/*
* __ckpt_last_addr --
* Return the cookie associated with the file's last checkpoint.
*/
static int
__ckpt_last_addr(
WT_SESSION_IMPL *session, const char *config, WT_ITEM *addr)
{
WT_CONFIG ckptconf;
WT_CONFIG_ITEM a, k, v;
int64_t found;
WT_RET(__wt_config_getones(session, config, "checkpoint", &v));
WT_RET(__wt_config_subinit(session, &ckptconf, &v));
for (found = 0; __wt_config_next(&ckptconf, &k, &v) == 0;) {
/* Ignore checkpoints before the ones we've already seen. */
WT_RET(__wt_config_subgets(session, &v, "order", &a));
if (found && a.val < found)
continue;
found = a.val;
/*
* Copy out the address; our caller wants the raw cookie, not
* the hex.
*/
WT_RET(__wt_config_subgets(session, &v, "addr", &a));
if (a.len != 0)
WT_RET(__wt_nhex_to_raw(session, a.str, a.len, addr));
}
return (found ? 0 : WT_NOTFOUND);
}
/*
* __verify_config_offsets --
* Debugging: optionally dump specific blocks from the file.
*/
static int
__verify_config_offsets(
WT_SESSION_IMPL *session, const char *cfg[], bool *quitp)
{
WT_CONFIG list;
WT_CONFIG_ITEM cval, k, v;
WT_DECL_RET;
u_long offset;
*quitp = false;
WT_RET(__wt_config_gets(session, cfg, "dump_offsets", &cval));
WT_RET(__wt_config_subinit(session, &list, &cval));
while ((ret = __wt_config_next(&list, &k, &v)) == 0) {
/*
* Quit after dumping the requested blocks. (That's hopefully
* what the user wanted, all of this stuff is just hooked into
* verify because that's where we "dump blocks" for debugging.)
*/
*quitp = true;
if (v.len != 0 || sscanf(k.str, "%lu", &offset) != 1)
WT_RET_MSG(session, EINVAL,
"unexpected dump offset format");
#if !defined(HAVE_DIAGNOSTIC)
WT_RET_MSG(session, ENOTSUP,
"the WiredTiger library was not built in diagnostic mode");
#else
WT_TRET(
__wt_debug_offset_blind(session, (wt_off_t)offset, NULL));
#endif
}
return (ret == WT_NOTFOUND ? 0 : ret);
}
/*
* __find_column_format --
* Find the format of the named column.
*/
static int
__find_column_format(WT_SESSION_IMPL *session, WT_TABLE *table,
WT_CONFIG_ITEM *colname, bool value_only, WT_PACK_VALUE *pv)
{
WT_CONFIG conf;
WT_CONFIG_ITEM k, v;
WT_DECL_RET;
WT_PACK pack;
bool inkey;
__wt_config_subinit(session, &conf, &table->colconf);
WT_RET(__pack_init(session, &pack, table->key_format));
inkey = true;
while ((ret = __wt_config_next(&conf, &k, &v)) == 0) {
if ((ret = __pack_next(&pack, pv)) == WT_NOTFOUND && inkey) {
ret = __pack_init(session, &pack, table->value_format);
if (ret == 0)
ret = __pack_next(&pack, pv);
inkey = false;
}
if (ret != 0)
return (ret);
if (k.len == colname->len &&
strncmp(colname->str, k.str, k.len) == 0) {
if (value_only && inkey)
return (__wt_set_return(session, EINVAL));
return (0);
}
}
return (ret);
}
/*
* __wt_meta_ckptlist_get --
* Load all available checkpoint information for a file.
*/
int
__wt_meta_ckptlist_get(
WT_SESSION_IMPL *session, const char *fname, WT_CKPT **ckptbasep)
{
WT_CKPT *ckpt, *ckptbase;
WT_CONFIG ckptconf;
WT_CONFIG_ITEM k, v;
WT_DECL_ITEM(buf);
WT_DECL_RET;
size_t allocated, slot;
char *config;
*ckptbasep = NULL;
ckptbase = NULL;
allocated = slot = 0;
config = NULL;
/* Retrieve the metadata information for the file. */
WT_RET(__wt_metadata_search(session, fname, &config));
/* Load any existing checkpoints into the array. */
WT_ERR(__wt_scr_alloc(session, 0, &buf));
if (__wt_config_getones(session, config, "checkpoint", &v) == 0) {
__wt_config_subinit(session, &ckptconf, &v);
for (; __wt_config_next(&ckptconf, &k, &v) == 0; ++slot) {
WT_ERR(__wt_realloc_def(
session, &allocated, slot + 1, &ckptbase));
ckpt = &ckptbase[slot];
WT_ERR(__ckpt_load(session, &k, &v, ckpt));
}
}
/*
* Allocate an extra slot for a new value, plus a slot to mark the end.
*
* This isn't very clean, but there's necessary cooperation between the
* schema layer (that maintains the list of checkpoints), the btree
* layer (that knows when the root page is written, creating a new
* checkpoint), and the block manager (which actually creates the
* checkpoint). All of that cooperation is handled in the WT_CKPT
* structure referenced from the WT_BTREE structure.
*/
WT_ERR(__wt_realloc_def(session, &allocated, slot + 2, &ckptbase));
/* Sort in creation-order. */
qsort(ckptbase, slot, sizeof(WT_CKPT), __ckpt_compare_order);
/* Return the array to our caller. */
*ckptbasep = ckptbase;
if (0) {
err: __wt_meta_ckptlist_free(session, &ckptbase);
}
__wt_free(session, config);
__wt_scr_free(session, &buf);
return (ret);
}
/*
* __snap_get_last --
* Return the cookie associated with the file's last snapshot.
*/
static int
__snap_get_last(
WT_SESSION_IMPL *session, const char *config, WT_ITEM *addr)
{
WT_CONFIG snapconf;
WT_CONFIG_ITEM a, k, v;
int64_t found;
WT_RET(__wt_config_getones(session, config, "snapshot", &v));
WT_RET(__wt_config_subinit(session, &snapconf, &v));
for (found = 0; __wt_config_next(&snapconf, &k, &v) == 0;) {
if (found) {
WT_RET(__wt_config_subgets(session, &v, "order", &a));
if (a.val < found)
continue;
}
WT_RET(__wt_config_subgets(session, &v, "addr", &a));
if (a.len == 0)
WT_RET(EINVAL);
/* Our caller wants the raw cookie, not the hex. */
WT_RET(__wt_nhex_to_raw(session, a.str, a.len, addr));
WT_RET(__wt_config_subgets(session, &v, "order", &a));
found = a.val;
}
return (found ? 0 : WT_NOTFOUND);
}
/*
* __backup_uri --
* Backup a list of objects.
*/
static int
__backup_uri(WT_SESSION_IMPL *session,
WT_CURSOR_BACKUP *cb, const char *cfg[], int *foundp, int *log_only)
{
WT_CONFIG targetconf;
WT_CONFIG_ITEM cval, k, v;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
int target_list;
const char *uri;
*foundp = 0;
*log_only = 0;
/*
* If we find a non-empty target configuration string, we have a job,
* otherwise it's not our problem.
*/
WT_RET(__wt_config_gets(session, cfg, "target", &cval));
WT_RET(__wt_config_subinit(session, &targetconf, &cval));
for (cb->list_next = 0, target_list = 0;
(ret = __wt_config_next(&targetconf, &k, &v)) == 0; ++target_list) {
/* If it is our first time through, allocate. */
if (target_list == 0) {
*foundp = 1;
WT_ERR(__wt_scr_alloc(session, 512, &tmp));
}
WT_ERR(__wt_buf_fmt(session, tmp, "%.*s", (int)k.len, k.str));
uri = tmp->data;
if (v.len != 0)
WT_ERR_MSG(session, EINVAL,
"%s: invalid backup target: URIs may need quoting",
uri);
/*
* Handle log targets. We do not need to go through the
* schema worker, just call the function to append them.
