void
FtsCondSetTxnReadOnly(bool *XactFlag)
{
if (!isFTSEnabled())
return;
if (*ftsReadOnlyFlag && Gp_role != GP_ROLE_UTILITY)
*XactFlag = true;
}
/*
* Creates a new gang by logging on a session to each segDB involved.
*
* call this function in GangContext memory context.
* elog ERROR or return a non-NULL gang.
*/
static Gang*
createGang_async(GangType type, int gang_id, int size, int content)
{
Gang *newGangDefinition;
SegmentDatabaseDescriptor *segdbDesc = NULL;
int i = 0;
int create_gang_retry_counter = 0;
int in_recovery_mode_count = 0;
int successful_connections = 0;
bool retry = false;
int poll_timeout = 0;
struct timeval startTS;
PostgresPollingStatusType *pollingStatus = NULL;
/* true means connection status is confirmed, either established or in recovery mode */
bool *connStatusDone = NULL;
ELOG_DISPATCHER_DEBUG("createGang type = %d, gang_id = %d, size = %d, content = %d",
type, gang_id, size, content);
/* check arguments */
Assert(size == 1 || size == getgpsegmentCount());
Assert(CurrentResourceOwner != NULL);
Assert(CurrentMemoryContext == GangContext);
/* Writer gang is created before reader gangs. */
if (type == GANGTYPE_PRIMARY_WRITER)
Insist(!GangsExist());
/* Check writer gang firstly*/
if (type != GANGTYPE_PRIMARY_WRITER && !isPrimaryWriterGangAlive())
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("writer gang got broken before creating reader gangs")));
create_gang_retry:
/* If we're in a retry, we may need to reset our initial state, a bit */
newGangDefinition = NULL;
successful_connections = 0;
in_recovery_mode_count = 0;
retry = false;
/* allocate and initialize a gang structure */
newGangDefinition = buildGangDefinition(type, gang_id, size, content);
Assert(newGangDefinition != NULL);
Assert(newGangDefinition->size == size);
Assert(newGangDefinition->perGangContext != NULL);
MemoryContextSwitchTo(newGangDefinition->perGangContext);
/* allocate memory within perGangContext and will be freed automatically when gang is destroyed */
pollingStatus = palloc(sizeof(PostgresPollingStatusType) * size);
connStatusDone = palloc(sizeof(bool) * size);
struct pollfd *fds;
PG_TRY();
{
for (i = 0; i < size; i++)
{
char gpqeid[100];
char *options;
/*
* Create the connection requests. If we find a segment without a
* valid segdb we error out. Also, if this segdb is invalid, we must
* fail the connection.
*/
segdbDesc = &newGangDefinition->db_descriptors[i];
/*
* Build the connection string. Writer-ness needs to be processed
* early enough now some locks are taken before command line options
* are recognized.
*/
build_gpqeid_param(gpqeid, sizeof(gpqeid),
segdbDesc->segindex,
type == GANGTYPE_PRIMARY_WRITER,
gang_id);
options = makeOptions();
/* start connection in asynchronous way */
cdbconn_doConnectStart(segdbDesc, gpqeid, options);
if(cdbconn_isBadConnection(segdbDesc))
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("%s (%s)", PQerrorMessage(segdbDesc->conn), segdbDesc->whoami)));
connStatusDone[i] = false;
/*
* If connection status is not CONNECTION_BAD after PQconnectStart(), we must
* act as if the PQconnectPoll() had returned PGRES_POLLING_WRITING
*/
pollingStatus[i] = PGRES_POLLING_WRITING;
}
/*
* Ok, we've now launched all the connection attempts. Start the
* timeout clock (= get the start timestamp), and poll until they're
* all completed or we reach timeout.
