citus/src/backend/distributed/transaction/transaction_recovery.c

556 lines
17 KiB
C

/*-------------------------------------------------------------------------
*
* transaction_recovery.c
*
* Routines for recovering two-phase commits started by this node if a
* failure occurs between prepare and commit/abort.
*
* Copyright (c) Citus Data, Inc.
*
* $Id$
*
*-------------------------------------------------------------------------
*/
#include "postgres.h"
#include "pg_version_constants.h"
#include "miscadmin.h"
#include "libpq-fe.h"
#include <sys/stat.h>
#include <unistd.h>
#include "access/genam.h"
#include "access/heapam.h"
#include "access/htup_details.h"
#include "access/relscan.h"
#include "access/xact.h"
#include "catalog/indexing.h"
#include "distributed/backend_data.h"
#include "distributed/connection_management.h"
#include "distributed/listutils.h"
#include "distributed/metadata_cache.h"
#include "distributed/pg_dist_transaction.h"
#include "distributed/remote_commands.h"
#include "distributed/resource_lock.h"
#include "distributed/transaction_recovery.h"
#include "distributed/worker_manager.h"
#include "distributed/version_compat.h"
#include "lib/stringinfo.h"
#include "storage/lmgr.h"
#include "storage/lock.h"
#include "utils/builtins.h"
#include "utils/fmgroids.h"
#include "utils/memutils.h"
#include "utils/rel.h"
/* exports for SQL callable functions */
PG_FUNCTION_INFO_V1(recover_prepared_transactions);
/* Local functions forward declarations */
static int RecoverWorkerTransactions(WorkerNode *workerNode);
static List * PendingWorkerTransactionList(MultiConnection *connection);
static bool IsTransactionInProgress(HTAB *activeTransactionNumberSet,
char *preparedTransactionName);
static bool RecoverPreparedTransactionOnWorker(MultiConnection *connection,
char *transactionName, bool shouldCommit);
/*
* recover_prepared_transactions recovers any pending prepared
* transactions started by this node on other nodes.
*/
Datum
recover_prepared_transactions(PG_FUNCTION_ARGS)
{
CheckCitusVersion(ERROR);
int recoveredTransactionCount = RecoverTwoPhaseCommits();
PG_RETURN_INT32(recoveredTransactionCount);
}
/*
* LogTransactionRecord registers the fact that a transaction has been
* prepared on a worker. The presence of this record indicates that the
* prepared transaction should be committed.
*/
void
LogTransactionRecord(int32 groupId, char *transactionName)
{
Datum values[Natts_pg_dist_transaction];
bool isNulls[Natts_pg_dist_transaction];
/* form new transaction tuple */
memset(values, 0, sizeof(values));
memset(isNulls, false, sizeof(isNulls));
values[Anum_pg_dist_transaction_groupid - 1] = Int32GetDatum(groupId);
values[Anum_pg_dist_transaction_gid - 1] = CStringGetTextDatum(transactionName);
/* open transaction relation and insert new tuple */
Relation pgDistTransaction = table_open(DistTransactionRelationId(),
RowExclusiveLock);
TupleDesc tupleDescriptor = RelationGetDescr(pgDistTransaction);
HeapTuple heapTuple = heap_form_tuple(tupleDescriptor, values, isNulls);
CatalogTupleInsert(pgDistTransaction, heapTuple);
CommandCounterIncrement();
/* close relation and invalidate previous cache entry */
table_close(pgDistTransaction, NoLock);
}
/*
* RecoverTwoPhaseCommits recovers any pending prepared
* transactions started by this node on other nodes.
*/
int
RecoverTwoPhaseCommits(void)
{
int recoveredTransactionCount = 0;
/* take advisory lock first to avoid running concurrently */
LockTransactionRecovery(ShareUpdateExclusiveLock);
List *workerList = ActivePrimaryNodeList(NoLock);
WorkerNode *workerNode = NULL;
foreach_ptr(workerNode, workerList)
{
recoveredTransactionCount += RecoverWorkerTransactions(workerNode);
}
return recoveredTransactionCount;
}
/*
* RecoverWorkerTransactions recovers any pending prepared transactions
* started by this node on the specified worker.