* Set log_only only if it is our only URI target.
*/
if (WT_PREFIX_MATCH(uri, "log:")) {
if (target_list == 0)
*log_only = 1;
else
*log_only = 0;
WT_ERR(__wt_backup_list_uri_append(
session, uri, NULL));
} else
WT_ERR(__wt_schema_worker(session,
uri, NULL, __wt_backup_list_uri_append, cfg, 0));
}
WT_ERR_NOTFOUND_OK(ret);
err:
__wt_scr_free(session, &tmp);
return (ret);
}
/*
* __snap_get_name --
* Return the cookie associated with a file's named snapshot.
*/
static int
__snap_get_name(WT_SESSION_IMPL *session,
const char *name, const char *config, WT_ITEM *addr)
{
WT_CONFIG snapconf;
WT_CONFIG_ITEM a, k, v;
WT_RET(__wt_config_getones(session, config, "snapshot", &v));
WT_RET(__wt_config_subinit(session, &snapconf, &v));
while (__wt_config_next(&snapconf, &k, &v) == 0)
if (strlen(name) == k.len && strncmp(name, k.str, k.len) == 0) {
WT_RET(__wt_config_subgets(session, &v, "addr", &a));
WT_RET(__wt_nhex_to_raw(session, a.str, a.len, addr));
return (0);
}
return (WT_NOTFOUND);
}
/*
* __backup_uri --
* Backup a list of objects.
*/
static int
__backup_uri(WT_SESSION_IMPL *session,
WT_CURSOR_BACKUP *cb, const char *cfg[], int *foundp)
{
WT_CONFIG targetconf;
WT_CONFIG_ITEM cval, k, v;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
int target_list;
const char *uri;
*foundp = target_list = 0;
/*
* If we find a non-empty target configuration string, we have a job,
* otherwise it's not our problem.
*/
WT_RET(__wt_config_gets(session, cfg, "target", &cval));
WT_RET(__wt_config_subinit(session, &targetconf, &cval));
for (cb->list_next = 0;
(ret = __wt_config_next(&targetconf, &k, &v)) == 0;) {
if (!target_list) {
target_list = *foundp = 1;
WT_ERR(__wt_scr_alloc(session, 512, &tmp));
}
WT_ERR(__wt_buf_fmt(session, tmp, "%.*s", (int)k.len, k.str));
uri = tmp->data;
if (v.len != 0)
WT_ERR_MSG(session, EINVAL,
"%s: invalid backup target: URIs may need quoting",
uri);
WT_ERR(__wt_schema_worker(
session, uri, NULL, __wt_backup_list_uri_append, cfg, 0));
}
WT_ERR_NOTFOUND_OK(ret);
err: __wt_scr_free(&tmp);
return (ret);
}
/*
* __ckpt_named --
* Return the information associated with a file's named checkpoint.
*/
static int
__ckpt_named(WT_SESSION_IMPL *session,
const char *checkpoint, const char *config, WT_CKPT *ckpt)
{
WT_CONFIG ckptconf;
WT_CONFIG_ITEM k, v;
WT_RET(__wt_config_getones(session, config, "checkpoint", &v));
WT_RET(__wt_config_subinit(session, &ckptconf, &v));
/*
* Take the first match: there should never be more than a single
* checkpoint of any name.
*/
while (__wt_config_next(&ckptconf, &k, &v) == 0)
if (WT_STRING_MATCH(checkpoint, k.str, k.len))
return (__ckpt_load(session, &k, &v, ckpt));
return (WT_NOTFOUND);
}
/*
* __ckpt_last_name --
* Return the name associated with the file's last unnamed checkpoint.
*/
static int
__ckpt_last_name(
WT_SESSION_IMPL *session, const char *config, const char **namep)
{
WT_CONFIG ckptconf;
WT_CONFIG_ITEM a, k, v;
WT_DECL_RET;
int64_t found;
*namep = NULL;
WT_ERR(__wt_config_getones(session, config, "checkpoint", &v));
WT_ERR(__wt_config_subinit(session, &ckptconf, &v));
for (found = 0; __wt_config_next(&ckptconf, &k, &v) == 0;) {
/*
* We only care about unnamed checkpoints; applications may not
* use any matching prefix as a checkpoint name, the comparison
* is pretty simple.
*/
if (k.len < strlen(WT_CHECKPOINT) ||
strncmp(k.str, WT_CHECKPOINT, strlen(WT_CHECKPOINT)) != 0)
continue;
/* Ignore checkpoints before the ones we've already seen. */
WT_ERR(__wt_config_subgets(session, &v, "order", &a));
if (found && a.val < found)
continue;
if (*namep != NULL)
__wt_free(session, *namep);
WT_ERR(__wt_strndup(session, k.str, k.len, namep));
found = a.val;
}
if (!found)
ret = WT_NOTFOUND;
if (0) {
err: __wt_free(session, namep);
}
return (ret);
}
/*
* __wt_schema_colcheck --
* Check that a list of columns matches a (key,value) format pair.
*/
int
__wt_schema_colcheck(WT_SESSION_IMPL *session,
const char *key_format, const char *value_format, WT_CONFIG_ITEM *colconf,
u_int *kcolsp, u_int *vcolsp)
{
WT_CONFIG conf;
WT_CONFIG_ITEM k, v;
WT_DECL_PACK_VALUE(pv);
WT_DECL_RET;
WT_PACK pack;
u_int kcols, ncols, vcols;
WT_RET(__pack_init(session, &pack, key_format));
for (kcols = 0; (ret = __pack_next(&pack, &pv)) == 0; kcols++)
;
WT_RET_NOTFOUND_OK(ret);
WT_RET(__pack_init(session, &pack, value_format));
for (vcols = 0; (ret = __pack_next(&pack, &pv)) == 0; vcols++)
;
WT_RET_TEST(ret != WT_NOTFOUND, ret);
/* Walk through the named columns. */
__wt_config_subinit(session, &conf, colconf);
for (ncols = 0; (ret = __wt_config_next(&conf, &k, &v)) == 0; ncols++)
;
WT_RET_TEST(ret != WT_NOTFOUND, ret);
if (ncols != 0 && ncols != kcols + vcols)
WT_RET_MSG(session, EINVAL, "Number of columns in '%.*s' "
"does not match key format '%s' plus value format '%s'",
(int)colconf->len, colconf->str, key_format, value_format);
if (kcolsp != NULL)
*kcolsp = kcols;
if (vcolsp != NULL)
*vcolsp = vcols;
return (0);
}
/*
* __wt_table_check --
* Make sure all columns appear in a column group.
*/
int
__wt_table_check(WT_SESSION_IMPL *session, WT_TABLE *table)
{
WT_CONFIG conf;
WT_CONFIG_ITEM k, v;
WT_DECL_RET;
u_int cg, col, i;
char coltype;
if (table->is_simple)
return (0);
/* Walk through the columns. */
__wt_config_subinit(session, &conf, &table->colconf);
/* Skip over the key columns. */
for (i = 0; i < table->nkey_columns; i++)
WT_RET(__wt_config_next(&conf, &k, &v));
cg = col = 0;
coltype = 0;
while ((ret = __wt_config_next(&conf, &k, &v)) == 0) {
if (__find_next_col(
session, table, &k, &cg, &col, &coltype) != 0)
WT_RET_MSG(session, EINVAL,
"Column '%.*s' in '%s' does not appear in a "
"column group",
(int)k.len, k.str, table->iface.name);
/*
* Column groups can't store key columns in their value:
* __wt_struct_reformat should have already detected this case.
