/*
* __wt_txn_config --
* Configure a transaction.
*/
int
__wt_txn_config(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_TXN *txn;
txn = &session->txn;
WT_RET(__wt_config_gets_def(session, cfg, "isolation", 0, &cval));
if (cval.len != 0)
txn->isolation =
WT_STRING_MATCH("snapshot", cval.str, cval.len) ?
WT_ISO_SNAPSHOT :
WT_STRING_MATCH("read-committed", cval.str, cval.len) ?
WT_ISO_READ_COMMITTED : WT_ISO_READ_UNCOMMITTED;
/*
* The default sync setting is inherited from the connection, but can
* be overridden by an explicit "sync" setting for this transaction.
*
* We want to distinguish between inheriting implicitly and explicitly.
*/
F_CLR(txn, WT_TXN_SYNC_SET);
WT_RET(__wt_config_gets_def(
session, cfg, "sync", (int)UINT_MAX, &cval));
if (cval.val == 0 || cval.val == 1)
/*
* This is an explicit setting of sync. Set the flag so
* that we know not to overwrite it in commit_transaction.
*/
F_SET(txn, WT_TXN_SYNC_SET);
/*
* If sync is turned off explicitly, clear the transaction's sync field.
*/
if (cval.val == 0)
txn->txn_logsync = 0;
WT_RET(__wt_config_gets_def(session, cfg, "snapshot", 0, &cval));
if (cval.len > 0)
/*
* The layering here isn't ideal - the named snapshot get
* function does both validation and setup. Otherwise we'd
* need to walk the list of named snapshots twice during
* transaction open.
*/
WT_RET(__wt_txn_named_snapshot_get(session, &cval));
return (0);
}
/*
* __wt_txn_begin --
* Begin a transaction.
*/
int
__wt_txn_begin(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_TXN *txn;
WT_TXN_GLOBAL *txn_global;
WT_TXN_STATE *txn_state;
conn = S2C(session);
txn = &session->txn;
txn_global = &conn->txn_global;
txn_state = &txn_global->states[session->id];
WT_ASSERT(session, txn_state->id == WT_TXN_NONE);
WT_RET(__wt_config_gets_def(session, cfg, "isolation", 0, &cval));
if (cval.len == 0)
txn->isolation = session->isolation;
else
txn->isolation =
WT_STRING_MATCH("snapshot", cval.str, cval.len) ?
TXN_ISO_SNAPSHOT :
WT_STRING_MATCH("read-committed", cval.str, cval.len) ?
TXN_ISO_READ_COMMITTED : TXN_ISO_READ_UNCOMMITTED;
/*
* Allocate a transaction ID.
*
* We use an atomic compare and swap to ensure that we get a
* unique ID that is published before the global counter is
* updated.
*
* If two threads race to allocate an ID, only the latest ID
* will proceed. The winning thread can be sure its snapshot
* contains all of the earlier active IDs. Threads that race
* and get an earlier ID may not appear in the snapshot, but
* they will loop and allocate a new ID before proceeding to
* make any updates.
*
* This potentially wastes transaction IDs when threads race to
* begin transactions: that is the price we pay to keep this
* path latch free.
*/
do {
txn_state->id = txn->id = txn_global->current;
} while (!WT_ATOMIC_CAS(txn_global->current, txn->id, txn->id + 1));
/*
* If we have used 64-bits of transaction IDs, there is nothing
* more we can do.
*/
if (txn->id == WT_TXN_ABORTED)
WT_RET_MSG(session, ENOMEM, "Out of transaction IDs");
F_SET(txn, TXN_RUNNING);
if (txn->isolation == TXN_ISO_SNAPSHOT)
__wt_txn_refresh(session, WT_TXN_NONE, 1);
return (0);
}
/*
* __wt_curfile_open --
* WT_SESSION->open_cursor method for the btree cursor type.
*/
int
__wt_curfile_open(WT_SESSION_IMPL *session, const char *uri,
WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
uint32_t flags;
bool bitmap, bulk;
bitmap = bulk = false;
flags = 0;
WT_RET(__wt_config_gets_def(session, cfg, "bulk", 0, &cval));
if (cval.type == WT_CONFIG_ITEM_BOOL ||
(cval.type == WT_CONFIG_ITEM_NUM &&
(cval.val == 0 || cval.val == 1))) {
bitmap = false;
bulk = cval.val != 0;
} else if (WT_STRING_MATCH("bitmap", cval.str, cval.len))
bitmap = bulk = true;
else
WT_RET_MSG(session, EINVAL,
"Value for 'bulk' must be a boolean or 'bitmap'");
/* Bulk handles require exclusive access. */
if (bulk)
LF_SET(WT_BTREE_BULK | WT_DHANDLE_EXCLUSIVE);
/* Get the handle and lock it while the cursor is using it. */
if (WT_PREFIX_MATCH(uri, "file:")) {
/*
* If we are opening a bulk cursor, get the handle while
* holding the checkpoint lock. This prevents a bulk cursor
* open failing with EBUSY due to a database-wide checkpoint.
