MBox, some improvements

Some changes reverted (warnings), multi-thread safety fixed, MBox MMIO
modified

rpcs3/Emu/Cell/MFC.h

@@ -163,14 +163,17 @@ struct DMAC
 			return MFC_PPU_DMA_QUEUE_FULL;
 		}
 
-		while (_InterlockedCompareExchange(&proxy_lock,  1, 0));
+		while (_InterlockedExchange(&proxy_lock, 1));
+		_mm_lfence();
 		DMAC_Proxy& p = proxy[proxy_pos];
 		p.cmd = cmd;
 		p.tag = tag;
 		p.lsa = lsa;
 		p.ea = ea;
 		p.size = size;
+		_mm_sfence(); //for DoCmd()
 		proxy_pos++;
+		_mm_sfence();
 		proxy_lock = 0;
 
 		return MFC_PPU_DMA_CMD_ENQUEUE_SUCCESSFUL;
@@ -178,8 +181,10 @@ struct DMAC
 
 	void ClearCmd()
 	{
-		while (_InterlockedCompareExchange(&proxy_lock,  1, 0));
+		while (_InterlockedExchange(&proxy_lock, 1));
+		_mm_lfence();
 		memcpy(proxy, proxy + 1, --proxy_pos * sizeof(DMAC_Proxy));
+		_mm_sfence();
 		proxy_lock = 0; //release lock
 	}
 

rpcs3/Emu/Cell/RawSPUThread.cpp

@@ -62,11 +62,17 @@ bool RawSPUThread::Read32(const u64 addr, u32* value)
 	case MFC_EAL_offs:				ConLog.Warning("RawSPUThread[%d]: Read32(MFC_EAL)", m_index);			*value = MFC2.EAL.GetValue(); break;
 	case MFC_Size_Tag_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(MFC_Size_Tag)", m_index);		*value = MFC2.Size_Tag.GetValue(); break;
 	case MFC_CMDStatus_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(MFC_CMDStatus)", m_index);		*value = MFC2.CMDStatus.GetValue(); break;
-	case MFC_QStatus_offs:		*value = MFC2.QStatus.GetValue(); break;
+	case MFC_QStatus_offs:		
+		ConLog.Warning("RawSPUThread[%d]: Read32(MFC_QStatus)", m_index);
+		*value = MFC2.QStatus.GetValue(); 
+	break;
 	case Prxy_QueryType_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(Prxy_QueryType)", m_index);	*value = Prxy.QueryType.GetValue(); break;
 	case Prxy_QueryMask_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(Prxy_QueryMask)", m_index);	*value = Prxy.QueryMask.GetValue(); break;
 	case Prxy_TagStatus_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(Prxy_TagStatus)", m_index);	*value = Prxy.TagStatus.GetValue(); break;
-	case SPU_Out_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(SPU_Out_MBox)", m_index);		while(!SPU.Out_MBox.Pop(*value) && !Emu.IsStopped()) Sleep(1); break;
+	case SPU_Out_MBox_offs:			
+		ConLog.Warning("RawSPUThread[%d]: Read32(SPU_Out_MBox)", m_index);
+		SPU.Out_MBox.PopUncond(*value); //if Out_MBox is empty yet, the result will be undefined 
+	break;
 	case SPU_In_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(SPU_In_MBox)", m_index);		while(!SPU.In_MBox.Pop(*value)  && !Emu.IsStopped()) Sleep(1); break;
 	case SPU_MBox_Status_offs:		//ConLog.Warning("RawSPUThread[%d]: Read32(SPU_MBox_Status)", m_index);
 		//SPU.MBox_Status.SetValue(SPU.Out_MBox.GetCount() ? SPU.MBox_Status.GetValue() | 1 : SPU.MBox_Status.GetValue() & ~1);
@@ -183,7 +189,10 @@ bool RawSPUThread::Write32(const u64 addr, const u32 value)
 	case Prxy_QueryMask_offs:		ConLog.Warning("RawSPUThread[%d]: Write32(Prxy_QueryMask, 0x%x)", m_index, value);		Prxy.QueryMask.SetValue(value); break;
 	case Prxy_TagStatus_offs:		ConLog.Warning("RawSPUThread[%d]: Write32(Prxy_TagStatus, 0x%x)", m_index, value);		Prxy.TagStatus.SetValue(value); break;
 	case SPU_Out_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_Out_MBox, 0x%x)", m_index, value);		while(!SPU.Out_MBox.Push(value) && !Emu.IsStopped()) Sleep(1); break;
-	case SPU_In_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_In_MBox, 0x%x)", m_index, value);			while(!SPU.In_MBox.Push(value) && !Emu.IsStopped()) Sleep(1); break;
+	case SPU_In_MBox_offs:			
+		ConLog.Warning("RawSPUThread[%d]: Write32(SPU_In_MBox, 0x%x)", m_index, value);
+		SPU.In_MBox.PushUncond(value); //if In_MBox is already full, the last message will be overwritten  
+	break;
 	case SPU_MBox_Status_offs:		ConLog.Warning("RawSPUThread[%d]: Write32(SPU_MBox_Status, 0x%x)", m_index, value);		SPU.MBox_Status.SetValue(value); break;
 	case SPU_RunCntl_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_RunCntl, 0x%x)", m_index, value);			SPU.RunCntl.SetValue(value); break;
 	case SPU_Status_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_Status, 0x%x)", m_index, value);			SPU.Status.SetValue(value); break;

