/*
* __backup_uri --
* Backup a list of objects.
*/
static int
__backup_uri(WT_SESSION_IMPL *session,
WT_CURSOR_BACKUP *cb, 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_RET(__wt_config_subinit(session, &targetconf, &cval));
for (cb->list_next = 0, 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_only = !target_list;
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);
}
/*
* __verify_config --
* Verification supports dumping pages in various formats.
*/
static int
__verify_config(WT_SESSION_IMPL *session, const char *cfg[], WT_VSTUFF *vs)
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
ret = __wt_config_gets(session, cfg, "dump_address", &cval);
if (ret != 0 && ret != WT_NOTFOUND)
WT_RET(ret);
if (ret == 0 && cval.val != 0)
vs->dump_address = 1;
ret = __wt_config_gets(session, cfg, "dump_blocks", &cval);
if (ret != 0 && ret != WT_NOTFOUND)
WT_RET(ret);
if (ret == 0 && cval.val != 0)
vs->dump_blocks = 1;
ret = __wt_config_gets(session, cfg, "dump_pages", &cval);
if (ret != 0 && ret != WT_NOTFOUND)
WT_RET(ret);
if (ret == 0 && cval.val != 0)
vs->dump_pages = 1;
#ifdef HAVE_DIAGNOSTIC
/*
* We use the verification code to do debugging dumps because if we're
* dumping in debugging mode, we want to confirm the page is OK before
* walking it.
*/
#else
if (vs->dump_address || vs->dump_blocks || vs->dump_pages)
WT_RET_MSG(session, ENOTSUP,
"the WiredTiger library was not built in diagnostic mode");
#endif
return (0);
}
/*
* __ckpt_server_config --
* Parse and setup the checkpoint server options.
*/
static int
__ckpt_server_config(WT_SESSION_IMPL *session, const char **cfg, int *startp)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
conn = S2C(session);
/*
* The checkpoint configuration requires a wait time -- if it's not set,
* we're not running at all.
*/
WT_RET(__wt_config_gets(session, cfg, "checkpoint.wait", &cval));
if (cval.val == 0) {
*startp = 0;
return (0);
}
conn->ckpt_usecs = (long)cval.val * 1000000;
*startp = 1;
WT_RET(__wt_config_gets(session, cfg, "checkpoint.name", &cval));
if (!WT_STRING_MATCH(WT_CHECKPOINT, cval.str, cval.len)) {
WT_RET(__wt_scr_alloc(session, cval.len + 20, &tmp));
strcpy((char *)tmp->data, "name=");
strncat((char *)tmp->data, cval.str, cval.len);
ret = __wt_strndup(session,
tmp->data, strlen("name=") + cval.len, &conn->ckpt_config);
__wt_scr_free(&tmp);
WT_RET(ret);
}
return (0);
}
/*
* __logmgr_sync_cfg --
* Interpret the transaction_sync config.
*/
static int
__logmgr_sync_cfg(WT_SESSION_IMPL *session, const char **cfg)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
conn = S2C(session);
WT_RET(
__wt_config_gets(session, cfg, "transaction_sync.enabled", &cval));
if (cval.val)
FLD_SET(conn->txn_logsync, WT_LOG_FLUSH);
else
FLD_CLR(conn->txn_logsync, WT_LOG_FLUSH);
WT_RET(
__wt_config_gets(session, cfg, "transaction_sync.method", &cval));
FLD_CLR(conn->txn_logsync, WT_LOG_DSYNC | WT_LOG_FSYNC);
if (WT_STRING_MATCH("dsync", cval.str, cval.len))
FLD_SET(conn->txn_logsync, WT_LOG_DSYNC);
else if (WT_STRING_MATCH("fsync", cval.str, cval.len))
FLD_SET(conn->txn_logsync, WT_LOG_FSYNC);
return (0);
}
/*
* __verify_config --
* Debugging: verification supports dumping pages in various formats.
*/
static int
__verify_config(WT_SESSION_IMPL *session, const char *cfg[], WT_VSTUFF *vs)
{
WT_CONFIG_ITEM cval;
WT_RET(__wt_config_gets(session, cfg, "dump_address", &cval));
vs->dump_address = cval.val != 0;
WT_RET(__wt_config_gets(session, cfg, "dump_blocks", &cval));
vs->dump_blocks = cval.val != 0;
WT_RET(__wt_config_gets(session, cfg, "dump_pages", &cval));
vs->dump_pages = cval.val != 0;
WT_RET(__wt_config_gets(session, cfg, "dump_shape", &cval));
vs->dump_shape = cval.val != 0;
#if !defined(HAVE_DIAGNOSTIC)
if (vs->dump_blocks || vs->dump_pages)
WT_RET_MSG(session, ENOTSUP,
"the WiredTiger library was not built in diagnostic mode");
#endif
return (0);
}
/*从cfg_arg中读取key对应的value值，并用cval返回*/
int __wt_ext_config_get(WT_EXTENSION_API *wt_api, WT_SESSION *wt_session, WT_CONFIG_ARG *cfg_arg, const char *key, WT_CONFIG_ITEM *cval)
{
WT_CONNECTION_IMPL *conn;
WT_SESSION_IMPL *session;
const char **cfg;
conn = (WT_CONNECTION_IMPL *)wt_api->conn;
if ((session = (WT_SESSION_IMPL *)wt_session) == NULL)
session = conn->default_session;
if ((cfg = (const char **)cfg_arg) == NULL)
return WT_NOTFOUND;
return __wt_config_gets(session, cfg, key, cval);
}
/*
* __conn_load_extension --
* WT_CONNECTION->load_extension method.
*/
static int
__conn_load_extension(
WT_CONNECTION *wt_conn, const char *path, const char *config)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_DLH *dlh;
WT_SESSION_IMPL *session;
int (*entry)(WT_SESSION *, WT_EXTENSION_API *, const char *);
const char *entry_name;
dlh = NULL;
conn = (WT_CONNECTION_IMPL *)wt_conn;
CONNECTION_API_CALL(conn, session, load_extension, config, cfg);
entry_name = NULL;
WT_ERR(__wt_config_gets(session, cfg, "entry", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &entry_name));
/*
* This assumes the underlying shared libraries are reference counted,
* that is, that re-opening a shared library simply increments a ref
* count, and closing it simply decrements the ref count, and the last
* close discards the reference entirely -- in other words, we do not
* check to see if we've already opened this shared library.
