/*
* __wt_btcur_iterate_setup --
* Initialize a cursor for iteration, usually based on a search.
*/
void
__wt_btcur_iterate_setup(WT_CURSOR_BTREE *cbt, int next)
{
WT_PAGE *page;
WT_UNUSED(next);
/*
* We don't currently have to do any setup when we switch between next
* and prev calls, but I'm sure we will someday -- I'm leaving support
* here for both flags for that reason.
*/
F_SET(cbt, WT_CBT_ITERATE_NEXT | WT_CBT_ITERATE_PREV);
/*
* Clear the count of deleted items on the page.
*/
cbt->page_deleted_count = 0;
/*
* If we don't have a search page, then we're done, we're starting at
* the beginning or end of the tree, not as a result of a search.
*/
if (cbt->ref == NULL)
return;
page = cbt->ref->page;
if (page->type == WT_PAGE_ROW_LEAF) {
/*
* 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.
*/
cbt->row_iteration_slot = (cbt->slot + 1) * 2;
if (cbt->ins_head != NULL) {
if (cbt->ins_head == WT_ROW_INSERT_SMALLEST(page))
cbt->row_iteration_slot = 1;
else
cbt->row_iteration_slot += 1;
}
} else {
/*
* For column-store pages, calculate the largest record on the
* page.
*/
cbt->last_standard_recno = page->type == WT_PAGE_COL_VAR ?
__col_var_last_recno(page) : __col_fix_last_recno(page);
/* If we're traversing the append list, set the reference. */
if (cbt->ins_head != NULL &&
cbt->ins_head == WT_COL_APPEND(page))
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
}
/*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, bool newpage)
{
WT_BTREE *btree;
WT_ITEM *val;
WT_PAGE *page;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
page = cbt->ref->page;
btree = S2BT(session);
val = &cbt->iface.value;
/* Initialize for each new page. */
if (newpage) {
cbt->last_standard_recno = __col_fix_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. */
if (cbt->recno == page->pg_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(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(page, cbt->recno, btree->bitcnt);
val->data = &cbt->v;
} else
val->data = WT_UPDATE_DATA(upd);
val->size = 1;
return (0);
}
/*
* __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);
}
/*为一个检索操作初始化一个查询的cursor*/
void __wt_btcur_iterate_setup(WT_CURSOR_BTREE* cbt, int next)
{
WT_PAGE *page;
WT_UNUSED(next);
/*
* We don't currently have to do any setup when we switch between next
* and prev calls, but I'm sure we will someday -- I'm leaving support
* here for both flags for that reason.
*/
F_SET(cbt, WT_CBT_ITERATE_NEXT | WT_CBT_ITERATE_PREV);
/*初始化del count计数器，在检索的过程需要统计del count*/
cbt->page_deleted_count = 0;
if (cbt->ref == NULL)
return;
page = cbt->ref->page;
/*行存储叶子节点*/
if (page->type == WT_PAGE_ROW_LEAF){
/*指定iteration slot*/
/*
* 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.
*/
cbt->row_iteration_slot = (cbt->slot + 1) * 2;
if (cbt->ins_head != NULL){
if (cbt->ins_head == WT_ROW_INSERT_SMALLEST(page))
cbt->row_iteration_slot = 1;
else
cbt->row_iteration_slot += 1;
}
}
else{ /*column store page, 计算这个page中最大的记录序号*/
cbt->last_standard_recno = page->type == WT_PAGE_COL_VAR ? __col_var_last_recno(page) : __col_fix_last_recno(page);
if (cbt->ins_head != NULL && cbt->ins_head == WT_COL_APPEND(page))
F_SET(cbt, WT_CBT_ITERATE_APPEND);
}
}
/*
* __wt_col_modify --
* Column-store delete, insert, and update.
*/
int
__wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
uint64_t recno, WT_ITEM *value, WT_UPDATE *upd, int is_remove)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_INSERT *ins;
WT_INSERT_HEAD *ins_head, **ins_headp;
WT_ITEM _value;
WT_PAGE *page;
WT_UPDATE *old_upd;
size_t ins_size, upd_size;
u_int i, skipdepth;
int append, logged;
btree = cbt->btree;
ins = NULL;
page = cbt->ref->page;
append = logged = 0;
/* This code expects a remove to have a NULL value. */
if (is_remove) {
if (btree->type == BTREE_COL_FIX) {
value = &_value;
value->data = "";
value->size = 1;
} else
value = NULL;
} else {
/*
* There's some chance the application specified a record past
* the last record on the page. If that's the case, and we're
* inserting a new WT_INSERT/WT_UPDATE pair, it goes on the
* append list, not the update list. In addition, a recno of 0
* implies an append operation, we're allocating a new row.
