]> granicus.if.org Git - postgresql/commitdiff
Handle parallel index builds on mapped relations.
authorPeter Geoghegan <pg@bowt.ie>
Fri, 10 Aug 2018 20:01:34 +0000 (13:01 -0700)
committerPeter Geoghegan <pg@bowt.ie>
Fri, 10 Aug 2018 20:01:34 +0000 (13:01 -0700)
Commit 9da0cc35284, which introduced parallel CREATE INDEX, failed to
propagate relmapper.c backend local cache state to parallel worker
processes.  This could result in parallel index builds against mapped
catalog relations where the leader process (participating as a worker)
scans the new, pristine relfilenode, while worker processes scan the
obsolescent relfilenode.  When this happened, the final index structure
was typically not consistent with the owning table's structure.  The
final index structure could contain entries formed from both heap
relfilenodes.  Only rebuilds on mapped catalog relations that occur as
part of a VACUUM FULL or CLUSTER could become corrupt in practice, since
their mapped relation relfilenode swap is what allows the inconsistency
to arise.

On master, fix the problem by propagating the required relmapper.c
backend state as part of standard parallel initialization (Cf. commit
29d58fd3).  On v11, simply disallow builds against mapped catalog
relations by deeming them parallel unsafe.

Author: Peter Geoghegan
Reported-By: "death lock"
Reviewed-By: Tom Lane, Amit Kapila
Bug: #15309
Discussion: https://postgr.es/m/153329671686.1405.18298309097348420351@wrigleys.postgresql.org
Backpatch: 11-, where parallel CREATE INDEX was introduced.

src/backend/access/transam/README.parallel
src/backend/access/transam/parallel.c
src/backend/access/transam/xact.c
src/backend/utils/cache/relmapper.c
src/include/utils/relmapper.h

index f09a580634582c520df863979f91dba89e060709..85e5840feba5424c9e709337f9a46248f57f1eb0 100644 (file)
@@ -125,6 +125,10 @@ worker.  This includes:
   - State related to pending REINDEX operations, which prevents access to
     an index that is currently being rebuilt.
 
+  - Active relmapper.c mapping state.  This is needed to allow consistent
+    answers when fetching the current relfilenode for relation oids of
+    mapped relations.
+
 To prevent unprincipled deadlocks when running in parallel mode, this code
 also arranges for the leader and all workers to participate in group
 locking.  See src/backend/storage/lmgr/README for more details.
index 30ddf94c952dfc2ac2874fc03d3072b21e947d0e..c1681184670de99ccd81a552001a732d699d67a2 100644 (file)
@@ -37,6 +37,7 @@
 #include "utils/guc.h"
 #include "utils/inval.h"
 #include "utils/memutils.h"
+#include "utils/relmapper.h"
 #include "utils/snapmgr.h"
 #include "utils/typcache.h"
 
@@ -69,6 +70,7 @@
 #define PARALLEL_KEY_ENTRYPOINT                                UINT64CONST(0xFFFFFFFFFFFF0009)
 #define PARALLEL_KEY_SESSION_DSM                       UINT64CONST(0xFFFFFFFFFFFF000A)
 #define PARALLEL_KEY_REINDEX_STATE                     UINT64CONST(0xFFFFFFFFFFFF000B)
+#define PARALLEL_KEY_RELMAPPER_STATE           UINT64CONST(0xFFFFFFFFFFFF000C)
 
 /* Fixed-size parallel state. */
 typedef struct FixedParallelState
@@ -205,6 +207,7 @@ InitializeParallelDSM(ParallelContext *pcxt)
        Size            asnaplen = 0;
        Size            tstatelen = 0;
        Size            reindexlen = 0;
+       Size            relmapperlen = 0;
        Size            segsize = 0;
        int                     i;
        FixedParallelState *fps;
@@ -256,8 +259,10 @@ InitializeParallelDSM(ParallelContext *pcxt)
                shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
                reindexlen = EstimateReindexStateSpace();
                shm_toc_estimate_chunk(&pcxt->estimator, reindexlen);
+               relmapperlen = EstimateRelationMapSpace();
+               shm_toc_estimate_chunk(&pcxt->estimator, relmapperlen);
                /* If you add more chunks here, you probably need to add keys. */
-               shm_toc_estimate_keys(&pcxt->estimator, 8);
+               shm_toc_estimate_keys(&pcxt->estimator, 9);
 
