_delete; 152 dbp->fd = __bt_fd; 153 dbp->get = __bt_get; 154 dbp->put = __bt_put; 155 dbp->seq = __bt_seq; 156 dbp->sync = __bt_sync; 157 158 /* 159 * If no file name was supplied, this is an in-memory btree and we 160 * open a backing temporary file. Otherwise, it's a disk-based tree. 161 */ 162 if (fname) { 163 switch(flags & O_ACCMODE) { 164 case O_RDONLY: 165 SET(t, B_RDONLY); 166 break; 167 case O_RDWR: 168 break; 169 case O_WRONLY: 170 default: 171 goto einval; 172 } 173 174 if ((t->bt_fd = open(fname, flags, mode)) < 0) 175 goto err; 176 177 } else { 178 if ((flags & O_ACCMODE) != O_RDWR) 179 goto einval; 180 if ((t->bt_fd = tmp()) == -1) 181 goto err; 182 SET(t, B_INMEM); 183 } 184 185 if (fcntl(t->bt_fd, F_SETFD, 1) == -1) 186 goto err; 187 188 if (fstat(t->bt_fd, &sb)) 189 goto err; 190 if (sb.st_size) { 191 nr = read(t->bt_fd, &m, sizeof(BTMETA)); 192 if (nr < 0) 193 goto err; 194 if (nr != sizeof(BTMETA)) 195 goto eftype; 196 197 /* 198 * Read in the meta-data. This can change the notion of what 199 * the lorder, page size and flags are, and, when the page size 200 * changes, the cachesize value can change too. If the user 201 * specified the wrong byte order for an existing database, we 202 * don't bother to return an error, we just clear the NEEDSWAP 203 * bit. 204 */ 205 if (m.m_magic == BTREEMAGIC) 206 CLR(t, B_NEEDSWAP); 207 else { 208 SET(t, B_NEEDSWAP); 209 M_32_SWAP(m.m_magic); 210 M_32_SWAP(m.m_version); 211 M_32_SWAP(m.m_psize); 212 M_32_SWAP(m.m_free); 213 M_32_SWAP(m.m_nrecs); 214 M_32_SWAP(m.m_flags); 215 } 216 if (m.m_magic != BTREEMAGIC || m.m_version != BTREEVERSION) 217 goto eftype; 218 if (m.m_psize < MINPSIZE || m.m_psize > MAX_PAGE_OFFSET + 1 || 219 m.m_psize & sizeof(indx_t) - 1) 220 goto eftype; 221 if (m.m_flags & ~SAVEMETA) 222 goto eftype; 223 b.psize = m.m_psize; 224 t->bt_flags |= m.m_flags; 225 t->bt_free = m.m_free; 226 t->bt_nrecs = m.m_nrecs; 227 } else { 228 /* 229 * Set the page size to the best value for I/O to this file. 230 * Don't overflow the page offset type. 231 */ 232 if (b.psize == 0) { 233 b.psize = sb.st_blksize; 234 if (b.psize < MINPSIZE) 235 b.psize = MINPSIZE; 236 if (b.psize > MAX_PAGE_OFFSET + 1) 237 b.psize = MAX_PAGE_OFFSET + 1; 238 } 239 240 /* Set flag if duplicates permitted. */ 241 if (!(b.flags & R_DUP)) 242 SET(t, B_NODUPS); 243 244 t->bt_free = P_INVALID; 245 t->bt_nrecs = 0; 246 SET(t, B_METADIRTY); 247 } 248 249 t->bt_psize = b.psize; 250 251 /* Set the cache size; must be a multiple of the page size. */ 252 if (b.cachesize && b.cachesize & b.psize - 1) 253 b.cachesize += (~b.cachesize & b.psize - 1) + 1; 254 if (b.cachesize < b.psize * MINCACHE) 255 b.cachesize = b.psize * MINCACHE; 256 257 /* Calculate number of pages to cache. */ 258 ncache = (b.cachesize + t->bt_psize - 1) / t->bt_psize; 259 260 /* 261 * The btree data structure requires that at least two keys can fit on 262 * a page, but other than that there's no fixed requirement. The user 263 * specified a minimum number per page, and we translated that into the 264 * number of bytes a key/data pair can use before being placed on an 265 * overflow page. This calculation includes the page header, the size 266 * of the index referencing the leaf item and the size of the leaf item 267 * structure. Also, don't let the user specify a minkeypage such that 268 * a key/data pair won't fit even if both key and data are on overflow 269 * pages. 