/*
* __log_truncate --
* Truncate the log to the given LSN. If this_log is set, it will only
* truncate the log file indicated in the given LSN. If not set,
* it will truncate between the given LSN and the trunc_lsn. That is,
* since we pre-allocate log files, it will free that space and allow the
* log to be traversed. We use the trunc_lsn because logging has already
* opened the new/next log file before recovery ran. This function assumes
* we are in recovery or other dedicated time and not during live running.
*/
static int
__log_truncate(WT_SESSION_IMPL *session, WT_LSN *lsn, uint32_t this_log)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *log_fh, *tmp_fh;
WT_LOG *log;
uint32_t lognum;
u_int i, logcount;
char **logfiles;
conn = S2C(session);
log = conn->log;
log_fh = NULL;
logcount = 0;
logfiles = NULL;
/*
* Truncate the log file to the given LSN.
*/
WT_ERR(__log_openfile(session, 0, &log_fh, lsn->file));
WT_ERR(__wt_ftruncate(session, log_fh, lsn->offset));
tmp_fh = log_fh;
log_fh = NULL;
WT_ERR(__wt_close(session, tmp_fh));
/*
* If we just want to truncate the current log, return and skip
* looking for intervening logs.
*/
if (this_log)
goto err;
WT_ERR(__wt_log_get_files(session, &logfiles, &logcount));
for (i = 0; i < logcount; i++) {
WT_ERR(__wt_log_extract_lognum(session, logfiles[i], &lognum));
if (lognum > lsn->file && lognum < log->trunc_lsn.file) {
WT_ERR(__log_openfile(session, 0, &log_fh, lognum));
/*
* If there are intervening files pre-allocated,
* truncate them to the end of the log file header.
*/
WT_ERR(__wt_ftruncate(session,
log_fh, LOG_FIRST_RECORD));
tmp_fh = log_fh;
log_fh = NULL;
WT_ERR(__wt_close(session, tmp_fh));
}
}
err: if (log_fh != NULL)
WT_TRET(__wt_close(session, log_fh));
if (logfiles != NULL)
__wt_log_files_free(session, logfiles, logcount);
return (ret);
}
/*
* __wt_connection_destroy --
* Destroy the connection's underlying WT_CONNECTION_IMPL structure.
*/
int
__wt_connection_destroy(WT_CONNECTION_IMPL *conn)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
/* Check there's something to destroy. */
if (conn == NULL)
return (0);
session = conn->default_session;
/*
* Close remaining open files (before discarding the mutex, the
* underlying file-close code uses the mutex to guard lists of
* open files.
*/
if (conn->lock_fh != NULL)
WT_TRET(__wt_close(session, conn->lock_fh));
if (conn->log_fh != NULL)
WT_TRET(__wt_close(session, conn->log_fh));
/* Remove from the list of connections. */
__wt_spin_lock(session, &__wt_process.spinlock);
TAILQ_REMOVE(&__wt_process.connqh, conn, q);
__wt_spin_unlock(session, &__wt_process.spinlock);
/* Configuration */
__wt_conn_config_discard(session); /* configuration */
__wt_conn_foc_discard(session); /* free-on-close */
__wt_spin_destroy(session, &conn->api_lock);
__wt_spin_destroy(session, &conn->block_lock);
__wt_spin_destroy(session, &conn->checkpoint_lock);
__wt_spin_destroy(session, &conn->fh_lock);
__wt_spin_destroy(session, &conn->hot_backup_lock);
__wt_spin_destroy(session, &conn->schema_lock);
__wt_spin_destroy(session, &conn->serial_lock);
/* Free allocated memory. */
__wt_free(session, conn->home);
__wt_free(session, conn->error_prefix);
__wt_free(session, conn->sessions);
__wt_free(NULL, conn);
return (ret);
}
/*
* __wt_close_connection_close --
* Close any open file handles at connection close.
*/
int
__wt_close_connection_close(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
WT_FH *fh;
WT_CONNECTION_IMPL *conn;
conn = S2C(session);
while ((fh = TAILQ_FIRST(&conn->fhqh)) != NULL) {
/*
* In-memory configurations will have open files, but the ref
* counts should be zero.
*/
if (!F_ISSET(conn, WT_CONN_IN_MEMORY) || fh->ref != 0) {
ret = EBUSY;
__wt_errx(session,
"Connection has open file handles: %s", fh->name);
}
fh->ref = 1;
F_CLR(fh, WT_FH_IN_MEMORY);
WT_TRET(__wt_close(session, &fh));
}
return (ret);
}
/*
* __wt_connection_destroy --
* Destroy the connection's underlying WT_CONNECTION_IMPL structure.
*/
int
__wt_connection_destroy(WT_CONNECTION_IMPL *conn)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
/* Check there's something to destroy. */
if (conn == NULL)
return (0);
session = conn->default_session;
/*
* Close remaining open files (before discarding the mutex, the
* underlying file-close code uses the mutex to guard lists of
* open files.
*/
WT_TRET(__wt_close(session, &conn->lock_fh));
/* Remove from the list of connections. */
__wt_spin_lock(session, &__wt_process.spinlock);
TAILQ_REMOVE(&__wt_process.connqh, conn, q);
__wt_spin_unlock(session, &__wt_process.spinlock);
/* Configuration */
__wt_conn_config_discard(session); /* configuration */
__wt_conn_foc_discard(session); /* free-on-close */
__wt_spin_destroy(session, &conn->api_lock);
__wt_spin_destroy(session, &conn->block_lock);
__wt_spin_destroy(session, &conn->checkpoint_lock);
__wt_spin_destroy(session, &conn->dhandle_lock);
__wt_spin_destroy(session, &conn->encryptor_lock);
__wt_spin_destroy(session, &conn->fh_lock);
WT_TRET(__wt_rwlock_destroy(session, &conn->hot_backup_lock));
__wt_spin_destroy(session, &conn->las_lock);
__wt_spin_destroy(session, &conn->metadata_lock);
__wt_spin_destroy(session, &conn->reconfig_lock);
__wt_spin_destroy(session, &conn->schema_lock);
__wt_spin_destroy(session, &conn->table_lock);
__wt_spin_destroy(session, &conn->turtle_lock);
for (i = 0; i < WT_PAGE_LOCKS; ++i)
__wt_spin_destroy(session, &conn->page_lock[i]);
__wt_free(session, conn->page_lock);
/* Free allocated memory. */
__wt_free(session, conn->cfg);
__wt_free(session, conn->home);
__wt_free(session, conn->error_prefix);
__wt_free(session, conn->sessions);
__wt_free(NULL, conn);
return (ret);
}
/*
* __wt_log_close --
* Close the log file.
*/
int
__wt_log_close(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
conn = S2C(session);
log = conn->log;
if (log->log_close_fh != NULL && log->log_close_fh != log->log_fh) {
WT_RET(__wt_verbose(session, WT_VERB_LOG,
"closing old log %s", log->log_close_fh->name));
WT_RET(__wt_close(session, log->log_close_fh));
}
if (log->log_fh != NULL) {
WT_RET(__wt_verbose(session, WT_VERB_LOG,
"closing log %s", log->log_fh->name));
WT_RET(__wt_close(session, log->log_fh));
log->log_fh = NULL;
}
return (0);
}
/*
* __fstream_close --
* Close a stream handle.