*/
WT_ERR(__wt_dlopen(session, path, &dlh));
WT_ERR(__wt_dlsym(session, dlh, entry_name, &entry));
/* Call the entry function. */
WT_ERR(entry(&session->iface, &__api, config));
/* Link onto the environment's list of open libraries. */
__wt_spin_lock(session, &conn->api_lock);
TAILQ_INSERT_TAIL(&conn->dlhqh, dlh, q);
__wt_spin_unlock(session, &conn->api_lock);
if (0) {
err: if (dlh != NULL)
WT_TRET(__wt_dlclose(session, dlh));
}
__wt_free(session, entry_name);
API_END_NOTFOUND_MAP(session, ret);
}
/*
* __wt_config_gets_def --
* Performance hack: skip parsing config strings by hard-coding defaults.
*
* It's expensive to repeatedly parse configuration strings, so don't do
* it unless it's necessary in performance paths like cursor creation.
* Assume the second configuration string is the application's
* configuration string, and if it's not set (which is true most of the
* time), then use the supplied default value. This makes it faster to
* open cursors when checking for obscure open configuration strings like
* "next_random".
*/
int
__wt_config_gets_def(WT_SESSION_IMPL *session,
const char **cfg, const char *key, int def, WT_CONFIG_ITEM *value)
{
static const WT_CONFIG_ITEM false_value = {
"", 0, 0, WT_CONFIG_ITEM_NUM
};
*value = false_value;
value->val = def;
if (cfg == NULL || cfg[0] == NULL || cfg[1] == NULL)
return (0);
else if (cfg[2] == NULL)
WT_RET_NOTFOUND_OK(
__wt_config_getones(session, cfg[1], key, value));
return (__wt_config_gets(session, cfg, key, value));
}
/*
* __wt_conn_compat_config --
* Configure compatibility version.
*/
int
__wt_conn_compat_config(WT_SESSION_IMPL *session, const char **cfg)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
uint16_t patch;
bool txn_active;
conn = S2C(session);
WT_RET(__wt_config_gets(session, cfg, "compatibility.release", &cval));
if (cval.len == 0) {
conn->compat_major = WIREDTIGER_VERSION_MAJOR;
conn->compat_minor = WIREDTIGER_VERSION_MINOR;
return (0);
}
/*
* Accept either a major.minor release string or a major.minor.patch
* release string. We ignore the patch value, but allow it in the
* string.
*/
if (sscanf(cval.str, "%" SCNu16 ".%" SCNu16,
&conn->compat_major, &conn->compat_minor) != 2 &&
sscanf(cval.str, "%" SCNu16 ".%" SCNu16 ".%" SCNu16,
&conn->compat_major, &conn->compat_minor, &patch) != 3)
WT_RET_MSG(session, EINVAL, "illegal compatibility release");
if (conn->compat_major > WIREDTIGER_VERSION_MAJOR)
WT_RET_MSG(session, ENOTSUP, "unsupported major version");
if (conn->compat_major == WIREDTIGER_VERSION_MAJOR &&
conn->compat_minor > WIREDTIGER_VERSION_MINOR)
WT_RET_MSG(session, ENOTSUP, "unsupported minor version");
/*
* We're doing an upgrade or downgrade, check whether transactions are
* active.
*/
WT_RET(__wt_txn_activity_check(session, &txn_active));
if (txn_active)
WT_RET_MSG(session, ENOTSUP,
"system must be quiescent for upgrade or downgrade");
return (0);
}
/*
* __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);
}
/*
* __conn_reconfigure --
* WT_CONNECTION->reconfigure method.
*/
static int
__conn_reconfigure(WT_CONNECTION *wt_conn, const char *config)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_SESSION_IMPL *session;
/*
* Special version of cfg that doesn't include the default config: used
* to limit changes to values that the application sets explicitly.
* Note that any function using this value has to be prepared to handle
* not-found as a valid option return.
*/
const char *raw_cfg[] = { config, NULL };
conn = (WT_CONNECTION_IMPL *)wt_conn;
CONNECTION_API_CALL(conn, session, reconfigure, config, cfg);
/* Turning on statistics clears any existing values. */
if ((ret =
__wt_config_gets(session, raw_cfg, "statistics", &cval)) == 0) {
conn->statistics = cval.val == 0 ? 0 : 1;
if (conn->statistics)
__wt_stat_clear_connection_stats(&conn->stats);
}
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_conn_cache_pool_config(session, cfg));
WT_ERR(__wt_cache_config(conn, raw_cfg));
WT_ERR(__conn_verbose_config(session, raw_cfg));
/* Wake up the cache pool server so any changes are noticed. */
if (F_ISSET(conn, WT_CONN_CACHE_POOL))
WT_ERR(__wt_cond_signal(
session, __wt_process.cache_pool->cache_pool_cond));
err: API_END(session);
return (ret);
}
/*
* __conn_verbose_config --
* Set verbose configuration.