*/
if (bulk)
__wt_spin_lock(
session, &S2C(session)->checkpoint_lock);
ret = __wt_session_get_btree_ckpt(session, uri, cfg, flags);
if (bulk)
__wt_spin_unlock(
session, &S2C(session)->checkpoint_lock);
WT_RET(ret);
} else
WT_RET(__wt_bad_object_type(session, uri));
WT_ERR(__wt_curfile_create(session, owner, cfg, bulk, bitmap, cursorp));
/* Increment the data-source's in-use counter. */
__wt_cursor_dhandle_incr_use(session);
return (0);
err: /* If the cursor could not be opened, release the handle. */
WT_TRET(__wt_session_release_btree(session));
return (ret);
}
/*
* __logmgr_config --
* Parse and setup the logging server options.
*/
static int
__logmgr_config(WT_SESSION_IMPL *session, const char **cfg, int *runp)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
conn = S2C(session);
/*
* The logging configuration is off by default.
*/
WT_RET(__wt_config_gets(session, cfg, "log.enabled", &cval));
*runp = cval.val != 0;
/*
* Setup a log path, compression and encryption even if logging is
* disabled in case we are going to print a log.
*/
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 == 0)
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);
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);
WT_RET(__wt_config_gets(session, cfg, "log.prealloc", &cval));
/*
* If pre-allocation is configured, set the initial number to one.
* We'll adapt as load dictates.
*/
if (cval.val != 0) {
FLD_SET(conn->log_flags, WT_CONN_LOG_PREALLOC);
conn->log_prealloc = 1;
}
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(__logmgr_sync_cfg(session, cfg));
return (0);
}
/*
* __wt_schema_drop --
* Process a WT_SESSION::drop operation for all supported types.
*/
int
__wt_schema_drop(WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_DATA_SOURCE *dsrc;
WT_DECL_RET;
bool force;
WT_RET(__wt_config_gets_def(session, cfg, "force", 0, &cval));
force = cval.val != 0;
WT_RET(__wt_meta_track_on(session));
/* Paranoia: clear any handle from our caller. */
session->dhandle = NULL;
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, cfg);
else if (WT_PREFIX_MATCH(uri, "index:"))
ret = __drop_index(session, uri, force, cfg);
else if (WT_PREFIX_MATCH(uri, "lsm:"))
ret = __wt_lsm_tree_drop(session, uri, cfg);
else if (WT_PREFIX_MATCH(uri, "table:"))
ret = __drop_table(session, uri, cfg);
else if ((dsrc = __wt_schema_get_source(session, uri)) != NULL)
ret = dsrc->drop == NULL ?
__wt_object_unsupported(session, uri) :
dsrc->drop(
dsrc, &session->iface, uri, (WT_CONFIG_ARG *)cfg);
else
ret = __wt_bad_object_type(session, uri);
/*
* Map WT_NOTFOUND to ENOENT, based on the assumption WT_NOTFOUND means
* there was no metadata entry. Map ENOENT to zero if force is set.
*/
if (ret == WT_NOTFOUND || ret == ENOENT)
ret = force ? 0 : ENOENT;
/* Bump the schema generation so that stale data is ignored. */
++S2C(session)->schema_gen;
WT_TRET(__wt_meta_track_off(session, true, ret != 0));
return (ret);
}
/*
* __wt_curfile_create --
* Open a cursor for a given btree handle.
