/*
* BT_RROOT -- Fix up the recno root page after it has been split.
*
* Parameters:
* t: tree
* h: root page
* l: left page
* r: right page
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
static int
bt_rroot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
{
char *dest;
uint32_t sz;
size_t temp;
temp = t->bt_psize - NRINTERNAL;
_DBFIT(temp, uint32_t);
sz = (uint32_t)temp;
/* Insert the left and right keys, set the header information. */
_DBFIT(sz, indx_t);
h->linp[0] = h->upper = (indx_t)sz;
dest = (char *)(void *)h + h->upper;
WR_RINTERNAL(dest,
l->flags & P_RLEAF ? NEXTINDEX(l) : rec_total(l), l->pgno);
h->linp[1] = h->upper -= NRINTERNAL;
dest = (char *)(void *)h + h->upper;
WR_RINTERNAL(dest,
r->flags & P_RLEAF ? NEXTINDEX(r) : rec_total(r), r->pgno);
h->lower = BTDATAOFF + 2 * sizeof(indx_t);
/* Unpin the root page, set to recno internal page. */
h->flags &= ~P_TYPE;
h->flags |= P_RINTERNAL;
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
/*
* __OVFL_PUT -- Store an overflow key/data item.
*
* Parameters:
* t: tree
* data: DBT to store
* pgno: storage page number
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
int
__ovfl_put(BTREE *t, const DBT *dbt, pgno_t *pg)
{
PAGE *h, *last;
void *p;
pgno_t npg;
uint32_t sz, nb, plen;
size_t temp;
/*
* Allocate pages and copy the key/data record into them. Store the
* number of the first page in the chain.
*/
temp = t->bt_psize - BTDATAOFF;
_DBFIT(temp, uint32_t);
plen = (uint32_t)temp;
last = NULL;
p = dbt->data;
temp = dbt->size;
_DBFIT(temp, uint32_t);
sz = temp;
for (;; p = (char *)p + plen, last = h) {
if ((h = __bt_new(t, &npg)) == NULL)
return (RET_ERROR);
h->pgno = npg;
h->nextpg = h->prevpg = P_INVALID;
h->flags = P_OVERFLOW;
h->lower = h->upper = 0;
nb = MIN(sz, plen);
(void)memmove((char *)(void *)h + BTDATAOFF, p, (size_t)nb);
if (last) {
last->nextpg = h->pgno;
mpool_put(t->bt_mp, last, MPOOL_DIRTY);
} else
*pg = h->pgno;
if ((sz -= nb) == 0) {
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
break;
}
}
return (RET_SUCCESS);
}
/*
* __OVFL_GET -- Get an overflow key/data item.
*
* Parameters:
* t: tree
* p: pointer to { pgno_t, uint32_t }
* buf: storage address
* bufsz: storage size
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
int
__ovfl_get(BTREE *t, void *p, size_t *ssz, void **buf, size_t *bufsz)
{
PAGE *h;
pgno_t pg;
uint32_t sz, nb, plen;
size_t temp;
memmove(&pg, p, sizeof(pg));
memmove(&sz, (char *)p + sizeof(pgno_t), sizeof(uint32_t));
*ssz = sz;
#ifdef DEBUG
if (pg == P_INVALID || sz == 0)
abort();
#endif
/* Make the buffer bigger as necessary. */
if (*bufsz < sz) {
void *nbuf = realloc(*buf, sz);
if (nbuf == NULL)
return (RET_ERROR);
*buf = nbuf;
*bufsz = sz;
}
/*
* Step through the linked list of pages, copying the data on each one
* into the buffer. Never copy more than the data's length.
*/
temp = t->bt_psize - BTDATAOFF;
_DBFIT(temp, uint32_t);
plen = (uint32_t)temp;
for (p = *buf;; p = (char *)p + nb, pg = h->nextpg) {
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
nb = MIN(sz, plen);
memmove(p, (char *)(void *)h + BTDATAOFF, nb);
mpool_put(t->bt_mp, h, 0);
if ((sz -= nb) == 0)
break;
}
return (RET_SUCCESS);
}
/*
* This calls alloc_segs which may run out of memory. Alloc_segs will destroy
* the table and set errno, so we just pass the error information along.