*/
WT_ASSERT(session, coltype == WT_PROJ_VALUE);
}
WT_RET_TEST(ret != WT_NOTFOUND, ret);
return (0);
}
/*
* __ckpt_last --
* Return the information associated with the file's last checkpoint.
*/
static int
__ckpt_last(WT_SESSION_IMPL *session, const char *config, WT_CKPT *ckpt)
{
WT_CONFIG ckptconf;
WT_CONFIG_ITEM a, k, v;
int64_t found;
WT_RET(__wt_config_getones(session, config, "checkpoint", &v));
WT_RET(__wt_config_subinit(session, &ckptconf, &v));
for (found = 0; __wt_config_next(&ckptconf, &k, &v) == 0;) {
/* Ignore checkpoints before the ones we've already seen. */
WT_RET(__wt_config_subgets(session, &v, "order", &a));
if (found) {
if (a.val < found)
continue;
__wt_meta_checkpoint_free(session, ckpt);
}
found = a.val;
WT_RET(__ckpt_load(session, &k, &v, ckpt));
}
return (found ? 0 : WT_NOTFOUND);
}
/*
* __ckpt_named_addr --
* Return the cookie associated with a file's named checkpoint.
*/
static int
__ckpt_named_addr(WT_SESSION_IMPL *session,
const char *checkpoint, const char *config, WT_ITEM *addr)
{
WT_CONFIG ckptconf;
WT_CONFIG_ITEM a, k, v;
WT_RET(__wt_config_getones(session, config, "checkpoint", &v));
WT_RET(__wt_config_subinit(session, &ckptconf, &v));
/*
* Take the first match: there should never be more than a single
* checkpoint of any name.
*/
while (__wt_config_next(&ckptconf, &k, &v) == 0)
if (WT_STRING_MATCH(checkpoint, k.str, k.len)) {
WT_RET(__wt_config_subgets(session, &v, "addr", &a));
if (a.len != 0)
WT_RET(__wt_nhex_to_raw(
session, a.str, a.len, addr));
return (0);
}
return (WT_NOTFOUND);
}
static int
__find_next_col(WT_SESSION_IMPL *session, WT_TABLE *table,
WT_CONFIG_ITEM *colname, int *cgnump, int *colnump, char *coltype)
{
WT_BTREE *cgtree;
WT_CONFIG conf;
WT_CONFIG_ITEM cval, k, v;
WT_DECL_RET;
int cg, col, foundcg, foundcol, getnext;
foundcg = foundcol = -1;
getnext = 1;
for (cgtree = NULL, cg = 0; cg < WT_COLGROUPS(table); cg++) {
WT_RET(__wt_schema_get_btree(session,
table->cg_name[cg], strlen(table->cg_name[cg]),
NULL, 0));
cgtree = session->btree;
/*
* If there is only one column group, we just scan through all
* of the columns. For tables with multiple column groups, we
* look at the key columns once, then go through the value
* columns for each group.
*/
if (cg == 0) {
cval = table->colconf;
col = 0;
} else {
cgcols: WT_ERR(__wt_config_getones(session,
cgtree->config, "columns", &cval));
col = table->nkey_columns;
}
WT_ERR(__wt_config_subinit(session, &conf, &cval));
for (; __wt_config_next(&conf, &k, &v) == 0; col++) {
if (cg == *cgnump && col == *colnump)
getnext = 1;
if (getnext && k.len == colname->len &&
strncmp(colname->str, k.str, k.len) == 0) {
foundcg = cg;
foundcol = col;
getnext = 0;
}
if (cg == 0 && table->ncolgroups > 0 &&
col == table->nkey_columns - 1)
goto cgcols;
}
cgtree = NULL;
WT_ERR(__wt_session_release_btree(session));
}
err: if (cgtree != NULL)
WT_TRET(__wt_session_release_btree(session));
WT_RET(ret);
if (foundcg == -1)
return (WT_NOTFOUND);
*cgnump = foundcg;
if (foundcol < table->nkey_columns) {
*coltype = WT_PROJ_KEY;
*colnump = foundcol;
} else {
*coltype = WT_PROJ_VALUE;
*colnump = foundcol - table->nkey_columns;
}
return (0);
}
/*
* __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);
}
/*
* config_check --
* Check the keys in an application-supplied config string match what is
* specified in an array of check strings.
*/
static int
config_check(WT_SESSION_IMPL *session,
const WT_CONFIG_CHECK *checks, const char *config, size_t config_len)
{
WT_CONFIG parser, cparser, sparser;
WT_CONFIG_ITEM k, v, ck, cv, dummy;
WT_DECL_RET;
int i;
bool badtype, found;
/*
* The config_len parameter is optional, and allows passing in strings
* that are not nul-terminated.
*/
if (config_len == 0)
WT_RET(__wt_config_init(session, &parser, config));
else
WT_RET(__wt_config_initn(session, &parser, config, config_len));
while ((ret = __wt_config_next(&parser, &k, &v)) == 0) {
if (k.type != WT_CONFIG_ITEM_STRING &&
k.type != WT_CONFIG_ITEM_ID)
WT_RET_MSG(session, EINVAL,
"Invalid configuration key found: '%.*s'",
(int)k.len, k.str);
/* Search for a matching entry. */
for (i = 0; checks[i].name != NULL; i++)
if (WT_STRING_MATCH(checks[i].name, k.str, k.len))
break;
if (checks[i].name == NULL)
WT_RET_MSG(session, EINVAL,
"unknown configuration key: '%.*s'",
(int)k.len, k.str);
if (strcmp(checks[i].type, "boolean") == 0) {
badtype = v.type != WT_CONFIG_ITEM_BOOL &&
(v.type != WT_CONFIG_ITEM_NUM ||
(v.val != 0 && v.val != 1));
} else if (strcmp(checks[i].type, "category") == 0) {
/* Deal with categories of the form: XXX=(XXX=blah). */
ret = config_check(session,
checks[i].subconfigs,
k.str + strlen(checks[i].name) + 1, v.len);
if (ret != EINVAL)
badtype = false;
else
badtype = true;
} else if (strcmp(checks[i].type, "format") == 0) {
badtype = false;
} else if (strcmp(checks[i].type, "int") == 0) {
badtype = v.type != WT_CONFIG_ITEM_NUM;
} else if (strcmp(checks[i].type, "list") == 0) {
badtype = v.len > 0 && v.type != WT_CONFIG_ITEM_STRUCT;
} else if (strcmp(checks[i].type, "string") == 0) {
badtype = false;
} else
WT_RET_MSG(session, EINVAL,
"unknown configuration type: '%s'",
checks[i].type);
if (badtype)
WT_RET_MSG(session, EINVAL,
"Invalid value for key '%.*s': expected a %s",
(int)k.len, k.str, checks[i].type);
if (checks[i].checkf != NULL)
WT_RET(checks[i].checkf(session, &v));
if (checks[i].checks == NULL)
continue;
/* Setup an iterator for the check string. */
WT_RET(__wt_config_init(session, &cparser, checks[i].checks));
while ((ret = __wt_config_next(&cparser, &ck, &cv)) == 0) {
if (WT_STRING_MATCH("min", ck.str, ck.len)) {
if (v.val < cv.val)
WT_RET_MSG(session, EINVAL,
"Value too small for key '%.*s' "
"the minimum is %.*s",
(int)k.len, k.str,
(int)cv.len, cv.str);
} else if (WT_STRING_MATCH("max", ck.str, ck.len)) {
if (v.val > cv.val)
WT_RET_MSG(session, EINVAL,
"Value too large for key '%.*s' "
"the maximum is %.*s",
(int)k.len, k.str,
(int)cv.len, cv.str);
} else if (WT_STRING_MATCH("choices", ck.str, ck.len)) {
if (v.len == 0)
WT_RET_MSG(session, EINVAL,
"Key '%.*s' requires a value",
(int)k.len, k.str);
if (v.type == WT_CONFIG_ITEM_STRUCT) {
/*
* Handle the 'verbose' case of a list
* containing restricted choices.
