TLS for ARMv7 threads

rpcs3/Emu/ARMv7/ARMv7Context.h

@@ -111,6 +111,8 @@ struct ARMv7Context
 
 	} ITSTATE;
 
+	u32 TLS;
+
 	u32 R_ADDR;
 	u64 R_DATA;
 

rpcs3/Emu/ARMv7/ARMv7Interpreter.cpp

@@ -336,15 +336,17 @@ void ARMv7_instrs::MRC_(ARMv7Context& context, const ARMv7Code code, const ARMv7
 
 	if (ConditionPassed(context, cond))
 	{
-		if (cp == 15 && opc1 == 0 && cn == 13 && cm == 0 && opc2 == 3)
+		// APSR flags are written if t = 15
-		{
-			LOG_ERROR(ARMv7, "TODO: TLS requested");
 
-			if (t < 15)
+		if (t < 15 && cp == 15 && opc1 == 0 && cn == 13 && cm == 0 && opc2 == 3)
+		{
+			if (!context.TLS)
 			{
-				context.GPR[t] = 0;
+				throw "TLS not initialized";
-				return;
 			}
+
+			context.GPR[t] = context.TLS;
+			return;
 		}
 
 		throw fmt::format("Bad instruction: mrc p%d,%d,r%d,c%d,c%d,%d", cp, opc1, t, cn, cm, opc2);

rpcs3/Emu/ARMv7/ARMv7Thread.cpp

@@ -30,6 +30,69 @@ void ARMv7Context::fast_call(u32 addr)
 	return thread.FastCall(addr);
 }
 
+#define TLS_MAX 128
+
+u32 g_armv7_tls_start;
+
+std::array<std::atomic<u32>, TLS_MAX> g_armv7_tls_owners;
+
+void armv7_init_tls()
+{
+	g_armv7_tls_start = Emu.GetTLSMemsz() ? vm::cast(Memory.PSV.RAM.AllocAlign(Emu.GetTLSMemsz() * TLS_MAX, 4096)) : 0;
+
+	for (auto& v : g_armv7_tls_owners)
+	{
+		v.store(0, std::memory_order_relaxed);
+	}
+}
+
+u32 armv7_get_tls(u32 thread)
+{
+	if (!Emu.GetTLSMemsz())
+	{
+		return 0;
+	}
+
+	for (u32 i = 0; i < TLS_MAX; i++)
+	{
+		if (g_armv7_tls_owners[i] == thread)
+		{
+			return g_armv7_tls_start + i * Emu.GetTLSMemsz(); // if already initialized, return TLS address
+		}
+	}
+
+	for (u32 i = 0; i < TLS_MAX; i++)
+	{
+		u32 old = 0;
+		if (g_armv7_tls_owners[i].compare_exchange_strong(old, thread))
+		{
+			const u32 addr = g_armv7_tls_start + i * Emu.GetTLSMemsz(); // get TLS address
+			memset(vm::get_ptr(addr), 0, Emu.GetTLSMemsz()); // fill TLS area with zeros
+			memcpy(vm::get_ptr(addr), vm::get_ptr(Emu.GetTLSAddr()), Emu.GetTLSFilesz()); // initialize from TLS image
+			return addr;
+		}
+	}
+
+	throw "Out of TLS memory";
+}
+
+void armv7_free_tls(u32 thread)
+{
+	if (!Emu.GetTLSMemsz())
+	{
+		return;
+	}
+
+	for (auto& v : g_armv7_tls_owners)
+	{
+		u32 old = thread;
+		if (v.compare_exchange_strong(old, 0))
+		{
+			return;
+		}
+	}
+}
+
 ARMv7Thread::ARMv7Thread()
 	: CPUThread(CPU_THREAD_ARMv7)
 	, context(*this)
@@ -39,6 +102,11 @@ ARMv7Thread::ARMv7Thread()
 {
 }
 
+ARMv7Thread::~ARMv7Thread()
+{
+	armv7_free_tls(GetId());
+}
+
 void ARMv7Thread::InitRegs()
 {
 	memset(context.GPR, 0, sizeof(context.GPR[0]) * 15);
@@ -47,6 +115,7 @@ void ARMv7Thread::InitRegs()
 	context.ISET = Thumb;
 	context.ITSTATE.IT = 0;
 	context.SP = m_stack_addr + m_stack_size;
+	context.TLS = armv7_get_tls(GetId());
 }
 
 void ARMv7Thread::InitStack()

rpcs3/Emu/ARMv7/ARMv7Thread.h

@@ -12,6 +12,7 @@ public:
 	//const char* m_last_instr_name;
 