*
* Returns 0 on No Error
*/
static int
init_htab(HTAB *hashp, size_t nelem)
{
int nbuckets;
uint32_t nsegs;
int l2;
/*
* Divide number of elements by the fill factor and determine a
* desired number of buckets. Allocate space for the next greater
* power of two number of buckets.
*/
nelem = (nelem - 1) / hashp->FFACTOR + 1;
_DBFIT(nelem, uint32_t);
l2 = __log2(MAX((uint32_t)nelem, 2));
nbuckets = 1 << l2;
hashp->SPARES[l2] = l2 + 1;
hashp->SPARES[l2 + 1] = l2 + 1;
hashp->OVFL_POINT = l2;
hashp->LAST_FREED = 2;
/* First bitmap page is at: splitpoint l2 page offset 1 */
if (__ibitmap(hashp, (int)OADDR_OF(l2, 1), l2 + 1, 0))
return (-1);
hashp->MAX_BUCKET = hashp->LOW_MASK = nbuckets - 1;
hashp->HIGH_MASK = (nbuckets << 1) - 1;
/* LINTED constant in conditional context */
hashp->HDRPAGES = ((MAX(sizeof(HASHHDR), MINHDRSIZE) - 1) >>
hashp->BSHIFT) + 1;
nsegs = (nbuckets - 1) / hashp->SGSIZE + 1;
nsegs = 1 << __log2(nsegs);
if (nsegs > hashp->DSIZE)
hashp->DSIZE = nsegs;
return (alloc_segs(hashp, (int)nsegs));
}
/*
* __OVFL_DELETE -- Delete an overflow chain.
*
* Parameters:
* t: tree
* p: pointer to { pgno_t, uint32_t }
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
int
__ovfl_delete(BTREE *t, void *p)
{
PAGE *h;
pgno_t pg;
uint32_t sz, plen;
size_t temp;
(void)memmove(&pg, p, sizeof(pgno_t));
(void)memmove(&sz, (char *)p + sizeof(pgno_t), sizeof(uint32_t));
#ifdef DEBUG
if (pg == P_INVALID || sz == 0)
abort();
#endif
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
/* Don't delete chains used by internal pages. */
if (h->flags & P_PRESERVE) {
mpool_put(t->bt_mp, h, 0);
return (RET_SUCCESS);
}
/* Step through the chain, calling the free routine for each page. */
temp = t->bt_psize - BTDATAOFF;
_DBFIT(temp, uint32_t);
plen = (uint32_t)temp;
for (;; sz -= plen) {
pg = h->nextpg;
__bt_free(t, h);
if (sz <= plen)
break;
if ((h = mpool_get(t->bt_mp, pg, 0)) == NULL)
return (RET_ERROR);
}
return (RET_SUCCESS);
}
/*
* __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 = (BTREE *)malloc(sizeof(BTREE))) == 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 = (DB *)malloc(sizeof(DB))) == 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 = open(fname, flags, mode)) == -1)
goto err;
if (fcntl(t->bt_fd, F_SETFD, FD_CLOEXEC) == -1)
goto err;
} else {
if ((flags & O_ACCMODE) != O_RDWR)
goto einval;
if ((t->bt_fd = tmp()) == -1)
goto err;
F_SET(t, B_INMEM);
}
if (fcntl(t->bt_fd, F_SETFD, FD_CLOEXEC) == -1)
goto err;
if (fstat(t->bt_fd, &sb))
goto err;
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))
t->bt_ovflsize =
NBLEAFDBT(NOVFLSIZE, NOVFLSIZE) + sizeof(indx_t);
/* 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);
}
/*
* __REC_IPUT -- Add a recno item to the tree.
*
* Parameters:
* t: tree
* nrec: record number
* data: data
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
int
__rec_iput(BTREE *t, recno_t nrec, const DBT *data, u_int flags)
{
DBT tdata;
EPG *e;
PAGE *h;
indx_t idx, nxtindex;
pgno_t pg;
uint32_t nbytes;
int dflags, status;
char *dest, db[NOVFLSIZE];
/*
* If the data won't fit on a page, store it on indirect pages.
*
* XXX
* If the insert fails later on, these pages aren't recovered.