*/
static int
__conn_verbose_config(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval, sval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
static struct {
const char *name;
uint32_t flag;
} *ft, verbtypes[] = {
{ "block", WT_VERB_block },
{ "shared_cache",WT_VERB_shared_cache },
{ "ckpt", WT_VERB_ckpt },
{ "evict", WT_VERB_evict },
{ "evictserver",WT_VERB_evictserver },
{ "fileops", WT_VERB_fileops },
{ "hazard", WT_VERB_hazard },
{ "lsm", WT_VERB_lsm },
{ "mutex", WT_VERB_mutex },
{ "read", WT_VERB_read },
{ "reconcile", WT_VERB_reconcile },
{ "salvage", WT_VERB_salvage },
{ "verify", WT_VERB_verify },
{ "write", WT_VERB_write },
{ NULL, 0 }
};
conn = S2C(session);
if ((ret = __wt_config_gets(session, cfg, "verbose", &cval)) != 0)
return (ret == WT_NOTFOUND ? 0 : ret);
for (ft = verbtypes; ft->name != NULL; ft++) {
if ((ret = __wt_config_subgets(
session, &cval, ft->name, &sval)) == 0 && sval.val != 0)
FLD_SET(conn->verbose, ft->flag);
else
FLD_CLR(conn->verbose, ft->flag);
WT_RET_NOTFOUND_OK(ret);
}
return (0);
}
/*读取evict LRU cache的配置*/
static int __cache_config_local(WT_SESSION_IMPL* session, int shared, const char* cfg[])
{
WT_CACHE* cache;
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
uint32_t evict_workers_max, evict_workers_min;
conn = S2C(session);
cache = conn->cache;
/*读取cache size*/
if (!shared){
WT_RET(__wt_config_gets(session, cfg, "cache_size", &cval));
conn->cache_size = (uint64_t)cval.val;
}
WT_RET(__wt_config_gets(session, cfg, "cache_overhead", &cval));
cache->overhead_pct = (u_int)cval.val;
WT_RET(__wt_config_gets(session, cfg, "eviction_target", &cval));
cache->eviction_target = (u_int)cval.val;
WT_RET(__wt_config_gets(session, cfg, "eviction_trigger", &cval));
cache->eviction_trigger = (u_int)cval.val;
WT_RET(__wt_config_gets(session, cfg, "eviction_dirty_target", &cval));
cache->eviction_dirty_target = (u_int)cval.val;
WT_RET(__wt_config_gets(session, cfg, "eviction.threads_max", &cval));
WT_ASSERT(session, cval.val > 0);
evict_workers_max = (uint32_t)cval.val - 1;
WT_RET(__wt_config_gets(session, cfg, "eviction.threads_min", &cval));
WT_ASSERT(session, cval.val > 0);
evict_workers_min = (uint32_t)cval.val - 1;
/*合法性校验*/
if (evict_workers_min > evict_workers_max)
WT_RET_MSG(session, EINVAL, "eviction=(threads_min) cannot be greater than eviction=(threads_max)");
conn->evict_workers_max = evict_workers_max;
conn->evict_workers_min = evict_workers_min;
return 0;
}
/*
* __wt_snapshot --
* Snapshot the tree.
*/
int
__wt_snapshot(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
char *name;
name = NULL;
/* This may be a named snapshot, check the configuration. */
if ((ret = __wt_config_gets(
session, cfg, "snapshot", &cval)) != 0 && ret != WT_NOTFOUND)
WT_RET(ret);
if (cval.len != 0)
WT_RET(__wt_strndup(session, cval.str, cval.len, &name));
ret = __snapshot_worker(session, name, 0, SNAPSHOT);
__wt_free(session, name);
return (ret);
}
/*
* __lsm_tree_open_check --
* Validate the configuration of an LSM tree.
*/
static int
__lsm_tree_open_check(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_CONFIG_ITEM cval;
uint64_t maxleafpage, required;
const char *cfg[] = { WT_CONFIG_BASE(
session, session_create), lsm_tree->file_config, NULL };
WT_RET(__wt_config_gets(session, cfg, "leaf_page_max", &cval));
maxleafpage = (uint64_t)cval.val;
/* Three chunks, plus one page for each participant in a merge. */
required = 3 * lsm_tree->chunk_size +
lsm_tree->merge_threads * (lsm_tree->merge_max * maxleafpage);
if (S2C(session)->cache_size < required)
WT_RET_MSG(session, EINVAL,
"The LSM configuration requires a cache size of at least %"
PRIu64 ". Configured size is %" PRIu64,
required, S2C(session)->cache_size);
return (0);
}
/*
* __drop_file --
* Drop a file.
*/
static int
__drop_file(
WT_SESSION_IMPL *session, const char *uri, int force, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
int exist, remove_files;
const char *filename;
WT_RET(__wt_config_gets(session, cfg, "remove_files", &cval));
remove_files = (cval.val != 0);
filename = uri;
if (!WT_PREFIX_SKIP(filename, "file:"))
return (EINVAL);
/* Close all btree handles associated with this file. */
WT_WITH_DHANDLE_LOCK(session,
ret = __wt_conn_dhandle_close_all(session, uri, force));
WT_RET(ret);
/* Remove the metadata entry (ignore missing items). */
WT_TRET(__wt_metadata_remove(session, uri));
if (!remove_files)
return (ret);
/* Remove the underlying physical file. */
exist = 0;
WT_TRET(__wt_exist(session, filename, &exist));
if (exist) {
/*
* There is no point tracking this operation: there is no going
* back from here.
*/
WT_TRET(__wt_remove(session, filename));
}
return (ret);
}
/*
* __conn_config_env --
* Read configuration from an environment variable, if set.
*/
static int
__conn_config_env(WT_SESSION_IMPL *session, const char **cfg)
{
WT_CONFIG_ITEM cval;
const char *env_config;
if ((env_config = getenv("WIREDTIGER_CONFIG")) == NULL ||
strlen(env_config) == 0)
return (0);
/*
* Security stuff:
*
* If the "use_environment_priv" configuration string is set, use the
* environment variable if the process has appropriate privileges.
*/
WT_RET(__wt_config_gets(session, cfg, "use_environment_priv", &cval));
if (cval.val == 0 && __wt_has_priv())
WT_RET_MSG(session, WT_ERROR, "%s",
"WIREDTIGER_CONFIG environment variable set but process "
"lacks privileges to use that environment variable");
/* Check the configuration string. */
WT_RET(__wt_config_check(
session, __wt_confchk_wiredtiger_open, env_config, 0));
/*
* The environment setting comes second-to-last: it overrides the
* WiredTiger.config file (if any), but not the config passed by the
* application.
