示例#1
0
/*
 * 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);
}
示例#2
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, 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);
}