*/
static int
__fstream_close(WT_SESSION_IMPL *session, WT_FSTREAM *fstr)
{
WT_DECL_RET;
if (!F_ISSET(fstr, WT_STREAM_READ))
WT_TRET(fstr->fstr_flush(session, fstr));
WT_TRET(__wt_close(session, &fstr->fh));
__wt_buf_free(session, &fstr->buf);
__wt_free(session, fstr);
return (ret);
}
/*
* __optrack_open_file --
* Open the per-session operation-tracking file.
*/
static int
__optrack_open_file(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_OPTRACK_HEADER optrack_header = { WT_OPTRACK_VERSION, 0,
(uint32_t)WT_TSC_DEFAULT_RATIO * WT_THOUSAND };
conn = S2C(session);
if (!F_ISSET(conn, WT_CONN_OPTRACK))
WT_RET_MSG(session, WT_ERROR, "WT_CONN_OPTRACK not set");
WT_RET(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_filename_construct(session, conn->optrack_path,
"optrack", conn->optrack_pid, session->id, buf));
WT_ERR(__wt_open(session,
(const char *)buf->data, WT_FS_OPEN_FILE_TYPE_REGULAR,
WT_FS_OPEN_CREATE, &session->optrack_fh));
/* Indicate whether this is an internal session */
if (F_ISSET(session, WT_SESSION_INTERNAL))
optrack_header.optrack_session_internal = 1;
/*
* Record the clock ticks to nanoseconds ratio. Multiply it by one
* thousand, so we can use a fixed width integer.
*/
optrack_header.optrack_tsc_nsec_ratio =
(uint32_t)(__wt_process.tsc_nsec_ratio * WT_THOUSAND);
/* Write the header into the operation-tracking file. */
WT_ERR(session->optrack_fh->handle->fh_write(
session->optrack_fh->handle, (WT_SESSION *)session,
0, sizeof(WT_OPTRACK_HEADER), &optrack_header));
session->optrack_offset = sizeof(WT_OPTRACK_HEADER);
if (0) {
err: WT_TRET(__wt_close(session, &session->optrack_fh));
}
__wt_scr_free(session, &buf);
return (ret);
}
/*
* __wt_fopen --
* Open a stream handle.
*/
int
__wt_fopen(WT_SESSION_IMPL *session,
const char *name, uint32_t open_flags, uint32_t flags, WT_FSTREAM **fstrp)
{
WT_DECL_RET;
WT_FH *fh;
WT_FSTREAM *fstr;
*fstrp = NULL;
fstr = NULL;
WT_RET(__wt_open(
session, name, WT_OPEN_FILE_TYPE_REGULAR, open_flags, &fh));
WT_ERR(__wt_calloc_one(session, &fstr));
fstr->fh = fh;
fstr->name = fh->name;
fstr->flags = flags;
fstr->close = __fstream_close;
WT_ERR(__wt_filesize(session, fh, &fstr->size));
if (LF_ISSET(WT_STREAM_APPEND))
fstr->off = fstr->size;
if (LF_ISSET(WT_STREAM_APPEND | WT_STREAM_WRITE)) {
fstr->fstr_flush = __fstream_flush;
fstr->fstr_getline = __fstream_getline_notsup;
fstr->fstr_printf = __fstream_printf;
} else {
WT_ASSERT(session, LF_ISSET(WT_STREAM_READ));
fstr->fstr_flush = __fstream_flush_notsup;
fstr->fstr_getline = __fstream_getline;
fstr->fstr_printf = __fstream_printf_notsup;
}
*fstrp = fstr;
return (0);
err: WT_TRET(__wt_close(session, &fh));
__wt_free(session, fstr);
return (ret);
}
/*
* __wt_block_manager_truncate --
* Truncate a file.
*/
int
__wt_block_manager_truncate(
WT_SESSION_IMPL *session, const char *filename, uint32_t allocsize)
{
WT_DECL_RET;
WT_FH *fh;
/* Open the underlying file handle. */
WT_RET(__wt_open(session, filename, 0, 0, WT_FILE_TYPE_DATA, &fh));
/* Truncate the file. */
WT_ERR(__wt_ftruncate(session, fh, (off_t)0));
/* Write out the file's meta-data. */
ret = __wt_desc_init(session, fh, allocsize);
/* Close the file handle. */
err: WT_TRET(__wt_close(session, fh));
return (ret);
}
/*
* __block_destroy --
* Destroy a block handle.
*/
static int
__block_destroy(WT_SESSION_IMPL *session, WT_BLOCK *block)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
conn = S2C(session);
TAILQ_REMOVE(&conn->blockqh, block, q);
if (block->name != NULL)
__wt_free(session, block->name);
if (block->fh != NULL)
WT_TRET(__wt_close(session, block->fh));
__wt_spin_destroy(session, &block->live_lock);
__wt_overwrite_and_free(session, block);
return (ret);
}
/*
* __wt_sync_and_rename_fh --
* Sync and close a file, and swap it into place.
*/
int
__wt_sync_and_rename_fh(
WT_SESSION_IMPL *session, WT_FH **fhp, const char *from, const char *to)
{
WT_DECL_RET;
WT_FH *fh;
fh = *fhp;
*fhp = NULL;
/* Flush to disk and close the handle. */
ret = __wt_fsync(session, fh);
WT_TRET(__wt_close(session, &fh));
WT_RET(ret);
/* Rename the source file to the target. */
WT_RET(__wt_rename(session, from, to));
/* Flush the backing directory to guarantee the rename. */
return (__wt_directory_sync(session, NULL));
}
/*
* __wt_close_connection_close --
* Close any open file handles at connection close.
*/
int
__wt_close_connection_close(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
WT_FH *fh;
WT_CONNECTION_IMPL *conn;
conn = S2C(session);
while ((fh = TAILQ_FIRST(&conn->fhqh)) != NULL) {
if (fh->ref != 0) {
ret = EBUSY;
__wt_errx(session,
"Connection has open file handles: %s", fh->name);
}
fh->ref = 1;
WT_TRET(__wt_close(session, &fh));
}
return (ret);
}
/*
* __wt_turtle_update --
* Update the turtle file.
*/
int
__wt_turtle_update(
WT_SESSION_IMPL *session, const char *key, const char *value)
{
WT_FH *fh;
WT_DECL_ITEM(buf);
WT_DECL_RET;
int vmajor, vminor, vpatch;
const char *version;
fh = NULL;
/*
* Create the turtle setup file: we currently re-write it from scratch
* every time.