*/
gettimeofday(&startTS, NULL);
fds = (struct pollfd *) palloc0(sizeof(struct pollfd) * size);
for(;;)
{
int nready;
int nfds = 0;
poll_timeout = getPollTimeout(&startTS);
for (i = 0; i < size; i++)
{
segdbDesc = &newGangDefinition->db_descriptors[i];
/* Skip established connections and in-recovery-mode connections*/
if (connStatusDone[i])
continue;
switch (pollingStatus[i])
{
case PGRES_POLLING_OK:
cdbconn_doConnectComplete(segdbDesc);
if (segdbDesc->motionListener == -1 || segdbDesc->motionListener == 0)
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("Internal error: No motion listener port (%s)", segdbDesc->whoami)));
successful_connections++;
connStatusDone[i] = true;
continue;
case PGRES_POLLING_READING:
fds[nfds].fd = PQsocket(segdbDesc->conn);
fds[nfds].events = POLLIN;
nfds++;
break;
case PGRES_POLLING_WRITING:
fds[nfds].fd = PQsocket(segdbDesc->conn);
fds[nfds].events = POLLOUT;
nfds++;
break;
case PGRES_POLLING_FAILED:
if (segment_failure_due_to_recovery(&segdbDesc->conn->errorMessage))
{
in_recovery_mode_count++;
connStatusDone[i] = true;
elog(LOG, "segment is in recovery mode (%s)", segdbDesc->whoami);
}
else
{
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("%s (%s)", PQerrorMessage(segdbDesc->conn), segdbDesc->whoami)));
}
break;
default:
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("unknow pollstatus (%s)", segdbDesc->whoami)));
break;
}
if (poll_timeout == 0)
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("timeout expired\n (%s)", segdbDesc->whoami)));
}
if (nfds == 0)
break;
CHECK_FOR_INTERRUPTS();
/* Wait until something happens */
nready = poll(fds, nfds, poll_timeout);
if (nready < 0)
{
int sock_errno = SOCK_ERRNO;
if (sock_errno == EINTR)
continue;
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("poll() failed: errno = %d", sock_errno)));
}
else if (nready > 0)
{
int currentFdNumber = 0;
for (i = 0; i < size; i++)
{
segdbDesc = &newGangDefinition->db_descriptors[i];
if (connStatusDone[i])
continue;
Assert(PQsocket(segdbDesc->conn) > 0);
Assert(PQsocket(segdbDesc->conn) == fds[currentFdNumber].fd);
if (fds[currentFdNumber].revents & fds[currentFdNumber].events)
pollingStatus[i] = PQconnectPoll(segdbDesc->conn);
currentFdNumber++;
}
}
}
ELOG_DISPATCHER_DEBUG("createGang: %d processes requested; %d successful connections %d in recovery",
size, successful_connections, in_recovery_mode_count);
MemoryContextSwitchTo(GangContext);
/* some segments are in recovery mode*/
if (successful_connections != size)
{
Assert(successful_connections + in_recovery_mode_count == size);
/* FTS shows some segment DBs are down */
if (isFTSEnabled() &&
FtsTestSegmentDBIsDown(newGangDefinition->db_descriptors, size))
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("FTS detected one or more segments are down")));
if ( gp_gang_creation_retry_count <= 0 ||
create_gang_retry_counter++ >= gp_gang_creation_retry_count ||
type != GANGTYPE_PRIMARY_WRITER)
ereport(ERROR, (errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("segments is in recovery mode")));
ELOG_DISPATCHER_DEBUG("createGang: gang creation failed, but retryable.");
DisconnectAndDestroyGang(newGangDefinition);
newGangDefinition = NULL;
retry = true;
}
}
PG_CATCH();
{
MemoryContextSwitchTo(GangContext);
DisconnectAndDestroyGang(newGangDefinition);
newGangDefinition = NULL;
if (type == GANGTYPE_PRIMARY_WRITER)
{
DisconnectAndDestroyAllGangs(true);
CheckForResetSession();
}
PG_RE_THROW();
}
PG_END_TRY();
if (retry)
{
CHECK_FOR_INTERRUPTS();
pg_usleep(gp_gang_creation_retry_timer * 1000);
CHECK_FOR_INTERRUPTS();
goto create_gang_retry;
}
setLargestGangsize(size);
return newGangDefinition;
}
/*
* Creates a new gang by logging on a session to each segDB involved.
*
* call this function in GangContext memory context.
* elog ERROR or return a non-NULL gang.
*/
static Gang *
createGang_thread(GangType type, int gang_id, int size, int content)
{
Gang *newGangDefinition = NULL;
SegmentDatabaseDescriptor *segdbDesc = NULL;
DoConnectParms *doConnectParmsAr = NULL;
DoConnectParms *pParms = NULL;
int parmIndex = 0;
int threadCount = 0;
int i = 0;
int create_gang_retry_counter = 0;
int in_recovery_mode_count = 0;
int successful_connections = 0;
PQExpBufferData create_gang_error;
ELOG_DISPATCHER_DEBUG("createGang type = %d, gang_id = %d, size = %d, content = %d",
type, gang_id, size, content);
/* check arguments */
Assert(size == 1 || size == getgpsegmentCount());
Assert(CurrentResourceOwner != NULL);
Assert(CurrentMemoryContext == GangContext);
Assert(gp_connections_per_thread > 0);
/* Writer gang is created before reader gangs. */
if (type == GANGTYPE_PRIMARY_WRITER)
Insist(!GangsExist());
initPQExpBuffer(&create_gang_error);
Assert(CurrentGangCreating == NULL);
create_gang_retry:
/*
* If we're in a retry, we may need to reset our initial state a bit. We
* also want to ensure that all resources have been released.