*/
while (cfg[1] != NULL)
++cfg;
cfg[1] = cfg[0];
cfg[0] = env_config;
return (0);
}
/*
* __wt_config_gets_def --
* Performance hack: skip parsing config strings by hard-coding defaults.
*
* It's expensive to repeatedly parse configuration strings, so don't do
* it unless it's necessary in performance paths like cursor creation.
* Assume the second configuration string is the application's
* configuration string, and if it's not set (which is true most of the
* time), then use the supplied default value. This makes it faster to
* open cursors when checking for obscure open configuration strings like
* "next_random".
*/
int
__wt_config_gets_def(WT_SESSION_IMPL *session,
const char **cfg, const char *key, int def, WT_CONFIG_ITEM *value)
{
WT_CONFIG_ITEM_STATIC_INIT(false_value);
const char **end;
*value = false_value;
value->val = def;
if (cfg == NULL)
return (0);
/*
* Checking the "length" of the pointer array is a little odd, but it's
* deliberate. The reason is because we pass variable length arrays of
* pointers as the configuration argument, some of which have only one
* element and the NULL termination. Static analyzers (like Coverity)
* complain if we read from an offset past the end of the array, even
* if we check there's no NULL slots before the offset.
*/
for (end = cfg; *end != NULL; ++end)
;
switch ((int)(end - cfg)) {
case 0: /* cfg[0] == NULL */
case 1: /* cfg[1] == NULL */
return (0);
case 2: /* cfg[2] == NULL */
WT_RET_NOTFOUND_OK(
__wt_config_getones(session, cfg[1], key, value));
return (0);
default:
return (__wt_config_gets(session, cfg, key, value));
}
/* NOTREACHED */
}
int
__wt_schema_drop(WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_CONFIG_ITEM cval;
int force, ret;
cval.val = 0;
ret = __wt_config_gets(session, cfg, "force", &cval);
if (ret != 0 && ret != WT_NOTFOUND)
WT_RET(ret);
force = cval.val == 0 ? 0 : 1;
/* Disallow drops from the WiredTiger name space. */
WT_RET(__wt_schema_name_check(session, uri));
if (WT_PREFIX_MATCH(uri, "colgroup:"))
ret = __drop_colgroup(session, uri, force, cfg);
else if (WT_PREFIX_MATCH(uri, "file:"))
ret = __drop_file(session, uri, force);
else if (WT_PREFIX_MATCH(uri, "index:"))
ret = __drop_index(session, uri, force, cfg);
else if (WT_PREFIX_MATCH(uri, "table:"))
ret = __drop_table(session, uri, force);
else
return (__wt_unknown_object_type(session, uri));
/*
* Map WT_NOTFOUND to ENOENT (or to 0 if "force" is set), based on the
* assumption WT_NOTFOUND means there was no schema file entry. The
* underlying drop functions should handle this case (we passed them
* the "force" value), but better safe than sorry.
*/
if (ret == WT_NOTFOUND)
ret = force ? 0 : ENOENT;
return (ret);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] = API_CONF_DEFAULTS(session, create, config);
const char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_read(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_RET(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_newest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_NEWEST);
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
(FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_NEWEST) ||
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST)))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_config", &cval));
if (cval.type == ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_chunk_size", &cval));
lsm_tree->chunk_size = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge_threads", &cval));
lsm_tree->merge_threads = (uint32_t)cval.val;
/* Sanity check that api_data.py is in sync with lsm.h */
WT_ASSERT(session, lsm_tree->merge_threads <= WT_LSM_MAX_WORKERS);
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s,key_format=u,value_format=u", config));
lsm_tree->file_config = __wt_buf_steal(session, buf, NULL);
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}
/*
* __wt_conn_compat_config --
* Configure compatibility version.
*/
int
__wt_conn_compat_config(
WT_SESSION_IMPL *session, const char **cfg, bool reconfig)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
uint16_t max_major, max_minor, min_major, min_minor;
uint16_t rel_major, rel_minor;
char *value;
bool txn_active;
conn = S2C(session);
value = NULL;
max_major = WT_CONN_COMPAT_NONE;
max_minor = WT_CONN_COMPAT_NONE;
min_major = WT_CONN_COMPAT_NONE;
min_minor = WT_CONN_COMPAT_NONE;
WT_RET(__wt_config_gets(session, cfg, "compatibility.release", &cval));
if (cval.len == 0) {
rel_major = WIREDTIGER_VERSION_MAJOR;
rel_minor = WIREDTIGER_VERSION_MINOR;
F_CLR(conn, WT_CONN_COMPATIBILITY);
} else {
WT_RET(__conn_compat_parse(
session, &cval, &rel_major, &rel_minor));
/*
* We're doing an upgrade or downgrade, check whether
* transactions are active.
*/
WT_RET(__wt_txn_activity_check(session, &txn_active));
if (txn_active)
WT_RET_MSG(session, ENOTSUP,
"system must be quiescent"
" for upgrade or downgrade");
F_SET(conn, WT_CONN_COMPATIBILITY);
}
/*
* If we're a reconfigure and the user did not set any compatibility,
* we're done.
*/
if (reconfig && !F_ISSET(conn, WT_CONN_COMPATIBILITY))
goto done;
/*
* The maximum and minimum required version for existing files
* is only available on opening the connection, not reconfigure.
*/
WT_RET(__wt_config_gets(session,
cfg, "compatibility.require_min", &cval));
if (cval.len != 0)
WT_RET(__conn_compat_parse(
session, &cval, &min_major, &min_minor));
WT_RET(__wt_config_gets(session,
cfg, "compatibility.require_max", &cval));
if (cval.len != 0)
WT_RET(__conn_compat_parse(
session, &cval, &max_major, &max_minor));
/*
* The maximum required must be greater than or equal to the
* compatibility release we're using now. This is on an open and we're
* checking the two against each other. We'll check against what was
* saved on a restart later.