*/
WT_RET(__wt_config_subinit(session,
&sparser, &v));
found = true;
while (found &&
(ret = __wt_config_next(&sparser,
&v, &dummy)) == 0) {
ret = __wt_config_subgetraw(
session, &cv, &v, &dummy);
found = ret == 0;
}
} else {
ret = __wt_config_subgetraw(session,
&cv, &v, &dummy);
found = ret == 0;
}
if (ret != 0 && ret != WT_NOTFOUND)
return (ret);
if (!found)
WT_RET_MSG(session, EINVAL,
"Value '%.*s' not a "
"permitted choice for key '%.*s'",
(int)v.len, v.str,
(int)k.len, k.str);
} else
WT_RET_MSG(session, EINVAL,
"unexpected configuration description "
"keyword %.*s", (int)ck.len, ck.str);
}
}
if (ret == WT_NOTFOUND)
ret = 0;
return (ret);
}
/*
* __wt_meta_ckptlist_get --
* Load all available checkpoint information for a file.
*/
int
__wt_meta_ckptlist_get(
WT_SESSION_IMPL *session, const char *fname, WT_CKPT **ckptbasep)
{
WT_CKPT *ckpt, *ckptbase;
WT_CONFIG ckptconf;
WT_CONFIG_ITEM a, k, v;
WT_DECL_RET;
WT_ITEM *buf;
size_t allocated, slot;
const char *config;
char timebuf[64];
*ckptbasep = NULL;
buf = NULL;
ckptbase = NULL;
allocated = slot = 0;
config = NULL;
/* Retrieve the metadata information for the file. */
WT_RET(__wt_metadata_read(session, fname, &config));
/* Load any existing checkpoints into the array. */
WT_ERR(__wt_scr_alloc(session, 0, &buf));
if (__wt_config_getones(session, config, "checkpoint", &v) == 0 &&
__wt_config_subinit(session, &ckptconf, &v) == 0)
for (; __wt_config_next(&ckptconf, &k, &v) == 0; ++slot) {
if (slot * sizeof(WT_CKPT) == allocated)
WT_ERR(__wt_realloc(session, &allocated,
(slot + 50) * sizeof(WT_CKPT), &ckptbase));
ckpt = &ckptbase[slot];
/*
* Copy the name, address (raw and hex), order and time
* into the slot. If there's no address, it's a fake.
*/
WT_ERR(
__wt_strndup(session, k.str, k.len, &ckpt->name));
WT_ERR(__wt_config_subgets(session, &v, "addr", &a));
WT_ERR(
__wt_buf_set(session, &ckpt->addr, a.str, a.len));
if (a.len == 0)
F_SET(ckpt, WT_CKPT_FAKE);
else
WT_ERR(__wt_nhex_to_raw(
session, a.str, a.len, &ckpt->raw));
WT_ERR(__wt_config_subgets(session, &v, "order", &a));
if (a.val == 0)
goto format;
ckpt->order = a.val;
WT_ERR(__wt_config_subgets(session, &v, "time", &a));
if (a.len == 0)
goto format;
if (a.len > sizeof(timebuf) - 1)
goto format;
memcpy(timebuf, a.str, a.len);
timebuf[a.len] = '\0';
if (sscanf(timebuf, "%" SCNuMAX, &ckpt->sec) != 1)
goto format;
WT_ERR(__wt_config_subgets(session, &v, "size", &a));
ckpt->ckpt_size = (uint64_t)a.val;
}
/*
* Allocate an extra slot for a new value, plus a slot to mark the end.
*
* This isn't very clean, but there's necessary cooperation between the
* schema layer (that maintains the list of checkpoints), the btree
* layer (that knows when the root page is written, creating a new
* checkpoint), and the block manager (which actually creates the
* checkpoint). All of that cooperation is handled in the WT_CKPT
* structure referenced from the WT_BTREE structure.
*/
if ((slot + 2) * sizeof(WT_CKPT) > allocated)
WT_ERR(__wt_realloc(session, &allocated,
(slot + 2) * sizeof(WT_CKPT), &ckptbase));
/* Sort in creation-order. */
qsort(ckptbase, slot, sizeof(WT_CKPT), __ckpt_compare_order);
/* Return the array to our caller. */
*ckptbasep = ckptbase;
if (0) {
format: WT_ERR_MSG(session, WT_ERROR, "corrupted checkpoint list");
err: __wt_meta_ckptlist_free(session, ckptbase);
}
__wt_free(session, config);
__wt_scr_free(&buf);
return (ret);
}
/*
* wiredtiger_open --
* Main library entry point: open a new connection to a WiredTiger
* database.
*/
int
wiredtiger_open(const char *home, WT_EVENT_HANDLER *event_handler,
const char *config, WT_CONNECTION **wt_connp)
{
static WT_CONNECTION stdc = {
__conn_close,
__conn_reconfigure,
__conn_get_home,
__conn_is_new,
__conn_open_session,
__conn_load_extension,
__conn_add_data_source,
__conn_add_collator,
__conn_add_compressor,
__conn_add_extractor
};
static struct {
const char *name;
uint32_t flag;
} *ft, directio_types[] = {
{ "data", WT_DIRECTIO_DATA },
{ "log", WT_DIRECTIO_LOG },
{ NULL, 0 }
};
WT_CONFIG subconfig;
WT_CONFIG_ITEM cval, skey, sval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_ITEM *cbuf, expath, exconfig;
WT_SESSION_IMPL *session;
const char *cfg[] =
{ __wt_confdfl_wiredtiger_open, config, NULL, NULL, NULL };
int exist;
*wt_connp = NULL;
session = NULL;
cbuf = NULL;
WT_CLEAR(expath);
WT_CLEAR(exconfig);
WT_RET(__wt_library_init());
WT_RET(__wt_calloc_def(NULL, 1, &conn));
conn->iface = stdc;
/*
* Immediately link the structure into the connection structure list:
* the only thing ever looked at on that list is the database name,
* and a NULL value is fine.
*/
__wt_spin_lock(NULL, &__wt_process.spinlock);
TAILQ_INSERT_TAIL(&__wt_process.connqh, conn, q);
__wt_spin_unlock(NULL, &__wt_process.spinlock);
session = conn->default_session = &conn->dummy_session;
session->iface.connection = &conn->iface;
session->name = "wiredtiger_open";
__wt_event_handler_set(session, event_handler);
/* Remaining basic initialization of the connection structure. */
WT_ERR(__wt_connection_init(conn));
/* Check the configuration strings. */
WT_ERR(__wt_config_check(
session, __wt_confchk_wiredtiger_open, config, 0));
/* Get the database home. */
WT_ERR(__conn_home(session, home, cfg));
/* Make sure no other thread of control already owns this database. */
WT_ERR(__conn_single(session, cfg));
/* Read the database-home configuration file. */
WT_ERR(__conn_config_file(session, cfg, &cbuf));
/* Read the environment variable configuration. */
WT_ERR(__conn_config_env(session, cfg));
WT_ERR(__wt_config_gets(session, cfg, "hazard_max", &cval));
conn->hazard_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "session_max", &cval));
conn->session_size = (uint32_t)cval.val + WT_NUM_INTERNAL_SESSIONS;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge", &cval));
if (cval.val)
F_SET(conn, WT_CONN_LSM_MERGE);
WT_ERR(__wt_config_gets(session, cfg, "sync", &cval));
if (cval.val)
F_SET(conn, WT_CONN_SYNC);
WT_ERR(__wt_config_gets(session, cfg, "transactional", &cval));
if (cval.val)
F_SET(conn, WT_CONN_TRANSACTIONAL);
/* Configure verbose flags. */
WT_ERR(__conn_verbose_config(session, cfg));
WT_ERR(__wt_conn_cache_pool_config(session, cfg));
WT_ERR(__wt_config_gets(session, cfg, "logging", &cval));
if (cval.val != 0)
WT_ERR(__wt_open(
session, WT_LOG_FILENAME, 1, 0, 0, &conn->log_fh));
/* Configure direct I/O and buffer alignment. */
WT_ERR(__wt_config_gets(session, cfg, "buffer_alignment", &cval));
if (cval.val == -1)
conn->buffer_alignment = WT_BUFFER_ALIGNMENT_DEFAULT;
else
conn->buffer_alignment = (size_t)cval.val;
#ifndef HAVE_POSIX_MEMALIGN
if (conn->buffer_alignment != 0)
WT_ERR_MSG(session, EINVAL,
"buffer_alignment requires posix_memalign");
#endif
/*
* Configuration: direct_io, mmap, statistics.
