/*
* __cursor_fix_prev --
* Move to the previous, fixed-length column-store item.
*/
static inline int
__cursor_fix_prev(WT_CURSOR_BTREE *cbt, int newpage)
{
WT_BTREE *btree;
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = session->btree;
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);
__cursor_set_recno(cbt, cbt->last_standard_recno);
goto new_page;
}
/* Move to the previous entry and return the item. */
for (;;) {
if (cbt->recno == cbt->page->u.col_fix.recno)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, cbt->recno - 1);
new_page: /* Check any insert list for a matching record. */
cbt->ins_head = WT_COL_UPDATE_SINGLE(cbt->page);
cbt->ins = __col_insert_search(
cbt->ins_head, cbt->ins_stack, cbt->next_stack, cbt->recno);
if (cbt->ins != NULL &&
cbt->recno != WT_INSERT_RECNO(cbt->ins))
cbt->ins = NULL;
upd = cbt->ins == NULL ?
NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd != NULL) {
val->data = WT_UPDATE_DATA(upd);
val->size = 1;
return (0);
}
cbt->v = __bit_getv_recno(cbt->page, cbt->recno, btree->bitcnt);
val->data = &cbt->v;
val->size = 1;
return (0);
}
/* NOTREACHED */
}
/*btree cursor移向下一个记录(fix col方式存储)，在btree树空间上移动*/
static inline int __cursor_fix_next(WT_CURSOR_BTREE* cbt, int newpage)
{
WT_BTREE *btree;
WT_ITEM *val;
WT_PAGE *page;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = S2BT(session);
page = cbt->ref->page;
val = &cbt->iface.value;
/*切换到新的page上做next操作*/
if (newpage){
cbt->last_standard_recno = __col_fix_last_recno(page);
if (cbt->last_standard_recno == 0)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, page->pg_fix_recno);
goto new_page;
}
/*记录序号超出最后一个序号，到末尾了*/
if (cbt->recno >= cbt->last_standard_recno)
return WT_NOTFOUND;
__cursor_set_recno(cbt, cbt->recno + 1);
new_page:
/*获得第一个修改列表*/
cbt->ins_head = WT_COL_UPDATE_SINGLE(page);
/*定位recno所在的修改条目*/
cbt->ins = __col_insert_search(cbt->ins_head, cbt->ins_stack, cbt->next_stack, cbt->recno);
if (cbt->ins != NULL && cbt->recno != WT_INSERT_RECNO(cbt->ins)) /*不正确的修改定位，将ins设置为NULL表示定位失败*/
cbt->ins = NULL;
/*做事务隔离读取记录版本*/
upd = cbt->ins == NULL ? NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd == NULL){
cbt->v = __bit_getv_recno(page, cbt->recno, btree->bitcnt); /*V赋值*/
val->data = &cbt->v;
}
else
val->data = WT_UPDATE_DATA(upd);
val->size = 1;
return 0;
}
/*
* __cursor_fix_prev --
* Move to the previous, fixed-length column-store item.
*/
static inline int
__cursor_fix_prev(WT_CURSOR_BTREE *cbt, int newpage)
{
WT_BTREE *btree;
WT_INSERT *ins;
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
uint64_t *recnop;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = session->btree;
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 (;;) {
if (cbt->recno == cbt->page->u.col_fix.recno)
return (WT_NOTFOUND);
--cbt->recno;
new_page: *recnop = cbt->recno;
/* Check any insert list for a matching record. */
if ((ins = cbt->ins = __col_insert_search_match(
WT_COL_UPDATE_SINGLE(cbt->page), cbt->recno)) != NULL &&
(upd = __wt_txn_read(session, ins->upd)) != NULL) {
val->data = WT_UPDATE_DATA(upd);
val->size = 1;
return (0);
}
cbt->v = __bit_getv_recno(cbt->page, cbt->recno, btree->bitcnt);
val->data = &cbt->v;
val->size = 1;
return (0);
}
/* NOTREACHED */
}
/*
* __cursor_fix_next --
* Move to the next, fixed-length column-store item.
