/*
* Initialize workspace for a worker process: create the schema if it doesn't
* already exist.
*/
static void
initialize_worker_spi(worktable *table)
{
int ret;
int ntup;
bool isnull;
StringInfoData buf;
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, "initializing spi_worker schema");
/* XXX could we use CREATE SCHEMA IF NOT EXISTS? */
initStringInfo(&buf);
appendStringInfo(&buf, "select count(*) from pg_namespace where nspname = '%s'",
table->schema);
ret = SPI_execute(buf.data, true, 0);
if (ret != SPI_OK_SELECT)
elog(FATAL, "SPI_execute failed: error code %d", ret);
if (SPI_processed != 1)
elog(FATAL, "not a singleton result");
ntup = DatumGetInt64(SPI_getbinval(SPI_tuptable->vals[0],
SPI_tuptable->tupdesc,
1, &isnull));
if (isnull)
elog(FATAL, "null result");
if (ntup == 0)
{
resetStringInfo(&buf);
appendStringInfo(&buf,
"CREATE SCHEMA \"%s\" "
"CREATE TABLE \"%s\" ("
" type text CHECK (type IN ('total', 'delta')), "
" value integer)"
"CREATE UNIQUE INDEX \"%s_unique_total\" ON \"%s\" (type) "
"WHERE type = 'total'",
table->schema, table->name, table->name, table->name);
/* set statement start time */
SetCurrentStatementStartTimestamp();
ret = SPI_execute(buf.data, false, 0);
if (ret != SPI_OK_UTILITY)
elog(FATAL, "failed to create my schema");
}
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
}
static void
execute_pg_settings_logger(config_log_objects *objects) {
int ret;
bool isnull;
StringInfoData buf;
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, "executing configuration logger function");
initStringInfo(&buf);
appendStringInfo(
&buf,
"SELECT %s.%s()",
config_log_schema,
objects->function_name
);
ret = SPI_execute(buf.data, false, 0);
if (ret != SPI_OK_SELECT)
{
elog(FATAL, "SPI_execute failed: error code %d", ret);
}
if (SPI_processed != 1)
{
elog(FATAL, "not a singleton result");
}
log_info("pg_settings_logger() executed");
if(DatumGetBool(SPI_getbinval(SPI_tuptable->vals[0],
SPI_tuptable->tupdesc,
1, &isnull)))
{
log_info("Configuration changes recorded");
}
else
{
log_info("No configuration changes detected");
}
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
}
void worker_main(Datum arg)
{
int ret;
StringInfoData buf;
uint32 segment = UInt32GetDatum(arg);
/* Setup signal handlers */
pqsignal(SIGHUP, worker_sighup);
pqsignal(SIGTERM, worker_sigterm);
/* Allow signals */
BackgroundWorkerUnblockSignals();
initialize_worker(segment);
/* Connect to the database */
BackgroundWorkerInitializeConnection(job->datname, job->rolname);
elog(LOG, "%s initialized running job id %d", MyBgworkerEntry->bgw_name, job->job_id);
pgstat_report_appname(MyBgworkerEntry->bgw_name);
/* Initialize the query text */
initStringInfo(&buf);
appendStringInfo(&buf,
"SELECT * FROM %s.%s(%d, NULL)",
job_run_function.schema,
job_run_function.name,
job->job_id);
/* Initialize the SPI subsystem */
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, buf.data);
SetCurrentStatementStartTimestamp();
/* And run the query */
ret = SPI_execute(buf.data, true, 0);
if (ret < 0)
elog(FATAL, "errors while executing %s", buf.data);
/* Commmit the transaction */
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
proc_exit(0);
}
static void
cq_bgproc_main(Datum arg)
{
void (*run) (void);
int default_priority = getpriority(PRIO_PROCESS, MyProcPid);
int priority;
MyContQueryProc = (ContQueryProc *) DatumGetPointer(arg);
BackgroundWorkerUnblockSignals();
BackgroundWorkerInitializeConnection(NameStr(MyContQueryProc->group->db_name), NULL);
/* if we got a cancel signal in prior command, quit */
CHECK_FOR_INTERRUPTS();
MyContQueryProc->latch = &MyProc->procLatch;
ereport(LOG, (errmsg("continuous query process \"%s\" running with pid %d", GetContQueryProcName(MyContQueryProc), MyProcPid)));
pgstat_report_activity(STATE_RUNNING, GetContQueryProcName(MyContQueryProc));
/*
* The BackgroundWorkerHandle's slot is always in the range [0, max_worker_processes)
* and will be unique for any background process being run. We use this knowledge to
* assign our continuous query processes's a unique ID that fits within any TupleBuffer's
* waiters Bitmapset.
*/
MyContQueryProc->id = MyContQueryProc->handle.slot;
/*
* Be nice!
*
* More is less here. A higher number indicates a lower scheduling priority.
*/
priority = Max(default_priority, MAX_PRIORITY - ceil(continuous_query_proc_priority * (MAX_PRIORITY - default_priority)));
priority = nice(priority);
switch (MyContQueryProc->type)
{
case Combiner:
am_cont_combiner = true;
run = &ContinuousQueryCombinerMain;
break;
case Worker:
am_cont_worker = true;
run = &ContinuousQueryWorkerMain;
break;
default:
ereport(ERROR, (errmsg("continuous queries can only be run as worker or combiner processes")));
}
/* initialize process level CQ stats */
cq_stat_init(&MyProcCQStats, 0, MyProcPid);
run();
/* purge proc level CQ stats */
cq_stat_send_purge(0, MyProcPid, IsContQueryWorkerProcess() ? CQ_STAT_WORKER : CQ_STAT_COMBINER);
}
/*
* Ensure that the environment is sane.
