/*
* __wt_schema_get_index --
* Find an index by URI.
*/
int
__wt_schema_get_index(WT_SESSION_IMPL *session,
const char *uri, bool quiet, WT_TABLE **tablep, WT_INDEX **indexp)
{
WT_DECL_RET;
WT_INDEX *idx;
WT_TABLE *table;
const char *tablename, *tend;
u_int i;
*indexp = NULL;
tablename = uri;
if (!WT_PREFIX_SKIP(tablename, "index:") ||
(tend = strchr(tablename, ':')) == NULL)
return (__wt_bad_object_type(session, uri));
WT_RET(__wt_schema_get_table(session,
tablename, WT_PTRDIFF(tend, tablename), false, &table));
/* Try to find the index in the table. */
for (i = 0; i < table->nindices; i++) {
idx = table->indices[i];
if (idx != NULL && strcmp(idx->name, uri) == 0) {
if (tablep != NULL)
*tablep = table;
else
__wt_schema_release_table(session, table);
*indexp = idx;
return (0);
}
}
/* Otherwise, open it. */
WT_ERR(__wt_schema_open_index(
session, table, tend + 1, strlen(tend + 1), indexp));
if (tablep != NULL)
*tablep = table;
err: __wt_schema_release_table(session, table);
WT_RET(ret);
if (*indexp != NULL)
return (0);
if (quiet)
WT_RET(ENOENT);
WT_RET_MSG(session, ENOENT, "%s not found in table", uri);
}
/*
* __truncate_table --
* WT_SESSION::truncate for a table.
*/
static int
__truncate_table(WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_DECL_RET;
WT_TABLE *table;
u_int i;
WT_RET(__wt_schema_get_table(
session, uri, strlen(uri), false, 0, &table));
WT_STAT_DATA_INCR(session, cursor_truncate);
/* Truncate the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++)
WT_ERR(__wt_schema_truncate(
session, table->cgroups[i]->source, cfg));
/* Truncate the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++)
WT_ERR(__wt_schema_truncate(
session, table->indices[i]->source, cfg));
err: WT_TRET(__wt_schema_release_table(session, table));
return (ret);
}
/*
* __drop_table --
* WT_SESSION::drop for a table.
*/
static int
__drop_table(WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_COLGROUP *colgroup;
WT_DECL_RET;
WT_INDEX *idx;
WT_TABLE *table;
const char *name;
u_int i;
name = uri;
(void)WT_PREFIX_SKIP(name, "table:");
table = NULL;
WT_ERR(__wt_schema_get_table(
session, name, strlen(name), true, &table));
/* Drop the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
if ((colgroup = table->cgroups[i]) == NULL)
continue;
/*
* Drop the column group before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, colgroup->source, cfg));
WT_ERR(__wt_metadata_remove(session, colgroup->name));
}
/* Drop the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
if ((idx = table->indices[i]) == NULL)
continue;
/*
* Drop the column group before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, idx->source, cfg));
WT_ERR(__wt_metadata_remove(session, idx->name));
}
WT_ERR(__wt_schema_remove_table(session, table));
table = NULL;
/* Remove the metadata entry (ignore missing items). */
WT_ERR(__wt_metadata_remove(session, uri));
err:
if (table != NULL)
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __curjoin_close --
* WT_CURSOR::close for join cursors.
*/
static int
__curjoin_close(WT_CURSOR *cursor)
{
WT_CURSOR_JOIN *cjoin;
WT_CURSOR_JOIN_ENDPOINT *end;
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
cjoin = (WT_CURSOR_JOIN *)cursor;
JOINABLE_CURSOR_API_CALL_PREPARE_ALLOWED(cursor, session, close, NULL);
err:
WT_TRET(__wt_schema_release_table(session, &cjoin->table));
/* This is owned by the table */
cursor->key_format = NULL;
if (cjoin->projection != NULL) {
__wt_free(session, cjoin->projection);
__wt_free(session, cursor->value_format);
}
for (entry = cjoin->entries, i = 0; i < cjoin->entries_next;
entry++, i++) {
if (entry->subjoin != NULL) {
F_CLR(&entry->subjoin->iface, WT_CURSTD_JOINED);
entry->subjoin->parent = NULL;
}
if (entry->main != NULL)
WT_TRET(entry->main->close(entry->main));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_OWN_BLOOM))
WT_TRET(__wt_bloom_close(entry->bloom));
for (end = &entry->ends[0];
end < &entry->ends[entry->ends_next]; end++) {
F_CLR(end->cursor, WT_CURSTD_JOINED);
if (F_ISSET(end, WT_CURJOIN_END_OWN_CURSOR))
WT_TRET(end->cursor->close(end->cursor));
}
__wt_free(session, entry->ends);
__wt_free(session, entry->repack_format);
}
if (cjoin->iter != NULL)
WT_TRET(__curjoin_iter_close_all(cjoin->iter));
if (cjoin->main != NULL)
WT_TRET(cjoin->main->close(cjoin->main));
__wt_free(session, cjoin->entries);
__wt_cursor_close(cursor);
API_END_RET(session, ret);
}
/*
* __wt_schema_get_colgroup --
* Find a column group by URI.
