/*
* __wt_btcur_update_check --
* Check whether an update would conflict.
*
* This can be used to replace WT_CURSOR::insert or WT_CURSOR::update, so
* they only check for conflicts without updating the tree. It is used to
* maintain snapshot isolation for transactions that span multiple chunks
* in an LSM tree.
*/
int
__wt_btcur_update_check(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cursor = &cbt->iface;
btree = cbt->btree;
session = (WT_SESSION_IMPL *)cursor->session;
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, NULL, 1));
/*
* Just check for conflicts.
*/
ret = __curfile_update_check(cbt);
break;
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART)
goto retry;
WT_TRET(__curfile_leave(cbt));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_next_random --
* Move to a random record in the tree.
*/
int
__wt_btcur_next_random(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = cbt->btree;
/*
* Only supports row-store: applications can trivially select a random
* value from a column-store, if there were any reason to do so.
*/
if (btree->type != BTREE_ROW)
WT_RET(ENOTSUP);
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
WT_RET(__cursor_func_init(cbt, 1));
WT_ERR(__wt_row_random(session, cbt));
if (__cursor_valid(cbt, &upd))
WT_ERR(__wt_kv_return(session, cbt, upd));
else
WT_ERR(__wt_btcur_search_near(cbt, 0));
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_reset --
* Invalidate the cursor position.
*/
int
__wt_btcur_reset(WT_CURSOR_BTREE *cbt)
{
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_CONN_INCR(session, cursor_reset);
WT_STAT_DATA_INCR(session, cursor_reset);
return (__cursor_reset(cbt));
}
/*
* __wt_btcur_search --
* Search for a matching record in the tree.
*/
int
__wt_btcur_search(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_search);
WT_STAT_FAST_DATA_INCR(session, cursor_search);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_func_init(cbt, 1));
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 0) :
__cursor_col_search(session, cbt));
if (cbt->compare == 0 && __cursor_valid(cbt, &upd))
ret = __wt_kv_return(session, cbt, upd);
else if (__cursor_fix_implicit(btree, cbt)) {
/*
* Creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
*/
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
} else
ret = WT_NOTFOUND;
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_next_random --
* Move to a random record in the tree. There are two algorithms, one
* where we select a record at random from the whole tree on each
* retrieval and one where we first select a record at random from the
* whole tree, and then subsequently sample forward from that location.
* The sampling approach allows us to select reasonably uniform random
* points from unbalanced trees.
*/
int
__wt_btcur_next_random(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
wt_off_t size;
uint64_t skip;
session = (WT_SESSION_IMPL *)cbt->iface.session;
btree = cbt->btree;
/*
* Only supports row-store: applications can trivially select a random
* value from a column-store, if there were any reason to do so.
*/
if (btree->type != BTREE_ROW)
WT_RET_MSG(session, ENOTSUP,
"WT_CURSOR.next_random only supported by row-store tables");
WT_STAT_CONN_INCR(session, cursor_next);
WT_STAT_DATA_INCR(session, cursor_next);
/*
* If retrieving random values without sampling, or we don't have a
* page reference, pick a roughly random leaf page in the tree.
*/
if (cbt->ref == NULL || cbt->next_random_sample_size == 0) {
/*
* Skip past the sample size of the leaf pages in the tree
* between each random key return to compensate for unbalanced
* trees.
*
* Use the underlying file size divided by its block allocation
* size as our guess of leaf pages in the file (this can be
* entirely wrong, as it depends on how many pages are in this
* particular checkpoint, how large the leaf and internal pages
* really are, and other factors). Then, divide that value by
* the configured sample size and increment the final result to
* make sure tiny files don't leave us with a skip value of 0.
*
* !!!
* Ideally, the number would be prime to avoid restart issues.
*/
if (cbt->next_random_sample_size != 0) {
WT_ERR(btree->bm->size(btree->bm, session, &size));
cbt->next_random_leaf_skip = (uint64_t)
((size / btree->allocsize) /
cbt->next_random_sample_size) + 1;
}
/*
* Choose a leaf page from the tree.
*/
WT_ERR(__cursor_func_init(cbt, true));
WT_WITH_PAGE_INDEX(
session, ret = __wt_row_random_descent(session, cbt));
WT_ERR(ret);
} else {
/*
* Read through the tree, skipping leaf pages. Be cautious about
* the skip count: if the last leaf page skipped was also the
* last leaf page in the tree, it may be set to zero on return
* with the end-of-walk condition.
