/*
* __cursor_truncate_fix --
* Discard a cursor range from fixed-width column-store tree.
*/
static int
__cursor_truncate_fix(WT_SESSION_IMPL *session,
WT_CURSOR_BTREE *start, WT_CURSOR_BTREE *stop,
int (*rmfunc)(WT_SESSION_IMPL *, WT_CURSOR_BTREE *, int))
{
WT_DECL_RET;
uint8_t *value;
/*
* Handle fixed-length column-store objects separately: for row-store
* and variable-length column-store objects we have "deleted" values
* and so returned objects actually exist: fixed-length column-store
* objects are filled-in if they don't exist, that is, if you create
* record 37, records 1-36 magically appear. Those records can't be
* deleted, which means we have to ignore already "deleted" records.
*
* First, call the standard cursor remove method to do a full search and
* re-position the cursor because we don't have a saved copy of the
* page's write generation information, which we need to remove records.
* Once that's done, we can delete records without a full search, unless
* we encounter a restart error because the page was modified by some
* other thread of control; in that case, repeat the full search to
* refresh the page's modification information.
*/
if (start == NULL) {
do {
WT_RET(__wt_btcur_remove(stop));
for (;;) {
if ((ret = __wt_btcur_prev(stop, 1)) != 0)
break;
stop->compare = 0; /* Exact match */
value = (uint8_t *)stop->iface.value.data;
if (*value != 0 &&
(ret = rmfunc(session, stop, 1)) != 0)
break;
}
} while (ret == WT_RESTART);
} else {
do {
WT_RET(__wt_btcur_remove(start));
for (;;) {
if (stop != NULL &&
__cursor_equals(start, stop))
break;
if ((ret = __wt_btcur_next(start, 1)) != 0)
break;
start->compare = 0; /* Exact match */
value = (uint8_t *)start->iface.value.data;
if (*value != 0 &&
(ret = rmfunc(session, start, 1)) != 0)
break;
}
} while (ret == WT_RESTART);
}
WT_RET_NOTFOUND_OK(ret);
return (0);
}
/*
* __cursor_truncate --
* Discard a cursor range from row-store or variable-width column-store
* tree.
*/
static int
__cursor_truncate(WT_SESSION_IMPL *session,
WT_CURSOR_BTREE *start, WT_CURSOR_BTREE *stop,
int (*rmfunc)(WT_SESSION_IMPL *, WT_CURSOR_BTREE *, int))
{
WT_DECL_RET;
/*
* First, call the standard cursor remove method to do a full search and
* re-position the cursor because we don't have a saved copy of the
* page's write generation information, which we need to remove records.
* Once that's done, we can delete records without a full search, unless
* we encounter a restart error because the page was modified by some
* other thread of control; in that case, repeat the full search to
* refresh the page's modification information.
*
* If this is a row-store, we delete leaf pages having no overflow items
* without reading them; for that to work, we have to ensure we read the
* page referenced by the ending cursor, since we may be deleting only a
* partial page at the end of the truncation. Our caller already fully
* instantiated the end cursor, so we know that page is pinned in memory
* and we can proceed without concern.
*/
if (start == NULL) {
do {
WT_RET(__wt_btcur_remove(stop));
for (;;) {
if ((ret = __wt_btcur_prev(stop, 1)) != 0)
break;
stop->compare = 0; /* Exact match */
if ((ret = rmfunc(session, stop, 1)) != 0)
break;
}
} while (ret == WT_RESTART);
} else {
do {
WT_RET(__wt_btcur_remove(start));
/*
* Reset ret each time through so that we don't loop
* forever in the cursor equals case.
*/
for (ret = 0;;) {
if (stop != NULL &&
__cursor_equals(start, stop))
break;
if ((ret = __wt_btcur_next(start, 1)) != 0)
break;
start->compare = 0; /* Exact match */
if ((ret = rmfunc(session, start, 1)) != 0)
break;
}
} while (ret == WT_RESTART);
}
WT_RET_NOTFOUND_OK(ret);
return (0);
}
/*
* __curfile_next --
* WT_CURSOR->next method for the btree cursor type.