*/
WT_RET(__wt_open(
session, WT_METADATA_TURTLE_SET, 1, 1, WT_FILE_TYPE_TURTLE, &fh));
version = wiredtiger_version(&vmajor, &vminor, &vpatch);
WT_ERR(__wt_scr_alloc(session, 2 * 1024, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s\n%s\n%s\n" "major=%d,minor=%d,patch=%d\n%s\n%s\n",
WT_METADATA_VERSION_STR, version,
WT_METADATA_VERSION, vmajor, vminor, vpatch,
key, value));
WT_ERR(__wt_write(session, fh, 0, buf->size, buf->data));
/* Flush the handle and rename the file into place. */
ret = __wt_sync_and_rename_fh(
session, &fh, WT_METADATA_TURTLE_SET, WT_METADATA_TURTLE);
/* Close any file handle left open, remove any temporary file. */
err: WT_TRET(__wt_close(session, &fh));
WT_TRET(__wt_remove_if_exists(session, WT_METADATA_TURTLE_SET));
__wt_scr_free(session, &buf);
return (ret);
}
/*
* __wt_conn_optrack_teardown --
* Clean up connection-wide resources used for operation logging.
*/
int
__wt_conn_optrack_teardown(WT_SESSION_IMPL *session, bool reconfig)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
conn = S2C(session);
if (!reconfig)
/* Looks like we are shutting down */
__wt_free(session, conn->optrack_path);
if (!F_ISSET(conn, WT_CONN_OPTRACK))
return (0);
__wt_spin_destroy(session, &conn->optrack_map_spinlock);
WT_TRET(__wt_close(session, &conn->optrack_map_fh));
__wt_free(session, conn->dummy_session.optrack_buf);
return (ret);
}
/*
* __wt_block_manager_create --
* Create a file.
*/
int
__wt_block_manager_create(
WT_SESSION_IMPL *session, const char *filename, uint32_t allocsize)
{
WT_DECL_RET;
WT_FH *fh;
/* Create the underlying file and open a handle. */
WT_RET(__wt_open(session, filename, 1, 1, WT_FILE_TYPE_DATA, &fh));
/* Write out the file's meta-data. */
ret = __wt_desc_init(session, fh, allocsize);
/* Close the file handle. */
WT_TRET(__wt_close(session, fh));
/* Undo any create on error. */
if (ret != 0)
WT_TRET(__wt_remove(session, filename));
return (ret);
}
/*
* __log_file_server --
* The log file server thread. This worker thread manages
* log file operations such as closing and syncing.
*/
static WT_THREAD_RET
__log_file_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *close_fh;
WT_LOG *log;
WT_LSN close_end_lsn, close_lsn, min_lsn;
WT_SESSION_IMPL *session;
int locked;
session = arg;
conn = S2C(session);
log = conn->log;
locked = 0;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* If there is a log file to close, make sure any outstanding
* write operations have completed, then fsync and close it.
*/
if ((close_fh = log->log_close_fh) != NULL &&
(ret = __wt_log_extract_lognum(session, close_fh->name,
&close_lsn.file)) == 0 &&
close_lsn.file < log->write_lsn.file) {
/*
* We've copied the file handle, clear out the one in
* log structure to allow it to be set again.
*/
log->log_close_fh = NULL;
/*
* Set the close_end_lsn to the LSN immediately after
* ours. That is, the beginning of the next log file.
* We need to know the LSN file number of our own close
* in case earlier calls are still in progress and the
* next one to move the sync_lsn into the next file for
* later syncs.
*/
close_lsn.offset = 0;
close_end_lsn = close_lsn;
close_end_lsn.file++;
WT_ERR(__wt_fsync(session, close_fh));
__wt_spin_lock(session, &log->log_sync_lock);
locked = 1;
WT_ERR(__wt_close(session, &close_fh));
WT_ASSERT(session,
WT_LOG_CMP(&close_end_lsn, &log->sync_lsn) >= 0);
log->sync_lsn = close_end_lsn;
WT_ERR(__wt_cond_signal(session, log->log_sync_cond));
locked = 0;
__wt_spin_unlock(session, &log->log_sync_lock);
}
/*
* If a later thread asked for a background sync, do it now.
*/
if (WT_LOG_CMP(&log->bg_sync_lsn, &log->sync_lsn) > 0) {
/*
* Save the latest write LSN which is the minimum
* we will have written to disk.
*/
min_lsn = log->write_lsn;
/*
* The sync LSN we asked for better be smaller than
* the current written LSN.
*/
WT_ASSERT(session,
WT_LOG_CMP(&log->bg_sync_lsn, &min_lsn) <= 0);
WT_ERR(__wt_fsync(session, log->log_fh));
__wt_spin_lock(session, &log->log_sync_lock);
locked = 1;
/*
* The sync LSN could have advanced while we were
* writing to disk.
*/
if (WT_LOG_CMP(&log->sync_lsn, &min_lsn) <= 0) {
log->sync_lsn = min_lsn;
WT_ERR(__wt_cond_signal(
session, log->log_sync_cond));
}
locked = 0;
__wt_spin_unlock(session, &log->log_sync_lock);
}
/* Wait until the next event. */
WT_ERR(__wt_cond_wait(
session, conn->log_file_cond, WT_MILLION));
}
if (0) {
err: __wt_err(session, ret, "log close server error");
}
if (locked)
__wt_spin_unlock(session, &log->log_sync_lock);
return (WT_THREAD_RET_VALUE);
}
/*
* __wt_connection_close --
* Close a connection handle.
*/
int
__wt_connection_close(WT_CONNECTION_IMPL *conn)
{
WT_CONNECTION *wt_conn;
WT_DECL_RET;
WT_DLH *dlh;
WT_SESSION_IMPL *s, *session;
WT_TXN_GLOBAL *txn_global;
u_int i;
wt_conn = &conn->iface;
txn_global = &conn->txn_global;
session = conn->default_session;
/*
* We're shutting down. Make sure everything gets freed.
*
* It's possible that the eviction server is in the middle of a long
* operation, with a transaction ID pinned. In that case, we will loop
* here until the transaction ID is released, when the oldest
* transaction ID will catch up with the current ID.
*/
for (;;) {
WT_TRET(__wt_txn_update_oldest(session,
WT_TXN_OLDEST_STRICT | WT_TXN_OLDEST_WAIT));
if (txn_global->oldest_id == txn_global->current)
break;
__wt_yield();
}
/* Clear any pending async ops. */
WT_TRET(__wt_async_flush(session));
/*
* Shut down server threads other than the eviction server, which is
* needed later to close btree handles. Some of these threads access
* btree handles, so take care in ordering shutdown to make sure they
* exit before files are closed.
