/*
* __curmetadata_insert --
* WT_CURSOR->insert method for the metadata cursor type.
*/
static int
__curmetadata_insert(WT_CURSOR *cursor)
{
WT_CURSOR *file_cursor;
WT_CURSOR_METADATA *mdc;
WT_DECL_RET;
WT_SESSION_IMPL *session;
mdc = (WT_CURSOR_METADATA *)cursor;
file_cursor = mdc->file_cursor;
CURSOR_API_CALL(cursor, session,
insert, ((WT_CURSOR_BTREE *)file_cursor)->btree);
WT_MD_CURSOR_NEEDKEY(cursor);
WT_MD_CURSOR_NEEDVALUE(cursor);
/*
* Since the key/value formats are 's' the WT_ITEMs must contain a
* NULL terminated string.
*/
ret =
__wt_metadata_insert(session, cursor->key.data, cursor->value.data);
err: API_END(session, ret);
return (ret);
}
/*
* __rename_file --
* WT_SESSION::rename for a file.
*/
static int
__rename_file(
WT_SESSION_IMPL *session, const char *uri, const char *newuri)
{
WT_DECL_RET;
bool exist;
const char *filename, *newfile;
char *newvalue, *oldvalue;
newvalue = oldvalue = NULL;
filename = uri;
newfile = newuri;
if (!WT_PREFIX_SKIP(filename, "file:") ||
!WT_PREFIX_SKIP(newfile, "file:"))
return (EINVAL);
/* Close any btree handles in the file. */
WT_WITH_HANDLE_LIST_LOCK(session,
ret = __wt_conn_dhandle_close_all(session, uri, false));
WT_ERR(ret);
/*
* First, check if the file being renamed exists in the system. Doing
* this check first matches the table rename behavior because we return
* WT_NOTFOUND when the renamed file doesn't exist (subsequently mapped
* to ENOENT by the session layer).
*/
WT_ERR(__wt_metadata_search(session, uri, &oldvalue));
/*
* Check to see if the proposed name is already in use, in either the
* metadata or the filesystem.
*/
switch (ret = __wt_metadata_search(session, newuri, &newvalue)) {
case 0:
WT_ERR_MSG(session, EEXIST, "%s", newuri);
/* NOTREACHED */
case WT_NOTFOUND:
break;
default:
WT_ERR(ret);
}
WT_ERR(__wt_fs_exist(session, newfile, &exist));
if (exist)
WT_ERR_MSG(session, EEXIST, "%s", newfile);
/* Replace the old file entries with new file entries. */
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, oldvalue));
/* Rename the underlying file. */
WT_ERR(__wt_fs_rename(session, filename, newfile, false));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, uri, newuri));
err: __wt_free(session, newvalue);
__wt_free(session, oldvalue);
return (ret);
}
/*修改文件名操作*/
static int __rename_file(WT_SESSION_IMPL* session, const char* uri, const char* newuri)
{
WT_DECL_RET;
int exist;
const char *filename, *newfile;
char *newvalue, *oldvalue;
newvalue = oldvalue = NULL;
filename = uri;
newfile = newuri;
if (!WT_PREFIX_SKIP(filename, "file:") || !WT_PREFIX_SKIP(newfile, "file:"))
return (EINVAL);
/*关闭这个文件对应的data source handler*/
WT_WITH_DHANDLE_LOCK(session, ret = __wt_conn_dhandle_close_all(session, uri, 0));
WT_ERR(ret);
/*
* First, check if the file being renamed exists in the system. Doing
* this check first matches the table rename behavior because we return
* WT_NOTFOUND when the renamed file doesn't exist (subsequently mapped
* to ENOENT by the session layer).
