/*
* __ckpt_server_start --
* Start the checkpoint server thread.
*/
static int
__ckpt_server_start(WT_CONNECTION_IMPL *conn)
{
WT_SESSION_IMPL *session;
/* Nothing to do if the server is already running. */
if (conn->ckpt_session != NULL)
return (0);
F_SET(conn, WT_CONN_SERVER_CHECKPOINT);
/* The checkpoint server gets its own session. */
WT_RET(__wt_open_internal_session(
conn, "checkpoint-server", 1, 1, &conn->ckpt_session));
session = conn->ckpt_session;
/*
* Checkpoint does enough I/O it may be called upon to perform slow
* operations for the block manager.
*/
F_SET(session, WT_SESSION_CAN_WAIT);
WT_RET(
__wt_cond_alloc(session, "checkpoint server", 0, &conn->ckpt_cond));
/*
* Start the thread.
*/
WT_RET(__wt_thread_create(
session, &conn->ckpt_tid, __ckpt_server, session));
conn->ckpt_tid_set = 1;
return (0);
}
/*
* __wt_las_create --
* Initialize the database's lookaside store.
*/
int
__wt_las_create(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
uint32_t session_flags;
const char *drop_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_drop), "force=true", NULL };
conn = S2C(session);
/*
* Done at startup: we cannot do it on demand because we require the
* schema lock to create and drop the table, and it may not always be
* available.
*
* Discard any previous incarnation of the table.
*/
WT_RET(__wt_session_drop(session, WT_LAS_URI, drop_cfg));
/* Re-create the table. */
WT_RET(__wt_session_create(session, WT_LAS_URI, WT_LAS_FORMAT));
/*
* Open a shared internal session used to access the lookaside table.
* This session should never be tapped for eviction.
*/
session_flags = WT_SESSION_LOOKASIDE_CURSOR | WT_SESSION_NO_EVICTION;
WT_RET(__wt_open_internal_session(
conn, "lookaside table", true, session_flags, &conn->las_session));
/* Flag that the lookaside table has been created. */
F_SET(conn, WT_CONN_LAS_OPEN);
return (0);
}
/*
* __wt_sweep_create --
* Start the handle sweep thread.
*/
int
__wt_sweep_create(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
conn = S2C(session);
/* Set first, the thread might run before we finish up. */
F_SET(conn, WT_CONN_SERVER_SWEEP);
WT_RET(__wt_open_internal_session(
conn, "sweep-server", 1, 1, &conn->sweep_session));
session = conn->sweep_session;
/*
* Handle sweep does enough I/O it may be called upon to perform slow
* operations for the block manager.
*/
F_SET(session, WT_SESSION_CAN_WAIT);
WT_RET(__wt_cond_alloc(
session, "handle sweep server", 0, &conn->sweep_cond));
WT_RET(__wt_thread_create(
session, &conn->sweep_tid, __sweep_server, session));
conn->sweep_tid_set = 1;
return (0);
}
/*
* __ckpt_server_start --
* Start the checkpoint server thread.
*/
static int
__ckpt_server_start(WT_CONNECTION_IMPL *conn)
{
WT_SESSION_IMPL *session;
session = conn->default_session;
/* Nothing to do if the server is already running. */
if (conn->ckpt_session != NULL)
return (0);
F_SET(conn, WT_CONN_SERVER_CHECKPOINT);
/* The checkpoint server gets its own session. */
WT_RET(__wt_open_internal_session(
conn, "checkpoint-server", 1, 1, &conn->ckpt_session));
WT_RET(
__wt_cond_alloc(session, "checkpoint server", 0, &conn->ckpt_cond));
/*
* Start the thread.
*/
WT_RET(__wt_thread_create(
session, &conn->ckpt_tid, __ckpt_server, conn->ckpt_session));
conn->ckpt_tid_set = 1;
return (0);
}
/*
* __wt_connection_open --
* Open a connection.
*/
int
__wt_connection_open(WT_CONNECTION_IMPL *conn, const char *cfg[])
{
WT_SESSION_IMPL *session;
/* Default session. */
session = conn->default_session;
WT_ASSERT(session, session->iface.connection == &conn->iface);
/*
* Tell internal server threads to run: this must be set before opening
* any sessions.
