*
* Michael J. Cahill, Uwe Röhm, and Alan D. Fekete. 2008.
* Serializable isolation for snapshot databases.
- * In SIGMOD ’08: Proceedings of the 2008 ACM SIGMOD
+ * In SIGMOD '08: Proceedings of the 2008 ACM SIGMOD
* international conference on Management of data,
- * pages 729–738, New York, NY, USA. ACM.
+ * pages 729-738, New York, NY, USA. ACM.
* http://doi.acm.org/10.1145/1376616.1376690
*
* and further elaborated in Cahill's doctoral thesis:
* examining the MVCC data.)
*
* (1) Besides tuples actually read, they must cover ranges of tuples
- * which would have been read based on the predicate. This will
+ * which would have been read based on the predicate. This will
* require modelling the predicates through locks against database
* objects such as pages, index ranges, or entire tables.
*
- * (2) They must be kept in RAM for quick access. Because of this, it
+ * (2) They must be kept in RAM for quick access. Because of this, it
* isn't possible to always maintain tuple-level granularity -- when
* the space allocated to store these approaches exhaustion, a
* request for a lock may need to scan for situations where a single
*
* (4) While they are associated with a transaction, they must survive
* a successful COMMIT of that transaction, and remain until all
- * overlapping transactions complete. This even means that they
+ * overlapping transactions complete. This even means that they
* must survive termination of the transaction's process. If a
* top level transaction is rolled back, however, it is immediately
* flagged so that it can be ignored, and its SIREAD locks can be
* an existing SIREAD lock for the same transaction, the SIREAD lock
* can be deleted.
*
- * (7) A write from a serializable transaction must ensure that a xact
+ * (7) A write from a serializable transaction must ensure that an xact
* record exists for the transaction, with the same lifespan (until
* all concurrent transaction complete or the transaction is rolled
* back) so that rw-dependencies to that transaction can be
* may yet matter because they overlap still-active transactions.
*
* SerializablePredicateLockListLock
- * - Protects the linked list of locks held by a transaction. Note
+ * - Protects the linked list of locks held by a transaction. Note
* that the locks themselves are also covered by the partition
* locks of their respective lock targets; this lock only affects
* the linked list connecting the locks related to a transaction.
* - All transactions share this single lock (with no partitioning).
* - There is never a need for a process other than the one running
* an active transaction to walk the list of locks held by that
- * transaction.
+ * transaction, except parallel query workers sharing the leader's
+ * transaction. In the parallel case, an extra per-sxact lock is
+ * taken; see below.
* - It is relatively infrequent that another process needs to
* modify the list for a transaction, but it does happen for such
* things as index page splits for pages with predicate locks and
- * freeing of predicate locked pages by a vacuum process. When
+ * freeing of predicate locked pages by a vacuum process. When
* removing a lock in such cases, the lock itself contains the
* pointers needed to remove it from the list. When adding a
* lock in such cases, the lock can be added using the anchor in
- * the transaction structure. Neither requires walking the list.
+ * the transaction structure. Neither requires walking the list.
* - Cleaning up the list for a terminated transaction is sometimes
* not done on a retail basis, in which case no lock is required.
* - Due to the above, a process accessing its active transaction's
* than its own active transaction must acquire an exclusive
* lock.
*
- * FirstPredicateLockMgrLock based partition locks
+ * SERIALIZABLEXACT's member 'predicateLockListLock'
+ * - Protects the linked list of locks held by a transaction. Only
+ * needed for parallel mode, where multiple backends share the
+ * same SERIALIZABLEXACT object. Not needed if
+ * SerializablePredicateLockListLock is held exclusively.
+ *
+ * PredicateLockHashPartitionLock(hashcode)
* - The same lock protects a target, all locks on that target, and
- * the linked list of locks on the target..
- * - When more than one is needed, acquire in ascending order.
+ * the linked list of locks on the target.
+ * - When more than one is needed, acquire in ascending address order.
+ * - When all are needed (rare), acquire in ascending index order with
+ * PredicateLockHashPartitionLockByIndex(index).
*
* SerializableXactHashLock
* - Protects both PredXact and SerializableXidHash.
*
*
- * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
*
* PageIsPredicateLocked(Relation relation, BlockNumber blkno)
*
* predicate lock maintenance
- * RegisterSerializableTransaction(Snapshot snapshot)
+ * GetSerializableTransactionSnapshot(Snapshot snapshot)
+ * SetSerializableTransactionSnapshot(Snapshot snapshot,
+ * VirtualTransactionId *sourcevxid)
* RegisterPredicateLockingXid(void)
- * PredicateLockRelation(Relation relation)
- * PredicateLockPage(Relation relation, BlockNumber blkno)
- * PredicateLockTuple(Relation relation, HeapTuple tuple)
+ * PredicateLockRelation(Relation relation, Snapshot snapshot)
+ * PredicateLockPage(Relation relation, BlockNumber blkno,
+ * Snapshot snapshot)
+ * PredicateLockTuple(Relation relation, HeapTuple tuple,
+ * Snapshot snapshot)
* PredicateLockPageSplit(Relation relation, BlockNumber oldblkno,
- * BlockNumber newblkno);
+ * BlockNumber newblkno)
* PredicateLockPageCombine(Relation relation, BlockNumber oldblkno,
- * BlockNumber newblkno);
- * PredicateLockTupleRowVersionLink(const Relation relation,
- * const HeapTuple oldTuple,
- * const HeapTuple newTuple)
- * ReleasePredicateLocks(bool isCommit)
+ * BlockNumber newblkno)
+ * TransferPredicateLocksToHeapRelation(Relation relation)
+ * ReleasePredicateLocks(bool isCommit, bool isReadOnlySafe)
*
* conflict detection (may also trigger rollback)
* CheckForSerializableConflictOut(bool visible, Relation relation,
- * HeapTupleData *tup, Buffer buffer)
+ * HeapTupleData *tup, Buffer buffer,
+ * Snapshot snapshot)
* CheckForSerializableConflictIn(Relation relation, HeapTupleData *tup,
* Buffer buffer)
+ * CheckTableForSerializableConflictIn(Relation relation)
*
* final rollback checking
* PreCommit_CheckForSerializationFailure(void)
#include "postgres.h"
+#include "access/heapam.h"
+#include "access/htup_details.h"
+#include "access/parallel.h"
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
#include "access/twophase.h"
#include "access/twophase_rmgr.h"
#include "access/xact.h"
+#include "access/xlog.h"
#include "miscadmin.h"
+#include "pgstat.h"
#include "storage/bufmgr.h"
#include "storage/predicate.h"
#include "storage/predicate_internals.h"
+#include "storage/proc.h"
#include "storage/procarray.h"
#include "utils/rel.h"
#include "utils/snapmgr.h"
-#include "utils/tqual.h"
/* Uncomment the next line to test the graceful degradation code. */
/* #define TEST_OLDSERXID */
#define PredicateLockHashPartition(hashcode) \
((hashcode) % NUM_PREDICATELOCK_PARTITIONS)
#define PredicateLockHashPartitionLock(hashcode) \
- ((LWLockId) (FirstPredicateLockMgrLock + PredicateLockHashPartition(hashcode)))
+ (&MainLWLockArray[PREDICATELOCK_MANAGER_LWLOCK_OFFSET + \
+ PredicateLockHashPartition(hashcode)].lock)
+#define PredicateLockHashPartitionLockByIndex(i) \
+ (&MainLWLockArray[PREDICATELOCK_MANAGER_LWLOCK_OFFSET + (i)].lock)
#define NPREDICATELOCKTARGETENTS() \
mul_size(max_predicate_locks_per_xact, add_size(MaxBackends, max_prepared_xacts))
#define SxactIsOnFinishedList(sxact) (!SHMQueueIsDetached(&((sxact)->finishedLink)))
-#define SxactIsPrepared(sxact) (((sxact)->flags & SXACT_FLAG_PREPARED) != 0)
+/*
+ * Note that a sxact is marked "prepared" once it has passed
+ * PreCommit_CheckForSerializationFailure, even if it isn't using
+ * 2PC. This is the point at which it can no longer be aborted.
+ *
+ * The PREPARED flag remains set after commit, so SxactIsCommitted
+ * implies SxactIsPrepared.
+ */
#define SxactIsCommitted(sxact) (((sxact)->flags & SXACT_FLAG_COMMITTED) != 0)
+#define SxactIsPrepared(sxact) (((sxact)->flags & SXACT_FLAG_PREPARED) != 0)
#define SxactIsRolledBack(sxact) (((sxact)->flags & SXACT_FLAG_ROLLED_BACK) != 0)
+#define SxactIsDoomed(sxact) (((sxact)->flags & SXACT_FLAG_DOOMED) != 0)
#define SxactIsReadOnly(sxact) (((sxact)->flags & SXACT_FLAG_READ_ONLY) != 0)
#define SxactHasSummaryConflictIn(sxact) (((sxact)->flags & SXACT_FLAG_SUMMARY_CONFLICT_IN) != 0)
#define SxactHasSummaryConflictOut(sxact) (((sxact)->flags & SXACT_FLAG_SUMMARY_CONFLICT_OUT) != 0)
+/*
+ * The following macro actually means that the specified transaction has a
+ * conflict out *to a transaction which committed ahead of it*. It's hard
+ * to get that into a name of a reasonable length.
+ */
#define SxactHasConflictOut(sxact) (((sxact)->flags & SXACT_FLAG_CONFLICT_OUT) != 0)
#define SxactIsDeferrableWaiting(sxact) (((sxact)->flags & SXACT_FLAG_DEFERRABLE_WAITING) != 0)
#define SxactIsROSafe(sxact) (((sxact)->flags & SXACT_FLAG_RO_SAFE) != 0)
#define SxactIsROUnsafe(sxact) (((sxact)->flags & SXACT_FLAG_RO_UNSAFE) != 0)
-#define SxactIsMarkedForDeath(sxact) (((sxact)->flags & SXACT_FLAG_MARKED_FOR_DEATH) != 0)
-
-/*
- * When a public interface method is called for a split on an index relation,
- * this is the test to see if we should do a quick return.
- */
-#define SkipSplitTracking(relation) \
- (((relation)->rd_id < FirstBootstrapObjectId) \
- || RelationUsesLocalBuffers(relation))
-
-/*
- * When a public interface method is called for serializing a relation within
- * the current transaction, this is the test to see if we should do a quick
- * return.
- */
-#define SkipSerialization(relation) \
- ((!IsolationIsSerializable()) \
- || ((MySerializableXact == InvalidSerializableXact)) \
- || ReleasePredicateLocksIfROSafe() \
- || SkipSplitTracking(relation))
-
+#define SxactIsPartiallyReleased(sxact) (((sxact)->flags & SXACT_FLAG_PARTIALLY_RELEASED) != 0)
/*
* Compute the hash code associated with a PREDICATELOCKTARGETTAG.
* the lock partition number from the hashcode.
*/
#define PredicateLockTargetTagHashCode(predicatelocktargettag) \
- (tag_hash((predicatelocktargettag), sizeof(PREDICATELOCKTARGETTAG)))
+ get_hash_value(PredicateLockTargetHash, predicatelocktargettag)
/*
* Given a predicate lock tag, and the hash for its target,
#define OLDSERXID_PAGESIZE BLCKSZ
#define OLDSERXID_ENTRYSIZE sizeof(SerCommitSeqNo)
#define OLDSERXID_ENTRIESPERPAGE (OLDSERXID_PAGESIZE / OLDSERXID_ENTRYSIZE)
-#define OLDSERXID_MAX_PAGE (SLRU_PAGES_PER_SEGMENT * 0x10000 - 1)
+
+/*
+ * Set maximum pages based on the number needed to track all transactions.
+ */
+#define OLDSERXID_MAX_PAGE (MaxTransactionId / OLDSERXID_ENTRIESPERPAGE)
#define OldSerXidNextPage(page) (((page) >= OLDSERXID_MAX_PAGE) ? 0 : (page) + 1)
(OldSerXidSlruCtl->shared->page_buffer[slotno] + \
((((uint32) (xid)) % OLDSERXID_ENTRIESPERPAGE) * OLDSERXID_ENTRYSIZE))))
-#define OldSerXidPage(xid) ((((uint32) (xid)) / OLDSERXID_ENTRIESPERPAGE) % (OLDSERXID_MAX_PAGE + 1))
-#define OldSerXidSegment(page) ((page) / SLRU_PAGES_PER_SEGMENT)
+#define OldSerXidPage(xid) (((uint32) (xid)) / OLDSERXID_ENTRIESPERPAGE)
typedef struct OldSerXidControlData
{
int headPage; /* newest initialized page */
TransactionId headXid; /* newest valid Xid in the SLRU */
TransactionId tailXid; /* oldest xmin we might be interested in */
- bool warningIssued;
-} OldSerXidControlData;
+} OldSerXidControlData;
typedef struct OldSerXidControlData *OldSerXidControl;
static SERIALIZABLEXACT *OldCommittedSxact;
-/* This configuration variable is used to set the predicate lock table size */
-int max_predicate_locks_per_xact; /* set by guc.c */
+/*
+ * These configuration variables are used to set the predicate lock table size
+ * and to control promotion of predicate locks to coarser granularity in an
+ * attempt to degrade performance (mostly as false positive serialization
+ * failure) gracefully in the face of memory pressurel
+ */
+int max_predicate_locks_per_xact; /* set by guc.c */
+int max_predicate_locks_per_relation; /* set by guc.c */
+int max_predicate_locks_per_page; /* set by guc.c */
/*
* This provides a list of objects in order to track transactions
- * participating in predicate locking. Entries in the list are fixed size,
+ * participating in predicate locking. Entries in the list are fixed size,
* and reside in shared memory. The memory address of an entry must remain
* fixed during its lifetime. The list will be protected from concurrent
* update externally; no provision is made in this code to manage that. The
static SHM_QUEUE *FinishedSerializableTransactions;
/*
- * Tag for a reserved entry in PredicateLockTargetHash; used to ensure
- * there's an element available for scratch space if we need it,
- * e.g. in PredicateLockPageSplit. This is an otherwise-invalid tag.
+ * Tag for a dummy entry in PredicateLockTargetHash. By temporarily removing
+ * this entry, you can ensure that there's enough scratch space available for
+ * inserting one entry in the hash table. This is an otherwise-invalid tag.
*/
-static const PREDICATELOCKTARGETTAG ReservedTargetTag = {0, 0, 0, 0, 0};
+static const PREDICATELOCKTARGETTAG ScratchTargetTag = {0, 0, 0, 0};
+static uint32 ScratchTargetTagHash;
+static LWLock *ScratchPartitionLock;
/*
* The local hash table used to determine when to combine multiple fine-
/*
* Keep a pointer to the currently-running serializable transaction (if any)
- * for quick reference.
- * TODO SSI: Remove volatile qualifier and the then-unnecessary casts?
+ * for quick reference. Also, remember if we have written anything that could
+ * cause a rw-conflict.
*/
-static volatile SERIALIZABLEXACT *MySerializableXact = InvalidSerializableXact;
+static SERIALIZABLEXACT *MySerializableXact = InvalidSerializableXact;
+static bool MyXactDidWrite = false;
+
+/*
+ * The SXACT_FLAG_RO_UNSAFE optimization might lead us to release
+ * MySerializableXact early. If that happens in a parallel query, the leader
+ * needs to defer the destruction of the SERIALIZABLEXACT until end of
+ * transaction, because the workers still have a reference to it. In that
+ * case, the leader stores it here.
+ */
+static SERIALIZABLEXACT *SavedSerializableXact = InvalidSerializableXact;
/* local functions */
static uint32 predicatelock_hash(const void *key, Size keysize);
static void SummarizeOldestCommittedSxact(void);
static Snapshot GetSafeSnapshot(Snapshot snapshot);
-static Snapshot RegisterSerializableTransactionInt(Snapshot snapshot);
+static Snapshot GetSerializableTransactionSnapshotInt(Snapshot snapshot,
+ VirtualTransactionId *sourcevxid,
+ int sourcepid);
static bool PredicateLockExists(const PREDICATELOCKTARGETTAG *targettag);
static bool GetParentPredicateLockTag(const PREDICATELOCKTARGETTAG *tag,
- PREDICATELOCKTARGETTAG *parent);
+ PREDICATELOCKTARGETTAG *parent);
static bool CoarserLockCovers(const PREDICATELOCKTARGETTAG *newtargettag);
+static void RemoveScratchTarget(bool lockheld);
+static void RestoreScratchTarget(bool lockheld);
static void RemoveTargetIfNoLongerUsed(PREDICATELOCKTARGET *target,
- uint32 targettaghash);
+ uint32 targettaghash);
static void DeleteChildTargetLocks(const PREDICATELOCKTARGETTAG *newtargettag);
-static int PredicateLockPromotionThreshold(const PREDICATELOCKTARGETTAG *tag);
+static int MaxPredicateChildLocks(const PREDICATELOCKTARGETTAG *tag);
static bool CheckAndPromotePredicateLockRequest(const PREDICATELOCKTARGETTAG *reqtag);
static void DecrementParentLocks(const PREDICATELOCKTARGETTAG *targettag);
static void CreatePredicateLock(const PREDICATELOCKTARGETTAG *targettag,
- uint32 targettaghash,
- SERIALIZABLEXACT *sxact);
+ uint32 targettaghash,
+ SERIALIZABLEXACT *sxact);
static void DeleteLockTarget(PREDICATELOCKTARGET *target, uint32 targettaghash);
-static bool TransferPredicateLocksToNewTarget(const PREDICATELOCKTARGETTAG oldtargettag,
- const PREDICATELOCKTARGETTAG newtargettag,
- bool removeOld);
+static bool TransferPredicateLocksToNewTarget(PREDICATELOCKTARGETTAG oldtargettag,
+ PREDICATELOCKTARGETTAG newtargettag,
+ bool removeOld);
static void PredicateLockAcquire(const PREDICATELOCKTARGETTAG *targettag);
+static void DropAllPredicateLocksFromTable(Relation relation,
+ bool transfer);
static void SetNewSxactGlobalXmin(void);
-static bool ReleasePredicateLocksIfROSafe(void);
static void ClearOldPredicateLocks(void);
static void ReleaseOneSerializableXact(SERIALIZABLEXACT *sxact, bool partial,
- bool summarize);
+ bool summarize);
static bool XidIsConcurrent(TransactionId xid);
static void CheckTargetForConflictsIn(PREDICATELOCKTARGETTAG *targettag);
static void FlagRWConflict(SERIALIZABLEXACT *reader, SERIALIZABLEXACT *writer);
static void OnConflict_CheckForSerializationFailure(const SERIALIZABLEXACT *reader,
- SERIALIZABLEXACT *writer);
+ SERIALIZABLEXACT *writer);
+static void CreateLocalPredicateLockHash(void);
+static void ReleasePredicateLocksLocal(void);
+
+
+/*------------------------------------------------------------------------*/
+
+/*
+ * Does this relation participate in predicate locking? Temporary and system
+ * relations are exempt, as are materialized views.
