/*
* __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);
}
/*
* __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 */
}
/*
* __wt_update_alloc --
* Allocate a WT_UPDATE structure and associated value and fill it in.
*/
int
__wt_update_alloc(WT_SESSION_IMPL *session,
WT_ITEM *value, WT_UPDATE **updp, size_t *sizep)
{
WT_UPDATE *upd;
size_t size;
/*
* Allocate the WT_UPDATE structure and room for the value, then copy
* the value into place.
*/
size = value == NULL ? 0 : value->size;
WT_RET(__wt_calloc(session, 1, sizeof(WT_UPDATE) + size, &upd));
if (value == NULL)
WT_UPDATE_DELETED_SET(upd);
else {
upd->size = WT_STORE_SIZE(size);
memcpy(WT_UPDATE_DATA(upd), value->data, size);
}
*updp = upd;
if (sizep != NULL)
*sizep = sizeof(WT_UPDATE) + size;
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_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 */
}
/*
* __txn_op_log --
* Log an operation for the current transaction.
*/
static int
__txn_op_log(WT_SESSION_IMPL *session,
WT_ITEM *logrec, WT_TXN_OP *op, WT_CURSOR_BTREE *cbt)
{
WT_DECL_RET;
WT_ITEM key, value;
WT_UPDATE *upd;
uint64_t recno;
WT_CLEAR(key);
upd = op->u.upd;
value.data = WT_UPDATE_DATA(upd);
value.size = upd->size;
/*
* Log the operation. It must be one of the following:
* 1) column store remove;
* 2) column store insert/update;
* 3) row store remove; or
* 4) row store insert/update.
*/
if (cbt->btree->type == BTREE_ROW) {
WT_ERR(__wt_cursor_row_leaf_key(cbt, &key));
if (WT_UPDATE_DELETED_ISSET(upd))
WT_ERR(__wt_logop_row_remove_pack(session, logrec,
op->fileid, &key));
else
WT_ERR(__wt_logop_row_put_pack(session, logrec,
op->fileid, &key, &value));
} else {
recno = WT_INSERT_RECNO(cbt->ins);
WT_ASSERT(session, recno != WT_RECNO_OOB);
if (WT_UPDATE_DELETED_ISSET(upd))
WT_ERR(__wt_logop_col_remove_pack(session, logrec,
op->fileid, recno));
else
WT_ERR(__wt_logop_col_put_pack(session, logrec,
op->fileid, recno, &value));
}
err: __wt_buf_free(session, &key);
return (ret);
}
/*
* __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, 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) {
if ((cbt->ins = WT_SKIP_LAST(cbt->ins_head)) == NULL)
return (WT_NOTFOUND);
cbt->recno = WT_INSERT_RECNO(cbt->ins);
} else {
if (cbt->recno == WT_INSERT_RECNO(cbt->ins)) {
__cursor_skip_prev(cbt);
if (cbt->ins == NULL)
return (WT_NOTFOUND);
}
--cbt->recno;
}
/*
* 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.
*/
cbt->iface.recno = cbt->recno;
if (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);
}
/*
* __txn_op_log --
* Log an operation for the current transaction.
*/
static int
__txn_op_log(WT_SESSION_IMPL *session, WT_ITEM *logrec, WT_TXN_OP *op)
{
WT_ITEM value;
uint64_t recno;
value.data = WT_UPDATE_DATA(op->u.op.upd);
value.size = op->u.op.upd->size;
/*
* Log the operation. It must be one of the following:
* 1) column store remove;
* 2) column store insert/update;
* 3) row store remove; or
* 4) row store insert/update.
*/
if (op->u.op.key.data == NULL) {
WT_ASSERT(session, op->u.op.ins != NULL);
recno = op->u.op.ins->u.recno;
if (WT_UPDATE_DELETED_ISSET(op->u.op.upd))
WT_RET(__wt_logop_col_remove_pack(session, logrec,
op->fileid, recno));
else
WT_RET(__wt_logop_col_put_pack(session, logrec,
op->fileid, recno, &value));
} else {
if (WT_UPDATE_DELETED_ISSET(op->u.op.upd))
WT_RET(__wt_logop_row_remove_pack(session, logrec,
op->fileid, &op->u.op.key));
else
WT_RET(__wt_logop_row_put_pack(session, logrec,
op->fileid, &op->u.op.key, &value));
}
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;
}
}
/*
* __wt_update_alloc --
* Allocate a WT_UPDATE structure and associated value from the session's
* buffer and fill it in.
*/
int
__wt_update_alloc(WT_SESSION_IMPL *session,
WT_ITEM *value, WT_UPDATE **updp, size_t *sizep)
{
WT_DECL_RET;
WT_UPDATE *upd;
size_t size;
/*
* Allocate the WT_UPDATE structure and room for the value, then copy
* the value into place.