*/
/*先查出旧的文件名uri对应的meta 信息，再用新的文件名uri旧版本的meta信息*/
WT_ERR(__wt_metadata_search(session, uri, &oldvalue));
switch (ret = __wt_metadata_search(session, newuri, &newvalue)) {
case 0:
WT_ERR_MSG(session, EEXIST, "%s", newuri);
/* NOTREACHED */
case WT_NOTFOUND:
break;
default:
WT_ERR(ret);
}
WT_ERR(__wt_exist(session, newfile, &exist));
if (exist)
WT_ERR_MSG(session, EEXIST, "%s", newfile);
/* Replace the old file entries with new file entries. */
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, oldvalue));
WT_ERR(__wt_rename(session, filename, newfile));
if(WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, uri, newuri));
err:
__wt_free(session, newvalue);
__wt_free(session, oldvalue);
return ret;
}
/*
* __metadata_rename --
* Rename an entry in the metadata table.
*/
static int
__metadata_rename(WT_SESSION_IMPL *session, const char *uri, const char *newuri)
{
WT_DECL_RET;
char *value;
WT_RET(__wt_metadata_search(session, uri, &value));
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, value));
err: __wt_free(session, value);
return (ret);
}
/*
* __rename_file --
* WT_SESSION::rename for a file.
*/
static int
__rename_file(
WT_SESSION_IMPL *session, const char *uri, const char *newuri)
{
WT_DECL_RET;
int exist;
const char *filename, *newfile, *value;
value = NULL;
filename = uri;
newfile = newuri;
if (!WT_PREFIX_SKIP(filename, "file:") ||
!WT_PREFIX_SKIP(newfile, "file:"))
return (EINVAL);
/* Close any btree handles in the file. */
WT_RET(__wt_conn_btree_close_all(session, uri));
/*
* Check to see if the proposed name is already in use, in either
* the metadata or the filesystem.
*/
switch (ret = __wt_metadata_read(session, newuri, &value)) {
case 0:
WT_ERR_MSG(session, EEXIST, "%s", newuri);
case WT_NOTFOUND:
ret = 0;
break;
default:
WT_ERR(ret);
}
WT_ERR(__wt_exist(session, newfile, &exist));
if (exist)
WT_ERR_MSG(session, EEXIST, "%s", newfile);
/* Replace the old file entries with new file entries. */
WT_ERR(__wt_metadata_read(session, uri, &value));
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, value));
/* Rename the underlying file. */
WT_ERR(__wt_rename(session, filename, newfile));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, uri, newuri));
err: __wt_free(session, value);
return (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);
}
/*
* __rename_table --
* WT_SESSION::rename for a table.
*/
static int
__rename_table(
WT_SESSION_IMPL *session, const char *oldname, const char *newname)
{
WT_DECL_RET;
WT_ITEM *buf;
WT_TABLE *table;
int i;
const char *value;
buf = NULL;
WT_RET(
__wt_schema_get_table(session, oldname, strlen(oldname), &table));
/* Rename the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++)
WT_RET(__rename_tree(session, table->cg_name[i], newname));
/* Rename the indices. */
WT_RET(__wt_schema_open_index(session, table, NULL, 0));
for (i = 0; i < table->nindices; i++)
WT_RET(__rename_tree(session, table->idx_name[i], newname));
WT_RET(__wt_schema_remove_table(session, table));
/* Rename the table. */
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf, "table:%s", oldname));
WT_ERR(__wt_metadata_read(session, buf->data, &value));
WT_ERR(__wt_metadata_remove(session, buf->data));
WT_ERR(__wt_buf_fmt(session, buf, "table:%s", newname));
WT_ERR(__wt_metadata_insert(session, buf->data, value));
err: __wt_scr_free(&buf);
return (ret);
}
/*
* __rename_tree --
* Rename an index or colgroup reference.
*/
static int
__rename_tree(WT_SESSION_IMPL *session,
WT_TABLE *table, const char *newuri, const char *name, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(nn);
WT_DECL_ITEM(ns);
WT_DECL_ITEM(nv);
WT_DECL_ITEM(os);
WT_DECL_RET;
bool is_colgroup;
const char *newname, *olduri, *suffix;
char *value;
olduri = table->name;
value = NULL;
newname = newuri;
(void)WT_PREFIX_SKIP(newname, "table:");
/*
* Create the new data source URI and update the schema value.