+ */
+static inline bool
+PredicateLockingNeededForRelation(Relation relation)
+{
+ return !(relation->rd_id < FirstBootstrapObjectId ||
+ RelationUsesLocalBuffers(relation) ||
+ relation->rd_rel->relkind == RELKIND_MATVIEW);
+}
+
+/*
+ * When a public interface method is called for a read, this is the test to
+ * see if we should do a quick return.
+ *
+ * Note: this function has side-effects! If this transaction has been flagged
+ * as RO-safe since the last call, we release all predicate locks and reset
+ * MySerializableXact. That makes subsequent calls to return quickly.
+ *
+ * This is marked as 'inline' to eliminate the function call overhead in the
+ * common case that serialization is not needed.
+ */
+static inline bool
+SerializationNeededForRead(Relation relation, Snapshot snapshot)
+{
+ /* Nothing to do if this is not a serializable transaction */
+ if (MySerializableXact == InvalidSerializableXact)
+ return false;
+
+ /*
+ * Don't acquire locks or conflict when scanning with a special snapshot.
+ * This excludes things like CLUSTER and REINDEX. They use the wholesale
+ * functions TransferPredicateLocksToHeapRelation() and
+ * CheckTableForSerializableConflictIn() to participate in serialization,
+ * but the scans involved don't need serialization.
+ */
+ if (!IsMVCCSnapshot(snapshot))
+ return false;
+
+ /*
+ * Check if we have just become "RO-safe". If we have, immediately release
+ * all locks as they're not needed anymore. This also resets
+ * MySerializableXact, so that subsequent calls to this function can exit
+ * quickly.
+ *
+ * A transaction is flagged as RO_SAFE if all concurrent R/W transactions
+ * commit without having conflicts out to an earlier snapshot, thus
+ * ensuring that no conflicts are possible for this transaction.
+ */
+ if (SxactIsROSafe(MySerializableXact))
+ {
+ ReleasePredicateLocks(false, true);
+ return false;
+ }
+
+ /* Check if the relation doesn't participate in predicate locking */
+ if (!PredicateLockingNeededForRelation(relation))
+ return false;
+
+ return true; /* no excuse to skip predicate locking */
+}
+
+/*
+ * Like SerializationNeededForRead(), but called on writes.
+ * The logic is the same, but there is no snapshot and we can't be RO-safe.
+ */
+static inline bool
+SerializationNeededForWrite(Relation relation)
+{
+ /* Nothing to do if this is not a serializable transaction */
+ if (MySerializableXact == InvalidSerializableXact)
+ return false;
+
+ /* Check if the relation doesn't participate in predicate locking */
+ if (!PredicateLockingNeededForRelation(relation))
+ return false;
+
+ return true; /* no excuse to skip predicate locking */
+}
+
/*------------------------------------------------------------------------*/
/*
* These functions are a simple implementation of a list for this specific
- * type of struct. If there is ever a generalized shared memory list, we
+ * type of struct. If there is ever a generalized shared memory list, we
* should probably switch to that.
*/
static SERIALIZABLEXACT *
ptle = (PredXactListElement)
(((char *) sxact)
- offsetof(PredXactListElementData, sxact)
- +offsetof(PredXactListElementData, link));
+ + offsetof(PredXactListElementData, link));
SHMQueueDelete(&ptle->link);
SHMQueueInsertBefore(&PredXact->availableList, &ptle->link);
}
ptle = (PredXactListElement)
(((char *) sxact)
- offsetof(PredXactListElementData, sxact)
- +offsetof(PredXactListElementData, link));
+ + offsetof(PredXactListElementData, link));
ptle = (PredXactListElement)
SHMQueueNext(&PredXact->activeList,
&ptle->link,
Assert(reader != writer);
/* Check the ends of the purported conflict first. */
- if (SxactIsRolledBack(reader)
- || SxactIsRolledBack(writer)
+ if (SxactIsDoomed(reader)
+ || SxactIsDoomed(writer)
|| SHMQueueEmpty(&reader->outConflicts)
|| SHMQueueEmpty(&writer->inConflicts))
return false;
if (!conflict)
ereport(ERROR,
(errcode(ERRCODE_OUT_OF_MEMORY),
- errmsg("not enough elements in RWConflictPool to record a rw-conflict"),
+ errmsg("not enough elements in RWConflictPool to record a read/write conflict"),
errhint("You might need to run fewer transactions at a time or increase max_connections.")));
SHMQueueDelete(&conflict->outLink);
if (!conflict)
ereport(ERROR,
(errcode(ERRCODE_OUT_OF_MEMORY),
- errmsg("not enough elements in RWConflictPool to record a potential rw-conflict"),
+ errmsg("not enough elements in RWConflictPool to record a potential read/write conflict"),
errhint("You might need to run fewer transactions at a time or increase max_connections.")));
SHMQueueDelete(&conflict->outLink);
int diff;
/*
- * We have to compare modulo (OLDSERXID_MAX_PAGE+1)/2. Both inputs should
+ * We have to compare modulo (OLDSERXID_MAX_PAGE+1)/2. Both inputs should
* be in the range 0..OLDSERXID_MAX_PAGE.
*/
Assert(p >= 0 && p <= OLDSERXID_MAX_PAGE);
diff = p - q;
if (diff >= ((OLDSERXID_MAX_PAGE + 1) / 2))
diff -= OLDSERXID_MAX_PAGE + 1;
- else if (diff < -((OLDSERXID_MAX_PAGE + 1) / 2))
+ else if (diff < -((int) (OLDSERXID_MAX_PAGE + 1) / 2))
diff += OLDSERXID_MAX_PAGE + 1;
return diff < 0;
}
* Set up SLRU management of the pg_serial data.
*/
OldSerXidSlruCtl->PagePrecedes = OldSerXidPagePrecedesLogically;
- SimpleLruInit(OldSerXidSlruCtl, "OldSerXid SLRU Ctl", NUM_OLDSERXID_BUFFERS, 0,
- OldSerXidLock, "pg_serial");
+ SimpleLruInit(OldSerXidSlruCtl, "oldserxid",
+ NUM_OLDSERXID_BUFFERS, 0, OldSerXidLock, "pg_serial",
+ LWTRANCHE_OLDSERXID_BUFFERS);
/* Override default assumption that writes should be fsync'd */
OldSerXidSlruCtl->do_fsync = false;
oldSerXidControl = (OldSerXidControl)
ShmemInitStruct("OldSerXidControlData", sizeof(OldSerXidControlData), &found);
+ Assert(found == IsUnderPostmaster);
if (!found)
{
/*
oldSerXidControl->headPage = -1;
oldSerXidControl->headXid = InvalidTransactionId;
oldSerXidControl->tailXid = InvalidTransactionId;
- oldSerXidControl->warningIssued = false;
}
}
/*
* Record a committed read write serializable xid and the minimum
* commitSeqNo of any transactions to which this xid had a rw-conflict out.
- * A zero seqNo means that there were no conflicts out from xid.
+ * An invalid seqNo means that there were no conflicts out from xid.
*/
static void
OldSerXidAdd(TransactionId xid, SerCommitSeqNo minConflictCommitSeqNo)
int targetPage;
int slotno;
int firstZeroPage;
- int xidSpread;
bool isNewPage;
Assert(TransactionIdIsValid(xid));
Assert(TransactionIdIsValid(tailXid));
/*
- * If the SLRU is currently unused, zero out the whole active region
- * from tailXid to headXid before taking it into use. Otherwise zero
- * out only any new pages that enter the tailXid-headXid range as we
- * advance headXid.
+ * If the SLRU is currently unused, zero out the whole active region from
+ * tailXid to headXid before taking it into use. Otherwise zero out only
+ * any new pages that enter the tailXid-headXid range as we advance
+ * headXid.
*/
if (oldSerXidControl->headPage < 0)
{
if (isNewPage)
oldSerXidControl->headPage = targetPage;
- xidSpread = (((uint32) xid) - ((uint32) tailXid));
- if (oldSerXidControl->warningIssued)
- {
- if (xidSpread < 800000000)
- oldSerXidControl->warningIssued = false;
- }
- else if (xidSpread >= 1000000000)
- {
- oldSerXidControl->warningIssued = true;
- ereport(WARNING,
- (errmsg("memory for serializable conflict tracking is nearly exhausted"),
- errhint("There may be an idle transaction or a forgotten prepared transaction causing this.")));
- }
-
if (isNewPage)
{
/* Initialize intervening pages. */
slotno = SimpleLruReadPage(OldSerXidSlruCtl, targetPage, true, xid);
OldSerXidValue(slotno, xid) = minConflictCommitSeqNo;
+ OldSerXidSlruCtl->shared->page_dirty[slotno] = true;
LWLockRelease(OldSerXidLock);
}
/*
- * Get the minimum commitSeqNo for any conflict out for the given xid. For
+ * Get the minimum commitSeqNo for any conflict out for the given xid. For
* a transaction which exists but has no conflict out, InvalidSerCommitSeqNo
* will be returned.
*/
/*
* When no sxacts are active, nothing overlaps, set the xid values to
* invalid to show that there are no valid entries. Don't clear headPage,
- * though. A new xmin might still land on that page, and we don't want
- * to repeatedly zero out the same page.
+ * though. A new xmin might still land on that page, and we don't want to
+ * repeatedly zero out the same page.
*/
if (!TransactionIdIsValid(xid))
{
void
CheckPointPredicate(void)
{
- int tailPage;
+ int tailPage;
LWLockAcquire(OldSerXidLock, LW_EXCLUSIVE);
else
{
/*
- * The SLRU is no longer needed. Truncate everything but the last
- * page. We don't dare to touch the last page in case the SLRU is
- * taken back to use, and the new tail falls on the same page.
+ * The SLRU is no longer needed. Truncate to head before we set head
+ * invalid.
+ *
+ * XXX: It's possible that the SLRU is not needed again until XID
+ * wrap-around has happened, so that the segment containing headPage
+ * that we leave behind will appear to be new again. In that case it
+ * won't be removed until XID horizon advances enough to make it
+ * current again.
*/
tailPage = oldSerXidControl->headPage;
oldSerXidControl->headPage = -1;
LWLockRelease(OldSerXidLock);
+ /* Truncate away pages that are no longer required */
+ SimpleLruTruncate(OldSerXidSlruCtl, tailPage);
+
/*
* Flush dirty SLRU pages to disk
*
* This is not actually necessary from a correctness point of view. We do
* it merely as a debugging aid.
+ *
+ * We're doing this after the truncation to avoid writing pages right
+ * before deleting the file in which they sit, which would be completely
+ * pointless.
*/
SimpleLruFlush(OldSerXidSlruCtl, true);
-
- /* Truncate away pages that are no longer required */
- SimpleLruTruncate(OldSerXidSlruCtl, tailPage);
}
/*------------------------------------------------------------------------*/
InitPredicateLocks(void)
{
HASHCTL info;
- int hash_flags;
- long init_table_size,
- max_table_size;
+ long max_table_size;
Size requestSize;
bool found;
+#ifndef EXEC_BACKEND
+ Assert(!IsUnderPostmaster);
+#endif
+
/*
- * Compute init/max size to request for predicate lock target hashtable.
- * Note these calculations must agree with PredicateLockShmemSize!
+ * Compute size of predicate lock target hashtable. Note these
+ * calculations must agree with PredicateLockShmemSize!
*/
max_table_size = NPREDICATELOCKTARGETENTS();
- init_table_size = max_table_size / 2;
/*
* Allocate hash table for PREDICATELOCKTARGET structs. This stores
MemSet(&info, 0, sizeof(info));
info.keysize = sizeof(PREDICATELOCKTARGETTAG);
info.entrysize = sizeof(PREDICATELOCKTARGET);
- info.hash = tag_hash;
info.num_partitions = NUM_PREDICATELOCK_PARTITIONS;
- hash_flags = (HASH_ELEM | HASH_FUNCTION | HASH_PARTITION);
PredicateLockTargetHash = ShmemInitHash("PREDICATELOCKTARGET hash",
- init_table_size,
+ max_table_size,
max_table_size,
&info,
- hash_flags);
-
- /* Assume an average of 2 xacts per target */
- max_table_size *= 2;
- init_table_size *= 2;
+ HASH_ELEM | HASH_BLOBS |
+ HASH_PARTITION | HASH_FIXED_SIZE);
/*
- * Reserve an entry in the hash table; we use it to make sure there's
+ * Reserve a dummy entry in the hash table; we use it to make sure there's
* always one entry available when we need to split or combine a page,
* because running out of space there could mean aborting a
* non-serializable transaction.
*/
- hash_search(PredicateLockTargetHash, &ReservedTargetTag,
- HASH_ENTER, NULL);
+ if (!IsUnderPostmaster)
+ {
+ (void) hash_search(PredicateLockTargetHash, &ScratchTargetTag,
+ HASH_ENTER, &found);
+ Assert(!found);
+ }
+ /* Pre-calculate the hash and partition lock of the scratch entry */
+ ScratchTargetTagHash = PredicateLockTargetTagHashCode(&ScratchTargetTag);
+ ScratchPartitionLock = PredicateLockHashPartitionLock(ScratchTargetTagHash);
/*
* Allocate hash table for PREDICATELOCK structs. This stores per
info.entrysize = sizeof(PREDICATELOCK);
info.hash = predicatelock_hash;
info.num_partitions = NUM_PREDICATELOCK_PARTITIONS;
- hash_flags = (HASH_ELEM | HASH_FUNCTION | HASH_PARTITION);
+
+ /* Assume an average of 2 xacts per target */
+ max_table_size *= 2;
PredicateLockHash = ShmemInitHash("PREDICATELOCK hash",
- init_table_size,
+ max_table_size,
max_table_size,
&info,
- hash_flags);
+ HASH_ELEM | HASH_FUNCTION |
+ HASH_PARTITION | HASH_FIXED_SIZE);
/*
- * Compute init/max size to request for serializable transaction
- * hashtable. Note these calculations must agree with
- * PredicateLockShmemSize!
+ * Compute size for serializable transaction hashtable. Note these
+ * calculations must agree with PredicateLockShmemSize!
*/
max_table_size = (MaxBackends + max_prepared_xacts);
PredXact = ShmemInitStruct("PredXactList",
PredXactListDataSize,
&found);
+ Assert(found == IsUnderPostmaster);
if (!found)
{
int i;
requestSize = mul_size((Size) max_table_size,
PredXactListElementDataSize);
PredXact->element = ShmemAlloc(requestSize);
- if (PredXact->element == NULL)
- ereport(ERROR,
- (errcode(ERRCODE_OUT_OF_MEMORY),
- errmsg("not enough shared memory for elements of data structure"
- " \"%s\" (%lu bytes requested)",
- "PredXactList", (unsigned long) requestSize)));
/* Add all elements to available list, clean. */
memset(PredXact->element, 0, requestSize);
for (i = 0; i < max_table_size; i++)
{
+ LWLockInitialize(&PredXact->element[i].sxact.predicateLockListLock,
+ LWTRANCHE_SXACT);
SHMQueueInsertBefore(&(PredXact->availableList),
&(PredXact->element[i].link));
}
PredXact->OldCommittedSxact = CreatePredXact();
SetInvalidVirtualTransactionId(PredXact->OldCommittedSxact->vxid);
+ PredXact->OldCommittedSxact->prepareSeqNo = 0;
PredXact->OldCommittedSxact->commitSeqNo = 0;
PredXact->OldCommittedSxact->SeqNo.lastCommitBeforeSnapshot = 0;
SHMQueueInit(&PredXact->OldCommittedSxact->outConflicts);
MemSet(&info, 0, sizeof(info));
info.keysize = sizeof(SERIALIZABLEXIDTAG);
info.entrysize = sizeof(SERIALIZABLEXID);
- info.hash = tag_hash;
- hash_flags = (HASH_ELEM | HASH_FUNCTION);
SerializableXidHash = ShmemInitHash("SERIALIZABLEXID hash",
max_table_size,
max_table_size,
&info,
- hash_flags);
+ HASH_ELEM | HASH_BLOBS |
+ HASH_FIXED_SIZE);
/*
* Allocate space for tracking rw-conflicts in lists attached to the
* that this will prevent resource exhaustion in even the most pessimal
* loads up to max_connections = 200 with all 200 connections pounding the
* database with serializable transactions. Beyond that, there may be
- * occassional transactions canceled when trying to flag conflicts. That's
+ * occasional transactions canceled when trying to flag conflicts. That's
* probably OK.
*/
max_table_size *= 5;
RWConflictPool = ShmemInitStruct("RWConflictPool",
RWConflictPoolHeaderDataSize,
&found);
+ Assert(found == IsUnderPostmaster);
if (!found)
{
int i;
requestSize = mul_size((Size) max_table_size,
RWConflictDataSize);
RWConflictPool->element = ShmemAlloc(requestSize);
- if (RWConflictPool->element == NULL)
- ereport(ERROR,
- (errcode(ERRCODE_OUT_OF_MEMORY),
- errmsg("not enough shared memory for elements of data structure"
- " \"%s\" (%lu bytes requested)",
- "RWConflictPool", (unsigned long) requestSize)));
/* Add all elements to available list, clean. */
memset(RWConflictPool->element, 0, requestSize);
for (i = 0; i < max_table_size; i++)
ShmemInitStruct("FinishedSerializableTransactions",
sizeof(SHM_QUEUE),
&found);
+ Assert(found == IsUnderPostmaster);
if (!found)
SHMQueueInit(FinishedSerializableTransactions);
* in ascending order, then SerializableXactHashLock.
