/*
* __recovery_file_scan --
* Scan the files referenced from the metadata and gather information
* about them for recovery.
*/
static int
__recovery_file_scan(WT_RECOVERY *r)
{
WT_DECL_RET;
WT_CURSOR *c;
const char *uri, *config;
int cmp;
/* Scan through all files in the metadata. */
c = r->files[0].c;
c->set_key(c, "file:");
if ((ret = c->search_near(c, &cmp)) != 0) {
/* Is the metadata empty? */
if (ret == WT_NOTFOUND)
ret = 0;
goto err;
}
if (cmp < 0)
WT_ERR_NOTFOUND_OK(c->next(c));
for (; ret == 0; ret = c->next(c)) {
WT_ERR(c->get_key(c, &uri));
if (!WT_PREFIX_MATCH(uri, "file:"))
break;
WT_ERR(c->get_value(c, &config));
WT_ERR(__recovery_setup_file(r, uri, config));
}
WT_ERR_NOTFOUND_OK(ret);
err: return (ret);
}
/*
* __backup_all --
* Backup all objects in the database.
*/
static int
__backup_all(WT_SESSION_IMPL *session, WT_CURSOR_BACKUP *cb)
{
WT_CONFIG_ITEM cval;
WT_CURSOR *cursor;
WT_DECL_RET;
const char *key, *value;
cursor = NULL;
/*
* Open a cursor on the metadata file and copy all of the entries to
* the hot backup file.
*/
WT_ERR(__wt_metadata_cursor(session, NULL, &cursor));
while ((ret = cursor->next(cursor)) == 0) {
WT_ERR(cursor->get_key(cursor, &key));
WT_ERR(cursor->get_value(cursor, &value));
WT_ERR_TEST((fprintf(
cb->bfp, "%s\n%s\n", key, value) < 0), __wt_errno());
/*
* While reading the metadata file, check there are no "sources"
* or "types" which can't support hot backup. This checks for
* a data source that's non-standard, which can't be backed up,
* but is also sanity checking: if there's an entry backed by
* anything other than a file or lsm entry, we're confused.
*/
if ((ret = __wt_config_getones(
session, value, "type", &cval)) == 0 &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "file") &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "lsm"))
WT_ERR_MSG(session, ENOTSUP,
"hot backup is not supported for objects of "
"type %.*s", (int)cval.len, cval.str);
WT_ERR_NOTFOUND_OK(ret);
if ((ret =__wt_config_getones(
session, value, "source", &cval)) == 0 &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "file:") &&
!WT_PREFIX_MATCH_LEN(cval.str, cval.len, "lsm:"))
WT_ERR_MSG(session, ENOTSUP,
"hot backup is not supported for objects of "
"source %.*s", (int)cval.len, cval.str);
WT_ERR_NOTFOUND_OK(ret);
}
WT_ERR_NOTFOUND_OK(ret);
/* Build a list of the file objects that need to be copied. */
WT_WITH_DHANDLE_LOCK(session,
ret = __wt_meta_btree_apply(
session, __backup_list_all_append, NULL));
err: if (cursor != NULL)
WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __recovery_set_checkpoint_timestamp --
* Set the checkpoint timestamp as retrieved from the metadata file.
*/
static int
__recovery_set_checkpoint_timestamp(WT_RECOVERY *r)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_SESSION_IMPL *session;
wt_timestamp_t ckpt_timestamp;
char ts_string[WT_TS_INT_STRING_SIZE], *sys_config;
sys_config = NULL;
session = r->session;
conn = S2C(session);
/*
* Read the system checkpoint information from the metadata file and
* save the stable timestamp of the last checkpoint for later query.
* This gets saved in the connection.
*/
ckpt_timestamp = 0;
/* Search in the metadata for the system information. */
WT_ERR_NOTFOUND_OK(
__wt_metadata_search(session, WT_SYSTEM_CKPT_URI, &sys_config));
if (sys_config != NULL) {
WT_CLEAR(cval);
WT_ERR_NOTFOUND_OK(__wt_config_getones(
session, sys_config, "checkpoint_timestamp", &cval));
if (cval.len != 0) {
__wt_verbose(session, WT_VERB_RECOVERY,
"Recovery timestamp %.*s",
(int)cval.len, cval.str);
WT_ERR(__wt_txn_parse_timestamp_raw(session,
"recovery", &ckpt_timestamp, &cval));
}
}
/*
* Set the recovery checkpoint timestamp and the metadata checkpoint
* timestamp so that the checkpoint after recovery writes the correct
* value into the metadata.
*/
conn->txn_global.meta_ckpt_timestamp =
conn->txn_global.recovery_timestamp = ckpt_timestamp;
if (WT_VERBOSE_ISSET(session,
WT_VERB_RECOVERY | WT_VERB_RECOVERY_PROGRESS)) {
__wt_timestamp_to_string(
conn->txn_global.recovery_timestamp, ts_string);
__wt_verbose(session,
WT_VERB_RECOVERY | WT_VERB_RECOVERY_PROGRESS,
"Set global recovery timestamp: %s", ts_string);
}
err: __wt_free(session, sys_config);
return (ret);
}
/*
* __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);
}
/*
* __wt_las_remove_block --
* Remove all records matching a key prefix from the lookaside store.
*/
int
__wt_las_remove_block(WT_SESSION_IMPL *session,
WT_CURSOR *cursor, uint32_t btree_id, const uint8_t *addr, size_t addr_size)
{
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_RET;
uint64_t las_counter, las_txnid;
uint32_t las_id;
int exact;
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
/*
* Search for the block's unique prefix and step through all matching
* records, removing them.
*/
las_addr->data = addr;
las_addr->size = addr_size;
las_key->size = 0;
cursor->set_key(
cursor, btree_id, las_addr, (uint64_t)0, (uint32_t)0, las_key);
if ((ret = cursor->search_near(cursor, &exact)) == 0 && exact < 0)
ret = cursor->next(cursor);
for (; ret == 0; ret = cursor->next(cursor)) {
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* Confirm the search using the unique prefix; if not a match,
* we're done searching for records for this page.
*/
if (las_id != btree_id ||
las_addr->size != addr_size ||
memcmp(las_addr->data, addr, addr_size) != 0)
break;
/*
* Cursor opened overwrite=true: won't return WT_NOTFOUND should
* another thread remove the record before we do, and the cursor
* remains positioned in that case.
*/
WT_ERR(cursor->remove(cursor));
}
WT_ERR_NOTFOUND_OK(ret);
err: __wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
return (ret);
}
/*
* __clsm_next --
* WT_CURSOR->next method for the LSM cursor type.
*/
static int
__clsm_next(WT_CURSOR *cursor)
{
WT_CURSOR_LSM *clsm;
WT_CURSOR *c;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
int cmp;
bool check, deleted;
clsm = (WT_CURSOR_LSM *)cursor;
CURSOR_API_CALL(cursor, session, next, NULL);
WT_CURSOR_NOVALUE(cursor);
WT_ERR(__clsm_enter(clsm, false, false));
/* If we aren't positioned for a forward scan, get started. */
if (clsm->current == NULL || !F_ISSET(clsm, WT_CLSM_ITERATE_NEXT)) {
F_CLR(clsm, WT_CLSM_MULTIPLE);
WT_FORALL_CURSORS(clsm, c, i) {
if (!F_ISSET(cursor, WT_CURSTD_KEY_SET)) {
WT_ERR(c->reset(c));
ret = c->next(c);
} else if (c != clsm->current) {
c->set_key(c, &cursor->key);
if ((ret = c->search_near(c, &cmp)) == 0) {
if (cmp < 0)
ret = c->next(c);
else if (cmp == 0) {
if (clsm->current == NULL)
clsm->current = c;
else
F_SET(clsm,
WT_CLSM_MULTIPLE);
}
} else
F_CLR(c, WT_CURSTD_KEY_SET);
}
WT_ERR_NOTFOUND_OK(ret);
}
F_SET(clsm, WT_CLSM_ITERATE_NEXT);
F_CLR(clsm, WT_CLSM_ITERATE_PREV);
/* We just positioned *at* the key, now move. */
if (clsm->current != NULL)
goto retry;
} else {
/*
* __truncate_dsrc --
* WT_SESSION::truncate for a data-source without a truncate operation.