*
* 'name' has the format (colgroup|index):<tablename>[:<suffix>];
* we need the suffix.
*/
is_colgroup = WT_PREFIX_MATCH(name, "colgroup:");
if (!is_colgroup && !WT_PREFIX_MATCH(name, "index:"))
WT_ERR_MSG(session, EINVAL,
"expected a 'colgroup:' or 'index:' source: '%s'", name);
suffix = strchr(name, ':');
/* An existing table should have a well formed name. */
WT_ASSERT(session, suffix != NULL);
suffix = strchr(suffix + 1, ':');
WT_ERR(__wt_scr_alloc(session, 0, &nn));
WT_ERR(__wt_buf_fmt(session, nn, "%s%s%s",
is_colgroup ? "colgroup:" : "index:",
newname,
(suffix == NULL) ? "" : suffix));
/* Skip the colon, if any. */
if (suffix != NULL)
++suffix;
/* Read the old schema value. */
WT_ERR(__wt_metadata_search(session, name, &value));
/*
* Calculate the new data source URI. Use the existing table structure
* and substitute the new name temporarily.
*/
WT_ERR(__wt_scr_alloc(session, 0, &ns));
table->name = newuri;
if (is_colgroup)
WT_ERR(__wt_schema_colgroup_source(
session, table, suffix, value, ns));
else
WT_ERR(__wt_schema_index_source(
session, table, suffix, value, ns));
if ((ret = __wt_config_getones(session, value, "source", &cval)) != 0)
WT_ERR_MSG(session, EINVAL,
"index or column group has no data source: %s", value);
/* Take a copy of the old data source. */
WT_ERR(__wt_scr_alloc(session, 0, &os));
WT_ERR(__wt_buf_fmt(session, os, "%.*s", (int)cval.len, cval.str));
/* Overwrite it with the new data source. */
WT_ERR(__wt_scr_alloc(session, 0, &nv));
WT_ERR(__wt_buf_fmt(session, nv, "%.*s%s%s",
(int)WT_PTRDIFF(cval.str, value), value,
(const char *)ns->data, cval.str + cval.len));
/*
* Do the rename before updating the metadata to avoid leaving the
* metadata inconsistent if the rename fails.
*/
WT_ERR(__wt_schema_rename(session, os->data, ns->data, cfg));
/*
* Remove the old metadata entry.
* Insert the new metadata entry.
*/
WT_ERR(__wt_metadata_remove(session, name));
WT_ERR(__wt_metadata_insert(session, nn->data, nv->data));
err: __wt_scr_free(session, &nn);
__wt_scr_free(session, &ns);
__wt_scr_free(session, &nv);
__wt_scr_free(session, &os);
__wt_free(session, value);
table->name = olduri;
return (ret);
}
/*
* __create_file --
* Create a new 'file:' object.
*/
static int
__create_file(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_DECL_ITEM(val);
WT_DECL_RET;
uint32_t allocsize;
int is_metadata;
const char *filename, **p, *filecfg[] =
{ WT_CONFIG_BASE(session, file_meta), config, NULL, NULL };
char *fileconf;
fileconf = NULL;
is_metadata = strcmp(uri, WT_METAFILE_URI) == 0;
filename = uri;
if (!WT_PREFIX_SKIP(filename, "file:"))
WT_RET_MSG(session, EINVAL, "Expected a 'file:' URI: %s", uri);
/* Check if the file already exists. */
if (!is_metadata && (ret =
__wt_metadata_search(session, uri, &fileconf)) != WT_NOTFOUND) {
if (exclusive)
WT_TRET(EEXIST);
goto err;
}
/* Sanity check the allocation size. */
WT_RET(__wt_direct_io_size_check(
session, filecfg, "allocation_size", &allocsize));
/* Create the file. */
WT_ERR(__wt_block_manager_create(session, filename, allocsize));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, NULL, uri));
/*
* If creating an ordinary file, append the file ID and current version
* numbers to the passed-in configuration and insert the resulting
* configuration into the metadata.