*/
if (recno == 0 ||
recno > (btree->type == BTREE_COL_VAR ?
__col_var_last_recno(page) : __col_fix_last_recno(page)))
append = 1;
}
/* If we don't yet have a modify structure, we'll need one. */
WT_RET(__wt_page_modify_init(session, page));
/*
* Delete, insert or update a column-store entry.
*
* If modifying a previously modified record, create a new WT_UPDATE
* entry and have a serialized function link it into an existing
* WT_INSERT entry's WT_UPDATE list.
*
* Else, allocate an insert array as necessary, build a WT_INSERT and
* WT_UPDATE structure pair, and call a serialized function to insert
* the WT_INSERT structure.
*/
if (cbt->compare == 0 && cbt->ins != NULL) {
/*
* If we are restoring updates that couldn't be evicted, the
* key must not exist on the new page.
*/
WT_ASSERT(session, upd == NULL);
/* Make sure the update can proceed. */
WT_ERR(__wt_txn_update_check(
session, old_upd = cbt->ins->upd));
/* Allocate a WT_UPDATE structure and transaction ID. */
WT_ERR(__wt_update_alloc(session, value, &upd, &upd_size));
WT_ERR(__wt_txn_modify(session, upd));
logged = 1;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
/*
* Point the new WT_UPDATE item to the next element in the list.
* If we get it right, the serialization function lock acts as
* our memory barrier to flush this write.
*/
upd->next = old_upd;
/* Serialize the update. */
WT_ERR(__wt_update_serial(
session, page, &cbt->ins->upd, &upd, upd_size));
} else {
/* Allocate the append/update list reference as necessary. */
if (append) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, page->modify->mod_append, ins_headp, 1);
ins_headp = &page->modify->mod_append[0];
} else if (page->type == WT_PAGE_COL_FIX) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, page->modify->mod_update, ins_headp, 1);
ins_headp = &page->modify->mod_update[0];
} else {
WT_PAGE_ALLOC_AND_SWAP(session,
page, page->modify->mod_update, ins_headp,
page->pg_var_entries);
ins_headp = &page->modify->mod_update[cbt->slot];
}
/* Allocate the WT_INSERT_HEAD structure as necessary. */
WT_PAGE_ALLOC_AND_SWAP(session, page, *ins_headp, ins_head, 1);
ins_head = *ins_headp;
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth(session);
/*
* Allocate a WT_INSERT/WT_UPDATE pair and transaction ID, and
* update the cursor to reference it (the WT_INSERT_HEAD might
* be allocated, the WT_INSERT was allocated).
*/
WT_ERR(__col_insert_alloc(
session, recno, skipdepth, &ins, &ins_size));
cbt->ins_head = ins_head;
cbt->ins = ins;
if (upd == NULL) {
WT_ERR(
__wt_update_alloc(session, value, &upd, &upd_size));
WT_ERR(__wt_txn_modify(session, upd));
logged = 1;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
} else
upd_size = __wt_update_list_memsize(upd);
ins->upd = upd;
ins_size += upd_size;
/*
* If there was no insert list during the search, or there was
* no search because the record number has not been allocated
* yet, the cursor's information cannot be correct, search
* couldn't have initialized it.
*
* Otherwise, point the new WT_INSERT item's skiplist to the
* next elements in the insert list (which we will check are
* still valid inside the serialization function).
*
* The serial mutex acts as our memory barrier to flush these
* writes before inserting them into the list.
*/
if (WT_SKIP_FIRST(ins_head) == NULL || recno == 0)
for (i = 0; i < skipdepth; i++) {
cbt->ins_stack[i] = &ins_head->head[i];
ins->next[i] = cbt->next_stack[i] = NULL;
}
else
for (i = 0; i < skipdepth; i++)
ins->next[i] = cbt->next_stack[i];
/* Append or insert the WT_INSERT structure. */
if (append)
WT_ERR(__wt_col_append_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, &cbt->recno, skipdepth));
else
WT_ERR(__wt_insert_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, skipdepth));
}
/* If the update was successful, add it to the in-memory log. */
if (logged)
WT_ERR(__wt_txn_log_op(session, cbt));
if (0) {
err: /*
* Remove the update from the current transaction, so we don't
* try to modify it on rollback.