*/
if (!reconfig && max_major != WT_CONN_COMPAT_NONE &&
(max_major < rel_major ||
(max_major == rel_major && max_minor < rel_minor)))
WT_RET_MSG(session, ENOTSUP,
WT_COMPAT_MSG_PREFIX
"required max of %" PRIu16 ".%" PRIu16
"cannot be smaller than compatibility release %"
PRIu16 ".%" PRIu16,
max_major, max_minor, rel_major, rel_minor);
/*
* The minimum required must be less than or equal to the compatibility
* release we're using now. This is on an open and we're checking the
* two against each other. We'll check against what was saved on a
* restart later.
*/
if (!reconfig && min_major != WT_CONN_COMPAT_NONE &&
(min_major > rel_major ||
(min_major == rel_major && min_minor > rel_minor)))
WT_RET_MSG(session, ENOTSUP,
WT_COMPAT_MSG_PREFIX
"required min of %" PRIu16 ".%" PRIu16
"cannot be larger than compatibility release %"
PRIu16 ".%" PRIu16,
min_major, min_minor, rel_major, rel_minor);
/*
* On a reconfigure, check the new release version against any
* required maximum version set on open.
*/
if (reconfig && conn->req_max_major != WT_CONN_COMPAT_NONE &&
(conn->req_max_major < rel_major ||
(conn->req_max_major == rel_major &&
conn->req_max_minor < rel_minor)))
WT_RET_MSG(session, ENOTSUP,
WT_COMPAT_MSG_PREFIX
"required max of %" PRIu16 ".%" PRIu16
"cannot be smaller than requested compatibility release %"
PRIu16 ".%" PRIu16,
conn->req_max_major, conn->req_max_minor,
rel_major, rel_minor);
/*
* On a reconfigure, check the new release version against any
* required minimum version set on open.
*/
if (reconfig && conn->req_min_major != WT_CONN_COMPAT_NONE &&
(conn->req_min_major > rel_major ||
(conn->req_min_major == rel_major &&
conn->req_min_minor > rel_minor)))
WT_RET_MSG(session, ENOTSUP,
WT_COMPAT_MSG_PREFIX
"required min of %" PRIu16 ".%" PRIu16
"cannot be larger than requested compatibility release %"
PRIu16 ".%" PRIu16,
conn->req_min_major, conn->req_min_minor,
rel_major, rel_minor);
conn->compat_major = rel_major;
conn->compat_minor = rel_minor;
/*
* Only rewrite the turtle file if this is a reconfig. On startup
* it will get written as part of creating the connection. We do this
* after checking the required minimum version so that we don't rewrite
* the turtle file if there is an error.
*/
if (reconfig)
WT_RET(__wt_metadata_turtle_rewrite(session));
/*
* The required maximum and minimum cannot be set via reconfigure and
* they are meaningless on a newly created database. We're done in
* those cases.
*/
if (reconfig || conn->is_new ||
(min_major == WT_CONN_COMPAT_NONE &&
max_major == WT_CONN_COMPAT_NONE))
goto done;
/*
* Check the minimum required against any saved compatibility version
* in the turtle file saved from an earlier run.
*/
rel_major = rel_minor = WT_CONN_COMPAT_NONE;
if ((ret =
__wt_metadata_search(session, WT_METADATA_COMPAT, &value)) == 0) {
WT_ERR(__wt_config_getones(session, value, "major", &cval));
rel_major = (uint16_t)cval.val;
WT_ERR(__wt_config_getones(session, value, "minor", &cval));
rel_minor = (uint16_t)cval.val;
if (max_major != WT_CONN_COMPAT_NONE &&
(max_major < rel_major ||
(max_major == rel_major && max_minor < rel_minor)))
WT_ERR_MSG(session, ENOTSUP,
WT_COMPAT_MSG_PREFIX
"required max of %" PRIu16 ".%" PRIu16
"cannot be larger than saved release %"
PRIu16 ".%" PRIu16,
max_major, max_minor, rel_major, rel_minor);
if (min_major != WT_CONN_COMPAT_NONE &&
(min_major > rel_major ||
(min_major == rel_major && min_minor > rel_minor)))
WT_ERR_MSG(session, ENOTSUP,
WT_COMPAT_MSG_PREFIX
"required min of %" PRIu16 ".%" PRIu16
"cannot be larger than saved release %"
PRIu16 ".%" PRIu16,
min_major, min_minor, rel_major, rel_minor);
} else if (ret == WT_NOTFOUND)
ret = 0;
else
WT_ERR(ret);
done: conn->req_max_major = max_major;
conn->req_max_minor = max_minor;
conn->req_min_major = min_major;
conn->req_min_minor = min_minor;
err: __wt_free(session, value);
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);
}
/*
* __logmgr_config --
* Parse and setup the logging server options.
*/
static int
__logmgr_config(
WT_SESSION_IMPL *session, const char **cfg, bool *runp, bool reconfig)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
bool enabled;
conn = S2C(session);
WT_RET(__wt_config_gets(session, cfg, "log.enabled", &cval));
enabled = cval.val != 0;
/*
* If we're reconfiguring, enabled must match the already
* existing setting.
*
* If it is off and the user it turning it on, or it is on
* and the user is turning it off, return an error.
*/
if (reconfig &&
((enabled && !FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED)) ||
(!enabled && FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED))))
return (EINVAL);
/* Logging is incompatible with in-memory */
if (enabled) {
WT_RET(__wt_config_gets(session, cfg, "in_memory", &cval));
if (cval.val != 0)
WT_RET_MSG(session, EINVAL,
"In memory configuration incompatible with "
"log=(enabled=true)");
}
*runp = enabled;
/*
* Setup a log path and compression even if logging is disabled in case
* we are going to print a log. Only do this on creation. Once a
* compressor or log path are set they cannot be changed.
*/
if (!reconfig) {
conn->log_compressor = NULL;
WT_RET(__wt_config_gets_none(
session, cfg, "log.compressor", &cval));
WT_RET(__wt_compressor_config(
session, &cval, &conn->log_compressor));
WT_RET(__wt_config_gets(session, cfg, "log.path", &cval));
WT_RET(__wt_strndup(
session, cval.str, cval.len, &conn->log_path));
}
/* We are done if logging isn't enabled. */
if (!*runp)
return (0);
WT_RET(__wt_config_gets(session, cfg, "log.archive", &cval));
if (cval.val != 0)
FLD_SET(conn->log_flags, WT_CONN_LOG_ARCHIVE);
if (!reconfig) {
/*
* Ignore if the user tries to change the file size. The
* amount of memory allocated to the log slots may be based
* on the log file size at creation and we don't want to
* re-allocate that memory while running.
