Splitted the undo log into two parts: in-memory one for initial undo …

…items (but they're much more dense now) and on-disk (cached up to TempSpaceLimit) one for full record versions (used in update-in-place).

This dramatically reduces the memory usage for bulk updates, thus solving CORE-1575 (serious memory bug on multiple update a table in a single transaction) and CORE-1477 (Avoiding excessive memory consumption by undo-log after series of updates) and predicting out-of-memory conditions that could often lead to record-level corruptions.

src/jrd/blb.cpp

@@ -217,7 +217,7 @@ void BLB_close(thread_db* tdbb, Jrd::blb* blob)
 		{
 			blob->blb_temp_size += BLP_SIZE;
 			jrd_tra* transaction = blob->blb_transaction;
-			TempSpace* tempSpace = transaction->getTempSpace();
+			TempSpace* const tempSpace = transaction->getBlobSpace();
 
 			blob->blb_temp_offset = tempSpace->allocateSpace(blob->blb_temp_size);
 			tempSpace->write(blob->blb_temp_offset,
@@ -1310,7 +1310,7 @@ blb* BLB_open2(thread_db* tdbb,
 
 				if (new_blob->blb_temp_size > 0)
 				{
-					transaction->getTempSpace()->read(new_blob->blb_temp_offset,
+					transaction->getBlobSpace()->read(new_blob->blb_temp_offset,
 						blob->getBuffer(), new_blob->blb_temp_size);
 				}
 
@@ -2641,7 +2641,7 @@ static void release_blob(blb* blob, const bool purge_flag)
 
 	if ((blob->blb_flags & BLB_temporary) && blob->blb_temp_size > 0)
 	{
-		blob->blb_transaction->getTempSpace()->releaseSpace(
+		blob->blb_transaction->getBlobSpace()->releaseSpace(
 			blob->blb_temp_offset, blob->blb_temp_size);
 	}
 

src/jrd/dpm.epp

@@ -2010,7 +2010,7 @@ RecordNumber DPM_store_blob(thread_db* tdbb, blb* blob, Record* record)
 		{
 			if (blob->blb_temp_size > 0)
 			{
-				blob->blb_transaction->getTempSpace()->read(
+				blob->blb_transaction->getBlobSpace()->read(
 					blob->blb_temp_offset, buffer.getBuffer(blob->blb_temp_size),
 					blob->blb_temp_size);
 				q = buffer.begin();