*/
static int
__truncate_dsrc(WT_SESSION_IMPL *session, const char *uri)
{
WT_CURSOR *cursor;
WT_DECL_RET;
const char *cfg[2];
/* Open a cursor and traverse the object, removing every entry. */
cfg[0] = WT_CONFIG_BASE(session, session_open_cursor);
cfg[1] = NULL;
WT_RET(__wt_open_cursor(session, uri, NULL, cfg, &cursor));
while ((ret = cursor->next(cursor)) == 0)
WT_ERR(cursor->remove(cursor));
WT_ERR_NOTFOUND_OK(ret);
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __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);
}
/*
* __metadata_load_bulk --
* Create any bulk-loaded file stubs.
*/
static int
__metadata_load_bulk(WT_SESSION_IMPL *session)
{
WT_CURSOR *cursor;
WT_DECL_RET;
uint32_t allocsize;
bool exist;
const char *filecfg[] = {
WT_CONFIG_BASE(session, file_meta), NULL, NULL };
const char *key, *value;
/*
* If a file was being bulk-loaded during the hot backup, it will appear
* in the metadata file, but the file won't exist. Create on demand.
*/
WT_RET(__wt_metadata_cursor(session, &cursor));
while ((ret = cursor->next(cursor)) == 0) {
WT_ERR(cursor->get_key(cursor, &key));
if (!WT_PREFIX_SKIP(key, "file:"))
continue;
/* If the file exists, it's all good. */
WT_ERR(__wt_fs_exist(session, key, &exist));
if (exist)
continue;
/*
* If the file doesn't exist, assume it's a bulk-loaded file;
* retrieve the allocation size and re-create the file.
*/
WT_ERR(cursor->get_value(cursor, &value));
filecfg[1] = value;
WT_ERR(__wt_direct_io_size_check(
session, filecfg, "allocation_size", &allocsize));
WT_ERR(__wt_block_manager_create(session, key, allocsize));
}
WT_ERR_NOTFOUND_OK(ret);
err: WT_TRET(__wt_metadata_cursor_release(session, &cursor));
return (ret);
}
/*
* __fill_index --
* Fill the index from the current contents of the table.
*/
static int
__fill_index(WT_SESSION_IMPL *session, WT_TABLE *table, WT_INDEX *idx)
{
WT_DECL_RET;
WT_CURSOR *tcur, *icur;
WT_SESSION *wt_session;
wt_session = &session->iface;
tcur = NULL;
icur = NULL;
WT_RET(__wt_schema_open_colgroups(session, table));
/*
* If the column groups have not been completely created,
* there cannot be data inserted yet, and we're done.
*/
if (!table->cg_complete)
return (0);
WT_ERR(wt_session->open_cursor(wt_session,
idx->source, NULL, "bulk=unordered", &icur));
WT_ERR(wt_session->open_cursor(wt_session,
table->name, NULL, "readonly", &tcur));
while ((ret = tcur->next(tcur)) == 0)
WT_ERR(__wt_apply_single_idx(session, idx,
icur, (WT_CURSOR_TABLE *)tcur, icur->insert));
WT_ERR_NOTFOUND_OK(ret);
err:
if (icur)
WT_TRET(icur->close(icur));
if (tcur)
WT_TRET(tcur->close(tcur));
return (ret);
}
/*
* __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);
}
/*
* __config_merge_scan --
* Walk a configuration string, inserting entries into the merged array.
*/
static int
__config_merge_scan(WT_SESSION_IMPL *session,
const char *key, const char *value, WT_CONFIG_MERGE *cp)
{
WT_CONFIG cparser;
WT_CONFIG_ITEM k, v;
WT_DECL_ITEM(kb);
WT_DECL_ITEM(vb);
WT_DECL_RET;
size_t len;
WT_ERR(__wt_scr_alloc(session, 0, &kb));
WT_ERR(__wt_scr_alloc(session, 0, &vb));
WT_ERR(__wt_config_init(session, &cparser, value));
while ((ret = __wt_config_next(&cparser, &k, &v)) == 0) {
if (k.type != WT_CONFIG_ITEM_STRING &&
k.type != WT_CONFIG_ITEM_ID)
WT_ERR_MSG(session, EINVAL,
"Invalid configuration key found: '%s'\n", k.str);
/* Include the quotes around string keys/values. */
if (k.type == WT_CONFIG_ITEM_STRING) {
--k.str;
k.len += 2;
}
if (v.type == WT_CONFIG_ITEM_STRING) {
--v.str;
v.len += 2;
}
/*
* !!!
* We're using a JSON quote character to separate the names we
* create for nested structures. That's not completely safe as
* it's possible to quote characters in JSON such that a quote
* character appears as a literal character in a key name. In
* a few cases, applications can create their own key namespace
* (for example, shared library extension names), and therefore
* it's possible for an application to confuse us. Error if we
* we ever see a key with a magic character.
*/
for (len = 0; len < k.len; ++len)
if (k.str[len] == SEPC)
WT_ERR_MSG(session, EINVAL,
"key %.*s contains a '%c' separator "
"character",
(int)k.len, (char *)k.str, SEPC);
/* Build the key/value strings. */
WT_ERR(__wt_buf_fmt(session,
kb, "%s%s%.*s",
key == NULL ? "" : key,
key == NULL ? "" : SEP,
(int)k.len, k.str));
WT_ERR(__wt_buf_fmt(session,
vb, "%.*s", (int)v.len, v.str));
/*
* If the value is a structure, recursively parse it.
*
* !!!
* Don't merge unless the structure has field names. WiredTiger
* stores checkpoint LSNs in the metadata file using nested
* structures without field names: "checkpoint_lsn=(1,0)", not
* "checkpoint_lsn=(file=1,offset=0)". The value type is still
* WT_CONFIG_ITEM_STRUCT, so we check for a field name in the
* value.
*/
if (v.type == WT_CONFIG_ITEM_STRUCT &&
strchr(vb->data, '=') != NULL) {
WT_ERR(__config_merge_scan(
session, kb->data, vb->data, cp));
continue;
}
/* Insert the value into the array. */
WT_ERR(__wt_realloc_def(session,
&cp->entries_allocated,
cp->entries_next + 1, &cp->entries));
WT_ERR(__wt_strndup(session,
kb->data, kb->size, &cp->entries[cp->entries_next].k));
WT_ERR(__wt_strndup(session,
vb->data, vb->size, &cp->entries[cp->entries_next].v));
cp->entries[cp->entries_next].gen = cp->entries_next;
++cp->entries_next;
}
WT_ERR_NOTFOUND_OK(ret);
err: __wt_scr_free(session, &kb);
__wt_scr_free(session, &vb);
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);
}
/*
* __wt_lsm_merge --
* Merge a set of chunks of an LSM tree.
*/
int
__wt_lsm_merge(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree, u_int id)
{
WT_BLOOM *bloom;
WT_CURSOR *dest, *src;
WT_DECL_RET;
WT_ITEM key, value;
WT_LSM_CHUNK *chunk;
uint32_t generation;
uint64_t insert_count, record_count;
u_int dest_id, end_chunk, i, nchunks, start_chunk, start_id, verb;
int tret;
bool created_chunk, create_bloom, locked, in_sync;
const char *cfg[3];
const char *drop_cfg[] =
{ WT_CONFIG_BASE(session, WT_SESSION_drop), "force", NULL };
bloom = NULL;
chunk = NULL;
dest = src = NULL;
start_id = 0;
created_chunk = create_bloom = locked = in_sync = false;
/* Fast path if it's obvious no merges could be done. */
if (lsm_tree->nchunks < lsm_tree->merge_min &&
lsm_tree->merge_aggressiveness < WT_LSM_AGGRESSIVE_THRESHOLD)
return (WT_NOTFOUND);
/*
* Use the lsm_tree lock to read the chunks (so no switches occur), but
* avoid holding it while the merge is in progress: that may take a
* long time.
*/
WT_RET(__wt_lsm_tree_writelock(session, lsm_tree));
locked = true;
WT_ERR(__lsm_merge_span(session,
lsm_tree, id, &start_chunk, &end_chunk, &record_count));
nchunks = (end_chunk + 1) - start_chunk;
WT_ASSERT(session, nchunks > 0);
start_id = lsm_tree->chunk[start_chunk]->id;
/* Find the merge generation. */
for (generation = 0, i = 0; i < nchunks; i++)
generation = WT_MAX(generation,
lsm_tree->chunk[start_chunk + i]->generation + 1);
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
locked = false;
/* Allocate an ID for the merge. */
dest_id = __wt_atomic_add32(&lsm_tree->last, 1);
/*
* We only want to do the chunk loop if we're running with verbose,
* so we wrap these statements in the conditional. Avoid the loop
* in the normal path.