*/
F_SET(conn, WT_CONN_SERVER_RUN | WT_CONN_LOG_SERVER_RUN);
/* WT_SESSION_IMPL array. */
WT_RET(__wt_calloc(session,
conn->session_size, sizeof(WT_SESSION_IMPL), &conn->sessions));
WT_CACHE_LINE_ALIGNMENT_VERIFY(session, conn->sessions);
/*
* Open the default session. We open this before starting service
* threads because those may allocate and use session resources that
* need to get cleaned up on close.
*/
WT_RET(__wt_open_internal_session(
conn, "connection", false, 0, &session));
/*
* The connection's default session is originally a static structure,
* swap that out for a more fully-functional session. It's necessary
* to have this step: the session allocation code uses the connection's
* session, and if we pass a reference to the default session as the
* place to store the allocated session, things get confused and error
* handling can be corrupted. So, we allocate into a stack variable
* and then assign it on success.
*/
conn->default_session = session;
/*
* Publish: there must be a barrier to ensure the connection structure
* fields are set before other threads read from the pointer.
*/
WT_WRITE_BARRIER();
/* Create the cache. */
WT_RET(__wt_cache_create(session, cfg));
/* Initialize transaction support. */
WT_RET(__wt_txn_global_init(session, cfg));
return (0);
}
/*
* __wt_meta_track_init --
* Initialize metadata tracking.
*/
int
__wt_meta_track_init(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
conn = S2C(session);
if (!FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED)) {
WT_RET(__wt_open_internal_session(conn,
"metadata-ckpt", false, WT_SESSION_NO_DATA_HANDLES,
&conn->meta_ckpt_session));
/*
* Sessions default to read-committed isolation, we rely on
* that for the correctness of metadata checkpoints.
*/
WT_ASSERT(session, conn->meta_ckpt_session->txn.isolation ==
WT_ISO_READ_COMMITTED);
}
return (0);
}
/*
* __logmgr_force_ckpt --
* Force a checkpoint out, waiting for the checkpoint LSN in the log
* is up to the given log number.
*/
static int
__logmgr_force_ckpt(WT_SESSION_IMPL *session, uint32_t lognum)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
WT_SESSION_IMPL *tmp_session;
int yield;
conn = S2C(session);
log = conn->log;
yield = 0;
WT_RET(__wt_open_internal_session(conn,
"compatibility-reconfig", true, 0, &tmp_session));
while (log->ckpt_lsn.l.file < lognum) {
/*
* Force a checkpoint to be written in the new log file and
* force the archiving of all previous log files. We do the
* checkpoint in the loop because the checkpoint LSN in the
* log record could still reflect the previous log file in
* cases such as the write LSN has not yet advanced into the
* new log file due to another group of threads still in
* progress with their slot copies or writes.
*/
WT_RET(tmp_session->iface.checkpoint(
&tmp_session->iface, "force=1"));
WT_RET(WT_SESSION_CHECK_PANIC(tmp_session));
/*
* Only sleep in the rare case that we had to come through
* this loop more than once.
*/
if (yield++) {
WT_STAT_CONN_INCR(session, log_force_ckpt_sleep);
__wt_sleep(0, WT_THOUSAND);
}
}
WT_RET(tmp_session->iface.close(&tmp_session->iface, NULL));
return (0);
}
/*
* __wt_las_create --
* Initialize the database's lookaside store.
*/
int
__wt_las_create(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
const char *drop_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_drop), "force=true", NULL };
conn = S2C(session);
/*
* Done at startup: we cannot do it on demand because we require the
* schema lock to create and drop the file, and it may not always be
* available.
*
* Open an internal session, used for the shared lookaside cursor.
*
* Sessions associated with a lookaside cursor should never be tapped
* for eviction.
*/
WT_RET(__wt_open_internal_session(
conn, "lookaside table", 1, 1, &conn->las_session));
session = conn->las_session;
F_SET(session, WT_SESSION_LOOKASIDE_CURSOR | WT_SESSION_NO_EVICTION);
/* Discard any previous incarnation of the file. */
WT_RET(__wt_session_drop(session, WT_LAS_URI, drop_cfg));
/* Re-create the file. */
WT_RET(__wt_session_create(session, WT_LAS_URI, WT_LAS_FORMAT));
/* Open the shared cursor. */
WT_WITHOUT_DHANDLE(session,
ret = __las_cursor_create(session, &conn->las_cursor));
return (ret);
}
/*
* __wt_logmgr_open --
* Start the log service threads.
