agc: reserve driver backing before guest direct-memory allocation

src/libs/agc.cpp

@@ -57,7 +57,6 @@ struct DriverDmem {
 };
 
 static DriverDmem g_driver_dmem;
-static int InitializeDmem();
 static void ReleaseDmem();
 
 } // namespace Gen5Driver
@@ -82,6 +81,8 @@ void Initialize() {
 	auto& video_out = VideoOut::VideoOutInit(width, height, presenter);
 	g_renderer      = &presenter.Renderer();
 	g_renderer->InitializeGpu(&video_out);
+	// Native driver module startup reserves its DMEM before the guest can allocate it.
+	EXIT_IF(Gen5Driver::Initialize() != OK);
 	ShaderInit();
 }
 
@@ -356,38 +357,6 @@ int KYTY_SYSV_ABI AgcInit(uint32_t* state, uint32_t ver) {
 	}
 
 	printf("version = %u\n", ver);
-	std::scoped_lock lock {Gen5Driver::g_driver_dmem.mutex};
-	if (Gen5Driver::g_driver_dmem.physical_offset >= 0) {
-		return OK;
-	}
-	if (const auto result = Gen5Driver::InitializeDmem(); result != OK) {
-		return result;
-	}
-	struct FunctionShaderBindings {
-		uint64_t code;
-		uint32_t user_data[2];
-	};
-	static_assert(sizeof(FunctionShaderBindings) == 16);
-	// Ordered-count queue validator, including its return to the caller.
-	static constexpr std::array<uint32_t, 16> validator {
-	    0xbeeb03ff, 0x00000021, 0xb96a1818, 0xbf06106a,
-	    0xbf800000, 0x8590807e, 0xb96a0a18, 0xbf066a80,
-	    0xbf800000, 0x85ea807e, 0x88ea106a, 0xbf870003,
-	    0xbefc03ff, 0x8a6ca000, 0xbf920009, 0xbefd210e,
-	};
-	const auto base = Gen5Driver::DriverDmemBase + Gen5Driver::AgcDmemOffset;
-	const auto function_table = base + 0x60;
-	const auto code_address = base + 0x100; // Shader code requires 256-byte alignment.
-	std::array<FunctionShaderBindings, 8> functions {};
-	functions[2].code = code_address;
-	const std::array<uint32_t, 4> descriptor {
-	    static_cast<uint32_t>(function_table),
-	    static_cast<uint32_t>(function_table >> 32u) | 0x00100000u,
-	    static_cast<uint32_t>(functions.size()), 0x5204u,
-	};
-	std::memcpy(reinterpret_cast<void*>(function_table), functions.data(), sizeof(functions));
-	std::memcpy(reinterpret_cast<void*>(code_address), validator.data(), sizeof(validator));
-	std::memcpy(reinterpret_cast<void*>(base), descriptor.data(), sizeof(descriptor));
 
 	return OK;
 }
@@ -4275,7 +4244,11 @@ namespace Gen5Driver {
 
 LIB_NAME("Graphics5Driver", "Graphics5Driver");
 
-static int InitializeDmem() {
+int Initialize() {
+	std::scoped_lock lock {g_driver_dmem.mutex};
+	if (g_driver_dmem.physical_offset >= 0) {
+		return OK;
+	}
 	int64_t physical_offset = -1;
 	auto result = LibKernel::Memory::KernelAllocateDirectMemory(
 	    0, LibKernel::Memory::KernelGetDirectMemorySize(), DriverDmemSize, DriverDmemSize,
@@ -4292,6 +4265,31 @@ static int InitializeDmem() {
 	}
 	EXIT_IF(reinterpret_cast<uint64_t>(address) != DriverDmemBase);
 	std::memset(address, 0, DriverDmemSize);
+	struct FunctionShaderBindings {
+		uint64_t code;
+		uint32_t user_data[2];
+	};
+	static_assert(sizeof(FunctionShaderBindings) == 16);
+	// Ordered-count queue validator, including its return to the caller.
+	static constexpr std::array<uint32_t, 16> validator {
+	    0xbeeb03ff, 0x00000021, 0xb96a1818, 0xbf06106a,
+	    0xbf800000, 0x8590807e, 0xb96a0a18, 0xbf066a80,
+	    0xbf800000, 0x85ea807e, 0x88ea106a, 0xbf870003,
+	    0xbefc03ff, 0x8a6ca000, 0xbf920009, 0xbefd210e,
+	};
+	const auto base = DriverDmemBase + AgcDmemOffset;
+	const auto function_table = base + 0x60;
+	const auto code_address = base + 0x100; // Shader code requires 256-byte alignment.
+	std::array<FunctionShaderBindings, 8> functions {};
+	functions[2].code = code_address;
+	const std::array<uint32_t, 4> descriptor {
+	    static_cast<uint32_t>(function_table),
+	    static_cast<uint32_t>(function_table >> 32u) | 0x00100000u,
+	    static_cast<uint32_t>(functions.size()), 0x5204u,
+	};
+	std::memcpy(reinterpret_cast<void*>(function_table), functions.data(), sizeof(functions));
+	std::memcpy(reinterpret_cast<void*>(code_address), validator.data(), sizeof(validator));
+	std::memcpy(reinterpret_cast<void*>(base), descriptor.data(), sizeof(descriptor));
 	g_driver_dmem.physical_offset = physical_offset;
 	return OK;
 }