 	ARMv7Thread();
+	~ARMv7Thread();
 
 	//void update_code(const u32 address)
 	//{

rpcs3/Emu/Cell/PPUThread.cpp

@@ -55,7 +55,7 @@ void PPUThread::DoReset()
 	XER.XER     = 0;
 	FPSCR.FPSCR = 0;
 	VSCR.VSCR   = 0;
-        VRSAVE      = 0;
+	VRSAVE      = 0;
 
 	cycle = 0;
 }
@@ -65,33 +65,8 @@ void PPUThread::InitRegs()
 	const u32 pc = entry ? vm::read32(entry) : 0;
 	const u32 rtoc = entry ? vm::read32(entry + 4) : 0;
 
-	//ConLog.Write("entry = 0x%x", entry);
-	//ConLog.Write("rtoc = 0x%x", rtoc);
-
 	SetPc(pc);
 
-	/*
-	const s32 thread_num = Emu.GetCPU().GetThreadNumById(GetType(), GetId());
-
-	if(thread_num < 0)
-	{
-		LOG_ERROR(PPU, "GetThreadNumById failed.");
-		Emu.Pause();
-		return;
-	}
-	*/
-
-	/*
-	const s32 tls_size = Emu.GetTLSFilesz() * thread_num;
-
-	if(tls_size >= Emu.GetTLSMemsz())
-	{
-		LOG_ERROR(PPU, "Out of TLS memory.");
-		Emu.Pause();
-		return;
-	}
-	*/
-
 	GPR[1] = align(m_stack_addr + m_stack_size, 0x200) - 0x200;
 	GPR[2] = rtoc;
 	//GPR[11] = entry;

rpcs3/Emu/SysCalls/Modules/sysPrxForUser.cpp

@@ -23,12 +23,10 @@
 
 Module *sysPrxForUser = nullptr;
 
-u32 g_tls_size; // size of every thread's storage
+#define TLS_MAX 128
-u32 g_tls_start; // start of TLS memory area
+
-u32 g_tls_image_addr; // address of TLS initialization area
+u32 g_tls_start; // start of TLS memory 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()
@@ -40,19 +38,16 @@ 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.MainMem.AllocAlign(Emu.GetTLSMemsz() * TLS_MAX, 4096)); // memory for up to TLS_MAX threads
-		g_tls_start = vm::cast(Memory.Alloc(g_tls_size * TLS_MAX, 4096)); // memory for up to TLS_MAX threads
+		sysPrxForUser->Notice("Thread Local Storage initialized (g_tls_start=0x%x, size = 0x%x)\n*** TLS segment addr: 0x%08x\n*** TLS segment size: 0x%08x",
-		g_tls_image_addr = vm::cast(Emu.GetTLSAddr(), "Emu.GetTLSAddr");
+			g_tls_start, Emu.GetTLSMemsz(), Emu.GetTLSAddr(), Emu.GetTLSFilesz());
-		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 (u32 i = 0; i < TLS_MAX; i++)
 	{
 		if (g_tls_owners[i] == thread)
 		{
-			return g_tls_start + i * g_tls_size; // if already initialized, return TLS address
+			return g_tls_start + i * Emu.GetTLSMemsz(); // if already initialized, return TLS address
 		}
 	}
 
@@ -61,9 +56,9 @@ u32 ppu_get_tls(u32 thread)
 		u32 old = 0;
 		if (g_tls_owners[i].compare_exchange_strong(old, thread))
 		{
-			const u32 addr = g_tls_start + i * g_tls_size; // get TLS address
+			const u32 addr = g_tls_start + i * Emu.GetTLSMemsz(); // get TLS address
-			memset(vm::get_ptr(addr), 0, g_tls_size);      // fill TLS area with zeros
+			memset(vm::get_ptr(addr), 0, Emu.GetTLSMemsz()); // 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
+			memcpy(vm::get_ptr(addr), vm::get_ptr(Emu.GetTLSAddr()), Emu.GetTLSFilesz()); // initialize from TLS image
 			return addr;
 		}
 	}
@@ -420,10 +415,7 @@ 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);

rpcs3/Emu/System.h

@@ -28,9 +28,9 @@ struct VFS;
 struct EmuInfo
 {
 private:
-	u64 tls_addr;
+	u32 tls_addr;
-	u64 tls_filesz;
+	u32 tls_filesz;
-	u64 tls_memsz;
+	u32 tls_memsz;
 
 	sys_process_param_info proc_param;
 