*/
int
__wt_curfile_create(WT_SESSION_IMPL *session,
WT_CURSOR *owner, const char *cfg[], int bulk, int bitmap,
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 */
__curfile_compare, /* compare */
__curfile_equals, /* equals */
__curfile_next, /* next */
__curfile_prev, /* prev */
__curfile_reset, /* reset */
__curfile_search, /* search */
__curfile_search_near, /* search-near */
__curfile_insert, /* insert */
__curfile_update, /* update */
__curfile_remove, /* remove */
__wt_cursor_reconfigure, /* reconfigure */
__curfile_close); /* close */
WT_BTREE *btree;
WT_CONFIG_ITEM cval;
WT_CURSOR *cursor;
WT_CURSOR_BTREE *cbt;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
size_t csize;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_BTREE, iface) == 0);
cbt = NULL;
btree = S2BT(session);
WT_ASSERT(session, btree != NULL);
csize = bulk ? sizeof(WT_CURSOR_BULK) : sizeof(WT_CURSOR_BTREE);
WT_RET(__wt_calloc(session, 1, csize, &cbt));
cursor = &cbt->iface;
*cursor = iface;
cursor->session = &session->iface;
cursor->internal_uri = btree->dhandle->name;
cursor->key_format = btree->key_format;
cursor->value_format = btree->value_format;
cbt->btree = btree;
if (bulk) {
F_SET(cursor, WT_CURSTD_BULK);
cbulk = (WT_CURSOR_BULK *)cbt;
/* Optionally skip the validation of each bulk-loaded key. */
WT_ERR(__wt_config_gets_def(
session, cfg, "skip_sort_check", 0, &cval));
WT_ERR(__wt_curbulk_init(
session, cbulk, bitmap, cval.val == 0 ? 0 : 1));
}
/*
* random_retrieval
* Random retrieval cursors only support next, reset and close.
*/
WT_ERR(__wt_config_gets_def(session, cfg, "next_random", 0, &cval));
if (cval.val != 0) {
__wt_cursor_set_notsup(cursor);
cursor->next = __curfile_next_random;
cursor->reset = __curfile_reset;
}
/* Underlying btree initialization. */
__wt_btcur_open(cbt);
/* __wt_cursor_init is last so we don't have to clean up on error. */
WT_ERR(__wt_cursor_init(
cursor, cursor->internal_uri, owner, cfg, cursorp));
WT_STAT_FAST_CONN_INCR(session, cursor_create);
WT_STAT_FAST_DATA_INCR(session, cursor_create);
if (0) {
err: __wt_free(session, cbt);
}
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);
}
/*
* __wt_txn_commit --
* Commit the current transaction.
*/
int
__wt_txn_commit(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_TXN *txn;
WT_TXN_OP *op;
u_int i;
txn = &session->txn;
conn = S2C(session);
WT_ASSERT(session, !F_ISSET(txn, WT_TXN_ERROR) || txn->mod_count == 0);
if (!F_ISSET(txn, WT_TXN_RUNNING))
WT_RET_MSG(session, EINVAL, "No transaction is active");
/*
* The default sync setting is inherited from the connection, but can
* be overridden by an explicit "sync" setting for this transaction.
*/
WT_RET(__wt_config_gets_def(session, cfg, "sync", 0, &cval));
/*
* If the user chose the default setting, check whether sync is enabled
* for this transaction (either inherited or via begin_transaction).
* If sync is disabled, clear the field to avoid the log write being
* flushed.
*
* Otherwise check for specific settings. We don't need to check for
* "on" because that is the default inherited from the connection. If
* the user set anything in begin_transaction, we only override with an
* explicit setting.
*/
if (cval.len == 0) {
if (!FLD_ISSET(txn->txn_logsync, WT_LOG_FLUSH) &&
!F_ISSET(txn, WT_TXN_SYNC_SET))
txn->txn_logsync = 0;
} else {
/*
* If the caller already set sync on begin_transaction then
* they should not be using sync on commit_transaction.
* Flag that as an error.
*/
if (F_ISSET(txn, WT_TXN_SYNC_SET))
WT_RET_MSG(session, EINVAL,
"Sync already set during begin_transaction.");
if (WT_STRING_MATCH("background", cval.str, cval.len))
txn->txn_logsync = WT_LOG_BACKGROUND;
else if (WT_STRING_MATCH("off", cval.str, cval.len))
txn->txn_logsync = 0;
/*
* We don't need to check for "on" here because that is the
* default to inherit from the connection setting.
*/
}
/* Commit notification. */
if (txn->notify != NULL)
WT_TRET(txn->notify->notify(txn->notify,
(WT_SESSION *)session, txn->id, 1));
/* If we are logging, write a commit log record. */
if (ret == 0 && txn->mod_count > 0 &&
FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED) &&
!F_ISSET(session, WT_SESSION_NO_LOGGING)) {
/*
* We are about to block on I/O writing the log.
* Release our snapshot in case it is keeping data pinned.
* This is particularly important for checkpoints.
*/
__wt_txn_release_snapshot(session);
ret = __wt_txn_log_commit(session, cfg);
WT_ASSERT(session, ret == 0);
}
/*
* If anything went wrong, roll back.
*
* !!!
* Nothing can fail after this point.
*/
if (ret != 0) {
WT_TRET(__wt_txn_rollback(session, cfg));
return (ret);
}
/* Free memory associated with updates. */
for (i = 0, op = txn->mod; i < txn->mod_count; i++, op++)
__wt_txn_op_free(session, op);
txn->mod_count = 0;
/*
* We are about to release the snapshot: copy values into any
* positioned cursors so they don't point to updates that could be
* freed once we don't have a transaction ID pinned.