*/
int
__wt_logmgr_open(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
conn = S2C(session);
/* If no log thread services are configured, we're done. */
if (!FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED))
return (0);
/*
* Start the log close thread. It is not configurable.
* If logging is enabled, this thread runs.
*/
WT_RET(__wt_open_internal_session(
conn, "log-close-server", 0, 0, &conn->log_file_session));
WT_RET(__wt_cond_alloc(conn->log_file_session,
"log close server", 0, &conn->log_file_cond));
/*
* Start the log file close thread.
*/
WT_RET(__wt_thread_create(conn->log_file_session,
&conn->log_file_tid, __log_file_server, conn->log_file_session));
conn->log_file_tid_set = 1;
/*
* Start the log write LSN thread. It is not configurable.
* If logging is enabled, this thread runs.
*/
WT_RET(__wt_open_internal_session(
conn, "log-wrlsn-server", 0, 0, &conn->log_wrlsn_session));
WT_RET(__wt_cond_alloc(conn->log_wrlsn_session,
"log write lsn server", 0, &conn->log_wrlsn_cond));
WT_RET(__wt_thread_create(conn->log_wrlsn_session,
&conn->log_wrlsn_tid, __log_wrlsn_server, conn->log_wrlsn_session));
conn->log_wrlsn_tid_set = 1;
/* If no log thread services are configured, we're done. */
if (!FLD_ISSET(conn->log_flags,
(WT_CONN_LOG_ARCHIVE | WT_CONN_LOG_PREALLOC)))
return (0);
/*
* If a log server thread exists, the user may have reconfigured
* archiving or pre-allocation. Signal the thread. Otherwise the
* user wants archiving and/or allocation and we need to start up
* the thread.
*/
if (conn->log_session != NULL) {
WT_ASSERT(session, conn->log_cond != NULL);
WT_ASSERT(session, conn->log_tid_set != 0);
WT_RET(__wt_cond_signal(session, conn->log_cond));
} else {
/* The log server gets its own session. */
WT_RET(__wt_open_internal_session(
conn, "log-server", 0, 0, &conn->log_session));
WT_RET(__wt_cond_alloc(conn->log_session,
"log server", 0, &conn->log_cond));
/*
* Start the thread.
*/
WT_RET(__wt_thread_create(conn->log_session,
&conn->log_tid, __log_server, conn->log_session));
conn->log_tid_set = 1;
}
return (0);
}
/*
* __wt_txn_recover --
* Run recovery.
*/
int
__wt_txn_recover(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR *metac;
WT_DECL_RET;
WT_RECOVERY r;
struct WT_RECOVERY_FILE *metafile;
char *config;
bool eviction_started, needs_rec, was_backup;
conn = S2C(session);
WT_CLEAR(r);
WT_INIT_LSN(&r.ckpt_lsn);
eviction_started = false;
was_backup = F_ISSET(conn, WT_CONN_WAS_BACKUP);
/* We need a real session for recovery. */
WT_RET(__wt_open_internal_session(conn, "txn-recover",
false, WT_SESSION_NO_LOGGING, &session));
r.session = session;
F_SET(conn, WT_CONN_RECOVERING);
WT_ERR(__wt_metadata_search(session, WT_METAFILE_URI, &config));
WT_ERR(__recovery_setup_file(&r, WT_METAFILE_URI, config));
WT_ERR(__wt_metadata_cursor_open(session, NULL, &metac));
metafile = &r.files[WT_METAFILE_ID];
metafile->c = metac;
/*
* If no log was found (including if logging is disabled), or if the
* last checkpoint was done with logging disabled, recovery should not
* run. Scan the metadata to figure out the largest file ID.