*/
static int
RecoverWorkerTransactions(WorkerNode *workerNode)
{
int recoveredTransactionCount = 0;
int32 groupId = workerNode->groupId;
char *nodeName = workerNode->workerName;
int nodePort = workerNode->workerPort;
ScanKeyData scanKey[1];
int scanKeyCount = 1;
bool indexOK = true;
HeapTuple heapTuple = NULL;
HASH_SEQ_STATUS status;
bool recoveryFailed = false;
int connectionFlags = 0;
MultiConnection *connection = GetNodeConnection(connectionFlags, nodeName, nodePort);
if (connection->pgConn == NULL || PQstatus(connection->pgConn) != CONNECTION_OK)
{
ereport(WARNING, (errmsg("transaction recovery cannot connect to %s:%d",
nodeName, nodePort)));
return 0;
}
MemoryContext localContext = AllocSetContextCreateInternal(CurrentMemoryContext,
"RecoverWorkerTransactions",
ALLOCSET_DEFAULT_MINSIZE,
ALLOCSET_DEFAULT_INITSIZE,
ALLOCSET_DEFAULT_MAXSIZE);
MemoryContext oldContext = MemoryContextSwitchTo(localContext);
Relation pgDistTransaction = table_open(DistTransactionRelationId(),
RowExclusiveLock);
TupleDesc tupleDescriptor = RelationGetDescr(pgDistTransaction);
/*
* We're going to check the list of prepared transactions on the worker,
* but some of those prepared transactions might belong to ongoing
* distributed transactions.
*
* We could avoid this by temporarily blocking new prepared transactions
* from being created by taking an ExclusiveLock on pg_dist_transaction.
* However, this hurts write performance, so instead we avoid blocking
* by consulting the list of active distributed transactions, and follow
* a carefully chosen order to avoid race conditions:
*
* 1) P = prepared transactions on worker
* 2) A = active distributed transactions
* 3) T = pg_dist_transaction snapshot
* 4) Q = prepared transactions on worker
*
* By observing A after P, we get a conclusive answer to which distributed
* transactions we observed in P are still in progress. It is safe to recover
* the transactions in P - A based on the presence or absence of a record
* in T.
*
* We also remove records in T if there is no prepared transaction, which
* we assume means the transaction committed. However, a transaction could
* have left prepared transactions and committed between steps 1 and 2.
* In that case, we would incorrectly remove the records, while the
* prepared transaction is still in place.
*
* We therefore observe the set of prepared transactions one more time in
* step 4. The aforementioned transactions would show up in Q, but not in
* P. We can skip those transactions and recover them later.
*/
/* find stale prepared transactions on the remote node */
List *pendingTransactionList = PendingWorkerTransactionList(connection);
HTAB *pendingTransactionSet = ListToHashSet(pendingTransactionList, NAMEDATALEN,
true);
/* find in-progress distributed transactions */
List *activeTransactionNumberList = ActiveDistributedTransactionNumbers();
HTAB *activeTransactionNumberSet = ListToHashSet(activeTransactionNumberList,
sizeof(uint64), false);
/* scan through all recovery records of the current worker */
ScanKeyInit(&scanKey[0], Anum_pg_dist_transaction_groupid,
BTEqualStrategyNumber, F_INT4EQ, Int32GetDatum(groupId));
/* get a snapshot of pg_dist_transaction */
SysScanDesc scanDescriptor = systable_beginscan(pgDistTransaction,
DistTransactionGroupIndexId(),
indexOK,
NULL, scanKeyCount, scanKey);
/* find stale prepared transactions on the remote node */
List *recheckTransactionList = PendingWorkerTransactionList(connection);
HTAB *recheckTransactionSet = ListToHashSet(recheckTransactionList, NAMEDATALEN,
true);
while (HeapTupleIsValid(heapTuple = systable_getnext(scanDescriptor)))
{
bool isNull = false;
bool foundPreparedTransactionBeforeCommit = false;
bool foundPreparedTransactionAfterCommit = false;
Datum transactionNameDatum = heap_getattr(heapTuple,
Anum_pg_dist_transaction_gid,
tupleDescriptor, &isNull);
char *transactionName = TextDatumGetCString(transactionNameDatum);
bool isTransactionInProgress = IsTransactionInProgress(activeTransactionNumberSet,
transactionName);
if (isTransactionInProgress)
{
/*
* Do not touch in progress transactions as we might mistakenly
* commit a transaction that is actually in the process of
* aborting or vice-versa.
