Merge pull request #12074 from GPUCode/yuwu-on-the-metal

Implement Native Code Execution (NCE)

9 files changed, 109 insertions, 21 deletions

diff --git a/src/core/hle/kernel/code_set.h b/src/core/hle/kernel/code_set.h
index af1af2b78..4d2d0098e 100644
--- a/src/core/hle/kernel/code_set.h
+++ b/src/core/hle/kernel/code_set.h
@@ -75,12 +75,26 @@ struct CodeSet final {
         return segments[2];
     }
 
+#ifdef HAS_NCE
+    Segment& PatchSegment() {
+        return patch_segment;
+    }
+
+    const Segment& PatchSegment() const {
+        return patch_segment;
+    }
+#endif
+
     /// The overall data that backs this code set.
     Kernel::PhysicalMemory memory;
 
     /// The segments that comprise this code set.
     std::array<Segment, 3> segments;
 
+#ifdef HAS_NCE
+    Segment patch_segment;
+#endif
+
     /// The entry point address for this code set.
     KProcessAddress entrypoint = 0;
 };
diff --git a/src/core/hle/kernel/k_address_space_info.cpp b/src/core/hle/kernel/k_address_space_info.cpp
index 32173e52b..23258071e 100644
--- a/src/core/hle/kernel/k_address_space_info.cpp
+++ b/src/core/hle/kernel/k_address_space_info.cpp
@@ -25,8 +25,8 @@ constexpr std::array<KAddressSpaceInfo, 13> AddressSpaceInfos{{
    { .bit_width = 36, .address = 2_GiB       , .size = 64_GiB  - 2_GiB  , .type = KAddressSpaceInfo::Type::MapLarge, },
    { .bit_width = 36, .address = Size_Invalid, .size = 8_GiB            , .type = KAddressSpaceInfo::Type::Heap,     },
    { .bit_width = 36, .address = Size_Invalid, .size = 6_GiB            , .type = KAddressSpaceInfo::Type::Alias,    },
-#ifdef ANDROID
-   // With Android, we use a 38-bit address space due to memory limitations. This should (safely) truncate ASLR region.
+#ifdef HAS_NCE
+   // With NCE, we use a 38-bit address space due to memory limitations. This should (safely) truncate ASLR region.
    { .bit_width = 39, .address = 128_MiB     , .size = 256_GiB - 128_MiB, .type = KAddressSpaceInfo::Type::Map39Bit, },
 #else
    { .bit_width = 39, .address = 128_MiB     , .size = 512_GiB - 128_MiB, .type = KAddressSpaceInfo::Type::Map39Bit, },
diff --git a/src/core/hle/kernel/k_page_table_base.cpp b/src/core/hle/kernel/k_page_table_base.cpp
index 47dc8fd35..6691586ed 100644
--- a/src/core/hle/kernel/k_page_table_base.cpp
+++ b/src/core/hle/kernel/k_page_table_base.cpp
@@ -88,6 +88,22 @@ Result FlushDataCache(AddressType addr, u64 size) {
     R_SUCCEED();
 }
 
+constexpr Common::MemoryPermission ConvertToMemoryPermission(KMemoryPermission perm) {
+    Common::MemoryPermission perms{};
+    if (True(perm & KMemoryPermission::UserRead)) {
+        perms |= Common::MemoryPermission::Read;
+    }
+    if (True(perm & KMemoryPermission::UserWrite)) {
+        perms |= Common::MemoryPermission::Write;
+    }
+#ifdef HAS_NCE
+    if (True(perm & KMemoryPermission::UserExecute)) {
+        perms |= Common::MemoryPermission::Execute;
+    }
+#endif
+    return perms;
+}
+
 } // namespace
 