* This involves checking the Postgresql version, and if in network mode
* also establishing a connection to a receiver.
*/
int ensure_valid_environment(void) {
StringInfoData buf;
int retval;
char* pgversion;
SPITupleTable *coltuptable;
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
/* Ensure compatible version */
pgstat_report_activity(STATE_RUNNING, "verifying compatible postgres version");
initStringInfo(&buf);
appendStringInfo(&buf,
"select version();"
);
retval = SPI_execute(buf.data, false, 0);
if (retval != SPI_OK_SELECT) {
elog(FATAL, "Unable to query postgres version %d", retval);
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
return 1;
}
coltuptable = SPI_tuptable;
pgversion = SPI_getvalue(coltuptable->vals[0], coltuptable->tupdesc, 1);
if(strstr(pgversion, "PostgreSQL 9.3") == NULL) {
elog(FATAL, "Unsupported Postgresql version");
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
return 1;
}
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
/*
* Attempt to establish a connection if the output mode is network.
*/
if (strcmp(output_mode, "network") == 0) {
retval = establish_connection();
if (retval == 2) {
elog(LOG, "Error : Failed to connect to antenna please check domain is available from host.");
}
}
//TODO verify logging directory is accessible when csv mode.
elog(LOG, "Pgsampler Initialized");
return 0;
}
/*
* Handle BEGIN message.
*/
static void
apply_handle_begin(StringInfo s)
{
LogicalRepBeginData begin_data;
logicalrep_read_begin(s, &begin_data);
remote_final_lsn = begin_data.final_lsn;
in_remote_transaction = true;
pgstat_report_activity(STATE_RUNNING, NULL);
}
/*
* Handle COMMIT message.
*
* TODO, support tracking of multiple origins
*/
static void
apply_handle_commit(StringInfo s)
{
LogicalRepCommitData commit_data;
logicalrep_read_commit(s, &commit_data);
Assert(commit_data.commit_lsn == remote_final_lsn);
/* The synchronization worker runs in single transaction. */
if (IsTransactionState() && !am_tablesync_worker())
{
/*
* Update origin state so we can restart streaming from correct
* position in case of crash.
*/
replorigin_session_origin_lsn = commit_data.end_lsn;
replorigin_session_origin_timestamp = commit_data.committime;
CommitTransactionCommand();
pgstat_report_stat(false);
store_flush_position(commit_data.end_lsn);
}
else
{
/* Process any invalidation messages that might have accumulated. */
AcceptInvalidationMessages();
maybe_reread_subscription();
}
in_remote_transaction = false;
/* Process any tables that are being synchronized in parallel. */
process_syncing_tables(commit_data.end_lsn);
pgstat_report_activity(STATE_IDLE, NULL);
}
void
worker_spi_main(Datum main_arg)
{
int index = DatumGetInt32(main_arg);
worktable *table;
StringInfoData buf;
char name[20];
table = palloc(sizeof(worktable));
sprintf(name, "schema%d", index);
table->schema = pstrdup(name);
table->name = pstrdup("counted");
/* Establish signal handlers before unblocking signals. */
pqsignal(SIGHUP, worker_spi_sighup);
pqsignal(SIGTERM, worker_spi_sigterm);
/* We're now ready to receive signals */
BackgroundWorkerUnblockSignals();
/* Connect to our database */
BackgroundWorkerInitializeConnection("postgres", NULL);
elog(LOG, "%s initialized with %s.%s",
MyBgworkerEntry->bgw_name, table->schema, table->name);
initialize_worker_spi(table);
/*
* Quote identifiers passed to us. Note that this must be done after
* initialize_worker_spi, because that routine assumes the names are not
* quoted.
*
* Note some memory might be leaked here.
*/
table->schema = quote_identifier(table->schema);
table->name = quote_identifier(table->name);
initStringInfo(&buf);
appendStringInfo(&buf,
"WITH deleted AS (DELETE "
"FROM %s.%s "
"WHERE type = 'delta' RETURNING value), "
"total AS (SELECT coalesce(sum(value), 0) as sum "
"FROM deleted) "
"UPDATE %s.%s "
"SET value = %s.value + total.sum "
"FROM total WHERE type = 'total' "
"RETURNING %s.value",
table->schema, table->name,
table->schema, table->name,
table->name,
table->name);
/*
* Main loop: do this until the SIGTERM handler tells us to terminate
*/
while (!got_sigterm)
{
int ret;
int rc;
/*
* Background workers mustn't call usleep() or any direct equivalent:
* instead, they may wait on their process latch, which sleeps as
* necessary, but is awakened if postmaster dies. That way the
* background process goes away immediately in an emergency.
*/
rc = WaitLatch(&MyProc->procLatch,
WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH,
worker_spi_naptime * 1000L);
ResetLatch(&MyProc->procLatch);
/* emergency bailout if postmaster has died */
if (rc & WL_POSTMASTER_DEATH)
proc_exit(1);
/*
* In case of a SIGHUP, just reload the configuration.
*/
if (got_sighup)
{
got_sighup = false;
ProcessConfigFile(PGC_SIGHUP);
}
/*
* Start a transaction on which we can run queries. Note that each
* StartTransactionCommand() call should be preceded by a
* SetCurrentStatementStartTimestamp() call, which sets both the time
* for the statement we're about the run, and also the transaction
* start time. Also, each other query sent to SPI should probably be
* preceded by SetCurrentStatementStartTimestamp(), so that statement
* start time is always up to date.
