/*
* __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_var_append_prev --
* Return the previous variable-length entry on the append list.
*/
static inline int
__cursor_var_append_prev(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
val = &cbt->iface.value;
if (newpage) {
cbt->ins = WT_SKIP_LAST(cbt->ins_head);
goto new_page;
}
for (;;) {
WT_RET(__cursor_skip_prev(cbt));
new_page: if (cbt->ins == NULL)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, WT_INSERT_RECNO(cbt->ins));
if ((upd = __wt_txn_read(session, cbt->ins->upd)) == NULL)
continue;
if (WT_UPDATE_DELETED_ISSET(upd)) {
if (__wt_txn_visible_all(session, upd->txnid))
++cbt->page_deleted_count;
continue;
}
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
return (0);
}
/* NOTREACHED */
}
/*
* __cursor_var_append_next --
* Return the next variable-length entry on the append list.
*/
static inline int
__cursor_var_append_next(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
if (newpage) {
cbt->ins = WT_SKIP_FIRST(cbt->ins_head);
goto new_page;
}
for (;;) {
cbt->ins = WT_SKIP_NEXT(cbt->ins);
new_page: if (cbt->ins == NULL)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, WT_INSERT_RECNO(cbt->ins));
if ((upd = __wt_txn_read(session, cbt->ins->upd)) == NULL)
continue;
if (upd->type == WT_UPDATE_DELETED) {
if (__wt_txn_upd_visible_all(session, upd))
++cbt->page_deleted_count;
continue;
}
return (__wt_value_return(session, cbt, upd));
}
/* NOTREACHED */
}
/*
* __cursor_var_append_prev --
* Return the previous variable-length entry on the append list.
*/
static inline int
__cursor_var_append_prev(WT_CURSOR_BTREE *cbt, int newpage)
{
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
val = &cbt->iface.value;
if (newpage) {
cbt->ins = WT_SKIP_LAST(cbt->ins_head);
goto new_page;
}
for (;;) {
WT_RET(__cursor_skip_prev(cbt));
new_page: if (cbt->ins == NULL)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, WT_INSERT_RECNO(cbt->ins));
if ((upd = __wt_txn_read(session, cbt->ins->upd)) == NULL ||
WT_UPDATE_DELETED_ISSET(upd))
continue;
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
break;
}
return (0);
}
/*btree cursor移向下一个记录，仅仅在append list上移动*/
static inline int __cursor_fix_append_next(WT_CURSOR_BTREE* cbt, int newpage)
{
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
val = &cbt->iface.value;
/*新载入的page,判断ins_head是否为空，如果为空表示没有append的记录*/
if (newpage){
if ((cbt->ins = WT_SKIP_FIRST(cbt->ins_head)) == NULL)
return (WT_NOTFOUND);
}
else{
/*已经到append list的最后一条记录了，后面没有记录*/
if (cbt->recno >= WT_INSERT_RECNO(cbt->ins) && (cbt->ins = WT_SKIP_NEXT(cbt->ins)) == NULL)
return (WT_NOTFOUND);
}
/*
* This code looks different from the cursor-previous code. The append
* list appears on the last page of the tree, but it may be preceded by
* other rows, which means the cursor's recno will be set to a value and
* we simply want to increment it. If the cursor's recno is NOT set,
* we're starting our iteration in a tree that has only appended items.
* In that case, recno will be 0 and happily enough the increment will
* set it to 1, which is correct.
*/
__cursor_set_recno(cbt, cbt->recno + 1);
/*
* Fixed-width column store appends are inherently non-transactional.
* Even a non-visible update by a concurrent or aborted transaction
* changes the effective end of the data. The effect is subtle because
* of the blurring between deleted and empty values, but ideally we
* would skip all uncommitted changes at the end of the data. This
* doesn't apply to variable-width column stores because the implicitly
* created records written by reconciliation are deleted and so can be
* never seen by a read.
