/*
* __wt_col_leaf_obsolete --
* Discard all obsolete updates on a column-store leaf page.
*/
void
__wt_col_leaf_obsolete(WT_SESSION_IMPL *session, WT_PAGE *page)
{
uint32_t i;
WT_COL *cip;
WT_INSERT *ins;
WT_UPDATE *upd;
switch (page->type) {
case WT_PAGE_COL_FIX:
WT_SKIP_FOREACH(ins, WT_COL_UPDATE_SINGLE(page))
if ((upd = __wt_update_obsolete_check(
session, ins->upd)) != NULL)
__wt_update_obsolete_free(session, page, upd);
break;
case WT_PAGE_COL_VAR:
WT_COL_FOREACH(page, cip, i)
WT_SKIP_FOREACH(ins, WT_COL_UPDATE(page, cip))
if ((upd = __wt_update_obsolete_check(
session, ins->upd)) != NULL)
__wt_update_obsolete_free(
session, page, upd);
break;
}
/* Walk any append list. */
WT_SKIP_FOREACH(ins, WT_COL_APPEND(page)) {
if ((upd =
__wt_update_obsolete_check(session, ins->upd)) != NULL)
__wt_update_obsolete_free(session, page, upd);
}
}
/*
* __stat_page_col_var --
* Stat a WT_PAGE_COL_VAR page.
*/
static int
__stat_page_col_var(WT_PAGE *page, WT_DSRC_STATS *stats)
{
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COL *cip;
WT_INSERT *ins;
WT_UPDATE *upd;
uint32_t i;
int orig_deleted;
unpack = &_unpack;
WT_STAT_INCR(stats, btree_column_variable);
/*
* Walk the page, counting regular and overflow data items, and checking
* to be sure any updates weren't deletions. If the item was updated,
* assume it was updated by an item of the same size (it's expensive to
* figure out if it will require the same space or not, especially if
* there's Huffman encoding).
*/
WT_COL_FOREACH(page, cip, i) {
if ((cell = WT_COL_PTR(page, cip)) == NULL) {
orig_deleted = 1;
WT_STAT_INCR(stats, btree_column_deleted);
} else {
orig_deleted = 0;
__wt_cell_unpack(cell, unpack);
WT_STAT_INCRV(
stats, btree_entries, __wt_cell_rle(unpack));
}
/*
* Walk the insert list, checking for changes. For each insert
* we find, correct the original count based on its state.
*/
WT_SKIP_FOREACH(ins, WT_COL_UPDATE(page, cip)) {
upd = ins->upd;
if (WT_UPDATE_DELETED_ISSET(upd)) {
if (orig_deleted)
continue;
WT_STAT_INCR(stats, btree_column_deleted);
WT_STAT_DECR(stats, btree_entries);
} else {
if (!orig_deleted)
continue;
WT_STAT_DECR(stats, btree_column_deleted);
WT_STAT_INCR(stats, btree_entries);
}
}
}
/*
* __stat_page_col_var --
* Stat a WT_PAGE_COL_VAR page.
*/
static void
__stat_page_col_var(
WT_SESSION_IMPL *session, WT_PAGE *page, WT_DSRC_STATS **stats)
{
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COL *cip;
WT_INSERT *ins;
WT_UPDATE *upd;
uint64_t deleted_cnt, entry_cnt, ovfl_cnt, rle_cnt;
uint32_t i;
bool orig_deleted;
unpack = &_unpack;
deleted_cnt = entry_cnt = ovfl_cnt = rle_cnt = 0;
WT_STAT_INCR(session, stats, btree_column_variable);
/*
* Walk the page counting regular items, adjusting if the item has been
* subsequently deleted or not. This is a mess because 10-item RLE might
* have 3 of the items subsequently deleted. Overflow items are harder,
* we can't know if an updated item will be an overflow item or not; do
* our best, and simply count every overflow item (or RLE set of items)
* we see.
*/
WT_COL_FOREACH(page, cip, i) {
if ((cell = WT_COL_PTR(page, cip)) == NULL) {
orig_deleted = true;
++deleted_cnt;
} else {
orig_deleted = false;
__wt_cell_unpack(cell, unpack);
if (unpack->type == WT_CELL_ADDR_DEL)
orig_deleted = true;
else {
entry_cnt += __wt_cell_rle(unpack);
rle_cnt += __wt_cell_rle(unpack) - 1;
}
if (unpack->ovfl)
++ovfl_cnt;
}
/*
* Walk the insert list, checking for changes. For each insert
* we find, correct the original count based on its state.
*/
WT_SKIP_FOREACH(ins, WT_COL_UPDATE(page, cip)) {
upd = ins->upd;
if (WT_UPDATE_DELETED_ISSET(upd)) {
if (!orig_deleted) {
++deleted_cnt;
--entry_cnt;
}
} else
if (orig_deleted) {
--deleted_cnt;
++entry_cnt;
}
}
}
/*
* __cursor_var_prev --
* Move to the previous, variable-length column-store item.
*/
static inline int
__cursor_var_prev(WT_CURSOR_BTREE *cbt, int newpage)
{
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_COL *cip;
WT_INSERT *ins;
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
uint64_t *recnop;
session = (WT_SESSION_IMPL *)cbt->iface.session;
recnop = &cbt->iface.recno;
val = &cbt->iface.value;
/* Initialize for each new page. */
if (newpage) {
cbt->last_standard_recno = __col_last_recno(cbt->page);
if (cbt->last_standard_recno == 0)
return (WT_NOTFOUND);
cbt->recno = cbt->last_standard_recno;
goto new_page;
}
/* Move to the previous entry and return the item. */
for (;;) {
--cbt->recno;
new_page: *recnop = cbt->recno;
if (cbt->recno < cbt->page->u.col_var.recno)
return (WT_NOTFOUND);
/* Find the matching WT_COL slot. */
if ((cip = __col_var_search(cbt->page, cbt->recno)) == NULL)
return (WT_NOTFOUND);
/* Check any insert list for a matching record. */
if ((ins = __col_insert_search_match(
WT_COL_UPDATE(cbt->page, cip), cbt->recno)) != NULL &&
(upd = __wt_txn_read(session, ins->upd)) != NULL) {
if (WT_UPDATE_DELETED_ISSET(upd))
continue;
cbt->ins = ins;
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
return (0);
}
/*
* If we're at the same slot as the last reference and there's
* no matching insert list item, re-use the return information
* (so encoded items with large repeat counts aren't repeatedly
* decoded). Otherwise, unpack the cell and build the return
* information.
*/
if (cbt->cip_saved != cip) {
if ((cell = WT_COL_PTR(cbt->page, cip)) == NULL)
continue;
__wt_cell_unpack(cell, &unpack);
switch (unpack.type) {
case WT_CELL_DEL:
continue;
case WT_CELL_VALUE:
if (session->btree->huffman_value == NULL) {
cbt->tmp.data = unpack.data;
cbt->tmp.size = unpack.size;
break;
}
/* FALLTHROUGH */
default:
WT_RET(__wt_cell_unpack_copy(
session, &unpack, &cbt->tmp));
}
cbt->cip_saved = cip;
}
val->data = cbt->tmp.data;
val->size = cbt->tmp.size;
return (0);
}
/* NOTREACHED */
}