*
* Pages read for data sampling aren't "useful"; don't update
* the read generation of pages already in memory, and if a page
* is read, set its generation to a low value so it is evicted
* quickly.
*/
for (skip =
cbt->next_random_leaf_skip; cbt->ref == NULL || skip > 0;)
WT_ERR(__wt_tree_walk_skip(session, &cbt->ref, &skip,
WT_READ_NO_GEN |
WT_READ_SKIP_INTL | WT_READ_WONT_NEED));
}
/*
* Select a random entry from the leaf page. If it's not valid, move to
* the next entry, if that doesn't work, move to the previous entry.
*/
WT_ERR(__wt_row_random_leaf(session, cbt));
if (__cursor_valid(cbt, &upd))
WT_ERR(__wt_kv_return(session, cbt, upd));
else {
if ((ret = __wt_btcur_next(cbt, false)) == WT_NOTFOUND)
ret = __wt_btcur_prev(cbt, false);
WT_ERR(ret);
}
return (0);
err: WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_update --
* Update a record in the tree.
*/
int
__wt_btcur_update(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_CONN_INCR(session, cursor_update);
WT_STAT_DATA_INCR(session, cursor_update);
WT_STAT_DATA_INCRV(session, cursor_update_bytes, cursor->value.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
/*
* The tree is no longer empty: eviction should pay attention to it,
* and it's no longer possible to bulk-load into it.
*/
if (btree->bulk_load_ok) {
btree->bulk_load_ok = false;
__wt_btree_evictable(session, true);
}
retry: WT_RET(__cursor_func_init(cbt, true));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ERR(__cursor_col_search(session, cbt, NULL));
/*
* If not overwriting, fail if the key doesn't exist. If we
* find an update for the key, check for conflicts. Update the
* record if it exists. Creating a record past the end of the
* tree in a fixed-length column-store implicitly fills the gap
* with empty records. Update the record in that case, the
* record exists.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE)) {
WT_ERR(__curfile_update_check(cbt));
if ((cbt->compare != 0 || !__cursor_valid(cbt, NULL)) &&
!__cursor_fix_implicit(btree, cbt))
WT_ERR(WT_NOTFOUND);
}
ret = __cursor_col_modify(session, cbt, false);
break;
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, NULL, true));
/*
* If not overwriting, check for conflicts and fail if the key
* does not exist.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE)) {
WT_ERR(__curfile_update_check(cbt));
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL))
WT_ERR(WT_NOTFOUND);
}
ret = __cursor_row_modify(session, cbt, false);
break;
}
err: if (ret == WT_RESTART) {
WT_STAT_CONN_INCR(session, cursor_restart);
WT_STAT_DATA_INCR(session, cursor_restart);
goto retry;
}
/*
* If successful, point the cursor at internal copies of the data. We
* could shuffle memory in the cursor so the key/value pair are in local
* buffer memory, but that's a data copy. We don't want to do another
* search (and we might get a different update structure if we race).
* To make this work, we add a field to the btree cursor to pass back a
* pointer to the modify function's allocated update structure.
*/
if (ret == 0)
WT_TRET(__wt_kv_return(session, cbt, cbt->modify_update));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_remove --
* Remove a record from the tree.
*/
int
__wt_btcur_remove(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_CONN_INCR(session, cursor_remove);
WT_STAT_DATA_INCR(session, cursor_remove);
WT_STAT_DATA_INCRV(session, cursor_remove_bytes, cursor->key.size);
retry: WT_RET(__cursor_func_init(cbt, true));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ERR(__cursor_col_search(session, cbt, NULL));
/*
* If we find a matching record, check whether an update would
* conflict. Do this before checking if the update is visible
* in __cursor_valid, or we can miss conflict.
*/
WT_ERR(__curfile_update_check(cbt));
/* Remove the record if it exists. */
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL)) {
if (!__cursor_fix_implicit(btree, cbt))
WT_ERR(WT_NOTFOUND);
/*
* Creating a record past the end of the tree in a
* fixed-length column-store implicitly fills the
* gap with empty records. Return success in that
* case, the record was deleted successfully.
*
* Correct the btree cursor's location: the search
* will have pointed us at the previous/next item,
* and that's not correct.
