/*
* repalloc
* Adjust the size of a previously allocated chunk.
*/
void *
repalloc(void *pointer, Size size)
{
MemoryContext context;
void *ret;
if (!AllocSizeIsValid(size))
elog(ERROR, "invalid memory alloc request size %zu", size);
/* pgpool hack by Muhammad Usama <*****@*****.**> */
if(pointer == NULL)
return palloc(size);
context = GetMemoryChunkContext(pointer);
AssertNotInCriticalSection(context);
/* isReset must be false already */
Assert(!context->isReset);
ret = (*context->methods->realloc) (context, pointer, size);
if (ret == NULL)
{
MemoryContextStats(TopMemoryContext);
ereport(ERROR,
(errcode(ERRCODE_OUT_OF_MEMORY),
errmsg("out of memory"),
errdetail("Failed on request of size %zu.", size)));
}
VALGRIND_MEMPOOL_CHANGE(context, pointer, ret, size);
return ret;
}
Type Type_getCoerceIn(Type self, Type other)
{
Oid funcId;
Type coerce;
Oid fromOid = other->typeId;
Oid toOid = self->typeId;
if(self->inCoercions != 0)
{
coerce = HashMap_getByOid(self->inCoercions, fromOid);
if(coerce != 0)
return coerce;
}
if (!find_coercion_pathway(toOid, fromOid, COERCION_EXPLICIT, &funcId))
{
elog(ERROR, "no conversion function from %s to %s",
format_type_be(fromOid),
format_type_be(toOid));
}
if(funcId == InvalidOid)
/*
* Binary compatible type. No need for a special coercer
*/
return self;
if(self->inCoercions == 0)
self->inCoercions = HashMap_create(7, GetMemoryChunkContext(self));
coerce = Coerce_createIn(self, other, funcId);
HashMap_putByOid(self->inCoercions, fromOid, coerce);
return coerce;
}
/*
* pfree
* Release an allocated chunk.
*/
void
pfree(void *pointer)
{
MemoryContext context = GetMemoryChunkContext(pointer);
(*context->methods->free_p) (context, pointer);
VALGRIND_MEMPOOL_FREE(context, pointer);
}
/*
* GetMemoryChunkSpace
* Given a currently-allocated chunk, determine the total space
* it occupies (including all memory-allocation overhead).
*
* This is useful for measuring the total space occupied by a set of
* allocated chunks.
*/
Size
GetMemoryChunkSpace(void *pointer)
{
MemoryContext context = GetMemoryChunkContext(pointer);
return (context->methods->get_chunk_space) (context,
pointer);
}
Iterator Iterator_create(HashMap source)
{
Iterator self = (Iterator)PgObjectClass_allocInstance(s_IteratorClass, GetMemoryChunkContext(source));
self->source = source;
self->sourceTableSize = source->tableSize;
self->currentBucket = 0;
self->nextEntry = 0;
return self;
}
static String String_create(TypeClass cls, Oid typeId)
{
HeapTuple typeTup = PgObject_getValidTuple(TYPEOID, typeId, "type");
Form_pg_type pgType = (Form_pg_type)GETSTRUCT(typeTup);
String self = (String)TypeClass_allocInstance(cls, typeId);
MemoryContext ctx = GetMemoryChunkContext(self);
fmgr_info_cxt(pgType->typoutput, &self->textOutput, ctx);
fmgr_info_cxt(pgType->typinput, &self->textInput, ctx);
self->elementType = pgType->typelem;
ReleaseSysCache(typeTup);
return self;
}
static void setupTriggerParams(Function self, ParseResult info)
{
if(info->parameters != 0)
ereport(ERROR, (
errcode(ERRCODE_SYNTAX_ERROR),
errmsg("Triggers can not have a java parameter declaration")));
self->func.nonudt.returnType = Type_fromJavaType(InvalidOid, "void");
/* Parameters are not used when calling triggers.