*/
int
__wt_schema_get_colgroup(WT_SESSION_IMPL *session,
const char *uri, bool quiet, WT_TABLE **tablep, WT_COLGROUP **colgroupp)
{
WT_COLGROUP *colgroup;
WT_TABLE *table;
const char *tablename, *tend;
u_int i;
*colgroupp = NULL;
tablename = uri;
if (!WT_PREFIX_SKIP(tablename, "colgroup:"))
return (__wt_bad_object_type(session, uri));
if ((tend = strchr(tablename, ':')) == NULL)
tend = tablename + strlen(tablename);
WT_RET(__wt_schema_get_table(session,
tablename, WT_PTRDIFF(tend, tablename), false, &table));
for (i = 0; i < WT_COLGROUPS(table); i++) {
colgroup = table->cgroups[i];
if (strcmp(colgroup->name, uri) == 0) {
*colgroupp = colgroup;
if (tablep != NULL)
*tablep = table;
else
__wt_schema_release_table(session, table);
return (0);
}
}
__wt_schema_release_table(session, table);
if (quiet)
WT_RET(ENOENT);
WT_RET_MSG(session, ENOENT, "%s not found in table", uri);
}
/*
* __truncate_table --
* WT_SESSION::truncate for a table.
*/
static int
__truncate_table(WT_SESSION_IMPL *session, const char *name)
{
WT_BTREE *btree;
WT_DECL_ITEM(namebuf);
WT_DECL_RET;
WT_TABLE *table;
u_int i;
WT_RET(__wt_scr_alloc(session, 0, &namebuf));
WT_ERR(__wt_schema_get_table(session, name, strlen(name), 0, &table));
/* Truncate the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
/*
* Get an exclusive lock on the handle: it will be released by
* __wt_conn_btree_close_all.
*/
WT_ERR(__wt_session_get_btree(session,
table->cgroups[i]->source, NULL, NULL, WT_BTREE_EXCLUSIVE));
btree = session->btree;
WT_ERR(__wt_buf_set(
session, namebuf, btree->name, strlen(btree->name) + 1));
WT_ERR(__truncate_file(session, namebuf->data));
}
/* Truncate the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
/*
* Get an exclusive lock on the handle: it will be released by
* __wt_conn_btree_close_all.
*/
WT_ERR(__wt_session_get_btree(session,
table->indices[i]->source, NULL, NULL, WT_BTREE_EXCLUSIVE));
btree = session->btree;
WT_ERR(__wt_buf_set(
session, namebuf, btree->name, strlen(btree->name) + 1));
WT_ERR(__truncate_file(session, namebuf->data));
}
err: __wt_scr_free(&namebuf);
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __curjoin_close --
* WT_CURSOR::close for join cursors.
*/
static int
__curjoin_close(WT_CURSOR *cursor)
{
WT_CURSOR_JOIN *cjoin;
WT_CURSOR_JOIN_ENDPOINT *end;
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
cjoin = (WT_CURSOR_JOIN *)cursor;
CURSOR_API_CALL(cursor, session, close, NULL);
__wt_schema_release_table(session, cjoin->table);
/* These are owned by the table */
cursor->internal_uri = NULL;
cursor->key_format = NULL;
if (cjoin->projection != NULL) {
__wt_free(session, cjoin->projection);
__wt_free(session, cursor->value_format);
}
for (entry = cjoin->entries, i = 0; i < cjoin->entries_next;
entry++, i++) {
if (entry->main != NULL)
WT_TRET(entry->main->close(entry->main));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_OWN_BLOOM))
WT_TRET(__wt_bloom_close(entry->bloom));
for (end = &entry->ends[0];
end < &entry->ends[entry->ends_next]; end++) {
F_CLR(end->cursor, WT_CURSTD_JOINED);
if (F_ISSET(end, WT_CURJOIN_END_OWN_KEY))
__wt_free(session, end->key.data);
}
__wt_free(session, entry->ends);
}
if (cjoin->iter != NULL)
WT_TRET(__curjoin_entry_iter_close(cjoin->iter));
__wt_free(session, cjoin->entries);
WT_TRET(__wt_cursor_close(cursor));
err: API_END_RET(session, ret);
}
/*
* __curindex_close --
* WT_CURSOR->close method for index cursors.
*/
static int
__curindex_close(WT_CURSOR *cursor)
{
WT_CURSOR_INDEX *cindex;
WT_CURSOR **cp;
WT_DECL_RET;
WT_INDEX *idx;
WT_SESSION_IMPL *session;
u_int i;
cindex = (WT_CURSOR_INDEX *)cursor;
idx = cindex->index;
JOINABLE_CURSOR_API_CALL(cursor, session, close, NULL);
if ((cp = cindex->cg_cursors) != NULL)
for (i = 0, cp = cindex->cg_cursors;
i < WT_COLGROUPS(cindex->table); i++, cp++)
if (*cp != NULL) {
WT_TRET((*cp)->close(*cp));
*cp = NULL;
}
__wt_free(session, cindex->cg_needvalue);
__wt_free(session, cindex->cg_cursors);
if (cindex->key_plan != idx->key_plan)
__wt_free(session, cindex->key_plan);
if (cursor->value_format != cindex->table->value_format)
__wt_free(session, cursor->value_format);
if (cindex->value_plan != idx->value_plan)
__wt_free(session, cindex->value_plan);
if (cindex->child != NULL)
WT_TRET(cindex->child->close(cindex->child));
__wt_schema_release_table(session, cindex->table);
/* The URI is owned by the index. */
cursor->internal_uri = NULL;
WT_TRET(__wt_cursor_close(cursor));
err: API_END_RET(session, ret);
}
/*
* __rename_table --
* WT_SESSION::rename for a table.
