/*
* __wt_page_obsolete --
* Discard all obsolete updates on a row-store leaf page.
*/
void
__wt_row_leaf_obsolete(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_INSERT *ins;
WT_ROW *rip;
WT_UPDATE *upd;
uint32_t i;
/* For entries before the first on-page record... */
WT_SKIP_FOREACH(ins, WT_ROW_INSERT_SMALLEST(page))
if ((upd =
__wt_update_obsolete_check(session, ins->upd)) != NULL)
__wt_update_obsolete_free(session, page, upd);
/* For each entry on the page... */
WT_ROW_FOREACH(page, rip, i) {
if ((upd = __wt_update_obsolete_check(
session, WT_ROW_UPDATE(page, rip))) != NULL)
__wt_update_obsolete_free(session, page, upd);
WT_SKIP_FOREACH(ins, WT_ROW_INSERT(page, rip))
if ((upd = __wt_update_obsolete_check(
session, ins->upd)) != NULL)
__wt_update_obsolete_free(session, page, upd);
}
}
/*
* __txn_abort_newer_row_leaf --
* Abort updates on a row leaf page with timestamps newer than the
* rollback timestamp.
*/
static void
__txn_abort_newer_row_leaf(
WT_SESSION_IMPL *session, WT_PAGE *page, wt_timestamp_t *rollback_timestamp)
{
WT_INSERT_HEAD *insert;
WT_ROW *rip;
WT_UPDATE *upd;
uint32_t i;
/*
* Review the insert list for keys before the first entry on the disk
* page.
*/
if ((insert = WT_ROW_INSERT_SMALLEST(page)) != NULL)
__txn_abort_newer_row_skip(
session, insert, rollback_timestamp);
/*
* Review updates that belong to keys that are on the disk image,
* as well as for keys inserted since the page was read from disk.
*/
WT_ROW_FOREACH(page, rip, i) {
if ((upd = WT_ROW_UPDATE(page, rip)) != NULL)
__txn_abort_newer_update(
session, upd, rollback_timestamp);
if ((insert = WT_ROW_INSERT(page, rip)) != NULL)
__txn_abort_newer_row_skip(
session, insert, rollback_timestamp);
}
}
/*
* __ovfl_cache_row_visible --
* row-store: check for a globally visible update.
*/
static bool
__ovfl_cache_row_visible(WT_SESSION_IMPL *session, WT_PAGE *page, WT_ROW *rip)
{
WT_UPDATE *upd;
/* Check to see if there's a globally visible update. */
for (upd = WT_ROW_UPDATE(page, rip); upd != NULL; upd = upd->next)
if (__wt_txn_visible_all(session, upd->txnid))
return (true);
return (false);
}
/*
* __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_row_next --
* Move to the next row-store item.
*/
static inline int
__cursor_row_next(WT_CURSOR_BTREE *cbt, bool newpage)
{
WT_INSERT *ins;
WT_ITEM *key;
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;
/*
* 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.
*
* Initialize for each new page.
*/
if (newpage) {
cbt->ins_head = WT_ROW_INSERT_SMALLEST(page);
cbt->ins = WT_SKIP_FIRST(cbt->ins_head);
cbt->row_iteration_slot = 1;
cbt->rip_saved = NULL;
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)
continue;
if (upd->type == WT_UPDATE_DELETED) {
if (__wt_txn_upd_visible_all(session, upd))
++cbt->page_deleted_count;
continue;
}
key->data = WT_INSERT_KEY(ins);
key->size = WT_INSERT_KEY_SIZE(ins);
return (__wt_value_return(session, cbt, upd));
}
/* Check for the end of the page. */
if (cbt->row_iteration_slot >= 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(
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 = &page->pg_row[cbt->slot];
upd = __wt_txn_read(session, WT_ROW_UPDATE(page, rip));
if (upd != NULL && upd->type == WT_UPDATE_DELETED) {
if (__wt_txn_upd_visible_all(session, upd))
++cbt->page_deleted_count;
continue;
}
return (__cursor_row_slot_return(cbt, rip, upd));
}
/* 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);
}
}
/*
* __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);
}