*/
if (WT_VERBOSE_ISSET(session, WT_VERB_LSM)) {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Merging %s chunks %u-%u into %u (%" PRIu64 " records)"
", generation %" PRIu32,
lsm_tree->name,
start_chunk, end_chunk, dest_id, record_count, generation));
for (verb = start_chunk; verb <= end_chunk; verb++)
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Merging %s: Chunk[%u] id %u, gen: %" PRIu32
", size: %" PRIu64 ", records: %" PRIu64,
lsm_tree->name, verb, lsm_tree->chunk[verb]->id,
lsm_tree->chunk[verb]->generation,
lsm_tree->chunk[verb]->size,
lsm_tree->chunk[verb]->count));
}
WT_ERR(__wt_calloc_one(session, &chunk));
created_chunk = true;
chunk->id = dest_id;
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_MERGED) &&
(FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST) ||
start_chunk > 0) && record_count > 0)
create_bloom = true;
/*
* Special setup for the merge cursor:
* first, reset to open the dependent cursors;
* then restrict the cursor to a specific number of chunks;
* then set MERGE so the cursor doesn't track updates to the tree.
*/
WT_ERR(__wt_open_cursor(session, lsm_tree->name, NULL, NULL, &src));
F_SET(src, WT_CURSTD_RAW);
WT_ERR(__wt_clsm_init_merge(src, start_chunk, start_id, nchunks));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
WT_ERR(ret);
if (create_bloom) {
WT_ERR(__wt_lsm_tree_setup_bloom(session, lsm_tree, chunk));
WT_ERR(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config,
record_count, lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count, &bloom));
}
/* Discard pages we read as soon as we're done with them. */
F_SET(session, WT_SESSION_NO_CACHE);
cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
cfg[1] = "bulk,raw,skip_sort_check";
cfg[2] = NULL;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
#define LSM_MERGE_CHECK_INTERVAL WT_THOUSAND
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
if (insert_count % LSM_MERGE_CHECK_INTERVAL == 0) {
if (!F_ISSET(lsm_tree, WT_LSM_TREE_ACTIVE))
WT_ERR(EINTR);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
}
WT_ERR(src->get_key(src, &key));
dest->set_key(dest, &key);
WT_ERR(src->get_value(src, &value));
dest->set_value(dest, &value);
WT_ERR(dest->insert(dest));
if (create_bloom)
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, insert_count % LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Bloom size for %" PRIu64 " has %" PRIu64 " items inserted.",
record_count, insert_count));
/*
* Closing and syncing the files can take a while. Set the
* merge_syncing field so that compact knows it is still in
* progress.
*/
(void)__wt_atomic_add32(&lsm_tree->merge_syncing, 1);
in_sync = true;
/*
* We've successfully created the new chunk. Now install it. We need
* to ensure that the NO_CACHE flag is cleared and the bloom filter
* is closed (even if a step fails), so track errors but don't return
* until we've cleaned up.
*/
WT_TRET(src->close(src));
WT_TRET(dest->close(dest));
src = dest = NULL;
F_CLR(session, WT_SESSION_NO_CACHE);
/*
* We're doing advisory reads to fault the new trees into cache.
* Don't block if the cache is full: our next unit of work may be to
* discard some trees to free space.
*/
F_SET(session, WT_SESSION_NO_EVICTION);
if (create_bloom) {
if (ret == 0)
WT_TRET(__wt_bloom_finalize(bloom));
/*
* Read in a key to make sure the Bloom filters btree handle is
* open before it becomes visible to application threads.
* Otherwise application threads will stall while it is opened
* and internal pages are read into cache.
*/
if (ret == 0) {
WT_CLEAR(key);
WT_TRET_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
}
WT_TRET(__wt_bloom_close(bloom));
bloom = NULL;
}
WT_ERR(ret);
/*
* Open a handle on the new chunk before application threads attempt
* to access it, opening it pre-loads internal pages into the file
* system cache.
*/
cfg[1] = "checkpoint=" WT_CHECKPOINT;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
WT_TRET(dest->close(dest));
dest = NULL;
++lsm_tree->merge_progressing;
(void)__wt_atomic_sub32(&lsm_tree->merge_syncing, 1);
in_sync = false;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
locked = true;
/*
* Check whether we raced with another merge, and adjust the chunk
* array offset as necessary.
*/
if (start_chunk >= lsm_tree->nchunks ||
lsm_tree->chunk[start_chunk]->id != start_id)
for (start_chunk = 0;
start_chunk < lsm_tree->nchunks;
start_chunk++)
if (lsm_tree->chunk[start_chunk]->id == start_id)
break;
/*
* It is safe to error out here - since the update can only fail
* prior to making updates to the tree.
*/
WT_ERR(__wt_lsm_merge_update_tree(
session, lsm_tree, start_chunk, nchunks, chunk));
if (create_bloom)
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
chunk->count = insert_count;
chunk->generation = generation;
F_SET(chunk, WT_LSM_CHUNK_ONDISK);
/*
* We have no current way of continuing if the metadata update fails,
* so we will panic in that case. Put some effort into cleaning up
* after ourselves here - so things have a chance of shutting down.
*
* Any errors that happened after the tree was locked are
* fatal - we can't guarantee the state of the tree.
*/
if ((ret = __wt_lsm_meta_write(session, lsm_tree)) != 0)
WT_PANIC_ERR(session, ret, "Failed finalizing LSM merge");
lsm_tree->dsk_gen++;
/* Update the throttling while holding the tree lock. */
__wt_lsm_tree_throttle(session, lsm_tree, true);
/* Schedule a pass to discard old chunks */
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_DROP, 0, lsm_tree));
err: if (locked)
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (in_sync)
(void)__wt_atomic_sub32(&lsm_tree->merge_syncing, 1);
if (src != NULL)
WT_TRET(src->close(src));
if (dest != NULL)
WT_TRET(dest->close(dest));
if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
if (ret != 0 && created_chunk) {
/* Drop the newly-created files on error. */
if (chunk->uri != NULL) {
WT_WITH_SCHEMA_LOCK(session, tret =
__wt_schema_drop(session, chunk->uri, drop_cfg));
WT_TRET(tret);
}
if (create_bloom && chunk->bloom_uri != NULL) {
WT_WITH_SCHEMA_LOCK(session,
tret = __wt_schema_drop(
session, chunk->bloom_uri, drop_cfg));
WT_TRET(tret);
}
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, chunk);
if (ret == EINTR)
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge aborted due to close"));
else
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge failed with %s",
__wt_strerror(session, ret, NULL, 0)));
}
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
return (ret);
}
/*
* __txn_op_apply --
* Apply a transactional operation during recovery.
*/
static int
__txn_op_apply(
WT_RECOVERY *r, WT_LSN *lsnp, const uint8_t **pp, const uint8_t *end)
{
WT_CURSOR *cursor, *start, *stop;
WT_DECL_RET;
WT_ITEM key, start_key, stop_key, value;
WT_SESSION_IMPL *session;
uint64_t recno, start_recno, stop_recno;
uint32_t fileid, mode, optype, opsize;
session = r->session;
cursor = NULL;
/* Peek at the size and the type. */
WT_ERR(__wt_logop_read(session, pp, end, &optype, &opsize));
end = *pp + opsize;
switch (optype) {
case WT_LOGOP_COL_MODIFY:
WT_ERR(__wt_logop_col_modify_unpack(session, pp, end,
&fileid, &recno, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
if ((ret = cursor->search(cursor)) != 0)
WT_ERR_NOTFOUND_OK(ret);
else {
/*
* Build/insert a complete value during recovery rather
* than using cursor modify to create a partial update
* (for no particular reason than simplicity).