*/
static int
__rename_table(WT_SESSION_IMPL *session,
const char *uri, const char *newuri, const char *cfg[])
{
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_TABLE *table;
u_int i;
const char *oldname, *value;
oldname = uri;
(void)WT_PREFIX_SKIP(oldname, "table:");
WT_RET(__wt_schema_get_table(
session, oldname, strlen(oldname), 0, &table));
/* Rename the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++)
WT_ERR(__rename_tree(session, table, newuri,
table->cgroups[i]->name, cfg));
/* Rename the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++)
WT_ERR(__rename_tree(session, table, newuri,
table->indices[i]->name, cfg));
__wt_schema_remove_table(session, table);
table = NULL;
/* Rename the table. */
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_metadata_read(session, uri, &value));
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, value));
err: __wt_scr_free(&buf);
if (table != NULL)
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __drop_table --
* WT_SESSION::drop for a table.
*/
static int
__drop_table(
WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_COLGROUP *colgroup;
WT_DECL_RET;
WT_INDEX *idx;
WT_TABLE *table;
u_int i;
const char *name;
bool tracked;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_TABLE_WRITE));
name = uri;
WT_PREFIX_SKIP_REQUIRED(session, name, "table:");
table = NULL;
tracked = false;
/*
* Open the table so we can drop its column groups and indexes.
*
* Ideally we would keep the table locked exclusive across the drop,
* but for now we rely on the global table lock to prevent the table
* being reopened while it is being dropped. One issue is that the
* WT_WITHOUT_LOCKS macro can drop and reacquire the global table lock,
* avoiding deadlocks while waiting for LSM operation to quiesce.
*
* Temporarily getting the table exclusively serves the purpose
* of ensuring that cursors on the table that are already open
* must at least be closed before this call proceeds.
*/
WT_ERR(__wt_schema_get_table_uri(session, uri, true,
WT_DHANDLE_EXCLUSIVE, &table));
WT_ERR(__wt_schema_release_table(session, table));
WT_ERR(__wt_schema_get_table_uri(session, uri, true, 0, &table));
/* Drop the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
if ((colgroup = table->cgroups[i]) == NULL)
continue;
/*
* Drop the column group before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, colgroup->source, cfg));
WT_ERR(__wt_metadata_remove(session, colgroup->name));
}
/* Drop the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
if ((idx = table->indices[i]) == NULL)
continue;
/*
* Drop the index before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, idx->source, cfg));
WT_ERR(__wt_metadata_remove(session, idx->name));
}
/* Make sure the table data handle is closed. */
WT_TRET(__wt_schema_release_table(session, table));
WT_ERR(__wt_schema_get_table_uri(
session, uri, true, WT_DHANDLE_EXCLUSIVE, &table));
F_SET(&table->iface, WT_DHANDLE_DISCARD);
if (WT_META_TRACKING(session)) {
WT_WITH_DHANDLE(session, &table->iface,
ret = __wt_meta_track_handle_lock(session, false));
WT_ERR(ret);
tracked = true;
}
/* Remove the metadata entry (ignore missing items). */
WT_ERR(__wt_metadata_remove(session, uri));
err: if (table != NULL && !tracked)
WT_TRET(__wt_schema_release_table(session, table));
return (ret);
}
/*
* __create_index --
* Create an index.
*/
static int
__create_index(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG pkcols;
WT_CONFIG_ITEM ckey, cval, icols;
WT_DECL_RET;
WT_ITEM confbuf, extra_cols, fmt, namebuf;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, index_meta), NULL, NULL, NULL };
const char *sourcecfg[] = { config, NULL, NULL };
const char *sourceconf, *source, *idxconf, *idxname;
const char *tablename;
size_t tlen;
u_int i;
idxconf = sourceconf = NULL;
WT_CLEAR(confbuf);
WT_CLEAR(fmt);
WT_CLEAR(extra_cols);
WT_CLEAR(namebuf);
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "index:"))
return (EINVAL);
idxname = strchr(tablename, ':');
if (idxname == NULL)
WT_RET_MSG(session, EINVAL, "Invalid index name, "
"should be <table name>:<index name>: %s", name);
tlen = (size_t)(idxname++ - tablename);
if ((ret =
__wt_schema_get_table(session, tablename, tlen, 1, &table)) != 0)
WT_RET_MSG(session, ret,
"Can't create an index for a non-existent table: %.*s",
(int)tlen, tablename);
if (__wt_config_getones(session, config, "source", &cval) == 0) {
WT_ERR(__wt_buf_fmt(session, &namebuf,
"%.*s", (int)cval.len, cval.str));
source = namebuf.data;
} else {
WT_ERR(__wt_schema_index_source(
session, table, idxname, config, &namebuf));
source = namebuf.data;
/* Add the source name to the index config before collapsing. */
WT_ERR(__wt_buf_catfmt(session, &confbuf,
",source=\"%s\"", source));
}
/* Calculate the key/value formats. */
if (__wt_config_getones(session, config, "columns", &icols) != 0)
WT_ERR_MSG(session, EINVAL,
"No 'columns' configuration for '%s'", name);
/*
* The key format for an index is somewhat subtle: the application
* specifies a set of columns that it will use for the key, but the
* engine usually adds some hidden columns in order to derive the
* primary key. These hidden columns are part of the source's
* key_format, which we are calculating now, but not part of an index
* cursor's key_format.