*/
static int
__curfile_next(WT_CURSOR *cursor)
{
WT_CURSOR_BTREE *cbt;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cbt = (WT_CURSOR_BTREE *)cursor;
CURSOR_API_CALL(cursor, session, next, cbt->btree);
ret = __wt_btcur_next((WT_CURSOR_BTREE *)cursor, 0);
API_END(session);
return (ret);
}
/*
* __curfile_next --
* WT_CURSOR->next method for the btree cursor type.
*/
static int
__curfile_next(WT_CURSOR *cursor)
{
WT_CURSOR_BTREE *cbt;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cbt = (WT_CURSOR_BTREE *)cursor;
CURSOR_API_CALL(cursor, session, next, cbt->btree);
F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET);
if ((ret = __wt_btcur_next(cbt, 0)) == 0)
F_SET(cursor, WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT);
err: API_END_RET(session, ret);
}
/*
* __curfile_next --
* WT_CURSOR->next method for the btree cursor type.
*/
static int
__curfile_next(WT_CURSOR *cursor)
{
WT_CURSOR_BTREE *cbt;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cbt = (WT_CURSOR_BTREE *)cursor;
CURSOR_API_CALL(cursor, session, next, cbt->btree);
WT_ERR(__wt_btcur_next(cbt, false));
/* Next maintains a position, key and value. */
WT_ASSERT(session,
F_MASK(cursor, WT_CURSTD_KEY_SET) == WT_CURSTD_KEY_INT &&
F_MASK(cursor, WT_CURSTD_VALUE_SET) == WT_CURSTD_VALUE_INT);
err: API_END_RET(session, 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_WITH_PAGE_INDEX(session,
ret = __wt_row_random(session, cbt));
WT_ERR(ret);
if (__cursor_valid(cbt, &upd))
WT_ERR(__wt_kv_return(session, cbt, upd));
else {
if ((ret = __wt_btcur_next(cbt, 0)) == WT_NOTFOUND)
ret = __wt_btcur_prev(cbt, 0);
WT_ERR(ret);
}
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_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_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_search_near --
* Search for a record in the tree.
*/
int
__wt_btcur_search_near(WT_CURSOR_BTREE *cbt, int *exact)
{
WT_BTREE *btree;
WT_ITEM *val;
WT_CURSOR *cursor;
WT_SESSION_IMPL *session;
int ret;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_BSTAT_INCR(session, cursor_read_near);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
__cursor_func_init(cbt, 1);
WT_ERR(btree->type == BTREE_ROW ?
__wt_row_search(session, cbt, 0) :
__wt_col_search(session, cbt, 0));
/*
* 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, if we find a valid key (one that wasn't deleted), return it.
*
* Else, if we found a deleted key, try to move to the next key in the
* tree (bias for prefix searches). Cursor next skips deleted records,
* so we don't have to test for them again.
*
* Else if there's no larger tree key, redo the search and try and find
* an earlier record. If that fails, quit, there's no record to return.
*/
if (cbt->compare != 0 && __cursor_fix_implicit(btree, cbt)) {
cbt->v = 0;
val = &cbt->iface.value;
val->data = &cbt->v;
val->size = 1;
*exact = 0;
} else if (!__cursor_invalid(cbt)) {
*exact = cbt->compare;
ret = __wt_kv_return(session, cbt, cbt->compare == 0 ? 0 : 1);
} else if ((ret = __wt_btcur_next(cbt)) != WT_NOTFOUND)
*exact = 1;
else {
WT_ERR(btree->type == BTREE_ROW ?
__wt_row_search(session, cbt, 0) :
__wt_col_search(session, cbt, 0));
if (!__cursor_invalid(cbt)) {
*exact = cbt->compare;
ret = __wt_kv_return(
session, cbt, cbt->compare == 0 ? 0 : 1);
} else if ((ret = __wt_btcur_prev(cbt)) != WT_NOTFOUND)
*exact = -1;
}
err: __cursor_func_resolve(cbt, ret);
return (ret);
}