/*
* BT_BROOT -- Fix up the btree 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_broot(BTREE *t, PAGE *h, PAGE *l, PAGE *r)
{
BINTERNAL *bi;
BLEAF *bl;
u_int32_t nbytes;
char *dest;
/*
* If the root page was a leaf page, change it into an internal page.
* We copy the key we split on (but not the key's data, in the case of
* a leaf page) to the new root page.
*
* The btree comparison code guarantees that the left-most key on any
* level of the tree is never used, so it doesn't need to be filled in.
*/
nbytes = NBINTERNAL(0);
h->linp[0] = h->upper = t->bt_psize - nbytes;
dest = (char *)h + h->upper;
WR_BINTERNAL(dest, 0, l->pgno, 0);
switch (h->flags & P_TYPE) {
case P_BLEAF:
bl = GETBLEAF(r, 0);
nbytes = NBINTERNAL(bl->ksize);
__PAST_END(h->linp, 1) = h->upper -= nbytes;
dest = (char *)h + h->upper;
WR_BINTERNAL(dest, bl->ksize, r->pgno, 0);
memmove(dest, bl->bytes, bl->ksize);
/*
* If the key is on an overflow page, mark the overflow chain
* so it isn't deleted when the leaf copy of the key is deleted.
*/
if (bl->flags & P_BIGKEY &&
bt_preserve(t, *(pgno_t *)bl->bytes) == RET_ERROR)
return (RET_ERROR);
break;
case P_BINTERNAL:
bi = GETBINTERNAL(r, 0);
nbytes = NBINTERNAL(bi->ksize);
__PAST_END(h->linp, 1) = h->upper -= nbytes;
dest = (char *)h + h->upper;
memmove(dest, bi, nbytes);
((BINTERNAL *)dest)->pgno = r->pgno;
break;
default:
abort();
}
/* There are two keys on the page. */
h->lower = BTDATAOFF + 2 * sizeof(indx_t);
/* Unpin the root page, set to btree internal page. */
h->flags &= ~P_TYPE;
h->flags |= P_BINTERNAL;
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
/*
* __bt_pdelete --
* Delete a single page from the tree.
*
* Parameters:
* t: tree
* h: leaf page
*
* Returns:
* RET_SUCCESS, RET_ERROR.
*
* Side-effects:
* mpool_put's the page
*/
static int
__bt_pdelete(BTREE *t, PAGE *h)
{
BINTERNAL *bi;
PAGE *pg;
EPGNO *parent;
indx_t cnt, idx, *ip, offset;
u_int32_t nksize;
char *from;
/*
* Walk the parent page stack -- a LIFO stack of the pages that were
* traversed when we searched for the page where the delete occurred.
* 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 deleted
* a page, so we have to delete the key from the parent page.
*
* If the delete from the parent page makes it empty, this process may
* continue all the way up the tree. We stop if we reach the root page
* (which is never deleted, it's just not worth the effort) or if the
* delete does not empty the page.
*/
while ((parent = BT_POP(t)) != NULL) {
/* Get the parent page. */
if ((pg = mpool_get(t->bt_mp, parent->pgno, 0)) == NULL)
return (RET_ERROR);
idx = parent->index;
bi = GETBINTERNAL(pg, idx);
/* Free any overflow pages. */
if (bi->flags & P_BIGKEY &&
__ovfl_delete(t, bi->bytes) == RET_ERROR) {
mpool_put(t->bt_mp, pg, 0);
return (RET_ERROR);
}
/*
* Free the parent if it has only the one key and it's not the
* root page. If it's the rootpage, turn it back into an empty
* leaf page.