*/
Assert(newGangDefinition == NULL);
Assert(doConnectParmsAr == NULL);
successful_connections = 0;
in_recovery_mode_count = 0;
threadCount = 0;
/* allocate and initialize a gang structure */
newGangDefinition = buildGangDefinition(type, gang_id, size, content);
CurrentGangCreating = newGangDefinition;
Assert(newGangDefinition != NULL);
Assert(newGangDefinition->size == size);
Assert(newGangDefinition->perGangContext != NULL);
MemoryContextSwitchTo(newGangDefinition->perGangContext);
resetPQExpBuffer(&create_gang_error);
/*
* The most threads we could have is segdb_count /
* gp_connections_per_thread, rounded up. This is equivalent to 1 +
* (segdb_count-1) / gp_connections_per_thread. We allocate enough memory
* for this many DoConnectParms structures, even though we may not use
* them all.
*/
threadCount = 1 + (size - 1) / gp_connections_per_thread;
Assert(threadCount > 0);
/* initialize connect parameters */
doConnectParmsAr = makeConnectParms(threadCount, type, gang_id);
for (i = 0; i < size; i++)
{
parmIndex = i / gp_connections_per_thread;
pParms = &doConnectParmsAr[parmIndex];
segdbDesc = &newGangDefinition->db_descriptors[i];
pParms->segdbDescPtrArray[pParms->db_count++] = segdbDesc;
}
/* start threads and doing the connect */
for (i = 0; i < threadCount; i++)
{
int pthread_err;
pParms = &doConnectParmsAr[i];
ELOG_DISPATCHER_DEBUG("createGang creating thread %d of %d for libpq connections",
i + 1, threadCount);
pthread_err = gp_pthread_create(&pParms->thread, thread_DoConnect, pParms, "createGang");
if (pthread_err != 0)
{
int j;
/*
* Error during thread create (this should be caused by resource
* constraints). If we leave the threads running, they'll
* immediately have some problems -- so we need to join them, and
* *then* we can issue our FATAL error
*/
for (j = 0; j < i; j++)
{
pthread_join(doConnectParmsAr[j].thread, NULL);
}
ereport(FATAL, (errcode(ERRCODE_INTERNAL_ERROR),
errmsg("failed to create thread %d of %d", i + 1, threadCount),
errdetail("pthread_create() failed with err %d", pthread_err)));
}
}
/*
* wait for all of the DoConnect threads to complete.
*/
for (i = 0; i < threadCount; i++)
{
ELOG_DISPATCHER_DEBUG("joining to thread %d of %d for libpq connections",
i + 1, threadCount);
if (0 != pthread_join(doConnectParmsAr[i].thread, NULL))
{
elog(FATAL, "could not create segworker group");
}
}
/*
* Free the memory allocated for the threadParms array
*/
destroyConnectParms(doConnectParmsAr, threadCount);
doConnectParmsAr = NULL;
SIMPLE_FAULT_INJECTOR(GangCreated);
/* find out the successful connections and the failed ones */
checkConnectionStatus(newGangDefinition, &in_recovery_mode_count,
&successful_connections, &create_gang_error);
ELOG_DISPATCHER_DEBUG("createGang: %d processes requested; %d successful connections %d in recovery",
size, successful_connections, in_recovery_mode_count);
MemoryContextSwitchTo(GangContext);
if (size == successful_connections)
{
setLargestGangsize(size);
termPQExpBuffer(&create_gang_error);
CurrentGangCreating = NULL;
return newGangDefinition;
}
/* there'er failed connections */
/* FTS shows some segment DBs are down, destroy all gangs. */
if (isFTSEnabled() &&
FtsTestSegmentDBIsDown(newGangDefinition->db_descriptors, size))
{
appendPQExpBuffer(&create_gang_error, "FTS detected one or more segments are down\n");
goto exit;
}
/* failure due to recovery */
if (successful_connections + in_recovery_mode_count == size)
{
if (gp_gang_creation_retry_count &&
create_gang_retry_counter++ < gp_gang_creation_retry_count &&
