diff --git a/Src/stm32f4xx_it.c b/Src/stm32f4xx_it.c
index ceb5b6b..c18a418 100644
--- a/Src/stm32f4xx_it.c
+++ b/Src/stm32f4xx_it.c
@@ -35,6 +35,7 @@ extern "C" {
 }
 
 #include "globals_cpp.h"
+#include "ll_memcpydma.h"
 
 extern "C" {
 
@@ -427,7 +428,7 @@ void DMA2_Stream2_IRQHandler(void)
 void DMA2_Stream3_IRQHandler(void)
 {
   /* USER CODE BEGIN DMA2_Stream3_IRQn 0 */
-  HandleMemcpyDmaIrq();
+  f4ll::LL_MemcpyDma::Instance().DmaTransferCompleted();
 
   /* USER CODE END DMA2_Stream3_IRQn 0 */
   
diff --git a/app/globals.c b/app/globals.c
index 540cb61..edf988c 100644
--- a/app/globals.c
+++ b/app/globals.c
@@ -12,5 +12,5 @@ USARTSTATUS g_uartStatuses[USARTCOUNT];
 CRCSTATUS g_crcStatus;
 
 DMAINFO g_ConsoleTxDmaInfo;
-uint8_t g_statsBuf[128];
+uint8_t g_statsBuf[256];
 
diff --git a/app/globals.h b/app/globals.h
index 526fd13..439f19c 100644
--- a/app/globals.h
+++ b/app/globals.h
@@ -19,6 +19,6 @@ extern USARTSTATUS g_uartStatuses[USARTCOUNT];
 extern CRCSTATUS g_crcStatus;
 
 extern DMAINFO g_ConsoleTxDmaInfo;
-extern uint8_t g_statsBuf[128];
+extern uint8_t g_statsBuf[256];
 
 #endif /* GLOBALS_H_ */
diff --git a/app/globals_cpp.cpp b/app/globals_cpp.cpp
index 6f0e59e..81e55f9 100644
--- a/app/globals_cpp.cpp
+++ b/app/globals_cpp.cpp
@@ -7,7 +7,7 @@
 #include "globals.h"
 
 #include "ll_hsusart.h"
+#include "ll_memcpydma.h"
 
 f4ll::LL_HsUsart *g_usarts[4];
 
-
diff --git a/app/ll_testbed.cpp b/app/ll_testbed.cpp
index d588b62..309846b 100644
--- a/app/ll_testbed.cpp
+++ b/app/ll_testbed.cpp
@@ -8,10 +8,12 @@
 #include <stdlib.h>
 #include "ll_hsusart.h"
 #include "ll_crchandler.h"
+#include "ll_memcpydma.h"
 extern "C" {
 	#include "main.h"
 	#include "globals.h"
 	#include "strutil.h"
+	#include "config.h"
 }
 #include "globals_cpp.h"
 
@@ -59,9 +61,9 @@ extern "C" void MainLoop()
 			"Megszentsegtelenithetetlensegeskedeseitekert\r\n"
 			"--------------------------------------------\r\n\0\0\0";
 
+	f4ll::LL_MemcpyDma::Init(MEMCPY_DMA_ENGINE, MEMCPY_DMA_STREAM);
 	f4ll::LL_CrcHandler::Init(DMA2, LL_DMA_STREAM_4);
 