*/
F_CLR(conn, WT_CONN_SERVER_RUN);
WT_TRET(__wt_async_destroy(session));
WT_TRET(__wt_lsm_manager_destroy(session));
WT_TRET(__wt_sweep_destroy(session));
F_SET(conn, WT_CONN_CLOSING);
WT_TRET(__wt_checkpoint_server_destroy(session));
WT_TRET(__wt_statlog_destroy(session, true));
WT_TRET(__wt_evict_destroy(session));
/* Shut down the lookaside table, after all eviction is complete. */
WT_TRET(__wt_las_destroy(session));
/* Close open data handles. */
WT_TRET(__wt_conn_dhandle_discard(session));
/* Shut down metadata tracking, required before creating tables. */
WT_TRET(__wt_meta_track_destroy(session));
/*
* Now that all data handles are closed, tell logging that a checkpoint
* has completed then shut down the log manager (only after closing
* data handles). The call to destroy the log manager is outside the
* conditional because we allocate the log path so that printlog can
* run without running logging or recovery.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED) &&
FLD_ISSET(conn->log_flags, WT_CONN_LOG_RECOVER_DONE))
WT_TRET(__wt_txn_checkpoint_log(
session, true, WT_TXN_LOG_CKPT_STOP, NULL));
F_CLR(conn, WT_CONN_LOG_SERVER_RUN);
WT_TRET(__wt_logmgr_destroy(session));
/* Free memory for collators, compressors, data sources. */
WT_TRET(__wt_conn_remove_collator(session));
WT_TRET(__wt_conn_remove_compressor(session));
WT_TRET(__wt_conn_remove_data_source(session));
WT_TRET(__wt_conn_remove_encryptor(session));
WT_TRET(__wt_conn_remove_extractor(session));
/* Disconnect from shared cache - must be before cache destroy. */
WT_TRET(__wt_conn_cache_pool_destroy(session));
/* Discard the cache. */
WT_TRET(__wt_cache_destroy(session));
/* Discard transaction state. */
WT_TRET(__wt_txn_global_destroy(session));
/* Close extensions, first calling any unload entry point. */
while ((dlh = TAILQ_FIRST(&conn->dlhqh)) != NULL) {
TAILQ_REMOVE(&conn->dlhqh, dlh, q);
if (dlh->terminate != NULL)
WT_TRET(dlh->terminate(wt_conn));
WT_TRET(__wt_dlclose(session, dlh));
}
/* Close the lock file, opening up the database to other connections. */
if (conn->lock_fh != NULL)
WT_TRET(__wt_close(session, &conn->lock_fh));
/* Close any file handles left open. */
WT_TRET(__wt_close_connection_close(session));
/*
* Close the internal (default) session, and switch back to the dummy
* session in case of any error messages from the remaining operations
* while destroying the connection handle.
*/
if (session != &conn->dummy_session) {
WT_TRET(session->iface.close(&session->iface, NULL));
session = conn->default_session = &conn->dummy_session;
}
/*
* The session's split stash isn't discarded during normal session close
* because it may persist past the life of the session. Discard it now.
*/
if ((s = conn->sessions) != NULL)
for (i = 0; i < conn->session_size; ++s, ++i)
__wt_split_stash_discard_all(session, s);
/*
* The session's hazard pointer memory isn't discarded during normal
* session close because access to it isn't serialized. Discard it
* now.
*/
if ((s = conn->sessions) != NULL)
for (i = 0; i < conn->session_size; ++s, ++i) {
/*
* If hash arrays were allocated, free them now.
*/
__wt_free(session, s->dhhash);
__wt_free(session, s->tablehash);
__wt_free(session, s->hazard);
}
/* Destroy the handle. */
WT_TRET(__wt_connection_destroy(conn));
return (ret);
}
/*
* __wt_connection_close --
* Close a connection handle.
*/
int
__wt_connection_close(WT_CONNECTION_IMPL *conn)
{
WT_CONNECTION *wt_conn;
WT_DECL_RET;
WT_DLH *dlh;
WT_SESSION_IMPL *s, *session;
u_int i;
wt_conn = &conn->iface;
session = conn->default_session;
/* Shut down transactions (wait for in-flight operations to complete. */
WT_TRET(__wt_txn_global_shutdown(session));
/* Shut down the subsystems, ensuring workers see the state change. */
F_SET(conn, WT_CONN_CLOSING);
WT_FULL_BARRIER();
/*
* Clear any pending async operations and shut down the async worker
* threads and system before closing LSM.
*/
WT_TRET(__wt_async_flush(session));
WT_TRET(__wt_async_destroy(session));
/*
* Shut down server threads other than the eviction server, which is
* needed later to close btree handles. Some of these threads access
* btree handles, so take care in ordering shutdown to make sure they
* exit before files are closed.
*/
WT_TRET(__wt_lsm_manager_destroy(session));
/*
* Once the async and LSM threads exit, we shouldn't be opening any
* more files.
*/
F_SET(conn, WT_CONN_CLOSING_NO_MORE_OPENS);
WT_FULL_BARRIER();
WT_TRET(__wt_checkpoint_server_destroy(session));
WT_TRET(__wt_statlog_destroy(session, true));
WT_TRET(__wt_sweep_destroy(session));
/* The eviction server is shut down last. */
WT_TRET(__wt_evict_destroy(session));
/* Shut down the lookaside table, after all eviction is complete. */
WT_TRET(__wt_las_destroy(session));
/* Close open data handles. */
WT_TRET(__wt_conn_dhandle_discard(session));
/* Shut down metadata tracking. */
WT_TRET(__wt_meta_track_destroy(session));
/*
* Now that all data handles are closed, tell logging that a checkpoint
* has completed then shut down the log manager (only after closing
* data handles). The call to destroy the log manager is outside the
* conditional because we allocate the log path so that printlog can
* run without running logging or recovery.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED) &&
FLD_ISSET(conn->log_flags, WT_CONN_LOG_RECOVER_DONE))
WT_TRET(__wt_txn_checkpoint_log(
session, true, WT_TXN_LOG_CKPT_STOP, NULL));
WT_TRET(__wt_logmgr_destroy(session));
/* Free memory for collators, compressors, data sources. */
WT_TRET(__wt_conn_remove_collator(session));
WT_TRET(__wt_conn_remove_compressor(session));
WT_TRET(__wt_conn_remove_data_source(session));
WT_TRET(__wt_conn_remove_encryptor(session));
WT_TRET(__wt_conn_remove_extractor(session));
/* Disconnect from shared cache - must be before cache destroy. */
WT_TRET(__wt_conn_cache_pool_destroy(session));
/* Discard the cache. */
WT_TRET(__wt_cache_destroy(session));
/* Discard transaction state. */
__wt_txn_global_destroy(session);
/* Close the lock file, opening up the database to other connections. */
if (conn->lock_fh != NULL)
WT_TRET(__wt_close(session, &conn->lock_fh));
/* Close any file handles left open. */
WT_TRET(__wt_close_connection_close(session));
/*
* Close the internal (default) session, and switch back to the dummy
* session in case of any error messages from the remaining operations
* while destroying the connection handle.
*/
if (session != &conn->dummy_session) {
WT_TRET(session->iface.close(&session->iface, NULL));
session = conn->default_session = &conn->dummy_session;
}
/*
* The session split stash, hazard information and handle arrays aren't
* discarded during normal session close, they persist past the life of
* the session. Discard them now.
*/
if (!F_ISSET(conn, WT_CONN_LEAK_MEMORY))
if ((s = conn->sessions) != NULL)
for (i = 0; i < conn->session_size; ++s, ++i) {
__wt_free(session, s->dhhash);
__wt_stash_discard_all(session, s);
__wt_free(session, s->hazard);
}
/* Destroy the file-system configuration. */
if (conn->file_system != NULL && conn->file_system->terminate != NULL)
WT_TRET(conn->file_system->terminate(
conn->file_system, (WT_SESSION *)session));
/* Close extensions, first calling any unload entry point. */
while ((dlh = TAILQ_FIRST(&conn->dlhqh)) != NULL) {
TAILQ_REMOVE(&conn->dlhqh, dlh, q);
if (dlh->terminate != NULL)
WT_TRET(dlh->terminate(wt_conn));
WT_TRET(__wt_dlclose(session, dlh));
}
/* Destroy the handle. */
__wt_connection_destroy(conn);
return (ret);
}
/*
* __wt_log_scan --
* Scan the logs, calling a function on each record found.