*
* The SPI_connect() call lets us run queries through the SPI manager,
* and the PushActiveSnapshot() call creates an "active" snapshot
* which is necessary for queries to have MVCC data to work on.
*
* The pgstat_report_activity() call makes our activity visible
* through the pgstat views.
*/
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, buf.data);
/* We can now execute queries via SPI */
ret = SPI_execute(buf.data, false, 0);
if (ret != SPI_OK_UPDATE_RETURNING)
elog(FATAL, "cannot select from table %s.%s: error code %d",
table->schema, table->name, ret);
if (SPI_processed > 0)
{
bool isnull;
int32 val;
val = DatumGetInt32(SPI_getbinval(SPI_tuptable->vals[0],
SPI_tuptable->tupdesc,
1, &isnull));
if (!isnull)
elog(LOG, "%s: count in %s.%s is now %d",
MyBgworkerEntry->bgw_name,
table->schema, table->name, val);
}
/*
* And finish our transaction.
*/
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
}
proc_exit(1);
}
/*
* This is the function called from the main polling loop.
*
* It calls command functions to get strings of data, and sends them to the server.
*
*/
int collect_and_send_metrics(int cycle) {
int retval;
char* command;
StringInfoData commands;
pgstat_report_activity(STATE_RUNNING, "Collecting metrics");
initStringInfo(&commands);
/*
* Populate first cycle command data. These are executed on the first cycle
* of a restart. The bgworker restars every N cycles, as listed at the bottom
* of the main loop in pgsampler.c.
*/
if (cycle == 0) {
command = restart_gucs();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("restart_gucs", command);
pfree(command);
command = relation_info();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_class", command);
pfree(command);
command = database_list();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("databases", command);
pfree(command);
command = column_info();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("columns", command);
pfree(command);
command = index_info();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("indexes", command);
pfree(command);
command = column_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_column", command);
pfree(command);
command = db_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_database", command);
pfree(command);
}
/* HEARTBEAT */
if (cycle % heartbeat_seconds == 0) {
command = heartbeat();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("heartbeats", command);
pfree(command);
}
/* SYSTEM INFO */
if (cycle % system_seconds == 0) {
command = system_info();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_system", command);
pfree(command);
command = fs_info();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_filesystem", command);
pfree(command);
}
/* */
if (cycle % activity_seconds == 0) {
command = activity_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_activity", command);
pfree(command);
}
if (cycle % replication_seconds == 0) {
command = replication_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_replication", command);
pfree(command);
}
/* */
if (cycle % bgwriter_seconds == 0) {
command = bgwriter_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_bgwriter", command);
pfree(command);
}
if (cycle % guc_seconds == 0) {
command = transient_gucs();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("transient_gucs", command);
pfree(command);
}
if (cycle % statements_seconds == 0) {
command = stat_statements();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_statements", command);
pfree(command);
}
/* */
if (cycle % relation_seconds == 0) {
command = table_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_table", command);
pfree(command);
command = index_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_index", command);
pfree(command);
command = table_io_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("statio_user_tables", command);
pfree(command);
command = index_io_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("statio_user_indexes", command);
pfree(command);
command = function_stats();
appendStringInfoString(&commands, command);
if (strcmp(output_mode, "csv") == 0)
write_to_csv("stat_function", command);
pfree(command);
}
/* Send / Write metrics based on output_mode */
if (strcmp(output_mode, "network") == 0) {
pgstat_report_activity(STATE_RUNNING, "Sending metrics to antenna");
retval = send_data(commands.data);
if (retval == NO_DATA_SENT) { //close socket and retry establishing connection and sending data.
// elog(LOG, "reseting..."); //just a note to say reseting socket
if (sockfd != 0)
shutdown(sockfd, SHUT_RDWR);
sockfd = 0;
retval = send_data(commands.data); // we ignore success or failure here. drops data if fails.
}
}
return 0;
}
void
worker_test_main(Datum main_arg)
{
dsm_segment *seg;
volatile test_shm_mq_header *hdr;
PGPROC *registrant;
pqsignal(SIGHUP, handle_sighup);
pqsignal(SIGTERM, handle_sigterm);
BackgroundWorkerUnblockSignals();
printf("worker_test_main: %d\n", DatumGetInt32(main_arg));
CurrentResourceOwner = ResourceOwnerCreate(NULL, "worker test");
seg = dsm_attach(DatumGetInt32(main_arg));
if (seg == NULL)
ereport(ERROR,
(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
errmsg("unable to map dynamic shared memory segment")));
hdr = dsm_segment_address(seg);
/* 開始 */
SpinLockAcquire(&hdr->mutex);
hdr->workers_ready++;
hdr->workers_attached++;
SpinLockRelease(&hdr->mutex);
registrant = BackendPidGetProc(MyBgworkerEntry->bgw_notify_pid);
if (registrant == NULL)
{
elog(DEBUG1, "registrant backend has exited prematurely");
proc_exit(1);
}
SetLatch(®istrant->procLatch);
/* Do the work */
BackgroundWorkerInitializeConnection(hdr->dbname, NULL);
printf("DSM: %p\n", dsm_segment_address);
#if 0
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, "initializing spi_worker schema");
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
#endif
dsm_detach(seg);
proc_exit(0);
}
/*
* ContinuousQueryWorkerStartup
*
* Launches a CQ worker, which continuously generates partial query results to send
* back to the combiner process.