*/
if (data->size > t->bt_ovflsize) {
if (__ovfl_put(t, data, &pg) == RET_ERROR)
return (RET_ERROR);
tdata.data = db;
tdata.size = NOVFLSIZE;
*(pgno_t *)(void *)db = pg;
_DBFIT(data->size, uint32_t);
*(uint32_t *)(void *)(db + sizeof(pgno_t)) =
(uint32_t)data->size;
dflags = P_BIGDATA;
data = &tdata;
} else
dflags = 0;
/* __rec_search pins the returned page. */
if ((e = __rec_search(t, nrec,
nrec > t->bt_nrecs || flags == R_IAFTER || flags == R_IBEFORE ?
SINSERT : SEARCH)) == NULL)
return (RET_ERROR);
h = e->page;
idx = e->index;
/*
* Add the specified key/data pair to the tree. The R_IAFTER and
* R_IBEFORE flags insert the key after/before the specified key.
*
* Pages are split as required.
*/
switch (flags) {
case R_IAFTER:
++idx;
break;
case R_IBEFORE:
break;
default:
if (nrec < t->bt_nrecs &&
__rec_dleaf(t, h, (uint32_t)idx) == RET_ERROR) {
mpool_put(t->bt_mp, h, 0);
return (RET_ERROR);
}
break;
}
/*
* If not enough room, split the page. The split code will insert
* the key and data and unpin the current page. If inserting into
* the offset array, shift the pointers up.
*/
nbytes = NRLEAFDBT(data->size);
if ((uint32_t) (h->upper - h->lower) < nbytes + sizeof(indx_t)) {
status = __bt_split(t, h, NULL, data, dflags, nbytes,
(uint32_t)idx);
if (status == RET_SUCCESS)
++t->bt_nrecs;
return (status);
}
if (idx < (nxtindex = NEXTINDEX(h)))
memmove(h->linp + idx + 1, h->linp + idx,
(nxtindex - idx) * sizeof(indx_t));
h->lower += sizeof(indx_t);
h->linp[idx] = h->upper -= nbytes;
dest = (char *)(void *)h + h->upper;
WR_RLEAF(dest, data, dflags);
++t->bt_nrecs;
F_SET(t, B_MODIFIED);
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
/*
* Big_insert
*
* You need to do an insert and the key/data pair is too big
*
* Returns:
* 0 ==> OK
*-1 ==> ERROR
*/
int
__big_insert(HTAB *hashp, BUFHEAD *bufp, const DBT *key, const DBT *val)
{
uint16_t *p, n;
size_t key_size, val_size;
uint16_t space, move_bytes, off;
char *cp, *key_data, *val_data;
size_t temp;
cp = bufp->page; /* Character pointer of p. */
p = (uint16_t *)(void *)cp;
key_data = (char *)key->data;
_DBFIT(key->size, int);
key_size = key->size;
val_data = (char *)val->data;
_DBFIT(val->size, int);
val_size = val->size;
/* First move the Key */
temp = FREESPACE(p) - BIGOVERHEAD;
_DBFIT(temp, uint16_t);
space = (uint16_t)temp;
while (key_size) {
move_bytes = MIN(space, key_size);
off = OFFSET(p) - move_bytes;
memmove(cp + off, key_data, (size_t)move_bytes);
key_size -= move_bytes;
key_data += move_bytes;
n = p[0];
p[++n] = off;
p[0] = ++n;
temp = off - PAGE_META(n);
_DBFIT(temp, uint16_t);
FREESPACE(p) = (uint16_t)temp;
OFFSET(p) = off;
p[n] = PARTIAL_KEY;
bufp = __add_ovflpage(hashp, bufp);
if (!bufp)
return (-1);
n = p[0];
if (!key_size) {
space = FREESPACE(p);
if (space) {
move_bytes = MIN(space, val_size);
/*
* If the data would fit exactly in the
* remaining space, we must overflow it to the
* next page; otherwise the invariant that the
* data must end on a page with FREESPACE
* non-zero would fail.