 void KPageTableBase::MemoryRange::Open() {
@@ -170,7 +186,8 @@ Result KPageTableBase::InitializeForProcess(Svc::CreateProcessFlag as_type, bool
                                             KMemoryManager::Pool pool, KProcessAddress code_address,
                                             size_t code_size, KSystemResource* system_resource,
                                             KResourceLimit* resource_limit,
-                                            Core::Memory::Memory& memory) {
+                                            Core::Memory::Memory& memory,
+                                            KProcessAddress aslr_space_start) {
     // Calculate region extents.
     const size_t as_width = GetAddressSpaceWidth(as_type);
     const KProcessAddress start = 0;
@@ -211,7 +228,8 @@ Result KPageTableBase::InitializeForProcess(Svc::CreateProcessFlag as_type, bool
         heap_region_size = GetSpaceSize(KAddressSpaceInfo::Type::Heap);
         stack_region_size = GetSpaceSize(KAddressSpaceInfo::Type::Stack);
         kernel_map_region_size = GetSpaceSize(KAddressSpaceInfo::Type::MapSmall);
-        m_code_region_start = GetSpaceStart(KAddressSpaceInfo::Type::Map39Bit);
+        m_code_region_start = m_address_space_start + aslr_space_start +
+                              GetSpaceStart(KAddressSpaceInfo::Type::Map39Bit);
         m_code_region_end = m_code_region_start + GetSpaceSize(KAddressSpaceInfo::Type::Map39Bit);
         m_alias_code_region_start = m_code_region_start;
         m_alias_code_region_end = m_code_region_end;
@@ -5643,7 +5661,8 @@ Result KPageTableBase::Operate(PageLinkedList* page_list, KProcessAddress virt_a
     case OperationType::Map: {
         ASSERT(virt_addr != 0);
         ASSERT(Common::IsAligned(GetInteger(virt_addr), PageSize));
-        m_memory->MapMemoryRegion(*m_impl, virt_addr, num_pages * PageSize, phys_addr);
+        m_memory->MapMemoryRegion(*m_impl, virt_addr, num_pages * PageSize, phys_addr,
+                                  ConvertToMemoryPermission(properties.perm));
 
         // Open references to pages, if we should.
         if (this->IsHeapPhysicalAddress(phys_addr)) {
@@ -5658,8 +5677,11 @@ Result KPageTableBase::Operate(PageLinkedList* page_list, KProcessAddress virt_a
     }
     case OperationType::ChangePermissions:
     case OperationType::ChangePermissionsAndRefresh:
-    case OperationType::ChangePermissionsAndRefreshAndFlush:
+    case OperationType::ChangePermissionsAndRefreshAndFlush: {
+        m_memory->ProtectRegion(*m_impl, virt_addr, num_pages * PageSize,
+                                ConvertToMemoryPermission(properties.perm));
         R_SUCCEED();
+    }
     default:
         UNREACHABLE();
     }
@@ -5687,7 +5709,8 @@ Result KPageTableBase::Operate(PageLinkedList* page_list, KProcessAddress virt_a
             const size_t size{node.GetNumPages() * PageSize};
 
             // Map the pages.
-            m_memory->MapMemoryRegion(*m_impl, virt_addr, size, node.GetAddress());
+            m_memory->MapMemoryRegion(*m_impl, virt_addr, size, node.GetAddress(),
+                                      ConvertToMemoryPermission(properties.perm));
 
             virt_addr += size;
         }
diff --git a/src/core/hle/kernel/k_page_table_base.h b/src/core/hle/kernel/k_page_table_base.h
index ee2c41e67..556d230b3 100644
--- a/src/core/hle/kernel/k_page_table_base.h
+++ b/src/core/hle/kernel/k_page_table_base.h
@@ -235,7 +235,8 @@ public:
                                 bool enable_device_address_space_merge, bool from_back,
                                 KMemoryManager::Pool pool, KProcessAddress code_address,
                                 size_t code_size, KSystemResource* system_resource,
-                                KResourceLimit* resource_limit, Core::Memory::Memory& memory);
+                                KResourceLimit* resource_limit, Core::Memory::Memory& memory,
+                                KProcessAddress aslr_space_start);
 