type == GANGTYPE_PRIMARY_WRITER)
{
/*
* Retry for non-writer gangs is meaningless because writer gang
* must be gone when QE is in recovery mode
*/
DisconnectAndDestroyGang(newGangDefinition);
newGangDefinition = NULL;
CurrentGangCreating = NULL;
ELOG_DISPATCHER_DEBUG("createGang: gang creation failed, but retryable.");
CHECK_FOR_INTERRUPTS();
pg_usleep(gp_gang_creation_retry_timer * 1000);
CHECK_FOR_INTERRUPTS();
goto create_gang_retry;
}
appendPQExpBuffer(&create_gang_error, "segment(s) are in recovery mode\n");
}
exit:
if (newGangDefinition != NULL)
DisconnectAndDestroyGang(newGangDefinition);
if (type == GANGTYPE_PRIMARY_WRITER)
{
DisconnectAndDestroyAllGangs(true);
CheckForResetSession();
}
CurrentGangCreating = NULL;
ereport(ERROR,
(errcode(ERRCODE_GP_INTERCONNECTION_ERROR),
errmsg("failed to acquire resources on one or more segments"),
errdetail("%s", create_gang_error.data)));
return NULL;
}
/*
* FtsHandleGangConnectionFailure is called by createGang during
* creating connections return true if error need to be thrown
*/
bool
FtsHandleGangConnectionFailure(SegmentDatabaseDescriptor * segdbDesc, int size)
{
int i;
bool dtx_active;
bool reportError = false;
bool realFaultFound = false;
bool forceRescan=true;
for (i = 0; i < size; i++)
{
if (PQstatus(segdbDesc[i].conn) != CONNECTION_OK)
{
CdbComponentDatabaseInfo *segInfo = segdbDesc[i].segment_database_info;
elog(DEBUG2, "FtsHandleGangConnectionFailure: looking for real fault on segment dbid %d", segInfo->dbid);
if (!FtsTestConnection(segInfo, forceRescan))
{
elog(DEBUG2, "found fault with segment dbid %d", segInfo->dbid);
realFaultFound = true;
/* that at least one fault exists is enough, for now */
break;
}
forceRescan = false; /* only force the rescan on the first call. */
}
}
if (!realFaultFound)
{
/* If we successfully tested the gang and didn't notice a
* failure, our caller must've seen some kind of transient
* failure when the gang was originally constructed ... */
elog(DEBUG2, "FtsHandleGangConnectionFailure: no real fault found!");
return false;
}
if (!isFTSEnabled())
{
return false;
}
ereport(LOG, (errmsg_internal("FTS: reconfiguration is in progress")));
forceRescan = true;
for (i = 0; i < size; i++)
{
CdbComponentDatabaseInfo *segInfo = segdbDesc[i].segment_database_info;
if (PQstatus(segdbDesc[i].conn) != CONNECTION_OK)
{
if (!FtsTestConnection(segInfo, forceRescan))
{
ereport(LOG, (errmsg_internal("FTS: found bad segment with dbid %d", segInfo->dbid),
errSendAlert(true)));
/* probe process has already marked segment down. */
}
forceRescan = false; /* only force rescan on first call. */
}
}
if (gangsExist())
{
reportError = true;
disconnectAndDestroyAllGangs();
}
/*
* KLUDGE: Do not error out if we are attempting a DTM protocol retry
*/
if (DistributedTransactionContext == DTX_CONTEXT_QD_RETRY_PHASE_2)
{
return false;
}
/* is there a transaction active ? */
dtx_active = isCurrentDtxActive();
/* When the error is raised, it will abort the current DTM transaction */
if (dtx_active)
{
elog((Debug_print_full_dtm ? LOG : DEBUG5),
"FtsHandleGangConnectionFailure found an active DTM transaction (returning true).");
return true;
}
/*
* error out if this sets read only flag, at this stage the read only
* transaction checking has passed, so error out, but do not error out if
* tm is in recovery
*/
if ((*ftsReadOnlyFlag && !isTMInRecovery()) || reportError)
return true;
elog((Debug_print_full_dtm ? LOG : DEBUG5),
"FtsHandleGangConnectionFailure returning false.");
return false;
}