*/
static inline int
__cursor_fix_next(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_BTREE *btree;
WT_PAGE *page;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = S2BT(session);
page = cbt->ref->page;
/* Initialize for each new page. */
if (newpage) {
cbt->last_standard_recno = __col_fix_last_recno(cbt->ref);
if (cbt->last_standard_recno == 0)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, cbt->ref->ref_recno);
goto new_page;
}
/* Move to the next entry and return the item. */
if (cbt->recno >= cbt->last_standard_recno)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, cbt->recno + 1);
new_page:
/* Check any insert list for a matching record. */
cbt->ins_head = WT_COL_UPDATE_SINGLE(page);
cbt->ins = __col_insert_search(
cbt->ins_head, cbt->ins_stack, cbt->next_stack, cbt->recno);
if (cbt->ins != NULL && cbt->recno != WT_INSERT_RECNO(cbt->ins))
cbt->ins = NULL;
upd = cbt->ins == NULL ? NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd == NULL) {
cbt->v = __bit_getv_recno(cbt->ref, cbt->recno, btree->bitcnt);
cbt->iface.value.data = &cbt->v;
} else
cbt->iface.value.data = upd->data;
cbt->iface.value.size = 1;
return (0);
}
int __wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_UPDATE *upd)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_CURSOR *cursor;
WT_PAGE *page;
WT_ROW *rip;
uint8_t v;
switch (page->type){
case WT_PAGE_COL_FIX:
cursor->recno = cbt->recno;
/*cursor对应的是一个upd,直接返回value*/
if (upd != NULL){
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return 0;
}
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return __wt_buf_set(session, &cursor->value, &v, 1);
case WT_PAGE_COL_VAR:
cursor->recno = cbt->recno;
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/*获得对应的cell,并通过cell得到K/V值*/
cell = WT_COL_PTR(page, &page->pg_var_d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->pg_row_d[cbt->slot];
if (cbt->ins != NULL){ /*插入的k/v对*/
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
}
else if (cbt->compare == 0){/*比较器定位到了对应的k/v对*/
cursor->key.data = cbt->search_key.data;
cursor->key.size = cbt->search_key.size;
}
else
WT_RET(__wt_row_leaf_key(session, page, rip, &cursor->key, 0)); /*设置key的值*/
/*值是在append/update list当中，从当中取*/
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/*可以直接通过rip指针获得value,K/V是存储在cell空间之内*/
if (__wt_row_leaf_value(page, rip, &cursor->value))
return 0;
/*不是连续存储的，需要通过解析cell来定位到value*/
if (cell = __wt_row_leaf_value_cell(page, rip, NULL) == NULL){
cursor->value.size = 0;
return 0;
}
break;
WT_ILLEGAL_VALUE(session);
}
/*通过cell解析到对应的value值, ovfl item*/
__wt_cell_unpack(cell, &unpack);
WT_RET(__wt_page_cell_data_ref(session, page, &unpack, &cursor->value));
return 0;
}
/*
* __wt_kv_return --
* Return a page referenced key/value pair to the application.
*/
int
__wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_UPDATE *upd)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_CURSOR *cursor;
WT_ITEM *tmp;
WT_PAGE *page;
WT_ROW *rip;
uint8_t v;
btree = S2BT(session);
page = cbt->ref->page;
cursor = &cbt->iface;
switch (page->type) {
case WT_PAGE_COL_FIX:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/* If the cursor references a WT_UPDATE item, return it. */
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the value from the original page. */
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return (__wt_buf_set(session, &cursor->value, &v, 1));
case WT_PAGE_COL_VAR:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/* If the cursor references a WT_UPDATE item, return it. */
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the value from the original page cell. */
cell = WT_COL_PTR(page, &page->pg_var_d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->pg_row_d[cbt->slot];
/*
* If the cursor references a WT_INSERT item, take its key.
* Else, if we have an exact match, we copied the key in the
* search function, take it from there.
* If we don't have an exact match, take the key from the
* original page.
*/
if (cbt->ins != NULL) {
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
} else if (cbt->compare == 0) {
/*
* If not in an insert list and there's an exact match,
* the row-store search function built the key we want
* to return in the cursor's temporary buffer. Swap the
* cursor's search-key and temporary buffers so we can
* return it (it's unsafe to return the temporary buffer
* itself because our caller might do another search in
* this table using the key we return, and we'd corrupt
* the search key during any subsequent search that used
* the temporary buffer.
*/
tmp = cbt->row_key;
cbt->row_key = cbt->tmp;
cbt->tmp = tmp;
cursor->key.data = cbt->row_key->data;
cursor->key.size = cbt->row_key->size;
} else
WT_RET(__wt_row_leaf_key(
session, page, rip, &cursor->key, false));
/* If the cursor references a WT_UPDATE item, return it. */
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Simple values have their location encoded in the WT_ROW. */
if (__wt_row_leaf_value(page, rip, &cursor->value))
return (0);
/*
* Take the value from the original page cell (which may be
* empty).