*/
WT_RET(__wt_config_gets(session, cfg, "log.file_max", &cval));
conn->log_file_max = (wt_off_t)cval.val;
WT_STAT_FAST_CONN_SET(session,
log_max_filesize, conn->log_file_max);
}
/*
* If pre-allocation is configured, set the initial number to a few.
* We'll adapt as load dictates.
*/
WT_RET(__wt_config_gets(session, cfg, "log.prealloc", &cval));
if (cval.val != 0)
conn->log_prealloc = 1;
/*
* Note that it is meaningless to reconfigure this value during
* runtime. It only matters on create before recovery runs.
*/
WT_RET(__wt_config_gets_def(session, cfg, "log.recover", 0, &cval));
if (cval.len != 0 && WT_STRING_MATCH("error", cval.str, cval.len))
FLD_SET(conn->log_flags, WT_CONN_LOG_RECOVER_ERR);
WT_RET(__wt_config_gets(session, cfg, "log.zero_fill", &cval));
if (cval.val != 0) {
if (F_ISSET(conn, WT_CONN_READONLY))
WT_RET_MSG(session, EINVAL,
"Read-only configuration incompatible with "
"zero-filling log files");
FLD_SET(conn->log_flags, WT_CONN_LOG_ZERO_FILL);
}
WT_RET(__logmgr_sync_cfg(session, cfg));
if (conn->log_cond != NULL)
WT_RET(__wt_cond_auto_signal(session, conn->log_cond));
return (0);
}
/*
* __conn_single --
* Confirm that no other thread of control is using this database.
*/
static int
__conn_single(WT_SESSION_IMPL *session, const char **cfg)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn, *t;
WT_DECL_RET;
off_t size;
uint32_t len;
int created;
char buf[256];
conn = S2C(session);
/*
* Optionally create the wiredtiger flag file if it doesn't already
* exist. We don't actually care if we create it or not, the "am I the
* only locker" tests are all that matter.
*/
WT_RET(__wt_config_gets(session, cfg, "create", &cval));
WT_RET(__wt_open(session,
WT_SINGLETHREAD, cval.val == 0 ? 0 : 1, 0, 0, &conn->lock_fh));
/*
* Lock a byte of the file: if we don't get the lock, some other process
* is holding it, we're done. Note the file may be zero-length length,
* and that's OK, the underlying call supports acquisition of locks past
* the end-of-file.
*/
if (__wt_bytelock(conn->lock_fh, (off_t)0, 1) != 0)
WT_ERR_MSG(session, EBUSY, "%s",
"WiredTiger database is already being managed by another "
"process");
/* Check to see if another thread of control has this database open. */
__wt_spin_lock(session, &__wt_process.spinlock);
TAILQ_FOREACH(t, &__wt_process.connqh, q)
if (t->home != NULL &&
t != conn && strcmp(t->home, conn->home) == 0) {
ret = EBUSY;
break;
}
__wt_spin_unlock(session, &__wt_process.spinlock);
if (ret != 0)
WT_ERR_MSG(session, EBUSY, "%s",
"WiredTiger database is already being managed by another "
"thread in this process");
/*
* If the size of the file is 0, we created it (or we're racing with
* the thread that created it, it doesn't matter), write some bytes
* into the file. Strictly speaking, this isn't even necessary, but
* zero-length files always make me nervous.
*/
WT_ERR(__wt_filesize(session, conn->lock_fh, &size));
if (size == 0) {
len = (uint32_t)snprintf(buf, sizeof(buf), "%s\n%s\n",
WT_SINGLETHREAD, wiredtiger_version(NULL, NULL, NULL));
WT_ERR(__wt_write(
session, conn->lock_fh, (off_t)0, (uint32_t)len, buf));
created = 1;
} else
created = 0;
/*
* If we found a zero-length WiredTiger lock file, and eventually ended
* as the database owner, return that we created the database. (There
* is a theoretical chance that another process created the WiredTiger
* lock file but we won the race to add the WT_CONNECTION_IMPL structure
* to the process' list. It doesn't much matter, only one thread will
* be told it created the database.)
*/
conn->is_new = created;
return (0);
err: if (conn->lock_fh != NULL) {
WT_TRET(__wt_close(session, conn->lock_fh));
conn->lock_fh = NULL;
}
return (ret);
}
/*
* __wt_session_compact --
* WT_SESSION.compact method.
*/
int
__wt_session_compact(
WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_COMPACT_STATE compact;
WT_CONFIG_ITEM cval;
WT_DATA_SOURCE *dsrc;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
bool ignore_cache_size_set;
ignore_cache_size_set = false;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, compact, config, cfg);
/*
* The compaction thread should not block when the cache is full: it is
* holding locks blocking checkpoints and once the cache is full, it can
* spend a long time doing eviction.
*/
if (!F_ISSET(session, WT_SESSION_IGNORE_CACHE_SIZE)) {
ignore_cache_size_set = true;
F_SET(session, WT_SESSION_IGNORE_CACHE_SIZE);
}
/* In-memory ignores compaction operations. */
if (F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
goto err;
/*
* Non-LSM object compaction requires checkpoints, which are impossible
* in transactional contexts. Disallow in all contexts (there's no
* reason for LSM to allow this, possible or not), and check now so the
* error message isn't confusing.
*/
WT_ERR(__wt_txn_context_check(session, false));
/* Disallow objects in the WiredTiger name space. */
WT_ERR(__wt_str_name_check(session, uri));
if (!WT_PREFIX_MATCH(uri, "colgroup:") &&
!WT_PREFIX_MATCH(uri, "file:") &&
!WT_PREFIX_MATCH(uri, "index:") &&
!WT_PREFIX_MATCH(uri, "lsm:") &&
!WT_PREFIX_MATCH(uri, "table:")) {
if ((dsrc = __wt_schema_get_source(session, uri)) != NULL)
ret = dsrc->compact == NULL ?
