/*
* BT_ROOT -- Split the root page of a btree.
*
* Parameters:
* t: tree
* h: root page
* lp: pointer to left page pointer
* rp: pointer to right page pointer
* skip: pointer to index to leave open
* ilen: insert length
*
* Returns:
* Pointer to page in which to insert or NULL on error.
*/
static PAGE *
bt_root(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
{
PAGE *l, *r, *tp;
pgno_t lnpg, rnpg;
#ifdef STATISTICS
++bt_split;
++bt_rootsplit;
#endif
/* Put the new left and right pages for the split into place. */
if ((l = __bt_new(t, &lnpg)) == NULL ||
(r = __bt_new(t, &rnpg)) == NULL)
return (NULL);
l->pgno = lnpg;
r->pgno = rnpg;
l->nextpg = r->pgno;
r->prevpg = l->pgno;
l->prevpg = r->nextpg = P_INVALID;
l->lower = r->lower = BTDATAOFF;
l->upper = r->upper = t->bt_psize;
l->flags = r->flags = h->flags & P_TYPE;
/* Split the root page. */
tp = bt_psplit(t, h, l, r, skip, ilen);
*lp = l;
*rp = r;
return (tp);
}
/*
* __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);
}
/*
* BT_PAGE -- Split a non-root page of a btree.
*
* Parameters:
* t: tree
* h: root page
* lp: pointer to left page pointer
* rp: pointer to right page pointer
* skip: pointer to index to leave open
* ilen: insert length
*
* Returns:
* Pointer to page in which to insert or NULL on error.
*/
static PAGE *
bt_page(BTREE *t, PAGE *h, PAGE **lp, PAGE **rp, indx_t *skip, size_t ilen)
{
PAGE *l, *r, *tp;
pgno_t npg;
#ifdef STATISTICS
++bt_split;
#endif
/* Put the new right page for the split into place. */
if ((r = __bt_new(t, &npg)) == NULL)
return (NULL);
r->pgno = npg;
r->lower = BTDATAOFF;
r->upper = t->bt_psize;
r->nextpg = h->nextpg;
r->prevpg = h->pgno;
r->flags = h->flags & P_TYPE;
/*
* If we're splitting the last page on a level because we're appending
* a key to it (skip is NEXTINDEX()), it's likely that the data is
* sorted. Adding an empty page on the side of the level is less work
* and can push the fill factor much higher than normal. If we're
* wrong it's no big deal, we'll just do the split the right way next
* time. It may look like it's equally easy to do a similar hack for
* reverse sorted data, that is, split the tree left, but it's not.
* Don't even try.
*/
if (h->nextpg == P_INVALID && *skip == NEXTINDEX(h)) {
#ifdef STATISTICS
++bt_sortsplit;
#endif
h->nextpg = r->pgno;
r->lower = BTDATAOFF + sizeof(indx_t);
*skip = 0;
*lp = h;
*rp = r;
return (r);
}
/* Put the new left page for the split into place. */
if ((l = (PAGE *)calloc(1, t->bt_psize)) == NULL) {
mpool_put(t->bt_mp, r, 0);
return (NULL);
}
l->pgno = h->pgno;
l->nextpg = r->pgno;
l->prevpg = h->prevpg;
l->lower = BTDATAOFF;
l->upper = t->bt_psize;
l->flags = h->flags & P_TYPE;
/* Fix up the previous pointer of the page after the split page. */
if (h->nextpg != P_INVALID) {
if ((tp = mpool_get(t->bt_mp, h->nextpg, 0)) == NULL) {
free(l);
/* XXX mpool_free(t->bt_mp, r->pgno); */
return (NULL);
}
tp->prevpg = r->pgno;
mpool_put(t->bt_mp, tp, MPOOL_DIRTY);
}
/*
* Split right. The key/data pairs aren't sorted in the btree page so
* it's simpler to copy the data from the split page onto two new pages
* instead of copying half the data to the right page and compacting
* the left page in place. Since the left page can't change, we have
* to swap the original and the allocated left page after the split.
*/
tp = bt_psplit(t, h, l, r, skip, ilen);
/* Move the new left page onto the old left page. */
memmove(h, l, t->bt_psize);
if (tp == l)
tp = h;
free(l);
*lp = h;
*rp = r;
return (tp);
}