*/
if (NEXTINDEX(pg) == 1) {
if (pg->pgno == P_ROOT) {
pg->lower = BTDATAOFF;
pg->upper = t->bt_psize;
pg->flags = P_BLEAF;
} else {
if (__bt_relink(t, pg) || __bt_free(t, pg))
return (RET_ERROR);
continue;
}
} else {
/* Pack remaining key items at the end of the page. */
nksize = NBINTERNAL(bi->ksize);
from = (char *)pg + pg->upper;
memmove(from + nksize, from, (char *)bi - from);
pg->upper += nksize;
/* Adjust indices' offsets, shift the indices down. */
offset = pg->linp[idx];
for (cnt = idx, ip = &pg->linp[0]; cnt--; ++ip)
if (ip[0] < offset)
ip[0] += nksize;
for (cnt = NEXTINDEX(pg) - idx; --cnt; ++ip)
ip[0] = ip[1] < offset ? ip[1] + nksize : ip[1];
pg->lower -= sizeof(indx_t);
}
mpool_put(t->bt_mp, pg, MPOOL_DIRTY);
break;
}
/* Free the leaf page, as long as it wasn't the root. */
if (h->pgno == P_ROOT) {
mpool_put(t->bt_mp, h, MPOOL_DIRTY);
return (RET_SUCCESS);
}
return (__bt_relink(t, h) || __bt_free(t, h));
}
/*
* __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, u_int32_t argskip)
{
BINTERNAL *bi;
BLEAF *bl, *tbl;
DBT a, b;
EPGNO *parent;
PAGE *h, *l, *r, *lchild, *rchild;
indx_t nxtindex;
u_int16_t skip;
u_int32_t n, nbytes, nksize;
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.)
*/
h->linp[skip] = h->upper -= ilen;
dest = (char *)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)) {
tbl = GETBLEAF(lchild, NEXTINDEX(lchild) - 1);
a.size = tbl->ksize;
a.data = tbl->bytes;
b.size = bl->ksize;
b.data = bl->bytes;
nksize = t->bt_pfx(&a, &b);
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 ((u_int32_t)(h->upper - 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 *)h + h->linp[skip];
memmove(dest, bi, nbytes);
((BINTERNAL *)dest)->pgno = rchild->pgno;
break;
case P_BLEAF:
h->linp[skip] = h->upper -= nbytes;
dest = (char *)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) {
pgno_t pgno;
memcpy(&pgno, bl->bytes, sizeof(pgno));
if (bt_preserve(t, pgno) == 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 *)h + h->linp[skip - 1];
else
dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
((RINTERNAL *)dest)->nrecs = rec_total(lchild);
((RINTERNAL *)dest)->pgno = lchild->pgno;
/* Update the right page count. */
h->linp[skip] = h->upper -= nbytes;
dest = (char *)h + h->linp[skip];
((RINTERNAL *)dest)->nrecs = rec_total(rchild);
((RINTERNAL *)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 *)h + h->linp[skip - 1];
else
dest = (char *)l + l->linp[NEXTINDEX(l) - 1];
((RINTERNAL *)dest)->nrecs = NEXTINDEX(lchild);
((RINTERNAL *)dest)->pgno = lchild->pgno;
/* Update the right page count. */
h->linp[skip] = h->upper -= nbytes;
dest = (char *)h + h->linp[skip];
((RINTERNAL *)dest)->nrecs = NEXTINDEX(rchild);
((RINTERNAL *)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;
indx_t full, half, nxt, off, skip, top, used;
u_int32_t nbytes;
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;
full = t->bt_psize - BTDATAOFF;
half = full / 2;
used = 0;
for (nxt = off = 0, top = NEXTINDEX(h); nxt < top; ++off) {
if (skip == off) {
nbytes = 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 *)l + l->upper, src, nbytes);
}
used += nbytes + sizeof(indx_t);
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.
*/
l->lower += (off + 1) * sizeof(indx_t);
/*
* 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 *)r + r->upper, src, nbytes);
}
r->lower += off * sizeof(indx_t);
/* If the key is being appended to the page, adjust the index. */
if (skip == top)
r->lower += sizeof(indx_t);
return (rval);
}