__wt_object_unsupported(session, uri) :
dsrc->compact(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg);
else
ret = __wt_bad_object_type(session, uri);
goto err;
}
/* Setup the session handle's compaction state structure. */
memset(&compact, 0, sizeof(WT_COMPACT_STATE));
session->compact = &compact;
/* Compaction can be time-limited. */
WT_ERR(__wt_config_gets(session, cfg, "timeout", &cval));
session->compact->max_time = (uint64_t)cval.val;
__wt_epoch(session, &session->compact->begin);
/* Find the types of data sources being compacted. */
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, uri,
__compact_handle_append, __compact_uri_analyze, cfg, 0));
WT_ERR(ret);
if (session->compact->lsm_count != 0)
WT_ERR(__wt_schema_worker(
session, uri, NULL, __wt_lsm_compact, cfg, 0));
if (session->compact->file_count != 0)
WT_ERR(__compact_worker(session));
err: session->compact = NULL;
for (i = 0; i < session->op_handle_next; ++i) {
WT_WITH_DHANDLE(session, session->op_handle[i],
WT_TRET(__compact_end(session)));
WT_WITH_DHANDLE(session, session->op_handle[i],
WT_TRET(__wt_session_release_dhandle(session)));
}
__wt_free(session, session->op_handle);
session->op_handle_allocated = session->op_handle_next = 0;
/*
* Release common session resources (for example, checkpoint may acquire
* significant reconciliation structures/memory).
*/
WT_TRET(__wt_session_release_resources(session));
if (ignore_cache_size_set)
F_CLR(session, WT_SESSION_IGNORE_CACHE_SIZE);
if (ret != 0)
WT_STAT_CONN_INCR(session, session_table_compact_fail);
else
WT_STAT_CONN_INCR(session, session_table_compact_success);
API_END_RET_NOTFOUND_MAP(session, ret);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_create), config, NULL };
char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_search(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "collator", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->collator_name));
WT_ERR(__wt_config_gets(session, cfg, "lsm.auto_throttle", &cval));
if (cval.val)
F_SET(lsm_tree, WT_LSM_TREE_THROTTLE);
else
F_CLR(lsm_tree, WT_LSM_TREE_THROTTLE);
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_config", &cval));
if (cval.type == WT_CONFIG_ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_max", &cval));
lsm_tree->chunk_max = (uint64_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_size", &cval));
lsm_tree->chunk_size = (uint64_t)cval.val;
if (lsm_tree->chunk_size > lsm_tree->chunk_max)
WT_ERR_MSG(session, EINVAL,
"Chunk size (chunk_size) must be smaller than or equal to "
"the maximum chunk size (chunk_max)");
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_min", &cval));
lsm_tree->merge_min = (uint32_t)cval.val;
if (lsm_tree->merge_min > lsm_tree->merge_max)
WT_ERR_MSG(session, EINVAL,
"LSM merge_min must be less than or equal to merge_max");
/*
* Set up the config for each chunk.
*
* Make the memory_page_max double the chunk size, so application
* threads don't immediately try to force evict the chunk when the
* worker thread clears the NO_EVICTION flag.
*/
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s,key_format=u,value_format=u,memory_page_max=%" PRIu64,
config, 2 * lsm_tree->chunk_max));
WT_ERR(__wt_strndup(
session, buf->data, buf->size, &lsm_tree->file_config));
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}
/*
* __wt_block_open --
* Open a block handle.
*/
int
__wt_block_open(WT_SESSION_IMPL *session,
const char *filename, const char *cfg[],
int forced_salvage, int readonly, uint32_t allocsize, WT_BLOCK **blockp)
{
WT_BLOCK *block;
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_TRET(__wt_verbose(session, WT_VERB_BLOCK, "open: %s", filename));
conn = S2C(session);
*blockp = NULL;
__wt_spin_lock(session, &conn->block_lock);
TAILQ_FOREACH(block, &conn->blockqh, q)
if (strcmp(filename, block->name) == 0) {
++block->ref;
*blockp = block;
__wt_spin_unlock(session, &conn->block_lock);
return (0);
}
/* Basic structure allocation, initialization. */
WT_ERR(__wt_calloc_def(session, 1, &block));
block->ref = 1;
TAILQ_INSERT_HEAD(&conn->blockqh, block, q);
WT_ERR(__wt_strdup(session, filename, &block->name));
block->allocsize = allocsize;
WT_ERR(__wt_config_gets(session, cfg, "block_allocation", &cval));
block->allocfirst =
WT_STRING_MATCH("first", cval.str, cval.len) ? 1 : 0;
/* Configuration: optional OS buffer cache maximum size. */
WT_ERR(__wt_config_gets(session, cfg, "os_cache_max", &cval));
block->os_cache_max = (size_t)cval.val;
#ifdef HAVE_POSIX_FADVISE
if (conn->direct_io && block->os_cache_max)
WT_ERR_MSG(session, EINVAL,
"os_cache_max not supported in combination with direct_io");
#else
if (block->os_cache_max)
WT_ERR_MSG(session, EINVAL,
"os_cache_max not supported if posix_fadvise not "
"available");
#endif
/* Configuration: optional immediate write scheduling flag. */
WT_ERR(__wt_config_gets(session, cfg, "os_cache_dirty_max", &cval));
block->os_cache_dirty_max = (size_t)cval.val;
#ifdef HAVE_SYNC_FILE_RANGE
if (conn->direct_io && block->os_cache_dirty_max)
WT_ERR_MSG(session, EINVAL,
"os_cache_dirty_max not supported in combination with "
"direct_io");
#else
if (block->os_cache_dirty_max) {
/*
* Ignore any setting if it is not supported.