src/libs/agc.h

@@ -307,6 +307,8 @@ uint32_t* KYTY_SYSV_ABI AgcDcbSetFlip(CommandBuffer* buf, uint32_t video_out_han
 
 namespace Gen5Driver {
 
+int Initialize();
+
 struct Packet {
 	uint32_t* addr;
 	uint32_t  dw_num;

tests/ShaderRecompilerComputeTests.cpp

@@ -41903,9 +41903,24 @@ void CheckAgcSystemTable(RenderContext &renderer) {
       0xbf800000, 0x85ea807e, 0x88ea106a, 0xbf870003,
       0xbefc03ff, 0x8a6ca000, 0xbf920009, 0xbefd210e};
   LibKernel::Memory::InstallGpuResources(&renderer);
+  Require(name, "driver module startup", Gen5Driver::Initialize() == 0,
+          "driver module startup failed before guest initialization");
+  const auto direct_memory_size = LibKernel::Memory::KernelGetDirectMemorySize();
+  int64_t available_offset = -1;
+  size_t available_size = 0;
+  Require(name, "remaining direct pool",
+          LibKernel::Memory::KernelAvailableDirectMemorySize(
+              0, direct_memory_size, 0, &available_offset, &available_size) == 0 &&
+              available_size == direct_memory_size - 0x200000,
+          "driver startup did not reserve its backing in the reported direct pool");
+  int64_t guest_allocation = -1;
+  Require(name, "guest consumes remaining direct pool",
+          LibKernel::Memory::KernelAllocateDirectMemory(
+              0, direct_memory_size, available_size, 0x4000, 0, &guest_allocation) == 0,
+          "guest could not allocate all remaining direct memory");
   uint32_t state = 0xabcdef01u;
   Require(name, "driver startup", Gen5::AgcInit(&state, 13) == 0 && state == 0xabcdef01u,
-          "Agc startup failed or changed its ignored public state pointer");
+          "Agc startup required free guest memory or changed its ignored public state pointer");
   LibKernel::Memory::VirtualQueryInfo mapping{};
   Require(name, "shared driver backing",
           LibKernel::Memory::KernelVirtualQuery(reinterpret_cast<void *>(driver_base), 0,
@@ -41932,10 +41947,14 @@ void CheckAgcSystemTable(RenderContext &renderer) {
   Require(name, "default instrumentation", *instrumentation == 0,
           "unused vertex validation was enabled at startup");
   *instrumentation = 0x12340000u;
-  Require(name, "repeat startup", Gen5::AgcInit(nullptr, 13) == 0 &&
+  Require(name, "repeat startup", Gen5Driver::Initialize() == 0 &&
+              Gen5::AgcInit(nullptr, 13) == 0 &&
               *instrumentation == 0x12340000u,
           "repeat Agc initialization reset existing system state");
   *instrumentation = 0;
+  Require(name, "release guest allocation",
+          LibKernel::Memory::KernelCheckedReleaseDirectMemory(guest_allocation, available_size) == 0,
+          "guest allocation release failed");
   LibKernel::Memory::InstallGpuResources(nullptr);
   std::printf("[host]    %-32s ok\n", name);
 }