*/
for (i = 0; i < NUM_PREDICATELOCK_PARTITIONS; i++)
- LWLockAcquire(FirstPredicateLockMgrLock + i, LW_SHARED);
+ LWLockAcquire(PredicateLockHashPartitionLockByIndex(i), LW_SHARED);
LWLockAcquire(SerializableXactHashLock, LW_SHARED);
/* Get number of locks and allocate appropriately-sized arrays. */
/* Release locks in reverse order */
LWLockRelease(SerializableXactHashLock);
for (i = NUM_PREDICATELOCK_PARTITIONS - 1; i >= 0; i--)
- LWLockRelease(FirstPredicateLockMgrLock + i);
+ LWLockRelease(PredicateLockHashPartitionLockByIndex(i));
return data;
}
/*
* This function is only called if there are no sxact slots available.
* Some of them must belong to old, already-finished transactions, so
- * there should be something in FinishedSerializableTransactions list
- * that we can summarize. However, there's a race condition: while we
- * were not holding any locks, a transaction might have ended and cleaned
- * up all the finished sxact entries already, freeing up their sxact
- * slots. In that case, we have nothing to do here. The caller will find
- * one of the slots released by the other backend when it retries.
+ * there should be something in FinishedSerializableTransactions list that
+ * we can summarize. However, there's a race condition: while we were not
+ * holding any locks, a transaction might have ended and cleaned up all
+ * the finished sxact entries already, freeing up their sxact slots. In
+ * that case, we have nothing to do here. The caller will find one of the
+ * slots released by the other backend when it retries.
*/
if (SHMQueueEmpty(FinishedSerializableTransactions))
{
/*
* Grab the first sxact off the finished list -- this will be the earliest
- * commit. Remove it from the list.
+ * commit. Remove it from the list.
*/
sxact = (SERIALIZABLEXACT *)
SHMQueueNext(FinishedSerializableTransactions,
/* Add to SLRU summary information. */
if (TransactionIdIsValid(sxact->topXid) && !SxactIsReadOnly(sxact))
OldSerXidAdd(sxact->topXid, SxactHasConflictOut(sxact)
- ? sxact->SeqNo.earliestOutConflictCommit : InvalidSerCommitSeqNo);
+ ? sxact->SeqNo.earliestOutConflictCommit : InvalidSerCommitSeqNo);
/* Summarize and release the detail. */
ReleaseOneSerializableXact(sxact, false, true);
* without further checks. This requires waiting for concurrent
* transactions to complete, and retrying with a new snapshot if
* one of them could possibly create a conflict.
+ *
+ * As with GetSerializableTransactionSnapshot (which this is a subroutine
+ * for), the passed-in Snapshot pointer should reference a static data
+ * area that can safely be passed to GetSnapshotData.
*/
static Snapshot
GetSafeSnapshot(Snapshot origSnapshot)
while (true)
{
/*
- * RegisterSerializableTransactionInt is going to call
- * GetSnapshotData, so we need to provide it the static snapshot our
- * caller passed to us. It returns a copy of that snapshot and
- * registers it on TopTransactionResourceOwner.
+ * GetSerializableTransactionSnapshotInt is going to call
+ * GetSnapshotData, so we need to provide it the static snapshot area
+ * our caller passed to us. The pointer returned is actually the same
+ * one passed to it, but we avoid assuming that here.
*/
- snapshot = RegisterSerializableTransactionInt(origSnapshot);
+ snapshot = GetSerializableTransactionSnapshotInt(origSnapshot,
+ NULL, InvalidPid);
if (MySerializableXact == InvalidSerializableXact)
return snapshot; /* no concurrent r/w xacts; it's safe */
- MySerializableXact->flags |= SXACT_FLAG_DEFERRABLE_WAITING;
+ LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
/*
* Wait for concurrent transactions to finish. Stop early if one of
* them marked us as conflicted.
*/
- while (!(SHMQueueEmpty((SHM_QUEUE *)
- &MySerializableXact->possibleUnsafeConflicts) ||
+ MySerializableXact->flags |= SXACT_FLAG_DEFERRABLE_WAITING;
+ while (!(SHMQueueEmpty(&MySerializableXact->possibleUnsafeConflicts) ||
SxactIsROUnsafe(MySerializableXact)))
- ProcWaitForSignal();
-
+ {
+ LWLockRelease(SerializableXactHashLock);
+ ProcWaitForSignal(WAIT_EVENT_SAFE_SNAPSHOT);
+ LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
+ }
MySerializableXact->flags &= ~SXACT_FLAG_DEFERRABLE_WAITING;
+
if (!SxactIsROUnsafe(MySerializableXact))
+ {
+ LWLockRelease(SerializableXactHashLock);
break; /* success */
+ }
+
+ LWLockRelease(SerializableXactHashLock);
/* else, need to retry... */
ereport(DEBUG2,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("deferrable snapshot was unsafe; trying a new one")));
- ReleasePredicateLocks(false);
- UnregisterSnapshotFromOwner(snapshot,
- TopTransactionResourceOwner);
+ ReleasePredicateLocks(false, false);
}
/*
* Now we have a safe snapshot, so we don't need to do any further checks.
*/
Assert(SxactIsROSafe(MySerializableXact));
- ReleasePredicateLocks(false);
+ ReleasePredicateLocks(false, true);
return snapshot;
}
/*
- * Acquire and register a snapshot which can be used for this transaction..
+ * GetSafeSnapshotBlockingPids
+ * If the specified process is currently blocked in GetSafeSnapshot,
+ * write the process IDs of all processes that it is blocked by
+ * into the caller-supplied buffer output[]. The list is truncated at
+ * output_size, and the number of PIDs written into the buffer is
+ * returned. Returns zero if the given PID is not currently blocked
+ * in GetSafeSnapshot.
+ */
+int
+GetSafeSnapshotBlockingPids(int blocked_pid, int *output, int output_size)
+{
+ int num_written = 0;
+ SERIALIZABLEXACT *sxact;
+
+ LWLockAcquire(SerializableXactHashLock, LW_SHARED);
+
+ /* Find blocked_pid's SERIALIZABLEXACT by linear search. */
+ for (sxact = FirstPredXact(); sxact != NULL; sxact = NextPredXact(sxact))
+ {
+ if (sxact->pid == blocked_pid)
+ break;
+ }
+
+ /* Did we find it, and is it currently waiting in GetSafeSnapshot? */
+ if (sxact != NULL && SxactIsDeferrableWaiting(sxact))
+ {
+ RWConflict possibleUnsafeConflict;
+
+ /* Traverse the list of possible unsafe conflicts collecting PIDs. */
+ possibleUnsafeConflict = (RWConflict)
+ SHMQueueNext(&sxact->possibleUnsafeConflicts,
+ &sxact->possibleUnsafeConflicts,
+ offsetof(RWConflictData, inLink));
+
+ while (possibleUnsafeConflict != NULL && num_written < output_size)
+ {
+ output[num_written++] = possibleUnsafeConflict->sxactOut->pid;
+ possibleUnsafeConflict = (RWConflict)
+ SHMQueueNext(&sxact->possibleUnsafeConflicts,
+ &possibleUnsafeConflict->inLink,
+ offsetof(RWConflictData, inLink));
+ }
+ }
+
+ LWLockRelease(SerializableXactHashLock);
+
+ return num_written;
+}
+
+/*
+ * Acquire a snapshot that can be used for the current transaction.
+ *
* Make sure we have a SERIALIZABLEXACT reference in MySerializableXact.
* It should be current for this process and be contained in PredXact.
+ *
+ * The passed-in Snapshot pointer should reference a static data area that
+ * can safely be passed to GetSnapshotData. The return value is actually
+ * always this same pointer; no new snapshot data structure is allocated
+ * within this function.
*/
Snapshot
-RegisterSerializableTransaction(Snapshot snapshot)
+GetSerializableTransactionSnapshot(Snapshot snapshot)
{
Assert(IsolationIsSerializable());
+ /*
+ * Can't use serializable mode while recovery is still active, as it is,
+ * for example, on a hot standby. We could get here despite the check in
+ * check_XactIsoLevel() if default_transaction_isolation is set to
+ * serializable, so phrase the hint accordingly.
+ */
+ if (RecoveryInProgress())
+ ereport(ERROR,
+ (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("cannot use serializable mode in a hot standby"),
+ errdetail("\"default_transaction_isolation\" is set to \"serializable\"."),
+ errhint("You can use \"SET default_transaction_isolation = 'repeatable read'\" to change the default.")));
+
/*
* A special optimization is available for SERIALIZABLE READ ONLY
* DEFERRABLE transactions -- we can wait for a suitable snapshot and
- * thereby avoid all SSI overhead once it's running..
+ * thereby avoid all SSI overhead once it's running.
*/
if (XactReadOnly && XactDeferrable)
return GetSafeSnapshot(snapshot);
- return RegisterSerializableTransactionInt(snapshot);
+ return GetSerializableTransactionSnapshotInt(snapshot,
+ NULL, InvalidPid);
+}
+
+/*
+ * Import a snapshot to be used for the current transaction.
+ *
+ * This is nearly the same as GetSerializableTransactionSnapshot, except that
+ * we don't take a new snapshot, but rather use the data we're handed.
+ *
+ * The caller must have verified that the snapshot came from a serializable
+ * transaction; and if we're read-write, the source transaction must not be
+ * read-only.
+ */
+void
+SetSerializableTransactionSnapshot(Snapshot snapshot,
+ VirtualTransactionId *sourcevxid,
+ int sourcepid)
+{
+ Assert(IsolationIsSerializable());
+
+ /*
+ * If this is called by parallel.c in a parallel worker, we don't want to
+ * create a SERIALIZABLEXACT just yet because the leader's
+ * SERIALIZABLEXACT will be installed with AttachSerializableXact(). We
+ * also don't want to reject SERIALIZABLE READ ONLY DEFERRABLE in this
+ * case, because the leader has already determined that the snapshot it
+ * has passed us is safe. So there is nothing for us to do.
+ */
+ if (IsParallelWorker())
+ return;
+
+ /*
+ * We do not allow SERIALIZABLE READ ONLY DEFERRABLE transactions to
+ * import snapshots, since there's no way to wait for a safe snapshot when
+ * we're using the snap we're told to. (XXX instead of throwing an error,
+ * we could just ignore the XactDeferrable flag?)
+ */
+ if (XactReadOnly && XactDeferrable)
+ ereport(ERROR,
+ (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("a snapshot-importing transaction must not be READ ONLY DEFERRABLE")));
+
+ (void) GetSerializableTransactionSnapshotInt(snapshot, sourcevxid,
+ sourcepid);
}
+/*
+ * Guts of GetSerializableTransactionSnapshot
+ *
+ * If sourcexid is valid, this is actually an import operation and we should
+ * skip calling GetSnapshotData, because the snapshot contents are already
+ * loaded up. HOWEVER: to avoid race conditions, we must check that the
+ * source xact is still running after we acquire SerializableXactHashLock.
+ * We do that by calling ProcArrayInstallImportedXmin.
+ */
static Snapshot
-RegisterSerializableTransactionInt(Snapshot snapshot)
+GetSerializableTransactionSnapshotInt(Snapshot snapshot,
+ VirtualTransactionId *sourcevxid,
+ int sourcepid)
{
PGPROC *proc;
VirtualTransactionId vxid;
SERIALIZABLEXACT *sxact,
*othersxact;
- HASHCTL hash_ctl;
/* We only do this for serializable transactions. Once. */
Assert(MySerializableXact == InvalidSerializableXact);
Assert(!RecoveryInProgress());
+ /*
+ * Since all parts of a serializable transaction must use the same
+ * snapshot, it is too late to establish one after a parallel operation
+ * has begun.
+ */
+ if (IsInParallelMode())
+ elog(ERROR, "cannot establish serializable snapshot during a parallel operation");
+
proc = MyProc;
Assert(proc != NULL);
GET_VXID_FROM_PGPROC(vxid, *proc);
/*
* First we get the sxact structure, which may involve looping and access
* to the "finished" list to free a structure for use.
+ *
+ * We must hold SerializableXactHashLock when taking/checking the snapshot
+ * to avoid race conditions, for much the same reasons that
+ * GetSnapshotData takes the ProcArrayLock. Since we might have to
+ * release SerializableXactHashLock to call SummarizeOldestCommittedSxact,
+ * this means we have to create the sxact first, which is a bit annoying
+ * (in particular, an elog(ERROR) in procarray.c would cause us to leak
+ * the sxact). Consider refactoring to avoid this.
*/
#ifdef TEST_OLDSERXID
SummarizeOldestCommittedSxact();
}
} while (!sxact);
- /* Get and register a snapshot */
- snapshot = GetSnapshotData(snapshot);
- snapshot = RegisterSnapshotOnOwner(snapshot, TopTransactionResourceOwner);
+ /* Get the snapshot, or check that it's safe to use */
+ if (!sourcevxid)
+ snapshot = GetSnapshotData(snapshot);
+ else if (!ProcArrayInstallImportedXmin(snapshot->xmin, sourcevxid))
+ {
+ ReleasePredXact(sxact);
+ LWLockRelease(SerializableXactHashLock);
+ ereport(ERROR,
+ (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
+ errmsg("could not import the requested snapshot"),
+ errdetail("The source process with PID %d is not running anymore.",
+ sourcepid)));
+ }
/*
* If there are no serializable transactions which are not read-only, we
/* Initialize the structure. */
sxact->vxid = vxid;
sxact->SeqNo.lastCommitBeforeSnapshot = PredXact->LastSxactCommitSeqNo;
+ sxact->prepareSeqNo = InvalidSerCommitSeqNo;
sxact->commitSeqNo = InvalidSerCommitSeqNo;
SHMQueueInit(&(sxact->outConflicts));
SHMQueueInit(&(sxact->inConflicts));
othersxact != NULL;
othersxact = NextPredXact(othersxact))
{
- if (!SxactIsOnFinishedList(othersxact) &&
- !SxactIsReadOnly(othersxact))
+ if (!SxactIsCommitted(othersxact)
+ && !SxactIsDoomed(othersxact)
+ && !SxactIsReadOnly(othersxact))
{
SetPossibleUnsafeConflict(sxact, othersxact);
}
}
MySerializableXact = sxact;
+ MyXactDidWrite = false; /* haven't written anything yet */
LWLockRelease(SerializableXactHashLock);
+ CreateLocalPredicateLockHash();
+
+ return snapshot;
+}
+
+static void
+CreateLocalPredicateLockHash(void)
+{
+ HASHCTL hash_ctl;
+
/* Initialize the backend-local hash table of parent locks */
Assert(LocalPredicateLockHash == NULL);
MemSet(&hash_ctl, 0, sizeof(hash_ctl));
hash_ctl.keysize = sizeof(PREDICATELOCKTARGETTAG);
hash_ctl.entrysize = sizeof(LOCALPREDICATELOCK);
- hash_ctl.hash = tag_hash;
LocalPredicateLockHash = hash_create("Local predicate lock",
max_predicate_locks_per_xact,
&hash_ctl,
- HASH_ELEM | HASH_FUNCTION);
-
- return snapshot;
+ HASH_ELEM | HASH_BLOBS);
}
/*
* Also store it for easy reference in MySerializableXact.
*/
void
-RegisterPredicateLockingXid(const TransactionId xid)
+RegisterPredicateLockingXid(TransactionId xid)
{
SERIALIZABLEXIDTAG sxidtag;
SERIALIZABLEXID *sxid;
if (MySerializableXact == InvalidSerializableXact)
return;
- /* This should only be done once per transaction. */
- Assert(MySerializableXact->topXid == InvalidTransactionId);
-
/* We should have a valid XID and be at the top level. */
Assert(TransactionIdIsValid(xid));
+ LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
+
+ /* This should only be done once per transaction. */
+ Assert(MySerializableXact->topXid == InvalidTransactionId);
+
MySerializableXact->topXid = xid;
sxidtag.xid = xid;
- LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
sxid = (SERIALIZABLEXID *) hash_search(SerializableXidHash,
&sxidtag,
HASH_ENTER, &found);
- if (!sxid)
- /* This should not be possible, based on allocation. */
- ereport(ERROR,
- (errcode(ERRCODE_OUT_OF_MEMORY),
- errmsg("out of shared memory")));
-
Assert(!found);
/* Initialize the structure. */
- sxid->myXact = (SERIALIZABLEXACT *) MySerializableXact;
+ sxid->myXact = MySerializableXact;
LWLockRelease(SerializableXactHashLock);
}
* One use is to support proper behavior during GiST index vacuum.
*/
bool
-PageIsPredicateLocked(const Relation relation, const BlockNumber blkno)
+PageIsPredicateLocked(Relation relation, BlockNumber blkno)
{
PREDICATELOCKTARGETTAG targettag;
uint32 targettaghash;
- LWLockId partitionLock;
+ LWLock *partitionLock;
PREDICATELOCKTARGET *target;
SET_PREDICATELOCKTARGETTAG_PAGE(targettag,
case PREDLOCKTAG_PAGE:
/* parent lock is relation lock */
SET_PREDICATELOCKTARGETTAG_RELATION(*parent,
- GET_PREDICATELOCKTARGETTAG_DB(*tag),
- GET_PREDICATELOCKTARGETTAG_RELATION(*tag));
+ GET_PREDICATELOCKTARGETTAG_DB(*tag),
+ GET_PREDICATELOCKTARGETTAG_RELATION(*tag));
return true;
case PREDLOCKTAG_TUPLE:
/* parent lock is page lock */
SET_PREDICATELOCKTARGETTAG_PAGE(*parent,
- GET_PREDICATELOCKTARGETTAG_DB(*tag),
- GET_PREDICATELOCKTARGETTAG_RELATION(*tag),
- GET_PREDICATELOCKTARGETTAG_PAGE(*tag));
+ GET_PREDICATELOCKTARGETTAG_DB(*tag),
+ GET_PREDICATELOCKTARGETTAG_RELATION(*tag),
+ GET_PREDICATELOCKTARGETTAG_PAGE(*tag));
return true;
}
}
/*
- * Check whether both the list of related predicate locks and the pointer to
- * a prior version of the row (if this is a tuple lock target) are empty for
- * a predicate lock target, and remove the target if they are.