*/
WT_ERR(__wt_modify_apply(session, cursor, value.data));
WT_ERR(cursor->insert(cursor));
}
break;
case WT_LOGOP_COL_PUT:
WT_ERR(__wt_logop_col_put_unpack(session, pp, end,
&fileid, &recno, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
__wt_cursor_set_raw_value(cursor, &value);
WT_ERR(cursor->insert(cursor));
break;
case WT_LOGOP_COL_REMOVE:
WT_ERR(__wt_logop_col_remove_unpack(session, pp, end,
&fileid, &recno));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
cursor->set_key(cursor, recno);
WT_ERR(cursor->remove(cursor));
break;
case WT_LOGOP_COL_TRUNCATE:
WT_ERR(__wt_logop_col_truncate_unpack(session, pp, end,
&fileid, &start_recno, &stop_recno));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
/* Set up the cursors. */
if (start_recno == WT_RECNO_OOB) {
start = NULL;
stop = cursor;
} else if (stop_recno == WT_RECNO_OOB) {
start = cursor;
stop = NULL;
} else {
start = cursor;
WT_ERR(__recovery_cursor(
session, r, lsnp, fileid, true, &stop));
}
/* Set the keys. */
if (start != NULL)
start->set_key(start, start_recno);
if (stop != NULL)
stop->set_key(stop, stop_recno);
WT_TRET(session->iface.truncate(&session->iface, NULL,
start, stop, NULL));
/* If we opened a duplicate cursor, close it now. */
if (stop != NULL && stop != cursor)
WT_TRET(stop->close(stop));
WT_ERR(ret);
break;
case WT_LOGOP_ROW_MODIFY:
WT_ERR(__wt_logop_row_modify_unpack(session, pp, end,
&fileid, &key, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
if ((ret = cursor->search(cursor)) != 0)
WT_ERR_NOTFOUND_OK(ret);
else {
/*
* Build/insert a complete value during recovery rather
* than using cursor modify to create a partial update
* (for no particular reason than simplicity).
*/
WT_ERR(__wt_modify_apply(session, cursor, value.data));
WT_ERR(cursor->insert(cursor));
}
break;
case WT_LOGOP_ROW_PUT:
WT_ERR(__wt_logop_row_put_unpack(session, pp, end,
&fileid, &key, &value));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
__wt_cursor_set_raw_value(cursor, &value);
WT_ERR(cursor->insert(cursor));
break;
case WT_LOGOP_ROW_REMOVE:
WT_ERR(__wt_logop_row_remove_unpack(session, pp, end,
&fileid, &key));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
__wt_cursor_set_raw_key(cursor, &key);
WT_ERR(cursor->remove(cursor));
break;
case WT_LOGOP_ROW_TRUNCATE:
WT_ERR(__wt_logop_row_truncate_unpack(session, pp, end,
&fileid, &start_key, &stop_key, &mode));
GET_RECOVERY_CURSOR(session, r, lsnp, fileid, &cursor);
/* Set up the cursors. */
start = stop = NULL;
switch (mode) {
case WT_TXN_TRUNC_ALL:
/* Both cursors stay NULL. */
break;
case WT_TXN_TRUNC_BOTH:
start = cursor;
WT_ERR(__recovery_cursor(
session, r, lsnp, fileid, true, &stop));
break;
case WT_TXN_TRUNC_START:
start = cursor;
break;
case WT_TXN_TRUNC_STOP:
stop = cursor;
break;
WT_ILLEGAL_VALUE_ERR(session, mode);
}
/* Set the keys. */
if (start != NULL)
__wt_cursor_set_raw_key(start, &start_key);
if (stop != NULL)
__wt_cursor_set_raw_key(stop, &stop_key);
WT_TRET(session->iface.truncate(&session->iface, NULL,
start, stop, NULL));
/* If we opened a duplicate cursor, close it now. */
if (stop != NULL && stop != cursor)
WT_TRET(stop->close(stop));
WT_ERR(ret);
break;
WT_ILLEGAL_VALUE_ERR(session, optype);
}
/* Reset the cursor so it doesn't block eviction. */
if (cursor != NULL)
WT_ERR(cursor->reset(cursor));
return (0);
err: __wt_err(session, ret,
"operation apply failed during recovery: operation type %"
PRIu32 " at LSN %" PRIu32 "/%" PRIu32,
optype, lsnp->l.file, lsnp->l.offset);
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);
}
/*
* __wt_las_sweep --
* Sweep the lookaside table.
*/
int
__wt_las_sweep(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR *cursor;
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_RET;
WT_ITEM *key;
uint64_t cnt, las_counter, las_txnid;
int64_t remove_cnt;
uint32_t las_id, session_flags;
int notused;
conn = S2C(session);
cursor = NULL;
key = &conn->las_sweep_key;
remove_cnt = 0;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/*
* If we're not starting a new sweep, position the cursor using the key
* from the last call (we don't care if we're before or after the key,
* just roughly in the same spot is fine).
*/
if (key->size != 0) {
__wt_cursor_set_raw_key(cursor, key);
ret = cursor->search_near(cursor, ¬used);
/*
* Don't search for the same key twice; if we don't set a new
* key below, it's because we've reached the end of the table
* and we want the next pass to start at the beginning of the
* table. Searching for the same key could leave us stuck at
* the end of the table, repeatedly checking the same rows.
*/
key->size = 0;
if (ret != 0)
goto srch_notfound;
}
/*
* The sweep server wakes up every 10 seconds (by default), it's a slow
* moving thread. Try to review the entire lookaside table once every 5
* minutes, or every 30 calls.
*
* The reason is because the lookaside table exists because we're seeing
* cache/eviction pressure (it allows us to trade performance and disk
* space for cache space), and it's likely lookaside blocks are being
* evicted, and reading them back in doesn't help things. A trickier,
* but possibly better, alternative might be to review all lookaside
* blocks in the cache in order to get rid of them, and slowly review
* lookaside blocks that have already been evicted.
*/
cnt = (uint64_t)WT_MAX(100, conn->las_record_cnt / 30);
/* Discard pages we read as soon as we're done with them. */
F_SET(session, WT_SESSION_NO_CACHE);
/* Walk the file. */
for (; cnt > 0 && (ret = cursor->next(cursor)) == 0; --cnt) {
/*
* If the loop terminates after completing a work unit, we will
* continue the table sweep next time. Get a local copy of the
* sweep key, we're going to reset the cursor; do so before
* calling cursor.remove, cursor.remove can discard our hazard
* pointer and the page could be evicted from underneath us.
*/
if (cnt == 1) {
WT_ERR(__wt_cursor_get_raw_key(cursor, key));
if (!WT_DATA_IN_ITEM(key))
WT_ERR(__wt_buf_set(
session, key, key->data, key->size));
}
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* If the on-page record transaction ID associated with the
* record is globally visible, the record can be discarded.
*
* Cursor opened overwrite=true: won't return WT_NOTFOUND should
* another thread remove the record before we do, and the cursor
* remains positioned in that case.
*/
if (__wt_txn_visible_all(session, las_txnid)) {
WT_ERR(cursor->remove(cursor));
++remove_cnt;
}
}
srch_notfound:
WT_ERR_NOTFOUND_OK(ret);
if (0) {
err: __wt_buf_free(session, key);
}
WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
/*
* If there were races to remove records, we can over-count. All
* arithmetic is signed, so underflow isn't fatal, but check anyway so
* we don't skew low over time.
*/
if (remove_cnt > S2C(session)->las_record_cnt)
S2C(session)->las_record_cnt = 0;
else if (remove_cnt > 0)
(void)__wt_atomic_subi64(&conn->las_record_cnt, remove_cnt);
F_CLR(session, WT_SESSION_NO_CACHE);
__wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
return (ret);
}
/*
* __txn_rollback_to_stable_lookaside_fixup --
* Remove any updates that need to be rolled back from the lookaside file.
*/
static int
__txn_rollback_to_stable_lookaside_fixup(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_DECL_TIMESTAMP(rollback_timestamp)
WT_ITEM las_addr, las_key, las_timestamp;
WT_TXN_GLOBAL *txn_global;
uint64_t las_counter, las_txnid, remove_cnt;
uint32_t las_id, session_flags;
conn = S2C(session);
cursor = NULL;
remove_cnt = 0;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
WT_CLEAR(las_timestamp);
/*
* Copy the stable timestamp, otherwise we'd need to lock it each time
* it's accessed. Even though the stable timestamp isn't supposed to be
* updated while rolling back, accessing it without a lock would
* violate protocol.
*/
txn_global = &S2C(session)->txn_global;
__wt_readlock(session, &txn_global->rwlock);
__wt_timestamp_set(&rollback_timestamp, &txn_global->stable_timestamp);
__wt_readunlock(session, &txn_global->rwlock);
__wt_las_cursor(session, &cursor, &session_flags);
/* Discard pages we read as soon as we're done with them. */
F_SET(session, WT_SESSION_NO_CACHE);
/* Walk the file. */
for (; (ret = cursor->next(cursor)) == 0; ) {
WT_ERR(cursor->get_key(cursor, &las_id, &las_addr, &las_counter,
&las_txnid, &las_timestamp, &las_key));
/* Check the file ID so we can skip durable tables */
if (__bit_test(conn->stable_rollback_bitstring, las_id))
continue;
/*
* Entries with no timestamp will have a timestamp of zero,
* which will fail the following check and cause them to never
* be removed.