*/
WT_ERR(__wt_config_gets(session, cfg, "direct_io", &cval));
for (ft = directio_types; ft->name != NULL; ft++) {
ret = __wt_config_subgets(session, &cval, ft->name, &sval);
if (ret == 0) {
if (sval.val)
FLD_SET(conn->direct_io, ft->flag);
} else if (ret != WT_NOTFOUND)
goto err;
}
WT_ERR(__wt_config_gets(session, cfg, "mmap", &cval));
conn->mmap = cval.val == 0 ? 0 : 1;
WT_ERR(__wt_config_gets(session, cfg, "statistics", &cval));
conn->statistics = cval.val == 0 ? 0 : 1;
/* Load any extensions referenced in the config. */
WT_ERR(__wt_config_gets(session, cfg, "extensions", &cval));
WT_ERR(__wt_config_subinit(session, &subconfig, &cval));
while ((ret = __wt_config_next(&subconfig, &skey, &sval)) == 0) {
WT_ERR(__wt_buf_fmt(
session, &expath, "%.*s", (int)skey.len, skey.str));
if (sval.len > 0)
WT_ERR(__wt_buf_fmt(session, &exconfig,
"entry=%.*s\n", (int)sval.len, sval.str));
WT_ERR(conn->iface.load_extension(&conn->iface,
expath.data, (sval.len > 0) ? exconfig.data : NULL));
}
WT_ERR_NOTFOUND_OK(ret);
/*
* Open the connection; if that fails, the connection handle has been
* destroyed by the time the open function returns.
*/
if ((ret = __wt_connection_open(conn, cfg)) != 0) {
conn = NULL;
WT_ERR(ret);
}
/* Open the default session. */
WT_ERR(__wt_open_session(conn, 1, NULL, NULL, &conn->default_session));
session = conn->default_session;
/*
* Check on the turtle and metadata files, creating them if necessary
* (which avoids application threads racing to create the metadata file
* later).
*/
WT_ERR(__wt_meta_turtle_init(session, &exist));
if (!exist) {
/*
* We're single-threaded, but acquire the schema lock
* regardless: the lower level code checks that it is
* appropriately synchronized.
*/
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_create(session, WT_METADATA_URI, NULL));
WT_ERR(ret);
}
WT_ERR(__wt_metadata_open(session));
/* If there's a hot-backup file, load it. */
WT_ERR(__wt_metadata_load_backup(session));
/*
* XXX LSM initialization.
* This is structured so that it could be moved to an extension.
*/
WT_ERR(__wt_lsm_init(&conn->iface, NULL));
STATIC_ASSERT(offsetof(WT_CONNECTION_IMPL, iface) == 0);
*wt_connp = &conn->iface;
/*
* Destroying the connection on error will destroy our session handle,
* cleanup using the session handle first, then discard the connection.
*/
err: if (cbuf != NULL)
__wt_buf_free(session, cbuf);
__wt_buf_free(session, &expath);
__wt_buf_free(session, &exconfig);
if (ret != 0 && conn != NULL)
WT_TRET(__wt_connection_destroy(conn));
/* Let the server threads proceed. */
if (ret == 0)
conn->connection_initialized = 1;
return (ret);
}
/*
* __find_next_col --
* Find the next column to use for a plan.
*/
static int
__find_next_col(WT_SESSION_IMPL *session, WT_TABLE *table,
WT_CONFIG_ITEM *colname, u_int *cgnump, u_int *colnump, char *coltype)
{
WT_COLGROUP *colgroup;
WT_CONFIG conf;
WT_CONFIG_ITEM cval, k, v;
WT_DECL_RET;
u_int cg, col, foundcg, foundcol, matchcg, matchcol;
bool getnext;
foundcg = foundcol = UINT_MAX;
matchcg = *cgnump;
matchcol = (*coltype == WT_PROJ_KEY) ?
*colnump : *colnump + table->nkey_columns;
getnext = true;
for (colgroup = NULL, cg = 0; cg < WT_COLGROUPS(table); cg++) {
colgroup = table->cgroups[cg];
/*
* If there is only one column group, we just scan through all
* of the columns. For tables with multiple column groups, we
* look at the key columns once, then go through the value
* columns for each group.
*/
if (cg == 0) {
cval = table->colconf;
col = 0;
} else {
cgcols: cval = colgroup->colconf;
col = table->nkey_columns;
}
__wt_config_subinit(session, &conf, &cval);
for (; (ret = __wt_config_next(&conf, &k, &v)) == 0; col++) {
if (k.len == colname->len &&
strncmp(colname->str, k.str, k.len) == 0) {
if (getnext) {
foundcg = cg;
foundcol = col;
}
getnext = cg == matchcg && col == matchcol;
}
if (cg == 0 && table->ncolgroups > 0 &&
col == table->nkey_columns - 1)
goto cgcols;
}
WT_RET_TEST(ret != WT_NOTFOUND, ret);
colgroup = NULL;
}
if (foundcg == UINT_MAX)
return (WT_NOTFOUND);
*cgnump = foundcg;
if (foundcol < table->nkey_columns) {
*coltype = WT_PROJ_KEY;
*colnump = foundcol;
} else {
*coltype = WT_PROJ_VALUE;
*colnump = foundcol - table->nkey_columns;
}
return (0);
}
/*
* __backup_uri --
* Backup a list of objects.
*/
static int
__backup_uri(WT_SESSION_IMPL *session,
const char *cfg[], bool *foundp, bool *log_only)
{
WT_CONFIG targetconf;
WT_CONFIG_ITEM cval, k, v;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
bool target_list;
const char *uri;
*foundp = *log_only = false;
/*
* If we find a non-empty target configuration string, we have a job,
* otherwise it's not our problem.
*/
WT_RET(__wt_config_gets(session, cfg, "target", &cval));
__wt_config_subinit(session, &targetconf, &cval);
for (target_list = false;
(ret = __wt_config_next(&targetconf, &k, &v)) == 0;
target_list = true) {
/* If it is our first time through, allocate. */
if (!target_list) {
*foundp = true;
WT_ERR(__wt_scr_alloc(session, 512, &tmp));
}
WT_ERR(__wt_buf_fmt(session, tmp, "%.*s", (int)k.len, k.str));
uri = tmp->data;
if (v.len != 0)
WT_ERR_MSG(session, EINVAL,
"%s: invalid backup target: URIs may need quoting",
uri);
/*
* Handle log targets. We do not need to go through the schema
* worker, just call the function to append them. Set log_only
* only if it is our only URI target.
*/
if (WT_PREFIX_MATCH(uri, "log:")) {
/*
* Log archive cannot mix with incremental backup, don't
* let that happen.
