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);
}
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 BgwPoolMainLoop(Datum arg)
{
BgwPoolExecutorCtx* ctx = (BgwPoolExecutorCtx*)arg;
int id = ctx->id;
BgwPool* pool = ctx->constructor();
int size;
void* work;
BackgroundWorkerUnblockSignals();
BackgroundWorkerInitializeConnection(pool->dbname, NULL);
while(true) {
PGSemaphoreLock(&pool->available);
SpinLockAcquire(&pool->lock);
size = *(int*)&pool->queue[pool->head];
Assert(size < pool->size);
work = malloc(size);
pool->pending -= 1;
pool->active += 1;
if (pool->lastPeakTime == 0 && pool->active == pool->nWorkers && pool->pending != 0) {
pool->lastPeakTime = MtmGetSystemTime();
}
if (pool->head + size + 4 > pool->size) {
memcpy(work, pool->queue, size);
pool->head = INTALIGN(size);
} else {
memcpy(work, &pool->queue[pool->head+4], size);
pool->head += 4 + INTALIGN(size);
}
if (pool->size == pool->head) {
pool->head = 0;
}
if (pool->producerBlocked) {
pool->producerBlocked = false;
PGSemaphoreUnlock(&pool->overflow);
pool->lastPeakTime = 0;
}
SpinLockRelease(&pool->lock);
pool->executor(id, work, size);
free(work);
SpinLockAcquire(&pool->lock);
pool->active -= 1;
pool->lastPeakTime = 0;
SpinLockRelease(&pool->lock);
}
}
static void
config_log_main(Datum main_arg)
{
config_log_objects *objects;
pqsignal(SIGTERM, config_log_sigterm);
pqsignal(SIGHUP, config_log_sighup);
/* We're now ready to receive signals */
BackgroundWorkerUnblockSignals();
/* Connect to database */
BackgroundWorkerInitializeConnection(config_log_database, NULL);
/* Verify expected objects exist */
objects = initialize_objects();
while (!got_sigterm)
{
int rc;
rc = WaitLatch(&MyProc->procLatch,
WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH,
100000L);
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);
execute_pg_settings_logger(objects);
}
}
proc_exit(0);
}
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);
}
/*
* kafka_consume_main
*
* Main function for Kafka consumers running as background workers
*/
void
kafka_consume_main(Datum arg)
{
char err_msg[512];
rd_kafka_topic_conf_t *topic_conf;
rd_kafka_t *kafka;
rd_kafka_topic_t *topic;
rd_kafka_message_t **messages;
const struct rd_kafka_metadata *meta;
struct rd_kafka_metadata_topic topic_meta;
rd_kafka_resp_err_t err;
bool found;
Oid id = (Oid) arg;
ListCell *lc;
KafkaConsumerProc *proc = hash_search(consumer_procs, &id, HASH_FIND, &found);
KafkaConsumer consumer;
CopyStmt *copy;
int valid_brokers = 0;
int i;
int my_partitions = 0;
if (!found)
elog(ERROR, "kafka consumer %d not found", id);
pqsignal(SIGTERM, kafka_consume_main_sigterm);
#define BACKTRACE_SEGFAULTS
#ifdef BACKTRACE_SEGFAULTS
pqsignal(SIGSEGV, debug_segfault);
#endif
/* we're now ready to receive signals */
BackgroundWorkerUnblockSignals();
/* give this proc access to the database */
BackgroundWorkerInitializeConnection(NameStr(proc->dbname), NULL);
/* load saved consumer state */
StartTransactionCommand();
load_consumer_state(proc->consumer_id, &consumer);
copy = get_copy_statement(&consumer);
topic_conf = rd_kafka_topic_conf_new();
kafka = rd_kafka_new(RD_KAFKA_CONSUMER, NULL, err_msg, sizeof(err_msg));
rd_kafka_set_logger(kafka, logger);
/*
* Add all brokers currently in pipeline_kafka_brokers
*/
if (consumer.brokers == NIL)
elog(ERROR, "no valid brokers were found");
foreach(lc, consumer.brokers)
valid_brokers += rd_kafka_brokers_add(kafka, lfirst(lc));
if (!valid_brokers)
elog(ERROR, "no valid brokers were found");
/*
* Set up our topic to read from
*/
topic = rd_kafka_topic_new(kafka, consumer.topic, topic_conf);
err = rd_kafka_metadata(kafka, false, topic, &meta, CONSUMER_TIMEOUT);
if (err != RD_KAFKA_RESP_ERR_NO_ERROR)
elog(ERROR, "failed to acquire metadata: %s", rd_kafka_err2str(err));
Assert(meta->topic_cnt == 1);
topic_meta = meta->topics[0];
load_consumer_offsets(&consumer, &topic_meta, proc->offset);
CommitTransactionCommand();
/*
* Begin consuming all partitions that this process is responsible for
*/
for (i = 0; i < topic_meta.partition_cnt; i++)
{
int partition = topic_meta.partitions[i].id;
Assert(partition <= consumer.num_partitions);
if (partition % consumer.parallelism != proc->partition_group)
continue;
elog(LOG, "[kafka consumer] %s <- %s consuming partition %d from offset %ld",
consumer.rel->relname, consumer.topic, partition, consumer.offsets[partition]);
if (rd_kafka_consume_start(topic, partition, consumer.offsets[partition]) == -1)
elog(ERROR, "failed to start consuming: %s", rd_kafka_err2str(rd_kafka_errno2err(errno)));
my_partitions++;
}
/*
* No point doing anything if we don't have any partitions assigned to us
*/
if (my_partitions == 0)
{