+ * Remove the dummy entry from the predicate lock target hash, to free up some
+ * scratch space. The caller must be holding SerializablePredicateLockListLock,
+ * and must restore the entry with RestoreScratchTarget() before releasing the
+ * lock.
+ *
+ * If lockheld is true, the caller is already holding the partition lock
+ * of the partition containing the scratch entry.
+ */
+static void
+RemoveScratchTarget(bool lockheld)
+{
+ bool found;
+
+ Assert(LWLockHeldByMe(SerializablePredicateLockListLock));
+
+ if (!lockheld)
+ LWLockAcquire(ScratchPartitionLock, LW_EXCLUSIVE);
+ hash_search_with_hash_value(PredicateLockTargetHash,
+ &ScratchTargetTag,
+ ScratchTargetTagHash,
+ HASH_REMOVE, &found);
+ Assert(found);
+ if (!lockheld)
+ LWLockRelease(ScratchPartitionLock);
+}
+
+/*
+ * Re-insert the dummy entry in predicate lock target hash.
+ */
+static void
+RestoreScratchTarget(bool lockheld)
+{
+ bool found;
+
+ Assert(LWLockHeldByMe(SerializablePredicateLockListLock));
+
+ if (!lockheld)
+ LWLockAcquire(ScratchPartitionLock, LW_EXCLUSIVE);
+ hash_search_with_hash_value(PredicateLockTargetHash,
+ &ScratchTargetTag,
+ ScratchTargetTagHash,
+ HASH_ENTER, &found);
+ Assert(!found);
+ if (!lockheld)
+ LWLockRelease(ScratchPartitionLock);
+}
+
+/*
+ * Check whether the list of related predicate locks is empty for a
+ * predicate lock target, and remove the target if it is.
*/
static void
RemoveTargetIfNoLongerUsed(PREDICATELOCKTARGET *target, uint32 targettaghash)
{
- PREDICATELOCKTARGET *rmtarget;
+ PREDICATELOCKTARGET *rmtarget PG_USED_FOR_ASSERTS_ONLY;
Assert(LWLockHeldByMe(SerializablePredicateLockListLock));
/*
* Delete child target locks owned by this process.
* This implementation is assuming that the usage of each target tag field
- * is uniform. No need to make this hard if we don't have to.
+ * is uniform. No need to make this hard if we don't have to.
*
- * We aren't acquiring lightweight locks for the predicate lock or lock
+ * We acquire an LWLock in the case of parallel mode, because worker
+ * backends have access to the leader's SERIALIZABLEXACT. Otherwise,
+ * we aren't acquiring LWLocks for the predicate lock or lock
* target structures associated with this transaction unless we're going
* to modify them, because no other process is permitted to modify our
* locks.
PREDICATELOCK *predlock;
LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
- sxact = (SERIALIZABLEXACT *) MySerializableXact;
+ sxact = MySerializableXact;
+ if (IsInParallelMode())
+ LWLockAcquire(&sxact->predicateLockListLock, LW_EXCLUSIVE);
predlock = (PREDICATELOCK *)
SHMQueueNext(&(sxact->predicateLocks),
&(sxact->predicateLocks),
if (TargetTagIsCoveredBy(oldtargettag, *newtargettag))
{
uint32 oldtargettaghash;
- LWLockId partitionLock;
- PREDICATELOCK *rmpredlock;
+ LWLock *partitionLock;
+ PREDICATELOCK *rmpredlock PG_USED_FOR_ASSERTS_ONLY;
oldtargettaghash = PredicateLockTargetTagHashCode(&oldtargettag);
partitionLock = PredicateLockHashPartitionLock(oldtargettaghash);
predlock = nextpredlock;
}
+ if (IsInParallelMode())
+ LWLockRelease(&sxact->predicateLockListLock);
LWLockRelease(SerializablePredicateLockListLock);
}
/*
- * Returns the promotion threshold for a given predicate lock
- * target. This is the number of descendant locks required to promote
- * to the specified tag. Note that the threshold includes non-direct
- * descendants, e.g. both tuples and pages for a relation lock.
+ * Returns the promotion limit for a given predicate lock target. This is the
+ * max number of descendant locks allowed before promoting to the specified
+ * tag. Note that the limit includes non-direct descendants (e.g., both tuples
+ * and pages for a relation lock).
+ *
+ * Currently the default limit is 2 for a page lock, and half of the value of
+ * max_pred_locks_per_transaction - 1 for a relation lock, to match behavior
+ * of earlier releases when upgrading.
*
- * TODO SSI: We should do something more intelligent about what the
- * thresholds are, either making it proportional to the number of
- * tuples in a page & pages in a relation, or at least making it a
- * GUC. Currently the threshold is 3 for a page lock, and
- * max_pred_locks_per_transaction/2 for a relation lock, chosen
- * entirely arbitrarily (and without benchmarking).
+ * TODO SSI: We should probably add additional GUCs to allow a maximum ratio
+ * of page and tuple locks based on the pages in a relation, and the maximum
+ * ratio of tuple locks to tuples in a page. This would provide more
+ * generally "balanced" allocation of locks to where they are most useful,
+ * while still allowing the absolute numbers to prevent one relation from
+ * tying up all predicate lock resources.
*/
static int
-PredicateLockPromotionThreshold(const PREDICATELOCKTARGETTAG *tag)
+MaxPredicateChildLocks(const PREDICATELOCKTARGETTAG *tag)
{
switch (GET_PREDICATELOCKTARGETTAG_TYPE(*tag))
{
case PREDLOCKTAG_RELATION:
- return max_predicate_locks_per_xact / 2;
+ return max_predicate_locks_per_relation < 0
+ ? (max_predicate_locks_per_xact
+ / (-max_predicate_locks_per_relation)) - 1
+ : max_predicate_locks_per_relation;
case PREDLOCKTAG_PAGE:
- return 3;
+ return max_predicate_locks_per_page;
case PREDLOCKTAG_TUPLE:
else
parentlock->childLocks++;
- if (parentlock->childLocks >=
- PredicateLockPromotionThreshold(&targettag))
+ if (parentlock->childLocks >
+ MaxPredicateChildLocks(&targettag))
{
/*
* We should promote to this parent lock. Continue to check its
{
uint32 targettaghash;
LOCALPREDICATELOCK *parentlock,
- *rmlock;
+ *rmlock PG_USED_FOR_ASSERTS_ONLY;
parenttag = nexttag;
targettaghash = PredicateLockTargetTagHashCode(&parenttag);
PREDICATELOCKTARGET *target;
PREDICATELOCKTAG locktag;
PREDICATELOCK *lock;
- LWLockId partitionLock;
+ LWLock *partitionLock;
bool found;
partitionLock = PredicateLockHashPartitionLock(targettaghash);
LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
+ if (IsInParallelMode())
+ LWLockAcquire(&sxact->predicateLockListLock, LW_EXCLUSIVE);
LWLockAcquire(partitionLock, LW_EXCLUSIVE);
/* Make sure that the target is represented. */
target = (PREDICATELOCKTARGET *)
hash_search_with_hash_value(PredicateLockTargetHash,
targettag, targettaghash,
- HASH_ENTER, &found);
+ HASH_ENTER_NULL, &found);
if (!target)
ereport(ERROR,
(errcode(ERRCODE_OUT_OF_MEMORY),
locktag.myXact = sxact;
lock = (PREDICATELOCK *)
hash_search_with_hash_value(PredicateLockHash, &locktag,
- PredicateLockHashCodeFromTargetHashCode(&locktag, targettaghash),
- HASH_ENTER, &found);
+ PredicateLockHashCodeFromTargetHashCode(&locktag, targettaghash),
+ HASH_ENTER_NULL, &found);
if (!lock)
ereport(ERROR,
(errcode(ERRCODE_OUT_OF_MEMORY),
SHMQueueInsertBefore(&(target->predicateLocks), &(lock->targetLink));
SHMQueueInsertBefore(&(sxact->predicateLocks),
&(lock->xactLink));
- lock->commitSeqNo = 0;
+ lock->commitSeqNo = InvalidSerCommitSeqNo;
}
LWLockRelease(partitionLock);
+ if (IsInParallelMode())
+ LWLockRelease(&sxact->predicateLockListLock);
LWLockRelease(SerializablePredicateLockListLock);
}
locallock->childLocks = 0;
/* Actually create the lock */
- CreatePredicateLock(targettag, targettaghash,
- (SERIALIZABLEXACT *) MySerializableXact);
+ CreatePredicateLock(targettag, targettaghash, MySerializableXact);
/*
* Lock has been acquired. Check whether it should be promoted to a
* Clear any finer-grained predicate locks this session has on the relation.
*/
void
-PredicateLockRelation(const Relation relation)
+PredicateLockRelation(Relation relation, Snapshot snapshot)
{
PREDICATELOCKTARGETTAG tag;
- if (SkipSerialization(relation))
+ if (!SerializationNeededForRead(relation, snapshot))
return;
SET_PREDICATELOCKTARGETTAG_RELATION(tag,
* Clear any finer-grained predicate locks this session has on the relation.
*/
void
-PredicateLockPage(const Relation relation, const BlockNumber blkno)
+PredicateLockPage(Relation relation, BlockNumber blkno, Snapshot snapshot)
{
PREDICATELOCKTARGETTAG tag;
- if (SkipSerialization(relation))
+ if (!SerializationNeededForRead(relation, snapshot))
return;
SET_PREDICATELOCKTARGETTAG_PAGE(tag,
* Skip if this is a temporary table.
*/
void
-PredicateLockTuple(const Relation relation, const HeapTuple tuple)
+PredicateLockTuple(Relation relation, HeapTuple tuple, Snapshot snapshot)
{
PREDICATELOCKTARGETTAG tag;
ItemPointer tid;
TransactionId targetxmin;
- if (SkipSerialization(relation))
+ if (!SerializationNeededForRead(relation, snapshot))
return;
/*
*/
if (relation->rd_index == NULL)
{
- TransactionId myxid;
+ TransactionId myxid;
targetxmin = HeapTupleHeaderGetXmin(tuple->t_data);
if (TransactionIdFollowsOrEquals(targetxmin, TransactionXmin))
{
TransactionId xid = SubTransGetTopmostTransaction(targetxmin);
+
if (TransactionIdEquals(xid, myxid))
{
/* We wrote it; we already have a write lock. */
}
}
}
- else
- targetxmin = InvalidTransactionId;
/*
- * Do quick-but-not-definitive test for a relation lock first. This will
+ * Do quick-but-not-definitive test for a relation lock first. This will
* never cause a return when the relation is *not* locked, but will
* occasionally let the check continue when there really *is* a relation
* level lock.
relation->rd_node.dbNode,
relation->rd_id,
ItemPointerGetBlockNumber(tid),
- ItemPointerGetOffsetNumber(tid),
- targetxmin);
+ ItemPointerGetOffsetNumber(tid));
PredicateLockAcquire(&tag);
}
-/*
- * If the old tuple has any predicate locks, copy them to the new target.
- *
- * This is called at an UPDATE, where any predicate locks held on the old
- * tuple need to be copied to the new tuple, because logically they both
- * represent the same row. A lock taken before the update must conflict
- * with anyone locking the same row after the update.
- */
-void
-PredicateLockTupleRowVersionLink(const Relation relation,
- const HeapTuple oldTuple,
- const HeapTuple newTuple)
-{
- PREDICATELOCKTARGETTAG oldtupletag;
- PREDICATELOCKTARGETTAG oldpagetag;
- PREDICATELOCKTARGETTAG newtupletag;
- BlockNumber oldblk,
- newblk;
- OffsetNumber oldoff,
- newoff;
- TransactionId oldxmin,
- newxmin;
-
- oldblk = ItemPointerGetBlockNumber(&(oldTuple->t_self));
- oldoff = ItemPointerGetOffsetNumber(&(oldTuple->t_self));
- oldxmin = HeapTupleHeaderGetXmin(oldTuple->t_data);
-
- newblk = ItemPointerGetBlockNumber(&(newTuple->t_self));
- newoff = ItemPointerGetOffsetNumber(&(newTuple->t_self));
- newxmin = HeapTupleHeaderGetXmin(newTuple->t_data);
-
- SET_PREDICATELOCKTARGETTAG_TUPLE(oldtupletag,
- relation->rd_node.dbNode,
- relation->rd_id,
- oldblk,
- oldoff,
- oldxmin);
-
- SET_PREDICATELOCKTARGETTAG_PAGE(oldpagetag,
- relation->rd_node.dbNode,
- relation->rd_id,
- oldblk);
-
- SET_PREDICATELOCKTARGETTAG_TUPLE(newtupletag,
- relation->rd_node.dbNode,
- relation->rd_id,
- newblk,
- newoff,
- newxmin);
-
- /*
- * A page-level lock on the page containing the old tuple counts too.
- * Anyone holding a lock on the page is logically holding a lock on
- * the old tuple, so we need to acquire a lock on his behalf on the
- * new tuple too. However, if the new tuple is on the same page as the
- * old one, the old page-level lock already covers the new tuple.
- *
- * A relation-level lock always covers both tuple versions, so we don't
- * need to worry about those here.
- */
- LWLockAcquire(SerializablePredicateLockListLock, LW_EXCLUSIVE);
-
- TransferPredicateLocksToNewTarget(oldtupletag, newtupletag, false);
- if (newblk != oldblk)
- TransferPredicateLocksToNewTarget(oldpagetag, newtupletag, false);
-
- LWLockRelease(SerializablePredicateLockListLock);
-}
-
/*
* DeleteLockTarget
PREDICATELOCK *nextpredlock;
bool found;
- Assert(LWLockHeldByMe(SerializablePredicateLockListLock));
+ Assert(LWLockHeldByMeInMode(SerializablePredicateLockListLock,
+ LW_EXCLUSIVE));
Assert(LWLockHeldByMe(PredicateLockHashPartitionLock(targettaghash)));
predlock = (PREDICATELOCK *)
*
* Returns true on success, or false if we ran out of shared memory to
* allocate the new target or locks. Guaranteed to always succeed if
- * removeOld is set (by using the reserved entry in
- * PredicateLockTargetHash for scratch space).
+ * removeOld is set (by using the scratch entry in PredicateLockTargetHash
+ * for scratch space).
*
* Warning: the "removeOld" option should be used only with care,
* because this function does not (indeed, can not) update other
* covers it, or if we are absolutely certain that no one will need to
* refer to that lock in the future.
*
- * Caller must hold SerializablePredicateLockListLock.
+ * Caller must hold SerializablePredicateLockListLock exclusively.
*/
static bool
-TransferPredicateLocksToNewTarget(const PREDICATELOCKTARGETTAG oldtargettag,
- const PREDICATELOCKTARGETTAG newtargettag,
+TransferPredicateLocksToNewTarget(PREDICATELOCKTARGETTAG oldtargettag,
+ PREDICATELOCKTARGETTAG newtargettag,
bool removeOld)
{
uint32 oldtargettaghash;
- LWLockId oldpartitionLock;
+ LWLock *oldpartitionLock;
PREDICATELOCKTARGET *oldtarget;
uint32 newtargettaghash;
- LWLockId newpartitionLock;
+ LWLock *newpartitionLock;
bool found;
bool outOfShmem = false;
- uint32 reservedtargettaghash;
- LWLockId reservedpartitionLock;
-
- Assert(LWLockHeldByMe(SerializablePredicateLockListLock));
+ Assert(LWLockHeldByMeInMode(SerializablePredicateLockListLock,
+ LW_EXCLUSIVE));
oldtargettaghash = PredicateLockTargetTagHashCode(&oldtargettag);
newtargettaghash = PredicateLockTargetTagHashCode(&newtargettag);
oldpartitionLock = PredicateLockHashPartitionLock(oldtargettaghash);
newpartitionLock = PredicateLockHashPartitionLock(newtargettaghash);
- reservedtargettaghash = 0; /* Quiet compiler warnings. */
- reservedpartitionLock = 0; /* Quiet compiler warnings. */
-
if (removeOld)
{
/*
- * Remove the reserved entry to give us scratch space, so we know
- * we'll be able to create the new lock target.
+ * Remove the dummy entry to give us scratch space, so we know we'll
+ * be able to create the new lock target.
*/
- reservedtargettaghash = PredicateLockTargetTagHashCode(&ReservedTargetTag);
- reservedpartitionLock = PredicateLockHashPartitionLock(reservedtargettaghash);
- LWLockAcquire(reservedpartitionLock, LW_EXCLUSIVE);
- hash_search_with_hash_value(PredicateLockTargetHash,
- &ReservedTargetTag,
- reservedtargettaghash,
- HASH_REMOVE, &found);
- Assert(found);
- LWLockRelease(reservedpartitionLock);
+ RemoveScratchTarget(false);
}
/*
newpredlocktag.myTarget = newtarget;
+ /*
+ * Loop through all the locks on the old target, replacing them with
+ * locks on the new target.