*/
if (space == val_size && val_size == val->size)
goto toolarge;
off = OFFSET(p) - move_bytes;
memmove(cp + off, val_data, (size_t)move_bytes);
val_data += move_bytes;
val_size -= move_bytes;
p[n] = off;
p[n - 2] = FULL_KEY_DATA;
FREESPACE(p) = FREESPACE(p) - move_bytes;
OFFSET(p) = off;
} else {
toolarge:
p[n - 2] = FULL_KEY;
}
}
p = (uint16_t *)(void *)bufp->page;
cp = bufp->page;
bufp->flags |= BUF_MOD;
temp = FREESPACE(p) - BIGOVERHEAD;
_DBFIT(temp, uint16_t);
space = (uint16_t)temp;
}
/* Now move the data */
temp = FREESPACE(p) - BIGOVERHEAD;
_DBFIT(temp, uint16_t);
space = (uint16_t)temp;
while (val_size) {
move_bytes = MIN(space, val_size);
/*
* Here's the hack to make sure that if the data ends on the
* same page as the key ends, FREESPACE is at least one.
*/
if (space == val_size && val_size == val->size)
move_bytes--;
off = OFFSET(p) - move_bytes;
memmove(cp + off, val_data, (size_t)move_bytes);
val_size -= move_bytes;
val_data += move_bytes;
n = p[0];
p[++n] = off;
p[0] = ++n;
temp = off - PAGE_META(n);
_DBFIT(temp, uint16_t);
FREESPACE(p) = (uint16_t)temp;
OFFSET(p) = off;
if (val_size) {
p[n] = FULL_KEY;
bufp = __add_ovflpage(hashp, bufp);
if (!bufp)
return (-1);
cp = bufp->page;
p = (uint16_t *)(void *)cp;
} else
p[n] = FULL_KEY_DATA;
bufp->flags |= BUF_MOD;
temp = FREESPACE(p) - BIGOVERHEAD;
_DBFIT(temp, uint16_t);
space = (uint16_t)temp;
}
return (0);
}
/*
* Returns:
* 0 => OK
* -1 => error
*/
int
__big_split(
HTAB *hashp,
BUFHEAD *op, /* Pointer to where to put keys that go in old bucket */
BUFHEAD *np, /* Pointer to new bucket page */
/* Pointer to first page containing the big key/data */
BUFHEAD *big_keyp,
int addr, /* Address of big_keyp */
uint32_t obucket,/* Old Bucket */
SPLIT_RETURN *ret
)
{
BUFHEAD *tmpp;
uint16_t *tp;
BUFHEAD *bp;
DBT key, val;
uint32_t change;
uint16_t free_space, n, off;
size_t temp;
bp = big_keyp;
/* Now figure out where the big key/data goes */
if (__big_keydata(hashp, big_keyp, &key, &val, 0))
return (-1);
change = (__call_hash(hashp, key.data, (int)key.size) != obucket);
if ((ret->next_addr = __find_last_page(hashp, &big_keyp)) != 0) {
if (!(ret->nextp =
__get_buf(hashp, (uint32_t)ret->next_addr, big_keyp, 0)))
return (-1);
} else
ret->nextp = NULL;
/* Now make one of np/op point to the big key/data pair */
_DIAGASSERT(np->ovfl == NULL);
if (change)
tmpp = np;
else
tmpp = op;
tmpp->flags |= BUF_MOD;
#ifdef DEBUG1
(void)fprintf(stderr,
"BIG_SPLIT: %d->ovfl was %d is now %d\n", tmpp->addr,
(tmpp->ovfl ? tmpp->ovfl->addr : 0), (bp ? bp->addr : 0));
#endif
tmpp->ovfl = bp; /* one of op/np point to big_keyp */
tp = (uint16_t *)(void *)tmpp->page;
_DIAGASSERT(FREESPACE(tp) >= OVFLSIZE);
n = tp[0];
off = OFFSET(tp);
free_space = FREESPACE(tp);
tp[++n] = (uint16_t)addr;
tp[++n] = OVFLPAGE;
tp[0] = n;
OFFSET(tp) = off;
temp = free_space - OVFLSIZE;
_DBFIT(temp, uint16_t);
FREESPACE(tp) = (uint16_t)temp;
/*
* Finally, set the new and old return values. BIG_KEYP contains a
* pointer to the last page of the big key_data pair. Make sure that
* big_keyp has no following page (2 elements) or create an empty
* following page.
*/
ret->newp = np;
ret->oldp = op;
tp = (uint16_t *)(void *)big_keyp->page;
big_keyp->flags |= BUF_MOD;
if (tp[0] > 2) {
/*
* There may be either one or two offsets on this page. If
* there is one, then the overflow page is linked on normally
* and tp[4] is OVFLPAGE. If there are two, tp[4] contains
* the second offset and needs to get stuffed in after the
* next overflow page is added.