*/
if (!FLD_ISSET(S2C(session)->log_flags, WT_CONN_LOG_EXISTED) ||
WT_IS_MAX_LSN(&metafile->ckpt_lsn)) {
WT_ERR(__recovery_file_scan(&r));
conn->next_file_id = r.max_fileid;
goto done;
}
/*
* First, do a pass through the log to recover the metadata, and
* establish the last checkpoint LSN. Skip this when opening a hot
* backup: we already have the correct metadata in that case.
*/
if (!was_backup) {
r.metadata_only = true;
/*
* If this is a read-only connection, check if the checkpoint
* LSN in the metadata file is up to date, indicating a clean
* shutdown.
*/
if (F_ISSET(conn, WT_CONN_READONLY)) {
WT_ERR(__wt_log_needs_recovery(
session, &metafile->ckpt_lsn, &needs_rec));
if (needs_rec)
WT_ERR_MSG(session, WT_RUN_RECOVERY,
"Read-only database needs recovery");
}
if (WT_IS_INIT_LSN(&metafile->ckpt_lsn))
WT_ERR(__wt_log_scan(session,
NULL, WT_LOGSCAN_FIRST, __txn_log_recover, &r));
else {
/*
* Start at the last checkpoint LSN referenced in the
* metadata. If we see the end of a checkpoint while
* scanning, we will change the full scan to start from
* there.
*/
r.ckpt_lsn = metafile->ckpt_lsn;
ret = __wt_log_scan(session,
&metafile->ckpt_lsn, 0, __txn_log_recover, &r);
if (ret == ENOENT)
ret = 0;
WT_ERR(ret);
}
}
/* Scan the metadata to find the live files and their IDs. */
WT_ERR(__recovery_file_scan(&r));
/*
* We no longer need the metadata cursor: close it to avoid pinning any
* resources that could block eviction during recovery.
*/
r.files[0].c = NULL;
WT_ERR(metac->close(metac));
/*
* Now, recover all the files apart from the metadata.
* Pass WT_LOGSCAN_RECOVER so that old logs get truncated.
*/
r.metadata_only = false;
WT_ERR(__wt_verbose(session, WT_VERB_RECOVERY,
"Main recovery loop: starting at %" PRIu32 "/%" PRIu32,
r.ckpt_lsn.l.file, r.ckpt_lsn.l.offset));
WT_ERR(__wt_log_needs_recovery(session, &r.ckpt_lsn, &needs_rec));
/*
* Check if the database was shut down cleanly. If not
* return an error if the user does not want automatic
* recovery.
*/
if (needs_rec &&
(FLD_ISSET(conn->log_flags, WT_CONN_LOG_RECOVER_ERR) ||
F_ISSET(conn, WT_CONN_READONLY))) {
if (F_ISSET(conn, WT_CONN_READONLY))
WT_ERR_MSG(session, WT_RUN_RECOVERY,
"Read-only database needs recovery");
WT_ERR(WT_RUN_RECOVERY);
}
if (F_ISSET(conn, WT_CONN_READONLY))
goto done;
/*
* Recovery can touch more data than fits in cache, so it relies on
* regular eviction to manage paging. Start eviction threads for
* recovery without LAS cursors.
*/
WT_ERR(__wt_evict_create(session));
eviction_started = true;
/*
* Always run recovery even if it was a clean shutdown only if
* this is not a read-only connection.
* We can consider skipping it in the future.
*/
if (WT_IS_INIT_LSN(&r.ckpt_lsn))
WT_ERR(__wt_log_scan(session, NULL,
WT_LOGSCAN_FIRST | WT_LOGSCAN_RECOVER,
__txn_log_recover, &r));
else {
ret = __wt_log_scan(session, &r.ckpt_lsn,
WT_LOGSCAN_RECOVER, __txn_log_recover, &r);
if (ret == ENOENT)
ret = 0;
WT_ERR(ret);
}
conn->next_file_id = r.max_fileid;
/*
* If recovery ran successfully forcibly log a checkpoint so the next
* open is fast and keep the metadata up to date with the checkpoint
* LSN and archiving.
*/
WT_ERR(session->iface.checkpoint(&session->iface, "force=1"));
done: FLD_SET(conn->log_flags, WT_CONN_LOG_RECOVER_DONE);
err: WT_TRET(__recovery_free(&r));
__wt_free(session, config);
if (ret != 0)
__wt_err(session, ret, "Recovery failed");
/*
* Destroy the eviction threads that were started in support of
* recovery. They will be restarted once the lookaside table is
* created.