rpcs3/Emu/Cell/SPUThread.h

@@ -282,37 +282,69 @@ public:
 
 		__forceinline bool Pop(u32& res)
 		{
-			while (_InterlockedCompareExchange(&m_lock,  1, 0));
+			while (_InterlockedExchange(&m_lock, 1));
+			_mm_lfence();
 			if(!m_index) 
 			{
 				m_lock = 0; //release lock
 				return false;
 			}
 			res = m_value[--m_index];
+			_mm_sfence();
 			m_lock = 0;
 			return true;
 		}
 
 		__forceinline bool Push(u32 value)
 		{
-			while (_InterlockedCompareExchange(&m_lock,  1, 0));
+			while (_InterlockedExchange(&m_lock, 1));
+			_mm_lfence();
 			if(m_index >= max_count) 
 			{
 				m_lock = 0; //release lock
 				return false;
 			}
 			m_value[m_index++] = value;
+			_mm_sfence();
 			m_lock = 0;
 			return true;
 		}
 
+		__forceinline void PushUncond(u32 value)
+		{
+			while (_InterlockedExchange(&m_lock, 1));
+			_mm_lfence();
+			if(m_index >= max_count) 
+				m_value[max_count-1] = value; //last message is overwritten
+			else
+				m_value[m_index++] = value;
+			_mm_sfence();
+			m_lock = 0;
+		}
+
+		__forceinline void PopUncond(u32& res)
+		{
+			while (_InterlockedExchange(&m_lock, 1));
+			_mm_lfence();
+			if(!m_index) 
+				res = 0; //result is undefined
+			else
+				res = m_value[--m_index];
+			_mm_sfence();
+			m_lock = 0;
+		}
+
 		u32 GetCount() const
 		{
+			while (m_lock);
+			_mm_lfence();
 			return m_index;
 		}
 
 		u32 GetFreeCount() const
 		{
+			while (m_lock);
+			_mm_lfence();
 			return max_count - m_index;
 		}
 
@@ -383,6 +415,14 @@ public:
 			u16 tag = (u16)size_tag;
 			u16 size = size_tag >> 16;
 
+			ConLog.Warning("DMA %s:", op & MFC_PUT_CMD ? "PUT" : "GET");
+			ConLog.Warning("*** lsa  = 0x%x", lsa);
+			ConLog.Warning("*** ea   = 0x%llx", ea);
+			ConLog.Warning("*** tag  = 0x%x", tag);
+			ConLog.Warning("*** size = 0x%x", size);
+			ConLog.Warning("*** cmd  = 0x%x", cmd);
+			ConLog.SkipLn();
+
 			MFCArgs.CMDStatus.SetValue(dmac.Cmd(cmd, tag, lsa, ea, size));
 		}
 		break;
@@ -406,22 +446,6 @@ public:
 		case SPU_WrOutIntrMbox:
 			return 0;//return SPU.OutIntr_Mbox.GetFreeCount();
 
-		case MFC_LSA:
-			return MFC1.LSA.max_count;
-
-		case MFC_EAH:
-			return MFC1.EAH.max_count;
-
-		case MFC_EAL:
-			return MFC1.EAL.max_count;
-
-		case MFC_Size:
-		case MFC_TagID:
-			return MFC1.Size_Tag.max_count;
-		
-		case MFC_Cmd:
-			return MFC1.CMDStatus.max_count;
-
 		default:
 			ConLog.Error("%s error: unknown/illegal channel (%d [%s]).", __FUNCTION__, ch, spu_ch_name[ch]);
 		break;