*/
cbt->recno = cursor->recno;
} else
ret = __cursor_col_modify(session, cbt, true);
break;
case BTREE_ROW:
/* Remove the record if it exists. */
WT_ERR(__cursor_row_search(session, cbt, NULL, false));
/* Check whether an update would conflict. */
WT_ERR(__curfile_update_check(cbt));
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL))
WT_ERR(WT_NOTFOUND);
ret = __cursor_row_modify(session, cbt, true);
break;
}
err: if (ret == WT_RESTART) {
WT_STAT_CONN_INCR(session, cursor_restart);
WT_STAT_DATA_INCR(session, cursor_restart);
goto retry;
}
/*
* If the cursor is configured to overwrite and the record is not
* found, that is exactly what we want.
*/
if (F_ISSET(cursor, WT_CURSTD_OVERWRITE) && ret == WT_NOTFOUND)
ret = 0;
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_insert --
* Insert a record into the tree.
*/
int
__wt_btcur_insert(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_CONN_INCR(session, cursor_insert);
WT_STAT_DATA_INCR(session, cursor_insert);
WT_STAT_DATA_INCRV(session,
cursor_insert_bytes, cursor->key.size + cursor->value.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
/*
* The tree is no longer empty: eviction should pay attention to it,
* and it's no longer possible to bulk-load into it.
*/
if (btree->bulk_load_ok) {
btree->bulk_load_ok = false;
__wt_btree_evictable(session, true);
}
retry: WT_RET(__cursor_func_init(cbt, true));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
/*
* If WT_CURSTD_APPEND is set, insert a new record (ignoring
* the application's record number). The real record number
* is assigned by the serialized append operation.
*/
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = WT_RECNO_OOB;
WT_ERR(__cursor_col_search(session, cbt, NULL));
/*
* If not overwriting, fail if the key exists. Creating a
* record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
* Fail in that case, the record exists.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
((cbt->compare == 0 && __cursor_valid(cbt, NULL)) ||
(cbt->compare != 0 && __cursor_fix_implicit(btree, cbt))))
WT_ERR(WT_DUPLICATE_KEY);
WT_ERR(__cursor_col_modify(session, cbt, false));
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = cbt->recno;
break;
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, NULL, true));
/*
* If not overwriting, fail if the key exists, else insert the
* key/value pair.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
cbt->compare == 0 && __cursor_valid(cbt, NULL))
WT_ERR(WT_DUPLICATE_KEY);
ret = __cursor_row_modify(session, cbt, false);
break;
}
err: if (ret == WT_RESTART) {
WT_STAT_CONN_INCR(session, cursor_restart);
WT_STAT_DATA_INCR(session, cursor_restart);
goto retry;
}
/* Insert doesn't maintain a position across calls, clear resources. */
if (ret == 0)
WT_TRET(__curfile_leave(cbt));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_search_near --
* Search for a record in the tree.
*/
int
__wt_btcur_search_near(WT_CURSOR_BTREE *cbt, int *exactp)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
int exact;
bool valid;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
upd = NULL; /* -Wuninitialized */
exact = 0;
WT_STAT_CONN_INCR(session, cursor_search_near);
WT_STAT_DATA_INCR(session, cursor_search_near);
/*
* If we have a row-store page pinned, search it; if we don't have a
* page pinned, or the search of the pinned page doesn't find an exact
* match, search from the root. Unlike WT_CURSOR.search, ignore pinned
* pages in the case of column-store, search-near isn't an interesting
* enough case for column-store to add the complexity needed to avoid
* the tree search.
*
* Set the "insert" flag for the btree row-store search; we may intend
* to position the cursor at the end of the tree, rather than match an
* existing record.
*/
valid = false;
if (btree->type == BTREE_ROW &&
F_ISSET(cbt, WT_CBT_ACTIVE) &&
cbt->ref->page->read_gen != WT_READGEN_OLDEST) {
__wt_txn_cursor_op(session);
WT_ERR(__cursor_row_search(session, cbt, cbt->ref, true));
/*
* Search-near is trickier than search when searching an already
* pinned page. If search returns the first or last page slots,
* discard the results and search the full tree as the neighbor
* pages might offer better matches. This test is simplistic as
* we're ignoring append lists (there may be no page slots or we
* might be legitimately positioned after the last page slot).