*/
if (eviction_started)
WT_TRET(__wt_evict_destroy(session));
WT_TRET(session->iface.close(&session->iface, NULL));
F_CLR(conn, WT_CONN_RECOVERING);
return (ret);
}
/*
* __wt_txn_recover --
* Run recovery.
*/
int
__wt_txn_recover(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR *metac;
WT_DECL_RET;
WT_RECOVERY r;
WT_RECOVERY_FILE *metafile;
char *config;
bool do_checkpoint, eviction_started, needs_rec, was_backup;
conn = S2C(session);
WT_CLEAR(r);
WT_INIT_LSN(&r.ckpt_lsn);
config = NULL;
do_checkpoint = true;
eviction_started = false;
was_backup = F_ISSET(conn, WT_CONN_WAS_BACKUP);
/* We need a real session for recovery. */
WT_RET(__wt_open_internal_session(conn, "txn-recover",
false, WT_SESSION_NO_LOGGING, &session));
r.session = session;
WT_MAX_LSN(&r.max_ckpt_lsn);
WT_MAX_LSN(&r.max_rec_lsn);
conn->txn_global.recovery_timestamp =
conn->txn_global.meta_ckpt_timestamp = 0;
F_SET(conn, WT_CONN_RECOVERING);
WT_ERR(__wt_metadata_search(session, WT_METAFILE_URI, &config));
WT_ERR(__recovery_setup_file(&r, WT_METAFILE_URI, config));
WT_ERR(__wt_metadata_cursor_open(session, NULL, &metac));
metafile = &r.files[WT_METAFILE_ID];
metafile->c = metac;
/*
* If no log was found (including if logging is disabled), or if the
* last checkpoint was done with logging disabled, recovery should not
* run. Scan the metadata to figure out the largest file ID.
*/
if (!FLD_ISSET(conn->log_flags, WT_CONN_LOG_EXISTED) ||
WT_IS_MAX_LSN(&metafile->ckpt_lsn)) {
/*
* Detect if we're going from logging disabled to enabled.
* We need to know this to verify LSNs and start at the correct
* log file later. If someone ran with logging, then disabled
* it and removed all the log files and then turned logging back
* on, we have to start logging in the log file number that is
* larger than any checkpoint LSN we have from the earlier time.
*/
WT_ERR(__recovery_file_scan(&r));
/*
* The array can be re-allocated in recovery_file_scan. Reset
* our pointer after scanning all the files.
*/
metafile = &r.files[WT_METAFILE_ID];
conn->next_file_id = r.max_fileid;
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED) &&
WT_IS_MAX_LSN(&metafile->ckpt_lsn) &&
!WT_IS_MAX_LSN(&r.max_ckpt_lsn))
WT_ERR(__wt_log_reset(session, r.max_ckpt_lsn.l.file));
else
do_checkpoint = false;
goto done;
}
/*
* First, do a pass through the log to recover the metadata, and
* establish the last checkpoint LSN. Skip this when opening a hot
* backup: we already have the correct metadata in that case.
*
* If we're running with salvage and we hit an error, we ignore it
* and continue. In salvage we want to recover whatever part of the
* data we can from the last checkpoint up until whatever problem we
* detect in the log file. In salvage, we ignore errors from scanning
* the log so recovery can continue. Other errors remain errors.
*/
if (!was_backup) {
r.metadata_only = true;
/*
* If this is a read-only connection, check if the checkpoint
* LSN in the metadata file is up to date, indicating a clean
* shutdown.
*/
if (F_ISSET(conn, WT_CONN_READONLY)) {
WT_ERR(__wt_log_needs_recovery(
session, &metafile->ckpt_lsn, &needs_rec));
if (needs_rec)
WT_ERR_MSG(session, WT_RUN_RECOVERY,
"Read-only database needs recovery");
}
if (WT_IS_INIT_LSN(&metafile->ckpt_lsn))
ret = __wt_log_scan(session,
NULL, WT_LOGSCAN_FIRST, __txn_log_recover, &r);
else {
/*
* Start at the last checkpoint LSN referenced in the
* metadata. If we see the end of a checkpoint while
* scanning, we will change the full scan to start from
* there.