*/
int
__wt_log_scan(WT_SESSION_IMPL *session, WT_LSN *lsnp, uint32_t flags,
int (*func)(WT_SESSION_IMPL *session,
WT_ITEM *record, WT_LSN *lsnp, void *cookie), void *cookie)
{
WT_CONNECTION_IMPL *conn;
WT_ITEM buf;
WT_DECL_RET;
WT_FH *log_fh;
WT_LOG *log;
WT_LOG_RECORD *logrec;
WT_LSN end_lsn, rd_lsn, start_lsn;
off_t log_size;
uint32_t allocsize, cksum, firstlog, lastlog, lognum, rdup_len, reclen;
u_int i, logcount;
int eol;
char **logfiles;
conn = S2C(session);
log = conn->log;
log_fh = NULL;
logcount = 0;
logfiles = NULL;
eol = 0;
WT_CLEAR(buf);
/*
* If the caller did not give us a callback function there is nothing
* to do.
*/
if (func == NULL)
return (0);
if (LF_ISSET(WT_LOGSCAN_RECOVER))
WT_RET(__wt_verbose(session, WT_VERB_LOG,
"__wt_log_scan truncating to %u/%" PRIuMAX,
log->trunc_lsn.file, (uintmax_t)log->trunc_lsn.offset));
if (log != NULL) {
allocsize = log->allocsize;
if (lsnp == NULL) {
if (LF_ISSET(WT_LOGSCAN_FIRST))
start_lsn = log->first_lsn;
else if (LF_ISSET(WT_LOGSCAN_FROM_CKP))
start_lsn = log->ckpt_lsn;
else
return (WT_ERROR); /* Illegal usage */
} else {
if (LF_ISSET(WT_LOGSCAN_FIRST|WT_LOGSCAN_FROM_CKP))
WT_RET_MSG(session, WT_ERROR,
"choose either a start LSN or a start flag");
/* Offsets must be on allocation boundaries. */
if (lsnp->offset % allocsize != 0 ||
lsnp->file > log->fileid)
return (WT_NOTFOUND);
/*
* Log cursors may not know the starting LSN. If an
* LSN pointer is passed in, but it is the INIT_LSN,
* start from the first_lsn.
*/
start_lsn = *lsnp;
if (IS_INIT_LSN(&start_lsn))
start_lsn = log->first_lsn;
}
end_lsn = log->alloc_lsn;
} else {
/*
* If logging is not configured, we can still print out the log
* if log files exist. We just need to set the LSNs from what
* is in the files versus what is in the live connection.
*/
/*
* Set allocsize to the minimum alignment it could be. Larger
* records and larger allocation boundaries should always be
* a multiple of this.
*/
allocsize = LOG_ALIGN;
lastlog = 0;
firstlog = UINT32_MAX;
WT_RET(__wt_log_get_files(session, &logfiles, &logcount));
if (logcount == 0)
/*
* Return it is not supported if none don't exist.
*/
return (ENOTSUP);
for (i = 0; i < logcount; i++) {
WT_ERR(__wt_log_extract_lognum(session, logfiles[i],
&lognum));
lastlog = WT_MAX(lastlog, lognum);
firstlog = WT_MIN(firstlog, lognum);
}
start_lsn.file = firstlog;
end_lsn.file = lastlog;
start_lsn.offset = end_lsn.offset = 0;
__wt_log_files_free(session, logfiles, logcount);
logfiles = NULL;
}
WT_ERR(__log_openfile(session, 0, &log_fh, start_lsn.file));
WT_ERR(__log_filesize(session, log_fh, &log_size));
rd_lsn = start_lsn;
WT_ERR(__wt_buf_initsize(session, &buf, LOG_ALIGN));
for (;;) {
if (rd_lsn.offset + allocsize > log_size) {
advance:
/*
* If we read the last record, go to the next file.
*/
WT_ERR(__wt_close(session, log_fh));
log_fh = NULL;
eol = 1;
/*
* Truncate this log file before we move to the next.
*/
if (LF_ISSET(WT_LOGSCAN_RECOVER))
WT_ERR(__log_truncate(session, &rd_lsn, 1));
rd_lsn.file++;
rd_lsn.offset = 0;
/*
* Avoid an error message when we reach end of log
* by checking here.
*/
if (rd_lsn.file > end_lsn.file)
break;
WT_ERR(__log_openfile(
session, 0, &log_fh, rd_lsn.file));
WT_ERR(__log_filesize(session, log_fh, &log_size));
continue;
}
/*
* Read the minimum allocation size a record could be.
*/
WT_ASSERT(session, buf.memsize >= allocsize);
WT_ERR(__wt_read(session,
log_fh, rd_lsn.offset, (size_t)allocsize, buf.mem));
/*
* First 8 bytes is the real record length. See if we
* need to read more than the allocation size. We expect
* that we rarely will have to read more. Most log records
* will be fairly small.
*/
reclen = *(uint32_t *)buf.mem;
/*
* Log files are pre-allocated. We never expect a zero length
* unless we've reached the end of the log. The log can be
* written out of order, so when recovery finds the end of
* the log, truncate the file and remove any later log files
* that may exist.
*/
if (reclen == 0) {
/* This LSN is the end. */
break;
}
rdup_len = __wt_rduppo2(reclen, allocsize);
if (reclen > allocsize) {
/*
* The log file end could be the middle of this
* log record.
*/
if (rd_lsn.offset + rdup_len > log_size)
goto advance;
/*
* We need to round up and read in the full padded
* record, especially for direct I/O.
*/
WT_ERR(__wt_buf_grow(session, &buf, rdup_len));
WT_ERR(__wt_read(session,
log_fh, rd_lsn.offset, (size_t)rdup_len, buf.mem));
WT_STAT_FAST_CONN_INCR(session, log_scan_rereads);
}
/*
* We read in the record, verify checksum.
*/
buf.size = reclen;
logrec = (WT_LOG_RECORD *)buf.mem;
cksum = logrec->checksum;
logrec->checksum = 0;
logrec->checksum = __wt_cksum(logrec, logrec->len);
if (logrec->checksum != cksum) {
/*
* A checksum mismatch means we have reached the end of
* the useful part of the log. This should be found on
* the first pass through recovery. In the second pass
* where we truncate the log, this is where it should
* end.
*/
if (log != NULL)
log->trunc_lsn = rd_lsn;
break;
}
/*
* We have a valid log record. If it is not the log file
* header, invoke the callback.