*/
void
ContinuousQueryWorkerRun(Portal portal, ContinuousViewState *state, QueryDesc *queryDesc, ResourceOwner owner)
{
EState *estate = NULL;
DestReceiver *dest;
CmdType operation;
MemoryContext oldcontext;
int timeoutms = state->maxwaitms;
MemoryContext runcontext;
CQProcEntry *entry = GetCQProcEntry(MyCQId);
ResourceOwner cqowner = ResourceOwnerCreate(NULL, "CQResourceOwner");
bool savereadonly = XactReadOnly;
cq_stat_initialize(state->viewid, MyProcPid);
dest = CreateDestReceiver(DestCombiner);
SetCombinerDestReceiverParams(dest, MyCQId);
/* workers only need read-only transactions */
XactReadOnly = true;
runcontext = AllocSetContextCreate(TopMemoryContext, "CQRunContext",
ALLOCSET_DEFAULT_MINSIZE,
ALLOCSET_DEFAULT_INITSIZE,
ALLOCSET_DEFAULT_MAXSIZE);
elog(LOG, "\"%s\" worker %d running", queryDesc->plannedstmt->cq_target->relname, MyProcPid);
MarkWorkerAsRunning(MyCQId, MyWorkerId);
pgstat_report_activity(STATE_RUNNING, queryDesc->sourceText);
TupleBufferInitLatch(WorkerTupleBuffer, MyCQId, MyWorkerId, &MyProc->procLatch);
oldcontext = MemoryContextSwitchTo(runcontext);
retry:
PG_TRY();
{
bool xact_commit = true;
TimestampTz last_process = GetCurrentTimestamp();
TimestampTz last_commit = GetCurrentTimestamp();
start_executor(queryDesc, runcontext, cqowner);
CurrentResourceOwner = cqowner;
estate = queryDesc->estate;
operation = queryDesc->operation;
/*
* Initialize context that lives for the duration of a single iteration
* of the main worker loop
*/
CQExecutionContext = AllocSetContextCreate(estate->es_query_cxt, "CQExecutionContext",
ALLOCSET_DEFAULT_MINSIZE,
ALLOCSET_DEFAULT_INITSIZE,
ALLOCSET_DEFAULT_MAXSIZE);
estate->es_lastoid = InvalidOid;
/*
* Startup combiner receiver
*/
(*dest->rStartup) (dest, operation, queryDesc->tupDesc);
for (;;)
{
if (!TupleBufferHasUnreadSlots())
{
if (TimestampDifferenceExceeds(last_process, GetCurrentTimestamp(), state->emptysleepms))
{
/* force stats flush */
cq_stat_report(true);
pgstat_report_activity(STATE_IDLE, queryDesc->sourceText);
TupleBufferWait(WorkerTupleBuffer, MyCQId, MyWorkerId);
pgstat_report_activity(STATE_RUNNING, queryDesc->sourceText);
}
else
pg_usleep(Min(WAIT_SLEEP_MS, state->emptysleepms) * 1000);
}
TupleBufferResetNotify(WorkerTupleBuffer, MyCQId, MyWorkerId);
if (xact_commit)
StartTransactionCommand();
set_snapshot(estate, cqowner);
CurrentResourceOwner = cqowner;
MemoryContextSwitchTo(estate->es_query_cxt);
estate->es_processed = 0;
estate->es_filtered = 0;
/*
* Run plan on a microbatch
*/
ExecutePlan(estate, queryDesc->planstate, operation,
true, 0, timeoutms, ForwardScanDirection, dest);
IncrementCQExecutions(1);
TupleBufferClearPinnedSlots();
if (state->long_xact)
{
if (TimestampDifferenceExceeds(last_commit, GetCurrentTimestamp(), LONG_RUNNING_XACT_DURATION))
xact_commit = true;
else
xact_commit = false;
}
unset_snapshot(estate, cqowner);
if (xact_commit)
{
CommitTransactionCommand();
last_commit = GetCurrentTimestamp();
}
MemoryContextResetAndDeleteChildren(CQExecutionContext);
MemoryContextSwitchTo(runcontext);
CurrentResourceOwner = cqowner;
if (estate->es_processed || estate->es_filtered)
{
/*
* If the CV query is such that the select does not return any tuples
* ex: select id where id=99; and id=99 does not exist, then this reset
* will fail. What will happen is that the worker will block at the latch for every
* allocated slot, TILL a cv returns a non-zero tuple, at which point
* the worker will resume a simple sleep for the threshold time.