*/
r.ckpt_lsn = metafile->ckpt_lsn;
ret = __wt_log_scan(session,
&metafile->ckpt_lsn, 0, __txn_log_recover, &r);
}
if (F_ISSET(conn, WT_CONN_SALVAGE))
ret = 0;
/*
* If log scan couldn't find a file we expected to be around,
* this indicates a corruption of some sort.
*/
if (ret == ENOENT) {
F_SET(conn, WT_CONN_DATA_CORRUPTION);
ret = WT_ERROR;
}
WT_ERR(ret);
}
/* Scan the metadata to find the live files and their IDs. */
WT_ERR(__recovery_file_scan(&r));
/*
* Clear this out. We no longer need it and it could have been
* re-allocated when scanning the files.
*/
WT_NOT_READ(metafile, NULL);
/*
* We no longer need the metadata cursor: close it to avoid pinning any
* resources that could block eviction during recovery.
*/
r.files[0].c = NULL;
WT_ERR(metac->close(metac));
/*
* Now, recover all the files apart from the metadata.
* Pass WT_LOGSCAN_RECOVER so that old logs get truncated.
*/
r.metadata_only = false;
__wt_verbose(session, WT_VERB_RECOVERY | WT_VERB_RECOVERY_PROGRESS,
"Main recovery loop: starting at %" PRIu32 "/%" PRIu32
" to %" PRIu32 "/%" PRIu32, r.ckpt_lsn.l.file, r.ckpt_lsn.l.offset,
r.max_rec_lsn.l.file, r.max_rec_lsn.l.offset);
WT_ERR(__wt_log_needs_recovery(session, &r.ckpt_lsn, &needs_rec));
/*
* Check if the database was shut down cleanly. If not
* return an error if the user does not want automatic
* recovery.
*/
if (needs_rec &&
(FLD_ISSET(conn->log_flags, WT_CONN_LOG_RECOVER_ERR) ||
F_ISSET(conn, WT_CONN_READONLY))) {
if (F_ISSET(conn, WT_CONN_READONLY))
WT_ERR_MSG(session, WT_RUN_RECOVERY,
"Read-only database needs recovery");
WT_ERR_MSG(session, WT_RUN_RECOVERY, "Database needs recovery");
}
if (F_ISSET(conn, WT_CONN_READONLY)) {
do_checkpoint = false;
goto done;
}
/*
* Recovery can touch more data than fits in cache, so it relies on
* regular eviction to manage paging. Start eviction threads for
* recovery without LAS cursors.
*/
WT_ERR(__wt_evict_create(session));
eviction_started = true;
/*
* Always run recovery even if it was a clean shutdown only if
* this is not a read-only connection.
* We can consider skipping it in the future.
*/
if (needs_rec)
FLD_SET(conn->log_flags, WT_CONN_LOG_RECOVER_DIRTY);
if (WT_IS_INIT_LSN(&r.ckpt_lsn))
ret = __wt_log_scan(session, NULL,
WT_LOGSCAN_FIRST | WT_LOGSCAN_RECOVER,
__txn_log_recover, &r);
else
ret = __wt_log_scan(session, &r.ckpt_lsn,
WT_LOGSCAN_RECOVER, __txn_log_recover, &r);
if (F_ISSET(conn, WT_CONN_SALVAGE))
ret = 0;
WT_ERR(ret);
conn->next_file_id = r.max_fileid;
done: WT_ERR(__recovery_set_checkpoint_timestamp(&r));
if (do_checkpoint)
/*
* Forcibly log a checkpoint so the next open is fast and keep
* the metadata up to date with the checkpoint LSN and
* archiving.
