Fixed bug CORE-4444 : Engine could hung and block all attachments in out of disk space condition during physical backup

Improvement CORE-4445 : Extend main database file faster when physical backup state changed from stalled to merge

src/jrd/GlobalRWLock.cpp

@@ -188,7 +188,7 @@ bool GlobalRWLock::lockWrite(thread_db* tdbb, SSHORT wait)
 	}
 }
 
-void GlobalRWLock::unlockWrite(thread_db* tdbb)
+void GlobalRWLock::unlockWrite(thread_db* tdbb, const bool release)
 {
 	SET_TDBB(tdbb);
 
@@ -201,13 +201,12 @@ void GlobalRWLock::unlockWrite(thread_db* tdbb)
 
 	currentWriter = false;
 
-	if (!lockCaching)
+	if (!lockCaching || release)
 		LCK_release(tdbb, cachedLock);
 	else if (blocking)
-	{
 		LCK_downgrade(tdbb, cachedLock);
-		blocking = false;
-	}
+
+	blocking = false;
 
 	if (cachedLock->lck_physical < LCK_read)
 		invalidate(tdbb);

src/jrd/GlobalRWLock.h

@@ -65,7 +65,7 @@ public:
 
 	// This function returns false if it cannot take the lock
 	bool lockWrite(thread_db* tdbb, SSHORT wait);
-	void unlockWrite(thread_db* tdbb);
+	void unlockWrite(thread_db* tdbb, const bool release = false);
 	bool lockRead(thread_db* tdbb, SSHORT wait, const bool queueJump = false);
 	void unlockRead(thread_db* tdbb);
 	bool tryReleaseLock(thread_db* tdbb);

src/jrd/nbak.cpp

@@ -360,6 +360,46 @@ ULONG BackupManager::getPageCount()
 }
 
 
+bool BackupManager::extendDatabase(thread_db* tdbb)
+{
+	ULONG maxPage = 0;
+	{
+		LocalAllocReadGuard localAllocGuard(this);
+		AllocItemTree::Accessor all(alloc_table);
+
+		if (all.getFirst()) {
+			do
+			{
+				const ULONG pg = all.current().db_page;
+				if (maxPage < pg)
+					maxPage = pg;
+			} while (all.getNext());
+		}
+	}
+
+	PageSpace *pgSpace = database->dbb_page_manager.findPageSpace(DB_PAGE_SPACE);
+	ULONG maxAllocPage = pgSpace->maxAlloc(database->dbb_page_size);
+	if (maxAllocPage >= maxPage)
+		return true;
+
+	if (!pgSpace->extend(tdbb, maxPage, true))
+		return false;
+
+	maxAllocPage = pgSpace->maxAlloc(database->dbb_page_size);
+	while (maxAllocPage < maxPage)
+	{
+		const USHORT ret = PIO_init_data(database, pgSpace->file, tdbb->tdbb_status_vector, 
+										 maxAllocPage, 256);
+
+		if (ret != 256)
+			return false;
+
+		maxAllocPage += ret;
+	}
+	return true;
+}
+
+
 // Merge difference file to main files (if needed) and unlink() difference
 // file then. If merge is already in progress method silently returns and
 // does nothing (so it can be used for recovery on database startup).
@@ -398,6 +438,9 @@ void BackupManager::endBackup(thread_db* tdbb, bool recover)
 				endLock.unlockWrite(tdbb);
 				return;
 			}
+
+			if (backup_state == nbak_state_stalled && !extendDatabase(tdbb))
+				status_exception::raise(tdbb->tdbb_status_vector);
 		}
 
 		// Here backup state can be changed. Need to check it again after lock
@@ -410,6 +453,13 @@ void BackupManager::endBackup(thread_db* tdbb, bool recover)
 			endLock.unlockWrite(tdbb);
 			return;
 		}
+
+		if (!extendDatabase(tdbb))
+		{
+			stateGuard.setSuccess();
+			status_exception::raise(tdbb->tdbb_status_vector);
+		}
+
 		header = (Ods::header_page*) window.win_buffer;
 
 		NBAK_TRACE(("difference file %s, current backup state is %d", diff_name.c_str(), backup_state));

src/jrd/nbak.h

@@ -365,7 +365,7 @@ public:
 	void unlockStateWrite(thread_db* tdbb)
 	{
 		tdbb->tdbb_flags &= ~TDBB_backup_write_locked;
-		stateLock->unlockWrite(tdbb);
+		stateLock->unlockWrite(tdbb, backup_state == nbak_state_unknown);
 	}
 
 	bool lockStateRead(thread_db* tdbb, SSHORT wait)
@@ -464,6 +464,7 @@ private:
 
 	ULONG findPageIndex(thread_db* tdbb, ULONG db_page);
 	void generateFilename();
+	bool extendDatabase(thread_db* tdbb);
 
 	void lockAllocWrite(thread_db* tdbb)
 	{