+ */
oldpredlock = (PREDICATELOCK *)
SHMQueueNext(&(oldtarget->predicateLocks),
&(oldtarget->predicateLocks),
SHM_QUEUE *predlocktargetlink;
PREDICATELOCK *nextpredlock;
PREDICATELOCK *newpredlock;
+ SerCommitSeqNo oldCommitSeqNo = oldpredlock->commitSeqNo;
predlocktargetlink = &(oldpredlock->targetLink);
nextpredlock = (PREDICATELOCK *)
hash_search_with_hash_value
(PredicateLockHash,
&oldpredlock->tag,
- PredicateLockHashCodeFromTargetHashCode(&oldpredlock->tag,
- oldtargettaghash),
+ PredicateLockHashCodeFromTargetHashCode(&oldpredlock->tag,
+ oldtargettaghash),
HASH_REMOVE, &found);
Assert(found);
}
-
newpredlock = (PREDICATELOCK *)
- hash_search_with_hash_value
- (PredicateLockHash,
- &newpredlocktag,
- PredicateLockHashCodeFromTargetHashCode(&newpredlocktag,
- newtargettaghash),
- HASH_ENTER_NULL, &found);
+ hash_search_with_hash_value(PredicateLockHash,
+ &newpredlocktag,
+ PredicateLockHashCodeFromTargetHashCode(&newpredlocktag,
+ newtargettaghash),
+ HASH_ENTER_NULL,
+ &found);
if (!newpredlock)
{
/* Out of shared memory. Undo what we've done so far. */
&(newpredlock->targetLink));
SHMQueueInsertBefore(&(newpredlocktag.myXact->predicateLocks),
&(newpredlock->xactLink));
- newpredlock->commitSeqNo = InvalidSerCommitSeqNo;
+ newpredlock->commitSeqNo = oldCommitSeqNo;
+ }
+ else
+ {
+ if (newpredlock->commitSeqNo < oldCommitSeqNo)
+ newpredlock->commitSeqNo = oldCommitSeqNo;
}
+ Assert(newpredlock->commitSeqNo != 0);
+ Assert((newpredlock->commitSeqNo == InvalidSerCommitSeqNo)
+ || (newpredlock->tag.myXact == OldCommittedSxact));
+
oldpredlock = nextpredlock;
}
LWLockRelease(SerializableXactHashLock);
/* We shouldn't run out of memory if we're moving locks */
Assert(!outOfShmem);
- /* Put the reserved entry back */
- LWLockAcquire(reservedpartitionLock, LW_EXCLUSIVE);
- hash_search_with_hash_value(PredicateLockTargetHash,
- &ReservedTargetTag,
- reservedtargettaghash,
- HASH_ENTER, &found);
- Assert(!found);
- LWLockRelease(reservedpartitionLock);
+ /* Put the scratch entry back */
+ RestoreScratchTarget(false);
}
return !outOfShmem;
}
+/*
+ * Drop all predicate locks of any granularity from the specified relation,
+ * which can be a heap relation or an index relation. If 'transfer' is true,
+ * acquire a relation lock on the heap for any transactions with any lock(s)
+ * on the specified relation.
+ *
+ * This requires grabbing a lot of LW locks and scanning the entire lock
+ * target table for matches. That makes this more expensive than most
+ * predicate lock management functions, but it will only be called for DDL
+ * type commands that are expensive anyway, and there are fast returns when
+ * no serializable transactions are active or the relation is temporary.
+ *
+ * We don't use the TransferPredicateLocksToNewTarget function because it
+ * acquires its own locks on the partitions of the two targets involved,
+ * and we'll already be holding all partition locks.
+ *
+ * We can't throw an error from here, because the call could be from a
+ * transaction which is not serializable.
+ *
+ * NOTE: This is currently only called with transfer set to true, but that may
+ * change. If we decide to clean up the locks from a table on commit of a
+ * transaction which executed DROP TABLE, the false condition will be useful.
+ */
+static void
+DropAllPredicateLocksFromTable(Relation relation, bool transfer)
+{
+ HASH_SEQ_STATUS seqstat;
+ PREDICATELOCKTARGET *oldtarget;
+ PREDICATELOCKTARGET *heaptarget;
+ Oid dbId;
+ Oid relId;
+ Oid heapId;
+ int i;
+ bool isIndex;
+ bool found;
+ uint32 heaptargettaghash;
+
+ /*
+ * Bail out quickly if there are no serializable transactions running.
+ * It's safe to check this without taking locks because the caller is
+ * holding an ACCESS EXCLUSIVE lock on the relation. No new locks which
+ * would matter here can be acquired while that is held.
+ */
+ if (!TransactionIdIsValid(PredXact->SxactGlobalXmin))
+ return;
+
+ if (!PredicateLockingNeededForRelation(relation))
+ return;
+
+ dbId = relation->rd_node.dbNode;
+ relId = relation->rd_id;
+ if (relation->rd_index == NULL)
+ {
+ isIndex = false;
+ heapId = relId;
+ }
+ else
+ {
+ isIndex = true;
+ heapId = relation->rd_index->indrelid;
+ }
+ Assert(heapId != InvalidOid);
+ Assert(transfer || !isIndex); /* index OID only makes sense with
+ * transfer */
+
+ /* Retrieve first time needed, then keep. */
+ heaptargettaghash = 0;
+ heaptarget = NULL;
+
+ /* Acquire locks on all lock partitions */
+ LWLockAcquire(SerializablePredicateLockListLock, LW_EXCLUSIVE);
+ for (i = 0; i < NUM_PREDICATELOCK_PARTITIONS; i++)
+ LWLockAcquire(PredicateLockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
+ LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
+
+ /*
+ * Remove the dummy entry to give us scratch space, so we know we'll be
+ * able to create the new lock target.
+ */
+ if (transfer)
+ RemoveScratchTarget(true);
+
+ /* Scan through target map */
+ hash_seq_init(&seqstat, PredicateLockTargetHash);
+
+ while ((oldtarget = (PREDICATELOCKTARGET *) hash_seq_search(&seqstat)))
+ {
+ PREDICATELOCK *oldpredlock;
+
+ /*
+ * Check whether this is a target which needs attention.
+ */
+ if (GET_PREDICATELOCKTARGETTAG_RELATION(oldtarget->tag) != relId)
+ continue; /* wrong relation id */
+ if (GET_PREDICATELOCKTARGETTAG_DB(oldtarget->tag) != dbId)
+ continue; /* wrong database id */
+ if (transfer && !isIndex
+ && GET_PREDICATELOCKTARGETTAG_TYPE(oldtarget->tag) == PREDLOCKTAG_RELATION)
+ continue; /* already the right lock */
+
+ /*
+ * If we made it here, we have work to do. We make sure the heap
+ * relation lock exists, then we walk the list of predicate locks for
+ * the old target we found, moving all locks to the heap relation lock
+ * -- unless they already hold that.
+ */
+
+ /*
+ * First make sure we have the heap relation target. We only need to
+ * do this once.
+ */
+ if (transfer && heaptarget == NULL)
+ {
+ PREDICATELOCKTARGETTAG heaptargettag;
+
+ SET_PREDICATELOCKTARGETTAG_RELATION(heaptargettag, dbId, heapId);
+ heaptargettaghash = PredicateLockTargetTagHashCode(&heaptargettag);
+ heaptarget = hash_search_with_hash_value(PredicateLockTargetHash,
+ &heaptargettag,
+ heaptargettaghash,
+ HASH_ENTER, &found);
+ if (!found)
+ SHMQueueInit(&heaptarget->predicateLocks);
+ }
+
+ /*
+ * Loop through all the locks on the old target, replacing them with
+ * locks on the new target.
+ */
+ oldpredlock = (PREDICATELOCK *)
+ SHMQueueNext(&(oldtarget->predicateLocks),
+ &(oldtarget->predicateLocks),
+ offsetof(PREDICATELOCK, targetLink));
+ while (oldpredlock)
+ {
+ PREDICATELOCK *nextpredlock;
+ PREDICATELOCK *newpredlock;
+ SerCommitSeqNo oldCommitSeqNo;
+ SERIALIZABLEXACT *oldXact;
+
+ nextpredlock = (PREDICATELOCK *)
+ SHMQueueNext(&(oldtarget->predicateLocks),
+ &(oldpredlock->targetLink),
+ offsetof(PREDICATELOCK, targetLink));
+
+ /*
+ * Remove the old lock first. This avoids the chance of running
+ * out of lock structure entries for the hash table.
+ */
+ oldCommitSeqNo = oldpredlock->commitSeqNo;
+ oldXact = oldpredlock->tag.myXact;
+
+ SHMQueueDelete(&(oldpredlock->xactLink));
+
+ /*
+ * No need for retail delete from oldtarget list, we're removing
+ * the whole target anyway.
+ */
+ hash_search(PredicateLockHash,
+ &oldpredlock->tag,
+ HASH_REMOVE, &found);
+ Assert(found);
+
+ if (transfer)
+ {
+ PREDICATELOCKTAG newpredlocktag;
+
+ newpredlocktag.myTarget = heaptarget;
+ newpredlocktag.myXact = oldXact;
+ newpredlock = (PREDICATELOCK *)
+ hash_search_with_hash_value(PredicateLockHash,
+ &newpredlocktag,
+ PredicateLockHashCodeFromTargetHashCode(&newpredlocktag,
+ heaptargettaghash),
+ HASH_ENTER,
+ &found);
+ if (!found)
+ {
+ SHMQueueInsertBefore(&(heaptarget->predicateLocks),
+ &(newpredlock->targetLink));
+ SHMQueueInsertBefore(&(newpredlocktag.myXact->predicateLocks),
+ &(newpredlock->xactLink));
+ newpredlock->commitSeqNo = oldCommitSeqNo;
+ }
+ else
+ {
+ if (newpredlock->commitSeqNo < oldCommitSeqNo)
+ newpredlock->commitSeqNo = oldCommitSeqNo;
+ }
+
+ Assert(newpredlock->commitSeqNo != 0);
+ Assert((newpredlock->commitSeqNo == InvalidSerCommitSeqNo)
+ || (newpredlock->tag.myXact == OldCommittedSxact));
+ }
+
+ oldpredlock = nextpredlock;
+ }
+
+ hash_search(PredicateLockTargetHash, &oldtarget->tag, HASH_REMOVE,
+ &found);
+ Assert(found);
+ }
+
+ /* Put the scratch entry back */
+ if (transfer)
+ RestoreScratchTarget(true);
+
+ /* Release locks in reverse order */
+ LWLockRelease(SerializableXactHashLock);
+ for (i = NUM_PREDICATELOCK_PARTITIONS - 1; i >= 0; i--)
+ LWLockRelease(PredicateLockHashPartitionLockByIndex(i));
+ LWLockRelease(SerializablePredicateLockListLock);
+}
+
+/*
+ * TransferPredicateLocksToHeapRelation
+ * For all transactions, transfer all predicate locks for the given
+ * relation to a single relation lock on the heap.
+ */
+void
+TransferPredicateLocksToHeapRelation(Relation relation)
+{
+ DropAllPredicateLocksFromTable(relation, true);
+}
+
/*
* PredicateLockPageSplit
* which hold the locks getting in and noticing.
*/
void
-PredicateLockPageSplit(const Relation relation, const BlockNumber oldblkno,
- const BlockNumber newblkno)
+PredicateLockPageSplit(Relation relation, BlockNumber oldblkno,
+ BlockNumber newblkno)
{
PREDICATELOCKTARGETTAG oldtargettag;
PREDICATELOCKTARGETTAG newtargettag;
bool success;
- if (SkipSplitTracking(relation))
+ /*
+ * Bail out quickly if there are no serializable transactions running.
+ *
+ * It's safe to do this check without taking any additional locks. Even if
+ * a serializable transaction starts concurrently, we know it can't take
+ * any SIREAD locks on the page being split because the caller is holding
+ * the associated buffer page lock. Memory reordering isn't an issue; the
+ * memory barrier in the LWLock acquisition guarantees that this read
+ * occurs while the buffer page lock is held.
+ */
+ if (!TransactionIdIsValid(PredXact->SxactGlobalXmin))
+ return;
+
+ if (!PredicateLockingNeededForRelation(relation))
return;
Assert(oldblkno != newblkno);
/*
* Move the locks to the parent. This shouldn't fail.
*
- * Note that here we are removing locks held by other
- * backends, leading to a possible inconsistency in their
- * local lock hash table. This is OK because we're replacing
- * it with a lock that covers the old one.
+ * Note that here we are removing locks held by other backends,
+ * leading to a possible inconsistency in their local lock hash table.
+ * This is OK because we're replacing it with a lock that covers the
+ * old one.
*/
success = TransferPredicateLocksToNewTarget(oldtargettag,
newtargettag,
* occurs in the context of another transaction isolation level.
*/
void
-PredicateLockPageCombine(const Relation relation, const BlockNumber oldblkno,
- const BlockNumber newblkno)
+PredicateLockPageCombine(Relation relation, BlockNumber oldblkno,
+ BlockNumber newblkno)
{
/*
- * Page combines differ from page splits in that we ought to be
- * able to remove the locks on the old page after transferring
- * them to the new page, instead of duplicating them. However,
- * because we can't edit other backends' local lock tables,
- * removing the old lock would leave them with an entry in their
- * LocalPredicateLockHash for a lock they're not holding, which
- * isn't acceptable. So we wind up having to do the same work as a
- * page split, acquiring a lock on the new page and keeping the old
- * page locked too. That can lead to some false positives, but
- * should be rare in practice.
+ * Page combines differ from page splits in that we ought to be able to
+ * remove the locks on the old page after transferring them to the new
+ * page, instead of duplicating them. However, because we can't edit other
+ * backends' local lock tables, removing the old lock would leave them
+ * with an entry in their LocalPredicateLockHash for a lock they're not
+ * holding, which isn't acceptable. So we wind up having to do the same
+ * work as a page split, acquiring a lock on the new page and keeping the
+ * old page locked too. That can lead to some false positives, but should
+ * be rare in practice.
*/
PredicateLockPageSplit(relation, oldblkno, newblkno);
}
/*
- * Walk the hash table and find the new xmin.
+ * Walk the list of in-progress serializable transactions and find the new
+ * xmin.
*/
static void
SetNewSxactGlobalXmin(void)
* up in some relatively timely fashion.
*
* If this transaction is committing and is holding any predicate locks,
- * it must be added to a list of completed serializable transaction still
+ * it must be added to a list of completed serializable transactions still
* holding locks.
+ *
+ * If isReadOnlySafe is true, then predicate locks are being released before
+ * the end of the transaction because MySerializableXact has been determined
+ * to be RO_SAFE. In non-parallel mode we can release it completely, but it
+ * in parallel mode we partially release the SERIALIZABLEXACT and keep it
+ * around until the end of the transaction, allowing each backend to clear its
+ * MySerializableXact variable and benefit from the optimization in its own
+ * time.
*/
void
-ReleasePredicateLocks(const bool isCommit)
+ReleasePredicateLocks(bool isCommit, bool isReadOnlySafe)
{
bool needToClear;
RWConflict conflict,
/*
* We can't trust XactReadOnly here, because a transaction which started
* as READ WRITE can show as READ ONLY later, e.g., within
- * substransactions. We want to flag a transaction as READ ONLY if it
+ * subtransactions. We want to flag a transaction as READ ONLY if it
* commits without writing so that de facto READ ONLY transactions get the
* benefit of some RO optimizations, so we will use this local variable to
* get some cleanup logic right which is based on whether the transaction
*/
bool topLevelIsDeclaredReadOnly;
+ /* We can't be both committing and releasing early due to RO_SAFE. */
+ Assert(!(isCommit && isReadOnlySafe));
+
+ /* Are we at the end of a transaction, that is, a commit or abort? */
+ if (!isReadOnlySafe)
+ {
+ /*
+ * Parallel workers mustn't release predicate locks at the end of
+ * their transaction. The leader will do that at the end of its
+ * transaction.
+ */
+ if (IsParallelWorker())
+ {
+ ReleasePredicateLocksLocal();
+ return;
+ }
+
+ /*
+ * By the time the leader in a parallel query reaches end of
+ * transaction, it has waited for all workers to exit.
+ */
+ Assert(!ParallelContextActive());
+
+ /*
+ * If the leader in a parallel query earlier stashed a partially
+ * released SERIALIZABLEXACT for final clean-up at end of transaction
+ * (because workers might still have been accessing it), then it's
+ * time to restore it.
+ */
+ if (SavedSerializableXact != InvalidSerializableXact)
+ {
+ Assert(MySerializableXact == InvalidSerializableXact);
+ MySerializableXact = SavedSerializableXact;
+ SavedSerializableXact = InvalidSerializableXact;
+ Assert(SxactIsPartiallyReleased(MySerializableXact));
+ }
+ }
+
if (MySerializableXact == InvalidSerializableXact)
{
Assert(LocalPredicateLockHash == NULL);
return;
}
+ LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
+
+ /*
+ * If the transaction is committing, but it has been partially released
+ * already, then treat this as a roll back. It was marked as rolled back.
+ */
+ if (isCommit && SxactIsPartiallyReleased(MySerializableXact))
+ isCommit = false;
+
+ /*
+ * If we're called in the middle of a transaction because we discovered
+ * that the SXACT_FLAG_RO_SAFE flag was set, then we'll partially release
+ * it (that is, release the predicate locks and conflicts, but not the
+ * SERIALIZABLEXACT itself) if we're the first backend to have noticed.
+ */
+ if (isReadOnlySafe && IsInParallelMode())
+ {
+ /*
+ * The leader needs to stash a pointer to it, so that it can
+ * completely release it at end-of-transaction.
+ */
+ if (!IsParallelWorker())
+ SavedSerializableXact = MySerializableXact;
+
+ /*
+ * The first backend to reach this condition will partially release
+ * the SERIALIZABLEXACT. All others will just clear their
+ * backend-local state so that they stop doing SSI checks for the rest
+ * of the transaction.