     void Finalize();
 
diff --git a/src/core/hle/kernel/k_process.cpp b/src/core/hle/kernel/k_process.cpp
index 3cfb414e5..6c29eb72c 100644
--- a/src/core/hle/kernel/k_process.cpp
+++ b/src/core/hle/kernel/k_process.cpp
@@ -300,7 +300,7 @@ Result KProcess::Initialize(const Svc::CreateProcessParameter& params, const KPa
             False(params.flags & Svc::CreateProcessFlag::DisableDeviceAddressSpaceMerge);
         R_TRY(m_page_table.Initialize(as_type, enable_aslr, enable_das_merge, !enable_aslr, pool,
                                       params.code_address, params.code_num_pages * PageSize,
-                                      m_system_resource, res_limit, this->GetMemory()));
+                                      m_system_resource, res_limit, this->GetMemory(), 0));
     }
     ON_RESULT_FAILURE_2 {
         m_page_table.Finalize();
@@ -332,7 +332,7 @@ Result KProcess::Initialize(const Svc::CreateProcessParameter& params, const KPa
 
 Result KProcess::Initialize(const Svc::CreateProcessParameter& params,
                             std::span<const u32> user_caps, KResourceLimit* res_limit,
-                            KMemoryManager::Pool pool) {
+                            KMemoryManager::Pool pool, KProcessAddress aslr_space_start) {
     ASSERT(res_limit != nullptr);
 
     // Set members.
@@ -393,7 +393,7 @@ Result KProcess::Initialize(const Svc::CreateProcessParameter& params,
             False(params.flags & Svc::CreateProcessFlag::DisableDeviceAddressSpaceMerge);
         R_TRY(m_page_table.Initialize(as_type, enable_aslr, enable_das_merge, !enable_aslr, pool,
                                       params.code_address, code_size, m_system_resource, res_limit,
-                                      this->GetMemory()));
+                                      this->GetMemory(), aslr_space_start));
     }
     ON_RESULT_FAILURE_2 {
         m_page_table.Finalize();
@@ -1128,7 +1128,7 @@ KProcess::KProcess(KernelCore& kernel)
 KProcess::~KProcess() = default;
 
 Result KProcess::LoadFromMetadata(const FileSys::ProgramMetadata& metadata, std::size_t code_size,
-                                  bool is_hbl) {
+                                  KProcessAddress aslr_space_start, bool is_hbl) {
     // Create a resource limit for the process.
     const auto physical_memory_size =
         m_kernel.MemoryManager().GetSize(Kernel::KMemoryManager::Pool::Application);
@@ -1179,7 +1179,7 @@ Result KProcess::LoadFromMetadata(const FileSys::ProgramMetadata& metadata, std:
         .name = {},
         .version = {},
         .program_id = metadata.GetTitleID(),
-        .code_address = code_address,
+        .code_address = code_address + GetInteger(aslr_space_start),
         .code_num_pages = static_cast<s32>(code_size / PageSize),
         .flags = flag,
         .reslimit = Svc::InvalidHandle,
@@ -1193,7 +1193,7 @@ Result KProcess::LoadFromMetadata(const FileSys::ProgramMetadata& metadata, std:
 
     // Initialize for application process.
     R_TRY(this->Initialize(params, metadata.GetKernelCapabilities(), res_limit,
-                           KMemoryManager::Pool::Application));
+                           KMemoryManager::Pool::Application, aslr_space_start));
 
