Merge pull request #370 from Nekotekina/master

Fixed some samples freeze

Utilities/SSemaphore.cpp

@@ -1,29 +1,45 @@
 #include "stdafx.h"
 #include "Utilities/SSemaphore.h"
 
-bool SSemaphore::wait(u64 timeout)
+void SSemaphore::wait()
 {
-	std::unique_lock<std::mutex> lock(m_cv_mutex);
+	u32 order;
+	{
+		std::lock_guard<std::mutex> lock(m_mutex);
+		if (m_count && m_out_order == m_in_order)
+		{
+			m_count--;
+			return;
+		}
+		order = m_in_order++;
+	}
+
+	std::unique_lock<std::mutex> cv_lock(m_cv_mutex);
 	
-	u64 counter = 0;
 	while (true)
 	{
 		if (Emu.IsStopped())
 		{
-			return false;
+			return;
 		}
-		if (timeout && counter >= timeout)
-		{
-			return false;
-		}
-		m_cond.wait_for(lock, std::chrono::milliseconds(1));
-		counter++;
 
+		m_cond.wait_for(cv_lock, std::chrono::milliseconds(1));
+		
+		{
 			std::lock_guard<std::mutex> lock(m_mutex);
 			if (m_count)
+			{
+				if (m_out_order == order)
 				{
 					m_count--;
-			return true;
+					m_out_order++;
+					return;
+				}
+				else
+				{
+					m_cond.notify_one();
+				}
+			}
 		}
 	}
 }
@@ -32,7 +48,7 @@ bool SSemaphore::try_wait()
 {
 	std::lock_guard<std::mutex> lock(m_mutex);
 
-	if (m_count)
+	if (m_count && m_in_order == m_out_order)
 	{
 		m_count--;
 		return true;
@@ -47,18 +63,21 @@ void SSemaphore::post()
 {
 	std::lock_guard<std::mutex> lock(m_mutex);
 
-	if (m_count >= m_max)
+	if (m_count < m_max)
+	{
+		m_count++;
+	}
+	else
 	{
 		return;
 	}
 
-	m_count++;
 	m_cond.notify_one();
 }
 
 bool SSemaphore::post_and_wait()
 {
-	// TODO: ???
+	// TODO: merge these functions? Probably has a race condition.
 	if (try_wait()) return false;
 
 	post();

Utilities/SSemaphore.h

@@ -4,19 +4,26 @@ class SSemaphore
 {
 	const u32 m_max;
 	u32 m_count;
-	std::mutex m_mutex, m_cv_mutex;
+	u32 m_in_order;
+	u32 m_out_order;
+	std::mutex m_cv_mutex;
+	std::mutex m_mutex;
 	std::condition_variable m_cond;
 
 public:
 	SSemaphore(u32 value, u32 max = 1)
 		: m_max(max > 0 ? max : 0xffffffff)
 		, m_count(value > m_max ? m_max : value)
+		, m_in_order(0)
+		, m_out_order(0)
 	{
 	}
 
 	SSemaphore()
 		: m_max(0xffffffff)
 		, m_count(0)
+		, m_in_order(0)
+		, m_out_order(0)
 	{
 	}
 
@@ -24,7 +31,7 @@ public:
 	{
 	}
 
-	bool wait(u64 timeout = 0);
+	void wait();
 
 	bool try_wait();
 

rpcs3/Emu/Cell/RawSPUThread.h

@@ -2,36 +2,6 @@
 #include "SPUThread.h"
 #include "Emu/event.h"
 
-enum
-{
-	MFC_LSA_offs          = 0x3004,
-	MFC_EAH_offs          = 0x3008,
-	MFC_EAL_offs          = 0x300C,
-	MFC_Size_Tag_offs     = 0x3010,
-	MFC_Class_CMD_offs    = 0x3014,
-	MFC_CMDStatus_offs    = 0x3014,
-	MFC_QStatus_offs      = 0x3104,
-	Prxy_QueryType_offs   = 0x3204,
-	Prxy_QueryMask_offs   = 0x321C,
-	Prxy_TagStatus_offs   = 0x322C,
-	SPU_Out_MBox_offs     = 0x4004,
-	SPU_In_MBox_offs      = 0x400C,
-	SPU_MBox_Status_offs  = 0x4014,
-	SPU_RunCntl_offs      = 0x401C,
-	SPU_Status_offs       = 0x4024,
-	SPU_NPC_offs          = 0x4034,
-	SPU_RdSigNotify1_offs = 0x1400C,
-	SPU_RdSigNotify2_offs = 0x1C00C,
-};
-
-enum : u64
-{
-	RAW_SPU_OFFSET      = 0x0000000000100000,
-	RAW_SPU_BASE_ADDR   = 0x00000000E0000000,
-	RAW_SPU_LS_OFFSET   = 0x0000000000000000,
-	RAW_SPU_PROB_OFFSET = 0x0000000000040000,
-};
-
 __forceinline static u32 GetRawSPURegAddrByNum(int num, int offset)
 {
 	return RAW_SPU_OFFSET * num + RAW_SPU_BASE_ADDR + RAW_SPU_PROB_OFFSET + offset;

rpcs3/Emu/Cell/SPUThread.h

@@ -129,6 +129,36 @@ enum : u32
 	SYS_SPU_THREAD_SNR2      = 0x05C00c,
 };
 
