/*
* Execute the CREATE BARRIER command. Write a BARRIER WAL record and flush the
* WAL buffers to disk before returning to the caller. Writing the WAL record
* does not guarantee successful completion of the barrier command.
*/
void
ProcessCreateBarrierExecute(const char *id)
{
StringInfoData buf;
if (!IsConnFromCoord())
ereport(ERROR,
(errcode(ERRCODE_INTERNAL_ERROR),
errmsg("The CREATE BARRIER EXECUTE message is expected to "
"arrive from a Coordinator")));
{
XLogRecData rdata[1];
XLogRecPtr recptr;
rdata[0].data = (char *) id;
rdata[0].len = strlen(id) + 1;
rdata[0].buffer = InvalidBuffer;
rdata[0].next = NULL;
recptr = XLogInsert(RM_BARRIER_ID, XLOG_BARRIER_CREATE, rdata);
XLogFlush(recptr);
}
pq_beginmessage(&buf, 'b');
pq_sendstring(&buf, id);
pq_endmessage(&buf);
pq_flush();
}
void
RequestBarrier(const char *id, char *completionTag)
{
PGXCNodeAllHandles *prepared_handles;
const char *barrier_id;
elog(DEBUG2, "CREATE BARRIER request received");
/*
* Ensure that we are a Coordinator and the request is not from another
* coordinator
*/
if (!IS_PGXC_COORDINATOR)
ereport(ERROR,
(errcode(ERRCODE_INTERNAL_ERROR),
errmsg("CREATE BARRIER command must be sent to a Coordinator")));
if (IsConnFromCoord())
ereport(ERROR,
(errcode(ERRCODE_INTERNAL_ERROR),
errmsg("CREATE BARRIER command is not expected from another Coordinator")));
/*
* Get a barrier id if the user has not supplied it
*/
barrier_id = generate_barrier_id(id);
elog(DEBUG2, "CREATE BARRIER <%s>", barrier_id);
/*
* Step One. Prepare all Coordinators for upcoming barrier request
*/
prepared_handles = PrepareBarrier(barrier_id);
/*
* Step two. Issue BARRIER command to all involved components, including
* Coordinators and Datanodes
*/
ExecuteBarrier(barrier_id);
/*
* Step three. Inform Coordinators about a successfully completed barrier
*/
EndBarrier(prepared_handles, barrier_id);
/* Finally report the barrier to GTM to backup its restart point */
ReportBarrierGTM(barrier_id);
/* Free the handles */
pfree_pgxc_all_handles(prepared_handles);
if (completionTag)
snprintf(completionTag, COMPLETION_TAG_BUFSIZE, "BARRIER %s", barrier_id);
}
/*
* GetActiveSnapshot
* Return the topmost snapshot in the Active stack.
*/
Snapshot
GetActiveSnapshot(void)
{
#ifdef PGXC
/*
* Check if topmost snapshot is null or not,
* if it is, a new one will be taken from GTM.
*/
if (!ActiveSnapshot && IS_PGXC_COORDINATOR && !IsConnFromCoord())
return NULL;
#endif
Assert(ActiveSnapshot != NULL);
return ActiveSnapshot->as_snap;
}
/*
* Mark the completion of an on-going barrier. We must have remembered the
* barrier ID when we received the CREATE BARRIER PREPARE command
*/
void
ProcessCreateBarrierEnd(const char *id)
{
StringInfoData buf;
if (!IS_PGXC_COORDINATOR || !IsConnFromCoord())
ereport(ERROR,
(errcode(ERRCODE_INTERNAL_ERROR),
errmsg("The CREATE BARRIER END message is expected to "
"arrive at a Coordinator from another Coordinator")));
LWLockRelease(BarrierLock);
pq_beginmessage(&buf, 'b');
pq_sendstring(&buf, id);
pq_endmessage(&buf);
pq_flush();
/*
* TODO Stop the timer
*/
}
/*
* Prepare ourselves for an incoming BARRIER. We must disable all new 2PC
* commits and let the ongoing commits to finish. We then remember the
* barrier id (so that it can be matched with the final END message) and
* tell the driving Coordinator to proceed with the next step.