*/
WT_ERR(__wt_config_subinit(session, &pkcols, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&pkcols, &ckey, &cval)) == 0;
i++) {
/*
* If the primary key column is already in the secondary key,
* don't add it again.
*/
if (__wt_config_subgetraw(session, &icols, &ckey, &cval) == 0)
continue;
WT_ERR(__wt_buf_catfmt(
session, &extra_cols, "%.*s,", (int)ckey.len, ckey.str));
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
/*
* Index values are normally empty: all columns are packed into the
* index key. The exception is LSM, which (currently) reserves empty
* values as tombstones. Use a single padding byte in that case.
*/
if (WT_PREFIX_MATCH(source, "lsm:"))
WT_ERR(__wt_buf_fmt(session, &fmt, "value_format=x,"));
else
WT_ERR(__wt_buf_fmt(session, &fmt, "value_format=,"));
WT_ERR(__wt_buf_fmt(session, &fmt, "value_format=,key_format="));
WT_ERR(__wt_struct_reformat(session, table,
icols.str, icols.len, (const char *)extra_cols.data, 0, &fmt));
/* Check for a record number index key, which makes no sense. */
WT_ERR(__wt_config_getones(session, fmt.data, "key_format", &cval));
if (cval.len == 1 && cval.str[0] == 'r')
WT_ERR_MSG(session, EINVAL,
"column-store index may not use the record number as its "
"index key");
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_concat(session, sourcecfg, &sourceconf));
WT_ERR(__wt_schema_create(session, source, sourceconf));
cfg[1] = sourceconf;
cfg[2] = confbuf.data;
WT_ERR(__wt_config_collapse(session, cfg, &idxconf));
if ((ret = __wt_metadata_insert(session, name, idxconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
err: __wt_free(session, idxconf);
__wt_free(session, sourceconf);
__wt_buf_free(session, &confbuf);
__wt_buf_free(session, &extra_cols);
__wt_buf_free(session, &fmt);
__wt_buf_free(session, &namebuf);
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __create_index --
* Create an index.
*/
static int
__create_index(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG kcols, pkcols;
WT_CONFIG_ITEM ckey, cval, icols, kval;
WT_DECL_PACK_VALUE(pv);
WT_DECL_RET;
WT_ITEM confbuf, extra_cols, fmt, namebuf;
WT_PACK pack;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, index_meta), NULL, NULL, NULL };
const char *sourcecfg[] = { config, NULL, NULL };
const char *source, *sourceconf, *idxname, *tablename;
char *idxconf;
size_t tlen;
int have_extractor;
u_int i, npublic_cols;
sourceconf = NULL;
idxconf = NULL;
WT_CLEAR(confbuf);
WT_CLEAR(fmt);
WT_CLEAR(extra_cols);
WT_CLEAR(namebuf);
have_extractor = 0;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "index:"))
return (EINVAL);
idxname = strchr(tablename, ':');
if (idxname == NULL)
WT_RET_MSG(session, EINVAL, "Invalid index name, "
"should be <table name>:<index name>: %s", name);
tlen = (size_t)(idxname++ - tablename);
if ((ret =
__wt_schema_get_table(session, tablename, tlen, 1, &table)) != 0)
WT_RET_MSG(session, ret,
"Can't create an index for a non-existent table: %.*s",
(int)tlen, tablename);
if (table->is_simple)
WT_RET_MSG(session, EINVAL,
"%s requires a table with named columns", name);
if (__wt_config_getones(session, config, "source", &cval) == 0) {
WT_ERR(__wt_buf_fmt(session, &namebuf,
"%.*s", (int)cval.len, cval.str));
source = namebuf.data;
} else {
WT_ERR(__wt_schema_index_source(
session, table, idxname, config, &namebuf));
source = namebuf.data;
/* Add the source name to the index config before collapsing. */
WT_ERR(__wt_buf_catfmt(session, &confbuf,
",source=\"%s\"", source));
}
if (__wt_config_getones_none(
session, config, "extractor", &cval) == 0 && cval.len != 0) {
have_extractor = 1;
/* Custom extractors must supply a key format. */
if ((ret = __wt_config_getones(
session, config, "key_format", &kval)) != 0)
WT_ERR_MSG(session, EINVAL,
"%s: custom extractors require a key_format", name);
}
/* Calculate the key/value formats. */
WT_CLEAR(icols);
if (__wt_config_getones(session, config, "columns", &icols) != 0 &&
!have_extractor)
WT_ERR_MSG(session, EINVAL,
"%s: requires 'columns' configuration", name);
/*
* Count the public columns using the declared columns for normal
* indices or the key format for custom extractors.