*/
if (session->ncursors > 0)
WT_RET(__wt_session_copy_values(session));
__wt_txn_release(session);
return (0);
}
/*
* __wt_curfile_open --
* WT_SESSION->open_cursor method for the btree cursor type.
*/
int
__wt_curfile_open(WT_SESSION_IMPL *session, const char *uri,
WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
uint32_t flags;
bool bitmap, bulk, checkpoint_wait;
bitmap = bulk = false;
checkpoint_wait = true;
flags = 0;
/*
* Decode the bulk configuration settings. In memory databases
* ignore bulk load.
*/
if (!F_ISSET(S2C(session), WT_CONN_IN_MEMORY)) {
WT_RET(__wt_config_gets_def(session, cfg, "bulk", 0, &cval));
if (cval.type == WT_CONFIG_ITEM_BOOL ||
(cval.type == WT_CONFIG_ITEM_NUM &&
(cval.val == 0 || cval.val == 1))) {
bitmap = false;
bulk = cval.val != 0;
} else if (WT_STRING_MATCH("bitmap", cval.str, cval.len))
bitmap = bulk = true;
/*
* Unordered bulk insert is a special case used
* internally by index creation on existing tables. It
* doesn't enforce any special semantics at the file
* level. It primarily exists to avoid some locking
* problems between LSM and index creation.
*/
else if (!WT_STRING_MATCH("unordered", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"Value for 'bulk' must be a boolean or 'bitmap'");
if (bulk) {
WT_RET(__wt_config_gets(session,
cfg, "checkpoint_wait", &cval));
checkpoint_wait = cval.val != 0;
}
}
/* Bulk handles require exclusive access. */
if (bulk)
LF_SET(WT_BTREE_BULK | WT_DHANDLE_EXCLUSIVE);
/* Get the handle and lock it while the cursor is using it. */
if (WT_PREFIX_MATCH(uri, "file:")) {
/*
* If we are opening exclusive and don't want a bulk cursor
* open to fail with EBUSY due to a database-wide checkpoint,
* get the handle while holding the checkpoint lock.
*/
if (LF_ISSET(WT_DHANDLE_EXCLUSIVE) && checkpoint_wait)
WT_WITH_CHECKPOINT_LOCK(session,
ret = __wt_session_get_btree_ckpt(
session, uri, cfg, flags));
else
ret = __wt_session_get_btree_ckpt(
session, uri, cfg, flags);
WT_RET(ret);
} else
WT_RET(__wt_bad_object_type(session, uri));
WT_ERR(__curfile_create(session, owner, cfg, bulk, bitmap, cursorp));
return (0);
err: /* If the cursor could not be opened, release the handle. */
WT_TRET(__wt_session_release_btree(session));
return (ret);
}
/*
* __curfile_create --
* Open a cursor for a given btree handle.
*/
static int
__curfile_create(WT_SESSION_IMPL *session,
WT_CURSOR *owner, const char *cfg[], bool bulk, bool bitmap,
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 */
__curfile_compare, /* compare */
__curfile_equals, /* equals */
__curfile_next, /* next */
__curfile_prev, /* prev */
__curfile_reset, /* reset */
__curfile_search, /* search */
__curfile_search_near, /* search-near */
__curfile_insert, /* insert */
__wt_cursor_modify_notsup, /* modify */
__curfile_update, /* update */
__curfile_remove, /* remove */
__curfile_reserve, /* reserve */
__wt_cursor_reconfigure, /* reconfigure */
__curfile_close); /* close */
WT_BTREE *btree;
WT_CONFIG_ITEM cval;
WT_CURSOR *cursor;
WT_CURSOR_BTREE *cbt;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
size_t csize;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_BTREE, iface) == 0);
cbt = NULL;
btree = S2BT(session);
WT_ASSERT(session, btree != NULL);
csize = bulk ? sizeof(WT_CURSOR_BULK) : sizeof(WT_CURSOR_BTREE);
WT_RET(__wt_calloc(session, 1, csize, &cbt));
cursor = &cbt->iface;
*cursor = iface;
cursor->session = &session->iface;
cursor->internal_uri = btree->dhandle->name;
cursor->key_format = btree->key_format;
cursor->value_format = btree->value_format;
cbt->btree = btree;
/*
* Increment the data-source's in-use counter; done now because closing
* the cursor will decrement it, and all failure paths from here close
* the cursor.