* Ignore those cases, it makes things too complicated.
*/
if (cbt->slot != 0 &&
cbt->slot != cbt->ref->page->pg_row_entries - 1)
valid = __cursor_valid(cbt, &upd);
}
if (!valid) {
WT_ERR(__cursor_func_init(cbt, true));
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, NULL, true) :
__cursor_col_search(session, cbt, NULL));
valid = __cursor_valid(cbt, &upd);
}
/*
* If we find a valid key, return it.
*
* Else, creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records. In this
* case, we instantiate the empty record, it's an exact match.
*
* Else, move to the next key in the tree (bias for prefix searches).
* Cursor next skips invalid rows, so we don't have to test for them
* again.
*
* Else, redo the search and move to the previous key in the tree.
* Cursor previous skips invalid rows, so we don't have to test for
* them again.
*
* If that fails, quit, there's no record to return.
*/
if (valid) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if (__cursor_fix_implicit(btree, cbt)) {
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
exact = 0;
} else if ((ret = __wt_btcur_next(cbt, false)) != WT_NOTFOUND)
exact = 1;
else {
WT_ERR(__cursor_func_init(cbt, true));
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, NULL, true) :
__cursor_col_search(session, cbt, NULL));
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if ((ret = __wt_btcur_prev(cbt, false)) != WT_NOTFOUND)
exact = -1;
}
#ifdef HAVE_DIAGNOSTIC
if (ret == 0)
WT_ERR(__wt_cursor_key_order_init(session, cbt));
#endif
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
if (exactp != NULL && (ret == 0 || ret == WT_NOTFOUND))
*exactp = exact;
return (ret);
}
/*
* __wt_btcur_search --
* Search for a matching record in the tree.
*/
int
__wt_btcur_search(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
bool valid;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
upd = NULL; /* -Wuninitialized */
WT_STAT_CONN_INCR(session, cursor_search);
WT_STAT_DATA_INCR(session, cursor_search);
/*
* If we have a page pinned, search it; if we don't have a page pinned,
* or the search of the pinned page doesn't find an exact match, search
* from the root.
*/
valid = false;
if (F_ISSET(cbt, WT_CBT_ACTIVE) &&
cbt->ref->page->read_gen != WT_READGEN_OLDEST) {
__wt_txn_cursor_op(session);
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, cbt->ref, false) :
__cursor_col_search(session, cbt, cbt->ref));
valid = cbt->compare == 0 && __cursor_valid(cbt, &upd);
}
if (!valid) {
WT_ERR(__cursor_func_init(cbt, true));
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, NULL, false) :
__cursor_col_search(session, cbt, NULL));
valid = cbt->compare == 0 && __cursor_valid(cbt, &upd);
}
if (valid)
ret = __wt_kv_return(session, cbt, upd);
else if (__cursor_fix_implicit(btree, cbt)) {
/*
* Creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
*/
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
} else
ret = WT_NOTFOUND;
#ifdef HAVE_DIAGNOSTIC
if (ret == 0)
WT_ERR(__wt_cursor_key_order_init(session, cbt));
#endif
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_prev --
* Move to the previous record in the tree.
*/
int
__wt_btcur_prev(WT_CURSOR_BTREE *cbt, bool truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
bool newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_prev);
WT_STAT_FAST_DATA_INCR(session, cursor_prev);
flags = WT_READ_PREV | WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, false));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_PREV))
__wt_btcur_iterate_setup(cbt);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the previous page, until we reach the start
* of the file.
*/
for (newpage = false;; newpage = true) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (newpage && page != NULL && page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL)
F_SET(cbt, WT_CBT_ITERATE_APPEND);
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
newpage = true;
}
if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_prev(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_prev(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_prev(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
}
/*
* If we saw a lot of deleted records on this page, or we went
* all the way through a page and only saw deleted records, try
* to evict the page when we release it. Otherwise repeatedly
* deleting from the beginning of a tree can have quadratic
* performance. Take care not to force eviction of pages that
* are genuinely empty, in new trees.
*/
if (page != NULL &&
(cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD ||
(newpage && cbt->page_deleted_count > 0)))
__wt_page_evict_soon(page);
cbt->page_deleted_count = 0;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_search_near --
* Search for a record in the tree.