*/
if (__wt_timestamp_cmp(
&rollback_timestamp, las_timestamp.data) < 0) {
WT_ERR(cursor->remove(cursor));
++remove_cnt;
}
}
WT_ERR_NOTFOUND_OK(ret);
err: WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
/*
* If there were races to remove records, we can over-count. Underflow
* isn't fatal, but check anyway so we don't skew low over time.
*/
if (remove_cnt > conn->las_record_cnt)
conn->las_record_cnt = 0;
else if (remove_cnt > 0)
(void)__wt_atomic_sub64(&conn->las_record_cnt, remove_cnt);
F_CLR(session, WT_SESSION_NO_CACHE);
return (ret);
}
/*
* __create_colgroup --
* Create a column group.
*/
static int
__create_colgroup(WT_SESSION_IMPL *session,
const char *name, bool exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
WT_ITEM confbuf, fmt, namebuf;
WT_TABLE *table;
size_t tlen;
const char **cfgp, *cfg[4] =
{ WT_CONFIG_BASE(session, colgroup_meta), config, NULL, NULL };
const char *sourcecfg[] = { config, NULL, NULL };
const char *cgname, *source, *sourceconf, *tablename;
char *cgconf, *origconf;
bool exists;
sourceconf = NULL;
cgconf = origconf = NULL;
WT_CLEAR(fmt);
WT_CLEAR(confbuf);
WT_CLEAR(namebuf);
exists = false;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "colgroup:"))
return (
__wt_unexpected_object_type(session, name, "colgroup:"));
cgname = strchr(tablename, ':');
if (cgname != NULL) {
tlen = (size_t)(cgname - tablename);
++cgname;
} else
tlen = strlen(tablename);
if ((ret =
__wt_schema_get_table(session, tablename, tlen, true, &table)) != 0)
WT_RET_MSG(session, (ret == WT_NOTFOUND) ? ENOENT : ret,
"Can't create '%s' for non-existent table '%.*s'",
name, (int)tlen, tablename);
/* Make sure the column group is referenced from the table. */
if (cgname != NULL && (ret =
__wt_config_subgets(session, &table->cgconf, cgname, &cval)) != 0)
WT_ERR_MSG(session, EINVAL,
"Column group '%s' not found in table '%.*s'",
cgname, (int)tlen, tablename);
/* Check if the column group already exists. */
if ((ret = __wt_metadata_search(session, name, &origconf)) == 0) {
if (exclusive)
WT_ERR(EEXIST);
exists = true;
}
WT_ERR_NOTFOUND_OK(ret);
/* Find the first NULL entry in the cfg stack. */
for (cfgp = &cfg[1]; *cfgp; cfgp++)
;
/* Add the source to the colgroup config before collapsing. */
if (__wt_config_getones(
session, config, "source", &cval) == 0 && cval.len != 0) {
WT_ERR(__wt_buf_fmt(
session, &namebuf, "%.*s", (int)cval.len, cval.str));
source = namebuf.data;
} else {
WT_ERR(__wt_schema_colgroup_source(
session, table, cgname, config, &namebuf));
source = namebuf.data;
WT_ERR(__wt_buf_fmt(
session, &confbuf, "source=\"%s\"", source));
*cfgp++ = confbuf.data;
}
/* Calculate the key/value formats: these go into the source config. */
WT_ERR(__wt_buf_fmt(session, &fmt, "key_format=%s", table->key_format));
if (cgname == NULL)
WT_ERR(__wt_buf_catfmt
(session, &fmt, ",value_format=%s", table->value_format));
else {
if (__wt_config_getones(session, config, "columns", &cval) != 0)
WT_ERR_MSG(session, EINVAL,
"No 'columns' configuration for '%s'", name);
WT_ERR(__wt_buf_catfmt(session, &fmt, ",value_format="));
WT_ERR(__wt_struct_reformat(session,
table, cval.str, cval.len, NULL, true, &fmt));
}
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_merge(session, sourcecfg, NULL, &sourceconf));
WT_ERR(__wt_schema_create(session, source, sourceconf));
WT_ERR(__wt_config_collapse(session, cfg, &cgconf));
if (!exists) {
WT_ERR(__wt_metadata_insert(session, name, cgconf));
WT_ERR(__wt_schema_open_colgroups(session, table));
}
err: __wt_free(session, cgconf);
__wt_free(session, sourceconf);
__wt_free(session, origconf);
__wt_buf_free(session, &confbuf);
__wt_buf_free(session, &fmt);
__wt_buf_free(session, &namebuf);
__wt_schema_release_table(session, table);
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);
}
/*
* __wt_las_sweep --
* Sweep the lookaside table.
*/
int
__wt_las_sweep(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR *cursor;
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_RET;
WT_ITEM *key;
uint64_t cnt, las_counter, las_txnid;
uint32_t las_id, session_flags;
int notused;
conn = S2C(session);
cursor = NULL;
key = &conn->las_sweep_key;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/*
* If we're not starting a new sweep, position the cursor using the key
* from the last call (we don't care if we're before or after the key,
* just roughly in the same spot is fine).
*/
if (conn->las_sweep_call != 0 && key->data != NULL) {
__wt_cursor_set_raw_key(cursor, key);
if ((ret = cursor->search_near(cursor, ¬used)) != 0)
goto srch_notfound;
}
/*
* The sweep server wakes up every 10 seconds (by default), it's a slow
* moving thread. Try to review the entire lookaside table once every 5
* minutes, or every 30 calls.
*
* The reason is because the lookaside table exists because we're seeing
* cache/eviction pressure (it allows us to trade performance and disk
* space for cache space), and it's likely lookaside blocks are being
* evicted, and reading them back in doesn't help things. A trickier,
* but possibly better, alternative might be to review all lookaside
* blocks in the cache in order to get rid of them, and slowly review
* lookaside blocks that have already been evicted.
*
* We can't know for sure how many records are in the lookaside table,
* the cursor insert and remove statistics aren't updated atomically.
* Start with reviewing 100 rows, and if it takes more than the target
* number of calls to finish, increase the number of rows checked on
* each call; if it takes less than the target calls to finish, then
* decrease the number of rows reviewed on each call (but never less
* than 100).
*/
#define WT_SWEEP_LOOKASIDE_MIN_CNT 100
#define WT_SWEEP_LOOKASIDE_PASS_TARGET 30
++conn->las_sweep_call;
if ((cnt = conn->las_sweep_cnt) < WT_SWEEP_LOOKASIDE_MIN_CNT)
cnt = conn->las_sweep_cnt = WT_SWEEP_LOOKASIDE_MIN_CNT;
/* Walk the file. */
for (; cnt > 0 && (ret = cursor->next(cursor)) == 0; --cnt) {
/*
* If the loop terminates after completing a work unit, we will
* continue the table sweep next time. Get a local copy of the
* sweep key, we're going to reset the cursor; do so before
* calling cursor.remove, cursor.remove can discard our hazard
* pointer and the page could be evicted from underneath us.
*/
if (cnt == 1) {
WT_ERR(__wt_cursor_get_raw_key(cursor, key));
if (!WT_DATA_IN_ITEM(key))
WT_ERR(__wt_buf_set(
session, key, key->data, key->size));
}
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* If the on-page record transaction ID associated with the
* record is globally visible, the record can be discarded.
*
* Cursor opened overwrite=true: won't return WT_NOTFOUND should
* another thread remove the record before we do, and the cursor
* remains positioned in that case.
*/
if (__wt_txn_visible_all(session, las_txnid))
WT_ERR(cursor->remove(cursor));
}
/*
* When reaching the lookaside table end or the target number of calls,
* adjust the row count. Decrease/increase the row count depending on
* if the number of calls is less/more than the target.
*/
if (ret == WT_NOTFOUND ||
conn->las_sweep_call > WT_SWEEP_LOOKASIDE_PASS_TARGET) {
if (conn->las_sweep_call < WT_SWEEP_LOOKASIDE_PASS_TARGET &&
conn->las_sweep_cnt > WT_SWEEP_LOOKASIDE_MIN_CNT)
conn->las_sweep_cnt -= WT_SWEEP_LOOKASIDE_MIN_CNT;
if (conn->las_sweep_call > WT_SWEEP_LOOKASIDE_PASS_TARGET)
conn->las_sweep_cnt += WT_SWEEP_LOOKASIDE_MIN_CNT;
}
srch_notfound:
if (ret == WT_NOTFOUND)
conn->las_sweep_call = 0;
WT_ERR_NOTFOUND_OK(ret);
if (0) {
err: __wt_buf_free(session, key);
}
WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
__wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
return (ret);
}
/*
* __wt_curds_open --
* Initialize a data-source cursor.