*/
__wt_cursor_dhandle_incr_use(session);
if (session->dhandle->checkpoint != NULL)
F_SET(cbt, WT_CBT_NO_TXN);
if (bulk) {
F_SET(cursor, WT_CURSTD_BULK);
cbulk = (WT_CURSOR_BULK *)cbt;
/* Optionally skip the validation of each bulk-loaded key. */
WT_ERR(__wt_config_gets_def(
session, cfg, "skip_sort_check", 0, &cval));
WT_ERR(__wt_curbulk_init(
session, cbulk, bitmap, cval.val == 0 ? 0 : 1));
}
/*
* Random retrieval, row-store only.
* Random retrieval cursors support a limited set of methods.
*/
WT_ERR(__wt_config_gets_def(session, cfg, "next_random", 0, &cval));
if (cval.val != 0) {
if (WT_CURSOR_RECNO(cursor))
WT_ERR_MSG(session, ENOTSUP,
"next_random configuration not supported for "
"column-store objects");
__wt_cursor_set_notsup(cursor);
cursor->next = __wt_curfile_next_random;
cursor->reset = __curfile_reset;
WT_ERR(__wt_config_gets_def(
session, cfg, "next_random_sample_size", 0, &cval));
if (cval.val != 0)
cbt->next_random_sample_size = (u_int)cval.val;
}
/* Underlying btree initialization. */
__wt_btcur_open(cbt);
/*
* WT_CURSOR.modify supported on 'u' value formats, but the fast-path
* through the btree code requires log file format changes, it's not
* available in all versions.
*/
if (WT_STREQ(cursor->value_format, "u") &&
S2C(session)->compat_major >= WT_LOG_V2)
cursor->modify = __curfile_modify;
WT_ERR(__wt_cursor_init(
cursor, cursor->internal_uri, owner, cfg, cursorp));
WT_STAT_CONN_INCR(session, cursor_create);
WT_STAT_DATA_INCR(session, cursor_create);
if (0) {
err: /*
* Our caller expects to release the data handle if we fail.
* Disconnect it from the cursor before closing.
*/
if (session->dhandle != NULL)
__wt_cursor_dhandle_decr_use(session);
cbt->btree = NULL;
WT_TRET(__curfile_close(cursor));
*cursorp = NULL;
}
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;
/*
* A note on reconfiguration: the standard "is this configuration string
* allowed" checks should fail if reconfiguration has invalid strings,
* for example, "log=(enabled)", or "statistics_log=(path=XXX)", because
* the connection reconfiguration method doesn't allow those strings.
* Additionally, the base configuration values during reconfiguration
* are the currently configured values (so we don't revert to default
* values when repeatedly reconfiguring), and configuration processing
* of a currently set value should not change the currently set value.
*
* In this code path, log server reconfiguration does not stop/restart
* the log server, so there's no point in re-evaluating configuration
* strings that cannot be reconfigured, risking bugs in configuration
* setup, and depending on evaluation of currently set values to always
* result in the currently set value. Skip tests for any configuration
* strings which don't make sense during reconfiguration, but don't
* worry about error reporting because it should never happen.
*/
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.
*
* See above: should never happen.
*/
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.
*
* See above: should never happen.
*/
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);
/*
* The file size cannot be reconfigured. 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.
*
* See above: should never happen.
*/
if (!reconfig) {
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 it's meaningless to reconfigure this value during runtime, it
* only matters on create before recovery runs.
*
* See above: should never happen.
*/
if (!reconfig) {
WT_RET(__wt_config_gets_def(
session, cfg, "log.recover", 0, &cval));
if (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);
}
/*
* __wt_txn_commit --
* Commit the current transaction.
*/
int
__wt_txn_commit(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_TXN *txn;
WT_TXN_GLOBAL *txn_global;
WT_TXN_OP *op;
u_int i;
bool locked, readonly;
#ifdef HAVE_TIMESTAMPS
wt_timestamp_t prev_commit_timestamp, ts;
bool update_timestamp;
#endif
txn = &session->txn;
conn = S2C(session);
txn_global = &conn->txn_global;
locked = false;
WT_ASSERT(session, F_ISSET(txn, WT_TXN_RUNNING));
WT_ASSERT(session, !F_ISSET(txn, WT_TXN_ERROR) ||
txn->mod_count == 0);
readonly = txn->mod_count == 0;
/*
* Look for a commit timestamp.
*/
WT_ERR(
__wt_config_gets_def(session, cfg, "commit_timestamp", 0, &cval));
if (cval.len != 0) {
#ifdef HAVE_TIMESTAMPS
WT_ERR(__wt_txn_parse_timestamp(session, "commit", &ts, &cval));
WT_ERR(__wt_timestamp_validate(session,
"commit", &ts, &cval, true, true, true));
__wt_timestamp_set(&txn->commit_timestamp, &ts);
__wt_txn_set_commit_timestamp(session);
#else
WT_ERR_MSG(session, EINVAL, "commit_timestamp requires a "
"version of WiredTiger built with timestamp support");
#endif
}
#ifdef HAVE_TIMESTAMPS
/*
* Debugging checks on timestamps, if user requested them.