*/
int
__wt_btcur_search_near(WT_CURSOR_BTREE *cbt, int *exactp)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
int exact;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
exact = 0;
WT_STAT_FAST_CONN_INCR(session, cursor_search_near);
WT_STAT_FAST_DATA_INCR(session, cursor_search_near);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_func_init(cbt, 1));
/*
* Set the "insert" flag for the btree row-store search; we may intend
* to position our cursor at the end of the tree, rather than match an
* existing record.
*/
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 1) :
__cursor_col_search(session, cbt));
/*
* If we find an valid key, return it.
*
* Else, creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records. In this
* case, we instantiate the empty record, it's an exact match.
*
* Else, move to the next key in the tree (bias for prefix searches).
* Cursor next skips invalid rows, so we don't have to test for them
* again.
*
* Else, redo the search and move to the previous key in the tree.
* Cursor previous skips invalid rows, so we don't have to test for
* them again.
*
* If that fails, quit, there's no record to return.
*/
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if (__cursor_fix_implicit(btree, cbt)) {
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
exact = 0;
} else if ((ret = __wt_btcur_next(cbt, 0)) != WT_NOTFOUND)
exact = 1;
else {
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 1) :
__cursor_col_search(session, cbt));
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if ((ret = __wt_btcur_prev(cbt, 0)) != WT_NOTFOUND)
exact = -1;
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
if (exactp != NULL && (ret == 0 || ret == WT_NOTFOUND))
*exactp = exact;
return (ret);
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, bool truncating)
{
WT_CURSOR *cursor;
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
bool newpage;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_CONN_INCR(session, cursor_next);
WT_STAT_DATA_INCR(session, cursor_next);
F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET);
WT_RET(__cursor_func_init(cbt, false));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
flags = WT_READ_SKIP_INTL; /* tree walk flags */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
for (newpage = false;; newpage = true) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* Column-store pages may have appended entries. Handle
* it separately from the usual cursor code, it's in a
* simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*
* If we saw a lot of deleted records on this page, or we went
* all the way through a page and only saw deleted records, try
* to evict the page when we release it. Otherwise repeatedly
* deleting from the beginning of a tree can have quadratic
* performance. Take care not to force eviction of pages that
* are genuinely empty, in new trees.
*/
if (page != NULL &&
(cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD ||
(newpage && cbt->page_deleted_count > 0)))
__wt_page_evict_soon(session, cbt->ref);
cbt->page_deleted_count = 0;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
#ifdef HAVE_DIAGNOSTIC
if (ret == 0)
WT_ERR(__wt_cursor_key_order_check(session, cbt, true));
#endif
if (ret == 0)
F_SET(cursor, WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT);
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*将btree cursor移动到下一个记录*/
int __wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
/*btree 扫描标示*/
flags = WT_READ_SKIP_INTL;
if (truncating)
LF_SET(WT_READ_TRUNCATE);
/*激活一个btree cursor*/
WT_RET(__cursor_func_init(cbt, 0));
/*初始化cursor*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*对btree的扫描*/
for (;;){
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
/*column store append方式，在insert header上做扫描*/
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)){
switch (page->type){
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
/*清除掉column store的标记*/
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
}
else if (page != NULL){
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/*找到对应的记录了，直接返回*/
if (ret != WT_NOTFOUND)
break;
/*假如是column store方式，检查是否要扫描insert header list*/
if (page->type != WT_PAGE_ROW_LEAF && (cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*删除的记录太多，对page进行重组，增大page的填充因子*/
if (page != NULL && (cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD || (newpage && cbt->page_deleted_count > 0))){
__wt_page_evict_soon(page);
}
cbt->page_deleted_count = 0;
/*btree cursor跳转到下一个page上*/
WT_ERR(__wt_tree_walk(session, &cbt->ref, NULL, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err:
if (ret != 0) /*失败了，恢复cursor的状态*/
WT_TRET(__cursor_reset(cbt));
return ret;
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int skipped, newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
flags = WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, 0));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
for (skipped = newpage = 0;; skipped = 0, newpage = 1) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(
cbt, newpage, &skipped);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage, &skipped);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage, &skipped);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*
* If we scanned all the way through a page and only saw
* deleted records, try to evict the page as we release it.
* Otherwise repeatedly deleting from the beginning of a tree
* can have quadratic performance.
*/
if (newpage && skipped)
page->read_gen = WT_READGEN_OLDEST;
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}