Merge cac372eb01 into 1b952bf173

src/core/cpu_patches.cpp

@@ -402,10 +402,7 @@ static const std::unordered_map<ZydisMnemonic, PatchInfo> Patches = {
     {ZYDIS_MNEMONIC_INSERTQ, {FilterNoSSE4a, GenerateINSERTQ, true}},
 
 #if defined(_WIN32)
-    // Windows needs a trampoline.
     {ZYDIS_MNEMONIC_MOV, {FilterTcbAccess, GenerateTcbAccess, true}},
-#elif !defined(__APPLE__)
-    {ZYDIS_MNEMONIC_MOV, {FilterTcbAccess, GenerateTcbAccess, false}},
 #endif
 };
 
@@ -674,24 +671,12 @@ static bool TryPatchJit(void* code_address) {
     return TryPatch(code, module).first;
 }
 
-static void TryPatchAot(void* code_address, u64 code_size) {
+#ifdef _WIN32
-    auto* code = static_cast<u8*>(code_address);
+// Used for patching TCB accesses on Windows.
-    auto* module = GetModule(code);
-    if (module == nullptr) {
-        return;
-    }
-
-    std::unique_lock lock{module->mutex};
-
-    const auto* end = code + code_size;
-    while (code < end) {
-        code += TryPatch(code, module).second;
-    }
-}
-
 static bool PatchesAccessViolationHandler(void* context, void* /* fault_address */) {
     return TryPatchJit(Common::GetRip(context));
 }
+#endif
 
 static bool PatchesIllegalInstructionHandler(void* context) {
     void* code_address = Common::GetRip(context);
@@ -719,7 +704,9 @@ static void PatchesInit() {
         auto* signals = Signals::Instance();
         // Should be called last.
         constexpr auto priority = std::numeric_limits<u32>::max();
+#ifdef _WIN32
         signals->RegisterAccessViolationHandler(PatchesAccessViolationHandler, priority);
+#endif
         signals->RegisterIllegalInstructionHandler(PatchesIllegalInstructionHandler, priority);
     }
 }
@@ -735,14 +722,4 @@ void RegisterPatchModule(void* module_ptr, u64 module_size, void* trampoline_are
                                           trampoline_area_size));
 }
 
-void PrePatchInstructions(u64 segment_addr, u64 segment_size) {
-#if !defined(_WIN32) && !defined(__APPLE__)
-    // Linux and others have an FS segment pointing to valid memory, so continue to do full
-    // ahead-of-time patching for now until a better solution is worked out.
-    if (!Patches.empty()) {
-        TryPatchAot(reinterpret_cast<void*>(segment_addr), segment_size);
-    }
-#endif
-}
-
 } // namespace Core

src/core/cpu_patches.h

@@ -11,7 +11,4 @@ namespace Core {
 void RegisterPatchModule(void* module_ptr, u64 module_size, void* trampoline_area_ptr,
                          u64 trampoline_area_size);
 
-/// Applies CPU patches that need to be done before beginning executions.
-void PrePatchInstructions(u64 segment_addr, u64 segment_size);
-
 } // namespace Core

src/core/module.cpp

@@ -163,11 +163,6 @@ void Module::LoadModuleToMemory(u32& max_tls_index) {
             LOG_INFO(Core_Linker, "segment_mode ..........: {}", segment_mode);
 
             add_segment(elf_pheader[i]);
-#ifdef ARCH_X86_64
-            if (elf_pheader[i].p_flags & PF_EXEC) {
-                PrePatchInstructions(segment_addr, segment_file_size);
-            }
-#endif
             break;
         }
         case PT_DYNAMIC:

src/core/tls.cpp

@@ -46,6 +46,8 @@ Tcb* GetTcbBase() {
     return reinterpret_cast<Tcb*>(TlsGetValue(GetTcbKey()));
 }
 
+void SwapTcb() {}
+
 #elif defined(__APPLE__) && defined(ARCH_X86_64)
 
 // Apple x86_64
@@ -151,6 +153,8 @@ Tcb* GetTcbBase() {
     return tcb;
 }
 
+void SwapTcb() {}
+
 #elif defined(ARCH_X86_64)
 
 // Other POSIX x86_64
@@ -165,6 +169,17 @@ Tcb* GetTcbBase() {
     return tcb;
 }
 
+void SwapTcb() {
+    void* fs;
+    void* gs;
+    asm volatile("rdfsbase %0" : "=r"(fs)::"memory");
+    asm volatile("rdgsbase %0" : "=r"(gs)::"memory");
+
+    // Swap FS and GS
+    asm volatile("wrfsbase %0" ::"r"(gs) : "memory");
+    asm volatile("wrgsbase %0" ::"r"(fs) : "memory");
+}
+
 #else
 
 // POSIX non-x86_64
@@ -191,6 +206,8 @@ Tcb* GetTcbBase() {
     return static_cast<Tcb*>(pthread_getspecific(GetTcbKey()));
 }
 
+void SwapTcb() {}
+
 #endif
 
 thread_local std::once_flag init_tls_flag;

src/core/tls.h

@@ -39,6 +39,9 @@ void SetTcbBase(void* image_address);
 /// Retrieves Tcb structure for the calling thread.
 Tcb* GetTcbBase();
 
+/// Swaps the TCB in or out of the fs register, if required by the platform.
+void SwapTcb();
+
 /// Makes sure TLS is initialized for the thread before entering guest.
 void EnsureThreadInitialized();
 
@@ -54,7 +57,17 @@ ReturnType ExecuteGuest(PS4_SYSV_ABI ReturnType (*func)(FuncArgs...), CallArgs&&
     EnsureThreadInitialized();
     // clear stack to avoid trash from EnsureThreadInitialized
     ClearStack<13_KB>();
-    return func(std::forward<CallArgs>(args)...);
+
+    if constexpr (std::is_same_v<ReturnType, void>) {
+        SwapTcb();
+        func(std::forward<CallArgs>(args)...);
+        SwapTcb();
+    } else {
+        SwapTcb();
+        auto ret = func(std::forward<CallArgs>(args)...);
+        SwapTcb();
+        return ret;
+    }
 }
 
 template <class F, F f>
@@ -63,7 +76,16 @@ struct HostCallWrapperImpl;
 template <class ReturnType, class... Args, PS4_SYSV_ABI ReturnType (*func)(Args...)>
 struct HostCallWrapperImpl<PS4_SYSV_ABI ReturnType (*)(Args...), func> {
     static ReturnType PS4_SYSV_ABI wrap(Args... args) {
-        return func(args...);
+        if constexpr (std::is_same_v<ReturnType, void>) {
+            SwapTcb();
+            func(args...);
+            SwapTcb();
+        } else {
+            SwapTcb();
+            auto ret = func(args...);
+            SwapTcb();
+            return ret;
+        }
     }
 };