DB *
__rec_open(const char *fname, int flags, mode_t mode, const RECNOINFO *openinfo,
int dflags)
{
BTREE *t;
BTREEINFO btopeninfo;
DB *dbp;
PAGE *h;
struct stat sb;
int rfd = -1; /* pacify gcc */
int sverrno;
dbp = NULL;
/* Open the user's file -- if this fails, we're done. */
if (fname != NULL) {
if ((rfd = __dbopen(fname, flags, mode, NULL)) == -1)
return NULL;
}
/* Create a btree in memory (backed by disk). */
if (openinfo) {
if (openinfo->flags & ~(R_FIXEDLEN | R_NOKEY | R_SNAPSHOT))
goto einval;
btopeninfo.flags = 0;
btopeninfo.cachesize = openinfo->cachesize;
btopeninfo.maxkeypage = 0;
btopeninfo.minkeypage = 0;
btopeninfo.psize = openinfo->psize;
btopeninfo.compare = NULL;
btopeninfo.prefix = NULL;
btopeninfo.lorder = openinfo->lorder;
dbp = __bt_open(openinfo->bfname,
O_RDWR, S_IRUSR | S_IWUSR, &btopeninfo, dflags);
} else
dbp = __bt_open(NULL, O_RDWR, S_IRUSR | S_IWUSR, NULL, dflags);
if (dbp == NULL)
goto err;
/*
* Some fields in the tree structure are recno specific. Fill them
* in and make the btree structure look like a recno structure. We
* don't change the bt_ovflsize value, it's close enough and slightly
* bigger.
*/
t = dbp->internal;
if (openinfo) {
if (openinfo->flags & R_FIXEDLEN) {
F_SET(t, R_FIXLEN);
t->bt_reclen = openinfo->reclen;
if (t->bt_reclen == 0)
goto einval;
}
t->bt_bval = openinfo->bval;
} else
t->bt_bval = '\n';
F_SET(t, R_RECNO);
if (fname == NULL)
F_SET(t, R_EOF | R_INMEM);
else
t->bt_rfd = rfd;
if (fname != NULL) {
/*
* In 4.4BSD, stat(2) returns true for ISSOCK on pipes.
* Unfortunately, that's not portable, so we use lseek
* and check the errno values.
*/
errno = 0;
if (lseek(rfd, (off_t)0, SEEK_CUR) == -1 && errno == ESPIPE) {
switch (flags & O_ACCMODE) {
case O_RDONLY:
F_SET(t, R_RDONLY);
break;
default:
goto einval;
}
slow: if ((t->bt_rfp = fdopen(rfd, "r")) == NULL)
goto err;
F_SET(t, R_CLOSEFP);
t->bt_irec =
F_ISSET(t, R_FIXLEN) ? __rec_fpipe : __rec_vpipe;
} else {
switch (flags & O_ACCMODE) {
case O_RDONLY:
F_SET(t, R_RDONLY);
break;
case O_RDWR:
break;
default:
goto einval;
}
if (fstat(rfd, &sb))
goto err;
/*
* Kluge -- we'd like to test to see if the file is too
* big to mmap. Since, we don't know what size or type
* off_t's or size_t's are, what the largest unsigned
* integral type is, or what random insanity the local
* C compiler will perpetrate, doing the comparison in
* a portable way is flatly impossible. Hope that mmap
* fails if the file is too large.
*/
if (sb.st_size == 0)
F_SET(t, R_EOF);
else {
#ifdef MMAP_NOT_AVAILABLE
/*
* XXX
* Mmap doesn't work correctly on many current
* systems. In particular, it can fail subtly,
* with cache coherency problems. Don't use it
* for now.
*/
t->bt_msize = sb.st_size;
if ((t->bt_smap = mmap(NULL, t->bt_msize,
PROT_READ, MAP_FILE | MAP_PRIVATE, rfd,
(off_t)0)) == (caddr_t)-1)
goto slow;
t->bt_cmap = t->bt_smap;
t->bt_emap = t->bt_smap + sb.st_size;
t->bt_irec = F_ISSET(t, R_FIXLEN) ?