     // Assign remaining properties.
     m_is_hbl = is_hbl;
@@ -1214,6 +1214,17 @@ void KProcess::LoadModule(CodeSet code_set, KProcessAddress base_addr) {
     ReprotectSegment(code_set.CodeSegment(), Svc::MemoryPermission::ReadExecute);
     ReprotectSegment(code_set.RODataSegment(), Svc::MemoryPermission::Read);
     ReprotectSegment(code_set.DataSegment(), Svc::MemoryPermission::ReadWrite);
+
+#ifdef HAS_NCE
+    if (Settings::IsNceEnabled()) {
+        auto& buffer = m_kernel.System().DeviceMemory().buffer;
+        const auto& code = code_set.CodeSegment();
+        const auto& patch = code_set.PatchSegment();
+        buffer.Protect(GetInteger(base_addr + code.addr), code.size, true, true, true);
+        buffer.Protect(GetInteger(base_addr + patch.addr), patch.size, true, true, true);
+        ReprotectSegment(code_set.PatchSegment(), Svc::MemoryPermission::None);
+    }
+#endif
 }
 
 bool KProcess::InsertWatchpoint(KProcessAddress addr, u64 size, DebugWatchpointType type) {
diff --git a/src/core/hle/kernel/k_process.h b/src/core/hle/kernel/k_process.h
index 8339465fd..d8cd0fdde 100644
--- a/src/core/hle/kernel/k_process.h
+++ b/src/core/hle/kernel/k_process.h
@@ -120,6 +120,9 @@ private:
     std::atomic<s64> m_num_ipc_messages{};
     std::atomic<s64> m_num_ipc_replies{};
     std::atomic<s64> m_num_ipc_receives{};
+#ifdef HAS_NCE
+    std::unordered_map<u64, u64> m_post_handlers{};
+#endif
 
 private:
     Result StartTermination();
@@ -150,7 +153,8 @@ public:
                       std::span<const u32> caps, KResourceLimit* res_limit,
                       KMemoryManager::Pool pool, bool immortal);
     Result Initialize(const Svc::CreateProcessParameter& params, std::span<const u32> user_caps,
-                      KResourceLimit* res_limit, KMemoryManager::Pool pool);
+                      KResourceLimit* res_limit, KMemoryManager::Pool pool,
+                      KProcessAddress aslr_space_start);
     void Exit();
 
     const char* GetName() const {
@@ -466,6 +470,12 @@ public:
 
     static void Switch(KProcess* cur_process, KProcess* next_process);
 
+#ifdef HAS_NCE
+    std::unordered_map<u64, u64>& GetPostHandlers() noexcept {
+        return m_post_handlers;
+    }
+#endif
+
 public:
     // Attempts to insert a watchpoint into a free slot. Returns false if none are available.
     bool InsertWatchpoint(KProcessAddress addr, u64 size, DebugWatchpointType type);
@@ -479,7 +489,7 @@ public:
 
 public:
     Result LoadFromMetadata(const FileSys::ProgramMetadata& metadata, std::size_t code_size,
-                            bool is_hbl);
+                            KProcessAddress aslr_space_start, bool is_hbl);
 
     void LoadModule(CodeSet code_set, KProcessAddress base_addr);
 
diff --git a/src/core/hle/kernel/k_process_page_table.h b/src/core/hle/kernel/k_process_page_table.h
index b7ae5abd0..9e40f68bc 100644
--- a/src/core/hle/kernel/k_process_page_table.h
+++ b/src/core/hle/kernel/k_process_page_table.h
@@ -23,10 +23,11 @@ public:
     Result Initialize(Svc::CreateProcessFlag as_type, bool enable_aslr, bool enable_das_merge,
                       bool from_back, KMemoryManager::Pool pool, KProcessAddress code_address,
                       size_t code_size, KSystemResource* system_resource,
-                      KResourceLimit* resource_limit, Core::Memory::Memory& memory) {
-        R_RETURN(m_page_table.InitializeForProcess(as_type, enable_aslr, enable_das_merge,
-                                                   from_back, pool, code_address, code_size,
-                                                   system_resource, resource_limit, memory));
+                      KResourceLimit* resource_limit, Core::Memory::Memory& memory,
+                      KProcessAddress aslr_space_start) {
+        R_RETURN(m_page_table.InitializeForProcess(
+            as_type, enable_aslr, enable_das_merge, from_back, pool, code_address, code_size,
+            system_resource, resource_limit, memory, aslr_space_start));
     }
 