*/
npublic_cols = 0;
if (!have_extractor) {
WT_ERR(__wt_config_subinit(session, &kcols, &icols));
while ((ret = __wt_config_next(&kcols, &ckey, &cval)) == 0)
++npublic_cols;
WT_ERR_NOTFOUND_OK(ret);
} else {
WT_ERR(__pack_initn(session, &pack, kval.str, kval.len));
while ((ret = __pack_next(&pack, &pv)) == 0)
++npublic_cols;
WT_ERR_NOTFOUND_OK(ret);
}
/*
* The key format for an index is somewhat subtle: the application
* specifies a set of columns that it will use for the key, but the
* engine usually adds some hidden columns in order to derive the
* primary key. These hidden columns are part of the source's
* key_format, which we are calculating now, but not part of an index
* cursor's key_format.
*/
WT_ERR(__wt_config_subinit(session, &pkcols, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&pkcols, &ckey, &cval)) == 0;
i++) {
/*
* If the primary key column is already in the secondary key,
* don't add it again.
*/
if (__wt_config_subgetraw(session, &icols, &ckey, &cval) == 0) {
if (have_extractor)
WT_ERR_MSG(session, EINVAL,
"an index with a custom extractor may not "
"include primary key columns");
continue;
}
WT_ERR(__wt_buf_catfmt(
session, &extra_cols, "%.*s,", (int)ckey.len, ckey.str));
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
/* Index values are empty: all columns are packed into the index key. */
WT_ERR(__wt_buf_fmt(session, &fmt, "value_format=,key_format="));
if (have_extractor) {
WT_ERR(__wt_buf_catfmt(session, &fmt, "%.*s",
(int)kval.len, kval.str));
WT_CLEAR(icols);
}
/*
* Construct the index key format, or append the primary key columns
* for custom extractors.
*/
WT_ERR(__wt_struct_reformat(session, table,
icols.str, icols.len, (const char *)extra_cols.data, 0, &fmt));
/* Check for a record number index key, which makes no sense. */
WT_ERR(__wt_config_getones(session, fmt.data, "key_format", &cval));
if (cval.len == 1 && cval.str[0] == 'r')
WT_ERR_MSG(session, EINVAL,
"column-store index may not use the record number as its "
"index key");
WT_ERR(__wt_buf_catfmt(
session, &fmt, ",index_key_columns=%u", npublic_cols));
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_merge(session, sourcecfg, NULL, &sourceconf));
WT_ERR(__wt_schema_create(session, source, sourceconf));
cfg[1] = sourceconf;
cfg[2] = confbuf.data;
WT_ERR(__wt_config_collapse(session, cfg, &idxconf));
if ((ret = __wt_metadata_insert(session, name, idxconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
/* Make sure that the configuration is valid. */
WT_ERR(__wt_schema_open_index(
session, table, idxname, strlen(idxname), NULL));
err: __wt_free(session, idxconf);
__wt_free(session, sourceconf);
__wt_buf_free(session, &confbuf);
__wt_buf_free(session, &extra_cols);
__wt_buf_free(session, &fmt);
__wt_buf_free(session, &namebuf);
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __create_table --
* Create a table.
*/
static int
__create_table(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG conf;
WT_CONFIG_ITEM cgkey, cgval, cval;
WT_DECL_RET;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, table_meta), config, NULL, NULL };
const char *tablename;
char *tableconf, *cgname;
size_t cgsize;
int ncolgroups;
cgname = NULL;
table = NULL;
tableconf = NULL;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "table:"))
return (EINVAL);
if ((ret = __wt_schema_get_table(session,
tablename, strlen(tablename), 0, &table)) == 0) {
__wt_schema_release_table(session, table);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_ERR(__wt_config_gets(session, cfg, "colgroups", &cval));
WT_ERR(__wt_config_subinit(session, &conf, &cval));
for (ncolgroups = 0;
(ret = __wt_config_next(&conf, &cgkey, &cgval)) == 0;
ncolgroups++)
;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_config_collapse(session, cfg, &tableconf));
if ((ret = __wt_metadata_insert(session, name, tableconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
/* Attempt to open the table now to catch any errors. */
WT_ERR(__wt_schema_get_table(
session, tablename, strlen(tablename), 1, &table));
if (ncolgroups == 0) {
cgsize = strlen("colgroup:") + strlen(tablename) + 1;
WT_ERR(__wt_calloc_def(session, cgsize, &cgname));
snprintf(cgname, cgsize, "colgroup:%s", tablename);
WT_ERR(__create_colgroup(session, cgname, exclusive, config));
}
if (0) {
err: if (table != NULL) {
WT_TRET(__wt_schema_remove_table(session, table));
table = NULL;
}
}
if (table != NULL)
__wt_schema_release_table(session, table);
__wt_free(session, cgname);
__wt_free(session, tableconf);
return (ret);
}
/*
* __create_colgroup --
* Create a column group.