*/
n = tp[4];
free_space = FREESPACE(tp);
off = OFFSET(tp);
tp[0] -= 2;
temp = free_space + OVFLSIZE;
_DBFIT(temp, uint16_t);
FREESPACE(tp) = (uint16_t)temp;
OFFSET(tp) = off;
tmpp = __add_ovflpage(hashp, big_keyp);
if (!tmpp)
return (-1);
tp[4] = n;
} else
tmpp = big_keyp;
if (change)
ret->newp = tmpp;
else
ret->oldp = tmpp;
return (0);
}
/*
* Called when bufp's page contains a partial key (index should be 1)
*
* All pages in the big key/data pair except bufp are freed. We cannot
* free bufp because the page pointing to it is lost and we can't get rid
* of its pointer.
*
* Returns:
* 0 => OK
*-1 => ERROR
*/
int
__big_delete(HTAB *hashp, BUFHEAD *bufp)
{
BUFHEAD *last_bfp, *rbufp;
uint16_t *bp, pageno;
int key_done, n;
size_t temp;
rbufp = bufp;
last_bfp = NULL;
bp = (uint16_t *)(void *)bufp->page;
pageno = 0;
key_done = 0;
while (!key_done || (bp[2] != FULL_KEY_DATA)) {
if (bp[2] == FULL_KEY || bp[2] == FULL_KEY_DATA)
key_done = 1;
/*
* If there is freespace left on a FULL_KEY_DATA page, then
* the data is short and fits entirely on this page, and this
* is the last page.
*/
if (bp[2] == FULL_KEY_DATA && FREESPACE(bp))
break;
pageno = bp[bp[0] - 1];
rbufp->flags |= BUF_MOD;
rbufp = __get_buf(hashp, (uint32_t)pageno, rbufp, 0);
if (last_bfp)
__free_ovflpage(hashp, last_bfp);
last_bfp = rbufp;
if (!rbufp)
return (-1); /* Error. */
bp = (uint16_t *)(void *)rbufp->page;
}
/*
* If we get here then rbufp points to the last page of the big
* key/data pair. Bufp points to the first one -- it should now be
* empty pointing to the next page after this pair. Can't free it
* because we don't have the page pointing to it.
*/
/* This is information from the last page of the pair. */
n = bp[0];
pageno = bp[n - 1];
/* Now, bp is the first page of the pair. */
bp = (uint16_t *)(void *)bufp->page;
if (n > 2) {
/* There is an overflow page. */
bp[1] = pageno;
bp[2] = OVFLPAGE;
bufp->ovfl = rbufp->ovfl;
} else
/* This is the last page. */
bufp->ovfl = NULL;
n -= 2;
bp[0] = n;
temp = hashp->BSIZE - PAGE_META(n);
_DBFIT(temp, uint16_t);
FREESPACE(bp) = (uint16_t)temp;
OFFSET(bp) = hashp->BSIZE;
bufp->flags |= BUF_MOD;
if (rbufp)
__free_ovflpage(hashp, rbufp);
if (last_bfp && last_bfp != rbufp)
__free_ovflpage(hashp, last_bfp);
hashp->NKEYS--;
return (0);
}
/*
* __BT_SPLIT -- Split the tree.
*
* Parameters:
* t: tree
* sp: page to split
* key: key to insert
* data: data to insert
* flags: BIGKEY/BIGDATA flags
* ilen: insert length
* skip: index to leave open
*
* Returns:
* RET_ERROR, RET_SUCCESS
*/
int
__bt_split(BTREE *t, PAGE *sp, const DBT *key, const DBT *data, int flags,
size_t ilen, uint32_t argskip)
{
BINTERNAL *bi = NULL; /* pacify gcc */
BLEAF *bl = NULL, *tbl; /* pacify gcc */
DBT a, b;
EPGNO *parent;
PAGE *h, *l, *r, *lchild, *rchild;
indx_t nxtindex;
uint16_t skip;
uint32_t n, nbytes, nksize = 0; /* pacify gcc */
int parentsplit;
char *dest;
/*
* Split the page into two pages, l and r. The split routines return
* a pointer to the page into which the key should be inserted and with
* skip set to the offset which should be used. Additionally, l and r
* are pinned.