*
* The problem is that we don't know at this point whether there may be
* multiple uncommitted changes at the end of the data, and it would be
* expensive to check every time we hit an aborted update. If an
* insert is aborted, we simply return zero (empty), regardless of
* whether we are at the end of the data.
*/
if (cbt->recno < WT_INSERT_RECNO(cbt->ins) || (upd = __wt_txn_read(session, cbt->ins->upd)) == NULL){ /*没有可见的记录值，直接返回0*/
cbt->v = 0;
val->data = &cbt->v;
}
else
val->data = WT_UPDATE_DATA(upd);
val->size = 1;
return 0;
}
/*
* __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);
}
/*在append list上移动variable-length类型的btree cursor*/
static inline int __cursor_var_append_next(WT_CURSOR_BTREE* cbt, int newpage)
{
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
val = &cbt->iface.value;
if (newpage){
cbt->ins = WT_SKIP_FIRST(cbt->ins_head);
goto new_page;
}
for (;;){
cbt->ins = WT_SKIP_NEXT(cbt->ins);
new_page:
if (cbt->ins == NULL)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, WT_INSERT_RECNO(cbt->ins));
/*事务隔离读，对本事务不可见，继续向前*/
if ((upd = __wt_txn_read(session, cbt->ins->upd)) == NULL)
continue;
/*删除集合，不做指向这条记录，继续向下移动*/
if (WT_UPDATE_DELETED_ISSET(upd)) {
++cbt->page_deleted_count;
continue;
}
/*赋值value*/
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
return 0;
}
}
/*
* __cursor_var_prev --
* Move to the previous, variable-length column-store item.
*/
static inline int
__cursor_var_prev(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_COL *cip;
WT_INSERT *ins;
WT_ITEM *val;
WT_PAGE *page;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
uint64_t rle_start;
session = (WT_SESSION_IMPL *)cbt->iface.session;
page = cbt->ref->page;
val = &cbt->iface.value;
rle_start = 0; /* -Werror=maybe-uninitialized */
/* Initialize for each new page. */
if (newpage) {
cbt->last_standard_recno = __col_var_last_recno(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 (;;) {
__cursor_set_recno(cbt, cbt->recno - 1);
new_page: if (cbt->recno < page->pg_var_recno)
return (WT_NOTFOUND);
/* Find the matching WT_COL slot. */
if ((cip =
__col_var_search(page, cbt->recno, &rle_start)) == NULL)
return (WT_NOTFOUND);
cbt->slot = WT_COL_SLOT(page, cip);
/* Check any insert list for a matching record. */
cbt->ins_head = WT_COL_UPDATE_SLOT(page, cbt->slot);
cbt->ins = __col_insert_search_match(cbt->ins_head, cbt->recno);
upd = cbt->ins == NULL ?
NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd != NULL) {
if (WT_UPDATE_DELETED_ISSET(upd)) {
if (__wt_txn_visible_all(session, upd->txnid))
++cbt->page_deleted_count;
continue;
}
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(page, cip)) == NULL)
continue;
__wt_cell_unpack(cell, &unpack);
if (unpack.type == WT_CELL_DEL) {
if (__wt_cell_rle(&unpack) == 1)
continue;
/*
* There can be huge gaps in the variable-length
* column-store name space appearing as deleted
* records. If more than one deleted record, do
* the work of finding the next record to return
* instead of looping through the records.
*
* First, find the largest record in the update
* list that's smaller than the current record.
*/
ins = __col_insert_search_lt(
cbt->ins_head, cbt->recno);
/*
* Second, for records with RLEs greater than 1,
* the above call to __col_var_search located
* this record in the page's list of repeating
* records, and returned the starting record.
* The starting record - 1 is the record to
* which we could skip, if there was no larger
* record in the update list.