*/
WT_STAT_FAST_CONN_INCR(session, log_scan_records);
if (rd_lsn.offset != 0) {
WT_ERR((*func)(session, &buf, &rd_lsn, cookie));
if (LF_ISSET(WT_LOGSCAN_ONE))
break;
}
rd_lsn.offset += (off_t)rdup_len;
}
/* Truncate if we're in recovery. */
if (LF_ISSET(WT_LOGSCAN_RECOVER) &&
LOG_CMP(&rd_lsn, &log->trunc_lsn) < 0)
WT_ERR(__log_truncate(session, &rd_lsn, 0));
err: WT_STAT_FAST_CONN_INCR(session, log_scans);
if (logfiles != NULL)
__wt_log_files_free(session, logfiles, logcount);
__wt_buf_free(session, &buf);
/*
* If the caller wants one record and it is at the end of log,
* return WT_NOTFOUND.
*/
if (LF_ISSET(WT_LOGSCAN_ONE) && eol && ret == 0)
ret = WT_NOTFOUND;
if (ret == ENOENT)
ret = 0;
if (log_fh != NULL)
WT_TRET(__wt_close(session, log_fh));
return (ret);
}
/*
* __log_file_server --
* The log file server thread. This worker thread manages
* log file operations such as closing and syncing.
*/
static WT_THREAD_RET
__log_file_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *close_fh;
WT_LOG *log;
WT_LSN close_end_lsn, min_lsn;
WT_SESSION_IMPL *session;
uint32_t filenum;
bool locked;
session = arg;
conn = S2C(session);
log = conn->log;
locked = false;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* If there is a log file to close, make sure any outstanding
* write operations have completed, then fsync and close it.
*/
if ((close_fh = log->log_close_fh) != NULL) {
WT_ERR(__wt_log_extract_lognum(session, close_fh->name,
&filenum));
/*
* We update the close file handle before updating the
* close LSN when changing files. It is possible we
* could see mismatched settings. If we do, yield
* until it is set. This should rarely happen.
*/
while (log->log_close_lsn.l.file < filenum)
__wt_yield();
if (__wt_log_cmp(
&log->write_lsn, &log->log_close_lsn) >= 0) {
/*
* We've copied the file handle, clear out the
* one in the log structure to allow it to be
* set again. Copy the LSN before clearing
* the file handle.
* Use a barrier to make sure the compiler does
* not reorder the following two statements.
*/
close_end_lsn = log->log_close_lsn;
WT_FULL_BARRIER();
log->log_close_fh = NULL;
/*
* Set the close_end_lsn to the LSN immediately
* after ours. That is, the beginning of the
* next log file. We need to know the LSN
* file number of our own close in case earlier
* calls are still in progress and the next one
* to move the sync_lsn into the next file for
* later syncs.
*/
WT_ERR(__wt_fsync(session, close_fh, true));
/*
* We want to have the file size reflect actual
* data with minimal pre-allocated zeroed space.
* We can't truncate the file during hot backup,
* or the underlying file system may not support
* truncate: both are OK, it's just more work
* during cursor traversal.
*/
if (!conn->hot_backup) {
__wt_readlock(
session, conn->hot_backup_lock);
if (!conn->hot_backup)
WT_ERR_ERROR_OK(
__wt_ftruncate(session,
close_fh,
close_end_lsn.l.offset),
ENOTSUP);
__wt_readunlock(
session, conn->hot_backup_lock);
}
WT_SET_LSN(&close_end_lsn,
close_end_lsn.l.file + 1, 0);
__wt_spin_lock(session, &log->log_sync_lock);
locked = true;
WT_ERR(__wt_close(session, &close_fh));
WT_ASSERT(session, __wt_log_cmp(
&close_end_lsn, &log->sync_lsn) >= 0);
log->sync_lsn = close_end_lsn;
__wt_cond_signal(session, log->log_sync_cond);
locked = false;
__wt_spin_unlock(session, &log->log_sync_lock);
}
}
/*
* If a later thread asked for a background sync, do it now.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &log->sync_lsn) > 0) {
/*
* Save the latest write LSN which is the minimum
* we will have written to disk.
*/
min_lsn = log->write_lsn;
/*
* We have to wait until the LSN we asked for is
* written. If it isn't signal the wrlsn thread
* to get it written.
*
* We also have to wait for the written LSN and the
* sync LSN to be in the same file so that we know we
* have synchronized all earlier log files.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &min_lsn) <= 0) {
/*
* If the sync file is behind either the one
* wanted for a background sync or the write LSN
* has moved to another file continue to let
* this worker thread process that older file
* immediately.
*/
if ((log->sync_lsn.l.file <
log->bg_sync_lsn.l.file) ||
(log->sync_lsn.l.file < min_lsn.l.file))
continue;
WT_ERR(__wt_fsync(session, log->log_fh, true));
__wt_spin_lock(session, &log->log_sync_lock);
locked = true;
/*
* The sync LSN could have advanced while we
* were writing to disk.
*/
if (__wt_log_cmp(
&log->sync_lsn, &min_lsn) <= 0) {
WT_ASSERT(session,
min_lsn.l.file ==
log->sync_lsn.l.file);
log->sync_lsn = min_lsn;
__wt_cond_signal(
session, log->log_sync_cond);
}
locked = false;
__wt_spin_unlock(session, &log->log_sync_lock);
} else {
__wt_cond_auto_signal(
session, conn->log_wrlsn_cond);
/*
* We do not want to wait potentially a second
* to process this. Yield to give the wrlsn
* thread a chance to run and try again in
* this case.
*/
__wt_yield();
continue;
}
}
/* Wait until the next event. */
__wt_cond_wait(session, conn->log_file_cond, WT_MILLION / 10);
}
if (0) {
err: __wt_err(session, ret, "log close server error");
}
if (locked)
__wt_spin_unlock(session, &log->log_sync_lock);
return (WT_THREAD_RET_VALUE);
}
/*
* __conn_config_file --
* Read in any WiredTiger_config file in the home directory.
*/
static int
__conn_config_file(WT_SESSION_IMPL *session, const char **cfg, WT_ITEM **cbufp)
{
WT_DECL_ITEM(cbuf);
WT_DECL_RET;
WT_FH *fh;
off_t size;
uint32_t len;
int exist, quoted;
uint8_t *p, *t;
*cbufp = NULL; /* Returned buffer */
fh = NULL;
/* Check for an optional configuration file. */
#define WT_CONFIGFILE "WiredTiger.config"
WT_RET(__wt_exist(session, WT_CONFIGFILE, &exist));
if (!exist)
return (0);
/* Open the configuration file. */
WT_RET(__wt_open(session, WT_CONFIGFILE, 0, 0, 0, &fh));
WT_ERR(__wt_filesize(session, fh, &size));
if (size == 0)
goto err;
/*
* Sanity test: a 100KB configuration file would be insane. (There's
* no practical reason to limit the file size, but I can either limit
* the file size to something rational, or I can add code to test if
* the off_t size is larger than a uint32_t, which is more complicated
* and a waste of time.)
*/
if (size > 100 * 1024)
WT_ERR_MSG(session, EFBIG, WT_CONFIGFILE);
len = (uint32_t)size;
/*
* Copy the configuration file into memory, with a little slop, I'm not
* interested in debugging off-by-ones.
*
* The beginning of a file is the same as if we run into an unquoted
* newline character, simplify the parsing loop by pretending that's
* what we're doing.