*/
skip = argskip;
h = sp->pgno == P_ROOT ?
bt_root(t, sp, &l, &r, &skip, ilen) :
bt_page(t, sp, &l, &r, &skip, ilen);
if (h == NULL)
return (RET_ERROR);
/*
* Insert the new key/data pair into the leaf page. (Key inserts
* always cause a leaf page to split first.)
*/
_DBFIT(ilen, indx_t);
h->upper -= (indx_t)ilen;
h->linp[skip] = h->upper;
dest = (char *)(void *)h + h->upper;
if (F_ISSET(t, R_RECNO))
WR_RLEAF(dest, data, flags);
else
WR_BLEAF(dest, key, data, flags);
/* If the root page was split, make it look right. */
if (sp->pgno == P_ROOT &&
(F_ISSET(t, R_RECNO) ?
bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
goto err2;
/*
* Now we walk the parent page stack -- a LIFO stack of the pages that
* were traversed when we searched for the page that split. Each stack
* entry is a page number and a page index offset. The offset is for
* the page traversed on the search. We've just split a page, so we
* have to insert a new key into the parent page.
*
* If the insert into the parent page causes it to split, may have to
* continue splitting all the way up the tree. We stop if the root
* splits or the page inserted into didn't have to split to hold the
* new key. Some algorithms replace the key for the old page as well
* as the new page. We don't, as there's no reason to believe that the
* first key on the old page is any better than the key we have, and,
* in the case of a key being placed at index 0 causing the split, the
* key is unavailable.
*
* There are a maximum of 5 pages pinned at any time. We keep the left
* and right pages pinned while working on the parent. The 5 are the
* two children, left parent and right parent (when the parent splits)
* and the root page or the overflow key page when calling bt_preserve.
* This code must make sure that all pins are released other than the
* root page or overflow page which is unlocked elsewhere.
*/
while ((parent = BT_POP(t)) != NULL) {
lchild = l;
rchild = r;
/* Get the parent page. */
if ((h = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
goto err2;
/*
* The new key goes ONE AFTER the index, because the split
* was to the right.
*/
skip = parent->index + 1;
/*
* Calculate the space needed on the parent page.
*
* Prefix trees: space hack when inserting into BINTERNAL
* pages. Retain only what's needed to distinguish between
* the new entry and the LAST entry on the page to its left.
* If the keys compare equal, retain the entire key. Note,
* we don't touch overflow keys, and the entire key must be
* retained for the next-to-left most key on the leftmost
* page of each level, or the search will fail. Applicable
* ONLY to internal pages that have leaf pages as children.
* Further reduction of the key between pairs of internal
* pages loses too much information.
*/
switch (rchild->flags & P_TYPE) {
case P_BINTERNAL:
bi = GETBINTERNAL(rchild, 0);
nbytes = NBINTERNAL(bi->ksize);
break;
case P_BLEAF:
bl = GETBLEAF(rchild, 0);
nbytes = NBINTERNAL(bl->ksize);
if (t->bt_pfx && !(bl->flags & P_BIGKEY) &&
(h->prevpg != P_INVALID || skip > 1)) {
size_t temp;
tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
a.size = tbl->ksize;
a.data = tbl->bytes;
b.size = bl->ksize;
b.data = bl->bytes;
temp = t->bt_pfx(&a, &b);
_DBFIT(temp, uint32_t);
nksize = (uint32_t)temp;
n = NBINTERNAL(nksize);
if (n < nbytes) {
#ifdef STATISTICS
bt_pfxsaved += nbytes - n;
#endif
nbytes = n;
} else
nksize = 0;
} else
nksize = 0;
break;
case P_RINTERNAL:
case P_RLEAF:
nbytes = NRINTERNAL;
break;
default:
abort();
}
/* Split the parent page if necessary or shift the indices. */
if ((uint32_t)h->upper - (uint32_t)h->lower < nbytes + sizeof(indx_t)) {
sp = h;
h = h->pgno == P_ROOT ?