*/
block->os_cache_dirty_max = 0;
WT_ERR(__wt_verbose(session, WT_VERB_BLOCK,
"os_cache_dirty_max ignored when sync_file_range not "
"available"));
}
#endif
/* Open the underlying file handle. */
WT_ERR(__wt_open(session, filename, 0, 0,
readonly ? WT_FILE_TYPE_CHECKPOINT : WT_FILE_TYPE_DATA,
&block->fh));
/* Initialize the live checkpoint's lock. */
WT_ERR(__wt_spin_init(session, &block->live_lock, "block manager"));
/*
* Read the description information from the first block.
*
* Salvage is a special case: if we're forcing the salvage, we don't
* look at anything, including the description information.
*/
if (!forced_salvage)
WT_ERR(__desc_read(session, block));
*blockp = block;
__wt_spin_unlock(session, &conn->block_lock);
return (0);
err: WT_TRET(__block_destroy(session, block));
__wt_spin_unlock(session, &conn->block_lock);
return (ret);
}
/*
* __wt_conn_optrack_setup --
* Set up operation logging.
*/
int
__wt_conn_optrack_setup(WT_SESSION_IMPL *session,
const char *cfg[], bool reconfig)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_ITEM(buf);
WT_DECL_RET;
conn = S2C(session);
/* Once an operation tracking path has been set it can't be changed. */
if (!reconfig) {
WT_RET(__wt_config_gets(session,
cfg, "operation_tracking.path", &cval));
WT_RET(__wt_strndup(session,
cval.str, cval.len, &conn->optrack_path));
}
WT_RET(__wt_config_gets(session,
cfg, "operation_tracking.enabled", &cval));
if (cval.val == 0) {
if (F_ISSET(conn, WT_CONN_OPTRACK)) {
WT_RET(__wt_conn_optrack_teardown(session, reconfig));
F_CLR(conn, WT_CONN_OPTRACK);
}
return (0);
}
if (F_ISSET(conn, WT_CONN_READONLY))
/* Operation tracking isn't supported in read-only mode */
WT_RET_MSG(session, EINVAL,
"Operation tracking is incompatible with read only "
"configuration.");
if (F_ISSET(conn, WT_CONN_OPTRACK))
/* Already enabled, nothing else to do */
return (0);
/*
* Operation tracking files will include the ID of the creating process
* in their name, so we can distinguish between log files created by
* different WiredTiger processes in the same directory. We cache the
* process id for future use.
*/
conn->optrack_pid = __wt_process_id();
/*
* Open the file in the same directory that will hold a map of
* translations between function names and function IDs. If the file
* exists, remove it.
*/
WT_RET(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_filename_construct(session, conn->optrack_path,
"optrack-map", conn->optrack_pid, UINT32_MAX, buf));
WT_ERR(__wt_open(session,
(const char *)buf->data, WT_FS_OPEN_FILE_TYPE_REGULAR,
WT_FS_OPEN_CREATE, &conn->optrack_map_fh));
WT_ERR(__wt_spin_init(session,
&conn->optrack_map_spinlock, "optrack map spinlock"));
WT_ERR(__wt_malloc(session, WT_OPTRACK_BUFSIZE,
&conn->dummy_session.optrack_buf));
/* Set operation tracking on */
F_SET(conn, WT_CONN_OPTRACK);
err: __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);
}
/*
* __wt_conn_statistics_config --
* Set statistics configuration.
*/
int
__wt_conn_statistics_config(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval, sval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
uint32_t flags;
int set;
conn = S2C(session);
WT_RET(__wt_config_gets(session, cfg, "statistics", &cval));
flags = 0;
set = 0;
if ((ret = __wt_config_subgets(
session, &cval, "none", &sval)) == 0 && sval.val != 0) {
flags = 0;
++set;
}
WT_RET_NOTFOUND_OK(ret);
if ((ret = __wt_config_subgets(
session, &cval, "fast", &sval)) == 0 && sval.val != 0) {
LF_SET(WT_STAT_TYPE_FAST);
++set;
}
WT_RET_NOTFOUND_OK(ret);
if ((ret = __wt_config_subgets(
session, &cval, "all", &sval)) == 0 && sval.val != 0) {
LF_SET(
WT_STAT_TYPE_ALL | WT_STAT_TYPE_CACHE_WALK |
WT_STAT_TYPE_FAST | WT_STAT_TYPE_TREE_WALK);
++set;
}
WT_RET_NOTFOUND_OK(ret);
if (set > 1)
WT_RET_MSG(session, EINVAL,
"Only one of all, fast, none configuration values should "
"be specified");
/*
* Now that we've parsed general statistics categories, process
* sub-categories.
*/
if ((ret = __wt_config_subgets(
session, &cval, "cache_walk", &sval)) == 0 && sval.val != 0)
/*
* Configuring cache walk statistics implies fast statistics.
* Keep that knowledge internal for now - it may change in the
* future.
*/
LF_SET(WT_STAT_TYPE_FAST | WT_STAT_TYPE_CACHE_WALK);
WT_RET_NOTFOUND_OK(ret);
if ((ret = __wt_config_subgets(
session, &cval, "tree_walk", &sval)) == 0 && sval.val != 0)
/*
* Configuring tree walk statistics implies fast statistics.
* Keep that knowledge internal for now - it may change in the
* future.
*/
LF_SET(WT_STAT_TYPE_FAST | WT_STAT_TYPE_TREE_WALK);
WT_RET_NOTFOUND_OK(ret);
if ((ret = __wt_config_subgets(
session, &cval, "clear", &sval)) == 0 && sval.val != 0) {
if (!LF_ISSET(WT_STAT_TYPE_ALL | WT_STAT_TYPE_CACHE_WALK |
WT_STAT_TYPE_FAST | WT_STAT_TYPE_TREE_WALK))
WT_RET_MSG(session, EINVAL,
"the value \"clear\" can only be specified if "
"statistics are enabled");
LF_SET(WT_STAT_CLEAR);
}
WT_RET_NOTFOUND_OK(ret);
/* Configuring statistics clears any existing values. */
conn->stat_flags = flags;
return (0);
}