Fixed TLS for PPU threads

rpcs3/Emu/ARMv7/Modules/sceLibc.cpp

@@ -18,6 +18,8 @@ namespace sce_libc_func
 	{
 		sceLibc.Error("__cxa_atexit(func=0x%x, arg=0x%x, dso=0x%x)", func, arg, dso);
 		
+		LV2_LOCK(0);
+
 		g_atexit.insert(g_atexit.begin(), [func, arg, dso](ARMv7Context& context)
 		{
 			func(context, arg);
@@ -28,6 +30,8 @@ namespace sce_libc_func
 	{
 		sceLibc.Error("__aeabi_atexit(arg=0x%x, func=0x%x, dso=0x%x)", arg, func, dso);
 
+		LV2_LOCK(0);
+
 		g_atexit.insert(g_atexit.begin(), [func, arg, dso](ARMv7Context& context)
 		{
 			func(context, arg);
@@ -38,6 +42,8 @@ namespace sce_libc_func
 	{
 		sceLibc.Error("exit()");
 		
+		LV2_LOCK(0);
+
 		for (auto func : g_atexit)
 		{
 			func(context);

rpcs3/Emu/Cell/PPUThread.cpp

@@ -15,6 +15,9 @@
 
 u64 rotate_mask[64][64];
 
+extern u32 ppu_get_tls(u32 thread);
+extern void ppu_free_tls(u32 thread);
+
 PPUThread& GetCurrentPPUThread()
 {
 	PPCThread* thread = GetCurrentPPCThread();
@@ -32,6 +35,7 @@ PPUThread::PPUThread() : PPCThread(CPU_THREAD_PPU)
 
 PPUThread::~PPUThread()
 {
+	ppu_free_tls(GetId());
 }
 
 void PPUThread::DoReset()
@@ -92,7 +96,7 @@ void PPUThread::InitRegs()
 	GPR[2] = rtoc;
 	//GPR[11] = entry;
 	//GPR[12] = Emu.GetMallocPageSize();
-	GPR[13] = Memory.PRXMem.GetStartAddr() + 0x7060;
+	GPR[13] = ppu_get_tls(GetId()) + 0x7000; // 0x7000 is usually subtracted from r13 to access first TLS element (details are not clear)
 
 	LR = Emu.GetCPUThreadExit();
 	CTR = PC;

rpcs3/Emu/SysCalls/Modules/sysPrxForUser.cpp

@@ -23,11 +23,62 @@
 
 Module *sysPrxForUser = nullptr;
 
+u32 g_tls_size; // size of every thread's storage
+u32 g_tls_start; // start of TLS memory area
+u32 g_tls_image_addr; // address of TLS initialization area
+u32 g_tls_image_size; // size of TLS initialization area
+
+const u32 TLS_MAX = 256;
+std::array<std::atomic<u32>, TLS_MAX> g_tls_owners;
+
 void sys_initialize_tls()
 {
 	sysPrxForUser->Log("sys_initialize_tls()");
 }
 
+u32 ppu_get_tls(u32 thread)
+{
+	if (!g_tls_start)
+	{
+		g_tls_size = vm::cast(Emu.GetTLSMemsz(), "Emu.GetTLSMemsz"); // (not an address for vm::cast, but fine)
+		g_tls_start = vm::cast(Memory.Alloc(g_tls_size * TLS_MAX, 4096)); // memory for up to TLS_MAX threads
+		g_tls_image_addr = vm::cast(Emu.GetTLSAddr(), "Emu.GetTLSAddr");
+		g_tls_image_size = vm::cast(Emu.GetTLSFilesz(), "Emu.GetTLSFilesz");
+
+		sysPrxForUser->Warning("TLS initialized (g_tls_size=0x%x, g_tls_start=0x%x, g_tls_image_addr=0x%x, g_tls_image_size=0x%x)", g_tls_size, g_tls_start, g_tls_image_addr, g_tls_image_size);
+	}
+	
+	for (auto& v : g_tls_owners)
+	{
+		u32 old = 0;
+		if (v.compare_exchange_strong(old, thread))
+		{
+			const u32 addr = g_tls_start + (&v - g_tls_owners.data()) * g_tls_size;     // get TLS address
+			memset(vm::get_ptr(addr), 0, g_tls_size);                                   // fill TLS area with zeros
+			memcpy(vm::get_ptr(addr), vm::get_ptr(g_tls_image_addr), g_tls_image_size); // initialize from TLS image
+			return addr;
+		}
+		else if (old == thread)
+		{
+			return g_tls_start + (&v - g_tls_owners.data()) * g_tls_size; // if already initialized, return TLS address again
+		}
+	}
+
+	throw "Out of TLS memory";
+}
+
+void ppu_free_tls(u32 thread)
+{
+	for (auto& v : g_tls_owners)
+	{
+		u32 old = thread;
+		if (v.compare_exchange_strong(old, 0))
+		{
+			return;
+		}
+	}
+}
+
 int _sys_heap_create_heap(const u32 heap_addr, const u32 align, const u32 size)
 {
 	sysPrxForUser->Warning("_sys_heap_create_heap(heap_addr=0x%x, align=0x%x, size=0x%x)", heap_addr, align, size);
@@ -365,6 +416,15 @@ void sysPrxForUser_init(Module *pxThis)
 {
 	sysPrxForUser = pxThis;
 
+	g_tls_size = 0;
+	g_tls_start = 0;
+	g_tls_image_addr = 0;
+	g_tls_image_size = 0;
+	for (auto& v : g_tls_owners)
+	{
+		v.store(0, std::memory_order_relaxed);
+	}
+
 	// Setup random number generator
 	srand(time(NULL));
 

rpcs3/Loader/ELF64.cpp

@@ -356,7 +356,7 @@ namespace loader
 			ppu_thr_stop_data[1] = BLR();
 			Emu.SetCPUThreadStop(ppu_thr_stop_data.addr());
 
-			vm::write64(Memory.PRXMem.AllocAlign(0x10000), 0xDEADBEEFABADCAFE);
+			//vm::write64(Memory.PRXMem.AllocAlign(0x10000), 0xDEADBEEFABADCAFE);
 			/*
 			//TODO
 			static const int branch_size = 6 * 4;
@@ -423,7 +423,6 @@ namespace loader
 
 				case 0x00000007: //TLS
 					Emu.SetTLSData(phdr.p_vaddr.addr(), phdr.p_filesz.value(), phdr.p_memsz.value());
-					LOG_ERROR(LOADER, "TLS: addr=0x%x, filesz=0x%x, memsz=0x%x", Emu.GetTLSAddr(), Emu.GetTLSFilesz(), Emu.GetTLSMemsz());
 					break;
 
 				case 0x60000001: //LOOS+1