*/
self->func.nonudt.numParams = 1;
self->func.nonudt.paramTypes = (Type*)MemoryContextAlloc(GetMemoryChunkContext(self), sizeof(Type));
self->func.nonudt.paramTypes[0] = Type_fromJavaType(InvalidOid, "org.postgresql.pljava.TriggerData");
}
static void setupFunctionParams(Function self, ParseResult info, Form_pg_proc procStruct, PG_FUNCTION_ARGS)
{
Oid* paramOids;
MemoryContext ctx = GetMemoryChunkContext(self);
int32 top = (int32)procStruct->pronargs;;
self->func.nonudt.numParams = top;
self->func.nonudt.isMultiCall = procStruct->proretset;
self->func.nonudt.returnType = Type_fromOid(procStruct->prorettype, self->func.nonudt.typeMap);
if(top > 0)
{
int idx;
paramOids = PARAM_OIDS(procStruct);
self->func.nonudt.paramTypes = (Type*)MemoryContextAlloc(ctx, top * sizeof(Type));
for(idx = 0; idx < top; ++idx)
self->func.nonudt.paramTypes[idx] = Type_fromOid(paramOids[idx], self->func.nonudt.typeMap);
}
else
{
self->func.nonudt.paramTypes = 0;
paramOids = 0;
}
if(info->parameters != 0)
parseParameters(self, paramOids, info->parameters);
if(info->returnType != 0)
{
const char* jtName = Type_getJavaTypeName(self->func.nonudt.returnType);
if(strcmp(jtName, info->returnType) != 0)
{
Type repl = Type_fromJavaType(Type_getOid(self->func.nonudt.returnType), info->returnType);
if(!Type_canReplaceType(repl, self->func.nonudt.returnType))
repl = Type_getCoerceOut(repl, self->func.nonudt.returnType);
self->func.nonudt.returnType = repl;
}
}
}
/*
* Compose and dispatch the MPPEXEC commands corresponding to a plan tree
* within a complete parallel plan. (A plan tree will correspond either
* to an initPlan or to the main plan.)
*
* If cancelOnError is true, then any dispatching error, a cancellation
* request from the client, or an error from any of the associated QEs,
* may cause the unfinished portion of the plan to be abandoned or canceled;
* and in the event this occurs before all gangs have been dispatched, this
* function does not return, but waits for all QEs to stop and exits to
* the caller's error catcher via ereport(ERROR,...).Otherwise this
* function returns normally and errors are not reported until later.
*
* If cancelOnError is false, the plan is to be dispatched as fully as
* possible and the QEs allowed to proceed regardless of cancellation
* requests, errors or connection failures from other QEs, etc.
*
* The CdbDispatchResults objects allocated for the plan are returned
* in *pPrimaryResults. The caller, after calling
* CdbCheckDispatchResult(), can examine the CdbDispatchResults
* objects, can keep them as long as needed, and ultimately must free
* them with cdbdisp_destroyDispatcherState() prior to deallocation of
* the caller's memory context. Callers should use PG_TRY/PG_CATCH to
* ensure proper cleanup.
*
* To wait for completion, check for errors, and clean up, it is
* suggested that the caller use cdbdisp_finishCommand().
*
* Note that the slice tree dispatched is the one specified in the EState
* of the argument QueryDesc as es_cur__slice.
*
* Note that the QueryDesc params must include PARAM_EXEC_REMOTE parameters
* containing the values of any initplans required by the slice to be run.
* (This is handled by calls to addRemoteExecParamsToParamList() from the
* functions preprocess_initplans() and ExecutorRun().)
*
* Each QE receives its assignment as a message of type 'M' in PostgresMain().
* The message is deserialized and processed by exec_mpp_query() in postgres.c.