     void Finalize() {
diff --git a/src/core/hle/kernel/k_thread.h b/src/core/hle/kernel/k_thread.h
index e1f80b04f..e9ca5dfca 100644
--- a/src/core/hle/kernel/k_thread.h
+++ b/src/core/hle/kernel/k_thread.h
@@ -655,6 +655,21 @@ public:
         return m_stack_top;
     }
 
+public:
+    // TODO: This shouldn't be defined in kernel namespace
+    struct NativeExecutionParameters {
+        u64 tpidr_el0{};
+        u64 tpidrro_el0{};
+        void* native_context{};
+        std::atomic<u32> lock{1};
+        bool is_running{};
+        u32 magic{Common::MakeMagic('Y', 'U', 'Z', 'U')};
+    };
+
+    NativeExecutionParameters& GetNativeExecutionParameters() {
+        return m_native_execution_parameters;
+    }
+
 private:
     KThread* RemoveWaiterByKey(bool* out_has_waiters, KProcessAddress key,
                                bool is_kernel_address_key);
@@ -914,6 +929,7 @@ private:
     ThreadWaitReasonForDebugging m_wait_reason_for_debugging{};
     uintptr_t m_argument{};
     KProcessAddress m_stack_top{};
+    NativeExecutionParameters m_native_execution_parameters{};
 
 public:
     using ConditionVariableThreadTreeType = ConditionVariableThreadTree;
diff --git a/src/core/hle/kernel/physical_core.cpp b/src/core/hle/kernel/physical_core.cpp
index 5ee869fa2..073039825 100644
--- a/src/core/hle/kernel/physical_core.cpp
+++ b/src/core/hle/kernel/physical_core.cpp
@@ -1,8 +1,12 @@
 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
+#include "common/settings.h"
 #include "core/arm/dynarmic/arm_dynarmic_32.h"
 #include "core/arm/dynarmic/arm_dynarmic_64.h"
+#ifdef HAS_NCE
+#include "core/arm/nce/arm_nce.h"
+#endif
 #include "core/core.h"
 #include "core/hle/kernel/k_scheduler.h"
 #include "core/hle/kernel/kernel.h"
@@ -14,7 +18,8 @@ PhysicalCore::PhysicalCore(std::size_t core_index, Core::System& system, KSchedu
     : m_core_index{core_index}, m_system{system}, m_scheduler{scheduler} {
 #if defined(ARCHITECTURE_x86_64) || defined(ARCHITECTURE_arm64)
     // TODO(bunnei): Initialization relies on a core being available. We may later replace this with
-    // a 32-bit instance of Dynarmic. This should be abstracted out to a CPU manager.
+    // an NCE interface or a 32-bit instance of Dynarmic. This should be abstracted out to a CPU
+    // manager.
     auto& kernel = system.Kernel();
     m_arm_interface = std::make_unique<Core::ARM_Dynarmic_64>(
         system, kernel.IsMulticore(),
@@ -28,6 +33,13 @@ PhysicalCore::PhysicalCore(std::size_t core_index, Core::System& system, KSchedu
 PhysicalCore::~PhysicalCore() = default;
 
 void PhysicalCore::Initialize(bool is_64_bit) {
+#if defined(HAS_NCE)
+    if (Settings::IsNceEnabled()) {
+        m_arm_interface = std::make_unique<Core::ARM_NCE>(m_system, m_system.Kernel().IsMulticore(),
+                                                          m_core_index);
+        return;
+    }
+#endif
 #if defined(ARCHITECTURE_x86_64) || defined(ARCHITECTURE_arm64)
     auto& kernel = m_system.Kernel();
     if (!is_64_bit) {