__rec_fmap : __rec_vmap;
F_SET(t, R_MEMMAPPED);
#else
goto slow;
#endif
}
}
}
/* Use the recno routines. */
dbp->close = __rec_close;
dbp->del = __rec_delete;
dbp->fd = __rec_fd;
dbp->get = __rec_get;
dbp->put = __rec_put;
dbp->seq = __rec_seq;
dbp->sync = __rec_sync;
/* If the root page was created, reset the flags. */
if ((h = mpool_get(t->bt_mp, P_ROOT, 0)) == NULL)
goto err;
if ((h->flags & P_TYPE) == P_BLEAF) {
F_CLR(h, P_TYPE);
F_SET(h, P_RLEAF);
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
} else
mpool_put(t->bt_mp, h, 0);
if (openinfo && openinfo->flags & R_SNAPSHOT &&
!F_ISSET(t, R_EOF | R_INMEM) &&
t->bt_irec(t, MAX_REC_NUMBER) == RET_ERROR)
goto err;
return (dbp);
einval: errno = EINVAL;
err: sverrno = errno;
if (dbp != NULL)
(void)__bt_close(dbp);
if (fname != NULL)
(void)close(rfd);
errno = sverrno;
return (NULL);
}
/*
* __BT_OPEN -- Open a btree.
*
* Creates and fills a DB struct, and calls the routine that actually
* opens the btree.
*
* Parameters:
* fname: filename (NULL for in-memory trees)
* flags: open flag bits
* mode: open permission bits
* b: BTREEINFO pointer
*
* Returns:
* NULL on failure, pointer to DB on success.
*
*/
DB *
__bt_open(const char *fname, int flags, mode_t mode, const BTREEINFO *openinfo,
int dflags)
{
struct stat sb;
BTMETA m;
BTREE *t;
BTREEINFO b;
DB *dbp;
pgno_t ncache;
ssize_t nr;
size_t temp;
int machine_lorder;
t = NULL;
/*
* Intention is to make sure all of the user's selections are okay
* here and then use them without checking. Can't be complete, since
* we don't know the right page size, lorder or flags until the backing
* file is opened. Also, the file's page size can cause the cachesize
* to change.
*/
machine_lorder = byteorder();
if (openinfo) {
b = *openinfo;
/* Flags: R_DUP. */
if (b.flags & ~(R_DUP))
goto einval;
/*
* Page size must be indx_t aligned and >= MINPSIZE. Default
* page size is set farther on, based on the underlying file
* transfer size.
*/
if (b.psize &&
(b.psize < MINPSIZE || b.psize > MAX_PAGE_OFFSET + 1 ||
b.psize & (sizeof(indx_t) - 1)))
goto einval;
/* Minimum number of keys per page; absolute minimum is 2. */
if (b.minkeypage) {
if (b.minkeypage < 2)
goto einval;
} else
b.minkeypage = DEFMINKEYPAGE;
/* If no comparison, use default comparison and prefix. */
if (b.compare == NULL) {
b.compare = __bt_defcmp;
if (b.prefix == NULL)
b.prefix = __bt_defpfx;
}
if (b.lorder == 0)
b.lorder = machine_lorder;
} else {
b.compare = __bt_defcmp;
b.cachesize = 0;
b.flags = 0;
b.lorder = machine_lorder;
b.minkeypage = DEFMINKEYPAGE;
b.prefix = __bt_defpfx;
b.psize = 0;
}
/* Check for the ubiquitous PDP-11. */
if (b.lorder != BIG_ENDIAN && b.lorder != LITTLE_ENDIAN)
goto einval;
/* Allocate and initialize DB and BTREE structures. */
if ((t = malloc(sizeof(*t))) == NULL)
goto err;
memset(t, 0, sizeof(BTREE));
t->bt_fd = -1; /* Don't close unopened fd on error. */
t->bt_lorder = b.lorder;
t->bt_order = NOT;
t->bt_cmp = b.compare;
t->bt_pfx = b.prefix;
t->bt_rfd = -1;
if ((t->bt_dbp = dbp = malloc(sizeof(*dbp))) == NULL)
goto err;
memset(t->bt_dbp, 0, sizeof(DB));
if (t->bt_lorder != machine_lorder)
F_SET(t, B_NEEDSWAP);
dbp->type = DB_BTREE;
dbp->internal = t;
dbp->close = __bt_close;
dbp->del = __bt_delete;
dbp->fd = __bt_fd;
dbp->get = __bt_get;
dbp->put = __bt_put;
dbp->seq = __bt_seq;
dbp->sync = __bt_sync;
/*
* If no file name was supplied, this is an in-memory btree and we
* open a backing temporary file. Otherwise, it's a disk-based tree.