*/
void
cdbdisp_dispatchPlan(struct QueryDesc *queryDesc,
bool planRequiresTxn,
bool cancelOnError, struct CdbDispatcherState *ds)
{
char *splan,
*sddesc,
*sparams;
int splan_len,
splan_len_uncompressed,
sddesc_len,
sparams_len;
SliceTable *sliceTbl;
int rootIdx;
int oldLocalSlice;
PlannedStmt *stmt;
bool is_SRI;
DispatchCommandQueryParms queryParms;
CdbComponentDatabaseInfo *qdinfo;
ds->primaryResults = NULL;
ds->dispatchThreads = NULL;
Assert(Gp_role == GP_ROLE_DISPATCH);
Assert(queryDesc != NULL && queryDesc->estate != NULL);
/*
* Later we'll need to operate with the slice table provided via the
* EState structure in the argument QueryDesc. Cache this information
* locally and assert our expectations about it.
*/
sliceTbl = queryDesc->estate->es_sliceTable;
rootIdx = RootSliceIndex(queryDesc->estate);
Assert(sliceTbl != NULL);
Assert(rootIdx == 0 ||
(rootIdx > sliceTbl->nMotions
&& rootIdx <= sliceTbl->nMotions + sliceTbl->nInitPlans));
/*
* Keep old value so we can restore it. We use this field as a parameter.
*/
oldLocalSlice = sliceTbl->localSlice;
/*
* This function is called only for planned statements.
*/
stmt = queryDesc->plannedstmt;
Assert(stmt);
/*
* Let's evaluate STABLE functions now, so we get consistent values on the QEs
*
* Also, if this is a single-row INSERT statement, let's evaluate
* nextval() and currval() now, so that we get the QD's values, and a
* consistent value for everyone
*
*/
is_SRI = false;
if (queryDesc->operation == CMD_INSERT)
{
Assert(stmt->commandType == CMD_INSERT);
/*
* We might look for constant input relation (instead of SRI), but I'm afraid
* * that wouldn't scale.
*/
is_SRI = IsA(stmt->planTree, Result)
&& stmt->planTree->lefttree == NULL;
}
if (!is_SRI)
clear_relsize_cache();
if (queryDesc->operation == CMD_INSERT ||
queryDesc->operation == CMD_SELECT ||
queryDesc->operation == CMD_UPDATE ||
queryDesc->operation == CMD_DELETE)
{
MemoryContext oldContext;
oldContext = CurrentMemoryContext;
if (stmt->qdContext)
{
oldContext = MemoryContextSwitchTo(stmt->qdContext);
}
else
/*
* memory context of plan tree should not change
*/
{
MemoryContext mc = GetMemoryChunkContext(stmt->planTree);
oldContext = MemoryContextSwitchTo(mc);
}
stmt->planTree = (Plan *) exec_make_plan_constant(stmt, is_SRI);
MemoryContextSwitchTo(oldContext);
}
/*
* Cursor queries and bind/execute path queries don't run on the
* writer-gang QEs; but they require snapshot-synchronization to
* get started.
*
* initPlans, and other work (see the function pre-evaluation
* above) may advance the snapshot "segmateSync" value, so we're
* best off setting the shared-snapshot-ready value here. This
* will dispatch to the writer gang and force it to set its
* snapshot; we'll then be able to serialize the same snapshot
* version (see qdSerializeDtxContextInfo() below).
*/
if (queryDesc->extended_query)
{
verify_shared_snapshot_ready();
}
/*
* serialized plan tree. Note that we're called for a single
* slice tree (corresponding to an initPlan or the main plan), so the
* parameters are fixed and we can include them in the prefix.