*/
last_process = GetCurrentTimestamp();
/*
* Send stats to the collector
*/
cq_stat_report(false);
}
/* Has the CQ been deactivated? */
if (!entry->active)
{
if (ActiveSnapshotSet())
unset_snapshot(estate, cqowner);
if (IsTransactionState())
CommitTransactionCommand();
break;
}
}
CurrentResourceOwner = cqowner;
/*
* The cleanup functions below expect these things to be registered
*/
RegisterSnapshotOnOwner(estate->es_snapshot, cqowner);
RegisterSnapshotOnOwner(queryDesc->snapshot, cqowner);
RegisterSnapshotOnOwner(queryDesc->crosscheck_snapshot, cqowner);
/* cleanup */
ExecutorFinish(queryDesc);
ExecutorEnd(queryDesc);
FreeQueryDesc(queryDesc);
}
PG_CATCH();
{
EmitErrorReport();
FlushErrorState();
/* Since the worker is read-only, we can simply commit the transaction. */
if (ActiveSnapshotSet())
unset_snapshot(estate, cqowner);
if (IsTransactionState())
CommitTransactionCommand();
TupleBufferUnpinAllPinnedSlots();
TupleBufferClearReaders();
/* This resets the es_query_ctx and in turn the CQExecutionContext */
MemoryContextResetAndDeleteChildren(runcontext);
IncrementCQErrors(1);
if (continuous_query_crash_recovery)
goto retry;
}
PG_END_TRY();
(*dest->rShutdown) (dest);
MemoryContextSwitchTo(oldcontext);
MemoryContextDelete(runcontext);
XactReadOnly = savereadonly;
/*
* Remove proc-level stats
*/
cq_stat_report(true);
cq_stat_send_purge(state->viewid, MyProcPid, CQ_STAT_WORKER);
CurrentResourceOwner = owner;
}
static void
kill_idle_main(Datum main_arg)
{
StringInfoData buf;
/* Register functions for SIGTERM/SIGHUP management */
pqsignal(SIGHUP, kill_idle_sighup);
pqsignal(SIGTERM, kill_idle_sigterm);
/* We're now ready to receive signals */
BackgroundWorkerUnblockSignals();
/* Connect to a database */
BackgroundWorkerInitializeConnection("postgres", NULL);
/* Build query for process */
initStringInfo(&buf);
kill_idle_build_query(&buf);
while (!got_sigterm)
{
int rc, ret, i;
/* Wait necessary amount of time */
rc = WaitLatch(&MyProc->procLatch,
WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH,
kill_max_idle_time * 1000L,
PG_WAIT_EXTENSION);
ResetLatch(&MyProc->procLatch);
/* Emergency bailout if postmaster has died */
if (rc & WL_POSTMASTER_DEATH)
proc_exit(1);
/* Process signals */
if (got_sighup)
{
int old_interval;
/* Save old value of kill interval */
old_interval = kill_max_idle_time;
/* Process config file */
ProcessConfigFile(PGC_SIGHUP);
got_sighup = false;
ereport(LOG, (errmsg("bgworker kill_idle signal: processed SIGHUP")));
/* Rebuild query if necessary */
if (old_interval != kill_max_idle_time)
{
resetStringInfo(&buf);
initStringInfo(&buf);
kill_idle_build_query(&buf);
}
}
if (got_sigterm)
{
/* Simply exit */
ereport(LOG, (errmsg("bgworker kill_idle signal: processed SIGTERM")));
proc_exit(0);
}
/* Process idle connection kill */
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, buf.data);
/* Statement start time */
SetCurrentStatementStartTimestamp();
/* Execute query */
ret = SPI_execute(buf.data, false, 0);
/* Some error handling */
if (ret != SPI_OK_SELECT)
elog(FATAL, "Error when trying to kill idle connections");
/* Do some processing and log stuff disconnected */
for (i = 0; i < SPI_processed; i++)
{
int32 pidValue;
bool isnull;
char *datname = NULL;
char *usename = NULL;
char *client_addr = NULL;
/* Fetch values */
pidValue = DatumGetInt32(SPI_getbinval(SPI_tuptable->vals[i],
SPI_tuptable->tupdesc,
1, &isnull));
usename = DatumGetCString(SPI_getbinval(SPI_tuptable->vals[i],
SPI_tuptable->tupdesc,
3, &isnull));
datname = DatumGetCString(SPI_getbinval(SPI_tuptable->vals[i],
SPI_tuptable->tupdesc,
4, &isnull));
client_addr = DatumGetCString(SPI_getbinval(SPI_tuptable->vals[i],
SPI_tuptable->tupdesc,
5, &isnull));
/* Log what has been disconnected */
elog(LOG, "Disconnected idle connection: PID %d %s/%s/%s",
pidValue, datname ? datname : "none",
usename ? usename : "none",
client_addr ? client_addr : "none");
}
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
}
/* No problems, so clean exit */
proc_exit(0);
}
/*
* worker logic
*/
void
wed_worker_main(Datum main_arg)
{
StringInfoData buf;
/* Establish signal handlers before unblocking signals. */
pqsignal(SIGHUP, wed_worker_sighup);
pqsignal(SIGTERM, wed_worker_sigterm);
/* We're now ready to receive signals */
BackgroundWorkerUnblockSignals();
/* Connect to our database */
BackgroundWorkerInitializeConnection(wed_worker_db_name, NULL);
elog(LOG, "%s initialized in: %s",
MyBgworkerEntry->bgw_name, wed_worker_db_name);
initStringInfo(&buf);
appendStringInfo(&buf, "SELECT trcheck()");
/*
* Main loop: do this until the SIGTERM handler tells us to terminate
*/
while (!got_sigterm)
{
int ret;
int rc;
/*
* Background workers mustn't call usleep() or any direct equivalent:
* instead, they may wait on their process latch, which sleeps as
* necessary, but is awakened if postmaster dies. That way the
* background process goes away immediately in an emergency.
*/
rc = WaitLatch(&MyProc->procLatch,
WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH,
wed_worker_naptime * 1000L);
ResetLatch(&MyProc->procLatch);
/* emergency bailout if postmaster has died */
if (rc & WL_POSTMASTER_DEATH)
proc_exit(1);
/*
* In case of a SIGHUP, just reload the configuration.
*/
if (got_sighup)
{
got_sighup = false;
ProcessConfigFile(PGC_SIGHUP);
}
/*
* Start a transaction on which we can run queries. Note that each
* StartTransactionCommand() call should be preceded by a
* SetCurrentStatementStartTimestamp() call, which sets both the time
* for the statement we're about the run, and also the transaction
* start time. Also, each other query sent to SPI should probably be
* preceded by SetCurrentStatementStartTimestamp(), so that statement
* start time is always up to date.