elog(LOG, "[kafka consumer] %s <- %s consumer %d doesn't have any partitions to read from",
consumer.rel->relname, consumer.topic, MyProcPid);
goto done;
}
messages = palloc0(sizeof(rd_kafka_message_t) * consumer.batch_size);
/*
* Consume messages until we are terminated
*/
while (!got_sigterm)
{
ssize_t num_consumed;
int i;
int messages_buffered = 0;
int partition;
StringInfoData buf;
bool xact = false;
for (partition = 0; partition < consumer.num_partitions; partition++)
{
if (partition % consumer.parallelism != proc->partition_group)
continue;
num_consumed = rd_kafka_consume_batch(topic, partition,
CONSUMER_TIMEOUT, messages, consumer.batch_size);
if (num_consumed <= 0)
continue;
if (!xact)
{
StartTransactionCommand();
xact = true;
}
initStringInfo(&buf);
for (i = 0; i < num_consumed; i++)
{
if (messages[i]->payload != NULL)
{
appendBinaryStringInfo(&buf, messages[i]->payload, messages[i]->len);
if (buf.len > 0 && buf.data[buf.len - 1] != '\n')
appendStringInfoChar(&buf, '\n');
messages_buffered++;
}
consumer.offsets[partition] = messages[i]->offset;
rd_kafka_message_destroy(messages[i]);
}
}
if (!xact)
{
pg_usleep(1 * 1000);
continue;
}
/* we don't want to die in the event of any errors */
PG_TRY();
{
if (messages_buffered)
execute_copy(copy, &buf);
}
PG_CATCH();
{
elog(LOG, "[kafka consumer] %s <- %s failed to process batch, dropped %d message%s:",
consumer.rel->relname, consumer.topic, (int) num_consumed, (num_consumed == 1 ? "" : "s"));
EmitErrorReport();
FlushErrorState();
AbortCurrentTransaction();
xact = false;
}
PG_END_TRY();
if (!xact)
StartTransactionCommand();
if (messages_buffered)
save_consumer_state(&consumer, proc->partition_group);
CommitTransactionCommand();
}
done:
hash_search(consumer_procs, &id, HASH_REMOVE, NULL);
rd_kafka_topic_destroy(topic);
rd_kafka_destroy(kafka);
rd_kafka_wait_destroyed(CONSUMER_TIMEOUT);
}
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);
}
/*
* PgOctopusWorkerMain is the main entry-point for the background worker
* that performs health checks.
*/
static void
PgOctopusWorkerMain(Datum arg)
{
MemoryContext healthCheckContext = NULL;
/* Establish signal handlers before unblocking signals. */
pqsignal(SIGHUP, pg_octopus_sighup);
pqsignal(SIGINT, SIG_IGN);
pqsignal(SIGTERM, pg_octopus_sigterm);
/* We're now ready to receive signals */
BackgroundWorkerUnblockSignals();
/* Connect to our database */
BackgroundWorkerInitializeConnection("postgres", NULL);
healthCheckContext = AllocSetContextCreate(CurrentMemoryContext,
"Health check context",
ALLOCSET_DEFAULT_MINSIZE,
ALLOCSET_DEFAULT_INITSIZE,
ALLOCSET_DEFAULT_MAXSIZE);
MemoryContextSwitchTo(healthCheckContext);
elog(LOG, "pg_octopus monitor started");
/*
* Main loop: do this until the SIGTERM handler tells us to terminate
*/
while (!got_sigterm)
{
struct timeval currentTime = {0, 0};
struct timeval roundEndTime = {0, 0};
struct timeval timeout = {0, 0};
List *nodeHealthList = NIL;
List *healthCheckList = NIL;
gettimeofday(¤tTime, NULL);
roundEndTime = AddTimeMillis(currentTime, HealthCheckPeriod);
nodeHealthList = LoadNodeHealthList();
healthCheckList = CreateHealthChecks(nodeHealthList);
DoHealthChecks(healthCheckList);
MemoryContextReset(healthCheckContext);
gettimeofday(¤tTime, NULL);
timeout = SubtractTimes(roundEndTime, currentTime);
if (timeout.tv_sec >= 0 && timeout.tv_usec >= 0)
{
LatchWait(timeout);
}
if (got_sighup)
{
got_sighup = false;
ProcessConfigFile(PGC_SIGHUP);
}
}
elog(LOG, "pg_octopus monitor exiting");
proc_exit(0);
}
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);
}
static void
worker_spi_main(Datum main_arg)
{
/* Register functions for SIGTERM/SIGHUP management */
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);
while (!got_sigterm)
{
int ret;
int rc;
StringInfoData buf;
/*
* 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,
1000L);
ResetLatch(&MyProc->procLatch);
/* emergency bailout if postmaster has died */
if (rc & WL_POSTMASTER_DEATH)
proc_exit(1);
StartTransactionCommand();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
initStringInfo(&buf);
/* Build the query string */
appendStringInfo(&buf,
"SELECT count(*) FROM pg_class;");
ret = SPI_execute(buf.data, true, 0);
/* Some error messages in case of incorrect handling */
if (ret != SPI_OK_SELECT)
elog(FATAL, "SPI_execute failed: error code %d", ret);
if (SPI_processed > 0)
{
int32 count;
bool isnull;
count = DatumGetInt32(SPI_getbinval(SPI_tuptable->vals[0],
SPI_tuptable->tupdesc,
1, &isnull));
elog(LOG, "Currently %d relations in database",
count);
}
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
}
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);
}
/*
* Entry point for pg_keeper.