*/
if (!is_metadata) {
WT_ERR(__wt_scr_alloc(session, 0, &val));
WT_ERR(__wt_buf_fmt(session, val,
"id=%" PRIu32 ",version=(major=%d,minor=%d)",
++S2C(session)->next_file_id,
WT_BTREE_MAJOR_VERSION_MAX, WT_BTREE_MINOR_VERSION_MAX));
for (p = filecfg; *p != NULL; ++p)
;
*p = val->data;
WT_ERR(__wt_config_collapse(session, filecfg, &fileconf));
WT_ERR(__wt_metadata_insert(session, uri, fileconf));
}
/*
* Open the file to check that it was setup correctly. We don't need to
* pass the configuration, we just wrote the collapsed configuration
* into the metadata file, and it's going to be read/used by underlying
* functions.
*
* Keep the handle exclusive until it is released at the end of the
* call, otherwise we could race with a drop.
*/
WT_ERR(__wt_session_get_btree(
session, uri, NULL, NULL, WT_DHANDLE_EXCLUSIVE));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_handle_lock(session, 1));
else
WT_ERR(__wt_session_release_btree(session));
err: __wt_scr_free(session, &val);
__wt_free(session, fileconf);
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_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_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);
}
/*
* __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);
}
/*
* __rename_tree --
* Rename an index or colgroup reference.
*/
static int
__rename_tree(WT_SESSION_IMPL *session, const char *name, const char *newname)
{
WT_DECL_RET;
WT_ITEM *of, *nf, *nk, *nv;
const char *newfile, *p, *t, *value;
nf = nk = nv = of = NULL;
/* Read the old schema value. */
WT_ERR(__wt_metadata_read(session, name, &value));
/*
* Create the new file name, new schema key, new schema value.
*
* Names are of the form "prefix.oldname:suffix", where suffix is
* optional; we need prefix and suffix.
*/
if ((p = strchr(name, ':')) == NULL)
WT_ERR_MSG(session, EINVAL,
"invalid index or column-group name: %s", name);
t = strchr(p + 1, ':');
WT_ERR(__wt_scr_alloc(session, 0, &nf));
WT_ERR(__wt_buf_fmt(session, nf, "file:%s%s%s.wt", newname,
t == NULL ? "" : "_", t == NULL ? "" : t + 1));
newfile = (const char *)nf->data + strlen("file:");
WT_ERR(__wt_scr_alloc(session, 0, &nk));
WT_ERR(__wt_buf_fmt(session, nk, "%.*s:%s%s%s",
(int)WT_PTRDIFF(p, name), name, newname,
t == NULL ? "" : ":", t == NULL ? "" : t + 1));
if ((p = strstr(value, "filename=")) == NULL)
WT_ERR_MSG(session, EINVAL,
"index or column-group value has no file name: %s", value);
p += strlen("filename=");
t = strchr(p, ',');
/* Take a copy of the old filename. */
WT_ERR(__wt_scr_alloc(session, 0, &of));
WT_ERR(__wt_buf_fmt(session, of, "file:%.*s",
(int)((t == NULL) ? strlen(p) : WT_PTRDIFF(t, p)), p));
/* Overwrite it with the new filename. */
WT_ERR(__wt_scr_alloc(session, 0, &nv));
WT_ERR(__wt_buf_fmt(session, nv, "%.*s%s%s",
(int)WT_PTRDIFF(p, value), value,
newfile, t == NULL ? "" : t));
/*
* Remove the old metadata entry.
* Insert the new metadata entry.
*/
WT_ERR(__wt_metadata_remove(session, name));
WT_ERR(__wt_metadata_insert(session, nk->data, nv->data));
/* Rename the file. */
WT_ERR(__rename_file(session, of->data, nf->data));
err: __wt_scr_free(&nf);
__wt_scr_free(&nk);
__wt_scr_free(&nv);
__wt_scr_free(&of);
__wt_free(session, value);
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);
}