+ */
+ if (SxactIsPartiallyReleased(MySerializableXact))
+ {
+ LWLockRelease(SerializableXactHashLock);
+ ReleasePredicateLocksLocal();
+ return;
+ }
+ else
+ {
+ MySerializableXact->flags |= SXACT_FLAG_PARTIALLY_RELEASED;
+ /* ... and proceed to perform the partial release below. */
+ }
+ }
Assert(!isCommit || SxactIsPrepared(MySerializableXact));
- Assert(!SxactIsRolledBack(MySerializableXact));
+ Assert(!isCommit || !SxactIsDoomed(MySerializableXact));
Assert(!SxactIsCommitted(MySerializableXact));
+ Assert(SxactIsPartiallyReleased(MySerializableXact)
+ || !SxactIsRolledBack(MySerializableXact));
/* may not be serializable during COMMIT/ROLLBACK PREPARED */
- if (MySerializableXact->pid != 0)
- Assert(IsolationIsSerializable());
+ Assert(MySerializableXact->pid == 0 || IsolationIsSerializable());
/* We'd better not already be on the cleanup list. */
- Assert(!SxactIsOnFinishedList((SERIALIZABLEXACT *) MySerializableXact));
+ Assert(!SxactIsOnFinishedList(MySerializableXact));
topLevelIsDeclaredReadOnly = SxactIsReadOnly(MySerializableXact);
- LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
-
/*
- * We don't hold a lock here, assuming that TransactionId is atomic!
+ * We don't hold XidGenLock lock here, assuming that TransactionId is
+ * atomic!
*
* If this value is changing, we don't care that much whether we get the
* old or new value -- it is just used to determine how far
- * GlobalSerizableXmin must advance before this transaction can be cleaned
- * fully cleaned up. The worst that could happen is we wait for ome more
+ * GlobalSerializableXmin must advance before this transaction can be
+ * fully cleaned up. The worst that could happen is we wait for one more
* transaction to complete before freeing some RAM; correctness of visible
* behavior is not affected.
*/
- MySerializableXact->finishedBefore = ShmemVariableCache->nextXid;
+ MySerializableXact->finishedBefore = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
/*
- * If it's not a commit it's a rollback, and we can clear our locks
- * immediately.
+ * If it's not a commit it's either a rollback or a read-only transaction
+ * flagged SXACT_FLAG_RO_SAFE, and we can clear our locks immediately.
*/
if (isCommit)
{
MySerializableXact->flags |= SXACT_FLAG_COMMITTED;
MySerializableXact->commitSeqNo = ++(PredXact->LastSxactCommitSeqNo);
/* Recognize implicit read-only transaction (commit without write). */
- if (!(MySerializableXact->flags & SXACT_FLAG_DID_WRITE))
+ if (!MyXactDidWrite)
MySerializableXact->flags |= SXACT_FLAG_READ_ONLY;
}
else
{
+ /*
+ * The DOOMED flag indicates that we intend to roll back this
+ * transaction and so it should not cause serialization failures for
+ * other transactions that conflict with it. Note that this flag might
+ * already be set, if another backend marked this transaction for
+ * abort.
+ *
+ * The ROLLED_BACK flag further indicates that ReleasePredicateLocks
+ * has been called, and so the SerializableXact is eligible for
+ * cleanup. This means it should not be considered when calculating
+ * SxactGlobalXmin.
+ */
+ MySerializableXact->flags |= SXACT_FLAG_DOOMED;
MySerializableXact->flags |= SXACT_FLAG_ROLLED_BACK;
+
+ /*
+ * If the transaction was previously prepared, but is now failing due
+ * to a ROLLBACK PREPARED or (hopefully very rare) error after the
+ * prepare, clear the prepared flag. This simplifies conflict
+ * checking.
+ */
+ MySerializableXact->flags &= ~SXACT_FLAG_PREPARED;
}
if (!topLevelIsDeclaredReadOnly)
* opposed to 'outLink' for the r/w xacts.
*/
possibleUnsafeConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->possibleUnsafeConflicts,
- (SHM_QUEUE *) &MySerializableXact->possibleUnsafeConflicts,
+ SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
+ &MySerializableXact->possibleUnsafeConflicts,
offsetof(RWConflictData, inLink));
while (possibleUnsafeConflict)
{
nextConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->possibleUnsafeConflicts,
+ SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
&possibleUnsafeConflict->inLink,
offsetof(RWConflictData, inLink));
&& !SxactIsReadOnly(MySerializableXact)
&& SxactHasSummaryConflictOut(MySerializableXact))
{
+ /*
+ * we don't know which old committed transaction we conflicted with,
+ * so be conservative and use FirstNormalSerCommitSeqNo here
+ */
MySerializableXact->SeqNo.earliestOutConflictCommit =
FirstNormalSerCommitSeqNo;
MySerializableXact->flags |= SXACT_FLAG_CONFLICT_OUT;
}
/*
- * Release all outConflicts to committed transactions. If we're rolling
+ * Release all outConflicts to committed transactions. If we're rolling
* back clear them all. Set SXACT_FLAG_CONFLICT_OUT if any point to
* previously committed transactions.
*/
conflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->outConflicts,
- (SHM_QUEUE *) &MySerializableXact->outConflicts,
+ SHMQueueNext(&MySerializableXact->outConflicts,
+ &MySerializableXact->outConflicts,
offsetof(RWConflictData, outLink));
while (conflict)
{
nextConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->outConflicts,
+ SHMQueueNext(&MySerializableXact->outConflicts,
&conflict->outLink,
offsetof(RWConflictData, outLink));
&& SxactIsCommitted(conflict->sxactIn))
{
if ((MySerializableXact->flags & SXACT_FLAG_CONFLICT_OUT) == 0
- || conflict->sxactIn->commitSeqNo < MySerializableXact->SeqNo.earliestOutConflictCommit)
- MySerializableXact->SeqNo.earliestOutConflictCommit = conflict->sxactIn->commitSeqNo;
+ || conflict->sxactIn->prepareSeqNo < MySerializableXact->SeqNo.earliestOutConflictCommit)
+ MySerializableXact->SeqNo.earliestOutConflictCommit = conflict->sxactIn->prepareSeqNo;
MySerializableXact->flags |= SXACT_FLAG_CONFLICT_OUT;
}
* we're rolling back, clear them all.
*/
conflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->inConflicts,
- (SHM_QUEUE *) &MySerializableXact->inConflicts,
+ SHMQueueNext(&MySerializableXact->inConflicts,
+ &MySerializableXact->inConflicts,
offsetof(RWConflictData, inLink));
while (conflict)
{
nextConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->inConflicts,
+ SHMQueueNext(&MySerializableXact->inConflicts,
&conflict->inLink,
offsetof(RWConflictData, inLink));
* up if they are known safe or known unsafe.
*/
possibleUnsafeConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->possibleUnsafeConflicts,
- (SHM_QUEUE *) &MySerializableXact->possibleUnsafeConflicts,
+ SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
+ &MySerializableXact->possibleUnsafeConflicts,
offsetof(RWConflictData, outLink));
while (possibleUnsafeConflict)
{
nextConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->possibleUnsafeConflicts,
+ SHMQueueNext(&MySerializableXact->possibleUnsafeConflicts,
&possibleUnsafeConflict->outLink,
offsetof(RWConflictData, outLink));
/* Mark conflicted if necessary. */
if (isCommit
- && (MySerializableXact->flags & SXACT_FLAG_DID_WRITE)
+ && MyXactDidWrite
&& SxactHasConflictOut(MySerializableXact)
&& (MySerializableXact->SeqNo.earliestOutConflictCommit
<= roXact->SeqNo.lastCommitBeforeSnapshot))
/* Add this to the list of transactions to check for later cleanup. */
if (isCommit)
SHMQueueInsertBefore(FinishedSerializableTransactions,
- (SHM_QUEUE *) &(MySerializableXact->finishedLink));
+ &MySerializableXact->finishedLink);
+ /*
+ * If we're releasing a RO_SAFE transaction in parallel mode, we'll only
+ * partially release it. That's necessary because other backends may have
+ * a reference to it. The leader will release the SERIALIZABLEXACT itself
+ * at the end of the transaction after workers have stopped running.
+ */
if (!isCommit)
- ReleaseOneSerializableXact((SERIALIZABLEXACT *) MySerializableXact,
- false, false);
+ ReleaseOneSerializableXact(MySerializableXact,
+ isReadOnlySafe && IsInParallelMode(),
+ false);
LWLockRelease(SerializableFinishedListLock);
if (needToClear)
ClearOldPredicateLocks();
+ ReleasePredicateLocksLocal();
+}
+
+static void
+ReleasePredicateLocksLocal(void)
+{
MySerializableXact = InvalidSerializableXact;
+ MyXactDidWrite = false;
/* Delete per-transaction lock table */
if (LocalPredicateLockHash != NULL)
}
/*
- * ReleasePredicateLocksIfROSafe
- * Check if the current transaction is read only and operating on
- * a safe snapshot. If so, release predicate locks and return
- * true.
- *
- * A transaction is flagged as RO_SAFE if all concurrent R/W
- * transactions commit without having conflicts out to an earlier
- * snapshot, thus ensuring that no conflicts are possible for this
- * transaction. Thus, we call this function as part of the
- * SkipSerialization check on all public interface methods.
- */
-static bool
-ReleasePredicateLocksIfROSafe(void)
-{
- if (SxactIsROSafe(MySerializableXact))
- {
- ReleasePredicateLocks(false);
- return true;
- }
- else
- return false;
-}
-
-/*
- * Clear old predicate locks.
+ * Clear old predicate locks, belonging to committed transactions that are no
+ * longer interesting to any in-progress transaction.
*/
static void
ClearOldPredicateLocks(void)
SERIALIZABLEXACT *finishedSxact;
PREDICATELOCK *predlock;
+ /*
+ * Loop through finished transactions. They are in commit order, so we can
+ * stop as soon as we find one that's still interesting.
+ */
LWLockAcquire(SerializableFinishedListLock, LW_EXCLUSIVE);
finishedSxact = (SERIALIZABLEXACT *)
SHMQueueNext(FinishedSerializableTransactions,
|| TransactionIdPrecedesOrEquals(finishedSxact->finishedBefore,
PredXact->SxactGlobalXmin))
{
+ /*
+ * This transaction committed before any in-progress transaction
+ * took its snapshot. It's no longer interesting.
+ */
LWLockRelease(SerializableXactHashLock);
SHMQueueDelete(&(finishedSxact->finishedLink));
ReleaseOneSerializableXact(finishedSxact, false, false);
LWLockAcquire(SerializableXactHashLock, LW_SHARED);
}
else if (finishedSxact->commitSeqNo > PredXact->HavePartialClearedThrough
- && finishedSxact->commitSeqNo <= PredXact->CanPartialClearThrough)
+ && finishedSxact->commitSeqNo <= PredXact->CanPartialClearThrough)
{
+ /*
+ * Any active transactions that took their snapshot before this
+ * transaction committed are read-only, so we can clear part of
+ * its state.
+ */
LWLockRelease(SerializableXactHashLock);
- ReleaseOneSerializableXact(finishedSxact,
- !SxactIsReadOnly(finishedSxact),
- false);
+
+ if (SxactIsReadOnly(finishedSxact))
+ {
+ /* A read-only transaction can be removed entirely */
+ SHMQueueDelete(&(finishedSxact->finishedLink));
+ ReleaseOneSerializableXact(finishedSxact, false, false);
+ }
+ else
+ {
+ /*
+ * A read-write transaction can only be partially cleared. We
+ * need to keep the SERIALIZABLEXACT but can release the
+ * SIREAD locks and conflicts in.
+ */
+ ReleaseOneSerializableXact(finishedSxact, true, false);
+ }
+
PredXact->HavePartialClearedThrough = finishedSxact->commitSeqNo;
LWLockAcquire(SerializableXactHashLock, LW_SHARED);
}
else
+ {
+ /* Still interesting. */
break;
+ }
finishedSxact = nextSxact;
}
LWLockRelease(SerializableXactHashLock);
/*
* Loop through predicate locks on dummy transaction for summarized data.
*/
+ LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
predlock = (PREDICATELOCK *)
SHMQueueNext(&OldCommittedSxact->predicateLocks,
&OldCommittedSxact->predicateLocks,
offsetof(PREDICATELOCK, xactLink));
- LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
while (predlock)
{
PREDICATELOCK *nextpredlock;
offsetof(PREDICATELOCK, xactLink));
LWLockAcquire(SerializableXactHashLock, LW_SHARED);
+ Assert(predlock->commitSeqNo != 0);
+ Assert(predlock->commitSeqNo != InvalidSerCommitSeqNo);
canDoPartialCleanup = (predlock->commitSeqNo <= PredXact->CanPartialClearThrough);
LWLockRelease(SerializableXactHashLock);
+ /*
+ * If this lock originally belonged to an old enough transaction, we
+ * can release it.
+ */
if (canDoPartialCleanup)
{
PREDICATELOCKTAG tag;
- SHM_QUEUE *targetLink;
PREDICATELOCKTARGET *target;
PREDICATELOCKTARGETTAG targettag;
uint32 targettaghash;
- LWLockId partitionLock;
+ LWLock *partitionLock;
tag = predlock->tag;
- targetLink = &(predlock->targetLink);
target = tag.myTarget;
targettag = target->tag;
targettaghash = PredicateLockTargetTagHashCode(&targettag);
LWLockAcquire(partitionLock, LW_EXCLUSIVE);
- SHMQueueDelete(targetLink);
+ SHMQueueDelete(&(predlock->targetLink));
SHMQueueDelete(&(predlock->xactLink));
hash_search_with_hash_value(PredicateLockHash, &tag,
- PredicateLockHashCodeFromTargetHashCode(&tag,
- targettaghash),
+ PredicateLockHashCodeFromTargetHashCode(&tag,
+ targettaghash),
HASH_REMOVE, NULL);
RemoveTargetIfNoLongerUsed(target, targettaghash);
* delete the transaction.
*
* When the partial flag is set, we can release all predicate locks and
- * out-conflict information -- we've established that there are no longer
+ * in-conflict information -- we've established that there are no longer
* any overlapping read write transactions for which this transaction could
- * matter.
+ * matter -- but keep the transaction entry itself and any outConflicts.
*
* When the summarize flag is set, we've run short of room for sxact data
- * and must summarize to the SLRU. Predicate locks are transferred to a
+ * and must summarize to the SLRU. Predicate locks are transferred to a
* dummy "old" transaction, with duplicate locks on a single target
* collapsing to a single lock with the "latest" commitSeqNo from among
* the conflicting locks..
Assert(sxact != NULL);
Assert(SxactIsRolledBack(sxact) || SxactIsCommitted(sxact));
+ Assert(partial || !SxactIsOnFinishedList(sxact));
Assert(LWLockHeldByMe(SerializableFinishedListLock));
+ /*
+ * First release all the predicate locks held by this xact (or transfer
+ * them to OldCommittedSxact if summarize is true)
+ */
LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
+ if (IsInParallelMode())
+ LWLockAcquire(&sxact->predicateLockListLock, LW_EXCLUSIVE);
predlock = (PREDICATELOCK *)
SHMQueueNext(&(sxact->predicateLocks),
&(sxact->predicateLocks),
PREDICATELOCKTARGET *target;
PREDICATELOCKTARGETTAG targettag;
uint32 targettaghash;
- LWLockId partitionLock;
+ LWLock *partitionLock;
nextpredlock = (PREDICATELOCK *)
SHMQueueNext(&(sxact->predicateLocks),
SHMQueueDelete(targetLink);
hash_search_with_hash_value(PredicateLockHash, &tag,
- PredicateLockHashCodeFromTargetHashCode(&tag,
- targettaghash),
+ PredicateLockHashCodeFromTargetHashCode(&tag,
+ targettaghash),
HASH_REMOVE, NULL);
if (summarize)
{
/* Fold into dummy transaction list. */
tag.myXact = OldCommittedSxact;
predlock = hash_search_with_hash_value(PredicateLockHash, &tag,
- PredicateLockHashCodeFromTargetHashCode(&tag,
- targettaghash),
- HASH_ENTER, &found);
+ PredicateLockHashCodeFromTargetHashCode(&tag,
+ targettaghash),
+ HASH_ENTER_NULL, &found);
if (!predlock)
ereport(ERROR,
(errcode(ERRCODE_OUT_OF_MEMORY),
errhint("You might need to increase max_pred_locks_per_transaction.")));
if (found)
{
+ Assert(predlock->commitSeqNo != 0);
+ Assert(predlock->commitSeqNo != InvalidSerCommitSeqNo);
if (predlock->commitSeqNo < sxact->commitSeqNo)
predlock->commitSeqNo = sxact->commitSeqNo;
}
*/
SHMQueueInit(&sxact->predicateLocks);
+ if (IsInParallelMode())
+ LWLockRelease(&sxact->predicateLockListLock);
LWLockRelease(SerializablePredicateLockListLock);
sxidtag.xid = sxact->topXid;
LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
+ /* Release all outConflicts (unless 'partial' is true) */
if (!partial)
{
- /* Release all outConflicts. */
conflict = (RWConflict)
SHMQueueNext(&sxact->outConflicts,
&sxact->outConflicts,
conflict = nextConflict;
}
+ /* Finally, get rid of the xid and the record of the transaction itself. */
if (!partial)
{
- /* Get rid of the xid and the record of the transaction itself. */
if (sxidtag.xid != InvalidTransactionId)
hash_search(SerializableXidHash, &sxidtag, HASH_REMOVE, NULL);
ReleasePredXact(sxact);
/*
* CheckForSerializableConflictOut
* We are reading a tuple which has been modified. If it is visible to
- * us but has been deleted, that indicates a rw-conflict out. If it's
+ * us but has been deleted, that indicates a rw-conflict out. If it's
* not visible and was created by a concurrent (overlapping)
* serializable transaction, that is also a rw-conflict out,
*
* currently no known reason to call this function from an index AM.
*/
void
-CheckForSerializableConflictOut(const bool visible, const Relation relation,
- const HeapTuple tuple, const Buffer buffer)
+CheckForSerializableConflictOut(bool visible, Relation relation,
+ HeapTuple tuple, Buffer buffer,
+ Snapshot snapshot)
{
TransactionId xid;
SERIALIZABLEXIDTAG sxidtag;
SERIALIZABLEXACT *sxact;
HTSV_Result htsvResult;
- if (SkipSerialization(relation))
+ if (!SerializationNeededForRead(relation, snapshot))
return;
- if (SxactIsMarkedForDeath(MySerializableXact))
+ /* Check if someone else has already decided that we need to die */
+ if (SxactIsDoomed(MySerializableXact))
{
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on identification as a pivot, during conflict out checking."),
+ errdetail_internal("Reason code: Canceled on identification as a pivot, during conflict out checking."),
errhint("The transaction might succeed if retried.")));
}
* tuple is visible to us, while HeapTupleSatisfiesVacuum checks what else
* is going on with it.