                /* Estimate space need for error queues. */
                StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
@@ -327,6 +332,7 @@ InitializeParallelDSM(ParallelContext *pcxt)
                char       *asnapspace;
                char       *tstatespace;
                char       *reindexspace;
+               char       *relmapperspace;
                char       *error_queue_space;
                char       *session_dsm_handle_space;
                char       *entrypointstate;
@@ -373,6 +379,12 @@ InitializeParallelDSM(ParallelContext *pcxt)
                SerializeReindexState(reindexlen, reindexspace);
                shm_toc_insert(pcxt->toc, PARALLEL_KEY_REINDEX_STATE, reindexspace);
 
+               /* Serialize relmapper state. */
+               relmapperspace = shm_toc_allocate(pcxt->toc, relmapperlen);
+               SerializeRelationMap(relmapperlen, relmapperspace);
+               shm_toc_insert(pcxt->toc, PARALLEL_KEY_RELMAPPER_STATE,
+                                          relmapperspace);
+
                /* Allocate space for worker information. */
                pcxt->worker = palloc0(sizeof(ParallelWorkerInfo) * pcxt->nworkers);
 
@@ -1205,6 +1217,7 @@ ParallelWorkerMain(Datum main_arg)
        char       *asnapspace;
        char       *tstatespace;
        char       *reindexspace;
+       char       *relmapperspace;
        StringInfoData msgbuf;
        char       *session_dsm_handle_space;
 
@@ -1380,6 +1393,10 @@ ParallelWorkerMain(Datum main_arg)
        reindexspace = shm_toc_lookup(toc, PARALLEL_KEY_REINDEX_STATE, false);
        RestoreReindexState(reindexspace);
 
+       /* Restore relmapper state. */
+       relmapperspace = shm_toc_lookup(toc, PARALLEL_KEY_RELMAPPER_STATE, false);
+       RestoreRelationMap(relmapperspace);
+
        /*
         * We've initialized all of our state now; nothing should change
         * hereafter.
index 9aa63c8792be56e4186ef2bbfc12198c00d3b6c3..cd8270d5fb04b165d101cb444ef591436ab71879 100644 (file)
@@ -2018,7 +2018,7 @@ CommitTransaction(void)
        HOLD_INTERRUPTS();
 
        /* Commit updates to the relation map --- do this as late as possible */
-       AtEOXact_RelationMap(true);
+       AtEOXact_RelationMap(true, is_parallel_worker);
 
        /*
         * set the current transaction state information appropriately during
@@ -2539,7 +2539,7 @@ AbortTransaction(void)
        AtAbort_Portals();
        AtEOXact_LargeObject(false);
        AtAbort_Notify();
-       AtEOXact_RelationMap(false);
+       AtEOXact_RelationMap(false, is_parallel_worker);
        AtAbort_Twophase();
 
        /*
index c7f0e6f6d4af8b72a2c96bf3eb6418bf7ae15457..905867dc767f70f57b173b40cba4e439a400285b 100644 (file)
@@ -91,6 +91,16 @@ typedef struct RelMapFile
        int32           pad;                    /* to make the struct size be 512 exactly */
 } RelMapFile;
 
+/*
+ * State for serializing local and shared relmappings for parallel workers
+ * (active states only).  See notes on active_* and pending_* updates state.
+ */
+typedef struct SerializedActiveRelMaps
+{
+       RelMapFile      active_shared_updates;
+       RelMapFile      active_local_updates;
+} SerializedActiveRelMaps;
+
 /*
  * The currently known contents of the shared map file and our database's
  * local map file are stored here.  These can be reloaded from disk
@@ -111,6 +121,9 @@ static RelMapFile local_map;
  * they will become active at the next CommandCounterIncrement.  This setup
  * lets map updates act similarly to updates of pg_class rows, ie, they
  * become visible only at the next CommandCounterIncrement boundary.
+ *
+ * Active shared and active local updates are serialized by the parallel
+ * infrastructure, and deserialized within parallel workers.
  */
 static RelMapFile active_shared_updates;
 static RelMapFile active_local_updates;
@@ -263,13 +276,16 @@ RelationMapUpdateMap(Oid relationId, Oid fileNode, bool shared,
        else
        {
                /*
-                * We don't currently support map changes within subtransactions. This
-                * could be done with more bookkeeping infrastructure, but it doesn't
-                * presently seem worth it.
+                * We don't currently support map changes within subtransactions, or
+                * when in parallel mode.  This could be done with more bookkeeping
+                * infrastructure, but it doesn't presently seem worth it.
                 */
                if (GetCurrentTransactionNestLevel() > 1)
                        elog(ERROR, "cannot change relation mapping within subtransaction");
 