@@ -50,16 +50,16 @@ public:
 		proc_param.primary_prio = be_t<s32>::make(0x50);
 	}
 
-	void SetTLSData(const u64 addr, const u64 filesz, const u64 memsz)
+	void SetTLSData(u32 addr, u32 filesz, u32 memsz)
 	{
 		tls_addr = addr;
 		tls_filesz = filesz;
 		tls_memsz = memsz;
 	}
 
-	u64 GetTLSAddr() const { return tls_addr; }
+	u32 GetTLSAddr() const { return tls_addr; }
-	u64 GetTLSFilesz() const { return tls_filesz; }
+	u32 GetTLSFilesz() const { return tls_filesz; }
-	u64 GetTLSMemsz() const { return tls_memsz; }
+	u32 GetTLSMemsz() const { return tls_memsz; }
 };
 
 class ModuleInitializer
@@ -173,7 +173,7 @@ public:
 		m_modules_init.push_back(std::move(m));
 	}
 
-	void SetTLSData(const u64 addr, const u64 filesz, const u64 memsz)
+	void SetTLSData(u32 addr, u32 filesz, u32 memsz)
 	{
 		m_info.SetTLSData(addr, filesz, memsz);
 	}
@@ -195,9 +195,9 @@ public:
 
 	EmuInfo& GetInfo() { return m_info; }
 
-	u64 GetTLSAddr() const { return m_info.GetTLSAddr(); }
+	u32 GetTLSAddr() const { return m_info.GetTLSAddr(); }
-	u64 GetTLSFilesz() const { return m_info.GetTLSFilesz(); }
+	u32 GetTLSFilesz() const { return m_info.GetTLSFilesz(); }
-	u64 GetTLSMemsz() const { return m_info.GetTLSMemsz(); }
+	u32 GetTLSMemsz() const { return m_info.GetTLSMemsz(); }
 
 	u32 GetMallocPageSize() { return m_info.GetProcParam().malloc_pagesize; }
 

rpcs3/Loader/ELF32.cpp

@@ -11,6 +11,8 @@
 #include "Emu/ARMv7/PSVFuncList.h"
 #include "Emu/System.h"
 
+extern void armv7_init_tls();
+
 namespace loader
 {
 	namespace handlers
@@ -235,9 +237,13 @@ namespace loader
 					}
 					else if (!strcmp(name.c_str(), ".tbss"))
 					{
-						LOG_NOTICE(LOADER, ".tbss analysis");
+						LOG_NOTICE(LOADER, ".tbss analysis...");
+						const u32 img_addr = shdr.data_le.sh_addr;                  // start address of TLS initialization image
+						const u32 img_size = (&shdr)[1].data_le.sh_addr - img_addr; // calculate its size as the difference between sections
+						const u32 tls_size = shdr.data_le.sh_size;                  // full size of TLS
 
-						LOG_ERROR(LOADER, "TLS: size=0x%08x", shdr.data_le.sh_size);
+						LOG_WARNING(LOADER, "TLS: img_addr=0x%08x, img_size=0x%x, tls_size=0x%x", img_addr, img_size, tls_size);
+						Emu.SetTLSData(img_addr, img_size, tls_size);
 					}
 					else if (!strcmp(name.c_str(), ".sceRefs.rodata"))
 					{
@@ -314,6 +320,7 @@ namespace loader
 					}
 				}
 
+				armv7_init_tls();
 				armv7_decoder_initialize(code_start, code_end);
 
 				arm7_thread(entry & ~1 /* TODO: Thumb/ARM encoding selection */, "main_thread").args({ Emu.GetPath()/*, "-emu"*/ }).run();

rpcs3/Loader/ELF64.cpp

@@ -422,7 +422,10 @@ namespace loader
 					break;
 
 				case 0x00000007: //TLS
-					Emu.SetTLSData(phdr.p_vaddr.addr(), phdr.p_filesz.value(), phdr.p_memsz.value());
+					Emu.SetTLSData(
+						vm::cast(phdr.p_vaddr.addr(), "TLS: phdr.p_vaddr"),
+						vm::cast(phdr.p_filesz.value(), "TLS: phdr.p_filesz"),
+						vm::cast(phdr.p_memsz.value(), "TLS: phdr.p_memsz"));
 					break;
 
 				case 0x60000001: //LOOS+1