*/
WT_ERR(__wt_scr_alloc(session, len + 10, &cbuf));
WT_ERR(
__wt_read(session, fh, (off_t)0, len, ((uint8_t *)cbuf->mem) + 1));
((uint8_t *)cbuf->mem)[0] = '\n';
cbuf->size = len + 1;
/*
* Collapse the file's lines into a single string: newline characters
* are replaced with commas unless the newline is quoted or backslash
* escaped. Comment lines (an unescaped newline where the next non-
* white-space character is a hash), are discarded.
*/
for (quoted = 0, p = t = cbuf->mem; len > 0;) {
/*
* Backslash pairs pass through untouched, unless immediately
* preceding a newline, in which case both the backslash and
* the newline are discarded. Backslash characters escape
* quoted characters, too, that is, a backslash followed by a
* quote doesn't start or end a quoted string.
*/
if (*p == '\\' && len > 1) {
if (p[1] != '\n') {
*t++ = p[0];
*t++ = p[1];
}
p += 2;
len -= 2;
continue;
}
/*
* If we're in a quoted string, or starting a quoted string,
* take all characters, including white-space and newlines.
*/
if (quoted || *p == '"') {
if (*p == '"')
quoted = !quoted;
*t++ = *p++;
--len;
continue;
}
/* Everything else gets taken, except for newline characters. */
if (*p != '\n') {
*t++ = *p++;
--len;
continue;
}
/*
* Replace any newline characters with commas (and strings of
* commas are safe).
*
* After any newline, skip to a non-white-space character; if
* the next character is a hash mark, skip to the next newline.
*/
for (;;) {
for (*t++ = ','; --len > 0 && isspace(*++p);)
;
if (len == 0)
break;
if (*p != '#')
break;
while (--len > 0 && *++p != '\n')
;
if (len == 0)
break;
}
}
*t = '\0';
#if 0
fprintf(stderr, "file config: {%s}\n", (const char *)cbuf->data);
#endif
/* Check the configuration string. */
WT_ERR(__wt_config_check(
session, __wt_confchk_wiredtiger_open, cbuf->data, 0));
/*
* The configuration file falls between the default configuration and
* the wiredtiger_open() configuration, overriding the defaults but not
* overriding the wiredtiger_open() configuration.
*/
while (cfg[1] != NULL)
++cfg;
cfg[1] = cfg[0];
cfg[0] = cbuf->data;
*cbufp = cbuf;
if (0) {
err: if (cbuf != NULL)
__wt_buf_free(session, cbuf);
}
if (fh != NULL)
WT_TRET(__wt_close(session, fh));
return (ret);
}
/*
* __conn_single --
* Confirm that no other thread of control is using this database.
*/
static int
__conn_single(WT_SESSION_IMPL *session, const char **cfg)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn, *t;
WT_DECL_RET;
off_t size;
uint32_t len;
int created;
char buf[256];
conn = S2C(session);
/*
* Optionally create the wiredtiger flag file if it doesn't already
* exist. We don't actually care if we create it or not, the "am I the
* only locker" tests are all that matter.
*/
WT_RET(__wt_config_gets(session, cfg, "create", &cval));
WT_RET(__wt_open(session,
WT_SINGLETHREAD, cval.val == 0 ? 0 : 1, 0, 0, &conn->lock_fh));
/*
* Lock a byte of the file: if we don't get the lock, some other process
* is holding it, we're done. Note the file may be zero-length length,
* and that's OK, the underlying call supports acquisition of locks past
* the end-of-file.
*/
if (__wt_bytelock(conn->lock_fh, (off_t)0, 1) != 0)
WT_ERR_MSG(session, EBUSY, "%s",
"WiredTiger database is already being managed by another "
"process");
/* Check to see if another thread of control has this database open. */
__wt_spin_lock(session, &__wt_process.spinlock);
TAILQ_FOREACH(t, &__wt_process.connqh, q)
if (t->home != NULL &&
t != conn && strcmp(t->home, conn->home) == 0) {
ret = EBUSY;
break;
}
__wt_spin_unlock(session, &__wt_process.spinlock);
if (ret != 0)
WT_ERR_MSG(session, EBUSY, "%s",
"WiredTiger database is already being managed by another "
"thread in this process");
/*
* If the size of the file is 0, we created it (or we're racing with
* the thread that created it, it doesn't matter), write some bytes
* into the file. Strictly speaking, this isn't even necessary, but
* zero-length files always make me nervous.
*/
WT_ERR(__wt_filesize(session, conn->lock_fh, &size));
if (size == 0) {
len = (uint32_t)snprintf(buf, sizeof(buf), "%s\n%s\n",
WT_SINGLETHREAD, wiredtiger_version(NULL, NULL, NULL));
WT_ERR(__wt_write(
session, conn->lock_fh, (off_t)0, (uint32_t)len, buf));
created = 1;
} else
created = 0;
/*
* If we found a zero-length WiredTiger lock file, and eventually ended
* as the database owner, return that we created the database. (There
* is a theoretical chance that another process created the WiredTiger
* lock file but we won the race to add the WT_CONNECTION_IMPL structure
* to the process' list. It doesn't much matter, only one thread will
* be told it created the database.)
*/
conn->is_new = created;
return (0);
err: if (conn->lock_fh != NULL) {
WT_TRET(__wt_close(session, conn->lock_fh));
conn->lock_fh = NULL;
}
return (ret);
}
/*
* __log_release --
* Release a log slot.
*/
static int
__log_release(WT_SESSION_IMPL *session, WT_LOGSLOT *slot)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *close_fh;
WT_LOG *log;
WT_LSN sync_lsn;
size_t write_size;
WT_DECL_SPINLOCK_ID(id); /* Must appear last */
conn = S2C(session);
log = conn->log;
/*
* If we're going to have to close our log file, make a local copy
* of the file handle structure.
*/
close_fh = NULL;
if (F_ISSET(slot, SLOT_CLOSEFH)) {
close_fh = log->log_close_fh;
log->log_close_fh = NULL;
F_CLR(slot, SLOT_CLOSEFH);
}
/* Write the buffered records */
if (F_ISSET(slot, SLOT_BUFFERED)) {
write_size = (size_t)
(slot->slot_end_lsn.offset - slot->slot_start_offset);
WT_ERR(__wt_write(session, slot->slot_fh,
slot->slot_start_offset, write_size, slot->slot_buf.mem));
}
/*
* Wait for earlier groups to finish, otherwise there could be holes
* in the log file.
*/
while (LOG_CMP(&log->write_lsn, &slot->slot_release_lsn) != 0)
__wt_yield();
log->write_lsn = slot->slot_end_lsn;
/*
* Try to consolidate calls to fsync to wait less. Acquire a spin lock
* so that threads finishing writing to the log will wait while the
* current fsync completes and advance log->write_lsn.
*/
while (F_ISSET(slot, SLOT_SYNC) &&
LOG_CMP(&log->sync_lsn, &slot->slot_end_lsn) < 0) {
if (__wt_spin_trylock(session, &log->log_sync_lock, &id) != 0) {
(void)__wt_cond_wait(
session, log->log_sync_cond, 10000);
continue;
}
/*
* Record the current end of log after we grabbed the lock.
* That is how far our fsync call with guarantee.