*/
size = value == NULL ? 0 : value->size;
WT_RET(__wt_calloc(session, 1, sizeof(WT_UPDATE) + size, &upd));
if (value == NULL)
WT_UPDATE_DELETED_SET(upd);
else {
upd->size = WT_STORE_SIZE(size);
memcpy(WT_UPDATE_DATA(upd), value->data, size);
}
/*
* This must come last: after __wt_txn_modify succeeds, we must return
* a non-NULL upd so our callers can call __wt_txn_unmodify on any
* subsequent failure.
*/
if ((ret = __wt_txn_modify(session, &upd->txnid)) != 0) {
__wt_free(session, upd);
return (ret);
}
*updp = upd;
if (sizep != NULL)
*sizep = sizeof(WT_UPDATE) + size;
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_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);
}
/*
* __cursor_row_prev --
* Move to the previous row-store item.
*/
static inline int
__cursor_row_prev(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_INSERT *ins;
WT_ITEM *key, *val;
WT_PAGE *page;
WT_ROW *rip;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
page = cbt->ref->page;
key = &cbt->iface.key;
val = &cbt->iface.value;
/*
* For row-store pages, we need a single item that tells us the part
* of the page we're walking (otherwise switching from next to prev
* and vice-versa is just too complicated), so we map the WT_ROW and
* WT_INSERT_HEAD insert array slots into a single name space: slot 1
* is the "smallest key insert list", slot 2 is WT_ROW[0], slot 3 is
* WT_INSERT_HEAD[0], and so on. This means WT_INSERT lists are
* odd-numbered slots, and WT_ROW array slots are even-numbered slots.
*
* New page configuration.
*/
if (newpage) {
/*
* If we haven't instantiated keys on this page, do so, else it
* is a very, very slow traversal.
*/
if (!F_ISSET_ATOMIC(page, WT_PAGE_BUILD_KEYS))
WT_RET(__wt_row_leaf_keys(session, page));
if (page->pg_row_entries == 0)
cbt->ins_head = WT_ROW_INSERT_SMALLEST(page);
else
cbt->ins_head =
WT_ROW_INSERT_SLOT(page, page->pg_row_entries - 1);
cbt->ins = WT_SKIP_LAST(cbt->ins_head);
cbt->row_iteration_slot = page->pg_row_entries * 2 + 1;
goto new_insert;
}
/* Move to the previous entry and return the item. */
for (;;) {
/*
* Continue traversing any insert list. Maintain the reference
* to the current insert element in case we switch to a cursor
* next movement.
*/
if (cbt->ins != NULL)
WT_RET(__cursor_skip_prev(cbt));
new_insert: if ((ins = cbt->ins) != NULL) {
if ((upd = __wt_txn_read(session, ins->upd)) == NULL)
continue;
if (WT_UPDATE_DELETED_ISSET(upd)) {
if (__wt_txn_visible_all(session, upd->txnid))
++cbt->page_deleted_count;
continue;
}
key->data = WT_INSERT_KEY(ins);
key->size = WT_INSERT_KEY_SIZE(ins);
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
return (0);
}
/* Check for the beginning of the page. */
if (cbt->row_iteration_slot == 1)
return (WT_NOTFOUND);
--cbt->row_iteration_slot;
/*
* Odd-numbered slots configure as WT_INSERT_HEAD entries,
* even-numbered slots configure as WT_ROW entries.
*/
if (cbt->row_iteration_slot & 0x01) {
cbt->ins_head = cbt->row_iteration_slot == 1 ?
WT_ROW_INSERT_SMALLEST(page) :
WT_ROW_INSERT_SLOT(
page, cbt->row_iteration_slot / 2 - 1);
cbt->ins = WT_SKIP_LAST(cbt->ins_head);
goto new_insert;
}
cbt->ins_head = NULL;
cbt->ins = NULL;
cbt->slot = cbt->row_iteration_slot / 2 - 1;
rip = &page->pg_row_d[cbt->slot];
upd = __wt_txn_read(session, WT_ROW_UPDATE(page, rip));
if (upd != NULL && WT_UPDATE_DELETED_ISSET(upd)) {
if (__wt_txn_visible_all(session, upd->txnid))
++cbt->page_deleted_count;
continue;
}
return (__cursor_row_slot_return(cbt, rip, upd));
}
/* NOTREACHED */
}
/*
* __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 */
}
/*
* __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));
}
/*
* __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 */
}
/*
* __cursor_row_next --
* Move to the next row-store item.
*/
static inline int
__cursor_row_next(WT_CURSOR_BTREE *cbt, int newpage)
{
WT_INSERT *ins;
WT_ITEM *key, *val;
WT_ROW *rip;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
key = &cbt->iface.key;
val = &cbt->iface.value;
session = (WT_SESSION_IMPL *)cbt->iface.session;
/*
* For row-store pages, we need a single item that tells us the part
* of the page we're walking (otherwise switching from next to prev
* and vice-versa is just too complicated), so we map the WT_ROW and
* WT_INSERT_HEAD insert array slots into a single name space: slot 1
* is the "smallest key insert list", slot 2 is WT_ROW[0], slot 3 is
* WT_INSERT_HEAD[0], and so on. This means WT_INSERT lists are
* odd-numbered slots, and WT_ROW array slots are even-numbered slots.