*/
int
__wt_curds_open(
WT_SESSION_IMPL *session, const char *uri, WT_CURSOR *owner,
const char *cfg[], WT_DATA_SOURCE *dsrc, WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__wt_cursor_set_value, /* set-value */
__curds_compare, /* compare */
__wt_cursor_equals, /* equals */
__curds_next, /* next */
__curds_prev, /* prev */
__curds_reset, /* reset */
__curds_search, /* search */
__curds_search_near, /* search-near */
__curds_insert, /* insert */
__wt_cursor_modify_notsup, /* modify */
__curds_update, /* update */
__curds_remove, /* remove */
__curds_reserve, /* reserve */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__curds_close); /* close */
WT_CONFIG_ITEM cval, metadata;
WT_CURSOR *cursor, *source;
WT_CURSOR_DATA_SOURCE *data_source;
WT_DECL_RET;
char *metaconf;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_DATA_SOURCE, iface) == 0);
data_source = NULL;
metaconf = NULL;
WT_RET(__wt_calloc_one(session, &data_source));
cursor = &data_source->iface;
*cursor = iface;
cursor->session = &session->iface;
/*
* XXX
* The underlying data-source may require the object's key and value
* formats. This isn't a particularly elegant way of getting that
* information to the data-source, this feels like a layering problem
* to me.
*/
WT_ERR(__wt_metadata_search(session, uri, &metaconf));
WT_ERR(__wt_config_getones(session, metaconf, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &cursor->key_format));
WT_ERR(__wt_config_getones(session, metaconf, "value_format", &cval));
WT_ERR(
__wt_strndup(session, cval.str, cval.len, &cursor->value_format));
WT_ERR(__wt_cursor_init(cursor, uri, owner, cfg, cursorp));
/* Data-source cursors may have a custom collator. */
ret = __wt_config_getones(session, metaconf, "collator", &cval);
if (ret == 0 && cval.len != 0) {
WT_CLEAR(metadata);
WT_ERR_NOTFOUND_OK(__wt_config_getones(
session, metaconf, "app_metadata", &metadata));
WT_ERR(__wt_collator_config(session, uri, &cval, &metadata,
&data_source->collator, &data_source->collator_owned));
}
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(dsrc->open_cursor(dsrc,
&session->iface, uri, (WT_CONFIG_ARG *)cfg, &data_source->source));
source = data_source->source;
source->session = (WT_SESSION *)session;
memset(&source->q, 0, sizeof(source->q));
source->recno = WT_RECNO_OOB;
memset(source->raw_recno_buf, 0, sizeof(source->raw_recno_buf));
memset(&source->key, 0, sizeof(source->key));
memset(&source->value, 0, sizeof(source->value));
source->saved_err = 0;
source->flags = 0;
if (0) {
err: WT_TRET(__curds_close(cursor));
*cursorp = NULL;
}
__wt_free(session, metaconf);
return (ret);
}
/*
* __las_page_instantiate --
* Instantiate lookaside update records in a recently read page.
*/
static int
__las_page_instantiate(WT_SESSION_IMPL *session,
WT_REF *ref, uint32_t read_id, const uint8_t *addr, size_t addr_size)
{
WT_CURSOR *cursor;
WT_CURSOR_BTREE cbt;
WT_DECL_ITEM(current_key);
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_ITEM(las_value);
WT_DECL_RET;
WT_PAGE *page;
WT_UPDATE *first_upd, *last_upd, *upd;
size_t incr, total_incr;
uint64_t current_recno, las_counter, las_txnid, recno, upd_txnid;
uint32_t las_id, upd_size, session_flags;
int exact;
const uint8_t *p;
cursor = NULL;
page = ref->page;
first_upd = last_upd = upd = NULL;
total_incr = 0;
current_recno = recno = WT_RECNO_OOB;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
__wt_btcur_init(session, &cbt);
__wt_btcur_open(&cbt);
WT_ERR(__wt_scr_alloc(session, 0, ¤t_key));
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
WT_ERR(__wt_scr_alloc(session, 0, &las_value));
/* Open a lookaside table cursor. */
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/*
* The lookaside records are in key and update order, that is, there
* will be a set of in-order updates for a key, then another set of
* in-order updates for a subsequent key. We process all of the updates
* for a key and then insert those updates into the page, then all the
* updates for the next key, and so on.
*
* Search for the block's unique prefix, stepping through any matching
* records.
*/
las_addr->data = addr;
las_addr->size = addr_size;
las_key->size = 0;
cursor->set_key(
cursor, read_id, las_addr, (uint64_t)0, (uint32_t)0, las_key);
if ((ret = cursor->search_near(cursor, &exact)) == 0 && exact < 0)
ret = cursor->next(cursor);
for (; ret == 0; ret = cursor->next(cursor)) {
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* Confirm the search using the unique prefix; if not a match,
* we're done searching for records for this page.
*/
if (las_id != read_id ||
las_addr->size != addr_size ||
memcmp(las_addr->data, addr, addr_size) != 0)
break;
/*
* If the on-page value has become globally visible, this record
* is no longer needed.
*/
if (__wt_txn_visible_all(session, las_txnid))
continue;
/* Allocate the WT_UPDATE structure. */
WT_ERR(cursor->get_value(
cursor, &upd_txnid, &upd_size, las_value));
WT_ERR(__wt_update_alloc(session,
(upd_size == WT_UPDATE_DELETED_VALUE) ? NULL : las_value,
&upd, &incr));
total_incr += incr;
upd->txnid = upd_txnid;
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
p = las_key->data;
WT_ERR(__wt_vunpack_uint(&p, 0, &recno));
if (current_recno == recno)
break;
WT_ASSERT(session, current_recno < recno);
if (first_upd != NULL) {
WT_ERR(__col_instantiate(session,
current_recno, ref, &cbt, first_upd));
first_upd = NULL;
}
current_recno = recno;
break;
case WT_PAGE_ROW_LEAF:
if (current_key->size == las_key->size &&
memcmp(current_key->data,
las_key->data, las_key->size) == 0)
break;
if (first_upd != NULL) {
WT_ERR(__row_instantiate(session,
current_key, ref, &cbt, first_upd));
first_upd = NULL;
}
WT_ERR(__wt_buf_set(session,
current_key, las_key->data, las_key->size));
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Append the latest update to the list. */
if (first_upd == NULL)
first_upd = last_upd = upd;
else {
last_upd->next = upd;
last_upd = upd;
}
upd = NULL;
}
WT_ERR_NOTFOUND_OK(ret);
/* Insert the last set of updates, if any. */
if (first_upd != NULL)
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
WT_ERR(__col_instantiate(session,
current_recno, ref, &cbt, first_upd));
first_upd = NULL;
break;
case WT_PAGE_ROW_LEAF:
WT_ERR(__row_instantiate(session,
current_key, ref, &cbt, first_upd));
first_upd = NULL;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Discard the cursor. */
WT_ERR(__wt_las_cursor_close(session, &cursor, session_flags));
if (total_incr != 0) {
__wt_cache_page_inmem_incr(session, page, total_incr);
/*
* We've modified/dirtied the page, but that's not necessary and
* if we keep the page clean, it's easier to evict. We leave the
* lookaside table updates in place, so if we evict this page
* without dirtying it, any future instantiation of it will find
* the records it needs. If the page is dirtied before eviction,
* then we'll write any needed lookaside table records for the
* new location of the page.
*/
__wt_page_modify_clear(session, page);
}
err: WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
WT_TRET(__wt_btcur_close(&cbt, 1));
/*
* On error, upd points to a single unlinked WT_UPDATE structure,
* first_upd points to a list.
*/
if (upd != NULL)
__wt_free(session, upd);
if (first_upd != NULL)
__wt_free_update_list(session, first_upd);
__wt_scr_free(session, ¤t_key);
__wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
__wt_scr_free(session, &las_value);
return (ret);
}
/*
* __create_index --
* Create an index.