*
* The SPI_connect() call lets us run queries through the SPI manager,
* and the PushActiveSnapshot() call creates an "active" snapshot
* which is necessary for queries to have MVCC data to work on.
*
* The pgstat_report_activity() call makes our activity visible
* through the pgstat views.
*/
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, buf.data);
/* We can now execute queries via SPI */
ret = SPI_execute(buf.data, false, 0);
if (ret != SPI_OK_SELECT)
elog(FATAL, "stored procedure trcheck() not found: error code %d", ret);
elog(LOG, "%s : trcheck() done !", MyBgworkerEntry->bgw_name);
/*
* And finish our transaction.
*/
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
}
proc_exit(1);
}
static config_log_objects *
initialize_objects(void)
{
config_log_objects *objects;
int ret;
int ntup;
bool isnull;
StringInfoData buf;
objects = palloc(sizeof(config_log_objects));
objects->table_name = pstrdup("pg_settings_log");
objects->function_name = pstrdup("pg_settings_logger");
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, "Verifying config log objects");
initStringInfo(&buf);
appendStringInfo(
&buf,
"SELECT COUNT(*)\
FROM information_schema.tables\
WHERE table_schema='%s'\
AND table_name ='%s'\
AND table_type='BASE TABLE'",
config_log_schema,
objects->table_name
);
ret = SPI_execute(buf.data, true, 0);
if (ret != SPI_OK_SELECT)
{
ereport(FATAL,
(errmsg("SPI_execute failed: SPI error code %d", ret)
));
}
/* This should never happen */
if (SPI_processed != 1)
{
elog(FATAL, "not a singleton result");
}
ntup = DatumGetInt32(SPI_getbinval(SPI_tuptable->vals[0],
SPI_tuptable->tupdesc,
1, &isnull));
/* This should never happen */
if (isnull)
{
elog(FATAL, "null result");
}
if (ntup == 0)
{
ereport(FATAL,
(
errmsg("Expected config log table '%s.%s' not found", config_log_schema,
objects->table_name),
errhint("Ensure superuser search_path includes the schema used by config_log; "
"check config_log.* GUC settings")
));
}
/* check function pg_settings_logger() exists */
resetStringInfo(&buf);
appendStringInfo(
&buf,
"SELECT COUNT(*) FROM pg_catalog.pg_proc p \
INNER JOIN pg_catalog.pg_namespace n ON n.oid = p.pronamespace \
WHERE p.proname='%s' \
AND n.nspname='%s' \
AND p.pronargs = 0",
objects->function_name,
config_log_schema
);
ret = SPI_execute(buf.data, true, 0);
if (ret != SPI_OK_SELECT)
{
ereport(FATAL,
(errmsg("SPI_execute failed: SPI error code %d", ret)));
}
if (SPI_processed != 1)
{
elog(FATAL, "not a singleton result");
}
ntup = DatumGetInt32(SPI_getbinval(SPI_tuptable->vals[0],
SPI_tuptable->tupdesc,
1, &isnull));
if (isnull)
{
elog(FATAL, "null result");
}
if (ntup == 0)
{
ereport(FATAL,
(
errmsg("Expected config log function '%s.%s' not found", config_log_schema,
objects->function_name),
errhint("Ensure superuser search_path includes the schema used by config_log; "
"check config_log.* GUC settings")
));
}
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
log_info("initialized, database objects validated");
/* execute pg_settings_logger() here to catch any settings which have changed after server restart */
execute_pg_settings_logger(objects);
return objects;
}
static void
SaveBuffers(void)
{
int i;
int num_buffers;
int log_level = DEBUG3;
SavedBuffer *saved_buffers;
volatile BufferDesc *bufHdr; // XXX: Do we really need volatile here?
FILE *file = NULL;
int database_counter= 0;
Oid prev_database = InvalidOid;
Oid prev_filenode = InvalidOid;
ForkNumber prev_forknum = InvalidForkNumber;
BlockNumber prev_blocknum = InvalidBlockNumber;
BlockNumber range_counter = 0;
const char *savefile_path;
/*
* XXX: If the memory request fails, ask for a smaller memory chunk, and use
* it to create chunks of save-files, and make the workers read those chunks.
*
* This is not a concern as of now, so deferred; there's at least one other
* place that allocates (NBuffers * (much_bigger_struct)), so this seems to
* be an acceptable practice.
*/
saved_buffers = (SavedBuffer *) palloc(sizeof(SavedBuffer) * NBuffers);
/* Lock the buffer partitions for reading. */
for (i = 0; i < NUM_BUFFER_PARTITIONS; ++i)
LWLockAcquire(FirstBufMappingLock + i, LW_SHARED);
/* Scan and save a list of valid buffers. */
for (num_buffers = 0, i = 0, bufHdr = BufferDescriptors; i < NBuffers; ++i, ++bufHdr)
{
/* Lock each buffer header before inspecting. */
LockBufHdr(bufHdr);
/* Skip invalid buffers */
if ((bufHdr->flags & BM_VALID) && (bufHdr->flags & BM_TAG_VALID))
{
saved_buffers[num_buffers].database = bufHdr->tag.rnode.dbNode;
saved_buffers[num_buffers].filenode = bufHdr->tag.rnode.relNode;
saved_buffers[num_buffers].forknum = bufHdr->tag.forkNum;
saved_buffers[num_buffers].blocknum = bufHdr->tag.blockNum;
++num_buffers;
}
UnlockBufHdr(bufHdr);
}
/* Unlock the buffer partitions in reverse order, to avoid a deadlock. */
for (i = NUM_BUFFER_PARTITIONS - 1; i >= 0; --i)
LWLockRelease(FirstBufMappingLock + i);
/*
* Sort the list, so that we can optimize the storage of these buffers.