*/
splan = serializeNode((Node *) queryDesc->plannedstmt,
&splan_len, &splan_len_uncompressed);
uint64 plan_size_in_kb = ((uint64) splan_len_uncompressed) / (uint64) 1024;
if (0 < gp_max_plan_size && plan_size_in_kb > gp_max_plan_size)
{
ereport(ERROR,
(errcode(ERRCODE_STATEMENT_TOO_COMPLEX),
(errmsg("Query plan size limit exceeded, current size: "
UINT64_FORMAT "KB, max allowed size: %dKB",
plan_size_in_kb, gp_max_plan_size),
errhint("Size controlled by gp_max_plan_size"))));
}
Assert(splan != NULL && splan_len > 0 && splan_len_uncompressed > 0);
if (queryDesc->params != NULL && queryDesc->params->numParams > 0)
{
ParamListInfoData *pli;
ParamExternData *pxd;
StringInfoData parambuf;
Size length;
int plioff;
int32 iparam;
/*
* Allocate buffer for params
*/
initStringInfo(¶mbuf);
/*
* Copy ParamListInfoData header and ParamExternData array
*/
pli = queryDesc->params;
length = (char *) &pli->params[pli->numParams] - (char *) pli;
plioff = parambuf.len;
Assert(plioff == MAXALIGN(plioff));
appendBinaryStringInfo(¶mbuf, pli, length);
/*
* Copy pass-by-reference param values.
*/
for (iparam = 0; iparam < queryDesc->params->numParams; iparam++)
{
int16 typlen;
bool typbyval;
/*
* Recompute pli each time in case parambuf.data is repalloc'ed
*/
pli = (ParamListInfoData *) (parambuf.data + plioff);
pxd = &pli->params[iparam];
if (pxd->ptype == InvalidOid)
continue;
/*
* Does pxd->value contain the value itself, or a pointer?
*/
get_typlenbyval(pxd->ptype, &typlen, &typbyval);
if (!typbyval)
{
char *s = DatumGetPointer(pxd->value);
if (pxd->isnull || !PointerIsValid(s))
{
pxd->isnull = true;
pxd->value = 0;
}
else
{
length = datumGetSize(pxd->value, typbyval, typlen);
/*
* We *must* set this before we
* append. Appending may realloc, which will
* invalidate our pxd ptr. (obviously we could
* append first if we recalculate pxd from the new
* base address)
*/
pxd->value = Int32GetDatum(length);
appendBinaryStringInfo(¶mbuf, &iparam, sizeof(iparam));
appendBinaryStringInfo(¶mbuf, s, length);
}
}
}
sparams = parambuf.data;
sparams_len = parambuf.len;
}
else
{
sparams = NULL;
sparams_len = 0;
}
sddesc = serializeNode((Node *) queryDesc->ddesc, &sddesc_len, NULL /*uncompressed_size */ );
MemSet(&queryParms, 0, sizeof(queryParms));
queryParms.strCommand = queryDesc->sourceText;
queryParms.serializedQuerytree = NULL;
queryParms.serializedQuerytreelen = 0;
queryParms.serializedPlantree = splan;
queryParms.serializedPlantreelen = splan_len;
queryParms.serializedParams = sparams;
queryParms.serializedParamslen = sparams_len;
queryParms.serializedQueryDispatchDesc = sddesc;
queryParms.serializedQueryDispatchDesclen = sddesc_len;
queryParms.rootIdx = rootIdx;
/*
* sequence server info
*/
qdinfo = &(getComponentDatabases()->entry_db_info[0]);
Assert(qdinfo != NULL && qdinfo->hostip != NULL);
queryParms.seqServerHost = pstrdup(qdinfo->hostip);
queryParms.seqServerHostlen = strlen(qdinfo->hostip) + 1;
queryParms.seqServerPort = seqServerCtl->seqServerPort;
/*
* serialized a version of our snapshot
*/
/*
* Generate our transction isolations. We generally want Plan
* based dispatch to be in a global transaction. The executor gets
* to decide if the special circumstances exist which allow us to
* dispatch without starting a global xact.
*/
queryParms.serializedDtxContextInfo =
qdSerializeDtxContextInfo(&queryParms.serializedDtxContextInfolen,
true /* wantSnapshot */ ,
queryDesc->extended_query,
mppTxnOptions(planRequiresTxn),
"cdbdisp_dispatchPlan");
cdbdisp_dispatchX(&queryParms, cancelOnError, sliceTbl, ds);
sliceTbl->localSlice = oldLocalSlice;
}