*/
static int
__create_colgroup(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
WT_ITEM confbuf, fmt, namebuf;
WT_TABLE *table;
size_t tlen;
const char **cfgp, *cfg[4] =
{ WT_CONFIG_BASE(session, colgroup_meta), config, NULL, NULL };
const char *sourcecfg[] = { config, NULL, NULL };
const char *cgname, *source, *sourceconf, *tablename;
char *cgconf, *oldconf;
sourceconf = NULL;
cgconf = oldconf = NULL;
WT_CLEAR(fmt);
WT_CLEAR(confbuf);
WT_CLEAR(namebuf);
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "colgroup:"))
return (EINVAL);
cgname = strchr(tablename, ':');
if (cgname != NULL) {
tlen = (size_t)(cgname - tablename);
++cgname;
} else
tlen = strlen(tablename);
if ((ret =
__wt_schema_get_table(session, tablename, tlen, 1, &table)) != 0)
WT_RET_MSG(session, (ret == WT_NOTFOUND) ? ENOENT : ret,
"Can't create '%s' for non-existent table '%.*s'",
name, (int)tlen, tablename);
/* Make sure the column group is referenced from the table. */
if (cgname != NULL && (ret =
__wt_config_subgets(session, &table->cgconf, cgname, &cval)) != 0)
WT_ERR_MSG(session, EINVAL,
"Column group '%s' not found in table '%.*s'",
cgname, (int)tlen, tablename);
/* Find the first NULL entry in the cfg stack. */
for (cfgp = &cfg[1]; *cfgp; cfgp++)
;
/* Add the source to the colgroup config before collapsing. */
if (__wt_config_getones(
session, config, "source", &cval) == 0 && cval.len != 0) {
WT_ERR(__wt_buf_fmt(
session, &namebuf, "%.*s", (int)cval.len, cval.str));
source = namebuf.data;
} else {
WT_ERR(__wt_schema_colgroup_source(
session, table, cgname, config, &namebuf));
source = namebuf.data;
WT_ERR(__wt_buf_fmt(
session, &confbuf, "source=\"%s\"", source));
*cfgp++ = confbuf.data;
}
/* Calculate the key/value formats: these go into the source config. */
WT_ERR(__wt_buf_fmt(session, &fmt, "key_format=%s", table->key_format));
if (cgname == NULL)
WT_ERR(__wt_buf_catfmt
(session, &fmt, ",value_format=%s", table->value_format));
else {
if (__wt_config_getones(session, config, "columns", &cval) != 0)
WT_ERR_MSG(session, EINVAL,
"No 'columns' configuration for '%s'", name);
WT_ERR(__wt_buf_catfmt(session, &fmt, ",value_format="));
WT_ERR(__wt_struct_reformat(session,
table, cval.str, cval.len, NULL, 1, &fmt));
}
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_merge(session, sourcecfg, NULL, &sourceconf));
WT_ERR(__wt_schema_create(session, source, sourceconf));
WT_ERR(__wt_config_collapse(session, cfg, &cgconf));
if ((ret = __wt_metadata_insert(session, name, cgconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
WT_ERR(__wt_schema_open_colgroups(session, table));
err: __wt_free(session, cgconf);
__wt_free(session, sourceconf);
__wt_free(session, oldconf);
__wt_buf_free(session, &confbuf);
__wt_buf_free(session, &fmt);
__wt_buf_free(session, &namebuf);
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __wt_schema_worker --
* Get Btree handles for the object and cycle through calls to an
* underlying worker function with each handle.
*/
int
__wt_schema_worker(WT_SESSION_IMPL *session,
const char *uri,
int (*file_func)(WT_SESSION_IMPL *, const char *[]),
int (*name_func)(WT_SESSION_IMPL *, const char *, bool *),
const char *cfg[], uint32_t open_flags)
{
WT_COLGROUP *colgroup;
WT_DATA_SOURCE *dsrc;
WT_DECL_RET;
WT_INDEX *idx;
WT_SESSION *wt_session;
WT_TABLE *table;
u_int i;
bool skip;
table = NULL;
skip = false;
if (name_func != NULL)
WT_ERR(name_func(session, uri, &skip));
/* If the callback said to skip this object, we're done. */
if (skip)
return (0);
/* Get the btree handle(s) and call the underlying function. */
if (WT_PREFIX_MATCH(uri, "file:")) {
if (file_func != NULL)
WT_ERR(__wt_exclusive_handle_operation(session,
uri, file_func, cfg, open_flags));
} else if (WT_PREFIX_MATCH(uri, "colgroup:")) {
WT_ERR(__wt_schema_get_colgroup(
session, uri, false, NULL, &colgroup));
WT_ERR(__wt_schema_worker(session,
colgroup->source, file_func, name_func, cfg, open_flags));
} else if (WT_PREFIX_MATCH(uri, "index:")) {
idx = NULL;
WT_ERR(__wt_schema_get_index(session, uri, false, false, &idx));
WT_ERR(__wt_schema_worker(session, idx->source,
file_func, name_func, cfg, open_flags));
} else if (WT_PREFIX_MATCH(uri, "lsm:")) {
WT_ERR(__wt_lsm_tree_worker(session,
uri, file_func, name_func, cfg, open_flags));
} else if (WT_PREFIX_MATCH(uri, "table:")) {
/*
* Note: we would like to use open_flags here (e.g., to lock
* the table exclusive during schema-changing operations), but
* that is currently problematic because we get the table again
* in order to discover column groups and indexes.