*/
if (F_ISSET(txn, WT_TXN_TS_COMMIT_ALWAYS) &&
!F_ISSET(txn, WT_TXN_HAS_TS_COMMIT) &&
txn->mod_count != 0)
WT_ERR_MSG(session, EINVAL, "commit_timestamp required and "
"none set on this transaction");
if (F_ISSET(txn, WT_TXN_TS_COMMIT_NEVER) &&
F_ISSET(txn, WT_TXN_HAS_TS_COMMIT) &&
txn->mod_count != 0)
WT_ERR_MSG(session, EINVAL, "no commit_timestamp required and "
"timestamp set on this transaction");
#endif
/*
* The default sync setting is inherited from the connection, but can
* be overridden by an explicit "sync" setting for this transaction.
*/
WT_ERR(__wt_config_gets_def(session, cfg, "sync", 0, &cval));
/*
* If the user chose the default setting, check whether sync is enabled
* for this transaction (either inherited or via begin_transaction).
* If sync is disabled, clear the field to avoid the log write being
* flushed.
*
* Otherwise check for specific settings. We don't need to check for
* "on" because that is the default inherited from the connection. If
* the user set anything in begin_transaction, we only override with an
* explicit setting.
*/
if (cval.len == 0) {
if (!FLD_ISSET(txn->txn_logsync, WT_LOG_SYNC_ENABLED) &&
!F_ISSET(txn, WT_TXN_SYNC_SET))
txn->txn_logsync = 0;
} else {
/*
* If the caller already set sync on begin_transaction then
* they should not be using sync on commit_transaction.
* Flag that as an error.
*/
if (F_ISSET(txn, WT_TXN_SYNC_SET))
WT_ERR_MSG(session, EINVAL,
"Sync already set during begin_transaction");
if (WT_STRING_MATCH("background", cval.str, cval.len))
txn->txn_logsync = WT_LOG_BACKGROUND;
else if (WT_STRING_MATCH("off", cval.str, cval.len))
txn->txn_logsync = 0;
/*
* We don't need to check for "on" here because that is the
* default to inherit from the connection setting.
*/
}
/* Commit notification. */
if (txn->notify != NULL)
WT_ERR(txn->notify->notify(txn->notify,
(WT_SESSION *)session, txn->id, 1));
/*
* We are about to release the snapshot: copy values into any
* positioned cursors so they don't point to updates that could be
* freed once we don't have a snapshot.
*/
if (session->ncursors > 0) {
WT_DIAGNOSTIC_YIELD;
WT_ERR(__wt_session_copy_values(session));
}
/* If we are logging, write a commit log record. */
if (txn->logrec != NULL &&
FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED) &&
!F_ISSET(session, WT_SESSION_NO_LOGGING)) {
/*
* We are about to block on I/O writing the log.
* Release our snapshot in case it is keeping data pinned.
* This is particularly important for checkpoints.
*/
__wt_txn_release_snapshot(session);
/*
* We hold the visibility lock for reading from the time
* we write our log record until the time we release our
* transaction so that the LSN any checkpoint gets will
* always reflect visible data.
*/
__wt_readlock(session, &txn_global->visibility_rwlock);
locked = true;
WT_ERR(__wt_txn_log_commit(session, cfg));
}
/* Note: we're going to commit: nothing can fail after this point. */
/* Process and free updates. */
for (i = 0, op = txn->mod; i < txn->mod_count; i++, op++) {
switch (op->type) {
case WT_TXN_OP_BASIC:
case WT_TXN_OP_BASIC_TS:
case WT_TXN_OP_INMEM:
/*
* Switch reserved operations to abort to
* simplify obsolete update list truncation.
*/
if (op->u.upd->type == WT_UPDATE_RESERVED) {
op->u.upd->txnid = WT_TXN_ABORTED;
break;
}
/*
* Writes to the lookaside file can be evicted as soon
* as they commit.