*/
static int
__create_index(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG kcols, pkcols;
WT_CONFIG_ITEM ckey, cval, icols, kval;
WT_DECL_PACK_VALUE(pv);
WT_DECL_RET;
WT_ITEM confbuf, extra_cols, fmt, namebuf;
WT_PACK pack;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, index_meta), NULL, NULL, NULL };
const char *sourcecfg[] = { config, NULL, NULL };
const char *source, *sourceconf, *idxname, *tablename;
char *idxconf;
size_t tlen;
int have_extractor;
u_int i, npublic_cols;
sourceconf = NULL;
idxconf = NULL;
WT_CLEAR(confbuf);
WT_CLEAR(fmt);
WT_CLEAR(extra_cols);
WT_CLEAR(namebuf);
have_extractor = 0;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "index:"))
return (EINVAL);
idxname = strchr(tablename, ':');
if (idxname == NULL)
WT_RET_MSG(session, EINVAL, "Invalid index name, "
"should be <table name>:<index name>: %s", name);
tlen = (size_t)(idxname++ - tablename);
if ((ret =
__wt_schema_get_table(session, tablename, tlen, 1, &table)) != 0)
WT_RET_MSG(session, ret,
"Can't create an index for a non-existent table: %.*s",
(int)tlen, tablename);
if (table->is_simple)
WT_RET_MSG(session, EINVAL,
"%s requires a table with named columns", name);
if (__wt_config_getones(session, config, "source", &cval) == 0) {
WT_ERR(__wt_buf_fmt(session, &namebuf,
"%.*s", (int)cval.len, cval.str));
source = namebuf.data;
} else {
WT_ERR(__wt_schema_index_source(
session, table, idxname, config, &namebuf));
source = namebuf.data;
/* Add the source name to the index config before collapsing. */
WT_ERR(__wt_buf_catfmt(session, &confbuf,
",source=\"%s\"", source));
}
if (__wt_config_getones_none(
session, config, "extractor", &cval) == 0 && cval.len != 0) {
have_extractor = 1;
/* Custom extractors must supply a key format. */
if ((ret = __wt_config_getones(
session, config, "key_format", &kval)) != 0)
WT_ERR_MSG(session, EINVAL,
"%s: custom extractors require a key_format", name);
}
/* Calculate the key/value formats. */
WT_CLEAR(icols);
if (__wt_config_getones(session, config, "columns", &icols) != 0 &&
!have_extractor)
WT_ERR_MSG(session, EINVAL,
"%s: requires 'columns' configuration", name);
/*
* Count the public columns using the declared columns for normal
* indices or the key format for custom extractors.
*/
npublic_cols = 0;
if (!have_extractor) {
WT_ERR(__wt_config_subinit(session, &kcols, &icols));
while ((ret = __wt_config_next(&kcols, &ckey, &cval)) == 0)
++npublic_cols;
WT_ERR_NOTFOUND_OK(ret);
} else {
WT_ERR(__pack_initn(session, &pack, kval.str, kval.len));
while ((ret = __pack_next(&pack, &pv)) == 0)
++npublic_cols;
WT_ERR_NOTFOUND_OK(ret);
}
/*
* The key format for an index is somewhat subtle: the application
* specifies a set of columns that it will use for the key, but the
* engine usually adds some hidden columns in order to derive the
* primary key. These hidden columns are part of the source's
* key_format, which we are calculating now, but not part of an index
* cursor's key_format.
*/
WT_ERR(__wt_config_subinit(session, &pkcols, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&pkcols, &ckey, &cval)) == 0;
i++) {
/*
* If the primary key column is already in the secondary key,
* don't add it again.
*/
if (__wt_config_subgetraw(session, &icols, &ckey, &cval) == 0) {
if (have_extractor)
WT_ERR_MSG(session, EINVAL,
"an index with a custom extractor may not "
"include primary key columns");
continue;
}
WT_ERR(__wt_buf_catfmt(
session, &extra_cols, "%.*s,", (int)ckey.len, ckey.str));
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
/* Index values are empty: all columns are packed into the index key. */
WT_ERR(__wt_buf_fmt(session, &fmt, "value_format=,key_format="));
if (have_extractor) {
WT_ERR(__wt_buf_catfmt(session, &fmt, "%.*s",
(int)kval.len, kval.str));
WT_CLEAR(icols);
}
/*
* Construct the index key format, or append the primary key columns
* for custom extractors.
*/
WT_ERR(__wt_struct_reformat(session, table,
icols.str, icols.len, (const char *)extra_cols.data, 0, &fmt));
/* Check for a record number index key, which makes no sense. */
WT_ERR(__wt_config_getones(session, fmt.data, "key_format", &cval));
if (cval.len == 1 && cval.str[0] == 'r')
WT_ERR_MSG(session, EINVAL,
"column-store index may not use the record number as its "
"index key");
WT_ERR(__wt_buf_catfmt(
session, &fmt, ",index_key_columns=%u", npublic_cols));
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_merge(session, sourcecfg, NULL, &sourceconf));
WT_ERR(__wt_schema_create(session, source, sourceconf));
cfg[1] = sourceconf;
cfg[2] = confbuf.data;
WT_ERR(__wt_config_collapse(session, cfg, &idxconf));
if ((ret = __wt_metadata_insert(session, name, idxconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
/* Make sure that the configuration is valid. */
WT_ERR(__wt_schema_open_index(
session, table, idxname, strlen(idxname), NULL));
err: __wt_free(session, idxconf);
__wt_free(session, sourceconf);
__wt_buf_free(session, &confbuf);
__wt_buf_free(session, &extra_cols);
__wt_buf_free(session, &fmt);
__wt_buf_free(session, &namebuf);
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __create_table --
* Create a table.
*/
static int
__create_table(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG conf;
WT_CONFIG_ITEM cgkey, cgval, cval;
WT_DECL_RET;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, table_meta), config, NULL, NULL };
const char *tablename;
char *tableconf, *cgname;
size_t cgsize;
int ncolgroups;
cgname = NULL;
table = NULL;
tableconf = NULL;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "table:"))
return (EINVAL);
if ((ret = __wt_schema_get_table(session,
tablename, strlen(tablename), 0, &table)) == 0) {
__wt_schema_release_table(session, table);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_ERR(__wt_config_gets(session, cfg, "colgroups", &cval));
WT_ERR(__wt_config_subinit(session, &conf, &cval));
for (ncolgroups = 0;
(ret = __wt_config_next(&conf, &cgkey, &cgval)) == 0;
ncolgroups++)
;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_config_collapse(session, cfg, &tableconf));
if ((ret = __wt_metadata_insert(session, name, tableconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
/* Attempt to open the table now to catch any errors. */
WT_ERR(__wt_schema_get_table(
session, tablename, strlen(tablename), 1, &table));
if (ncolgroups == 0) {
cgsize = strlen("colgroup:") + strlen(tablename) + 1;
WT_ERR(__wt_calloc_def(session, cgsize, &cgname));
snprintf(cgname, cgsize, "colgroup:%s", tablename);
WT_ERR(__create_colgroup(session, cgname, exclusive, config));
}
if (0) {
err: if (table != NULL) {
WT_TRET(__wt_schema_remove_table(session, table));
table = NULL;
}
}
if (table != NULL)
__wt_schema_release_table(session, table);
__wt_free(session, cgname);
__wt_free(session, tableconf);
return (ret);
}
/*
* __schema_open_index --
* Open one or more indices for a table (internal version).
*/
static int
__schema_open_index(WT_SESSION_IMPL *session,
WT_TABLE *table, const char *idxname, size_t len, WT_INDEX **indexp)
{
WT_CURSOR *cursor;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_INDEX *idx;
u_int i;
int cmp;
bool match;
const char *idxconf, *name, *tablename, *uri;
/* Check if we've already done the work. */
if (idxname == NULL && table->idx_complete)
return (0);
cursor = NULL;
idx = NULL;
match = false;
/* Build a search key. */
tablename = table->name;
(void)WT_PREFIX_SKIP(tablename, "table:");
WT_ERR(__wt_scr_alloc(session, 512, &tmp));
WT_ERR(__wt_buf_fmt(session, tmp, "index:%s:", tablename));
/* Find matching indices. */
WT_ERR(__wt_metadata_cursor(session, &cursor));
cursor->set_key(cursor, tmp->data);
if ((ret = cursor->search_near(cursor, &cmp)) == 0 && cmp < 0)
ret = cursor->next(cursor);
for (i = 0; ret == 0; i++, ret = cursor->next(cursor)) {
WT_ERR(cursor->get_key(cursor, &uri));
name = uri;
if (!WT_PREFIX_SKIP(name, tmp->data))
break;
/* Is this the index we are looking for? */
match = idxname == NULL || WT_STRING_MATCH(name, idxname, len);
/*
* Ensure there is space, including if we have to make room for
* a new entry in the middle of the list.