*/
void
KeeperMain(Datum main_arg)
{
int ret;
/* Sanity check */
checkParameter();
/* Initial setting */
KeeperMaster = keeper_node1_conninfo;
KeeperStandby = keeper_node2_conninfo;
/* Determine keeper mode of itself */
current_mode = RecoveryInProgress() ? KEEPER_STANDBY_MODE : KEEPER_MASTER_MODE;
/* Establish signal handlers before unblocking signals */
pqsignal(SIGHUP, pg_keeper_sighup);
pqsignal(SIGTERM, pg_keeper_sigterm);
/* We're now ready to receive signals */
BackgroundWorkerUnblockSignals();
/* Connect to our database */
BackgroundWorkerInitializeConnection("postgres", NULL);
exec:
if (current_mode == KEEPER_MASTER_MODE)
{
/* Routine for master_mode */
setupKeeperMaster();
ret = KeeperMainMaster();
}
else if (current_mode == KEEPER_STANDBY_MODE)
{
/* Routine for standby_mode */
setupKeeperStandby();
ret = KeeperMainStandby();
/*
* After promoting is sucessfully done, attempt to re-execute
* main routine as master mode in order to avoid to restart
* for invoking pg_keeper process again.
*/
if (ret)
{
/* Change mode to master mode */
current_mode = KEEPER_MASTER_MODE;
/* Switch master and standby connection information */
swtichMasterAndStandby();
ereport(LOG,
(errmsg("swtiched master and standby informations"),
errdetail("\"%s\" is regarded as master server, \"%s\" is regarded as standby server",
KeeperMaster, KeeperStandby)));
goto exec;
}
}
else
ereport(ERROR, (errmsg("invalid keeper mode : \"%d\"", current_mode)));
proc_exit(ret);
}
static void powa_main(Datum main_arg)
{
char *q1 = "SELECT powa_take_snapshot()";
static char *q2 = "SET application_name = 'POWA collector'";
instr_time begin;
instr_time end;
long time_to_wait;
die_on_too_small_frequency();
/*
Set up signal handlers, then unblock signalsl
*/
pqsignal(SIGHUP, powa_sighup);
pqsignal(SIGTERM, powa_sigterm);
BackgroundWorkerUnblockSignals();
/*
We only connect when powa_frequency >0. If not, powa has been deactivated
*/
if (powa_frequency < 0)
{
elog(LOG, "POWA is deactivated (powa.frequency = %i), exiting",
powa_frequency);
exit(1);
}
// We got here: it means powa_frequency > 0. Let's connect
/*
Connect to POWA database
*/
BackgroundWorkerInitializeConnection(powa_database, NULL);
elog(LOG, "POWA connected to %s", powa_database);
StartTransactionCommand();
SetCurrentStatementStartTimestamp();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
SPI_execute(q2, false, 0);
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
/*
let's store the current time. It will be used to
calculate a quite stable interval between each measure
*/
while (!got_sigterm)
{
/*
We can get here with a new value of powa_frequency
because of a reload. Let's suicide to disconnect
if this value is <0
*/
if (powa_frequency < 0)
{
elog(LOG, "POWA exits to disconnect from the database now");
exit(1);
}
INSTR_TIME_SET_CURRENT(begin);
ResetLatch(&MyProc->procLatch);
StartTransactionCommand();
SetCurrentStatementStartTimestamp();
SPI_connect();
PushActiveSnapshot(GetTransactionSnapshot());
SPI_execute(q1, false, 0);
SPI_finish();
PopActiveSnapshot();
CommitTransactionCommand();
INSTR_TIME_SET_CURRENT(end);
INSTR_TIME_SUBTRACT(end, begin);
/*
Wait powa.frequency, compensate for work time of last snapshot
*/
/*
If we got off schedule (because of a compact or delete,
just do another operation right now
*/
time_to_wait = powa_frequency - INSTR_TIME_GET_MILLISEC(end);
if (time_to_wait > 0)
{
WaitLatch(&MyProc->procLatch,
WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH,
time_to_wait);
}
}
proc_exit(0);
}
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)));
}