*/
sync_lsn = log->write_lsn;
if (LOG_CMP(&log->sync_lsn, &slot->slot_end_lsn) < 0) {
WT_STAT_FAST_CONN_INCR(session, log_sync);
ret = __wt_fsync(session, log->log_fh);
if (ret == 0) {
F_CLR(slot, SLOT_SYNC);
log->sync_lsn = sync_lsn;
ret = __wt_cond_signal(
session, log->log_sync_cond);
}
}
__wt_spin_unlock(session, &log->log_sync_lock);
WT_ERR(ret);
}
if (F_ISSET(slot, SLOT_BUF_GROW)) {
WT_STAT_FAST_CONN_INCR(session, log_buffer_grow);
F_CLR(slot, SLOT_BUF_GROW);
WT_STAT_FAST_CONN_INCRV(session,
log_buffer_size, slot->slot_buf.memsize);
WT_ERR(__wt_buf_grow(session,
&slot->slot_buf, slot->slot_buf.memsize * 2));
}
/*
* If we have a file to close, close it now.
*/
if (close_fh)
WT_ERR(__wt_close(session, close_fh));
err: if (ret != 0 && slot->slot_error == 0)
slot->slot_error = ret;
return (ret);
}
/*
* __log_file_server --
* The log file server thread. This worker thread manages
* log file operations such as closing and syncing.
*/
static WT_THREAD_RET
__log_file_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *close_fh;
WT_LOG *log;
WT_LSN close_end_lsn, min_lsn;
WT_SESSION_IMPL *session;
uint32_t filenum;
int locked;
session = arg;
conn = S2C(session);
log = conn->log;
locked = 0;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* If there is a log file to close, make sure any outstanding
* write operations have completed, then fsync and close it.
*/
if ((close_fh = log->log_close_fh) != NULL) {
WT_ERR(__wt_log_extract_lognum(session, close_fh->name,
&filenum));
/*
* We update the close file handle before updating the
* close LSN when changing files. It is possible we
* could see mismatched settings. If we do, yield
* until it is set. This should rarely happen.
*/
while (log->log_close_lsn.file < filenum)
__wt_yield();
if (__wt_log_cmp(
&log->write_lsn, &log->log_close_lsn) >= 0) {
/*
* We've copied the file handle, clear out the
* one in the log structure to allow it to be
* set again. Copy the LSN before clearing
* the file handle.
* Use a barrier to make sure the compiler does
* not reorder the following two statements.
*/
close_end_lsn = log->log_close_lsn;
WT_FULL_BARRIER();
log->log_close_fh = NULL;
/*
* Set the close_end_lsn to the LSN immediately
* after ours. That is, the beginning of the
* next log file. We need to know the LSN
* file number of our own close in case earlier
* calls are still in progress and the next one
* to move the sync_lsn into the next file for
* later syncs.
*/
close_end_lsn.file++;
close_end_lsn.offset = 0;
WT_ERR(__wt_fsync(session, close_fh));
__wt_spin_lock(session, &log->log_sync_lock);
locked = 1;
WT_ERR(__wt_close(session, &close_fh));
WT_ASSERT(session, __wt_log_cmp(
&close_end_lsn, &log->sync_lsn) >= 0);
log->sync_lsn = close_end_lsn;
WT_ERR(__wt_cond_signal(
session, log->log_sync_cond));
locked = 0;
__wt_spin_unlock(session, &log->log_sync_lock);
}
}
/*
* If a later thread asked for a background sync, do it now.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &log->sync_lsn) > 0) {
/*
* Save the latest write LSN which is the minimum
* we will have written to disk.
*/
min_lsn = log->write_lsn;
/*
* We have to wait until the LSN we asked for is
* written. If it isn't signal the wrlsn thread
* to get it written.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &min_lsn) <= 0) {
WT_ERR(__wt_fsync(session, log->log_fh));
__wt_spin_lock(session, &log->log_sync_lock);
locked = 1;
/*
* The sync LSN could have advanced while we
* were writing to disk.
*/
if (__wt_log_cmp(
&log->sync_lsn, &min_lsn) <= 0) {
log->sync_lsn = min_lsn;
WT_ERR(__wt_cond_signal(
session, log->log_sync_cond));
}
locked = 0;
__wt_spin_unlock(session, &log->log_sync_lock);
} else {
WT_ERR(__wt_cond_signal(
session, conn->log_wrlsn_cond));
/*
* We do not want to wait potentially a second
* to process this. Yield to give the wrlsn
* thread a chance to run and try again in
* this case.
*/
__wt_yield();
continue;
}
}
/* Wait until the next event. */
WT_ERR(__wt_cond_wait(
session, conn->log_file_cond, WT_MILLION));
}
if (0) {
err: __wt_err(session, ret, "log close server error");
}
if (locked)
__wt_spin_unlock(session, &log->log_sync_lock);
return (WT_THREAD_RET_VALUE);
}
/*
* __wt_log_read --
* Read the log record at the given LSN. Return the record (including
* the log header) in the WT_ITEM. Caller is responsible for freeing it.
*/
int
__wt_log_read(WT_SESSION_IMPL *session, WT_ITEM *record, WT_LSN *lsnp,
uint32_t flags)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *log_fh;
WT_LOG *log;
WT_LOG_RECORD *logrec;
uint32_t cksum, rdup_len, reclen;
WT_UNUSED(flags);
/*
* If the caller didn't give us an LSN or something to return,
* there's nothing to do.
*/
if (lsnp == NULL || record == NULL)
return (0);
conn = S2C(session);
log = conn->log;
/*
* If the offset isn't on an allocation boundary it must be wrong.
*/
if (lsnp->offset % log->allocsize != 0 || lsnp->file > log->fileid)
return (WT_NOTFOUND);
WT_RET(__log_openfile(session, 0, &log_fh, lsnp->file));
/*
* Read the minimum allocation size a record could be.
*/
WT_ERR(__wt_buf_init(session, record, log->allocsize));
WT_ERR(__wt_read(session,
log_fh, lsnp->offset, (size_t)log->allocsize, record->mem));
/*
* First 4 bytes is the real record length. See if we
* need to read more than the allocation size. We expect
* that we rarely will have to read more. Most log records
* will be fairly small.
*/
reclen = *(uint32_t *)record->mem;
if (reclen == 0) {
ret = WT_NOTFOUND;
goto err;
}
if (reclen > log->allocsize) {
rdup_len = __wt_rduppo2(reclen, log->allocsize);
WT_ERR(__wt_buf_grow(session, record, rdup_len));
WT_ERR(__wt_read(session,
log_fh, lsnp->offset, (size_t)rdup_len, record->mem));
}
/*
* We read in the record, verify checksum.
*/
logrec = (WT_LOG_RECORD *)record->mem;
cksum = logrec->checksum;
logrec->checksum = 0;
logrec->checksum = __wt_cksum(logrec, logrec->len);
if (logrec->checksum != cksum)
WT_ERR_MSG(session, WT_ERROR, "log_read: Bad checksum");
record->size = logrec->len;
WT_STAT_FAST_CONN_INCR(session, log_reads);
err:
WT_TRET(__wt_close(session, log_fh));
return (ret);
}