*/
if (conn->cache->las_fileid != 0 &&
op->fileid == conn->cache->las_fileid) {
op->u.upd->txnid = WT_TXN_NONE;
break;
}
#ifdef HAVE_TIMESTAMPS
if (F_ISSET(txn, WT_TXN_HAS_TS_COMMIT) &&
op->type != WT_TXN_OP_BASIC_TS) {
WT_ASSERT(session,
op->fileid != WT_METAFILE_ID);
__wt_timestamp_set(&op->u.upd->timestamp,
&txn->commit_timestamp);
}
#endif
break;
case WT_TXN_OP_REF:
#ifdef HAVE_TIMESTAMPS
if (F_ISSET(txn, WT_TXN_HAS_TS_COMMIT))
__wt_timestamp_set(
&op->u.ref->page_del->timestamp,
&txn->commit_timestamp);
#endif
break;
case WT_TXN_OP_TRUNCATE_COL:
case WT_TXN_OP_TRUNCATE_ROW:
/* Other operations don't need timestamps. */
break;
}
__wt_txn_op_free(session, op);
}
txn->mod_count = 0;
#ifdef HAVE_TIMESTAMPS
/*
* Track the largest commit timestamp we have seen.
*
* We don't actually clear the local commit timestamp, just the flag.
* That said, we can't update the global commit timestamp until this
* transaction is visible, which happens when we release it.
*/
update_timestamp = F_ISSET(txn, WT_TXN_HAS_TS_COMMIT);
#endif
__wt_txn_release(session);
if (locked)
__wt_readunlock(session, &txn_global->visibility_rwlock);
#ifdef HAVE_TIMESTAMPS
/* First check if we've already committed something in the future. */
if (update_timestamp) {
WT_WITH_TIMESTAMP_READLOCK(session, &txn_global->rwlock,
__wt_timestamp_set(
&prev_commit_timestamp, &txn_global->commit_timestamp));
update_timestamp = __wt_timestamp_cmp(
&txn->commit_timestamp, &prev_commit_timestamp) > 0;
}
/*
* If it looks like we need to move the global commit timestamp,
* write lock and re-check.
*/
if (update_timestamp) {
#if WT_TIMESTAMP_SIZE == 8
while (__wt_timestamp_cmp(
&txn->commit_timestamp, &prev_commit_timestamp) > 0) {
if (__wt_atomic_cas64(
&txn_global->commit_timestamp.val,
prev_commit_timestamp.val,
txn->commit_timestamp.val)) {
txn_global->has_commit_timestamp = true;
break;
}
__wt_timestamp_set(
&prev_commit_timestamp, &txn_global->commit_timestamp);
}
#else
__wt_writelock(session, &txn_global->rwlock);
if (__wt_timestamp_cmp(&txn->commit_timestamp,
&txn_global->commit_timestamp) > 0) {
__wt_timestamp_set(&txn_global->commit_timestamp,
&txn->commit_timestamp);
txn_global->has_commit_timestamp = true;
}
__wt_writeunlock(session, &txn_global->rwlock);
#endif
}
#endif
/*
* We're between transactions, if we need to block for eviction, it's
* a good time to do so. Note that we must ignore any error return
* because the user's data is committed.
*/
if (!readonly)
(void)__wt_cache_eviction_check(session, false, false, NULL);
return (0);
err: /*
* If anything went wrong, roll back.
*
* !!!
* Nothing can fail after this point.
*/
if (locked)
__wt_readunlock(session, &txn_global->visibility_rwlock);
WT_TRET(__wt_txn_rollback(session, cfg));
return (ret);
}
/*
* __wt_txn_config --
* Configure a transaction.
*/
int
__wt_txn_config(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_TXN *txn;
txn = &session->txn;
WT_RET(__wt_config_gets_def(session, cfg, "isolation", 0, &cval));
if (cval.len != 0)
txn->isolation =
WT_STRING_MATCH("snapshot", cval.str, cval.len) ?
WT_ISO_SNAPSHOT :
WT_STRING_MATCH("read-committed", cval.str, cval.len) ?
WT_ISO_READ_COMMITTED : WT_ISO_READ_UNCOMMITTED;
/*
* The default sync setting is inherited from the connection, but can
* be overridden by an explicit "sync" setting for this transaction.
*
* We want to distinguish between inheriting implicitly and explicitly.
*/
F_CLR(txn, WT_TXN_SYNC_SET);
WT_RET(__wt_config_gets_def(
session, cfg, "sync", (int)UINT_MAX, &cval));
if (cval.val == 0 || cval.val == 1)
/*
* This is an explicit setting of sync. Set the flag so
* that we know not to overwrite it in commit_transaction.
*/
F_SET(txn, WT_TXN_SYNC_SET);
/*
* If sync is turned off explicitly, clear the transaction's sync field.
*/
if (cval.val == 0)
txn->txn_logsync = 0;
WT_RET(__wt_config_gets_def(session, cfg, "snapshot", 0, &cval));
if (cval.len > 0)
/*
* The layering here isn't ideal - the named snapshot get
* function does both validation and setup. Otherwise we'd
* need to walk the list of named snapshots twice during
* transaction open.