+enum
+{
+	MFC_LSA_offs = 0x3004,
+	MFC_EAH_offs = 0x3008,
+	MFC_EAL_offs = 0x300C,
+	MFC_Size_Tag_offs = 0x3010,
+	MFC_Class_CMD_offs = 0x3014,
+	MFC_CMDStatus_offs = 0x3014,
+	MFC_QStatus_offs = 0x3104,
+	Prxy_QueryType_offs = 0x3204,
+	Prxy_QueryMask_offs = 0x321C,
+	Prxy_TagStatus_offs = 0x322C,
+	SPU_Out_MBox_offs = 0x4004,
+	SPU_In_MBox_offs = 0x400C,
+	SPU_MBox_Status_offs = 0x4014,
+	SPU_RunCntl_offs = 0x401C,
+	SPU_Status_offs = 0x4024,
+	SPU_NPC_offs = 0x4034,
+	SPU_RdSigNotify1_offs = 0x1400C,
+	SPU_RdSigNotify2_offs = 0x1C00C,
+};
+
+enum : u64
+{
+	RAW_SPU_OFFSET = 0x0000000000100000,
+	RAW_SPU_BASE_ADDR = 0x00000000E0000000,
+	RAW_SPU_LS_OFFSET = 0x0000000000000000,
+	RAW_SPU_PROB_OFFSET = 0x0000000000040000,
+};
+
 //Floating point status and control register.  Unsure if this is one of the GPRs or SPRs
 //Is 128 bits, but bits 0-19, 24-28, 32-49, 56-60, 64-81, 88-92, 96-115, 120-124 are unused
 class FPSCR
@@ -598,19 +628,20 @@ public:
 				return false;
 			}
 		}
-
+		else if (ea >= RAW_SPU_BASE_ADDR && size == 4)
-		//Sleep(1); // hack
+		{
-
+			switch (cmd & ~(MFC_BARRIER_MASK | MFC_FENCE_MASK | MFC_LIST_MASK | MFC_RESULT_MASK))
-		switch(cmd & ~(MFC_BARRIER_MASK | MFC_FENCE_MASK | MFC_LIST_MASK | MFC_RESULT_MASK))
 			{
 			case MFC_PUT_CMD:
 			{
-				return Memory.Copy(ea, dmac.ls_offset + lsa, size);
+				Memory.Write32(ea, ReadLS32(lsa));
+				return true; 
 			}
 
 			case MFC_GET_CMD:
 			{
-				return Memory.Copy(dmac.ls_offset + lsa, ea, size);
+				WriteLS32(lsa, Memory.Read32(ea));
+				return true;
 			}
 
 			default:
@@ -621,6 +652,44 @@ public:
 			}
 		}
 
+		//Sleep(1); // hack
+
+		switch (cmd & ~(MFC_BARRIER_MASK | MFC_FENCE_MASK | MFC_LIST_MASK | MFC_RESULT_MASK))
+		{
+		case MFC_PUT_CMD:
+		{
+			if (Memory.Copy(ea, dmac.ls_offset + lsa, size))
+			{
+				return true;
+			}
+			else
+			{
+				ConLog.Error("DMAC::ProcessCmd(): PUT* cmd failed (ea=0x%llx, lsa=0x%x, size=%d)", ea, lsa, size);
+				return false; // TODO: page fault (?)
+			}
+		}
+
+		case MFC_GET_CMD:
+		{
+			if (Memory.Copy(dmac.ls_offset + lsa, ea, size))
+			{
+				return true;
+			}
+			else
+			{
+				ConLog.Error("DMAC::ProcessCmd(): GET* cmd failed (ea=0x%llx, lsa=0x%x, size=%d)", ea, lsa, size);
+				return false; // TODO: page fault (?)
+			}
+		}
+
+		default:
+		{
+			ConLog.Error("DMAC::ProcessCmd(): Unknown DMA cmd.");
+			return false; // ???
+		}
+		}
+	}
+
 	u32 dmacCmd(u32 cmd, u32 tag, u32 lsa, u64 ea, u32 size)
 	{
 		/*if(proxy_pos >= MFC_PPU_MAX_QUEUE_SPACE)

rpcs3/Emu/GS/RSXThread.cpp

@@ -1882,8 +1882,13 @@ void RSXThread::Task()
 		inc=1;
 