*/
if ((cell =
__wt_row_leaf_value_cell(page, rip, NULL)) == NULL) {
cursor->value.size = 0;
return (0);
}
break;
WT_ILLEGAL_VALUE(session);
}
/* The value is an on-page cell, unpack and expand it as necessary. */
__wt_cell_unpack(cell, &unpack);
WT_RET(__wt_page_cell_data_ref(session, page, &unpack, &cursor->value));
return (0);
}
/*
* __wt_kv_return --
* Return a page referenced key/value pair to the application.
*/
int
__wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int key_ret)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_CURSOR *cursor;
WT_IKEY *ikey;
WT_PAGE *page;
WT_ROW *rip;
WT_UPDATE *upd;
uint8_t v;
btree = session->btree;
unpack = &_unpack;
page = cbt->page;
cursor = &cbt->iface;
switch (page->type) {
case WT_PAGE_COL_FIX:
if (key_ret)
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL) {
upd = cbt->ins->upd;
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return (__wt_buf_set(session, &cursor->value, &v, 1));
case WT_PAGE_COL_VAR:
if (key_ret)
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL) {
upd = cbt->ins->upd;
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
cell = WT_COL_PTR(page, &page->u.col_var.d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->u.row.d[cbt->slot];
/*
* If the cursor references a WT_INSERT item, take the key and
* related WT_UPDATE item. Otherwise, take the key from the
* original page, and the value from any related WT_UPDATE item,
* or the page if the key was never updated.
*/
if (cbt->ins == NULL) {
if (key_ret) {
if (__wt_off_page(page, rip->key)) {
ikey = rip->key;
cursor->key.data = WT_IKEY_DATA(ikey);
cursor->key.size = ikey->size;
} else
WT_RET(__wt_row_key(
session, page, rip, &cursor->key));
}
upd = WT_ROW_UPDATE(page, rip);
} else {
if (key_ret) {
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
}
upd = cbt->ins->upd;
}
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the original cell (which may be empty). */
if ((cell = __wt_row_value(page, rip)) == NULL) {
cursor->value.size = 0;
return (0);
}
break;
WT_ILLEGAL_VALUE(session);
}
/* It's a cell, unpack and expand it as necessary. */
__wt_cell_unpack(cell, unpack);
if (btree->huffman_value == NULL && unpack->type == WT_CELL_VALUE) {
cursor->value.data = unpack->data;
cursor->value.size = unpack->size;
return (0);
} else
return (__wt_cell_unpack_copy(session, unpack, &cursor->value));
}
/*
* __wt_kv_return --
* Return a page referenced key/value pair to the application.
*/
int
__wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_PAGE *page;
WT_ROW *rip;
WT_UPDATE *upd;
uint8_t v;
btree = S2BT(session);
page = cbt->page;
cursor = &cbt->iface;
switch (page->type) {
case WT_PAGE_COL_FIX:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return (__wt_buf_set(session, &cursor->value, &v, 1));
case WT_PAGE_COL_VAR:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
cell = WT_COL_PTR(page, &page->u.col_var.d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->u.row.d[cbt->slot];
/*
* If the cursor references a WT_INSERT item, take the key and
* related WT_UPDATE item. Otherwise, take the key from the
* original page, and the value from any related WT_UPDATE item,
* or the page if the key was never updated.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
} else {
WT_RET(
__wt_row_key(session, page, rip, &cursor->key, 0));
upd = __wt_txn_read(session, WT_ROW_UPDATE(page, rip));
}
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the original cell (which may be empty). */
if ((cell = __wt_row_value(page, rip)) == NULL) {
cursor->value.size = 0;
return (0);
}
break;
WT_ILLEGAL_VALUE(session);
}
/* The value is an on-page cell, unpack and expand it as necessary. */
__wt_cell_unpack(cell, &unpack);
ret = __wt_cell_unpack_ref(session, &unpack, &cursor->value);
/*
* Restart for a variable-length column-store. We could catch restart
* higher up the call-stack but there's no point to it: unlike row-store
* (where a normal search path finds cached overflow values), we have to
* access the page's reconciliation structures, and that's as easy here
* as higher up the stack.
*/
if (ret == WT_RESTART && page->type == WT_PAGE_COL_VAR)
ret = __wt_ovfl_cache_col_restart(
session, page, &unpack, &cursor->value);
return (ret);
}