*/
continue;
}
/*
* Remove the transaction from the pending list such that only transactions
* that need to be aborted remain at the end.
*/
hash_search(pendingTransactionSet, transactionName, HASH_REMOVE,
&foundPreparedTransactionBeforeCommit);
hash_search(recheckTransactionSet, transactionName, HASH_FIND,
&foundPreparedTransactionAfterCommit);
if (foundPreparedTransactionBeforeCommit && foundPreparedTransactionAfterCommit)
{
/*
* The transaction was committed, but the prepared transaction still exists
* on the worker. Try committing it.
*
* We double check that the recovery record exists both before and after
* checking ActiveDistributedTransactionNumbers(), since we may have
* observed a prepared transaction that was committed immediately after.
*/
bool shouldCommit = true;
bool commitSucceeded = RecoverPreparedTransactionOnWorker(connection,
transactionName,
shouldCommit);
if (!commitSucceeded)
{
/*
* Failed to commit on the current worker. Stop without throwing
* an error to allow recover_prepared_transactions to continue with
* other workers.
*/
recoveryFailed = true;
break;
}
recoveredTransactionCount++;
/*
* We successfully committed the prepared transaction, safe to delete
* the recovery record.
*/
}
else if (foundPreparedTransactionAfterCommit)
{
/*
* We found a committed pg_dist_transaction record that initially did
* not have a prepared transaction, but did when we checked again.
*
* If a transaction started and committed just after we observed the
* set of prepared transactions, and just before we called
* ActiveDistributedTransactionNumbers, then we would see a recovery
* record without a prepared transaction in pendingTransactionSet,
* but there may be prepared transactions that failed to commit.
* We should not delete the records for those prepared transactions,
* since we would otherwise roll back them on the next call to
* recover_prepared_transactions.
*
* In addition, if the transaction started after the call to
* ActiveDistributedTransactionNumbers and finished just before our
* pg_dist_transaction snapshot, then it may still be in the process
* of committing the prepared transactions in the post-commit callback
* and we should not touch the prepared transactions.
*
* To handle these cases, we just leave the records and prepared
* transactions for the next call to recover_prepared_transactions
* and skip them here.
*/
continue;
}
else
{
/*
* We found a recovery record without any prepared transaction. It
* must have already been committed, so it's safe to delete the
* recovery record.
*
* Transactions that started after we observed pendingTransactionSet,
* but successfully committed their prepared transactions before
* ActiveDistributedTransactionNumbers are indistinguishable from
* transactions that committed at an earlier time, in which case it's
* safe delete the recovery record as well.
*/
}
simple_heap_delete(pgDistTransaction, &heapTuple->t_self);
}
systable_endscan(scanDescriptor);
table_close(pgDistTransaction, NoLock);
if (!recoveryFailed)
{
char *pendingTransactionName = NULL;
bool abortSucceeded = true;
/*
* All remaining prepared transactions that are not part of an in-progress
* distributed transaction should be aborted since we did not find a recovery
* record, which implies the disributed transaction aborted.
*/
hash_seq_init(&status, pendingTransactionSet);
while ((pendingTransactionName = hash_seq_search(&status)) != NULL)
{
bool isTransactionInProgress = IsTransactionInProgress(
activeTransactionNumberSet,
pendingTransactionName);
if (isTransactionInProgress)
{
continue;
}
bool shouldCommit = false;
abortSucceeded = RecoverPreparedTransactionOnWorker(connection,
pendingTransactionName,
shouldCommit);
if (!abortSucceeded)
{
hash_seq_term(&status);
break;
}
recoveredTransactionCount++;
}
}
MemoryContextSwitchTo(oldContext);
MemoryContextDelete(localContext);
return recoveredTransactionCount;
}
/*
* PendingWorkerTransactionList returns a list of pending prepared
* transactions on a remote node that were started by this node.