*
* A simple way to implement this is to grab a lock in an exclusive mode
* while all other backend starting a 2PC will grab the lock in shared
* mode. So as long as we hold the exclusive lock, no other backend start a
* new 2PC and there can not be any 2PC in-progress. This technique would
* rely on assumption that an exclusive lock requester is not starved by
* share lock requesters.
*
* Note: To ensure that the 2PC are not blocked for a long time, we should
* set a timeout. The lock should be release after the timeout and the
* barrier should be canceled.
*/
void
ProcessCreateBarrierPrepare(const char *id)
{
StringInfoData buf;
if (!IS_PGXC_COORDINATOR || !IsConnFromCoord())
ereport(ERROR,
(errcode(ERRCODE_INTERNAL_ERROR),
errmsg("The CREATE BARRIER PREPARE message is expected to "
"arrive at a Coordinator from another Coordinator")));
LWLockAcquire(BarrierLock, LW_EXCLUSIVE);
pq_beginmessage(&buf, 'b');
pq_sendstring(&buf, id);
pq_endmessage(&buf);
pq_flush();
/*
* TODO Start a timer to terminate the pending barrier after a specified
* timeout
*/
}
/*
* GetTransactionSnapshot
* Get the appropriate snapshot for a new query in a transaction.
*
* Note that the return value may point at static storage that will be modified
* by future calls and by CommandCounterIncrement(). Callers should call
* RegisterSnapshot or PushActiveSnapshot on the returned snap if it is to be
* used very long.
*/
Snapshot
GetTransactionSnapshot(void)
{
/* First call in transaction? */
if (!FirstSnapshotSet)
{
Assert(RegisteredSnapshots == 0);
Assert(FirstXactSnapshot == NULL);
/*
* In transaction-snapshot mode, the first snapshot must live until
* end of xact regardless of what the caller does with it, so we must
* make a copy of it rather than returning CurrentSnapshotData
* directly. Furthermore, if we're running in serializable mode,
* predicate.c needs to wrap the snapshot fetch in its own processing.
*/
if (IsolationUsesXactSnapshot())
{
/* First, create the snapshot in CurrentSnapshotData */
if (IsolationIsSerializable())
CurrentSnapshot = GetSerializableTransactionSnapshot(&CurrentSnapshotData);
else
CurrentSnapshot = GetSnapshotData(&CurrentSnapshotData);
/* Make a saved copy */
CurrentSnapshot = CopySnapshot(CurrentSnapshot);
FirstXactSnapshot = CurrentSnapshot;
/* Mark it as "registered" in FirstXactSnapshot */
FirstXactSnapshot->regd_count++;
RegisteredSnapshots++;
}
else
CurrentSnapshot = GetSnapshotData(&CurrentSnapshotData);
FirstSnapshotSet = true;
return CurrentSnapshot;
}
if (IsolationUsesXactSnapshot())
{
#ifdef PGXC
/*
* Consider this test case taken from portals.sql
*
* CREATE TABLE cursor (a int, b int) distribute by replication;
* INSERT INTO cursor VALUES (10);
* BEGIN;
* SET TRANSACTION ISOLATION LEVEL SERIALIZABLE;
* DECLARE c1 NO SCROLL CURSOR FOR SELECT * FROM cursor FOR UPDATE;
* INSERT INTO cursor VALUES (2);
* FETCH ALL FROM c1;
* would result in
* ERROR: attempted to lock invisible tuple
* because FETCH would be sent as a select to the remote nodes
* with command id 0, whereas the command id would be 2
* in the current snapshot.
* (1 sent by Coordinator due to declare cursor &
* 2 because of the insert inside the transaction)
* The command id should therefore be updated in the
* current snapshot.
*/
if (IsConnFromCoord())
SnapshotSetCommandId(GetCurrentCommandId(false));
#endif
return CurrentSnapshot;
}
CurrentSnapshot = GetSnapshotData(&CurrentSnapshotData);
return CurrentSnapshot;
}