 		u32 put, get;
+		// this code produces only mov + bswap:
 		se_t<u32>::func(put, std::atomic_load((volatile std::atomic<u32>*)((u8*)m_ctrl + offsetof(CellGcmControl, put))));
 		se_t<u32>::func(get, std::atomic_load((volatile std::atomic<u32>*)((u8*)m_ctrl + offsetof(CellGcmControl, get))));
+		/*
+		se_t<u32>::func(put, InterlockedCompareExchange((volatile unsigned long*)((u8*)m_ctrl + offsetof(CellGcmControl, put)), 0, 0));
+		se_t<u32>::func(get, InterlockedCompareExchange((volatile unsigned long*)((u8*)m_ctrl + offsetof(CellGcmControl, get)), 0, 0));
+		*/
 
 		if(put == get || !Emu.IsRunning())
 		{

rpcs3/Emu/SysCalls/Modules/cellGcmSys.cpp

@@ -471,7 +471,7 @@ int cellGcmSetPrepareFlip(mem_ptr_t<CellGcmContextData> ctxt, u32 id)
 		return CELL_GCM_ERROR_FAILURE;
 	}
 
-	GSLockCurrent gslock(GS_LOCK_WAIT_FLUSH); // could freeze on exit
+	GSLockCurrent gslock(GS_LOCK_WAIT_FLUSH);
 
 	u32 current = ctxt->current;
 	u32 end = ctxt->end;
@@ -488,6 +488,8 @@ int cellGcmSetPrepareFlip(mem_ptr_t<CellGcmContextData> ctxt, u32 id)
  
 		if(res > 0) Memory.Copy(ctxt->begin, ctxt->current - res, res);
 		ctxt->current = ctxt->begin + res;
+
+		//InterlockedExchange64((volatile long long*)((u8*)&ctrl + offsetof(CellGcmControl, put)), (u64)(u32)re(res));
 		ctrl.put = res;
 		ctrl.get = 0;
 	}

rpcs3/Emu/SysCalls/Modules/cellResc.cpp

@@ -802,7 +802,7 @@ int cellRescSetConvertAndFlip(mem_ptr_t<CellGcmContextData> cntxt, s32 idx)
 int cellRescSetWaitFlip()
 {
 	cellResc->Log("cellRescSetWaitFlip()");
-	GSLockCurrent lock(GS_LOCK_WAIT_FLIP); // could stall on exit
+	GSLockCurrent lock(GS_LOCK_WAIT_FLIP);
 
 	return CELL_OK;
 }

rpcs3/Emu/SysCalls/lv2/SC_GCM.cpp

@@ -13,7 +13,7 @@ extern gcmInfo gcm_info;
 
 int cellGcmCallback(u32 context_addr, u32 count)
 {
-	GSLockCurrent gslock(GS_LOCK_WAIT_FLUSH); // could freeze on exit
+	GSLockCurrent gslock(GS_LOCK_WAIT_FLUSH);
 
 	CellGcmContextData& ctx = (CellGcmContextData&)Memory[context_addr];
 	CellGcmControl& ctrl = (CellGcmControl&)Memory[gcm_info.control_addr];
@@ -24,6 +24,7 @@ int cellGcmCallback(u32 context_addr, u32 count)
 
 	ctx.current = ctx.begin + res;
 
+	//InterlockedExchange64((volatile long long*)((u8*)&ctrl + offsetof(CellGcmControl, put)), (u64)(u32)re(res));
 	ctrl.put = res;
 	ctrl.get = 0;
 	

rpcs3/Emu/SysCalls/lv2/SC_Interrupt.cpp

@@ -13,7 +13,7 @@ static SysCallBase sc_int("sys_interrupt");
 
 int sys_interrupt_tag_destroy(u32 intrtag)
 {
-	sc_int.Error("sys_interrupt_tag_destroy(intrtag=%d)", intrtag);
+	sc_int.Warning("sys_interrupt_tag_destroy(intrtag=%d)", intrtag);
 
 	u32 id = intrtag & 0xff;
 	u32 class_id = intrtag >> 8;
@@ -40,7 +40,7 @@ int sys_interrupt_tag_destroy(u32 intrtag)
 
 int sys_interrupt_thread_establish(mem32_t ih, u32 intrtag, u64 intrthread, u64 arg)
 {
-	sc_int.Error("sys_interrupt_thread_establish(ih_addr=0x%x, intrtag=%d, intrthread=%lld, arg=0x%llx)", ih.GetAddr(), intrtag, intrthread, arg);
+	sc_int.Warning("sys_interrupt_thread_establish(ih_addr=0x%x, intrtag=%d, intrthread=%lld, arg=0x%llx)", ih.GetAddr(), intrtag, intrthread, arg);
 
 	if (!ih.IsGood())
 	{