*/
if (FLD_ISSET(
S2C(session)->log_flags, WT_CONN_LOG_ARCHIVE))
WT_ERR_MSG(session, EINVAL,
"incremental backup not possible when "
"automatic log archival configured");
*log_only = !target_list;
WT_ERR(__backup_log_append(
session, session->bkp_cursor, false));
} else {
*log_only = false;
WT_ERR(__wt_schema_worker(session,
uri, NULL, __backup_list_uri_append, cfg, 0));
}
}
WT_ERR_NOTFOUND_OK(ret);
err: __wt_scr_free(session, &tmp);
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);
}
/*
* ___open_index --
* Open an index.
*/
static int
__open_index(WT_SESSION_IMPL *session, WT_TABLE *table, WT_INDEX *idx)
{
WT_CONFIG colconf;
WT_CONFIG_ITEM ckey, cval;
WT_DECL_ITEM(buf);
WT_DECL_ITEM(plan);
WT_DECL_RET;
u_int cursor_key_cols, i;
WT_ERR(__wt_scr_alloc(session, 0, &buf));
/* Get the data source from the index config. */
WT_ERR(__wt_config_getones(session, idx->config, "source", &cval));
WT_ERR(__wt_buf_fmt(
session, buf, "%.*s", (int)cval.len, cval.str));
idx->source = __wt_buf_steal(session, buf, NULL);
idx->need_value = WT_PREFIX_MATCH(idx->source, "lsm:");
WT_ERR(__wt_config_getones(session, idx->config, "key_format", &cval));
WT_ERR(__wt_buf_fmt(
session, buf, "%.*s", (int)cval.len, cval.str));
idx->key_format = __wt_buf_steal(session, buf, NULL);
/*
* 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 file's key.
*
* The file's key_format is stored persistently, we need to calculate
* the index cursor key format (which will usually omit some of those
* keys).
*/
WT_ERR(__wt_config_getones(
session, idx->config, "columns", &idx->colconf));
/* Start with the declared index columns. */
WT_ERR(__wt_config_subinit(session, &colconf, &idx->colconf));
cursor_key_cols = 0;
while ((ret = __wt_config_next(&colconf, &ckey, &cval)) == 0) {
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
++cursor_key_cols;
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
/*
* Now add any primary key columns from the table that are not
* already part of the index key.
*/
WT_ERR(__wt_config_subinit(session, &colconf, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&colconf, &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, &idx->colconf, &ckey, &cval) == 0)
continue;
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
WT_ERR(__wt_scr_alloc(session, 0, &plan));
WT_ERR(__wt_struct_plan(session, table, buf->data, buf->size, 0, plan));
idx->key_plan = __wt_buf_steal(session, plan, NULL);
/* Set up the cursor key format (the visible columns). */
WT_ERR(__wt_buf_init(session, buf, 0));
WT_ERR(__wt_struct_truncate(session,
idx->key_format, cursor_key_cols, buf));
idx->idxkey_format = __wt_buf_steal(session, buf, NULL);
/* By default, index cursor values are the table value columns. */
/* TODO Optimize to use index columns in preference to table lookups. */
WT_ERR(__wt_struct_plan(session,
table, table->colconf.str, table->colconf.len, 1, plan));
idx->value_plan = __wt_buf_steal(session, plan, NULL);
err: __wt_scr_free(&buf);
__wt_scr_free(&plan);
return (ret);
}
/*
* __wt_schema_open_table --
* Open a named table.
*/
int
__wt_schema_open_table(WT_SESSION_IMPL *session,
const char *name, size_t namelen, WT_TABLE **tablep)
{
WT_CONFIG cparser;
WT_CONFIG_ITEM ckey, cval;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_ITEM buf;
WT_TABLE *table;
const char *tconfig;
char *tablename;
cursor = NULL;
table = NULL;
WT_CLEAR(buf);
WT_RET(__wt_buf_fmt(session, &buf, "table:%.*s", (int)namelen, name));
tablename = __wt_buf_steal(session, &buf, NULL);
WT_ERR(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, tablename);
WT_ERR(cursor->search(cursor));
WT_ERR(cursor->get_value(cursor, &tconfig));
WT_ERR(__wt_calloc_def(session, 1, &table));
table->name = tablename;
tablename = NULL;
WT_ERR(__wt_config_getones(session, tconfig, "columns", &cval));
WT_ERR(__wt_config_getones(session, tconfig, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &table->key_format));
WT_ERR(__wt_config_getones(session, tconfig, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &table->value_format));
WT_ERR(__wt_strdup(session, tconfig, &table->config));
/* Point to some items in the copy to save re-parsing. */
WT_ERR(__wt_config_getones(session, table->config,
"columns", &table->colconf));
/*
* Count the number of columns: tables are "simple" if the columns
* are not named.
*/
WT_ERR(__wt_config_subinit(session, &cparser, &table->colconf));
table->is_simple = 1;
while ((ret = __wt_config_next(&cparser, &ckey, &cval)) == 0)
table->is_simple = 0;
if (ret != WT_NOTFOUND)
goto err;
/* Check that the columns match the key and value formats. */
if (!table->is_simple)
WT_ERR(__wt_schema_colcheck(session,
table->key_format, table->value_format, &table->colconf,
&table->nkey_columns, NULL));
WT_ERR(__wt_config_getones(session, table->config,
"colgroups", &table->cgconf));
/* Count the number of column groups. */
WT_ERR(__wt_config_subinit(session, &cparser, &table->cgconf));
table->ncolgroups = 0;
while ((ret = __wt_config_next(&cparser, &ckey, &cval)) == 0)
++table->ncolgroups;
if (ret != WT_NOTFOUND)
goto err;
WT_ERR(__wt_calloc_def(session, WT_COLGROUPS(table), &table->cgroups));
WT_ERR(__wt_schema_open_colgroups(session, table));
*tablep = table;
if (0) {
err: if (table != NULL)
__wt_schema_destroy_table(session, table);
}
if (cursor != NULL)
WT_TRET(cursor->close(cursor));
__wt_free(session, tablename);
return (ret);
}
/*
* __wt_schema_open_colgroups --
* Open the column groups for a table.
*/
int
__wt_schema_open_colgroups(WT_SESSION_IMPL *session, WT_TABLE *table)
{
WT_COLGROUP *colgroup;
WT_CONFIG cparser;
WT_CONFIG_ITEM ckey, cval;
WT_DECL_RET;
WT_DECL_ITEM(buf);
char *cgconfig;
u_int i;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_TABLE));
if (table->cg_complete)
return (0);
colgroup = NULL;
cgconfig = NULL;
WT_RET(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_config_subinit(session, &cparser, &table->cgconf));
/* Open each column group. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
if (table->ncolgroups > 0)
WT_ERR(__wt_config_next(&cparser, &ckey, &cval));
else
WT_CLEAR(ckey);
/*
* Always open from scratch: we may have failed part of the way
* through opening a table, or column groups may have changed.
*/
__wt_schema_destroy_colgroup(session, &table->cgroups[i]);
WT_ERR(__wt_buf_init(session, buf, 0));
WT_ERR(__wt_schema_colgroup_name(session, table,
ckey.str, ckey.len, buf));
if ((ret = __wt_metadata_search(
session, buf->data, &cgconfig)) != 0) {
/* It is okay if the table is incomplete. */
if (ret == WT_NOTFOUND)
ret = 0;
goto err;
}
WT_ERR(__wt_calloc_one(session, &colgroup));
WT_ERR(__wt_strndup(
session, buf->data, buf->size, &colgroup->name));
colgroup->config = cgconfig;
cgconfig = NULL;
WT_ERR(__wt_config_getones(session,
colgroup->config, "columns", &colgroup->colconf));
WT_ERR(__wt_config_getones(
session, colgroup->config, "source", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &colgroup->source));
table->cgroups[i] = colgroup;
colgroup = NULL;
}
if (!table->is_simple) {
WT_ERR(__wt_table_check(session, table));
WT_ERR(__wt_buf_init(session, buf, 0));
WT_ERR(__wt_struct_plan(session,
table, table->colconf.str, table->colconf.len, true, buf));
WT_ERR(__wt_strndup(
session, buf->data, buf->size, &table->plan));
}
table->cg_complete = true;
err: __wt_scr_free(session, &buf);
__wt_schema_destroy_colgroup(session, &colgroup);
if (cgconfig != NULL)
__wt_free(session, cgconfig);
return (ret);
}
/*
* __wt_lsm_meta_read --
* Read the metadata for an LSM tree.