*
* New page configuration.
*/
if (newpage) {
cbt->ins_head = WT_ROW_INSERT_SMALLEST(cbt->page);
cbt->ins = WT_SKIP_FIRST(cbt->ins_head);
cbt->row_iteration_slot = 1;
goto new_insert;
}
/* Move to the next entry and return the item. */
for (;;) {
/*
* Continue traversing any insert list; maintain the insert list
* head reference and entry count in case we switch to a cursor
* previous movement.
*/
if (cbt->ins != NULL)
cbt->ins = WT_SKIP_NEXT(cbt->ins);
new_insert: if ((ins = cbt->ins) != NULL) {
if ((upd = __wt_txn_read(session, ins->upd)) == NULL ||
WT_UPDATE_DELETED_ISSET(upd))
continue;
key->data = WT_INSERT_KEY(ins);
key->size = WT_INSERT_KEY_SIZE(ins);
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
return (0);
}
/* Check for the end of the page. */
if (cbt->row_iteration_slot >= cbt->page->entries * 2 + 1)
return (WT_NOTFOUND);
++cbt->row_iteration_slot;
/*
* Odd-numbered slots configure as WT_INSERT_HEAD entries,
* even-numbered slots configure as WT_ROW entries.
*/
if (cbt->row_iteration_slot & 0x01) {
cbt->ins_head = WT_ROW_INSERT_SLOT(
cbt->page, cbt->row_iteration_slot / 2 - 1);
cbt->ins = WT_SKIP_FIRST(cbt->ins_head);
goto new_insert;
}
cbt->ins_head = NULL;
cbt->ins = NULL;
cbt->slot = cbt->row_iteration_slot / 2 - 1;
rip = &cbt->page->u.row.d[cbt->slot];
upd = __wt_txn_read(session, WT_ROW_UPDATE(cbt->page, rip));
if (upd != NULL && WT_UPDATE_DELETED_ISSET(upd))
continue;
return (__cursor_row_slot_return(cbt, rip, upd));
}
/* NOTREACHED */
}
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;
}
/*
* __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 */
}
/*移向行存储的下一个行对象*/
static inline int __cursor_row_next(WT_CURSOR_BTREE* cbt, int newpage)
{
WT_INSERT *ins;
WT_ITEM *key, *val;
WT_PAGE *page;
WT_ROW *rip;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
page = cbt->ref->page;
key = &cbt->iface.key;
val = &cbt->iface.value;
/*假如是newpage，定位到insert修改队列的头位置*/
if (newpage){
cbt->ins_head = WT_ROW_INSERT_SMALLEST(page);
cbt->ins = WT_SKIP_FIRST(cbt->ins_head);
cbt->row_iteration_slot = 1;
goto new_insert;
}
for (;;){
if (cbt->ins != NULL)
cbt->ins = WT_SKIP_NEXT(cbt->ins);
new_insert:
if ((ins = cbt->ins) != NULL) {
/*事务可见数据读取*/
if ((upd = __wt_txn_read(session, ins->upd)) == NULL)
continue;
/*判断是否删除，如果删除，跳过被删除的对象*/
if (WT_UPDATE_DELETED_ISSET(upd)) {
++cbt->page_deleted_count;
continue;
}
key->data = WT_INSERT_KEY(ins);
key->size = WT_INSERT_KEY_SIZE(ins);
val->data = WT_UPDATE_DATA(upd);
val->size = upd->size;
return 0;
}
/*检索page row entires数组， 到了page的末尾*/
if (cbt->row_iteration_slot >= page->pg_row_entries * 2 + 1)
return (WT_NOTFOUND);
++cbt->row_iteration_slot;
/*
* Odd-numbered slots configure as WT_INSERT_HEAD entries,
* even-numbered slots configure as WT_ROW entries.
*/
if (cbt->row_iteration_slot & 0x01) {
cbt->ins_head = WT_ROW_INSERT_SLOT(page, cbt->row_iteration_slot / 2 - 1);
cbt->ins = WT_SKIP_FIRST(cbt->ins_head);
goto new_insert;
}
cbt->ins_head = NULL;
cbt->ins = NULL;
/*计算定位slot*/
cbt->slot = cbt->row_iteration_slot / 2 - 1;
rip = &page->pg_row_d[cbt->slot];
upd = __wt_txn_read(session, WT_ROW_UPDATE(page, rip));
if (upd != NULL && WT_UPDATE_DELETED_ISSET(upd)) {
++cbt->page_deleted_count;
continue;
}
return __cursor_row_slot_return(cbt, rip, upd);
}
}
/*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 */
}
/*
* __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);
}