*/
static List *
PendingWorkerTransactionList(MultiConnection *connection)
{
StringInfo command = makeStringInfo();
bool raiseInterrupts = true;
List *transactionNames = NIL;
int32 coordinatorId = GetLocalGroupId();
appendStringInfo(command, "SELECT gid FROM pg_prepared_xacts "
"WHERE gid LIKE 'citus\\_%d\\_%%' and database = current_database()",
coordinatorId);
int querySent = SendRemoteCommand(connection, command->data);
if (querySent == 0)
{
ReportConnectionError(connection, ERROR);
}
PGresult *result = GetRemoteCommandResult(connection, raiseInterrupts);
if (!IsResponseOK(result))
{
ReportResultError(connection, result, ERROR);
}
int rowCount = PQntuples(result);
for (int rowIndex = 0; rowIndex < rowCount; rowIndex++)
{
const int columnIndex = 0;
char *transactionName = PQgetvalue(result, rowIndex, columnIndex);
transactionNames = lappend(transactionNames, pstrdup(transactionName));
}
PQclear(result);
ForgetResults(connection);
return transactionNames;
}
/*
* IsTransactionInProgress returns whether the distributed transaction to which
* preparedTransactionName belongs is still in progress, or false if the
* transaction name cannot be parsed. This can happen when the user manually
* inserts into pg_dist_transaction.
*/
static bool
IsTransactionInProgress(HTAB *activeTransactionNumberSet, char *preparedTransactionName)
{
int32 groupId = 0;
int procId = 0;
uint32 connectionNumber = 0;
uint64 transactionNumber = 0;
bool isTransactionInProgress = false;
bool isValidName = ParsePreparedTransactionName(preparedTransactionName, &groupId,
&procId,
&transactionNumber,
&connectionNumber);
if (isValidName)
{
hash_search(activeTransactionNumberSet, &transactionNumber, HASH_FIND,
&isTransactionInProgress);
}
return isTransactionInProgress;
}
/*
* RecoverPreparedTransactionOnWorker recovers a single prepared transaction over
* the given connection. If shouldCommit is true we send
*/
static bool
RecoverPreparedTransactionOnWorker(MultiConnection *connection, char *transactionName,
bool shouldCommit)
{
StringInfo command = makeStringInfo();
PGresult *result = NULL;
bool raiseInterrupts = false;
if (shouldCommit)
{
/* should have committed this prepared transaction */
appendStringInfo(command, "COMMIT PREPARED %s",
quote_literal_cstr(transactionName));
}
else
{
/* should have aborted this prepared transaction */
appendStringInfo(command, "ROLLBACK PREPARED %s",
quote_literal_cstr(transactionName));
}
int executeCommand = ExecuteOptionalRemoteCommand(connection, command->data, &result);
if (executeCommand == QUERY_SEND_FAILED)
{
return false;
}
if (executeCommand == RESPONSE_NOT_OKAY)
{
return false;
}
PQclear(result);
ClearResults(connection, raiseInterrupts);
ereport(LOG, (errmsg("recovered a prepared transaction on %s:%d",
connection->hostname, connection->port),
errcontext("%s", command->data)));
return true;
}
/*
* DeleteWorkerTransactions deletes the entries on pg_dist_transaction for a given
* worker node. It's implemented to be called at master_remove_node.
*/
void
DeleteWorkerTransactions(WorkerNode *workerNode)
{
if (workerNode == NULL)
{
/*
* We don't expect this, but let's be defensive since crashing is much worse
* than leaving pg_dist_transction entries.
*/
return;
}
bool indexOK = true;
int scanKeyCount = 1;
ScanKeyData scanKey[1];
int32 groupId = workerNode->groupId;
HeapTuple heapTuple = NULL;
Relation pgDistTransaction = table_open(DistTransactionRelationId(),
RowExclusiveLock);
ScanKeyInit(&scanKey[0], Anum_pg_dist_transaction_groupid,
BTEqualStrategyNumber, F_INT4EQ, Int32GetDatum(groupId));
SysScanDesc scanDescriptor = systable_beginscan(pgDistTransaction,
DistTransactionGroupIndexId(),
indexOK,
NULL, scanKeyCount, scanKey);
while (HeapTupleIsValid(heapTuple = systable_getnext(scanDescriptor)))
{
simple_heap_delete(pgDistTransaction, &heapTuple->t_self);
}
CommandCounterIncrement();
systable_endscan(scanDescriptor);
table_close(pgDistTransaction, NoLock);
}