*
* The side-effect of this storage optimization is that when reading the
* blocks back from relation forks, it leads to sequential reads, which
* improve the restore speeds quite considerably as compared to random reads
* from different blocks all over the data directory.
*/
pg_qsort(saved_buffers, num_buffers, sizeof(SavedBuffer), SavedBufferCmp);
/* Connect to the database and start a transaction for database name lookups. */
BackgroundWorkerInitializeConnection(guc_default_database, NULL);
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, "saving buffers");
for (i = 0; i < num_buffers; ++i)
{
int j;
SavedBuffer *buf = &saved_buffers[i];
if (i == 0)
{
/*
* Special case for global objects. The sort brings them to the
* front of the list.
*/
/* Make sure the first buffer we save belongs to global object. */
Assert(buf->database == InvalidOid);
/*
* Database number (and save-file name) 1 is reserverd for storing
* list of buffers of global objects.
*/
database_counter = 1;
savefile_path = getSavefileName(database_counter);
file = fileOpen(savefile_path, PG_BINARY_W);
writeDBName("", file, savefile_path);
prev_database = buf->database;
}
if (buf->database != prev_database)
{
char *dbname;
/*
* We are beginning to process a different database than the
* previous one; close the save-file of previous database, and open
* a new one.
*/
++database_counter;
dbname = get_database_name(buf->database);
Assert(dbname != NULL);
if (file != NULL)
fileClose(file, savefile_path);
savefile_path = getSavefileName(database_counter);
file = fileOpen(savefile_path, PG_BINARY_W);
writeDBName(dbname, file, savefile_path);
pfree(dbname);
/* Reset trackers appropriately */
prev_database = buf->database;
prev_filenode = InvalidOid;
prev_forknum = InvalidForkNumber;
prev_blocknum = InvalidBlockNumber;
range_counter = 0;
}
if (buf->filenode != prev_filenode)
{
/* We're beginning to process a new relation; emit a record for it. */
fileWrite("r", 1, file, savefile_path);
fileWrite(&(buf->filenode), sizeof(Oid), file, savefile_path);
/* Reset trackers appropriately */
prev_filenode = buf->filenode;
prev_forknum = InvalidForkNumber;
prev_blocknum = InvalidBlockNumber;
range_counter = 0;
}
if (buf->forknum != prev_forknum)
{
/*
* We're beginning to process a new fork of this relation; add a
* record for it.
*/
fileWrite("f", 1, file, savefile_path);
fileWrite(&(buf->forknum), sizeof(ForkNumber), file, savefile_path);
/* Reset trackers appropriately */
prev_forknum = buf->forknum;
prev_blocknum = InvalidBlockNumber;
range_counter = 0;
}
ereport(log_level,
(errmsg("writer: writing block db %d filenode %d forknum %d blocknum %d",
database_counter, prev_filenode, prev_forknum, buf->blocknum)));
fileWrite("b", 1, file, savefile_path);
fileWrite(&(buf->blocknum), sizeof(BlockNumber), file, savefile_path);
prev_blocknum = buf->blocknum;
/*
* If a continuous range of blocks follows this block, then emit one
* entry for the range, instead of one for each block.
*/
range_counter = 0;
for ( j = i+1; j < num_buffers; ++j)
{
SavedBuffer *tmp = &saved_buffers[j];
if (tmp->database == prev_database
&& tmp->filenode == prev_filenode
&& tmp->forknum == prev_forknum
&& tmp->blocknum == (prev_blocknum + range_counter + 1))
{
++range_counter;
}
}
if (range_counter != 0)
{
ereport(log_level,
(errmsg("writer: writing range db %d filenode %d forknum %d blocknum %d range %d",
database_counter, prev_filenode, prev_forknum, prev_blocknum, range_counter)));
fileWrite("N", 1, file, savefile_path);
fileWrite(&range_counter, sizeof(range_counter), file, savefile_path);
i += range_counter;
}
}
ereport(LOG,
(errmsg("Buffer Saver: saved metadata of %d blocks", num_buffers)));
Assert(file != NULL);
fileClose(file, savefile_path);
pfree(saved_buffers);
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
}
static void
ReadBlocks(int filenum)
{
FILE *file;
char record_type;
char *dbname;
Oid record_filenode;
ForkNumber record_forknum;
BlockNumber record_blocknum;
BlockNumber record_range;
int log_level = DEBUG3;
Oid relOid = InvalidOid;
Relation rel = NULL;
bool skip_relation = false;
bool skip_fork = false;
bool skip_block = false;
BlockNumber nblocks = 0;
BlockNumber blocks_restored = 0;
const char *filepath;
/*
* If this condition changes, then this code, and the code in the writer
* will need to be changed; especially the format specifiers in log and
* error messages.
*/
StaticAssertStmt(MaxBlockNumber == 0xFFFFFFFE, "Code may need review.");
filepath = getSavefileName(filenum);
file = fileOpen(filepath, PG_BINARY_R);
dbname = readDBName(file, filepath);
/*
* When restoring global objects, the dbname is zero-length string, and non-
* zero length otherwise. And filenum is never expected to be smaller than 1.