*/
int
__wt_lsm_meta_read(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_CONFIG cparser, lparser;
WT_CONFIG_ITEM ck, cv, fileconf, lk, lv, metadata;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
char *lsmconfig;
u_int nchunks;
chunk = NULL; /* -Wconditional-uninitialized */
/* LSM trees inherit the merge setting from the connection. */
if (F_ISSET(S2C(session), WT_CONN_LSM_MERGE))
F_SET(lsm_tree, WT_LSM_TREE_MERGES);
WT_RET(__wt_metadata_search(session, lsm_tree->name, &lsmconfig));
WT_ERR(__wt_config_init(session, &cparser, lsmconfig));
while ((ret = __wt_config_next(&cparser, &ck, &cv)) == 0) {
if (WT_STRING_MATCH("key_format", ck.str, ck.len)) {
__wt_free(session, lsm_tree->key_format);
WT_ERR(__wt_strndup(session,
cv.str, cv.len, &lsm_tree->key_format));
} else if (WT_STRING_MATCH("value_format", ck.str, ck.len)) {
__wt_free(session, lsm_tree->value_format);
WT_ERR(__wt_strndup(session,
cv.str, cv.len, &lsm_tree->value_format));
} else if (WT_STRING_MATCH("collator", ck.str, ck.len)) {
if (cv.len == 0 ||
WT_STRING_MATCH("none", cv.str, cv.len))
continue;
/*
* Extract the application-supplied metadata (if any)
* from the file configuration.
*/
WT_ERR(__wt_config_getones(
session, lsmconfig, "file_config", &fileconf));
WT_CLEAR(metadata);
WT_ERR_NOTFOUND_OK(__wt_config_subgets(
session, &fileconf, "app_metadata", &metadata));
WT_ERR(__wt_collator_config(session, lsm_tree->name,
&cv, &metadata,
&lsm_tree->collator, &lsm_tree->collator_owned));
WT_ERR(__wt_strndup(session,
cv.str, cv.len, &lsm_tree->collator_name));
} else if (WT_STRING_MATCH("bloom_config", ck.str, ck.len)) {
__wt_free(session, lsm_tree->bloom_config);
/* Don't include the brackets. */
WT_ERR(__wt_strndup(session,
cv.str + 1, cv.len - 2, &lsm_tree->bloom_config));
} else if (WT_STRING_MATCH("file_config", ck.str, ck.len)) {
__wt_free(session, lsm_tree->file_config);
/* Don't include the brackets. */
WT_ERR(__wt_strndup(session,
cv.str + 1, cv.len - 2, &lsm_tree->file_config));
} else if (WT_STRING_MATCH("auto_throttle", ck.str, ck.len)) {
if (cv.val)
F_SET(lsm_tree, WT_LSM_TREE_THROTTLE);
else
F_CLR(lsm_tree, WT_LSM_TREE_THROTTLE);
} else if (WT_STRING_MATCH("bloom", ck.str, ck.len))
lsm_tree->bloom = (uint32_t)cv.val;
else if (WT_STRING_MATCH("bloom_bit_count", ck.str, ck.len))
lsm_tree->bloom_bit_count = (uint32_t)cv.val;
else if (WT_STRING_MATCH("bloom_hash_count", ck.str, ck.len))
lsm_tree->bloom_hash_count = (uint32_t)cv.val;
else if (WT_STRING_MATCH("chunk_count_limit", ck.str, ck.len)) {
lsm_tree->chunk_count_limit = (uint32_t)cv.val;
if (cv.val != 0)
F_CLR(lsm_tree, WT_LSM_TREE_MERGES);
} else if (WT_STRING_MATCH("chunk_max", ck.str, ck.len))
lsm_tree->chunk_max = (uint64_t)cv.val;
else if (WT_STRING_MATCH("chunk_size", ck.str, ck.len))
lsm_tree->chunk_size = (uint64_t)cv.val;
else if (WT_STRING_MATCH("merge_max", ck.str, ck.len))
lsm_tree->merge_max = (uint32_t)cv.val;
else if (WT_STRING_MATCH("merge_min", ck.str, ck.len))
lsm_tree->merge_min = (uint32_t)cv.val;
else if (WT_STRING_MATCH("last", ck.str, ck.len))
lsm_tree->last = (u_int)cv.val;
else if (WT_STRING_MATCH("chunks", ck.str, ck.len)) {
WT_ERR(__wt_config_subinit(session, &lparser, &cv));
for (nchunks = 0; (ret =
__wt_config_next(&lparser, &lk, &lv)) == 0; ) {
if (WT_STRING_MATCH("id", lk.str, lk.len)) {
WT_ERR(__wt_realloc_def(session,
&lsm_tree->chunk_alloc,
nchunks + 1, &lsm_tree->chunk));
WT_ERR(
__wt_calloc_one(session, &chunk));
lsm_tree->chunk[nchunks++] = chunk;
chunk->id = (uint32_t)lv.val;
WT_ERR(__wt_lsm_tree_chunk_name(session,
lsm_tree, chunk->id, &chunk->uri));
F_SET(chunk,
WT_LSM_CHUNK_ONDISK |
WT_LSM_CHUNK_STABLE);
} else if (WT_STRING_MATCH(
"bloom", lk.str, lk.len)) {
WT_ERR(__wt_lsm_tree_bloom_name(
session, lsm_tree,
chunk->id, &chunk->bloom_uri));
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
continue;
} else if (WT_STRING_MATCH(
"chunk_size", lk.str, lk.len)) {
chunk->size = (uint64_t)lv.val;
continue;
} else if (WT_STRING_MATCH(
"count", lk.str, lk.len)) {
chunk->count = (uint64_t)lv.val;
continue;
} else if (WT_STRING_MATCH(
"generation", lk.str, lk.len)) {
chunk->generation = (uint32_t)lv.val;
continue;
}
}
WT_ERR_NOTFOUND_OK(ret);
lsm_tree->nchunks = nchunks;
} else if (WT_STRING_MATCH("old_chunks", ck.str, ck.len)) {
WT_ERR(__wt_config_subinit(session, &lparser, &cv));
for (nchunks = 0; (ret =
__wt_config_next(&lparser, &lk, &lv)) == 0; ) {
if (WT_STRING_MATCH("bloom", lk.str, lk.len)) {
WT_ERR(__wt_strndup(session,
lv.str, lv.len, &chunk->bloom_uri));
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
continue;
}
WT_ERR(__wt_realloc_def(session,
&lsm_tree->old_alloc, nchunks + 1,
&lsm_tree->old_chunks));
WT_ERR(__wt_calloc_one(session, &chunk));
lsm_tree->old_chunks[nchunks++] = chunk;
WT_ERR(__wt_strndup(session,
lk.str, lk.len, &chunk->uri));
F_SET(chunk, WT_LSM_CHUNK_ONDISK);
}
WT_ERR_NOTFOUND_OK(ret);
lsm_tree->nold_chunks = nchunks;
}
/*
* Ignore any other values: the metadata entry might have been
* created by a future release, with unknown options.
*/
}
WT_ERR_NOTFOUND_OK(ret);
/*
* If the default merge_min was not overridden, calculate it now. We
* do this here so that trees created before merge_min was added get a
* sane value.
*/
if (lsm_tree->merge_min < 2)
lsm_tree->merge_min = WT_MAX(2, lsm_tree->merge_max / 2);
err: __wt_free(session, lsmconfig);
return (ret);
}
/*
* __open_index --
* Open an index.