*/
cbt->recno = rle_start - 1;
if (ins != NULL &&
WT_INSERT_RECNO(ins) > cbt->recno)
cbt->recno = WT_INSERT_RECNO(ins);
/* Adjust for the outer loop decrement. */
++cbt->recno;
continue;
}
WT_RET(__wt_page_cell_data_ref(
session, page, &unpack, cbt->tmp));
cbt->cip_saved = cip;
}
val->data = cbt->tmp->data;
val->size = cbt->tmp->size;
return (0);
}
/* NOTREACHED */
}
/*
* __cursor_fix_append_prev --
* Return the previous fixed-length entry on the append list.
*/
static inline int
__cursor_fix_append_prev(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_ITEM *val;
WT_PAGE *page;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
page = cbt->ref->page;
val = &cbt->iface.value;
if (newpage) {
if ((cbt->ins = WT_SKIP_LAST(cbt->ins_head)) == NULL)
return (WT_NOTFOUND);
} else {
/*
* Handle the special case of leading implicit records, that is,
* there aren't any records in the tree not on the append list,
* and the first record on the append list isn't record 1.
*
* The "right" place to handle this is probably in our caller.
* The high-level cursor-previous routine would:
* -- call this routine to walk the append list
* -- call the routine to walk the standard page items
* -- call the tree walk routine looking for a previous page
* Each of them returns WT_NOTFOUND, at which point our caller
* checks the cursor record number, and if it's larger than 1,
* returns the implicit records. Instead, I'm trying to detect
* the case here, mostly because I don't want to put that code
* into our caller. Anyway, if this code breaks for any reason,
* that's the way I'd go.
*
* If we're not pointing to a WT_INSERT entry, or we can't find
* a WT_INSERT record that precedes our record name-space, check
* if there are any records on the page. If there aren't, then
* we're in the magic zone, keep going until we get to a record
* number of 1.
*/
if (cbt->ins != NULL &&
cbt->recno <= WT_INSERT_RECNO(cbt->ins))
WT_RET(__cursor_skip_prev(cbt));
if (cbt->ins == NULL &&
(cbt->recno == 1 || __col_fix_last_recno(page) != 0))
return (WT_NOTFOUND);
}
/*
* This code looks different from the cursor-next code. The append
* list appears on the last page of the tree and contains the last
* records in the tree. If we're iterating through the tree, starting
* at the last record in the tree, by definition we're starting a new
* iteration and we set the record number to the last record found in
* the tree. Otherwise, decrement the record.
*/
if (newpage)
__cursor_set_recno(cbt, WT_INSERT_RECNO(cbt->ins));
else
__cursor_set_recno(cbt, cbt->recno - 1);
/*
* Fixed-width column store appends are inherently non-transactional.
* Even a non-visible update by a concurrent or aborted transaction
* changes the effective end of the data. The effect is subtle because
* of the blurring between deleted and empty values, but ideally we
* would skip all uncommitted changes at the end of the data. This
* doesn't apply to variable-width column stores because the implicitly
* created records written by reconciliation are deleted and so can be
* never seen by a read.
*/
if (cbt->ins == NULL ||
cbt->recno > WT_INSERT_RECNO(cbt->ins) ||
(upd = __wt_txn_read(session, cbt->ins->upd)) == NULL) {
cbt->v = 0;
val->data = &cbt->v;
} else
val->data = WT_UPDATE_DATA(upd);
val->size = 1;
return (0);
}
/*
* __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_ITEM *val;
WT_PAGE *page;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
page = cbt->ref->page;
val = &cbt->iface.value;
/* Initialize for each new page. */
if (newpage) {
cbt->last_standard_recno = __col_var_last_recno(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 (;;) {
__cursor_set_recno(cbt, cbt->recno - 1);
new_page: if (cbt->recno < page->pg_var_recno)
return (WT_NOTFOUND);
/* Find the matching WT_COL slot. */
if ((cip = __col_var_search(page, cbt->recno)) == NULL)
return (WT_NOTFOUND);
cbt->slot = WT_COL_SLOT(page, cip);
/* Check any insert list for a matching record. */
cbt->ins_head = WT_COL_UPDATE_SLOT(page, cbt->slot);
cbt->ins = __col_insert_search_match(cbt->ins_head, cbt->recno);
upd = cbt->ins == NULL ?
NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd != NULL) {
if (WT_UPDATE_DELETED_ISSET(upd)) {
++cbt->page_deleted_count;
continue;
}
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(page, cip)) == NULL)
continue;
__wt_cell_unpack(cell, &unpack);
if (unpack.type == WT_CELL_DEL)
continue;
WT_RET(__wt_page_cell_data_ref(
session, page, &unpack, &cbt->tmp));
cbt->cip_saved = cip;
}
val->data = cbt->tmp.data;
val->size = cbt->tmp.size;
return (0);
}
/* NOTREACHED */
}
/*移向下条variable-length column-store 记录*/
static inline int __cursor_var_next(WT_CURSOR_BTREE* cbt, int newpage)
{
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_COL *cip;
WT_ITEM *val;
WT_INSERT *ins;
WT_PAGE *page;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
uint64_t rle, rle_start;
session = (WT_SESSION_IMPL *)cbt->iface.session;
page = cbt->ref->page;
val = &cbt->iface.value;
rle_start = 0; /* -Werror=maybe-uninitialized */
if (newpage){
cbt->last_standard_recno = __col_var_last_recno(page);
if (cbt->last_standard_recno == 0)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, page->pg_var_recno);
goto new_page;
}
for (;;){
if (cbt->recno >= cbt->last_standard_recno)
return (WT_NOTFOUND);
__cursor_set_recno(cbt, cbt->recno + 1);
new_page:
/*定位到recno对应的WT_COL slot*/
if ((cip = __col_var_search(page, cbt->recno, &rle_start)) == NULL)
return (WT_NOTFOUND);
cbt->slot = WT_COL_SLOT(page, cip);
/*读取内容值*/
cbt->ins_head = WT_COL_UPDATE_SLOT(page, cbt->slot);
cbt->ins = __col_insert_search_match(cbt->ins_head, cbt->recno);
upd = cbt->ins == NULL ? NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd != NULL) {
if (WT_UPDATE_DELETED_ISSET(upd)) {
++cbt->page_deleted_count;
continue;
}
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.
* upd == NULL, 记录可能被删除放入到了insert列表中，slot可能被重用了，那么需要进行cell unpack取值
*/
if (cbt->cip_saved != cip) {
if ((cell = WT_COL_PTR(page, cip)) == NULL)
continue;
__wt_cell_unpack(cell, &unpack);
if (unpack.type == WT_CELL_DEL) {
if ((rle = __wt_cell_rle(&unpack)) == 1)
continue;
/*定位到修改列表中的记录*/
ins = __col_insert_search_gt(cbt->ins_head, cbt->recno);
cbt->recno = rle_start + rle;
if (ins != NULL && WT_INSERT_RECNO(ins) < cbt->recno)
cbt->recno = WT_INSERT_RECNO(ins);
/* Adjust for the outer loop increment. */
--cbt->recno;
continue;
}
/*取出cell中的值到tmp中*/
WT_RET(__wt_page_cell_data_ref(session, page, &unpack, &cbt->tmp));
cbt->cip_saved = cip;
}
val->data = cbt->tmp.data;
val->size = cbt->tmp.size;
return 0;
}
}
/*
* __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_DECL_RET;
WT_ITEM *val;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
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 (;;) {
__cursor_set_recno(cbt, cbt->recno - 1);
new_page: 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);
cbt->slot = WT_COL_SLOT(cbt->page, cip);
/* Check any insert list for a matching record. */
cbt->ins_head = WT_COL_UPDATE_SLOT(cbt->page, cbt->slot);
cbt->ins = __col_insert_search_match(cbt->ins_head, cbt->recno);
upd = cbt->ins == NULL ?