*/
WT_ERR(__wt_schema_get_table_uri(
session, uri, false, 0, &table));
/*
* We could make a recursive call for each colgroup or index
* URI, but since we have already opened the table, we can take
* a short cut and skip straight to the sources. If we have a
* name function, it needs to know about the intermediate URIs.
*/
for (i = 0; i < WT_COLGROUPS(table); i++) {
colgroup = table->cgroups[i];
skip = false;
if (name_func != NULL)
WT_ERR(name_func(
session, colgroup->name, &skip));
if (!skip)
WT_ERR(__wt_schema_worker(
session, colgroup->source,
file_func, name_func, cfg, open_flags));
}
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
idx = table->indices[i];
skip = false;
if (name_func != NULL)
WT_ERR(name_func(session, idx->name, &skip));
if (!skip)
WT_ERR(__wt_schema_worker(session, idx->source,
file_func, name_func, cfg, open_flags));
}
} else if ((dsrc = __wt_schema_get_source(session, uri)) != NULL) {
wt_session = (WT_SESSION *)session;
if (file_func == __wt_salvage && dsrc->salvage != NULL)
WT_ERR(dsrc->salvage(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg));
else if (file_func == __wt_verify && dsrc->verify != NULL)
WT_ERR(dsrc->verify(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg));
else if (file_func == __wt_checkpoint)
;
else if (file_func == __wt_checkpoint_get_handles)
;
else if (file_func == __wt_checkpoint_sync)
;
else
WT_ERR(__wt_object_unsupported(session, uri));
} else
WT_ERR(__wt_bad_object_type(session, uri));
err: if (table != NULL)
WT_TRET(__wt_schema_release_table(session, table));
return (ret);
}
/*
* __wt_curindex_open --
* WT_SESSION->open_cursor method for index cursors.
*/
int
__wt_curindex_open(WT_SESSION_IMPL *session,
const char *uri, WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__curindex_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__curindex_set_value, /* set-value */
__curindex_compare, /* compare */
__wt_cursor_equals, /* equals */
__curindex_next, /* next */
__curindex_prev, /* prev */
__curindex_reset, /* reset */
__curindex_search, /* search */
__curindex_search_near, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__curindex_close); /* close */
WT_CURSOR_INDEX *cindex;
WT_CURSOR *cursor;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_INDEX *idx;
WT_TABLE *table;
const char *columns, *idxname, *tablename;
size_t namesize;
tablename = uri;
if (!WT_PREFIX_SKIP(tablename, "index:") ||
(idxname = strchr(tablename, ':')) == NULL)
WT_RET_MSG(session, EINVAL, "Invalid cursor URI: '%s'", uri);
namesize = (size_t)(idxname - tablename);
++idxname;
if ((ret = __wt_schema_get_table(session,
tablename, namesize, false, &table)) != 0) {
if (ret == WT_NOTFOUND)
WT_RET_MSG(session, EINVAL,
"Cannot open cursor '%s' on unknown table", uri);
return (ret);
}
columns = strchr(idxname, '(');
if (columns == NULL)
namesize = strlen(idxname);
else
namesize = (size_t)(columns - idxname);
if ((ret = __wt_schema_open_index(
session, table, idxname, namesize, &idx)) != 0) {
__wt_schema_release_table(session, table);
return (ret);
}
WT_RET(__wt_calloc_one(session, &cindex));
cursor = &cindex->iface;
*cursor = iface;
cursor->session = &session->iface;
cindex->table = table;
cindex->index = idx;
cindex->key_plan = idx->key_plan;
cindex->value_plan = idx->value_plan;
cursor->internal_uri = idx->name;
cursor->key_format = idx->idxkey_format;
cursor->value_format = table->value_format;
/*
* XXX
* A very odd corner case is an index with a recno key.
* The only way to get here is by creating an index on a column store
* using only the primary's recno as the index key. Disallow that for
* now.
*/
if (WT_CURSOR_RECNO(cursor))
WT_ERR_MSG(session, WT_ERROR,
"Column store indexes based on a record number primary "
"key are not supported");
/* Handle projections. */
if (columns != NULL) {
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__wt_struct_reformat(session, table,
columns, strlen(columns), NULL, false, tmp));
WT_ERR(__wt_strndup(
session, tmp->data, tmp->size, &cursor->value_format));
WT_ERR(__wt_buf_init(session, tmp, 0));
WT_ERR(__wt_struct_plan(session, table,
columns, strlen(columns), false, tmp));
WT_ERR(__wt_strndup(
session, tmp->data, tmp->size, &cindex->value_plan));
}
WT_ERR(__wt_cursor_init(
cursor, cursor->internal_uri, owner, cfg, cursorp));
WT_ERR(__wt_open_cursor(
session, idx->source, cursor, cfg, &cindex->child));
/* Open the column groups needed for this index cursor. */
WT_ERR(__curindex_open_colgroups(session, cindex, cfg));
if (F_ISSET(cursor, WT_CURSTD_DUMP_JSON))
__wt_json_column_init(
cursor, table->key_format, &idx->colconf, &table->colconf);
if (0) {
err: WT_TRET(__curindex_close(cursor));
*cursorp = NULL;
}
__wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_schema_worker --
* Get Btree handles for the object and cycle through calls to an
* underlying worker function with each handle.