*/
WT_ERR(session->iface.checkpoint(&session->iface, "force=1"));
/*
* If we're downgrading and have newer log files, force an archive,
* no matter what the archive setting is.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_FORCE_DOWNGRADE))
WT_ERR(__wt_log_truncate_files(session, NULL, true));
FLD_SET(conn->log_flags, WT_CONN_LOG_RECOVER_DONE);
err: WT_TRET(__recovery_free(&r));
__wt_free(session, config);
FLD_CLR(conn->log_flags, WT_CONN_LOG_RECOVER_DIRTY);
if (ret != 0) {
FLD_SET(conn->log_flags, WT_CONN_LOG_RECOVER_FAILED);
__wt_err(session, ret, "Recovery failed");
}
/*
* Destroy the eviction threads that were started in support of
* recovery. They will be restarted once the lookaside table is
* created.
*/
if (eviction_started)
WT_TRET(__wt_evict_destroy(session));
WT_TRET(session->iface.close(&session->iface, NULL));
F_CLR(conn, WT_CONN_RECOVERING);
return (ret);
}
/*
* __thread_group_resize --
* Resize an array of utility threads already holding the lock.
*/
static int
__thread_group_resize(
WT_SESSION_IMPL *session, WT_THREAD_GROUP *group,
uint32_t new_min, uint32_t new_max, uint32_t flags)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_THREAD *thread;
size_t alloc;
uint32_t i, session_flags;
conn = S2C(session);
session_flags = 0;
WT_ASSERT(session,
group->current_threads <= group->alloc &&
__wt_rwlock_islocked(session, group->lock));
if (new_min == group->min && new_max == group->max)
return (0);
/*
* Coll shrink to reduce the number of thread structures and running
* threads if required by the change in group size.
*/
WT_RET(__thread_group_shrink(session, group, new_max));
/*
* Only reallocate the thread array if it is the largest ever, since
* our realloc doesn't support shrinking the allocated size.
*/
if (group->alloc < new_max) {
alloc = group->alloc * sizeof(*group->threads);
WT_RET(__wt_realloc(session, &alloc,
new_max * sizeof(*group->threads), &group->threads));
group->alloc = new_max;
}
/*
* Initialize the structures based on the previous group size, not
* the previous allocated size.
*/
for (i = group->max; i < new_max; i++) {
WT_ERR(__wt_calloc_one(session, &thread));
/*
* Threads get their own session and lookaside table cursor
* if the lookaside table is open. Note that threads are
* started during recovery, before the lookaside table is
* created.
*/
if (LF_ISSET(WT_THREAD_CAN_WAIT))
session_flags = WT_SESSION_CAN_WAIT;
if (F_ISSET(conn, WT_CONN_LAS_OPEN))
FLD_SET(session_flags, WT_SESSION_LOOKASIDE_CURSOR);
WT_ERR(__wt_open_internal_session(conn, group->name,
false, session_flags, &thread->session));
if (LF_ISSET(WT_THREAD_PANIC_FAIL))
F_SET(thread, WT_THREAD_PANIC_FAIL);
thread->id = i;
thread->run_func = group->run_func;
WT_ASSERT(session, group->threads[i] == NULL);
group->threads[i] = thread;
}
if (group->current_threads < new_min)
WT_ERR(__thread_group_grow(session, group, new_min));
err: /*
* Update the thread group information even on failure to improve our
* chances of cleaning up properly.
*/
group->max = new_max;
group->min = new_min;
/*
* An error resizing a thread array is fatal, it should only happen
* in an out of memory situation.
*/
if (ret != 0) {
WT_TRET(__wt_thread_group_destroy(session, group));
WT_PANIC_RET(session, ret, "Error while resizing thread group");
}
return (ret);
}
/*
* __wt_logmgr_open --
* Start the log service threads.
*/
int
__wt_logmgr_open(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
uint32_t session_flags;
conn = S2C(session);
/* If no log thread services are configured, we're done. */
if (!FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED))
return (0);
F_SET(conn, WT_CONN_SERVER_LOG);
/*
* Start the log close thread. It is not configurable.
* If logging is enabled, this thread runs.
*/
session_flags = WT_SESSION_NO_DATA_HANDLES;
WT_RET(__wt_open_internal_session(conn,
"log-close-server", false, session_flags, &conn->log_file_session));
WT_RET(__wt_cond_alloc(
conn->log_file_session, "log close server", &conn->log_file_cond));
/*
* Start the log file close thread.