NULL : __wt_txn_read(session, cbt->ins->upd);
if (upd != NULL) {
if (WT_UPDATE_DELETED_ISSET(upd))
continue;
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);
if (unpack.type == WT_CELL_DEL)
continue;
/*
* 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_cell_unpack_ref(
session, &unpack, &cbt->tmp)) == WT_RESTART)
ret = __wt_ovfl_cache_col_restart(
session, cbt->page, &unpack, &cbt->tmp);
WT_RET(ret);
cbt->cip_saved = cip;
}
val->data = cbt->tmp.data;
val->size = cbt->tmp.size;
return (0);
}
/* NOTREACHED */
}
/*
* __cursor_fix_append_prev --
* Return the previous fixed-length entry on the append list.
*/
static inline int
__cursor_fix_append_prev(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
if (newpage) {
if ((cbt->ins = WT_SKIP_LAST(cbt->ins_head)) == NULL)
return (WT_NOTFOUND);
} else {
/* Move to the previous record in the append list, if any. */
if (cbt->ins != NULL &&
cbt->recno <= WT_INSERT_RECNO(cbt->ins))
WT_RET(__cursor_skip_prev(cbt));
/*
* Handle the special case of leading implicit records, that is,
* there aren't any records in the page not on the append list,
* and the append list's first record isn't the first record on
* the page. (Although implemented as a test of the page values,
* this is really a test for a tree where the first inserted
* record wasn't record 1, any other page with only an append
* list will have a first page record number matching the first
* record in the append list.)
*
* The "right" place to handle this is probably in our caller.
* The high-level cursor-previous routine would:
* -- call this routine to walk the append list
* -- call the routine to walk the standard page items
* -- call the tree walk routine looking for a previous page
* Each of them returns WT_NOTFOUND, at which point our caller
* checks the cursor record number, and if it's larger than 1,
* returns the implicit records. Instead, I'm trying to detect
* the case here, mostly because I don't want to put that code
* into our caller. Anyway, if this code breaks for any reason,
* that's the way I'd go.
*
* If we're not pointing to a WT_INSERT entry (we didn't find a
* WT_INSERT record preceding our record name-space), check if
* we've reached the beginning of this page, a possibility if a
* page had a large number of items appended, and then split.
* If not, check if there are any records on the page. If there
* aren't, then we're in the magic zone, keep going until we get
* to a record number matching the first record on the page.
*/
if (cbt->ins == NULL &&
(cbt->recno == cbt->ref->ref_recno ||
__col_fix_last_recno(cbt->ref) != 0))
return (WT_NOTFOUND);
}
/*
* This code looks different from the cursor-next code. The append list
* may be preceded by other rows. If we're iterating through the tree,
* starting at the last record in the tree, by definition we're starting
* a new iteration and we set the record number to the last record found
* on the page. Otherwise, decrement the record.
*/
if (newpage)
__cursor_set_recno(cbt, WT_INSERT_RECNO(cbt->ins));
else
__cursor_set_recno(cbt, cbt->recno - 1);
/*
* Fixed-width column store appends are inherently non-transactional.
* Even a non-visible update by a concurrent or aborted transaction
* changes the effective end of the data. The effect is subtle because
* of the blurring between deleted and empty values, but ideally we
* would skip all uncommitted changes at the end of the data. This
* doesn't apply to variable-width column stores because the implicitly
* created records written by reconciliation are deleted and so can be
* never seen by a read.
*/
if (cbt->ins == NULL ||
cbt->recno > WT_INSERT_RECNO(cbt->ins) ||
(upd = __wt_txn_read(session, cbt->ins->upd)) == NULL) {
cbt->v = 0;
cbt->iface.value.data = &cbt->v;
} else
cbt->iface.value.data = upd->data;
cbt->iface.value.size = 1;
return (0);
}