*/
- htsvResult = HeapTupleSatisfiesVacuum(tuple->t_data, TransactionXmin, buffer);
+ htsvResult = HeapTupleSatisfiesVacuum(tuple, TransactionXmin, buffer);
switch (htsvResult)
{
case HEAPTUPLE_LIVE:
case HEAPTUPLE_RECENTLY_DEAD:
if (!visible)
return;
- xid = HeapTupleHeaderGetXmax(tuple->t_data);
+ xid = HeapTupleHeaderGetUpdateXid(tuple->t_data);
break;
case HEAPTUPLE_DELETE_IN_PROGRESS:
- xid = HeapTupleHeaderGetXmax(tuple->t_data);
+ xid = HeapTupleHeaderGetUpdateXid(tuple->t_data);
break;
case HEAPTUPLE_INSERT_IN_PROGRESS:
xid = HeapTupleHeaderGetXmin(tuple->t_data);
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
/*
- * Find top level xid. Bail out if xid is too early to be a conflict, or
+ * Find top level xid. Bail out if xid is too early to be a conflict, or
* if it's our own xid.
*/
if (TransactionIdEquals(xid, GetTopTransactionIdIfAny()))
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on conflict out to old pivot %u.", xid),
- errhint("The transaction might succeed if retried.")));
+ errdetail_internal("Reason code: Canceled on conflict out to old pivot %u.", xid),
+ errhint("The transaction might succeed if retried.")));
if (SxactHasSummaryConflictIn(MySerializableXact)
- || !SHMQueueEmpty((SHM_QUEUE *) &MySerializableXact->inConflicts))
+ || !SHMQueueEmpty(&MySerializableXact->inConflicts))
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on identification as a pivot, with conflict out to old committed transaction %u.", xid),
- errhint("The transaction might succeed if retried.")));
+ errdetail_internal("Reason code: Canceled on identification as a pivot, with conflict out to old committed transaction %u.", xid),
+ errhint("The transaction might succeed if retried.")));
MySerializableXact->flags |= SXACT_FLAG_SUMMARY_CONFLICT_OUT;
}
}
sxact = sxid->myXact;
Assert(TransactionIdEquals(sxact->topXid, xid));
- if (sxact == MySerializableXact
- || SxactIsRolledBack(sxact)
- || SxactIsMarkedForDeath(sxact))
+ if (sxact == MySerializableXact || SxactIsDoomed(sxact))
{
- /* We can't conflict with our own transaction or one rolled back. */
+ /* Can't conflict with ourself or a transaction that will roll back. */
LWLockRelease(SerializableXactHashLock);
return;
}
/*
* We have a conflict out to a transaction which has a conflict out to a
- * summarized transaction. That summarized transaction must have
+ * summarized transaction. That summarized transaction must have
* committed first, and we can't tell when it committed in relation to our
- * snapshot acquisition, so something needs to be cancelled.
+ * snapshot acquisition, so something needs to be canceled.
*/
if (SxactHasSummaryConflictOut(sxact))
{
if (!SxactIsPrepared(sxact))
{
- sxact->flags |= SXACT_FLAG_MARKED_FOR_DEATH;
+ sxact->flags |= SXACT_FLAG_DOOMED;
LWLockRelease(SerializableXactHashLock);
return;
}
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on conflict out to old pivot."),
+ errdetail_internal("Reason code: Canceled on conflict out to old pivot."),
errhint("The transaction might succeed if retried.")));
}
}
&& (!SxactHasConflictOut(sxact)
|| MySerializableXact->SeqNo.lastCommitBeforeSnapshot < sxact->SeqNo.earliestOutConflictCommit))
{
- /* Read-only transaction will appear to run first. No conflict. */
+ /* Read-only transaction will appear to run first. No conflict. */
LWLockRelease(SerializableXactHashLock);
return;
}
return;
}
- if (RWConflictExists((SERIALIZABLEXACT *) MySerializableXact, sxact))
+ if (RWConflictExists(MySerializableXact, sxact))
{
/* We don't want duplicate conflict records in the list. */
LWLockRelease(SerializableXactHashLock);
* Flag the conflict. But first, if this conflict creates a dangerous
* structure, ereport an error.
*/
- FlagRWConflict((SERIALIZABLEXACT *) MySerializableXact, sxact);
+ FlagRWConflict(MySerializableXact, sxact);
LWLockRelease(SerializableXactHashLock);
}
/*
- * Check a particular target for rw-dependency conflict in.
+ * Check a particular target for rw-dependency conflict in. A subroutine of
+ * CheckForSerializableConflictIn().
*/
static void
CheckTargetForConflictsIn(PREDICATELOCKTARGETTAG *targettag)
{
uint32 targettaghash;
- LWLockId partitionLock;
+ LWLock *partitionLock;
PREDICATELOCKTARGET *target;
PREDICATELOCK *predlock;
+ PREDICATELOCK *mypredlock = NULL;
+ PREDICATELOCKTAG mypredlocktag;
Assert(MySerializableXact != InvalidSerializableXact);
if (sxact == MySerializableXact)
{
/*
- * If we're getting a write lock on the tuple, we don't need a
- * predicate (SIREAD) lock. At this point our transaction already
- * has an ExclusiveRowLock on the relation, so we are OK to drop
- * the predicate lock on the tuple, if found, without fearing that
- * another write against the tuple will occur before the MVCC
- * information makes it to the buffer.
+ * If we're getting a write lock on a tuple, we don't need a
+ * predicate (SIREAD) lock on the same tuple. We can safely remove
+ * our SIREAD lock, but we'll defer doing so until after the loop
+ * because that requires upgrading to an exclusive partition lock.
+ *
+ * We can't use this optimization within a subtransaction because
+ * the subtransaction could roll back, and we would be left
+ * without any lock at the top level.
*/
- if (GET_PREDICATELOCKTARGETTAG_OFFSET(*targettag))
+ if (!IsSubTransaction()
+ && GET_PREDICATELOCKTARGETTAG_OFFSET(*targettag))
{
- uint32 predlockhashcode;
- PREDICATELOCKTARGET *rmtarget = NULL;
- PREDICATELOCK *rmpredlock;
- LOCALPREDICATELOCK *locallock,
- *rmlocallock;
-
- /*
- * This is a tuple on which we have a tuple predicate lock. We
- * only have shared LW locks now; release those, and get
- * exclusive locks only while we modify things.
- */
- LWLockRelease(SerializableXactHashLock);
- LWLockRelease(partitionLock);
- LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
- LWLockAcquire(partitionLock, LW_EXCLUSIVE);
- LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
-
- /*
- * Remove the predicate lock from shared memory, if it wasn't
- * removed while the locks were released. One way that could
- * happen is from autovacuum cleaning up an index.
- */
- predlockhashcode = PredicateLockHashCodeFromTargetHashCode
- (&(predlock->tag), targettaghash);
- rmpredlock = (PREDICATELOCK *)
- hash_search_with_hash_value(PredicateLockHash,
- &(predlock->tag),
- predlockhashcode,
- HASH_FIND, NULL);
- if (rmpredlock)
- {
- Assert(rmpredlock == predlock);
-
- SHMQueueDelete(predlocktargetlink);
- SHMQueueDelete(&(predlock->xactLink));
-
- rmpredlock = (PREDICATELOCK *)
- hash_search_with_hash_value(PredicateLockHash,
- &(predlock->tag),
- predlockhashcode,
- HASH_REMOVE, NULL);
- Assert(rmpredlock == predlock);
-
- RemoveTargetIfNoLongerUsed(target, targettaghash);
-
- LWLockRelease(SerializableXactHashLock);
- LWLockRelease(partitionLock);
- LWLockRelease(SerializablePredicateLockListLock);
-
- locallock = (LOCALPREDICATELOCK *)
- hash_search_with_hash_value(LocalPredicateLockHash,
- targettag, targettaghash,
- HASH_FIND, NULL);
-
- /*
- * Remove entry in local lock table if it exists and has
- * no children. It's OK if it doesn't exist; that means
- * the lock was transferred to a new target by a
- * different backend.
- */
- if (locallock != NULL)
- {
- locallock->held = false;
-
- if (locallock->childLocks == 0)
- {
- rmlocallock = (LOCALPREDICATELOCK *)
- hash_search_with_hash_value(LocalPredicateLockHash,
- targettag, targettaghash,
- HASH_REMOVE, NULL);
- Assert(rmlocallock == locallock);
- }
- }
-
- DecrementParentLocks(targettag);
-
- /*
- * If we've cleaned up the last of the predicate locks for
- * the target, bail out before re-acquiring the locks.
- */
- if (rmtarget)
- return;
-
- /*
- * The list has been altered. Start over at the front.
- */
- LWLockAcquire(partitionLock, LW_SHARED);
- nextpredlock = (PREDICATELOCK *)
- SHMQueueNext(&(target->predicateLocks),
- &(target->predicateLocks),
- offsetof(PREDICATELOCK, targetLink));
-
- LWLockAcquire(SerializableXactHashLock, LW_SHARED);
- }
- else
- {
- /*
- * The predicate lock was cleared while we were attempting
- * to upgrade our lightweight locks. Revert to the shared
- * locks.
- */
- LWLockRelease(SerializableXactHashLock);
- LWLockRelease(partitionLock);
- LWLockRelease(SerializablePredicateLockListLock);
- LWLockAcquire(partitionLock, LW_SHARED);
- LWLockAcquire(SerializableXactHashLock, LW_SHARED);
- }
+ mypredlock = predlock;
+ mypredlocktag = predlock->tag;
}
}
- else if (!SxactIsRolledBack(sxact)
+ else if (!SxactIsDoomed(sxact)
&& (!SxactIsCommitted(sxact)
|| TransactionIdPrecedes(GetTransactionSnapshot()->xmin,
sxact->finishedBefore))
- && !RWConflictExists(sxact, (SERIALIZABLEXACT *) MySerializableXact))
+ && !RWConflictExists(sxact, MySerializableXact))
{
LWLockRelease(SerializableXactHashLock);
LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
- FlagRWConflict(sxact, (SERIALIZABLEXACT *) MySerializableXact);
+ /*
+ * Re-check after getting exclusive lock because the other
+ * transaction may have flagged a conflict.
+ */
+ if (!SxactIsDoomed(sxact)
+ && (!SxactIsCommitted(sxact)
+ || TransactionIdPrecedes(GetTransactionSnapshot()->xmin,
+ sxact->finishedBefore))
+ && !RWConflictExists(sxact, MySerializableXact))
+ {
+ FlagRWConflict(sxact, MySerializableXact);
+ }
LWLockRelease(SerializableXactHashLock);
LWLockAcquire(SerializableXactHashLock, LW_SHARED);
}
LWLockRelease(SerializableXactHashLock);
LWLockRelease(partitionLock);
+
+ /*
+ * If we found one of our own SIREAD locks to remove, remove it now.
+ *
+ * At this point our transaction already has a RowExclusiveLock on the
+ * relation, so we are OK to drop the predicate lock on the tuple, if
+ * found, without fearing that another write against the tuple will occur
+ * before the MVCC information makes it to the buffer.
+ */
+ if (mypredlock != NULL)
+ {
+ uint32 predlockhashcode;
+ PREDICATELOCK *rmpredlock;
+
+ LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
+ if (IsInParallelMode())
+ LWLockAcquire(&MySerializableXact->predicateLockListLock, LW_EXCLUSIVE);
+ LWLockAcquire(partitionLock, LW_EXCLUSIVE);
+ LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
+
+ /*
+ * Remove the predicate lock from shared memory, if it wasn't removed
+ * while the locks were released. One way that could happen is from
+ * autovacuum cleaning up an index.
+ */
+ predlockhashcode = PredicateLockHashCodeFromTargetHashCode
+ (&mypredlocktag, targettaghash);
+ rmpredlock = (PREDICATELOCK *)
+ hash_search_with_hash_value(PredicateLockHash,
+ &mypredlocktag,
+ predlockhashcode,
+ HASH_FIND, NULL);
+ if (rmpredlock != NULL)
+ {
+ Assert(rmpredlock == mypredlock);
+
+ SHMQueueDelete(&(mypredlock->targetLink));
+ SHMQueueDelete(&(mypredlock->xactLink));
+
+ rmpredlock = (PREDICATELOCK *)
+ hash_search_with_hash_value(PredicateLockHash,
+ &mypredlocktag,
+ predlockhashcode,
+ HASH_REMOVE, NULL);
+ Assert(rmpredlock == mypredlock);
+
+ RemoveTargetIfNoLongerUsed(target, targettaghash);
+ }
+
+ LWLockRelease(SerializableXactHashLock);
+ LWLockRelease(partitionLock);
+ if (IsInParallelMode())
+ LWLockRelease(&MySerializableXact->predicateLockListLock);
+ LWLockRelease(SerializablePredicateLockListLock);
+
+ if (rmpredlock != NULL)
+ {
+ /*
+ * Remove entry in local lock table if it exists. It's OK if it
+ * doesn't exist; that means the lock was transferred to a new
+ * target by a different backend.
+ */
+ hash_search_with_hash_value(LocalPredicateLockHash,
+ targettag, targettaghash,
+ HASH_REMOVE, NULL);
+
+ DecrementParentLocks(targettag);
+ }
+ }
}
/*
* tuple itself.
*/
void
-CheckForSerializableConflictIn(const Relation relation, const HeapTuple tuple,
- const Buffer buffer)
+CheckForSerializableConflictIn(Relation relation, HeapTuple tuple,
+ Buffer buffer)
{
PREDICATELOCKTARGETTAG targettag;
- if (SkipSerialization(relation))
+ if (!SerializationNeededForWrite(relation))
return;
- if (SxactIsMarkedForDeath(MySerializableXact))
+ /* Check if someone else has already decided that we need to die */
+ if (SxactIsDoomed(MySerializableXact))
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on identification as a pivot, during conflict in checking."),
+ errdetail_internal("Reason code: Canceled on identification as a pivot, during conflict in checking."),
errhint("The transaction might succeed if retried.")));
- MySerializableXact->flags |= SXACT_FLAG_DID_WRITE;
+ /*
+ * We're doing a write which might cause rw-conflicts now or later.
+ * Memorize that fact.
+ */
+ MyXactDidWrite = true;
/*
* It is important that we check for locks from the finest granularity to
SET_PREDICATELOCKTARGETTAG_TUPLE(targettag,
relation->rd_node.dbNode,
relation->rd_id,
- ItemPointerGetBlockNumber(&(tuple->t_data->t_ctid)),
- ItemPointerGetOffsetNumber(&(tuple->t_data->t_ctid)),
- HeapTupleHeaderGetXmin(tuple->t_data));
+ ItemPointerGetBlockNumber(&(tuple->t_self)),
+ ItemPointerGetOffsetNumber(&(tuple->t_self)));
CheckTargetForConflictsIn(&targettag);
}
CheckTargetForConflictsIn(&targettag);
}
+/*
+ * CheckTableForSerializableConflictIn
+ * The entire table is going through a DDL-style logical mass delete
+ * like TRUNCATE or DROP TABLE. If that causes a rw-conflict in from
+ * another serializable transaction, take appropriate action.
+ *
+ * While these operations do not operate entirely within the bounds of
+ * snapshot isolation, they can occur inside a serializable transaction, and
+ * will logically occur after any reads which saw rows which were destroyed
+ * by these operations, so we do what we can to serialize properly under
+ * SSI.
+ *
+ * The relation passed in must be a heap relation. Any predicate lock of any
+ * granularity on the heap will cause a rw-conflict in to this transaction.
+ * Predicate locks on indexes do not matter because they only exist to guard
+ * against conflicting inserts into the index, and this is a mass *delete*.
+ * When a table is truncated or dropped, the index will also be truncated
+ * or dropped, and we'll deal with locks on the index when that happens.
+ *
+ * Dropping or truncating a table also needs to drop any existing predicate
+ * locks on heap tuples or pages, because they're about to go away. This
+ * should be done before altering the predicate locks because the transaction
+ * could be rolled back because of a conflict, in which case the lock changes
+ * are not needed. (At the moment, we don't actually bother to drop the
+ * existing locks on a dropped or truncated table at the moment. That might
+ * lead to some false positives, but it doesn't seem worth the trouble.)
+ */
+void
+CheckTableForSerializableConflictIn(Relation relation)
+{
+ HASH_SEQ_STATUS seqstat;
+ PREDICATELOCKTARGET *target;
+ Oid dbId;
+ Oid heapId;
+ int i;
+
+ /*
+ * Bail out quickly if there are no serializable transactions running.
+ * It's safe to check this without taking locks because the caller is
+ * holding an ACCESS EXCLUSIVE lock on the relation. No new locks which
+ * would matter here can be acquired while that is held.
+ */
+ if (!TransactionIdIsValid(PredXact->SxactGlobalXmin))
+ return;
+
+ if (!SerializationNeededForWrite(relation))
+ return;
+
+ /*
+ * We're doing a write which might cause rw-conflicts now or later.
+ * Memorize that fact.
+ */
+ MyXactDidWrite = true;
+
+ Assert(relation->rd_index == NULL); /* not an index relation */
+
+ dbId = relation->rd_node.dbNode;
+ heapId = relation->rd_id;
+
+ LWLockAcquire(SerializablePredicateLockListLock, LW_EXCLUSIVE);
+ for (i = 0; i < NUM_PREDICATELOCK_PARTITIONS; i++)
+ LWLockAcquire(PredicateLockHashPartitionLockByIndex(i), LW_SHARED);
+ LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
+
+ /* Scan through target list */
+ hash_seq_init(&seqstat, PredicateLockTargetHash);
+
+ while ((target = (PREDICATELOCKTARGET *) hash_seq_search(&seqstat)))
+ {
+ PREDICATELOCK *predlock;
+
+ /*
+ * Check whether this is a target which needs attention.