*/
if (logged)
__wt_txn_unmodify(session);
__wt_free(session, ins);
__wt_free(session, upd);
}
return (ret);
}
/*
* __wt_col_modify --
* Column-store delete, insert, and update.
*/
int
__wt_col_modify(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt,
uint64_t recno, const WT_ITEM *value,
WT_UPDATE *upd_arg, u_int modify_type, bool exclusive)
{
static const WT_ITEM col_fix_remove = { "", 1, NULL, 0, 0 };
WT_BTREE *btree;
WT_DECL_RET;
WT_INSERT *ins;
WT_INSERT_HEAD *ins_head, **ins_headp;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_UPDATE *old_upd, *upd;
size_t ins_size, upd_size;
u_int i, skipdepth;
bool append, logged;
btree = cbt->btree;
ins = NULL;
page = cbt->ref->page;
upd = upd_arg;
append = logged = false;
if (upd_arg == NULL) {
if (modify_type == WT_UPDATE_RESERVE ||
modify_type == WT_UPDATE_TOMBSTONE) {
/*
* Fixed-size column-store doesn't have on-page deleted
* values, it's a nul byte.
*/
if (modify_type == WT_UPDATE_TOMBSTONE &&
btree->type == BTREE_COL_FIX) {
modify_type = WT_UPDATE_STANDARD;
value = &col_fix_remove;
}
}
/*
* There's a chance the application specified a record past the
* last record on the page. If that's the case and we're
* inserting a new WT_INSERT/WT_UPDATE pair, it goes on the
* append list, not the update list. Also, an out-of-band recno
* implies an append operation, we're allocating a new row.
* Ignore any information obtained from the search.
*/
WT_ASSERT(session, recno != WT_RECNO_OOB || cbt->compare != 0);
if (cbt->compare != 0 &&
(recno == WT_RECNO_OOB ||
recno > (btree->type == BTREE_COL_VAR ?
__col_var_last_recno(cbt->ref) :
__col_fix_last_recno(cbt->ref)))) {
append = true;
cbt->ins = NULL;
cbt->ins_head = NULL;
}
}
/* We're going to modify the page, we should have loaded history. */
WT_ASSERT(session, cbt->ref->state != WT_REF_LIMBO);
/* If we don't yet have a modify structure, we'll need one. */
WT_RET(__wt_page_modify_init(session, page));
mod = page->modify;
/*
* If modifying a record not previously modified, but which is in the
* same update slot as a previously modified record, cursor.ins will
* not be set because there's no list of update records for this recno,
* but cursor.ins_head will be set to point to the correct update slot.
* Acquire the necessary insert information, then create a new update
* entry and link it into the existing list. We get here if a page has
* a single cell representing multiple records (the records have the
* same value), and then a record in the cell is updated or removed,
* creating the update list for the cell, and then a cursor iterates
* into that same cell to update/remove a different record. We find the
* correct slot in the update array, but we don't find an update list
* (because it doesn't exist), and don't have the information we need
* to do the insert. Normally, we wouldn't care (we could fail and do
* a search for the record which would configure everything for the
* insert), but range truncation does this pattern for every record in
* the cell, and the performance is terrible. For that reason, catch it
* here.
*/
if (cbt->ins == NULL && cbt->ins_head != NULL) {
cbt->ins = __col_insert_search(
cbt->ins_head, cbt->ins_stack, cbt->next_stack, recno);
if (cbt->ins != NULL) {
if (WT_INSERT_RECNO(cbt->ins) == recno)
cbt->compare = 0;
else {
/*
* The test below is for cursor.compare set to 0
* and cursor.ins set: cursor.compare wasn't set
* by the search we just did, and has an unknown
* value. Clear cursor.ins to avoid the test.
*/
cbt->ins = NULL;
}
}
}
/*
* Delete, insert or update a column-store entry.
*
* If modifying a previously modified record, cursor.ins will be set to
* point to the correct update list. Create a new update entry and link
* it into the existing list.
*
* Else, allocate an insert array as necessary, build an insert/update
* structure pair, and link it into place.
*/
if (cbt->compare == 0 && cbt->ins != NULL) {
/*
* If we are restoring updates that couldn't be evicted, the
* key must not exist on the new page.