*/
WT_RET(__wt_txn_named_snapshot_get(session, &cval));
WT_RET(__wt_config_gets_def(session, cfg, "read_timestamp", 0, &cval));
if (cval.len > 0) {
#ifdef HAVE_TIMESTAMPS
wt_timestamp_t ts;
WT_TXN_GLOBAL *txn_global;
char timestamp_buf[2 * WT_TIMESTAMP_SIZE + 1];
bool round_to_oldest;
txn_global = &S2C(session)->txn_global;
WT_RET(__wt_txn_parse_timestamp(session, "read", &ts, &cval));
/*
* Read the configuration here to reduce the span of the
* critical section.
*/
WT_RET(__wt_config_gets_def(session,
cfg, "round_to_oldest", 0, &cval));
round_to_oldest = cval.val;
/*
* This code is not using the timestamp validate function to
* avoid a race between checking and setting transaction
* timestamp.
*/
__wt_readlock(session, &txn_global->rwlock);
if (__wt_timestamp_cmp(&ts, &txn_global->oldest_timestamp) < 0)
{
WT_RET(__wt_timestamp_to_hex_string(session,
timestamp_buf, &ts));
/*
* If given read timestamp is earlier than oldest
* timestamp then round the read timestamp to
* oldest timestamp.
*/
if (round_to_oldest)
__wt_timestamp_set(&txn->read_timestamp,
&txn_global->oldest_timestamp);
else {
__wt_readunlock(session, &txn_global->rwlock);
WT_RET_MSG(session, EINVAL, "read timestamp "
"%s older than oldest timestamp",
timestamp_buf);
}
} else {
__wt_timestamp_set(&txn->read_timestamp, &ts);
/*
* Reset to avoid a verbose message as read
* timestamp is not rounded to oldest timestamp.
*/
round_to_oldest = false;
}
__wt_txn_set_read_timestamp(session);
__wt_readunlock(session, &txn_global->rwlock);
txn->isolation = WT_ISO_SNAPSHOT;
if (round_to_oldest) {
/*
* This message is generated here to reduce the span of
* critical section.
*/
__wt_verbose(session, WT_VERB_TIMESTAMP, "Read "
"timestamp %s : Rounded to oldest timestamp",
timestamp_buf);
}
#else
WT_RET_MSG(session, EINVAL, "read_timestamp requires a "
"version of WiredTiger built with timestamp support");
#endif
}
return (0);
}
/*
* __wt_curmetadata_open --
* WT_SESSION->open_cursor method for metadata cursors.
*
* Metadata cursors are a similar to a file cursor on the special metadata
* table, except that the metadata for the metadata table (which is stored
* in the turtle file) can also be queried.
*
* Metadata cursors are read-only by default.
*/
int
__wt_curmetadata_open(WT_SESSION_IMPL *session,
const char *uri, WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__wt_cursor_set_value, /* set-value */
__curmetadata_compare, /* compare */
__wt_cursor_equals, /* equals */
__curmetadata_next, /* next */
__curmetadata_prev, /* prev */
__curmetadata_reset, /* reset */
__curmetadata_search, /* search */
__curmetadata_search_near, /* search-near */
__curmetadata_insert, /* insert */
__curmetadata_update, /* update */
__curmetadata_remove, /* remove */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__curmetadata_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_METADATA *mdc;
WT_DECL_RET;
WT_CONFIG_ITEM cval;
WT_RET(__wt_calloc_one(session, &mdc));
cursor = &mdc->iface;
*cursor = iface;
cursor->session = &session->iface;
cursor->key_format = "S";
cursor->value_format = "S";
/*
* Open the file cursor for operations on the regular metadata; don't
* use the existing, cached session metadata cursor, the configuration
* may not be the same.
*/
WT_ERR(__wt_metadata_cursor_open(session, cfg[1], &mdc->file_cursor));
WT_ERR(__wt_cursor_init(cursor, uri, owner, cfg, cursorp));
/* If we are only returning create config, strip internal metadata. */
if (WT_STREQ(uri, "metadata:create"))
F_SET(mdc, WT_MDC_CREATEONLY);
/*
* Metadata cursors default to readonly; if not set to not-readonly,
* they are permanently readonly and cannot be reconfigured.
*/
WT_ERR(__wt_config_gets_def(session, cfg, "readonly", 1, &cval));
if (cval.val != 0) {
cursor->insert = __wt_cursor_notsup;
cursor->update = __wt_cursor_notsup;
cursor->remove = __wt_cursor_notsup;
}
if (0) {
err: WT_TRET(__curmetadata_close(cursor));
*cursorp = NULL;
}
return (ret);
}