270 */ 271 t->bt_ovflsize = (t->bt_psize - BTDATAOFF) / b.minkeypage - 272 (sizeof(indx_t) + NBLEAFDBT(0, 0)); 273 if (t->bt_ovflsize < NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(indx_t)) 274 t->bt_ovflsize = 275 NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(indx_t); 276 277 /* Initialize the buffer pool. */ 278 if ((t->bt_mp = 279 mpool_open(NULL, t->bt_fd, t->bt_psize, ncache)) == NULL) 280 goto err; 281 if (!ISSET(t, B_INMEM)) 282 mpool_filter(t->bt_mp, __bt_pgin, __bt_pgout, t); 283 284 /* Create a root page if new tree. */ 285 if (nroot(t) == RET_ERROR) 286 goto err; 287 288 /* Global flags. */ 289 if (dflags & DB_LOCK) 290 SET(t, B_DB_LOCK); 291 if (dflags & DB_SHMEM) 292 SET(t, B_DB_SHMEM); 293 if (dflags & DB_TXN) 294 SET(t, B_DB_TXN); 295 296 return (dbp); 297 298 einval: errno = EINVAL; 299 goto err; 300 301 eftype: errno = EFTYPE; 302 goto err; 303 304 err: if (t) { 305 if (t->bt_dbp) 306 free(t->bt_dbp); 307 if (t->bt_fd != -1) 308 (void)close(t->bt_fd); 309 free(t); 310 } 311 return (NULL); 312 } 313 314 /* 315 * NROOT -- Create the root of a new tree. 316 * 317 * Parameters: 318 * t: tree 319 * 320 * Returns: 321 * RET_ERROR, RET_SUCCESS 322 */ 323 static int 324 nroot(t) 325 BTREE *t; 326 { 327 PAGE *meta, *root; 328 pgno_t npg; 329 330 if ((meta = mpool_get(t->bt_mp, 0, 0)) != NULL) { 331 mpool_put(t->bt_mp, meta, 0); 332 return (RET_SUCCESS); 333 } 334 if (errno != EINVAL) 335 return (RET_ERROR); 336 337 if ((meta = mpool_new(t->bt_mp, &npg)) == NULL) 338 return (RET_ERROR); 339 340 if ((root = mpool_new(t->bt_mp, &npg)) == NULL) 341 return (RET_ERROR); 342 343 if (npg != P_ROOT) 344 return (RET_ERROR); 345 root->pgno = npg; 346 root->prevpg = root->nextpg = P_INVALID; 347 root->lower = BTDATAOFF; 348 root->upper = t->bt_psize; 349 root->flags = P_BLEAF; 350 memset(meta, 0, t->bt_psize); 351 mpool_put(t->bt_mp, meta, MPOOL_DIRTY); 352 mpool_put(t->bt_mp, root, MPOOL_DIRTY); 353 return (RET_SUCCESS); 354 } 355 356 static int 357 tmp() 358 { 359 sigset_t set, oset; 360 int fd; 361 char *envtmp; 362 char path[MAXPATHLEN]; 363 364 envtmp = getenv("TMPDIR"); 365 (void)snprintf(path, 366 sizeof(path), "%s/bt.XXXXXX", envtmp ? envtmp : "/tmp"); 367 368 (void)sigfillset(&set); 369 (void)sigprocmask(SIG_BLOCK, &set, &oset); 370 if ((fd = mkstemp(path)) != -1) 371 (void)unlink(path); 372 (void)sigprocmask(SIG_SETMASK, &oset, NULL); 373 return(fd); 374 } 375 376 static int 377 byteorder() 378 { 379 u_int32_t x; 380 u_char *p; 381 382 x = 0x01020304; 383 p = (u_char *)&x; 384 switch (*p) { 385 case 1: 386 return (BIG_ENDIAN); 387 case 4: 388 return (LITTLE_ENDIAN); 389 default: 390 return (0); 391 } 392 } 393 394 int 395 __bt_fd(dbp) 396 const DB *dbp; 397 { 398 BTREE *t; 399 400 t = dbp->internal; 401 402 /* Toss any page pinned across calls. */ 403 if (t->bt_pinned != NULL) { 404 mpool_put(t->bt_mp, t->bt_pinned, 0); 405 t->bt_pinned = NULL; 406 } 407 408 /* In-memory database can't have a file descriptor. */ 409 if (ISSET(t, B_INMEM)) { 410 errno = ENOENT; 411 return (-1); 412 } 413 return (t->bt_fd); 414 } 415