*/
WT_RET(__wt_thread_create(conn->log_file_session,
&conn->log_file_tid, __log_file_server, conn->log_file_session));
conn->log_file_tid_set = true;
/*
* Start the log write LSN thread. It is not configurable.
* If logging is enabled, this thread runs.
*/
WT_RET(__wt_open_internal_session(conn, "log-wrlsn-server",
false, session_flags, &conn->log_wrlsn_session));
WT_RET(__wt_cond_auto_alloc(conn->log_wrlsn_session,
"log write lsn server", 10000, WT_MILLION, &conn->log_wrlsn_cond));
WT_RET(__wt_thread_create(conn->log_wrlsn_session,
&conn->log_wrlsn_tid, __log_wrlsn_server, conn->log_wrlsn_session));
conn->log_wrlsn_tid_set = true;
/*
* If a log server thread exists, the user may have reconfigured
* archiving or pre-allocation. Signal the thread. Otherwise the
* user wants archiving and/or allocation and we need to start up
* the thread.
*/
if (conn->log_session != NULL) {
WT_ASSERT(session, conn->log_cond != NULL);
WT_ASSERT(session, conn->log_tid_set == true);
__wt_cond_signal(session, conn->log_cond);
} else {
/* The log server gets its own session. */
WT_RET(__wt_open_internal_session(conn,
"log-server", false, session_flags, &conn->log_session));
WT_RET(__wt_cond_auto_alloc(conn->log_session,
"log server", 50000, WT_MILLION, &conn->log_cond));
/*
* Start the thread.
*/
WT_RET(__wt_thread_create(conn->log_session,
&conn->log_tid, __log_server, conn->log_session));
conn->log_tid_set = true;
}
return (0);
}
/*
* __wt_cache_create --
* Create the underlying cache.
*/
int
__wt_cache_create(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_CACHE *cache;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
int i;
conn = S2C(session);
WT_ASSERT(session, conn->cache == NULL);
WT_RET(__wt_calloc_one(session, &conn->cache));
cache = conn->cache;
/* Use a common routine for run-time configuration options. */
WT_RET(__wt_cache_config(session, false, cfg));
/*
* The lowest possible page read-generation has a special meaning, it
* marks a page for forcible eviction; don't let it happen by accident.
*/
cache->read_gen = WT_READGEN_START_VALUE;
/*
* The target size must be lower than the trigger size or we will never
* get any work done.
*/
if (cache->eviction_target >= cache->eviction_trigger)
WT_ERR_MSG(session, EINVAL,
"eviction target must be lower than the eviction trigger");
WT_ERR(__wt_cond_auto_alloc(session, "cache eviction server",
false, 10000, WT_MILLION, &cache->evict_cond));
WT_ERR(__wt_spin_init(session, &cache->evict_pass_lock, "evict pass"));
WT_ERR(__wt_spin_init(session,
&cache->evict_queue_lock, "cache eviction queue"));
WT_ERR(__wt_spin_init(session, &cache->evict_walk_lock, "cache walk"));
if ((ret = __wt_open_internal_session(conn, "evict pass",
false, WT_SESSION_NO_DATA_HANDLES, &cache->walk_session)) != 0)
WT_ERR_MSG(NULL, ret,
"Failed to create session for eviction walks");
/* Allocate the LRU eviction queue. */
cache->evict_slots = WT_EVICT_WALK_BASE + WT_EVICT_WALK_INCR;
for (i = 0; i < WT_EVICT_QUEUE_MAX; ++i) {
WT_ERR(__wt_calloc_def(session,
cache->evict_slots, &cache->evict_queues[i].evict_queue));
WT_ERR(__wt_spin_init(session,
&cache->evict_queues[i].evict_lock, "cache eviction"));
}
/* Ensure there are always non-NULL queues. */
cache->evict_current_queue = cache->evict_fill_queue =
&cache->evict_queues[0];
cache->evict_other_queue = &cache->evict_queues[1];
cache->evict_urgent_queue = &cache->evict_queues[WT_EVICT_URGENT_QUEUE];
/*
* We get/set some values in the cache statistics (rather than have
* two copies), configure them.
*/
__wt_cache_stats_update(session);
return (0);
err: WT_RET(__wt_cache_destroy(session));
return (ret);
}