*/
WT_ERR(__wt_realloc_def(session, &table->idx_alloc,
WT_MAX(i, table->nindices) + 1, &table->indices));
/* Keep the in-memory list in sync with the metadata. */
cmp = 0;
while (table->indices[i] != NULL &&
(cmp = strcmp(uri, table->indices[i]->name)) > 0) {
/* Index no longer exists, remove it. */
__wt_free(session, table->indices[i]);
memmove(&table->indices[i], &table->indices[i + 1],
(table->nindices - i) * sizeof(WT_INDEX *));
table->indices[--table->nindices] = NULL;
}
if (cmp < 0) {
/* Make room for a new index. */
memmove(&table->indices[i + 1], &table->indices[i],
(table->nindices - i) * sizeof(WT_INDEX *));
table->indices[i] = NULL;
++table->nindices;
}
if (!match)
continue;
if (table->indices[i] == NULL) {
WT_ERR(cursor->get_value(cursor, &idxconf));
WT_ERR(__wt_calloc_one(session, &idx));
WT_ERR(__wt_strdup(session, uri, &idx->name));
WT_ERR(__wt_strdup(session, idxconf, &idx->config));
WT_ERR(__open_index(session, table, idx));
/*
* If we're checking the creation of an index before a
* table is fully created, don't save the index: it
* will need to be reopened once the table is complete.
*/
if (!table->cg_complete) {
WT_ERR(
__wt_schema_destroy_index(session, &idx));
if (idxname != NULL)
break;
continue;
}
table->indices[i] = idx;
idx = NULL;
/*
* If the slot is bigger than anything else we've seen,
* bump the number of indices.
*/
if (i >= table->nindices)
table->nindices = i + 1;
}
/* If we were looking for a single index, we're done. */
if (indexp != NULL)
*indexp = table->indices[i];
if (idxname != NULL)
break;
}
WT_ERR_NOTFOUND_OK(ret);
if (idxname != NULL && !match)
ret = WT_NOTFOUND;
/* If we did a full pass, we won't need to do it again. */
if (idxname == NULL) {
table->nindices = i;
table->idx_complete = true;
}
err: WT_TRET(__wt_metadata_cursor_release(session, &cursor));
WT_TRET(__wt_schema_destroy_index(session, &idx));
__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);
}
/*
* __lsm_bloom_create --
* Create a bloom filter for a chunk of the LSM tree that has been
* checkpointed but not yet been merged.
*/
static int
__lsm_bloom_create(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk, u_int chunk_off)
{
WT_BLOOM *bloom;
WT_CURSOR *src;
WT_DECL_RET;
WT_ITEM key;
uint64_t insert_count;
WT_RET(__wt_lsm_tree_setup_bloom(session, lsm_tree, chunk));
bloom = NULL;
/*
* This is merge-like activity, and we don't want compacts to give up
* because we are creating a bunch of bloom filters before merging.
*/
++lsm_tree->merge_progressing;
WT_RET(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config, chunk->count,
lsm_tree->bloom_bit_count, lsm_tree->bloom_hash_count, &bloom));
/* Open a special merge cursor just on this chunk. */
WT_ERR(__wt_open_cursor(session, lsm_tree->name, NULL, NULL, &src));
F_SET(src, WT_CURSTD_RAW);
WT_ERR(__wt_clsm_init_merge(src, chunk_off, chunk->id, 1));
/*
* Setup so that we don't hold pages we read into cache, and so
* that we don't get stuck if the cache is full. If we allow
* ourselves to get stuck creating bloom filters, the entire tree
* can stall since there may be no worker threads available to flush.
*/
F_SET(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
WT_ERR(src->get_key(src, &key));
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_TRET(src->close(src));
WT_TRET(__wt_bloom_finalize(bloom));
WT_ERR(ret);
F_CLR(session, WT_SESSION_NO_CACHE);
/* Load the new Bloom filter into cache. */
WT_CLEAR(key);
WT_ERR_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"LSM worker created bloom filter %s. "
"Expected %" PRIu64 " items, got %" PRIu64,
chunk->bloom_uri, chunk->count, insert_count));
/* Ensure the bloom filter is in the metadata. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM bloom worker metadata write");
err: if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
return (ret);
}
/*
* __curjoin_init_bloom --
* Populate Bloom filters
*/
static int
__curjoin_init_bloom(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry, WT_BLOOM *bloom)
{
WT_COLLATOR *collator;
WT_CURSOR *c;
WT_CURSOR_JOIN_ENDPOINT *end, *endmax;
WT_DECL_ITEM(uribuf);
WT_DECL_RET;
WT_ITEM curkey, curvalue;
size_t size;
u_int skip;
int cmp;
const char *uri;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
c = NULL;
skip = 0;
if (entry->index != NULL)
/*
* Open the raw index. We're avoiding any references
* to the main table, they may be expensive.
*/
uri = entry->index->source;
else {
/*
* For joins on the main table, we just need the primary
* key for comparison, we don't need any values.
*/
size = strlen(cjoin->table->iface.name) + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s()",
cjoin->table->iface.name));
uri = uribuf->data;
}
WT_ERR(__wt_open_cursor(session, uri, &cjoin->iface, raw_cfg, &c));
/* Initially position the cursor if necessary. */
endmax = &entry->ends[entry->ends_next];
if ((end = &entry->ends[0]) < endmax) {
if (F_ISSET(end, WT_CURJOIN_END_GT) ||
WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_EQ) {
WT_ERR(__wt_cursor_dup_position(end->cursor, c));
if (WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_GE)
skip = 1;
} else if (F_ISSET(end, WT_CURJOIN_END_LT)) {
if ((ret = c->next(c)) == WT_NOTFOUND)
goto done;
WT_ERR(ret);
} else
WT_PANIC_ERR(session, EINVAL,
"fatal error in join cursor position state");
}
collator = (entry->index == NULL) ? NULL : entry->index->collator;
while (ret == 0) {
WT_ERR(c->get_key(c, &curkey));
entry->stats.iterated++;
if (entry->index != NULL) {
/*
* Repack so it's comparable to the
* reference endpoints.
*/
WT_ERR(__wt_struct_repack(session,
c->key_format,
(entry->repack_format != NULL ?
entry->repack_format : entry->index->idxkey_format),
&c->key, &curkey));
}
for (end = &entry->ends[skip]; end < endmax; end++) {
WT_ERR(__wt_compare(session, collator, &curkey,
&end->key, &cmp));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION)) {
/* if condition satisfied, insert immediately */
switch (WT_CURJOIN_END_RANGE(end)) {
case WT_CURJOIN_END_EQ:
if (cmp == 0)
goto insert;
break;
case WT_CURJOIN_END_GT:
if (cmp > 0) {
/* skip this check next time */
skip = entry->ends_next;
goto insert;
}
break;
case WT_CURJOIN_END_GE:
if (cmp >= 0)
goto insert;
break;
case WT_CURJOIN_END_LT:
if (cmp < 0)
goto insert;
break;
case WT_CURJOIN_END_LE:
if (cmp <= 0)
goto insert;
break;
}
} else if (!F_ISSET(end, WT_CURJOIN_END_LT)) {
if (cmp < 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto advance;
if (cmp > 0) {
if (F_ISSET(end, WT_CURJOIN_END_GT))
skip = 1;
else
goto done;
}
} else {
if (cmp > 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto done;
}
}
/*
* Either it's a disjunction that hasn't satisfied any
* condition, or it's a conjunction that has satisfied all
* conditions.
*/
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION))
goto advance;
insert:
if (entry->index != NULL) {
curvalue.data =
(unsigned char *)curkey.data + curkey.size;
WT_ASSERT(session, c->key.size > curkey.size);
curvalue.size = c->key.size - curkey.size;
}
else
WT_ERR(c->get_key(c, &curvalue));
__wt_bloom_insert(bloom, &curvalue);
entry->stats.bloom_insert++;
advance:
if ((ret = c->next(c)) == WT_NOTFOUND)
break;
}
done:
WT_ERR_NOTFOUND_OK(ret);
err: if (c != NULL)
WT_TRET(c->close(c));
__wt_scr_free(session, &uribuf);
return (ret);
}