bt_root(t, h, &l, &r, &skip, nbytes) :
bt_page(t, h, &l, &r, &skip, nbytes);
if (h == NULL)
goto err1;
parentsplit = 1;
} else {
if (skip < (nxtindex = NEXTINDEX(h)))
memmove(h->linp + skip + 1, h->linp + skip,
(nxtindex - skip) * sizeof(indx_t));
h->lower += sizeof(indx_t);
parentsplit = 0;
}
/* Insert the key into the parent page. */
switch (rchild->flags & P_TYPE) {
case P_BINTERNAL:
h->linp[skip] = h->upper -= nbytes;
dest = (char *)(void *)h + h->linp[skip];
memmove(dest, bi, nbytes);
((BINTERNAL *)(void *)dest)->pgno = rchild->pgno;
break;
case P_BLEAF:
h->linp[skip] = h->upper -= nbytes;
dest = (char *)(void *)h + h->linp[skip];
WR_BINTERNAL(dest, nksize ? nksize : bl->ksize,
rchild->pgno, bl->flags & P_BIGKEY);
memmove(dest, bl->bytes, nksize ? nksize : bl->ksize);
if (bl->flags & P_BIGKEY &&
bt_preserve(t, *(pgno_t *)(void *)bl->bytes) ==
RET_ERROR)
goto err1;
break;
case P_RINTERNAL:
/*
* Update the left page count. If split
* added at index 0, fix the correct page.
*/
if (skip > 0)
dest = (char *)(void *)h + h->linp[skip - 1];
else
dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
((RINTERNAL *)(void *)dest)->nrecs = rec_total(lchild);
((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
/* Update the right page count. */
h->linp[skip] = h->upper -= nbytes;
dest = (char *)(void *)h + h->linp[skip];
((RINTERNAL *)(void *)dest)->nrecs = rec_total(rchild);
((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
break;
case P_RLEAF:
/*
* Update the left page count. If split
* added at index 0, fix the correct page.
*/
if (skip > 0)
dest = (char *)(void *)h + h->linp[skip - 1];
else
dest = (char *)(void *)l + l->linp[NEXTINDEX(l) - 1];
((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(lchild);
((RINTERNAL *)(void *)dest)->pgno = lchild->pgno;
/* Update the right page count. */
h->linp[skip] = h->upper -= nbytes;
dest = (char *)(void *)h + h->linp[skip];
((RINTERNAL *)(void *)dest)->nrecs = NEXTINDEX(rchild);
((RINTERNAL *)(void *)dest)->pgno = rchild->pgno;
break;
default:
abort();
}
/* Unpin the held pages. */
if (!parentsplit) {
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
break;
}
/* If the root page was split, make it look right. */
if (sp->pgno == P_ROOT &&
(F_ISSET(t, R_RECNO) ?
bt_rroot(t, sp, l, r) : bt_broot(t, sp, l, r)) == RET_ERROR)
goto err1;
mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
}
/* Unpin the held pages. */
mpool_put(t->bt_mp, l, MPOOL_DIRTY);
mpool_put(t->bt_mp, r, MPOOL_DIRTY);
/* Clear any pages left on the stack. */
return (RET_SUCCESS);
/*
* If something fails in the above loop we were already walking back
* up the tree and the tree is now inconsistent. Nothing much we can
* do about it but release any memory we're holding.
*/
err1: mpool_put(t->bt_mp, lchild, MPOOL_DIRTY);
mpool_put(t->bt_mp, rchild, MPOOL_DIRTY);
err2: mpool_put(t->bt_mp, l, 0);
mpool_put(t->bt_mp, r, 0);
__dbpanic(t->bt_dbp);
return (RET_ERROR);
}
/*
* BT_PSPLIT -- Do the real work of splitting the page.
*
* Parameters:
* t: tree
* h: page to be split
* l: page to put lower half of data
* r: page to put upper half of data
* pskip: pointer to index to leave open
* ilen: insert length
*
* Returns:
* Pointer to page in which to insert.
*/
static PAGE *
bt_psplit(BTREE *t, PAGE *h, PAGE *l, PAGE *r, indx_t *pskip, size_t ilen)
{
BINTERNAL *bi;
BLEAF *bl;
CURSOR *c;
RLEAF *rl;
PAGE *rval;
void *src = NULL; /* pacify gcc */
indx_t full, half, nxt, off, skip, top, used;
uint32_t nbytes;
size_t temp;
int bigkeycnt, isbigkey;
/*
* Split the data to the left and right pages. Leave the skip index
* open. Additionally, make some effort not to split on an overflow
* key. This makes internal page processing faster and can save
* space as overflow keys used by internal pages are never deleted.