+               if (IsInParallelMode())
+                       elog(ERROR, "cannot change relation mapping in parallel mode");
+
                if (immediate)
                {
                        /* Make it active, but only locally */
@@ -452,11 +468,14 @@ AtCCI_RelationMap(void)
  *
  * During abort, we just have to throw away any pending map changes.
  * Normal post-abort cleanup will take care of fixing relcache entries.
+ * Parallel worker commit/abort is handled by resetting active mappings
+ * that may have been received from the leader process.  (There should be
+ * no pending updates in parallel workers.)
  */
 void
-AtEOXact_RelationMap(bool isCommit)
+AtEOXact_RelationMap(bool isCommit, bool isParallelWorker)
 {
-       if (isCommit)
+       if (isCommit && !isParallelWorker)
        {
                /*
                 * We should not get here with any "pending" updates.  (We could
@@ -482,7 +501,10 @@ AtEOXact_RelationMap(bool isCommit)
        }
        else
        {
-               /* Abort --- drop all local and pending updates */
+               /* Abort or parallel worker --- drop all local and pending updates */
+               Assert(!isParallelWorker || pending_shared_updates.num_mappings == 0);
+               Assert(!isParallelWorker || pending_local_updates.num_mappings == 0);
+
                active_shared_updates.num_mappings = 0;
                active_local_updates.num_mappings = 0;
                pending_shared_updates.num_mappings = 0;
@@ -614,6 +636,56 @@ RelationMapInitializePhase3(void)
        load_relmap_file(false);
 }
 
+/*
+ * EstimateRelationMapSpace
+ *
+ * Estimate space needed to pass active shared and local relmaps to parallel
+ * workers.
+ */
+Size
+EstimateRelationMapSpace(void)
+{
+       return sizeof(SerializedActiveRelMaps);
+}
+
+/*
+ * SerializeRelationMap
+ *
+ * Serialize active shared and local relmap state for parallel workers.
+ */
+void
+SerializeRelationMap(Size maxSize, char *startAddress)
+{
+       SerializedActiveRelMaps    *relmaps;
+
+       Assert(maxSize >= EstimateRelationMapSpace());
+
+       relmaps = (SerializedActiveRelMaps *) startAddress;
+       relmaps->active_shared_updates = active_shared_updates;
+       relmaps->active_local_updates = active_local_updates;
+}
+
+/*
+ * RestoreRelationMap
+ *
+ * Restore active shared and local relmap state within a parallel worker.
+ */
+void
+RestoreRelationMap(char *startAddress)
+{
+       SerializedActiveRelMaps    *relmaps;
+
+       if (active_shared_updates.num_mappings != 0 ||
+               active_local_updates.num_mappings != 0 ||
+               pending_shared_updates.num_mappings != 0 ||
+               pending_local_updates.num_mappings != 0)
+               elog(ERROR, "parallel worker has existing mappings");
+
+       relmaps = (SerializedActiveRelMaps *) startAddress;
+       active_shared_updates = relmaps->active_shared_updates;
+       active_local_updates = relmaps->active_local_updates;
+}
+
 /*
  * load_relmap_file -- load data from the shared or local map file
  *
index f69b1006bf269a0f76dc2d5c00f7bc88d4dfd45e..fb51943498d736771a3a69dfaf09b8583e3a6ccc 100644 (file)
@@ -48,7 +48,7 @@ extern void RelationMapInvalidate(bool shared);
 extern void RelationMapInvalidateAll(void);
 
 extern void AtCCI_RelationMap(void);
-extern void AtEOXact_RelationMap(bool isCommit);
+extern void AtEOXact_RelationMap(bool isCommit, bool isParallelWorker);
 extern void AtPrepare_RelationMap(void);
 
 extern void CheckPointRelationMap(void);
@@ -59,6 +59,10 @@ extern void RelationMapInitialize(void);
 extern void RelationMapInitializePhase2(void);
 extern void RelationMapInitializePhase3(void);
 
+extern Size EstimateRelationMapSpace(void);
+extern void SerializeRelationMap(Size maxSize, char *startAddress);
+extern void RestoreRelationMap(char *startAddress);
+
 extern void relmap_redo(XLogReaderState *record);
 extern void relmap_desc(StringInfo buf, XLogReaderState *record);
 extern const char *relmap_identify(uint8 info);