*/
if (fname) {
switch (flags & O_ACCMODE) {
case O_RDONLY:
F_SET(t, B_RDONLY);
break;
case O_RDWR:
break;
case O_WRONLY:
default:
goto einval;
}
if ((t->bt_fd = __dbopen(fname, flags, mode, &sb)) == -1)
goto err;
} else {
if ((flags & O_ACCMODE) != O_RDWR)
goto einval;
if ((t->bt_fd = __dbtemp("bt.", &sb)) == -1)
goto err;
F_SET(t, B_INMEM);
}
if (sb.st_size) {
if ((nr = read(t->bt_fd, &m, sizeof(BTMETA))) < 0)
goto err;
if (nr != sizeof(BTMETA))
goto eftype;
/*
* Read in the meta-data. This can change the notion of what
* the lorder, page size and flags are, and, when the page size
* changes, the cachesize value can change too. If the user
* specified the wrong byte order for an existing database, we
* don't bother to return an error, we just clear the NEEDSWAP
* bit.
*/
if (m.magic == BTREEMAGIC)
F_CLR(t, B_NEEDSWAP);
else {
F_SET(t, B_NEEDSWAP);
M_32_SWAP(m.magic);
M_32_SWAP(m.version);
M_32_SWAP(m.psize);
M_32_SWAP(m.free);
M_32_SWAP(m.nrecs);
M_32_SWAP(m.flags);
}
if (m.magic != BTREEMAGIC || m.version != BTREEVERSION)
goto eftype;
if (m.psize < MINPSIZE || m.psize > MAX_PAGE_OFFSET + 1 ||
m.psize & (sizeof(indx_t) - 1))
goto eftype;
if (m.flags & ~SAVEMETA)
goto eftype;
b.psize = m.psize;
F_SET(t, m.flags);
t->bt_free = m.free;
t->bt_nrecs = m.nrecs;
} else {
/*
* Set the page size to the best value for I/O to this file.
* Don't overflow the page offset type.
*/
if (b.psize == 0) {
b.psize = sb.st_blksize;
if (b.psize < MINPSIZE)
b.psize = MINPSIZE;
if (b.psize > MAX_PAGE_OFFSET + 1)
b.psize = MAX_PAGE_OFFSET + 1;
}
/* Set flag if duplicates permitted. */
if (!(b.flags & R_DUP))
F_SET(t, B_NODUPS);
t->bt_free = P_INVALID;
t->bt_nrecs = 0;
F_SET(t, B_METADIRTY);
}
t->bt_psize = b.psize;
/* Set the cache size; must be a multiple of the page size. */
if (b.cachesize && b.cachesize & (b.psize - 1))
b.cachesize += (~b.cachesize & (b.psize - 1)) + 1;
if (b.cachesize < b.psize * MINCACHE)
b.cachesize = b.psize * MINCACHE;
/* Calculate number of pages to cache. */
ncache = (b.cachesize + t->bt_psize - 1) / t->bt_psize;
/*
* The btree data structure requires that at least two keys can fit on
* a page, but other than that there's no fixed requirement. The user
* specified a minimum number per page, and we translated that into the
* number of bytes a key/data pair can use before being placed on an
* overflow page. This calculation includes the page header, the size
* of the index referencing the leaf item and the size of the leaf item
* structure. Also, don't let the user specify a minkeypage such that
* a key/data pair won't fit even if both key and data are on overflow
* pages.
*/
temp = (t->bt_psize - BTDATAOFF) / b.minkeypage -
(sizeof(indx_t) + NBLEAFDBT(0, 0));
_DBFIT(temp, indx_t);
t->bt_ovflsize = (indx_t)temp;
if (t->bt_ovflsize < NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(indx_t)) {
size_t l = NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(indx_t);
_DBFIT(l, indx_t);
t->bt_ovflsize = (indx_t)l;
}
/* Initialize the buffer pool. */
if ((t->bt_mp =
mpool_open(NULL, t->bt_fd, t->bt_psize, ncache)) == NULL)
goto err;
if (!F_ISSET(t, B_INMEM))
mpool_filter(t->bt_mp, __bt_pgin, __bt_pgout, t);
/* Create a root page if new tree. */
if (nroot(t) == RET_ERROR)
goto err;
/* Global flags. */
if (dflags & DB_LOCK)
F_SET(t, B_DB_LOCK);
if (dflags & DB_SHMEM)
F_SET(t, B_DB_SHMEM);
if (dflags & DB_TXN)
F_SET(t, B_DB_TXN);
return (dbp);
einval: errno = EINVAL;
goto err;
eftype: errno = EFTYPE;
goto err;
err: if (t) {
if (t->bt_dbp)
free(t->bt_dbp);
if (t->bt_fd != -1)
(void)close(t->bt_fd);
free(t);
}
return (NULL);
}