*/
bigkeycnt = 0;
skip = *pskip;
temp = t->bt_psize - BTDATAOFF;
_DBFIT(temp, indx_t);
full = (indx_t)temp;
half = full / 2;
used = 0;
for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
if (skip == off) {
_DBFIT(ilen, uint32_t);
nbytes = (uint32_t)ilen;
isbigkey = 0; /* XXX: not really known. */
} else
switch (h->flags & P_TYPE) {
case P_BINTERNAL:
src = bi = GETBINTERNAL(h, nxt);
nbytes = NBINTERNAL(bi->ksize);
isbigkey = bi->flags & P_BIGKEY;
break;
case P_BLEAF:
src = bl = GETBLEAF(h, nxt);
nbytes = NBLEAF(bl);
isbigkey = bl->flags & P_BIGKEY;
break;
case P_RINTERNAL:
src = GETRINTERNAL(h, nxt);
nbytes = NRINTERNAL;
isbigkey = 0;
break;
case P_RLEAF:
src = rl = GETRLEAF(h, nxt);
nbytes = NRLEAF(rl);
isbigkey = 0;
break;
default:
abort();
}
/*
* If the key/data pairs are substantial fractions of the max
* possible size for the page, it's possible to get situations
* where we decide to try and copy too much onto the left page.
* Make sure that doesn't happen.
*/
if ((skip <= off && used + nbytes + sizeof(indx_t) >= full) ||
nxt == top - 1) {
--off;
break;
}
/* Copy the key/data pair, if not the skipped index. */
if (skip != off) {
++nxt;
l->linp[off] = l->upper -= nbytes;
memmove((char *)(void *)l + l->upper, src, nbytes);
}
temp = nbytes + sizeof(indx_t);
_DBFIT(temp, indx_t);
used += (indx_t)temp;
if (used >= half) {
if (!isbigkey || bigkeycnt == 3)
break;
else
++bigkeycnt;
}
}
/*
* Off is the last offset that's valid for the left page.
* Nxt is the first offset to be placed on the right page.
*/
temp = (off + 1) * sizeof(indx_t);
_DBFIT(temp, indx_t);
l->lower += (indx_t)temp;
/*
* If splitting the page that the cursor was on, the cursor has to be
* adjusted to point to the same record as before the split. If the
* cursor is at or past the skipped slot, the cursor is incremented by
* one. If the cursor is on the right page, it is decremented by the
* number of records split to the left page.
*/
c = &t->bt_cursor;
if (F_ISSET(c, CURS_INIT) && c->pg.pgno == h->pgno) {
if (c->pg.index >= skip)
++c->pg.index;
if (c->pg.index < nxt) /* Left page. */
c->pg.pgno = l->pgno;
else { /* Right page. */
c->pg.pgno = r->pgno;
c->pg.index -= nxt;
}
}
/*
* If the skipped index was on the left page, just return that page.
* Otherwise, adjust the skip index to reflect the new position on
* the right page.
*/
if (skip <= off) {
skip = MAX_PAGE_OFFSET;
rval = l;
} else {
rval = r;
*pskip -= nxt;
}
for (off = 0; nxt < top; ++off) {
if (skip == nxt) {
++off;
skip = MAX_PAGE_OFFSET;
}
switch (h->flags & P_TYPE) {
case P_BINTERNAL:
src = bi = GETBINTERNAL(h, nxt);
nbytes = NBINTERNAL(bi->ksize);
break;
case P_BLEAF:
src = bl = GETBLEAF(h, nxt);
nbytes = NBLEAF(bl);
break;
case P_RINTERNAL:
src = GETRINTERNAL(h, nxt);
nbytes = NRINTERNAL;
break;
case P_RLEAF:
src = rl = GETRLEAF(h, nxt);
nbytes = NRLEAF(rl);
break;
default:
abort();
}
++nxt;
r->linp[off] = r->upper -= nbytes;
memmove((char *)(void *)r + r->upper, src, nbytes);
}
temp = off * sizeof(indx_t);
_DBFIT(temp, indx_t);
r->lower += (indx_t)temp;
/* If the key is being appended to the page, adjust the index. */
if (skip == top)
r->lower += sizeof(indx_t);
return (rval);
}