+ */
+ if (GET_PREDICATELOCKTARGETTAG_RELATION(target->tag) != heapId)
+ continue; /* wrong relation id */
+ if (GET_PREDICATELOCKTARGETTAG_DB(target->tag) != dbId)
+ continue; /* wrong database id */
+
+ /*
+ * Loop through locks for this target and flag conflicts.
+ */
+ predlock = (PREDICATELOCK *)
+ SHMQueueNext(&(target->predicateLocks),
+ &(target->predicateLocks),
+ offsetof(PREDICATELOCK, targetLink));
+ while (predlock)
+ {
+ PREDICATELOCK *nextpredlock;
+
+ nextpredlock = (PREDICATELOCK *)
+ SHMQueueNext(&(target->predicateLocks),
+ &(predlock->targetLink),
+ offsetof(PREDICATELOCK, targetLink));
+
+ if (predlock->tag.myXact != MySerializableXact
+ && !RWConflictExists(predlock->tag.myXact, MySerializableXact))
+ {
+ FlagRWConflict(predlock->tag.myXact, MySerializableXact);
+ }
+
+ predlock = nextpredlock;
+ }
+ }
+
+ /* Release locks in reverse order */
+ LWLockRelease(SerializableXactHashLock);
+ for (i = NUM_PREDICATELOCK_PARTITIONS - 1; i >= 0; i--)
+ LWLockRelease(PredicateLockHashPartitionLockByIndex(i));
+ LWLockRelease(SerializablePredicateLockListLock);
+}
+
+
/*
* Flag a rw-dependency between two serializable transactions.
*
SetRWConflict(reader, writer);
}
-/*
- * Check whether we should roll back one of these transactions
- * instead of flagging a new rw-conflict.
+/*----------------------------------------------------------------------------
+ * We are about to add a RW-edge to the dependency graph - check that we don't
+ * introduce a dangerous structure by doing so, and abort one of the
+ * transactions if so.
+ *
+ * A serialization failure can only occur if there is a dangerous structure
+ * in the dependency graph:
+ *
+ * Tin ------> Tpivot ------> Tout
+ * rw rw
+ *
+ * Furthermore, Tout must commit first.
+ *
+ * One more optimization is that if Tin is declared READ ONLY (or commits
+ * without writing), we can only have a problem if Tout committed before Tin
+ * acquired its snapshot.
+ *----------------------------------------------------------------------------
*/
static void
OnConflict_CheckForSerializationFailure(const SERIALIZABLEXACT *reader,
failure = false;
- /*
- * Check for already-committed writer with rw-conflict out flagged. This
- * means that the reader must immediately fail.
+ /*------------------------------------------------------------------------
+ * Check for already-committed writer with rw-conflict out flagged
+ * (conflict-flag on W means that T2 committed before W):
+ *
+ * R ------> W ------> T2
+ * rw rw
+ *
+ * That is a dangerous structure, so we must abort. (Since the writer
+ * has already committed, we must be the reader)
+ *------------------------------------------------------------------------
*/
if (SxactIsCommitted(writer)
- && (SxactHasConflictOut(writer) || SxactHasSummaryConflictOut(writer)))
+ && (SxactHasConflictOut(writer) || SxactHasSummaryConflictOut(writer)))
failure = true;
- /*
- * Check whether the reader has become a pivot with a committed writer. If
- * so, we must roll back unless every in-conflict either committed before
- * the writer committed or is READ ONLY and overlaps the writer.
+ /*------------------------------------------------------------------------
+ * Check whether the writer has become a pivot with an out-conflict
+ * committed transaction (T2), and T2 committed first:
+ *
+ * R ------> W ------> T2
+ * rw rw
+ *
+ * Because T2 must've committed first, there is no anomaly if:
+ * - the reader committed before T2
+ * - the writer committed before T2
+ * - the reader is a READ ONLY transaction and the reader was concurrent
+ * with T2 (= reader acquired its snapshot before T2 committed)
+ *
+ * We also handle the case that T2 is prepared but not yet committed
+ * here. In that case T2 has already checked for conflicts, so if it
+ * commits first, making the above conflict real, it's too late for it
+ * to abort.
+ *------------------------------------------------------------------------
*/
- if (!failure && SxactIsCommitted(writer) && !SxactIsReadOnly(reader))
+ if (!failure)
{
- if (SxactHasSummaryConflictIn(reader))
+ if (SxactHasSummaryConflictOut(writer))
{
failure = true;
conflict = NULL;
}
else
conflict = (RWConflict)
- SHMQueueNext(&reader->inConflicts,
- &reader->inConflicts,
- offsetof(RWConflictData, inLink));
+ SHMQueueNext(&writer->outConflicts,
+ &writer->outConflicts,
+ offsetof(RWConflictData, outLink));
while (conflict)
{
- if (!SxactIsRolledBack(conflict->sxactOut)
- && (!SxactIsCommitted(conflict->sxactOut)
- || conflict->sxactOut->commitSeqNo >= writer->commitSeqNo)
- && (!SxactIsReadOnly(conflict->sxactOut)
- || conflict->sxactOut->SeqNo.lastCommitBeforeSnapshot >= writer->commitSeqNo))
+ SERIALIZABLEXACT *t2 = conflict->sxactIn;
+
+ if (SxactIsPrepared(t2)
+ && (!SxactIsCommitted(reader)
+ || t2->prepareSeqNo <= reader->commitSeqNo)
+ && (!SxactIsCommitted(writer)
+ || t2->prepareSeqNo <= writer->commitSeqNo)
+ && (!SxactIsReadOnly(reader)
+ || t2->prepareSeqNo <= reader->SeqNo.lastCommitBeforeSnapshot))
{
failure = true;
break;
}
conflict = (RWConflict)
- SHMQueueNext(&reader->inConflicts,
- &conflict->inLink,
- offsetof(RWConflictData, inLink));
+ SHMQueueNext(&writer->outConflicts,
+ &conflict->outLink,
+ offsetof(RWConflictData, outLink));
}
}
- /*
- * Check whether the writer has become a pivot with an out-conflict
- * committed transaction, while neither reader nor writer is committed. If
- * the reader is a READ ONLY transaction, there is only a serialization
- * failure if an out-conflict transaction causing the pivot committed
- * before the reader acquired its snapshot. (That is, the reader must not
- * have been concurrent with the out-conflict transaction.)
+ /*------------------------------------------------------------------------
+ * Check whether the reader has become a pivot with a writer
+ * that's committed (or prepared):
+ *
+ * T0 ------> R ------> W
+ * rw rw
+ *
+ * Because W must've committed first for an anomaly to occur, there is no
+ * anomaly if:
+ * - T0 committed before the writer
+ * - T0 is READ ONLY, and overlaps the writer
+ *------------------------------------------------------------------------
*/
- if (!failure && !SxactIsCommitted(writer))
+ if (!failure && SxactIsPrepared(writer) && !SxactIsReadOnly(reader))
{
- if (SxactHasSummaryConflictOut(reader))
+ if (SxactHasSummaryConflictIn(reader))
{
failure = true;
conflict = NULL;
}
else
conflict = (RWConflict)
- SHMQueueNext(&writer->outConflicts,
- &writer->outConflicts,
- offsetof(RWConflictData, outLink));
+ SHMQueueNext(&reader->inConflicts,
+ &reader->inConflicts,
+ offsetof(RWConflictData, inLink));
while (conflict)
{
- if ((reader == conflict->sxactIn && SxactIsCommitted(reader))
- || (SxactIsCommitted(conflict->sxactIn)
- && !SxactIsCommitted(reader)
- && (!SxactIsReadOnly(reader)
- || conflict->sxactIn->commitSeqNo <= reader->SeqNo.lastCommitBeforeSnapshot)))
+ SERIALIZABLEXACT *t0 = conflict->sxactOut;
+
+ if (!SxactIsDoomed(t0)
+ && (!SxactIsCommitted(t0)
+ || t0->commitSeqNo >= writer->prepareSeqNo)
+ && (!SxactIsReadOnly(t0)
+ || t0->SeqNo.lastCommitBeforeSnapshot >= writer->prepareSeqNo))
{
failure = true;
break;
}
conflict = (RWConflict)
- SHMQueueNext(&writer->outConflicts,
- &conflict->outLink,
- offsetof(RWConflictData, outLink));
+ SHMQueueNext(&reader->inConflicts,
+ &conflict->inLink,
+ offsetof(RWConflictData, inLink));
}
}
if (failure)
{
+ /*
+ * We have to kill a transaction to avoid a possible anomaly from
+ * occurring. If the writer is us, we can just ereport() to cause a
+ * transaction abort. Otherwise we flag the writer for termination,
+ * causing it to abort when it tries to commit. However, if the writer
+ * is a prepared transaction, already prepared, we can't abort it
+ * anymore, so we have to kill the reader instead.
+ */
if (MySerializableXact == writer)
{
LWLockRelease(SerializableXactHashLock);
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on identification as pivot, during write."),
+ errdetail_internal("Reason code: Canceled on identification as a pivot, during write."),
errhint("The transaction might succeed if retried.")));
}
else if (SxactIsPrepared(writer))
{
LWLockRelease(SerializableXactHashLock);
+
+ /* if we're not the writer, we have to be the reader */
+ Assert(MySerializableXact == reader);
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on conflict out to pivot %u, during read.", writer->topXid),
+ errdetail_internal("Reason code: Canceled on conflict out to pivot %u, during read.", writer->topXid),
errhint("The transaction might succeed if retried.")));
}
- writer->flags |= SXACT_FLAG_MARKED_FOR_DEATH;
+ writer->flags |= SXACT_FLAG_DOOMED;
}
}
*
* If a dangerous structure is found, the pivot (the near conflict) is
* marked for death, because rolling back another transaction might mean
- * that we flail without ever making progress. This transaction is
+ * that we flail without ever making progress. This transaction is
* committing writes, so letting it commit ensures progress. If we
- * cancelled the far conflict, it might immediately fail again on retry.
+ * canceled the far conflict, it might immediately fail again on retry.
*/
void
PreCommit_CheckForSerializationFailure(void)
LWLockAcquire(SerializableXactHashLock, LW_EXCLUSIVE);
- if (SxactIsMarkedForDeath(MySerializableXact))
+ /* Check if someone else has already decided that we need to die */
+ if (SxactIsDoomed(MySerializableXact))
{
+ Assert(!SxactIsPartiallyReleased(MySerializableXact));
LWLockRelease(SerializableXactHashLock);
ereport(ERROR,
(errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
errmsg("could not serialize access due to read/write dependencies among transactions"),
- errdetail("Cancelled on identification as a pivot, during commit attempt."),
+ errdetail_internal("Reason code: Canceled on identification as a pivot, during commit attempt."),
errhint("The transaction might succeed if retried.")));
}
nearConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->inConflicts,
- (SHM_QUEUE *) &MySerializableXact->inConflicts,
+ SHMQueueNext(&MySerializableXact->inConflicts,
+ &MySerializableXact->inConflicts,
offsetof(RWConflictData, inLink));
while (nearConflict)
{
if (!SxactIsCommitted(nearConflict->sxactOut)
- && !SxactIsRolledBack(nearConflict->sxactOut)
- && !SxactIsMarkedForDeath(nearConflict->sxactOut))
+ && !SxactIsDoomed(nearConflict->sxactOut))
{
RWConflict farConflict;
if (farConflict->sxactOut == MySerializableXact
|| (!SxactIsCommitted(farConflict->sxactOut)
&& !SxactIsReadOnly(farConflict->sxactOut)
- && !SxactIsRolledBack(farConflict->sxactOut)
- && !SxactIsMarkedForDeath(farConflict->sxactOut)))
+ && !SxactIsDoomed(farConflict->sxactOut)))
{
- nearConflict->sxactOut->flags |= SXACT_FLAG_MARKED_FOR_DEATH;
+ /*
+ * Normally, we kill the pivot transaction to make sure we
+ * make progress if the failing transaction is retried.
+ * However, we can't kill it if it's already prepared, so
+ * in that case we commit suicide instead.
+ */
+ if (SxactIsPrepared(nearConflict->sxactOut))
+ {
+ LWLockRelease(SerializableXactHashLock);
+ ereport(ERROR,
+ (errcode(ERRCODE_T_R_SERIALIZATION_FAILURE),
+ errmsg("could not serialize access due to read/write dependencies among transactions"),
+ errdetail_internal("Reason code: Canceled on commit attempt with conflict in from prepared pivot."),
+ errhint("The transaction might succeed if retried.")));
+ }
+ nearConflict->sxactOut->flags |= SXACT_FLAG_DOOMED;
break;
}
farConflict = (RWConflict)
}
nearConflict = (RWConflict)
- SHMQueueNext((SHM_QUEUE *) &MySerializableXact->inConflicts,
+ SHMQueueNext(&MySerializableXact->inConflicts,
&nearConflict->inLink,
offsetof(RWConflictData, inLink));
}
+ MySerializableXact->prepareSeqNo = ++(PredXact->LastSxactCommitSeqNo);
MySerializableXact->flags |= SXACT_FLAG_PREPARED;
LWLockRelease(SerializableXactHashLock);
TwoPhasePredicateXactRecord *xactRecord;
TwoPhasePredicateLockRecord *lockRecord;
- sxact = (SERIALIZABLEXACT *) MySerializableXact;
+ sxact = MySerializableXact;
xactRecord = &(record.data.xactRecord);
lockRecord = &(record.data.lockRecord);
if (MySerializableXact == InvalidSerializableXact)
return;
- /* Generate a xact record for our SERIALIZABLEXACT */
+ /* Generate an xact record for our SERIALIZABLEXACT */
record.type = TWOPHASEPREDICATERECORD_XACT;
xactRecord->xmin = MySerializableXact->xmin;
xactRecord->flags = MySerializableXact->flags;
/*
- * Tweak the flags. Since we're not going to output the inConflicts and
- * outConflicts lists, if they're non-empty we'll represent that by
- * setting the appropriate summary conflict flags.
+ * Note that we don't include the list of conflicts in our out in the
+ * statefile, because new conflicts can be added even after the
+ * transaction prepares. We'll just make a conservative assumption during
+ * recovery instead.
*/
- if (!SHMQueueEmpty((SHM_QUEUE *) &MySerializableXact->inConflicts))
- xactRecord->flags |= SXACT_FLAG_SUMMARY_CONFLICT_IN;
- if (!SHMQueueEmpty((SHM_QUEUE *) &MySerializableXact->outConflicts))
- xactRecord->flags |= SXACT_FLAG_SUMMARY_CONFLICT_OUT;
RegisterTwoPhaseRecord(TWOPHASE_RM_PREDICATELOCK_ID, 0,
&record, sizeof(record));
*/
LWLockAcquire(SerializablePredicateLockListLock, LW_SHARED);
+ /*
+ * No need to take sxact->predicateLockListLock in parallel mode because
+ * there cannot be any parallel workers running while we are preparing a
+ * transaction.
+ */
+ Assert(!IsParallelWorker() && !ParallelContextActive());
+
predlock = (PREDICATELOCK *)
SHMQueueNext(&(sxact->predicateLocks),
&(sxact->predicateLocks),
LocalPredicateLockHash = NULL;
MySerializableXact = InvalidSerializableXact;
+ MyXactDidWrite = false;
}
/*
/* Release its locks */
MySerializableXact = sxid->myXact;
- ReleasePredicateLocks(isCommit);
+ MyXactDidWrite = true; /* conservatively assume that we wrote
+ * something */
+ ReleasePredicateLocks(isCommit, false);
}
/*
sxact->pid = 0;
/* a prepared xact hasn't committed yet */
+ sxact->prepareSeqNo = RecoverySerCommitSeqNo;
sxact->commitSeqNo = InvalidSerCommitSeqNo;
sxact->finishedBefore = InvalidTransactionId;
sxact->SeqNo.lastCommitBeforeSnapshot = RecoverySerCommitSeqNo;
-
- /*
- * We don't need the details of a prepared transaction's conflicts,
- * just whether it had conflicts in or out (which we get from the
- * flags)
- */
- SHMQueueInit(&(sxact->outConflicts));
- SHMQueueInit(&(sxact->inConflicts));
-
/*
* Don't need to track this; no transactions running at the time the
* recovered xact started are still active, except possibly other
(MaxBackends + max_prepared_xacts));
}
+ /*
+ * We don't know whether the transaction had any conflicts or not, so
+ * we'll conservatively assume that it had both a conflict in and a
+ * conflict out, and represent that with the summary conflict flags.
+ */
+ SHMQueueInit(&(sxact->outConflicts));
+ SHMQueueInit(&(sxact->inConflicts));
+ sxact->flags |= SXACT_FLAG_SUMMARY_CONFLICT_IN;
+ sxact->flags |= SXACT_FLAG_SUMMARY_CONFLICT_OUT;
+
/* Register the transaction's xid */
sxidtag.xid = xid;
sxid = (SERIALIZABLEXID *) hash_search(SerializableXidHash,
&sxidtag,
HASH_ENTER, &found);
- if (!sxid)
- ereport(ERROR,
- (errcode(ERRCODE_OUT_OF_MEMORY),
- errmsg("out of shared memory")));
+ Assert(sxid != NULL);
Assert(!found);
sxid->myXact = (SERIALIZABLEXACT *) sxact;
CreatePredicateLock(&lockRecord->target, targettaghash, sxact);
}
}
+
+/*
+ * Prepare to share the current SERIALIZABLEXACT with parallel workers.
+ * Return a handle object that can be used by AttachSerializableXact() in a
+ * parallel worker.
+ */
+SerializableXactHandle
+ShareSerializableXact(void)
+{
+ return MySerializableXact;
+}
+
+/*
+ * Allow parallel workers to import the leader's SERIALIZABLEXACT.
+ */
+void
+AttachSerializableXact(SerializableXactHandle handle)
+{
+
+ Assert(MySerializableXact == InvalidSerializableXact);
+
+ MySerializableXact = (SERIALIZABLEXACT *) handle;
+ if (MySerializableXact != InvalidSerializableXact)
+ CreateLocalPredicateLockHash();
+}