*/
static int
__open_index(WT_SESSION_IMPL *session, WT_TABLE *table, WT_INDEX *idx)
{
WT_CONFIG colconf;
WT_CONFIG_ITEM ckey, cval, metadata;
WT_DECL_ITEM(buf);
WT_DECL_ITEM(plan);
WT_DECL_RET;
u_int npublic_cols, i;
WT_ERR(__wt_scr_alloc(session, 0, &buf));
/* Get the data source from the index config. */
WT_ERR(__wt_config_getones(session, idx->config, "source", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &idx->source));
WT_ERR(__wt_config_getones(session, idx->config, "immutable", &cval));
if (cval.val)
F_SET(idx, WT_INDEX_IMMUTABLE);
/*
* Compatibility: we didn't always maintain collator information in
* index metadata, cope when it isn't found.
*/
WT_CLEAR(cval);
WT_ERR_NOTFOUND_OK(__wt_config_getones(
session, idx->config, "collator", &cval));
if (cval.len != 0) {
WT_CLEAR(metadata);
WT_ERR_NOTFOUND_OK(__wt_config_getones(
session, idx->config, "app_metadata", &metadata));
WT_ERR(__wt_collator_config(
session, idx->name, &cval, &metadata,
&idx->collator, &idx->collator_owned));
}
WT_ERR(__wt_extractor_config(
session, idx->name, idx->config, &idx->extractor,
&idx->extractor_owned));
WT_ERR(__wt_config_getones(session, idx->config, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &idx->key_format));
/*
* 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 file's key.
*
* The file's key_format is stored persistently, we need to calculate
* the index cursor key format (which will usually omit some of those
* keys).
*/
WT_ERR(__wt_buf_init(session, buf, 0));
WT_ERR(__wt_config_getones(
session, idx->config, "columns", &idx->colconf));
/* Start with the declared index columns. */
WT_ERR(__wt_config_subinit(session, &colconf, &idx->colconf));
for (npublic_cols = 0;
(ret = __wt_config_next(&colconf, &ckey, &cval)) == 0;
++npublic_cols)
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
if (ret != WT_NOTFOUND)
goto err;
/*
* If we didn't find any columns, the index must have an extractor.
* We don't rely on this unconditionally because it was only added to
* the metadata after version 2.3.1.
*/
if (npublic_cols == 0) {
WT_ERR(__wt_config_getones(
session, idx->config, "index_key_columns", &cval));
npublic_cols = (u_int)cval.val;
WT_ASSERT(session, npublic_cols != 0);
for (i = 0; i < npublic_cols; i++)
WT_ERR(__wt_buf_catfmt(session, buf, "\"bad col\","));
}
/*
* Now add any primary key columns from the table that are not
* already part of the index key.
*/
WT_ERR(__wt_config_subinit(session, &colconf, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&colconf, &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, &idx->colconf, &ckey, &cval) == 0)
continue;
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
}
WT_ERR_NOTFOUND_OK(ret);
/*
* If the table doesn't yet have its column groups, don't try to
* calculate a plan: we are just checking that the index creation is
* sane.
*/
if (!table->cg_complete)
goto err;
WT_ERR(__wt_scr_alloc(session, 0, &plan));
WT_ERR(__wt_struct_plan(
session, table, buf->data, buf->size, false, plan));
WT_ERR(__wt_strndup(session, plan->data, plan->size, &idx->key_plan));
/* Set up the cursor key format (the visible columns). */
WT_ERR(__wt_buf_init(session, buf, 0));
WT_ERR(__wt_struct_truncate(session,
idx->key_format, npublic_cols, buf));
WT_ERR(__wt_strndup(
session, buf->data, buf->size, &idx->idxkey_format));
/*
* Add a trailing padding byte to the format. This ensures that there
* will be no special optimization of the last column, so the primary
* key columns can be simply appended.
*/
WT_ERR(__wt_buf_catfmt(session, buf, "x"));
WT_ERR(__wt_strndup(session, buf->data, buf->size, &idx->exkey_format));
/* By default, index cursor values are the table value columns. */
/* TODO Optimize to use index columns in preference to table lookups. */
WT_ERR(__wt_buf_init(session, plan, 0));
WT_ERR(__wt_struct_plan(session,
table, table->colconf.str, table->colconf.len, true, plan));
WT_ERR(__wt_strndup(session, plan->data, plan->size, &idx->value_plan));
err: __wt_scr_free(session, &buf);
__wt_scr_free(session, &plan);
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);
}
/*
* __schema_open_table --
* Open a named table (internal version).
*/
static int
__schema_open_table(WT_SESSION_IMPL *session,
const char *name, size_t namelen, bool ok_incomplete, WT_TABLE **tablep)
{
WT_CONFIG cparser;
WT_CONFIG_ITEM ckey, cval;
WT_CURSOR *cursor;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_TABLE *table;
const char *tconfig;
char *tablename;
cursor = NULL;
table = NULL;
tablename = NULL;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_TABLE));
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf, "table:%.*s", (int)namelen, name));
WT_ERR(__wt_strndup(session, buf->data, buf->size, &tablename));
WT_ERR(__wt_metadata_cursor(session, &cursor));
cursor->set_key(cursor, tablename);
WT_ERR(cursor->search(cursor));
WT_ERR(cursor->get_value(cursor, &tconfig));
WT_ERR(__wt_calloc_one(session, &table));
table->name = tablename;
tablename = NULL;
table->name_hash = __wt_hash_city64(name, namelen);
WT_ERR(__wt_config_getones(session, tconfig, "columns", &cval));
WT_ERR(__wt_config_getones(session, tconfig, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &table->key_format));
WT_ERR(__wt_config_getones(session, tconfig, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &table->value_format));
WT_ERR(__wt_strdup(session, tconfig, &table->config));
/* Point to some items in the copy to save re-parsing. */
WT_ERR(__wt_config_getones(session, table->config,
"columns", &table->colconf));
/*
* Count the number of columns: tables are "simple" if the columns
* are not named.
*/
WT_ERR(__wt_config_subinit(session, &cparser, &table->colconf));
table->is_simple = true;
while ((ret = __wt_config_next(&cparser, &ckey, &cval)) == 0)
table->is_simple = false;
if (ret != WT_NOTFOUND)
goto err;
/* Check that the columns match the key and value formats. */
if (!table->is_simple)
WT_ERR(__wt_schema_colcheck(session,
table->key_format, table->value_format, &table->colconf,
&table->nkey_columns, NULL));
WT_ERR(__wt_config_getones(session, table->config,
"colgroups", &table->cgconf));
/* Count the number of column groups. */
WT_ERR(__wt_config_subinit(session, &cparser, &table->cgconf));
table->ncolgroups = 0;
while ((ret = __wt_config_next(&cparser, &ckey, &cval)) == 0)
++table->ncolgroups;
if (ret != WT_NOTFOUND)
goto err;
if (table->ncolgroups > 0 && table->is_simple)
WT_ERR_MSG(session, EINVAL,
"%s requires a table with named columns", tablename);
WT_ERR(__wt_calloc_def(session, WT_COLGROUPS(table), &table->cgroups));
WT_ERR(__wt_schema_open_colgroups(session, table));
if (!ok_incomplete && !table->cg_complete)
WT_ERR_MSG(session, EINVAL, "'%s' cannot be used "
"until all column groups are created",
table->name);
/* Copy the schema generation into the new table. */
table->schema_gen = S2C(session)->schema_gen;
*tablep = table;
if (0) {
err: WT_TRET(__wt_schema_destroy_table(session, &table));
}
WT_TRET(__wt_metadata_cursor_release(session, &cursor));
__wt_free(session, tablename);
__wt_scr_free(session, &buf);
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);
}