*/
int
__wt_schema_worker(WT_SESSION_IMPL *session,
const char *uri,
int (*file_func)(WT_SESSION_IMPL *, const char *[]),
int (*name_func)(WT_SESSION_IMPL *, const char *),
const char *cfg[], uint32_t open_flags)
{
WT_COLGROUP *colgroup;
WT_DATA_SOURCE *dsrc;
WT_DECL_RET;
WT_INDEX *idx;
WT_SESSION *wt_session;
WT_TABLE *table;
const char *tablename;
u_int i;
table = NULL;
tablename = uri;
if (name_func != NULL)
WT_ERR(name_func(session, uri));
/* Get the btree handle(s) and call the underlying function. */
if (WT_PREFIX_MATCH(uri, "file:")) {
if (file_func != NULL) {
WT_ERR(__wt_session_get_btree_ckpt(
session, uri, cfg, open_flags));
ret = file_func(session, cfg);
WT_TRET(__wt_session_release_btree(session));
}
} else if (WT_PREFIX_MATCH(uri, "colgroup:")) {
WT_ERR(__wt_schema_get_colgroup(session, uri, NULL, &colgroup));
WT_ERR(__wt_schema_worker(session, colgroup->source,
file_func, name_func, cfg, open_flags));
} else if (WT_PREFIX_SKIP(tablename, "index:")) {
idx = NULL;
WT_ERR(__wt_schema_get_index(session, uri, NULL, &idx));
WT_ERR(__wt_schema_worker(session, idx->source,
file_func, name_func, cfg, open_flags));
} else if (WT_PREFIX_MATCH(uri, "lsm:")) {
WT_ERR(__wt_lsm_tree_worker(
session, uri, file_func, name_func, cfg, open_flags));
} else if (WT_PREFIX_SKIP(tablename, "table:")) {
WT_ERR(__wt_schema_get_table(session,
tablename, strlen(tablename), 0, &table));
WT_ASSERT(session, session->dhandle == NULL);
/*
* We could make a recursive call for each colgroup or index
* URI, but since we have already opened the table, we can take
* a short cut and skip straight to the sources. If we have a
* name function, it needs to know about the intermediate URIs.
*/
for (i = 0; i < WT_COLGROUPS(table); i++) {
colgroup = table->cgroups[i];
if (name_func != NULL)
WT_ERR(name_func(session, colgroup->name));
WT_ERR(__wt_schema_worker(session, colgroup->source,
file_func, name_func, cfg, open_flags));
}
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
idx = table->indices[i];
if (name_func != NULL)
WT_ERR(name_func(session, idx->name));
WT_ERR(__wt_schema_worker(session, idx->source,
file_func, name_func, cfg, open_flags));
}
} else if ((dsrc = __wt_schema_get_source(session, uri)) != NULL) {
wt_session = (WT_SESSION *)session;
if (file_func == __wt_compact && dsrc->compact != NULL)
WT_ERR(dsrc->compact(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg));
else if (file_func == __wt_salvage && dsrc->salvage != NULL)
WT_ERR(dsrc->salvage(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg));
else if (file_func == __wt_verify && dsrc->verify != NULL)
WT_ERR(dsrc->verify(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg));
else
WT_ERR(__wt_object_unsupported(session, uri));
} else
WT_ERR(__wt_bad_object_type(session, uri));
err: if (table != NULL)
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __create_table --
* Create a table.
*/
static int
__create_table(WT_SESSION_IMPL *session,
const char *name, bool exclusive, const char *config)
{
WT_CONFIG conf;
WT_CONFIG_ITEM cgkey, cgval, cval;
WT_DECL_RET;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, table_meta), config, NULL, NULL };
const char *tablename;
char *tableconf, *cgname;
size_t cgsize;
int ncolgroups;
bool exists;
cgname = NULL;
table = NULL;
tableconf = NULL;
exists = false;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "table:"))
return (__wt_unexpected_object_type(session, name, "table:"));
if ((ret = __wt_schema_get_table(session,
tablename, strlen(tablename), false, &table)) == 0) {
if (exclusive)
WT_ERR(EEXIST);
exists = true;
}
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_config_gets(session, cfg, "colgroups", &cval));
__wt_config_subinit(session, &conf, &cval);
for (ncolgroups = 0;
(ret = __wt_config_next(&conf, &cgkey, &cgval)) == 0;
ncolgroups++)
;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_config_collapse(session, cfg, &tableconf));
if (!exists) {
WT_ERR(__wt_metadata_insert(session, name, tableconf));
/* Attempt to open the table now to catch any errors. */
WT_ERR(__wt_schema_get_table(
session, tablename, strlen(tablename), true, &table));
if (ncolgroups == 0) {
cgsize = strlen("colgroup:") + strlen(tablename) + 1;
WT_ERR(__wt_calloc_def(session, cgsize, &cgname));
snprintf(cgname, cgsize, "colgroup:%s", tablename);
WT_ERR(__create_colgroup(
session, cgname, exclusive, config));
}
}
if (0) {
err: if (table != NULL) {
WT_TRET(__wt_schema_remove_table(session, table));
table = NULL;
}
}
if (table != NULL)
__wt_schema_release_table(session, table);
__wt_free(session, cgname);
__wt_free(session, tableconf);
return (ret);
}