*/
Assert(filenum >= 1);
Assert(filenum == 1 ? strlen(dbname) == 0 : strlen(dbname) > 0);
/* To restore the global objects, use default database */
BackgroundWorkerInitializeConnection(filenum == 1 ? guc_default_database : dbname, NULL);
SetCurrentStatementStartTimestamp();
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
pgstat_report_activity(STATE_RUNNING, "restoring buffers");
/*
* Note that in case of a read error, we will leak relcache entry that we may
* currently have open. In case of EOF, we close the relation after the loop.
*/
while (fileRead(&record_type, 1, file, true, filepath))
{
/*
* If we want to process the signals, this seems to be the best place
* to do it. Generally the backends refrain from processing config file
* while in transaction, but that's more for the fear of allowing GUC
* changes to affect expression evaluation, causing different results
* for the same expression in a transaction. Since this worker is not
* processing any queries, it is okay to process the config file here.
*
* Even though it's okay to process SIGHUP here, doing so doesn't add
* any value. The only reason we might want to process config file here
* would be to allow the user to interrupt the BlockReader's operation
* by changing this extenstion's GUC parameter. But the user can do that
* anyway, using SIGTERM or pg_terminate_backend().
*/
/* Stop processing the save-file if the Postmaster wants us to die. */
if (got_sigterm)
break;
ereport(log_level,
(errmsg("record type %x - %c", record_type, record_type)));
switch (record_type)
{
case 'r':
{
/* Close the previous relation, if any. */
if (rel)
{
relation_close(rel, AccessShareLock);
rel = NULL;
}
record_forknum = InvalidForkNumber;
record_blocknum = InvalidBlockNumber;
nblocks = 0;
fileRead(&record_filenode, sizeof(Oid), file, false, filepath);
relOid = GetRelOid(record_filenode);
ereport(log_level, (errmsg("processing filenode %u, relation %u",
record_filenode, relOid)));
/*
* If the relation has been rewritten/dropped since we saved it,
* just skip it and process the next relation.
*/
if (relOid == InvalidOid)
skip_relation = true;
else
{
skip_relation = false;
/* Open the relation */
rel = relation_open(relOid, AccessShareLock);
RelationOpenSmgr(rel);
}
}
break;
case 'f':
{
record_blocknum = InvalidBlockNumber;
nblocks = 0;
fileRead(&record_forknum, sizeof(ForkNumber), file, false, filepath);
if (skip_relation)
continue;
if (rel == NULL)
ereport(ERROR,
(errmsg("found a fork record without a preceeding relation record")));
ereport(log_level, (errmsg("processing fork %d", record_forknum)));
if (!smgrexists(rel->rd_smgr, record_forknum))
skip_fork = true;
else
{
skip_fork = false;
nblocks = RelationGetNumberOfBlocksInFork(rel, record_forknum);
}
}
break;
case 'b':
{
if (record_forknum == InvalidForkNumber)
ereport(ERROR,
(errmsg("found a block record without a preceeding fork record")));
fileRead(&record_blocknum, sizeof(BlockNumber), file, false, filepath);
if (skip_relation || skip_fork)
continue;
/*
* Don't try to read past the file; the file may have been shrunk
* by a vaccum/truncate operation.
*/
if (record_blocknum >= nblocks)
{
ereport(log_level,
(errmsg("reader %d skipping block filenode %u forknum %d blocknum %u",
filenum, record_filenode, record_forknum, record_blocknum)));
skip_block = true;
continue;
}
else
{
Buffer buf;
skip_block = false;
ereport(log_level,
(errmsg("reader %d reading block filenode %u forknum %d blocknum %u",
filenum, record_filenode, record_forknum, record_blocknum)));
buf = ReadBufferExtended(rel, record_forknum, record_blocknum, RBM_NORMAL, NULL);
ReleaseBuffer(buf);
++blocks_restored;
}
}
break;
case 'N':
{
BlockNumber block;
Assert(record_blocknum != InvalidBlockNumber);
if (record_blocknum == InvalidBlockNumber)
ereport(ERROR,
(errmsg("found a block range record without a preceeding block record")));
fileRead(&record_range, sizeof(int), file, false, filepath);
if (skip_relation || skip_fork || skip_block)
continue;
ereport(log_level,
(errmsg("reader %d reading range filenode %u forknum %d blocknum %u range %u",
filenum, record_filenode, record_forknum, record_blocknum, record_range)));
for (block = record_blocknum + 1; block <= (record_blocknum + record_range); ++block)
{
Buffer buf;
/*
* Don't try to read past the file; the file may have been
* shrunk by a vaccum operation.
*/
if (block >= nblocks)
{
ereport(log_level,
(errmsg("reader %d skipping block range filenode %u forknum %d start %u end %u",
filenum, record_filenode, record_forknum,
block, record_blocknum + record_range)));
break;
}
buf = ReadBufferExtended(rel, record_forknum, block, RBM_NORMAL, NULL);
ReleaseBuffer(buf);
++blocks_restored;
}
}
break;
default:
{
ereport(ERROR,
(errmsg("found unexpected save-file marker %x - %c)", record_type, record_type)));
Assert(false);
}
break;
}
}
if (rel)
relation_close(rel, AccessShareLock);
ereport(LOG,
(errmsg("Block Reader %d: restored %u blocks",
filenum, blocks_restored)));
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
pgstat_report_activity(STATE_IDLE, NULL);
fileClose(file, filepath);
/* Remove the save-file */
if (remove(filepath) != 0)
ereport(ERROR,
(errcode_for_file_access(),
errmsg("error removing file \"%s\" : %m", filepath)));
}