-	f4ll::LL_CrcHandler::Slot<2> slt;
 	f4ll::LL_HsUsart u1{ USART1, DMA2, LL_DMA_STREAM_2, LL_DMA_STREAM_7 };
 	f4ll::LL_HsUsart u2{ USART2, DMA1, LL_DMA_STREAM_5, LL_DMA_STREAM_6 };
 	f4ll::LL_HsUsart u3{ USART3, DMA1, LL_DMA_STREAM_1, LL_DMA_STREAM_3 };
@@ -101,25 +103,26 @@ extern "C" void MainLoop()
 		for(auto u : usarts) {
 			if(!u->IsTxBusy() && send) {
 				//DIAG_ENTER_BUSY();
-				u->PostPacket(text2Send, sizeof(text2Send) - 1 - (rand() & randmask));
+				auto len = sizeof(text2Send) - 1 - (rand() & randmask);
+				f4ll::LL_MemcpyDma::Instance().Copy(u->GetTxPacketBuffer(), text2Send, len);
+				u->PostPacket(nullptr, len);
 				//DIAG_EXIT_BUSY();
 			}
 			for(uint16_t rIdx = 0; rIdx < 2; ++rIdx)
-				if(u->IsRxBusy((bool)rIdx) || u->IsRxFailed(rIdx)) {
+				if(u->IsRxBusy((bool)rIdx) || u->IsRxFailed(rIdx))
 					u->RxProcessed((bool)rIdx);
-				}
 		}
 		if(tick - lastStatsTick > STATS_DELAY_MS) {
 			_PrintStats((char*)g_statsBuf, statId, *usarts[statId], UART4, &g_ConsoleTxDmaInfo);
 			lastStatsTick += STATS_DELAY_MS;
 			++statId;
-			if(statId >= USARTCOUNT)
+			if(statId >= sizeof(usarts) / sizeof(usarts[0]))
 				statId = 0;
 		}
-		uint32_t ein = LL_GPIO_ReadInputPort(KEY1_GPIO_Port);
-		if(!(ein & KEY1_Pin)) {
-			void (*fptr)(void) = (void (*)(void))(void*)0xa0000000;
-			fptr();
-		}
+//		uint32_t ein = LL_GPIO_ReadInputPort(KEY1_GPIO_Port);
+//		if(!(ein & KEY1_Pin)) {
+//			void (*fptr)(void) = (void (*)(void))(void*)0xa0000000;
+//			fptr();
+//		}
 	}
 }
diff --git a/lib/ll_memcpydma.cpp b/lib/ll_memcpydma.cpp
index 7ea564b..64c52c3 100644
--- a/lib/ll_memcpydma.cpp
+++ b/lib/ll_memcpydma.cpp
@@ -15,4 +15,23 @@ LL_MemcpyDma::LL_MemcpyDma(DMA_TypeDef *dma, uint32_t stream)
 	LL_DMA_EnableIT_TC(dma, stream);
 }
 
+void* LL_MemcpyDma::Copy(void *dst, void const *src, uint16_t length)
+{
+	LL_DMA_SetM2MSrcAddress(GetDma(), GetStream(), (uint32_t)src);
+	LL_DMA_SetM2MDstAddress(GetDma(), GetStream(), (uint32_t)dst);
+	LL_DMA_SetDataLength(GetDma(), GetStream(), (length+3)/4 );
+	m_busy = 1;
+	LL_DMA_EnableStream(GetDma(), GetStream());
+	while(m_busy);
+	return dst;
+}
+
+void LL_MemcpyDma::DmaTransferCompleted()
+{
+	if(*GetIsReg() & GetTcMask()) {	// DMA transfer complete
+		*GetIcfReg() = GetTcMask();
+		LL_DMA_DisableStream(GetDma(), GetStream());
+		m_busy = 0;
+	}
+}
 } /* namespace f4ll */
diff --git a/lib/ll_memcpydma.h b/lib/ll_memcpydma.h
index 23dd0b1..c00ee1d 100644
--- a/lib/ll_memcpydma.h
+++ b/lib/ll_memcpydma.h
@@ -8,15 +8,19 @@
 #ifndef LL_MEMCPY_DMA_H_
 #define LL_MEMCPY_DMA_H_
 #include "ll_dmahelper.h"
+#include "singleton.h"
 
 namespace f4ll {
 
-class LL_MemcpyDma : private LL_DmaHelper
+class LL_MemcpyDma : public Singleton<LL_MemcpyDma>, private LL_DmaHelper
 {
+	friend class Singleton<LL_MemcpyDma>;
 public:
+	void* Copy(void *dst, void const *src, uint16_t length);
+	void DmaTransferCompleted();
+private:
 	LL_MemcpyDma(DMA_TypeDef *dma, uint32_t stream);
-	void Copy(void *dst, void const *src, uint16_t length);
-
+	bool volatile m_busy = false;
 };
 
 } /* namespace f4ll */
diff --git a/lib/singleton.h b/lib/singleton.h
index 8cd9886..b64c345 100644
--- a/lib/singleton.h
+++ b/lib/singleton.h
@@ -10,8 +10,7 @@ public:
     static T &Init(Args &&... args)
     {
         static T instance{ std::forward<Args>(args)... };
-        if(!m_instance)
-        	m_instance = &instance;
+       	m_instance = &instance;
         return instance;
     }