*/
WT_ASSERT(session, upd_arg == NULL);
/* Make sure the update can proceed. */
WT_ERR(__wt_txn_update_check(session, old_upd = cbt->ins->upd));
/* Allocate a WT_UPDATE structure and transaction ID. */
WT_ERR(__wt_update_alloc(session,
value, &upd, &upd_size, modify_type));
WT_ERR(__wt_txn_modify(session, upd));
logged = true;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
/*
* Point the new WT_UPDATE item to the next element in the list.
* If we get it right, the serialization function lock acts as
* our memory barrier to flush this write.
*/
upd->next = old_upd;
/* Serialize the update. */
WT_ERR(__wt_update_serial(
session, page, &cbt->ins->upd, &upd, upd_size, false));
} else {
/* Allocate the append/update list reference as necessary. */
if (append) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, mod->mod_col_append, ins_headp, 1);
ins_headp = &mod->mod_col_append[0];
} else if (page->type == WT_PAGE_COL_FIX) {
WT_PAGE_ALLOC_AND_SWAP(session,
page, mod->mod_col_update, ins_headp, 1);
ins_headp = &mod->mod_col_update[0];
} else {
WT_PAGE_ALLOC_AND_SWAP(session, page,
mod->mod_col_update, ins_headp, page->entries);
ins_headp = &mod->mod_col_update[cbt->slot];
}
/* Allocate the WT_INSERT_HEAD structure as necessary. */
WT_PAGE_ALLOC_AND_SWAP(session, page, *ins_headp, ins_head, 1);
ins_head = *ins_headp;
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth(session);
/*
* Allocate a WT_INSERT/WT_UPDATE pair and transaction ID, and
* update the cursor to reference it (the WT_INSERT_HEAD might
* be allocated, the WT_INSERT was allocated).
*/
WT_ERR(__col_insert_alloc(
session, recno, skipdepth, &ins, &ins_size));
cbt->ins_head = ins_head;
cbt->ins = ins;
/*
* Check for insert split and checkpoint races in column-store:
* it's easy (as opposed to in row-store) and a difficult bug to
* otherwise diagnose.
*/
WT_ASSERT(session, mod->mod_col_split_recno == WT_RECNO_OOB ||
(recno != WT_RECNO_OOB &&
mod->mod_col_split_recno > recno));
if (upd_arg == NULL) {
WT_ERR(__wt_update_alloc(session,
value, &upd, &upd_size, modify_type));
WT_ERR(__wt_txn_modify(session, upd));
logged = true;
/* Avoid a data copy in WT_CURSOR.update. */
cbt->modify_update = upd;
} else
upd_size = __wt_update_list_memsize(upd);
ins->upd = upd;
ins_size += upd_size;
/*
* If there was no insert list during the search, or there was
* no search because the record number has not been allocated
* yet, the cursor's information cannot be correct, search
* couldn't have initialized it.
*
* Otherwise, point the new WT_INSERT item's skiplist to the
* next elements in the insert list (which we will check are
* still valid inside the serialization function).
*
* The serial mutex acts as our memory barrier to flush these
* writes before inserting them into the list.
*/
if (cbt->ins_stack[0] == NULL || recno == WT_RECNO_OOB)
for (i = 0; i < skipdepth; i++) {
cbt->ins_stack[i] = &ins_head->head[i];
ins->next[i] = cbt->next_stack[i] = NULL;
}
else
for (i = 0; i < skipdepth; i++)
ins->next[i] = cbt->next_stack[i];
/* Append or insert the WT_INSERT structure. */
if (append)
WT_ERR(__wt_col_append_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, &cbt->recno, skipdepth, exclusive));
else
WT_ERR(__wt_insert_serial(
session, page, cbt->ins_head, cbt->ins_stack,
&ins, ins_size, skipdepth, exclusive));
}
/* If the update was successful, add it to the in-memory log. */
if (logged && modify_type != WT_UPDATE_RESERVE) {
WT_ERR(__wt_txn_log_op(session, cbt));
/*
* In case of append, the recno (key) for the value is assigned
* now. Set the recno in the transaction operation to be used
* incase this transaction is prepared to retrieve the update
* corresponding to this operation.
*/
__wt_txn_op_set_recno(session, cbt->recno);
}
if (0) {
err: /*
* Remove the update from the current transaction, so we don't
* try to modify it on rollback.
*/
if (logged)
__wt_txn_unmodify(session);
__wt_free(session, ins);
if (upd_arg == NULL)
__wt_free(session, upd);
}
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, 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);
}
